2019-05-19 13:08:55 +01:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-12-28 00:34:29 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2008-03-04 14:28:32 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2005-09-13 01:25:16 -07:00
|
|
|
|
2015-02-13 14:39:28 -08:00
|
|
|
|
2022-09-15 17:03:48 +02:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2011-05-24 17:12:16 -07:00
|
|
|
|
2007-10-16 23:25:53 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-02-27 17:42:37 +00:00
|
|
|
|
2005-10-29 18:16:53 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2011-05-24 17:11:33 -07:00
|
|
|
|
2006-12-08 02:39:45 -08:00
|
|
|
|
2010-05-24 14:32:30 -07:00
|
|
|
|
2009-09-21 17:02:44 -07:00
|
|
|
|
2017-02-22 15:42:00 -08:00
|
|
|
|
2011-05-20 12:50:29 -07:00
|
|
|
|
2013-09-11 14:22:36 -07:00
|
|
|
|
2020-12-14 19:08:30 -08:00
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
2017-02-02 20:43:54 +01:00
|
|
|
|
mm/page_owner: keep track of page owners
This is the page owner tracking code which is introduced so far ago. It
is resident on Andrew's tree, though, nobody tried to upstream so it
remain as is. Our company uses this feature actively to debug memory leak
or to find a memory hogger so I decide to upstream this feature.
This functionality help us to know who allocates the page. When
allocating a page, we store some information about allocation in extra
memory. Later, if we need to know status of all pages, we can get and
analyze it from this stored information.
In previous version of this feature, extra memory is statically defined in
struct page, but, in this version, extra memory is allocated outside of
struct page. It enables us to turn on/off this feature at boottime
without considerable memory waste.
Although we already have tracepoint for tracing page allocation/free,
using it to analyze page owner is rather complex. We need to enlarge the
trace buffer for preventing overlapping until userspace program launched.
And, launched program continually dump out the trace buffer for later
analysis and it would change system behaviour with more possibility rather
than just keeping it in memory, so bad for debug.
Moreover, we can use page_owner feature further for various purposes. For
example, we can use it for fragmentation statistics implemented in this
patch. And, I also plan to implement some CMA failure debugging feature
using this interface.
I'd like to give the credit for all developers contributed this feature,
but, it's not easy because I don't know exact history. Sorry about that.
Below is people who has "Signed-off-by" in the patches in Andrew's tree.
Contributor:
Alexander Nyberg <alexn@dsv.su.se>
Mel Gorman <mgorman@suse.de>
Dave Hansen <dave@linux.vnet.ibm.com>
Minchan Kim <minchan@kernel.org>
Michal Nazarewicz <mina86@mina86.com>
Andrew Morton <akpm@linux-foundation.org>
Jungsoo Son <jungsoo.son@lge.com>
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Dave Hansen <dave@sr71.net>
Cc: Michal Nazarewicz <mina86@mina86.com>
Cc: Jungsoo Son <jungsoo.son@lge.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-12-12 16:56:01 -08:00
|
|
|
|
2022-01-14 14:06:37 -08:00
|
|
|
|
mm: charge/uncharge kmemcg from generic page allocator paths
Currently, to charge a non-slab allocation to kmemcg one has to use
alloc_kmem_pages helper with __GFP_ACCOUNT flag. A page allocated with
this helper should finally be freed using free_kmem_pages, otherwise it
won't be uncharged.
This API suits its current users fine, but it turns out to be impossible
to use along with page reference counting, i.e. when an allocation is
supposed to be freed with put_page, as it is the case with pipe or unix
socket buffers.
To overcome this limitation, this patch moves charging/uncharging to
generic page allocator paths, i.e. to __alloc_pages_nodemask and
free_pages_prepare, and zaps alloc/free_kmem_pages helpers. This way,
one can use any of the available page allocation functions to get the
allocated page charged to kmemcg - it's enough to pass __GFP_ACCOUNT,
just like in case of kmalloc and friends. A charged page will be
automatically uncharged on free.
To make it possible, we need to mark pages charged to kmemcg somehow.
To avoid introducing a new page flag, we make use of page->_mapcount for
marking such pages. Since pages charged to kmemcg are not supposed to
be mapped to userspace, it should work just fine. There are other
(ab)users of page->_mapcount - buddy and balloon pages - but we don't
conflict with them.
In case kmemcg is compiled out or not used at runtime, this patch
introduces no overhead to generic page allocator paths. If kmemcg is
used, it will be plus one gfp flags check on alloc and plus one
page->_mapcount check on free, which shouldn't hurt performance, because
the data accessed are hot.
Link: http://lkml.kernel.org/r/a9736d856f895bcb465d9f257b54efe32eda6f99.1464079538.git.vdavydov@virtuozzo.com
Signed-off-by: Vladimir Davydov <vdavydov@virtuozzo.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-07-26 15:24:24 -07:00
|
|
|
|
2017-03-03 16:15:39 -05:00
|
|
|
|
2017-03-03 10:13:38 +01:00
|
|
|
|
psi: pressure stall information for CPU, memory, and IO
When systems are overcommitted and resources become contended, it's hard
to tell exactly the impact this has on workload productivity, or how close
the system is to lockups and OOM kills. In particular, when machines work
multiple jobs concurrently, the impact of overcommit in terms of latency
and throughput on the individual job can be enormous.
In order to maximize hardware utilization without sacrificing individual
job health or risk complete machine lockups, this patch implements a way
to quantify resource pressure in the system.
A kernel built with CONFIG_PSI=y creates files in /proc/pressure/ that
expose the percentage of time the system is stalled on CPU, memory, or IO,
respectively. Stall states are aggregate versions of the per-task delay
accounting delays:
cpu: some tasks are runnable but not executing on a CPU
memory: tasks are reclaiming, or waiting for swapin or thrashing cache
io: tasks are waiting for io completions
These percentages of walltime can be thought of as pressure percentages,
and they give a general sense of system health and productivity loss
incurred by resource overcommit. They can also indicate when the system
is approaching lockup scenarios and OOMs.
To do this, psi keeps track of the task states associated with each CPU
and samples the time they spend in stall states. Every 2 seconds, the
samples are averaged across CPUs - weighted by the CPUs' non-idle time to
eliminate artifacts from unused CPUs - and translated into percentages of
walltime. A running average of those percentages is maintained over 10s,
1m, and 5m periods (similar to the loadaverage).
[hannes@cmpxchg.org: doc fixlet, per Randy]
Link: http://lkml.kernel.org/r/20180828205625.GA14030@cmpxchg.org
[hannes@cmpxchg.org: code optimization]
Link: http://lkml.kernel.org/r/20180907175015.GA8479@cmpxchg.org
[hannes@cmpxchg.org: rename psi_clock() to psi_update_work(), per Peter]
Link: http://lkml.kernel.org/r/20180907145404.GB11088@cmpxchg.org
[hannes@cmpxchg.org: fix build]
Link: http://lkml.kernel.org/r/20180913014222.GA2370@cmpxchg.org
Link: http://lkml.kernel.org/r/20180828172258.3185-9-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Daniel Drake <drake@endlessm.com>
Tested-by: Suren Baghdasaryan <surenb@google.com>
Cc: Christopher Lameter <cl@linux.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Johannes Weiner <jweiner@fb.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Enderborg <peter.enderborg@sony.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vinayak Menon <vinmenon@codeaurora.org>
Cc: Randy Dunlap <rdunlap@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-26 15:06:27 -07:00
|
|
|
|
2020-10-10 23:16:40 -07:00
|
|
|
|
2022-01-19 18:10:15 -08:00
|
|
|
|
mm, pcp: reduce lock contention for draining high-order pages
In commit f26b3fa04611 ("mm/page_alloc: limit number of high-order pages
on PCP during bulk free"), the PCP (Per-CPU Pageset) will be drained when
PCP is mostly used for high-order pages freeing to improve the cache-hot
pages reusing between page allocating and freeing CPUs.
On system with small per-CPU data cache slice, pages shouldn't be cached
before draining to guarantee cache-hot. But on a system with large
per-CPU data cache slice, some pages can be cached before draining to
reduce zone lock contention.
So, in this patch, instead of draining without any caching, "pcp->batch"
pages will be cached in PCP before draining if the size of the per-CPU
data cache slice is more than "3 * batch".
In theory, if the size of per-CPU data cache slice is more than "2 *
batch", we can reuse cache-hot pages between CPUs. But considering the
other usage of cache (code, other data accessing, etc.), "3 * batch" is
used.
Note: "3 * batch" is chosen to make sure the optimization works on recent
x86_64 server CPUs. If you want to increase it, please check whether it
breaks the optimization.
On a 2-socket Intel server with 128 logical CPU, with the patch, the
network bandwidth of the UNIX (AF_UNIX) test case of lmbench test suite
with 16-pair processes increase 70.5%. The cycles% of the spinlock
contention (mostly for zone lock) decreases from 46.1% to 21.3%. The
number of PCP draining for high order pages freeing (free_high) decreases
89.9%. The cache miss rate keeps 0.2%.
Link: https://lkml.kernel.org/r/20231016053002.756205-4-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:29:56 +08:00
|
|
|
|
2024-03-21 09:36:36 -07:00
|
|
|
|
2006-05-15 09:43:59 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
mm: shuffle initial free memory to improve memory-side-cache utilization
Patch series "mm: Randomize free memory", v10.
This patch (of 3):
Randomization of the page allocator improves the average utilization of
a direct-mapped memory-side-cache. Memory side caching is a platform
capability that Linux has been previously exposed to in HPC
(high-performance computing) environments on specialty platforms. In
that instance it was a smaller pool of high-bandwidth-memory relative to
higher-capacity / lower-bandwidth DRAM. Now, this capability is going
to be found on general purpose server platforms where DRAM is a cache in
front of higher latency persistent memory [1].
Robert offered an explanation of the state of the art of Linux
interactions with memory-side-caches [2], and I copy it here:
It's been a problem in the HPC space:
http://www.nersc.gov/research-and-development/knl-cache-mode-performance-coe/
A kernel module called zonesort is available to try to help:
https://software.intel.com/en-us/articles/xeon-phi-software
and this abandoned patch series proposed that for the kernel:
https://lkml.kernel.org/r/20170823100205.17311-1-lukasz.daniluk@intel.com
Dan's patch series doesn't attempt to ensure buffers won't conflict, but
also reduces the chance that the buffers will. This will make performance
more consistent, albeit slower than "optimal" (which is near impossible
to attain in a general-purpose kernel). That's better than forcing
users to deploy remedies like:
"To eliminate this gradual degradation, we have added a Stream
measurement to the Node Health Check that follows each job;
nodes are rebooted whenever their measured memory bandwidth
falls below 300 GB/s."
A replacement for zonesort was merged upstream in commit cc9aec03e58f
("x86/numa_emulation: Introduce uniform split capability"). With this
numa_emulation capability, memory can be split into cache sized
("near-memory" sized) numa nodes. A bind operation to such a node, and
disabling workloads on other nodes, enables full cache performance.
However, once the workload exceeds the cache size then cache conflicts
are unavoidable. While HPC environments might be able to tolerate
time-scheduling of cache sized workloads, for general purpose server
platforms, the oversubscribed cache case will be the common case.
The worst case scenario is that a server system owner benchmarks a
workload at boot with an un-contended cache only to see that performance
degrade over time, even below the average cache performance due to
excessive conflicts. Randomization clips the peaks and fills in the
valleys of cache utilization to yield steady average performance.
Here are some performance impact details of the patches:
1/ An Intel internal synthetic memory bandwidth measurement tool, saw a
3X speedup in a contrived case that tries to force cache conflicts.
The contrived cased used the numa_emulation capability to force an
instance of the benchmark to be run in two of the near-memory sized
numa nodes. If both instances were placed on the same emulated they
would fit and cause zero conflicts. While on separate emulated nodes
without randomization they underutilized the cache and conflicted
unnecessarily due to the in-order allocation per node.
2/ A well known Java server application benchmark was run with a heap
size that exceeded cache size by 3X. The cache conflict rate was 8%
for the first run and degraded to 21% after page allocator aging. With
randomization enabled the rate levelled out at 11%.
3/ A MongoDB workload did not observe measurable difference in
cache-conflict rates, but the overall throughput dropped by 7% with
randomization in one case.
4/ Mel Gorman ran his suite of performance workloads with randomization
enabled on platforms without a memory-side-cache and saw a mix of some
improvements and some losses [3].
While there is potentially significant improvement for applications that
depend on low latency access across a wide working-set, the performance
may be negligible to negative for other workloads. For this reason the
shuffle capability defaults to off unless a direct-mapped
memory-side-cache is detected. Even then, the page_alloc.shuffle=0
parameter can be specified to disable the randomization on those systems.
Outside of memory-side-cache utilization concerns there is potentially
security benefit from randomization. Some data exfiltration and
return-oriented-programming attacks rely on the ability to infer the
location of sensitive data objects. The kernel page allocator, especially
early in system boot, has predictable first-in-first out behavior for
physical pages. Pages are freed in physical address order when first
onlined.
Quoting Kees:
"While we already have a base-address randomization
(CONFIG_RANDOMIZE_MEMORY), attacks against the same hardware and
memory layouts would certainly be using the predictability of
allocation ordering (i.e. for attacks where the base address isn't
important: only the relative positions between allocated memory).
This is common in lots of heap-style attacks. They try to gain
control over ordering by spraying allocations, etc.
I'd really like to see this because it gives us something similar
to CONFIG_SLAB_FREELIST_RANDOM but for the page allocator."
While SLAB_FREELIST_RANDOM reduces the predictability of some local slab
caches it leaves vast bulk of memory to be predictably in order allocated.
However, it should be noted, the concrete security benefits are hard to
quantify, and no known CVE is mitigated by this randomization.
Introduce shuffle_free_memory(), and its helper shuffle_zone(), to perform
a Fisher-Yates shuffle of the page allocator 'free_area' lists when they
are initially populated with free memory at boot and at hotplug time. Do
this based on either the presence of a page_alloc.shuffle=Y command line
parameter, or autodetection of a memory-side-cache (to be added in a
follow-on patch).
The shuffling is done in terms of CONFIG_SHUFFLE_PAGE_ORDER sized free
pages where the default CONFIG_SHUFFLE_PAGE_ORDER is MAX_ORDER-1 i.e. 10,
4MB this trades off randomization granularity for time spent shuffling.
MAX_ORDER-1 was chosen to be minimally invasive to the page allocator
while still showing memory-side cache behavior improvements, and the
expectation that the security implications of finer granularity
randomization is mitigated by CONFIG_SLAB_FREELIST_RANDOM. The
performance impact of the shuffling appears to be in the noise compared to
other memory initialization work.
This initial randomization can be undone over time so a follow-on patch is
introduced to inject entropy on page free decisions. It is reasonable to
ask if the page free entropy is sufficient, but it is not enough due to
the in-order initial freeing of pages. At the start of that process
putting page1 in front or behind page0 still keeps them close together,
page2 is still near page1 and has a high chance of being adjacent. As
more pages are added ordering diversity improves, but there is still high
page locality for the low address pages and this leads to no significant
impact to the cache conflict rate.
[1]: https://itpeernetwork.intel.com/intel-optane-dc-persistent-memory-operating-modes/
[2]: https://lkml.kernel.org/r/AT5PR8401MB1169D656C8B5E121752FC0F8AB120@AT5PR8401MB1169.NAMPRD84.PROD.OUTLOOK.COM
[3]: https://lkml.org/lkml/2018/10/12/309
[dan.j.williams@intel.com: fix shuffle enable]
Link: http://lkml.kernel.org/r/154943713038.3858443.4125180191382062871.stgit@dwillia2-desk3.amr.corp.intel.com
[cai@lca.pw: fix SHUFFLE_PAGE_ALLOCATOR help texts]
Link: http://lkml.kernel.org/r/20190425201300.75650-1-cai@lca.pw
Link: http://lkml.kernel.org/r/154899811738.3165233.12325692939590944259.stgit@dwillia2-desk3.amr.corp.intel.com
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Qian Cai <cai@lca.pw>
Reviewed-by: Kees Cook <keescook@chromium.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Keith Busch <keith.busch@intel.com>
Cc: Robert Elliott <elliott@hpe.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 15:41:28 -07:00
|
|
|
|
2020-04-06 20:04:56 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
mm/page_alloc: convert "report" flag of __free_one_page() to a proper flag
Patch series "mm: place pages to the freelist tail when onlining and undoing isolation", v2.
When adding separate memory blocks via add_memory*() and onlining them
immediately, the metadata (especially the memmap) of the next block will
be placed onto one of the just added+onlined block. This creates a chain
of unmovable allocations: If the last memory block cannot get
offlined+removed() so will all dependent ones. We directly have unmovable
allocations all over the place.
This can be observed quite easily using virtio-mem, however, it can also
be observed when using DIMMs. The freshly onlined pages will usually be
placed to the head of the freelists, meaning they will be allocated next,
turning the just-added memory usually immediately un-removable. The fresh
pages are cold, prefering to allocate others (that might be hot) also
feels to be the natural thing to do.
It also applies to the hyper-v balloon xen-balloon, and ppc64 dlpar: when
adding separate, successive memory blocks, each memory block will have
unmovable allocations on them - for example gigantic pages will fail to
allocate.
While the ZONE_NORMAL doesn't provide any guarantees that memory can get
offlined+removed again (any kind of fragmentation with unmovable
allocations is possible), there are many scenarios (hotplugging a lot of
memory, running workload, hotunplug some memory/as much as possible) where
we can offline+remove quite a lot with this patchset.
a) To visualize the problem, a very simple example:
Start a VM with 4GB and 8GB of virtio-mem memory:
[root@localhost ~]# lsmem
RANGE SIZE STATE REMOVABLE BLOCK
0x0000000000000000-0x00000000bfffffff 3G online yes 0-23
0x0000000100000000-0x000000033fffffff 9G online yes 32-103
Memory block size: 128M
Total online memory: 12G
Total offline memory: 0B
Then try to unplug as much as possible using virtio-mem. Observe which
memory blocks are still around. Without this patch set:
[root@localhost ~]# lsmem
RANGE SIZE STATE REMOVABLE BLOCK
0x0000000000000000-0x00000000bfffffff 3G online yes 0-23
0x0000000100000000-0x000000013fffffff 1G online yes 32-39
0x0000000148000000-0x000000014fffffff 128M online yes 41
0x0000000158000000-0x000000015fffffff 128M online yes 43
0x0000000168000000-0x000000016fffffff 128M online yes 45
0x0000000178000000-0x000000017fffffff 128M online yes 47
0x0000000188000000-0x0000000197ffffff 256M online yes 49-50
0x00000001a0000000-0x00000001a7ffffff 128M online yes 52
0x00000001b0000000-0x00000001b7ffffff 128M online yes 54
0x00000001c0000000-0x00000001c7ffffff 128M online yes 56
0x00000001d0000000-0x00000001d7ffffff 128M online yes 58
0x00000001e0000000-0x00000001e7ffffff 128M online yes 60
0x00000001f0000000-0x00000001f7ffffff 128M online yes 62
0x0000000200000000-0x0000000207ffffff 128M online yes 64
0x0000000210000000-0x0000000217ffffff 128M online yes 66
0x0000000220000000-0x0000000227ffffff 128M online yes 68
0x0000000230000000-0x0000000237ffffff 128M online yes 70
0x0000000240000000-0x0000000247ffffff 128M online yes 72
0x0000000250000000-0x0000000257ffffff 128M online yes 74
0x0000000260000000-0x0000000267ffffff 128M online yes 76
0x0000000270000000-0x0000000277ffffff 128M online yes 78
0x0000000280000000-0x0000000287ffffff 128M online yes 80
0x0000000290000000-0x0000000297ffffff 128M online yes 82
0x00000002a0000000-0x00000002a7ffffff 128M online yes 84
0x00000002b0000000-0x00000002b7ffffff 128M online yes 86
0x00000002c0000000-0x00000002c7ffffff 128M online yes 88
0x00000002d0000000-0x00000002d7ffffff 128M online yes 90
0x00000002e0000000-0x00000002e7ffffff 128M online yes 92
0x00000002f0000000-0x00000002f7ffffff 128M online yes 94
0x0000000300000000-0x0000000307ffffff 128M online yes 96
0x0000000310000000-0x0000000317ffffff 128M online yes 98
0x0000000320000000-0x0000000327ffffff 128M online yes 100
0x0000000330000000-0x000000033fffffff 256M online yes 102-103
Memory block size: 128M
Total online memory: 8.1G
Total offline memory: 0B
With this patch set:
[root@localhost ~]# lsmem
RANGE SIZE STATE REMOVABLE BLOCK
0x0000000000000000-0x00000000bfffffff 3G online yes 0-23
0x0000000100000000-0x000000013fffffff 1G online yes 32-39
Memory block size: 128M
Total online memory: 4G
Total offline memory: 0B
All memory can get unplugged, all memory block can get removed. Of
course, no workload ran and the system was basically idle, but it
highlights the issue - the fairly deterministic chain of unmovable
allocations. When a huge page for the 2MB memmap is needed, a
just-onlined 4MB page will be split. The remaining 2MB page will be used
for the memmap of the next memory block. So one memory block will hold
the memmap of the two following memory blocks. Finally the pages of the
last-onlined memory block will get used for the next bigger allocations -
if any allocation is unmovable, all dependent memory blocks cannot get
unplugged and removed until that allocation is gone.
Note that with bigger memory blocks (e.g., 256MB), *all* memory
blocks are dependent and none can get unplugged again!
b) Experiment with memory intensive workload
I performed an experiment with an older version of this patch set (before
we used undo_isolate_page_range() in online_pages(): Hotplug 56GB to a VM
with an initial 4GB, onlining all memory to ZONE_NORMAL right from the
kernel when adding it. I then run various memory intensive workloads that
consume most system memory for a total of 45 minutes. Once finished, I
try to unplug as much memory as possible.
With this change, I am able to remove via virtio-mem (adding individual
128MB memory blocks) 413 out of 448 added memory blocks. Via individual
(256MB) DIMMs 380 out of 448 added memory blocks. (I don't have any
numbers without this patchset, but looking at the above example, it's at
most half of the 448 memory blocks for virtio-mem, and most probably none
for DIMMs).
Again, there are workloads that might behave very differently due to the
nature of ZONE_NORMAL.
This change also affects (besides memory onlining):
- Other users of undo_isolate_page_range(): Pages are always placed to the
tail.
-- When memory offlining fails
-- When memory isolation fails after having isolated some pageblocks
-- When alloc_contig_range() either succeeds or fails
- Other users of __putback_isolated_page(): Pages are always placed to the
tail.
-- Free page reporting
- Other users of __free_pages_core()
-- AFAIKs, any memory that is getting exposed to the buddy during boot.
IIUC we will now usually allocate memory from lower addresses within
a zone first (especially during boot).
- Other users of generic_online_page()
-- Hyper-V balloon
This patch (of 5):
Let's prepare for additional flags and avoid long parameter lists of
bools. Follow-up patches will also make use of the flags in
__free_pages_ok().
Signed-off-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Alexander Duyck <alexander.h.duyck@linux.intel.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Wei Yang <richard.weiyang@linux.alibaba.com>
Reviewed-by: Pankaj Gupta <pankaj.gupta.linux@gmail.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Haiyang Zhang <haiyangz@microsoft.com>
Cc: "K. Y. Srinivasan" <kys@microsoft.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Scott Cheloha <cheloha@linux.ibm.com>
Cc: Stephen Hemminger <sthemmin@microsoft.com>
Cc: Wei Liu <wei.liu@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Link: https://lkml.kernel.org/r/20201005121534.15649-1-david@redhat.com
Link: https://lkml.kernel.org/r/20201005121534.15649-2-david@redhat.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-10-15 20:09:20 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/page_alloc: place pages to tail in __putback_isolated_page()
__putback_isolated_page() already documents that pages will be placed to
the tail of the freelist - this is, however, not the case for "order >=
MAX_ORDER - 2" (see buddy_merge_likely()) - which should be the case for
all existing users.
This change affects two users:
- free page reporting
- page isolation, when undoing the isolation (including memory onlining).
This behavior is desirable for pages that haven't really been touched
lately, so exactly the two users that don't actually read/write page
content, but rather move untouched pages.
The new behavior is especially desirable for memory onlining, where we
allow allocation of newly onlined pages via undo_isolate_page_range() in
online_pages(). Right now, we always place them to the head of the
freelist, resulting in undesireable behavior: Assume we add individual
memory chunks via add_memory() and online them right away to the NORMAL
zone. We create a dependency chain of unmovable allocations e.g., via the
memmap. The memmap of the next chunk will be placed onto previous chunks
- if the last block cannot get offlined+removed, all dependent ones cannot
get offlined+removed. While this can already be observed with individual
DIMMs, it's more of an issue for virtio-mem (and I suspect also ppc
DLPAR).
Document that this should only be used for optimizations, and no code
should rely on this behavior for correction (if the order of the freelists
ever changes).
We won't care about page shuffling: memory onlining already properly
shuffles after onlining. free page reporting doesn't care about
physically contiguous ranges, and there are already cases where page
isolation will simply move (physically close) free pages to (currently)
the head of the freelists via move_freepages_block() instead of shuffling.
If this becomes ever relevant, we should shuffle the whole zone when
undoing isolation of larger ranges, and after free_contig_range().
Signed-off-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Alexander Duyck <alexander.h.duyck@linux.intel.com>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Wei Yang <richard.weiyang@linux.alibaba.com>
Reviewed-by: Pankaj Gupta <pankaj.gupta.linux@gmail.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Scott Cheloha <cheloha@linux.ibm.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Haiyang Zhang <haiyangz@microsoft.com>
Cc: "K. Y. Srinivasan" <kys@microsoft.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Stephen Hemminger <sthemmin@microsoft.com>
Cc: Wei Liu <wei.liu@kernel.org>
Link: https://lkml.kernel.org/r/20201005121534.15649-3-david@redhat.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-10-15 20:09:26 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-02-21 18:44:24 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-07-03 15:01:29 -07:00
|
|
|
|
|
|
|
|
|
2021-06-28 19:42:24 -07:00
|
|
|
|
2013-07-03 15:01:29 -07:00
|
|
|
|
2022-06-24 13:54:21 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-07-03 15:01:29 -07:00
|
|
|
|
2022-06-24 13:54:21 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-06-24 13:54:23 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-11-18 10:17:14 +00:00
|
|
|
|
2022-06-24 13:54:23 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-11-18 10:17:14 +00:00
|
|
|
|
2022-06-24 13:54:23 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-11-18 10:17:14 +00:00
|
|
|
|
|
|
|
|
|
2022-06-24 13:54:23 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-05-26 14:44:56 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-11-15 17:38:22 -08:00
|
|
|
|
|
|
|
|
|
2010-05-26 14:45:00 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-04-10 14:32:32 -07:00
|
|
|
|
2017-02-24 14:56:56 -08:00
|
|
|
|
gcc-plugins: Add latent_entropy plugin
This adds a new gcc plugin named "latent_entropy". It is designed to
extract as much possible uncertainty from a running system at boot time as
possible, hoping to capitalize on any possible variation in CPU operation
(due to runtime data differences, hardware differences, SMP ordering,
thermal timing variation, cache behavior, etc).
At the very least, this plugin is a much more comprehensive example for
how to manipulate kernel code using the gcc plugin internals.
The need for very-early boot entropy tends to be very architecture or
system design specific, so this plugin is more suited for those sorts
of special cases. The existing kernel RNG already attempts to extract
entropy from reliable runtime variation, but this plugin takes the idea to
a logical extreme by permuting a global variable based on any variation
in code execution (e.g. a different value (and permutation function)
is used to permute the global based on loop count, case statement,
if/then/else branching, etc).
To do this, the plugin starts by inserting a local variable in every
marked function. The plugin then adds logic so that the value of this
variable is modified by randomly chosen operations (add, xor and rol) and
random values (gcc generates separate static values for each location at
compile time and also injects the stack pointer at runtime). The resulting
value depends on the control flow path (e.g., loops and branches taken).
Before the function returns, the plugin mixes this local variable into
the latent_entropy global variable. The value of this global variable
is added to the kernel entropy pool in do_one_initcall() and _do_fork(),
though it does not credit any bytes of entropy to the pool; the contents
of the global are just used to mix the pool.
Additionally, the plugin can pre-initialize arrays with build-time
random contents, so that two different kernel builds running on identical
hardware will not have the same starting values.
Signed-off-by: Emese Revfy <re.emese@gmail.com>
[kees: expanded commit message and code comments]
Signed-off-by: Kees Cook <keescook@chromium.org>
2016-06-20 20:41:19 +02:00
|
|
|
|
2016-10-19 00:08:04 +02:00
|
|
|
|
gcc-plugins: Add latent_entropy plugin
This adds a new gcc plugin named "latent_entropy". It is designed to
extract as much possible uncertainty from a running system at boot time as
possible, hoping to capitalize on any possible variation in CPU operation
(due to runtime data differences, hardware differences, SMP ordering,
thermal timing variation, cache behavior, etc).
At the very least, this plugin is a much more comprehensive example for
how to manipulate kernel code using the gcc plugin internals.
The need for very-early boot entropy tends to be very architecture or
system design specific, so this plugin is more suited for those sorts
of special cases. The existing kernel RNG already attempts to extract
entropy from reliable runtime variation, but this plugin takes the idea to
a logical extreme by permuting a global variable based on any variation
in code execution (e.g. a different value (and permutation function)
is used to permute the global based on loop count, case statement,
if/then/else branching, etc).
To do this, the plugin starts by inserting a local variable in every
marked function. The plugin then adds logic so that the value of this
variable is modified by randomly chosen operations (add, xor and rol) and
random values (gcc generates separate static values for each location at
compile time and also injects the stack pointer at runtime). The resulting
value depends on the control flow path (e.g., loops and branches taken).
Before the function returns, the plugin mixes this local variable into
the latent_entropy global variable. The value of this global variable
is added to the kernel entropy pool in do_one_initcall() and _do_fork(),
though it does not credit any bytes of entropy to the pool; the contents
of the global are just used to mix the pool.
Additionally, the plugin can pre-initialize arrays with build-time
random contents, so that two different kernel builds running on identical
hardware will not have the same starting values.
Signed-off-by: Emese Revfy <re.emese@gmail.com>
[kees: expanded commit message and code comments]
Signed-off-by: Kees Cook <keescook@chromium.org>
2016-06-20 20:41:19 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
Memoryless nodes: Generic management of nodemasks for various purposes
Why do we need to support memoryless nodes?
KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> wrote:
> For fujitsu, problem is called "empty" node.
>
> When ACPI's SRAT table includes "possible nodes", ia64 bootstrap(acpi_numa_init)
> creates nodes, which includes no memory, no cpu.
>
> I tried to remove empty-node in past, but that was denied.
> It was because we can hot-add cpu to the empty node.
> (node-hotplug triggered by cpu is not implemented now. and it will be ugly.)
>
>
> For HP, (Lee can comment on this later), they have memory-less-node.
> As far as I hear, HP's machine can have following configration.
>
> (example)
> Node0: CPU0 memory AAA MB
> Node1: CPU1 memory AAA MB
> Node2: CPU2 memory AAA MB
> Node3: CPU3 memory AAA MB
> Node4: Memory XXX GB
>
> AAA is very small value (below 16MB) and will be omitted by ia64 bootstrap.
> After boot, only Node 4 has valid memory (but have no cpu.)
>
> Maybe this is memory-interleave by firmware config.
Christoph Lameter <clameter@sgi.com> wrote:
> Future SGI platforms (actually also current one can have but nothing like
> that is deployed to my knowledge) have nodes with only cpus. Current SGI
> platforms have nodes with just I/O that we so far cannot manage in the
> core. So the arch code maps them to the nearest memory node.
Lee Schermerhorn <Lee.Schermerhorn@hp.com> wrote:
> For the HP platforms, we can configure each cell with from 0% to 100%
> "cell local memory". When we configure with <100% CLM, the "missing
> percentages" are interleaved by hardware on a cache-line granularity to
> improve bandwidth at the expense of latency for numa-challenged
> applications [and OSes, but not our problem ;-)]. When we boot Linux on
> such a config, all of the real nodes have no memory--it all resides in a
> single interleaved pseudo-node.
>
> When we boot Linux on a 100% CLM configuration [== NUMA], we still have
> the interleaved pseudo-node. It contains a few hundred MB stolen from
> the real nodes to contain the DMA zone. [Interleaved memory resides at
> phys addr 0]. The memoryless-nodes patches, along with the zoneorder
> patches, support this config as well.
>
> Also, when we boot a NUMA config with the "mem=" command line,
> specifying less memory than actually exists, Linux takes the excluded
> memory "off the top" rather than distributing it across the nodes. This
> can result in memoryless nodes, as well.
>
This patch:
Preparation for memoryless node patches.
Provide a generic way to keep nodemasks describing various characteristics of
NUMA nodes.
Remove the node_online_map and the node_possible map and realize the same
functionality using two nodes stats: N_POSSIBLE and N_ONLINE.
[Lee.Schermerhorn@hp.com: Initialize N_*_MEMORY and N_CPU masks for non-NUMA config]
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Tested-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Acked-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Acked-by: Bob Picco <bob.picco@hp.com>
Cc: Nishanth Aravamudan <nacc@us.ibm.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Mel Gorman <mel@skynet.ie>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: "Serge E. Hallyn" <serge@hallyn.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 01:25:27 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
Memoryless nodes: Generic management of nodemasks for various purposes
Why do we need to support memoryless nodes?
KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> wrote:
> For fujitsu, problem is called "empty" node.
>
> When ACPI's SRAT table includes "possible nodes", ia64 bootstrap(acpi_numa_init)
> creates nodes, which includes no memory, no cpu.
>
> I tried to remove empty-node in past, but that was denied.
> It was because we can hot-add cpu to the empty node.
> (node-hotplug triggered by cpu is not implemented now. and it will be ugly.)
>
>
> For HP, (Lee can comment on this later), they have memory-less-node.
> As far as I hear, HP's machine can have following configration.
>
> (example)
> Node0: CPU0 memory AAA MB
> Node1: CPU1 memory AAA MB
> Node2: CPU2 memory AAA MB
> Node3: CPU3 memory AAA MB
> Node4: Memory XXX GB
>
> AAA is very small value (below 16MB) and will be omitted by ia64 bootstrap.
> After boot, only Node 4 has valid memory (but have no cpu.)
>
> Maybe this is memory-interleave by firmware config.
Christoph Lameter <clameter@sgi.com> wrote:
> Future SGI platforms (actually also current one can have but nothing like
> that is deployed to my knowledge) have nodes with only cpus. Current SGI
> platforms have nodes with just I/O that we so far cannot manage in the
> core. So the arch code maps them to the nearest memory node.
Lee Schermerhorn <Lee.Schermerhorn@hp.com> wrote:
> For the HP platforms, we can configure each cell with from 0% to 100%
> "cell local memory". When we configure with <100% CLM, the "missing
> percentages" are interleaved by hardware on a cache-line granularity to
> improve bandwidth at the expense of latency for numa-challenged
> applications [and OSes, but not our problem ;-)]. When we boot Linux on
> such a config, all of the real nodes have no memory--it all resides in a
> single interleaved pseudo-node.
>
> When we boot Linux on a 100% CLM configuration [== NUMA], we still have
> the interleaved pseudo-node. It contains a few hundred MB stolen from
> the real nodes to contain the DMA zone. [Interleaved memory resides at
> phys addr 0]. The memoryless-nodes patches, along with the zoneorder
> patches, support this config as well.
>
> Also, when we boot a NUMA config with the "mem=" command line,
> specifying less memory than actually exists, Linux takes the excluded
> memory "off the top" rather than distributing it across the nodes. This
> can result in memoryless nodes, as well.
>
This patch:
Preparation for memoryless node patches.
Provide a generic way to keep nodemasks describing various characteristics of
NUMA nodes.
Remove the node_online_map and the node_possible map and realize the same
functionality using two nodes stats: N_POSSIBLE and N_ONLINE.
[Lee.Schermerhorn@hp.com: Initialize N_*_MEMORY and N_CPU masks for non-NUMA config]
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Tested-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Acked-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Acked-by: Bob Picco <bob.picco@hp.com>
Cc: Nishanth Aravamudan <nacc@us.ibm.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Mel Gorman <mel@skynet.ie>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: "Serge E. Hallyn" <serge@hallyn.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 01:25:27 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-12-12 13:52:00 -08:00
|
|
|
|
|
|
|
|
|
Memoryless nodes: Generic management of nodemasks for various purposes
Why do we need to support memoryless nodes?
KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> wrote:
> For fujitsu, problem is called "empty" node.
>
> When ACPI's SRAT table includes "possible nodes", ia64 bootstrap(acpi_numa_init)
> creates nodes, which includes no memory, no cpu.
>
> I tried to remove empty-node in past, but that was denied.
> It was because we can hot-add cpu to the empty node.
> (node-hotplug triggered by cpu is not implemented now. and it will be ugly.)
>
>
> For HP, (Lee can comment on this later), they have memory-less-node.
> As far as I hear, HP's machine can have following configration.
>
> (example)
> Node0: CPU0 memory AAA MB
> Node1: CPU1 memory AAA MB
> Node2: CPU2 memory AAA MB
> Node3: CPU3 memory AAA MB
> Node4: Memory XXX GB
>
> AAA is very small value (below 16MB) and will be omitted by ia64 bootstrap.
> After boot, only Node 4 has valid memory (but have no cpu.)
>
> Maybe this is memory-interleave by firmware config.
Christoph Lameter <clameter@sgi.com> wrote:
> Future SGI platforms (actually also current one can have but nothing like
> that is deployed to my knowledge) have nodes with only cpus. Current SGI
> platforms have nodes with just I/O that we so far cannot manage in the
> core. So the arch code maps them to the nearest memory node.
Lee Schermerhorn <Lee.Schermerhorn@hp.com> wrote:
> For the HP platforms, we can configure each cell with from 0% to 100%
> "cell local memory". When we configure with <100% CLM, the "missing
> percentages" are interleaved by hardware on a cache-line granularity to
> improve bandwidth at the expense of latency for numa-challenged
> applications [and OSes, but not our problem ;-)]. When we boot Linux on
> such a config, all of the real nodes have no memory--it all resides in a
> single interleaved pseudo-node.
>
> When we boot Linux on a 100% CLM configuration [== NUMA], we still have
> the interleaved pseudo-node. It contains a few hundred MB stolen from
> the real nodes to contain the DMA zone. [Interleaved memory resides at
> phys addr 0]. The memoryless-nodes patches, along with the zoneorder
> patches, support this config as well.
>
> Also, when we boot a NUMA config with the "mem=" command line,
> specifying less memory than actually exists, Linux takes the excluded
> memory "off the top" rather than distributing it across the nodes. This
> can result in memoryless nodes, as well.
>
This patch:
Preparation for memoryless node patches.
Provide a generic way to keep nodemasks describing various characteristics of
NUMA nodes.
Remove the node_online_map and the node_possible map and realize the same
functionality using two nodes stats: N_POSSIBLE and N_ONLINE.
[Lee.Schermerhorn@hp.com: Initialize N_*_MEMORY and N_CPU masks for non-NUMA config]
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Tested-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Acked-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Acked-by: Bob Picco <bob.picco@hp.com>
Cc: Nishanth Aravamudan <nacc@us.ibm.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Mel Gorman <mel@skynet.ie>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: "Serge E. Hallyn" <serge@hallyn.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 01:25:27 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2009-06-18 13:24:12 +10:00
|
|
|
|
mm: security: introduce init_on_alloc=1 and init_on_free=1 boot options
Patch series "add init_on_alloc/init_on_free boot options", v10.
Provide init_on_alloc and init_on_free boot options.
These are aimed at preventing possible information leaks and making the
control-flow bugs that depend on uninitialized values more deterministic.
Enabling either of the options guarantees that the memory returned by the
page allocator and SL[AU]B is initialized with zeroes. SLOB allocator
isn't supported at the moment, as its emulation of kmem caches complicates
handling of SLAB_TYPESAFE_BY_RCU caches correctly.
Enabling init_on_free also guarantees that pages and heap objects are
initialized right after they're freed, so it won't be possible to access
stale data by using a dangling pointer.
As suggested by Michal Hocko, right now we don't let the heap users to
disable initialization for certain allocations. There's not enough
evidence that doing so can speed up real-life cases, and introducing ways
to opt-out may result in things going out of control.
This patch (of 2):
The new options are needed to prevent possible information leaks and make
control-flow bugs that depend on uninitialized values more deterministic.
This is expected to be on-by-default on Android and Chrome OS. And it
gives the opportunity for anyone else to use it under distros too via the
boot args. (The init_on_free feature is regularly requested by folks
where memory forensics is included in their threat models.)
init_on_alloc=1 makes the kernel initialize newly allocated pages and heap
objects with zeroes. Initialization is done at allocation time at the
places where checks for __GFP_ZERO are performed.
init_on_free=1 makes the kernel initialize freed pages and heap objects
with zeroes upon their deletion. This helps to ensure sensitive data
doesn't leak via use-after-free accesses.
Both init_on_alloc=1 and init_on_free=1 guarantee that the allocator
returns zeroed memory. The two exceptions are slab caches with
constructors and SLAB_TYPESAFE_BY_RCU flag. Those are never
zero-initialized to preserve their semantics.
Both init_on_alloc and init_on_free default to zero, but those defaults
can be overridden with CONFIG_INIT_ON_ALLOC_DEFAULT_ON and
CONFIG_INIT_ON_FREE_DEFAULT_ON.
If either SLUB poisoning or page poisoning is enabled, those options take
precedence over init_on_alloc and init_on_free: initialization is only
applied to unpoisoned allocations.
Slowdown for the new features compared to init_on_free=0, init_on_alloc=0:
hackbench, init_on_free=1: +7.62% sys time (st.err 0.74%)
hackbench, init_on_alloc=1: +7.75% sys time (st.err 2.14%)
Linux build with -j12, init_on_free=1: +8.38% wall time (st.err 0.39%)
Linux build with -j12, init_on_free=1: +24.42% sys time (st.err 0.52%)
Linux build with -j12, init_on_alloc=1: -0.13% wall time (st.err 0.42%)
Linux build with -j12, init_on_alloc=1: +0.57% sys time (st.err 0.40%)
The slowdown for init_on_free=0, init_on_alloc=0 compared to the baseline
is within the standard error.
The new features are also going to pave the way for hardware memory
tagging (e.g. arm64's MTE), which will require both on_alloc and on_free
hooks to set the tags for heap objects. With MTE, tagging will have the
same cost as memory initialization.
Although init_on_free is rather costly, there are paranoid use-cases where
in-memory data lifetime is desired to be minimized. There are various
arguments for/against the realism of the associated threat models, but
given that we'll need the infrastructure for MTE anyway, and there are
people who want wipe-on-free behavior no matter what the performance cost,
it seems reasonable to include it in this series.
[glider@google.com: v8]
Link: http://lkml.kernel.org/r/20190626121943.131390-2-glider@google.com
[glider@google.com: v9]
Link: http://lkml.kernel.org/r/20190627130316.254309-2-glider@google.com
[glider@google.com: v10]
Link: http://lkml.kernel.org/r/20190628093131.199499-2-glider@google.com
Link: http://lkml.kernel.org/r/20190617151050.92663-2-glider@google.com
Signed-off-by: Alexander Potapenko <glider@google.com>
Acked-by: Kees Cook <keescook@chromium.org>
Acked-by: Michal Hocko <mhocko@suse.cz> [page and dmapool parts
Acked-by: James Morris <jamorris@linux.microsoft.com>]
Cc: Christoph Lameter <cl@linux.com>
Cc: Masahiro Yamada <yamada.masahiro@socionext.com>
Cc: "Serge E. Hallyn" <serge@hallyn.com>
Cc: Nick Desaulniers <ndesaulniers@google.com>
Cc: Kostya Serebryany <kcc@google.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Sandeep Patil <sspatil@android.com>
Cc: Laura Abbott <labbott@redhat.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Jann Horn <jannh@google.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Marco Elver <elver@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-11 20:59:19 -07:00
|
|
|
|
2007-10-16 01:26:01 -07:00
|
|
|
|
2015-11-06 16:29:57 -08:00
|
|
|
|
2007-10-16 01:26:01 -07:00
|
|
|
|
|
|
|
|
|
2020-10-15 20:09:35 -07:00
|
|
|
|
|
|
|
|
|
2006-01-06 00:11:08 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-02-12 15:00:22 -08:00
|
|
|
|
2005-11-05 17:25:53 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2023-05-18 10:11:19 -04:00
|
|
|
|
2007-02-10 01:43:10 -08:00
|
|
|
|
2018-04-10 16:30:11 -07:00
|
|
|
|
2007-02-10 01:43:10 -08:00
|
|
|
|
2006-09-25 23:31:13 -07:00
|
|
|
|
2018-04-10 16:30:11 -07:00
|
|
|
|
2006-09-25 23:31:13 -07:00
|
|
|
|
2018-04-10 16:30:11 -07:00
|
|
|
|
2006-09-25 23:31:14 -07:00
|
|
|
|
2018-04-10 16:30:11 -07:00
|
|
|
|
2006-09-25 23:31:14 -07:00
|
|
|
|
2018-04-10 16:30:11 -07:00
|
|
|
|
2006-09-25 23:31:13 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2023-03-21 19:05:02 +02:00
|
|
|
|
2007-02-10 01:43:10 -08:00
|
|
|
|
2006-09-25 23:31:13 -07:00
|
|
|
|
2007-02-10 01:43:10 -08:00
|
|
|
|
2006-09-25 23:31:13 -07:00
|
|
|
|
2006-09-25 23:31:13 -07:00
|
|
|
|
2006-09-25 23:31:13 -07:00
|
|
|
|
2006-09-25 23:31:13 -07:00
|
|
|
|
2006-09-25 23:31:14 -07:00
|
|
|
|
2007-07-17 04:03:12 -07:00
|
|
|
|
2006-09-25 23:31:14 -07:00
|
|
|
|
2007-07-17 04:03:12 -07:00
|
|
|
|
2015-08-09 15:29:06 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-09-25 23:31:13 -07:00
|
|
|
|
|
|
|
|
|
2018-12-28 00:35:55 -08:00
|
|
|
|
mm, page_owner: print migratetype of page and pageblock, symbolic flags
The information in /sys/kernel/debug/page_owner includes the migratetype
of the pageblock the page belongs to. This is also checked against the
page's migratetype (as declared by gfp_flags during its allocation), and
the page is reported as Fallback if its migratetype differs from the
pageblock's one. t This is somewhat misleading because in fact fallback
allocation is not the only reason why these two can differ. It also
doesn't direcly provide the page's migratetype, although it's possible
to derive that from the gfp_flags.
It's arguably better to print both page and pageblock's migratetype and
leave the interpretation to the consumer than to suggest fallback
allocation as the only possible reason. While at it, we can print the
migratetypes as string the same way as /proc/pagetypeinfo does, as some
of the numeric values depend on kernel configuration. For that, this
patch moves the migratetype_names array from #ifdef CONFIG_PROC_FS part
of mm/vmstat.c to mm/page_alloc.c and exports it.
With the new format strings for flags, we can now also provide symbolic
page and gfp flags in the /sys/kernel/debug/page_owner file. This
replaces the positional printing of page flags as single letters, which
might have looked nicer, but was limited to a subset of flags, and
required the user to remember the letters.
Example page_owner entry after the patch:
Page allocated via order 0, mask 0x24213ca(GFP_HIGHUSER_MOVABLE|__GFP_COLD|__GFP_NOWARN|__GFP_NORETRY)
PFN 520 type Movable Block 1 type Movable Flags 0xfffff8001006c(referenced|uptodate|lru|active|mappedtodisk)
[<ffffffff811682c4>] __alloc_pages_nodemask+0x134/0x230
[<ffffffff811b4058>] alloc_pages_current+0x88/0x120
[<ffffffff8115e386>] __page_cache_alloc+0xe6/0x120
[<ffffffff8116ba6c>] __do_page_cache_readahead+0xdc/0x240
[<ffffffff8116bd05>] ondemand_readahead+0x135/0x260
[<ffffffff8116bfb1>] page_cache_sync_readahead+0x31/0x50
[<ffffffff81160523>] generic_file_read_iter+0x453/0x760
[<ffffffff811e0d57>] __vfs_read+0xa7/0xd0
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Sasha Levin <sasha.levin@oracle.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Mel Gorman <mgorman@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-03-15 14:56:08 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2014-01-23 15:53:28 -08:00
|
|
|
|
2023-05-16 14:38:20 +08:00
|
|
|
|
|
|
|
|
|
mm: page_alloc: defrag_mode kswapd/kcompactd watermarks
The previous patch added pageblock_order reclaim to kswapd/kcompactd,
which helps, but produces only one block at a time. Allocation stalls and
THP failure rates are still higher than they could be.
To adequately reflect ALLOC_NOFRAGMENT demand for pageblocks, change the
watermarking for kswapd & kcompactd: instead of targeting the high
watermark in order-0 pages and checking for one suitable block, simply
require that the high watermark is entirely met in pageblocks.
To this end, track the number of free pages within contiguous pageblocks,
then change pgdat_balanced() and compact_finished() to check watermarks
against this new value.
This further reduces THP latencies and allocation stalls, and improves THP
success rates against the previous patch:
DEFRAGMODE-ASYNC DEFRAGMODE-ASYNC-WMARKS
Hugealloc Time mean 34300.36 ( +0.00%) 28904.00 ( -15.73%)
Hugealloc Time stddev 36390.42 ( +0.00%) 33464.37 ( -8.04%)
Kbuild Real time 196.13 ( +0.00%) 196.59 ( +0.23%)
Kbuild User time 1234.74 ( +0.00%) 1231.67 ( -0.25%)
Kbuild System time 62.62 ( +0.00%) 59.10 ( -5.54%)
THP fault alloc 57054.53 ( +0.00%) 63223.67 ( +10.81%)
THP fault fallback 11581.40 ( +0.00%) 5412.47 ( -53.26%)
Direct compact fail 107.80 ( +0.00%) 59.07 ( -44.79%)
Direct compact success 4.53 ( +0.00%) 2.80 ( -31.33%)
Direct compact success rate % 3.20 ( +0.00%) 3.99 ( +18.66%)
Compact daemon scanned migrate 5461033.93 ( +0.00%) 2267500.33 ( -58.48%)
Compact daemon scanned free 5824897.93 ( +0.00%) 2339773.00 ( -59.83%)
Compact direct scanned migrate 58336.93 ( +0.00%) 47659.93 ( -18.30%)
Compact direct scanned free 32791.87 ( +0.00%) 40729.67 ( +24.21%)
Compact total migrate scanned 5519370.87 ( +0.00%) 2315160.27 ( -58.05%)
Compact total free scanned 5857689.80 ( +0.00%) 2380502.67 ( -59.36%)
Alloc stall 2424.60 ( +0.00%) 638.87 ( -73.62%)
Pages kswapd scanned 2657018.33 ( +0.00%) 4002186.33 ( +50.63%)
Pages kswapd reclaimed 559583.07 ( +0.00%) 718577.80 ( +28.41%)
Pages direct scanned 722094.07 ( +0.00%) 355172.73 ( -50.81%)
Pages direct reclaimed 107257.80 ( +0.00%) 31162.80 ( -70.95%)
Pages total scanned 3379112.40 ( +0.00%) 4357359.07 ( +28.95%)
Pages total reclaimed 666840.87 ( +0.00%) 749740.60 ( +12.43%)
Swap out 77238.20 ( +0.00%) 110084.33 ( +42.53%)
Swap in 11712.80 ( +0.00%) 24457.00 ( +108.80%)
File refaults 143438.80 ( +0.00%) 188226.93 ( +31.22%)
Also of note is that compaction work overall is reduced. The reason for
this is that when free pageblocks are more readily available, allocations
are also much more likely to get physically placed in LRU order, instead
of being forced to scavenge free space here and there. This means that
reclaim by itself has better chances of freeing up whole blocks, and the
system relies less on compaction.
Comparing all changes to the vanilla kernel:
VANILLA DEFRAGMODE-ASYNC-WMARKS
Hugealloc Time mean 52739.45 ( +0.00%) 28904.00 ( -45.19%)
Hugealloc Time stddev 56541.26 ( +0.00%) 33464.37 ( -40.81%)
Kbuild Real time 197.47 ( +0.00%) 196.59 ( -0.44%)
Kbuild User time 1240.49 ( +0.00%) 1231.67 ( -0.71%)
Kbuild System time 70.08 ( +0.00%) 59.10 ( -15.45%)
THP fault alloc 46727.07 ( +0.00%) 63223.67 ( +35.30%)
THP fault fallback 21910.60 ( +0.00%) 5412.47 ( -75.29%)
Direct compact fail 195.80 ( +0.00%) 59.07 ( -69.48%)
Direct compact success 7.93 ( +0.00%) 2.80 ( -57.46%)
Direct compact success rate % 3.51 ( +0.00%) 3.99 ( +10.49%)
Compact daemon scanned migrate 3369601.27 ( +0.00%) 2267500.33 ( -32.71%)
Compact daemon scanned free 5075474.47 ( +0.00%) 2339773.00 ( -53.90%)
Compact direct scanned migrate 161787.27 ( +0.00%) 47659.93 ( -70.54%)
Compact direct scanned free 163467.53 ( +0.00%) 40729.67 ( -75.08%)
Compact total migrate scanned 3531388.53 ( +0.00%) 2315160.27 ( -34.44%)
Compact total free scanned 5238942.00 ( +0.00%) 2380502.67 ( -54.56%)
Alloc stall 2371.07 ( +0.00%) 638.87 ( -73.02%)
Pages kswapd scanned 2160926.73 ( +0.00%) 4002186.33 ( +85.21%)
Pages kswapd reclaimed 533191.07 ( +0.00%) 718577.80 ( +34.77%)
Pages direct scanned 400450.33 ( +0.00%) 355172.73 ( -11.31%)
Pages direct reclaimed 94441.73 ( +0.00%) 31162.80 ( -67.00%)
Pages total scanned 2561377.07 ( +0.00%) 4357359.07 ( +70.12%)
Pages total reclaimed 627632.80 ( +0.00%) 749740.60 ( +19.46%)
Swap out 47959.53 ( +0.00%) 110084.33 ( +129.53%)
Swap in 7276.00 ( +0.00%) 24457.00 ( +236.10%)
File refaults 138043.00 ( +0.00%) 188226.93 ( +36.35%)
THP allocation latencies and %sys time are down dramatically.
THP allocation failures are down from nearly 50% to 8.5%. And to recall
previous data points, the success rates are steady and reliable without
the cumulative deterioration of fragmentation events.
Compaction work is down overall. Direct compaction work especially is
drastically reduced. As an aside, its success rate of 4% indicates there
is room for improvement. For now it's good to rely on it less.
Reclaim work is up overall, however direct reclaim work is down. Part of
the increase can be attributed to a higher use of THPs, which due to
internal fragmentation increase the memory footprint. This is not
necessarily an unexpected side-effect for users of THP.
However, taken both points together, there may well be some opportunities
for fine tuning in the reclaim/compaction coordination.
[hannes@cmpxchg.org: fix squawks from rebasing]
Link: https://lkml.kernel.org/r/20250314210558.GD1316033@cmpxchg.org
Link: https://lkml.kernel.org/r/20250313210647.1314586-6-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-03-13 17:05:36 -04:00
|
|
|
|
[PATCH] Introduce mechanism for registering active regions of memory
At a basic level, architectures define structures to record where active
ranges of page frames are located. Once located, the code to calculate zone
sizes and holes in each architecture is very similar. Some of this zone and
hole sizing code is difficult to read for no good reason. This set of patches
eliminates the similar-looking architecture-specific code.
The patches introduce a mechanism where architectures register where the
active ranges of page frames are with add_active_range(). When all areas have
been discovered, free_area_init_nodes() is called to initialise the pgdat and
zones. The zone sizes and holes are then calculated in an architecture
independent manner.
Patch 1 introduces the mechanism for registering and initialising PFN ranges
Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed
Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed
Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed
Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed
Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable.
It adjusts the watermarks slightly
Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig,
gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on
IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based
machine. These were on patches against 2.6.17-rc1 and release 3 of these
patches but there have been no ia64-changes since release 3.
There are differences in the zone sizes for x86_64 as the arch-specific code
for x86_64 accounts the kernel image and the starting mem_maps as memory holes
but the architecture-independent code accounts the memory as present.
The big benefit of this set of patches is a sizable reduction of
architecture-specific code, some of which is very hairy. There should be a
greater reduction when other architectures use the same mechanisms for zone
and hole sizing but I lack the hardware to test on.
Additional credit;
Dave Hansen for the initial suggestion and comments on early patches
Andy Whitcroft for reviewing early versions and catching numerous
errors
Tony Luck for testing and debugging on IA64
Bob Picco for fixing bugs related to pfn registration, reviewing a
number of patch revisions, providing a number of suggestions
on future direction and testing heavily
Jack Steiner and Robin Holt for testing on IA64 and clarifying
issues related to memory holes
Yasunori for testing on IA64
Andi Kleen for reviewing and feeding back about x86_64
Christian Kujau for providing valuable information related to ACPI
problems on x86_64 and testing potential fixes
This patch:
Define the structure to represent an active range of page frames within a node
in an architecture independent manner. Architectures are expected to register
active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call
free_area_init_nodes() passing the PFNs of the end of each zone.
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Bob Picco <bob.picco@hp.com>
Cc: Dave Hansen <haveblue@us.ibm.com>
Cc: Andy Whitcroft <apw@shadowen.org>
Cc: Andi Kleen <ak@muc.de>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: "Keith Mannthey" <kmannth@gmail.com>
Cc: "Luck, Tony" <tony.luck@intel.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 01:49:43 -07:00
|
|
|
|
2011-12-08 10:22:09 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
[PATCH] Introduce mechanism for registering active regions of memory
At a basic level, architectures define structures to record where active
ranges of page frames are located. Once located, the code to calculate zone
sizes and holes in each architecture is very similar. Some of this zone and
hole sizing code is difficult to read for no good reason. This set of patches
eliminates the similar-looking architecture-specific code.
The patches introduce a mechanism where architectures register where the
active ranges of page frames are with add_active_range(). When all areas have
been discovered, free_area_init_nodes() is called to initialise the pgdat and
zones. The zone sizes and holes are then calculated in an architecture
independent manner.
Patch 1 introduces the mechanism for registering and initialising PFN ranges
Patch 2 changes ppc to use the mechanism - 139 arch-specific LOC removed
Patch 3 changes x86 to use the mechanism - 136 arch-specific LOC removed
Patch 4 changes x86_64 to use the mechanism - 74 arch-specific LOC removed
Patch 5 changes ia64 to use the mechanism - 52 arch-specific LOC removed
Patch 6 accounts for mem_map as a memory hole as the pages are not reclaimable.
It adjusts the watermarks slightly
Tony Luck has successfully tested for ia64 on Itanium with tiger_defconfig,
gensparse_defconfig and defconfig. Bob Picco has also tested and debugged on
IA64. Jack Steiner successfully boot tested on a mammoth SGI IA64-based
machine. These were on patches against 2.6.17-rc1 and release 3 of these
patches but there have been no ia64-changes since release 3.
There are differences in the zone sizes for x86_64 as the arch-specific code
for x86_64 accounts the kernel image and the starting mem_maps as memory holes
but the architecture-independent code accounts the memory as present.
The big benefit of this set of patches is a sizable reduction of
architecture-specific code, some of which is very hairy. There should be a
greater reduction when other architectures use the same mechanisms for zone
and hole sizing but I lack the hardware to test on.
Additional credit;
Dave Hansen for the initial suggestion and comments on early patches
Andy Whitcroft for reviewing early versions and catching numerous
errors
Tony Luck for testing and debugging on IA64
Bob Picco for fixing bugs related to pfn registration, reviewing a
number of patch revisions, providing a number of suggestions
on future direction and testing heavily
Jack Steiner and Robin Holt for testing on IA64 and clarifying
issues related to memory holes
Yasunori for testing on IA64
Andi Kleen for reviewing and feeding back about x86_64
Christian Kujau for providing valuable information related to ACPI
problems on x86_64 and testing potential fixes
This patch:
Define the structure to represent an active range of page frames within a node
in an architecture independent manner. Architectures are expected to register
active ranges of PFNs using add_active_range(nid, start_pfn, end_pfn) and call
free_area_init_nodes() passing the PFNs of the end of each zone.
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Bob Picco <bob.picco@hp.com>
Cc: Dave Hansen <haveblue@us.ibm.com>
Cc: Andy Whitcroft <apw@shadowen.org>
Cc: Andi Kleen <ak@muc.de>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: "Keith Mannthey" <kmannth@gmail.com>
Cc: "Luck, Tony" <tony.luck@intel.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-09-27 01:49:43 -07:00
|
|
|
|
2007-05-23 13:57:55 -07:00
|
|
|
|
2019-03-05 15:48:26 -08:00
|
|
|
|
2019-03-05 15:48:29 -08:00
|
|
|
|
2007-05-23 13:57:55 -07:00
|
|
|
|
2009-06-16 15:32:15 -07:00
|
|
|
|
2007-05-23 13:57:55 -07:00
|
|
|
|
|
|
|
|
|
2023-06-06 17:26:29 +03:00
|
|
|
|
2025-05-06 14:25:08 +03:00
|
|
|
|
|
|
|
|
|
2023-06-06 17:26:29 +03:00
|
|
|
|
|
|
|
|
|
2007-10-16 01:25:54 -07:00
|
|
|
|
|
|
|
|
|
2015-06-30 14:57:02 -07:00
|
|
|
|
mm/page_alloc.c: don't call kasan_free_pages() at deferred mem init
When CONFIG_KASAN is enabled on large memory SMP systems, the deferrred
pages initialization can take a long time. Below were the reported init
times on a 8-socket 96-core 4TB IvyBridge system.
1) Non-debug kernel without CONFIG_KASAN
[ 8.764222] node 1 initialised, 132086516 pages in 7027ms
2) Debug kernel with CONFIG_KASAN
[ 146.288115] node 1 initialised, 132075466 pages in 143052ms
So the page init time in a debug kernel was 20X of the non-debug kernel.
The long init time can be problematic as the page initialization is done
with interrupt disabled. In this particular case, it caused the
appearance of following warning messages as well as NMI backtraces of all
the cores that were doing the initialization.
[ 68.240049] rcu: INFO: rcu_sched detected stalls on CPUs/tasks:
[ 68.241000] rcu: 25-...0: (100 ticks this GP) idle=b72/1/0x4000000000000000 softirq=915/915 fqs=16252
[ 68.241000] rcu: 44-...0: (95 ticks this GP) idle=49a/1/0x4000000000000000 softirq=788/788 fqs=16253
[ 68.241000] rcu: 54-...0: (104 ticks this GP) idle=03a/1/0x4000000000000000 softirq=721/825 fqs=16253
[ 68.241000] rcu: 60-...0: (103 ticks this GP) idle=cbe/1/0x4000000000000000 softirq=637/740 fqs=16253
[ 68.241000] rcu: 72-...0: (105 ticks this GP) idle=786/1/0x4000000000000000 softirq=536/641 fqs=16253
[ 68.241000] rcu: 84-...0: (99 ticks this GP) idle=292/1/0x4000000000000000 softirq=537/537 fqs=16253
[ 68.241000] rcu: 111-...0: (104 ticks this GP) idle=bde/1/0x4000000000000000 softirq=474/476 fqs=16253
[ 68.241000] rcu: (detected by 13, t=65018 jiffies, g=249, q=2)
The long init time was mainly caused by the call to kasan_free_pages() to
poison the newly initialized pages. On a 4TB system, we are talking about
almost 500GB of memory probably on the same node.
In reality, we may not need to poison the newly initialized pages before
they are ever allocated. So KASAN poisoning of freed pages before the
completion of deferred memory initialization is now disabled. Those pages
will be properly poisoned when they are allocated or freed after deferred
pages initialization is done.
With this change, the new page initialization time became:
[ 21.948010] node 1 initialised, 132075466 pages in 18702ms
This was still about double the non-debug kernel time, but was much
better than before.
Link: http://lkml.kernel.org/r/1544459388-8736-1-git-send-email-longman@redhat.com
Signed-off-by: Waiman Long <longman@redhat.com>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Pasha Tatashin <Pavel.Tatashin@microsoft.com>
Cc: Oscar Salvador <osalvador@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:38:51 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-03-21 19:05:02 +02:00
|
|
|
|
mm/page_alloc.c: don't call kasan_free_pages() at deferred mem init
When CONFIG_KASAN is enabled on large memory SMP systems, the deferrred
pages initialization can take a long time. Below were the reported init
times on a 8-socket 96-core 4TB IvyBridge system.
1) Non-debug kernel without CONFIG_KASAN
[ 8.764222] node 1 initialised, 132086516 pages in 7027ms
2) Debug kernel with CONFIG_KASAN
[ 146.288115] node 1 initialised, 132075466 pages in 143052ms
So the page init time in a debug kernel was 20X of the non-debug kernel.
The long init time can be problematic as the page initialization is done
with interrupt disabled. In this particular case, it caused the
appearance of following warning messages as well as NMI backtraces of all
the cores that were doing the initialization.
[ 68.240049] rcu: INFO: rcu_sched detected stalls on CPUs/tasks:
[ 68.241000] rcu: 25-...0: (100 ticks this GP) idle=b72/1/0x4000000000000000 softirq=915/915 fqs=16252
[ 68.241000] rcu: 44-...0: (95 ticks this GP) idle=49a/1/0x4000000000000000 softirq=788/788 fqs=16253
[ 68.241000] rcu: 54-...0: (104 ticks this GP) idle=03a/1/0x4000000000000000 softirq=721/825 fqs=16253
[ 68.241000] rcu: 60-...0: (103 ticks this GP) idle=cbe/1/0x4000000000000000 softirq=637/740 fqs=16253
[ 68.241000] rcu: 72-...0: (105 ticks this GP) idle=786/1/0x4000000000000000 softirq=536/641 fqs=16253
[ 68.241000] rcu: 84-...0: (99 ticks this GP) idle=292/1/0x4000000000000000 softirq=537/537 fqs=16253
[ 68.241000] rcu: 111-...0: (104 ticks this GP) idle=bde/1/0x4000000000000000 softirq=474/476 fqs=16253
[ 68.241000] rcu: (detected by 13, t=65018 jiffies, g=249, q=2)
The long init time was mainly caused by the call to kasan_free_pages() to
poison the newly initialized pages. On a 4TB system, we are talking about
almost 500GB of memory probably on the same node.
In reality, we may not need to poison the newly initialized pages before
they are ever allocated. So KASAN poisoning of freed pages before the
completion of deferred memory initialization is now disabled. Those pages
will be properly poisoned when they are allocated or freed after deferred
pages initialization is done.
With this change, the new page initialization time became:
[ 21.948010] node 1 initialised, 132075466 pages in 18702ms
This was still about double the non-debug kernel time, but was much
better than before.
Link: http://lkml.kernel.org/r/1544459388-8736-1-git-send-email-longman@redhat.com
Signed-off-by: Waiman Long <longman@redhat.com>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Pasha Tatashin <Pavel.Tatashin@microsoft.com>
Cc: Oscar Salvador <osalvador@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:38:51 -08:00
|
|
|
|
kasan, page_alloc: deduplicate should_skip_kasan_poison
Patch series "kasan, vmalloc, arm64: add vmalloc tagging support for SW/HW_TAGS", v6.
This patchset adds vmalloc tagging support for SW_TAGS and HW_TAGS
KASAN modes.
About half of patches are cleanups I went for along the way. None of them
seem to be important enough to go through stable, so I decided not to
split them out into separate patches/series.
The patchset is partially based on an early version of the HW_TAGS
patchset by Vincenzo that had vmalloc support. Thus, I added a
Co-developed-by tag into a few patches.
SW_TAGS vmalloc tagging support is straightforward. It reuses all of the
generic KASAN machinery, but uses shadow memory to store tags instead of
magic values. Naturally, vmalloc tagging requires adding a few
kasan_reset_tag() annotations to the vmalloc code.
HW_TAGS vmalloc tagging support stands out. HW_TAGS KASAN is based on Arm
MTE, which can only assigns tags to physical memory. As a result, HW_TAGS
KASAN only tags vmalloc() allocations, which are backed by page_alloc
memory. It ignores vmap() and others.
This patch (of 39):
Currently, should_skip_kasan_poison() has two definitions: one for when
CONFIG_DEFERRED_STRUCT_PAGE_INIT is enabled, one for when it's not.
Instead of duplicating the checks, add a deferred_pages_enabled() helper
and use it in a single should_skip_kasan_poison() definition.
Also move should_skip_kasan_poison() closer to its caller and clarify all
conditions in the comment.
Link: https://lkml.kernel.org/r/cover.1643047180.git.andreyknvl@google.com
Link: https://lkml.kernel.org/r/658b79f5fb305edaf7dc16bc52ea870d3220d4a8.1643047180.git.andreyknvl@google.com
Signed-off-by: Andrey Konovalov <andreyknvl@google.com>
Acked-by: Marco Elver <elver@google.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Andrey Ryabinin <ryabinin.a.a@gmail.com>
Cc: Vincenzo Frascino <vincenzo.frascino@arm.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Will Deacon <will@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Peter Collingbourne <pcc@google.com>
Cc: Evgenii Stepanov <eugenis@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-03-24 18:10:04 -07:00
|
|
|
|
mm/page_alloc.c: don't call kasan_free_pages() at deferred mem init
When CONFIG_KASAN is enabled on large memory SMP systems, the deferrred
pages initialization can take a long time. Below were the reported init
times on a 8-socket 96-core 4TB IvyBridge system.
1) Non-debug kernel without CONFIG_KASAN
[ 8.764222] node 1 initialised, 132086516 pages in 7027ms
2) Debug kernel with CONFIG_KASAN
[ 146.288115] node 1 initialised, 132075466 pages in 143052ms
So the page init time in a debug kernel was 20X of the non-debug kernel.
The long init time can be problematic as the page initialization is done
with interrupt disabled. In this particular case, it caused the
appearance of following warning messages as well as NMI backtraces of all
the cores that were doing the initialization.
[ 68.240049] rcu: INFO: rcu_sched detected stalls on CPUs/tasks:
[ 68.241000] rcu: 25-...0: (100 ticks this GP) idle=b72/1/0x4000000000000000 softirq=915/915 fqs=16252
[ 68.241000] rcu: 44-...0: (95 ticks this GP) idle=49a/1/0x4000000000000000 softirq=788/788 fqs=16253
[ 68.241000] rcu: 54-...0: (104 ticks this GP) idle=03a/1/0x4000000000000000 softirq=721/825 fqs=16253
[ 68.241000] rcu: 60-...0: (103 ticks this GP) idle=cbe/1/0x4000000000000000 softirq=637/740 fqs=16253
[ 68.241000] rcu: 72-...0: (105 ticks this GP) idle=786/1/0x4000000000000000 softirq=536/641 fqs=16253
[ 68.241000] rcu: 84-...0: (99 ticks this GP) idle=292/1/0x4000000000000000 softirq=537/537 fqs=16253
[ 68.241000] rcu: 111-...0: (104 ticks this GP) idle=bde/1/0x4000000000000000 softirq=474/476 fqs=16253
[ 68.241000] rcu: (detected by 13, t=65018 jiffies, g=249, q=2)
The long init time was mainly caused by the call to kasan_free_pages() to
poison the newly initialized pages. On a 4TB system, we are talking about
almost 500GB of memory probably on the same node.
In reality, we may not need to poison the newly initialized pages before
they are ever allocated. So KASAN poisoning of freed pages before the
completion of deferred memory initialization is now disabled. Those pages
will be properly poisoned when they are allocated or freed after deferred
pages initialization is done.
With this change, the new page initialization time became:
[ 21.948010] node 1 initialised, 132075466 pages in 18702ms
This was still about double the non-debug kernel time, but was much
better than before.
Link: http://lkml.kernel.org/r/1544459388-8736-1-git-send-email-longman@redhat.com
Signed-off-by: Waiman Long <longman@redhat.com>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Pasha Tatashin <Pavel.Tatashin@microsoft.com>
Cc: Oscar Salvador <osalvador@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:38:51 -08:00
|
|
|
|
kasan, page_alloc: deduplicate should_skip_kasan_poison
Patch series "kasan, vmalloc, arm64: add vmalloc tagging support for SW/HW_TAGS", v6.
This patchset adds vmalloc tagging support for SW_TAGS and HW_TAGS
KASAN modes.
About half of patches are cleanups I went for along the way. None of them
seem to be important enough to go through stable, so I decided not to
split them out into separate patches/series.
The patchset is partially based on an early version of the HW_TAGS
patchset by Vincenzo that had vmalloc support. Thus, I added a
Co-developed-by tag into a few patches.
SW_TAGS vmalloc tagging support is straightforward. It reuses all of the
generic KASAN machinery, but uses shadow memory to store tags instead of
magic values. Naturally, vmalloc tagging requires adding a few
kasan_reset_tag() annotations to the vmalloc code.
HW_TAGS vmalloc tagging support stands out. HW_TAGS KASAN is based on Arm
MTE, which can only assigns tags to physical memory. As a result, HW_TAGS
KASAN only tags vmalloc() allocations, which are backed by page_alloc
memory. It ignores vmap() and others.
This patch (of 39):
Currently, should_skip_kasan_poison() has two definitions: one for when
CONFIG_DEFERRED_STRUCT_PAGE_INIT is enabled, one for when it's not.
Instead of duplicating the checks, add a deferred_pages_enabled() helper
and use it in a single should_skip_kasan_poison() definition.
Also move should_skip_kasan_poison() closer to its caller and clarify all
conditions in the comment.
Link: https://lkml.kernel.org/r/cover.1643047180.git.andreyknvl@google.com
Link: https://lkml.kernel.org/r/658b79f5fb305edaf7dc16bc52ea870d3220d4a8.1643047180.git.andreyknvl@google.com
Signed-off-by: Andrey Konovalov <andreyknvl@google.com>
Acked-by: Marco Elver <elver@google.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Andrey Ryabinin <ryabinin.a.a@gmail.com>
Cc: Vincenzo Frascino <vincenzo.frascino@arm.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Will Deacon <will@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Peter Collingbourne <pcc@google.com>
Cc: Evgenii Stepanov <eugenis@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-03-24 18:10:04 -07:00
|
|
|
|
mm/page_alloc.c: don't call kasan_free_pages() at deferred mem init
When CONFIG_KASAN is enabled on large memory SMP systems, the deferrred
pages initialization can take a long time. Below were the reported init
times on a 8-socket 96-core 4TB IvyBridge system.
1) Non-debug kernel without CONFIG_KASAN
[ 8.764222] node 1 initialised, 132086516 pages in 7027ms
2) Debug kernel with CONFIG_KASAN
[ 146.288115] node 1 initialised, 132075466 pages in 143052ms
So the page init time in a debug kernel was 20X of the non-debug kernel.
The long init time can be problematic as the page initialization is done
with interrupt disabled. In this particular case, it caused the
appearance of following warning messages as well as NMI backtraces of all
the cores that were doing the initialization.
[ 68.240049] rcu: INFO: rcu_sched detected stalls on CPUs/tasks:
[ 68.241000] rcu: 25-...0: (100 ticks this GP) idle=b72/1/0x4000000000000000 softirq=915/915 fqs=16252
[ 68.241000] rcu: 44-...0: (95 ticks this GP) idle=49a/1/0x4000000000000000 softirq=788/788 fqs=16253
[ 68.241000] rcu: 54-...0: (104 ticks this GP) idle=03a/1/0x4000000000000000 softirq=721/825 fqs=16253
[ 68.241000] rcu: 60-...0: (103 ticks this GP) idle=cbe/1/0x4000000000000000 softirq=637/740 fqs=16253
[ 68.241000] rcu: 72-...0: (105 ticks this GP) idle=786/1/0x4000000000000000 softirq=536/641 fqs=16253
[ 68.241000] rcu: 84-...0: (99 ticks this GP) idle=292/1/0x4000000000000000 softirq=537/537 fqs=16253
[ 68.241000] rcu: 111-...0: (104 ticks this GP) idle=bde/1/0x4000000000000000 softirq=474/476 fqs=16253
[ 68.241000] rcu: (detected by 13, t=65018 jiffies, g=249, q=2)
The long init time was mainly caused by the call to kasan_free_pages() to
poison the newly initialized pages. On a 4TB system, we are talking about
almost 500GB of memory probably on the same node.
In reality, we may not need to poison the newly initialized pages before
they are ever allocated. So KASAN poisoning of freed pages before the
completion of deferred memory initialization is now disabled. Those pages
will be properly poisoned when they are allocated or freed after deferred
pages initialization is done.
With this change, the new page initialization time became:
[ 21.948010] node 1 initialised, 132075466 pages in 18702ms
This was still about double the non-debug kernel time, but was much
better than before.
Link: http://lkml.kernel.org/r/1544459388-8736-1-git-send-email-longman@redhat.com
Signed-off-by: Waiman Long <longman@redhat.com>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Pasha Tatashin <Pavel.Tatashin@microsoft.com>
Cc: Oscar Salvador <osalvador@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:38:51 -08:00
|
|
|
|
|
|
|
|
|
2015-06-30 14:57:02 -07:00
|
|
|
|
2023-03-21 19:05:02 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-06-30 14:57:02 -07:00
|
|
|
|
2023-03-21 19:05:02 +02:00
|
|
|
|
|
|
|
|
|
2015-06-30 14:57:02 -07:00
|
|
|
|
mm/page_alloc.c: don't show protection in zone's ->lowmem_reserve[] for empty zone
On one node, for lower zone's ->lowmem_reserve[], it will show how much
memory is reserved in this lower zone to avoid excessive page allocation
from the relevant higher zone's fallback allocation.
However, currently lower zone's lowmem_reserve[] element will be filled
even though the relevant higher zone is empty. That doesnt' make sense
and can cause confusion.
E.g on node 0 of one system as below, it has zone
DMA/DMA32/NORMAL/MOVABLE/DEVICE, among them zone MOVABLE/DEVICE are the
highest and both are empty. In zone DMA/DMA32's protection array, we can
see that it has value for zone MOVABLE and DEVICE.
Node 0, zone DMA
......
pages free 2816
boost 0
min 7
low 10
high 13
spanned 4095
present 3998
managed 3840
cma 0
protection: (0, 1582, 23716, 23716, 23716)
......
Node 0, zone DMA32
pages free 403269
boost 0
min 753
low 1158
high 1563
spanned 1044480
present 487039
managed 405070
cma 0
protection: (0, 0, 22134, 22134, 22134)
......
Node 0, zone Normal
pages free 5423879
boost 0
min 10539
low 16205
high 21871
spanned 5767168
present 5767168
managed 5666438
cma 0
protection: (0, 0, 0, 0, 0)
......
Node 0, zone Movable
pages free 0
boost 0
min 32
low 32
high 32
spanned 0
present 0
managed 0
cma 0
protection: (0, 0, 0, 0, 0)
Node 0, zone Device
pages free 0
boost 0
min 0
low 0
high 0
spanned 0
present 0
managed 0
cma 0
protection: (0, 0, 0, 0, 0)
Here, clear out the element value in lower zone's ->lowmem_reserve[] if the
relevant higher zone is empty.
And also replace space with tab in _deferred_grow_zone()
Link: https://lkml.kernel.org/r/20240326061134.1055295-7-bhe@redhat.com
Signed-off-by: Baoquan He <bhe@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: "Mike Rapoport (IBM)" <rppt@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-26 14:11:32 +08:00
|
|
|
|
2015-06-30 14:57:02 -07:00
|
|
|
|
|
|
|
|
|
kasan, page_alloc: deduplicate should_skip_kasan_poison
Patch series "kasan, vmalloc, arm64: add vmalloc tagging support for SW/HW_TAGS", v6.
This patchset adds vmalloc tagging support for SW_TAGS and HW_TAGS
KASAN modes.
About half of patches are cleanups I went for along the way. None of them
seem to be important enough to go through stable, so I decided not to
split them out into separate patches/series.
The patchset is partially based on an early version of the HW_TAGS
patchset by Vincenzo that had vmalloc support. Thus, I added a
Co-developed-by tag into a few patches.
SW_TAGS vmalloc tagging support is straightforward. It reuses all of the
generic KASAN machinery, but uses shadow memory to store tags instead of
magic values. Naturally, vmalloc tagging requires adding a few
kasan_reset_tag() annotations to the vmalloc code.
HW_TAGS vmalloc tagging support stands out. HW_TAGS KASAN is based on Arm
MTE, which can only assigns tags to physical memory. As a result, HW_TAGS
KASAN only tags vmalloc() allocations, which are backed by page_alloc
memory. It ignores vmap() and others.
This patch (of 39):
Currently, should_skip_kasan_poison() has two definitions: one for when
CONFIG_DEFERRED_STRUCT_PAGE_INIT is enabled, one for when it's not.
Instead of duplicating the checks, add a deferred_pages_enabled() helper
and use it in a single should_skip_kasan_poison() definition.
Also move should_skip_kasan_poison() closer to its caller and clarify all
conditions in the comment.
Link: https://lkml.kernel.org/r/cover.1643047180.git.andreyknvl@google.com
Link: https://lkml.kernel.org/r/658b79f5fb305edaf7dc16bc52ea870d3220d4a8.1643047180.git.andreyknvl@google.com
Signed-off-by: Andrey Konovalov <andreyknvl@google.com>
Acked-by: Marco Elver <elver@google.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Andrey Ryabinin <ryabinin.a.a@gmail.com>
Cc: Vincenzo Frascino <vincenzo.frascino@arm.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Will Deacon <will@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Peter Collingbourne <pcc@google.com>
Cc: Evgenii Stepanov <eugenis@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-03-24 18:10:04 -07:00
|
|
|
|
mm, kasan: don't poison boot memory with tag-based modes
During boot, all non-reserved memblock memory is exposed to page_alloc via
memblock_free_pages->__free_pages_core(). This results in
kasan_free_pages() being called, which poisons that memory.
Poisoning all that memory lengthens boot time. The most noticeable effect
is observed with the HW_TAGS mode. A boot-time impact may potentially
also affect systems with large amount of RAM.
This patch changes the tag-based modes to not poison the memory during the
memblock->page_alloc transition.
An exception is made for KASAN_GENERIC. Since it marks all new memory as
accessible, not poisoning the memory released from memblock will lead to
KASAN missing invalid boot-time accesses to that memory.
With KASAN_SW_TAGS, as it uses the invalid 0xFE tag as the default tag for
all memory, it won't miss bad boot-time accesses even if the poisoning of
memblock memory is removed.
With KASAN_HW_TAGS, the default memory tags values are unspecified.
Therefore, if memblock poisoning is removed, this KASAN mode will miss the
mentioned type of boot-time bugs with a 1/16 probability. This is taken
as an acceptable trafe-off.
Internally, the poisoning is removed as follows. __free_pages_core() is
used when exposing fresh memory during system boot and when onlining
memory during hotplug. This patch adds a new FPI_SKIP_KASAN_POISON flag
and passes it to __free_pages_ok() through free_pages_prepare() from
__free_pages_core(). If FPI_SKIP_KASAN_POISON is set, kasan_free_pages()
is not called.
All memory allocated normally when the boot is over keeps getting poisoned
as usual.
Link: https://lkml.kernel.org/r/a0570dc1e3a8f39a55aa343a1fc08cd5c2d4cad6.1613692950.git.andreyknvl@google.com
Signed-off-by: Andrey Konovalov <andreyknvl@google.com>
Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Will Deacon <will.deacon@arm.com>
Cc: Vincenzo Frascino <vincenzo.frascino@arm.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Marco Elver <elver@google.com>
Cc: Peter Collingbourne <pcc@google.com>
Cc: Evgenii Stepanov <eugenis@google.com>
Cc: Branislav Rankov <Branislav.Rankov@arm.com>
Cc: Kevin Brodsky <kevin.brodsky@arm.com>
Cc: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:59:52 -07:00
|
|
|
|
kasan, page_alloc: deduplicate should_skip_kasan_poison
Patch series "kasan, vmalloc, arm64: add vmalloc tagging support for SW/HW_TAGS", v6.
This patchset adds vmalloc tagging support for SW_TAGS and HW_TAGS
KASAN modes.
About half of patches are cleanups I went for along the way. None of them
seem to be important enough to go through stable, so I decided not to
split them out into separate patches/series.
The patchset is partially based on an early version of the HW_TAGS
patchset by Vincenzo that had vmalloc support. Thus, I added a
Co-developed-by tag into a few patches.
SW_TAGS vmalloc tagging support is straightforward. It reuses all of the
generic KASAN machinery, but uses shadow memory to store tags instead of
magic values. Naturally, vmalloc tagging requires adding a few
kasan_reset_tag() annotations to the vmalloc code.
HW_TAGS vmalloc tagging support stands out. HW_TAGS KASAN is based on Arm
MTE, which can only assigns tags to physical memory. As a result, HW_TAGS
KASAN only tags vmalloc() allocations, which are backed by page_alloc
memory. It ignores vmap() and others.
This patch (of 39):
Currently, should_skip_kasan_poison() has two definitions: one for when
CONFIG_DEFERRED_STRUCT_PAGE_INIT is enabled, one for when it's not.
Instead of duplicating the checks, add a deferred_pages_enabled() helper
and use it in a single should_skip_kasan_poison() definition.
Also move should_skip_kasan_poison() closer to its caller and clarify all
conditions in the comment.
Link: https://lkml.kernel.org/r/cover.1643047180.git.andreyknvl@google.com
Link: https://lkml.kernel.org/r/658b79f5fb305edaf7dc16bc52ea870d3220d4a8.1643047180.git.andreyknvl@google.com
Signed-off-by: Andrey Konovalov <andreyknvl@google.com>
Acked-by: Marco Elver <elver@google.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Andrey Ryabinin <ryabinin.a.a@gmail.com>
Cc: Vincenzo Frascino <vincenzo.frascino@arm.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Will Deacon <will@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Peter Collingbourne <pcc@google.com>
Cc: Evgenii Stepanov <eugenis@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-03-24 18:10:04 -07:00
|
|
|
|
mm, kasan: don't poison boot memory with tag-based modes
During boot, all non-reserved memblock memory is exposed to page_alloc via
memblock_free_pages->__free_pages_core(). This results in
kasan_free_pages() being called, which poisons that memory.
Poisoning all that memory lengthens boot time. The most noticeable effect
is observed with the HW_TAGS mode. A boot-time impact may potentially
also affect systems with large amount of RAM.
This patch changes the tag-based modes to not poison the memory during the
memblock->page_alloc transition.
An exception is made for KASAN_GENERIC. Since it marks all new memory as
accessible, not poisoning the memory released from memblock will lead to
KASAN missing invalid boot-time accesses to that memory.
With KASAN_SW_TAGS, as it uses the invalid 0xFE tag as the default tag for
all memory, it won't miss bad boot-time accesses even if the poisoning of
memblock memory is removed.
With KASAN_HW_TAGS, the default memory tags values are unspecified.
Therefore, if memblock poisoning is removed, this KASAN mode will miss the
mentioned type of boot-time bugs with a 1/16 probability. This is taken
as an acceptable trafe-off.
Internally, the poisoning is removed as follows. __free_pages_core() is
used when exposing fresh memory during system boot and when onlining
memory during hotplug. This patch adds a new FPI_SKIP_KASAN_POISON flag
and passes it to __free_pages_ok() through free_pages_prepare() from
__free_pages_core(). If FPI_SKIP_KASAN_POISON is set, kasan_free_pages()
is not called.
All memory allocated normally when the boot is over keeps getting poisoned
as usual.
Link: https://lkml.kernel.org/r/a0570dc1e3a8f39a55aa343a1fc08cd5c2d4cad6.1613692950.git.andreyknvl@google.com
Signed-off-by: Andrey Konovalov <andreyknvl@google.com>
Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Will Deacon <will.deacon@arm.com>
Cc: Vincenzo Frascino <vincenzo.frascino@arm.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Marco Elver <elver@google.com>
Cc: Peter Collingbourne <pcc@google.com>
Cc: Evgenii Stepanov <eugenis@google.com>
Cc: Branislav Rankov <Branislav.Rankov@arm.com>
Cc: Kevin Brodsky <kevin.brodsky@arm.com>
Cc: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:59:52 -07:00
|
|
|
|
2024-08-09 14:48:48 +03:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-03-21 19:05:02 +02:00
|
|
|
|
2015-06-30 14:57:02 -07:00
|
|
|
|
2016-05-19 17:14:27 -07:00
|
|
|
|
2021-06-28 19:41:22 -07:00
|
|
|
|
2016-05-19 17:14:27 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/sparsemem: introduce struct mem_section_usage
Patch series "mm: Sub-section memory hotplug support", v10.
The memory hotplug section is an arbitrary / convenient unit for memory
hotplug. 'Section-size' units have bled into the user interface
('memblock' sysfs) and can not be changed without breaking existing
userspace. The section-size constraint, while mostly benign for typical
memory hotplug, has and continues to wreak havoc with 'device-memory'
use cases, persistent memory (pmem) in particular. Recall that pmem
uses devm_memremap_pages(), and subsequently arch_add_memory(), to
allocate a 'struct page' memmap for pmem. However, it does not use the
'bottom half' of memory hotplug, i.e. never marks pmem pages online and
never exposes the userspace memblock interface for pmem. This leaves an
opening to redress the section-size constraint.
To date, the libnvdimm subsystem has attempted to inject padding to
satisfy the internal constraints of arch_add_memory(). Beyond
complicating the code, leading to bugs [2], wasting memory, and limiting
configuration flexibility, the padding hack is broken when the platform
changes this physical memory alignment of pmem from one boot to the
next. Device failure (intermittent or permanent) and physical
reconfiguration are events that can cause the platform firmware to
change the physical placement of pmem on a subsequent boot, and device
failure is an everyday event in a data-center.
It turns out that sections are only a hard requirement of the
user-facing interface for memory hotplug and with a bit more
infrastructure sub-section arch_add_memory() support can be added for
kernel internal usages like devm_memremap_pages(). Here is an analysis
of the current design assumptions in the current code and how they are
addressed in the new implementation:
Current design assumptions:
- Sections that describe boot memory (early sections) are never
unplugged / removed.
- pfn_valid(), in the CONFIG_SPARSEMEM_VMEMMAP=y, case devolves to a
valid_section() check
- __add_pages() and helper routines assume all operations occur in
PAGES_PER_SECTION units.
- The memblock sysfs interface only comprehends full sections
New design assumptions:
- Sections are instrumented with a sub-section bitmask to track (on
x86) individual 2MB sub-divisions of a 128MB section.
- Partially populated early sections can be extended with additional
sub-sections, and those sub-sections can be removed with
arch_remove_memory(). With this in place we no longer lose usable
memory capacity to padding.
- pfn_valid() is updated to look deeper than valid_section() to also
check the active-sub-section mask. This indication is in the same
cacheline as the valid_section() so the performance impact is
expected to be negligible. So far the lkp robot has not reported any
regressions.
- Outside of the core vmemmap population routines which are replaced,
other helper routines like shrink_{zone,pgdat}_span() are updated to
handle the smaller granularity. Core memory hotplug routines that
deal with online memory are not touched.
- The existing memblock sysfs user api guarantees / assumptions are not
touched since this capability is limited to !online
!memblock-sysfs-accessible sections.
Meanwhile the issue reports continue to roll in from users that do not
understand when and how the 128MB constraint will bite them. The current
implementation relied on being able to support at least one misaligned
namespace, but that immediately falls over on any moderately complex
namespace creation attempt. Beyond the initial problem of 'System RAM'
colliding with pmem, and the unsolvable problem of physical alignment
changes, Linux is now being exposed to platforms that collide pmem ranges
with other pmem ranges by default [3]. In short, devm_memremap_pages()
has pushed the venerable section-size constraint past the breaking point,
and the simplicity of section-aligned arch_add_memory() is no longer
tenable.
These patches are exposed to the kbuild robot on a subsection-v10 branch
[4], and a preview of the unit test for this functionality is available
on the 'subsection-pending' branch of ndctl [5].
[2]: https://lore.kernel.org/r/155000671719.348031.2347363160141119237.stgit@dwillia2-desk3.amr.corp.intel.com
[3]: https://github.com/pmem/ndctl/issues/76
[4]: https://git.kernel.org/pub/scm/linux/kernel/git/djbw/nvdimm.git/log/?h=subsection-v10
[5]: https://github.com/pmem/ndctl/commit/7c59b4867e1c
This patch (of 13):
Towards enabling memory hotplug to track partial population of a section,
introduce 'struct mem_section_usage'.
A pointer to a 'struct mem_section_usage' instance replaces the existing
pointer to a 'pageblock_flags' bitmap. Effectively it adds one more
'unsigned long' beyond the 'pageblock_flags' (usemap) allocation to house
a new 'subsection_map' bitmap. The new bitmap enables the memory
hot{plug,remove} implementation to act on incremental sub-divisions of a
section.
SUBSECTION_SHIFT is defined as global constant instead of per-architecture
value like SECTION_SIZE_BITS in order to allow cross-arch compatibility of
subsection users. Specifically a common subsection size allows for the
possibility that persistent memory namespace configurations be made
compatible across architectures.
The primary motivation for this functionality is to support platforms that
mix "System RAM" and "Persistent Memory" within a single section, or
multiple PMEM ranges with different mapping lifetimes within a single
section. The section restriction for hotplug has caused an ongoing saga
of hacks and bugs for devm_memremap_pages() users.
Beyond the fixups to teach existing paths how to retrieve the 'usemap'
from a section, and updates to usemap allocation path, there are no
expected behavior changes.
Link: http://lkml.kernel.org/r/156092349845.979959.73333291612799019.stgit@dwillia2-desk3.amr.corp.intel.com
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Wei Yang <richardw.yang@linux.intel.com>
Tested-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com> [ppc64]
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Logan Gunthorpe <logang@deltatee.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Jérôme Glisse <jglisse@redhat.com>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Jane Chu <jane.chu@oracle.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Qian Cai <cai@lca.pw>
Cc: Logan Gunthorpe <logang@deltatee.com>
Cc: Toshi Kani <toshi.kani@hpe.com>
Cc: Jeff Moyer <jmoyer@redhat.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Jason Gunthorpe <jgg@mellanox.com>
Cc: Christoph Hellwig <hch@lst.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-18 15:57:57 -07:00
|
|
|
|
2016-05-19 17:14:27 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-28 19:41:22 -07:00
|
|
|
|
2016-05-19 17:14:27 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-09-07 14:08:42 +08:00
|
|
|
|
2016-05-19 17:14:27 -07:00
|
|
|
|
2020-08-06 23:25:44 -07:00
|
|
|
|
2016-05-19 17:14:27 -07:00
|
|
|
|
|
|
|
|
|
mm/page_alloc: pageblock flags functions clean up
Patch series "Make MIGRATE_ISOLATE a standalone bit", v10.
This patchset moves MIGRATE_ISOLATE to a standalone bit to avoid being
overwritten during pageblock isolation process. Currently,
MIGRATE_ISOLATE is part of enum migratetype (in include/linux/mmzone.h),
thus, setting a pageblock to MIGRATE_ISOLATE overwrites its original
migratetype. This causes pageblock migratetype loss during
alloc_contig_range() and memory offline, especially when the process fails
due to a failed pageblock isolation and the code tries to undo the
finished pageblock isolations.
In terms of performance for changing pageblock types, no performance
change is observed:
1. I used perf to collect stats of offlining and onlining all memory
of a 40GB VM 10 times and see that get_pfnblock_flags_mask() and
set_pfnblock_flags_mask() take about 0.12% and 0.02% of the whole
process respectively with and without this patchset across 3 runs.
2. I used perf to collect stats of dd from /dev/random to a 40GB tmpfs
file and find get_pfnblock_flags_mask() takes about 0.05% of the
process with and without this patchset across 3 runs.
This patch (of 6):
No functional change is intended.
1. Add __NR_PAGEBLOCK_BITS for the number of pageblock flag bits and use
roundup_pow_of_two(__NR_PAGEBLOCK_BITS) as NR_PAGEBLOCK_BITS to take
right amount of bits for pageblock flags.
2. Rename PB_migrate_skip to PB_compact_skip.
3. Add {get,set,clear}_pfnblock_bit() to operate one a standalone bit,
like PB_compact_skip.
3. Make {get,set}_pfnblock_flags_mask() internal functions and use
{get,set}_pfnblock_migratetype() for pageblock migratetype operations.
4. Move pageblock flags common code to get_pfnblock_bitmap_bitidx().
3. Use MIGRATETYPE_MASK to get the migratetype of a pageblock from its
flags.
4. Use PB_migrate_end in the definition of MIGRATETYPE_MASK instead of
PB_migrate_bits.
5. Add a comment on is_migrate_cma_folio() to prevent one from changing it
to use get_pageblock_migratetype() and causing issues.
Link: https://lkml.kernel.org/r/20250617021115.2331563-1-ziy@nvidia.com
Link: https://lkml.kernel.org/r/20250617021115.2331563-2-ziy@nvidia.com
Signed-off-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kirill A. Shuemov <kirill.shutemov@linux.intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Richard Chang <richardycc@google.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-06-16 22:11:09 -04:00
|
|
|
|
|
|
|
|
|
2025-08-27 07:01:05 +00:00
|
|
|
|
mm/page_alloc: pageblock flags functions clean up
Patch series "Make MIGRATE_ISOLATE a standalone bit", v10.
This patchset moves MIGRATE_ISOLATE to a standalone bit to avoid being
overwritten during pageblock isolation process. Currently,
MIGRATE_ISOLATE is part of enum migratetype (in include/linux/mmzone.h),
thus, setting a pageblock to MIGRATE_ISOLATE overwrites its original
migratetype. This causes pageblock migratetype loss during
alloc_contig_range() and memory offline, especially when the process fails
due to a failed pageblock isolation and the code tries to undo the
finished pageblock isolations.
In terms of performance for changing pageblock types, no performance
change is observed:
1. I used perf to collect stats of offlining and onlining all memory
of a 40GB VM 10 times and see that get_pfnblock_flags_mask() and
set_pfnblock_flags_mask() take about 0.12% and 0.02% of the whole
process respectively with and without this patchset across 3 runs.
2. I used perf to collect stats of dd from /dev/random to a 40GB tmpfs
file and find get_pfnblock_flags_mask() takes about 0.05% of the
process with and without this patchset across 3 runs.
This patch (of 6):
No functional change is intended.
1. Add __NR_PAGEBLOCK_BITS for the number of pageblock flag bits and use
roundup_pow_of_two(__NR_PAGEBLOCK_BITS) as NR_PAGEBLOCK_BITS to take
right amount of bits for pageblock flags.
2. Rename PB_migrate_skip to PB_compact_skip.
3. Add {get,set,clear}_pfnblock_bit() to operate one a standalone bit,
like PB_compact_skip.
3. Make {get,set}_pfnblock_flags_mask() internal functions and use
{get,set}_pfnblock_migratetype() for pageblock migratetype operations.
4. Move pageblock flags common code to get_pfnblock_bitmap_bitidx().
3. Use MIGRATETYPE_MASK to get the migratetype of a pageblock from its
flags.
4. Use PB_migrate_end in the definition of MIGRATETYPE_MASK instead of
PB_migrate_bits.
5. Add a comment on is_migrate_cma_folio() to prevent one from changing it
to use get_pageblock_migratetype() and causing issues.
Link: https://lkml.kernel.org/r/20250617021115.2331563-1-ziy@nvidia.com
Link: https://lkml.kernel.org/r/20250617021115.2331563-2-ziy@nvidia.com
Signed-off-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kirill A. Shuemov <kirill.shutemov@linux.intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Richard Chang <richardycc@google.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-06-16 22:11:09 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-06-16 22:11:10 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/page_alloc: pageblock flags functions clean up
Patch series "Make MIGRATE_ISOLATE a standalone bit", v10.
This patchset moves MIGRATE_ISOLATE to a standalone bit to avoid being
overwritten during pageblock isolation process. Currently,
MIGRATE_ISOLATE is part of enum migratetype (in include/linux/mmzone.h),
thus, setting a pageblock to MIGRATE_ISOLATE overwrites its original
migratetype. This causes pageblock migratetype loss during
alloc_contig_range() and memory offline, especially when the process fails
due to a failed pageblock isolation and the code tries to undo the
finished pageblock isolations.
In terms of performance for changing pageblock types, no performance
change is observed:
1. I used perf to collect stats of offlining and onlining all memory
of a 40GB VM 10 times and see that get_pfnblock_flags_mask() and
set_pfnblock_flags_mask() take about 0.12% and 0.02% of the whole
process respectively with and without this patchset across 3 runs.
2. I used perf to collect stats of dd from /dev/random to a 40GB tmpfs
file and find get_pfnblock_flags_mask() takes about 0.05% of the
process with and without this patchset across 3 runs.
This patch (of 6):
No functional change is intended.
1. Add __NR_PAGEBLOCK_BITS for the number of pageblock flag bits and use
roundup_pow_of_two(__NR_PAGEBLOCK_BITS) as NR_PAGEBLOCK_BITS to take
right amount of bits for pageblock flags.
2. Rename PB_migrate_skip to PB_compact_skip.
3. Add {get,set,clear}_pfnblock_bit() to operate one a standalone bit,
like PB_compact_skip.
3. Make {get,set}_pfnblock_flags_mask() internal functions and use
{get,set}_pfnblock_migratetype() for pageblock migratetype operations.
4. Move pageblock flags common code to get_pfnblock_bitmap_bitidx().
3. Use MIGRATETYPE_MASK to get the migratetype of a pageblock from its
flags.
4. Use PB_migrate_end in the definition of MIGRATETYPE_MASK instead of
PB_migrate_bits.
5. Add a comment on is_migrate_cma_folio() to prevent one from changing it
to use get_pageblock_migratetype() and causing issues.
Link: https://lkml.kernel.org/r/20250617021115.2331563-1-ziy@nvidia.com
Link: https://lkml.kernel.org/r/20250617021115.2331563-2-ziy@nvidia.com
Signed-off-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kirill A. Shuemov <kirill.shutemov@linux.intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Richard Chang <richardycc@google.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-06-16 22:11:09 -04:00
|
|
|
|
2025-06-16 22:11:10 -04:00
|
|
|
|
2025-08-27 07:01:05 +00:00
|
|
|
|
mm/page_alloc: pageblock flags functions clean up
Patch series "Make MIGRATE_ISOLATE a standalone bit", v10.
This patchset moves MIGRATE_ISOLATE to a standalone bit to avoid being
overwritten during pageblock isolation process. Currently,
MIGRATE_ISOLATE is part of enum migratetype (in include/linux/mmzone.h),
thus, setting a pageblock to MIGRATE_ISOLATE overwrites its original
migratetype. This causes pageblock migratetype loss during
alloc_contig_range() and memory offline, especially when the process fails
due to a failed pageblock isolation and the code tries to undo the
finished pageblock isolations.
In terms of performance for changing pageblock types, no performance
change is observed:
1. I used perf to collect stats of offlining and onlining all memory
of a 40GB VM 10 times and see that get_pfnblock_flags_mask() and
set_pfnblock_flags_mask() take about 0.12% and 0.02% of the whole
process respectively with and without this patchset across 3 runs.
2. I used perf to collect stats of dd from /dev/random to a 40GB tmpfs
file and find get_pfnblock_flags_mask() takes about 0.05% of the
process with and without this patchset across 3 runs.
This patch (of 6):
No functional change is intended.
1. Add __NR_PAGEBLOCK_BITS for the number of pageblock flag bits and use
roundup_pow_of_two(__NR_PAGEBLOCK_BITS) as NR_PAGEBLOCK_BITS to take
right amount of bits for pageblock flags.
2. Rename PB_migrate_skip to PB_compact_skip.
3. Add {get,set,clear}_pfnblock_bit() to operate one a standalone bit,
like PB_compact_skip.
3. Make {get,set}_pfnblock_flags_mask() internal functions and use
{get,set}_pfnblock_migratetype() for pageblock migratetype operations.
4. Move pageblock flags common code to get_pfnblock_bitmap_bitidx().
3. Use MIGRATETYPE_MASK to get the migratetype of a pageblock from its
flags.
4. Use PB_migrate_end in the definition of MIGRATETYPE_MASK instead of
PB_migrate_bits.
5. Add a comment on is_migrate_cma_folio() to prevent one from changing it
to use get_pageblock_migratetype() and causing issues.
Link: https://lkml.kernel.org/r/20250617021115.2331563-1-ziy@nvidia.com
Link: https://lkml.kernel.org/r/20250617021115.2331563-2-ziy@nvidia.com
Signed-off-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kirill A. Shuemov <kirill.shutemov@linux.intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Richard Chang <richardycc@google.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-06-16 22:11:09 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-08-11 19:59:44 +08:00
|
|
|
|
mm/page_alloc: pageblock flags functions clean up
Patch series "Make MIGRATE_ISOLATE a standalone bit", v10.
This patchset moves MIGRATE_ISOLATE to a standalone bit to avoid being
overwritten during pageblock isolation process. Currently,
MIGRATE_ISOLATE is part of enum migratetype (in include/linux/mmzone.h),
thus, setting a pageblock to MIGRATE_ISOLATE overwrites its original
migratetype. This causes pageblock migratetype loss during
alloc_contig_range() and memory offline, especially when the process fails
due to a failed pageblock isolation and the code tries to undo the
finished pageblock isolations.
In terms of performance for changing pageblock types, no performance
change is observed:
1. I used perf to collect stats of offlining and onlining all memory
of a 40GB VM 10 times and see that get_pfnblock_flags_mask() and
set_pfnblock_flags_mask() take about 0.12% and 0.02% of the whole
process respectively with and without this patchset across 3 runs.
2. I used perf to collect stats of dd from /dev/random to a 40GB tmpfs
file and find get_pfnblock_flags_mask() takes about 0.05% of the
process with and without this patchset across 3 runs.
This patch (of 6):
No functional change is intended.
1. Add __NR_PAGEBLOCK_BITS for the number of pageblock flag bits and use
roundup_pow_of_two(__NR_PAGEBLOCK_BITS) as NR_PAGEBLOCK_BITS to take
right amount of bits for pageblock flags.
2. Rename PB_migrate_skip to PB_compact_skip.
3. Add {get,set,clear}_pfnblock_bit() to operate one a standalone bit,
like PB_compact_skip.
3. Make {get,set}_pfnblock_flags_mask() internal functions and use
{get,set}_pfnblock_migratetype() for pageblock migratetype operations.
4. Move pageblock flags common code to get_pfnblock_bitmap_bitidx().
3. Use MIGRATETYPE_MASK to get the migratetype of a pageblock from its
flags.
4. Use PB_migrate_end in the definition of MIGRATETYPE_MASK instead of
PB_migrate_bits.
5. Add a comment on is_migrate_cma_folio() to prevent one from changing it
to use get_pageblock_migratetype() and causing issues.
Link: https://lkml.kernel.org/r/20250617021115.2331563-1-ziy@nvidia.com
Link: https://lkml.kernel.org/r/20250617021115.2331563-2-ziy@nvidia.com
Signed-off-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kirill A. Shuemov <kirill.shutemov@linux.intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Richard Chang <richardycc@google.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-06-16 22:11:09 -04:00
|
|
|
|
|
|
|
|
|
2023-08-11 19:59:44 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/page_alloc: pageblock flags functions clean up
Patch series "Make MIGRATE_ISOLATE a standalone bit", v10.
This patchset moves MIGRATE_ISOLATE to a standalone bit to avoid being
overwritten during pageblock isolation process. Currently,
MIGRATE_ISOLATE is part of enum migratetype (in include/linux/mmzone.h),
thus, setting a pageblock to MIGRATE_ISOLATE overwrites its original
migratetype. This causes pageblock migratetype loss during
alloc_contig_range() and memory offline, especially when the process fails
due to a failed pageblock isolation and the code tries to undo the
finished pageblock isolations.
In terms of performance for changing pageblock types, no performance
change is observed:
1. I used perf to collect stats of offlining and onlining all memory
of a 40GB VM 10 times and see that get_pfnblock_flags_mask() and
set_pfnblock_flags_mask() take about 0.12% and 0.02% of the whole
process respectively with and without this patchset across 3 runs.
2. I used perf to collect stats of dd from /dev/random to a 40GB tmpfs
file and find get_pfnblock_flags_mask() takes about 0.05% of the
process with and without this patchset across 3 runs.
This patch (of 6):
No functional change is intended.
1. Add __NR_PAGEBLOCK_BITS for the number of pageblock flag bits and use
roundup_pow_of_two(__NR_PAGEBLOCK_BITS) as NR_PAGEBLOCK_BITS to take
right amount of bits for pageblock flags.
2. Rename PB_migrate_skip to PB_compact_skip.
3. Add {get,set,clear}_pfnblock_bit() to operate one a standalone bit,
like PB_compact_skip.
3. Make {get,set}_pfnblock_flags_mask() internal functions and use
{get,set}_pfnblock_migratetype() for pageblock migratetype operations.
4. Move pageblock flags common code to get_pfnblock_bitmap_bitidx().
3. Use MIGRATETYPE_MASK to get the migratetype of a pageblock from its
flags.
4. Use PB_migrate_end in the definition of MIGRATETYPE_MASK instead of
PB_migrate_bits.
5. Add a comment on is_migrate_cma_folio() to prevent one from changing it
to use get_pageblock_migratetype() and causing issues.
Link: https://lkml.kernel.org/r/20250617021115.2331563-1-ziy@nvidia.com
Link: https://lkml.kernel.org/r/20250617021115.2331563-2-ziy@nvidia.com
Signed-off-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kirill A. Shuemov <kirill.shutemov@linux.intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Richard Chang <richardycc@google.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-06-16 22:11:09 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-05-19 17:14:27 -07:00
|
|
|
|
mm/page_alloc: pageblock flags functions clean up
Patch series "Make MIGRATE_ISOLATE a standalone bit", v10.
This patchset moves MIGRATE_ISOLATE to a standalone bit to avoid being
overwritten during pageblock isolation process. Currently,
MIGRATE_ISOLATE is part of enum migratetype (in include/linux/mmzone.h),
thus, setting a pageblock to MIGRATE_ISOLATE overwrites its original
migratetype. This causes pageblock migratetype loss during
alloc_contig_range() and memory offline, especially when the process fails
due to a failed pageblock isolation and the code tries to undo the
finished pageblock isolations.
In terms of performance for changing pageblock types, no performance
change is observed:
1. I used perf to collect stats of offlining and onlining all memory
of a 40GB VM 10 times and see that get_pfnblock_flags_mask() and
set_pfnblock_flags_mask() take about 0.12% and 0.02% of the whole
process respectively with and without this patchset across 3 runs.
2. I used perf to collect stats of dd from /dev/random to a 40GB tmpfs
file and find get_pfnblock_flags_mask() takes about 0.05% of the
process with and without this patchset across 3 runs.
This patch (of 6):
No functional change is intended.
1. Add __NR_PAGEBLOCK_BITS for the number of pageblock flag bits and use
roundup_pow_of_two(__NR_PAGEBLOCK_BITS) as NR_PAGEBLOCK_BITS to take
right amount of bits for pageblock flags.
2. Rename PB_migrate_skip to PB_compact_skip.
3. Add {get,set,clear}_pfnblock_bit() to operate one a standalone bit,
like PB_compact_skip.
3. Make {get,set}_pfnblock_flags_mask() internal functions and use
{get,set}_pfnblock_migratetype() for pageblock migratetype operations.
4. Move pageblock flags common code to get_pfnblock_bitmap_bitidx().
3. Use MIGRATETYPE_MASK to get the migratetype of a pageblock from its
flags.
4. Use PB_migrate_end in the definition of MIGRATETYPE_MASK instead of
PB_migrate_bits.
5. Add a comment on is_migrate_cma_folio() to prevent one from changing it
to use get_pageblock_migratetype() and causing issues.
Link: https://lkml.kernel.org/r/20250617021115.2331563-1-ziy@nvidia.com
Link: https://lkml.kernel.org/r/20250617021115.2331563-2-ziy@nvidia.com
Signed-off-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kirill A. Shuemov <kirill.shutemov@linux.intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Richard Chang <richardycc@google.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-06-16 22:11:09 -04:00
|
|
|
|
|
|
|
|
|
2016-05-19 17:14:27 -07:00
|
|
|
|
|
|
|
|
|
mm/page_alloc: pageblock flags functions clean up
Patch series "Make MIGRATE_ISOLATE a standalone bit", v10.
This patchset moves MIGRATE_ISOLATE to a standalone bit to avoid being
overwritten during pageblock isolation process. Currently,
MIGRATE_ISOLATE is part of enum migratetype (in include/linux/mmzone.h),
thus, setting a pageblock to MIGRATE_ISOLATE overwrites its original
migratetype. This causes pageblock migratetype loss during
alloc_contig_range() and memory offline, especially when the process fails
due to a failed pageblock isolation and the code tries to undo the
finished pageblock isolations.
In terms of performance for changing pageblock types, no performance
change is observed:
1. I used perf to collect stats of offlining and onlining all memory
of a 40GB VM 10 times and see that get_pfnblock_flags_mask() and
set_pfnblock_flags_mask() take about 0.12% and 0.02% of the whole
process respectively with and without this patchset across 3 runs.
2. I used perf to collect stats of dd from /dev/random to a 40GB tmpfs
file and find get_pfnblock_flags_mask() takes about 0.05% of the
process with and without this patchset across 3 runs.
This patch (of 6):
No functional change is intended.
1. Add __NR_PAGEBLOCK_BITS for the number of pageblock flag bits and use
roundup_pow_of_two(__NR_PAGEBLOCK_BITS) as NR_PAGEBLOCK_BITS to take
right amount of bits for pageblock flags.
2. Rename PB_migrate_skip to PB_compact_skip.
3. Add {get,set,clear}_pfnblock_bit() to operate one a standalone bit,
like PB_compact_skip.
3. Make {get,set}_pfnblock_flags_mask() internal functions and use
{get,set}_pfnblock_migratetype() for pageblock migratetype operations.
4. Move pageblock flags common code to get_pfnblock_bitmap_bitidx().
3. Use MIGRATETYPE_MASK to get the migratetype of a pageblock from its
flags.
4. Use PB_migrate_end in the definition of MIGRATETYPE_MASK instead of
PB_migrate_bits.
5. Add a comment on is_migrate_cma_folio() to prevent one from changing it
to use get_pageblock_migratetype() and causing issues.
Link: https://lkml.kernel.org/r/20250617021115.2331563-1-ziy@nvidia.com
Link: https://lkml.kernel.org/r/20250617021115.2331563-2-ziy@nvidia.com
Signed-off-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kirill A. Shuemov <kirill.shutemov@linux.intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Richard Chang <richardycc@google.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-06-16 22:11:09 -04:00
|
|
|
|
2022-05-24 10:15:25 -07:00
|
|
|
|
mm/page_alloc: pageblock flags functions clean up
Patch series "Make MIGRATE_ISOLATE a standalone bit", v10.
This patchset moves MIGRATE_ISOLATE to a standalone bit to avoid being
overwritten during pageblock isolation process. Currently,
MIGRATE_ISOLATE is part of enum migratetype (in include/linux/mmzone.h),
thus, setting a pageblock to MIGRATE_ISOLATE overwrites its original
migratetype. This causes pageblock migratetype loss during
alloc_contig_range() and memory offline, especially when the process fails
due to a failed pageblock isolation and the code tries to undo the
finished pageblock isolations.
In terms of performance for changing pageblock types, no performance
change is observed:
1. I used perf to collect stats of offlining and onlining all memory
of a 40GB VM 10 times and see that get_pfnblock_flags_mask() and
set_pfnblock_flags_mask() take about 0.12% and 0.02% of the whole
process respectively with and without this patchset across 3 runs.
2. I used perf to collect stats of dd from /dev/random to a 40GB tmpfs
file and find get_pfnblock_flags_mask() takes about 0.05% of the
process with and without this patchset across 3 runs.
This patch (of 6):
No functional change is intended.
1. Add __NR_PAGEBLOCK_BITS for the number of pageblock flag bits and use
roundup_pow_of_two(__NR_PAGEBLOCK_BITS) as NR_PAGEBLOCK_BITS to take
right amount of bits for pageblock flags.
2. Rename PB_migrate_skip to PB_compact_skip.
3. Add {get,set,clear}_pfnblock_bit() to operate one a standalone bit,
like PB_compact_skip.
3. Make {get,set}_pfnblock_flags_mask() internal functions and use
{get,set}_pfnblock_migratetype() for pageblock migratetype operations.
4. Move pageblock flags common code to get_pfnblock_bitmap_bitidx().
3. Use MIGRATETYPE_MASK to get the migratetype of a pageblock from its
flags.
4. Use PB_migrate_end in the definition of MIGRATETYPE_MASK instead of
PB_migrate_bits.
5. Add a comment on is_migrate_cma_folio() to prevent one from changing it
to use get_pageblock_migratetype() and causing issues.
Link: https://lkml.kernel.org/r/20250617021115.2331563-1-ziy@nvidia.com
Link: https://lkml.kernel.org/r/20250617021115.2331563-2-ziy@nvidia.com
Signed-off-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kirill A. Shuemov <kirill.shutemov@linux.intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Richard Chang <richardycc@google.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-06-16 22:11:09 -04:00
|
|
|
|
2022-05-24 10:15:25 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/page_alloc: pageblock flags functions clean up
Patch series "Make MIGRATE_ISOLATE a standalone bit", v10.
This patchset moves MIGRATE_ISOLATE to a standalone bit to avoid being
overwritten during pageblock isolation process. Currently,
MIGRATE_ISOLATE is part of enum migratetype (in include/linux/mmzone.h),
thus, setting a pageblock to MIGRATE_ISOLATE overwrites its original
migratetype. This causes pageblock migratetype loss during
alloc_contig_range() and memory offline, especially when the process fails
due to a failed pageblock isolation and the code tries to undo the
finished pageblock isolations.
In terms of performance for changing pageblock types, no performance
change is observed:
1. I used perf to collect stats of offlining and onlining all memory
of a 40GB VM 10 times and see that get_pfnblock_flags_mask() and
set_pfnblock_flags_mask() take about 0.12% and 0.02% of the whole
process respectively with and without this patchset across 3 runs.
2. I used perf to collect stats of dd from /dev/random to a 40GB tmpfs
file and find get_pfnblock_flags_mask() takes about 0.05% of the
process with and without this patchset across 3 runs.
This patch (of 6):
No functional change is intended.
1. Add __NR_PAGEBLOCK_BITS for the number of pageblock flag bits and use
roundup_pow_of_two(__NR_PAGEBLOCK_BITS) as NR_PAGEBLOCK_BITS to take
right amount of bits for pageblock flags.
2. Rename PB_migrate_skip to PB_compact_skip.
3. Add {get,set,clear}_pfnblock_bit() to operate one a standalone bit,
like PB_compact_skip.
3. Make {get,set}_pfnblock_flags_mask() internal functions and use
{get,set}_pfnblock_migratetype() for pageblock migratetype operations.
4. Move pageblock flags common code to get_pfnblock_bitmap_bitidx().
3. Use MIGRATETYPE_MASK to get the migratetype of a pageblock from its
flags.
4. Use PB_migrate_end in the definition of MIGRATETYPE_MASK instead of
PB_migrate_bits.
5. Add a comment on is_migrate_cma_folio() to prevent one from changing it
to use get_pageblock_migratetype() and causing issues.
Link: https://lkml.kernel.org/r/20250617021115.2331563-1-ziy@nvidia.com
Link: https://lkml.kernel.org/r/20250617021115.2331563-2-ziy@nvidia.com
Signed-off-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kirill A. Shuemov <kirill.shutemov@linux.intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Richard Chang <richardycc@google.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-06-16 22:11:09 -04:00
|
|
|
|
2020-08-06 23:25:48 -07:00
|
|
|
|
2016-05-19 17:14:27 -07:00
|
|
|
|
|
|
|
|
|
mm/page_alloc: pageblock flags functions clean up
Patch series "Make MIGRATE_ISOLATE a standalone bit", v10.
This patchset moves MIGRATE_ISOLATE to a standalone bit to avoid being
overwritten during pageblock isolation process. Currently,
MIGRATE_ISOLATE is part of enum migratetype (in include/linux/mmzone.h),
thus, setting a pageblock to MIGRATE_ISOLATE overwrites its original
migratetype. This causes pageblock migratetype loss during
alloc_contig_range() and memory offline, especially when the process fails
due to a failed pageblock isolation and the code tries to undo the
finished pageblock isolations.
In terms of performance for changing pageblock types, no performance
change is observed:
1. I used perf to collect stats of offlining and onlining all memory
of a 40GB VM 10 times and see that get_pfnblock_flags_mask() and
set_pfnblock_flags_mask() take about 0.12% and 0.02% of the whole
process respectively with and without this patchset across 3 runs.
2. I used perf to collect stats of dd from /dev/random to a 40GB tmpfs
file and find get_pfnblock_flags_mask() takes about 0.05% of the
process with and without this patchset across 3 runs.
This patch (of 6):
No functional change is intended.
1. Add __NR_PAGEBLOCK_BITS for the number of pageblock flag bits and use
roundup_pow_of_two(__NR_PAGEBLOCK_BITS) as NR_PAGEBLOCK_BITS to take
right amount of bits for pageblock flags.
2. Rename PB_migrate_skip to PB_compact_skip.
3. Add {get,set,clear}_pfnblock_bit() to operate one a standalone bit,
like PB_compact_skip.
3. Make {get,set}_pfnblock_flags_mask() internal functions and use
{get,set}_pfnblock_migratetype() for pageblock migratetype operations.
4. Move pageblock flags common code to get_pfnblock_bitmap_bitidx().
3. Use MIGRATETYPE_MASK to get the migratetype of a pageblock from its
flags.
4. Use PB_migrate_end in the definition of MIGRATETYPE_MASK instead of
PB_migrate_bits.
5. Add a comment on is_migrate_cma_folio() to prevent one from changing it
to use get_pageblock_migratetype() and causing issues.
Link: https://lkml.kernel.org/r/20250617021115.2331563-1-ziy@nvidia.com
Link: https://lkml.kernel.org/r/20250617021115.2331563-2-ziy@nvidia.com
Signed-off-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kirill A. Shuemov <kirill.shutemov@linux.intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Richard Chang <richardycc@google.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-06-16 22:11:09 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-05-19 17:14:27 -07:00
|
|
|
|
mm/page_alloc: pageblock flags functions clean up
Patch series "Make MIGRATE_ISOLATE a standalone bit", v10.
This patchset moves MIGRATE_ISOLATE to a standalone bit to avoid being
overwritten during pageblock isolation process. Currently,
MIGRATE_ISOLATE is part of enum migratetype (in include/linux/mmzone.h),
thus, setting a pageblock to MIGRATE_ISOLATE overwrites its original
migratetype. This causes pageblock migratetype loss during
alloc_contig_range() and memory offline, especially when the process fails
due to a failed pageblock isolation and the code tries to undo the
finished pageblock isolations.
In terms of performance for changing pageblock types, no performance
change is observed:
1. I used perf to collect stats of offlining and onlining all memory
of a 40GB VM 10 times and see that get_pfnblock_flags_mask() and
set_pfnblock_flags_mask() take about 0.12% and 0.02% of the whole
process respectively with and without this patchset across 3 runs.
2. I used perf to collect stats of dd from /dev/random to a 40GB tmpfs
file and find get_pfnblock_flags_mask() takes about 0.05% of the
process with and without this patchset across 3 runs.
This patch (of 6):
No functional change is intended.
1. Add __NR_PAGEBLOCK_BITS for the number of pageblock flag bits and use
roundup_pow_of_two(__NR_PAGEBLOCK_BITS) as NR_PAGEBLOCK_BITS to take
right amount of bits for pageblock flags.
2. Rename PB_migrate_skip to PB_compact_skip.
3. Add {get,set,clear}_pfnblock_bit() to operate one a standalone bit,
like PB_compact_skip.
3. Make {get,set}_pfnblock_flags_mask() internal functions and use
{get,set}_pfnblock_migratetype() for pageblock migratetype operations.
4. Move pageblock flags common code to get_pfnblock_bitmap_bitidx().
3. Use MIGRATETYPE_MASK to get the migratetype of a pageblock from its
flags.
4. Use PB_migrate_end in the definition of MIGRATETYPE_MASK instead of
PB_migrate_bits.
5. Add a comment on is_migrate_cma_folio() to prevent one from changing it
to use get_pageblock_migratetype() and causing issues.
Link: https://lkml.kernel.org/r/20250617021115.2331563-1-ziy@nvidia.com
Link: https://lkml.kernel.org/r/20250617021115.2331563-2-ziy@nvidia.com
Signed-off-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kirill A. Shuemov <kirill.shutemov@linux.intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Richard Chang <richardycc@google.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-06-16 22:11:09 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-05-19 17:14:27 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/page_alloc: pageblock flags functions clean up
Patch series "Make MIGRATE_ISOLATE a standalone bit", v10.
This patchset moves MIGRATE_ISOLATE to a standalone bit to avoid being
overwritten during pageblock isolation process. Currently,
MIGRATE_ISOLATE is part of enum migratetype (in include/linux/mmzone.h),
thus, setting a pageblock to MIGRATE_ISOLATE overwrites its original
migratetype. This causes pageblock migratetype loss during
alloc_contig_range() and memory offline, especially when the process fails
due to a failed pageblock isolation and the code tries to undo the
finished pageblock isolations.
In terms of performance for changing pageblock types, no performance
change is observed:
1. I used perf to collect stats of offlining and onlining all memory
of a 40GB VM 10 times and see that get_pfnblock_flags_mask() and
set_pfnblock_flags_mask() take about 0.12% and 0.02% of the whole
process respectively with and without this patchset across 3 runs.
2. I used perf to collect stats of dd from /dev/random to a 40GB tmpfs
file and find get_pfnblock_flags_mask() takes about 0.05% of the
process with and without this patchset across 3 runs.
This patch (of 6):
No functional change is intended.
1. Add __NR_PAGEBLOCK_BITS for the number of pageblock flag bits and use
roundup_pow_of_two(__NR_PAGEBLOCK_BITS) as NR_PAGEBLOCK_BITS to take
right amount of bits for pageblock flags.
2. Rename PB_migrate_skip to PB_compact_skip.
3. Add {get,set,clear}_pfnblock_bit() to operate one a standalone bit,
like PB_compact_skip.
3. Make {get,set}_pfnblock_flags_mask() internal functions and use
{get,set}_pfnblock_migratetype() for pageblock migratetype operations.
4. Move pageblock flags common code to get_pfnblock_bitmap_bitidx().
3. Use MIGRATETYPE_MASK to get the migratetype of a pageblock from its
flags.
4. Use PB_migrate_end in the definition of MIGRATETYPE_MASK instead of
PB_migrate_bits.
5. Add a comment on is_migrate_cma_folio() to prevent one from changing it
to use get_pageblock_migratetype() and causing issues.
Link: https://lkml.kernel.org/r/20250617021115.2331563-1-ziy@nvidia.com
Link: https://lkml.kernel.org/r/20250617021115.2331563-2-ziy@nvidia.com
Signed-off-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kirill A. Shuemov <kirill.shutemov@linux.intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Richard Chang <richardycc@google.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-06-16 22:11:09 -04:00
|
|
|
|
2016-05-19 17:14:27 -07:00
|
|
|
|
|
|
|
|
|
mm/page_alloc: pageblock flags functions clean up
Patch series "Make MIGRATE_ISOLATE a standalone bit", v10.
This patchset moves MIGRATE_ISOLATE to a standalone bit to avoid being
overwritten during pageblock isolation process. Currently,
MIGRATE_ISOLATE is part of enum migratetype (in include/linux/mmzone.h),
thus, setting a pageblock to MIGRATE_ISOLATE overwrites its original
migratetype. This causes pageblock migratetype loss during
alloc_contig_range() and memory offline, especially when the process fails
due to a failed pageblock isolation and the code tries to undo the
finished pageblock isolations.
In terms of performance for changing pageblock types, no performance
change is observed:
1. I used perf to collect stats of offlining and onlining all memory
of a 40GB VM 10 times and see that get_pfnblock_flags_mask() and
set_pfnblock_flags_mask() take about 0.12% and 0.02% of the whole
process respectively with and without this patchset across 3 runs.
2. I used perf to collect stats of dd from /dev/random to a 40GB tmpfs
file and find get_pfnblock_flags_mask() takes about 0.05% of the
process with and without this patchset across 3 runs.
This patch (of 6):
No functional change is intended.
1. Add __NR_PAGEBLOCK_BITS for the number of pageblock flag bits and use
roundup_pow_of_two(__NR_PAGEBLOCK_BITS) as NR_PAGEBLOCK_BITS to take
right amount of bits for pageblock flags.
2. Rename PB_migrate_skip to PB_compact_skip.
3. Add {get,set,clear}_pfnblock_bit() to operate one a standalone bit,
like PB_compact_skip.
3. Make {get,set}_pfnblock_flags_mask() internal functions and use
{get,set}_pfnblock_migratetype() for pageblock migratetype operations.
4. Move pageblock flags common code to get_pfnblock_bitmap_bitidx().
3. Use MIGRATETYPE_MASK to get the migratetype of a pageblock from its
flags.
4. Use PB_migrate_end in the definition of MIGRATETYPE_MASK instead of
PB_migrate_bits.
5. Add a comment on is_migrate_cma_folio() to prevent one from changing it
to use get_pageblock_migratetype() and causing issues.
Link: https://lkml.kernel.org/r/20250617021115.2331563-1-ziy@nvidia.com
Link: https://lkml.kernel.org/r/20250617021115.2331563-2-ziy@nvidia.com
Signed-off-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kirill A. Shuemov <kirill.shutemov@linux.intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Richard Chang <richardycc@google.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-06-16 22:11:09 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-06-16 22:11:10 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/page_alloc: pageblock flags functions clean up
Patch series "Make MIGRATE_ISOLATE a standalone bit", v10.
This patchset moves MIGRATE_ISOLATE to a standalone bit to avoid being
overwritten during pageblock isolation process. Currently,
MIGRATE_ISOLATE is part of enum migratetype (in include/linux/mmzone.h),
thus, setting a pageblock to MIGRATE_ISOLATE overwrites its original
migratetype. This causes pageblock migratetype loss during
alloc_contig_range() and memory offline, especially when the process fails
due to a failed pageblock isolation and the code tries to undo the
finished pageblock isolations.
In terms of performance for changing pageblock types, no performance
change is observed:
1. I used perf to collect stats of offlining and onlining all memory
of a 40GB VM 10 times and see that get_pfnblock_flags_mask() and
set_pfnblock_flags_mask() take about 0.12% and 0.02% of the whole
process respectively with and without this patchset across 3 runs.
2. I used perf to collect stats of dd from /dev/random to a 40GB tmpfs
file and find get_pfnblock_flags_mask() takes about 0.05% of the
process with and without this patchset across 3 runs.
This patch (of 6):
No functional change is intended.
1. Add __NR_PAGEBLOCK_BITS for the number of pageblock flag bits and use
roundup_pow_of_two(__NR_PAGEBLOCK_BITS) as NR_PAGEBLOCK_BITS to take
right amount of bits for pageblock flags.
2. Rename PB_migrate_skip to PB_compact_skip.
3. Add {get,set,clear}_pfnblock_bit() to operate one a standalone bit,
like PB_compact_skip.
3. Make {get,set}_pfnblock_flags_mask() internal functions and use
{get,set}_pfnblock_migratetype() for pageblock migratetype operations.
4. Move pageblock flags common code to get_pfnblock_bitmap_bitidx().
3. Use MIGRATETYPE_MASK to get the migratetype of a pageblock from its
flags.
4. Use PB_migrate_end in the definition of MIGRATETYPE_MASK instead of
PB_migrate_bits.
5. Add a comment on is_migrate_cma_folio() to prevent one from changing it
to use get_pageblock_migratetype() and causing issues.
Link: https://lkml.kernel.org/r/20250617021115.2331563-1-ziy@nvidia.com
Link: https://lkml.kernel.org/r/20250617021115.2331563-2-ziy@nvidia.com
Signed-off-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kirill A. Shuemov <kirill.shutemov@linux.intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Richard Chang <richardycc@google.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-06-16 22:11:09 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-05-19 17:14:27 -07:00
|
|
|
|
|
|
|
|
|
mm/page_alloc: pageblock flags functions clean up
Patch series "Make MIGRATE_ISOLATE a standalone bit", v10.
This patchset moves MIGRATE_ISOLATE to a standalone bit to avoid being
overwritten during pageblock isolation process. Currently,
MIGRATE_ISOLATE is part of enum migratetype (in include/linux/mmzone.h),
thus, setting a pageblock to MIGRATE_ISOLATE overwrites its original
migratetype. This causes pageblock migratetype loss during
alloc_contig_range() and memory offline, especially when the process fails
due to a failed pageblock isolation and the code tries to undo the
finished pageblock isolations.
In terms of performance for changing pageblock types, no performance
change is observed:
1. I used perf to collect stats of offlining and onlining all memory
of a 40GB VM 10 times and see that get_pfnblock_flags_mask() and
set_pfnblock_flags_mask() take about 0.12% and 0.02% of the whole
process respectively with and without this patchset across 3 runs.
2. I used perf to collect stats of dd from /dev/random to a 40GB tmpfs
file and find get_pfnblock_flags_mask() takes about 0.05% of the
process with and without this patchset across 3 runs.
This patch (of 6):
No functional change is intended.
1. Add __NR_PAGEBLOCK_BITS for the number of pageblock flag bits and use
roundup_pow_of_two(__NR_PAGEBLOCK_BITS) as NR_PAGEBLOCK_BITS to take
right amount of bits for pageblock flags.
2. Rename PB_migrate_skip to PB_compact_skip.
3. Add {get,set,clear}_pfnblock_bit() to operate one a standalone bit,
like PB_compact_skip.
3. Make {get,set}_pfnblock_flags_mask() internal functions and use
{get,set}_pfnblock_migratetype() for pageblock migratetype operations.
4. Move pageblock flags common code to get_pfnblock_bitmap_bitidx().
3. Use MIGRATETYPE_MASK to get the migratetype of a pageblock from its
flags.
4. Use PB_migrate_end in the definition of MIGRATETYPE_MASK instead of
PB_migrate_bits.
5. Add a comment on is_migrate_cma_folio() to prevent one from changing it
to use get_pageblock_migratetype() and causing issues.
Link: https://lkml.kernel.org/r/20250617021115.2331563-1-ziy@nvidia.com
Link: https://lkml.kernel.org/r/20250617021115.2331563-2-ziy@nvidia.com
Signed-off-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kirill A. Shuemov <kirill.shutemov@linux.intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Richard Chang <richardycc@google.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-06-16 22:11:09 -04:00
|
|
|
|
|
|
|
|
|
2016-05-19 17:14:27 -07:00
|
|
|
|
mm/page_alloc: pageblock flags functions clean up
Patch series "Make MIGRATE_ISOLATE a standalone bit", v10.
This patchset moves MIGRATE_ISOLATE to a standalone bit to avoid being
overwritten during pageblock isolation process. Currently,
MIGRATE_ISOLATE is part of enum migratetype (in include/linux/mmzone.h),
thus, setting a pageblock to MIGRATE_ISOLATE overwrites its original
migratetype. This causes pageblock migratetype loss during
alloc_contig_range() and memory offline, especially when the process fails
due to a failed pageblock isolation and the code tries to undo the
finished pageblock isolations.
In terms of performance for changing pageblock types, no performance
change is observed:
1. I used perf to collect stats of offlining and onlining all memory
of a 40GB VM 10 times and see that get_pfnblock_flags_mask() and
set_pfnblock_flags_mask() take about 0.12% and 0.02% of the whole
process respectively with and without this patchset across 3 runs.
2. I used perf to collect stats of dd from /dev/random to a 40GB tmpfs
file and find get_pfnblock_flags_mask() takes about 0.05% of the
process with and without this patchset across 3 runs.
This patch (of 6):
No functional change is intended.
1. Add __NR_PAGEBLOCK_BITS for the number of pageblock flag bits and use
roundup_pow_of_two(__NR_PAGEBLOCK_BITS) as NR_PAGEBLOCK_BITS to take
right amount of bits for pageblock flags.
2. Rename PB_migrate_skip to PB_compact_skip.
3. Add {get,set,clear}_pfnblock_bit() to operate one a standalone bit,
like PB_compact_skip.
3. Make {get,set}_pfnblock_flags_mask() internal functions and use
{get,set}_pfnblock_migratetype() for pageblock migratetype operations.
4. Move pageblock flags common code to get_pfnblock_bitmap_bitidx().
3. Use MIGRATETYPE_MASK to get the migratetype of a pageblock from its
flags.
4. Use PB_migrate_end in the definition of MIGRATETYPE_MASK instead of
PB_migrate_bits.
5. Add a comment on is_migrate_cma_folio() to prevent one from changing it
to use get_pageblock_migratetype() and causing issues.
Link: https://lkml.kernel.org/r/20250617021115.2331563-1-ziy@nvidia.com
Link: https://lkml.kernel.org/r/20250617021115.2331563-2-ziy@nvidia.com
Signed-off-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kirill A. Shuemov <kirill.shutemov@linux.intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Richard Chang <richardycc@google.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-06-16 22:11:09 -04:00
|
|
|
|
|
|
|
|
|
mm/page_alloc: use try_cmpxchg in set_pfnblock_flags_mask
Use try_cmpxchg instead of cmpxchg in set_pfnblock_flags_mask. x86
CMPXCHG instruction returns success in ZF flag, so this change saves a
compare after cmpxchg (and related move instruction in front of cmpxchg).
The main loop improves from:
1c5d: 48 89 c2 mov %rax,%rdx
1c60: 48 89 c1 mov %rax,%rcx
1c63: 48 21 fa and %rdi,%rdx
1c66: 4c 09 c2 or %r8,%rdx
1c69: f0 48 0f b1 16 lock cmpxchg %rdx,(%rsi)
1c6e: 48 39 c1 cmp %rax,%rcx
1c71: 75 ea jne 1c5d <...>
to:
1c60: 48 89 ca mov %rcx,%rdx
1c63: 48 21 c2 and %rax,%rdx
1c66: 4c 09 c2 or %r8,%rdx
1c69: f0 48 0f b1 16 lock cmpxchg %rdx,(%rsi)
1c6e: 75 f0 jne 1c60 <...>
Link: https://lkml.kernel.org/r/20220708140736.8737-1-ubizjak@gmail.com
Signed-off-by: Uros Bizjak <ubizjak@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-07-08 16:07:36 +02:00
|
|
|
|
2016-05-19 17:14:27 -07:00
|
|
|
|
mm/page_alloc: pageblock flags functions clean up
Patch series "Make MIGRATE_ISOLATE a standalone bit", v10.
This patchset moves MIGRATE_ISOLATE to a standalone bit to avoid being
overwritten during pageblock isolation process. Currently,
MIGRATE_ISOLATE is part of enum migratetype (in include/linux/mmzone.h),
thus, setting a pageblock to MIGRATE_ISOLATE overwrites its original
migratetype. This causes pageblock migratetype loss during
alloc_contig_range() and memory offline, especially when the process fails
due to a failed pageblock isolation and the code tries to undo the
finished pageblock isolations.
In terms of performance for changing pageblock types, no performance
change is observed:
1. I used perf to collect stats of offlining and onlining all memory
of a 40GB VM 10 times and see that get_pfnblock_flags_mask() and
set_pfnblock_flags_mask() take about 0.12% and 0.02% of the whole
process respectively with and without this patchset across 3 runs.
2. I used perf to collect stats of dd from /dev/random to a 40GB tmpfs
file and find get_pfnblock_flags_mask() takes about 0.05% of the
process with and without this patchset across 3 runs.
This patch (of 6):
No functional change is intended.
1. Add __NR_PAGEBLOCK_BITS for the number of pageblock flag bits and use
roundup_pow_of_two(__NR_PAGEBLOCK_BITS) as NR_PAGEBLOCK_BITS to take
right amount of bits for pageblock flags.
2. Rename PB_migrate_skip to PB_compact_skip.
3. Add {get,set,clear}_pfnblock_bit() to operate one a standalone bit,
like PB_compact_skip.
3. Make {get,set}_pfnblock_flags_mask() internal functions and use
{get,set}_pfnblock_migratetype() for pageblock migratetype operations.
4. Move pageblock flags common code to get_pfnblock_bitmap_bitidx().
3. Use MIGRATETYPE_MASK to get the migratetype of a pageblock from its
flags.
4. Use PB_migrate_end in the definition of MIGRATETYPE_MASK instead of
PB_migrate_bits.
5. Add a comment on is_migrate_cma_folio() to prevent one from changing it
to use get_pageblock_migratetype() and causing issues.
Link: https://lkml.kernel.org/r/20250617021115.2331563-1-ziy@nvidia.com
Link: https://lkml.kernel.org/r/20250617021115.2331563-2-ziy@nvidia.com
Signed-off-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kirill A. Shuemov <kirill.shutemov@linux.intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Richard Chang <richardycc@google.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-06-16 22:11:09 -04:00
|
|
|
|
2016-05-19 17:14:27 -07:00
|
|
|
|
2020-08-06 23:25:48 -07:00
|
|
|
|
|
|
|
|
|
2016-05-19 17:14:27 -07:00
|
|
|
|
mm/page_alloc: pageblock flags functions clean up
Patch series "Make MIGRATE_ISOLATE a standalone bit", v10.
This patchset moves MIGRATE_ISOLATE to a standalone bit to avoid being
overwritten during pageblock isolation process. Currently,
MIGRATE_ISOLATE is part of enum migratetype (in include/linux/mmzone.h),
thus, setting a pageblock to MIGRATE_ISOLATE overwrites its original
migratetype. This causes pageblock migratetype loss during
alloc_contig_range() and memory offline, especially when the process fails
due to a failed pageblock isolation and the code tries to undo the
finished pageblock isolations.
In terms of performance for changing pageblock types, no performance
change is observed:
1. I used perf to collect stats of offlining and onlining all memory
of a 40GB VM 10 times and see that get_pfnblock_flags_mask() and
set_pfnblock_flags_mask() take about 0.12% and 0.02% of the whole
process respectively with and without this patchset across 3 runs.
2. I used perf to collect stats of dd from /dev/random to a 40GB tmpfs
file and find get_pfnblock_flags_mask() takes about 0.05% of the
process with and without this patchset across 3 runs.
This patch (of 6):
No functional change is intended.
1. Add __NR_PAGEBLOCK_BITS for the number of pageblock flag bits and use
roundup_pow_of_two(__NR_PAGEBLOCK_BITS) as NR_PAGEBLOCK_BITS to take
right amount of bits for pageblock flags.
2. Rename PB_migrate_skip to PB_compact_skip.
3. Add {get,set,clear}_pfnblock_bit() to operate one a standalone bit,
like PB_compact_skip.
3. Make {get,set}_pfnblock_flags_mask() internal functions and use
{get,set}_pfnblock_migratetype() for pageblock migratetype operations.
4. Move pageblock flags common code to get_pfnblock_bitmap_bitidx().
3. Use MIGRATETYPE_MASK to get the migratetype of a pageblock from its
flags.
4. Use PB_migrate_end in the definition of MIGRATETYPE_MASK instead of
PB_migrate_bits.
5. Add a comment on is_migrate_cma_folio() to prevent one from changing it
to use get_pageblock_migratetype() and causing issues.
Link: https://lkml.kernel.org/r/20250617021115.2331563-1-ziy@nvidia.com
Link: https://lkml.kernel.org/r/20250617021115.2331563-2-ziy@nvidia.com
Signed-off-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kirill A. Shuemov <kirill.shutemov@linux.intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Richard Chang <richardycc@google.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-06-16 22:11:09 -04:00
|
|
|
|
mm/page_alloc: use try_cmpxchg in set_pfnblock_flags_mask
Use try_cmpxchg instead of cmpxchg in set_pfnblock_flags_mask. x86
CMPXCHG instruction returns success in ZF flag, so this change saves a
compare after cmpxchg (and related move instruction in front of cmpxchg).
The main loop improves from:
1c5d: 48 89 c2 mov %rax,%rdx
1c60: 48 89 c1 mov %rax,%rcx
1c63: 48 21 fa and %rdi,%rdx
1c66: 4c 09 c2 or %r8,%rdx
1c69: f0 48 0f b1 16 lock cmpxchg %rdx,(%rsi)
1c6e: 48 39 c1 cmp %rax,%rcx
1c71: 75 ea jne 1c5d <...>
to:
1c60: 48 89 ca mov %rcx,%rdx
1c63: 48 21 c2 and %rax,%rdx
1c66: 4c 09 c2 or %r8,%rdx
1c69: f0 48 0f b1 16 lock cmpxchg %rdx,(%rsi)
1c6e: 75 f0 jne 1c60 <...>
Link: https://lkml.kernel.org/r/20220708140736.8737-1-ubizjak@gmail.com
Signed-off-by: Uros Bizjak <ubizjak@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-07-08 16:07:36 +02:00
|
|
|
|
mm/page_alloc: pageblock flags functions clean up
Patch series "Make MIGRATE_ISOLATE a standalone bit", v10.
This patchset moves MIGRATE_ISOLATE to a standalone bit to avoid being
overwritten during pageblock isolation process. Currently,
MIGRATE_ISOLATE is part of enum migratetype (in include/linux/mmzone.h),
thus, setting a pageblock to MIGRATE_ISOLATE overwrites its original
migratetype. This causes pageblock migratetype loss during
alloc_contig_range() and memory offline, especially when the process fails
due to a failed pageblock isolation and the code tries to undo the
finished pageblock isolations.
In terms of performance for changing pageblock types, no performance
change is observed:
1. I used perf to collect stats of offlining and onlining all memory
of a 40GB VM 10 times and see that get_pfnblock_flags_mask() and
set_pfnblock_flags_mask() take about 0.12% and 0.02% of the whole
process respectively with and without this patchset across 3 runs.
2. I used perf to collect stats of dd from /dev/random to a 40GB tmpfs
file and find get_pfnblock_flags_mask() takes about 0.05% of the
process with and without this patchset across 3 runs.
This patch (of 6):
No functional change is intended.
1. Add __NR_PAGEBLOCK_BITS for the number of pageblock flag bits and use
roundup_pow_of_two(__NR_PAGEBLOCK_BITS) as NR_PAGEBLOCK_BITS to take
right amount of bits for pageblock flags.
2. Rename PB_migrate_skip to PB_compact_skip.
3. Add {get,set,clear}_pfnblock_bit() to operate one a standalone bit,
like PB_compact_skip.
3. Make {get,set}_pfnblock_flags_mask() internal functions and use
{get,set}_pfnblock_migratetype() for pageblock migratetype operations.
4. Move pageblock flags common code to get_pfnblock_bitmap_bitidx().
3. Use MIGRATETYPE_MASK to get the migratetype of a pageblock from its
flags.
4. Use PB_migrate_end in the definition of MIGRATETYPE_MASK instead of
PB_migrate_bits.
5. Add a comment on is_migrate_cma_folio() to prevent one from changing it
to use get_pageblock_migratetype() and causing issues.
Link: https://lkml.kernel.org/r/20250617021115.2331563-1-ziy@nvidia.com
Link: https://lkml.kernel.org/r/20250617021115.2331563-2-ziy@nvidia.com
Signed-off-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kirill A. Shuemov <kirill.shutemov@linux.intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Richard Chang <richardycc@google.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-06-16 22:11:09 -04:00
|
|
|
|
2016-05-19 17:14:27 -07:00
|
|
|
|
2015-06-30 14:57:02 -07:00
|
|
|
|
mm/page_alloc: pageblock flags functions clean up
Patch series "Make MIGRATE_ISOLATE a standalone bit", v10.
This patchset moves MIGRATE_ISOLATE to a standalone bit to avoid being
overwritten during pageblock isolation process. Currently,
MIGRATE_ISOLATE is part of enum migratetype (in include/linux/mmzone.h),
thus, setting a pageblock to MIGRATE_ISOLATE overwrites its original
migratetype. This causes pageblock migratetype loss during
alloc_contig_range() and memory offline, especially when the process fails
due to a failed pageblock isolation and the code tries to undo the
finished pageblock isolations.
In terms of performance for changing pageblock types, no performance
change is observed:
1. I used perf to collect stats of offlining and onlining all memory
of a 40GB VM 10 times and see that get_pfnblock_flags_mask() and
set_pfnblock_flags_mask() take about 0.12% and 0.02% of the whole
process respectively with and without this patchset across 3 runs.
2. I used perf to collect stats of dd from /dev/random to a 40GB tmpfs
file and find get_pfnblock_flags_mask() takes about 0.05% of the
process with and without this patchset across 3 runs.
This patch (of 6):
No functional change is intended.
1. Add __NR_PAGEBLOCK_BITS for the number of pageblock flag bits and use
roundup_pow_of_two(__NR_PAGEBLOCK_BITS) as NR_PAGEBLOCK_BITS to take
right amount of bits for pageblock flags.
2. Rename PB_migrate_skip to PB_compact_skip.
3. Add {get,set,clear}_pfnblock_bit() to operate one a standalone bit,
like PB_compact_skip.
3. Make {get,set}_pfnblock_flags_mask() internal functions and use
{get,set}_pfnblock_migratetype() for pageblock migratetype operations.
4. Move pageblock flags common code to get_pfnblock_bitmap_bitidx().
3. Use MIGRATETYPE_MASK to get the migratetype of a pageblock from its
flags.
4. Use PB_migrate_end in the definition of MIGRATETYPE_MASK instead of
PB_migrate_bits.
5. Add a comment on is_migrate_cma_folio() to prevent one from changing it
to use get_pageblock_migratetype() and causing issues.
Link: https://lkml.kernel.org/r/20250617021115.2331563-1-ziy@nvidia.com
Link: https://lkml.kernel.org/r/20250617021115.2331563-2-ziy@nvidia.com
Signed-off-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kirill A. Shuemov <kirill.shutemov@linux.intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Richard Chang <richardycc@google.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-06-16 22:11:09 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-06-16 22:11:12 -04:00
|
|
|
|
|
|
|
|
|
2007-10-16 01:25:48 -07:00
|
|
|
|
2013-11-12 15:08:18 -08:00
|
|
|
|
|
|
|
|
|
2009-06-16 15:31:58 -07:00
|
|
|
|
|
|
|
|
|
2025-06-16 22:11:10 -04:00
|
|
|
|
|
|
|
|
|
2025-06-16 22:11:12 -04:00
|
|
|
|
|
|
|
|
|
2025-06-16 22:11:10 -04:00
|
|
|
|
|
|
|
|
|
2025-08-27 07:01:04 +00:00
|
|
|
|
2025-06-16 22:11:12 -04:00
|
|
|
|
2025-06-16 22:11:10 -04:00
|
|
|
|
|
|
|
|
|
mm/page_alloc: pageblock flags functions clean up
Patch series "Make MIGRATE_ISOLATE a standalone bit", v10.
This patchset moves MIGRATE_ISOLATE to a standalone bit to avoid being
overwritten during pageblock isolation process. Currently,
MIGRATE_ISOLATE is part of enum migratetype (in include/linux/mmzone.h),
thus, setting a pageblock to MIGRATE_ISOLATE overwrites its original
migratetype. This causes pageblock migratetype loss during
alloc_contig_range() and memory offline, especially when the process fails
due to a failed pageblock isolation and the code tries to undo the
finished pageblock isolations.
In terms of performance for changing pageblock types, no performance
change is observed:
1. I used perf to collect stats of offlining and onlining all memory
of a 40GB VM 10 times and see that get_pfnblock_flags_mask() and
set_pfnblock_flags_mask() take about 0.12% and 0.02% of the whole
process respectively with and without this patchset across 3 runs.
2. I used perf to collect stats of dd from /dev/random to a 40GB tmpfs
file and find get_pfnblock_flags_mask() takes about 0.05% of the
process with and without this patchset across 3 runs.
This patch (of 6):
No functional change is intended.
1. Add __NR_PAGEBLOCK_BITS for the number of pageblock flag bits and use
roundup_pow_of_two(__NR_PAGEBLOCK_BITS) as NR_PAGEBLOCK_BITS to take
right amount of bits for pageblock flags.
2. Rename PB_migrate_skip to PB_compact_skip.
3. Add {get,set,clear}_pfnblock_bit() to operate one a standalone bit,
like PB_compact_skip.
3. Make {get,set}_pfnblock_flags_mask() internal functions and use
{get,set}_pfnblock_migratetype() for pageblock migratetype operations.
4. Move pageblock flags common code to get_pfnblock_bitmap_bitidx().
3. Use MIGRATETYPE_MASK to get the migratetype of a pageblock from its
flags.
4. Use PB_migrate_end in the definition of MIGRATETYPE_MASK instead of
PB_migrate_bits.
5. Add a comment on is_migrate_cma_folio() to prevent one from changing it
to use get_pageblock_migratetype() and causing issues.
Link: https://lkml.kernel.org/r/20250617021115.2331563-1-ziy@nvidia.com
Link: https://lkml.kernel.org/r/20250617021115.2331563-2-ziy@nvidia.com
Signed-off-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kirill A. Shuemov <kirill.shutemov@linux.intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Richard Chang <richardycc@google.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-06-16 22:11:09 -04:00
|
|
|
|
2025-06-16 22:11:10 -04:00
|
|
|
|
|
|
|
|
|
2007-10-16 01:25:48 -07:00
|
|
|
|
|
|
|
|
|
2025-06-16 22:11:11 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-01-06 00:10:58 -08:00
|
|
|
|
2005-10-29 18:16:52 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2023-07-04 19:18:23 +08:00
|
|
|
|
2005-10-29 18:16:53 -07:00
|
|
|
|
|
|
|
|
|
2013-02-22 16:35:28 -08:00
|
|
|
|
2005-10-29 18:16:52 -07:00
|
|
|
|
2005-10-29 18:16:53 -07:00
|
|
|
|
|
|
|
|
|
2013-02-22 16:35:28 -08:00
|
|
|
|
|
|
|
|
|
2023-07-04 19:18:23 +08:00
|
|
|
|
2005-10-29 18:16:53 -07:00
|
|
|
|
|
|
|
|
|
2013-02-22 16:35:28 -08:00
|
|
|
|
2014-06-04 16:07:27 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-02-22 16:35:28 -08:00
|
|
|
|
2005-10-29 18:16:53 -07:00
|
|
|
|
2005-10-29 18:16:52 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-02-21 15:32:27 +08:00
|
|
|
|
2005-10-29 18:16:52 -07:00
|
|
|
|
|
|
|
|
|
2024-02-21 15:32:27 +08:00
|
|
|
|
2023-05-16 14:38:13 +08:00
|
|
|
|
2024-02-21 15:32:27 +08:00
|
|
|
|
2005-10-29 18:16:52 -07:00
|
|
|
|
2024-02-21 15:32:27 +08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2006-01-06 00:10:58 -08:00
|
|
|
|
2024-02-21 15:32:27 +08:00
|
|
|
|
2006-01-06 00:10:58 -08:00
|
|
|
|
2024-02-21 15:32:27 +08:00
|
|
|
|
2006-01-06 00:10:58 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-06-03 15:58:29 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2009-01-06 14:40:12 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-03-15 14:56:24 -07:00
|
|
|
|
2009-01-06 14:40:13 -08:00
|
|
|
|
2009-01-06 14:40:12 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-03-15 14:56:24 -07:00
|
|
|
|
badpage: replace page_remove_rmap Eeek and BUG
Now that bad pages are kept out of circulation, there is no need for the
infamous page_remove_rmap() BUG() - once that page is freed, its negative
mapcount will issue a "Bad page state" message and the page won't be
freed. Removing the BUG() allows more info, on subsequent pages, to be
gathered.
We do have more info about the page at this point than bad_page() can know
- notably, what the pmd is, which might pinpoint something like low 64kB
corruption - but page_remove_rmap() isn't given the address to find that.
In practice, there is only one call to page_remove_rmap() which has ever
reported anything, that from zap_pte_range() (usually on exit, sometimes
on munmap). It has all the info, so remove page_remove_rmap()'s "Eeek"
message and leave it all to zap_pte_range().
mm/memory.c already has a hardly used print_bad_pte() function, showing
some of the appropriate info: extend it to show what we want for the rmap
case: pte info, page info (when there is a page) and vma info to compare.
zap_pte_range() already knows the pmd, but print_bad_pte() is easier to
use if it works that out for itself.
Some of this info is also shown in bad_page()'s "Bad page state" message.
Keep them separate, but adjust them to match each other as far as
possible. Say "Bad page map" in print_bad_pte(), and add a TAINT_BAD_PAGE
there too.
print_bad_pte() show current->comm unconditionally (though it should get
repeated in the usually irrelevant stack trace): sorry, I misled Nick
Piggin to make it conditional on vm_mm == current->mm, but current->mm is
already NULL in the exit case. Usually current->comm is good, though
exceptionally it may not be that of the mm (when "swapoff" for example).
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-01-06 14:40:08 -08:00
|
|
|
|
2021-06-28 19:41:07 -07:00
|
|
|
|
2008-10-18 20:28:16 -07:00
|
|
|
|
2011-10-31 17:07:24 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2009-01-06 14:40:12 -08:00
|
|
|
|
2009-01-06 14:40:06 -08:00
|
|
|
|
2024-05-29 13:19:02 +02:00
|
|
|
|
|
|
|
|
|
2013-01-21 17:17:39 +10:30
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2021-06-28 19:43:08 -07:00
|
|
|
|
|
|
|
|
|
mm/page_alloc: Separate THP PCP into movable and non-movable categories
Since commit 5d0a661d808f ("mm/page_alloc: use only one PCP list for
THP-sized allocations") no longer differentiates the migration type of
pages in THP-sized PCP list, it's possible that non-movable allocation
requests may get a CMA page from the list, in some cases, it's not
acceptable.
If a large number of CMA memory are configured in system (for example, the
CMA memory accounts for 50% of the system memory), starting a virtual
machine with device passthrough will get stuck. During starting the
virtual machine, it will call pin_user_pages_remote(..., FOLL_LONGTERM,
...) to pin memory. Normally if a page is present and in CMA area,
pin_user_pages_remote() will migrate the page from CMA area to non-CMA
area because of FOLL_LONGTERM flag. But if non-movable allocation
requests return CMA memory, migrate_longterm_unpinnable_pages() will
migrate a CMA page to another CMA page, which will fail to pass the check
in check_and_migrate_movable_pages() and cause migration endless.
Call trace:
pin_user_pages_remote
--__gup_longterm_locked // endless loops in this function
----_get_user_pages_locked
----check_and_migrate_movable_pages
------migrate_longterm_unpinnable_pages
--------alloc_migration_target
This problem will also have a negative impact on CMA itself. For example,
when CMA is borrowed by THP, and we need to reclaim it through cma_alloc()
or dma_alloc_coherent(), we must move those pages out to ensure CMA's
users can retrieve that contigous memory. Currently, CMA's memory is
occupied by non-movable pages, meaning we can't relocate them. As a
result, cma_alloc() is more likely to fail.
To fix the problem above, we add one PCP list for THP, which will not
introduce a new cacheline for struct per_cpu_pages. THP will have 2 PCP
lists, one PCP list is used by MOVABLE allocation, and the other PCP list
is used by UNMOVABLE allocation. MOVABLE allocation contains GPF_MOVABLE,
and UNMOVABLE allocation contains GFP_UNMOVABLE and GFP_RECLAIMABLE.
Link: https://lkml.kernel.org/r/1718845190-4456-1-git-send-email-yangge1116@126.com
Fixes: 5d0a661d808f ("mm/page_alloc: use only one PCP list for THP-sized allocations")
Signed-off-by: yangge <yangge1116@126.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Barry Song <21cnbao@gmail.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-06-20 08:59:50 +08:00
|
|
|
|
2021-06-28 19:43:08 -07:00
|
|
|
|
2025-03-19 17:17:26 +08:00
|
|
|
|
2021-06-28 19:43:08 -07:00
|
|
|
|
2024-04-03 21:47:21 +08:00
|
|
|
|
mm/page_alloc: Separate THP PCP into movable and non-movable categories
Since commit 5d0a661d808f ("mm/page_alloc: use only one PCP list for
THP-sized allocations") no longer differentiates the migration type of
pages in THP-sized PCP list, it's possible that non-movable allocation
requests may get a CMA page from the list, in some cases, it's not
acceptable.
If a large number of CMA memory are configured in system (for example, the
CMA memory accounts for 50% of the system memory), starting a virtual
machine with device passthrough will get stuck. During starting the
virtual machine, it will call pin_user_pages_remote(..., FOLL_LONGTERM,
...) to pin memory. Normally if a page is present and in CMA area,
pin_user_pages_remote() will migrate the page from CMA area to non-CMA
area because of FOLL_LONGTERM flag. But if non-movable allocation
requests return CMA memory, migrate_longterm_unpinnable_pages() will
migrate a CMA page to another CMA page, which will fail to pass the check
in check_and_migrate_movable_pages() and cause migration endless.
Call trace:
pin_user_pages_remote
--__gup_longterm_locked // endless loops in this function
----_get_user_pages_locked
----check_and_migrate_movable_pages
------migrate_longterm_unpinnable_pages
--------alloc_migration_target
This problem will also have a negative impact on CMA itself. For example,
when CMA is borrowed by THP, and we need to reclaim it through cma_alloc()
or dma_alloc_coherent(), we must move those pages out to ensure CMA's
users can retrieve that contigous memory. Currently, CMA's memory is
occupied by non-movable pages, meaning we can't relocate them. As a
result, cma_alloc() is more likely to fail.
To fix the problem above, we add one PCP list for THP, which will not
introduce a new cacheline for struct per_cpu_pages. THP will have 2 PCP
lists, one PCP list is used by MOVABLE allocation, and the other PCP list
is used by UNMOVABLE allocation. MOVABLE allocation contains GPF_MOVABLE,
and UNMOVABLE allocation contains GFP_UNMOVABLE and GFP_RECLAIMABLE.
Link: https://lkml.kernel.org/r/1718845190-4456-1-git-send-email-yangge1116@126.com
Fixes: 5d0a661d808f ("mm/page_alloc: use only one PCP list for THP-sized allocations")
Signed-off-by: yangge <yangge1116@126.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Barry Song <21cnbao@gmail.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-06-20 08:59:50 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-28 19:43:08 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-08-03 19:49:34 +08:00
|
|
|
|
2021-06-28 19:43:08 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/page_alloc: Separate THP PCP into movable and non-movable categories
Since commit 5d0a661d808f ("mm/page_alloc: use only one PCP list for
THP-sized allocations") no longer differentiates the migration type of
pages in THP-sized PCP list, it's possible that non-movable allocation
requests may get a CMA page from the list, in some cases, it's not
acceptable.
If a large number of CMA memory are configured in system (for example, the
CMA memory accounts for 50% of the system memory), starting a virtual
machine with device passthrough will get stuck. During starting the
virtual machine, it will call pin_user_pages_remote(..., FOLL_LONGTERM,
...) to pin memory. Normally if a page is present and in CMA area,
pin_user_pages_remote() will migrate the page from CMA area to non-CMA
area because of FOLL_LONGTERM flag. But if non-movable allocation
requests return CMA memory, migrate_longterm_unpinnable_pages() will
migrate a CMA page to another CMA page, which will fail to pass the check
in check_and_migrate_movable_pages() and cause migration endless.
Call trace:
pin_user_pages_remote
--__gup_longterm_locked // endless loops in this function
----_get_user_pages_locked
----check_and_migrate_movable_pages
------migrate_longterm_unpinnable_pages
--------alloc_migration_target
This problem will also have a negative impact on CMA itself. For example,
when CMA is borrowed by THP, and we need to reclaim it through cma_alloc()
or dma_alloc_coherent(), we must move those pages out to ensure CMA's
users can retrieve that contigous memory. Currently, CMA's memory is
occupied by non-movable pages, meaning we can't relocate them. As a
result, cma_alloc() is more likely to fail.
To fix the problem above, we add one PCP list for THP, which will not
introduce a new cacheline for struct per_cpu_pages. THP will have 2 PCP
lists, one PCP list is used by MOVABLE allocation, and the other PCP list
is used by UNMOVABLE allocation. MOVABLE allocation contains GPF_MOVABLE,
and UNMOVABLE allocation contains GFP_UNMOVABLE and GFP_RECLAIMABLE.
Link: https://lkml.kernel.org/r/1718845190-4456-1-git-send-email-yangge1116@126.com
Fixes: 5d0a661d808f ("mm/page_alloc: use only one PCP list for THP-sized allocations")
Signed-off-by: yangge <yangge1116@126.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Barry Song <21cnbao@gmail.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-06-20 08:59:50 +08:00
|
|
|
|
2024-04-03 21:47:21 +08:00
|
|
|
|
2021-06-28 19:43:08 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-04-03 21:47:21 +08:00
|
|
|
|
2021-06-28 19:43:08 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-11-06 16:29:54 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2015-11-06 16:29:54 -08:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2015-11-06 16:29:54 -08:00
|
|
|
|
[PATCH] compound page: use page[1].lru
If a compound page has its own put_page_testzero destructor (the only current
example is free_huge_page), that is noted in page[1].mapping of the compound
page. But that's rather a poor place to keep it: functions which call
set_page_dirty_lock after get_user_pages (e.g. Infiniband's
__ib_umem_release) ought to be checking first, otherwise set_page_dirty is
liable to crash on what's not the address of a struct address_space.
And now I'm about to make that worse: it turns out that every compound page
needs a destructor, so we can no longer rely on hugetlb pages going their own
special way, to avoid further problems of page->mapping reuse. For example,
not many people know that: on 50% of i386 -Os builds, the first tail page of a
compound page purports to be PageAnon (when its destructor has an odd
address), which surprises page_add_file_rmap.
Keep the compound page destructor in page[1].lru.next instead. And to free up
the common pairing of mapping and index, also move compound page order from
index to lru.prev. Slab reuses page->lru too: but if we ever need slab to use
compound pages, it can easily stack its use above this.
(akpm: decoded version of the above: the tail pages of a compound page now
have ->mapping==NULL, so there's no need for the set_page_dirty[_lock]()
caller to check that they're not compund pages before doing the dirty).
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-02-14 13:52:58 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2006-02-14 13:52:59 -08:00
|
|
|
|
2015-11-06 16:29:57 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-05-06 14:49:40 -07:00
|
|
|
|
mm/page_alloc: split prep_compound_page into head and tail subparts
Patch series "mm, device-dax: Introduce compound pages in devmap", v7.
This series converts device-dax to use compound pages, and moves away
from the 'struct page per basepage on PMD/PUD' that is done today.
Doing so
1) unlocks a few noticeable improvements on unpin_user_pages() and
makes device-dax+altmap case 4x times faster in pinning (numbers
below and in last patch)
2) as mentioned in various other threads it's one important step
towards cleaning up ZONE_DEVICE refcounting.
I've split the compound pages on devmap part from the rest based on
recent discussions on devmap pending and future work planned[5][6].
There is consensus that device-dax should be using compound pages to
represent its PMD/PUDs just like HugeTLB and THP, and that leads to less
specialization of the dax parts. I will pursue the rest of the work in
parallel once this part is merged, particular the GUP-{slow,fast}
improvements [7] and the tail struct page deduplication memory savings
part[8].
To summarize what the series does:
Patch 1: Prepare hwpoisoning to work with dax compound pages.
Patches 2-3: Split the current utility function of prep_compound_page()
into head and tail and use those two helpers where appropriate to take
advantage of caches being warm after __init_single_page(). This is used
when initializing zone device when we bring up device-dax namespaces.
Patches 4-10: Add devmap support for compound pages in device-dax.
memmap_init_zone_device() initialize its metadata as compound pages, and
it introduces a new devmap property known as vmemmap_shift which
outlines how the vmemmap is structured (defaults to base pages as done
today). The property describe the page order of the metadata
essentially. While at it do a few cleanups in device-dax in patches
5-9. Finally enable device-dax usage of devmap @vmemmap_shift to a
value based on its own @align property. @vmemmap_shift returns 0 by
default (which is today's case of base pages in devmap, like fsdax or
the others) and the usage of compound devmap is optional. Starting with
device-dax (*not* fsdax) we enable it by default. There are a few
pinning improvements particular on the unpinning case and altmap, as
well as unpin_user_page_range_dirty_lock() being just as effective as
THP/hugetlb[0] pages.
$ gup_test -f /dev/dax1.0 -m 16384 -r 10 -S -a -n 512 -w
(pin_user_pages_fast 2M pages) put:~71 ms -> put:~22 ms
[altmap]
(pin_user_pages_fast 2M pages) get:~524ms put:~525 ms -> get: ~127ms put:~71ms
$ gup_test -f /dev/dax1.0 -m 129022 -r 10 -S -a -n 512 -w
(pin_user_pages_fast 2M pages) put:~513 ms -> put:~188 ms
[altmap with -m 127004]
(pin_user_pages_fast 2M pages) get:~4.1 secs put:~4.12 secs -> get:~1sec put:~563ms
Tested on x86 with 1Tb+ of pmem (alongside registering it with RDMA with
and without altmap), alongside gup_test selftests with dynamic dax
regions and static dax regions. Coupled with ndctl unit tests for
dynamic dax devices that exercise all of this. Note, for dynamic dax
regions I had to revert commit 8aa83e6395 ("x86/setup: Call
early_reserve_memory() earlier"), it is a known issue that this commit
broke efi_fake_mem=.
This patch (of 11):
Split the utility function prep_compound_page() into head and tail
counterparts, and use them accordingly.
This is in preparation for sharing the storage for compound page
metadata.
Link: https://lkml.kernel.org/r/20211202204422.26777-1-joao.m.martins@oracle.com
Link: https://lkml.kernel.org/r/20211202204422.26777-3-joao.m.martins@oracle.com
Signed-off-by: Joao Martins <joao.m.martins@oracle.com>
Acked-by: Mike Kravetz <mike.kravetz@oracle.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Reviewed-by: Muchun Song <songmuchun@bytedance.com>
Cc: Vishal Verma <vishal.l.verma@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Naoya Horiguchi <naoya.horiguchi@nec.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Jason Gunthorpe <jgg@ziepe.ca>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Jane Chu <jane.chu@oracle.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-14 14:04:15 -08:00
|
|
|
|
|
|
|
|
|
2020-08-14 17:30:23 -07:00
|
|
|
|
mm/page_alloc: split prep_compound_page into head and tail subparts
Patch series "mm, device-dax: Introduce compound pages in devmap", v7.
This series converts device-dax to use compound pages, and moves away
from the 'struct page per basepage on PMD/PUD' that is done today.
Doing so
1) unlocks a few noticeable improvements on unpin_user_pages() and
makes device-dax+altmap case 4x times faster in pinning (numbers
below and in last patch)
2) as mentioned in various other threads it's one important step
towards cleaning up ZONE_DEVICE refcounting.
I've split the compound pages on devmap part from the rest based on
recent discussions on devmap pending and future work planned[5][6].
There is consensus that device-dax should be using compound pages to
represent its PMD/PUDs just like HugeTLB and THP, and that leads to less
specialization of the dax parts. I will pursue the rest of the work in
parallel once this part is merged, particular the GUP-{slow,fast}
improvements [7] and the tail struct page deduplication memory savings
part[8].
To summarize what the series does:
Patch 1: Prepare hwpoisoning to work with dax compound pages.
Patches 2-3: Split the current utility function of prep_compound_page()
into head and tail and use those two helpers where appropriate to take
advantage of caches being warm after __init_single_page(). This is used
when initializing zone device when we bring up device-dax namespaces.
Patches 4-10: Add devmap support for compound pages in device-dax.
memmap_init_zone_device() initialize its metadata as compound pages, and
it introduces a new devmap property known as vmemmap_shift which
outlines how the vmemmap is structured (defaults to base pages as done
today). The property describe the page order of the metadata
essentially. While at it do a few cleanups in device-dax in patches
5-9. Finally enable device-dax usage of devmap @vmemmap_shift to a
value based on its own @align property. @vmemmap_shift returns 0 by
default (which is today's case of base pages in devmap, like fsdax or
the others) and the usage of compound devmap is optional. Starting with
device-dax (*not* fsdax) we enable it by default. There are a few
pinning improvements particular on the unpinning case and altmap, as
well as unpin_user_page_range_dirty_lock() being just as effective as
THP/hugetlb[0] pages.
$ gup_test -f /dev/dax1.0 -m 16384 -r 10 -S -a -n 512 -w
(pin_user_pages_fast 2M pages) put:~71 ms -> put:~22 ms
[altmap]
(pin_user_pages_fast 2M pages) get:~524ms put:~525 ms -> get: ~127ms put:~71ms
$ gup_test -f /dev/dax1.0 -m 129022 -r 10 -S -a -n 512 -w
(pin_user_pages_fast 2M pages) put:~513 ms -> put:~188 ms
[altmap with -m 127004]
(pin_user_pages_fast 2M pages) get:~4.1 secs put:~4.12 secs -> get:~1sec put:~563ms
Tested on x86 with 1Tb+ of pmem (alongside registering it with RDMA with
and without altmap), alongside gup_test selftests with dynamic dax
regions and static dax regions. Coupled with ndctl unit tests for
dynamic dax devices that exercise all of this. Note, for dynamic dax
regions I had to revert commit 8aa83e6395 ("x86/setup: Call
early_reserve_memory() earlier"), it is a known issue that this commit
broke efi_fake_mem=.
This patch (of 11):
Split the utility function prep_compound_page() into head and tail
counterparts, and use them accordingly.
This is in preparation for sharing the storage for compound page
metadata.
Link: https://lkml.kernel.org/r/20211202204422.26777-1-joao.m.martins@oracle.com
Link: https://lkml.kernel.org/r/20211202204422.26777-3-joao.m.martins@oracle.com
Signed-off-by: Joao Martins <joao.m.martins@oracle.com>
Acked-by: Mike Kravetz <mike.kravetz@oracle.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Reviewed-by: Muchun Song <songmuchun@bytedance.com>
Cc: Vishal Verma <vishal.l.verma@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Naoya Horiguchi <naoya.horiguchi@nec.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Jason Gunthorpe <jgg@ziepe.ca>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Jane Chu <jane.chu@oracle.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-14 14:04:15 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2020-10-15 20:10:15 -07:00
|
|
|
|
2006-04-18 22:20:52 -07:00
|
|
|
|
[PATCH] mm: split page table lock
Christoph Lameter demonstrated very poor scalability on the SGI 512-way, with
a many-threaded application which concurrently initializes different parts of
a large anonymous area.
This patch corrects that, by using a separate spinlock per page table page, to
guard the page table entries in that page, instead of using the mm's single
page_table_lock. (But even then, page_table_lock is still used to guard page
table allocation, and anon_vma allocation.)
In this implementation, the spinlock is tucked inside the struct page of the
page table page: with a BUILD_BUG_ON in case it overflows - which it would in
the case of 32-bit PA-RISC with spinlock debugging enabled.
Splitting the lock is not quite for free: another cacheline access. Ideally,
I suppose we would use split ptlock only for multi-threaded processes on
multi-cpu machines; but deciding that dynamically would have its own costs.
So for now enable it by config, at some number of cpus - since the Kconfig
language doesn't support inequalities, let preprocessor compare that with
NR_CPUS. But I don't think it's worth being user-configurable: for good
testing of both split and unsplit configs, split now at 4 cpus, and perhaps
change that to 8 later.
There is a benefit even for singly threaded processes: kswapd can be attacking
one part of the mm while another part is busy faulting.
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-10-29 18:16:40 -07:00
|
|
|
|
2006-04-10 11:21:48 +10:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2019-03-05 15:45:41 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-08-06 23:25:16 -07:00
|
|
|
|
2019-03-05 15:45:41 -08:00
|
|
|
|
|
|
|
|
|
2020-08-06 23:25:16 -07:00
|
|
|
|
2019-03-05 15:45:41 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-04-12 11:27:04 +08:00
|
|
|
|
|
|
|
|
|
2019-03-05 15:45:41 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-04-12 11:27:04 +08:00
|
|
|
|
|
|
|
|
|
2019-03-05 15:45:41 -08:00
|
|
|
|
|
|
|
|
|
2025-03-13 17:05:33 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-03-05 15:45:41 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
|
|
|
|
|
2020-04-06 20:04:49 -07:00
|
|
|
|
mm/page_alloc: keep track of free highatomic
OOM kills due to vastly overestimated free highatomic reserves were
observed:
... invoked oom-killer: gfp_mask=0x100cca(GFP_HIGHUSER_MOVABLE), order=0 ...
Node 0 Normal free:1482936kB boost:0kB min:410416kB low:739404kB high:1068392kB reserved_highatomic:1073152KB ...
Node 0 Normal: 1292*4kB (ME) 1920*8kB (E) 383*16kB (UE) 220*32kB (ME) 340*64kB (E) 2155*128kB (UE) 3243*256kB (UE) 615*512kB (U) 1*1024kB (M) 0*2048kB 0*4096kB = 1477408kB
The second line above shows that the OOM kill was due to the following
condition:
free (1482936kB) - reserved_highatomic (1073152kB) = 409784KB < min (410416kB)
And the third line shows there were no free pages in any
MIGRATE_HIGHATOMIC pageblocks, which otherwise would show up as type 'H'.
Therefore __zone_watermark_unusable_free() underestimated the usable free
memory by over 1GB, which resulted in the unnecessary OOM kill above.
The comments in __zone_watermark_unusable_free() warns about the potential
risk, i.e.,
If the caller does not have rights to reserves below the min
watermark then subtract the high-atomic reserves. This will
over-estimate the size of the atomic reserve but it avoids a search.
However, it is possible to keep track of free pages in reserved highatomic
pageblocks with a new per-zone counter nr_free_highatomic protected by the
zone lock, to avoid a search when calculating the usable free memory. And
the cost would be minimal, i.e., simple arithmetics in the highatomic
alloc/free/move paths.
Note that since nr_free_highatomic can be relatively small, using a
per-cpu counter might cause too much drift and defeat its purpose, in
addition to the extra memory overhead.
Dependson e0932b6c1f94 ("mm: page_alloc: consolidate free page accounting") - see [1]
[akpm@linux-foundation.org: s/if/else if/, per Johannes, stealth whitespace tweak]
Link: https://lkml.kernel.org/r/20241028182653.3420139-1-yuzhao@google.com
Link: https://lkml.kernel.org/r/0d0ddb33-fcdc-43e2-801f-0c1df2031afb@suse.cz [1]
Fixes: 0aaa29a56e4f ("mm, page_alloc: reserve pageblocks for high-order atomic allocations on demand")
Signed-off-by: Yu Zhao <yuzhao@google.com>
Reported-by: Link Lin <linkl@google.com>
Acked-by: David Rientjes <rientjes@google.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-10-28 12:26:53 -06:00
|
|
|
|
|
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
|
|
|
|
|
2020-04-06 20:04:49 -07:00
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-08-21 13:29:47 +00:00
|
|
|
|
mm/page_alloc: keep track of free highatomic
OOM kills due to vastly overestimated free highatomic reserves were
observed:
... invoked oom-killer: gfp_mask=0x100cca(GFP_HIGHUSER_MOVABLE), order=0 ...
Node 0 Normal free:1482936kB boost:0kB min:410416kB low:739404kB high:1068392kB reserved_highatomic:1073152KB ...
Node 0 Normal: 1292*4kB (ME) 1920*8kB (E) 383*16kB (UE) 220*32kB (ME) 340*64kB (E) 2155*128kB (UE) 3243*256kB (UE) 615*512kB (U) 1*1024kB (M) 0*2048kB 0*4096kB = 1477408kB
The second line above shows that the OOM kill was due to the following
condition:
free (1482936kB) - reserved_highatomic (1073152kB) = 409784KB < min (410416kB)
And the third line shows there were no free pages in any
MIGRATE_HIGHATOMIC pageblocks, which otherwise would show up as type 'H'.
Therefore __zone_watermark_unusable_free() underestimated the usable free
memory by over 1GB, which resulted in the unnecessary OOM kill above.
The comments in __zone_watermark_unusable_free() warns about the potential
risk, i.e.,
If the caller does not have rights to reserves below the min
watermark then subtract the high-atomic reserves. This will
over-estimate the size of the atomic reserve but it avoids a search.
However, it is possible to keep track of free pages in reserved highatomic
pageblocks with a new per-zone counter nr_free_highatomic protected by the
zone lock, to avoid a search when calculating the usable free memory. And
the cost would be minimal, i.e., simple arithmetics in the highatomic
alloc/free/move paths.
Note that since nr_free_highatomic can be relatively small, using a
per-cpu counter might cause too much drift and defeat its purpose, in
addition to the extra memory overhead.
Dependson e0932b6c1f94 ("mm: page_alloc: consolidate free page accounting") - see [1]
[akpm@linux-foundation.org: s/if/else if/, per Johannes, stealth whitespace tweak]
Link: https://lkml.kernel.org/r/20241028182653.3420139-1-yuzhao@google.com
Link: https://lkml.kernel.org/r/0d0ddb33-fcdc-43e2-801f-0c1df2031afb@suse.cz [1]
Fixes: 0aaa29a56e4f ("mm, page_alloc: reserve pageblocks for high-order atomic allocations on demand")
Signed-off-by: Yu Zhao <yuzhao@google.com>
Reported-by: Link Lin <linkl@google.com>
Acked-by: David Rientjes <rientjes@google.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-10-28 12:26:53 -06:00
|
|
|
|
|
|
|
|
|
2020-04-06 20:04:49 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-04-06 20:04:49 -07:00
|
|
|
|
|
|
|
|
|
mm: page_alloc: defrag_mode kswapd/kcompactd watermarks
The previous patch added pageblock_order reclaim to kswapd/kcompactd,
which helps, but produces only one block at a time. Allocation stalls and
THP failure rates are still higher than they could be.
To adequately reflect ALLOC_NOFRAGMENT demand for pageblocks, change the
watermarking for kswapd & kcompactd: instead of targeting the high
watermark in order-0 pages and checking for one suitable block, simply
require that the high watermark is entirely met in pageblocks.
To this end, track the number of free pages within contiguous pageblocks,
then change pgdat_balanced() and compact_finished() to check watermarks
against this new value.
This further reduces THP latencies and allocation stalls, and improves THP
success rates against the previous patch:
DEFRAGMODE-ASYNC DEFRAGMODE-ASYNC-WMARKS
Hugealloc Time mean 34300.36 ( +0.00%) 28904.00 ( -15.73%)
Hugealloc Time stddev 36390.42 ( +0.00%) 33464.37 ( -8.04%)
Kbuild Real time 196.13 ( +0.00%) 196.59 ( +0.23%)
Kbuild User time 1234.74 ( +0.00%) 1231.67 ( -0.25%)
Kbuild System time 62.62 ( +0.00%) 59.10 ( -5.54%)
THP fault alloc 57054.53 ( +0.00%) 63223.67 ( +10.81%)
THP fault fallback 11581.40 ( +0.00%) 5412.47 ( -53.26%)
Direct compact fail 107.80 ( +0.00%) 59.07 ( -44.79%)
Direct compact success 4.53 ( +0.00%) 2.80 ( -31.33%)
Direct compact success rate % 3.20 ( +0.00%) 3.99 ( +18.66%)
Compact daemon scanned migrate 5461033.93 ( +0.00%) 2267500.33 ( -58.48%)
Compact daemon scanned free 5824897.93 ( +0.00%) 2339773.00 ( -59.83%)
Compact direct scanned migrate 58336.93 ( +0.00%) 47659.93 ( -18.30%)
Compact direct scanned free 32791.87 ( +0.00%) 40729.67 ( +24.21%)
Compact total migrate scanned 5519370.87 ( +0.00%) 2315160.27 ( -58.05%)
Compact total free scanned 5857689.80 ( +0.00%) 2380502.67 ( -59.36%)
Alloc stall 2424.60 ( +0.00%) 638.87 ( -73.62%)
Pages kswapd scanned 2657018.33 ( +0.00%) 4002186.33 ( +50.63%)
Pages kswapd reclaimed 559583.07 ( +0.00%) 718577.80 ( +28.41%)
Pages direct scanned 722094.07 ( +0.00%) 355172.73 ( -50.81%)
Pages direct reclaimed 107257.80 ( +0.00%) 31162.80 ( -70.95%)
Pages total scanned 3379112.40 ( +0.00%) 4357359.07 ( +28.95%)
Pages total reclaimed 666840.87 ( +0.00%) 749740.60 ( +12.43%)
Swap out 77238.20 ( +0.00%) 110084.33 ( +42.53%)
Swap in 11712.80 ( +0.00%) 24457.00 ( +108.80%)
File refaults 143438.80 ( +0.00%) 188226.93 ( +31.22%)
Also of note is that compaction work overall is reduced. The reason for
this is that when free pageblocks are more readily available, allocations
are also much more likely to get physically placed in LRU order, instead
of being forced to scavenge free space here and there. This means that
reclaim by itself has better chances of freeing up whole blocks, and the
system relies less on compaction.
Comparing all changes to the vanilla kernel:
VANILLA DEFRAGMODE-ASYNC-WMARKS
Hugealloc Time mean 52739.45 ( +0.00%) 28904.00 ( -45.19%)
Hugealloc Time stddev 56541.26 ( +0.00%) 33464.37 ( -40.81%)
Kbuild Real time 197.47 ( +0.00%) 196.59 ( -0.44%)
Kbuild User time 1240.49 ( +0.00%) 1231.67 ( -0.71%)
Kbuild System time 70.08 ( +0.00%) 59.10 ( -15.45%)
THP fault alloc 46727.07 ( +0.00%) 63223.67 ( +35.30%)
THP fault fallback 21910.60 ( +0.00%) 5412.47 ( -75.29%)
Direct compact fail 195.80 ( +0.00%) 59.07 ( -69.48%)
Direct compact success 7.93 ( +0.00%) 2.80 ( -57.46%)
Direct compact success rate % 3.51 ( +0.00%) 3.99 ( +10.49%)
Compact daemon scanned migrate 3369601.27 ( +0.00%) 2267500.33 ( -32.71%)
Compact daemon scanned free 5075474.47 ( +0.00%) 2339773.00 ( -53.90%)
Compact direct scanned migrate 161787.27 ( +0.00%) 47659.93 ( -70.54%)
Compact direct scanned free 163467.53 ( +0.00%) 40729.67 ( -75.08%)
Compact total migrate scanned 3531388.53 ( +0.00%) 2315160.27 ( -34.44%)
Compact total free scanned 5238942.00 ( +0.00%) 2380502.67 ( -54.56%)
Alloc stall 2371.07 ( +0.00%) 638.87 ( -73.02%)
Pages kswapd scanned 2160926.73 ( +0.00%) 4002186.33 ( +85.21%)
Pages kswapd reclaimed 533191.07 ( +0.00%) 718577.80 ( +34.77%)
Pages direct scanned 400450.33 ( +0.00%) 355172.73 ( -11.31%)
Pages direct reclaimed 94441.73 ( +0.00%) 31162.80 ( -67.00%)
Pages total scanned 2561377.07 ( +0.00%) 4357359.07 ( +70.12%)
Pages total reclaimed 627632.80 ( +0.00%) 749740.60 ( +19.46%)
Swap out 47959.53 ( +0.00%) 110084.33 ( +129.53%)
Swap in 7276.00 ( +0.00%) 24457.00 ( +236.10%)
File refaults 138043.00 ( +0.00%) 188226.93 ( +36.35%)
THP allocation latencies and %sys time are down dramatically.
THP allocation failures are down from nearly 50% to 8.5%. And to recall
previous data points, the success rates are steady and reliable without
the cumulative deterioration of fragmentation events.
Compaction work is down overall. Direct compaction work especially is
drastically reduced. As an aside, its success rate of 4% indicates there
is room for improvement. For now it's good to rely on it less.
Reclaim work is up overall, however direct reclaim work is down. Part of
the increase can be attributed to a higher use of THPs, which due to
internal fragmentation increase the memory footprint. This is not
necessarily an unexpected side-effect for users of THP.
However, taken both points together, there may well be some opportunities
for fine tuning in the reclaim/compaction coordination.
[hannes@cmpxchg.org: fix squawks from rebasing]
Link: https://lkml.kernel.org/r/20250314210558.GD1316033@cmpxchg.org
Link: https://lkml.kernel.org/r/20250313210647.1314586-6-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-03-13 17:05:36 -04:00
|
|
|
|
2020-04-06 20:04:49 -07:00
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
mm/page_alloc: pageblock flags functions clean up
Patch series "Make MIGRATE_ISOLATE a standalone bit", v10.
This patchset moves MIGRATE_ISOLATE to a standalone bit to avoid being
overwritten during pageblock isolation process. Currently,
MIGRATE_ISOLATE is part of enum migratetype (in include/linux/mmzone.h),
thus, setting a pageblock to MIGRATE_ISOLATE overwrites its original
migratetype. This causes pageblock migratetype loss during
alloc_contig_range() and memory offline, especially when the process fails
due to a failed pageblock isolation and the code tries to undo the
finished pageblock isolations.
In terms of performance for changing pageblock types, no performance
change is observed:
1. I used perf to collect stats of offlining and onlining all memory
of a 40GB VM 10 times and see that get_pfnblock_flags_mask() and
set_pfnblock_flags_mask() take about 0.12% and 0.02% of the whole
process respectively with and without this patchset across 3 runs.
2. I used perf to collect stats of dd from /dev/random to a 40GB tmpfs
file and find get_pfnblock_flags_mask() takes about 0.05% of the
process with and without this patchset across 3 runs.
This patch (of 6):
No functional change is intended.
1. Add __NR_PAGEBLOCK_BITS for the number of pageblock flag bits and use
roundup_pow_of_two(__NR_PAGEBLOCK_BITS) as NR_PAGEBLOCK_BITS to take
right amount of bits for pageblock flags.
2. Rename PB_migrate_skip to PB_compact_skip.
3. Add {get,set,clear}_pfnblock_bit() to operate one a standalone bit,
like PB_compact_skip.
3. Make {get,set}_pfnblock_flags_mask() internal functions and use
{get,set}_pfnblock_migratetype() for pageblock migratetype operations.
4. Move pageblock flags common code to get_pfnblock_bitmap_bitidx().
3. Use MIGRATETYPE_MASK to get the migratetype of a pageblock from its
flags.
4. Use PB_migrate_end in the definition of MIGRATETYPE_MASK instead of
PB_migrate_bits.
5. Add a comment on is_migrate_cma_folio() to prevent one from changing it
to use get_pageblock_migratetype() and causing issues.
Link: https://lkml.kernel.org/r/20250617021115.2331563-1-ziy@nvidia.com
Link: https://lkml.kernel.org/r/20250617021115.2331563-2-ziy@nvidia.com
Signed-off-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kirill A. Shuemov <kirill.shutemov@linux.intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Richard Chang <richardycc@google.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-06-16 22:11:09 -04:00
|
|
|
|
mm: page_alloc: defrag_mode kswapd/kcompactd watermarks
The previous patch added pageblock_order reclaim to kswapd/kcompactd,
which helps, but produces only one block at a time. Allocation stalls and
THP failure rates are still higher than they could be.
To adequately reflect ALLOC_NOFRAGMENT demand for pageblocks, change the
watermarking for kswapd & kcompactd: instead of targeting the high
watermark in order-0 pages and checking for one suitable block, simply
require that the high watermark is entirely met in pageblocks.
To this end, track the number of free pages within contiguous pageblocks,
then change pgdat_balanced() and compact_finished() to check watermarks
against this new value.
This further reduces THP latencies and allocation stalls, and improves THP
success rates against the previous patch:
DEFRAGMODE-ASYNC DEFRAGMODE-ASYNC-WMARKS
Hugealloc Time mean 34300.36 ( +0.00%) 28904.00 ( -15.73%)
Hugealloc Time stddev 36390.42 ( +0.00%) 33464.37 ( -8.04%)
Kbuild Real time 196.13 ( +0.00%) 196.59 ( +0.23%)
Kbuild User time 1234.74 ( +0.00%) 1231.67 ( -0.25%)
Kbuild System time 62.62 ( +0.00%) 59.10 ( -5.54%)
THP fault alloc 57054.53 ( +0.00%) 63223.67 ( +10.81%)
THP fault fallback 11581.40 ( +0.00%) 5412.47 ( -53.26%)
Direct compact fail 107.80 ( +0.00%) 59.07 ( -44.79%)
Direct compact success 4.53 ( +0.00%) 2.80 ( -31.33%)
Direct compact success rate % 3.20 ( +0.00%) 3.99 ( +18.66%)
Compact daemon scanned migrate 5461033.93 ( +0.00%) 2267500.33 ( -58.48%)
Compact daemon scanned free 5824897.93 ( +0.00%) 2339773.00 ( -59.83%)
Compact direct scanned migrate 58336.93 ( +0.00%) 47659.93 ( -18.30%)
Compact direct scanned free 32791.87 ( +0.00%) 40729.67 ( +24.21%)
Compact total migrate scanned 5519370.87 ( +0.00%) 2315160.27 ( -58.05%)
Compact total free scanned 5857689.80 ( +0.00%) 2380502.67 ( -59.36%)
Alloc stall 2424.60 ( +0.00%) 638.87 ( -73.62%)
Pages kswapd scanned 2657018.33 ( +0.00%) 4002186.33 ( +50.63%)
Pages kswapd reclaimed 559583.07 ( +0.00%) 718577.80 ( +28.41%)
Pages direct scanned 722094.07 ( +0.00%) 355172.73 ( -50.81%)
Pages direct reclaimed 107257.80 ( +0.00%) 31162.80 ( -70.95%)
Pages total scanned 3379112.40 ( +0.00%) 4357359.07 ( +28.95%)
Pages total reclaimed 666840.87 ( +0.00%) 749740.60 ( +12.43%)
Swap out 77238.20 ( +0.00%) 110084.33 ( +42.53%)
Swap in 11712.80 ( +0.00%) 24457.00 ( +108.80%)
File refaults 143438.80 ( +0.00%) 188226.93 ( +31.22%)
Also of note is that compaction work overall is reduced. The reason for
this is that when free pageblocks are more readily available, allocations
are also much more likely to get physically placed in LRU order, instead
of being forced to scavenge free space here and there. This means that
reclaim by itself has better chances of freeing up whole blocks, and the
system relies less on compaction.
Comparing all changes to the vanilla kernel:
VANILLA DEFRAGMODE-ASYNC-WMARKS
Hugealloc Time mean 52739.45 ( +0.00%) 28904.00 ( -45.19%)
Hugealloc Time stddev 56541.26 ( +0.00%) 33464.37 ( -40.81%)
Kbuild Real time 197.47 ( +0.00%) 196.59 ( -0.44%)
Kbuild User time 1240.49 ( +0.00%) 1231.67 ( -0.71%)
Kbuild System time 70.08 ( +0.00%) 59.10 ( -15.45%)
THP fault alloc 46727.07 ( +0.00%) 63223.67 ( +35.30%)
THP fault fallback 21910.60 ( +0.00%) 5412.47 ( -75.29%)
Direct compact fail 195.80 ( +0.00%) 59.07 ( -69.48%)
Direct compact success 7.93 ( +0.00%) 2.80 ( -57.46%)
Direct compact success rate % 3.51 ( +0.00%) 3.99 ( +10.49%)
Compact daemon scanned migrate 3369601.27 ( +0.00%) 2267500.33 ( -32.71%)
Compact daemon scanned free 5075474.47 ( +0.00%) 2339773.00 ( -53.90%)
Compact direct scanned migrate 161787.27 ( +0.00%) 47659.93 ( -70.54%)
Compact direct scanned free 163467.53 ( +0.00%) 40729.67 ( -75.08%)
Compact total migrate scanned 3531388.53 ( +0.00%) 2315160.27 ( -34.44%)
Compact total free scanned 5238942.00 ( +0.00%) 2380502.67 ( -54.56%)
Alloc stall 2371.07 ( +0.00%) 638.87 ( -73.02%)
Pages kswapd scanned 2160926.73 ( +0.00%) 4002186.33 ( +85.21%)
Pages kswapd reclaimed 533191.07 ( +0.00%) 718577.80 ( +34.77%)
Pages direct scanned 400450.33 ( +0.00%) 355172.73 ( -11.31%)
Pages direct reclaimed 94441.73 ( +0.00%) 31162.80 ( -67.00%)
Pages total scanned 2561377.07 ( +0.00%) 4357359.07 ( +70.12%)
Pages total reclaimed 627632.80 ( +0.00%) 749740.60 ( +19.46%)
Swap out 47959.53 ( +0.00%) 110084.33 ( +129.53%)
Swap in 7276.00 ( +0.00%) 24457.00 ( +236.10%)
File refaults 138043.00 ( +0.00%) 188226.93 ( +36.35%)
THP allocation latencies and %sys time are down dramatically.
THP allocation failures are down from nearly 50% to 8.5%. And to recall
previous data points, the success rates are steady and reliable without
the cumulative deterioration of fragmentation events.
Compaction work is down overall. Direct compaction work especially is
drastically reduced. As an aside, its success rate of 4% indicates there
is room for improvement. For now it's good to rely on it less.
Reclaim work is up overall, however direct reclaim work is down. Part of
the increase can be attributed to a higher use of THPs, which due to
internal fragmentation increase the memory footprint. This is not
necessarily an unexpected side-effect for users of THP.
However, taken both points together, there may well be some opportunities
for fine tuning in the reclaim/compaction coordination.
[hannes@cmpxchg.org: fix squawks from rebasing]
Link: https://lkml.kernel.org/r/20250314210558.GD1316033@cmpxchg.org
Link: https://lkml.kernel.org/r/20250313210647.1314586-6-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-03-13 17:05:36 -04:00
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-04-06 20:04:49 -07:00
|
|
|
|
mm: page_alloc: defrag_mode kswapd/kcompactd watermarks
The previous patch added pageblock_order reclaim to kswapd/kcompactd,
which helps, but produces only one block at a time. Allocation stalls and
THP failure rates are still higher than they could be.
To adequately reflect ALLOC_NOFRAGMENT demand for pageblocks, change the
watermarking for kswapd & kcompactd: instead of targeting the high
watermark in order-0 pages and checking for one suitable block, simply
require that the high watermark is entirely met in pageblocks.
To this end, track the number of free pages within contiguous pageblocks,
then change pgdat_balanced() and compact_finished() to check watermarks
against this new value.
This further reduces THP latencies and allocation stalls, and improves THP
success rates against the previous patch:
DEFRAGMODE-ASYNC DEFRAGMODE-ASYNC-WMARKS
Hugealloc Time mean 34300.36 ( +0.00%) 28904.00 ( -15.73%)
Hugealloc Time stddev 36390.42 ( +0.00%) 33464.37 ( -8.04%)
Kbuild Real time 196.13 ( +0.00%) 196.59 ( +0.23%)
Kbuild User time 1234.74 ( +0.00%) 1231.67 ( -0.25%)
Kbuild System time 62.62 ( +0.00%) 59.10 ( -5.54%)
THP fault alloc 57054.53 ( +0.00%) 63223.67 ( +10.81%)
THP fault fallback 11581.40 ( +0.00%) 5412.47 ( -53.26%)
Direct compact fail 107.80 ( +0.00%) 59.07 ( -44.79%)
Direct compact success 4.53 ( +0.00%) 2.80 ( -31.33%)
Direct compact success rate % 3.20 ( +0.00%) 3.99 ( +18.66%)
Compact daemon scanned migrate 5461033.93 ( +0.00%) 2267500.33 ( -58.48%)
Compact daemon scanned free 5824897.93 ( +0.00%) 2339773.00 ( -59.83%)
Compact direct scanned migrate 58336.93 ( +0.00%) 47659.93 ( -18.30%)
Compact direct scanned free 32791.87 ( +0.00%) 40729.67 ( +24.21%)
Compact total migrate scanned 5519370.87 ( +0.00%) 2315160.27 ( -58.05%)
Compact total free scanned 5857689.80 ( +0.00%) 2380502.67 ( -59.36%)
Alloc stall 2424.60 ( +0.00%) 638.87 ( -73.62%)
Pages kswapd scanned 2657018.33 ( +0.00%) 4002186.33 ( +50.63%)
Pages kswapd reclaimed 559583.07 ( +0.00%) 718577.80 ( +28.41%)
Pages direct scanned 722094.07 ( +0.00%) 355172.73 ( -50.81%)
Pages direct reclaimed 107257.80 ( +0.00%) 31162.80 ( -70.95%)
Pages total scanned 3379112.40 ( +0.00%) 4357359.07 ( +28.95%)
Pages total reclaimed 666840.87 ( +0.00%) 749740.60 ( +12.43%)
Swap out 77238.20 ( +0.00%) 110084.33 ( +42.53%)
Swap in 11712.80 ( +0.00%) 24457.00 ( +108.80%)
File refaults 143438.80 ( +0.00%) 188226.93 ( +31.22%)
Also of note is that compaction work overall is reduced. The reason for
this is that when free pageblocks are more readily available, allocations
are also much more likely to get physically placed in LRU order, instead
of being forced to scavenge free space here and there. This means that
reclaim by itself has better chances of freeing up whole blocks, and the
system relies less on compaction.
Comparing all changes to the vanilla kernel:
VANILLA DEFRAGMODE-ASYNC-WMARKS
Hugealloc Time mean 52739.45 ( +0.00%) 28904.00 ( -45.19%)
Hugealloc Time stddev 56541.26 ( +0.00%) 33464.37 ( -40.81%)
Kbuild Real time 197.47 ( +0.00%) 196.59 ( -0.44%)
Kbuild User time 1240.49 ( +0.00%) 1231.67 ( -0.71%)
Kbuild System time 70.08 ( +0.00%) 59.10 ( -15.45%)
THP fault alloc 46727.07 ( +0.00%) 63223.67 ( +35.30%)
THP fault fallback 21910.60 ( +0.00%) 5412.47 ( -75.29%)
Direct compact fail 195.80 ( +0.00%) 59.07 ( -69.48%)
Direct compact success 7.93 ( +0.00%) 2.80 ( -57.46%)
Direct compact success rate % 3.51 ( +0.00%) 3.99 ( +10.49%)
Compact daemon scanned migrate 3369601.27 ( +0.00%) 2267500.33 ( -32.71%)
Compact daemon scanned free 5075474.47 ( +0.00%) 2339773.00 ( -53.90%)
Compact direct scanned migrate 161787.27 ( +0.00%) 47659.93 ( -70.54%)
Compact direct scanned free 163467.53 ( +0.00%) 40729.67 ( -75.08%)
Compact total migrate scanned 3531388.53 ( +0.00%) 2315160.27 ( -34.44%)
Compact total free scanned 5238942.00 ( +0.00%) 2380502.67 ( -54.56%)
Alloc stall 2371.07 ( +0.00%) 638.87 ( -73.02%)
Pages kswapd scanned 2160926.73 ( +0.00%) 4002186.33 ( +85.21%)
Pages kswapd reclaimed 533191.07 ( +0.00%) 718577.80 ( +34.77%)
Pages direct scanned 400450.33 ( +0.00%) 355172.73 ( -11.31%)
Pages direct reclaimed 94441.73 ( +0.00%) 31162.80 ( -67.00%)
Pages total scanned 2561377.07 ( +0.00%) 4357359.07 ( +70.12%)
Pages total reclaimed 627632.80 ( +0.00%) 749740.60 ( +19.46%)
Swap out 47959.53 ( +0.00%) 110084.33 ( +129.53%)
Swap in 7276.00 ( +0.00%) 24457.00 ( +236.10%)
File refaults 138043.00 ( +0.00%) 188226.93 ( +36.35%)
THP allocation latencies and %sys time are down dramatically.
THP allocation failures are down from nearly 50% to 8.5%. And to recall
previous data points, the success rates are steady and reliable without
the cumulative deterioration of fragmentation events.
Compaction work is down overall. Direct compaction work especially is
drastically reduced. As an aside, its success rate of 4% indicates there
is room for improvement. For now it's good to rely on it less.
Reclaim work is up overall, however direct reclaim work is down. Part of
the increase can be attributed to a higher use of THPs, which due to
internal fragmentation increase the memory footprint. This is not
necessarily an unexpected side-effect for users of THP.
However, taken both points together, there may well be some opportunities
for fine tuning in the reclaim/compaction coordination.
[hannes@cmpxchg.org: fix squawks from rebasing]
Link: https://lkml.kernel.org/r/20250314210558.GD1316033@cmpxchg.org
Link: https://lkml.kernel.org/r/20250313210647.1314586-6-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-03-13 17:05:36 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-04-06 20:04:49 -07:00
|
|
|
|
|
|
|
|
|
mm/page_alloc: move pages to tail in move_to_free_list()
Whenever we move pages between freelists via move_to_free_list()/
move_freepages_block(), we don't actually touch the pages:
1. Page isolation doesn't actually touch the pages, it simply isolates
pageblocks and moves all free pages to the MIGRATE_ISOLATE freelist.
When undoing isolation, we move the pages back to the target list.
2. Page stealing (steal_suitable_fallback()) moves free pages directly
between lists without touching them.
3. reserve_highatomic_pageblock()/unreserve_highatomic_pageblock() moves
free pages directly between freelists without touching them.
We already place pages to the tail of the freelists when undoing isolation
via __putback_isolated_page(), let's do it in any case (e.g., if order <=
pageblock_order) and document the behavior. To simplify, let's move the
pages to the tail for all move_to_free_list()/move_freepages_block() users.
In 2., the target list is empty, so there should be no change. In 3., we
might observe a change, however, highatomic is more concerned about
allocations succeeding than cache hotness - if we ever realize this change
degrades a workload, we can special-case this instance and add a proper
comment.
This change results in all pages getting onlined via online_pages() to be
placed to the tail of the freelist.
Signed-off-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Wei Yang <richard.weiyang@linux.alibaba.com>
Acked-by: Pankaj Gupta <pankaj.gupta.linux@gmail.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Alexander Duyck <alexander.h.duyck@linux.intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Scott Cheloha <cheloha@linux.ibm.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Haiyang Zhang <haiyangz@microsoft.com>
Cc: "K. Y. Srinivasan" <kys@microsoft.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Stephen Hemminger <sthemmin@microsoft.com>
Cc: Wei Liu <wei.liu@kernel.org>
Link: https://lkml.kernel.org/r/20201005121534.15649-4-david@redhat.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-10-15 20:09:30 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-04-06 20:04:49 -07:00
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
2020-04-06 20:04:49 -07:00
|
|
|
|
|
|
|
|
|
mm: page_alloc: defrag_mode kswapd/kcompactd watermarks
The previous patch added pageblock_order reclaim to kswapd/kcompactd,
which helps, but produces only one block at a time. Allocation stalls and
THP failure rates are still higher than they could be.
To adequately reflect ALLOC_NOFRAGMENT demand for pageblocks, change the
watermarking for kswapd & kcompactd: instead of targeting the high
watermark in order-0 pages and checking for one suitable block, simply
require that the high watermark is entirely met in pageblocks.
To this end, track the number of free pages within contiguous pageblocks,
then change pgdat_balanced() and compact_finished() to check watermarks
against this new value.
This further reduces THP latencies and allocation stalls, and improves THP
success rates against the previous patch:
DEFRAGMODE-ASYNC DEFRAGMODE-ASYNC-WMARKS
Hugealloc Time mean 34300.36 ( +0.00%) 28904.00 ( -15.73%)
Hugealloc Time stddev 36390.42 ( +0.00%) 33464.37 ( -8.04%)
Kbuild Real time 196.13 ( +0.00%) 196.59 ( +0.23%)
Kbuild User time 1234.74 ( +0.00%) 1231.67 ( -0.25%)
Kbuild System time 62.62 ( +0.00%) 59.10 ( -5.54%)
THP fault alloc 57054.53 ( +0.00%) 63223.67 ( +10.81%)
THP fault fallback 11581.40 ( +0.00%) 5412.47 ( -53.26%)
Direct compact fail 107.80 ( +0.00%) 59.07 ( -44.79%)
Direct compact success 4.53 ( +0.00%) 2.80 ( -31.33%)
Direct compact success rate % 3.20 ( +0.00%) 3.99 ( +18.66%)
Compact daemon scanned migrate 5461033.93 ( +0.00%) 2267500.33 ( -58.48%)
Compact daemon scanned free 5824897.93 ( +0.00%) 2339773.00 ( -59.83%)
Compact direct scanned migrate 58336.93 ( +0.00%) 47659.93 ( -18.30%)
Compact direct scanned free 32791.87 ( +0.00%) 40729.67 ( +24.21%)
Compact total migrate scanned 5519370.87 ( +0.00%) 2315160.27 ( -58.05%)
Compact total free scanned 5857689.80 ( +0.00%) 2380502.67 ( -59.36%)
Alloc stall 2424.60 ( +0.00%) 638.87 ( -73.62%)
Pages kswapd scanned 2657018.33 ( +0.00%) 4002186.33 ( +50.63%)
Pages kswapd reclaimed 559583.07 ( +0.00%) 718577.80 ( +28.41%)
Pages direct scanned 722094.07 ( +0.00%) 355172.73 ( -50.81%)
Pages direct reclaimed 107257.80 ( +0.00%) 31162.80 ( -70.95%)
Pages total scanned 3379112.40 ( +0.00%) 4357359.07 ( +28.95%)
Pages total reclaimed 666840.87 ( +0.00%) 749740.60 ( +12.43%)
Swap out 77238.20 ( +0.00%) 110084.33 ( +42.53%)
Swap in 11712.80 ( +0.00%) 24457.00 ( +108.80%)
File refaults 143438.80 ( +0.00%) 188226.93 ( +31.22%)
Also of note is that compaction work overall is reduced. The reason for
this is that when free pageblocks are more readily available, allocations
are also much more likely to get physically placed in LRU order, instead
of being forced to scavenge free space here and there. This means that
reclaim by itself has better chances of freeing up whole blocks, and the
system relies less on compaction.
Comparing all changes to the vanilla kernel:
VANILLA DEFRAGMODE-ASYNC-WMARKS
Hugealloc Time mean 52739.45 ( +0.00%) 28904.00 ( -45.19%)
Hugealloc Time stddev 56541.26 ( +0.00%) 33464.37 ( -40.81%)
Kbuild Real time 197.47 ( +0.00%) 196.59 ( -0.44%)
Kbuild User time 1240.49 ( +0.00%) 1231.67 ( -0.71%)
Kbuild System time 70.08 ( +0.00%) 59.10 ( -15.45%)
THP fault alloc 46727.07 ( +0.00%) 63223.67 ( +35.30%)
THP fault fallback 21910.60 ( +0.00%) 5412.47 ( -75.29%)
Direct compact fail 195.80 ( +0.00%) 59.07 ( -69.48%)
Direct compact success 7.93 ( +0.00%) 2.80 ( -57.46%)
Direct compact success rate % 3.51 ( +0.00%) 3.99 ( +10.49%)
Compact daemon scanned migrate 3369601.27 ( +0.00%) 2267500.33 ( -32.71%)
Compact daemon scanned free 5075474.47 ( +0.00%) 2339773.00 ( -53.90%)
Compact direct scanned migrate 161787.27 ( +0.00%) 47659.93 ( -70.54%)
Compact direct scanned free 163467.53 ( +0.00%) 40729.67 ( -75.08%)
Compact total migrate scanned 3531388.53 ( +0.00%) 2315160.27 ( -34.44%)
Compact total free scanned 5238942.00 ( +0.00%) 2380502.67 ( -54.56%)
Alloc stall 2371.07 ( +0.00%) 638.87 ( -73.02%)
Pages kswapd scanned 2160926.73 ( +0.00%) 4002186.33 ( +85.21%)
Pages kswapd reclaimed 533191.07 ( +0.00%) 718577.80 ( +34.77%)
Pages direct scanned 400450.33 ( +0.00%) 355172.73 ( -11.31%)
Pages direct reclaimed 94441.73 ( +0.00%) 31162.80 ( -67.00%)
Pages total scanned 2561377.07 ( +0.00%) 4357359.07 ( +70.12%)
Pages total reclaimed 627632.80 ( +0.00%) 749740.60 ( +19.46%)
Swap out 47959.53 ( +0.00%) 110084.33 ( +129.53%)
Swap in 7276.00 ( +0.00%) 24457.00 ( +236.10%)
File refaults 138043.00 ( +0.00%) 188226.93 ( +36.35%)
THP allocation latencies and %sys time are down dramatically.
THP allocation failures are down from nearly 50% to 8.5%. And to recall
previous data points, the success rates are steady and reliable without
the cumulative deterioration of fragmentation events.
Compaction work is down overall. Direct compaction work especially is
drastically reduced. As an aside, its success rate of 4% indicates there
is room for improvement. For now it's good to rely on it less.
Reclaim work is up overall, however direct reclaim work is down. Part of
the increase can be attributed to a higher use of THPs, which due to
internal fragmentation increase the memory footprint. This is not
necessarily an unexpected side-effect for users of THP.
However, taken both points together, there may well be some opportunities
for fine tuning in the reclaim/compaction coordination.
[hannes@cmpxchg.org: fix squawks from rebasing]
Link: https://lkml.kernel.org/r/20250314210558.GD1316033@cmpxchg.org
Link: https://lkml.kernel.org/r/20250313210647.1314586-6-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-03-13 17:05:36 -04:00
|
|
|
|
2020-04-06 20:04:49 -07:00
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
|
|
|
|
|
mm/page_alloc: pageblock flags functions clean up
Patch series "Make MIGRATE_ISOLATE a standalone bit", v10.
This patchset moves MIGRATE_ISOLATE to a standalone bit to avoid being
overwritten during pageblock isolation process. Currently,
MIGRATE_ISOLATE is part of enum migratetype (in include/linux/mmzone.h),
thus, setting a pageblock to MIGRATE_ISOLATE overwrites its original
migratetype. This causes pageblock migratetype loss during
alloc_contig_range() and memory offline, especially when the process fails
due to a failed pageblock isolation and the code tries to undo the
finished pageblock isolations.
In terms of performance for changing pageblock types, no performance
change is observed:
1. I used perf to collect stats of offlining and onlining all memory
of a 40GB VM 10 times and see that get_pfnblock_flags_mask() and
set_pfnblock_flags_mask() take about 0.12% and 0.02% of the whole
process respectively with and without this patchset across 3 runs.
2. I used perf to collect stats of dd from /dev/random to a 40GB tmpfs
file and find get_pfnblock_flags_mask() takes about 0.05% of the
process with and without this patchset across 3 runs.
This patch (of 6):
No functional change is intended.
1. Add __NR_PAGEBLOCK_BITS for the number of pageblock flag bits and use
roundup_pow_of_two(__NR_PAGEBLOCK_BITS) as NR_PAGEBLOCK_BITS to take
right amount of bits for pageblock flags.
2. Rename PB_migrate_skip to PB_compact_skip.
3. Add {get,set,clear}_pfnblock_bit() to operate one a standalone bit,
like PB_compact_skip.
3. Make {get,set}_pfnblock_flags_mask() internal functions and use
{get,set}_pfnblock_migratetype() for pageblock migratetype operations.
4. Move pageblock flags common code to get_pfnblock_bitmap_bitidx().
3. Use MIGRATETYPE_MASK to get the migratetype of a pageblock from its
flags.
4. Use PB_migrate_end in the definition of MIGRATETYPE_MASK instead of
PB_migrate_bits.
5. Add a comment on is_migrate_cma_folio() to prevent one from changing it
to use get_pageblock_migratetype() and causing issues.
Link: https://lkml.kernel.org/r/20250617021115.2331563-1-ziy@nvidia.com
Link: https://lkml.kernel.org/r/20250617021115.2331563-2-ziy@nvidia.com
Signed-off-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kirill A. Shuemov <kirill.shutemov@linux.intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Richard Chang <richardycc@google.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-06-16 22:11:09 -04:00
|
|
|
|
mm: page_alloc: defrag_mode kswapd/kcompactd watermarks
The previous patch added pageblock_order reclaim to kswapd/kcompactd,
which helps, but produces only one block at a time. Allocation stalls and
THP failure rates are still higher than they could be.
To adequately reflect ALLOC_NOFRAGMENT demand for pageblocks, change the
watermarking for kswapd & kcompactd: instead of targeting the high
watermark in order-0 pages and checking for one suitable block, simply
require that the high watermark is entirely met in pageblocks.
To this end, track the number of free pages within contiguous pageblocks,
then change pgdat_balanced() and compact_finished() to check watermarks
against this new value.
This further reduces THP latencies and allocation stalls, and improves THP
success rates against the previous patch:
DEFRAGMODE-ASYNC DEFRAGMODE-ASYNC-WMARKS
Hugealloc Time mean 34300.36 ( +0.00%) 28904.00 ( -15.73%)
Hugealloc Time stddev 36390.42 ( +0.00%) 33464.37 ( -8.04%)
Kbuild Real time 196.13 ( +0.00%) 196.59 ( +0.23%)
Kbuild User time 1234.74 ( +0.00%) 1231.67 ( -0.25%)
Kbuild System time 62.62 ( +0.00%) 59.10 ( -5.54%)
THP fault alloc 57054.53 ( +0.00%) 63223.67 ( +10.81%)
THP fault fallback 11581.40 ( +0.00%) 5412.47 ( -53.26%)
Direct compact fail 107.80 ( +0.00%) 59.07 ( -44.79%)
Direct compact success 4.53 ( +0.00%) 2.80 ( -31.33%)
Direct compact success rate % 3.20 ( +0.00%) 3.99 ( +18.66%)
Compact daemon scanned migrate 5461033.93 ( +0.00%) 2267500.33 ( -58.48%)
Compact daemon scanned free 5824897.93 ( +0.00%) 2339773.00 ( -59.83%)
Compact direct scanned migrate 58336.93 ( +0.00%) 47659.93 ( -18.30%)
Compact direct scanned free 32791.87 ( +0.00%) 40729.67 ( +24.21%)
Compact total migrate scanned 5519370.87 ( +0.00%) 2315160.27 ( -58.05%)
Compact total free scanned 5857689.80 ( +0.00%) 2380502.67 ( -59.36%)
Alloc stall 2424.60 ( +0.00%) 638.87 ( -73.62%)
Pages kswapd scanned 2657018.33 ( +0.00%) 4002186.33 ( +50.63%)
Pages kswapd reclaimed 559583.07 ( +0.00%) 718577.80 ( +28.41%)
Pages direct scanned 722094.07 ( +0.00%) 355172.73 ( -50.81%)
Pages direct reclaimed 107257.80 ( +0.00%) 31162.80 ( -70.95%)
Pages total scanned 3379112.40 ( +0.00%) 4357359.07 ( +28.95%)
Pages total reclaimed 666840.87 ( +0.00%) 749740.60 ( +12.43%)
Swap out 77238.20 ( +0.00%) 110084.33 ( +42.53%)
Swap in 11712.80 ( +0.00%) 24457.00 ( +108.80%)
File refaults 143438.80 ( +0.00%) 188226.93 ( +31.22%)
Also of note is that compaction work overall is reduced. The reason for
this is that when free pageblocks are more readily available, allocations
are also much more likely to get physically placed in LRU order, instead
of being forced to scavenge free space here and there. This means that
reclaim by itself has better chances of freeing up whole blocks, and the
system relies less on compaction.
Comparing all changes to the vanilla kernel:
VANILLA DEFRAGMODE-ASYNC-WMARKS
Hugealloc Time mean 52739.45 ( +0.00%) 28904.00 ( -45.19%)
Hugealloc Time stddev 56541.26 ( +0.00%) 33464.37 ( -40.81%)
Kbuild Real time 197.47 ( +0.00%) 196.59 ( -0.44%)
Kbuild User time 1240.49 ( +0.00%) 1231.67 ( -0.71%)
Kbuild System time 70.08 ( +0.00%) 59.10 ( -15.45%)
THP fault alloc 46727.07 ( +0.00%) 63223.67 ( +35.30%)
THP fault fallback 21910.60 ( +0.00%) 5412.47 ( -75.29%)
Direct compact fail 195.80 ( +0.00%) 59.07 ( -69.48%)
Direct compact success 7.93 ( +0.00%) 2.80 ( -57.46%)
Direct compact success rate % 3.51 ( +0.00%) 3.99 ( +10.49%)
Compact daemon scanned migrate 3369601.27 ( +0.00%) 2267500.33 ( -32.71%)
Compact daemon scanned free 5075474.47 ( +0.00%) 2339773.00 ( -53.90%)
Compact direct scanned migrate 161787.27 ( +0.00%) 47659.93 ( -70.54%)
Compact direct scanned free 163467.53 ( +0.00%) 40729.67 ( -75.08%)
Compact total migrate scanned 3531388.53 ( +0.00%) 2315160.27 ( -34.44%)
Compact total free scanned 5238942.00 ( +0.00%) 2380502.67 ( -54.56%)
Alloc stall 2371.07 ( +0.00%) 638.87 ( -73.02%)
Pages kswapd scanned 2160926.73 ( +0.00%) 4002186.33 ( +85.21%)
Pages kswapd reclaimed 533191.07 ( +0.00%) 718577.80 ( +34.77%)
Pages direct scanned 400450.33 ( +0.00%) 355172.73 ( -11.31%)
Pages direct reclaimed 94441.73 ( +0.00%) 31162.80 ( -67.00%)
Pages total scanned 2561377.07 ( +0.00%) 4357359.07 ( +70.12%)
Pages total reclaimed 627632.80 ( +0.00%) 749740.60 ( +19.46%)
Swap out 47959.53 ( +0.00%) 110084.33 ( +129.53%)
Swap in 7276.00 ( +0.00%) 24457.00 ( +236.10%)
File refaults 138043.00 ( +0.00%) 188226.93 ( +36.35%)
THP allocation latencies and %sys time are down dramatically.
THP allocation failures are down from nearly 50% to 8.5%. And to recall
previous data points, the success rates are steady and reliable without
the cumulative deterioration of fragmentation events.
Compaction work is down overall. Direct compaction work especially is
drastically reduced. As an aside, its success rate of 4% indicates there
is room for improvement. For now it's good to rely on it less.
Reclaim work is up overall, however direct reclaim work is down. Part of
the increase can be attributed to a higher use of THPs, which due to
internal fragmentation increase the memory footprint. This is not
necessarily an unexpected side-effect for users of THP.
However, taken both points together, there may well be some opportunities
for fine tuning in the reclaim/compaction coordination.
[hannes@cmpxchg.org: fix squawks from rebasing]
Link: https://lkml.kernel.org/r/20250314210558.GD1316033@cmpxchg.org
Link: https://lkml.kernel.org/r/20250313210647.1314586-6-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-03-13 17:05:36 -04:00
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: page_alloc: defrag_mode kswapd/kcompactd watermarks
The previous patch added pageblock_order reclaim to kswapd/kcompactd,
which helps, but produces only one block at a time. Allocation stalls and
THP failure rates are still higher than they could be.
To adequately reflect ALLOC_NOFRAGMENT demand for pageblocks, change the
watermarking for kswapd & kcompactd: instead of targeting the high
watermark in order-0 pages and checking for one suitable block, simply
require that the high watermark is entirely met in pageblocks.
To this end, track the number of free pages within contiguous pageblocks,
then change pgdat_balanced() and compact_finished() to check watermarks
against this new value.
This further reduces THP latencies and allocation stalls, and improves THP
success rates against the previous patch:
DEFRAGMODE-ASYNC DEFRAGMODE-ASYNC-WMARKS
Hugealloc Time mean 34300.36 ( +0.00%) 28904.00 ( -15.73%)
Hugealloc Time stddev 36390.42 ( +0.00%) 33464.37 ( -8.04%)
Kbuild Real time 196.13 ( +0.00%) 196.59 ( +0.23%)
Kbuild User time 1234.74 ( +0.00%) 1231.67 ( -0.25%)
Kbuild System time 62.62 ( +0.00%) 59.10 ( -5.54%)
THP fault alloc 57054.53 ( +0.00%) 63223.67 ( +10.81%)
THP fault fallback 11581.40 ( +0.00%) 5412.47 ( -53.26%)
Direct compact fail 107.80 ( +0.00%) 59.07 ( -44.79%)
Direct compact success 4.53 ( +0.00%) 2.80 ( -31.33%)
Direct compact success rate % 3.20 ( +0.00%) 3.99 ( +18.66%)
Compact daemon scanned migrate 5461033.93 ( +0.00%) 2267500.33 ( -58.48%)
Compact daemon scanned free 5824897.93 ( +0.00%) 2339773.00 ( -59.83%)
Compact direct scanned migrate 58336.93 ( +0.00%) 47659.93 ( -18.30%)
Compact direct scanned free 32791.87 ( +0.00%) 40729.67 ( +24.21%)
Compact total migrate scanned 5519370.87 ( +0.00%) 2315160.27 ( -58.05%)
Compact total free scanned 5857689.80 ( +0.00%) 2380502.67 ( -59.36%)
Alloc stall 2424.60 ( +0.00%) 638.87 ( -73.62%)
Pages kswapd scanned 2657018.33 ( +0.00%) 4002186.33 ( +50.63%)
Pages kswapd reclaimed 559583.07 ( +0.00%) 718577.80 ( +28.41%)
Pages direct scanned 722094.07 ( +0.00%) 355172.73 ( -50.81%)
Pages direct reclaimed 107257.80 ( +0.00%) 31162.80 ( -70.95%)
Pages total scanned 3379112.40 ( +0.00%) 4357359.07 ( +28.95%)
Pages total reclaimed 666840.87 ( +0.00%) 749740.60 ( +12.43%)
Swap out 77238.20 ( +0.00%) 110084.33 ( +42.53%)
Swap in 11712.80 ( +0.00%) 24457.00 ( +108.80%)
File refaults 143438.80 ( +0.00%) 188226.93 ( +31.22%)
Also of note is that compaction work overall is reduced. The reason for
this is that when free pageblocks are more readily available, allocations
are also much more likely to get physically placed in LRU order, instead
of being forced to scavenge free space here and there. This means that
reclaim by itself has better chances of freeing up whole blocks, and the
system relies less on compaction.
Comparing all changes to the vanilla kernel:
VANILLA DEFRAGMODE-ASYNC-WMARKS
Hugealloc Time mean 52739.45 ( +0.00%) 28904.00 ( -45.19%)
Hugealloc Time stddev 56541.26 ( +0.00%) 33464.37 ( -40.81%)
Kbuild Real time 197.47 ( +0.00%) 196.59 ( -0.44%)
Kbuild User time 1240.49 ( +0.00%) 1231.67 ( -0.71%)
Kbuild System time 70.08 ( +0.00%) 59.10 ( -15.45%)
THP fault alloc 46727.07 ( +0.00%) 63223.67 ( +35.30%)
THP fault fallback 21910.60 ( +0.00%) 5412.47 ( -75.29%)
Direct compact fail 195.80 ( +0.00%) 59.07 ( -69.48%)
Direct compact success 7.93 ( +0.00%) 2.80 ( -57.46%)
Direct compact success rate % 3.51 ( +0.00%) 3.99 ( +10.49%)
Compact daemon scanned migrate 3369601.27 ( +0.00%) 2267500.33 ( -32.71%)
Compact daemon scanned free 5075474.47 ( +0.00%) 2339773.00 ( -53.90%)
Compact direct scanned migrate 161787.27 ( +0.00%) 47659.93 ( -70.54%)
Compact direct scanned free 163467.53 ( +0.00%) 40729.67 ( -75.08%)
Compact total migrate scanned 3531388.53 ( +0.00%) 2315160.27 ( -34.44%)
Compact total free scanned 5238942.00 ( +0.00%) 2380502.67 ( -54.56%)
Alloc stall 2371.07 ( +0.00%) 638.87 ( -73.02%)
Pages kswapd scanned 2160926.73 ( +0.00%) 4002186.33 ( +85.21%)
Pages kswapd reclaimed 533191.07 ( +0.00%) 718577.80 ( +34.77%)
Pages direct scanned 400450.33 ( +0.00%) 355172.73 ( -11.31%)
Pages direct reclaimed 94441.73 ( +0.00%) 31162.80 ( -67.00%)
Pages total scanned 2561377.07 ( +0.00%) 4357359.07 ( +70.12%)
Pages total reclaimed 627632.80 ( +0.00%) 749740.60 ( +19.46%)
Swap out 47959.53 ( +0.00%) 110084.33 ( +129.53%)
Swap in 7276.00 ( +0.00%) 24457.00 ( +236.10%)
File refaults 138043.00 ( +0.00%) 188226.93 ( +36.35%)
THP allocation latencies and %sys time are down dramatically.
THP allocation failures are down from nearly 50% to 8.5%. And to recall
previous data points, the success rates are steady and reliable without
the cumulative deterioration of fragmentation events.
Compaction work is down overall. Direct compaction work especially is
drastically reduced. As an aside, its success rate of 4% indicates there
is room for improvement. For now it's good to rely on it less.
Reclaim work is up overall, however direct reclaim work is down. Part of
the increase can be attributed to a higher use of THPs, which due to
internal fragmentation increase the memory footprint. This is not
necessarily an unexpected side-effect for users of THP.
However, taken both points together, there may well be some opportunities
for fine tuning in the reclaim/compaction coordination.
[hannes@cmpxchg.org: fix squawks from rebasing]
Link: https://lkml.kernel.org/r/20250314210558.GD1316033@cmpxchg.org
Link: https://lkml.kernel.org/r/20250313210647.1314586-6-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-03-13 17:05:36 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-04-06 20:04:49 -07:00
|
|
|
|
|
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
|
|
|
|
|
2020-04-06 20:04:49 -07:00
|
|
|
|
mm: page_alloc: defrag_mode kswapd/kcompactd watermarks
The previous patch added pageblock_order reclaim to kswapd/kcompactd,
which helps, but produces only one block at a time. Allocation stalls and
THP failure rates are still higher than they could be.
To adequately reflect ALLOC_NOFRAGMENT demand for pageblocks, change the
watermarking for kswapd & kcompactd: instead of targeting the high
watermark in order-0 pages and checking for one suitable block, simply
require that the high watermark is entirely met in pageblocks.
To this end, track the number of free pages within contiguous pageblocks,
then change pgdat_balanced() and compact_finished() to check watermarks
against this new value.
This further reduces THP latencies and allocation stalls, and improves THP
success rates against the previous patch:
DEFRAGMODE-ASYNC DEFRAGMODE-ASYNC-WMARKS
Hugealloc Time mean 34300.36 ( +0.00%) 28904.00 ( -15.73%)
Hugealloc Time stddev 36390.42 ( +0.00%) 33464.37 ( -8.04%)
Kbuild Real time 196.13 ( +0.00%) 196.59 ( +0.23%)
Kbuild User time 1234.74 ( +0.00%) 1231.67 ( -0.25%)
Kbuild System time 62.62 ( +0.00%) 59.10 ( -5.54%)
THP fault alloc 57054.53 ( +0.00%) 63223.67 ( +10.81%)
THP fault fallback 11581.40 ( +0.00%) 5412.47 ( -53.26%)
Direct compact fail 107.80 ( +0.00%) 59.07 ( -44.79%)
Direct compact success 4.53 ( +0.00%) 2.80 ( -31.33%)
Direct compact success rate % 3.20 ( +0.00%) 3.99 ( +18.66%)
Compact daemon scanned migrate 5461033.93 ( +0.00%) 2267500.33 ( -58.48%)
Compact daemon scanned free 5824897.93 ( +0.00%) 2339773.00 ( -59.83%)
Compact direct scanned migrate 58336.93 ( +0.00%) 47659.93 ( -18.30%)
Compact direct scanned free 32791.87 ( +0.00%) 40729.67 ( +24.21%)
Compact total migrate scanned 5519370.87 ( +0.00%) 2315160.27 ( -58.05%)
Compact total free scanned 5857689.80 ( +0.00%) 2380502.67 ( -59.36%)
Alloc stall 2424.60 ( +0.00%) 638.87 ( -73.62%)
Pages kswapd scanned 2657018.33 ( +0.00%) 4002186.33 ( +50.63%)
Pages kswapd reclaimed 559583.07 ( +0.00%) 718577.80 ( +28.41%)
Pages direct scanned 722094.07 ( +0.00%) 355172.73 ( -50.81%)
Pages direct reclaimed 107257.80 ( +0.00%) 31162.80 ( -70.95%)
Pages total scanned 3379112.40 ( +0.00%) 4357359.07 ( +28.95%)
Pages total reclaimed 666840.87 ( +0.00%) 749740.60 ( +12.43%)
Swap out 77238.20 ( +0.00%) 110084.33 ( +42.53%)
Swap in 11712.80 ( +0.00%) 24457.00 ( +108.80%)
File refaults 143438.80 ( +0.00%) 188226.93 ( +31.22%)
Also of note is that compaction work overall is reduced. The reason for
this is that when free pageblocks are more readily available, allocations
are also much more likely to get physically placed in LRU order, instead
of being forced to scavenge free space here and there. This means that
reclaim by itself has better chances of freeing up whole blocks, and the
system relies less on compaction.
Comparing all changes to the vanilla kernel:
VANILLA DEFRAGMODE-ASYNC-WMARKS
Hugealloc Time mean 52739.45 ( +0.00%) 28904.00 ( -45.19%)
Hugealloc Time stddev 56541.26 ( +0.00%) 33464.37 ( -40.81%)
Kbuild Real time 197.47 ( +0.00%) 196.59 ( -0.44%)
Kbuild User time 1240.49 ( +0.00%) 1231.67 ( -0.71%)
Kbuild System time 70.08 ( +0.00%) 59.10 ( -15.45%)
THP fault alloc 46727.07 ( +0.00%) 63223.67 ( +35.30%)
THP fault fallback 21910.60 ( +0.00%) 5412.47 ( -75.29%)
Direct compact fail 195.80 ( +0.00%) 59.07 ( -69.48%)
Direct compact success 7.93 ( +0.00%) 2.80 ( -57.46%)
Direct compact success rate % 3.51 ( +0.00%) 3.99 ( +10.49%)
Compact daemon scanned migrate 3369601.27 ( +0.00%) 2267500.33 ( -32.71%)
Compact daemon scanned free 5075474.47 ( +0.00%) 2339773.00 ( -53.90%)
Compact direct scanned migrate 161787.27 ( +0.00%) 47659.93 ( -70.54%)
Compact direct scanned free 163467.53 ( +0.00%) 40729.67 ( -75.08%)
Compact total migrate scanned 3531388.53 ( +0.00%) 2315160.27 ( -34.44%)
Compact total free scanned 5238942.00 ( +0.00%) 2380502.67 ( -54.56%)
Alloc stall 2371.07 ( +0.00%) 638.87 ( -73.02%)
Pages kswapd scanned 2160926.73 ( +0.00%) 4002186.33 ( +85.21%)
Pages kswapd reclaimed 533191.07 ( +0.00%) 718577.80 ( +34.77%)
Pages direct scanned 400450.33 ( +0.00%) 355172.73 ( -11.31%)
Pages direct reclaimed 94441.73 ( +0.00%) 31162.80 ( -67.00%)
Pages total scanned 2561377.07 ( +0.00%) 4357359.07 ( +70.12%)
Pages total reclaimed 627632.80 ( +0.00%) 749740.60 ( +19.46%)
Swap out 47959.53 ( +0.00%) 110084.33 ( +129.53%)
Swap in 7276.00 ( +0.00%) 24457.00 ( +236.10%)
File refaults 138043.00 ( +0.00%) 188226.93 ( +36.35%)
THP allocation latencies and %sys time are down dramatically.
THP allocation failures are down from nearly 50% to 8.5%. And to recall
previous data points, the success rates are steady and reliable without
the cumulative deterioration of fragmentation events.
Compaction work is down overall. Direct compaction work especially is
drastically reduced. As an aside, its success rate of 4% indicates there
is room for improvement. For now it's good to rely on it less.
Reclaim work is up overall, however direct reclaim work is down. Part of
the increase can be attributed to a higher use of THPs, which due to
internal fragmentation increase the memory footprint. This is not
necessarily an unexpected side-effect for users of THP.
However, taken both points together, there may well be some opportunities
for fine tuning in the reclaim/compaction coordination.
[hannes@cmpxchg.org: fix squawks from rebasing]
Link: https://lkml.kernel.org/r/20250314210558.GD1316033@cmpxchg.org
Link: https://lkml.kernel.org/r/20250313210647.1314586-6-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-03-13 17:05:36 -04:00
|
|
|
|
|
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
mm/page_alloc: pageblock flags functions clean up
Patch series "Make MIGRATE_ISOLATE a standalone bit", v10.
This patchset moves MIGRATE_ISOLATE to a standalone bit to avoid being
overwritten during pageblock isolation process. Currently,
MIGRATE_ISOLATE is part of enum migratetype (in include/linux/mmzone.h),
thus, setting a pageblock to MIGRATE_ISOLATE overwrites its original
migratetype. This causes pageblock migratetype loss during
alloc_contig_range() and memory offline, especially when the process fails
due to a failed pageblock isolation and the code tries to undo the
finished pageblock isolations.
In terms of performance for changing pageblock types, no performance
change is observed:
1. I used perf to collect stats of offlining and onlining all memory
of a 40GB VM 10 times and see that get_pfnblock_flags_mask() and
set_pfnblock_flags_mask() take about 0.12% and 0.02% of the whole
process respectively with and without this patchset across 3 runs.
2. I used perf to collect stats of dd from /dev/random to a 40GB tmpfs
file and find get_pfnblock_flags_mask() takes about 0.05% of the
process with and without this patchset across 3 runs.
This patch (of 6):
No functional change is intended.
1. Add __NR_PAGEBLOCK_BITS for the number of pageblock flag bits and use
roundup_pow_of_two(__NR_PAGEBLOCK_BITS) as NR_PAGEBLOCK_BITS to take
right amount of bits for pageblock flags.
2. Rename PB_migrate_skip to PB_compact_skip.
3. Add {get,set,clear}_pfnblock_bit() to operate one a standalone bit,
like PB_compact_skip.
3. Make {get,set}_pfnblock_flags_mask() internal functions and use
{get,set}_pfnblock_migratetype() for pageblock migratetype operations.
4. Move pageblock flags common code to get_pfnblock_bitmap_bitidx().
3. Use MIGRATETYPE_MASK to get the migratetype of a pageblock from its
flags.
4. Use PB_migrate_end in the definition of MIGRATETYPE_MASK instead of
PB_migrate_bits.
5. Add a comment on is_migrate_cma_folio() to prevent one from changing it
to use get_pageblock_migratetype() and causing issues.
Link: https://lkml.kernel.org/r/20250617021115.2331563-1-ziy@nvidia.com
Link: https://lkml.kernel.org/r/20250617021115.2331563-2-ziy@nvidia.com
Signed-off-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kirill A. Shuemov <kirill.shutemov@linux.intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Richard Chang <richardycc@google.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-06-16 22:11:09 -04:00
|
|
|
|
mm: page_alloc: defrag_mode kswapd/kcompactd watermarks
The previous patch added pageblock_order reclaim to kswapd/kcompactd,
which helps, but produces only one block at a time. Allocation stalls and
THP failure rates are still higher than they could be.
To adequately reflect ALLOC_NOFRAGMENT demand for pageblocks, change the
watermarking for kswapd & kcompactd: instead of targeting the high
watermark in order-0 pages and checking for one suitable block, simply
require that the high watermark is entirely met in pageblocks.
To this end, track the number of free pages within contiguous pageblocks,
then change pgdat_balanced() and compact_finished() to check watermarks
against this new value.
This further reduces THP latencies and allocation stalls, and improves THP
success rates against the previous patch:
DEFRAGMODE-ASYNC DEFRAGMODE-ASYNC-WMARKS
Hugealloc Time mean 34300.36 ( +0.00%) 28904.00 ( -15.73%)
Hugealloc Time stddev 36390.42 ( +0.00%) 33464.37 ( -8.04%)
Kbuild Real time 196.13 ( +0.00%) 196.59 ( +0.23%)
Kbuild User time 1234.74 ( +0.00%) 1231.67 ( -0.25%)
Kbuild System time 62.62 ( +0.00%) 59.10 ( -5.54%)
THP fault alloc 57054.53 ( +0.00%) 63223.67 ( +10.81%)
THP fault fallback 11581.40 ( +0.00%) 5412.47 ( -53.26%)
Direct compact fail 107.80 ( +0.00%) 59.07 ( -44.79%)
Direct compact success 4.53 ( +0.00%) 2.80 ( -31.33%)
Direct compact success rate % 3.20 ( +0.00%) 3.99 ( +18.66%)
Compact daemon scanned migrate 5461033.93 ( +0.00%) 2267500.33 ( -58.48%)
Compact daemon scanned free 5824897.93 ( +0.00%) 2339773.00 ( -59.83%)
Compact direct scanned migrate 58336.93 ( +0.00%) 47659.93 ( -18.30%)
Compact direct scanned free 32791.87 ( +0.00%) 40729.67 ( +24.21%)
Compact total migrate scanned 5519370.87 ( +0.00%) 2315160.27 ( -58.05%)
Compact total free scanned 5857689.80 ( +0.00%) 2380502.67 ( -59.36%)
Alloc stall 2424.60 ( +0.00%) 638.87 ( -73.62%)
Pages kswapd scanned 2657018.33 ( +0.00%) 4002186.33 ( +50.63%)
Pages kswapd reclaimed 559583.07 ( +0.00%) 718577.80 ( +28.41%)
Pages direct scanned 722094.07 ( +0.00%) 355172.73 ( -50.81%)
Pages direct reclaimed 107257.80 ( +0.00%) 31162.80 ( -70.95%)
Pages total scanned 3379112.40 ( +0.00%) 4357359.07 ( +28.95%)
Pages total reclaimed 666840.87 ( +0.00%) 749740.60 ( +12.43%)
Swap out 77238.20 ( +0.00%) 110084.33 ( +42.53%)
Swap in 11712.80 ( +0.00%) 24457.00 ( +108.80%)
File refaults 143438.80 ( +0.00%) 188226.93 ( +31.22%)
Also of note is that compaction work overall is reduced. The reason for
this is that when free pageblocks are more readily available, allocations
are also much more likely to get physically placed in LRU order, instead
of being forced to scavenge free space here and there. This means that
reclaim by itself has better chances of freeing up whole blocks, and the
system relies less on compaction.
Comparing all changes to the vanilla kernel:
VANILLA DEFRAGMODE-ASYNC-WMARKS
Hugealloc Time mean 52739.45 ( +0.00%) 28904.00 ( -45.19%)
Hugealloc Time stddev 56541.26 ( +0.00%) 33464.37 ( -40.81%)
Kbuild Real time 197.47 ( +0.00%) 196.59 ( -0.44%)
Kbuild User time 1240.49 ( +0.00%) 1231.67 ( -0.71%)
Kbuild System time 70.08 ( +0.00%) 59.10 ( -15.45%)
THP fault alloc 46727.07 ( +0.00%) 63223.67 ( +35.30%)
THP fault fallback 21910.60 ( +0.00%) 5412.47 ( -75.29%)
Direct compact fail 195.80 ( +0.00%) 59.07 ( -69.48%)
Direct compact success 7.93 ( +0.00%) 2.80 ( -57.46%)
Direct compact success rate % 3.51 ( +0.00%) 3.99 ( +10.49%)
Compact daemon scanned migrate 3369601.27 ( +0.00%) 2267500.33 ( -32.71%)
Compact daemon scanned free 5075474.47 ( +0.00%) 2339773.00 ( -53.90%)
Compact direct scanned migrate 161787.27 ( +0.00%) 47659.93 ( -70.54%)
Compact direct scanned free 163467.53 ( +0.00%) 40729.67 ( -75.08%)
Compact total migrate scanned 3531388.53 ( +0.00%) 2315160.27 ( -34.44%)
Compact total free scanned 5238942.00 ( +0.00%) 2380502.67 ( -54.56%)
Alloc stall 2371.07 ( +0.00%) 638.87 ( -73.02%)
Pages kswapd scanned 2160926.73 ( +0.00%) 4002186.33 ( +85.21%)
Pages kswapd reclaimed 533191.07 ( +0.00%) 718577.80 ( +34.77%)
Pages direct scanned 400450.33 ( +0.00%) 355172.73 ( -11.31%)
Pages direct reclaimed 94441.73 ( +0.00%) 31162.80 ( -67.00%)
Pages total scanned 2561377.07 ( +0.00%) 4357359.07 ( +70.12%)
Pages total reclaimed 627632.80 ( +0.00%) 749740.60 ( +19.46%)
Swap out 47959.53 ( +0.00%) 110084.33 ( +129.53%)
Swap in 7276.00 ( +0.00%) 24457.00 ( +236.10%)
File refaults 138043.00 ( +0.00%) 188226.93 ( +36.35%)
THP allocation latencies and %sys time are down dramatically.
THP allocation failures are down from nearly 50% to 8.5%. And to recall
previous data points, the success rates are steady and reliable without
the cumulative deterioration of fragmentation events.
Compaction work is down overall. Direct compaction work especially is
drastically reduced. As an aside, its success rate of 4% indicates there
is room for improvement. For now it's good to rely on it less.
Reclaim work is up overall, however direct reclaim work is down. Part of
the increase can be attributed to a higher use of THPs, which due to
internal fragmentation increase the memory footprint. This is not
necessarily an unexpected side-effect for users of THP.
However, taken both points together, there may well be some opportunities
for fine tuning in the reclaim/compaction coordination.
[hannes@cmpxchg.org: fix squawks from rebasing]
Link: https://lkml.kernel.org/r/20250314210558.GD1316033@cmpxchg.org
Link: https://lkml.kernel.org/r/20250313210647.1314586-6-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-03-13 17:05:36 -04:00
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
2020-04-06 20:04:56 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-06-24 13:54:17 +01:00
|
|
|
|
2020-04-06 20:04:49 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: page_alloc: defrag_mode kswapd/kcompactd watermarks
The previous patch added pageblock_order reclaim to kswapd/kcompactd,
which helps, but produces only one block at a time. Allocation stalls and
THP failure rates are still higher than they could be.
To adequately reflect ALLOC_NOFRAGMENT demand for pageblocks, change the
watermarking for kswapd & kcompactd: instead of targeting the high
watermark in order-0 pages and checking for one suitable block, simply
require that the high watermark is entirely met in pageblocks.
To this end, track the number of free pages within contiguous pageblocks,
then change pgdat_balanced() and compact_finished() to check watermarks
against this new value.
This further reduces THP latencies and allocation stalls, and improves THP
success rates against the previous patch:
DEFRAGMODE-ASYNC DEFRAGMODE-ASYNC-WMARKS
Hugealloc Time mean 34300.36 ( +0.00%) 28904.00 ( -15.73%)
Hugealloc Time stddev 36390.42 ( +0.00%) 33464.37 ( -8.04%)
Kbuild Real time 196.13 ( +0.00%) 196.59 ( +0.23%)
Kbuild User time 1234.74 ( +0.00%) 1231.67 ( -0.25%)
Kbuild System time 62.62 ( +0.00%) 59.10 ( -5.54%)
THP fault alloc 57054.53 ( +0.00%) 63223.67 ( +10.81%)
THP fault fallback 11581.40 ( +0.00%) 5412.47 ( -53.26%)
Direct compact fail 107.80 ( +0.00%) 59.07 ( -44.79%)
Direct compact success 4.53 ( +0.00%) 2.80 ( -31.33%)
Direct compact success rate % 3.20 ( +0.00%) 3.99 ( +18.66%)
Compact daemon scanned migrate 5461033.93 ( +0.00%) 2267500.33 ( -58.48%)
Compact daemon scanned free 5824897.93 ( +0.00%) 2339773.00 ( -59.83%)
Compact direct scanned migrate 58336.93 ( +0.00%) 47659.93 ( -18.30%)
Compact direct scanned free 32791.87 ( +0.00%) 40729.67 ( +24.21%)
Compact total migrate scanned 5519370.87 ( +0.00%) 2315160.27 ( -58.05%)
Compact total free scanned 5857689.80 ( +0.00%) 2380502.67 ( -59.36%)
Alloc stall 2424.60 ( +0.00%) 638.87 ( -73.62%)
Pages kswapd scanned 2657018.33 ( +0.00%) 4002186.33 ( +50.63%)
Pages kswapd reclaimed 559583.07 ( +0.00%) 718577.80 ( +28.41%)
Pages direct scanned 722094.07 ( +0.00%) 355172.73 ( -50.81%)
Pages direct reclaimed 107257.80 ( +0.00%) 31162.80 ( -70.95%)
Pages total scanned 3379112.40 ( +0.00%) 4357359.07 ( +28.95%)
Pages total reclaimed 666840.87 ( +0.00%) 749740.60 ( +12.43%)
Swap out 77238.20 ( +0.00%) 110084.33 ( +42.53%)
Swap in 11712.80 ( +0.00%) 24457.00 ( +108.80%)
File refaults 143438.80 ( +0.00%) 188226.93 ( +31.22%)
Also of note is that compaction work overall is reduced. The reason for
this is that when free pageblocks are more readily available, allocations
are also much more likely to get physically placed in LRU order, instead
of being forced to scavenge free space here and there. This means that
reclaim by itself has better chances of freeing up whole blocks, and the
system relies less on compaction.
Comparing all changes to the vanilla kernel:
VANILLA DEFRAGMODE-ASYNC-WMARKS
Hugealloc Time mean 52739.45 ( +0.00%) 28904.00 ( -45.19%)
Hugealloc Time stddev 56541.26 ( +0.00%) 33464.37 ( -40.81%)
Kbuild Real time 197.47 ( +0.00%) 196.59 ( -0.44%)
Kbuild User time 1240.49 ( +0.00%) 1231.67 ( -0.71%)
Kbuild System time 70.08 ( +0.00%) 59.10 ( -15.45%)
THP fault alloc 46727.07 ( +0.00%) 63223.67 ( +35.30%)
THP fault fallback 21910.60 ( +0.00%) 5412.47 ( -75.29%)
Direct compact fail 195.80 ( +0.00%) 59.07 ( -69.48%)
Direct compact success 7.93 ( +0.00%) 2.80 ( -57.46%)
Direct compact success rate % 3.51 ( +0.00%) 3.99 ( +10.49%)
Compact daemon scanned migrate 3369601.27 ( +0.00%) 2267500.33 ( -32.71%)
Compact daemon scanned free 5075474.47 ( +0.00%) 2339773.00 ( -53.90%)
Compact direct scanned migrate 161787.27 ( +0.00%) 47659.93 ( -70.54%)
Compact direct scanned free 163467.53 ( +0.00%) 40729.67 ( -75.08%)
Compact total migrate scanned 3531388.53 ( +0.00%) 2315160.27 ( -34.44%)
Compact total free scanned 5238942.00 ( +0.00%) 2380502.67 ( -54.56%)
Alloc stall 2371.07 ( +0.00%) 638.87 ( -73.02%)
Pages kswapd scanned 2160926.73 ( +0.00%) 4002186.33 ( +85.21%)
Pages kswapd reclaimed 533191.07 ( +0.00%) 718577.80 ( +34.77%)
Pages direct scanned 400450.33 ( +0.00%) 355172.73 ( -11.31%)
Pages direct reclaimed 94441.73 ( +0.00%) 31162.80 ( -67.00%)
Pages total scanned 2561377.07 ( +0.00%) 4357359.07 ( +70.12%)
Pages total reclaimed 627632.80 ( +0.00%) 749740.60 ( +19.46%)
Swap out 47959.53 ( +0.00%) 110084.33 ( +129.53%)
Swap in 7276.00 ( +0.00%) 24457.00 ( +236.10%)
File refaults 138043.00 ( +0.00%) 188226.93 ( +36.35%)
THP allocation latencies and %sys time are down dramatically.
THP allocation failures are down from nearly 50% to 8.5%. And to recall
previous data points, the success rates are steady and reliable without
the cumulative deterioration of fragmentation events.
Compaction work is down overall. Direct compaction work especially is
drastically reduced. As an aside, its success rate of 4% indicates there
is room for improvement. For now it's good to rely on it less.
Reclaim work is up overall, however direct reclaim work is down. Part of
the increase can be attributed to a higher use of THPs, which due to
internal fragmentation increase the memory footprint. This is not
necessarily an unexpected side-effect for users of THP.
However, taken both points together, there may well be some opportunities
for fine tuning in the reclaim/compaction coordination.
[hannes@cmpxchg.org: fix squawks from rebasing]
Link: https://lkml.kernel.org/r/20250314210558.GD1316033@cmpxchg.org
Link: https://lkml.kernel.org/r/20250313210647.1314586-6-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-03-13 17:05:36 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-04-06 20:04:49 -07:00
|
|
|
|
|
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-03-19 13:42:14 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-06-21 16:14:28 +08:00
|
|
|
|
2023-03-19 13:42:14 +02:00
|
|
|
|
|
|
|
|
|
2020-04-06 20:04:45 -07:00
|
|
|
|
2024-06-19 01:06:12 +00:00
|
|
|
|
|
|
|
|
|
2020-04-06 20:04:45 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-06-19 01:06:12 +00:00
|
|
|
|
2020-04-06 20:04:45 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-04-28 23:16:01 -07:00
|
|
|
|
|
|
|
|
|
2020-04-06 20:04:45 -07:00
|
|
|
|
2023-12-28 17:47:04 +03:00
|
|
|
|
2020-04-06 20:04:45 -07:00
|
|
|
|
|
|
|
|
|
2022-04-28 23:16:01 -07:00
|
|
|
|
|
|
|
|
|
2020-04-06 20:04:45 -07:00
|
|
|
|
2022-04-28 23:16:01 -07:00
|
|
|
|
|
|
|
|
|
2020-04-06 20:04:45 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-06-07 17:08:18 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2012-01-11 15:16:11 +01:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2012-01-11 15:16:11 +01:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2012-12-06 10:39:54 +01:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2006-01-08 01:00:42 -08:00
|
|
|
|
2014-06-04 16:10:17 -07:00
|
|
|
|
2009-06-16 15:32:07 -07:00
|
|
|
|
mm/page_alloc: convert "report" flag of __free_one_page() to a proper flag
Patch series "mm: place pages to the freelist tail when onlining and undoing isolation", v2.
When adding separate memory blocks via add_memory*() and onlining them
immediately, the metadata (especially the memmap) of the next block will
be placed onto one of the just added+onlined block. This creates a chain
of unmovable allocations: If the last memory block cannot get
offlined+removed() so will all dependent ones. We directly have unmovable
allocations all over the place.
This can be observed quite easily using virtio-mem, however, it can also
be observed when using DIMMs. The freshly onlined pages will usually be
placed to the head of the freelists, meaning they will be allocated next,
turning the just-added memory usually immediately un-removable. The fresh
pages are cold, prefering to allocate others (that might be hot) also
feels to be the natural thing to do.
It also applies to the hyper-v balloon xen-balloon, and ppc64 dlpar: when
adding separate, successive memory blocks, each memory block will have
unmovable allocations on them - for example gigantic pages will fail to
allocate.
While the ZONE_NORMAL doesn't provide any guarantees that memory can get
offlined+removed again (any kind of fragmentation with unmovable
allocations is possible), there are many scenarios (hotplugging a lot of
memory, running workload, hotunplug some memory/as much as possible) where
we can offline+remove quite a lot with this patchset.
a) To visualize the problem, a very simple example:
Start a VM with 4GB and 8GB of virtio-mem memory:
[root@localhost ~]# lsmem
RANGE SIZE STATE REMOVABLE BLOCK
0x0000000000000000-0x00000000bfffffff 3G online yes 0-23
0x0000000100000000-0x000000033fffffff 9G online yes 32-103
Memory block size: 128M
Total online memory: 12G
Total offline memory: 0B
Then try to unplug as much as possible using virtio-mem. Observe which
memory blocks are still around. Without this patch set:
[root@localhost ~]# lsmem
RANGE SIZE STATE REMOVABLE BLOCK
0x0000000000000000-0x00000000bfffffff 3G online yes 0-23
0x0000000100000000-0x000000013fffffff 1G online yes 32-39
0x0000000148000000-0x000000014fffffff 128M online yes 41
0x0000000158000000-0x000000015fffffff 128M online yes 43
0x0000000168000000-0x000000016fffffff 128M online yes 45
0x0000000178000000-0x000000017fffffff 128M online yes 47
0x0000000188000000-0x0000000197ffffff 256M online yes 49-50
0x00000001a0000000-0x00000001a7ffffff 128M online yes 52
0x00000001b0000000-0x00000001b7ffffff 128M online yes 54
0x00000001c0000000-0x00000001c7ffffff 128M online yes 56
0x00000001d0000000-0x00000001d7ffffff 128M online yes 58
0x00000001e0000000-0x00000001e7ffffff 128M online yes 60
0x00000001f0000000-0x00000001f7ffffff 128M online yes 62
0x0000000200000000-0x0000000207ffffff 128M online yes 64
0x0000000210000000-0x0000000217ffffff 128M online yes 66
0x0000000220000000-0x0000000227ffffff 128M online yes 68
0x0000000230000000-0x0000000237ffffff 128M online yes 70
0x0000000240000000-0x0000000247ffffff 128M online yes 72
0x0000000250000000-0x0000000257ffffff 128M online yes 74
0x0000000260000000-0x0000000267ffffff 128M online yes 76
0x0000000270000000-0x0000000277ffffff 128M online yes 78
0x0000000280000000-0x0000000287ffffff 128M online yes 80
0x0000000290000000-0x0000000297ffffff 128M online yes 82
0x00000002a0000000-0x00000002a7ffffff 128M online yes 84
0x00000002b0000000-0x00000002b7ffffff 128M online yes 86
0x00000002c0000000-0x00000002c7ffffff 128M online yes 88
0x00000002d0000000-0x00000002d7ffffff 128M online yes 90
0x00000002e0000000-0x00000002e7ffffff 128M online yes 92
0x00000002f0000000-0x00000002f7ffffff 128M online yes 94
0x0000000300000000-0x0000000307ffffff 128M online yes 96
0x0000000310000000-0x0000000317ffffff 128M online yes 98
0x0000000320000000-0x0000000327ffffff 128M online yes 100
0x0000000330000000-0x000000033fffffff 256M online yes 102-103
Memory block size: 128M
Total online memory: 8.1G
Total offline memory: 0B
With this patch set:
[root@localhost ~]# lsmem
RANGE SIZE STATE REMOVABLE BLOCK
0x0000000000000000-0x00000000bfffffff 3G online yes 0-23
0x0000000100000000-0x000000013fffffff 1G online yes 32-39
Memory block size: 128M
Total online memory: 4G
Total offline memory: 0B
All memory can get unplugged, all memory block can get removed. Of
course, no workload ran and the system was basically idle, but it
highlights the issue - the fairly deterministic chain of unmovable
allocations. When a huge page for the 2MB memmap is needed, a
just-onlined 4MB page will be split. The remaining 2MB page will be used
for the memmap of the next memory block. So one memory block will hold
the memmap of the two following memory blocks. Finally the pages of the
last-onlined memory block will get used for the next bigger allocations -
if any allocation is unmovable, all dependent memory blocks cannot get
unplugged and removed until that allocation is gone.
Note that with bigger memory blocks (e.g., 256MB), *all* memory
blocks are dependent and none can get unplugged again!
b) Experiment with memory intensive workload
I performed an experiment with an older version of this patch set (before
we used undo_isolate_page_range() in online_pages(): Hotplug 56GB to a VM
with an initial 4GB, onlining all memory to ZONE_NORMAL right from the
kernel when adding it. I then run various memory intensive workloads that
consume most system memory for a total of 45 minutes. Once finished, I
try to unplug as much memory as possible.
With this change, I am able to remove via virtio-mem (adding individual
128MB memory blocks) 413 out of 448 added memory blocks. Via individual
(256MB) DIMMs 380 out of 448 added memory blocks. (I don't have any
numbers without this patchset, but looking at the above example, it's at
most half of the 448 memory blocks for virtio-mem, and most probably none
for DIMMs).
Again, there are workloads that might behave very differently due to the
nature of ZONE_NORMAL.
This change also affects (besides memory onlining):
- Other users of undo_isolate_page_range(): Pages are always placed to the
tail.
-- When memory offlining fails
-- When memory isolation fails after having isolated some pageblocks
-- When alloc_contig_range() either succeeds or fails
- Other users of __putback_isolated_page(): Pages are always placed to the
tail.
-- Free page reporting
- Other users of __free_pages_core()
-- AFAIKs, any memory that is getting exposed to the buddy during boot.
IIUC we will now usually allocate memory from lower addresses within
a zone first (especially during boot).
- Other users of generic_online_page()
-- Hyper-V balloon
This patch (of 5):
Let's prepare for additional flags and avoid long parameter lists of
bools. Follow-up patches will also make use of the flags in
__free_pages_ok().
Signed-off-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Alexander Duyck <alexander.h.duyck@linux.intel.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Wei Yang <richard.weiyang@linux.alibaba.com>
Reviewed-by: Pankaj Gupta <pankaj.gupta.linux@gmail.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Haiyang Zhang <haiyangz@microsoft.com>
Cc: "K. Y. Srinivasan" <kys@microsoft.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Scott Cheloha <cheloha@linux.ibm.com>
Cc: Stephen Hemminger <sthemmin@microsoft.com>
Cc: Wei Liu <wei.liu@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Link: https://lkml.kernel.org/r/20201005121534.15649-1-david@redhat.com
Link: https://lkml.kernel.org/r/20201005121534.15649-2-david@redhat.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-10-15 20:09:20 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2019-03-05 15:45:41 -08:00
|
|
|
|
2022-09-16 15:22:54 +08:00
|
|
|
|
2020-04-06 20:04:45 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/page_alloc: prevent merging between isolated and other pageblocks
Hanjun Guo has reported that a CMA stress test causes broken accounting of
CMA and free pages:
> Before the test, I got:
> -bash-4.3# cat /proc/meminfo | grep Cma
> CmaTotal: 204800 kB
> CmaFree: 195044 kB
>
>
> After running the test:
> -bash-4.3# cat /proc/meminfo | grep Cma
> CmaTotal: 204800 kB
> CmaFree: 6602584 kB
>
> So the freed CMA memory is more than total..
>
> Also the the MemFree is more than mem total:
>
> -bash-4.3# cat /proc/meminfo
> MemTotal: 16342016 kB
> MemFree: 22367268 kB
> MemAvailable: 22370528 kB
Laura Abbott has confirmed the issue and suspected the freepage accounting
rewrite around 3.18/4.0 by Joonsoo Kim. Joonsoo had a theory that this is
caused by unexpected merging between MIGRATE_ISOLATE and MIGRATE_CMA
pageblocks:
> CMA isolates MAX_ORDER aligned blocks, but, during the process,
> partialy isolated block exists. If MAX_ORDER is 11 and
> pageblock_order is 9, two pageblocks make up MAX_ORDER
> aligned block and I can think following scenario because pageblock
> (un)isolation would be done one by one.
>
> (each character means one pageblock. 'C', 'I' means MIGRATE_CMA,
> MIGRATE_ISOLATE, respectively.
>
> CC -> IC -> II (Isolation)
> II -> CI -> CC (Un-isolation)
>
> If some pages are freed at this intermediate state such as IC or CI,
> that page could be merged to the other page that is resident on
> different type of pageblock and it will cause wrong freepage count.
This was supposed to be prevented by CMA operating on MAX_ORDER blocks,
but since it doesn't hold the zone->lock between pageblocks, a race
window does exist.
It's also likely that unexpected merging can occur between
MIGRATE_ISOLATE and non-CMA pageblocks. This should be prevented in
__free_one_page() since commit 3c605096d315 ("mm/page_alloc: restrict
max order of merging on isolated pageblock"). However, we only check
the migratetype of the pageblock where buddy merging has been initiated,
not the migratetype of the buddy pageblock (or group of pageblocks)
which can be MIGRATE_ISOLATE.
Joonsoo has suggested checking for buddy migratetype as part of
page_is_buddy(), but that would add extra checks in allocator hotpath
and bloat-o-meter has shown significant code bloat (the function is
inline).
This patch reduces the bloat at some expense of more complicated code.
The buddy-merging while-loop in __free_one_page() is initially bounded
to pageblock_border and without any migratetype checks. The checks are
placed outside, bumping the max_order if merging is allowed, and
returning to the while-loop with a statement which can't be possibly
considered harmful.
This fixes the accounting bug and also removes the arguably weird state
in the original commit 3c605096d315 where buddies could be left
unmerged.
Fixes: 3c605096d315 ("mm/page_alloc: restrict max order of merging on isolated pageblock")
Link: https://lkml.org/lkml/2016/3/2/280
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Reported-by: Hanjun Guo <guohanjun@huawei.com>
Tested-by: Hanjun Guo <guohanjun@huawei.com>
Acked-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Debugged-by: Laura Abbott <labbott@redhat.com>
Debugged-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Cc: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Cc: Michal Nazarewicz <mina86@mina86.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Cc: <stable@vger.kernel.org> [3.18+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-03-25 14:21:50 -07:00
|
|
|
|
2013-02-22 16:35:25 -08:00
|
|
|
|
2025-08-05 18:22:51 +01:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2009-06-16 15:32:07 -07:00
|
|
|
|
2017-02-22 15:41:48 -08:00
|
|
|
|
2014-01-23 15:52:54 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
|
|
|
|
|
2023-12-28 17:47:04 +03:00
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
|
|
|
|
|
2019-03-05 15:45:41 -08:00
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
2019-03-05 15:45:41 -08:00
|
|
|
|
|
|
|
|
|
2017-02-22 15:41:51 -08:00
|
|
|
|
2022-04-28 23:16:01 -07:00
|
|
|
|
|
|
|
|
|
mm/page_alloc: prevent merging between isolated and other pageblocks
Hanjun Guo has reported that a CMA stress test causes broken accounting of
CMA and free pages:
> Before the test, I got:
> -bash-4.3# cat /proc/meminfo | grep Cma
> CmaTotal: 204800 kB
> CmaFree: 195044 kB
>
>
> After running the test:
> -bash-4.3# cat /proc/meminfo | grep Cma
> CmaTotal: 204800 kB
> CmaFree: 6602584 kB
>
> So the freed CMA memory is more than total..
>
> Also the the MemFree is more than mem total:
>
> -bash-4.3# cat /proc/meminfo
> MemTotal: 16342016 kB
> MemFree: 22367268 kB
> MemAvailable: 22370528 kB
Laura Abbott has confirmed the issue and suspected the freepage accounting
rewrite around 3.18/4.0 by Joonsoo Kim. Joonsoo had a theory that this is
caused by unexpected merging between MIGRATE_ISOLATE and MIGRATE_CMA
pageblocks:
> CMA isolates MAX_ORDER aligned blocks, but, during the process,
> partialy isolated block exists. If MAX_ORDER is 11 and
> pageblock_order is 9, two pageblocks make up MAX_ORDER
> aligned block and I can think following scenario because pageblock
> (un)isolation would be done one by one.
>
> (each character means one pageblock. 'C', 'I' means MIGRATE_CMA,
> MIGRATE_ISOLATE, respectively.
>
> CC -> IC -> II (Isolation)
> II -> CI -> CC (Un-isolation)
>
> If some pages are freed at this intermediate state such as IC or CI,
> that page could be merged to the other page that is resident on
> different type of pageblock and it will cause wrong freepage count.
This was supposed to be prevented by CMA operating on MAX_ORDER blocks,
but since it doesn't hold the zone->lock between pageblocks, a race
window does exist.
It's also likely that unexpected merging can occur between
MIGRATE_ISOLATE and non-CMA pageblocks. This should be prevented in
__free_one_page() since commit 3c605096d315 ("mm/page_alloc: restrict
max order of merging on isolated pageblock"). However, we only check
the migratetype of the pageblock where buddy merging has been initiated,
not the migratetype of the buddy pageblock (or group of pageblocks)
which can be MIGRATE_ISOLATE.
Joonsoo has suggested checking for buddy migratetype as part of
page_is_buddy(), but that would add extra checks in allocator hotpath
and bloat-o-meter has shown significant code bloat (the function is
inline).
This patch reduces the bloat at some expense of more complicated code.
The buddy-merging while-loop in __free_one_page() is initially bounded
to pageblock_border and without any migratetype checks. The checks are
placed outside, bumping the max_order if merging is allowed, and
returning to the while-loop with a statement which can't be possibly
considered harmful.
This fixes the accounting bug and also removes the arguably weird state
in the original commit 3c605096d315 where buddies could be left
unmerged.
Fixes: 3c605096d315 ("mm/page_alloc: restrict max order of merging on isolated pageblock")
Link: https://lkml.org/lkml/2016/3/2/280
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Reported-by: Hanjun Guo <guohanjun@huawei.com>
Tested-by: Hanjun Guo <guohanjun@huawei.com>
Acked-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Debugged-by: Laura Abbott <labbott@redhat.com>
Debugged-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Cc: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Cc: Michal Nazarewicz <mina86@mina86.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Cc: <stable@vger.kernel.org> [3.18+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-03-25 14:21:50 -07:00
|
|
|
|
2022-04-28 23:16:01 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
2022-04-28 23:16:01 -07:00
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-04-28 23:16:01 -07:00
|
|
|
|
|
|
|
|
|
2012-01-10 15:07:28 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-05-14 15:41:32 -07:00
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
2019-05-14 15:41:32 -07:00
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-02-22 15:41:48 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
mm/page_alloc: prevent merging between isolated and other pageblocks
Hanjun Guo has reported that a CMA stress test causes broken accounting of
CMA and free pages:
> Before the test, I got:
> -bash-4.3# cat /proc/meminfo | grep Cma
> CmaTotal: 204800 kB
> CmaFree: 195044 kB
>
>
> After running the test:
> -bash-4.3# cat /proc/meminfo | grep Cma
> CmaTotal: 204800 kB
> CmaFree: 6602584 kB
>
> So the freed CMA memory is more than total..
>
> Also the the MemFree is more than mem total:
>
> -bash-4.3# cat /proc/meminfo
> MemTotal: 16342016 kB
> MemFree: 22367268 kB
> MemAvailable: 22370528 kB
Laura Abbott has confirmed the issue and suspected the freepage accounting
rewrite around 3.18/4.0 by Joonsoo Kim. Joonsoo had a theory that this is
caused by unexpected merging between MIGRATE_ISOLATE and MIGRATE_CMA
pageblocks:
> CMA isolates MAX_ORDER aligned blocks, but, during the process,
> partialy isolated block exists. If MAX_ORDER is 11 and
> pageblock_order is 9, two pageblocks make up MAX_ORDER
> aligned block and I can think following scenario because pageblock
> (un)isolation would be done one by one.
>
> (each character means one pageblock. 'C', 'I' means MIGRATE_CMA,
> MIGRATE_ISOLATE, respectively.
>
> CC -> IC -> II (Isolation)
> II -> CI -> CC (Un-isolation)
>
> If some pages are freed at this intermediate state such as IC or CI,
> that page could be merged to the other page that is resident on
> different type of pageblock and it will cause wrong freepage count.
This was supposed to be prevented by CMA operating on MAX_ORDER blocks,
but since it doesn't hold the zone->lock between pageblocks, a race
window does exist.
It's also likely that unexpected merging can occur between
MIGRATE_ISOLATE and non-CMA pageblocks. This should be prevented in
__free_one_page() since commit 3c605096d315 ("mm/page_alloc: restrict
max order of merging on isolated pageblock"). However, we only check
the migratetype of the pageblock where buddy merging has been initiated,
not the migratetype of the buddy pageblock (or group of pageblocks)
which can be MIGRATE_ISOLATE.
Joonsoo has suggested checking for buddy migratetype as part of
page_is_buddy(), but that would add extra checks in allocator hotpath
and bloat-o-meter has shown significant code bloat (the function is
inline).
This patch reduces the bloat at some expense of more complicated code.
The buddy-merging while-loop in __free_one_page() is initially bounded
to pageblock_border and without any migratetype checks. The checks are
placed outside, bumping the max_order if merging is allowed, and
returning to the while-loop with a statement which can't be possibly
considered harmful.
This fixes the accounting bug and also removes the arguably weird state
in the original commit 3c605096d315 where buddies could be left
unmerged.
Fixes: 3c605096d315 ("mm/page_alloc: restrict max order of merging on isolated pageblock")
Link: https://lkml.org/lkml/2016/3/2/280
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Reported-by: Hanjun Guo <guohanjun@huawei.com>
Tested-by: Hanjun Guo <guohanjun@huawei.com>
Acked-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Debugged-by: Laura Abbott <labbott@redhat.com>
Debugged-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Cc: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Cc: Michal Nazarewicz <mina86@mina86.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Cc: <stable@vger.kernel.org> [3.18+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-03-25 14:21:50 -07:00
|
|
|
|
|
|
|
|
|
2020-10-15 20:10:15 -07:00
|
|
|
|
page allocator: reduce fragmentation in buddy allocator by adding buddies that are merging to the tail of the free lists
In order to reduce fragmentation, this patch classifies freed pages in two
groups according to their probability of being part of a high order merge.
Pages belonging to a compound whose next-highest buddy is free are more
likely to be part of a high order merge in the near future, so they will
be added at the tail of the freelist. The remaining pages are put at the
front of the freelist.
In this way, the pages that are more likely to cause a big merge are kept
free longer. Consequently there is a tendency to aggregate the
long-living allocations on a subset of the compounds, reducing the
fragmentation.
This heuristic was tested on three machines, x86, x86-64 and ppc64 with
3GB of RAM in each machine. The tests were kernbench, netperf, sysbench
and STREAM for performance and a high-order stress test for huge page
allocations.
KernBench X86
Elapsed mean 374.77 ( 0.00%) 375.10 (-0.09%)
User mean 649.53 ( 0.00%) 650.44 (-0.14%)
System mean 54.75 ( 0.00%) 54.18 ( 1.05%)
CPU mean 187.75 ( 0.00%) 187.25 ( 0.27%)
KernBench X86-64
Elapsed mean 94.45 ( 0.00%) 94.01 ( 0.47%)
User mean 323.27 ( 0.00%) 322.66 ( 0.19%)
System mean 36.71 ( 0.00%) 36.50 ( 0.57%)
CPU mean 380.75 ( 0.00%) 381.75 (-0.26%)
KernBench PPC64
Elapsed mean 173.45 ( 0.00%) 173.74 (-0.17%)
User mean 587.99 ( 0.00%) 587.95 ( 0.01%)
System mean 60.60 ( 0.00%) 60.57 ( 0.05%)
CPU mean 373.50 ( 0.00%) 372.75 ( 0.20%)
Nothing notable for kernbench.
NetPerf UDP X86
64 42.68 ( 0.00%) 42.77 ( 0.21%)
128 85.62 ( 0.00%) 85.32 (-0.35%)
256 170.01 ( 0.00%) 168.76 (-0.74%)
1024 655.68 ( 0.00%) 652.33 (-0.51%)
2048 1262.39 ( 0.00%) 1248.61 (-1.10%)
3312 1958.41 ( 0.00%) 1944.61 (-0.71%)
4096 2345.63 ( 0.00%) 2318.83 (-1.16%)
8192 4132.90 ( 0.00%) 4089.50 (-1.06%)
16384 6770.88 ( 0.00%) 6642.05 (-1.94%)*
NetPerf UDP X86-64
64 148.82 ( 0.00%) 154.92 ( 3.94%)
128 298.96 ( 0.00%) 312.95 ( 4.47%)
256 583.67 ( 0.00%) 626.39 ( 6.82%)
1024 2293.18 ( 0.00%) 2371.10 ( 3.29%)
2048 4274.16 ( 0.00%) 4396.83 ( 2.79%)
3312 6356.94 ( 0.00%) 6571.35 ( 3.26%)
4096 7422.68 ( 0.00%) 7635.42 ( 2.79%)*
8192 12114.81 ( 0.00%)* 12346.88 ( 1.88%)
16384 17022.28 ( 0.00%)* 17033.19 ( 0.06%)*
1.64% 2.73%
NetPerf UDP PPC64
64 49.98 ( 0.00%) 50.25 ( 0.54%)
128 98.66 ( 0.00%) 100.95 ( 2.27%)
256 197.33 ( 0.00%) 191.03 (-3.30%)
1024 761.98 ( 0.00%) 785.07 ( 2.94%)
2048 1493.50 ( 0.00%) 1510.85 ( 1.15%)
3312 2303.95 ( 0.00%) 2271.72 (-1.42%)
4096 2774.56 ( 0.00%) 2773.06 (-0.05%)
8192 4918.31 ( 0.00%) 4793.59 (-2.60%)
16384 7497.98 ( 0.00%) 7749.52 ( 3.25%)
The tests are run to have confidence limits within 1%. Results marked
with a * were not confident although in this case, it's only outside by
small amounts. Even with some results that were not confident, the
netperf UDP results were generally positive.
NetPerf TCP X86
64 652.25 ( 0.00%)* 648.12 (-0.64%)*
23.80% 22.82%
128 1229.98 ( 0.00%)* 1220.56 (-0.77%)*
21.03% 18.90%
256 2105.88 ( 0.00%) 1872.03 (-12.49%)*
1.00% 16.46%
1024 3476.46 ( 0.00%)* 3548.28 ( 2.02%)*
13.37% 11.39%
2048 4023.44 ( 0.00%)* 4231.45 ( 4.92%)*
9.76% 12.48%
3312 4348.88 ( 0.00%)* 4396.96 ( 1.09%)*
6.49% 8.75%
4096 4726.56 ( 0.00%)* 4877.71 ( 3.10%)*
9.85% 8.50%
8192 4732.28 ( 0.00%)* 5777.77 (18.10%)*
9.13% 13.04%
16384 5543.05 ( 0.00%)* 5906.24 ( 6.15%)*
7.73% 8.68%
NETPERF TCP X86-64
netperf-tcp-vanilla-netperf netperf-tcp
tcp-vanilla pgalloc-delay
64 1895.87 ( 0.00%)* 1775.07 (-6.81%)*
5.79% 4.78%
128 3571.03 ( 0.00%)* 3342.20 (-6.85%)*
3.68% 6.06%
256 5097.21 ( 0.00%)* 4859.43 (-4.89%)*
3.02% 2.10%
1024 8919.10 ( 0.00%)* 8892.49 (-0.30%)*
5.89% 6.55%
2048 10255.46 ( 0.00%)* 10449.39 ( 1.86%)*
7.08% 7.44%
3312 10839.90 ( 0.00%)* 10740.15 (-0.93%)*
6.87% 7.33%
4096 10814.84 ( 0.00%)* 10766.97 (-0.44%)*
6.86% 8.18%
8192 11606.89 ( 0.00%)* 11189.28 (-3.73%)*
7.49% 5.55%
16384 12554.88 ( 0.00%)* 12361.22 (-1.57%)*
7.36% 6.49%
NETPERF TCP PPC64
netperf-tcp-vanilla-netperf netperf-tcp
tcp-vanilla pgalloc-delay
64 594.17 ( 0.00%) 596.04 ( 0.31%)*
1.00% 2.29%
128 1064.87 ( 0.00%)* 1074.77 ( 0.92%)*
1.30% 1.40%
256 1852.46 ( 0.00%)* 1856.95 ( 0.24%)
1.25% 1.00%
1024 3839.46 ( 0.00%)* 3813.05 (-0.69%)
1.02% 1.00%
2048 4885.04 ( 0.00%)* 4881.97 (-0.06%)*
1.15% 1.04%
3312 5506.90 ( 0.00%) 5459.72 (-0.86%)
4096 6449.19 ( 0.00%) 6345.46 (-1.63%)
8192 7501.17 ( 0.00%) 7508.79 ( 0.10%)
16384 9618.65 ( 0.00%) 9490.10 (-1.35%)
There was a distinct lack of confidence in the X86* figures so I included
what the devation was where the results were not confident. Many of the
results, whether gains or losses were within the standard deviation so no
solid conclusion can be reached on performance impact. Looking at the
figures, only the X86-64 ones look suspicious with a few losses that were
outside the noise. However, the results were so unstable that without
knowing why they vary so much, a solid conclusion cannot be reached.
SYSBENCH X86
sysbench-vanilla pgalloc-delay
1 7722.85 ( 0.00%) 7756.79 ( 0.44%)
2 14901.11 ( 0.00%) 13683.44 (-8.90%)
3 15171.71 ( 0.00%) 14888.25 (-1.90%)
4 14966.98 ( 0.00%) 15029.67 ( 0.42%)
5 14370.47 ( 0.00%) 14865.00 ( 3.33%)
6 14870.33 ( 0.00%) 14845.57 (-0.17%)
7 14429.45 ( 0.00%) 14520.85 ( 0.63%)
8 14354.35 ( 0.00%) 14362.31 ( 0.06%)
SYSBENCH X86-64
1 17448.70 ( 0.00%) 17484.41 ( 0.20%)
2 34276.39 ( 0.00%) 34251.00 (-0.07%)
3 50805.25 ( 0.00%) 50854.80 ( 0.10%)
4 66667.10 ( 0.00%) 66174.69 (-0.74%)
5 66003.91 ( 0.00%) 65685.25 (-0.49%)
6 64981.90 ( 0.00%) 65125.60 ( 0.22%)
7 64933.16 ( 0.00%) 64379.23 (-0.86%)
8 63353.30 ( 0.00%) 63281.22 (-0.11%)
9 63511.84 ( 0.00%) 63570.37 ( 0.09%)
10 62708.27 ( 0.00%) 63166.25 ( 0.73%)
11 62092.81 ( 0.00%) 61787.75 (-0.49%)
12 61330.11 ( 0.00%) 61036.34 (-0.48%)
13 61438.37 ( 0.00%) 61994.47 ( 0.90%)
14 62304.48 ( 0.00%) 62064.90 (-0.39%)
15 63296.48 ( 0.00%) 62875.16 (-0.67%)
16 63951.76 ( 0.00%) 63769.09 (-0.29%)
SYSBENCH PPC64
-sysbench-pgalloc-delay-sysbench
sysbench-vanilla pgalloc-delay
1 7645.08 ( 0.00%) 7467.43 (-2.38%)
2 14856.67 ( 0.00%) 14558.73 (-2.05%)
3 21952.31 ( 0.00%) 21683.64 (-1.24%)
4 27946.09 ( 0.00%) 28623.29 ( 2.37%)
5 28045.11 ( 0.00%) 28143.69 ( 0.35%)
6 27477.10 ( 0.00%) 27337.45 (-0.51%)
7 26489.17 ( 0.00%) 26590.06 ( 0.38%)
8 26642.91 ( 0.00%) 25274.33 (-5.41%)
9 25137.27 ( 0.00%) 24810.06 (-1.32%)
10 24451.99 ( 0.00%) 24275.85 (-0.73%)
11 23262.20 ( 0.00%) 23674.88 ( 1.74%)
12 24234.81 ( 0.00%) 23640.89 (-2.51%)
13 24577.75 ( 0.00%) 24433.50 (-0.59%)
14 25640.19 ( 0.00%) 25116.52 (-2.08%)
15 26188.84 ( 0.00%) 26181.36 (-0.03%)
16 26782.37 ( 0.00%) 26255.99 (-2.00%)
Again, there is little to conclude here. While there are a few losses,
the results vary by +/- 8% in some cases. They are the results of most
concern as there are some large losses but it's also within the variance
typically seen between kernel releases.
The STREAM results varied so little and are so verbose that I didn't
include them here.
The final test stressed how many huge pages can be allocated. The
absolute number of huge pages allocated are the same with or without the
page. However, the "unusability free space index" which is a measure of
external fragmentation was slightly lower (lower is better) throughout the
lifetime of the system. I also measured the latency of how long it took
to successfully allocate a huge page. The latency was slightly lower and
on X86 and PPC64, more huge pages were allocated almost immediately from
the free lists. The improvement is slight but there.
[mel@csn.ul.ie: Tested, reworked for less branches]
[czoccolo@gmail.com: fix oops by checking pfn_valid_within()]
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Acked-by: Rik van Riel <riel@redhat.com>
Reviewed-by: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: Corrado Zoccolo <czoccolo@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-24 14:31:54 -07:00
|
|
|
|
mm/page_alloc: place pages to tail in __putback_isolated_page()
__putback_isolated_page() already documents that pages will be placed to
the tail of the freelist - this is, however, not the case for "order >=
MAX_ORDER - 2" (see buddy_merge_likely()) - which should be the case for
all existing users.
This change affects two users:
- free page reporting
- page isolation, when undoing the isolation (including memory onlining).
This behavior is desirable for pages that haven't really been touched
lately, so exactly the two users that don't actually read/write page
content, but rather move untouched pages.
The new behavior is especially desirable for memory onlining, where we
allow allocation of newly onlined pages via undo_isolate_page_range() in
online_pages(). Right now, we always place them to the head of the
freelist, resulting in undesireable behavior: Assume we add individual
memory chunks via add_memory() and online them right away to the NORMAL
zone. We create a dependency chain of unmovable allocations e.g., via the
memmap. The memmap of the next chunk will be placed onto previous chunks
- if the last block cannot get offlined+removed, all dependent ones cannot
get offlined+removed. While this can already be observed with individual
DIMMs, it's more of an issue for virtio-mem (and I suspect also ppc
DLPAR).
Document that this should only be used for optimizations, and no code
should rely on this behavior for correction (if the order of the freelists
ever changes).
We won't care about page shuffling: memory onlining already properly
shuffles after onlining. free page reporting doesn't care about
physically contiguous ranges, and there are already cases where page
isolation will simply move (physically close) free pages to (currently)
the head of the freelists via move_freepages_block() instead of shuffling.
If this becomes ever relevant, we should shuffle the whole zone when
undoing isolation of larger ranges, and after free_contig_range().
Signed-off-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Alexander Duyck <alexander.h.duyck@linux.intel.com>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Wei Yang <richard.weiyang@linux.alibaba.com>
Reviewed-by: Pankaj Gupta <pankaj.gupta.linux@gmail.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Scott Cheloha <cheloha@linux.ibm.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Haiyang Zhang <haiyangz@microsoft.com>
Cc: "K. Y. Srinivasan" <kys@microsoft.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Stephen Hemminger <sthemmin@microsoft.com>
Cc: Wei Liu <wei.liu@kernel.org>
Link: https://lkml.kernel.org/r/20201005121534.15649-3-david@redhat.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-10-15 20:09:26 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-04-06 20:04:45 -07:00
|
|
|
|
2019-05-14 15:41:35 -07:00
|
|
|
|
2020-04-06 20:04:45 -07:00
|
|
|
|
2019-05-14 15:41:35 -07:00
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
2020-04-06 20:04:56 -07:00
|
|
|
|
|
|
|
|
|
mm/page_alloc: convert "report" flag of __free_one_page() to a proper flag
Patch series "mm: place pages to the freelist tail when onlining and undoing isolation", v2.
When adding separate memory blocks via add_memory*() and onlining them
immediately, the metadata (especially the memmap) of the next block will
be placed onto one of the just added+onlined block. This creates a chain
of unmovable allocations: If the last memory block cannot get
offlined+removed() so will all dependent ones. We directly have unmovable
allocations all over the place.
This can be observed quite easily using virtio-mem, however, it can also
be observed when using DIMMs. The freshly onlined pages will usually be
placed to the head of the freelists, meaning they will be allocated next,
turning the just-added memory usually immediately un-removable. The fresh
pages are cold, prefering to allocate others (that might be hot) also
feels to be the natural thing to do.
It also applies to the hyper-v balloon xen-balloon, and ppc64 dlpar: when
adding separate, successive memory blocks, each memory block will have
unmovable allocations on them - for example gigantic pages will fail to
allocate.
While the ZONE_NORMAL doesn't provide any guarantees that memory can get
offlined+removed again (any kind of fragmentation with unmovable
allocations is possible), there are many scenarios (hotplugging a lot of
memory, running workload, hotunplug some memory/as much as possible) where
we can offline+remove quite a lot with this patchset.
a) To visualize the problem, a very simple example:
Start a VM with 4GB and 8GB of virtio-mem memory:
[root@localhost ~]# lsmem
RANGE SIZE STATE REMOVABLE BLOCK
0x0000000000000000-0x00000000bfffffff 3G online yes 0-23
0x0000000100000000-0x000000033fffffff 9G online yes 32-103
Memory block size: 128M
Total online memory: 12G
Total offline memory: 0B
Then try to unplug as much as possible using virtio-mem. Observe which
memory blocks are still around. Without this patch set:
[root@localhost ~]# lsmem
RANGE SIZE STATE REMOVABLE BLOCK
0x0000000000000000-0x00000000bfffffff 3G online yes 0-23
0x0000000100000000-0x000000013fffffff 1G online yes 32-39
0x0000000148000000-0x000000014fffffff 128M online yes 41
0x0000000158000000-0x000000015fffffff 128M online yes 43
0x0000000168000000-0x000000016fffffff 128M online yes 45
0x0000000178000000-0x000000017fffffff 128M online yes 47
0x0000000188000000-0x0000000197ffffff 256M online yes 49-50
0x00000001a0000000-0x00000001a7ffffff 128M online yes 52
0x00000001b0000000-0x00000001b7ffffff 128M online yes 54
0x00000001c0000000-0x00000001c7ffffff 128M online yes 56
0x00000001d0000000-0x00000001d7ffffff 128M online yes 58
0x00000001e0000000-0x00000001e7ffffff 128M online yes 60
0x00000001f0000000-0x00000001f7ffffff 128M online yes 62
0x0000000200000000-0x0000000207ffffff 128M online yes 64
0x0000000210000000-0x0000000217ffffff 128M online yes 66
0x0000000220000000-0x0000000227ffffff 128M online yes 68
0x0000000230000000-0x0000000237ffffff 128M online yes 70
0x0000000240000000-0x0000000247ffffff 128M online yes 72
0x0000000250000000-0x0000000257ffffff 128M online yes 74
0x0000000260000000-0x0000000267ffffff 128M online yes 76
0x0000000270000000-0x0000000277ffffff 128M online yes 78
0x0000000280000000-0x0000000287ffffff 128M online yes 80
0x0000000290000000-0x0000000297ffffff 128M online yes 82
0x00000002a0000000-0x00000002a7ffffff 128M online yes 84
0x00000002b0000000-0x00000002b7ffffff 128M online yes 86
0x00000002c0000000-0x00000002c7ffffff 128M online yes 88
0x00000002d0000000-0x00000002d7ffffff 128M online yes 90
0x00000002e0000000-0x00000002e7ffffff 128M online yes 92
0x00000002f0000000-0x00000002f7ffffff 128M online yes 94
0x0000000300000000-0x0000000307ffffff 128M online yes 96
0x0000000310000000-0x0000000317ffffff 128M online yes 98
0x0000000320000000-0x0000000327ffffff 128M online yes 100
0x0000000330000000-0x000000033fffffff 256M online yes 102-103
Memory block size: 128M
Total online memory: 8.1G
Total offline memory: 0B
With this patch set:
[root@localhost ~]# lsmem
RANGE SIZE STATE REMOVABLE BLOCK
0x0000000000000000-0x00000000bfffffff 3G online yes 0-23
0x0000000100000000-0x000000013fffffff 1G online yes 32-39
Memory block size: 128M
Total online memory: 4G
Total offline memory: 0B
All memory can get unplugged, all memory block can get removed. Of
course, no workload ran and the system was basically idle, but it
highlights the issue - the fairly deterministic chain of unmovable
allocations. When a huge page for the 2MB memmap is needed, a
just-onlined 4MB page will be split. The remaining 2MB page will be used
for the memmap of the next memory block. So one memory block will hold
the memmap of the two following memory blocks. Finally the pages of the
last-onlined memory block will get used for the next bigger allocations -
if any allocation is unmovable, all dependent memory blocks cannot get
unplugged and removed until that allocation is gone.
Note that with bigger memory blocks (e.g., 256MB), *all* memory
blocks are dependent and none can get unplugged again!
b) Experiment with memory intensive workload
I performed an experiment with an older version of this patch set (before
we used undo_isolate_page_range() in online_pages(): Hotplug 56GB to a VM
with an initial 4GB, onlining all memory to ZONE_NORMAL right from the
kernel when adding it. I then run various memory intensive workloads that
consume most system memory for a total of 45 minutes. Once finished, I
try to unplug as much memory as possible.
With this change, I am able to remove via virtio-mem (adding individual
128MB memory blocks) 413 out of 448 added memory blocks. Via individual
(256MB) DIMMs 380 out of 448 added memory blocks. (I don't have any
numbers without this patchset, but looking at the above example, it's at
most half of the 448 memory blocks for virtio-mem, and most probably none
for DIMMs).
Again, there are workloads that might behave very differently due to the
nature of ZONE_NORMAL.
This change also affects (besides memory onlining):
- Other users of undo_isolate_page_range(): Pages are always placed to the
tail.
-- When memory offlining fails
-- When memory isolation fails after having isolated some pageblocks
-- When alloc_contig_range() either succeeds or fails
- Other users of __putback_isolated_page(): Pages are always placed to the
tail.
-- Free page reporting
- Other users of __free_pages_core()
-- AFAIKs, any memory that is getting exposed to the buddy during boot.
IIUC we will now usually allocate memory from lower addresses within
a zone first (especially during boot).
- Other users of generic_online_page()
-- Hyper-V balloon
This patch (of 5):
Let's prepare for additional flags and avoid long parameter lists of
bools. Follow-up patches will also make use of the flags in
__free_pages_ok().
Signed-off-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Alexander Duyck <alexander.h.duyck@linux.intel.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Wei Yang <richard.weiyang@linux.alibaba.com>
Reviewed-by: Pankaj Gupta <pankaj.gupta.linux@gmail.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Haiyang Zhang <haiyangz@microsoft.com>
Cc: "K. Y. Srinivasan" <kys@microsoft.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Scott Cheloha <cheloha@linux.ibm.com>
Cc: Stephen Hemminger <sthemmin@microsoft.com>
Cc: Wei Liu <wei.liu@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Link: https://lkml.kernel.org/r/20201005121534.15649-1-david@redhat.com
Link: https://lkml.kernel.org/r/20201005121534.15649-2-david@redhat.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-10-15 20:09:20 -07:00
|
|
|
|
2020-04-06 20:04:56 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2016-05-19 17:14:15 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-04-29 22:56:45 -07:00
|
|
|
|
2023-11-24 11:16:52 +01:00
|
|
|
|
2025-04-09 12:41:36 +02:00
|
|
|
|
2025-08-05 18:22:51 +01:00
|
|
|
|
2016-05-19 17:14:15 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-06-03 15:58:39 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2020-06-03 15:58:29 -07:00
|
|
|
|
2014-01-23 15:52:49 -08:00
|
|
|
|
2016-01-15 16:53:42 -08:00
|
|
|
|
2014-01-23 15:52:49 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-03-17 14:19:26 -07:00
|
|
|
|
2016-05-19 17:10:49 -07:00
|
|
|
|
2025-08-05 18:22:51 +01:00
|
|
|
|
2020-06-03 15:58:39 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-01-23 15:52:49 -08:00
|
|
|
|
2014-12-10 15:44:58 -08:00
|
|
|
|
2021-04-29 22:56:45 -07:00
|
|
|
|
2014-12-10 15:44:58 -08:00
|
|
|
|
2023-11-24 11:16:52 +01:00
|
|
|
|
2025-04-09 12:41:36 +02:00
|
|
|
|
2023-11-24 11:16:52 +01:00
|
|
|
|
2020-06-03 15:58:39 -07:00
|
|
|
|
2016-05-19 17:14:18 -07:00
|
|
|
|
|
|
|
|
|
2022-09-13 15:20:48 -07:00
|
|
|
|
2016-05-19 17:14:18 -07:00
|
|
|
|
2016-05-19 17:14:21 -07:00
|
|
|
|
2022-09-13 15:20:48 -07:00
|
|
|
|
2016-05-19 17:14:18 -07:00
|
|
|
|
|
|
|
|
|
2025-03-28 09:20:31 +08:00
|
|
|
|
2022-09-13 15:20:48 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2023-05-16 14:38:21 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-04-05 16:28:40 +02:00
|
|
|
|
2016-05-19 17:14:32 -07:00
|
|
|
|
2023-01-11 14:28:47 +00:00
|
|
|
|
2016-05-19 17:14:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-05-16 14:38:21 +08:00
|
|
|
|
2016-05-19 17:14:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm,thp,rmap: simplify compound page mapcount handling
Compound page (folio) mapcount calculations have been different for anon
and file (or shmem) THPs, and involved the obscure PageDoubleMap flag.
And each huge mapping and unmapping of a file (or shmem) THP involved
atomically incrementing and decrementing the mapcount of every subpage of
that huge page, dirtying many struct page cachelines.
Add subpages_mapcount field to the struct folio and first tail page, so
that the total of subpage mapcounts is available in one place near the
head: then page_mapcount() and total_mapcount() and page_mapped(), and
their folio equivalents, are so quick that anon and file and hugetlb don't
need to be optimized differently. Delete the unloved PageDoubleMap.
page_add and page_remove rmap functions must now maintain the
subpages_mapcount as well as the subpage _mapcount, when dealing with pte
mappings of huge pages; and correct maintenance of NR_ANON_MAPPED and
NR_FILE_MAPPED statistics still needs reading through the subpages, using
nr_subpages_unmapped() - but only when first or last pmd mapping finds
subpages_mapcount raised (double-map case, not the common case).
But are those counts (used to decide when to split an anon THP, and in
vmscan's pagecache_reclaimable heuristic) correctly maintained? Not
quite: since page_remove_rmap() (and also split_huge_pmd()) is often
called without page lock, there can be races when a subpage pte mapcount
0<->1 while compound pmd mapcount 0<->1 is scanning - races which the
previous implementation had prevented. The statistics might become
inaccurate, and even drift down until they underflow through 0. That is
not good enough, but is better dealt with in a followup patch.
Update a few comments on first and second tail page overlaid fields.
hugepage_add_new_anon_rmap() has to "increment" compound_mapcount, but
subpages_mapcount and compound_pincount are already correctly at 0, so
delete its reinitialization of compound_pincount.
A simple 100 X munmap(mmap(2GB, MAP_SHARED|MAP_POPULATE, tmpfs), 2GB) took
18 seconds on small pages, and used to take 1 second on huge pages, but
now takes 119 milliseconds on huge pages. Mapping by pmds a second time
used to take 860ms and now takes 92ms; mapping by pmds after mapping by
ptes (when the scan is needed) used to take 870ms and now takes 495ms.
But there might be some benchmarks which would show a slowdown, because
tail struct pages now fall out of cache until final freeing checks them.
Link: https://lkml.kernel.org/r/47ad693-717-79c8-e1ba-46c3a6602e48@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: James Houghton <jthoughton@google.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mina Almasry <almasrymina@google.com>
Cc: Muchun Song <songmuchun@bytedance.com>
Cc: Naoya Horiguchi <naoya.horiguchi@linux.dev>
Cc: Peter Xu <peterx@redhat.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Zach O'Keefe <zokeefe@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-11-02 18:51:38 -07:00
|
|
|
|
mm: track mapcount of large folios in single value
Let's track the mapcount of large folios in a single value. The mapcount
of a large folio currently corresponds to the sum of the entire mapcount
and all page mapcounts.
This sum is what we actually want to know in folio_mapcount() and it is
also sufficient for implementing folio_mapped().
With PTE-mapped THP becoming more important and more widely used, we want
to avoid looping over all pages of a folio just to obtain the mapcount of
large folios. The comment "In the common case, avoid the loop when no
pages mapped by PTE" in folio_total_mapcount() does no longer hold for
mTHP that are always mapped by PTE.
Further, we are planning on using folio_mapcount() more frequently, and
might even want to remove page mapcounts for large folios in some kernel
configs. Therefore, allow for reading the mapcount of large folios
efficiently and atomically without looping over any pages.
Maintain the mapcount also for hugetlb pages for simplicity. Use the new
mapcount to implement folio_mapcount() and folio_mapped(). Make
page_mapped() simply call folio_mapped(). We can now get rid of
folio_large_is_mapped().
_nr_pages_mapped is now only used in rmap code and for debugging purposes.
Keep folio_nr_pages_mapped() around, but document that its use should be
limited to rmap internals and debugging purposes.
This change implies one additional atomic add/sub whenever
mapping/unmapping (parts of) a large folio.
As we now batch RMAP operations for PTE-mapped THP during fork(), during
unmap/zap, and when PTE-remapping a PMD-mapped THP, and we adjust the
large mapcount for a PTE batch only once, the added overhead in the common
case is small. Only when unmapping individual pages of a large folio
(e.g., during COW), the overhead might be bigger in comparison, but it's
essentially one additional atomic operation.
Note that before the new mapcount would overflow, already our refcount
would overflow: each mapping requires a folio reference. Extend the
focumentation of folio_mapcount().
Link: https://lkml.kernel.org/r/20240409192301.907377-5-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Reviewed-by: Yin Fengwei <fengwei.yin@intel.com>
Cc: Chris Zankel <chris@zankel.net>
Cc: Hugh Dickins <hughd@google.com>
Cc: John Paul Adrian Glaubitz <glaubitz@physik.fu-berlin.de>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Max Filippov <jcmvbkbc@gmail.com>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Naoya Horiguchi <nao.horiguchi@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Richard Chang <richardycc@google.com>
Cc: Rich Felker <dalias@libc.org>
Cc: Ryan Roberts <ryan.roberts@arm.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yoshinori Sato <ysato@users.sourceforge.jp>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-04-09 21:22:47 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: stop maintaining the per-page mapcount of large folios (CONFIG_NO_PAGE_MAPCOUNT)
Everything is in place to stop using the per-page mapcounts in large
folios: the mapcount of tail pages will always be logically 0 (-1 value),
just like it currently is for hugetlb folios already, and the page
mapcount of the head page is either 0 (-1 value) or contains a page type
(e.g., hugetlb).
Maintaining _nr_pages_mapped without per-page mapcounts is impossible, so
that one also has to go with CONFIG_NO_PAGE_MAPCOUNT.
There are two remaining implications:
(1) Per-node, per-cgroup and per-lruvec stats of "NR_ANON_MAPPED"
("mapped anonymous memory") and "NR_FILE_MAPPED"
("mapped file memory"):
As soon as any page of the folio is mapped -- folio_mapped() -- we
now account the complete folio as mapped. Once the last page is
unmapped -- !folio_mapped() -- we account the complete folio as
unmapped.
This implies that ...
* "AnonPages" and "Mapped" in /proc/meminfo and
/sys/devices/system/node/*/meminfo
* cgroup v2: "anon" and "file_mapped" in "memory.stat" and
"memory.numa_stat"
* cgroup v1: "rss" and "mapped_file" in "memory.stat" and
"memory.numa_stat
... can now appear higher than before. But note that these folios do
consume that memory, simply not all pages are actually currently
mapped.
It's worth nothing that other accounting in the kernel (esp. cgroup
charging on allocation) is not affected by this change.
[why oh why is "anon" called "rss" in cgroup v1]
(2) Detecting partial mappings
Detecting whether anon THPs are partially mapped gets a bit more
unreliable. As long as a single MM maps such a large folio
("exclusively mapped"), we can reliably detect it. Especially before
fork() / after a short-lived child process quit, we will detect
partial mappings reliably, which is the common case.
In essence, if the average per-page mapcount in an anon THP is < 1,
we know for sure that we have a partial mapping.
However, as soon as multiple MMs are involved, we might miss detecting
partial mappings: this might be relevant with long-lived child
processes. If we have a fully-mapped anon folio before fork(), once
our child processes and our parent all unmap (zap/COW) the same pages
(but not the complete folio), we might not detect the partial mapping.
However, once the child processes quit we would detect the partial
mapping.
How relevant this case is in practice remains to be seen.
Swapout/migration will likely mitigate this.
In the future, RMAP walkers could check for that for that case
(e.g., when collecting access bits during reclaim) and simply flag
them for deferred-splitting.
Link: https://lkml.kernel.org/r/20250303163014.1128035-21-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Cc: Andy Lutomirks^H^Hski <luto@kernel.org>
Cc: Borislav Betkov <bp@alien8.de>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jann Horn <jannh@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: Liam Howlett <liam.howlett@oracle.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Michal Koutn <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: tejun heo <tj@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-03-03 17:30:13 +01:00
|
|
|
|
|
|
|
|
|
2023-01-11 14:28:55 +00:00
|
|
|
|
mm,thp,rmap: simplify compound page mapcount handling
Compound page (folio) mapcount calculations have been different for anon
and file (or shmem) THPs, and involved the obscure PageDoubleMap flag.
And each huge mapping and unmapping of a file (or shmem) THP involved
atomically incrementing and decrementing the mapcount of every subpage of
that huge page, dirtying many struct page cachelines.
Add subpages_mapcount field to the struct folio and first tail page, so
that the total of subpage mapcounts is available in one place near the
head: then page_mapcount() and total_mapcount() and page_mapped(), and
their folio equivalents, are so quick that anon and file and hugetlb don't
need to be optimized differently. Delete the unloved PageDoubleMap.
page_add and page_remove rmap functions must now maintain the
subpages_mapcount as well as the subpage _mapcount, when dealing with pte
mappings of huge pages; and correct maintenance of NR_ANON_MAPPED and
NR_FILE_MAPPED statistics still needs reading through the subpages, using
nr_subpages_unmapped() - but only when first or last pmd mapping finds
subpages_mapcount raised (double-map case, not the common case).
But are those counts (used to decide when to split an anon THP, and in
vmscan's pagecache_reclaimable heuristic) correctly maintained? Not
quite: since page_remove_rmap() (and also split_huge_pmd()) is often
called without page lock, there can be races when a subpage pte mapcount
0<->1 while compound pmd mapcount 0<->1 is scanning - races which the
previous implementation had prevented. The statistics might become
inaccurate, and even drift down until they underflow through 0. That is
not good enough, but is better dealt with in a followup patch.
Update a few comments on first and second tail page overlaid fields.
hugepage_add_new_anon_rmap() has to "increment" compound_mapcount, but
subpages_mapcount and compound_pincount are already correctly at 0, so
delete its reinitialization of compound_pincount.
A simple 100 X munmap(mmap(2GB, MAP_SHARED|MAP_POPULATE, tmpfs), 2GB) took
18 seconds on small pages, and used to take 1 second on huge pages, but
now takes 119 milliseconds on huge pages. Mapping by pmds a second time
used to take 860ms and now takes 92ms; mapping by pmds after mapping by
ptes (when the scan is needed) used to take 870ms and now takes 495ms.
But there might be some benchmarks which would show a slowdown, because
tail struct pages now fall out of cache until final freeing checks them.
Link: https://lkml.kernel.org/r/47ad693-717-79c8-e1ba-46c3a6602e48@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: James Houghton <jthoughton@google.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mina Almasry <almasrymina@google.com>
Cc: Muchun Song <songmuchun@bytedance.com>
Cc: Naoya Horiguchi <naoya.horiguchi@linux.dev>
Cc: Peter Xu <peterx@redhat.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Zach O'Keefe <zokeefe@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-11-02 18:51:38 -07:00
|
|
|
|
|
|
|
|
|
mm/rmap: basic MM owner tracking for large folios (!hugetlb)
For small folios, we traditionally use the mapcount to decide whether it
was "certainly mapped exclusively" by a single MM (mapcount == 1) or
whether it "maybe mapped shared" by multiple MMs (mapcount > 1). For
PMD-sized folios that were PMD-mapped, we were able to use a similar
mechanism (single PMD mapping), but for PTE-mapped folios and in the
future folios that span multiple PMDs, this does not work.
So we need a different mechanism to handle large folios. Let's add a new
mechanism to detect whether a large folio is "certainly mapped
exclusively", or whether it is "maybe mapped shared".
We'll use this information next to optimize CoW reuse for PTE-mapped
anonymous THP, and to convert folio_likely_mapped_shared() to
folio_maybe_mapped_shared(), independent of per-page mapcounts.
For each large folio, we'll have two slots, whereby a slot stores:
(1) an MM id: unique id assigned to each MM
(2) a per-MM mapcount
If a slot is unoccupied, it can be taken by the next MM that maps folio
page.
In addition, we'll remember the current state -- "mapped exclusively" vs.
"maybe mapped shared" -- and use a bit spinlock to sync on updates and to
reduce the total number of atomic accesses on updates. In the future, it
might be possible to squeeze a proper spinlock into "struct folio". For
now, keep it simple, as we require the whole thing with THP only, that is
incompatible with RT.
As we have to squeeze this information into the "struct folio" of even
folios of order-1 (2 pages), and we generally want to reduce the required
metadata, we'll assign each MM a unique ID that can fit into an int. In
total, we can squeeze everything into 4x int (2x long) on 64bit.
32bit support is a bit challenging, because we only have 2x long == 2x int
in order-1 folios. But we can make it work for now, because we neither
expect many MMs nor very large folios on 32bit.
We will reliably detect folios as "mapped exclusively" vs. "mapped
shared" as long as only two MMs map pages of a folio at one point in time
-- for example with fork() and short-lived child processes, or with apps
that hand over state from one instance to another.
As soon as three MMs are involved at the same time, we might detect "maybe
mapped shared" although the folio is "mapped exclusively".
Example 1:
(1) App1 faults in a (shmem/file-backed) folio page -> Tracked as MM0
(2) App2 faults in a folio page -> Tracked as MM1
(4) App1 unmaps all folio pages
-> We will detect "mapped exclusively".
Example 2:
(1) App1 faults in a (shmem/file-backed) folio page -> Tracked as MM0
(2) App2 faults in a folio page -> Tracked as MM1
(3) App3 faults in a folio page -> No slot available, tracked as "unknown"
(4) App1 and App2 unmap all folio pages
-> We will detect "maybe mapped shared".
Make use of __always_inline to keep possible performance degradation when
(un)mapping large folios to a minimum.
Note: by squeezing the two flags into the "unsigned long" that stores the
MM ids, we can use non-atomic __bit_spin_unlock() and non-atomic
setting/clearing of the "maybe mapped shared" bit, effectively not adding
any new atomics on the hot path when updating the large mapcount + new
metadata, which further helps reduce the runtime overhead in
micro-benchmarks.
Link: https://lkml.kernel.org/r/20250303163014.1128035-13-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Cc: Andy Lutomirks^H^Hski <luto@kernel.org>
Cc: Borislav Betkov <bp@alien8.de>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jann Horn <jannh@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: Liam Howlett <liam.howlett@oracle.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Michal Koutn <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: tejun heo <tj@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-03-03 17:30:05 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: move _pincount in folio to page[2] on 32bit
Let's free up some space on 32bit in page[1] by moving the _pincount to
page[2].
For order-1 folios (never anon folios!) on 32bit, we will now also use the
GUP_PIN_COUNTING_BIAS approach. A fully-mapped order-1 folio requires 2
references. With GUP_PIN_COUNTING_BIAS being 1024, we'd detect such
folios as "maybe pinned" with 512 full mappings, instead of 1024 for
order-0. As anon folios are out of the picture (which are the most
relevant users of checking for pinnings on *mapped* pages) and we are
talking about 32bit, this is not expected to cause any trouble.
In __dump_page(), copy one additional folio page if we detect a folio with
an order > 1, so we can dump the pincount on order > 1 folios reliably.
Note that THPs on 32bit are not particularly common (and we don't care too
much about performance), but we want to keep it working reliably, because
likely we want to use large folios there as well in the future,
independent of PMD leaf support.
Once we dynamically allocate "struct folio", fortunately the 32bit
specifics will likely go away again; even small folios could then have a
pincount and folio_has_pincount() would essentially always return "true".
Link: https://lkml.kernel.org/r/20250303163014.1128035-6-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Cc: Andy Lutomirks^H^Hski <luto@kernel.org>
Cc: Borislav Betkov <bp@alien8.de>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jann Horn <jannh@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: Liam Howlett <liam.howlett@oracle.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Michal Koutn <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: tejun heo <tj@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-03-03 17:29:58 +01:00
|
|
|
|
2025-03-03 17:29:59 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: move _pincount in folio to page[2] on 32bit
Let's free up some space on 32bit in page[1] by moving the _pincount to
page[2].
For order-1 folios (never anon folios!) on 32bit, we will now also use the
GUP_PIN_COUNTING_BIAS approach. A fully-mapped order-1 folio requires 2
references. With GUP_PIN_COUNTING_BIAS being 1024, we'd detect such
folios as "maybe pinned" with 512 full mappings, instead of 1024 for
order-0. As anon folios are out of the picture (which are the most
relevant users of checking for pinnings on *mapped* pages) and we are
talking about 32bit, this is not expected to cause any trouble.
In __dump_page(), copy one additional folio page if we detect a folio with
an order > 1, so we can dump the pincount on order > 1 folios reliably.
Note that THPs on 32bit are not particularly common (and we don't care too
much about performance), but we want to keep it working reliably, because
likely we want to use large folios there as well in the future,
independent of PMD leaf support.
Once we dynamically allocate "struct folio", fortunately the 32bit
specifics will likely go away again; even small folios could then have a
pincount and folio_has_pincount() would essentially always return "true".
Link: https://lkml.kernel.org/r/20250303163014.1128035-6-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Cc: Andy Lutomirks^H^Hski <luto@kernel.org>
Cc: Borislav Betkov <bp@alien8.de>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jann Horn <jannh@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: Liam Howlett <liam.howlett@oracle.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Michal Koutn <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: tejun heo <tj@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-03-03 17:29:58 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm,thp,rmap: simplify compound page mapcount handling
Compound page (folio) mapcount calculations have been different for anon
and file (or shmem) THPs, and involved the obscure PageDoubleMap flag.
And each huge mapping and unmapping of a file (or shmem) THP involved
atomically incrementing and decrementing the mapcount of every subpage of
that huge page, dirtying many struct page cachelines.
Add subpages_mapcount field to the struct folio and first tail page, so
that the total of subpage mapcounts is available in one place near the
head: then page_mapcount() and total_mapcount() and page_mapped(), and
their folio equivalents, are so quick that anon and file and hugetlb don't
need to be optimized differently. Delete the unloved PageDoubleMap.
page_add and page_remove rmap functions must now maintain the
subpages_mapcount as well as the subpage _mapcount, when dealing with pte
mappings of huge pages; and correct maintenance of NR_ANON_MAPPED and
NR_FILE_MAPPED statistics still needs reading through the subpages, using
nr_subpages_unmapped() - but only when first or last pmd mapping finds
subpages_mapcount raised (double-map case, not the common case).
But are those counts (used to decide when to split an anon THP, and in
vmscan's pagecache_reclaimable heuristic) correctly maintained? Not
quite: since page_remove_rmap() (and also split_huge_pmd()) is often
called without page lock, there can be races when a subpage pte mapcount
0<->1 while compound pmd mapcount 0<->1 is scanning - races which the
previous implementation had prevented. The statistics might become
inaccurate, and even drift down until they underflow through 0. That is
not good enough, but is better dealt with in a followup patch.
Update a few comments on first and second tail page overlaid fields.
hugepage_add_new_anon_rmap() has to "increment" compound_mapcount, but
subpages_mapcount and compound_pincount are already correctly at 0, so
delete its reinitialization of compound_pincount.
A simple 100 X munmap(mmap(2GB, MAP_SHARED|MAP_POPULATE, tmpfs), 2GB) took
18 seconds on small pages, and used to take 1 second on huge pages, but
now takes 119 milliseconds on huge pages. Mapping by pmds a second time
used to take 860ms and now takes 92ms; mapping by pmds after mapping by
ptes (when the scan is needed) used to take 870ms and now takes 495ms.
But there might be some benchmarks which would show a slowdown, because
tail struct pages now fall out of cache until final freeing checks them.
Link: https://lkml.kernel.org/r/47ad693-717-79c8-e1ba-46c3a6602e48@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: James Houghton <jthoughton@google.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mina Almasry <almasrymina@google.com>
Cc: Muchun Song <songmuchun@bytedance.com>
Cc: Naoya Horiguchi <naoya.horiguchi@linux.dev>
Cc: Peter Xu <peterx@redhat.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Zach O'Keefe <zokeefe@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-11-02 18:51:38 -07:00
|
|
|
|
2016-05-19 17:14:32 -07:00
|
|
|
|
|
|
|
|
|
2024-03-21 14:24:39 +00:00
|
|
|
|
mm/thp: fix deferred split queue not partially_mapped
Recent changes are putting more pressure on THP deferred split queues:
under load revealing long-standing races, causing list_del corruptions,
"Bad page state"s and worse (I keep BUGs in both of those, so usually
don't get to see how badly they end up without). The relevant recent
changes being 6.8's mTHP, 6.10's mTHP swapout, and 6.12's mTHP swapin,
improved swap allocation, and underused THP splitting.
The new unlocked list_del_init() in deferred_split_scan() is buggy. I
gave bad advice, it looks plausible since that's a local on-stack list,
but the fact is that it can race with a third party freeing or migrating
the preceding folio (properly unqueueing it with refcount 0 while holding
split_queue_lock), thereby corrupting the list linkage.
The obvious answer would be to take split_queue_lock there: but it has a
long history of contention, so I'm reluctant to add to that. Instead,
make sure that there is always one safe (raised refcount) folio before, by
delaying its folio_put(). (And of course I was wrong to suggest updating
split_queue_len without the lock: leave that until the splice.)
And remove two over-eager partially_mapped checks, restoring those tests
to how they were before: if uncharge_folio() or free_tail_page_prepare()
finds _deferred_list non-empty, it's in trouble whether or not that folio
is partially_mapped (and the flag was already cleared in the latter case).
Link: https://lkml.kernel.org/r/81e34a8b-113a-0701-740e-2135c97eb1d7@google.com
Fixes: dafff3f4c850 ("mm: split underused THPs")
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Usama Arif <usamaarif642@gmail.com>
Reviewed-by: David Hildenbrand <david@redhat.com>
Reviewed-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Acked-by: Zi Yan <ziy@nvidia.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Chris Li <chrisl@kernel.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Ryan Roberts <ryan.roberts@arm.com>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Wei Yang <richard.weiyang@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-10-27 12:59:34 -07:00
|
|
|
|
|
|
|
|
|
2024-03-21 14:24:39 +00:00
|
|
|
|
|
|
|
|
|
mm: move _pincount in folio to page[2] on 32bit
Let's free up some space on 32bit in page[1] by moving the _pincount to
page[2].
For order-1 folios (never anon folios!) on 32bit, we will now also use the
GUP_PIN_COUNTING_BIAS approach. A fully-mapped order-1 folio requires 2
references. With GUP_PIN_COUNTING_BIAS being 1024, we'd detect such
folios as "maybe pinned" with 512 full mappings, instead of 1024 for
order-0. As anon folios are out of the picture (which are the most
relevant users of checking for pinnings on *mapped* pages) and we are
talking about 32bit, this is not expected to cause any trouble.
In __dump_page(), copy one additional folio page if we detect a folio with
an order > 1, so we can dump the pincount on order > 1 folios reliably.
Note that THPs on 32bit are not particularly common (and we don't care too
much about performance), but we want to keep it working reliably, because
likely we want to use large folios there as well in the future,
independent of PMD leaf support.
Once we dynamically allocate "struct folio", fortunately the 32bit
specifics will likely go away again; even small folios could then have a
pincount and folio_has_pincount() would essentially always return "true".
Link: https://lkml.kernel.org/r/20250303163014.1128035-6-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Cc: Andy Lutomirks^H^Hski <luto@kernel.org>
Cc: Borislav Betkov <bp@alien8.de>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jann Horn <jannh@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: Liam Howlett <liam.howlett@oracle.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Michal Koutn <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: tejun heo <tj@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-03-03 17:29:58 +01:00
|
|
|
|
2025-03-03 17:29:59 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: move _pincount in folio to page[2] on 32bit
Let's free up some space on 32bit in page[1] by moving the _pincount to
page[2].
For order-1 folios (never anon folios!) on 32bit, we will now also use the
GUP_PIN_COUNTING_BIAS approach. A fully-mapped order-1 folio requires 2
references. With GUP_PIN_COUNTING_BIAS being 1024, we'd detect such
folios as "maybe pinned" with 512 full mappings, instead of 1024 for
order-0. As anon folios are out of the picture (which are the most
relevant users of checking for pinnings on *mapped* pages) and we are
talking about 32bit, this is not expected to cause any trouble.
In __dump_page(), copy one additional folio page if we detect a folio with
an order > 1, so we can dump the pincount on order > 1 folios reliably.
Note that THPs on 32bit are not particularly common (and we don't care too
much about performance), but we want to keep it working reliably, because
likely we want to use large folios there as well in the future,
independent of PMD leaf support.
Once we dynamically allocate "struct folio", fortunately the 32bit
specifics will likely go away again; even small folios could then have a
pincount and folio_has_pincount() would essentially always return "true".
Link: https://lkml.kernel.org/r/20250303163014.1128035-6-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Cc: Andy Lutomirks^H^Hski <luto@kernel.org>
Cc: Borislav Betkov <bp@alien8.de>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jann Horn <jannh@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: Liam Howlett <liam.howlett@oracle.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Michal Koutn <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: tejun heo <tj@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-03-03 17:29:58 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-05-19 17:14:32 -07:00
|
|
|
|
2025-03-03 17:29:57 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-05-19 17:14:32 -07:00
|
|
|
|
|
|
|
|
|
2020-06-03 15:58:29 -07:00
|
|
|
|
2016-05-19 17:14:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-06-03 15:58:29 -07:00
|
|
|
|
2016-05-19 17:14:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-06-03 15:58:29 -07:00
|
|
|
|
2016-05-19 17:14:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
kasan, page_alloc: deduplicate should_skip_kasan_poison
Patch series "kasan, vmalloc, arm64: add vmalloc tagging support for SW/HW_TAGS", v6.
This patchset adds vmalloc tagging support for SW_TAGS and HW_TAGS
KASAN modes.
About half of patches are cleanups I went for along the way. None of them
seem to be important enough to go through stable, so I decided not to
split them out into separate patches/series.
The patchset is partially based on an early version of the HW_TAGS
patchset by Vincenzo that had vmalloc support. Thus, I added a
Co-developed-by tag into a few patches.
SW_TAGS vmalloc tagging support is straightforward. It reuses all of the
generic KASAN machinery, but uses shadow memory to store tags instead of
magic values. Naturally, vmalloc tagging requires adding a few
kasan_reset_tag() annotations to the vmalloc code.
HW_TAGS vmalloc tagging support stands out. HW_TAGS KASAN is based on Arm
MTE, which can only assigns tags to physical memory. As a result, HW_TAGS
KASAN only tags vmalloc() allocations, which are backed by page_alloc
memory. It ignores vmap() and others.
This patch (of 39):
Currently, should_skip_kasan_poison() has two definitions: one for when
CONFIG_DEFERRED_STRUCT_PAGE_INIT is enabled, one for when it's not.
Instead of duplicating the checks, add a deferred_pages_enabled() helper
and use it in a single should_skip_kasan_poison() definition.
Also move should_skip_kasan_poison() closer to its caller and clarify all
conditions in the comment.
Link: https://lkml.kernel.org/r/cover.1643047180.git.andreyknvl@google.com
Link: https://lkml.kernel.org/r/658b79f5fb305edaf7dc16bc52ea870d3220d4a8.1643047180.git.andreyknvl@google.com
Signed-off-by: Andrey Konovalov <andreyknvl@google.com>
Acked-by: Marco Elver <elver@google.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Andrey Ryabinin <ryabinin.a.a@gmail.com>
Cc: Vincenzo Frascino <vincenzo.frascino@arm.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Will Deacon <will@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Peter Collingbourne <pcc@google.com>
Cc: Evgenii Stepanov <eugenis@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-03-24 18:10:04 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-03-09 20:29:14 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
kasan: allow sampling page_alloc allocations for HW_TAGS
As Hardware Tag-Based KASAN is intended to be used in production, its
performance impact is crucial. As page_alloc allocations tend to be big,
tagging and checking all such allocations can introduce a significant
slowdown.
Add two new boot parameters that allow to alleviate that slowdown:
- kasan.page_alloc.sample, which makes Hardware Tag-Based KASAN tag only
every Nth page_alloc allocation with the order configured by the second
added parameter (default: tag every such allocation).
- kasan.page_alloc.sample.order, which makes sampling enabled by the first
parameter only affect page_alloc allocations with the order equal or
greater than the specified value (default: 3, see below).
The exact performance improvement caused by using the new parameters
depends on their values and the applied workload.
The chosen default value for kasan.page_alloc.sample.order is 3, which
matches both PAGE_ALLOC_COSTLY_ORDER and SKB_FRAG_PAGE_ORDER. This is
done for two reasons:
1. PAGE_ALLOC_COSTLY_ORDER is "the order at which allocations are deemed
costly to service", which corresponds to the idea that only large and
thus costly allocations are supposed to sampled.
2. One of the workloads targeted by this patch is a benchmark that sends
a large amount of data over a local loopback connection. Most multi-page
data allocations in the networking subsystem have the order of
SKB_FRAG_PAGE_ORDER (or PAGE_ALLOC_COSTLY_ORDER).
When running a local loopback test on a testing MTE-enabled device in sync
mode, enabling Hardware Tag-Based KASAN introduces a ~50% slowdown.
Applying this patch and setting kasan.page_alloc.sampling to a value
higher than 1 allows to lower the slowdown. The performance improvement
saturates around the sampling interval value of 10 with the default
sampling page order of 3. This lowers the slowdown to ~20%. The slowdown
in real scenarios involving the network will likely be better.
Enabling page_alloc sampling has a downside: KASAN misses bad accesses to
a page_alloc allocation that has not been tagged. This lowers the value
of KASAN as a security mitigation.
However, based on measuring the number of page_alloc allocations of
different orders during boot in a test build, sampling with the default
kasan.page_alloc.sample.order value affects only ~7% of allocations. The
rest ~93% of allocations are still checked deterministically.
Link: https://lkml.kernel.org/r/129da0614123bb85ed4dd61ae30842b2dd7c903f.1671471846.git.andreyknvl@google.com
Signed-off-by: Andrey Konovalov <andreyknvl@google.com>
Reviewed-by: Marco Elver <elver@google.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Andrey Ryabinin <ryabinin.a.a@gmail.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Evgenii Stepanov <eugenis@google.com>
Cc: Jann Horn <jannh@google.com>
Cc: Mark Brand <markbrand@google.com>
Cc: Peter Collingbourne <pcc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-19 19:09:18 +01:00
|
|
|
|
kasan, page_alloc: deduplicate should_skip_kasan_poison
Patch series "kasan, vmalloc, arm64: add vmalloc tagging support for SW/HW_TAGS", v6.
This patchset adds vmalloc tagging support for SW_TAGS and HW_TAGS
KASAN modes.
About half of patches are cleanups I went for along the way. None of them
seem to be important enough to go through stable, so I decided not to
split them out into separate patches/series.
The patchset is partially based on an early version of the HW_TAGS
patchset by Vincenzo that had vmalloc support. Thus, I added a
Co-developed-by tag into a few patches.
SW_TAGS vmalloc tagging support is straightforward. It reuses all of the
generic KASAN machinery, but uses shadow memory to store tags instead of
magic values. Naturally, vmalloc tagging requires adding a few
kasan_reset_tag() annotations to the vmalloc code.
HW_TAGS vmalloc tagging support stands out. HW_TAGS KASAN is based on Arm
MTE, which can only assigns tags to physical memory. As a result, HW_TAGS
KASAN only tags vmalloc() allocations, which are backed by page_alloc
memory. It ignores vmap() and others.
This patch (of 39):
Currently, should_skip_kasan_poison() has two definitions: one for when
CONFIG_DEFERRED_STRUCT_PAGE_INIT is enabled, one for when it's not.
Instead of duplicating the checks, add a deferred_pages_enabled() helper
and use it in a single should_skip_kasan_poison() definition.
Also move should_skip_kasan_poison() closer to its caller and clarify all
conditions in the comment.
Link: https://lkml.kernel.org/r/cover.1643047180.git.andreyknvl@google.com
Link: https://lkml.kernel.org/r/658b79f5fb305edaf7dc16bc52ea870d3220d4a8.1643047180.git.andreyknvl@google.com
Signed-off-by: Andrey Konovalov <andreyknvl@google.com>
Acked-by: Marco Elver <elver@google.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Andrey Ryabinin <ryabinin.a.a@gmail.com>
Cc: Vincenzo Frascino <vincenzo.frascino@arm.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Will Deacon <will@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Peter Collingbourne <pcc@google.com>
Cc: Evgenii Stepanov <eugenis@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-03-24 18:10:04 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/page_alloc: remove unused fpi_flags in free_pages_prepare()
Patch series "Enable >0 order folio memory compaction", v7.
This patchset enables >0 order folio memory compaction, which is one of
the prerequisitions for large folio support[1].
I am aware of that split free pages is necessary for folio migration in
compaction, since if >0 order free pages are never split and no order-0
free page is scanned, compaction will end prematurely due to migration
returns -ENOMEM. Free page split becomes a must instead of an
optimization.
lkp ncompare results (on a 8-CPU (Intel Xeon E5-2650 v4 @2.20GHz) 16G VM)
for default LRU (-no-mglru) and CONFIG_LRU_GEN are shown at the bottom,
copied from V3[4]. In sum, most of vm-scalability applications do not see
performance change, and the others see ~4% to ~26% performance boost under
default LRU and ~2% to ~6% performance boost under CONFIG_LRU_GEN.
Overview
===
To support >0 order folio compaction, the patchset changes how free pages
used for migration are kept during compaction. Free pages used to be
split into order-0 pages that are post allocation processed (i.e.,
PageBuddy flag cleared, page order stored in page->private is zeroed, and
page reference is set to 1). Now all free pages are kept in a
NR_PAGE_ORDER array of page lists based on their order without post
allocation process. When migrate_pages() asks for a new page, one of the
free pages, based on the requested page order, is then processed and given
out. And THP <2MB would need this feature.
[1] https://lore.kernel.org/linux-mm/f8d47176-03a8-99bf-a813-b5942830fd73@arm.com/
[2] https://lore.kernel.org/linux-mm/20231113170157.280181-1-zi.yan@sent.com/
[3] https://lore.kernel.org/linux-mm/20240123034636.1095672-1-zi.yan@sent.com/
[4] https://lore.kernel.org/linux-mm/20240202161554.565023-1-zi.yan@sent.com/
[5] https://lore.kernel.org/linux-mm/20240212163510.859822-1-zi.yan@sent.com/
[6] https://lore.kernel.org/linux-mm/20240214220420.1229173-1-zi.yan@sent.com/
[7] https://lore.kernel.org/linux-mm/20240216170432.1268753-1-zi.yan@sent.com/
This patch (of 4):
Commit 0a54864f8dfb ("kasan: remove PG_skip_kasan_poison flag") removes
the use of fpi_flags in should_skip_kasan_poison() and fpi_flags is only
passed to should_skip_kasan_poison() in free_pages_prepare(). Remove the
unused parameter.
Link: https://lkml.kernel.org/r/20240220183220.1451315-1-zi.yan@sent.com
Link: https://lkml.kernel.org/r/20240220183220.1451315-2-zi.yan@sent.com
Signed-off-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: David Hildenbrand <david@redhat.com>
Cc: Adam Manzanares <a.manzanares@samsung.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kemeng Shi <shikemeng@huaweicloud.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Luis Chamberlain <mcgrof@kernel.org>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Ryan Roberts <ryan.roberts@arm.com>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yin Fengwei <fengwei.yin@intel.com>
Cc: Yu Zhao <yuzhao@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-20 13:32:17 -05:00
|
|
|
|
kasan, page_alloc: deduplicate should_skip_kasan_poison
Patch series "kasan, vmalloc, arm64: add vmalloc tagging support for SW/HW_TAGS", v6.
This patchset adds vmalloc tagging support for SW_TAGS and HW_TAGS
KASAN modes.
About half of patches are cleanups I went for along the way. None of them
seem to be important enough to go through stable, so I decided not to
split them out into separate patches/series.
The patchset is partially based on an early version of the HW_TAGS
patchset by Vincenzo that had vmalloc support. Thus, I added a
Co-developed-by tag into a few patches.
SW_TAGS vmalloc tagging support is straightforward. It reuses all of the
generic KASAN machinery, but uses shadow memory to store tags instead of
magic values. Naturally, vmalloc tagging requires adding a few
kasan_reset_tag() annotations to the vmalloc code.
HW_TAGS vmalloc tagging support stands out. HW_TAGS KASAN is based on Arm
MTE, which can only assigns tags to physical memory. As a result, HW_TAGS
KASAN only tags vmalloc() allocations, which are backed by page_alloc
memory. It ignores vmap() and others.
This patch (of 39):
Currently, should_skip_kasan_poison() has two definitions: one for when
CONFIG_DEFERRED_STRUCT_PAGE_INIT is enabled, one for when it's not.
Instead of duplicating the checks, add a deferred_pages_enabled() helper
and use it in a single should_skip_kasan_poison() definition.
Also move should_skip_kasan_poison() closer to its caller and clarify all
conditions in the comment.
Link: https://lkml.kernel.org/r/cover.1643047180.git.andreyknvl@google.com
Link: https://lkml.kernel.org/r/658b79f5fb305edaf7dc16bc52ea870d3220d4a8.1643047180.git.andreyknvl@google.com
Signed-off-by: Andrey Konovalov <andreyknvl@google.com>
Acked-by: Marco Elver <elver@google.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Andrey Ryabinin <ryabinin.a.a@gmail.com>
Cc: Vincenzo Frascino <vincenzo.frascino@arm.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Will Deacon <will@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Peter Collingbourne <pcc@google.com>
Cc: Evgenii Stepanov <eugenis@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-03-24 18:10:04 -07:00
|
|
|
|
2023-03-09 20:29:14 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-12-21 21:04:44 +01:00
|
|
|
|
kasan, page_alloc: deduplicate should_skip_kasan_poison
Patch series "kasan, vmalloc, arm64: add vmalloc tagging support for SW/HW_TAGS", v6.
This patchset adds vmalloc tagging support for SW_TAGS and HW_TAGS
KASAN modes.
About half of patches are cleanups I went for along the way. None of them
seem to be important enough to go through stable, so I decided not to
split them out into separate patches/series.
The patchset is partially based on an early version of the HW_TAGS
patchset by Vincenzo that had vmalloc support. Thus, I added a
Co-developed-by tag into a few patches.
SW_TAGS vmalloc tagging support is straightforward. It reuses all of the
generic KASAN machinery, but uses shadow memory to store tags instead of
magic values. Naturally, vmalloc tagging requires adding a few
kasan_reset_tag() annotations to the vmalloc code.
HW_TAGS vmalloc tagging support stands out. HW_TAGS KASAN is based on Arm
MTE, which can only assigns tags to physical memory. As a result, HW_TAGS
KASAN only tags vmalloc() allocations, which are backed by page_alloc
memory. It ignores vmap() and others.
This patch (of 39):
Currently, should_skip_kasan_poison() has two definitions: one for when
CONFIG_DEFERRED_STRUCT_PAGE_INIT is enabled, one for when it's not.
Instead of duplicating the checks, add a deferred_pages_enabled() helper
and use it in a single should_skip_kasan_poison() definition.
Also move should_skip_kasan_poison() closer to its caller and clarify all
conditions in the comment.
Link: https://lkml.kernel.org/r/cover.1643047180.git.andreyknvl@google.com
Link: https://lkml.kernel.org/r/658b79f5fb305edaf7dc16bc52ea870d3220d4a8.1643047180.git.andreyknvl@google.com
Signed-off-by: Andrey Konovalov <andreyknvl@google.com>
Acked-by: Marco Elver <elver@google.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Andrey Ryabinin <ryabinin.a.a@gmail.com>
Cc: Vincenzo Frascino <vincenzo.frascino@arm.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Will Deacon <will@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Peter Collingbourne <pcc@google.com>
Cc: Evgenii Stepanov <eugenis@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-03-24 18:10:04 -07:00
|
|
|
|
|
|
|
|
|
2024-06-05 11:17:10 +02:00
|
|
|
|
mm: security: introduce init_on_alloc=1 and init_on_free=1 boot options
Patch series "add init_on_alloc/init_on_free boot options", v10.
Provide init_on_alloc and init_on_free boot options.
These are aimed at preventing possible information leaks and making the
control-flow bugs that depend on uninitialized values more deterministic.
Enabling either of the options guarantees that the memory returned by the
page allocator and SL[AU]B is initialized with zeroes. SLOB allocator
isn't supported at the moment, as its emulation of kmem caches complicates
handling of SLAB_TYPESAFE_BY_RCU caches correctly.
Enabling init_on_free also guarantees that pages and heap objects are
initialized right after they're freed, so it won't be possible to access
stale data by using a dangling pointer.
As suggested by Michal Hocko, right now we don't let the heap users to
disable initialization for certain allocations. There's not enough
evidence that doing so can speed up real-life cases, and introducing ways
to opt-out may result in things going out of control.
This patch (of 2):
The new options are needed to prevent possible information leaks and make
control-flow bugs that depend on uninitialized values more deterministic.
This is expected to be on-by-default on Android and Chrome OS. And it
gives the opportunity for anyone else to use it under distros too via the
boot args. (The init_on_free feature is regularly requested by folks
where memory forensics is included in their threat models.)
init_on_alloc=1 makes the kernel initialize newly allocated pages and heap
objects with zeroes. Initialization is done at allocation time at the
places where checks for __GFP_ZERO are performed.
init_on_free=1 makes the kernel initialize freed pages and heap objects
with zeroes upon their deletion. This helps to ensure sensitive data
doesn't leak via use-after-free accesses.
Both init_on_alloc=1 and init_on_free=1 guarantee that the allocator
returns zeroed memory. The two exceptions are slab caches with
constructors and SLAB_TYPESAFE_BY_RCU flag. Those are never
zero-initialized to preserve their semantics.
Both init_on_alloc and init_on_free default to zero, but those defaults
can be overridden with CONFIG_INIT_ON_ALLOC_DEFAULT_ON and
CONFIG_INIT_ON_FREE_DEFAULT_ON.
If either SLUB poisoning or page poisoning is enabled, those options take
precedence over init_on_alloc and init_on_free: initialization is only
applied to unpoisoned allocations.
Slowdown for the new features compared to init_on_free=0, init_on_alloc=0:
hackbench, init_on_free=1: +7.62% sys time (st.err 0.74%)
hackbench, init_on_alloc=1: +7.75% sys time (st.err 2.14%)
Linux build with -j12, init_on_free=1: +8.38% wall time (st.err 0.39%)
Linux build with -j12, init_on_free=1: +24.42% sys time (st.err 0.52%)
Linux build with -j12, init_on_alloc=1: -0.13% wall time (st.err 0.42%)
Linux build with -j12, init_on_alloc=1: +0.57% sys time (st.err 0.40%)
The slowdown for init_on_free=0, init_on_alloc=0 compared to the baseline
is within the standard error.
The new features are also going to pave the way for hardware memory
tagging (e.g. arm64's MTE), which will require both on_alloc and on_free
hooks to set the tags for heap objects. With MTE, tagging will have the
same cost as memory initialization.
Although init_on_free is rather costly, there are paranoid use-cases where
in-memory data lifetime is desired to be minimized. There are various
arguments for/against the realism of the associated threat models, but
given that we'll need the infrastructure for MTE anyway, and there are
people who want wipe-on-free behavior no matter what the performance cost,
it seems reasonable to include it in this series.
[glider@google.com: v8]
Link: http://lkml.kernel.org/r/20190626121943.131390-2-glider@google.com
[glider@google.com: v9]
Link: http://lkml.kernel.org/r/20190627130316.254309-2-glider@google.com
[glider@google.com: v10]
Link: http://lkml.kernel.org/r/20190628093131.199499-2-glider@google.com
Link: http://lkml.kernel.org/r/20190617151050.92663-2-glider@google.com
Signed-off-by: Alexander Potapenko <glider@google.com>
Acked-by: Kees Cook <keescook@chromium.org>
Acked-by: Michal Hocko <mhocko@suse.cz> [page and dmapool parts
Acked-by: James Morris <jamorris@linux.microsoft.com>]
Cc: Christoph Lameter <cl@linux.com>
Cc: Masahiro Yamada <yamada.masahiro@socionext.com>
Cc: "Serge E. Hallyn" <serge@hallyn.com>
Cc: Nick Desaulniers <ndesaulniers@google.com>
Cc: Kostya Serebryany <kcc@google.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Sandeep Patil <sspatil@android.com>
Cc: Laura Abbott <labbott@redhat.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Jann Horn <jannh@google.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Marco Elver <elver@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-11 20:59:19 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-08-06 23:25:54 -07:00
|
|
|
|
|
|
|
|
|
2022-06-09 20:18:46 +02:00
|
|
|
|
|
|
|
|
|
2020-08-06 23:25:54 -07:00
|
|
|
|
mm: security: introduce init_on_alloc=1 and init_on_free=1 boot options
Patch series "add init_on_alloc/init_on_free boot options", v10.
Provide init_on_alloc and init_on_free boot options.
These are aimed at preventing possible information leaks and making the
control-flow bugs that depend on uninitialized values more deterministic.
Enabling either of the options guarantees that the memory returned by the
page allocator and SL[AU]B is initialized with zeroes. SLOB allocator
isn't supported at the moment, as its emulation of kmem caches complicates
handling of SLAB_TYPESAFE_BY_RCU caches correctly.
Enabling init_on_free also guarantees that pages and heap objects are
initialized right after they're freed, so it won't be possible to access
stale data by using a dangling pointer.
As suggested by Michal Hocko, right now we don't let the heap users to
disable initialization for certain allocations. There's not enough
evidence that doing so can speed up real-life cases, and introducing ways
to opt-out may result in things going out of control.
This patch (of 2):
The new options are needed to prevent possible information leaks and make
control-flow bugs that depend on uninitialized values more deterministic.
This is expected to be on-by-default on Android and Chrome OS. And it
gives the opportunity for anyone else to use it under distros too via the
boot args. (The init_on_free feature is regularly requested by folks
where memory forensics is included in their threat models.)
init_on_alloc=1 makes the kernel initialize newly allocated pages and heap
objects with zeroes. Initialization is done at allocation time at the
places where checks for __GFP_ZERO are performed.
init_on_free=1 makes the kernel initialize freed pages and heap objects
with zeroes upon their deletion. This helps to ensure sensitive data
doesn't leak via use-after-free accesses.
Both init_on_alloc=1 and init_on_free=1 guarantee that the allocator
returns zeroed memory. The two exceptions are slab caches with
constructors and SLAB_TYPESAFE_BY_RCU flag. Those are never
zero-initialized to preserve their semantics.
Both init_on_alloc and init_on_free default to zero, but those defaults
can be overridden with CONFIG_INIT_ON_ALLOC_DEFAULT_ON and
CONFIG_INIT_ON_FREE_DEFAULT_ON.
If either SLUB poisoning or page poisoning is enabled, those options take
precedence over init_on_alloc and init_on_free: initialization is only
applied to unpoisoned allocations.
Slowdown for the new features compared to init_on_free=0, init_on_alloc=0:
hackbench, init_on_free=1: +7.62% sys time (st.err 0.74%)
hackbench, init_on_alloc=1: +7.75% sys time (st.err 2.14%)
Linux build with -j12, init_on_free=1: +8.38% wall time (st.err 0.39%)
Linux build with -j12, init_on_free=1: +24.42% sys time (st.err 0.52%)
Linux build with -j12, init_on_alloc=1: -0.13% wall time (st.err 0.42%)
Linux build with -j12, init_on_alloc=1: +0.57% sys time (st.err 0.40%)
The slowdown for init_on_free=0, init_on_alloc=0 compared to the baseline
is within the standard error.
The new features are also going to pave the way for hardware memory
tagging (e.g. arm64's MTE), which will require both on_alloc and on_free
hooks to set the tags for heap objects. With MTE, tagging will have the
same cost as memory initialization.
Although init_on_free is rather costly, there are paranoid use-cases where
in-memory data lifetime is desired to be minimized. There are various
arguments for/against the realism of the associated threat models, but
given that we'll need the infrastructure for MTE anyway, and there are
people who want wipe-on-free behavior no matter what the performance cost,
it seems reasonable to include it in this series.
[glider@google.com: v8]
Link: http://lkml.kernel.org/r/20190626121943.131390-2-glider@google.com
[glider@google.com: v9]
Link: http://lkml.kernel.org/r/20190627130316.254309-2-glider@google.com
[glider@google.com: v10]
Link: http://lkml.kernel.org/r/20190628093131.199499-2-glider@google.com
Link: http://lkml.kernel.org/r/20190617151050.92663-2-glider@google.com
Signed-off-by: Alexander Potapenko <glider@google.com>
Acked-by: Kees Cook <keescook@chromium.org>
Acked-by: Michal Hocko <mhocko@suse.cz> [page and dmapool parts
Acked-by: James Morris <jamorris@linux.microsoft.com>]
Cc: Christoph Lameter <cl@linux.com>
Cc: Masahiro Yamada <yamada.masahiro@socionext.com>
Cc: "Serge E. Hallyn" <serge@hallyn.com>
Cc: Nick Desaulniers <ndesaulniers@google.com>
Cc: Kostya Serebryany <kcc@google.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Sandeep Patil <sspatil@android.com>
Cc: Laura Abbott <labbott@redhat.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Jann Horn <jannh@google.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Marco Elver <elver@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-11 20:59:19 -07:00
|
|
|
|
|
|
|
|
|
2025-02-01 15:18:02 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-05-06 03:30:34 +08:00
|
|
|
|
|
|
|
|
|
2025-02-01 15:18:02 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-05-06 03:30:34 +08:00
|
|
|
|
2025-02-01 15:18:02 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-02-20 13:32:19 -05:00
|
|
|
|
mm/page_alloc: remove unused fpi_flags in free_pages_prepare()
Patch series "Enable >0 order folio memory compaction", v7.
This patchset enables >0 order folio memory compaction, which is one of
the prerequisitions for large folio support[1].
I am aware of that split free pages is necessary for folio migration in
compaction, since if >0 order free pages are never split and no order-0
free page is scanned, compaction will end prematurely due to migration
returns -ENOMEM. Free page split becomes a must instead of an
optimization.
lkp ncompare results (on a 8-CPU (Intel Xeon E5-2650 v4 @2.20GHz) 16G VM)
for default LRU (-no-mglru) and CONFIG_LRU_GEN are shown at the bottom,
copied from V3[4]. In sum, most of vm-scalability applications do not see
performance change, and the others see ~4% to ~26% performance boost under
default LRU and ~2% to ~6% performance boost under CONFIG_LRU_GEN.
Overview
===
To support >0 order folio compaction, the patchset changes how free pages
used for migration are kept during compaction. Free pages used to be
split into order-0 pages that are post allocation processed (i.e.,
PageBuddy flag cleared, page order stored in page->private is zeroed, and
page reference is set to 1). Now all free pages are kept in a
NR_PAGE_ORDER array of page lists based on their order without post
allocation process. When migrate_pages() asks for a new page, one of the
free pages, based on the requested page order, is then processed and given
out. And THP <2MB would need this feature.
[1] https://lore.kernel.org/linux-mm/f8d47176-03a8-99bf-a813-b5942830fd73@arm.com/
[2] https://lore.kernel.org/linux-mm/20231113170157.280181-1-zi.yan@sent.com/
[3] https://lore.kernel.org/linux-mm/20240123034636.1095672-1-zi.yan@sent.com/
[4] https://lore.kernel.org/linux-mm/20240202161554.565023-1-zi.yan@sent.com/
[5] https://lore.kernel.org/linux-mm/20240212163510.859822-1-zi.yan@sent.com/
[6] https://lore.kernel.org/linux-mm/20240214220420.1229173-1-zi.yan@sent.com/
[7] https://lore.kernel.org/linux-mm/20240216170432.1268753-1-zi.yan@sent.com/
This patch (of 4):
Commit 0a54864f8dfb ("kasan: remove PG_skip_kasan_poison flag") removes
the use of fpi_flags in should_skip_kasan_poison() and fpi_flags is only
passed to should_skip_kasan_poison() in free_pages_prepare(). Remove the
unused parameter.
Link: https://lkml.kernel.org/r/20240220183220.1451315-1-zi.yan@sent.com
Link: https://lkml.kernel.org/r/20240220183220.1451315-2-zi.yan@sent.com
Signed-off-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: David Hildenbrand <david@redhat.com>
Cc: Adam Manzanares <a.manzanares@samsung.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kemeng Shi <shikemeng@huaweicloud.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Luis Chamberlain <mcgrof@kernel.org>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Ryan Roberts <ryan.roberts@arm.com>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yin Fengwei <fengwei.yin@intel.com>
Cc: Yu Zhao <yuzhao@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-20 13:32:17 -05:00
|
|
|
|
2016-05-19 17:14:38 -07:00
|
|
|
|
|
|
|
|
|
mm/page_alloc: remove unused fpi_flags in free_pages_prepare()
Patch series "Enable >0 order folio memory compaction", v7.
This patchset enables >0 order folio memory compaction, which is one of
the prerequisitions for large folio support[1].
I am aware of that split free pages is necessary for folio migration in
compaction, since if >0 order free pages are never split and no order-0
free page is scanned, compaction will end prematurely due to migration
returns -ENOMEM. Free page split becomes a must instead of an
optimization.
lkp ncompare results (on a 8-CPU (Intel Xeon E5-2650 v4 @2.20GHz) 16G VM)
for default LRU (-no-mglru) and CONFIG_LRU_GEN are shown at the bottom,
copied from V3[4]. In sum, most of vm-scalability applications do not see
performance change, and the others see ~4% to ~26% performance boost under
default LRU and ~2% to ~6% performance boost under CONFIG_LRU_GEN.
Overview
===
To support >0 order folio compaction, the patchset changes how free pages
used for migration are kept during compaction. Free pages used to be
split into order-0 pages that are post allocation processed (i.e.,
PageBuddy flag cleared, page order stored in page->private is zeroed, and
page reference is set to 1). Now all free pages are kept in a
NR_PAGE_ORDER array of page lists based on their order without post
allocation process. When migrate_pages() asks for a new page, one of the
free pages, based on the requested page order, is then processed and given
out. And THP <2MB would need this feature.
[1] https://lore.kernel.org/linux-mm/f8d47176-03a8-99bf-a813-b5942830fd73@arm.com/
[2] https://lore.kernel.org/linux-mm/20231113170157.280181-1-zi.yan@sent.com/
[3] https://lore.kernel.org/linux-mm/20240123034636.1095672-1-zi.yan@sent.com/
[4] https://lore.kernel.org/linux-mm/20240202161554.565023-1-zi.yan@sent.com/
[5] https://lore.kernel.org/linux-mm/20240212163510.859822-1-zi.yan@sent.com/
[6] https://lore.kernel.org/linux-mm/20240214220420.1229173-1-zi.yan@sent.com/
[7] https://lore.kernel.org/linux-mm/20240216170432.1268753-1-zi.yan@sent.com/
This patch (of 4):
Commit 0a54864f8dfb ("kasan: remove PG_skip_kasan_poison flag") removes
the use of fpi_flags in should_skip_kasan_poison() and fpi_flags is only
passed to should_skip_kasan_poison() in free_pages_prepare(). Remove the
unused parameter.
Link: https://lkml.kernel.org/r/20240220183220.1451315-1-zi.yan@sent.com
Link: https://lkml.kernel.org/r/20240220183220.1451315-2-zi.yan@sent.com
Signed-off-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: David Hildenbrand <david@redhat.com>
Cc: Adam Manzanares <a.manzanares@samsung.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kemeng Shi <shikemeng@huaweicloud.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Luis Chamberlain <mcgrof@kernel.org>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Ryan Roberts <ryan.roberts@arm.com>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yin Fengwei <fengwei.yin@intel.com>
Cc: Yu Zhao <yuzhao@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-20 13:32:17 -05:00
|
|
|
|
2022-03-24 18:10:13 -07:00
|
|
|
|
2023-10-23 17:32:16 +09:00
|
|
|
|
2024-11-06 19:53:54 +00:00
|
|
|
|
2016-05-19 17:14:38 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-09-15 17:03:48 +02:00
|
|
|
|
2016-05-19 17:14:38 -07:00
|
|
|
|
2023-11-08 16:49:20 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-11-06 19:53:54 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm,hwpoison: rework soft offline for in-use pages
This patch changes the way we set and handle in-use poisoned pages. Until
now, poisoned pages were released to the buddy allocator, trusting that
the checks that take place at allocation time would act as a safe net and
would skip that page.
This has proved to be wrong, as we got some pfn walkers out there, like
compaction, that all they care is the page to be in a buddy freelist.
Although this might not be the only user, having poisoned pages in the
buddy allocator seems a bad idea as we should only have free pages that
are ready and meant to be used as such.
Before explaining the taken approach, let us break down the kind of pages
we can soft offline.
- Anonymous THP (after the split, they end up being 4K pages)
- Hugetlb
- Order-0 pages (that can be either migrated or invalited)
* Normal pages (order-0 and anon-THP)
- If they are clean and unmapped page cache pages, we invalidate
then by means of invalidate_inode_page().
- If they are mapped/dirty, we do the isolate-and-migrate dance.
Either way, do not call put_page directly from those paths. Instead, we
keep the page and send it to page_handle_poison to perform the right
handling.
page_handle_poison sets the HWPoison flag and does the last put_page.
Down the chain, we placed a check for HWPoison page in
free_pages_prepare, that just skips any poisoned page, so those pages
do not end up in any pcplist/freelist.
After that, we set the refcount on the page to 1 and we increment
the poisoned pages counter.
If we see that the check in free_pages_prepare creates trouble, we can
always do what we do for free pages:
- wait until the page hits buddy's freelists
- take it off, and flag it
The downside of the above approach is that we could race with an
allocation, so by the time we want to take the page off the buddy, the
page has been already allocated so we cannot soft offline it.
But the user could always retry it.
* Hugetlb pages
- We isolate-and-migrate them
After the migration has been successful, we call dissolve_free_huge_page,
and we set HWPoison on the page if we succeed.
Hugetlb has a slightly different handling though.
While for non-hugetlb pages we cared about closing the race with an
allocation, doing so for hugetlb pages requires quite some additional
and intrusive code (we would need to hook in free_huge_page and some other
places).
So I decided to not make the code overly complicated and just fail
normally if the page we allocated in the meantime.
We can always build on top of this.
As a bonus, because of the way we handle now in-use pages, we no longer
need the put-as-isolation-migratetype dance, that was guarding for poisoned
pages to end up in pcplists.
Signed-off-by: Oscar Salvador <osalvador@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Acked-by: Naoya Horiguchi <naoya.horiguchi@nec.com>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.ibm.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Cc: Aristeu Rozanski <aris@ruivo.org>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Dmitry Yakunin <zeil@yandex-team.ru>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Oscar Salvador <osalvador@suse.com>
Cc: Qian Cai <cai@lca.pw>
Cc: Tony Luck <tony.luck@intel.com>
Link: https://lkml.kernel.org/r/20200922135650.1634-10-osalvador@suse.de
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-10-15 20:07:09 -07:00
|
|
|
|
2023-11-08 16:49:20 +00:00
|
|
|
|
mm,hwpoison: rework soft offline for in-use pages
This patch changes the way we set and handle in-use poisoned pages. Until
now, poisoned pages were released to the buddy allocator, trusting that
the checks that take place at allocation time would act as a safe net and
would skip that page.
This has proved to be wrong, as we got some pfn walkers out there, like
compaction, that all they care is the page to be in a buddy freelist.
Although this might not be the only user, having poisoned pages in the
buddy allocator seems a bad idea as we should only have free pages that
are ready and meant to be used as such.
Before explaining the taken approach, let us break down the kind of pages
we can soft offline.
- Anonymous THP (after the split, they end up being 4K pages)
- Hugetlb
- Order-0 pages (that can be either migrated or invalited)
* Normal pages (order-0 and anon-THP)
- If they are clean and unmapped page cache pages, we invalidate
then by means of invalidate_inode_page().
- If they are mapped/dirty, we do the isolate-and-migrate dance.
Either way, do not call put_page directly from those paths. Instead, we
keep the page and send it to page_handle_poison to perform the right
handling.
page_handle_poison sets the HWPoison flag and does the last put_page.
Down the chain, we placed a check for HWPoison page in
free_pages_prepare, that just skips any poisoned page, so those pages
do not end up in any pcplist/freelist.
After that, we set the refcount on the page to 1 and we increment
the poisoned pages counter.
If we see that the check in free_pages_prepare creates trouble, we can
always do what we do for free pages:
- wait until the page hits buddy's freelists
- take it off, and flag it
The downside of the above approach is that we could race with an
allocation, so by the time we want to take the page off the buddy, the
page has been already allocated so we cannot soft offline it.
But the user could always retry it.
* Hugetlb pages
- We isolate-and-migrate them
After the migration has been successful, we call dissolve_free_huge_page,
and we set HWPoison on the page if we succeed.
Hugetlb has a slightly different handling though.
While for non-hugetlb pages we cared about closing the race with an
allocation, doing so for hugetlb pages requires quite some additional
and intrusive code (we would need to hook in free_huge_page and some other
places).
So I decided to not make the code overly complicated and just fail
normally if the page we allocated in the meantime.
We can always build on top of this.
As a bonus, because of the way we handle now in-use pages, we no longer
need the put-as-isolation-migratetype dance, that was guarding for poisoned
pages to end up in pcplists.
Signed-off-by: Oscar Salvador <osalvador@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Acked-by: Naoya Horiguchi <naoya.horiguchi@nec.com>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.ibm.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Cc: Aristeu Rozanski <aris@ruivo.org>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Dmitry Yakunin <zeil@yandex-team.ru>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Oscar Salvador <osalvador@suse.com>
Cc: Qian Cai <cai@lca.pw>
Cc: Tony Luck <tony.luck@intel.com>
Link: https://lkml.kernel.org/r/20200922135650.1634-10-osalvador@suse.de
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-10-15 20:07:09 -07:00
|
|
|
|
2022-01-14 14:06:37 -08:00
|
|
|
|
2024-03-21 09:36:36 -07:00
|
|
|
|
2024-08-26 00:36:49 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm,hwpoison: rework soft offline for in-use pages
This patch changes the way we set and handle in-use poisoned pages. Until
now, poisoned pages were released to the buddy allocator, trusting that
the checks that take place at allocation time would act as a safe net and
would skip that page.
This has proved to be wrong, as we got some pfn walkers out there, like
compaction, that all they care is the page to be in a buddy freelist.
Although this might not be the only user, having poisoned pages in the
buddy allocator seems a bad idea as we should only have free pages that
are ready and meant to be used as such.
Before explaining the taken approach, let us break down the kind of pages
we can soft offline.
- Anonymous THP (after the split, they end up being 4K pages)
- Hugetlb
- Order-0 pages (that can be either migrated or invalited)
* Normal pages (order-0 and anon-THP)
- If they are clean and unmapped page cache pages, we invalidate
then by means of invalidate_inode_page().
- If they are mapped/dirty, we do the isolate-and-migrate dance.
Either way, do not call put_page directly from those paths. Instead, we
keep the page and send it to page_handle_poison to perform the right
handling.
page_handle_poison sets the HWPoison flag and does the last put_page.
Down the chain, we placed a check for HWPoison page in
free_pages_prepare, that just skips any poisoned page, so those pages
do not end up in any pcplist/freelist.
After that, we set the refcount on the page to 1 and we increment
the poisoned pages counter.
If we see that the check in free_pages_prepare creates trouble, we can
always do what we do for free pages:
- wait until the page hits buddy's freelists
- take it off, and flag it
The downside of the above approach is that we could race with an
allocation, so by the time we want to take the page off the buddy, the
page has been already allocated so we cannot soft offline it.
But the user could always retry it.
* Hugetlb pages
- We isolate-and-migrate them
After the migration has been successful, we call dissolve_free_huge_page,
and we set HWPoison on the page if we succeed.
Hugetlb has a slightly different handling though.
While for non-hugetlb pages we cared about closing the race with an
allocation, doing so for hugetlb pages requires quite some additional
and intrusive code (we would need to hook in free_huge_page and some other
places).
So I decided to not make the code overly complicated and just fail
normally if the page we allocated in the meantime.
We can always build on top of this.
As a bonus, because of the way we handle now in-use pages, we no longer
need the put-as-isolation-migratetype dance, that was guarding for poisoned
pages to end up in pcplists.
Signed-off-by: Oscar Salvador <osalvador@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Acked-by: Naoya Horiguchi <naoya.horiguchi@nec.com>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.ibm.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Cc: Aristeu Rozanski <aris@ruivo.org>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Dmitry Yakunin <zeil@yandex-team.ru>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Oscar Salvador <osalvador@suse.com>
Cc: Qian Cai <cai@lca.pw>
Cc: Tony Luck <tony.luck@intel.com>
Link: https://lkml.kernel.org/r/20200922135650.1634-10-osalvador@suse.de
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-10-15 20:07:09 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-23 17:32:16 +09:00
|
|
|
|
|
|
|
|
|
2016-05-19 17:14:38 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-03-03 17:29:56 +01:00
|
|
|
|
2025-08-05 18:22:51 +01:00
|
|
|
|
2025-03-03 17:29:56 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-05-19 17:14:38 -07:00
|
|
|
|
|
|
|
|
|
2023-04-05 16:28:40 +02:00
|
|
|
|
2023-03-21 19:05:01 +02:00
|
|
|
|
2023-04-05 16:28:40 +02:00
|
|
|
|
mm, page_alloc: reduce page alloc/free sanity checks
Historically, we have performed sanity checks on all struct pages being
allocated or freed, making sure they have no unexpected page flags or
certain field values. This can detect insufficient cleanup and some cases
of use-after-free, although on its own it can't always identify the
culprit. The result is a warning and the "bad page" being leaked.
The checks do need some cpu cycles, so in 4.7 with commits 479f854a207c
("mm, page_alloc: defer debugging checks of pages allocated from the PCP")
and 4db7548ccbd9 ("mm, page_alloc: defer debugging checks of freed pages
until a PCP drain") they were no longer performed in the hot paths when
allocating and freeing from pcplists, but only when pcplists are bypassed,
refilled or drained. For debugging purposes, with CONFIG_DEBUG_VM enabled
the checks were instead still done in the hot paths and not when refilling
or draining pcplists.
With 4462b32c9285 ("mm, page_alloc: more extensive free page checking with
debug_pagealloc"), enabling debug_pagealloc also moved the sanity checks
back to hot pahs. When both debug_pagealloc and CONFIG_DEBUG_VM are
enabled, the checks are done both in hotpaths and pcplist refill/drain.
Even though the non-debug default today might seem to be a sensible
tradeoff between overhead and ability to detect bad pages, on closer look
it's arguably not. As most allocations go through the pcplists, catching
any bad pages when refilling or draining pcplists has only a small chance,
insufficient for debugging or serious hardening purposes. On the other
hand the cost of the checks is concentrated in the already expensive
drain/refill batching operations, and those are done under the often
contended zone lock. That was recently identified as an issue for page
allocation and the zone lock contention reduced by moving the checks
outside of the locked section with a patch "mm: reduce lock contention of
pcp buffer refill", but the cost of the checks is still visible compared
to their removal [1]. In the pcplist draining path free_pcppages_bulk()
the checks are still done under zone->lock.
Thus, remove the checks from pcplist refill and drain paths completely.
Introduce a static key check_pages_enabled to control checks during page
allocation a freeing (whether pcplist is used or bypassed). The static
key is enabled if either is true:
- kernel is built with CONFIG_DEBUG_VM=y (debugging)
- debug_pagealloc or page poisoning is boot-time enabled (debugging)
- init_on_alloc or init_on_free is boot-time enabled (hardening)
The resulting user visible changes:
- no checks when draining/refilling pcplists - less overhead, with
likely no practical reduction of ability to catch bad pages
- no checks when bypassing pcplists in default config (no
debugging/hardening) - less overhead etc. as above
- on typical hardened kernels [2], checks are now performed on each page
allocation/free (previously only when bypassing/draining/refilling
pcplists) - the init_on_alloc/init_on_free enabled should be sufficient
indication for preferring more costly alloc/free operations for
hardening purposes and we shouldn't need to introduce another toggle
- code (various wrappers) removal and simplification
[1] https://lore.kernel.org/all/68ba44d8-6899-c018-dcb3-36f3a96e6bea@sra.uni-hannover.de/
[2] https://lore.kernel.org/all/63ebc499.a70a0220.9ac51.29ea@mx.google.com/
[akpm@linux-foundation.org: coding-style cleanups]
[akpm@linux-foundation.org: make check_pages_enabled static]
Link: https://lkml.kernel.org/r/20230216095131.17336-1-vbabka@suse.cz
Reported-by: Alexander Halbuer <halbuer@sra.uni-hannover.de>
Reported-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Kees Cook <keescook@chromium.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-02-16 10:51:31 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-05-19 17:14:38 -07:00
|
|
|
|
2025-08-05 18:22:51 +01:00
|
|
|
|
2016-05-19 17:14:38 -07:00
|
|
|
|
|
|
|
|
|
2025-07-04 12:25:17 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: count the number of anonymous THPs per size
Patch series "mm: count the number of anonymous THPs per size", v4.
Knowing the number of transparent anon THPs in the system is crucial
for performance analysis. It helps in understanding the ratio and
distribution of THPs versus small folios throughout the system.
Additionally, partial unmapping by userspace can lead to significant waste
of THPs over time and increase memory reclamation pressure. We need this
information for comprehensive system tuning.
This patch (of 2):
Let's track for each anonymous THP size, how many of them are currently
allocated. We'll track the complete lifespan of an anon THP, starting
when it becomes an anon THP ("large anon folio") (->mapping gets set),
until it gets freed (->mapping gets cleared).
Introduce a new "nr_anon" counter per THP size and adjust the
corresponding counter in the following cases:
* We allocate a new THP and call folio_add_new_anon_rmap() to map
it the first time and turn it into an anon THP.
* We split an anon THP into multiple smaller ones.
* We migrate an anon THP, when we prepare the destination.
* We free an anon THP back to the buddy.
Note that AnonPages in /proc/meminfo currently tracks the total number of
*mapped* anonymous *pages*, and therefore has slightly different
semantics. In the future, we might also want to track "nr_anon_mapped"
for each THP size, which might be helpful when comparing it to the number
of allocated anon THPs (long-term pinning, stuck in swapcache, memory
leaks, ...).
Further note that for now, we only track anon THPs after they got their
->mapping set, for example via folio_add_new_anon_rmap(). If we would
allocate some in the swapcache, they will only show up in the statistics
for now after they have been mapped to user space the first time, where we
call folio_add_new_anon_rmap().
[akpm@linux-foundation.org: documentation fixups, per David]
Link: https://lkml.kernel.org/r/3e8add35-e26b-443b-8a04-1078f4bc78f6@redhat.com
Link: https://lkml.kernel.org/r/20240824010441.21308-1-21cnbao@gmail.com
Link: https://lkml.kernel.org/r/20240824010441.21308-2-21cnbao@gmail.com
Signed-off-by: Barry Song <v-songbaohua@oppo.com>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: Chuanhua Han <hanchuanhua@oppo.com>
Cc: Kairui Song <kasong@tencent.com>
Cc: Kalesh Singh <kaleshsingh@google.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: Ryan Roberts <ryan.roberts@arm.com>
Cc: Shuai Yuan <yuanshuai@oppo.com>
Cc: Usama Arif <usamaarif642@gmail.com>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-08-24 13:04:40 +12:00
|
|
|
|
2025-07-04 12:24:58 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-03-21 19:05:01 +02:00
|
|
|
|
mm, page_alloc: reduce page alloc/free sanity checks
Historically, we have performed sanity checks on all struct pages being
allocated or freed, making sure they have no unexpected page flags or
certain field values. This can detect insufficient cleanup and some cases
of use-after-free, although on its own it can't always identify the
culprit. The result is a warning and the "bad page" being leaked.
The checks do need some cpu cycles, so in 4.7 with commits 479f854a207c
("mm, page_alloc: defer debugging checks of pages allocated from the PCP")
and 4db7548ccbd9 ("mm, page_alloc: defer debugging checks of freed pages
until a PCP drain") they were no longer performed in the hot paths when
allocating and freeing from pcplists, but only when pcplists are bypassed,
refilled or drained. For debugging purposes, with CONFIG_DEBUG_VM enabled
the checks were instead still done in the hot paths and not when refilling
or draining pcplists.
With 4462b32c9285 ("mm, page_alloc: more extensive free page checking with
debug_pagealloc"), enabling debug_pagealloc also moved the sanity checks
back to hot pahs. When both debug_pagealloc and CONFIG_DEBUG_VM are
enabled, the checks are done both in hotpaths and pcplist refill/drain.
Even though the non-debug default today might seem to be a sensible
tradeoff between overhead and ability to detect bad pages, on closer look
it's arguably not. As most allocations go through the pcplists, catching
any bad pages when refilling or draining pcplists has only a small chance,
insufficient for debugging or serious hardening purposes. On the other
hand the cost of the checks is concentrated in the already expensive
drain/refill batching operations, and those are done under the often
contended zone lock. That was recently identified as an issue for page
allocation and the zone lock contention reduced by moving the checks
outside of the locked section with a patch "mm: reduce lock contention of
pcp buffer refill", but the cost of the checks is still visible compared
to their removal [1]. In the pcplist draining path free_pcppages_bulk()
the checks are still done under zone->lock.
Thus, remove the checks from pcplist refill and drain paths completely.
Introduce a static key check_pages_enabled to control checks during page
allocation a freeing (whether pcplist is used or bypassed). The static
key is enabled if either is true:
- kernel is built with CONFIG_DEBUG_VM=y (debugging)
- debug_pagealloc or page poisoning is boot-time enabled (debugging)
- init_on_alloc or init_on_free is boot-time enabled (hardening)
The resulting user visible changes:
- no checks when draining/refilling pcplists - less overhead, with
likely no practical reduction of ability to catch bad pages
- no checks when bypassing pcplists in default config (no
debugging/hardening) - less overhead etc. as above
- on typical hardened kernels [2], checks are now performed on each page
allocation/free (previously only when bypassing/draining/refilling
pcplists) - the init_on_alloc/init_on_free enabled should be sufficient
indication for preferring more costly alloc/free operations for
hardening purposes and we shouldn't need to introduce another toggle
- code (various wrappers) removal and simplification
[1] https://lore.kernel.org/all/68ba44d8-6899-c018-dcb3-36f3a96e6bea@sra.uni-hannover.de/
[2] https://lore.kernel.org/all/63ebc499.a70a0220.9ac51.29ea@mx.google.com/
[akpm@linux-foundation.org: coding-style cleanups]
[akpm@linux-foundation.org: make check_pages_enabled static]
Link: https://lkml.kernel.org/r/20230216095131.17336-1-vbabka@suse.cz
Reported-by: Alexander Halbuer <halbuer@sra.uni-hannover.de>
Reported-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Kees Cook <keescook@chromium.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-02-16 10:51:31 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-05-19 17:14:38 -07:00
|
|
|
|
|
|
|
|
|
2025-08-05 18:22:51 +01:00
|
|
|
|
2016-05-19 17:14:38 -07:00
|
|
|
|
2022-01-14 14:06:37 -08:00
|
|
|
|
2024-03-21 09:36:36 -07:00
|
|
|
|
2016-05-19 17:14:38 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: security: introduce init_on_alloc=1 and init_on_free=1 boot options
Patch series "add init_on_alloc/init_on_free boot options", v10.
Provide init_on_alloc and init_on_free boot options.
These are aimed at preventing possible information leaks and making the
control-flow bugs that depend on uninitialized values more deterministic.
Enabling either of the options guarantees that the memory returned by the
page allocator and SL[AU]B is initialized with zeroes. SLOB allocator
isn't supported at the moment, as its emulation of kmem caches complicates
handling of SLAB_TYPESAFE_BY_RCU caches correctly.
Enabling init_on_free also guarantees that pages and heap objects are
initialized right after they're freed, so it won't be possible to access
stale data by using a dangling pointer.
As suggested by Michal Hocko, right now we don't let the heap users to
disable initialization for certain allocations. There's not enough
evidence that doing so can speed up real-life cases, and introducing ways
to opt-out may result in things going out of control.
This patch (of 2):
The new options are needed to prevent possible information leaks and make
control-flow bugs that depend on uninitialized values more deterministic.
This is expected to be on-by-default on Android and Chrome OS. And it
gives the opportunity for anyone else to use it under distros too via the
boot args. (The init_on_free feature is regularly requested by folks
where memory forensics is included in their threat models.)
init_on_alloc=1 makes the kernel initialize newly allocated pages and heap
objects with zeroes. Initialization is done at allocation time at the
places where checks for __GFP_ZERO are performed.
init_on_free=1 makes the kernel initialize freed pages and heap objects
with zeroes upon their deletion. This helps to ensure sensitive data
doesn't leak via use-after-free accesses.
Both init_on_alloc=1 and init_on_free=1 guarantee that the allocator
returns zeroed memory. The two exceptions are slab caches with
constructors and SLAB_TYPESAFE_BY_RCU flag. Those are never
zero-initialized to preserve their semantics.
Both init_on_alloc and init_on_free default to zero, but those defaults
can be overridden with CONFIG_INIT_ON_ALLOC_DEFAULT_ON and
CONFIG_INIT_ON_FREE_DEFAULT_ON.
If either SLUB poisoning or page poisoning is enabled, those options take
precedence over init_on_alloc and init_on_free: initialization is only
applied to unpoisoned allocations.
Slowdown for the new features compared to init_on_free=0, init_on_alloc=0:
hackbench, init_on_free=1: +7.62% sys time (st.err 0.74%)
hackbench, init_on_alloc=1: +7.75% sys time (st.err 2.14%)
Linux build with -j12, init_on_free=1: +8.38% wall time (st.err 0.39%)
Linux build with -j12, init_on_free=1: +24.42% sys time (st.err 0.52%)
Linux build with -j12, init_on_alloc=1: -0.13% wall time (st.err 0.42%)
Linux build with -j12, init_on_alloc=1: +0.57% sys time (st.err 0.40%)
The slowdown for init_on_free=0, init_on_alloc=0 compared to the baseline
is within the standard error.
The new features are also going to pave the way for hardware memory
tagging (e.g. arm64's MTE), which will require both on_alloc and on_free
hooks to set the tags for heap objects. With MTE, tagging will have the
same cost as memory initialization.
Although init_on_free is rather costly, there are paranoid use-cases where
in-memory data lifetime is desired to be minimized. There are various
arguments for/against the realism of the associated threat models, but
given that we'll need the infrastructure for MTE anyway, and there are
people who want wipe-on-free behavior no matter what the performance cost,
it seems reasonable to include it in this series.
[glider@google.com: v8]
Link: http://lkml.kernel.org/r/20190626121943.131390-2-glider@google.com
[glider@google.com: v9]
Link: http://lkml.kernel.org/r/20190627130316.254309-2-glider@google.com
[glider@google.com: v10]
Link: http://lkml.kernel.org/r/20190628093131.199499-2-glider@google.com
Link: http://lkml.kernel.org/r/20190617151050.92663-2-glider@google.com
Signed-off-by: Alexander Potapenko <glider@google.com>
Acked-by: Kees Cook <keescook@chromium.org>
Acked-by: Michal Hocko <mhocko@suse.cz> [page and dmapool parts
Acked-by: James Morris <jamorris@linux.microsoft.com>]
Cc: Christoph Lameter <cl@linux.com>
Cc: Masahiro Yamada <yamada.masahiro@socionext.com>
Cc: "Serge E. Hallyn" <serge@hallyn.com>
Cc: Nick Desaulniers <ndesaulniers@google.com>
Cc: Kostya Serebryany <kcc@google.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Sandeep Patil <sspatil@android.com>
Cc: Laura Abbott <labbott@redhat.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Jann Horn <jannh@google.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Marco Elver <elver@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-11 20:59:19 -07:00
|
|
|
|
2020-12-14 19:13:34 -08:00
|
|
|
|
|
|
|
|
|
2021-03-12 21:08:10 -08:00
|
|
|
|
2021-04-29 23:00:02 -07:00
|
|
|
|
2022-03-24 18:10:10 -07:00
|
|
|
|
2021-04-29 23:00:02 -07:00
|
|
|
|
|
|
|
|
|
2021-03-12 21:08:10 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Revert "kasan: drop skip_kasan_poison variable in free_pages_prepare"
This reverts commit 487a32ec24be819e747af8c2ab0d5c515508086a.
should_skip_kasan_poison() reads the PG_skip_kasan_poison flag from
page->flags. However, this line of code in free_pages_prepare():
page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
clears most of page->flags, including PG_skip_kasan_poison, before calling
should_skip_kasan_poison(), which meant that it would never return true as
a result of the page flag being set. Therefore, fix the code to call
should_skip_kasan_poison() before clearing the flags, as we were doing
before the reverted patch.
This fixes a measurable performance regression introduced in the reverted
commit, where munmap() takes longer than intended if HW tags KASAN is
supported and enabled at runtime. Without this patch, we see a
single-digit percentage performance regression in a particular
mmap()-heavy benchmark when enabling HW tags KASAN, and with the patch,
there is no statistically significant performance impact when enabling HW
tags KASAN.
Link: https://lkml.kernel.org/r/20230310042914.3805818-2-pcc@google.com
Fixes: 487a32ec24be ("kasan: drop skip_kasan_poison variable in free_pages_prepare")
Link: https://linux-review.googlesource.com/id/Ic4f13affeebd20548758438bb9ed9ca40e312b79
Signed-off-by: Peter Collingbourne <pcc@google.com>
Reviewed-by: Andrey Konovalov <andreyknvl@gmail.com>
Cc: Andrey Ryabinin <ryabinin.a.a@gmail.com>
Cc: Catalin Marinas <catalin.marinas@arm.com> [arm64]
Cc: Evgenii Stepanov <eugenis@google.com>
Cc: Vincenzo Frascino <vincenzo.frascino@arm.com>
Cc: Will Deacon <will@kernel.org>
Cc: <stable@vger.kernel.org> [6.1]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-03-09 20:29:13 -08:00
|
|
|
|
2022-03-24 18:10:13 -07:00
|
|
|
|
2021-03-12 21:08:10 -08:00
|
|
|
|
2022-03-24 18:10:16 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-06-09 20:18:45 +02:00
|
|
|
|
2022-03-24 18:10:16 -07:00
|
|
|
|
mm/page_alloc.c: fix a crash in free_pages_prepare()
On architectures like s390, arch_free_page() could mark the page unused
(set_page_unused()) and any access later would trigger a kernel panic.
Fix it by moving arch_free_page() after all possible accessing calls.
Hardware name: IBM 2964 N96 400 (z/VM 6.4.0)
Krnl PSW : 0404e00180000000 0000000026c2b96e (__free_pages_ok+0x34e/0x5d8)
R:0 T:1 IO:0 EX:0 Key:0 M:1 W:0 P:0 AS:3 CC:2 PM:0 RI:0 EA:3
Krnl GPRS: 0000000088d43af7 0000000000484000 000000000000007c 000000000000000f
000003d080012100 000003d080013fc0 0000000000000000 0000000000100000
00000000275cca48 0000000000000100 0000000000000008 000003d080010000
00000000000001d0 000003d000000000 0000000026c2b78a 000000002717fdb0
Krnl Code: 0000000026c2b95c: ec1100b30659 risbgn %r1,%r1,0,179,6
0000000026c2b962: e32014000036 pfd 2,1024(%r1)
#0000000026c2b968: d7ff10001000 xc 0(256,%r1),0(%r1)
>0000000026c2b96e: 41101100 la %r1,256(%r1)
0000000026c2b972: a737fff8 brctg %r3,26c2b962
0000000026c2b976: d7ff10001000 xc 0(256,%r1),0(%r1)
0000000026c2b97c: e31003400004 lg %r1,832
0000000026c2b982: ebff1430016a asi 5168(%r1),-1
Call Trace:
__free_pages_ok+0x16a/0x5d8)
memblock_free_all+0x206/0x290
mem_init+0x58/0x120
start_kernel+0x2b0/0x570
startup_continue+0x6a/0xc0
INFO: lockdep is turned off.
Last Breaking-Event-Address:
__free_pages_ok+0x372/0x5d8
Kernel panic - not syncing: Fatal exception: panic_on_oops
00: HCPGIR450W CP entered; disabled wait PSW 00020001 80000000 00000000 26A2379C
In the past, only kernel_poison_pages() would trigger this but it needs
"page_poison=on" kernel cmdline, and I suspect nobody tested that on
s390. Recently, kernel_init_free_pages() (commit 6471384af2a6 ("mm:
security: introduce init_on_alloc=1 and init_on_free=1 boot options"))
was added and could trigger this as well.
[akpm@linux-foundation.org: add comment]
Link: http://lkml.kernel.org/r/1569613623-16820-1-git-send-email-cai@lca.pw
Fixes: 8823b1dbc05f ("mm/page_poison.c: enable PAGE_POISONING as a separate option")
Fixes: 6471384af2a6 ("mm: security: introduce init_on_alloc=1 and init_on_free=1 boot options")
Signed-off-by: Qian Cai <cai@lca.pw>
Reviewed-by: Heiko Carstens <heiko.carstens@de.ibm.com>
Acked-by: Christian Borntraeger <borntraeger@de.ibm.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Cc: Alexander Duyck <alexander.duyck@gmail.com>
Cc: <stable@vger.kernel.org> [5.3+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-10-06 17:58:25 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: introduce debug_pagealloc_{map,unmap}_pages() helpers
Patch series "arch, mm: improve robustness of direct map manipulation", v7.
During recent discussion about KVM protected memory, David raised a
concern about usage of __kernel_map_pages() outside of DEBUG_PAGEALLOC
scope [1].
Indeed, for architectures that define CONFIG_ARCH_HAS_SET_DIRECT_MAP it is
possible that __kernel_map_pages() would fail, but since this function is
void, the failure will go unnoticed.
Moreover, there's lack of consistency of __kernel_map_pages() semantics
across architectures as some guard this function with #ifdef
DEBUG_PAGEALLOC, some refuse to update the direct map if page allocation
debugging is disabled at run time and some allow modifying the direct map
regardless of DEBUG_PAGEALLOC settings.
This set straightens this out by restoring dependency of
__kernel_map_pages() on DEBUG_PAGEALLOC and updating the call sites
accordingly.
Since currently the only user of __kernel_map_pages() outside
DEBUG_PAGEALLOC is hibernation, it is updated to make direct map accesses
there more explicit.
[1] https://lore.kernel.org/lkml/2759b4bf-e1e3-d006-7d86-78a40348269d@redhat.com
This patch (of 4):
When CONFIG_DEBUG_PAGEALLOC is enabled, it unmaps pages from the kernel
direct mapping after free_pages(). The pages than need to be mapped back
before they could be used. Theese mapping operations use
__kernel_map_pages() guarded with with debug_pagealloc_enabled().
The only place that calls __kernel_map_pages() without checking whether
DEBUG_PAGEALLOC is enabled is the hibernation code that presumes
availability of this function when ARCH_HAS_SET_DIRECT_MAP is set. Still,
on arm64, __kernel_map_pages() will bail out when DEBUG_PAGEALLOC is not
enabled but set_direct_map_invalid_noflush() may render some pages not
present in the direct map and hibernation code won't be able to save such
pages.
To make page allocation debugging and hibernation interaction more robust,
the dependency on DEBUG_PAGEALLOC or ARCH_HAS_SET_DIRECT_MAP has to be
made more explicit.
Start with combining the guard condition and the call to
__kernel_map_pages() into debug_pagealloc_map_pages() and
debug_pagealloc_unmap_pages() functions to emphasize that
__kernel_map_pages() should not be called without DEBUG_PAGEALLOC and use
these new functions to map/unmap pages when page allocation debugging is
enabled.
Link: https://lkml.kernel.org/r/20201109192128.960-1-rppt@kernel.org
Link: https://lkml.kernel.org/r/20201109192128.960-2-rppt@kernel.org
Signed-off-by: Mike Rapoport <rppt@linux.ibm.com>
Reviewed-by: David Hildenbrand <david@redhat.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Albert Ou <aou@eecs.berkeley.edu>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Christian Borntraeger <borntraeger@de.ibm.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: "Edgecombe, Rick P" <rick.p.edgecombe@intel.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Len Brown <len.brown@intel.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Palmer Dabbelt <palmer@dabbelt.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Paul Walmsley <paul.walmsley@sifive.com>
Cc: Pavel Machek <pavel@ucw.cz>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: "Rafael J. Wysocki" <rjw@rjwysocki.net>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Cc: Will Deacon <will@kernel.org>
Cc: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-14 19:10:20 -08:00
|
|
|
|
2019-04-25 17:11:35 -07:00
|
|
|
|
2016-05-19 17:14:38 -07:00
|
|
|
|
|
|
|
|
|
2016-05-19 17:14:32 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
page-allocator: split per-cpu list into one-list-per-migrate-type
The following two patches remove searching in the page allocator fast-path
by maintaining multiple free-lists in the per-cpu structure. At the time
the search was introduced, increasing the per-cpu structures would waste a
lot of memory as per-cpu structures were statically allocated at
compile-time. This is no longer the case.
The patches are as follows. They are based on mmotm-2009-08-27.
Patch 1 adds multiple lists to struct per_cpu_pages, one per
migratetype that can be stored on the PCP lists.
Patch 2 notes that the pcpu drain path check empty lists multiple times. The
patch reduces the number of checks by maintaining a count of free
lists encountered. Lists containing pages will then free multiple
pages in batch
The patches were tested with kernbench, netperf udp/tcp, hackbench and
sysbench. The netperf tests were not bound to any CPU in particular and
were run such that the results should be 99% confidence that the reported
results are within 1% of the estimated mean. sysbench was run with a
postgres background and read-only tests. Similar to netperf, it was run
multiple times so that it's 99% confidence results are within 1%. The
patches were tested on x86, x86-64 and ppc64 as
x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine)
kernbench - No significant difference, variance well within noise
netperf-udp - 1.34% to 2.28% gain
netperf-tcp - 0.45% to 1.22% gain
hackbench - Small variances, very close to noise
sysbench - Very small gains
x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine)
kernbench - No significant difference, variance well within noise
netperf-udp - 1.83% to 10.42% gains
netperf-tcp - No conclusive until buffer >= PAGE_SIZE
4096 +15.83%
8192 + 0.34% (not significant)
16384 + 1%
hackbench - Small gains, very close to noise
sysbench - 0.79% to 1.6% gain
ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation)
kernbench - No significant difference, variance well within noise
netperf-udp - 2-3% gain for almost all buffer sizes tested
netperf-tcp - losses on small buffers, gains on larger buffers
possibly indicates some bad caching effect.
hackbench - No significant difference
sysbench - 2-4% gain
This patch:
Currently the per-cpu page allocator searches the PCP list for pages of
the correct migrate-type to reduce the possibility of pages being
inappropriate placed from a fragmentation perspective. This search is
potentially expensive in a fast-path and undesirable. Splitting the
per-cpu list into multiple lists increases the size of a per-cpu structure
and this was potentially a major problem at the time the search was
introduced. These problem has been mitigated as now only the necessary
number of structures is allocated for the running system.
This patch replaces a list search in the per-cpu allocator with one list
per migrate type. The potential snag with this approach is when bulk
freeing pages. We round-robin free pages based on migrate type which has
little bearing on the cache hotness of the page and potentially checks
empty lists repeatedly in the event the majority of PCP pages are of one
type.
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Acked-by: Nick Piggin <npiggin@suse.de>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-21 17:03:19 -07:00
|
|
|
|
2021-11-05 13:40:08 -07:00
|
|
|
|
2005-09-10 00:26:59 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
page-allocator: split per-cpu list into one-list-per-migrate-type
The following two patches remove searching in the page allocator fast-path
by maintaining multiple free-lists in the per-cpu structure. At the time
the search was introduced, increasing the per-cpu structures would waste a
lot of memory as per-cpu structures were statically allocated at
compile-time. This is no longer the case.
The patches are as follows. They are based on mmotm-2009-08-27.
Patch 1 adds multiple lists to struct per_cpu_pages, one per
migratetype that can be stored on the PCP lists.
Patch 2 notes that the pcpu drain path check empty lists multiple times. The
patch reduces the number of checks by maintaining a count of free
lists encountered. Lists containing pages will then free multiple
pages in batch
The patches were tested with kernbench, netperf udp/tcp, hackbench and
sysbench. The netperf tests were not bound to any CPU in particular and
were run such that the results should be 99% confidence that the reported
results are within 1% of the estimated mean. sysbench was run with a
postgres background and read-only tests. Similar to netperf, it was run
multiple times so that it's 99% confidence results are within 1%. The
patches were tested on x86, x86-64 and ppc64 as
x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine)
kernbench - No significant difference, variance well within noise
netperf-udp - 1.34% to 2.28% gain
netperf-tcp - 0.45% to 1.22% gain
hackbench - Small variances, very close to noise
sysbench - Very small gains
x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine)
kernbench - No significant difference, variance well within noise
netperf-udp - 1.83% to 10.42% gains
netperf-tcp - No conclusive until buffer >= PAGE_SIZE
4096 +15.83%
8192 + 0.34% (not significant)
16384 + 1%
hackbench - Small gains, very close to noise
sysbench - 0.79% to 1.6% gain
ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation)
kernbench - No significant difference, variance well within noise
netperf-udp - 2-3% gain for almost all buffer sizes tested
netperf-tcp - losses on small buffers, gains on larger buffers
possibly indicates some bad caching effect.
hackbench - No significant difference
sysbench - 2-4% gain
This patch:
Currently the per-cpu page allocator searches the PCP list for pages of
the correct migrate-type to reduce the possibility of pages being
inappropriate placed from a fragmentation perspective. This search is
potentially expensive in a fast-path and undesirable. Splitting the
per-cpu list into multiple lists increases the size of a per-cpu structure
and this was potentially a major problem at the time the search was
introduced. These problem has been mitigated as now only the necessary
number of structures is allocated for the running system.
This patch replaces a list search in the per-cpu allocator with one list
per migrate type. The potential snag with this approach is when bulk
freeing pages. We round-robin free pages based on migrate type which has
little bearing on the cache hotness of the page and potentially checks
empty lists repeatedly in the event the majority of PCP pages are of one
type.
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Acked-by: Nick Piggin <npiggin@suse.de>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-21 17:03:19 -07:00
|
|
|
|
2022-03-22 14:43:36 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2022-11-18 10:17:14 +00:00
|
|
|
|
2021-06-28 19:43:08 -07:00
|
|
|
|
2022-03-22 14:43:42 -07:00
|
|
|
|
2009-06-16 15:32:13 -07:00
|
|
|
|
mm, page_alloc: fix core hung in free_pcppages_bulk()
The following race is observed with the repeated online, offline and a
delay between two successive online of memory blocks of movable zone.
P1 P2
Online the first memory block in
the movable zone. The pcp struct
values are initialized to default
values,i.e., pcp->high = 0 &
pcp->batch = 1.
Allocate the pages from the
movable zone.
Try to Online the second memory
block in the movable zone thus it
entered the online_pages() but yet
to call zone_pcp_update().
This process is entered into
the exit path thus it tries
to release the order-0 pages
to pcp lists through
free_unref_page_commit().
As pcp->high = 0, pcp->count = 1
proceed to call the function
free_pcppages_bulk().
Update the pcp values thus the
new pcp values are like, say,
pcp->high = 378, pcp->batch = 63.
Read the pcp's batch value using
READ_ONCE() and pass the same to
free_pcppages_bulk(), pcp values
passed here are, batch = 63,
count = 1.
Since num of pages in the pcp
lists are less than ->batch,
then it will stuck in
while(list_empty(list)) loop
with interrupts disabled thus
a core hung.
Avoid this by ensuring free_pcppages_bulk() is called with proper count of
pcp list pages.
The mentioned race is some what easily reproducible without [1] because
pcp's are not updated for the first memory block online and thus there is
a enough race window for P2 between alloc+free and pcp struct values
update through onlining of second memory block.
With [1], the race still exists but it is very narrow as we update the pcp
struct values for the first memory block online itself.
This is not limited to the movable zone, it could also happen in cases
with the normal zone (e.g., hotplug to a node that only has DMA memory, or
no other memory yet).
[1]: https://patchwork.kernel.org/patch/11696389/
Fixes: 5f8dcc21211a ("page-allocator: split per-cpu list into one-list-per-migrate-type")
Signed-off-by: Charan Teja Reddy <charante@codeaurora.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Acked-by: David Hildenbrand <david@redhat.com>
Acked-by: David Rientjes <rientjes@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Vinayak Menon <vinmenon@codeaurora.org>
Cc: <stable@vger.kernel.org> [2.6+]
Link: http://lkml.kernel.org/r/1597150703-19003-1-git-send-email-charante@codeaurora.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-20 17:42:27 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-03-22 14:43:38 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-11-18 10:17:14 +00:00
|
|
|
|
2022-03-22 14:43:42 -07:00
|
|
|
|
2021-06-28 19:43:08 -07:00
|
|
|
|
page-allocator: split per-cpu list into one-list-per-migrate-type
The following two patches remove searching in the page allocator fast-path
by maintaining multiple free-lists in the per-cpu structure. At the time
the search was introduced, increasing the per-cpu structures would waste a
lot of memory as per-cpu structures were statically allocated at
compile-time. This is no longer the case.
The patches are as follows. They are based on mmotm-2009-08-27.
Patch 1 adds multiple lists to struct per_cpu_pages, one per
migratetype that can be stored on the PCP lists.
Patch 2 notes that the pcpu drain path check empty lists multiple times. The
patch reduces the number of checks by maintaining a count of free
lists encountered. Lists containing pages will then free multiple
pages in batch
The patches were tested with kernbench, netperf udp/tcp, hackbench and
sysbench. The netperf tests were not bound to any CPU in particular and
were run such that the results should be 99% confidence that the reported
results are within 1% of the estimated mean. sysbench was run with a
postgres background and read-only tests. Similar to netperf, it was run
multiple times so that it's 99% confidence results are within 1%. The
patches were tested on x86, x86-64 and ppc64 as
x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine)
kernbench - No significant difference, variance well within noise
netperf-udp - 1.34% to 2.28% gain
netperf-tcp - 0.45% to 1.22% gain
hackbench - Small variances, very close to noise
sysbench - Very small gains
x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine)
kernbench - No significant difference, variance well within noise
netperf-udp - 1.83% to 10.42% gains
netperf-tcp - No conclusive until buffer >= PAGE_SIZE
4096 +15.83%
8192 + 0.34% (not significant)
16384 + 1%
hackbench - Small gains, very close to noise
sysbench - 0.79% to 1.6% gain
ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation)
kernbench - No significant difference, variance well within noise
netperf-udp - 2-3% gain for almost all buffer sizes tested
netperf-tcp - losses on small buffers, gains on larger buffers
possibly indicates some bad caching effect.
hackbench - No significant difference
sysbench - 2-4% gain
This patch:
Currently the per-cpu page allocator searches the PCP list for pages of
the correct migrate-type to reduce the possibility of pages being
inappropriate placed from a fragmentation perspective. This search is
potentially expensive in a fast-path and undesirable. Splitting the
per-cpu list into multiple lists increases the size of a per-cpu structure
and this was potentially a major problem at the time the search was
introduced. These problem has been mitigated as now only the necessary
number of structures is allocated for the running system.
This patch replaces a list search in the per-cpu allocator with one list
per migrate type. The potential snag with this approach is when bulk
freeing pages. We round-robin free pages based on migrate type which has
little bearing on the cache hotness of the page and potentially checks
empty lists repeatedly in the event the majority of PCP pages are of one
type.
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Acked-by: Nick Piggin <npiggin@suse.de>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-21 17:03:19 -07:00
|
|
|
|
2022-03-22 14:43:36 -07:00
|
|
|
|
page-allocator: split per-cpu list into one-list-per-migrate-type
The following two patches remove searching in the page allocator fast-path
by maintaining multiple free-lists in the per-cpu structure. At the time
the search was introduced, increasing the per-cpu structures would waste a
lot of memory as per-cpu structures were statically allocated at
compile-time. This is no longer the case.
The patches are as follows. They are based on mmotm-2009-08-27.
Patch 1 adds multiple lists to struct per_cpu_pages, one per
migratetype that can be stored on the PCP lists.
Patch 2 notes that the pcpu drain path check empty lists multiple times. The
patch reduces the number of checks by maintaining a count of free
lists encountered. Lists containing pages will then free multiple
pages in batch
The patches were tested with kernbench, netperf udp/tcp, hackbench and
sysbench. The netperf tests were not bound to any CPU in particular and
were run such that the results should be 99% confidence that the reported
results are within 1% of the estimated mean. sysbench was run with a
postgres background and read-only tests. Similar to netperf, it was run
multiple times so that it's 99% confidence results are within 1%. The
patches were tested on x86, x86-64 and ppc64 as
x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine)
kernbench - No significant difference, variance well within noise
netperf-udp - 1.34% to 2.28% gain
netperf-tcp - 0.45% to 1.22% gain
hackbench - Small variances, very close to noise
sysbench - Very small gains
x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine)
kernbench - No significant difference, variance well within noise
netperf-udp - 1.83% to 10.42% gains
netperf-tcp - No conclusive until buffer >= PAGE_SIZE
4096 +15.83%
8192 + 0.34% (not significant)
16384 + 1%
hackbench - Small gains, very close to noise
sysbench - 0.79% to 1.6% gain
ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation)
kernbench - No significant difference, variance well within noise
netperf-udp - 2-3% gain for almost all buffer sizes tested
netperf-tcp - losses on small buffers, gains on larger buffers
possibly indicates some bad caching effect.
hackbench - No significant difference
sysbench - 2-4% gain
This patch:
Currently the per-cpu page allocator searches the PCP list for pages of
the correct migrate-type to reduce the possibility of pages being
inappropriate placed from a fragmentation perspective. This search is
potentially expensive in a fast-path and undesirable. Splitting the
per-cpu list into multiple lists increases the size of a per-cpu structure
and this was potentially a major problem at the time the search was
introduced. These problem has been mitigated as now only the necessary
number of structures is allocated for the running system.
This patch replaces a list search in the per-cpu allocator with one list
per migrate type. The potential snag with this approach is when bulk
freeing pages. We round-robin free pages based on migrate type which has
little bearing on the cache hotness of the page and potentially checks
empty lists repeatedly in the event the majority of PCP pages are of one
type.
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Acked-by: Nick Piggin <npiggin@suse.de>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-21 17:03:19 -07:00
|
|
|
|
2022-03-22 14:43:36 -07:00
|
|
|
|
page-allocator: split per-cpu list into one-list-per-migrate-type
The following two patches remove searching in the page allocator fast-path
by maintaining multiple free-lists in the per-cpu structure. At the time
the search was introduced, increasing the per-cpu structures would waste a
lot of memory as per-cpu structures were statically allocated at
compile-time. This is no longer the case.
The patches are as follows. They are based on mmotm-2009-08-27.
Patch 1 adds multiple lists to struct per_cpu_pages, one per
migratetype that can be stored on the PCP lists.
Patch 2 notes that the pcpu drain path check empty lists multiple times. The
patch reduces the number of checks by maintaining a count of free
lists encountered. Lists containing pages will then free multiple
pages in batch
The patches were tested with kernbench, netperf udp/tcp, hackbench and
sysbench. The netperf tests were not bound to any CPU in particular and
were run such that the results should be 99% confidence that the reported
results are within 1% of the estimated mean. sysbench was run with a
postgres background and read-only tests. Similar to netperf, it was run
multiple times so that it's 99% confidence results are within 1%. The
patches were tested on x86, x86-64 and ppc64 as
x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine)
kernbench - No significant difference, variance well within noise
netperf-udp - 1.34% to 2.28% gain
netperf-tcp - 0.45% to 1.22% gain
hackbench - Small variances, very close to noise
sysbench - Very small gains
x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine)
kernbench - No significant difference, variance well within noise
netperf-udp - 1.83% to 10.42% gains
netperf-tcp - No conclusive until buffer >= PAGE_SIZE
4096 +15.83%
8192 + 0.34% (not significant)
16384 + 1%
hackbench - Small gains, very close to noise
sysbench - 0.79% to 1.6% gain
ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation)
kernbench - No significant difference, variance well within noise
netperf-udp - 2-3% gain for almost all buffer sizes tested
netperf-tcp - losses on small buffers, gains on larger buffers
possibly indicates some bad caching effect.
hackbench - No significant difference
sysbench - 2-4% gain
This patch:
Currently the per-cpu page allocator searches the PCP list for pages of
the correct migrate-type to reduce the possibility of pages being
inappropriate placed from a fragmentation perspective. This search is
potentially expensive in a fast-path and undesirable. Splitting the
per-cpu list into multiple lists increases the size of a per-cpu structure
and this was potentially a major problem at the time the search was
introduced. These problem has been mitigated as now only the necessary
number of structures is allocated for the running system.
This patch replaces a list search in the per-cpu allocator with one list
per migrate type. The potential snag with this approach is when bulk
freeing pages. We round-robin free pages based on migrate type which has
little bearing on the cache hotness of the page and potentially checks
empty lists repeatedly in the event the majority of PCP pages are of one
type.
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Acked-by: Nick Piggin <npiggin@suse.de>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-21 17:03:19 -07:00
|
|
|
|
2023-08-09 18:07:53 +08:00
|
|
|
|
|
|
|
|
|
2021-06-28 19:43:08 -07:00
|
|
|
|
2023-08-09 18:07:53 +08:00
|
|
|
|
2006-01-08 01:00:42 -08:00
|
|
|
|
2021-06-28 19:43:08 -07:00
|
|
|
|
2022-03-22 14:43:36 -07:00
|
|
|
|
2009-09-21 17:03:20 -07:00
|
|
|
|
mm: page_alloc: remove pcppage migratetype caching
Patch series "mm: page_alloc: freelist migratetype hygiene", v4.
The page allocator's mobility grouping is intended to keep unmovable pages
separate from reclaimable/compactable ones to allow on-demand
defragmentation for higher-order allocations and huge pages.
Currently, there are several places where accidental type mixing occurs:
an allocation asks for a page of a certain migratetype and receives
another. This ruins pageblocks for compaction, which in turn makes
allocating huge pages more expensive and less reliable.
The series addresses those causes. The last patch adds type checks on all
freelist movements to prevent new violations being introduced.
The benefits can be seen in a mixed workload that stresses the machine
with a memcache-type workload and a kernel build job while periodically
attempting to allocate batches of THP. The following data is aggregated
over 50 consecutive defconfig builds:
VANILLA PATCHED
Hugealloc Time mean 165843.93 ( +0.00%) 113025.88 ( -31.85%)
Hugealloc Time stddev 158957.35 ( +0.00%) 114716.07 ( -27.83%)
Kbuild Real time 310.24 ( +0.00%) 300.73 ( -3.06%)
Kbuild User time 1271.13 ( +0.00%) 1259.42 ( -0.92%)
Kbuild System time 582.02 ( +0.00%) 559.79 ( -3.81%)
THP fault alloc 30585.14 ( +0.00%) 40853.62 ( +33.57%)
THP fault fallback 36626.46 ( +0.00%) 26357.62 ( -28.04%)
THP fault fail rate % 54.49 ( +0.00%) 39.22 ( -27.53%)
Pagealloc fallback 1328.00 ( +0.00%) 1.00 ( -99.85%)
Pagealloc type mismatch 181009.50 ( +0.00%) 0.00 ( -100.00%)
Direct compact stall 434.56 ( +0.00%) 257.66 ( -40.61%)
Direct compact fail 421.70 ( +0.00%) 249.94 ( -40.63%)
Direct compact success 12.86 ( +0.00%) 7.72 ( -37.09%)
Direct compact success rate % 2.86 ( +0.00%) 2.82 ( -0.96%)
Compact daemon scanned migrate 3370059.62 ( +0.00%) 3612054.76 ( +7.18%)
Compact daemon scanned free 7718439.20 ( +0.00%) 5386385.02 ( -30.21%)
Compact direct scanned migrate 309248.62 ( +0.00%) 176721.04 ( -42.85%)
Compact direct scanned free 433582.84 ( +0.00%) 315727.66 ( -27.18%)
Compact migrate scanned daemon % 91.20 ( +0.00%) 94.48 ( +3.56%)
Compact free scanned daemon % 94.58 ( +0.00%) 94.42 ( -0.16%)
Compact total migrate scanned 3679308.24 ( +0.00%) 3788775.80 ( +2.98%)
Compact total free scanned 8152022.04 ( +0.00%) 5702112.68 ( -30.05%)
Alloc stall 872.04 ( +0.00%) 5156.12 ( +490.71%)
Pages kswapd scanned 510645.86 ( +0.00%) 3394.94 ( -99.33%)
Pages kswapd reclaimed 134811.62 ( +0.00%) 2701.26 ( -98.00%)
Pages direct scanned 99546.06 ( +0.00%) 376407.52 ( +278.12%)
Pages direct reclaimed 62123.40 ( +0.00%) 289535.70 ( +366.06%)
Pages total scanned 610191.92 ( +0.00%) 379802.46 ( -37.76%)
Pages scanned kswapd % 76.36 ( +0.00%) 0.10 ( -98.58%)
Swap out 12057.54 ( +0.00%) 15022.98 ( +24.59%)
Swap in 209.16 ( +0.00%) 256.48 ( +22.52%)
File refaults 17701.64 ( +0.00%) 11765.40 ( -33.53%)
Huge page success rate is higher, allocation latencies are shorter and
more predictable.
Stealing (fallback) rate is drastically reduced. Notably, while the
vanilla kernel keeps doing fallbacks on an ongoing basis, the patched
kernel enters a steady state once the distribution of block types is
adequate for the workload. Steals over 50 runs:
VANILLA PATCHED
1504.0 227.0
1557.0 6.0
1391.0 13.0
1080.0 26.0
1057.0 40.0
1156.0 6.0
805.0 46.0
736.0 20.0
1747.0 2.0
1699.0 34.0
1269.0 13.0
1858.0 12.0
907.0 4.0
727.0 2.0
563.0 2.0
3094.0 2.0
10211.0 3.0
2621.0 1.0
5508.0 2.0
1060.0 2.0
538.0 3.0
5773.0 2.0
2199.0 0.0
3781.0 2.0
1387.0 1.0
4977.0 0.0
2865.0 1.0
1814.0 1.0
3739.0 1.0
6857.0 0.0
382.0 0.0
407.0 1.0
3784.0 0.0
297.0 0.0
298.0 0.0
6636.0 0.0
4188.0 0.0
242.0 0.0
9960.0 0.0
5816.0 0.0
354.0 0.0
287.0 0.0
261.0 0.0
140.0 1.0
2065.0 0.0
312.0 0.0
331.0 0.0
164.0 0.0
465.0 1.0
219.0 0.0
Type mismatches are down too. Those count every time an allocation
request asks for one migratetype and gets another. This can still occur
minimally in the patched kernel due to non-stealing fallbacks, but it's
quite rare and follows the pattern of overall fallbacks - once the block
type distribution settles, mismatches cease as well:
VANILLA: PATCHED:
182602.0 268.0
135794.0 20.0
88619.0 19.0
95973.0 0.0
129590.0 0.0
129298.0 0.0
147134.0 0.0
230854.0 0.0
239709.0 0.0
137670.0 0.0
132430.0 0.0
65712.0 0.0
57901.0 0.0
67506.0 0.0
63565.0 4.0
34806.0 0.0
42962.0 0.0
32406.0 0.0
38668.0 0.0
61356.0 0.0
57800.0 0.0
41435.0 0.0
83456.0 0.0
65048.0 0.0
28955.0 0.0
47597.0 0.0
75117.0 0.0
55564.0 0.0
38280.0 0.0
52404.0 0.0
26264.0 0.0
37538.0 0.0
19671.0 0.0
30936.0 0.0
26933.0 0.0
16962.0 0.0
44554.0 0.0
46352.0 0.0
24995.0 0.0
35152.0 0.0
12823.0 0.0
21583.0 0.0
18129.0 0.0
31693.0 0.0
28745.0 0.0
33308.0 0.0
31114.0 0.0
35034.0 0.0
12111.0 0.0
24885.0 0.0
Compaction work is markedly reduced despite much better THP rates.
In the vanilla kernel, reclaim seems to have been driven primarily by
watermark boosting that happens as a result of fallbacks. With those all
but eliminated, watermarks average lower and kswapd does less work. The
uptick in direct reclaim is because THP requests have to fend for
themselves more often - which is intended policy right now. Aggregate
reclaim activity is lowered significantly, though.
This patch (of 10):
The idea behind the cache is to save get_pageblock_migratetype() lookups
during bulk freeing. A microbenchmark suggests this isn't helping,
though. The pcp migratetype can get stale, which means that bulk freeing
has an extra branch to check if the pageblock was isolated while on the
pcp.
While the variance overlaps, the cache write and the branch seem to make
this a net negative. The following test allocates and frees batches of
10,000 pages (~3x the pcp high marks to trigger flushing):
Before:
8,668.48 msec task-clock # 99.735 CPUs utilized ( +- 2.90% )
19 context-switches # 4.341 /sec ( +- 3.24% )
0 cpu-migrations # 0.000 /sec
17,440 page-faults # 3.984 K/sec ( +- 2.90% )
41,758,692,473 cycles # 9.541 GHz ( +- 2.90% )
126,201,294,231 instructions # 5.98 insn per cycle ( +- 2.90% )
25,348,098,335 branches # 5.791 G/sec ( +- 2.90% )
33,436,921 branch-misses # 0.26% of all branches ( +- 2.90% )
0.0869148 +- 0.0000302 seconds time elapsed ( +- 0.03% )
After:
8,444.81 msec task-clock # 99.726 CPUs utilized ( +- 2.90% )
22 context-switches # 5.160 /sec ( +- 3.23% )
0 cpu-migrations # 0.000 /sec
17,443 page-faults # 4.091 K/sec ( +- 2.90% )
40,616,738,355 cycles # 9.527 GHz ( +- 2.90% )
126,383,351,792 instructions # 6.16 insn per cycle ( +- 2.90% )
25,224,985,153 branches # 5.917 G/sec ( +- 2.90% )
32,236,793 branch-misses # 0.25% of all branches ( +- 2.90% )
0.0846799 +- 0.0000412 seconds time elapsed ( +- 0.05% )
A side effect is that this also ensures that pages whose pageblock gets
stolen while on the pcplist end up on the right freelist and we don't
perform potentially type-incompatible buddy merges (or skip merges when we
shouldn't), which is likely beneficial to long-term fragmentation
management, although the effects would be harder to measure. Settle for
simpler and faster code as justification here.
Link: https://lkml.kernel.org/r/20240320180429.678181-1-hannes@cmpxchg.org
Link: https://lkml.kernel.org/r/20240320180429.678181-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Tested-by: "Huang, Ying" <ying.huang@intel.com>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-20 14:02:06 -04:00
|
|
|
|
2022-03-22 14:43:42 -07:00
|
|
|
|
|
|
|
|
|
2022-06-24 13:54:17 +01:00
|
|
|
|
mm: page_alloc: remove pcppage migratetype caching
Patch series "mm: page_alloc: freelist migratetype hygiene", v4.
The page allocator's mobility grouping is intended to keep unmovable pages
separate from reclaimable/compactable ones to allow on-demand
defragmentation for higher-order allocations and huge pages.
Currently, there are several places where accidental type mixing occurs:
an allocation asks for a page of a certain migratetype and receives
another. This ruins pageblocks for compaction, which in turn makes
allocating huge pages more expensive and less reliable.
The series addresses those causes. The last patch adds type checks on all
freelist movements to prevent new violations being introduced.
The benefits can be seen in a mixed workload that stresses the machine
with a memcache-type workload and a kernel build job while periodically
attempting to allocate batches of THP. The following data is aggregated
over 50 consecutive defconfig builds:
VANILLA PATCHED
Hugealloc Time mean 165843.93 ( +0.00%) 113025.88 ( -31.85%)
Hugealloc Time stddev 158957.35 ( +0.00%) 114716.07 ( -27.83%)
Kbuild Real time 310.24 ( +0.00%) 300.73 ( -3.06%)
Kbuild User time 1271.13 ( +0.00%) 1259.42 ( -0.92%)
Kbuild System time 582.02 ( +0.00%) 559.79 ( -3.81%)
THP fault alloc 30585.14 ( +0.00%) 40853.62 ( +33.57%)
THP fault fallback 36626.46 ( +0.00%) 26357.62 ( -28.04%)
THP fault fail rate % 54.49 ( +0.00%) 39.22 ( -27.53%)
Pagealloc fallback 1328.00 ( +0.00%) 1.00 ( -99.85%)
Pagealloc type mismatch 181009.50 ( +0.00%) 0.00 ( -100.00%)
Direct compact stall 434.56 ( +0.00%) 257.66 ( -40.61%)
Direct compact fail 421.70 ( +0.00%) 249.94 ( -40.63%)
Direct compact success 12.86 ( +0.00%) 7.72 ( -37.09%)
Direct compact success rate % 2.86 ( +0.00%) 2.82 ( -0.96%)
Compact daemon scanned migrate 3370059.62 ( +0.00%) 3612054.76 ( +7.18%)
Compact daemon scanned free 7718439.20 ( +0.00%) 5386385.02 ( -30.21%)
Compact direct scanned migrate 309248.62 ( +0.00%) 176721.04 ( -42.85%)
Compact direct scanned free 433582.84 ( +0.00%) 315727.66 ( -27.18%)
Compact migrate scanned daemon % 91.20 ( +0.00%) 94.48 ( +3.56%)
Compact free scanned daemon % 94.58 ( +0.00%) 94.42 ( -0.16%)
Compact total migrate scanned 3679308.24 ( +0.00%) 3788775.80 ( +2.98%)
Compact total free scanned 8152022.04 ( +0.00%) 5702112.68 ( -30.05%)
Alloc stall 872.04 ( +0.00%) 5156.12 ( +490.71%)
Pages kswapd scanned 510645.86 ( +0.00%) 3394.94 ( -99.33%)
Pages kswapd reclaimed 134811.62 ( +0.00%) 2701.26 ( -98.00%)
Pages direct scanned 99546.06 ( +0.00%) 376407.52 ( +278.12%)
Pages direct reclaimed 62123.40 ( +0.00%) 289535.70 ( +366.06%)
Pages total scanned 610191.92 ( +0.00%) 379802.46 ( -37.76%)
Pages scanned kswapd % 76.36 ( +0.00%) 0.10 ( -98.58%)
Swap out 12057.54 ( +0.00%) 15022.98 ( +24.59%)
Swap in 209.16 ( +0.00%) 256.48 ( +22.52%)
File refaults 17701.64 ( +0.00%) 11765.40 ( -33.53%)
Huge page success rate is higher, allocation latencies are shorter and
more predictable.
Stealing (fallback) rate is drastically reduced. Notably, while the
vanilla kernel keeps doing fallbacks on an ongoing basis, the patched
kernel enters a steady state once the distribution of block types is
adequate for the workload. Steals over 50 runs:
VANILLA PATCHED
1504.0 227.0
1557.0 6.0
1391.0 13.0
1080.0 26.0
1057.0 40.0
1156.0 6.0
805.0 46.0
736.0 20.0
1747.0 2.0
1699.0 34.0
1269.0 13.0
1858.0 12.0
907.0 4.0
727.0 2.0
563.0 2.0
3094.0 2.0
10211.0 3.0
2621.0 1.0
5508.0 2.0
1060.0 2.0
538.0 3.0
5773.0 2.0
2199.0 0.0
3781.0 2.0
1387.0 1.0
4977.0 0.0
2865.0 1.0
1814.0 1.0
3739.0 1.0
6857.0 0.0
382.0 0.0
407.0 1.0
3784.0 0.0
297.0 0.0
298.0 0.0
6636.0 0.0
4188.0 0.0
242.0 0.0
9960.0 0.0
5816.0 0.0
354.0 0.0
287.0 0.0
261.0 0.0
140.0 1.0
2065.0 0.0
312.0 0.0
331.0 0.0
164.0 0.0
465.0 1.0
219.0 0.0
Type mismatches are down too. Those count every time an allocation
request asks for one migratetype and gets another. This can still occur
minimally in the patched kernel due to non-stealing fallbacks, but it's
quite rare and follows the pattern of overall fallbacks - once the block
type distribution settles, mismatches cease as well:
VANILLA: PATCHED:
182602.0 268.0
135794.0 20.0
88619.0 19.0
95973.0 0.0
129590.0 0.0
129298.0 0.0
147134.0 0.0
230854.0 0.0
239709.0 0.0
137670.0 0.0
132430.0 0.0
65712.0 0.0
57901.0 0.0
67506.0 0.0
63565.0 4.0
34806.0 0.0
42962.0 0.0
32406.0 0.0
38668.0 0.0
61356.0 0.0
57800.0 0.0
41435.0 0.0
83456.0 0.0
65048.0 0.0
28955.0 0.0
47597.0 0.0
75117.0 0.0
55564.0 0.0
38280.0 0.0
52404.0 0.0
26264.0 0.0
37538.0 0.0
19671.0 0.0
30936.0 0.0
26933.0 0.0
16962.0 0.0
44554.0 0.0
46352.0 0.0
24995.0 0.0
35152.0 0.0
12823.0 0.0
21583.0 0.0
18129.0 0.0
31693.0 0.0
28745.0 0.0
33308.0 0.0
31114.0 0.0
35034.0 0.0
12111.0 0.0
24885.0 0.0
Compaction work is markedly reduced despite much better THP rates.
In the vanilla kernel, reclaim seems to have been driven primarily by
watermark boosting that happens as a result of fallbacks. With those all
but eliminated, watermarks average lower and kswapd does less work. The
uptick in direct reclaim is because THP requests have to fend for
themselves more often - which is intended policy right now. Aggregate
reclaim activity is lowered significantly, though.
This patch (of 10):
The idea behind the cache is to save get_pageblock_migratetype() lookups
during bulk freeing. A microbenchmark suggests this isn't helping,
though. The pcp migratetype can get stale, which means that bulk freeing
has an extra branch to check if the pageblock was isolated while on the
pcp.
While the variance overlaps, the cache write and the branch seem to make
this a net negative. The following test allocates and frees batches of
10,000 pages (~3x the pcp high marks to trigger flushing):
Before:
8,668.48 msec task-clock # 99.735 CPUs utilized ( +- 2.90% )
19 context-switches # 4.341 /sec ( +- 3.24% )
0 cpu-migrations # 0.000 /sec
17,440 page-faults # 3.984 K/sec ( +- 2.90% )
41,758,692,473 cycles # 9.541 GHz ( +- 2.90% )
126,201,294,231 instructions # 5.98 insn per cycle ( +- 2.90% )
25,348,098,335 branches # 5.791 G/sec ( +- 2.90% )
33,436,921 branch-misses # 0.26% of all branches ( +- 2.90% )
0.0869148 +- 0.0000302 seconds time elapsed ( +- 0.03% )
After:
8,444.81 msec task-clock # 99.726 CPUs utilized ( +- 2.90% )
22 context-switches # 5.160 /sec ( +- 3.23% )
0 cpu-migrations # 0.000 /sec
17,443 page-faults # 4.091 K/sec ( +- 2.90% )
40,616,738,355 cycles # 9.527 GHz ( +- 2.90% )
126,383,351,792 instructions # 6.16 insn per cycle ( +- 2.90% )
25,224,985,153 branches # 5.917 G/sec ( +- 2.90% )
32,236,793 branch-misses # 0.25% of all branches ( +- 2.90% )
0.0846799 +- 0.0000412 seconds time elapsed ( +- 0.05% )
A side effect is that this also ensures that pages whose pageblock gets
stolen while on the pcplist end up on the right freelist and we don't
perform potentially type-incompatible buddy merges (or skip merges when we
shouldn't), which is likely beneficial to long-term fragmentation
management, although the effects would be harder to measure. Settle for
simpler and faster code as justification here.
Link: https://lkml.kernel.org/r/20240320180429.678181-1-hannes@cmpxchg.org
Link: https://lkml.kernel.org/r/20240320180429.678181-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Tested-by: "Huang, Ying" <ying.huang@intel.com>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-20 14:02:06 -04:00
|
|
|
|
|
|
|
|
|
2022-03-22 14:43:42 -07:00
|
|
|
|
2018-04-05 16:24:10 -07:00
|
|
|
|
2022-06-24 13:54:17 +01:00
|
|
|
|
2022-03-22 14:43:36 -07:00
|
|
|
|
|
|
|
|
|
mm, page_isolation: remove bogus tests for isolated pages
The __test_page_isolated_in_pageblock() is used to verify whether all
pages in pageblock were either successfully isolated, or are hwpoisoned.
Two of the possible state of pages, that are tested, are however bogus
and misleading.
Both tests rely on get_freepage_migratetype(page), which however has no
guarantees about pages on freelists. Specifically, it doesn't guarantee
that the migratetype returned by the function actually matches the
migratetype of the freelist that the page is on. Such guarantee is not
its purpose and would have negative impact on allocator performance.
The first test checks whether the freepage_migratetype equals
MIGRATE_ISOLATE, supposedly to catch races between page isolation and
allocator activity. These races should be fixed nowadays with
51bb1a4093 ("mm/page_alloc: add freepage on isolate pageblock to correct
buddy list") and related patches. As explained above, the check
wouldn't be able to catch them reliably anyway. For the same reason
false positives can happen, although they are harmless, as the
move_freepages() call would just move the page to the same freelist it's
already on. So removing the test is not a bug fix, just cleanup. After
this patch, we assume that all PageBuddy pages are on the correct
freelist and that the races were really fixed. A truly reliable
verification in the form of e.g. VM_BUG_ON() would be complicated and
is arguably not needed.
The second test (page_count(page) == 0 && get_freepage_migratetype(page)
== MIGRATE_ISOLATE) is probably supposed (the code comes from a big
memory isolation patch from 2007) to catch pages on MIGRATE_ISOLATE
pcplists. However, pcplists don't contain MIGRATE_ISOLATE freepages
nowadays, those are freed directly to free lists, so the check is
obsolete. Remove it as well.
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Minchan Kim <minchan@kernel.org>
Acked-by: Michal Nazarewicz <mina86@mina86.com>
Cc: Laura Abbott <lauraa@codeaurora.org>
Reviewed-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Seungho Park <seungho1.park@lge.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-08 15:01:22 -07:00
|
|
|
|
mm: page_alloc: remove pcppage migratetype caching
Patch series "mm: page_alloc: freelist migratetype hygiene", v4.
The page allocator's mobility grouping is intended to keep unmovable pages
separate from reclaimable/compactable ones to allow on-demand
defragmentation for higher-order allocations and huge pages.
Currently, there are several places where accidental type mixing occurs:
an allocation asks for a page of a certain migratetype and receives
another. This ruins pageblocks for compaction, which in turn makes
allocating huge pages more expensive and less reliable.
The series addresses those causes. The last patch adds type checks on all
freelist movements to prevent new violations being introduced.
The benefits can be seen in a mixed workload that stresses the machine
with a memcache-type workload and a kernel build job while periodically
attempting to allocate batches of THP. The following data is aggregated
over 50 consecutive defconfig builds:
VANILLA PATCHED
Hugealloc Time mean 165843.93 ( +0.00%) 113025.88 ( -31.85%)
Hugealloc Time stddev 158957.35 ( +0.00%) 114716.07 ( -27.83%)
Kbuild Real time 310.24 ( +0.00%) 300.73 ( -3.06%)
Kbuild User time 1271.13 ( +0.00%) 1259.42 ( -0.92%)
Kbuild System time 582.02 ( +0.00%) 559.79 ( -3.81%)
THP fault alloc 30585.14 ( +0.00%) 40853.62 ( +33.57%)
THP fault fallback 36626.46 ( +0.00%) 26357.62 ( -28.04%)
THP fault fail rate % 54.49 ( +0.00%) 39.22 ( -27.53%)
Pagealloc fallback 1328.00 ( +0.00%) 1.00 ( -99.85%)
Pagealloc type mismatch 181009.50 ( +0.00%) 0.00 ( -100.00%)
Direct compact stall 434.56 ( +0.00%) 257.66 ( -40.61%)
Direct compact fail 421.70 ( +0.00%) 249.94 ( -40.63%)
Direct compact success 12.86 ( +0.00%) 7.72 ( -37.09%)
Direct compact success rate % 2.86 ( +0.00%) 2.82 ( -0.96%)
Compact daemon scanned migrate 3370059.62 ( +0.00%) 3612054.76 ( +7.18%)
Compact daemon scanned free 7718439.20 ( +0.00%) 5386385.02 ( -30.21%)
Compact direct scanned migrate 309248.62 ( +0.00%) 176721.04 ( -42.85%)
Compact direct scanned free 433582.84 ( +0.00%) 315727.66 ( -27.18%)
Compact migrate scanned daemon % 91.20 ( +0.00%) 94.48 ( +3.56%)
Compact free scanned daemon % 94.58 ( +0.00%) 94.42 ( -0.16%)
Compact total migrate scanned 3679308.24 ( +0.00%) 3788775.80 ( +2.98%)
Compact total free scanned 8152022.04 ( +0.00%) 5702112.68 ( -30.05%)
Alloc stall 872.04 ( +0.00%) 5156.12 ( +490.71%)
Pages kswapd scanned 510645.86 ( +0.00%) 3394.94 ( -99.33%)
Pages kswapd reclaimed 134811.62 ( +0.00%) 2701.26 ( -98.00%)
Pages direct scanned 99546.06 ( +0.00%) 376407.52 ( +278.12%)
Pages direct reclaimed 62123.40 ( +0.00%) 289535.70 ( +366.06%)
Pages total scanned 610191.92 ( +0.00%) 379802.46 ( -37.76%)
Pages scanned kswapd % 76.36 ( +0.00%) 0.10 ( -98.58%)
Swap out 12057.54 ( +0.00%) 15022.98 ( +24.59%)
Swap in 209.16 ( +0.00%) 256.48 ( +22.52%)
File refaults 17701.64 ( +0.00%) 11765.40 ( -33.53%)
Huge page success rate is higher, allocation latencies are shorter and
more predictable.
Stealing (fallback) rate is drastically reduced. Notably, while the
vanilla kernel keeps doing fallbacks on an ongoing basis, the patched
kernel enters a steady state once the distribution of block types is
adequate for the workload. Steals over 50 runs:
VANILLA PATCHED
1504.0 227.0
1557.0 6.0
1391.0 13.0
1080.0 26.0
1057.0 40.0
1156.0 6.0
805.0 46.0
736.0 20.0
1747.0 2.0
1699.0 34.0
1269.0 13.0
1858.0 12.0
907.0 4.0
727.0 2.0
563.0 2.0
3094.0 2.0
10211.0 3.0
2621.0 1.0
5508.0 2.0
1060.0 2.0
538.0 3.0
5773.0 2.0
2199.0 0.0
3781.0 2.0
1387.0 1.0
4977.0 0.0
2865.0 1.0
1814.0 1.0
3739.0 1.0
6857.0 0.0
382.0 0.0
407.0 1.0
3784.0 0.0
297.0 0.0
298.0 0.0
6636.0 0.0
4188.0 0.0
242.0 0.0
9960.0 0.0
5816.0 0.0
354.0 0.0
287.0 0.0
261.0 0.0
140.0 1.0
2065.0 0.0
312.0 0.0
331.0 0.0
164.0 0.0
465.0 1.0
219.0 0.0
Type mismatches are down too. Those count every time an allocation
request asks for one migratetype and gets another. This can still occur
minimally in the patched kernel due to non-stealing fallbacks, but it's
quite rare and follows the pattern of overall fallbacks - once the block
type distribution settles, mismatches cease as well:
VANILLA: PATCHED:
182602.0 268.0
135794.0 20.0
88619.0 19.0
95973.0 0.0
129590.0 0.0
129298.0 0.0
147134.0 0.0
230854.0 0.0
239709.0 0.0
137670.0 0.0
132430.0 0.0
65712.0 0.0
57901.0 0.0
67506.0 0.0
63565.0 4.0
34806.0 0.0
42962.0 0.0
32406.0 0.0
38668.0 0.0
61356.0 0.0
57800.0 0.0
41435.0 0.0
83456.0 0.0
65048.0 0.0
28955.0 0.0
47597.0 0.0
75117.0 0.0
55564.0 0.0
38280.0 0.0
52404.0 0.0
26264.0 0.0
37538.0 0.0
19671.0 0.0
30936.0 0.0
26933.0 0.0
16962.0 0.0
44554.0 0.0
46352.0 0.0
24995.0 0.0
35152.0 0.0
12823.0 0.0
21583.0 0.0
18129.0 0.0
31693.0 0.0
28745.0 0.0
33308.0 0.0
31114.0 0.0
35034.0 0.0
12111.0 0.0
24885.0 0.0
Compaction work is markedly reduced despite much better THP rates.
In the vanilla kernel, reclaim seems to have been driven primarily by
watermark boosting that happens as a result of fallbacks. With those all
but eliminated, watermarks average lower and kswapd does less work. The
uptick in direct reclaim is because THP requests have to fend for
themselves more often - which is intended policy right now. Aggregate
reclaim activity is lowered significantly, though.
This patch (of 10):
The idea behind the cache is to save get_pageblock_migratetype() lookups
during bulk freeing. A microbenchmark suggests this isn't helping,
though. The pcp migratetype can get stale, which means that bulk freeing
has an extra branch to check if the pageblock was isolated while on the
pcp.
While the variance overlaps, the cache write and the branch seem to make
this a net negative. The following test allocates and frees batches of
10,000 pages (~3x the pcp high marks to trigger flushing):
Before:
8,668.48 msec task-clock # 99.735 CPUs utilized ( +- 2.90% )
19 context-switches # 4.341 /sec ( +- 3.24% )
0 cpu-migrations # 0.000 /sec
17,440 page-faults # 3.984 K/sec ( +- 2.90% )
41,758,692,473 cycles # 9.541 GHz ( +- 2.90% )
126,201,294,231 instructions # 5.98 insn per cycle ( +- 2.90% )
25,348,098,335 branches # 5.791 G/sec ( +- 2.90% )
33,436,921 branch-misses # 0.26% of all branches ( +- 2.90% )
0.0869148 +- 0.0000302 seconds time elapsed ( +- 0.03% )
After:
8,444.81 msec task-clock # 99.726 CPUs utilized ( +- 2.90% )
22 context-switches # 5.160 /sec ( +- 3.23% )
0 cpu-migrations # 0.000 /sec
17,443 page-faults # 4.091 K/sec ( +- 2.90% )
40,616,738,355 cycles # 9.527 GHz ( +- 2.90% )
126,383,351,792 instructions # 6.16 insn per cycle ( +- 2.90% )
25,224,985,153 branches # 5.917 G/sec ( +- 2.90% )
32,236,793 branch-misses # 0.25% of all branches ( +- 2.90% )
0.0846799 +- 0.0000412 seconds time elapsed ( +- 0.05% )
A side effect is that this also ensures that pages whose pageblock gets
stolen while on the pcplist end up on the right freelist and we don't
perform potentially type-incompatible buddy merges (or skip merges when we
shouldn't), which is likely beneficial to long-term fragmentation
management, although the effects would be harder to measure. Settle for
simpler and faster code as justification here.
Link: https://lkml.kernel.org/r/20240320180429.678181-1-hannes@cmpxchg.org
Link: https://lkml.kernel.org/r/20240320180429.678181-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Tested-by: "Huang, Ying" <ying.huang@intel.com>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-20 14:02:06 -04:00
|
|
|
|
2022-03-22 14:43:42 -07:00
|
|
|
|
2022-03-22 14:43:36 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2018-04-05 16:24:10 -07:00
|
|
|
|
2022-11-18 10:17:14 +00:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2024-08-13 21:54:49 -06:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-12-10 10:34:37 +01:00
|
|
|
|
2024-08-13 21:54:49 -06:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-12-10 10:34:37 +01:00
|
|
|
|
|
|
|
|
|
2024-08-13 21:54:49 -06:00
|
|
|
|
2024-12-10 10:34:37 +01:00
|
|
|
|
2024-08-13 21:54:49 -06:00
|
|
|
|
|
|
|
|
|
2025-02-21 18:44:24 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-09-10 15:29:19 +01:00
|
|
|
|
2025-02-21 18:44:24 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-20 14:02:12 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-01-08 01:00:42 -08:00
|
|
|
|
2025-02-21 18:44:24 -08:00
|
|
|
|
2021-06-28 19:42:00 -07:00
|
|
|
|
|
|
|
|
|
2025-03-30 17:28:09 -07:00
|
|
|
|
|
|
|
|
|
2025-02-21 18:44:24 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-03-30 17:28:09 -07:00
|
|
|
|
2025-02-21 18:44:24 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-09-10 15:29:19 +01:00
|
|
|
|
2025-02-21 18:44:24 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-08-13 21:54:49 -06:00
|
|
|
|
2021-06-28 19:42:00 -07:00
|
|
|
|
2024-09-04 20:54:19 +00:00
|
|
|
|
|
|
|
|
|
2006-01-08 01:00:42 -08:00
|
|
|
|
|
|
|
|
|
2020-10-15 20:09:35 -07:00
|
|
|
|
|
|
|
|
|
2010-05-24 14:32:38 -07:00
|
|
|
|
2014-06-04 16:10:17 -07:00
|
|
|
|
2021-06-28 19:41:57 -07:00
|
|
|
|
2010-05-24 14:32:38 -07:00
|
|
|
|
2024-09-04 20:54:19 +00:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2024-07-06 06:16:15 +00:00
|
|
|
|
mm: pass meminit_context to __free_pages_core()
Patch series "mm/memory_hotplug: use PageOffline() instead of
PageReserved() for !ZONE_DEVICE".
This can be a considered a long-overdue follow-up to some parts of [1].
The patches are based on [2], but they are not strictly required -- just
makes it clearer why we can use adjust_managed_page_count() for memory
hotplug without going into details about highmem.
We stop initializing pages with PageReserved() in memory hotplug code --
except when dealing with ZONE_DEVICE for now. Instead, we use
PageOffline(): all pages are initialized to PageOffline() when onlining a
memory section, and only the ones actually getting exposed to the
system/page allocator will get PageOffline cleared.
This way, we enlighten memory hotplug more about PageOffline() pages and
can cleanup some hacks we have in virtio-mem code.
What about ZONE_DEVICE? PageOffline() is wrong, but we might just stop
using PageReserved() for them later by simply checking for
is_zone_device_page() at suitable places. That will be a separate patch
set / proposal.
This primarily affects virtio-mem, HV-balloon and XEN balloon. I only
briefly tested with virtio-mem, which benefits most from these cleanups.
[1] https://lore.kernel.org/all/20191024120938.11237-1-david@redhat.com/
[2] https://lkml.kernel.org/r/20240607083711.62833-1-david@redhat.com
This patch (of 3):
In preparation for further changes, let's teach __free_pages_core() about
the differences of memory hotplug handling.
Move the memory hotplug specific handling from generic_online_page() to
__free_pages_core(), use adjust_managed_page_count() on the memory hotplug
path, and spell out why memory freed via memblock cannot currently use
adjust_managed_page_count().
[david@redhat.com: add missed CONFIG_DEFERRED_STRUCT_PAGE_INIT]
Link: https://lkml.kernel.org/r/b72e6efd-fb0a-459c-b1a0-88a98e5b19e2@redhat.com
[david@redhat.com: fix up the memblock comment, per Oscar]
Link: https://lkml.kernel.org/r/2ed64218-7f3b-4302-a5dc-27f060654fe2@redhat.com
[david@redhat.com: add the parameter name also in the declaration]
Link: https://lkml.kernel.org/r/ca575956-f0dd-4fb9-a307-6b7621681ed9@redhat.com
Link: https://lkml.kernel.org/r/20240607090939.89524-1-david@redhat.com
Link: https://lkml.kernel.org/r/20240607090939.89524-2-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Dexuan Cui <decui@microsoft.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Eugenio Pérez <eperezma@redhat.com>
Cc: Haiyang Zhang <haiyangz@microsoft.com>
Cc: Jason Wang <jasowang@redhat.com>
Cc: Juergen Gross <jgross@suse.com>
Cc: "K. Y. Srinivasan" <kys@microsoft.com>
Cc: Marco Elver <elver@google.com>
Cc: Michael S. Tsirkin <mst@redhat.com>
Cc: Mike Rapoport (IBM) <rppt@kernel.org>
Cc: Oleksandr Tyshchenko <oleksandr_tyshchenko@epam.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Stefano Stabellini <sstabellini@kernel.org>
Cc: Wei Liu <wei.liu@kernel.org>
Cc: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-06-07 11:09:36 +02:00
|
|
|
|
2006-01-06 00:11:08 -08:00
|
|
|
|
2012-01-10 15:08:10 -08:00
|
|
|
|
2013-09-11 14:20:37 -07:00
|
|
|
|
2012-01-10 15:08:10 -08:00
|
|
|
|
2006-01-06 00:11:08 -08:00
|
|
|
|
2020-10-15 20:09:35 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-06-07 11:09:37 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-10-15 20:09:35 -07:00
|
|
|
|
mm: pass meminit_context to __free_pages_core()
Patch series "mm/memory_hotplug: use PageOffline() instead of
PageReserved() for !ZONE_DEVICE".
This can be a considered a long-overdue follow-up to some parts of [1].
The patches are based on [2], but they are not strictly required -- just
makes it clearer why we can use adjust_managed_page_count() for memory
hotplug without going into details about highmem.
We stop initializing pages with PageReserved() in memory hotplug code --
except when dealing with ZONE_DEVICE for now. Instead, we use
PageOffline(): all pages are initialized to PageOffline() when onlining a
memory section, and only the ones actually getting exposed to the
system/page allocator will get PageOffline cleared.
This way, we enlighten memory hotplug more about PageOffline() pages and
can cleanup some hacks we have in virtio-mem code.
What about ZONE_DEVICE? PageOffline() is wrong, but we might just stop
using PageReserved() for them later by simply checking for
is_zone_device_page() at suitable places. That will be a separate patch
set / proposal.
This primarily affects virtio-mem, HV-balloon and XEN balloon. I only
briefly tested with virtio-mem, which benefits most from these cleanups.
[1] https://lore.kernel.org/all/20191024120938.11237-1-david@redhat.com/
[2] https://lkml.kernel.org/r/20240607083711.62833-1-david@redhat.com
This patch (of 3):
In preparation for further changes, let's teach __free_pages_core() about
the differences of memory hotplug handling.
Move the memory hotplug specific handling from generic_online_page() to
__free_pages_core(), use adjust_managed_page_count() on the memory hotplug
path, and spell out why memory freed via memblock cannot currently use
adjust_managed_page_count().
[david@redhat.com: add missed CONFIG_DEFERRED_STRUCT_PAGE_INIT]
Link: https://lkml.kernel.org/r/b72e6efd-fb0a-459c-b1a0-88a98e5b19e2@redhat.com
[david@redhat.com: fix up the memblock comment, per Oscar]
Link: https://lkml.kernel.org/r/2ed64218-7f3b-4302-a5dc-27f060654fe2@redhat.com
[david@redhat.com: add the parameter name also in the declaration]
Link: https://lkml.kernel.org/r/ca575956-f0dd-4fb9-a307-6b7621681ed9@redhat.com
Link: https://lkml.kernel.org/r/20240607090939.89524-1-david@redhat.com
Link: https://lkml.kernel.org/r/20240607090939.89524-2-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Dexuan Cui <decui@microsoft.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Eugenio Pérez <eperezma@redhat.com>
Cc: Haiyang Zhang <haiyangz@microsoft.com>
Cc: Jason Wang <jasowang@redhat.com>
Cc: Juergen Gross <jgross@suse.com>
Cc: "K. Y. Srinivasan" <kys@microsoft.com>
Cc: Marco Elver <elver@google.com>
Cc: Michael S. Tsirkin <mst@redhat.com>
Cc: Mike Rapoport (IBM) <rppt@kernel.org>
Cc: Oleksandr Tyshchenko <oleksandr_tyshchenko@epam.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Stefano Stabellini <sstabellini@kernel.org>
Cc: Wei Liu <wei.liu@kernel.org>
Cc: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-06-07 11:09:36 +02:00
|
|
|
|
|
|
|
|
|
2024-07-02 02:09:31 +00:00
|
|
|
|
2024-06-07 11:09:37 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: pass meminit_context to __free_pages_core()
Patch series "mm/memory_hotplug: use PageOffline() instead of
PageReserved() for !ZONE_DEVICE".
This can be a considered a long-overdue follow-up to some parts of [1].
The patches are based on [2], but they are not strictly required -- just
makes it clearer why we can use adjust_managed_page_count() for memory
hotplug without going into details about highmem.
We stop initializing pages with PageReserved() in memory hotplug code --
except when dealing with ZONE_DEVICE for now. Instead, we use
PageOffline(): all pages are initialized to PageOffline() when onlining a
memory section, and only the ones actually getting exposed to the
system/page allocator will get PageOffline cleared.
This way, we enlighten memory hotplug more about PageOffline() pages and
can cleanup some hacks we have in virtio-mem code.
What about ZONE_DEVICE? PageOffline() is wrong, but we might just stop
using PageReserved() for them later by simply checking for
is_zone_device_page() at suitable places. That will be a separate patch
set / proposal.
This primarily affects virtio-mem, HV-balloon and XEN balloon. I only
briefly tested with virtio-mem, which benefits most from these cleanups.
[1] https://lore.kernel.org/all/20191024120938.11237-1-david@redhat.com/
[2] https://lkml.kernel.org/r/20240607083711.62833-1-david@redhat.com
This patch (of 3):
In preparation for further changes, let's teach __free_pages_core() about
the differences of memory hotplug handling.
Move the memory hotplug specific handling from generic_online_page() to
__free_pages_core(), use adjust_managed_page_count() on the memory hotplug
path, and spell out why memory freed via memblock cannot currently use
adjust_managed_page_count().
[david@redhat.com: add missed CONFIG_DEFERRED_STRUCT_PAGE_INIT]
Link: https://lkml.kernel.org/r/b72e6efd-fb0a-459c-b1a0-88a98e5b19e2@redhat.com
[david@redhat.com: fix up the memblock comment, per Oscar]
Link: https://lkml.kernel.org/r/2ed64218-7f3b-4302-a5dc-27f060654fe2@redhat.com
[david@redhat.com: add the parameter name also in the declaration]
Link: https://lkml.kernel.org/r/ca575956-f0dd-4fb9-a307-6b7621681ed9@redhat.com
Link: https://lkml.kernel.org/r/20240607090939.89524-1-david@redhat.com
Link: https://lkml.kernel.org/r/20240607090939.89524-2-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Dexuan Cui <decui@microsoft.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Eugenio Pérez <eperezma@redhat.com>
Cc: Haiyang Zhang <haiyangz@microsoft.com>
Cc: Jason Wang <jasowang@redhat.com>
Cc: Juergen Gross <jgross@suse.com>
Cc: "K. Y. Srinivasan" <kys@microsoft.com>
Cc: Marco Elver <elver@google.com>
Cc: Michael S. Tsirkin <mst@redhat.com>
Cc: Mike Rapoport (IBM) <rppt@kernel.org>
Cc: Oleksandr Tyshchenko <oleksandr_tyshchenko@epam.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Stefano Stabellini <sstabellini@kernel.org>
Cc: Wei Liu <wei.liu@kernel.org>
Cc: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-06-07 11:09:36 +02:00
|
|
|
|
|
|
|
|
|
2024-07-02 02:09:31 +00:00
|
|
|
|
2024-06-07 11:09:37 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: pass meminit_context to __free_pages_core()
Patch series "mm/memory_hotplug: use PageOffline() instead of
PageReserved() for !ZONE_DEVICE".
This can be a considered a long-overdue follow-up to some parts of [1].
The patches are based on [2], but they are not strictly required -- just
makes it clearer why we can use adjust_managed_page_count() for memory
hotplug without going into details about highmem.
We stop initializing pages with PageReserved() in memory hotplug code --
except when dealing with ZONE_DEVICE for now. Instead, we use
PageOffline(): all pages are initialized to PageOffline() when onlining a
memory section, and only the ones actually getting exposed to the
system/page allocator will get PageOffline cleared.
This way, we enlighten memory hotplug more about PageOffline() pages and
can cleanup some hacks we have in virtio-mem code.
What about ZONE_DEVICE? PageOffline() is wrong, but we might just stop
using PageReserved() for them later by simply checking for
is_zone_device_page() at suitable places. That will be a separate patch
set / proposal.
This primarily affects virtio-mem, HV-balloon and XEN balloon. I only
briefly tested with virtio-mem, which benefits most from these cleanups.
[1] https://lore.kernel.org/all/20191024120938.11237-1-david@redhat.com/
[2] https://lkml.kernel.org/r/20240607083711.62833-1-david@redhat.com
This patch (of 3):
In preparation for further changes, let's teach __free_pages_core() about
the differences of memory hotplug handling.
Move the memory hotplug specific handling from generic_online_page() to
__free_pages_core(), use adjust_managed_page_count() on the memory hotplug
path, and spell out why memory freed via memblock cannot currently use
adjust_managed_page_count().
[david@redhat.com: add missed CONFIG_DEFERRED_STRUCT_PAGE_INIT]
Link: https://lkml.kernel.org/r/b72e6efd-fb0a-459c-b1a0-88a98e5b19e2@redhat.com
[david@redhat.com: fix up the memblock comment, per Oscar]
Link: https://lkml.kernel.org/r/2ed64218-7f3b-4302-a5dc-27f060654fe2@redhat.com
[david@redhat.com: add the parameter name also in the declaration]
Link: https://lkml.kernel.org/r/ca575956-f0dd-4fb9-a307-6b7621681ed9@redhat.com
Link: https://lkml.kernel.org/r/20240607090939.89524-1-david@redhat.com
Link: https://lkml.kernel.org/r/20240607090939.89524-2-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Dexuan Cui <decui@microsoft.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Eugenio Pérez <eperezma@redhat.com>
Cc: Haiyang Zhang <haiyangz@microsoft.com>
Cc: Jason Wang <jasowang@redhat.com>
Cc: Juergen Gross <jgross@suse.com>
Cc: "K. Y. Srinivasan" <kys@microsoft.com>
Cc: Marco Elver <elver@google.com>
Cc: Michael S. Tsirkin <mst@redhat.com>
Cc: Mike Rapoport (IBM) <rppt@kernel.org>
Cc: Oleksandr Tyshchenko <oleksandr_tyshchenko@epam.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Stefano Stabellini <sstabellini@kernel.org>
Cc: Wei Liu <wei.liu@kernel.org>
Cc: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-06-07 11:09:36 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-10-15 20:09:35 -07:00
|
|
|
|
2023-06-06 17:26:29 +03:00
|
|
|
|
2023-12-28 17:47:04 +03:00
|
|
|
|
2023-06-06 17:26:29 +03:00
|
|
|
|
|
|
|
|
|
2024-08-09 14:48:51 +03:00
|
|
|
|
2023-06-06 17:26:29 +03:00
|
|
|
|
|
|
|
|
|
2020-10-15 20:09:35 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-03-09 20:29:14 -08:00
|
|
|
|
2006-01-06 00:11:08 -08:00
|
|
|
|
|
|
|
|
|
2016-03-15 14:57:51 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-09-07 19:54:52 -07:00
|
|
|
|
2016-03-15 14:57:51 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/page_alloc: add some comments to explain the possible hole in __pageblock_pfn_to_page()
Now the __pageblock_pfn_to_page() is used by set_zone_contiguous(), which
checks whether the given zone contains holes, and uses
pfn_to_online_page() to validate if the start pfn is online and valid, as
well as using pfn_valid() to validate the end pfn.
However, the __pageblock_pfn_to_page() function may return non-NULL even
if the end pfn of a pageblock is in a memory hole in some situations. For
example, if the pageblock order is MAX_ORDER, which will fall into 2
sub-sections, and the end pfn of the pageblock may be hole even though the
start pfn is online and valid.
See below memory layout as an example and suppose the pageblock order is
MAX_ORDER.
[ 0.000000] Zone ranges:
[ 0.000000] DMA [mem 0x0000000040000000-0x00000000ffffffff]
[ 0.000000] DMA32 empty
[ 0.000000] Normal [mem 0x0000000100000000-0x0000001fa7ffffff]
[ 0.000000] Movable zone start for each node
[ 0.000000] Early memory node ranges
[ 0.000000] node 0: [mem 0x0000000040000000-0x0000001fa3c7ffff]
[ 0.000000] node 0: [mem 0x0000001fa3c80000-0x0000001fa3ffffff]
[ 0.000000] node 0: [mem 0x0000001fa4000000-0x0000001fa402ffff]
[ 0.000000] node 0: [mem 0x0000001fa4030000-0x0000001fa40effff]
[ 0.000000] node 0: [mem 0x0000001fa40f0000-0x0000001fa73cffff]
[ 0.000000] node 0: [mem 0x0000001fa73d0000-0x0000001fa745ffff]
[ 0.000000] node 0: [mem 0x0000001fa7460000-0x0000001fa746ffff]
[ 0.000000] node 0: [mem 0x0000001fa7470000-0x0000001fa758ffff]
[ 0.000000] node 0: [mem 0x0000001fa7590000-0x0000001fa7dfffff]
Focus on the last memory range, and there is a hole for the range [mem
0x0000001fa7590000-0x0000001fa7dfffff]. That means the last pageblock
will contain the range from 0x1fa7c00000 to 0x1fa7ffffff, since the
pageblock must be 4M aligned. And in this pageblock, these pfns will fall
into 2 sub-section (the sub-section size is 2M aligned).
So, the 1st sub-section (indicates pfn range: 0x1fa7c00000 - 0x1fa7dfffff
) in this pageblock is valid by calling subsection_map_init() in
free_area_init(), but the 2nd sub-section (indicates pfn range:
0x1fa7e00000 - 0x1fa7ffffff ) in this pageblock is not valid.
This did not break anything until now, but the zone continuous is fragile
in this possible scenario. So as previous discussion[1], it is better to
add some comments to explain this possible issue in case there are some
future pfn walkers that rely on this.
[1] https://lore.kernel.org/all/87r0sdsmr6.fsf@yhuang6-desk2.ccr.corp.intel.com/
Link: https://lkml.kernel.org/r/5c26368865e79c743a453dea48d30670b19d2e4f.1682425534.git.baolin.wang@linux.alibaba.com
Link: https://lkml.kernel.org/r/5c26368865e79c743a453dea48d30670b19d2e4f.1682425534.git.baolin.wang@linux.alibaba.com
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Mike Rapoport (IBM) <rppt@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-04-25 20:44:53 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-12-28 17:47:04 +03:00
|
|
|
|
mm/page_alloc: add some comments to explain the possible hole in __pageblock_pfn_to_page()
Now the __pageblock_pfn_to_page() is used by set_zone_contiguous(), which
checks whether the given zone contains holes, and uses
pfn_to_online_page() to validate if the start pfn is online and valid, as
well as using pfn_valid() to validate the end pfn.
However, the __pageblock_pfn_to_page() function may return non-NULL even
if the end pfn of a pageblock is in a memory hole in some situations. For
example, if the pageblock order is MAX_ORDER, which will fall into 2
sub-sections, and the end pfn of the pageblock may be hole even though the
start pfn is online and valid.
See below memory layout as an example and suppose the pageblock order is
MAX_ORDER.
[ 0.000000] Zone ranges:
[ 0.000000] DMA [mem 0x0000000040000000-0x00000000ffffffff]
[ 0.000000] DMA32 empty
[ 0.000000] Normal [mem 0x0000000100000000-0x0000001fa7ffffff]
[ 0.000000] Movable zone start for each node
[ 0.000000] Early memory node ranges
[ 0.000000] node 0: [mem 0x0000000040000000-0x0000001fa3c7ffff]
[ 0.000000] node 0: [mem 0x0000001fa3c80000-0x0000001fa3ffffff]
[ 0.000000] node 0: [mem 0x0000001fa4000000-0x0000001fa402ffff]
[ 0.000000] node 0: [mem 0x0000001fa4030000-0x0000001fa40effff]
[ 0.000000] node 0: [mem 0x0000001fa40f0000-0x0000001fa73cffff]
[ 0.000000] node 0: [mem 0x0000001fa73d0000-0x0000001fa745ffff]
[ 0.000000] node 0: [mem 0x0000001fa7460000-0x0000001fa746ffff]
[ 0.000000] node 0: [mem 0x0000001fa7470000-0x0000001fa758ffff]
[ 0.000000] node 0: [mem 0x0000001fa7590000-0x0000001fa7dfffff]
Focus on the last memory range, and there is a hole for the range [mem
0x0000001fa7590000-0x0000001fa7dfffff]. That means the last pageblock
will contain the range from 0x1fa7c00000 to 0x1fa7ffffff, since the
pageblock must be 4M aligned. And in this pageblock, these pfns will fall
into 2 sub-section (the sub-section size is 2M aligned).
So, the 1st sub-section (indicates pfn range: 0x1fa7c00000 - 0x1fa7dfffff
) in this pageblock is valid by calling subsection_map_init() in
free_area_init(), but the 2nd sub-section (indicates pfn range:
0x1fa7e00000 - 0x1fa7ffffff ) in this pageblock is not valid.
This did not break anything until now, but the zone continuous is fragile
in this possible scenario. So as previous discussion[1], it is better to
add some comments to explain this possible issue in case there are some
future pfn walkers that rely on this.
[1] https://lore.kernel.org/all/87r0sdsmr6.fsf@yhuang6-desk2.ccr.corp.intel.com/
Link: https://lkml.kernel.org/r/5c26368865e79c743a453dea48d30670b19d2e4f.1682425534.git.baolin.wang@linux.alibaba.com
Link: https://lkml.kernel.org/r/5c26368865e79c743a453dea48d30670b19d2e4f.1682425534.git.baolin.wang@linux.alibaba.com
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Mike Rapoport (IBM) <rppt@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-04-25 20:44:53 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-03-15 14:57:51 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-04-24 21:45:39 +08:00
|
|
|
|
2016-03-15 14:57:51 -07:00
|
|
|
|
|
|
|
|
|
2017-07-06 15:37:56 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-03-15 14:57:51 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-12-06 10:39:54 +01:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2024-08-26 14:40:48 +08:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2024-08-26 14:40:48 +08:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-01-23 15:52:54 -08:00
|
|
|
|
2012-01-10 15:07:28 -08:00
|
|
|
|
2016-10-07 16:58:15 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
2012-01-10 15:07:28 -08:00
|
|
|
|
2016-10-07 16:58:15 -07:00
|
|
|
|
2024-03-27 15:01:11 -04:00
|
|
|
|
2020-10-15 20:10:15 -07:00
|
|
|
|
2024-03-27 15:01:11 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2024-08-26 14:40:48 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2016-05-19 17:14:41 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2025-03-20 14:33:46 +08:00
|
|
|
|
2016-05-20 16:58:50 -07:00
|
|
|
|
2024-05-29 13:19:02 +02:00
|
|
|
|
|
|
|
|
|
2016-05-20 16:58:50 -07:00
|
|
|
|
mm: check __PG_HWPOISON separately from PAGE_FLAGS_CHECK_AT_*
The race condition addressed in commit add05cecef80 ("mm: soft-offline:
don't free target page in successful page migration") was not closed
completely, because that can happen not only for soft-offline, but also
for hard-offline. Consider that a slab page is about to be freed into
buddy pool, and then an uncorrected memory error hits the page just
after entering __free_one_page(), then VM_BUG_ON_PAGE(page->flags &
PAGE_FLAGS_CHECK_AT_PREP) is triggered, despite the fact that it's not
necessary because the data on the affected page is not consumed.
To solve it, this patch drops __PG_HWPOISON from page flag checks at
allocation/free time. I think it's justified because __PG_HWPOISON
flags is defined to prevent the page from being reused, and setting it
outside the page's alloc-free cycle is a designed behavior (not a bug.)
For recent months, I was annoyed about BUG_ON when soft-offlined page
remains on lru cache list for a while, which is avoided by calling
put_page() instead of putback_lru_page() in page migration's success
path. This means that this patch reverts a major change from commit
add05cecef80 about the new refcounting rule of soft-offlined pages, so
"reuse window" revives. This will be closed by a subsequent patch.
Signed-off-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Dean Nelson <dnelson@redhat.com>
Cc: Tony Luck <tony.luck@intel.com>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Hugh Dickins <hughd@google.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-08-06 15:47:08 -07:00
|
|
|
|
2020-06-03 15:58:39 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-05-19 17:14:41 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-02-22 17:19:32 +08:00
|
|
|
|
2016-05-19 17:14:41 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-02-22 17:19:32 +08:00
|
|
|
|
2016-05-19 17:14:41 -07:00
|
|
|
|
|
|
|
|
|
2024-02-22 17:19:32 +08:00
|
|
|
|
2009-09-16 11:50:12 +02:00
|
|
|
|
|
|
|
|
|
mm, page_alloc: reduce page alloc/free sanity checks
Historically, we have performed sanity checks on all struct pages being
allocated or freed, making sure they have no unexpected page flags or
certain field values. This can detect insufficient cleanup and some cases
of use-after-free, although on its own it can't always identify the
culprit. The result is a warning and the "bad page" being leaked.
The checks do need some cpu cycles, so in 4.7 with commits 479f854a207c
("mm, page_alloc: defer debugging checks of pages allocated from the PCP")
and 4db7548ccbd9 ("mm, page_alloc: defer debugging checks of freed pages
until a PCP drain") they were no longer performed in the hot paths when
allocating and freeing from pcplists, but only when pcplists are bypassed,
refilled or drained. For debugging purposes, with CONFIG_DEBUG_VM enabled
the checks were instead still done in the hot paths and not when refilling
or draining pcplists.
With 4462b32c9285 ("mm, page_alloc: more extensive free page checking with
debug_pagealloc"), enabling debug_pagealloc also moved the sanity checks
back to hot pahs. When both debug_pagealloc and CONFIG_DEBUG_VM are
enabled, the checks are done both in hotpaths and pcplist refill/drain.
Even though the non-debug default today might seem to be a sensible
tradeoff between overhead and ability to detect bad pages, on closer look
it's arguably not. As most allocations go through the pcplists, catching
any bad pages when refilling or draining pcplists has only a small chance,
insufficient for debugging or serious hardening purposes. On the other
hand the cost of the checks is concentrated in the already expensive
drain/refill batching operations, and those are done under the often
contended zone lock. That was recently identified as an issue for page
allocation and the zone lock contention reduced by moving the checks
outside of the locked section with a patch "mm: reduce lock contention of
pcp buffer refill", but the cost of the checks is still visible compared
to their removal [1]. In the pcplist draining path free_pcppages_bulk()
the checks are still done under zone->lock.
Thus, remove the checks from pcplist refill and drain paths completely.
Introduce a static key check_pages_enabled to control checks during page
allocation a freeing (whether pcplist is used or bypassed). The static
key is enabled if either is true:
- kernel is built with CONFIG_DEBUG_VM=y (debugging)
- debug_pagealloc or page poisoning is boot-time enabled (debugging)
- init_on_alloc or init_on_free is boot-time enabled (hardening)
The resulting user visible changes:
- no checks when draining/refilling pcplists - less overhead, with
likely no practical reduction of ability to catch bad pages
- no checks when bypassing pcplists in default config (no
debugging/hardening) - less overhead etc. as above
- on typical hardened kernels [2], checks are now performed on each page
allocation/free (previously only when bypassing/draining/refilling
pcplists) - the init_on_alloc/init_on_free enabled should be sufficient
indication for preferring more costly alloc/free operations for
hardening purposes and we shouldn't need to introduce another toggle
- code (various wrappers) removal and simplification
[1] https://lore.kernel.org/all/68ba44d8-6899-c018-dcb3-36f3a96e6bea@sra.uni-hannover.de/
[2] https://lore.kernel.org/all/63ebc499.a70a0220.9ac51.29ea@mx.google.com/
[akpm@linux-foundation.org: coding-style cleanups]
[akpm@linux-foundation.org: make check_pages_enabled static]
Link: https://lkml.kernel.org/r/20230216095131.17336-1-vbabka@suse.cz
Reported-by: Alexander Halbuer <halbuer@sra.uni-hannover.de>
Reported-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Kees Cook <keescook@chromium.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-02-16 10:51:31 +01:00
|
|
|
|
2016-05-19 17:14:35 -07:00
|
|
|
|
2023-03-21 19:05:01 +02:00
|
|
|
|
mm, page_alloc: reduce page alloc/free sanity checks
Historically, we have performed sanity checks on all struct pages being
allocated or freed, making sure they have no unexpected page flags or
certain field values. This can detect insufficient cleanup and some cases
of use-after-free, although on its own it can't always identify the
culprit. The result is a warning and the "bad page" being leaked.
The checks do need some cpu cycles, so in 4.7 with commits 479f854a207c
("mm, page_alloc: defer debugging checks of pages allocated from the PCP")
and 4db7548ccbd9 ("mm, page_alloc: defer debugging checks of freed pages
until a PCP drain") they were no longer performed in the hot paths when
allocating and freeing from pcplists, but only when pcplists are bypassed,
refilled or drained. For debugging purposes, with CONFIG_DEBUG_VM enabled
the checks were instead still done in the hot paths and not when refilling
or draining pcplists.
With 4462b32c9285 ("mm, page_alloc: more extensive free page checking with
debug_pagealloc"), enabling debug_pagealloc also moved the sanity checks
back to hot pahs. When both debug_pagealloc and CONFIG_DEBUG_VM are
enabled, the checks are done both in hotpaths and pcplist refill/drain.
Even though the non-debug default today might seem to be a sensible
tradeoff between overhead and ability to detect bad pages, on closer look
it's arguably not. As most allocations go through the pcplists, catching
any bad pages when refilling or draining pcplists has only a small chance,
insufficient for debugging or serious hardening purposes. On the other
hand the cost of the checks is concentrated in the already expensive
drain/refill batching operations, and those are done under the often
contended zone lock. That was recently identified as an issue for page
allocation and the zone lock contention reduced by moving the checks
outside of the locked section with a patch "mm: reduce lock contention of
pcp buffer refill", but the cost of the checks is still visible compared
to their removal [1]. In the pcplist draining path free_pcppages_bulk()
the checks are still done under zone->lock.
Thus, remove the checks from pcplist refill and drain paths completely.
Introduce a static key check_pages_enabled to control checks during page
allocation a freeing (whether pcplist is used or bypassed). The static
key is enabled if either is true:
- kernel is built with CONFIG_DEBUG_VM=y (debugging)
- debug_pagealloc or page poisoning is boot-time enabled (debugging)
- init_on_alloc or init_on_free is boot-time enabled (hardening)
The resulting user visible changes:
- no checks when draining/refilling pcplists - less overhead, with
likely no practical reduction of ability to catch bad pages
- no checks when bypassing pcplists in default config (no
debugging/hardening) - less overhead etc. as above
- on typical hardened kernels [2], checks are now performed on each page
allocation/free (previously only when bypassing/draining/refilling
pcplists) - the init_on_alloc/init_on_free enabled should be sufficient
indication for preferring more costly alloc/free operations for
hardening purposes and we shouldn't need to introduce another toggle
- code (various wrappers) removal and simplification
[1] https://lore.kernel.org/all/68ba44d8-6899-c018-dcb3-36f3a96e6bea@sra.uni-hannover.de/
[2] https://lore.kernel.org/all/63ebc499.a70a0220.9ac51.29ea@mx.google.com/
[akpm@linux-foundation.org: coding-style cleanups]
[akpm@linux-foundation.org: make check_pages_enabled static]
Link: https://lkml.kernel.org/r/20230216095131.17336-1-vbabka@suse.cz
Reported-by: Alexander Halbuer <halbuer@sra.uni-hannover.de>
Reported-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Kees Cook <keescook@chromium.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-02-16 10:51:31 +01:00
|
|
|
|
|
|
|
|
|
2016-05-19 17:14:35 -07:00
|
|
|
|
2023-04-05 16:28:40 +02:00
|
|
|
|
mm, page_alloc: reduce page alloc/free sanity checks
Historically, we have performed sanity checks on all struct pages being
allocated or freed, making sure they have no unexpected page flags or
certain field values. This can detect insufficient cleanup and some cases
of use-after-free, although on its own it can't always identify the
culprit. The result is a warning and the "bad page" being leaked.
The checks do need some cpu cycles, so in 4.7 with commits 479f854a207c
("mm, page_alloc: defer debugging checks of pages allocated from the PCP")
and 4db7548ccbd9 ("mm, page_alloc: defer debugging checks of freed pages
until a PCP drain") they were no longer performed in the hot paths when
allocating and freeing from pcplists, but only when pcplists are bypassed,
refilled or drained. For debugging purposes, with CONFIG_DEBUG_VM enabled
the checks were instead still done in the hot paths and not when refilling
or draining pcplists.
With 4462b32c9285 ("mm, page_alloc: more extensive free page checking with
debug_pagealloc"), enabling debug_pagealloc also moved the sanity checks
back to hot pahs. When both debug_pagealloc and CONFIG_DEBUG_VM are
enabled, the checks are done both in hotpaths and pcplist refill/drain.
Even though the non-debug default today might seem to be a sensible
tradeoff between overhead and ability to detect bad pages, on closer look
it's arguably not. As most allocations go through the pcplists, catching
any bad pages when refilling or draining pcplists has only a small chance,
insufficient for debugging or serious hardening purposes. On the other
hand the cost of the checks is concentrated in the already expensive
drain/refill batching operations, and those are done under the often
contended zone lock. That was recently identified as an issue for page
allocation and the zone lock contention reduced by moving the checks
outside of the locked section with a patch "mm: reduce lock contention of
pcp buffer refill", but the cost of the checks is still visible compared
to their removal [1]. In the pcplist draining path free_pcppages_bulk()
the checks are still done under zone->lock.
Thus, remove the checks from pcplist refill and drain paths completely.
Introduce a static key check_pages_enabled to control checks during page
allocation a freeing (whether pcplist is used or bypassed). The static
key is enabled if either is true:
- kernel is built with CONFIG_DEBUG_VM=y (debugging)
- debug_pagealloc or page poisoning is boot-time enabled (debugging)
- init_on_alloc or init_on_free is boot-time enabled (hardening)
The resulting user visible changes:
- no checks when draining/refilling pcplists - less overhead, with
likely no practical reduction of ability to catch bad pages
- no checks when bypassing pcplists in default config (no
debugging/hardening) - less overhead etc. as above
- on typical hardened kernels [2], checks are now performed on each page
allocation/free (previously only when bypassing/draining/refilling
pcplists) - the init_on_alloc/init_on_free enabled should be sufficient
indication for preferring more costly alloc/free operations for
hardening purposes and we shouldn't need to introduce another toggle
- code (various wrappers) removal and simplification
[1] https://lore.kernel.org/all/68ba44d8-6899-c018-dcb3-36f3a96e6bea@sra.uni-hannover.de/
[2] https://lore.kernel.org/all/63ebc499.a70a0220.9ac51.29ea@mx.google.com/
[akpm@linux-foundation.org: coding-style cleanups]
[akpm@linux-foundation.org: make check_pages_enabled static]
Link: https://lkml.kernel.org/r/20230216095131.17336-1-vbabka@suse.cz
Reported-by: Alexander Halbuer <halbuer@sra.uni-hannover.de>
Reported-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Kees Cook <keescook@chromium.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-02-16 10:51:31 +01:00
|
|
|
|
|
|
|
|
|
2016-05-19 17:14:35 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-06-10 16:21:40 +01:00
|
|
|
|
2022-03-24 18:11:26 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-06-10 16:21:40 +01:00
|
|
|
|
2023-03-09 20:29:14 -08:00
|
|
|
|
2022-03-24 18:11:26 -07:00
|
|
|
|
2023-03-09 20:29:14 -08:00
|
|
|
|
2022-03-24 18:11:26 -07:00
|
|
|
|
|
|
|
|
|
2022-03-24 18:11:29 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-07-26 15:23:58 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-03-24 18:11:29 -07:00
|
|
|
|
|
|
|
|
|
kasan: allow sampling page_alloc allocations for HW_TAGS
As Hardware Tag-Based KASAN is intended to be used in production, its
performance impact is crucial. As page_alloc allocations tend to be big,
tagging and checking all such allocations can introduce a significant
slowdown.
Add two new boot parameters that allow to alleviate that slowdown:
- kasan.page_alloc.sample, which makes Hardware Tag-Based KASAN tag only
every Nth page_alloc allocation with the order configured by the second
added parameter (default: tag every such allocation).
- kasan.page_alloc.sample.order, which makes sampling enabled by the first
parameter only affect page_alloc allocations with the order equal or
greater than the specified value (default: 3, see below).
The exact performance improvement caused by using the new parameters
depends on their values and the applied workload.
The chosen default value for kasan.page_alloc.sample.order is 3, which
matches both PAGE_ALLOC_COSTLY_ORDER and SKB_FRAG_PAGE_ORDER. This is
done for two reasons:
1. PAGE_ALLOC_COSTLY_ORDER is "the order at which allocations are deemed
costly to service", which corresponds to the idea that only large and
thus costly allocations are supposed to sampled.
2. One of the workloads targeted by this patch is a benchmark that sends
a large amount of data over a local loopback connection. Most multi-page
data allocations in the networking subsystem have the order of
SKB_FRAG_PAGE_ORDER (or PAGE_ALLOC_COSTLY_ORDER).
When running a local loopback test on a testing MTE-enabled device in sync
mode, enabling Hardware Tag-Based KASAN introduces a ~50% slowdown.
Applying this patch and setting kasan.page_alloc.sampling to a value
higher than 1 allows to lower the slowdown. The performance improvement
saturates around the sampling interval value of 10 with the default
sampling page order of 3. This lowers the slowdown to ~20%. The slowdown
in real scenarios involving the network will likely be better.
Enabling page_alloc sampling has a downside: KASAN misses bad accesses to
a page_alloc allocation that has not been tagged. This lowers the value
of KASAN as a security mitigation.
However, based on measuring the number of page_alloc allocations of
different orders during boot in a test build, sampling with the default
kasan.page_alloc.sample.order value affects only ~7% of allocations. The
rest ~93% of allocations are still checked deterministically.
Link: https://lkml.kernel.org/r/129da0614123bb85ed4dd61ae30842b2dd7c903f.1671471846.git.andreyknvl@google.com
Signed-off-by: Andrey Konovalov <andreyknvl@google.com>
Reviewed-by: Marco Elver <elver@google.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Andrey Ryabinin <ryabinin.a.a@gmail.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Evgenii Stepanov <eugenis@google.com>
Cc: Jann Horn <jannh@google.com>
Cc: Mark Brand <markbrand@google.com>
Cc: Peter Collingbourne <pcc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-19 19:09:18 +01:00
|
|
|
|
2022-06-10 16:21:39 +01:00
|
|
|
|
2022-03-24 18:10:31 -07:00
|
|
|
|
2016-07-26 15:23:58 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: introduce debug_pagealloc_{map,unmap}_pages() helpers
Patch series "arch, mm: improve robustness of direct map manipulation", v7.
During recent discussion about KVM protected memory, David raised a
concern about usage of __kernel_map_pages() outside of DEBUG_PAGEALLOC
scope [1].
Indeed, for architectures that define CONFIG_ARCH_HAS_SET_DIRECT_MAP it is
possible that __kernel_map_pages() would fail, but since this function is
void, the failure will go unnoticed.
Moreover, there's lack of consistency of __kernel_map_pages() semantics
across architectures as some guard this function with #ifdef
DEBUG_PAGEALLOC, some refuse to update the direct map if page allocation
debugging is disabled at run time and some allow modifying the direct map
regardless of DEBUG_PAGEALLOC settings.
This set straightens this out by restoring dependency of
__kernel_map_pages() on DEBUG_PAGEALLOC and updating the call sites
accordingly.
Since currently the only user of __kernel_map_pages() outside
DEBUG_PAGEALLOC is hibernation, it is updated to make direct map accesses
there more explicit.
[1] https://lore.kernel.org/lkml/2759b4bf-e1e3-d006-7d86-78a40348269d@redhat.com
This patch (of 4):
When CONFIG_DEBUG_PAGEALLOC is enabled, it unmaps pages from the kernel
direct mapping after free_pages(). The pages than need to be mapped back
before they could be used. Theese mapping operations use
__kernel_map_pages() guarded with with debug_pagealloc_enabled().
The only place that calls __kernel_map_pages() without checking whether
DEBUG_PAGEALLOC is enabled is the hibernation code that presumes
availability of this function when ARCH_HAS_SET_DIRECT_MAP is set. Still,
on arm64, __kernel_map_pages() will bail out when DEBUG_PAGEALLOC is not
enabled but set_direct_map_invalid_noflush() may render some pages not
present in the direct map and hibernation code won't be able to save such
pages.
To make page allocation debugging and hibernation interaction more robust,
the dependency on DEBUG_PAGEALLOC or ARCH_HAS_SET_DIRECT_MAP has to be
made more explicit.
Start with combining the guard condition and the call to
__kernel_map_pages() into debug_pagealloc_map_pages() and
debug_pagealloc_unmap_pages() functions to emphasize that
__kernel_map_pages() should not be called without DEBUG_PAGEALLOC and use
these new functions to map/unmap pages when page allocation debugging is
enabled.
Link: https://lkml.kernel.org/r/20201109192128.960-1-rppt@kernel.org
Link: https://lkml.kernel.org/r/20201109192128.960-2-rppt@kernel.org
Signed-off-by: Mike Rapoport <rppt@linux.ibm.com>
Reviewed-by: David Hildenbrand <david@redhat.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Albert Ou <aou@eecs.berkeley.edu>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Christian Borntraeger <borntraeger@de.ibm.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: "Edgecombe, Rick P" <rick.p.edgecombe@intel.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Len Brown <len.brown@intel.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Palmer Dabbelt <palmer@dabbelt.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Paul Walmsley <paul.walmsley@sifive.com>
Cc: Pavel Machek <pavel@ucw.cz>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: "Rafael J. Wysocki" <rjw@rjwysocki.net>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Cc: Will Deacon <will@kernel.org>
Cc: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-14 19:10:20 -08:00
|
|
|
|
2020-12-14 19:11:15 -08:00
|
|
|
|
2021-04-29 23:00:02 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-03-24 18:10:31 -07:00
|
|
|
|
2021-04-29 23:00:02 -07:00
|
|
|
|
|
|
|
|
|
2022-03-24 18:10:34 -07:00
|
|
|
|
|
|
|
|
|
kasan: allow sampling page_alloc allocations for HW_TAGS
As Hardware Tag-Based KASAN is intended to be used in production, its
performance impact is crucial. As page_alloc allocations tend to be big,
tagging and checking all such allocations can introduce a significant
slowdown.
Add two new boot parameters that allow to alleviate that slowdown:
- kasan.page_alloc.sample, which makes Hardware Tag-Based KASAN tag only
every Nth page_alloc allocation with the order configured by the second
added parameter (default: tag every such allocation).
- kasan.page_alloc.sample.order, which makes sampling enabled by the first
parameter only affect page_alloc allocations with the order equal or
greater than the specified value (default: 3, see below).
The exact performance improvement caused by using the new parameters
depends on their values and the applied workload.
The chosen default value for kasan.page_alloc.sample.order is 3, which
matches both PAGE_ALLOC_COSTLY_ORDER and SKB_FRAG_PAGE_ORDER. This is
done for two reasons:
1. PAGE_ALLOC_COSTLY_ORDER is "the order at which allocations are deemed
costly to service", which corresponds to the idea that only large and
thus costly allocations are supposed to sampled.
2. One of the workloads targeted by this patch is a benchmark that sends
a large amount of data over a local loopback connection. Most multi-page
data allocations in the networking subsystem have the order of
SKB_FRAG_PAGE_ORDER (or PAGE_ALLOC_COSTLY_ORDER).
When running a local loopback test on a testing MTE-enabled device in sync
mode, enabling Hardware Tag-Based KASAN introduces a ~50% slowdown.
Applying this patch and setting kasan.page_alloc.sampling to a value
higher than 1 allows to lower the slowdown. The performance improvement
saturates around the sampling interval value of 10 with the default
sampling page order of 3. This lowers the slowdown to ~20%. The slowdown
in real scenarios involving the network will likely be better.
Enabling page_alloc sampling has a downside: KASAN misses bad accesses to
a page_alloc allocation that has not been tagged. This lowers the value
of KASAN as a security mitigation.
However, based on measuring the number of page_alloc allocations of
different orders during boot in a test build, sampling with the default
kasan.page_alloc.sample.order value affects only ~7% of allocations. The
rest ~93% of allocations are still checked deterministically.
Link: https://lkml.kernel.org/r/129da0614123bb85ed4dd61ae30842b2dd7c903f.1671471846.git.andreyknvl@google.com
Signed-off-by: Andrey Konovalov <andreyknvl@google.com>
Reviewed-by: Marco Elver <elver@google.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Andrey Ryabinin <ryabinin.a.a@gmail.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Evgenii Stepanov <eugenis@google.com>
Cc: Jann Horn <jannh@google.com>
Cc: Mark Brand <markbrand@google.com>
Cc: Peter Collingbourne <pcc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-19 19:09:18 +01:00
|
|
|
|
|
|
|
|
|
2022-03-24 18:10:34 -07:00
|
|
|
|
2025-11-18 08:21:27 -08:00
|
|
|
|
|
|
|
|
|
2022-03-24 18:10:34 -07:00
|
|
|
|
2023-03-09 20:29:14 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-06-10 16:21:39 +01:00
|
|
|
|
|
|
|
|
|
2021-06-02 16:52:28 -07:00
|
|
|
|
kasan: allow sampling page_alloc allocations for HW_TAGS
As Hardware Tag-Based KASAN is intended to be used in production, its
performance impact is crucial. As page_alloc allocations tend to be big,
tagging and checking all such allocations can introduce a significant
slowdown.
Add two new boot parameters that allow to alleviate that slowdown:
- kasan.page_alloc.sample, which makes Hardware Tag-Based KASAN tag only
every Nth page_alloc allocation with the order configured by the second
added parameter (default: tag every such allocation).
- kasan.page_alloc.sample.order, which makes sampling enabled by the first
parameter only affect page_alloc allocations with the order equal or
greater than the specified value (default: 3, see below).
The exact performance improvement caused by using the new parameters
depends on their values and the applied workload.
The chosen default value for kasan.page_alloc.sample.order is 3, which
matches both PAGE_ALLOC_COSTLY_ORDER and SKB_FRAG_PAGE_ORDER. This is
done for two reasons:
1. PAGE_ALLOC_COSTLY_ORDER is "the order at which allocations are deemed
costly to service", which corresponds to the idea that only large and
thus costly allocations are supposed to sampled.
2. One of the workloads targeted by this patch is a benchmark that sends
a large amount of data over a local loopback connection. Most multi-page
data allocations in the networking subsystem have the order of
SKB_FRAG_PAGE_ORDER (or PAGE_ALLOC_COSTLY_ORDER).
When running a local loopback test on a testing MTE-enabled device in sync
mode, enabling Hardware Tag-Based KASAN introduces a ~50% slowdown.
Applying this patch and setting kasan.page_alloc.sampling to a value
higher than 1 allows to lower the slowdown. The performance improvement
saturates around the sampling interval value of 10 with the default
sampling page order of 3. This lowers the slowdown to ~20%. The slowdown
in real scenarios involving the network will likely be better.
Enabling page_alloc sampling has a downside: KASAN misses bad accesses to
a page_alloc allocation that has not been tagged. This lowers the value
of KASAN as a security mitigation.
However, based on measuring the number of page_alloc allocations of
different orders during boot in a test build, sampling with the default
kasan.page_alloc.sample.order value affects only ~7% of allocations. The
rest ~93% of allocations are still checked deterministically.
Link: https://lkml.kernel.org/r/129da0614123bb85ed4dd61ae30842b2dd7c903f.1671471846.git.andreyknvl@google.com
Signed-off-by: Andrey Konovalov <andreyknvl@google.com>
Reviewed-by: Marco Elver <elver@google.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Andrey Ryabinin <ryabinin.a.a@gmail.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Evgenii Stepanov <eugenis@google.com>
Cc: Jann Horn <jannh@google.com>
Cc: Mark Brand <markbrand@google.com>
Cc: Peter Collingbourne <pcc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-19 19:09:18 +01:00
|
|
|
|
2022-03-24 18:10:40 -07:00
|
|
|
|
2022-06-09 20:18:45 +02:00
|
|
|
|
2021-04-29 23:00:02 -07:00
|
|
|
|
|
|
|
|
|
2022-01-14 14:06:37 -08:00
|
|
|
|
2024-03-21 09:36:36 -07:00
|
|
|
|
2016-07-26 15:23:58 -07:00
|
|
|
|
|
|
|
|
|
2016-05-19 17:14:35 -07:00
|
|
|
|
2016-05-19 17:13:38 -07:00
|
|
|
|
2009-09-16 11:50:12 +02:00
|
|
|
|
2016-07-26 15:23:58 -07:00
|
|
|
|
2006-03-22 00:08:41 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: set page->pfmemalloc in prep_new_page()
The possibility of replacing the numerous parameters of alloc_pages*
functions with a single structure has been discussed when Minchan proposed
to expand the x86 kernel stack [1]. This series implements the change,
along with few more cleanups/microoptimizations.
The series is based on next-20150108 and I used gcc 4.8.3 20140627 on
openSUSE 13.2 for compiling. Config includess NUMA and COMPACTION.
The core change is the introduction of a new struct alloc_context, which looks
like this:
struct alloc_context {
struct zonelist *zonelist;
nodemask_t *nodemask;
struct zone *preferred_zone;
int classzone_idx;
int migratetype;
enum zone_type high_zoneidx;
};
All the contents is mostly constant, except that __alloc_pages_slowpath()
changes preferred_zone, classzone_idx and potentially zonelist. But
that's not a problem in case control returns to retry_cpuset: in
__alloc_pages_nodemask(), those will be reset to initial values again
(although it's a bit subtle). On the other hand, gfp_flags and alloc_info
mutate so much that it doesn't make sense to put them into alloc_context.
Still, the result is one parameter instead of up to 7. This is all in
Patch 2.
Patch 3 is a step to expand alloc_context usage out of page_alloc.c
itself. The function try_to_compact_pages() can also much benefit from
the parameter reduction, but it means the struct definition has to be
moved to a shared header.
Patch 1 should IMHO be included even if the rest is deemed not useful
enough. It improves maintainability and also has some code/stack
reduction. Patch 4 is OTOH a tiny optimization.
Overall bloat-o-meter results:
add/remove: 0/0 grow/shrink: 0/4 up/down: 0/-460 (-460)
function old new delta
nr_free_zone_pages 129 115 -14
__alloc_pages_direct_compact 329 256 -73
get_page_from_freelist 2670 2576 -94
__alloc_pages_nodemask 2564 2285 -279
try_to_compact_pages 582 579 -3
Overall stack sizes per ./scripts/checkstack.pl:
old new delta
get_page_from_freelist: 184 184 0
__alloc_pages_nodemask 248 200 -48
__alloc_pages_direct_c 40 - -40
try_to_compact_pages 72 72 0
-88
[1] http://marc.info/?l=linux-mm&m=140142462528257&w=2
This patch (of 4):
prep_new_page() sets almost everything in the struct page of the page
being allocated, except page->pfmemalloc. This is not obvious and has at
least once led to a bug where page->pfmemalloc was forgotten to be set
correctly, see commit 8fb74b9fb2b1 ("mm: compaction: partially revert
capture of suitable high-order page").
This patch moves the pfmemalloc setting to prep_new_page(), which means it
needs to gain alloc_flags parameter. The call to prep_new_page is moved
from buffered_rmqueue() to get_page_from_freelist(), which also leads to
simpler code. An obsolete comment for buffered_rmqueue() is replaced.
In addition to better maintainability there is a small reduction of code
and stack usage for get_page_from_freelist(), which inlines the other
functions involved.
add/remove: 0/0 grow/shrink: 0/1 up/down: 0/-145 (-145)
function old new delta
get_page_from_freelist 2670 2525 -145
Stack usage is reduced from 184 to 168 bytes.
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:25:38 -08:00
|
|
|
|
2015-08-21 14:11:51 -07:00
|
|
|
|
mm: set page->pfmemalloc in prep_new_page()
The possibility of replacing the numerous parameters of alloc_pages*
functions with a single structure has been discussed when Minchan proposed
to expand the x86 kernel stack [1]. This series implements the change,
along with few more cleanups/microoptimizations.
The series is based on next-20150108 and I used gcc 4.8.3 20140627 on
openSUSE 13.2 for compiling. Config includess NUMA and COMPACTION.
The core change is the introduction of a new struct alloc_context, which looks
like this:
struct alloc_context {
struct zonelist *zonelist;
nodemask_t *nodemask;
struct zone *preferred_zone;
int classzone_idx;
int migratetype;
enum zone_type high_zoneidx;
};
All the contents is mostly constant, except that __alloc_pages_slowpath()
changes preferred_zone, classzone_idx and potentially zonelist. But
that's not a problem in case control returns to retry_cpuset: in
__alloc_pages_nodemask(), those will be reset to initial values again
(although it's a bit subtle). On the other hand, gfp_flags and alloc_info
mutate so much that it doesn't make sense to put them into alloc_context.
Still, the result is one parameter instead of up to 7. This is all in
Patch 2.
Patch 3 is a step to expand alloc_context usage out of page_alloc.c
itself. The function try_to_compact_pages() can also much benefit from
the parameter reduction, but it means the struct definition has to be
moved to a shared header.
Patch 1 should IMHO be included even if the rest is deemed not useful
enough. It improves maintainability and also has some code/stack
reduction. Patch 4 is OTOH a tiny optimization.
Overall bloat-o-meter results:
add/remove: 0/0 grow/shrink: 0/4 up/down: 0/-460 (-460)
function old new delta
nr_free_zone_pages 129 115 -14
__alloc_pages_direct_compact 329 256 -73
get_page_from_freelist 2670 2576 -94
__alloc_pages_nodemask 2564 2285 -279
try_to_compact_pages 582 579 -3
Overall stack sizes per ./scripts/checkstack.pl:
old new delta
get_page_from_freelist: 184 184 0
__alloc_pages_nodemask 248 200 -48
__alloc_pages_direct_c 40 - -40
try_to_compact_pages 72 72 0
-88
[1] http://marc.info/?l=linux-mm&m=140142462528257&w=2
This patch (of 4):
prep_new_page() sets almost everything in the struct page of the page
being allocated, except page->pfmemalloc. This is not obvious and has at
least once led to a bug where page->pfmemalloc was forgotten to be set
correctly, see commit 8fb74b9fb2b1 ("mm: compaction: partially revert
capture of suitable high-order page").
This patch moves the pfmemalloc setting to prep_new_page(), which means it
needs to gain alloc_flags parameter. The call to prep_new_page is moved
from buffered_rmqueue() to get_page_from_freelist(), which also leads to
simpler code. An obsolete comment for buffered_rmqueue() is replaced.
In addition to better maintainability there is a small reduction of code
and stack usage for get_page_from_freelist(), which inlines the other
functions involved.
add/remove: 0/0 grow/shrink: 0/1 up/down: 0/-145 (-145)
function old new delta
get_page_from_freelist 2670 2525 -145
Stack usage is reduced from 184 to 168 bytes.
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:25:38 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-08-21 14:11:51 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
Bias the location of pages freed for min_free_kbytes in the same MAX_ORDER_NR_PAGES blocks
The standard buddy allocator always favours the smallest block of pages.
The effect of this is that the pages free to satisfy min_free_kbytes tends
to be preserved since boot time at the same location of memory ffor a very
long time and as a contiguous block. When an administrator sets the
reserve at 16384 at boot time, it tends to be the same MAX_ORDER blocks
that remain free. This allows the occasional high atomic allocation to
succeed up until the point the blocks are split. In practice, it is
difficult to split these blocks but when they do split, the benefit of
having min_free_kbytes for contiguous blocks disappears. Additionally,
increasing min_free_kbytes once the system has been running for some time
has no guarantee of creating contiguous blocks.
On the other hand, CONFIG_PAGE_GROUP_BY_MOBILITY favours splitting large
blocks when there are no free pages of the appropriate type available. A
side-effect of this is that all blocks in memory tends to be used up and
the contiguous free blocks from boot time are not preserved like in the
vanilla allocator. This can cause a problem if a new caller is unwilling
to reclaim or does not reclaim for long enough.
A failure scenario was found for a wireless network device allocating
order-1 atomic allocations but the allocations were not intense or frequent
enough for a whole block of pages to be preserved for MIGRATE_HIGHALLOC.
This was reproduced on a desktop by booting with mem=256mb, forcing the
driver to allocate at order-1, running a bittorrent client (downloading a
debian ISO) and building a kernel with -j2.
This patch addresses the problem on the desktop machine booted with
mem=256mb. It works by setting aside a reserve of MAX_ORDER_NR_PAGES
blocks, the number of which depends on the value of min_free_kbytes. These
blocks are only fallen back to when there is no other free pages. Then the
smallest possible page is used just like the normal buddy allocator instead
of the largest possible page to preserve contiguous pages The pages in free
lists in the reserve blocks are never taken for another migrate type. The
results is that even if min_free_kbytes is set to a low value, contiguous
blocks will be preserved in the MIGRATE_RESERVE blocks.
This works better than the vanilla allocator because if min_free_kbytes is
increased, a new reserve block will be chosen based on the location of
reclaimable pages and the block will free up as contiguous pages. In the
vanilla allocator, no effort is made to target a block of pages to free as
contiguous pages and min_free_kbytes pages are scattered randomly.
This effect has been observed on the test machine. min_free_kbytes was set
initially low but it was kept as a contiguous free block within
MIGRATE_RESERVE. min_free_kbytes was then set to a higher value and over a
period of time, the free blocks were within the reserve and coalescing.
How long it takes to free up depends on how quickly LRU is rotating.
Amusingly, this means that more activity will free the blocks faster.
This mechanism potentially replaces MIGRATE_HIGHALLOC as it may be more
effective than grouping contiguous free pages together. It all depends on
whether the number of active atomic high allocations exceeds
min_free_kbytes or not. If the number of active allocations exceeds
min_free_kbytes, it's worth it but maybe in that situation, min_free_kbytes
should be set higher. Once there are no more reports of allocation
failures, a patch will be submitted that backs out MIGRATE_HIGHALLOC and
see if the reports stay missing.
Credit to Mariusz Kozlowski for discovering the problem, describing the
failure scenario and testing patches and scenarios.
[akpm@linux-foundation.org: cleanups]
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Acked-by: Andy Whitcroft <apw@shadowen.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 01:25:58 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-11-15 17:36:53 -08:00
|
|
|
|
2009-06-16 15:32:04 -07:00
|
|
|
|
Bias the location of pages freed for min_free_kbytes in the same MAX_ORDER_NR_PAGES blocks
The standard buddy allocator always favours the smallest block of pages.
The effect of this is that the pages free to satisfy min_free_kbytes tends
to be preserved since boot time at the same location of memory ffor a very
long time and as a contiguous block. When an administrator sets the
reserve at 16384 at boot time, it tends to be the same MAX_ORDER blocks
that remain free. This allows the occasional high atomic allocation to
succeed up until the point the blocks are split. In practice, it is
difficult to split these blocks but when they do split, the benefit of
having min_free_kbytes for contiguous blocks disappears. Additionally,
increasing min_free_kbytes once the system has been running for some time
has no guarantee of creating contiguous blocks.
On the other hand, CONFIG_PAGE_GROUP_BY_MOBILITY favours splitting large
blocks when there are no free pages of the appropriate type available. A
side-effect of this is that all blocks in memory tends to be used up and
the contiguous free blocks from boot time are not preserved like in the
vanilla allocator. This can cause a problem if a new caller is unwilling
to reclaim or does not reclaim for long enough.
A failure scenario was found for a wireless network device allocating
order-1 atomic allocations but the allocations were not intense or frequent
enough for a whole block of pages to be preserved for MIGRATE_HIGHALLOC.
This was reproduced on a desktop by booting with mem=256mb, forcing the
driver to allocate at order-1, running a bittorrent client (downloading a
debian ISO) and building a kernel with -j2.
This patch addresses the problem on the desktop machine booted with
mem=256mb. It works by setting aside a reserve of MAX_ORDER_NR_PAGES
blocks, the number of which depends on the value of min_free_kbytes. These
blocks are only fallen back to when there is no other free pages. Then the
smallest possible page is used just like the normal buddy allocator instead
of the largest possible page to preserve contiguous pages The pages in free
lists in the reserve blocks are never taken for another migrate type. The
results is that even if min_free_kbytes is set to a low value, contiguous
blocks will be preserved in the MIGRATE_RESERVE blocks.
This works better than the vanilla allocator because if min_free_kbytes is
increased, a new reserve block will be chosen based on the location of
reclaimable pages and the block will free up as contiguous pages. In the
vanilla allocator, no effort is made to target a block of pages to free as
contiguous pages and min_free_kbytes pages are scattered randomly.
This effect has been observed on the test machine. min_free_kbytes was set
initially low but it was kept as a contiguous free block within
MIGRATE_RESERVE. min_free_kbytes was then set to a higher value and over a
period of time, the free blocks were within the reserve and coalescing.
How long it takes to free up depends on how quickly LRU is rotating.
Amusingly, this means that more activity will free the blocks faster.
This mechanism potentially replaces MIGRATE_HIGHALLOC as it may be more
effective than grouping contiguous free pages together. It all depends on
whether the number of active atomic high allocations exceeds
min_free_kbytes or not. If the number of active allocations exceeds
min_free_kbytes, it's worth it but maybe in that situation, min_free_kbytes
should be set higher. Once there are no more reports of allocation
failures, a patch will be submitted that backs out MIGRATE_HIGHALLOC and
see if the reports stay missing.
Credit to Mariusz Kozlowski for discovering the problem, describing the
failure scenario and testing patches and scenarios.
[akpm@linux-foundation.org: cleanups]
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Acked-by: Andy Whitcroft <apw@shadowen.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 01:25:58 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-09-11 14:20:34 -07:00
|
|
|
|
Bias the location of pages freed for min_free_kbytes in the same MAX_ORDER_NR_PAGES blocks
The standard buddy allocator always favours the smallest block of pages.
The effect of this is that the pages free to satisfy min_free_kbytes tends
to be preserved since boot time at the same location of memory ffor a very
long time and as a contiguous block. When an administrator sets the
reserve at 16384 at boot time, it tends to be the same MAX_ORDER blocks
that remain free. This allows the occasional high atomic allocation to
succeed up until the point the blocks are split. In practice, it is
difficult to split these blocks but when they do split, the benefit of
having min_free_kbytes for contiguous blocks disappears. Additionally,
increasing min_free_kbytes once the system has been running for some time
has no guarantee of creating contiguous blocks.
On the other hand, CONFIG_PAGE_GROUP_BY_MOBILITY favours splitting large
blocks when there are no free pages of the appropriate type available. A
side-effect of this is that all blocks in memory tends to be used up and
the contiguous free blocks from boot time are not preserved like in the
vanilla allocator. This can cause a problem if a new caller is unwilling
to reclaim or does not reclaim for long enough.
A failure scenario was found for a wireless network device allocating
order-1 atomic allocations but the allocations were not intense or frequent
enough for a whole block of pages to be preserved for MIGRATE_HIGHALLOC.
This was reproduced on a desktop by booting with mem=256mb, forcing the
driver to allocate at order-1, running a bittorrent client (downloading a
debian ISO) and building a kernel with -j2.
This patch addresses the problem on the desktop machine booted with
mem=256mb. It works by setting aside a reserve of MAX_ORDER_NR_PAGES
blocks, the number of which depends on the value of min_free_kbytes. These
blocks are only fallen back to when there is no other free pages. Then the
smallest possible page is used just like the normal buddy allocator instead
of the largest possible page to preserve contiguous pages The pages in free
lists in the reserve blocks are never taken for another migrate type. The
results is that even if min_free_kbytes is set to a low value, contiguous
blocks will be preserved in the MIGRATE_RESERVE blocks.
This works better than the vanilla allocator because if min_free_kbytes is
increased, a new reserve block will be chosen based on the location of
reclaimable pages and the block will free up as contiguous pages. In the
vanilla allocator, no effort is made to target a block of pages to free as
contiguous pages and min_free_kbytes pages are scattered randomly.
This effect has been observed on the test machine. min_free_kbytes was set
initially low but it was kept as a contiguous free block within
MIGRATE_RESERVE. min_free_kbytes was then set to a higher value and over a
period of time, the free blocks were within the reserve and coalescing.
How long it takes to free up depends on how quickly LRU is rotating.
Amusingly, this means that more activity will free the blocks faster.
This mechanism potentially replaces MIGRATE_HIGHALLOC as it may be more
effective than grouping contiguous free pages together. It all depends on
whether the number of active atomic high allocations exceeds
min_free_kbytes or not. If the number of active allocations exceeds
min_free_kbytes, it's worth it but maybe in that situation, min_free_kbytes
should be set higher. Once there are no more reports of allocation
failures, a patch will be submitted that backs out MIGRATE_HIGHALLOC and
see if the reports stay missing.
Credit to Mariusz Kozlowski for discovering the problem, describing the
failure scenario and testing patches and scenarios.
[akpm@linux-foundation.org: cleanups]
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Acked-by: Andy Whitcroft <apw@shadowen.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 01:25:58 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-12-28 17:47:03 +03:00
|
|
|
|
Bias the location of pages freed for min_free_kbytes in the same MAX_ORDER_NR_PAGES blocks
The standard buddy allocator always favours the smallest block of pages.
The effect of this is that the pages free to satisfy min_free_kbytes tends
to be preserved since boot time at the same location of memory ffor a very
long time and as a contiguous block. When an administrator sets the
reserve at 16384 at boot time, it tends to be the same MAX_ORDER blocks
that remain free. This allows the occasional high atomic allocation to
succeed up until the point the blocks are split. In practice, it is
difficult to split these blocks but when they do split, the benefit of
having min_free_kbytes for contiguous blocks disappears. Additionally,
increasing min_free_kbytes once the system has been running for some time
has no guarantee of creating contiguous blocks.
On the other hand, CONFIG_PAGE_GROUP_BY_MOBILITY favours splitting large
blocks when there are no free pages of the appropriate type available. A
side-effect of this is that all blocks in memory tends to be used up and
the contiguous free blocks from boot time are not preserved like in the
vanilla allocator. This can cause a problem if a new caller is unwilling
to reclaim or does not reclaim for long enough.
A failure scenario was found for a wireless network device allocating
order-1 atomic allocations but the allocations were not intense or frequent
enough for a whole block of pages to be preserved for MIGRATE_HIGHALLOC.
This was reproduced on a desktop by booting with mem=256mb, forcing the
driver to allocate at order-1, running a bittorrent client (downloading a
debian ISO) and building a kernel with -j2.
This patch addresses the problem on the desktop machine booted with
mem=256mb. It works by setting aside a reserve of MAX_ORDER_NR_PAGES
blocks, the number of which depends on the value of min_free_kbytes. These
blocks are only fallen back to when there is no other free pages. Then the
smallest possible page is used just like the normal buddy allocator instead
of the largest possible page to preserve contiguous pages The pages in free
lists in the reserve blocks are never taken for another migrate type. The
results is that even if min_free_kbytes is set to a low value, contiguous
blocks will be preserved in the MIGRATE_RESERVE blocks.
This works better than the vanilla allocator because if min_free_kbytes is
increased, a new reserve block will be chosen based on the location of
reclaimable pages and the block will free up as contiguous pages. In the
vanilla allocator, no effort is made to target a block of pages to free as
contiguous pages and min_free_kbytes pages are scattered randomly.
This effect has been observed on the test machine. min_free_kbytes was set
initially low but it was kept as a contiguous free block within
MIGRATE_RESERVE. min_free_kbytes was then set to a higher value and over a
period of time, the free blocks were within the reserve and coalescing.
How long it takes to free up depends on how quickly LRU is rotating.
Amusingly, this means that more activity will free the blocks faster.
This mechanism potentially replaces MIGRATE_HIGHALLOC as it may be more
effective than grouping contiguous free pages together. It all depends on
whether the number of active atomic high allocations exceeds
min_free_kbytes or not. If the number of active allocations exceeds
min_free_kbytes, it's worth it but maybe in that situation, min_free_kbytes
should be set higher. Once there are no more reports of allocation
failures, a patch will be submitted that backs out MIGRATE_HIGHALLOC and
see if the reports stay missing.
Credit to Mariusz Kozlowski for discovering the problem, describing the
failure scenario and testing patches and scenarios.
[akpm@linux-foundation.org: cleanups]
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Acked-by: Andy Whitcroft <apw@shadowen.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 01:25:58 -07:00
|
|
|
|
2019-05-14 15:41:32 -07:00
|
|
|
|
2016-01-14 15:20:30 -08:00
|
|
|
|
|
|
|
|
|
2024-08-26 14:40:48 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-05-19 14:08:54 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Bias the location of pages freed for min_free_kbytes in the same MAX_ORDER_NR_PAGES blocks
The standard buddy allocator always favours the smallest block of pages.
The effect of this is that the pages free to satisfy min_free_kbytes tends
to be preserved since boot time at the same location of memory ffor a very
long time and as a contiguous block. When an administrator sets the
reserve at 16384 at boot time, it tends to be the same MAX_ORDER blocks
that remain free. This allows the occasional high atomic allocation to
succeed up until the point the blocks are split. In practice, it is
difficult to split these blocks but when they do split, the benefit of
having min_free_kbytes for contiguous blocks disappears. Additionally,
increasing min_free_kbytes once the system has been running for some time
has no guarantee of creating contiguous blocks.
On the other hand, CONFIG_PAGE_GROUP_BY_MOBILITY favours splitting large
blocks when there are no free pages of the appropriate type available. A
side-effect of this is that all blocks in memory tends to be used up and
the contiguous free blocks from boot time are not preserved like in the
vanilla allocator. This can cause a problem if a new caller is unwilling
to reclaim or does not reclaim for long enough.
A failure scenario was found for a wireless network device allocating
order-1 atomic allocations but the allocations were not intense or frequent
enough for a whole block of pages to be preserved for MIGRATE_HIGHALLOC.
This was reproduced on a desktop by booting with mem=256mb, forcing the
driver to allocate at order-1, running a bittorrent client (downloading a
debian ISO) and building a kernel with -j2.
This patch addresses the problem on the desktop machine booted with
mem=256mb. It works by setting aside a reserve of MAX_ORDER_NR_PAGES
blocks, the number of which depends on the value of min_free_kbytes. These
blocks are only fallen back to when there is no other free pages. Then the
smallest possible page is used just like the normal buddy allocator instead
of the largest possible page to preserve contiguous pages The pages in free
lists in the reserve blocks are never taken for another migrate type. The
results is that even if min_free_kbytes is set to a low value, contiguous
blocks will be preserved in the MIGRATE_RESERVE blocks.
This works better than the vanilla allocator because if min_free_kbytes is
increased, a new reserve block will be chosen based on the location of
reclaimable pages and the block will free up as contiguous pages. In the
vanilla allocator, no effort is made to target a block of pages to free as
contiguous pages and min_free_kbytes pages are scattered randomly.
This effect has been observed on the test machine. min_free_kbytes was set
initially low but it was kept as a contiguous free block within
MIGRATE_RESERVE. min_free_kbytes was then set to a higher value and over a
period of time, the free blocks were within the reserve and coalescing.
How long it takes to free up depends on how quickly LRU is rotating.
Amusingly, this means that more activity will free the blocks faster.
This mechanism potentially replaces MIGRATE_HIGHALLOC as it may be more
effective than grouping contiguous free pages together. It all depends on
whether the number of active atomic high allocations exceeds
min_free_kbytes or not. If the number of active allocations exceeds
min_free_kbytes, it's worth it but maybe in that situation, min_free_kbytes
should be set higher. Once there are no more reports of allocation
failures, a patch will be submitted that backs out MIGRATE_HIGHALLOC and
see if the reports stay missing.
Credit to Mariusz Kozlowski for discovering the problem, describing the
failure scenario and testing patches and scenarios.
[akpm@linux-foundation.org: cleanups]
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Acked-by: Andy Whitcroft <apw@shadowen.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 01:25:58 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-10-16 01:25:48 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-03-22 14:43:05 -07:00
|
|
|
|
|
|
|
|
|
2007-10-16 01:25:48 -07:00
|
|
|
|
2024-03-26 14:11:33 +08:00
|
|
|
|
2023-02-03 18:01:32 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-10-16 01:25:48 -07:00
|
|
|
|
|
|
|
|
|
2015-04-14 15:45:15 -07:00
|
|
|
|
2017-11-15 17:36:53 -08:00
|
|
|
|
2015-04-14 15:45:15 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-10-16 01:25:51 -07:00
|
|
|
|
2025-06-16 22:11:12 -04:00
|
|
|
|
|
|
|
|
|
2007-10-16 01:25:51 -07:00
|
|
|
|
2024-04-25 20:56:04 -07:00
|
|
|
|
|
|
|
|
|
2007-10-16 01:25:51 -07:00
|
|
|
|
|
|
|
|
|
2024-04-25 20:56:04 -07:00
|
|
|
|
2015-11-06 16:29:57 -08:00
|
|
|
|
2007-10-16 01:26:00 -07:00
|
|
|
|
2007-10-16 01:25:51 -07:00
|
|
|
|
2024-03-20 14:02:13 -04:00
|
|
|
|
2024-04-25 20:56:04 -07:00
|
|
|
|
2024-03-20 14:02:13 -04:00
|
|
|
|
2024-04-25 20:56:04 -07:00
|
|
|
|
2021-04-29 23:01:36 -07:00
|
|
|
|
2007-10-16 01:25:51 -07:00
|
|
|
|
2021-04-29 23:01:36 -07:00
|
|
|
|
2007-10-16 01:25:51 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, page_alloc: move_freepages should not examine struct page of reserved memory
After commit 907ec5fca3dc ("mm: zero remaining unavailable struct
pages"), struct page of reserved memory is zeroed. This causes
page->flags to be 0 and fixes issues related to reading
/proc/kpageflags, for example, of reserved memory.
The VM_BUG_ON() in move_freepages_block(), however, assumes that
page_zone() is meaningful even for reserved memory. That assumption is
no longer true after the aforementioned commit.
There's no reason why move_freepages_block() should be testing the
legitimacy of page_zone() for reserved memory; its scope is limited only
to pages on the zone's freelist.
Note that pfn_valid() can be true for reserved memory: there is a
backing struct page. The check for page_to_nid(page) is also buggy but
reserved memory normally only appears on node 0 so the zeroing doesn't
affect this.
Move the debug checks to after verifying PageBuddy is true. This
isolates the scope of the checks to only be for buddy pages which are on
the zone's freelist which move_freepages_block() is operating on. In
this case, an incorrect node or zone is a bug worthy of being warned
about (and the examination of struct page is acceptable bcause this
memory is not reserved).
Why does move_freepages_block() gets called on reserved memory? It's
simply math after finding a valid free page from the per-zone free area
to use as fallback. We find the beginning and end of the pageblock of
the valid page and that can bring us into memory that was reserved per
the e820. pfn_valid() is still true (it's backed by a struct page), but
since it's zero'd we shouldn't make any inferences here about comparing
its node or zone. The current node check just happens to succeed most
of the time by luck because reserved memory typically appears on node 0.
The fix here is to validate that we actually have buddy pages before
testing if there's any type of zone or node strangeness going on.
We noticed it almost immediately after bringing 907ec5fca3dc in on
CONFIG_DEBUG_VM builds. It depends on finding specific free pages in
the per-zone free area where the math in move_freepages() will bring the
start or end pfn into reserved memory and wanting to claim that entire
pageblock as a new migratetype. So the path will be rare, require
CONFIG_DEBUG_VM, and require fallback to a different migratetype.
Some struct pages were already zeroed from reserve pages before
907ec5fca3c so it theoretically could trigger before this commit. I
think it's rare enough under a config option that most people don't run
that others may not have noticed. I wouldn't argue against a stable tag
and the backport should be easy enough, but probably wouldn't single out
a commit that this is fixing.
Mel said:
: The overhead of the debugging check is higher with this patch although
: it'll only affect debug builds and the path is not particularly hot.
: If this was a concern, I think it would be reasonable to simply remove
: the debugging check as the zone boundaries are checked in
: move_freepages_block and we never expect a zone/node to be smaller than
: a pageblock and stuck in the middle of another zone.
Link: http://lkml.kernel.org/r/alpine.DEB.2.21.1908122036560.10779@chino.kir.corp.google.com
Signed-off-by: David Rientjes <rientjes@google.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Masayoshi Mizuma <m.mizuma@jp.fujitsu.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Pavel Tatashin <pavel.tatashin@microsoft.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-08-24 17:54:40 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-10-15 20:10:15 -07:00
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-04-29 23:01:36 -07:00
|
|
|
|
2007-10-16 01:26:00 -07:00
|
|
|
|
2007-10-16 01:25:51 -07:00
|
|
|
|
|
|
|
|
|
2007-10-16 01:26:00 -07:00
|
|
|
|
2007-10-16 01:25:51 -07:00
|
|
|
|
|
|
|
|
|
mm: page_alloc: fix freelist movement during block conversion
Currently, page block type conversion during fallbacks, atomic
reservations and isolation can strand various amounts of free pages on
incorrect freelists.
For example, fallback stealing moves free pages in the block to the new
type's freelists, but then may not actually claim the block for that type
if there aren't enough compatible pages already allocated.
In all cases, free page moving might fail if the block straddles more than
one zone, in which case no free pages are moved at all, but the block type
is changed anyway.
This is detrimental to type hygiene on the freelists. It encourages
incompatible page mixing down the line (ask for one type, get another) and
thus contributes to long-term fragmentation.
Split the process into a proper transaction: check first if conversion
will happen, then try to move the free pages, and only if that was
successful convert the block to the new type.
[baolin.wang@linux.alibaba.com: fix allocation failures with CONFIG_CMA]
Link: https://lkml.kernel.org/r/a97697e0-45b0-4f71-b087-fdc7a1d43c0e@linux.alibaba.com
Link: https://lkml.kernel.org/r/20240320180429.678181-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Tested-by: "Huang, Ying" <ying.huang@intel.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-20 14:02:11 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-10-16 01:25:51 -07:00
|
|
|
|
mm: page_alloc: fix freelist movement during block conversion
Currently, page block type conversion during fallbacks, atomic
reservations and isolation can strand various amounts of free pages on
incorrect freelists.
For example, fallback stealing moves free pages in the block to the new
type's freelists, but then may not actually claim the block for that type
if there aren't enough compatible pages already allocated.
In all cases, free page moving might fail if the block straddles more than
one zone, in which case no free pages are moved at all, but the block type
is changed anyway.
This is detrimental to type hygiene on the freelists. It encourages
incompatible page mixing down the line (ask for one type, get another) and
thus contributes to long-term fragmentation.
Split the process into a proper transaction: check first if conversion
will happen, then try to move the free pages, and only if that was
successful convert the block to the new type.
[baolin.wang@linux.alibaba.com: fix allocation failures with CONFIG_CMA]
Link: https://lkml.kernel.org/r/a97697e0-45b0-4f71-b087-fdc7a1d43c0e@linux.alibaba.com
Link: https://lkml.kernel.org/r/20240320180429.678181-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Tested-by: "Huang, Ying" <ying.huang@intel.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-20 14:02:11 -04:00
|
|
|
|
2018-10-26 15:09:24 -07:00
|
|
|
|
2021-04-29 23:01:36 -07:00
|
|
|
|
mm: page_alloc: fix freelist movement during block conversion
Currently, page block type conversion during fallbacks, atomic
reservations and isolation can strand various amounts of free pages on
incorrect freelists.
For example, fallback stealing moves free pages in the block to the new
type's freelists, but then may not actually claim the block for that type
if there aren't enough compatible pages already allocated.
In all cases, free page moving might fail if the block straddles more than
one zone, in which case no free pages are moved at all, but the block type
is changed anyway.
This is detrimental to type hygiene on the freelists. It encourages
incompatible page mixing down the line (ask for one type, get another) and
thus contributes to long-term fragmentation.
Split the process into a proper transaction: check first if conversion
will happen, then try to move the free pages, and only if that was
successful convert the block to the new type.
[baolin.wang@linux.alibaba.com: fix allocation failures with CONFIG_CMA]
Link: https://lkml.kernel.org/r/a97697e0-45b0-4f71-b087-fdc7a1d43c0e@linux.alibaba.com
Link: https://lkml.kernel.org/r/20240320180429.678181-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Tested-by: "Huang, Ying" <ying.huang@intel.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-20 14:02:11 -04:00
|
|
|
|
2024-04-25 20:56:04 -07:00
|
|
|
|
2007-10-16 01:25:51 -07:00
|
|
|
|
2024-03-20 14:02:10 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: page_alloc: fix freelist movement during block conversion
Currently, page block type conversion during fallbacks, atomic
reservations and isolation can strand various amounts of free pages on
incorrect freelists.
For example, fallback stealing moves free pages in the block to the new
type's freelists, but then may not actually claim the block for that type
if there aren't enough compatible pages already allocated.
In all cases, free page moving might fail if the block straddles more than
one zone, in which case no free pages are moved at all, but the block type
is changed anyway.
This is detrimental to type hygiene on the freelists. It encourages
incompatible page mixing down the line (ask for one type, get another) and
thus contributes to long-term fragmentation.
Split the process into a proper transaction: check first if conversion
will happen, then try to move the free pages, and only if that was
successful convert the block to the new type.
[baolin.wang@linux.alibaba.com: fix allocation failures with CONFIG_CMA]
Link: https://lkml.kernel.org/r/a97697e0-45b0-4f71-b087-fdc7a1d43c0e@linux.alibaba.com
Link: https://lkml.kernel.org/r/20240320180429.678181-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Tested-by: "Huang, Ying" <ying.huang@intel.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-20 14:02:11 -04:00
|
|
|
|
|
|
|
|
|
2024-04-25 20:56:04 -07:00
|
|
|
|
mm: page_alloc: fix freelist movement during block conversion
Currently, page block type conversion during fallbacks, atomic
reservations and isolation can strand various amounts of free pages on
incorrect freelists.
For example, fallback stealing moves free pages in the block to the new
type's freelists, but then may not actually claim the block for that type
if there aren't enough compatible pages already allocated.
In all cases, free page moving might fail if the block straddles more than
one zone, in which case no free pages are moved at all, but the block type
is changed anyway.
This is detrimental to type hygiene on the freelists. It encourages
incompatible page mixing down the line (ask for one type, get another) and
thus contributes to long-term fragmentation.
Split the process into a proper transaction: check first if conversion
will happen, then try to move the free pages, and only if that was
successful convert the block to the new type.
[baolin.wang@linux.alibaba.com: fix allocation failures with CONFIG_CMA]
Link: https://lkml.kernel.org/r/a97697e0-45b0-4f71-b087-fdc7a1d43c0e@linux.alibaba.com
Link: https://lkml.kernel.org/r/20240320180429.678181-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Tested-by: "Huang, Ying" <ying.huang@intel.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-20 14:02:11 -04:00
|
|
|
|
2007-10-16 01:25:51 -07:00
|
|
|
|
mm: page_alloc: fix freelist movement during block conversion
Currently, page block type conversion during fallbacks, atomic
reservations and isolation can strand various amounts of free pages on
incorrect freelists.
For example, fallback stealing moves free pages in the block to the new
type's freelists, but then may not actually claim the block for that type
if there aren't enough compatible pages already allocated.
In all cases, free page moving might fail if the block straddles more than
one zone, in which case no free pages are moved at all, but the block type
is changed anyway.
This is detrimental to type hygiene on the freelists. It encourages
incompatible page mixing down the line (ask for one type, get another) and
thus contributes to long-term fragmentation.
Split the process into a proper transaction: check first if conversion
will happen, then try to move the free pages, and only if that was
successful convert the block to the new type.
[baolin.wang@linux.alibaba.com: fix allocation failures with CONFIG_CMA]
Link: https://lkml.kernel.org/r/a97697e0-45b0-4f71-b087-fdc7a1d43c0e@linux.alibaba.com
Link: https://lkml.kernel.org/r/20240320180429.678181-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Tested-by: "Huang, Ying" <ying.huang@intel.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-20 14:02:11 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-04-25 20:56:04 -07:00
|
|
|
|
mm: page_alloc: fix freelist movement during block conversion
Currently, page block type conversion during fallbacks, atomic
reservations and isolation can strand various amounts of free pages on
incorrect freelists.
For example, fallback stealing moves free pages in the block to the new
type's freelists, but then may not actually claim the block for that type
if there aren't enough compatible pages already allocated.
In all cases, free page moving might fail if the block straddles more than
one zone, in which case no free pages are moved at all, but the block type
is changed anyway.
This is detrimental to type hygiene on the freelists. It encourages
incompatible page mixing down the line (ask for one type, get another) and
thus contributes to long-term fragmentation.
Split the process into a proper transaction: check first if conversion
will happen, then try to move the free pages, and only if that was
successful convert the block to the new type.
[baolin.wang@linux.alibaba.com: fix allocation failures with CONFIG_CMA]
Link: https://lkml.kernel.org/r/a97697e0-45b0-4f71-b087-fdc7a1d43c0e@linux.alibaba.com
Link: https://lkml.kernel.org/r/20240320180429.678181-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Tested-by: "Huang, Ying" <ying.huang@intel.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-20 14:02:11 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-07-04 12:25:10 +02:00
|
|
|
|
mm: page_alloc: fix freelist movement during block conversion
Currently, page block type conversion during fallbacks, atomic
reservations and isolation can strand various amounts of free pages on
incorrect freelists.
For example, fallback stealing moves free pages in the block to the new
type's freelists, but then may not actually claim the block for that type
if there aren't enough compatible pages already allocated.
In all cases, free page moving might fail if the block straddles more than
one zone, in which case no free pages are moved at all, but the block type
is changed anyway.
This is detrimental to type hygiene on the freelists. It encourages
incompatible page mixing down the line (ask for one type, get another) and
thus contributes to long-term fragmentation.
Split the process into a proper transaction: check first if conversion
will happen, then try to move the free pages, and only if that was
successful convert the block to the new type.
[baolin.wang@linux.alibaba.com: fix allocation failures with CONFIG_CMA]
Link: https://lkml.kernel.org/r/a97697e0-45b0-4f71-b087-fdc7a1d43c0e@linux.alibaba.com
Link: https://lkml.kernel.org/r/20240320180429.678181-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Tested-by: "Huang, Ying" <ying.huang@intel.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-20 14:02:11 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-20 14:02:14 -04:00
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
mm: page_alloc: fix freelist movement during block conversion
Currently, page block type conversion during fallbacks, atomic
reservations and isolation can strand various amounts of free pages on
incorrect freelists.
For example, fallback stealing moves free pages in the block to the new
type's freelists, but then may not actually claim the block for that type
if there aren't enough compatible pages already allocated.
In all cases, free page moving might fail if the block straddles more than
one zone, in which case no free pages are moved at all, but the block type
is changed anyway.
This is detrimental to type hygiene on the freelists. It encourages
incompatible page mixing down the line (ask for one type, get another) and
thus contributes to long-term fragmentation.
Split the process into a proper transaction: check first if conversion
will happen, then try to move the free pages, and only if that was
successful convert the block to the new type.
[baolin.wang@linux.alibaba.com: fix allocation failures with CONFIG_CMA]
Link: https://lkml.kernel.org/r/a97697e0-45b0-4f71-b087-fdc7a1d43c0e@linux.alibaba.com
Link: https://lkml.kernel.org/r/20240320180429.678181-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Tested-by: "Huang, Ying" <ying.huang@intel.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-20 14:02:11 -04:00
|
|
|
|
2024-04-25 20:56:04 -07:00
|
|
|
|
2025-06-16 22:11:12 -04:00
|
|
|
|
mm: page_alloc: fix freelist movement during block conversion
Currently, page block type conversion during fallbacks, atomic
reservations and isolation can strand various amounts of free pages on
incorrect freelists.
For example, fallback stealing moves free pages in the block to the new
type's freelists, but then may not actually claim the block for that type
if there aren't enough compatible pages already allocated.
In all cases, free page moving might fail if the block straddles more than
one zone, in which case no free pages are moved at all, but the block type
is changed anyway.
This is detrimental to type hygiene on the freelists. It encourages
incompatible page mixing down the line (ask for one type, get another) and
thus contributes to long-term fragmentation.
Split the process into a proper transaction: check first if conversion
will happen, then try to move the free pages, and only if that was
successful convert the block to the new type.
[baolin.wang@linux.alibaba.com: fix allocation failures with CONFIG_CMA]
Link: https://lkml.kernel.org/r/a97697e0-45b0-4f71-b087-fdc7a1d43c0e@linux.alibaba.com
Link: https://lkml.kernel.org/r/20240320180429.678181-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Tested-by: "Huang, Ying" <ying.huang@intel.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-20 14:02:11 -04:00
|
|
|
|
2024-04-25 20:56:04 -07:00
|
|
|
|
mm: page_alloc: fix freelist movement during block conversion
Currently, page block type conversion during fallbacks, atomic
reservations and isolation can strand various amounts of free pages on
incorrect freelists.
For example, fallback stealing moves free pages in the block to the new
type's freelists, but then may not actually claim the block for that type
if there aren't enough compatible pages already allocated.
In all cases, free page moving might fail if the block straddles more than
one zone, in which case no free pages are moved at all, but the block type
is changed anyway.
This is detrimental to type hygiene on the freelists. It encourages
incompatible page mixing down the line (ask for one type, get another) and
thus contributes to long-term fragmentation.
Split the process into a proper transaction: check first if conversion
will happen, then try to move the free pages, and only if that was
successful convert the block to the new type.
[baolin.wang@linux.alibaba.com: fix allocation failures with CONFIG_CMA]
Link: https://lkml.kernel.org/r/a97697e0-45b0-4f71-b087-fdc7a1d43c0e@linux.alibaba.com
Link: https://lkml.kernel.org/r/20240320180429.678181-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Tested-by: "Huang, Ying" <ying.huang@intel.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-20 14:02:11 -04:00
|
|
|
|
|
|
|
|
|
2025-06-16 22:11:12 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-10-16 01:25:51 -07:00
|
|
|
|
|
|
|
|
|
2024-03-20 14:02:14 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-08-28 09:16:18 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-20 14:02:14 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-06-16 22:11:12 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-08-27 07:01:04 +00:00
|
|
|
|
2025-06-16 22:11:12 -04:00
|
|
|
|
2025-08-27 07:01:04 +00:00
|
|
|
|
2025-06-16 22:11:12 -04:00
|
|
|
|
|
|
|
|
|
2024-03-20 14:02:14 -04:00
|
|
|
|
2025-06-16 22:11:12 -04:00
|
|
|
|
2024-03-20 14:02:14 -04:00
|
|
|
|
|
|
|
|
|
2025-06-16 22:11:12 -04:00
|
|
|
|
2024-03-20 14:02:14 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-06-16 22:11:12 -04:00
|
|
|
|
|
|
|
|
|
2024-03-20 14:02:14 -04:00
|
|
|
|
mm/page_alloc: check the correct buddy if it is a starting block
find_large_buddy() search buddy based on start_pfn, which maybe different
from page's pfn, e.g. when page is not pageblock aligned, because
prep_move_freepages_block() always align start_pfn to pageblock.
This means when we found a starting block at start_pfn, it may check on
the wrong page theoretically. And not split the free page as it is
supposed to, causing a freelist migratetype mismatch.
The good news is the page passed to __move_freepages_block_isolate() has
only two possible cases:
* page is pageblock aligned
* page is __first_valid_page() of this block
So it is safe for the first case, and it won't get a buddy larger than
pageblock for the second case.
To fix the issue, check the returned pfn of find_large_buddy() to decide
whether to split the free page:
1. if it is not a PageBuddy pfn, no split;
2. if it is a PageBuddy pfn but order <= pageblock_order, no split;
3. if it is a PageBuddy pfn with order > pageblock_order, start_pfn is
either in the starting block or tail block, split the PageBuddy at
pageblock_order level.
Link: https://lkml.kernel.org/r/20250905140358.28849-1-richard.weiyang@gmail.com
Signed-off-by: Wei Yang <richard.weiyang@gmail.com>
Reviewed-by: Zi Yan <ziy@nvidia.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: David Hildenbrand <david@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-09-05 14:03:58 +00:00
|
|
|
|
2025-06-16 22:11:12 -04:00
|
|
|
|
|
|
|
|
|
mm/page_alloc: check the correct buddy if it is a starting block
find_large_buddy() search buddy based on start_pfn, which maybe different
from page's pfn, e.g. when page is not pageblock aligned, because
prep_move_freepages_block() always align start_pfn to pageblock.
This means when we found a starting block at start_pfn, it may check on
the wrong page theoretically. And not split the free page as it is
supposed to, causing a freelist migratetype mismatch.
The good news is the page passed to __move_freepages_block_isolate() has
only two possible cases:
* page is pageblock aligned
* page is __first_valid_page() of this block
So it is safe for the first case, and it won't get a buddy larger than
pageblock for the second case.
To fix the issue, check the returned pfn of find_large_buddy() to decide
whether to split the free page:
1. if it is not a PageBuddy pfn, no split;
2. if it is a PageBuddy pfn but order <= pageblock_order, no split;
3. if it is a PageBuddy pfn with order > pageblock_order, start_pfn is
either in the starting block or tail block, split the PageBuddy at
pageblock_order level.
Link: https://lkml.kernel.org/r/20250905140358.28849-1-richard.weiyang@gmail.com
Signed-off-by: Wei Yang <richard.weiyang@gmail.com>
Reviewed-by: Zi Yan <ziy@nvidia.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: David Hildenbrand <david@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-09-05 14:03:58 +00:00
|
|
|
|
2025-06-16 22:11:12 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-20 14:02:14 -04:00
|
|
|
|
2024-04-25 20:56:04 -07:00
|
|
|
|
2024-03-20 14:02:14 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/page_alloc: check the correct buddy if it is a starting block
find_large_buddy() search buddy based on start_pfn, which maybe different
from page's pfn, e.g. when page is not pageblock aligned, because
prep_move_freepages_block() always align start_pfn to pageblock.
This means when we found a starting block at start_pfn, it may check on
the wrong page theoretically. And not split the free page as it is
supposed to, causing a freelist migratetype mismatch.
The good news is the page passed to __move_freepages_block_isolate() has
only two possible cases:
* page is pageblock aligned
* page is __first_valid_page() of this block
So it is safe for the first case, and it won't get a buddy larger than
pageblock for the second case.
To fix the issue, check the returned pfn of find_large_buddy() to decide
whether to split the free page:
1. if it is not a PageBuddy pfn, no split;
2. if it is a PageBuddy pfn but order <= pageblock_order, no split;
3. if it is a PageBuddy pfn with order > pageblock_order, start_pfn is
either in the starting block or tail block, split the PageBuddy at
pageblock_order level.
Link: https://lkml.kernel.org/r/20250905140358.28849-1-richard.weiyang@gmail.com
Signed-off-by: Wei Yang <richard.weiyang@gmail.com>
Reviewed-by: Zi Yan <ziy@nvidia.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: David Hildenbrand <david@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-09-05 14:03:58 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-20 14:02:14 -04:00
|
|
|
|
|
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
mm/page_alloc: check the correct buddy if it is a starting block
find_large_buddy() search buddy based on start_pfn, which maybe different
from page's pfn, e.g. when page is not pageblock aligned, because
prep_move_freepages_block() always align start_pfn to pageblock.
This means when we found a starting block at start_pfn, it may check on
the wrong page theoretically. And not split the free page as it is
supposed to, causing a freelist migratetype mismatch.
The good news is the page passed to __move_freepages_block_isolate() has
only two possible cases:
* page is pageblock aligned
* page is __first_valid_page() of this block
So it is safe for the first case, and it won't get a buddy larger than
pageblock for the second case.
To fix the issue, check the returned pfn of find_large_buddy() to decide
whether to split the free page:
1. if it is not a PageBuddy pfn, no split;
2. if it is a PageBuddy pfn but order <= pageblock_order, no split;
3. if it is a PageBuddy pfn with order > pageblock_order, start_pfn is
either in the starting block or tail block, split the PageBuddy at
pageblock_order level.
Link: https://lkml.kernel.org/r/20250905140358.28849-1-richard.weiyang@gmail.com
Signed-off-by: Wei Yang <richard.weiyang@gmail.com>
Reviewed-by: Zi Yan <ziy@nvidia.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: David Hildenbrand <david@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-09-05 14:03:58 +00:00
|
|
|
|
2025-06-16 22:11:12 -04:00
|
|
|
|
mm/page_alloc: check the correct buddy if it is a starting block
find_large_buddy() search buddy based on start_pfn, which maybe different
from page's pfn, e.g. when page is not pageblock aligned, because
prep_move_freepages_block() always align start_pfn to pageblock.
This means when we found a starting block at start_pfn, it may check on
the wrong page theoretically. And not split the free page as it is
supposed to, causing a freelist migratetype mismatch.
The good news is the page passed to __move_freepages_block_isolate() has
only two possible cases:
* page is pageblock aligned
* page is __first_valid_page() of this block
So it is safe for the first case, and it won't get a buddy larger than
pageblock for the second case.
To fix the issue, check the returned pfn of find_large_buddy() to decide
whether to split the free page:
1. if it is not a PageBuddy pfn, no split;
2. if it is a PageBuddy pfn but order <= pageblock_order, no split;
3. if it is a PageBuddy pfn with order > pageblock_order, start_pfn is
either in the starting block or tail block, split the PageBuddy at
pageblock_order level.
Link: https://lkml.kernel.org/r/20250905140358.28849-1-richard.weiyang@gmail.com
Signed-off-by: Wei Yang <richard.weiyang@gmail.com>
Reviewed-by: Zi Yan <ziy@nvidia.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: David Hildenbrand <david@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-09-05 14:03:58 +00:00
|
|
|
|
2024-03-20 14:02:14 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-06-16 22:11:12 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-20 14:02:14 -04:00
|
|
|
|
|
|
|
|
|
2025-06-16 22:11:12 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-20 14:02:14 -04:00
|
|
|
|
|
|
|
|
|
2009-09-21 17:02:31 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-12-14 19:12:15 -08:00
|
|
|
|
mm: reclaim small amounts of memory when an external fragmentation event occurs
An external fragmentation event was previously described as
When the page allocator fragments memory, it records the event using
the mm_page_alloc_extfrag event. If the fallback_order is smaller
than a pageblock order (order-9 on 64-bit x86) then it's considered
an event that will cause external fragmentation issues in the future.
The kernel reduces the probability of such events by increasing the
watermark sizes by calling set_recommended_min_free_kbytes early in the
lifetime of the system. This works reasonably well in general but if
there are enough sparsely populated pageblocks then the problem can still
occur as enough memory is free overall and kswapd stays asleep.
This patch introduces a watermark_boost_factor sysctl that allows a zone
watermark to be temporarily boosted when an external fragmentation causing
events occurs. The boosting will stall allocations that would decrease
free memory below the boosted low watermark and kswapd is woken if the
calling context allows to reclaim an amount of memory relative to the size
of the high watermark and the watermark_boost_factor until the boost is
cleared. When kswapd finishes, it wakes kcompactd at the pageblock order
to clean some of the pageblocks that may have been affected by the
fragmentation event. kswapd avoids any writeback, slab shrinkage and swap
from reclaim context during this operation to avoid excessive system
disruption in the name of fragmentation avoidance. Care is taken so that
kswapd will do normal reclaim work if the system is really low on memory.
This was evaluated using the same workloads as "mm, page_alloc: Spread
allocations across zones before introducing fragmentation".
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
4.20-rc3+patch1-4: 18421 (98% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-1 653.58 ( 0.00%) 652.71 ( 0.13%)
Amean fault-huge-1 0.00 ( 0.00%) 178.93 * -99.00%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 0.00 ( 0.00%) 5.12 ( 100.00%)
Note that external fragmentation causing events are massively reduced by
this path whether in comparison to the previous kernel or the vanilla
kernel. The fault latency for huge pages appears to be increased but that
is only because THP allocations were successful with the patch applied.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
4.20-rc3+patch1-4: 13464 (95% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Min fault-base-1 912.00 ( 0.00%) 905.00 ( 0.77%)
Min fault-huge-1 127.00 ( 0.00%) 135.00 ( -6.30%)
Amean fault-base-1 1467.55 ( 0.00%) 1481.67 ( -0.96%)
Amean fault-huge-1 1127.11 ( 0.00%) 1063.88 * 5.61%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 77.64 ( 0.00%) 83.46 ( 7.49%)
As before, massive reduction in external fragmentation events, some jitter
on latencies and an increase in THP allocation success rates.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
4.20-rc3+patch1-4: 14263 (93% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 1346.45 ( 0.00%) 1306.87 ( 2.94%)
Amean fault-huge-5 3418.60 ( 0.00%) 1348.94 ( 60.54%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 0.78 ( 0.00%) 7.91 ( 910.64%)
There is a 93% reduction in fragmentation causing events, there is a big
reduction in the huge page fault latency and allocation success rate is
higher.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
4.20-rc3+patch1-4: 11095 (93% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 6217.43 ( 0.00%) 7419.67 * -19.34%*
Amean fault-huge-5 3163.33 ( 0.00%) 3263.80 ( -3.18%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 95.14 ( 0.00%) 87.98 ( -7.53%)
There is a large reduction in fragmentation events with some jitter around
the latencies and success rates. As before, the high THP allocation
success rate does mean the system is under a lot of pressure. However, as
the fragmentation events are reduced, it would be expected that the
long-term allocation success rate would be higher.
Link: http://lkml.kernel.org/r/20181123114528.28802-5-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:52 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-12-14 19:12:15 -08:00
|
|
|
|
mm: limit boost_watermark on small zones
Commit 1c30844d2dfe ("mm: reclaim small amounts of memory when an
external fragmentation event occurs") adds a boost_watermark() function
which increases the min watermark in a zone by at least
pageblock_nr_pages or the number of pages in a page block.
On Arm64, with 64K pages and 512M huge pages, this is 8192 pages or
512M. It does this regardless of the number of managed pages managed in
the zone or the likelihood of success.
This can put the zone immediately under water in terms of allocating
pages from the zone, and can cause a small machine to fail immediately
due to OoM. Unlike set_recommended_min_free_kbytes(), which
substantially increases min_free_kbytes and is tied to THP,
boost_watermark() can be called even if THP is not active.
The problem is most likely to appear on architectures such as Arm64
where pageblock_nr_pages is very large.
It is desirable to run the kdump capture kernel in as small a space as
possible to avoid wasting memory. In some architectures, such as Arm64,
there are restrictions on where the capture kernel can run, and
therefore, the space available. A capture kernel running in 768M can
fail due to OoM immediately after boost_watermark() sets the min in zone
DMA32, where most of the memory is, to 512M. It fails even though there
is over 500M of free memory. With boost_watermark() suppressed, the
capture kernel can run successfully in 448M.
This patch limits boost_watermark() to boosting a zone's min watermark
only when there are enough pages that the boost will produce positive
results. In this case that is estimated to be four times as many pages
as pageblock_nr_pages.
Mel said:
: There is no harm in marking it stable. Clearly it does not happen very
: often but it's not impossible. 32-bit x86 is a lot less common now
: which would previously have been vulnerable to triggering this easily.
: ppc64 has a larger base page size but typically only has one zone.
: arm64 is likely the most vulnerable, particularly when CMA is
: configured with a small movable zone.
Fixes: 1c30844d2dfe ("mm: reclaim small amounts of memory when an external fragmentation event occurs")
Signed-off-by: Henry Willard <henry.willard@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: David Hildenbrand <david@redhat.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: <stable@vger.kernel.org>
Link: http://lkml.kernel.org/r/1588294148-6586-1-git-send-email-henry.willard@oracle.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-05-07 18:36:27 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-12-14 19:12:15 -08:00
|
|
|
|
mm: reclaim small amounts of memory when an external fragmentation event occurs
An external fragmentation event was previously described as
When the page allocator fragments memory, it records the event using
the mm_page_alloc_extfrag event. If the fallback_order is smaller
than a pageblock order (order-9 on 64-bit x86) then it's considered
an event that will cause external fragmentation issues in the future.
The kernel reduces the probability of such events by increasing the
watermark sizes by calling set_recommended_min_free_kbytes early in the
lifetime of the system. This works reasonably well in general but if
there are enough sparsely populated pageblocks then the problem can still
occur as enough memory is free overall and kswapd stays asleep.
This patch introduces a watermark_boost_factor sysctl that allows a zone
watermark to be temporarily boosted when an external fragmentation causing
events occurs. The boosting will stall allocations that would decrease
free memory below the boosted low watermark and kswapd is woken if the
calling context allows to reclaim an amount of memory relative to the size
of the high watermark and the watermark_boost_factor until the boost is
cleared. When kswapd finishes, it wakes kcompactd at the pageblock order
to clean some of the pageblocks that may have been affected by the
fragmentation event. kswapd avoids any writeback, slab shrinkage and swap
from reclaim context during this operation to avoid excessive system
disruption in the name of fragmentation avoidance. Care is taken so that
kswapd will do normal reclaim work if the system is really low on memory.
This was evaluated using the same workloads as "mm, page_alloc: Spread
allocations across zones before introducing fragmentation".
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
4.20-rc3+patch1-4: 18421 (98% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-1 653.58 ( 0.00%) 652.71 ( 0.13%)
Amean fault-huge-1 0.00 ( 0.00%) 178.93 * -99.00%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 0.00 ( 0.00%) 5.12 ( 100.00%)
Note that external fragmentation causing events are massively reduced by
this path whether in comparison to the previous kernel or the vanilla
kernel. The fault latency for huge pages appears to be increased but that
is only because THP allocations were successful with the patch applied.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
4.20-rc3+patch1-4: 13464 (95% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Min fault-base-1 912.00 ( 0.00%) 905.00 ( 0.77%)
Min fault-huge-1 127.00 ( 0.00%) 135.00 ( -6.30%)
Amean fault-base-1 1467.55 ( 0.00%) 1481.67 ( -0.96%)
Amean fault-huge-1 1127.11 ( 0.00%) 1063.88 * 5.61%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 77.64 ( 0.00%) 83.46 ( 7.49%)
As before, massive reduction in external fragmentation events, some jitter
on latencies and an increase in THP allocation success rates.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
4.20-rc3+patch1-4: 14263 (93% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 1346.45 ( 0.00%) 1306.87 ( 2.94%)
Amean fault-huge-5 3418.60 ( 0.00%) 1348.94 ( 60.54%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 0.78 ( 0.00%) 7.91 ( 910.64%)
There is a 93% reduction in fragmentation causing events, there is a big
reduction in the huge page fault latency and allocation success rate is
higher.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
4.20-rc3+patch1-4: 11095 (93% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 6217.43 ( 0.00%) 7419.67 * -19.34%*
Amean fault-huge-5 3163.33 ( 0.00%) 3263.80 ( -3.18%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 95.14 ( 0.00%) 87.98 ( -7.53%)
There is a large reduction in fragmentation events with some jitter around
the latencies and success rates. As before, the high THP allocation
success rate does mean the system is under a lot of pressure. However, as
the fragmentation events are reduced, it would be expected that the
long-term allocation success rate would be higher.
Link: http://lkml.kernel.org/r/20181123114528.28802-5-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:52 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-02-20 22:19:49 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-12-14 19:12:15 -08:00
|
|
|
|
2019-02-20 22:19:49 -08:00
|
|
|
|
mm: reclaim small amounts of memory when an external fragmentation event occurs
An external fragmentation event was previously described as
When the page allocator fragments memory, it records the event using
the mm_page_alloc_extfrag event. If the fallback_order is smaller
than a pageblock order (order-9 on 64-bit x86) then it's considered
an event that will cause external fragmentation issues in the future.
The kernel reduces the probability of such events by increasing the
watermark sizes by calling set_recommended_min_free_kbytes early in the
lifetime of the system. This works reasonably well in general but if
there are enough sparsely populated pageblocks then the problem can still
occur as enough memory is free overall and kswapd stays asleep.
This patch introduces a watermark_boost_factor sysctl that allows a zone
watermark to be temporarily boosted when an external fragmentation causing
events occurs. The boosting will stall allocations that would decrease
free memory below the boosted low watermark and kswapd is woken if the
calling context allows to reclaim an amount of memory relative to the size
of the high watermark and the watermark_boost_factor until the boost is
cleared. When kswapd finishes, it wakes kcompactd at the pageblock order
to clean some of the pageblocks that may have been affected by the
fragmentation event. kswapd avoids any writeback, slab shrinkage and swap
from reclaim context during this operation to avoid excessive system
disruption in the name of fragmentation avoidance. Care is taken so that
kswapd will do normal reclaim work if the system is really low on memory.
This was evaluated using the same workloads as "mm, page_alloc: Spread
allocations across zones before introducing fragmentation".
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
4.20-rc3+patch1-4: 18421 (98% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-1 653.58 ( 0.00%) 652.71 ( 0.13%)
Amean fault-huge-1 0.00 ( 0.00%) 178.93 * -99.00%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 0.00 ( 0.00%) 5.12 ( 100.00%)
Note that external fragmentation causing events are massively reduced by
this path whether in comparison to the previous kernel or the vanilla
kernel. The fault latency for huge pages appears to be increased but that
is only because THP allocations were successful with the patch applied.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
4.20-rc3+patch1-4: 13464 (95% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Min fault-base-1 912.00 ( 0.00%) 905.00 ( 0.77%)
Min fault-huge-1 127.00 ( 0.00%) 135.00 ( -6.30%)
Amean fault-base-1 1467.55 ( 0.00%) 1481.67 ( -0.96%)
Amean fault-huge-1 1127.11 ( 0.00%) 1063.88 * 5.61%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 77.64 ( 0.00%) 83.46 ( 7.49%)
As before, massive reduction in external fragmentation events, some jitter
on latencies and an increase in THP allocation success rates.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
4.20-rc3+patch1-4: 14263 (93% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 1346.45 ( 0.00%) 1306.87 ( 2.94%)
Amean fault-huge-5 3418.60 ( 0.00%) 1348.94 ( 60.54%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 0.78 ( 0.00%) 7.91 ( 910.64%)
There is a 93% reduction in fragmentation causing events, there is a big
reduction in the huge page fault latency and allocation success rate is
higher.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
4.20-rc3+patch1-4: 11095 (93% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 6217.43 ( 0.00%) 7419.67 * -19.34%*
Amean fault-huge-5 3163.33 ( 0.00%) 3263.80 ( -3.18%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 95.14 ( 0.00%) 87.98 ( -7.53%)
There is a large reduction in fragmentation events with some jitter around
the latencies and success rates. As before, the high THP allocation
success rate does mean the system is under a lot of pressure. However, as
the fragmentation events are reduced, it would be expected that the
long-term allocation success rate would be higher.
Link: http://lkml.kernel.org/r/20181123114528.28802-5-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:52 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-12-14 19:12:15 -08:00
|
|
|
|
|
|
|
|
|
mm: reclaim small amounts of memory when an external fragmentation event occurs
An external fragmentation event was previously described as
When the page allocator fragments memory, it records the event using
the mm_page_alloc_extfrag event. If the fallback_order is smaller
than a pageblock order (order-9 on 64-bit x86) then it's considered
an event that will cause external fragmentation issues in the future.
The kernel reduces the probability of such events by increasing the
watermark sizes by calling set_recommended_min_free_kbytes early in the
lifetime of the system. This works reasonably well in general but if
there are enough sparsely populated pageblocks then the problem can still
occur as enough memory is free overall and kswapd stays asleep.
This patch introduces a watermark_boost_factor sysctl that allows a zone
watermark to be temporarily boosted when an external fragmentation causing
events occurs. The boosting will stall allocations that would decrease
free memory below the boosted low watermark and kswapd is woken if the
calling context allows to reclaim an amount of memory relative to the size
of the high watermark and the watermark_boost_factor until the boost is
cleared. When kswapd finishes, it wakes kcompactd at the pageblock order
to clean some of the pageblocks that may have been affected by the
fragmentation event. kswapd avoids any writeback, slab shrinkage and swap
from reclaim context during this operation to avoid excessive system
disruption in the name of fragmentation avoidance. Care is taken so that
kswapd will do normal reclaim work if the system is really low on memory.
This was evaluated using the same workloads as "mm, page_alloc: Spread
allocations across zones before introducing fragmentation".
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
4.20-rc3+patch1-4: 18421 (98% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-1 653.58 ( 0.00%) 652.71 ( 0.13%)
Amean fault-huge-1 0.00 ( 0.00%) 178.93 * -99.00%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 0.00 ( 0.00%) 5.12 ( 100.00%)
Note that external fragmentation causing events are massively reduced by
this path whether in comparison to the previous kernel or the vanilla
kernel. The fault latency for huge pages appears to be increased but that
is only because THP allocations were successful with the patch applied.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
4.20-rc3+patch1-4: 13464 (95% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Min fault-base-1 912.00 ( 0.00%) 905.00 ( 0.77%)
Min fault-huge-1 127.00 ( 0.00%) 135.00 ( -6.30%)
Amean fault-base-1 1467.55 ( 0.00%) 1481.67 ( -0.96%)
Amean fault-huge-1 1127.11 ( 0.00%) 1063.88 * 5.61%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 77.64 ( 0.00%) 83.46 ( 7.49%)
As before, massive reduction in external fragmentation events, some jitter
on latencies and an increase in THP allocation success rates.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
4.20-rc3+patch1-4: 14263 (93% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 1346.45 ( 0.00%) 1306.87 ( 2.94%)
Amean fault-huge-5 3418.60 ( 0.00%) 1348.94 ( 60.54%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 0.78 ( 0.00%) 7.91 ( 910.64%)
There is a 93% reduction in fragmentation causing events, there is a big
reduction in the huge page fault latency and allocation success rate is
higher.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
4.20-rc3+patch1-4: 11095 (93% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 6217.43 ( 0.00%) 7419.67 * -19.34%*
Amean fault-huge-5 3163.33 ( 0.00%) 3263.80 ( -3.18%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 95.14 ( 0.00%) 87.98 ( -7.53%)
There is a large reduction in fragmentation events with some jitter around
the latencies and success rates. As before, the high THP allocation
success rate does mean the system is under a lot of pressure. However, as
the fragmentation events are reduced, it would be expected that the
long-term allocation success rate would be higher.
Link: http://lkml.kernel.org/r/20181123114528.28802-5-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:52 -08:00
|
|
|
|
|
|
|
|
|
2015-04-14 15:45:18 -07:00
|
|
|
|
2025-02-28 09:52:18 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/page_allo.c: restructure free-page stealing code and fix a bug
The free-page stealing code in __rmqueue_fallback() is somewhat hard to
follow, and has an incredible amount of subtlety hidden inside!
First off, there is a minor bug in the reporting of change-of-ownership of
pageblocks. Under some conditions, we try to move upto
'pageblock_nr_pages' no. of pages to the preferred allocation list. But
we change the ownership of that pageblock to the preferred type only if we
manage to successfully move atleast half of that pageblock (or if
page_group_by_mobility_disabled is set).
However, the current code ignores the latter part and sets the
'migratetype' variable to the preferred type, irrespective of whether we
actually changed the pageblock migratetype of that block or not. So, the
page_alloc_extfrag tracepoint can end up printing incorrect info (i.e.,
'change_ownership' might be shown as 1 when it must have been 0).
So fixing this involves moving the update of the 'migratetype' variable to
the right place. But looking closer, we observe that the 'migratetype'
variable is used subsequently for checks such as "is_migrate_cma()".
Obviously the intent there is to check if the *fallback* type is
MIGRATE_CMA, but since we already set the 'migratetype' variable to
start_migratetype, we end up checking if the *preferred* type is
MIGRATE_CMA!!
To make things more interesting, this actually doesn't cause a bug in
practice, because we never change *anything* if the fallback type is CMA.
So, restructure the code in such a way that it is trivial to understand
what is going on, and also fix the above mentioned bug. And while at it,
also add a comment explaining the subtlety behind the migratetype used in
the call to expand().
[akpm@linux-foundation.org: remove unneeded `inline', small coding-style fix]
Signed-off-by: Srivatsa S. Bhat <srivatsa.bhat@linux.vnet.ibm.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Cody P Schafer <cody@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 14:20:35 -07:00
|
|
|
|
2025-02-28 09:52:17 +00:00
|
|
|
|
2015-04-14 15:45:18 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-02-28 09:52:17 +00:00
|
|
|
|
2015-04-14 15:45:18 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-09-20 20:20:30 +08:00
|
|
|
|
2025-02-28 09:52:18 +00:00
|
|
|
|
|
|
|
|
|
2024-09-20 20:20:30 +08:00
|
|
|
|
2025-02-28 09:52:18 +00:00
|
|
|
|
2015-04-14 15:45:18 -07:00
|
|
|
|
|
|
|
|
|
2025-02-28 09:52:18 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-04-14 15:45:18 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-02-24 19:08:26 -05:00
|
|
|
|
|
|
|
|
|
mm: page_alloc: tighten up find_suitable_fallback()
find_suitable_fallback() is not as efficient as it could be, and somewhat
difficult to follow.
1. should_try_claim_block() is a loop invariant. There is no point in
checking fallback areas if the caller is interested in claimable
blocks but the order and the migratetype don't allow for that.
2. __rmqueue_steal() doesn't care about claimability, so it shouldn't
have to run those tests.
Different callers want different things from this helper:
1. __compact_finished() scans orders up until it finds a claimable block
2. __rmqueue_claim() scans orders down as long as blocks are claimable
3. __rmqueue_steal() doesn't care about claimability at all
Move should_try_claim_block() out of the loop. Only test it for the
two callers who care in the first place. Distinguish "no blocks" from
"order + mt are not claimable" in the return value; __rmqueue_claim()
can stop once order becomes unclaimable, __compact_finished() can keep
advancing until order becomes claimable.
Before:
Performance counter stats for './run case-lru-file-mmap-read' (5 runs):
85,294.85 msec task-clock # 5.644 CPUs utilized ( +- 0.32% )
15,968 context-switches # 187.209 /sec ( +- 3.81% )
153 cpu-migrations # 1.794 /sec ( +- 3.29% )
801,808 page-faults # 9.400 K/sec ( +- 0.10% )
733,358,331,786 instructions # 1.87 insn per cycle ( +- 0.20% ) (64.94%)
392,622,904,199 cycles # 4.603 GHz ( +- 0.31% ) (64.84%)
148,563,488,531 branches # 1.742 G/sec ( +- 0.18% ) (63.86%)
152,143,228 branch-misses # 0.10% of all branches ( +- 1.19% ) (62.82%)
15.1128 +- 0.0637 seconds time elapsed ( +- 0.42% )
After:
Performance counter stats for './run case-lru-file-mmap-read' (5 runs):
84,380.21 msec task-clock # 5.664 CPUs utilized ( +- 0.21% )
16,656 context-switches # 197.392 /sec ( +- 3.27% )
151 cpu-migrations # 1.790 /sec ( +- 3.28% )
801,703 page-faults # 9.501 K/sec ( +- 0.09% )
731,914,183,060 instructions # 1.88 insn per cycle ( +- 0.38% ) (64.90%)
388,673,535,116 cycles # 4.606 GHz ( +- 0.24% ) (65.06%)
148,251,482,143 branches # 1.757 G/sec ( +- 0.37% ) (63.92%)
149,766,550 branch-misses # 0.10% of all branches ( +- 1.22% ) (62.88%)
14.8968 +- 0.0486 seconds time elapsed ( +- 0.33% )
Link: https://lkml.kernel.org/r/20250407180154.63348-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Tested-by: Shivank Garg <shivankg@amd.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Carlos Song <carlos.song@nxp.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-07 14:01:54 -04:00
|
|
|
|
2025-02-28 09:52:17 +00:00
|
|
|
|
2025-02-24 19:08:26 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: page_alloc: tighten up find_suitable_fallback()
find_suitable_fallback() is not as efficient as it could be, and somewhat
difficult to follow.
1. should_try_claim_block() is a loop invariant. There is no point in
checking fallback areas if the caller is interested in claimable
blocks but the order and the migratetype don't allow for that.
2. __rmqueue_steal() doesn't care about claimability, so it shouldn't
have to run those tests.
Different callers want different things from this helper:
1. __compact_finished() scans orders up until it finds a claimable block
2. __rmqueue_claim() scans orders down as long as blocks are claimable
3. __rmqueue_steal() doesn't care about claimability at all
Move should_try_claim_block() out of the loop. Only test it for the
two callers who care in the first place. Distinguish "no blocks" from
"order + mt are not claimable" in the return value; __rmqueue_claim()
can stop once order becomes unclaimable, __compact_finished() can keep
advancing until order becomes claimable.
Before:
Performance counter stats for './run case-lru-file-mmap-read' (5 runs):
85,294.85 msec task-clock # 5.644 CPUs utilized ( +- 0.32% )
15,968 context-switches # 187.209 /sec ( +- 3.81% )
153 cpu-migrations # 1.794 /sec ( +- 3.29% )
801,808 page-faults # 9.400 K/sec ( +- 0.10% )
733,358,331,786 instructions # 1.87 insn per cycle ( +- 0.20% ) (64.94%)
392,622,904,199 cycles # 4.603 GHz ( +- 0.31% ) (64.84%)
148,563,488,531 branches # 1.742 G/sec ( +- 0.18% ) (63.86%)
152,143,228 branch-misses # 0.10% of all branches ( +- 1.19% ) (62.82%)
15.1128 +- 0.0637 seconds time elapsed ( +- 0.42% )
After:
Performance counter stats for './run case-lru-file-mmap-read' (5 runs):
84,380.21 msec task-clock # 5.664 CPUs utilized ( +- 0.21% )
16,656 context-switches # 197.392 /sec ( +- 3.27% )
151 cpu-migrations # 1.790 /sec ( +- 3.28% )
801,703 page-faults # 9.501 K/sec ( +- 0.09% )
731,914,183,060 instructions # 1.88 insn per cycle ( +- 0.38% ) (64.90%)
388,673,535,116 cycles # 4.606 GHz ( +- 0.24% ) (65.06%)
148,251,482,143 branches # 1.757 G/sec ( +- 0.37% ) (63.92%)
149,766,550 branch-misses # 0.10% of all branches ( +- 1.22% ) (62.88%)
14.8968 +- 0.0486 seconds time elapsed ( +- 0.33% )
Link: https://lkml.kernel.org/r/20250407180154.63348-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Tested-by: Shivank Garg <shivankg@amd.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Carlos Song <carlos.song@nxp.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-07 14:01:54 -04:00
|
|
|
|
mm: reclaim small amounts of memory when an external fragmentation event occurs
An external fragmentation event was previously described as
When the page allocator fragments memory, it records the event using
the mm_page_alloc_extfrag event. If the fallback_order is smaller
than a pageblock order (order-9 on 64-bit x86) then it's considered
an event that will cause external fragmentation issues in the future.
The kernel reduces the probability of such events by increasing the
watermark sizes by calling set_recommended_min_free_kbytes early in the
lifetime of the system. This works reasonably well in general but if
there are enough sparsely populated pageblocks then the problem can still
occur as enough memory is free overall and kswapd stays asleep.
This patch introduces a watermark_boost_factor sysctl that allows a zone
watermark to be temporarily boosted when an external fragmentation causing
events occurs. The boosting will stall allocations that would decrease
free memory below the boosted low watermark and kswapd is woken if the
calling context allows to reclaim an amount of memory relative to the size
of the high watermark and the watermark_boost_factor until the boost is
cleared. When kswapd finishes, it wakes kcompactd at the pageblock order
to clean some of the pageblocks that may have been affected by the
fragmentation event. kswapd avoids any writeback, slab shrinkage and swap
from reclaim context during this operation to avoid excessive system
disruption in the name of fragmentation avoidance. Care is taken so that
kswapd will do normal reclaim work if the system is really low on memory.
This was evaluated using the same workloads as "mm, page_alloc: Spread
allocations across zones before introducing fragmentation".
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
4.20-rc3+patch1-4: 18421 (98% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-1 653.58 ( 0.00%) 652.71 ( 0.13%)
Amean fault-huge-1 0.00 ( 0.00%) 178.93 * -99.00%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 0.00 ( 0.00%) 5.12 ( 100.00%)
Note that external fragmentation causing events are massively reduced by
this path whether in comparison to the previous kernel or the vanilla
kernel. The fault latency for huge pages appears to be increased but that
is only because THP allocations were successful with the patch applied.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
4.20-rc3+patch1-4: 13464 (95% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Min fault-base-1 912.00 ( 0.00%) 905.00 ( 0.77%)
Min fault-huge-1 127.00 ( 0.00%) 135.00 ( -6.30%)
Amean fault-base-1 1467.55 ( 0.00%) 1481.67 ( -0.96%)
Amean fault-huge-1 1127.11 ( 0.00%) 1063.88 * 5.61%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 77.64 ( 0.00%) 83.46 ( 7.49%)
As before, massive reduction in external fragmentation events, some jitter
on latencies and an increase in THP allocation success rates.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
4.20-rc3+patch1-4: 14263 (93% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 1346.45 ( 0.00%) 1306.87 ( 2.94%)
Amean fault-huge-5 3418.60 ( 0.00%) 1348.94 ( 60.54%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 0.78 ( 0.00%) 7.91 ( 910.64%)
There is a 93% reduction in fragmentation causing events, there is a big
reduction in the huge page fault latency and allocation success rate is
higher.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
4.20-rc3+patch1-4: 11095 (93% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 6217.43 ( 0.00%) 7419.67 * -19.34%*
Amean fault-huge-5 3163.33 ( 0.00%) 3263.80 ( -3.18%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 95.14 ( 0.00%) 87.98 ( -7.53%)
There is a large reduction in fragmentation events with some jitter around
the latencies and success rates. As before, the high THP allocation
success rate does mean the system is under a lot of pressure. However, as
the fragmentation events are reduced, it would be expected that the
long-term allocation success rate would be higher.
Link: http://lkml.kernel.org/r/20181123114528.28802-5-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:52 -08:00
|
|
|
|
2025-02-24 19:08:26 -05:00
|
|
|
|
mm: page_alloc: tighten up find_suitable_fallback()
find_suitable_fallback() is not as efficient as it could be, and somewhat
difficult to follow.
1. should_try_claim_block() is a loop invariant. There is no point in
checking fallback areas if the caller is interested in claimable
blocks but the order and the migratetype don't allow for that.
2. __rmqueue_steal() doesn't care about claimability, so it shouldn't
have to run those tests.
Different callers want different things from this helper:
1. __compact_finished() scans orders up until it finds a claimable block
2. __rmqueue_claim() scans orders down as long as blocks are claimable
3. __rmqueue_steal() doesn't care about claimability at all
Move should_try_claim_block() out of the loop. Only test it for the
two callers who care in the first place. Distinguish "no blocks" from
"order + mt are not claimable" in the return value; __rmqueue_claim()
can stop once order becomes unclaimable, __compact_finished() can keep
advancing until order becomes claimable.
Before:
Performance counter stats for './run case-lru-file-mmap-read' (5 runs):
85,294.85 msec task-clock # 5.644 CPUs utilized ( +- 0.32% )
15,968 context-switches # 187.209 /sec ( +- 3.81% )
153 cpu-migrations # 1.794 /sec ( +- 3.29% )
801,808 page-faults # 9.400 K/sec ( +- 0.10% )
733,358,331,786 instructions # 1.87 insn per cycle ( +- 0.20% ) (64.94%)
392,622,904,199 cycles # 4.603 GHz ( +- 0.31% ) (64.84%)
148,563,488,531 branches # 1.742 G/sec ( +- 0.18% ) (63.86%)
152,143,228 branch-misses # 0.10% of all branches ( +- 1.19% ) (62.82%)
15.1128 +- 0.0637 seconds time elapsed ( +- 0.42% )
After:
Performance counter stats for './run case-lru-file-mmap-read' (5 runs):
84,380.21 msec task-clock # 5.664 CPUs utilized ( +- 0.21% )
16,656 context-switches # 197.392 /sec ( +- 3.27% )
151 cpu-migrations # 1.790 /sec ( +- 3.28% )
801,703 page-faults # 9.501 K/sec ( +- 0.09% )
731,914,183,060 instructions # 1.88 insn per cycle ( +- 0.38% ) (64.90%)
388,673,535,116 cycles # 4.606 GHz ( +- 0.24% ) (65.06%)
148,251,482,143 branches # 1.757 G/sec ( +- 0.37% ) (63.92%)
149,766,550 branch-misses # 0.10% of all branches ( +- 1.22% ) (62.88%)
14.8968 +- 0.0486 seconds time elapsed ( +- 0.33% )
Link: https://lkml.kernel.org/r/20250407180154.63348-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Tested-by: Shivank Garg <shivankg@amd.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Carlos Song <carlos.song@nxp.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-07 14:01:54 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: reclaim small amounts of memory when an external fragmentation event occurs
An external fragmentation event was previously described as
When the page allocator fragments memory, it records the event using
the mm_page_alloc_extfrag event. If the fallback_order is smaller
than a pageblock order (order-9 on 64-bit x86) then it's considered
an event that will cause external fragmentation issues in the future.
The kernel reduces the probability of such events by increasing the
watermark sizes by calling set_recommended_min_free_kbytes early in the
lifetime of the system. This works reasonably well in general but if
there are enough sparsely populated pageblocks then the problem can still
occur as enough memory is free overall and kswapd stays asleep.
This patch introduces a watermark_boost_factor sysctl that allows a zone
watermark to be temporarily boosted when an external fragmentation causing
events occurs. The boosting will stall allocations that would decrease
free memory below the boosted low watermark and kswapd is woken if the
calling context allows to reclaim an amount of memory relative to the size
of the high watermark and the watermark_boost_factor until the boost is
cleared. When kswapd finishes, it wakes kcompactd at the pageblock order
to clean some of the pageblocks that may have been affected by the
fragmentation event. kswapd avoids any writeback, slab shrinkage and swap
from reclaim context during this operation to avoid excessive system
disruption in the name of fragmentation avoidance. Care is taken so that
kswapd will do normal reclaim work if the system is really low on memory.
This was evaluated using the same workloads as "mm, page_alloc: Spread
allocations across zones before introducing fragmentation".
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
4.20-rc3+patch1-4: 18421 (98% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-1 653.58 ( 0.00%) 652.71 ( 0.13%)
Amean fault-huge-1 0.00 ( 0.00%) 178.93 * -99.00%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 0.00 ( 0.00%) 5.12 ( 100.00%)
Note that external fragmentation causing events are massively reduced by
this path whether in comparison to the previous kernel or the vanilla
kernel. The fault latency for huge pages appears to be increased but that
is only because THP allocations were successful with the patch applied.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
4.20-rc3+patch1-4: 13464 (95% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Min fault-base-1 912.00 ( 0.00%) 905.00 ( 0.77%)
Min fault-huge-1 127.00 ( 0.00%) 135.00 ( -6.30%)
Amean fault-base-1 1467.55 ( 0.00%) 1481.67 ( -0.96%)
Amean fault-huge-1 1127.11 ( 0.00%) 1063.88 * 5.61%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 77.64 ( 0.00%) 83.46 ( 7.49%)
As before, massive reduction in external fragmentation events, some jitter
on latencies and an increase in THP allocation success rates.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
4.20-rc3+patch1-4: 14263 (93% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 1346.45 ( 0.00%) 1306.87 ( 2.94%)
Amean fault-huge-5 3418.60 ( 0.00%) 1348.94 ( 60.54%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 0.78 ( 0.00%) 7.91 ( 910.64%)
There is a 93% reduction in fragmentation causing events, there is a big
reduction in the huge page fault latency and allocation success rate is
higher.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
4.20-rc3+patch1-4: 11095 (93% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 6217.43 ( 0.00%) 7419.67 * -19.34%*
Amean fault-huge-5 3163.33 ( 0.00%) 3263.80 ( -3.18%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 95.14 ( 0.00%) 87.98 ( -7.53%)
There is a large reduction in fragmentation events with some jitter around
the latencies and success rates. As before, the high THP allocation
success rate does mean the system is under a lot of pressure. However, as
the fragmentation events are reduced, it would be expected that the
long-term allocation success rate would be higher.
Link: http://lkml.kernel.org/r/20181123114528.28802-5-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:52 -08:00
|
|
|
|
2025-02-24 19:08:26 -05:00
|
|
|
|
|
|
|
|
|
mm: reclaim small amounts of memory when an external fragmentation event occurs
An external fragmentation event was previously described as
When the page allocator fragments memory, it records the event using
the mm_page_alloc_extfrag event. If the fallback_order is smaller
than a pageblock order (order-9 on 64-bit x86) then it's considered
an event that will cause external fragmentation issues in the future.
The kernel reduces the probability of such events by increasing the
watermark sizes by calling set_recommended_min_free_kbytes early in the
lifetime of the system. This works reasonably well in general but if
there are enough sparsely populated pageblocks then the problem can still
occur as enough memory is free overall and kswapd stays asleep.
This patch introduces a watermark_boost_factor sysctl that allows a zone
watermark to be temporarily boosted when an external fragmentation causing
events occurs. The boosting will stall allocations that would decrease
free memory below the boosted low watermark and kswapd is woken if the
calling context allows to reclaim an amount of memory relative to the size
of the high watermark and the watermark_boost_factor until the boost is
cleared. When kswapd finishes, it wakes kcompactd at the pageblock order
to clean some of the pageblocks that may have been affected by the
fragmentation event. kswapd avoids any writeback, slab shrinkage and swap
from reclaim context during this operation to avoid excessive system
disruption in the name of fragmentation avoidance. Care is taken so that
kswapd will do normal reclaim work if the system is really low on memory.
This was evaluated using the same workloads as "mm, page_alloc: Spread
allocations across zones before introducing fragmentation".
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
4.20-rc3+patch1-4: 18421 (98% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-1 653.58 ( 0.00%) 652.71 ( 0.13%)
Amean fault-huge-1 0.00 ( 0.00%) 178.93 * -99.00%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 0.00 ( 0.00%) 5.12 ( 100.00%)
Note that external fragmentation causing events are massively reduced by
this path whether in comparison to the previous kernel or the vanilla
kernel. The fault latency for huge pages appears to be increased but that
is only because THP allocations were successful with the patch applied.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
4.20-rc3+patch1-4: 13464 (95% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Min fault-base-1 912.00 ( 0.00%) 905.00 ( 0.77%)
Min fault-huge-1 127.00 ( 0.00%) 135.00 ( -6.30%)
Amean fault-base-1 1467.55 ( 0.00%) 1481.67 ( -0.96%)
Amean fault-huge-1 1127.11 ( 0.00%) 1063.88 * 5.61%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 77.64 ( 0.00%) 83.46 ( 7.49%)
As before, massive reduction in external fragmentation events, some jitter
on latencies and an increase in THP allocation success rates.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
4.20-rc3+patch1-4: 14263 (93% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 1346.45 ( 0.00%) 1306.87 ( 2.94%)
Amean fault-huge-5 3418.60 ( 0.00%) 1348.94 ( 60.54%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 0.78 ( 0.00%) 7.91 ( 910.64%)
There is a 93% reduction in fragmentation causing events, there is a big
reduction in the huge page fault latency and allocation success rate is
higher.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
4.20-rc3+patch1-4: 11095 (93% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 6217.43 ( 0.00%) 7419.67 * -19.34%*
Amean fault-huge-5 3163.33 ( 0.00%) 3263.80 ( -3.18%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 95.14 ( 0.00%) 87.98 ( -7.53%)
There is a large reduction in fragmentation events with some jitter around
the latencies and success rates. As before, the high THP allocation
success rate does mean the system is under a lot of pressure. However, as
the fragmentation events are reduced, it would be expected that the
long-term allocation success rate would be higher.
Link: http://lkml.kernel.org/r/20181123114528.28802-5-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:52 -08:00
|
|
|
|
2025-02-24 19:08:26 -05:00
|
|
|
|
mm: page_alloc: tighten up find_suitable_fallback()
find_suitable_fallback() is not as efficient as it could be, and somewhat
difficult to follow.
1. should_try_claim_block() is a loop invariant. There is no point in
checking fallback areas if the caller is interested in claimable
blocks but the order and the migratetype don't allow for that.
2. __rmqueue_steal() doesn't care about claimability, so it shouldn't
have to run those tests.
Different callers want different things from this helper:
1. __compact_finished() scans orders up until it finds a claimable block
2. __rmqueue_claim() scans orders down as long as blocks are claimable
3. __rmqueue_steal() doesn't care about claimability at all
Move should_try_claim_block() out of the loop. Only test it for the
two callers who care in the first place. Distinguish "no blocks" from
"order + mt are not claimable" in the return value; __rmqueue_claim()
can stop once order becomes unclaimable, __compact_finished() can keep
advancing until order becomes claimable.
Before:
Performance counter stats for './run case-lru-file-mmap-read' (5 runs):
85,294.85 msec task-clock # 5.644 CPUs utilized ( +- 0.32% )
15,968 context-switches # 187.209 /sec ( +- 3.81% )
153 cpu-migrations # 1.794 /sec ( +- 3.29% )
801,808 page-faults # 9.400 K/sec ( +- 0.10% )
733,358,331,786 instructions # 1.87 insn per cycle ( +- 0.20% ) (64.94%)
392,622,904,199 cycles # 4.603 GHz ( +- 0.31% ) (64.84%)
148,563,488,531 branches # 1.742 G/sec ( +- 0.18% ) (63.86%)
152,143,228 branch-misses # 0.10% of all branches ( +- 1.19% ) (62.82%)
15.1128 +- 0.0637 seconds time elapsed ( +- 0.42% )
After:
Performance counter stats for './run case-lru-file-mmap-read' (5 runs):
84,380.21 msec task-clock # 5.664 CPUs utilized ( +- 0.21% )
16,656 context-switches # 197.392 /sec ( +- 3.27% )
151 cpu-migrations # 1.790 /sec ( +- 3.28% )
801,703 page-faults # 9.501 K/sec ( +- 0.09% )
731,914,183,060 instructions # 1.88 insn per cycle ( +- 0.38% ) (64.90%)
388,673,535,116 cycles # 4.606 GHz ( +- 0.24% ) (65.06%)
148,251,482,143 branches # 1.757 G/sec ( +- 0.37% ) (63.92%)
149,766,550 branch-misses # 0.10% of all branches ( +- 1.22% ) (62.88%)
14.8968 +- 0.0486 seconds time elapsed ( +- 0.33% )
Link: https://lkml.kernel.org/r/20250407180154.63348-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Tested-by: Shivank Garg <shivankg@amd.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Carlos Song <carlos.song@nxp.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-07 14:01:54 -04:00
|
|
|
|
2019-02-20 22:19:49 -08:00
|
|
|
|
mm: page_alloc: tighten up find_suitable_fallback()
find_suitable_fallback() is not as efficient as it could be, and somewhat
difficult to follow.
1. should_try_claim_block() is a loop invariant. There is no point in
checking fallback areas if the caller is interested in claimable
blocks but the order and the migratetype don't allow for that.
2. __rmqueue_steal() doesn't care about claimability, so it shouldn't
have to run those tests.
Different callers want different things from this helper:
1. __compact_finished() scans orders up until it finds a claimable block
2. __rmqueue_claim() scans orders down as long as blocks are claimable
3. __rmqueue_steal() doesn't care about claimability at all
Move should_try_claim_block() out of the loop. Only test it for the
two callers who care in the first place. Distinguish "no blocks" from
"order + mt are not claimable" in the return value; __rmqueue_claim()
can stop once order becomes unclaimable, __compact_finished() can keep
advancing until order becomes claimable.
Before:
Performance counter stats for './run case-lru-file-mmap-read' (5 runs):
85,294.85 msec task-clock # 5.644 CPUs utilized ( +- 0.32% )
15,968 context-switches # 187.209 /sec ( +- 3.81% )
153 cpu-migrations # 1.794 /sec ( +- 3.29% )
801,808 page-faults # 9.400 K/sec ( +- 0.10% )
733,358,331,786 instructions # 1.87 insn per cycle ( +- 0.20% ) (64.94%)
392,622,904,199 cycles # 4.603 GHz ( +- 0.31% ) (64.84%)
148,563,488,531 branches # 1.742 G/sec ( +- 0.18% ) (63.86%)
152,143,228 branch-misses # 0.10% of all branches ( +- 1.19% ) (62.82%)
15.1128 +- 0.0637 seconds time elapsed ( +- 0.42% )
After:
Performance counter stats for './run case-lru-file-mmap-read' (5 runs):
84,380.21 msec task-clock # 5.664 CPUs utilized ( +- 0.21% )
16,656 context-switches # 197.392 /sec ( +- 3.27% )
151 cpu-migrations # 1.790 /sec ( +- 3.28% )
801,703 page-faults # 9.501 K/sec ( +- 0.09% )
731,914,183,060 instructions # 1.88 insn per cycle ( +- 0.38% ) (64.90%)
388,673,535,116 cycles # 4.606 GHz ( +- 0.24% ) (65.06%)
148,251,482,143 branches # 1.757 G/sec ( +- 0.37% ) (63.92%)
149,766,550 branch-misses # 0.10% of all branches ( +- 1.22% ) (62.88%)
14.8968 +- 0.0486 seconds time elapsed ( +- 0.33% )
Link: https://lkml.kernel.org/r/20250407180154.63348-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Tested-by: Shivank Garg <shivankg@amd.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Carlos Song <carlos.song@nxp.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-07 14:01:54 -04:00
|
|
|
|
2025-02-24 19:08:26 -05:00
|
|
|
|
|
|
|
|
|
2020-12-14 19:12:15 -08:00
|
|
|
|
2025-02-24 19:08:26 -05:00
|
|
|
|
mm: reclaim small amounts of memory when an external fragmentation event occurs
An external fragmentation event was previously described as
When the page allocator fragments memory, it records the event using
the mm_page_alloc_extfrag event. If the fallback_order is smaller
than a pageblock order (order-9 on 64-bit x86) then it's considered
an event that will cause external fragmentation issues in the future.
The kernel reduces the probability of such events by increasing the
watermark sizes by calling set_recommended_min_free_kbytes early in the
lifetime of the system. This works reasonably well in general but if
there are enough sparsely populated pageblocks then the problem can still
occur as enough memory is free overall and kswapd stays asleep.
This patch introduces a watermark_boost_factor sysctl that allows a zone
watermark to be temporarily boosted when an external fragmentation causing
events occurs. The boosting will stall allocations that would decrease
free memory below the boosted low watermark and kswapd is woken if the
calling context allows to reclaim an amount of memory relative to the size
of the high watermark and the watermark_boost_factor until the boost is
cleared. When kswapd finishes, it wakes kcompactd at the pageblock order
to clean some of the pageblocks that may have been affected by the
fragmentation event. kswapd avoids any writeback, slab shrinkage and swap
from reclaim context during this operation to avoid excessive system
disruption in the name of fragmentation avoidance. Care is taken so that
kswapd will do normal reclaim work if the system is really low on memory.
This was evaluated using the same workloads as "mm, page_alloc: Spread
allocations across zones before introducing fragmentation".
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
4.20-rc3+patch1-4: 18421 (98% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-1 653.58 ( 0.00%) 652.71 ( 0.13%)
Amean fault-huge-1 0.00 ( 0.00%) 178.93 * -99.00%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 0.00 ( 0.00%) 5.12 ( 100.00%)
Note that external fragmentation causing events are massively reduced by
this path whether in comparison to the previous kernel or the vanilla
kernel. The fault latency for huge pages appears to be increased but that
is only because THP allocations were successful with the patch applied.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
4.20-rc3+patch1-4: 13464 (95% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Min fault-base-1 912.00 ( 0.00%) 905.00 ( 0.77%)
Min fault-huge-1 127.00 ( 0.00%) 135.00 ( -6.30%)
Amean fault-base-1 1467.55 ( 0.00%) 1481.67 ( -0.96%)
Amean fault-huge-1 1127.11 ( 0.00%) 1063.88 * 5.61%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 77.64 ( 0.00%) 83.46 ( 7.49%)
As before, massive reduction in external fragmentation events, some jitter
on latencies and an increase in THP allocation success rates.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
4.20-rc3+patch1-4: 14263 (93% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 1346.45 ( 0.00%) 1306.87 ( 2.94%)
Amean fault-huge-5 3418.60 ( 0.00%) 1348.94 ( 60.54%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 0.78 ( 0.00%) 7.91 ( 910.64%)
There is a 93% reduction in fragmentation causing events, there is a big
reduction in the huge page fault latency and allocation success rate is
higher.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
4.20-rc3+patch1-4: 11095 (93% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 6217.43 ( 0.00%) 7419.67 * -19.34%*
Amean fault-huge-5 3163.33 ( 0.00%) 3263.80 ( -3.18%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 95.14 ( 0.00%) 87.98 ( -7.53%)
There is a large reduction in fragmentation events with some jitter around
the latencies and success rates. As before, the high THP allocation
success rate does mean the system is under a lot of pressure. However, as
the fragmentation events are reduced, it would be expected that the
long-term allocation success rate would be higher.
Link: http://lkml.kernel.org/r/20181123114528.28802-5-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:52 -08:00
|
|
|
|
|
|
|
|
|
2015-04-14 15:45:18 -07:00
|
|
|
|
2025-02-28 09:52:17 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-04-14 15:45:18 -07:00
|
|
|
|
mm: page_alloc: fix freelist movement during block conversion
Currently, page block type conversion during fallbacks, atomic
reservations and isolation can strand various amounts of free pages on
incorrect freelists.
For example, fallback stealing moves free pages in the block to the new
type's freelists, but then may not actually claim the block for that type
if there aren't enough compatible pages already allocated.
In all cases, free page moving might fail if the block straddles more than
one zone, in which case no free pages are moved at all, but the block type
is changed anyway.
This is detrimental to type hygiene on the freelists. It encourages
incompatible page mixing down the line (ask for one type, get another) and
thus contributes to long-term fragmentation.
Split the process into a proper transaction: check first if conversion
will happen, then try to move the free pages, and only if that was
successful convert the block to the new type.
[baolin.wang@linux.alibaba.com: fix allocation failures with CONFIG_CMA]
Link: https://lkml.kernel.org/r/a97697e0-45b0-4f71-b087-fdc7a1d43c0e@linux.alibaba.com
Link: https://lkml.kernel.org/r/20240320180429.678181-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Tested-by: "Huang, Ying" <ying.huang@intel.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-20 14:02:11 -04:00
|
|
|
|
2025-02-28 09:52:17 +00:00
|
|
|
|
mm: page_alloc: don't steal single pages from biggest buddy
The fallback code searches for the biggest buddy first in an attempt to
steal the whole block and encourage type grouping down the line.
The approach used to be this:
- Non-movable requests will split the largest buddy and steal the
remainder. This splits up contiguity, but it allows subsequent
requests of this type to fall back into adjacent space.
- Movable requests go and look for the smallest buddy instead. The
thinking is that movable requests can be compacted, so grouping is
less important than retaining contiguity.
c0cd6f557b90 ("mm: page_alloc: fix freelist movement during block
conversion") enforces freelist type hygiene, which restricts stealing to
either claiming the whole block or just taking the requested chunk; no
additional pages or buddy remainders can be stolen any more.
The patch mishandled when to switch to finding the smallest buddy in that
new reality. As a result, it may steal the exact request size, but from
the biggest buddy. This causes fracturing for no good reason.
Fix this by committing to the new behavior: either steal the whole block,
or fall back to the smallest buddy.
Remove single-page stealing from steal_suitable_fallback(). Rename it to
try_to_steal_block() to make the intentions clear. If this fails, always
fall back to the smallest buddy.
The following is from 4 runs of mmtest's thpchallenge. "Pollute" is
single page fallback, "steal" is conversion of a partially used block.
The numbers for free block conversions (omitted) are comparable.
vanilla patched
@pollute[unmovable from reclaimable]: 27 106
@pollute[unmovable from movable]: 82 46
@pollute[reclaimable from unmovable]: 256 83
@pollute[reclaimable from movable]: 46 8
@pollute[movable from unmovable]: 4841 868
@pollute[movable from reclaimable]: 5278 12568
@steal[unmovable from reclaimable]: 11 12
@steal[unmovable from movable]: 113 49
@steal[reclaimable from unmovable]: 19 34
@steal[reclaimable from movable]: 47 21
@steal[movable from unmovable]: 250 183
@steal[movable from reclaimable]: 81 93
The allocator appears to do a better job at keeping stealing and polluting
to the first fallback preference. As a result, the numbers for "from
movable" - the least preferred fallback option, and most detrimental to
compactability - are down across the board.
Link: https://lkml.kernel.org/r/20250225001023.1494422-2-hannes@cmpxchg.org
Fixes: c0cd6f557b90 ("mm: page_alloc: fix freelist movement during block conversion")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-02-24 19:08:24 -05:00
|
|
|
|
2025-02-24 19:08:25 -05:00
|
|
|
|
mm/page_allo.c: restructure free-page stealing code and fix a bug
The free-page stealing code in __rmqueue_fallback() is somewhat hard to
follow, and has an incredible amount of subtlety hidden inside!
First off, there is a minor bug in the reporting of change-of-ownership of
pageblocks. Under some conditions, we try to move upto
'pageblock_nr_pages' no. of pages to the preferred allocation list. But
we change the ownership of that pageblock to the preferred type only if we
manage to successfully move atleast half of that pageblock (or if
page_group_by_mobility_disabled is set).
However, the current code ignores the latter part and sets the
'migratetype' variable to the preferred type, irrespective of whether we
actually changed the pageblock migratetype of that block or not. So, the
page_alloc_extfrag tracepoint can end up printing incorrect info (i.e.,
'change_ownership' might be shown as 1 when it must have been 0).
So fixing this involves moving the update of the 'migratetype' variable to
the right place. But looking closer, we observe that the 'migratetype'
variable is used subsequently for checks such as "is_migrate_cma()".
Obviously the intent there is to check if the *fallback* type is
MIGRATE_CMA, but since we already set the 'migratetype' variable to
start_migratetype, we end up checking if the *preferred* type is
MIGRATE_CMA!!
To make things more interesting, this actually doesn't cause a bug in
practice, because we never change *anything* if the fallback type is CMA.
So, restructure the code in such a way that it is trivial to understand
what is going on, and also fix the above mentioned bug. And while at it,
also add a comment explaining the subtlety behind the migratetype used in
the call to expand().
[akpm@linux-foundation.org: remove unneeded `inline', small coding-style fix]
Signed-off-by: Srivatsa S. Bhat <srivatsa.bhat@linux.vnet.ibm.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Cody P Schafer <cody@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 14:20:35 -07:00
|
|
|
|
mm, page_alloc: count movable pages when stealing from pageblock
When stealing pages from pageblock of a different migratetype, we count
how many free pages were stolen, and change the pageblock's migratetype
if more than half of the pageblock was free. This might be too
conservative, as there might be other pages that are not free, but were
allocated with the same migratetype as our allocation requested.
While we cannot determine the migratetype of allocated pages precisely
(at least without the page_owner functionality enabled), we can count
pages that compaction would try to isolate for migration - those are
either on LRU or __PageMovable(). The rest can be assumed to be
MIGRATE_RECLAIMABLE or MIGRATE_UNMOVABLE, which we cannot easily
distinguish. This counting can be done as part of free page stealing
with little additional overhead.
The page stealing code is changed so that it considers free pages plus
pages of the "good" migratetype for the decision whether to change
pageblock's migratetype.
The result should be more accurate migratetype of pageblocks wrt the
actual pages in the pageblocks, when stealing from semi-occupied
pageblocks. This should help the efficiency of page grouping by
mobility.
In testing based on 4.9 kernel with stress-highalloc from mmtests
configured for order-4 GFP_KERNEL allocations, this patch has reduced
the number of unmovable allocations falling back to movable pageblocks
by 47%. The number of movable allocations falling back to other
pageblocks are increased by 55%, but these events don't cause permanent
fragmentation, so the tradeoff should be positive. Later patches also
offset the movable fallback increase to some extent.
[akpm@linux-foundation.org: merge fix]
Link: http://lkml.kernel.org/r/20170307131545.28577-5-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-08 15:54:40 -07:00
|
|
|
|
2024-04-25 20:56:04 -07:00
|
|
|
|
2017-05-08 15:54:37 -07:00
|
|
|
|
mm/page_allo.c: restructure free-page stealing code and fix a bug
The free-page stealing code in __rmqueue_fallback() is somewhat hard to
follow, and has an incredible amount of subtlety hidden inside!
First off, there is a minor bug in the reporting of change-of-ownership of
pageblocks. Under some conditions, we try to move upto
'pageblock_nr_pages' no. of pages to the preferred allocation list. But
we change the ownership of that pageblock to the preferred type only if we
manage to successfully move atleast half of that pageblock (or if
page_group_by_mobility_disabled is set).
However, the current code ignores the latter part and sets the
'migratetype' variable to the preferred type, irrespective of whether we
actually changed the pageblock migratetype of that block or not. So, the
page_alloc_extfrag tracepoint can end up printing incorrect info (i.e.,
'change_ownership' might be shown as 1 when it must have been 0).
So fixing this involves moving the update of the 'migratetype' variable to
the right place. But looking closer, we observe that the 'migratetype'
variable is used subsequently for checks such as "is_migrate_cma()".
Obviously the intent there is to check if the *fallback* type is
MIGRATE_CMA, but since we already set the 'migratetype' variable to
start_migratetype, we end up checking if the *preferred* type is
MIGRATE_CMA!!
To make things more interesting, this actually doesn't cause a bug in
practice, because we never change *anything* if the fallback type is CMA.
So, restructure the code in such a way that it is trivial to understand
what is going on, and also fix the above mentioned bug. And while at it,
also add a comment explaining the subtlety behind the migratetype used in
the call to expand().
[akpm@linux-foundation.org: remove unneeded `inline', small coding-style fix]
Signed-off-by: Srivatsa S. Bhat <srivatsa.bhat@linux.vnet.ibm.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Cody P Schafer <cody@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 14:20:35 -07:00
|
|
|
|
|
|
|
|
|
2024-08-26 14:40:48 +08:00
|
|
|
|
|
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
mm/page_allo.c: restructure free-page stealing code and fix a bug
The free-page stealing code in __rmqueue_fallback() is somewhat hard to
follow, and has an incredible amount of subtlety hidden inside!
First off, there is a minor bug in the reporting of change-of-ownership of
pageblocks. Under some conditions, we try to move upto
'pageblock_nr_pages' no. of pages to the preferred allocation list. But
we change the ownership of that pageblock to the preferred type only if we
manage to successfully move atleast half of that pageblock (or if
page_group_by_mobility_disabled is set).
However, the current code ignores the latter part and sets the
'migratetype' variable to the preferred type, irrespective of whether we
actually changed the pageblock migratetype of that block or not. So, the
page_alloc_extfrag tracepoint can end up printing incorrect info (i.e.,
'change_ownership' might be shown as 1 when it must have been 0).
So fixing this involves moving the update of the 'migratetype' variable to
the right place. But looking closer, we observe that the 'migratetype'
variable is used subsequently for checks such as "is_migrate_cma()".
Obviously the intent there is to check if the *fallback* type is
MIGRATE_CMA, but since we already set the 'migratetype' variable to
start_migratetype, we end up checking if the *preferred* type is
MIGRATE_CMA!!
To make things more interesting, this actually doesn't cause a bug in
practice, because we never change *anything* if the fallback type is CMA.
So, restructure the code in such a way that it is trivial to understand
what is going on, and also fix the above mentioned bug. And while at it,
also add a comment explaining the subtlety behind the migratetype used in
the call to expand().
[akpm@linux-foundation.org: remove unneeded `inline', small coding-style fix]
Signed-off-by: Srivatsa S. Bhat <srivatsa.bhat@linux.vnet.ibm.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Cody P Schafer <cody@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 14:20:35 -07:00
|
|
|
|
2024-08-26 14:40:48 +08:00
|
|
|
|
|
|
|
|
|
mm: page_alloc: fix freelist movement during block conversion
Currently, page block type conversion during fallbacks, atomic
reservations and isolation can strand various amounts of free pages on
incorrect freelists.
For example, fallback stealing moves free pages in the block to the new
type's freelists, but then may not actually claim the block for that type
if there aren't enough compatible pages already allocated.
In all cases, free page moving might fail if the block straddles more than
one zone, in which case no free pages are moved at all, but the block type
is changed anyway.
This is detrimental to type hygiene on the freelists. It encourages
incompatible page mixing down the line (ask for one type, get another) and
thus contributes to long-term fragmentation.
Split the process into a proper transaction: check first if conversion
will happen, then try to move the free pages, and only if that was
successful convert the block to the new type.
[baolin.wang@linux.alibaba.com: fix allocation failures with CONFIG_CMA]
Link: https://lkml.kernel.org/r/a97697e0-45b0-4f71-b087-fdc7a1d43c0e@linux.alibaba.com
Link: https://lkml.kernel.org/r/20240320180429.678181-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Tested-by: "Huang, Ying" <ying.huang@intel.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-20 14:02:11 -04:00
|
|
|
|
mm/page_allo.c: restructure free-page stealing code and fix a bug
The free-page stealing code in __rmqueue_fallback() is somewhat hard to
follow, and has an incredible amount of subtlety hidden inside!
First off, there is a minor bug in the reporting of change-of-ownership of
pageblocks. Under some conditions, we try to move upto
'pageblock_nr_pages' no. of pages to the preferred allocation list. But
we change the ownership of that pageblock to the preferred type only if we
manage to successfully move atleast half of that pageblock (or if
page_group_by_mobility_disabled is set).
However, the current code ignores the latter part and sets the
'migratetype' variable to the preferred type, irrespective of whether we
actually changed the pageblock migratetype of that block or not. So, the
page_alloc_extfrag tracepoint can end up printing incorrect info (i.e.,
'change_ownership' might be shown as 1 when it must have been 0).
So fixing this involves moving the update of the 'migratetype' variable to
the right place. But looking closer, we observe that the 'migratetype'
variable is used subsequently for checks such as "is_migrate_cma()".
Obviously the intent there is to check if the *fallback* type is
MIGRATE_CMA, but since we already set the 'migratetype' variable to
start_migratetype, we end up checking if the *preferred* type is
MIGRATE_CMA!!
To make things more interesting, this actually doesn't cause a bug in
practice, because we never change *anything* if the fallback type is CMA.
So, restructure the code in such a way that it is trivial to understand
what is going on, and also fix the above mentioned bug. And while at it,
also add a comment explaining the subtlety behind the migratetype used in
the call to expand().
[akpm@linux-foundation.org: remove unneeded `inline', small coding-style fix]
Signed-off-by: Srivatsa S. Bhat <srivatsa.bhat@linux.vnet.ibm.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Cody P Schafer <cody@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 14:20:35 -07:00
|
|
|
|
|
|
|
|
|
mm: reclaim small amounts of memory when an external fragmentation event occurs
An external fragmentation event was previously described as
When the page allocator fragments memory, it records the event using
the mm_page_alloc_extfrag event. If the fallback_order is smaller
than a pageblock order (order-9 on 64-bit x86) then it's considered
an event that will cause external fragmentation issues in the future.
The kernel reduces the probability of such events by increasing the
watermark sizes by calling set_recommended_min_free_kbytes early in the
lifetime of the system. This works reasonably well in general but if
there are enough sparsely populated pageblocks then the problem can still
occur as enough memory is free overall and kswapd stays asleep.
This patch introduces a watermark_boost_factor sysctl that allows a zone
watermark to be temporarily boosted when an external fragmentation causing
events occurs. The boosting will stall allocations that would decrease
free memory below the boosted low watermark and kswapd is woken if the
calling context allows to reclaim an amount of memory relative to the size
of the high watermark and the watermark_boost_factor until the boost is
cleared. When kswapd finishes, it wakes kcompactd at the pageblock order
to clean some of the pageblocks that may have been affected by the
fragmentation event. kswapd avoids any writeback, slab shrinkage and swap
from reclaim context during this operation to avoid excessive system
disruption in the name of fragmentation avoidance. Care is taken so that
kswapd will do normal reclaim work if the system is really low on memory.
This was evaluated using the same workloads as "mm, page_alloc: Spread
allocations across zones before introducing fragmentation".
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
4.20-rc3+patch1-4: 18421 (98% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-1 653.58 ( 0.00%) 652.71 ( 0.13%)
Amean fault-huge-1 0.00 ( 0.00%) 178.93 * -99.00%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 0.00 ( 0.00%) 5.12 ( 100.00%)
Note that external fragmentation causing events are massively reduced by
this path whether in comparison to the previous kernel or the vanilla
kernel. The fault latency for huge pages appears to be increased but that
is only because THP allocations were successful with the patch applied.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
4.20-rc3+patch1-4: 13464 (95% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Min fault-base-1 912.00 ( 0.00%) 905.00 ( 0.77%)
Min fault-huge-1 127.00 ( 0.00%) 135.00 ( -6.30%)
Amean fault-base-1 1467.55 ( 0.00%) 1481.67 ( -0.96%)
Amean fault-huge-1 1127.11 ( 0.00%) 1063.88 * 5.61%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 77.64 ( 0.00%) 83.46 ( 7.49%)
As before, massive reduction in external fragmentation events, some jitter
on latencies and an increase in THP allocation success rates.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
4.20-rc3+patch1-4: 14263 (93% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 1346.45 ( 0.00%) 1306.87 ( 2.94%)
Amean fault-huge-5 3418.60 ( 0.00%) 1348.94 ( 60.54%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 0.78 ( 0.00%) 7.91 ( 910.64%)
There is a 93% reduction in fragmentation causing events, there is a big
reduction in the huge page fault latency and allocation success rate is
higher.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
4.20-rc3+patch1-4: 11095 (93% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 6217.43 ( 0.00%) 7419.67 * -19.34%*
Amean fault-huge-5 3163.33 ( 0.00%) 3263.80 ( -3.18%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 95.14 ( 0.00%) 87.98 ( -7.53%)
There is a large reduction in fragmentation events with some jitter around
the latencies and success rates. As before, the high THP allocation
success rate does mean the system is under a lot of pressure. However, as
the fragmentation events are reduced, it would be expected that the
long-term allocation success rate would be higher.
Link: http://lkml.kernel.org/r/20181123114528.28802-5-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:52 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-12-14 19:12:15 -08:00
|
|
|
|
mm, page_alloc: do not wake kswapd with zone lock held
syzbot reported the following regression in the latest merge window and
it was confirmed by Qian Cai that a similar bug was visible from a
different context.
======================================================
WARNING: possible circular locking dependency detected
4.20.0+ #297 Not tainted
------------------------------------------------------
syz-executor0/8529 is trying to acquire lock:
000000005e7fb829 (&pgdat->kswapd_wait){....}, at:
__wake_up_common_lock+0x19e/0x330 kernel/sched/wait.c:120
but task is already holding lock:
000000009bb7bae0 (&(&zone->lock)->rlock){-.-.}, at: spin_lock
include/linux/spinlock.h:329 [inline]
000000009bb7bae0 (&(&zone->lock)->rlock){-.-.}, at: rmqueue_bulk
mm/page_alloc.c:2548 [inline]
000000009bb7bae0 (&(&zone->lock)->rlock){-.-.}, at: __rmqueue_pcplist
mm/page_alloc.c:3021 [inline]
000000009bb7bae0 (&(&zone->lock)->rlock){-.-.}, at: rmqueue_pcplist
mm/page_alloc.c:3050 [inline]
000000009bb7bae0 (&(&zone->lock)->rlock){-.-.}, at: rmqueue
mm/page_alloc.c:3072 [inline]
000000009bb7bae0 (&(&zone->lock)->rlock){-.-.}, at:
get_page_from_freelist+0x1bae/0x52a0 mm/page_alloc.c:3491
It appears to be a false positive in that the only way the lock ordering
should be inverted is if kswapd is waking itself and the wakeup
allocates debugging objects which should already be allocated if it's
kswapd doing the waking. Nevertheless, the possibility exists and so
it's best to avoid the problem.
This patch flags a zone as needing a kswapd using the, surprisingly,
unused zone flag field. The flag is read without the lock held to do
the wakeup. It's possible that the flag setting context is not the same
as the flag clearing context or for small races to occur. However, each
race possibility is harmless and there is no visible degredation in
fragmentation treatment.
While zone->flag could have continued to be unused, there is potential
for moving some existing fields into the flags field instead.
Particularly read-mostly ones like zone->initialized and
zone->contiguous.
Link: http://lkml.kernel.org/r/20190103225712.GJ31517@techsingularity.net
Fixes: 1c30844d2dfe ("mm: reclaim small amounts of memory when an external fragmentation event occurs")
Reported-by: syzbot+93d94a001cfbce9e60e1@syzkaller.appspotmail.com
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Tested-by: Qian Cai <cai@lca.pw>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-01-08 15:23:39 -08:00
|
|
|
|
mm: reclaim small amounts of memory when an external fragmentation event occurs
An external fragmentation event was previously described as
When the page allocator fragments memory, it records the event using
the mm_page_alloc_extfrag event. If the fallback_order is smaller
than a pageblock order (order-9 on 64-bit x86) then it's considered
an event that will cause external fragmentation issues in the future.
The kernel reduces the probability of such events by increasing the
watermark sizes by calling set_recommended_min_free_kbytes early in the
lifetime of the system. This works reasonably well in general but if
there are enough sparsely populated pageblocks then the problem can still
occur as enough memory is free overall and kswapd stays asleep.
This patch introduces a watermark_boost_factor sysctl that allows a zone
watermark to be temporarily boosted when an external fragmentation causing
events occurs. The boosting will stall allocations that would decrease
free memory below the boosted low watermark and kswapd is woken if the
calling context allows to reclaim an amount of memory relative to the size
of the high watermark and the watermark_boost_factor until the boost is
cleared. When kswapd finishes, it wakes kcompactd at the pageblock order
to clean some of the pageblocks that may have been affected by the
fragmentation event. kswapd avoids any writeback, slab shrinkage and swap
from reclaim context during this operation to avoid excessive system
disruption in the name of fragmentation avoidance. Care is taken so that
kswapd will do normal reclaim work if the system is really low on memory.
This was evaluated using the same workloads as "mm, page_alloc: Spread
allocations across zones before introducing fragmentation".
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
4.20-rc3+patch1-4: 18421 (98% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-1 653.58 ( 0.00%) 652.71 ( 0.13%)
Amean fault-huge-1 0.00 ( 0.00%) 178.93 * -99.00%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 0.00 ( 0.00%) 5.12 ( 100.00%)
Note that external fragmentation causing events are massively reduced by
this path whether in comparison to the previous kernel or the vanilla
kernel. The fault latency for huge pages appears to be increased but that
is only because THP allocations were successful with the patch applied.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
4.20-rc3+patch1-4: 13464 (95% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Min fault-base-1 912.00 ( 0.00%) 905.00 ( 0.77%)
Min fault-huge-1 127.00 ( 0.00%) 135.00 ( -6.30%)
Amean fault-base-1 1467.55 ( 0.00%) 1481.67 ( -0.96%)
Amean fault-huge-1 1127.11 ( 0.00%) 1063.88 * 5.61%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 77.64 ( 0.00%) 83.46 ( 7.49%)
As before, massive reduction in external fragmentation events, some jitter
on latencies and an increase in THP allocation success rates.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
4.20-rc3+patch1-4: 14263 (93% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 1346.45 ( 0.00%) 1306.87 ( 2.94%)
Amean fault-huge-5 3418.60 ( 0.00%) 1348.94 ( 60.54%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 0.78 ( 0.00%) 7.91 ( 910.64%)
There is a 93% reduction in fragmentation causing events, there is a big
reduction in the huge page fault latency and allocation success rate is
higher.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
4.20-rc3+patch1-4: 11095 (93% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 6217.43 ( 0.00%) 7419.67 * -19.34%*
Amean fault-huge-5 3163.33 ( 0.00%) 3263.80 ( -3.18%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 95.14 ( 0.00%) 87.98 ( -7.53%)
There is a large reduction in fragmentation events with some jitter around
the latencies and success rates. As before, the high THP allocation
success rate does mean the system is under a lot of pressure. However, as
the fragmentation events are reduced, it would be expected that the
long-term allocation success rate would be higher.
Link: http://lkml.kernel.org/r/20181123114528.28802-5-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:52 -08:00
|
|
|
|
2023-08-01 20:37:23 +08:00
|
|
|
|
2024-04-25 20:56:04 -07:00
|
|
|
|
|
|
|
|
|
mm: page_alloc: don't steal single pages from biggest buddy
The fallback code searches for the biggest buddy first in an attempt to
steal the whole block and encourage type grouping down the line.
The approach used to be this:
- Non-movable requests will split the largest buddy and steal the
remainder. This splits up contiguity, but it allows subsequent
requests of this type to fall back into adjacent space.
- Movable requests go and look for the smallest buddy instead. The
thinking is that movable requests can be compacted, so grouping is
less important than retaining contiguity.
c0cd6f557b90 ("mm: page_alloc: fix freelist movement during block
conversion") enforces freelist type hygiene, which restricts stealing to
either claiming the whole block or just taking the requested chunk; no
additional pages or buddy remainders can be stolen any more.
The patch mishandled when to switch to finding the smallest buddy in that
new reality. As a result, it may steal the exact request size, but from
the biggest buddy. This causes fracturing for no good reason.
Fix this by committing to the new behavior: either steal the whole block,
or fall back to the smallest buddy.
Remove single-page stealing from steal_suitable_fallback(). Rename it to
try_to_steal_block() to make the intentions clear. If this fails, always
fall back to the smallest buddy.
The following is from 4 runs of mmtest's thpchallenge. "Pollute" is
single page fallback, "steal" is conversion of a partially used block.
The numbers for free block conversions (omitted) are comparable.
vanilla patched
@pollute[unmovable from reclaimable]: 27 106
@pollute[unmovable from movable]: 82 46
@pollute[reclaimable from unmovable]: 256 83
@pollute[reclaimable from movable]: 46 8
@pollute[movable from unmovable]: 4841 868
@pollute[movable from reclaimable]: 5278 12568
@steal[unmovable from reclaimable]: 11 12
@steal[unmovable from movable]: 113 49
@steal[reclaimable from unmovable]: 19 34
@steal[reclaimable from movable]: 47 21
@steal[movable from unmovable]: 250 183
@steal[movable from reclaimable]: 81 93
The allocator appears to do a better job at keeping stealing and polluting
to the first fallback preference. As a result, the numbers for "from
movable" - the least preferred fallback option, and most detrimental to
compactability - are down across the board.
Link: https://lkml.kernel.org/r/20250225001023.1494422-2-hannes@cmpxchg.org
Fixes: c0cd6f557b90 ("mm: page_alloc: fix freelist movement during block conversion")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-02-24 19:08:24 -05:00
|
|
|
|
2023-08-01 20:37:23 +08:00
|
|
|
|
mm, page_alloc: count movable pages when stealing from pageblock
When stealing pages from pageblock of a different migratetype, we count
how many free pages were stolen, and change the pageblock's migratetype
if more than half of the pageblock was free. This might be too
conservative, as there might be other pages that are not free, but were
allocated with the same migratetype as our allocation requested.
While we cannot determine the migratetype of allocated pages precisely
(at least without the page_owner functionality enabled), we can count
pages that compaction would try to isolate for migration - those are
either on LRU or __PageMovable(). The rest can be assumed to be
MIGRATE_RECLAIMABLE or MIGRATE_UNMOVABLE, which we cannot easily
distinguish. This counting can be done as part of free page stealing
with little additional overhead.
The page stealing code is changed so that it considers free pages plus
pages of the "good" migratetype for the decision whether to change
pageblock's migratetype.
The result should be more accurate migratetype of pageblocks wrt the
actual pages in the pageblocks, when stealing from semi-occupied
pageblocks. This should help the efficiency of page grouping by
mobility.
In testing based on 4.9 kernel with stress-highalloc from mmtests
configured for order-4 GFP_KERNEL allocations, this patch has reduced
the number of unmovable allocations falling back to movable pageblocks
by 47%. The number of movable allocations falling back to other
pageblocks are increased by 55%, but these events don't cause permanent
fragmentation, so the tradeoff should be positive. Later patches also
offset the movable fallback increase to some extent.
[akpm@linux-foundation.org: merge fix]
Link: http://lkml.kernel.org/r/20170307131545.28577-5-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-08 15:54:40 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: page_alloc: fix freelist movement during block conversion
Currently, page block type conversion during fallbacks, atomic
reservations and isolation can strand various amounts of free pages on
incorrect freelists.
For example, fallback stealing moves free pages in the block to the new
type's freelists, but then may not actually claim the block for that type
if there aren't enough compatible pages already allocated.
In all cases, free page moving might fail if the block straddles more than
one zone, in which case no free pages are moved at all, but the block type
is changed anyway.
This is detrimental to type hygiene on the freelists. It encourages
incompatible page mixing down the line (ask for one type, get another) and
thus contributes to long-term fragmentation.
Split the process into a proper transaction: check first if conversion
will happen, then try to move the free pages, and only if that was
successful convert the block to the new type.
[baolin.wang@linux.alibaba.com: fix allocation failures with CONFIG_CMA]
Link: https://lkml.kernel.org/r/a97697e0-45b0-4f71-b087-fdc7a1d43c0e@linux.alibaba.com
Link: https://lkml.kernel.org/r/20240320180429.678181-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Tested-by: "Huang, Ying" <ying.huang@intel.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-20 14:02:11 -04:00
|
|
|
|
mm, page_alloc: count movable pages when stealing from pageblock
When stealing pages from pageblock of a different migratetype, we count
how many free pages were stolen, and change the pageblock's migratetype
if more than half of the pageblock was free. This might be too
conservative, as there might be other pages that are not free, but were
allocated with the same migratetype as our allocation requested.
While we cannot determine the migratetype of allocated pages precisely
(at least without the page_owner functionality enabled), we can count
pages that compaction would try to isolate for migration - those are
either on LRU or __PageMovable(). The rest can be assumed to be
MIGRATE_RECLAIMABLE or MIGRATE_UNMOVABLE, which we cannot easily
distinguish. This counting can be done as part of free page stealing
with little additional overhead.
The page stealing code is changed so that it considers free pages plus
pages of the "good" migratetype for the decision whether to change
pageblock's migratetype.
The result should be more accurate migratetype of pageblocks wrt the
actual pages in the pageblocks, when stealing from semi-occupied
pageblocks. This should help the efficiency of page grouping by
mobility.
In testing based on 4.9 kernel with stress-highalloc from mmtests
configured for order-4 GFP_KERNEL allocations, this patch has reduced
the number of unmovable allocations falling back to movable pageblocks
by 47%. The number of movable allocations falling back to other
pageblocks are increased by 55%, but these events don't cause permanent
fragmentation, so the tradeoff should be positive. Later patches also
offset the movable fallback increase to some extent.
[akpm@linux-foundation.org: merge fix]
Link: http://lkml.kernel.org/r/20170307131545.28577-5-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-08 15:54:40 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-08-01 20:37:23 +08:00
|
|
|
|
mm, page_alloc: count movable pages when stealing from pageblock
When stealing pages from pageblock of a different migratetype, we count
how many free pages were stolen, and change the pageblock's migratetype
if more than half of the pageblock was free. This might be too
conservative, as there might be other pages that are not free, but were
allocated with the same migratetype as our allocation requested.
While we cannot determine the migratetype of allocated pages precisely
(at least without the page_owner functionality enabled), we can count
pages that compaction would try to isolate for migration - those are
either on LRU or __PageMovable(). The rest can be assumed to be
MIGRATE_RECLAIMABLE or MIGRATE_UNMOVABLE, which we cannot easily
distinguish. This counting can be done as part of free page stealing
with little additional overhead.
The page stealing code is changed so that it considers free pages plus
pages of the "good" migratetype for the decision whether to change
pageblock's migratetype.
The result should be more accurate migratetype of pageblocks wrt the
actual pages in the pageblocks, when stealing from semi-occupied
pageblocks. This should help the efficiency of page grouping by
mobility.
In testing based on 4.9 kernel with stress-highalloc from mmtests
configured for order-4 GFP_KERNEL allocations, this patch has reduced
the number of unmovable allocations falling back to movable pageblocks
by 47%. The number of movable allocations falling back to other
pageblocks are increased by 55%, but these events don't cause permanent
fragmentation, so the tradeoff should be positive. Later patches also
offset the movable fallback increase to some extent.
[akpm@linux-foundation.org: merge fix]
Link: http://lkml.kernel.org/r/20170307131545.28577-5-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-08 15:54:40 -07:00
|
|
|
|
|
|
|
|
|
mm: page_alloc: fix freelist movement during block conversion
Currently, page block type conversion during fallbacks, atomic
reservations and isolation can strand various amounts of free pages on
incorrect freelists.
For example, fallback stealing moves free pages in the block to the new
type's freelists, but then may not actually claim the block for that type
if there aren't enough compatible pages already allocated.
In all cases, free page moving might fail if the block straddles more than
one zone, in which case no free pages are moved at all, but the block type
is changed anyway.
This is detrimental to type hygiene on the freelists. It encourages
incompatible page mixing down the line (ask for one type, get another) and
thus contributes to long-term fragmentation.
Split the process into a proper transaction: check first if conversion
will happen, then try to move the free pages, and only if that was
successful convert the block to the new type.
[baolin.wang@linux.alibaba.com: fix allocation failures with CONFIG_CMA]
Link: https://lkml.kernel.org/r/a97697e0-45b0-4f71-b087-fdc7a1d43c0e@linux.alibaba.com
Link: https://lkml.kernel.org/r/20240320180429.678181-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Tested-by: "Huang, Ying" <ying.huang@intel.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-20 14:02:11 -04:00
|
|
|
|
2024-04-25 20:56:04 -07:00
|
|
|
|
2025-06-16 22:11:12 -04:00
|
|
|
|
mm: page_alloc: fix freelist movement during block conversion
Currently, page block type conversion during fallbacks, atomic
reservations and isolation can strand various amounts of free pages on
incorrect freelists.
For example, fallback stealing moves free pages in the block to the new
type's freelists, but then may not actually claim the block for that type
if there aren't enough compatible pages already allocated.
In all cases, free page moving might fail if the block straddles more than
one zone, in which case no free pages are moved at all, but the block type
is changed anyway.
This is detrimental to type hygiene on the freelists. It encourages
incompatible page mixing down the line (ask for one type, get another) and
thus contributes to long-term fragmentation.
Split the process into a proper transaction: check first if conversion
will happen, then try to move the free pages, and only if that was
successful convert the block to the new type.
[baolin.wang@linux.alibaba.com: fix allocation failures with CONFIG_CMA]
Link: https://lkml.kernel.org/r/a97697e0-45b0-4f71-b087-fdc7a1d43c0e@linux.alibaba.com
Link: https://lkml.kernel.org/r/20240320180429.678181-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Tested-by: "Huang, Ying" <ying.huang@intel.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-20 14:02:11 -04:00
|
|
|
|
|
|
|
|
|
2017-05-08 15:54:37 -07:00
|
|
|
|
mm: page_alloc: don't steal single pages from biggest buddy
The fallback code searches for the biggest buddy first in an attempt to
steal the whole block and encourage type grouping down the line.
The approach used to be this:
- Non-movable requests will split the largest buddy and steal the
remainder. This splits up contiguity, but it allows subsequent
requests of this type to fall back into adjacent space.
- Movable requests go and look for the smallest buddy instead. The
thinking is that movable requests can be compacted, so grouping is
less important than retaining contiguity.
c0cd6f557b90 ("mm: page_alloc: fix freelist movement during block
conversion") enforces freelist type hygiene, which restricts stealing to
either claiming the whole block or just taking the requested chunk; no
additional pages or buddy remainders can be stolen any more.
The patch mishandled when to switch to finding the smallest buddy in that
new reality. As a result, it may steal the exact request size, but from
the biggest buddy. This causes fracturing for no good reason.
Fix this by committing to the new behavior: either steal the whole block,
or fall back to the smallest buddy.
Remove single-page stealing from steal_suitable_fallback(). Rename it to
try_to_steal_block() to make the intentions clear. If this fails, always
fall back to the smallest buddy.
The following is from 4 runs of mmtest's thpchallenge. "Pollute" is
single page fallback, "steal" is conversion of a partially used block.
The numbers for free block conversions (omitted) are comparable.
vanilla patched
@pollute[unmovable from reclaimable]: 27 106
@pollute[unmovable from movable]: 82 46
@pollute[reclaimable from unmovable]: 256 83
@pollute[reclaimable from movable]: 46 8
@pollute[movable from unmovable]: 4841 868
@pollute[movable from reclaimable]: 5278 12568
@steal[unmovable from reclaimable]: 11 12
@steal[unmovable from movable]: 113 49
@steal[reclaimable from unmovable]: 19 34
@steal[reclaimable from movable]: 47 21
@steal[movable from unmovable]: 250 183
@steal[movable from reclaimable]: 81 93
The allocator appears to do a better job at keeping stealing and polluting
to the first fallback preference. As a result, the numbers for "from
movable" - the least preferred fallback option, and most detrimental to
compactability - are down across the board.
Link: https://lkml.kernel.org/r/20250225001023.1494422-2-hannes@cmpxchg.org
Fixes: c0cd6f557b90 ("mm: page_alloc: fix freelist movement during block conversion")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-02-24 19:08:24 -05:00
|
|
|
|
2015-04-14 15:45:18 -07:00
|
|
|
|
mm/page_allo.c: restructure free-page stealing code and fix a bug
The free-page stealing code in __rmqueue_fallback() is somewhat hard to
follow, and has an incredible amount of subtlety hidden inside!
First off, there is a minor bug in the reporting of change-of-ownership of
pageblocks. Under some conditions, we try to move upto
'pageblock_nr_pages' no. of pages to the preferred allocation list. But
we change the ownership of that pageblock to the preferred type only if we
manage to successfully move atleast half of that pageblock (or if
page_group_by_mobility_disabled is set).
However, the current code ignores the latter part and sets the
'migratetype' variable to the preferred type, irrespective of whether we
actually changed the pageblock migratetype of that block or not. So, the
page_alloc_extfrag tracepoint can end up printing incorrect info (i.e.,
'change_ownership' might be shown as 1 when it must have been 0).
So fixing this involves moving the update of the 'migratetype' variable to
the right place. But looking closer, we observe that the 'migratetype'
variable is used subsequently for checks such as "is_migrate_cma()".
Obviously the intent there is to check if the *fallback* type is
MIGRATE_CMA, but since we already set the 'migratetype' variable to
start_migratetype, we end up checking if the *preferred* type is
MIGRATE_CMA!!
To make things more interesting, this actually doesn't cause a bug in
practice, because we never change *anything* if the fallback type is CMA.
So, restructure the code in such a way that it is trivial to understand
what is going on, and also fix the above mentioned bug. And while at it,
also add a comment explaining the subtlety behind the migratetype used in
the call to expand().
[akpm@linux-foundation.org: remove unneeded `inline', small coding-style fix]
Signed-off-by: Srivatsa S. Bhat <srivatsa.bhat@linux.vnet.ibm.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Cody P Schafer <cody@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 14:20:35 -07:00
|
|
|
|
mm/compaction: enhance compaction finish condition
Compaction has anti fragmentation algorithm. It is that freepage should
be more than pageblock order to finish the compaction if we don't find any
freepage in requested migratetype buddy list. This is for mitigating
fragmentation, but, there is a lack of migratetype consideration and it is
too excessive compared to page allocator's anti fragmentation algorithm.
Not considering migratetype would cause premature finish of compaction.
For example, if allocation request is for unmovable migratetype, freepage
with CMA migratetype doesn't help that allocation and compaction should
not be stopped. But, current logic regards this situation as compaction
is no longer needed, so finish the compaction.
Secondly, condition is too excessive compared to page allocator's logic.
We can steal freepage from other migratetype and change pageblock
migratetype on more relaxed conditions in page allocator. This is
designed to prevent fragmentation and we can use it here. Imposing hard
constraint only to the compaction doesn't help much in this case since
page allocator would cause fragmentation again.
To solve these problems, this patch borrows anti fragmentation logic from
page allocator. It will reduce premature compaction finish in some cases
and reduce excessive compaction work.
stress-highalloc test in mmtests with non movable order 7 allocation shows
considerable increase of compaction success rate.
Compaction success rate (Compaction success * 100 / Compaction stalls, %)
31.82 : 42.20
I tested it on non-reboot 5 runs stress-highalloc benchmark and found that
there is no more degradation on allocation success rate than before. That
roughly means that this patch doesn't result in more fragmentations.
Vlastimil suggests additional idea that we only test for fallbacks when
migration scanner has scanned a whole pageblock. It looked good for
fragmentation because chance of stealing increase due to making more free
pages in certain pageblock. So, I tested it, but, it results in decreased
compaction success rate, roughly 38.00. I guess the reason that if system
is low memory condition, watermark check could be failed due to not enough
order 0 free page and so, sometimes, we can't reach a fallback check
although migrate_pfn is aligned to pageblock_nr_pages. I can insert code
to cope with this situation but it makes code more complicated so I don't
include his idea at this patch.
[akpm@linux-foundation.org: fix CONFIG_CMA=n build]
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Mel Gorman <mgorman@suse.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-04-14 15:45:21 -07:00
|
|
|
|
mm: page_alloc: speed up fallbacks in rmqueue_bulk()
The test robot identified c2f6ea38fc1b ("mm: page_alloc: don't steal
single pages from biggest buddy") as the root cause of a 56.4% regression
in vm-scalability::lru-file-mmap-read.
Carlos reports an earlier patch, c0cd6f557b90 ("mm: page_alloc: fix
freelist movement during block conversion"), as the root cause for a
regression in worst-case zone->lock+irqoff hold times.
Both of these patches modify the page allocator's fallback path to be less
greedy in an effort to stave off fragmentation. The flip side of this is
that fallbacks are also less productive each time around, which means the
fallback search can run much more frequently.
Carlos' traces point to rmqueue_bulk() specifically, which tries to refill
the percpu cache by allocating a large batch of pages in a loop. It
highlights how once the native freelists are exhausted, the fallback code
first scans orders top-down for whole blocks to claim, then falls back to
a bottom-up search for the smallest buddy to steal. For the next batch
page, it goes through the same thing again.
This can be made more efficient. Since rmqueue_bulk() holds the
zone->lock over the entire batch, the freelists are not subject to outside
changes; when the search for a block to claim has already failed, there is
no point in trying again for the next page.
Modify __rmqueue() to remember the last successful fallback mode, and
restart directly from there on the next rmqueue_bulk() iteration.
Oliver confirms that this improves beyond the regression that the test
robot reported against c2f6ea38fc1b:
commit:
f3b92176f4 ("tools/selftests: add guard region test for /proc/$pid/pagemap")
c2f6ea38fc ("mm: page_alloc: don't steal single pages from biggest buddy")
acc4d5ff0b ("Merge tag 'net-6.15-rc0' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net")
2c847f27c3 ("mm: page_alloc: speed up fallbacks in rmqueue_bulk()") <--- your patch
f3b92176f4f7100f c2f6ea38fc1b640aa7a2e155cc1 acc4d5ff0b61eb1715c498b6536 2c847f27c37da65a93d23c237c5
---------------- --------------------------- --------------------------- ---------------------------
%stddev %change %stddev %change %stddev %change %stddev
\ | \ | \ | \
25525364 ± 3% -56.4% 11135467 -57.8% 10779336 +31.6% 33581409 vm-scalability.throughput
Carlos confirms that worst-case times are almost fully recovered
compared to before the earlier culprit patch:
2dd482ba627d (before freelist hygiene): 1ms
c0cd6f557b90 (after freelist hygiene): 90ms
next-20250319 (steal smallest buddy): 280ms
this patch : 8ms
[jackmanb@google.com: comment updates]
Link: https://lkml.kernel.org/r/D92AC0P9594X.3BML64MUKTF8Z@google.com
[hannes@cmpxchg.org: reset rmqueue_mode in rmqueue_buddy() error loop, per Yunsheng Lin]
Link: https://lkml.kernel.org/r/20250409140023.GA2313@cmpxchg.org
Link: https://lkml.kernel.org/r/20250407180154.63348-1-hannes@cmpxchg.org
Fixes: c0cd6f557b90 ("mm: page_alloc: fix freelist movement during block conversion")
Fixes: c2f6ea38fc1b ("mm: page_alloc: don't steal single pages from biggest buddy")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Brendan Jackman <jackmanb@google.com>
Reported-by: kernel test robot <oliver.sang@intel.com>
Reported-by: Carlos Song <carlos.song@nxp.com>
Tested-by: Carlos Song <carlos.song@nxp.com>
Tested-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202503271547.fc08b188-lkp@intel.com
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Tested-by: Shivank Garg <shivankg@amd.com>
Acked-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: <stable@vger.kernel.org> [6.10+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-07 14:01:53 -04:00
|
|
|
|
|
|
|
|
|
2017-07-10 15:49:26 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-05-08 15:54:37 -07:00
|
|
|
|
mm: page_alloc: fix freelist movement during block conversion
Currently, page block type conversion during fallbacks, atomic
reservations and isolation can strand various amounts of free pages on
incorrect freelists.
For example, fallback stealing moves free pages in the block to the new
type's freelists, but then may not actually claim the block for that type
if there aren't enough compatible pages already allocated.
In all cases, free page moving might fail if the block straddles more than
one zone, in which case no free pages are moved at all, but the block type
is changed anyway.
This is detrimental to type hygiene on the freelists. It encourages
incompatible page mixing down the line (ask for one type, get another) and
thus contributes to long-term fragmentation.
Split the process into a proper transaction: check first if conversion
will happen, then try to move the free pages, and only if that was
successful convert the block to the new type.
[baolin.wang@linux.alibaba.com: fix allocation failures with CONFIG_CMA]
Link: https://lkml.kernel.org/r/a97697e0-45b0-4f71-b087-fdc7a1d43c0e@linux.alibaba.com
Link: https://lkml.kernel.org/r/20240320180429.678181-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Tested-by: "Huang, Ying" <ying.huang@intel.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-20 14:02:11 -04:00
|
|
|
|
mm: page_alloc: speed up fallbacks in rmqueue_bulk()
The test robot identified c2f6ea38fc1b ("mm: page_alloc: don't steal
single pages from biggest buddy") as the root cause of a 56.4% regression
in vm-scalability::lru-file-mmap-read.
Carlos reports an earlier patch, c0cd6f557b90 ("mm: page_alloc: fix
freelist movement during block conversion"), as the root cause for a
regression in worst-case zone->lock+irqoff hold times.
Both of these patches modify the page allocator's fallback path to be less
greedy in an effort to stave off fragmentation. The flip side of this is
that fallbacks are also less productive each time around, which means the
fallback search can run much more frequently.
Carlos' traces point to rmqueue_bulk() specifically, which tries to refill
the percpu cache by allocating a large batch of pages in a loop. It
highlights how once the native freelists are exhausted, the fallback code
first scans orders top-down for whole blocks to claim, then falls back to
a bottom-up search for the smallest buddy to steal. For the next batch
page, it goes through the same thing again.
This can be made more efficient. Since rmqueue_bulk() holds the
zone->lock over the entire batch, the freelists are not subject to outside
changes; when the search for a block to claim has already failed, there is
no point in trying again for the next page.
Modify __rmqueue() to remember the last successful fallback mode, and
restart directly from there on the next rmqueue_bulk() iteration.
Oliver confirms that this improves beyond the regression that the test
robot reported against c2f6ea38fc1b:
commit:
f3b92176f4 ("tools/selftests: add guard region test for /proc/$pid/pagemap")
c2f6ea38fc ("mm: page_alloc: don't steal single pages from biggest buddy")
acc4d5ff0b ("Merge tag 'net-6.15-rc0' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net")
2c847f27c3 ("mm: page_alloc: speed up fallbacks in rmqueue_bulk()") <--- your patch
f3b92176f4f7100f c2f6ea38fc1b640aa7a2e155cc1 acc4d5ff0b61eb1715c498b6536 2c847f27c37da65a93d23c237c5
---------------- --------------------------- --------------------------- ---------------------------
%stddev %change %stddev %change %stddev %change %stddev
\ | \ | \ | \
25525364 ± 3% -56.4% 11135467 -57.8% 10779336 +31.6% 33581409 vm-scalability.throughput
Carlos confirms that worst-case times are almost fully recovered
compared to before the earlier culprit patch:
2dd482ba627d (before freelist hygiene): 1ms
c0cd6f557b90 (after freelist hygiene): 90ms
next-20250319 (steal smallest buddy): 280ms
this patch : 8ms
[jackmanb@google.com: comment updates]
Link: https://lkml.kernel.org/r/D92AC0P9594X.3BML64MUKTF8Z@google.com
[hannes@cmpxchg.org: reset rmqueue_mode in rmqueue_buddy() error loop, per Yunsheng Lin]
Link: https://lkml.kernel.org/r/20250409140023.GA2313@cmpxchg.org
Link: https://lkml.kernel.org/r/20250407180154.63348-1-hannes@cmpxchg.org
Fixes: c0cd6f557b90 ("mm: page_alloc: fix freelist movement during block conversion")
Fixes: c2f6ea38fc1b ("mm: page_alloc: don't steal single pages from biggest buddy")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Brendan Jackman <jackmanb@google.com>
Reported-by: kernel test robot <oliver.sang@intel.com>
Reported-by: Carlos Song <carlos.song@nxp.com>
Tested-by: Carlos Song <carlos.song@nxp.com>
Tested-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202503271547.fc08b188-lkp@intel.com
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Tested-by: Shivank Garg <shivankg@amd.com>
Acked-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: <stable@vger.kernel.org> [6.10+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-07 14:01:53 -04:00
|
|
|
|
mm, page_alloc: spread allocations across zones before introducing fragmentation
Patch series "Fragmentation avoidance improvements", v5.
It has been noted before that fragmentation avoidance (aka
anti-fragmentation) is not perfect. Given sufficient time or an adverse
workload, memory gets fragmented and the long-term success of high-order
allocations degrades. This series defines an adverse workload, a definition
of external fragmentation events (including serious) ones and a series
that reduces the level of those fragmentation events.
The details of the workload and the consequences are described in more
detail in the changelogs. However, from patch 1, this is a high-level
summary of the adverse workload. The exact details are found in the
mmtests implementation.
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch)
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameterr create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed
3. Warm up a number of fio read-only threads accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll fault back in old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup
Overall the series reduces external fragmentation causing events by over 94%
on 1 and 2 socket machines, which in turn impacts high-order allocation
success rates over the long term. There are differences in latencies and
high-order allocation success rates. Latencies are a mixed bag as they
are vulnerable to exact system state and whether allocations succeeded
so they are treated as a secondary metric.
Patch 1 uses lower zones if they are populated and have free memory
instead of fragmenting a higher zone. It's special cased to
handle a Normal->DMA32 fallback with the reasons explained
in the changelog.
Patch 2-4 boosts watermarks temporarily when an external fragmentation
event occurs. kswapd wakes to reclaim a small amount of old memory
and then wakes kcompactd on completion to recover the system
slightly. This introduces some overhead in the slowpath. The level
of boosting can be tuned or disabled depending on the tolerance
for fragmentation vs allocation latency.
Patch 5 stalls some movable allocation requests to let kswapd from patch 4
make some progress. The duration of the stalls is very low but it
is possible to tune the system to avoid fragmentation events if
larger stalls can be tolerated.
The bulk of the improvement in fragmentation avoidance is from patches
1-4 but patch 5 can deal with a rare corner case and provides the option
of tuning a system for THP allocation success rates in exchange for
some stalls to control fragmentation.
This patch (of 5):
The page allocator zone lists are iterated based on the watermarks of each
zone which does not take anti-fragmentation into account. On x86, node 0
may have multiple zones while other nodes have one zone. A consequence is
that tasks running on node 0 may fragment ZONE_NORMAL even though
ZONE_DMA32 has plenty of free memory. This patch special cases the
allocator fast path such that it'll try an allocation from a lower local
zone before fragmenting a higher zone. In this case, stealing of
pageblocks or orders larger than a pageblock are still allowed in the fast
path as they are uninteresting from a fragmentation point of view.
This was evaluated using a benchmark designed to fragment memory before
attempting THP allocations. It's implemented in mmtests as the following
configurations
configs/config-global-dhp__workload_thpfioscale
configs/config-global-dhp__workload_thpfioscale-defrag
configs/config-global-dhp__workload_thpfioscale-madvhugepage
e.g. from mmtests
./run-mmtests.sh --run-monitor --config configs/config-global-dhp__workload_thpfioscale test-run-1
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch).
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameter create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed.
3. Warm up a number of fio read-only processes accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll refault old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds.
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup the test files.
Note that due to the use of IO and page cache that this benchmark is not
suitable for running on large machines where the time to fragment memory
may be excessive. Also note that while this is one mix that generates
fragmentation that it's not the only mix that generates fragmentation.
Differences in workload that are more slab-intensive or whether SLUB is
used with high-order pages may yield different results.
When the page allocator fragments memory, it records the event using the
mm_page_alloc_extfrag ftrace event. If the fallback_order is smaller than
a pageblock order (order-9 on 64-bit x86) then it's considered to be an
"external fragmentation event" that may cause issues in the future.
Hence, the primary metric here is the number of external fragmentation
events that occur with order < 9. The secondary metric is allocation
latency and huge page allocation success rates but note that differences
in latencies and what the success rate also can affect the number of
external fragmentation event which is why it's a secondary metric.
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 662.92 ( 0.00%) 653.58 * 1.41%*
Amean fault-huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
Fault latencies are slightly reduced while allocation success rates remain
at zero as this configuration does not make any special effort to allocate
THP and fio is heavily active at the time and either filling memory or
keeping pages resident. However, a 49% reduction of serious fragmentation
events reduces the changes of external fragmentation being a problem in
the future.
Vlastimil asked during review for a breakdown of the allocation types
that are falling back.
vanilla
3816 MIGRATE_UNMOVABLE
800845 MIGRATE_MOVABLE
33 MIGRATE_UNRECLAIMABLE
patch
735 MIGRATE_UNMOVABLE
408135 MIGRATE_MOVABLE
42 MIGRATE_UNRECLAIMABLE
The majority of the fallbacks are due to movable allocations and this is
consistent for the workload throughout the series so will not be presented
again as the primary source of fallbacks are movable allocations.
Movable fallbacks are sometimes considered "ok" to fallback because they
can be migrated. The problem is that they can fill an
unmovable/reclaimable pageblock causing those allocations to fallback
later and polluting pageblocks with pages that cannot move. If there is a
movable fallback, it is pretty much guaranteed to affect an
unmovable/reclaimable pageblock and while it might not be enough to
actually cause a unmovable/reclaimable fallback in the future, we cannot
know that in advance so the patch takes the only option available to it.
Hence, it's important to control them. This point is also consistent
throughout the series and will not be repeated.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 1495.14 ( 0.00%) 1467.55 ( 1.85%)
Amean fault-huge-1 1098.48 ( 0.00%) 1127.11 ( -2.61%)
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 78.57 ( 0.00%) 77.64 ( -1.18%)
Fragmentation events were reduced quite a bit although this is known
to be a little variable. The latencies and allocation success rates
are similar but they were already quite high.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 1350.05 ( 0.00%) 1346.45 ( 0.27%)
Amean fault-huge-5 4181.01 ( 0.00%) 3418.60 ( 18.24%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 1.15 ( 0.00%) 0.78 ( -31.88%)
The reduction of external fragmentation events is slight and this is
partially due to the removal of __GFP_THISNODE in commit ac5b2c18911f
("mm: thp: relax __GFP_THISNODE for MADV_HUGEPAGE mappings") as THP
allocations can now spill over to remote nodes instead of fragmenting
local memory.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 6138.97 ( 0.00%) 6217.43 ( -1.28%)
Amean fault-huge-5 2294.28 ( 0.00%) 3163.33 * -37.88%*
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 96.82 ( 0.00%) 95.14 ( -1.74%)
There was a slight reduction in external fragmentation events although the
latencies were higher. The allocation success rate is high enough that
the system is struggling and there is quite a lot of parallel reclaim and
compaction activity. There is also a certain degree of luck on whether
processes start on node 0 or not for this patch but the relevance is
reduced later in the series.
Overall, the patch reduces the number of external fragmentation causing
events so the success of THP over long periods of time would be improved
for this adverse workload.
Link: http://lkml.kernel.org/r/20181123114528.28802-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:41 -08:00
|
|
|
|
2007-10-16 01:25:48 -07:00
|
|
|
|
2013-09-11 14:20:34 -07:00
|
|
|
|
2017-07-10 15:49:26 -07:00
|
|
|
|
mm, page_alloc: spread allocations across zones before introducing fragmentation
Patch series "Fragmentation avoidance improvements", v5.
It has been noted before that fragmentation avoidance (aka
anti-fragmentation) is not perfect. Given sufficient time or an adverse
workload, memory gets fragmented and the long-term success of high-order
allocations degrades. This series defines an adverse workload, a definition
of external fragmentation events (including serious) ones and a series
that reduces the level of those fragmentation events.
The details of the workload and the consequences are described in more
detail in the changelogs. However, from patch 1, this is a high-level
summary of the adverse workload. The exact details are found in the
mmtests implementation.
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch)
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameterr create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed
3. Warm up a number of fio read-only threads accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll fault back in old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup
Overall the series reduces external fragmentation causing events by over 94%
on 1 and 2 socket machines, which in turn impacts high-order allocation
success rates over the long term. There are differences in latencies and
high-order allocation success rates. Latencies are a mixed bag as they
are vulnerable to exact system state and whether allocations succeeded
so they are treated as a secondary metric.
Patch 1 uses lower zones if they are populated and have free memory
instead of fragmenting a higher zone. It's special cased to
handle a Normal->DMA32 fallback with the reasons explained
in the changelog.
Patch 2-4 boosts watermarks temporarily when an external fragmentation
event occurs. kswapd wakes to reclaim a small amount of old memory
and then wakes kcompactd on completion to recover the system
slightly. This introduces some overhead in the slowpath. The level
of boosting can be tuned or disabled depending on the tolerance
for fragmentation vs allocation latency.
Patch 5 stalls some movable allocation requests to let kswapd from patch 4
make some progress. The duration of the stalls is very low but it
is possible to tune the system to avoid fragmentation events if
larger stalls can be tolerated.
The bulk of the improvement in fragmentation avoidance is from patches
1-4 but patch 5 can deal with a rare corner case and provides the option
of tuning a system for THP allocation success rates in exchange for
some stalls to control fragmentation.
This patch (of 5):
The page allocator zone lists are iterated based on the watermarks of each
zone which does not take anti-fragmentation into account. On x86, node 0
may have multiple zones while other nodes have one zone. A consequence is
that tasks running on node 0 may fragment ZONE_NORMAL even though
ZONE_DMA32 has plenty of free memory. This patch special cases the
allocator fast path such that it'll try an allocation from a lower local
zone before fragmenting a higher zone. In this case, stealing of
pageblocks or orders larger than a pageblock are still allowed in the fast
path as they are uninteresting from a fragmentation point of view.
This was evaluated using a benchmark designed to fragment memory before
attempting THP allocations. It's implemented in mmtests as the following
configurations
configs/config-global-dhp__workload_thpfioscale
configs/config-global-dhp__workload_thpfioscale-defrag
configs/config-global-dhp__workload_thpfioscale-madvhugepage
e.g. from mmtests
./run-mmtests.sh --run-monitor --config configs/config-global-dhp__workload_thpfioscale test-run-1
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch).
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameter create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed.
3. Warm up a number of fio read-only processes accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll refault old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds.
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup the test files.
Note that due to the use of IO and page cache that this benchmark is not
suitable for running on large machines where the time to fragment memory
may be excessive. Also note that while this is one mix that generates
fragmentation that it's not the only mix that generates fragmentation.
Differences in workload that are more slab-intensive or whether SLUB is
used with high-order pages may yield different results.
When the page allocator fragments memory, it records the event using the
mm_page_alloc_extfrag ftrace event. If the fallback_order is smaller than
a pageblock order (order-9 on 64-bit x86) then it's considered to be an
"external fragmentation event" that may cause issues in the future.
Hence, the primary metric here is the number of external fragmentation
events that occur with order < 9. The secondary metric is allocation
latency and huge page allocation success rates but note that differences
in latencies and what the success rate also can affect the number of
external fragmentation event which is why it's a secondary metric.
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 662.92 ( 0.00%) 653.58 * 1.41%*
Amean fault-huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
Fault latencies are slightly reduced while allocation success rates remain
at zero as this configuration does not make any special effort to allocate
THP and fio is heavily active at the time and either filling memory or
keeping pages resident. However, a 49% reduction of serious fragmentation
events reduces the changes of external fragmentation being a problem in
the future.
Vlastimil asked during review for a breakdown of the allocation types
that are falling back.
vanilla
3816 MIGRATE_UNMOVABLE
800845 MIGRATE_MOVABLE
33 MIGRATE_UNRECLAIMABLE
patch
735 MIGRATE_UNMOVABLE
408135 MIGRATE_MOVABLE
42 MIGRATE_UNRECLAIMABLE
The majority of the fallbacks are due to movable allocations and this is
consistent for the workload throughout the series so will not be presented
again as the primary source of fallbacks are movable allocations.
Movable fallbacks are sometimes considered "ok" to fallback because they
can be migrated. The problem is that they can fill an
unmovable/reclaimable pageblock causing those allocations to fallback
later and polluting pageblocks with pages that cannot move. If there is a
movable fallback, it is pretty much guaranteed to affect an
unmovable/reclaimable pageblock and while it might not be enough to
actually cause a unmovable/reclaimable fallback in the future, we cannot
know that in advance so the patch takes the only option available to it.
Hence, it's important to control them. This point is also consistent
throughout the series and will not be repeated.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 1495.14 ( 0.00%) 1467.55 ( 1.85%)
Amean fault-huge-1 1098.48 ( 0.00%) 1127.11 ( -2.61%)
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 78.57 ( 0.00%) 77.64 ( -1.18%)
Fragmentation events were reduced quite a bit although this is known
to be a little variable. The latencies and allocation success rates
are similar but they were already quite high.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 1350.05 ( 0.00%) 1346.45 ( 0.27%)
Amean fault-huge-5 4181.01 ( 0.00%) 3418.60 ( 18.24%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 1.15 ( 0.00%) 0.78 ( -31.88%)
The reduction of external fragmentation events is slight and this is
partially due to the removal of __GFP_THISNODE in commit ac5b2c18911f
("mm: thp: relax __GFP_THISNODE for MADV_HUGEPAGE mappings") as THP
allocations can now spill over to remote nodes instead of fragmenting
local memory.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 6138.97 ( 0.00%) 6217.43 ( -1.28%)
Amean fault-huge-5 2294.28 ( 0.00%) 3163.33 * -37.88%*
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 96.82 ( 0.00%) 95.14 ( -1.74%)
There was a slight reduction in external fragmentation events although the
latencies were higher. The allocation success rate is high enough that
the system is struggling and there is quite a lot of parallel reclaim and
compaction activity. There is also a certain degree of luck on whether
processes start on node 0 or not for this patch but the relevance is
reduced later in the series.
Overall, the patch reduces the number of external fragmentation causing
events so the success of THP over long periods of time would be improved
for this adverse workload.
Link: http://lkml.kernel.org/r/20181123114528.28802-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:41 -08:00
|
|
|
|
2007-10-16 01:25:48 -07:00
|
|
|
|
2015-04-14 15:45:18 -07:00
|
|
|
|
2007-10-16 01:25:48 -07:00
|
|
|
|
mm, page_alloc: spread allocations across zones before introducing fragmentation
Patch series "Fragmentation avoidance improvements", v5.
It has been noted before that fragmentation avoidance (aka
anti-fragmentation) is not perfect. Given sufficient time or an adverse
workload, memory gets fragmented and the long-term success of high-order
allocations degrades. This series defines an adverse workload, a definition
of external fragmentation events (including serious) ones and a series
that reduces the level of those fragmentation events.
The details of the workload and the consequences are described in more
detail in the changelogs. However, from patch 1, this is a high-level
summary of the adverse workload. The exact details are found in the
mmtests implementation.
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch)
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameterr create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed
3. Warm up a number of fio read-only threads accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll fault back in old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup
Overall the series reduces external fragmentation causing events by over 94%
on 1 and 2 socket machines, which in turn impacts high-order allocation
success rates over the long term. There are differences in latencies and
high-order allocation success rates. Latencies are a mixed bag as they
are vulnerable to exact system state and whether allocations succeeded
so they are treated as a secondary metric.
Patch 1 uses lower zones if they are populated and have free memory
instead of fragmenting a higher zone. It's special cased to
handle a Normal->DMA32 fallback with the reasons explained
in the changelog.
Patch 2-4 boosts watermarks temporarily when an external fragmentation
event occurs. kswapd wakes to reclaim a small amount of old memory
and then wakes kcompactd on completion to recover the system
slightly. This introduces some overhead in the slowpath. The level
of boosting can be tuned or disabled depending on the tolerance
for fragmentation vs allocation latency.
Patch 5 stalls some movable allocation requests to let kswapd from patch 4
make some progress. The duration of the stalls is very low but it
is possible to tune the system to avoid fragmentation events if
larger stalls can be tolerated.
The bulk of the improvement in fragmentation avoidance is from patches
1-4 but patch 5 can deal with a rare corner case and provides the option
of tuning a system for THP allocation success rates in exchange for
some stalls to control fragmentation.
This patch (of 5):
The page allocator zone lists are iterated based on the watermarks of each
zone which does not take anti-fragmentation into account. On x86, node 0
may have multiple zones while other nodes have one zone. A consequence is
that tasks running on node 0 may fragment ZONE_NORMAL even though
ZONE_DMA32 has plenty of free memory. This patch special cases the
allocator fast path such that it'll try an allocation from a lower local
zone before fragmenting a higher zone. In this case, stealing of
pageblocks or orders larger than a pageblock are still allowed in the fast
path as they are uninteresting from a fragmentation point of view.
This was evaluated using a benchmark designed to fragment memory before
attempting THP allocations. It's implemented in mmtests as the following
configurations
configs/config-global-dhp__workload_thpfioscale
configs/config-global-dhp__workload_thpfioscale-defrag
configs/config-global-dhp__workload_thpfioscale-madvhugepage
e.g. from mmtests
./run-mmtests.sh --run-monitor --config configs/config-global-dhp__workload_thpfioscale test-run-1
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch).
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameter create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed.
3. Warm up a number of fio read-only processes accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll refault old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds.
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup the test files.
Note that due to the use of IO and page cache that this benchmark is not
suitable for running on large machines where the time to fragment memory
may be excessive. Also note that while this is one mix that generates
fragmentation that it's not the only mix that generates fragmentation.
Differences in workload that are more slab-intensive or whether SLUB is
used with high-order pages may yield different results.
When the page allocator fragments memory, it records the event using the
mm_page_alloc_extfrag ftrace event. If the fallback_order is smaller than
a pageblock order (order-9 on 64-bit x86) then it's considered to be an
"external fragmentation event" that may cause issues in the future.
Hence, the primary metric here is the number of external fragmentation
events that occur with order < 9. The secondary metric is allocation
latency and huge page allocation success rates but note that differences
in latencies and what the success rate also can affect the number of
external fragmentation event which is why it's a secondary metric.
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 662.92 ( 0.00%) 653.58 * 1.41%*
Amean fault-huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
Fault latencies are slightly reduced while allocation success rates remain
at zero as this configuration does not make any special effort to allocate
THP and fio is heavily active at the time and either filling memory or
keeping pages resident. However, a 49% reduction of serious fragmentation
events reduces the changes of external fragmentation being a problem in
the future.
Vlastimil asked during review for a breakdown of the allocation types
that are falling back.
vanilla
3816 MIGRATE_UNMOVABLE
800845 MIGRATE_MOVABLE
33 MIGRATE_UNRECLAIMABLE
patch
735 MIGRATE_UNMOVABLE
408135 MIGRATE_MOVABLE
42 MIGRATE_UNRECLAIMABLE
The majority of the fallbacks are due to movable allocations and this is
consistent for the workload throughout the series so will not be presented
again as the primary source of fallbacks are movable allocations.
Movable fallbacks are sometimes considered "ok" to fallback because they
can be migrated. The problem is that they can fill an
unmovable/reclaimable pageblock causing those allocations to fallback
later and polluting pageblocks with pages that cannot move. If there is a
movable fallback, it is pretty much guaranteed to affect an
unmovable/reclaimable pageblock and while it might not be enough to
actually cause a unmovable/reclaimable fallback in the future, we cannot
know that in advance so the patch takes the only option available to it.
Hence, it's important to control them. This point is also consistent
throughout the series and will not be repeated.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 1495.14 ( 0.00%) 1467.55 ( 1.85%)
Amean fault-huge-1 1098.48 ( 0.00%) 1127.11 ( -2.61%)
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 78.57 ( 0.00%) 77.64 ( -1.18%)
Fragmentation events were reduced quite a bit although this is known
to be a little variable. The latencies and allocation success rates
are similar but they were already quite high.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 1350.05 ( 0.00%) 1346.45 ( 0.27%)
Amean fault-huge-5 4181.01 ( 0.00%) 3418.60 ( 18.24%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 1.15 ( 0.00%) 0.78 ( -31.88%)
The reduction of external fragmentation events is slight and this is
partially due to the removal of __GFP_THISNODE in commit ac5b2c18911f
("mm: thp: relax __GFP_THISNODE for MADV_HUGEPAGE mappings") as THP
allocations can now spill over to remote nodes instead of fragmenting
local memory.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 6138.97 ( 0.00%) 6217.43 ( -1.28%)
Amean fault-huge-5 2294.28 ( 0.00%) 3163.33 * -37.88%*
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 96.82 ( 0.00%) 95.14 ( -1.74%)
There was a slight reduction in external fragmentation events although the
latencies were higher. The allocation success rate is high enough that
the system is struggling and there is quite a lot of parallel reclaim and
compaction activity. There is also a certain degree of luck on whether
processes start on node 0 or not for this patch but the relevance is
reduced later in the series.
Overall, the patch reduces the number of external fragmentation causing
events so the success of THP over long periods of time would be improved
for this adverse workload.
Link: http://lkml.kernel.org/r/20181123114528.28802-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:41 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-08-03 10:51:21 +08:00
|
|
|
|
mm, page_alloc: spread allocations across zones before introducing fragmentation
Patch series "Fragmentation avoidance improvements", v5.
It has been noted before that fragmentation avoidance (aka
anti-fragmentation) is not perfect. Given sufficient time or an adverse
workload, memory gets fragmented and the long-term success of high-order
allocations degrades. This series defines an adverse workload, a definition
of external fragmentation events (including serious) ones and a series
that reduces the level of those fragmentation events.
The details of the workload and the consequences are described in more
detail in the changelogs. However, from patch 1, this is a high-level
summary of the adverse workload. The exact details are found in the
mmtests implementation.
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch)
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameterr create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed
3. Warm up a number of fio read-only threads accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll fault back in old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup
Overall the series reduces external fragmentation causing events by over 94%
on 1 and 2 socket machines, which in turn impacts high-order allocation
success rates over the long term. There are differences in latencies and
high-order allocation success rates. Latencies are a mixed bag as they
are vulnerable to exact system state and whether allocations succeeded
so they are treated as a secondary metric.
Patch 1 uses lower zones if they are populated and have free memory
instead of fragmenting a higher zone. It's special cased to
handle a Normal->DMA32 fallback with the reasons explained
in the changelog.
Patch 2-4 boosts watermarks temporarily when an external fragmentation
event occurs. kswapd wakes to reclaim a small amount of old memory
and then wakes kcompactd on completion to recover the system
slightly. This introduces some overhead in the slowpath. The level
of boosting can be tuned or disabled depending on the tolerance
for fragmentation vs allocation latency.
Patch 5 stalls some movable allocation requests to let kswapd from patch 4
make some progress. The duration of the stalls is very low but it
is possible to tune the system to avoid fragmentation events if
larger stalls can be tolerated.
The bulk of the improvement in fragmentation avoidance is from patches
1-4 but patch 5 can deal with a rare corner case and provides the option
of tuning a system for THP allocation success rates in exchange for
some stalls to control fragmentation.
This patch (of 5):
The page allocator zone lists are iterated based on the watermarks of each
zone which does not take anti-fragmentation into account. On x86, node 0
may have multiple zones while other nodes have one zone. A consequence is
that tasks running on node 0 may fragment ZONE_NORMAL even though
ZONE_DMA32 has plenty of free memory. This patch special cases the
allocator fast path such that it'll try an allocation from a lower local
zone before fragmenting a higher zone. In this case, stealing of
pageblocks or orders larger than a pageblock are still allowed in the fast
path as they are uninteresting from a fragmentation point of view.
This was evaluated using a benchmark designed to fragment memory before
attempting THP allocations. It's implemented in mmtests as the following
configurations
configs/config-global-dhp__workload_thpfioscale
configs/config-global-dhp__workload_thpfioscale-defrag
configs/config-global-dhp__workload_thpfioscale-madvhugepage
e.g. from mmtests
./run-mmtests.sh --run-monitor --config configs/config-global-dhp__workload_thpfioscale test-run-1
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch).
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameter create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed.
3. Warm up a number of fio read-only processes accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll refault old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds.
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup the test files.
Note that due to the use of IO and page cache that this benchmark is not
suitable for running on large machines where the time to fragment memory
may be excessive. Also note that while this is one mix that generates
fragmentation that it's not the only mix that generates fragmentation.
Differences in workload that are more slab-intensive or whether SLUB is
used with high-order pages may yield different results.
When the page allocator fragments memory, it records the event using the
mm_page_alloc_extfrag ftrace event. If the fallback_order is smaller than
a pageblock order (order-9 on 64-bit x86) then it's considered to be an
"external fragmentation event" that may cause issues in the future.
Hence, the primary metric here is the number of external fragmentation
events that occur with order < 9. The secondary metric is allocation
latency and huge page allocation success rates but note that differences
in latencies and what the success rate also can affect the number of
external fragmentation event which is why it's a secondary metric.
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 662.92 ( 0.00%) 653.58 * 1.41%*
Amean fault-huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
Fault latencies are slightly reduced while allocation success rates remain
at zero as this configuration does not make any special effort to allocate
THP and fio is heavily active at the time and either filling memory or
keeping pages resident. However, a 49% reduction of serious fragmentation
events reduces the changes of external fragmentation being a problem in
the future.
Vlastimil asked during review for a breakdown of the allocation types
that are falling back.
vanilla
3816 MIGRATE_UNMOVABLE
800845 MIGRATE_MOVABLE
33 MIGRATE_UNRECLAIMABLE
patch
735 MIGRATE_UNMOVABLE
408135 MIGRATE_MOVABLE
42 MIGRATE_UNRECLAIMABLE
The majority of the fallbacks are due to movable allocations and this is
consistent for the workload throughout the series so will not be presented
again as the primary source of fallbacks are movable allocations.
Movable fallbacks are sometimes considered "ok" to fallback because they
can be migrated. The problem is that they can fill an
unmovable/reclaimable pageblock causing those allocations to fallback
later and polluting pageblocks with pages that cannot move. If there is a
movable fallback, it is pretty much guaranteed to affect an
unmovable/reclaimable pageblock and while it might not be enough to
actually cause a unmovable/reclaimable fallback in the future, we cannot
know that in advance so the patch takes the only option available to it.
Hence, it's important to control them. This point is also consistent
throughout the series and will not be repeated.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 1495.14 ( 0.00%) 1467.55 ( 1.85%)
Amean fault-huge-1 1098.48 ( 0.00%) 1127.11 ( -2.61%)
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 78.57 ( 0.00%) 77.64 ( -1.18%)
Fragmentation events were reduced quite a bit although this is known
to be a little variable. The latencies and allocation success rates
are similar but they were already quite high.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 1350.05 ( 0.00%) 1346.45 ( 0.27%)
Amean fault-huge-5 4181.01 ( 0.00%) 3418.60 ( 18.24%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 1.15 ( 0.00%) 0.78 ( -31.88%)
The reduction of external fragmentation events is slight and this is
partially due to the removal of __GFP_THISNODE in commit ac5b2c18911f
("mm: thp: relax __GFP_THISNODE for MADV_HUGEPAGE mappings") as THP
allocations can now spill over to remote nodes instead of fragmenting
local memory.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 6138.97 ( 0.00%) 6217.43 ( -1.28%)
Amean fault-huge-5 2294.28 ( 0.00%) 3163.33 * -37.88%*
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 96.82 ( 0.00%) 95.14 ( -1.74%)
There was a slight reduction in external fragmentation events although the
latencies were higher. The allocation success rate is high enough that
the system is struggling and there is quite a lot of parallel reclaim and
compaction activity. There is also a certain degree of luck on whether
processes start on node 0 or not for this patch but the relevance is
reduced later in the series.
Overall, the patch reduces the number of external fragmentation causing
events so the success of THP over long periods of time would be improved
for this adverse workload.
Link: http://lkml.kernel.org/r/20181123114528.28802-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:41 -08:00
|
|
|
|
|
|
|
|
|
mm, page_alloc: fallback to smallest page when not stealing whole pageblock
Since commit 3bc48f96cf11 ("mm, page_alloc: split smallest stolen page
in fallback") we pick the smallest (but sufficient) page of all that
have been stolen from a pageblock of different migratetype. However,
there are cases when we decide not to steal the whole pageblock.
Practically in the current implementation it means that we are trying to
fallback for a MIGRATE_MOVABLE allocation of order X, go through the
freelists from MAX_ORDER-1 down to X, and find free page of order Y. If
Y is less than pageblock_order / 2, we decide not to steal all pages
from the pageblock. When Y > X, it means we are potentially splitting a
larger page than we need, as there might be other pages of order Z,
where X <= Z < Y. Since Y is already too small to steal whole
pageblock, picking smallest available Z will result in the same decision
and we avoid splitting a higher-order page in a MIGRATE_UNMOVABLE or
MIGRATE_RECLAIMABLE pageblock.
This patch therefore changes the fallback algorithm so that in the
situation described above, we switch the fallback search strategy to go
from order X upwards to find the smallest suitable fallback. In theory
there shouldn't be a downside of this change wrt fragmentation.
This has been tested with mmtests' stress-highalloc performing
GFP_KERNEL order-4 allocations, here is the relevant extfrag tracepoint
statistics:
4.12.0-rc2 4.12.0-rc2
1-kernel4 2-kernel4
Page alloc extfrag event 25640976 69680977
Extfrag fragmenting 25621086 69661364
Extfrag fragmenting for unmovable 74409 73204
Extfrag fragmenting unmovable placed with movable 69003 67684
Extfrag fragmenting unmovable placed with reclaim. 5406 5520
Extfrag fragmenting for reclaimable 6398 8467
Extfrag fragmenting reclaimable placed with movable 869 884
Extfrag fragmenting reclaimable placed with unmov. 5529 7583
Extfrag fragmenting for movable 25540279 69579693
Since we force movable allocations to steal the smallest available page
(which we then practially always split), we steal less per fallback, so
the number of fallbacks increases and steals potentially happen from
different pageblocks. This is however not an issue for movable pages
that can be compacted.
Importantly, the "unmovable placed with movable" statistics is lower,
which is the result of less fragmentation in the unmovable pageblocks.
The effect on reclaimable allocation is a bit unclear.
Link: http://lkml.kernel.org/r/20170529093947.22618-1-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-10 15:47:14 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-12-28 17:47:04 +03:00
|
|
|
|
2014-06-04 16:10:21 -07:00
|
|
|
|
2015-04-14 15:45:18 -07:00
|
|
|
|
|
|
|
|
|
mm: page_alloc: tighten up find_suitable_fallback()
find_suitable_fallback() is not as efficient as it could be, and somewhat
difficult to follow.
1. should_try_claim_block() is a loop invariant. There is no point in
checking fallback areas if the caller is interested in claimable
blocks but the order and the migratetype don't allow for that.
2. __rmqueue_steal() doesn't care about claimability, so it shouldn't
have to run those tests.
Different callers want different things from this helper:
1. __compact_finished() scans orders up until it finds a claimable block
2. __rmqueue_claim() scans orders down as long as blocks are claimable
3. __rmqueue_steal() doesn't care about claimability at all
Move should_try_claim_block() out of the loop. Only test it for the
two callers who care in the first place. Distinguish "no blocks" from
"order + mt are not claimable" in the return value; __rmqueue_claim()
can stop once order becomes unclaimable, __compact_finished() can keep
advancing until order becomes claimable.
Before:
Performance counter stats for './run case-lru-file-mmap-read' (5 runs):
85,294.85 msec task-clock # 5.644 CPUs utilized ( +- 0.32% )
15,968 context-switches # 187.209 /sec ( +- 3.81% )
153 cpu-migrations # 1.794 /sec ( +- 3.29% )
801,808 page-faults # 9.400 K/sec ( +- 0.10% )
733,358,331,786 instructions # 1.87 insn per cycle ( +- 0.20% ) (64.94%)
392,622,904,199 cycles # 4.603 GHz ( +- 0.31% ) (64.84%)
148,563,488,531 branches # 1.742 G/sec ( +- 0.18% ) (63.86%)
152,143,228 branch-misses # 0.10% of all branches ( +- 1.19% ) (62.82%)
15.1128 +- 0.0637 seconds time elapsed ( +- 0.42% )
After:
Performance counter stats for './run case-lru-file-mmap-read' (5 runs):
84,380.21 msec task-clock # 5.664 CPUs utilized ( +- 0.21% )
16,656 context-switches # 197.392 /sec ( +- 3.27% )
151 cpu-migrations # 1.790 /sec ( +- 3.28% )
801,703 page-faults # 9.501 K/sec ( +- 0.09% )
731,914,183,060 instructions # 1.88 insn per cycle ( +- 0.38% ) (64.90%)
388,673,535,116 cycles # 4.606 GHz ( +- 0.24% ) (65.06%)
148,251,482,143 branches # 1.757 G/sec ( +- 0.37% ) (63.92%)
149,766,550 branch-misses # 0.10% of all branches ( +- 1.22% ) (62.88%)
14.8968 +- 0.0486 seconds time elapsed ( +- 0.33% )
Link: https://lkml.kernel.org/r/20250407180154.63348-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Tested-by: Shivank Garg <shivankg@amd.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Carlos Song <carlos.song@nxp.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-07 14:01:54 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-04-14 15:45:18 -07:00
|
|
|
|
|
|
|
|
|
2007-10-16 01:25:48 -07:00
|
|
|
|
mm: page_alloc: tighten up find_suitable_fallback()
find_suitable_fallback() is not as efficient as it could be, and somewhat
difficult to follow.
1. should_try_claim_block() is a loop invariant. There is no point in
checking fallback areas if the caller is interested in claimable
blocks but the order and the migratetype don't allow for that.
2. __rmqueue_steal() doesn't care about claimability, so it shouldn't
have to run those tests.
Different callers want different things from this helper:
1. __compact_finished() scans orders up until it finds a claimable block
2. __rmqueue_claim() scans orders down as long as blocks are claimable
3. __rmqueue_steal() doesn't care about claimability at all
Move should_try_claim_block() out of the loop. Only test it for the
two callers who care in the first place. Distinguish "no blocks" from
"order + mt are not claimable" in the return value; __rmqueue_claim()
can stop once order becomes unclaimable, __compact_finished() can keep
advancing until order becomes claimable.
Before:
Performance counter stats for './run case-lru-file-mmap-read' (5 runs):
85,294.85 msec task-clock # 5.644 CPUs utilized ( +- 0.32% )
15,968 context-switches # 187.209 /sec ( +- 3.81% )
153 cpu-migrations # 1.794 /sec ( +- 3.29% )
801,808 page-faults # 9.400 K/sec ( +- 0.10% )
733,358,331,786 instructions # 1.87 insn per cycle ( +- 0.20% ) (64.94%)
392,622,904,199 cycles # 4.603 GHz ( +- 0.31% ) (64.84%)
148,563,488,531 branches # 1.742 G/sec ( +- 0.18% ) (63.86%)
152,143,228 branch-misses # 0.10% of all branches ( +- 1.19% ) (62.82%)
15.1128 +- 0.0637 seconds time elapsed ( +- 0.42% )
After:
Performance counter stats for './run case-lru-file-mmap-read' (5 runs):
84,380.21 msec task-clock # 5.664 CPUs utilized ( +- 0.21% )
16,656 context-switches # 197.392 /sec ( +- 3.27% )
151 cpu-migrations # 1.790 /sec ( +- 3.28% )
801,703 page-faults # 9.501 K/sec ( +- 0.09% )
731,914,183,060 instructions # 1.88 insn per cycle ( +- 0.38% ) (64.90%)
388,673,535,116 cycles # 4.606 GHz ( +- 0.24% ) (65.06%)
148,251,482,143 branches # 1.757 G/sec ( +- 0.37% ) (63.92%)
149,766,550 branch-misses # 0.10% of all branches ( +- 1.22% ) (62.88%)
14.8968 +- 0.0486 seconds time elapsed ( +- 0.33% )
Link: https://lkml.kernel.org/r/20250407180154.63348-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Tested-by: Shivank Garg <shivankg@amd.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Carlos Song <carlos.song@nxp.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-07 14:01:54 -04:00
|
|
|
|
|
|
|
|
|
mm: page_alloc: don't steal single pages from biggest buddy
The fallback code searches for the biggest buddy first in an attempt to
steal the whole block and encourage type grouping down the line.
The approach used to be this:
- Non-movable requests will split the largest buddy and steal the
remainder. This splits up contiguity, but it allows subsequent
requests of this type to fall back into adjacent space.
- Movable requests go and look for the smallest buddy instead. The
thinking is that movable requests can be compacted, so grouping is
less important than retaining contiguity.
c0cd6f557b90 ("mm: page_alloc: fix freelist movement during block
conversion") enforces freelist type hygiene, which restricts stealing to
either claiming the whole block or just taking the requested chunk; no
additional pages or buddy remainders can be stolen any more.
The patch mishandled when to switch to finding the smallest buddy in that
new reality. As a result, it may steal the exact request size, but from
the biggest buddy. This causes fracturing for no good reason.
Fix this by committing to the new behavior: either steal the whole block,
or fall back to the smallest buddy.
Remove single-page stealing from steal_suitable_fallback(). Rename it to
try_to_steal_block() to make the intentions clear. If this fails, always
fall back to the smallest buddy.
The following is from 4 runs of mmtest's thpchallenge. "Pollute" is
single page fallback, "steal" is conversion of a partially used block.
The numbers for free block conversions (omitted) are comparable.
vanilla patched
@pollute[unmovable from reclaimable]: 27 106
@pollute[unmovable from movable]: 82 46
@pollute[reclaimable from unmovable]: 256 83
@pollute[reclaimable from movable]: 46 8
@pollute[movable from unmovable]: 4841 868
@pollute[movable from reclaimable]: 5278 12568
@steal[unmovable from reclaimable]: 11 12
@steal[unmovable from movable]: 113 49
@steal[reclaimable from unmovable]: 19 34
@steal[reclaimable from movable]: 47 21
@steal[movable from unmovable]: 250 183
@steal[movable from reclaimable]: 81 93
The allocator appears to do a better job at keeping stealing and polluting
to the first fallback preference. As a result, the numbers for "from
movable" - the least preferred fallback option, and most detrimental to
compactability - are down across the board.
Link: https://lkml.kernel.org/r/20250225001023.1494422-2-hannes@cmpxchg.org
Fixes: c0cd6f557b90 ("mm: page_alloc: fix freelist movement during block conversion")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-02-24 19:08:24 -05:00
|
|
|
|
2007-10-16 01:25:53 -07:00
|
|
|
|
mm: page_alloc: don't steal single pages from biggest buddy
The fallback code searches for the biggest buddy first in an attempt to
steal the whole block and encourage type grouping down the line.
The approach used to be this:
- Non-movable requests will split the largest buddy and steal the
remainder. This splits up contiguity, but it allows subsequent
requests of this type to fall back into adjacent space.
- Movable requests go and look for the smallest buddy instead. The
thinking is that movable requests can be compacted, so grouping is
less important than retaining contiguity.
c0cd6f557b90 ("mm: page_alloc: fix freelist movement during block
conversion") enforces freelist type hygiene, which restricts stealing to
either claiming the whole block or just taking the requested chunk; no
additional pages or buddy remainders can be stolen any more.
The patch mishandled when to switch to finding the smallest buddy in that
new reality. As a result, it may steal the exact request size, but from
the biggest buddy. This causes fracturing for no good reason.
Fix this by committing to the new behavior: either steal the whole block,
or fall back to the smallest buddy.
Remove single-page stealing from steal_suitable_fallback(). Rename it to
try_to_steal_block() to make the intentions clear. If this fails, always
fall back to the smallest buddy.
The following is from 4 runs of mmtest's thpchallenge. "Pollute" is
single page fallback, "steal" is conversion of a partially used block.
The numbers for free block conversions (omitted) are comparable.
vanilla patched
@pollute[unmovable from reclaimable]: 27 106
@pollute[unmovable from movable]: 82 46
@pollute[reclaimable from unmovable]: 256 83
@pollute[reclaimable from movable]: 46 8
@pollute[movable from unmovable]: 4841 868
@pollute[movable from reclaimable]: 5278 12568
@steal[unmovable from reclaimable]: 11 12
@steal[unmovable from movable]: 113 49
@steal[reclaimable from unmovable]: 19 34
@steal[reclaimable from movable]: 47 21
@steal[movable from unmovable]: 250 183
@steal[movable from reclaimable]: 81 93
The allocator appears to do a better job at keeping stealing and polluting
to the first fallback preference. As a result, the numbers for "from
movable" - the least preferred fallback option, and most detrimental to
compactability - are down across the board.
Link: https://lkml.kernel.org/r/20250225001023.1494422-2-hannes@cmpxchg.org
Fixes: c0cd6f557b90 ("mm: page_alloc: fix freelist movement during block conversion")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-02-24 19:08:24 -05:00
|
|
|
|
2025-02-28 09:52:17 +00:00
|
|
|
|
2025-02-24 19:08:25 -05:00
|
|
|
|
|
|
|
|
|
mm: page_alloc: speed up fallbacks in rmqueue_bulk()
The test robot identified c2f6ea38fc1b ("mm: page_alloc: don't steal
single pages from biggest buddy") as the root cause of a 56.4% regression
in vm-scalability::lru-file-mmap-read.
Carlos reports an earlier patch, c0cd6f557b90 ("mm: page_alloc: fix
freelist movement during block conversion"), as the root cause for a
regression in worst-case zone->lock+irqoff hold times.
Both of these patches modify the page allocator's fallback path to be less
greedy in an effort to stave off fragmentation. The flip side of this is
that fallbacks are also less productive each time around, which means the
fallback search can run much more frequently.
Carlos' traces point to rmqueue_bulk() specifically, which tries to refill
the percpu cache by allocating a large batch of pages in a loop. It
highlights how once the native freelists are exhausted, the fallback code
first scans orders top-down for whole blocks to claim, then falls back to
a bottom-up search for the smallest buddy to steal. For the next batch
page, it goes through the same thing again.
This can be made more efficient. Since rmqueue_bulk() holds the
zone->lock over the entire batch, the freelists are not subject to outside
changes; when the search for a block to claim has already failed, there is
no point in trying again for the next page.
Modify __rmqueue() to remember the last successful fallback mode, and
restart directly from there on the next rmqueue_bulk() iteration.
Oliver confirms that this improves beyond the regression that the test
robot reported against c2f6ea38fc1b:
commit:
f3b92176f4 ("tools/selftests: add guard region test for /proc/$pid/pagemap")
c2f6ea38fc ("mm: page_alloc: don't steal single pages from biggest buddy")
acc4d5ff0b ("Merge tag 'net-6.15-rc0' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net")
2c847f27c3 ("mm: page_alloc: speed up fallbacks in rmqueue_bulk()") <--- your patch
f3b92176f4f7100f c2f6ea38fc1b640aa7a2e155cc1 acc4d5ff0b61eb1715c498b6536 2c847f27c37da65a93d23c237c5
---------------- --------------------------- --------------------------- ---------------------------
%stddev %change %stddev %change %stddev %change %stddev
\ | \ | \ | \
25525364 ± 3% -56.4% 11135467 -57.8% 10779336 +31.6% 33581409 vm-scalability.throughput
Carlos confirms that worst-case times are almost fully recovered
compared to before the earlier culprit patch:
2dd482ba627d (before freelist hygiene): 1ms
c0cd6f557b90 (after freelist hygiene): 90ms
next-20250319 (steal smallest buddy): 280ms
this patch : 8ms
[jackmanb@google.com: comment updates]
Link: https://lkml.kernel.org/r/D92AC0P9594X.3BML64MUKTF8Z@google.com
[hannes@cmpxchg.org: reset rmqueue_mode in rmqueue_buddy() error loop, per Yunsheng Lin]
Link: https://lkml.kernel.org/r/20250409140023.GA2313@cmpxchg.org
Link: https://lkml.kernel.org/r/20250407180154.63348-1-hannes@cmpxchg.org
Fixes: c0cd6f557b90 ("mm: page_alloc: fix freelist movement during block conversion")
Fixes: c2f6ea38fc1b ("mm: page_alloc: don't steal single pages from biggest buddy")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Brendan Jackman <jackmanb@google.com>
Reported-by: kernel test robot <oliver.sang@intel.com>
Reported-by: Carlos Song <carlos.song@nxp.com>
Tested-by: Carlos Song <carlos.song@nxp.com>
Tested-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202503271547.fc08b188-lkp@intel.com
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Tested-by: Shivank Garg <shivankg@amd.com>
Acked-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: <stable@vger.kernel.org> [6.10+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-07 14:01:53 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-10-16 01:25:48 -07:00
|
|
|
|
|
|
|
|
|
mm: page_alloc: speed up fallbacks in rmqueue_bulk()
The test robot identified c2f6ea38fc1b ("mm: page_alloc: don't steal
single pages from biggest buddy") as the root cause of a 56.4% regression
in vm-scalability::lru-file-mmap-read.
Carlos reports an earlier patch, c0cd6f557b90 ("mm: page_alloc: fix
freelist movement during block conversion"), as the root cause for a
regression in worst-case zone->lock+irqoff hold times.
Both of these patches modify the page allocator's fallback path to be less
greedy in an effort to stave off fragmentation. The flip side of this is
that fallbacks are also less productive each time around, which means the
fallback search can run much more frequently.
Carlos' traces point to rmqueue_bulk() specifically, which tries to refill
the percpu cache by allocating a large batch of pages in a loop. It
highlights how once the native freelists are exhausted, the fallback code
first scans orders top-down for whole blocks to claim, then falls back to
a bottom-up search for the smallest buddy to steal. For the next batch
page, it goes through the same thing again.
This can be made more efficient. Since rmqueue_bulk() holds the
zone->lock over the entire batch, the freelists are not subject to outside
changes; when the search for a block to claim has already failed, there is
no point in trying again for the next page.
Modify __rmqueue() to remember the last successful fallback mode, and
restart directly from there on the next rmqueue_bulk() iteration.
Oliver confirms that this improves beyond the regression that the test
robot reported against c2f6ea38fc1b:
commit:
f3b92176f4 ("tools/selftests: add guard region test for /proc/$pid/pagemap")
c2f6ea38fc ("mm: page_alloc: don't steal single pages from biggest buddy")
acc4d5ff0b ("Merge tag 'net-6.15-rc0' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net")
2c847f27c3 ("mm: page_alloc: speed up fallbacks in rmqueue_bulk()") <--- your patch
f3b92176f4f7100f c2f6ea38fc1b640aa7a2e155cc1 acc4d5ff0b61eb1715c498b6536 2c847f27c37da65a93d23c237c5
---------------- --------------------------- --------------------------- ---------------------------
%stddev %change %stddev %change %stddev %change %stddev
\ | \ | \ | \
25525364 ± 3% -56.4% 11135467 -57.8% 10779336 +31.6% 33581409 vm-scalability.throughput
Carlos confirms that worst-case times are almost fully recovered
compared to before the earlier culprit patch:
2dd482ba627d (before freelist hygiene): 1ms
c0cd6f557b90 (after freelist hygiene): 90ms
next-20250319 (steal smallest buddy): 280ms
this patch : 8ms
[jackmanb@google.com: comment updates]
Link: https://lkml.kernel.org/r/D92AC0P9594X.3BML64MUKTF8Z@google.com
[hannes@cmpxchg.org: reset rmqueue_mode in rmqueue_buddy() error loop, per Yunsheng Lin]
Link: https://lkml.kernel.org/r/20250409140023.GA2313@cmpxchg.org
Link: https://lkml.kernel.org/r/20250407180154.63348-1-hannes@cmpxchg.org
Fixes: c0cd6f557b90 ("mm: page_alloc: fix freelist movement during block conversion")
Fixes: c2f6ea38fc1b ("mm: page_alloc: don't steal single pages from biggest buddy")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Brendan Jackman <jackmanb@google.com>
Reported-by: kernel test robot <oliver.sang@intel.com>
Reported-by: Carlos Song <carlos.song@nxp.com>
Tested-by: Carlos Song <carlos.song@nxp.com>
Tested-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202503271547.fc08b188-lkp@intel.com
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Tested-by: Shivank Garg <shivankg@amd.com>
Acked-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: <stable@vger.kernel.org> [6.10+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-07 14:01:53 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, page_alloc: fallback to smallest page when not stealing whole pageblock
Since commit 3bc48f96cf11 ("mm, page_alloc: split smallest stolen page
in fallback") we pick the smallest (but sufficient) page of all that
have been stolen from a pageblock of different migratetype. However,
there are cases when we decide not to steal the whole pageblock.
Practically in the current implementation it means that we are trying to
fallback for a MIGRATE_MOVABLE allocation of order X, go through the
freelists from MAX_ORDER-1 down to X, and find free page of order Y. If
Y is less than pageblock_order / 2, we decide not to steal all pages
from the pageblock. When Y > X, it means we are potentially splitting a
larger page than we need, as there might be other pages of order Z,
where X <= Z < Y. Since Y is already too small to steal whole
pageblock, picking smallest available Z will result in the same decision
and we avoid splitting a higher-order page in a MIGRATE_UNMOVABLE or
MIGRATE_RECLAIMABLE pageblock.
This patch therefore changes the fallback algorithm so that in the
situation described above, we switch the fallback search strategy to go
from order X upwards to find the smallest suitable fallback. In theory
there shouldn't be a downside of this change wrt fragmentation.
This has been tested with mmtests' stress-highalloc performing
GFP_KERNEL order-4 allocations, here is the relevant extfrag tracepoint
statistics:
4.12.0-rc2 4.12.0-rc2
1-kernel4 2-kernel4
Page alloc extfrag event 25640976 69680977
Extfrag fragmenting 25621086 69661364
Extfrag fragmenting for unmovable 74409 73204
Extfrag fragmenting unmovable placed with movable 69003 67684
Extfrag fragmenting unmovable placed with reclaim. 5406 5520
Extfrag fragmenting for reclaimable 6398 8467
Extfrag fragmenting reclaimable placed with movable 869 884
Extfrag fragmenting reclaimable placed with unmov. 5529 7583
Extfrag fragmenting for movable 25540279 69579693
Since we force movable allocations to steal the smallest available page
(which we then practially always split), we steal less per fallback, so
the number of fallbacks increases and steals potentially happen from
different pageblocks. This is however not an issue for movable pages
that can be compacted.
Importantly, the "unmovable placed with movable" statistics is lower,
which is the result of less fragmentation in the unmovable pageblocks.
The effect on reclaimable allocation is a bit unclear.
Link: http://lkml.kernel.org/r/20170529093947.22618-1-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-10 15:47:14 -07:00
|
|
|
|
2023-12-28 17:47:03 +03:00
|
|
|
|
mm, page_alloc: fallback to smallest page when not stealing whole pageblock
Since commit 3bc48f96cf11 ("mm, page_alloc: split smallest stolen page
in fallback") we pick the smallest (but sufficient) page of all that
have been stolen from a pageblock of different migratetype. However,
there are cases when we decide not to steal the whole pageblock.
Practically in the current implementation it means that we are trying to
fallback for a MIGRATE_MOVABLE allocation of order X, go through the
freelists from MAX_ORDER-1 down to X, and find free page of order Y. If
Y is less than pageblock_order / 2, we decide not to steal all pages
from the pageblock. When Y > X, it means we are potentially splitting a
larger page than we need, as there might be other pages of order Z,
where X <= Z < Y. Since Y is already too small to steal whole
pageblock, picking smallest available Z will result in the same decision
and we avoid splitting a higher-order page in a MIGRATE_UNMOVABLE or
MIGRATE_RECLAIMABLE pageblock.
This patch therefore changes the fallback algorithm so that in the
situation described above, we switch the fallback search strategy to go
from order X upwards to find the smallest suitable fallback. In theory
there shouldn't be a downside of this change wrt fragmentation.
This has been tested with mmtests' stress-highalloc performing
GFP_KERNEL order-4 allocations, here is the relevant extfrag tracepoint
statistics:
4.12.0-rc2 4.12.0-rc2
1-kernel4 2-kernel4
Page alloc extfrag event 25640976 69680977
Extfrag fragmenting 25621086 69661364
Extfrag fragmenting for unmovable 74409 73204
Extfrag fragmenting unmovable placed with movable 69003 67684
Extfrag fragmenting unmovable placed with reclaim. 5406 5520
Extfrag fragmenting for reclaimable 6398 8467
Extfrag fragmenting reclaimable placed with movable 869 884
Extfrag fragmenting reclaimable placed with unmov. 5529 7583
Extfrag fragmenting for movable 25540279 69579693
Since we force movable allocations to steal the smallest available page
(which we then practially always split), we steal less per fallback, so
the number of fallbacks increases and steals potentially happen from
different pageblocks. This is however not an issue for movable pages
that can be compacted.
Importantly, the "unmovable placed with movable" statistics is lower,
which is the result of less fragmentation in the unmovable pageblocks.
The effect on reclaimable allocation is a bit unclear.
Link: http://lkml.kernel.org/r/20170529093947.22618-1-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-10 15:47:14 -07:00
|
|
|
|
|
|
|
|
|
mm: page_alloc: tighten up find_suitable_fallback()
find_suitable_fallback() is not as efficient as it could be, and somewhat
difficult to follow.
1. should_try_claim_block() is a loop invariant. There is no point in
checking fallback areas if the caller is interested in claimable
blocks but the order and the migratetype don't allow for that.
2. __rmqueue_steal() doesn't care about claimability, so it shouldn't
have to run those tests.
Different callers want different things from this helper:
1. __compact_finished() scans orders up until it finds a claimable block
2. __rmqueue_claim() scans orders down as long as blocks are claimable
3. __rmqueue_steal() doesn't care about claimability at all
Move should_try_claim_block() out of the loop. Only test it for the
two callers who care in the first place. Distinguish "no blocks" from
"order + mt are not claimable" in the return value; __rmqueue_claim()
can stop once order becomes unclaimable, __compact_finished() can keep
advancing until order becomes claimable.
Before:
Performance counter stats for './run case-lru-file-mmap-read' (5 runs):
85,294.85 msec task-clock # 5.644 CPUs utilized ( +- 0.32% )
15,968 context-switches # 187.209 /sec ( +- 3.81% )
153 cpu-migrations # 1.794 /sec ( +- 3.29% )
801,808 page-faults # 9.400 K/sec ( +- 0.10% )
733,358,331,786 instructions # 1.87 insn per cycle ( +- 0.20% ) (64.94%)
392,622,904,199 cycles # 4.603 GHz ( +- 0.31% ) (64.84%)
148,563,488,531 branches # 1.742 G/sec ( +- 0.18% ) (63.86%)
152,143,228 branch-misses # 0.10% of all branches ( +- 1.19% ) (62.82%)
15.1128 +- 0.0637 seconds time elapsed ( +- 0.42% )
After:
Performance counter stats for './run case-lru-file-mmap-read' (5 runs):
84,380.21 msec task-clock # 5.664 CPUs utilized ( +- 0.21% )
16,656 context-switches # 197.392 /sec ( +- 3.27% )
151 cpu-migrations # 1.790 /sec ( +- 3.28% )
801,703 page-faults # 9.501 K/sec ( +- 0.09% )
731,914,183,060 instructions # 1.88 insn per cycle ( +- 0.38% ) (64.90%)
388,673,535,116 cycles # 4.606 GHz ( +- 0.24% ) (65.06%)
148,251,482,143 branches # 1.757 G/sec ( +- 0.37% ) (63.92%)
149,766,550 branch-misses # 0.10% of all branches ( +- 1.22% ) (62.88%)
14.8968 +- 0.0486 seconds time elapsed ( +- 0.33% )
Link: https://lkml.kernel.org/r/20250407180154.63348-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Tested-by: Shivank Garg <shivankg@amd.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Carlos Song <carlos.song@nxp.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-07 14:01:54 -04:00
|
|
|
|
mm: page_alloc: don't steal single pages from biggest buddy
The fallback code searches for the biggest buddy first in an attempt to
steal the whole block and encourage type grouping down the line.
The approach used to be this:
- Non-movable requests will split the largest buddy and steal the
remainder. This splits up contiguity, but it allows subsequent
requests of this type to fall back into adjacent space.
- Movable requests go and look for the smallest buddy instead. The
thinking is that movable requests can be compacted, so grouping is
less important than retaining contiguity.
c0cd6f557b90 ("mm: page_alloc: fix freelist movement during block
conversion") enforces freelist type hygiene, which restricts stealing to
either claiming the whole block or just taking the requested chunk; no
additional pages or buddy remainders can be stolen any more.
The patch mishandled when to switch to finding the smallest buddy in that
new reality. As a result, it may steal the exact request size, but from
the biggest buddy. This causes fracturing for no good reason.
Fix this by committing to the new behavior: either steal the whole block,
or fall back to the smallest buddy.
Remove single-page stealing from steal_suitable_fallback(). Rename it to
try_to_steal_block() to make the intentions clear. If this fails, always
fall back to the smallest buddy.
The following is from 4 runs of mmtest's thpchallenge. "Pollute" is
single page fallback, "steal" is conversion of a partially used block.
The numbers for free block conversions (omitted) are comparable.
vanilla patched
@pollute[unmovable from reclaimable]: 27 106
@pollute[unmovable from movable]: 82 46
@pollute[reclaimable from unmovable]: 256 83
@pollute[reclaimable from movable]: 46 8
@pollute[movable from unmovable]: 4841 868
@pollute[movable from reclaimable]: 5278 12568
@steal[unmovable from reclaimable]: 11 12
@steal[unmovable from movable]: 113 49
@steal[reclaimable from unmovable]: 19 34
@steal[reclaimable from movable]: 47 21
@steal[movable from unmovable]: 250 183
@steal[movable from reclaimable]: 81 93
The allocator appears to do a better job at keeping stealing and polluting
to the first fallback preference. As a result, the numbers for "from
movable" - the least preferred fallback option, and most detrimental to
compactability - are down across the board.
Link: https://lkml.kernel.org/r/20250225001023.1494422-2-hannes@cmpxchg.org
Fixes: c0cd6f557b90 ("mm: page_alloc: fix freelist movement during block conversion")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-02-24 19:08:24 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: page_alloc: speed up fallbacks in rmqueue_bulk()
The test robot identified c2f6ea38fc1b ("mm: page_alloc: don't steal
single pages from biggest buddy") as the root cause of a 56.4% regression
in vm-scalability::lru-file-mmap-read.
Carlos reports an earlier patch, c0cd6f557b90 ("mm: page_alloc: fix
freelist movement during block conversion"), as the root cause for a
regression in worst-case zone->lock+irqoff hold times.
Both of these patches modify the page allocator's fallback path to be less
greedy in an effort to stave off fragmentation. The flip side of this is
that fallbacks are also less productive each time around, which means the
fallback search can run much more frequently.
Carlos' traces point to rmqueue_bulk() specifically, which tries to refill
the percpu cache by allocating a large batch of pages in a loop. It
highlights how once the native freelists are exhausted, the fallback code
first scans orders top-down for whole blocks to claim, then falls back to
a bottom-up search for the smallest buddy to steal. For the next batch
page, it goes through the same thing again.
This can be made more efficient. Since rmqueue_bulk() holds the
zone->lock over the entire batch, the freelists are not subject to outside
changes; when the search for a block to claim has already failed, there is
no point in trying again for the next page.
Modify __rmqueue() to remember the last successful fallback mode, and
restart directly from there on the next rmqueue_bulk() iteration.
Oliver confirms that this improves beyond the regression that the test
robot reported against c2f6ea38fc1b:
commit:
f3b92176f4 ("tools/selftests: add guard region test for /proc/$pid/pagemap")
c2f6ea38fc ("mm: page_alloc: don't steal single pages from biggest buddy")
acc4d5ff0b ("Merge tag 'net-6.15-rc0' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net")
2c847f27c3 ("mm: page_alloc: speed up fallbacks in rmqueue_bulk()") <--- your patch
f3b92176f4f7100f c2f6ea38fc1b640aa7a2e155cc1 acc4d5ff0b61eb1715c498b6536 2c847f27c37da65a93d23c237c5
---------------- --------------------------- --------------------------- ---------------------------
%stddev %change %stddev %change %stddev %change %stddev
\ | \ | \ | \
25525364 ± 3% -56.4% 11135467 -57.8% 10779336 +31.6% 33581409 vm-scalability.throughput
Carlos confirms that worst-case times are almost fully recovered
compared to before the earlier culprit patch:
2dd482ba627d (before freelist hygiene): 1ms
c0cd6f557b90 (after freelist hygiene): 90ms
next-20250319 (steal smallest buddy): 280ms
this patch : 8ms
[jackmanb@google.com: comment updates]
Link: https://lkml.kernel.org/r/D92AC0P9594X.3BML64MUKTF8Z@google.com
[hannes@cmpxchg.org: reset rmqueue_mode in rmqueue_buddy() error loop, per Yunsheng Lin]
Link: https://lkml.kernel.org/r/20250409140023.GA2313@cmpxchg.org
Link: https://lkml.kernel.org/r/20250407180154.63348-1-hannes@cmpxchg.org
Fixes: c0cd6f557b90 ("mm: page_alloc: fix freelist movement during block conversion")
Fixes: c2f6ea38fc1b ("mm: page_alloc: don't steal single pages from biggest buddy")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Brendan Jackman <jackmanb@google.com>
Reported-by: kernel test robot <oliver.sang@intel.com>
Reported-by: Carlos Song <carlos.song@nxp.com>
Tested-by: Carlos Song <carlos.song@nxp.com>
Tested-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202503271547.fc08b188-lkp@intel.com
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Tested-by: Shivank Garg <shivankg@amd.com>
Acked-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: <stable@vger.kernel.org> [6.10+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-07 14:01:53 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-10-16 01:25:48 -07:00
|
|
|
|
|
|
|
|
|
mm: page_alloc: fix freelist movement during block conversion
Currently, page block type conversion during fallbacks, atomic
reservations and isolation can strand various amounts of free pages on
incorrect freelists.
For example, fallback stealing moves free pages in the block to the new
type's freelists, but then may not actually claim the block for that type
if there aren't enough compatible pages already allocated.
In all cases, free page moving might fail if the block straddles more than
one zone, in which case no free pages are moved at all, but the block type
is changed anyway.
This is detrimental to type hygiene on the freelists. It encourages
incompatible page mixing down the line (ask for one type, get another) and
thus contributes to long-term fragmentation.
Split the process into a proper transaction: check first if conversion
will happen, then try to move the free pages, and only if that was
successful convert the block to the new type.
[baolin.wang@linux.alibaba.com: fix allocation failures with CONFIG_CMA]
Link: https://lkml.kernel.org/r/a97697e0-45b0-4f71-b087-fdc7a1d43c0e@linux.alibaba.com
Link: https://lkml.kernel.org/r/20240320180429.678181-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Tested-by: "Huang, Ying" <ying.huang@intel.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-20 14:02:11 -04:00
|
|
|
|
2007-10-16 01:25:48 -07:00
|
|
|
|
|
|
|
|
|
mm: page_alloc: speed up fallbacks in rmqueue_bulk()
The test robot identified c2f6ea38fc1b ("mm: page_alloc: don't steal
single pages from biggest buddy") as the root cause of a 56.4% regression
in vm-scalability::lru-file-mmap-read.
Carlos reports an earlier patch, c0cd6f557b90 ("mm: page_alloc: fix
freelist movement during block conversion"), as the root cause for a
regression in worst-case zone->lock+irqoff hold times.
Both of these patches modify the page allocator's fallback path to be less
greedy in an effort to stave off fragmentation. The flip side of this is
that fallbacks are also less productive each time around, which means the
fallback search can run much more frequently.
Carlos' traces point to rmqueue_bulk() specifically, which tries to refill
the percpu cache by allocating a large batch of pages in a loop. It
highlights how once the native freelists are exhausted, the fallback code
first scans orders top-down for whole blocks to claim, then falls back to
a bottom-up search for the smallest buddy to steal. For the next batch
page, it goes through the same thing again.
This can be made more efficient. Since rmqueue_bulk() holds the
zone->lock over the entire batch, the freelists are not subject to outside
changes; when the search for a block to claim has already failed, there is
no point in trying again for the next page.
Modify __rmqueue() to remember the last successful fallback mode, and
restart directly from there on the next rmqueue_bulk() iteration.
Oliver confirms that this improves beyond the regression that the test
robot reported against c2f6ea38fc1b:
commit:
f3b92176f4 ("tools/selftests: add guard region test for /proc/$pid/pagemap")
c2f6ea38fc ("mm: page_alloc: don't steal single pages from biggest buddy")
acc4d5ff0b ("Merge tag 'net-6.15-rc0' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net")
2c847f27c3 ("mm: page_alloc: speed up fallbacks in rmqueue_bulk()") <--- your patch
f3b92176f4f7100f c2f6ea38fc1b640aa7a2e155cc1 acc4d5ff0b61eb1715c498b6536 2c847f27c37da65a93d23c237c5
---------------- --------------------------- --------------------------- ---------------------------
%stddev %change %stddev %change %stddev %change %stddev
\ | \ | \ | \
25525364 ± 3% -56.4% 11135467 -57.8% 10779336 +31.6% 33581409 vm-scalability.throughput
Carlos confirms that worst-case times are almost fully recovered
compared to before the earlier culprit patch:
2dd482ba627d (before freelist hygiene): 1ms
c0cd6f557b90 (after freelist hygiene): 90ms
next-20250319 (steal smallest buddy): 280ms
this patch : 8ms
[jackmanb@google.com: comment updates]
Link: https://lkml.kernel.org/r/D92AC0P9594X.3BML64MUKTF8Z@google.com
[hannes@cmpxchg.org: reset rmqueue_mode in rmqueue_buddy() error loop, per Yunsheng Lin]
Link: https://lkml.kernel.org/r/20250409140023.GA2313@cmpxchg.org
Link: https://lkml.kernel.org/r/20250407180154.63348-1-hannes@cmpxchg.org
Fixes: c0cd6f557b90 ("mm: page_alloc: fix freelist movement during block conversion")
Fixes: c2f6ea38fc1b ("mm: page_alloc: don't steal single pages from biggest buddy")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Brendan Jackman <jackmanb@google.com>
Reported-by: kernel test robot <oliver.sang@intel.com>
Reported-by: Carlos Song <carlos.song@nxp.com>
Tested-by: Carlos Song <carlos.song@nxp.com>
Tested-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202503271547.fc08b188-lkp@intel.com
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Tested-by: Shivank Garg <shivankg@amd.com>
Acked-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: <stable@vger.kernel.org> [6.10+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-07 14:01:53 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Bias the location of pages freed for min_free_kbytes in the same MAX_ORDER_NR_PAGES blocks
The standard buddy allocator always favours the smallest block of pages.
The effect of this is that the pages free to satisfy min_free_kbytes tends
to be preserved since boot time at the same location of memory ffor a very
long time and as a contiguous block. When an administrator sets the
reserve at 16384 at boot time, it tends to be the same MAX_ORDER blocks
that remain free. This allows the occasional high atomic allocation to
succeed up until the point the blocks are split. In practice, it is
difficult to split these blocks but when they do split, the benefit of
having min_free_kbytes for contiguous blocks disappears. Additionally,
increasing min_free_kbytes once the system has been running for some time
has no guarantee of creating contiguous blocks.
On the other hand, CONFIG_PAGE_GROUP_BY_MOBILITY favours splitting large
blocks when there are no free pages of the appropriate type available. A
side-effect of this is that all blocks in memory tends to be used up and
the contiguous free blocks from boot time are not preserved like in the
vanilla allocator. This can cause a problem if a new caller is unwilling
to reclaim or does not reclaim for long enough.
A failure scenario was found for a wireless network device allocating
order-1 atomic allocations but the allocations were not intense or frequent
enough for a whole block of pages to be preserved for MIGRATE_HIGHALLOC.
This was reproduced on a desktop by booting with mem=256mb, forcing the
driver to allocate at order-1, running a bittorrent client (downloading a
debian ISO) and building a kernel with -j2.
This patch addresses the problem on the desktop machine booted with
mem=256mb. It works by setting aside a reserve of MAX_ORDER_NR_PAGES
blocks, the number of which depends on the value of min_free_kbytes. These
blocks are only fallen back to when there is no other free pages. Then the
smallest possible page is used just like the normal buddy allocator instead
of the largest possible page to preserve contiguous pages The pages in free
lists in the reserve blocks are never taken for another migrate type. The
results is that even if min_free_kbytes is set to a low value, contiguous
blocks will be preserved in the MIGRATE_RESERVE blocks.
This works better than the vanilla allocator because if min_free_kbytes is
increased, a new reserve block will be chosen based on the location of
reclaimable pages and the block will free up as contiguous pages. In the
vanilla allocator, no effort is made to target a block of pages to free as
contiguous pages and min_free_kbytes pages are scattered randomly.
This effect has been observed on the test machine. min_free_kbytes was set
initially low but it was kept as a contiguous free block within
MIGRATE_RESERVE. min_free_kbytes was then set to a higher value and over a
period of time, the free blocks were within the reserve and coalescing.
How long it takes to free up depends on how quickly LRU is rotating.
Amusingly, this means that more activity will free the blocks faster.
This mechanism potentially replaces MIGRATE_HIGHALLOC as it may be more
effective than grouping contiguous free pages together. It all depends on
whether the number of active atomic high allocations exceeds
min_free_kbytes or not. If the number of active allocations exceeds
min_free_kbytes, it's worth it but maybe in that situation, min_free_kbytes
should be set higher. Once there are no more reports of allocation
failures, a patch will be submitted that backs out MIGRATE_HIGHALLOC and
see if the reports stay missing.
Credit to Mariusz Kozlowski for discovering the problem, describing the
failure scenario and testing patches and scenarios.
[akpm@linux-foundation.org: cleanups]
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Acked-by: Andy Whitcroft <apw@shadowen.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 01:25:58 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-11-15 17:36:53 -08:00
|
|
|
|
mm, page_alloc: spread allocations across zones before introducing fragmentation
Patch series "Fragmentation avoidance improvements", v5.
It has been noted before that fragmentation avoidance (aka
anti-fragmentation) is not perfect. Given sufficient time or an adverse
workload, memory gets fragmented and the long-term success of high-order
allocations degrades. This series defines an adverse workload, a definition
of external fragmentation events (including serious) ones and a series
that reduces the level of those fragmentation events.
The details of the workload and the consequences are described in more
detail in the changelogs. However, from patch 1, this is a high-level
summary of the adverse workload. The exact details are found in the
mmtests implementation.
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch)
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameterr create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed
3. Warm up a number of fio read-only threads accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll fault back in old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup
Overall the series reduces external fragmentation causing events by over 94%
on 1 and 2 socket machines, which in turn impacts high-order allocation
success rates over the long term. There are differences in latencies and
high-order allocation success rates. Latencies are a mixed bag as they
are vulnerable to exact system state and whether allocations succeeded
so they are treated as a secondary metric.
Patch 1 uses lower zones if they are populated and have free memory
instead of fragmenting a higher zone. It's special cased to
handle a Normal->DMA32 fallback with the reasons explained
in the changelog.
Patch 2-4 boosts watermarks temporarily when an external fragmentation
event occurs. kswapd wakes to reclaim a small amount of old memory
and then wakes kcompactd on completion to recover the system
slightly. This introduces some overhead in the slowpath. The level
of boosting can be tuned or disabled depending on the tolerance
for fragmentation vs allocation latency.
Patch 5 stalls some movable allocation requests to let kswapd from patch 4
make some progress. The duration of the stalls is very low but it
is possible to tune the system to avoid fragmentation events if
larger stalls can be tolerated.
The bulk of the improvement in fragmentation avoidance is from patches
1-4 but patch 5 can deal with a rare corner case and provides the option
of tuning a system for THP allocation success rates in exchange for
some stalls to control fragmentation.
This patch (of 5):
The page allocator zone lists are iterated based on the watermarks of each
zone which does not take anti-fragmentation into account. On x86, node 0
may have multiple zones while other nodes have one zone. A consequence is
that tasks running on node 0 may fragment ZONE_NORMAL even though
ZONE_DMA32 has plenty of free memory. This patch special cases the
allocator fast path such that it'll try an allocation from a lower local
zone before fragmenting a higher zone. In this case, stealing of
pageblocks or orders larger than a pageblock are still allowed in the fast
path as they are uninteresting from a fragmentation point of view.
This was evaluated using a benchmark designed to fragment memory before
attempting THP allocations. It's implemented in mmtests as the following
configurations
configs/config-global-dhp__workload_thpfioscale
configs/config-global-dhp__workload_thpfioscale-defrag
configs/config-global-dhp__workload_thpfioscale-madvhugepage
e.g. from mmtests
./run-mmtests.sh --run-monitor --config configs/config-global-dhp__workload_thpfioscale test-run-1
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch).
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameter create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed.
3. Warm up a number of fio read-only processes accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll refault old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds.
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup the test files.
Note that due to the use of IO and page cache that this benchmark is not
suitable for running on large machines where the time to fragment memory
may be excessive. Also note that while this is one mix that generates
fragmentation that it's not the only mix that generates fragmentation.
Differences in workload that are more slab-intensive or whether SLUB is
used with high-order pages may yield different results.
When the page allocator fragments memory, it records the event using the
mm_page_alloc_extfrag ftrace event. If the fallback_order is smaller than
a pageblock order (order-9 on 64-bit x86) then it's considered to be an
"external fragmentation event" that may cause issues in the future.
Hence, the primary metric here is the number of external fragmentation
events that occur with order < 9. The secondary metric is allocation
latency and huge page allocation success rates but note that differences
in latencies and what the success rate also can affect the number of
external fragmentation event which is why it's a secondary metric.
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 662.92 ( 0.00%) 653.58 * 1.41%*
Amean fault-huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
Fault latencies are slightly reduced while allocation success rates remain
at zero as this configuration does not make any special effort to allocate
THP and fio is heavily active at the time and either filling memory or
keeping pages resident. However, a 49% reduction of serious fragmentation
events reduces the changes of external fragmentation being a problem in
the future.
Vlastimil asked during review for a breakdown of the allocation types
that are falling back.
vanilla
3816 MIGRATE_UNMOVABLE
800845 MIGRATE_MOVABLE
33 MIGRATE_UNRECLAIMABLE
patch
735 MIGRATE_UNMOVABLE
408135 MIGRATE_MOVABLE
42 MIGRATE_UNRECLAIMABLE
The majority of the fallbacks are due to movable allocations and this is
consistent for the workload throughout the series so will not be presented
again as the primary source of fallbacks are movable allocations.
Movable fallbacks are sometimes considered "ok" to fallback because they
can be migrated. The problem is that they can fill an
unmovable/reclaimable pageblock causing those allocations to fallback
later and polluting pageblocks with pages that cannot move. If there is a
movable fallback, it is pretty much guaranteed to affect an
unmovable/reclaimable pageblock and while it might not be enough to
actually cause a unmovable/reclaimable fallback in the future, we cannot
know that in advance so the patch takes the only option available to it.
Hence, it's important to control them. This point is also consistent
throughout the series and will not be repeated.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 1495.14 ( 0.00%) 1467.55 ( 1.85%)
Amean fault-huge-1 1098.48 ( 0.00%) 1127.11 ( -2.61%)
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 78.57 ( 0.00%) 77.64 ( -1.18%)
Fragmentation events were reduced quite a bit although this is known
to be a little variable. The latencies and allocation success rates
are similar but they were already quite high.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 1350.05 ( 0.00%) 1346.45 ( 0.27%)
Amean fault-huge-5 4181.01 ( 0.00%) 3418.60 ( 18.24%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 1.15 ( 0.00%) 0.78 ( -31.88%)
The reduction of external fragmentation events is slight and this is
partially due to the removal of __GFP_THISNODE in commit ac5b2c18911f
("mm: thp: relax __GFP_THISNODE for MADV_HUGEPAGE mappings") as THP
allocations can now spill over to remote nodes instead of fragmenting
local memory.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 6138.97 ( 0.00%) 6217.43 ( -1.28%)
Amean fault-huge-5 2294.28 ( 0.00%) 3163.33 * -37.88%*
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 96.82 ( 0.00%) 95.14 ( -1.74%)
There was a slight reduction in external fragmentation events although the
latencies were higher. The allocation success rate is high enough that
the system is struggling and there is quite a lot of parallel reclaim and
compaction activity. There is also a certain degree of luck on whether
processes start on node 0 or not for this patch but the relevance is
reduced later in the series.
Overall, the patch reduces the number of external fragmentation causing
events so the success of THP over long periods of time would be improved
for this adverse workload.
Link: http://lkml.kernel.org/r/20181123114528.28802-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:41 -08:00
|
|
|
|
mm: page_alloc: speed up fallbacks in rmqueue_bulk()
The test robot identified c2f6ea38fc1b ("mm: page_alloc: don't steal
single pages from biggest buddy") as the root cause of a 56.4% regression
in vm-scalability::lru-file-mmap-read.
Carlos reports an earlier patch, c0cd6f557b90 ("mm: page_alloc: fix
freelist movement during block conversion"), as the root cause for a
regression in worst-case zone->lock+irqoff hold times.
Both of these patches modify the page allocator's fallback path to be less
greedy in an effort to stave off fragmentation. The flip side of this is
that fallbacks are also less productive each time around, which means the
fallback search can run much more frequently.
Carlos' traces point to rmqueue_bulk() specifically, which tries to refill
the percpu cache by allocating a large batch of pages in a loop. It
highlights how once the native freelists are exhausted, the fallback code
first scans orders top-down for whole blocks to claim, then falls back to
a bottom-up search for the smallest buddy to steal. For the next batch
page, it goes through the same thing again.
This can be made more efficient. Since rmqueue_bulk() holds the
zone->lock over the entire batch, the freelists are not subject to outside
changes; when the search for a block to claim has already failed, there is
no point in trying again for the next page.
Modify __rmqueue() to remember the last successful fallback mode, and
restart directly from there on the next rmqueue_bulk() iteration.
Oliver confirms that this improves beyond the regression that the test
robot reported against c2f6ea38fc1b:
commit:
f3b92176f4 ("tools/selftests: add guard region test for /proc/$pid/pagemap")
c2f6ea38fc ("mm: page_alloc: don't steal single pages from biggest buddy")
acc4d5ff0b ("Merge tag 'net-6.15-rc0' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net")
2c847f27c3 ("mm: page_alloc: speed up fallbacks in rmqueue_bulk()") <--- your patch
f3b92176f4f7100f c2f6ea38fc1b640aa7a2e155cc1 acc4d5ff0b61eb1715c498b6536 2c847f27c37da65a93d23c237c5
---------------- --------------------------- --------------------------- ---------------------------
%stddev %change %stddev %change %stddev %change %stddev
\ | \ | \ | \
25525364 ± 3% -56.4% 11135467 -57.8% 10779336 +31.6% 33581409 vm-scalability.throughput
Carlos confirms that worst-case times are almost fully recovered
compared to before the earlier culprit patch:
2dd482ba627d (before freelist hygiene): 1ms
c0cd6f557b90 (after freelist hygiene): 90ms
next-20250319 (steal smallest buddy): 280ms
this patch : 8ms
[jackmanb@google.com: comment updates]
Link: https://lkml.kernel.org/r/D92AC0P9594X.3BML64MUKTF8Z@google.com
[hannes@cmpxchg.org: reset rmqueue_mode in rmqueue_buddy() error loop, per Yunsheng Lin]
Link: https://lkml.kernel.org/r/20250409140023.GA2313@cmpxchg.org
Link: https://lkml.kernel.org/r/20250407180154.63348-1-hannes@cmpxchg.org
Fixes: c0cd6f557b90 ("mm: page_alloc: fix freelist movement during block conversion")
Fixes: c2f6ea38fc1b ("mm: page_alloc: don't steal single pages from biggest buddy")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Brendan Jackman <jackmanb@google.com>
Reported-by: kernel test robot <oliver.sang@intel.com>
Reported-by: Carlos Song <carlos.song@nxp.com>
Tested-by: Carlos Song <carlos.song@nxp.com>
Tested-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202503271547.fc08b188-lkp@intel.com
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Tested-by: Shivank Garg <shivankg@amd.com>
Acked-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: <stable@vger.kernel.org> [6.10+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-07 14:01:53 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-01-12 15:49:08 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-05-19 14:08:54 -07:00
|
|
|
|
2021-01-12 15:49:08 -08:00
|
|
|
|
2020-06-03 15:58:42 -07:00
|
|
|
|
mm: page_alloc: fix freelist movement during block conversion
Currently, page block type conversion during fallbacks, atomic
reservations and isolation can strand various amounts of free pages on
incorrect freelists.
For example, fallback stealing moves free pages in the block to the new
type's freelists, but then may not actually claim the block for that type
if there aren't enough compatible pages already allocated.
In all cases, free page moving might fail if the block straddles more than
one zone, in which case no free pages are moved at all, but the block type
is changed anyway.
This is detrimental to type hygiene on the freelists. It encourages
incompatible page mixing down the line (ask for one type, get another) and
thus contributes to long-term fragmentation.
Split the process into a proper transaction: check first if conversion
will happen, then try to move the free pages, and only if that was
successful convert the block to the new type.
[baolin.wang@linux.alibaba.com: fix allocation failures with CONFIG_CMA]
Link: https://lkml.kernel.org/r/a97697e0-45b0-4f71-b087-fdc7a1d43c0e@linux.alibaba.com
Link: https://lkml.kernel.org/r/20240320180429.678181-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Tested-by: "Huang, Ying" <ying.huang@intel.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-20 14:02:11 -04:00
|
|
|
|
mm: page_alloc: speed up fallbacks in rmqueue_bulk()
The test robot identified c2f6ea38fc1b ("mm: page_alloc: don't steal
single pages from biggest buddy") as the root cause of a 56.4% regression
in vm-scalability::lru-file-mmap-read.
Carlos reports an earlier patch, c0cd6f557b90 ("mm: page_alloc: fix
freelist movement during block conversion"), as the root cause for a
regression in worst-case zone->lock+irqoff hold times.
Both of these patches modify the page allocator's fallback path to be less
greedy in an effort to stave off fragmentation. The flip side of this is
that fallbacks are also less productive each time around, which means the
fallback search can run much more frequently.
Carlos' traces point to rmqueue_bulk() specifically, which tries to refill
the percpu cache by allocating a large batch of pages in a loop. It
highlights how once the native freelists are exhausted, the fallback code
first scans orders top-down for whole blocks to claim, then falls back to
a bottom-up search for the smallest buddy to steal. For the next batch
page, it goes through the same thing again.
This can be made more efficient. Since rmqueue_bulk() holds the
zone->lock over the entire batch, the freelists are not subject to outside
changes; when the search for a block to claim has already failed, there is
no point in trying again for the next page.
Modify __rmqueue() to remember the last successful fallback mode, and
restart directly from there on the next rmqueue_bulk() iteration.
Oliver confirms that this improves beyond the regression that the test
robot reported against c2f6ea38fc1b:
commit:
f3b92176f4 ("tools/selftests: add guard region test for /proc/$pid/pagemap")
c2f6ea38fc ("mm: page_alloc: don't steal single pages from biggest buddy")
acc4d5ff0b ("Merge tag 'net-6.15-rc0' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net")
2c847f27c3 ("mm: page_alloc: speed up fallbacks in rmqueue_bulk()") <--- your patch
f3b92176f4f7100f c2f6ea38fc1b640aa7a2e155cc1 acc4d5ff0b61eb1715c498b6536 2c847f27c37da65a93d23c237c5
---------------- --------------------------- --------------------------- ---------------------------
%stddev %change %stddev %change %stddev %change %stddev
\ | \ | \ | \
25525364 ± 3% -56.4% 11135467 -57.8% 10779336 +31.6% 33581409 vm-scalability.throughput
Carlos confirms that worst-case times are almost fully recovered
compared to before the earlier culprit patch:
2dd482ba627d (before freelist hygiene): 1ms
c0cd6f557b90 (after freelist hygiene): 90ms
next-20250319 (steal smallest buddy): 280ms
this patch : 8ms
[jackmanb@google.com: comment updates]
Link: https://lkml.kernel.org/r/D92AC0P9594X.3BML64MUKTF8Z@google.com
[hannes@cmpxchg.org: reset rmqueue_mode in rmqueue_buddy() error loop, per Yunsheng Lin]
Link: https://lkml.kernel.org/r/20250409140023.GA2313@cmpxchg.org
Link: https://lkml.kernel.org/r/20250407180154.63348-1-hannes@cmpxchg.org
Fixes: c0cd6f557b90 ("mm: page_alloc: fix freelist movement during block conversion")
Fixes: c2f6ea38fc1b ("mm: page_alloc: don't steal single pages from biggest buddy")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Brendan Jackman <jackmanb@google.com>
Reported-by: kernel test robot <oliver.sang@intel.com>
Reported-by: Carlos Song <carlos.song@nxp.com>
Tested-by: Carlos Song <carlos.song@nxp.com>
Tested-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202503271547.fc08b188-lkp@intel.com
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Tested-by: Shivank Garg <shivankg@amd.com>
Acked-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: <stable@vger.kernel.org> [6.10+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-07 14:01:53 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-04-14 15:45:15 -07:00
|
|
|
|
mm: page_alloc: speed up fallbacks in rmqueue_bulk()
The test robot identified c2f6ea38fc1b ("mm: page_alloc: don't steal
single pages from biggest buddy") as the root cause of a 56.4% regression
in vm-scalability::lru-file-mmap-read.
Carlos reports an earlier patch, c0cd6f557b90 ("mm: page_alloc: fix
freelist movement during block conversion"), as the root cause for a
regression in worst-case zone->lock+irqoff hold times.
Both of these patches modify the page allocator's fallback path to be less
greedy in an effort to stave off fragmentation. The flip side of this is
that fallbacks are also less productive each time around, which means the
fallback search can run much more frequently.
Carlos' traces point to rmqueue_bulk() specifically, which tries to refill
the percpu cache by allocating a large batch of pages in a loop. It
highlights how once the native freelists are exhausted, the fallback code
first scans orders top-down for whole blocks to claim, then falls back to
a bottom-up search for the smallest buddy to steal. For the next batch
page, it goes through the same thing again.
This can be made more efficient. Since rmqueue_bulk() holds the
zone->lock over the entire batch, the freelists are not subject to outside
changes; when the search for a block to claim has already failed, there is
no point in trying again for the next page.
Modify __rmqueue() to remember the last successful fallback mode, and
restart directly from there on the next rmqueue_bulk() iteration.
Oliver confirms that this improves beyond the regression that the test
robot reported against c2f6ea38fc1b:
commit:
f3b92176f4 ("tools/selftests: add guard region test for /proc/$pid/pagemap")
c2f6ea38fc ("mm: page_alloc: don't steal single pages from biggest buddy")
acc4d5ff0b ("Merge tag 'net-6.15-rc0' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net")
2c847f27c3 ("mm: page_alloc: speed up fallbacks in rmqueue_bulk()") <--- your patch
f3b92176f4f7100f c2f6ea38fc1b640aa7a2e155cc1 acc4d5ff0b61eb1715c498b6536 2c847f27c37da65a93d23c237c5
---------------- --------------------------- --------------------------- ---------------------------
%stddev %change %stddev %change %stddev %change %stddev
\ | \ | \ | \
25525364 ± 3% -56.4% 11135467 -57.8% 10779336 +31.6% 33581409 vm-scalability.throughput
Carlos confirms that worst-case times are almost fully recovered
compared to before the earlier culprit patch:
2dd482ba627d (before freelist hygiene): 1ms
c0cd6f557b90 (after freelist hygiene): 90ms
next-20250319 (steal smallest buddy): 280ms
this patch : 8ms
[jackmanb@google.com: comment updates]
Link: https://lkml.kernel.org/r/D92AC0P9594X.3BML64MUKTF8Z@google.com
[hannes@cmpxchg.org: reset rmqueue_mode in rmqueue_buddy() error loop, per Yunsheng Lin]
Link: https://lkml.kernel.org/r/20250409140023.GA2313@cmpxchg.org
Link: https://lkml.kernel.org/r/20250407180154.63348-1-hannes@cmpxchg.org
Fixes: c0cd6f557b90 ("mm: page_alloc: fix freelist movement during block conversion")
Fixes: c2f6ea38fc1b ("mm: page_alloc: don't steal single pages from biggest buddy")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Brendan Jackman <jackmanb@google.com>
Reported-by: kernel test robot <oliver.sang@intel.com>
Reported-by: Carlos Song <carlos.song@nxp.com>
Tested-by: Carlos Song <carlos.song@nxp.com>
Tested-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202503271547.fc08b188-lkp@intel.com
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Tested-by: Shivank Garg <shivankg@amd.com>
Acked-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: <stable@vger.kernel.org> [6.10+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-07 14:01:53 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2009-06-16 15:32:04 -07:00
|
|
|
|
mm: page_alloc: speed up fallbacks in rmqueue_bulk()
The test robot identified c2f6ea38fc1b ("mm: page_alloc: don't steal
single pages from biggest buddy") as the root cause of a 56.4% regression
in vm-scalability::lru-file-mmap-read.
Carlos reports an earlier patch, c0cd6f557b90 ("mm: page_alloc: fix
freelist movement during block conversion"), as the root cause for a
regression in worst-case zone->lock+irqoff hold times.
Both of these patches modify the page allocator's fallback path to be less
greedy in an effort to stave off fragmentation. The flip side of this is
that fallbacks are also less productive each time around, which means the
fallback search can run much more frequently.
Carlos' traces point to rmqueue_bulk() specifically, which tries to refill
the percpu cache by allocating a large batch of pages in a loop. It
highlights how once the native freelists are exhausted, the fallback code
first scans orders top-down for whole blocks to claim, then falls back to
a bottom-up search for the smallest buddy to steal. For the next batch
page, it goes through the same thing again.
This can be made more efficient. Since rmqueue_bulk() holds the
zone->lock over the entire batch, the freelists are not subject to outside
changes; when the search for a block to claim has already failed, there is
no point in trying again for the next page.
Modify __rmqueue() to remember the last successful fallback mode, and
restart directly from there on the next rmqueue_bulk() iteration.
Oliver confirms that this improves beyond the regression that the test
robot reported against c2f6ea38fc1b:
commit:
f3b92176f4 ("tools/selftests: add guard region test for /proc/$pid/pagemap")
c2f6ea38fc ("mm: page_alloc: don't steal single pages from biggest buddy")
acc4d5ff0b ("Merge tag 'net-6.15-rc0' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net")
2c847f27c3 ("mm: page_alloc: speed up fallbacks in rmqueue_bulk()") <--- your patch
f3b92176f4f7100f c2f6ea38fc1b640aa7a2e155cc1 acc4d5ff0b61eb1715c498b6536 2c847f27c37da65a93d23c237c5
---------------- --------------------------- --------------------------- ---------------------------
%stddev %change %stddev %change %stddev %change %stddev
\ | \ | \ | \
25525364 ± 3% -56.4% 11135467 -57.8% 10779336 +31.6% 33581409 vm-scalability.throughput
Carlos confirms that worst-case times are almost fully recovered
compared to before the earlier culprit patch:
2dd482ba627d (before freelist hygiene): 1ms
c0cd6f557b90 (after freelist hygiene): 90ms
next-20250319 (steal smallest buddy): 280ms
this patch : 8ms
[jackmanb@google.com: comment updates]
Link: https://lkml.kernel.org/r/D92AC0P9594X.3BML64MUKTF8Z@google.com
[hannes@cmpxchg.org: reset rmqueue_mode in rmqueue_buddy() error loop, per Yunsheng Lin]
Link: https://lkml.kernel.org/r/20250409140023.GA2313@cmpxchg.org
Link: https://lkml.kernel.org/r/20250407180154.63348-1-hannes@cmpxchg.org
Fixes: c0cd6f557b90 ("mm: page_alloc: fix freelist movement during block conversion")
Fixes: c2f6ea38fc1b ("mm: page_alloc: don't steal single pages from biggest buddy")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Brendan Jackman <jackmanb@google.com>
Reported-by: kernel test robot <oliver.sang@intel.com>
Reported-by: Carlos Song <carlos.song@nxp.com>
Tested-by: Carlos Song <carlos.song@nxp.com>
Tested-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202503271547.fc08b188-lkp@intel.com
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Tested-by: Shivank Garg <shivankg@amd.com>
Acked-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: <stable@vger.kernel.org> [6.10+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-07 14:01:53 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2012-01-11 15:16:11 +01:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-01-11 15:16:11 +01:00
|
|
|
|
2007-10-16 01:25:48 -07:00
|
|
|
|
mm, page_alloc: spread allocations across zones before introducing fragmentation
Patch series "Fragmentation avoidance improvements", v5.
It has been noted before that fragmentation avoidance (aka
anti-fragmentation) is not perfect. Given sufficient time or an adverse
workload, memory gets fragmented and the long-term success of high-order
allocations degrades. This series defines an adverse workload, a definition
of external fragmentation events (including serious) ones and a series
that reduces the level of those fragmentation events.
The details of the workload and the consequences are described in more
detail in the changelogs. However, from patch 1, this is a high-level
summary of the adverse workload. The exact details are found in the
mmtests implementation.
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch)
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameterr create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed
3. Warm up a number of fio read-only threads accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll fault back in old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup
Overall the series reduces external fragmentation causing events by over 94%
on 1 and 2 socket machines, which in turn impacts high-order allocation
success rates over the long term. There are differences in latencies and
high-order allocation success rates. Latencies are a mixed bag as they
are vulnerable to exact system state and whether allocations succeeded
so they are treated as a secondary metric.
Patch 1 uses lower zones if they are populated and have free memory
instead of fragmenting a higher zone. It's special cased to
handle a Normal->DMA32 fallback with the reasons explained
in the changelog.
Patch 2-4 boosts watermarks temporarily when an external fragmentation
event occurs. kswapd wakes to reclaim a small amount of old memory
and then wakes kcompactd on completion to recover the system
slightly. This introduces some overhead in the slowpath. The level
of boosting can be tuned or disabled depending on the tolerance
for fragmentation vs allocation latency.
Patch 5 stalls some movable allocation requests to let kswapd from patch 4
make some progress. The duration of the stalls is very low but it
is possible to tune the system to avoid fragmentation events if
larger stalls can be tolerated.
The bulk of the improvement in fragmentation avoidance is from patches
1-4 but patch 5 can deal with a rare corner case and provides the option
of tuning a system for THP allocation success rates in exchange for
some stalls to control fragmentation.
This patch (of 5):
The page allocator zone lists are iterated based on the watermarks of each
zone which does not take anti-fragmentation into account. On x86, node 0
may have multiple zones while other nodes have one zone. A consequence is
that tasks running on node 0 may fragment ZONE_NORMAL even though
ZONE_DMA32 has plenty of free memory. This patch special cases the
allocator fast path such that it'll try an allocation from a lower local
zone before fragmenting a higher zone. In this case, stealing of
pageblocks or orders larger than a pageblock are still allowed in the fast
path as they are uninteresting from a fragmentation point of view.
This was evaluated using a benchmark designed to fragment memory before
attempting THP allocations. It's implemented in mmtests as the following
configurations
configs/config-global-dhp__workload_thpfioscale
configs/config-global-dhp__workload_thpfioscale-defrag
configs/config-global-dhp__workload_thpfioscale-madvhugepage
e.g. from mmtests
./run-mmtests.sh --run-monitor --config configs/config-global-dhp__workload_thpfioscale test-run-1
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch).
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameter create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed.
3. Warm up a number of fio read-only processes accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll refault old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds.
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup the test files.
Note that due to the use of IO and page cache that this benchmark is not
suitable for running on large machines where the time to fragment memory
may be excessive. Also note that while this is one mix that generates
fragmentation that it's not the only mix that generates fragmentation.
Differences in workload that are more slab-intensive or whether SLUB is
used with high-order pages may yield different results.
When the page allocator fragments memory, it records the event using the
mm_page_alloc_extfrag ftrace event. If the fallback_order is smaller than
a pageblock order (order-9 on 64-bit x86) then it's considered to be an
"external fragmentation event" that may cause issues in the future.
Hence, the primary metric here is the number of external fragmentation
events that occur with order < 9. The secondary metric is allocation
latency and huge page allocation success rates but note that differences
in latencies and what the success rate also can affect the number of
external fragmentation event which is why it's a secondary metric.
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 662.92 ( 0.00%) 653.58 * 1.41%*
Amean fault-huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
Fault latencies are slightly reduced while allocation success rates remain
at zero as this configuration does not make any special effort to allocate
THP and fio is heavily active at the time and either filling memory or
keeping pages resident. However, a 49% reduction of serious fragmentation
events reduces the changes of external fragmentation being a problem in
the future.
Vlastimil asked during review for a breakdown of the allocation types
that are falling back.
vanilla
3816 MIGRATE_UNMOVABLE
800845 MIGRATE_MOVABLE
33 MIGRATE_UNRECLAIMABLE
patch
735 MIGRATE_UNMOVABLE
408135 MIGRATE_MOVABLE
42 MIGRATE_UNRECLAIMABLE
The majority of the fallbacks are due to movable allocations and this is
consistent for the workload throughout the series so will not be presented
again as the primary source of fallbacks are movable allocations.
Movable fallbacks are sometimes considered "ok" to fallback because they
can be migrated. The problem is that they can fill an
unmovable/reclaimable pageblock causing those allocations to fallback
later and polluting pageblocks with pages that cannot move. If there is a
movable fallback, it is pretty much guaranteed to affect an
unmovable/reclaimable pageblock and while it might not be enough to
actually cause a unmovable/reclaimable fallback in the future, we cannot
know that in advance so the patch takes the only option available to it.
Hence, it's important to control them. This point is also consistent
throughout the series and will not be repeated.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 1495.14 ( 0.00%) 1467.55 ( 1.85%)
Amean fault-huge-1 1098.48 ( 0.00%) 1127.11 ( -2.61%)
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 78.57 ( 0.00%) 77.64 ( -1.18%)
Fragmentation events were reduced quite a bit although this is known
to be a little variable. The latencies and allocation success rates
are similar but they were already quite high.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 1350.05 ( 0.00%) 1346.45 ( 0.27%)
Amean fault-huge-5 4181.01 ( 0.00%) 3418.60 ( 18.24%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 1.15 ( 0.00%) 0.78 ( -31.88%)
The reduction of external fragmentation events is slight and this is
partially due to the removal of __GFP_THISNODE in commit ac5b2c18911f
("mm: thp: relax __GFP_THISNODE for MADV_HUGEPAGE mappings") as THP
allocations can now spill over to remote nodes instead of fragmenting
local memory.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 6138.97 ( 0.00%) 6217.43 ( -1.28%)
Amean fault-huge-5 2294.28 ( 0.00%) 3163.33 * -37.88%*
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 96.82 ( 0.00%) 95.14 ( -1.74%)
There was a slight reduction in external fragmentation events although the
latencies were higher. The allocation success rate is high enough that
the system is struggling and there is quite a lot of parallel reclaim and
compaction activity. There is also a certain degree of luck on whether
processes start on node 0 or not for this patch but the relevance is
reduced later in the series.
Overall, the patch reduces the number of external fragmentation causing
events so the success of THP over long periods of time would be improved
for this adverse workload.
Link: http://lkml.kernel.org/r/20181123114528.28802-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:41 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
mm: page_alloc: speed up fallbacks in rmqueue_bulk()
The test robot identified c2f6ea38fc1b ("mm: page_alloc: don't steal
single pages from biggest buddy") as the root cause of a 56.4% regression
in vm-scalability::lru-file-mmap-read.
Carlos reports an earlier patch, c0cd6f557b90 ("mm: page_alloc: fix
freelist movement during block conversion"), as the root cause for a
regression in worst-case zone->lock+irqoff hold times.
Both of these patches modify the page allocator's fallback path to be less
greedy in an effort to stave off fragmentation. The flip side of this is
that fallbacks are also less productive each time around, which means the
fallback search can run much more frequently.
Carlos' traces point to rmqueue_bulk() specifically, which tries to refill
the percpu cache by allocating a large batch of pages in a loop. It
highlights how once the native freelists are exhausted, the fallback code
first scans orders top-down for whole blocks to claim, then falls back to
a bottom-up search for the smallest buddy to steal. For the next batch
page, it goes through the same thing again.
This can be made more efficient. Since rmqueue_bulk() holds the
zone->lock over the entire batch, the freelists are not subject to outside
changes; when the search for a block to claim has already failed, there is
no point in trying again for the next page.
Modify __rmqueue() to remember the last successful fallback mode, and
restart directly from there on the next rmqueue_bulk() iteration.
Oliver confirms that this improves beyond the regression that the test
robot reported against c2f6ea38fc1b:
commit:
f3b92176f4 ("tools/selftests: add guard region test for /proc/$pid/pagemap")
c2f6ea38fc ("mm: page_alloc: don't steal single pages from biggest buddy")
acc4d5ff0b ("Merge tag 'net-6.15-rc0' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net")
2c847f27c3 ("mm: page_alloc: speed up fallbacks in rmqueue_bulk()") <--- your patch
f3b92176f4f7100f c2f6ea38fc1b640aa7a2e155cc1 acc4d5ff0b61eb1715c498b6536 2c847f27c37da65a93d23c237c5
---------------- --------------------------- --------------------------- ---------------------------
%stddev %change %stddev %change %stddev %change %stddev
\ | \ | \ | \
25525364 ± 3% -56.4% 11135467 -57.8% 10779336 +31.6% 33581409 vm-scalability.throughput
Carlos confirms that worst-case times are almost fully recovered
compared to before the earlier culprit patch:
2dd482ba627d (before freelist hygiene): 1ms
c0cd6f557b90 (after freelist hygiene): 90ms
next-20250319 (steal smallest buddy): 280ms
this patch : 8ms
[jackmanb@google.com: comment updates]
Link: https://lkml.kernel.org/r/D92AC0P9594X.3BML64MUKTF8Z@google.com
[hannes@cmpxchg.org: reset rmqueue_mode in rmqueue_buddy() error loop, per Yunsheng Lin]
Link: https://lkml.kernel.org/r/20250409140023.GA2313@cmpxchg.org
Link: https://lkml.kernel.org/r/20250407180154.63348-1-hannes@cmpxchg.org
Fixes: c0cd6f557b90 ("mm: page_alloc: fix freelist movement during block conversion")
Fixes: c2f6ea38fc1b ("mm: page_alloc: don't steal single pages from biggest buddy")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Brendan Jackman <jackmanb@google.com>
Reported-by: kernel test robot <oliver.sang@intel.com>
Reported-by: Carlos Song <carlos.song@nxp.com>
Tested-by: Carlos Song <carlos.song@nxp.com>
Tested-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202503271547.fc08b188-lkp@intel.com
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Tested-by: Shivank Garg <shivankg@amd.com>
Acked-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: <stable@vger.kernel.org> [6.10+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-07 14:01:53 -04:00
|
|
|
|
2022-11-18 10:17:14 +00:00
|
|
|
|
mm, page_alloc: reduce page alloc/free sanity checks
Historically, we have performed sanity checks on all struct pages being
allocated or freed, making sure they have no unexpected page flags or
certain field values. This can detect insufficient cleanup and some cases
of use-after-free, although on its own it can't always identify the
culprit. The result is a warning and the "bad page" being leaked.
The checks do need some cpu cycles, so in 4.7 with commits 479f854a207c
("mm, page_alloc: defer debugging checks of pages allocated from the PCP")
and 4db7548ccbd9 ("mm, page_alloc: defer debugging checks of freed pages
until a PCP drain") they were no longer performed in the hot paths when
allocating and freeing from pcplists, but only when pcplists are bypassed,
refilled or drained. For debugging purposes, with CONFIG_DEBUG_VM enabled
the checks were instead still done in the hot paths and not when refilling
or draining pcplists.
With 4462b32c9285 ("mm, page_alloc: more extensive free page checking with
debug_pagealloc"), enabling debug_pagealloc also moved the sanity checks
back to hot pahs. When both debug_pagealloc and CONFIG_DEBUG_VM are
enabled, the checks are done both in hotpaths and pcplist refill/drain.
Even though the non-debug default today might seem to be a sensible
tradeoff between overhead and ability to detect bad pages, on closer look
it's arguably not. As most allocations go through the pcplists, catching
any bad pages when refilling or draining pcplists has only a small chance,
insufficient for debugging or serious hardening purposes. On the other
hand the cost of the checks is concentrated in the already expensive
drain/refill batching operations, and those are done under the often
contended zone lock. That was recently identified as an issue for page
allocation and the zone lock contention reduced by moving the checks
outside of the locked section with a patch "mm: reduce lock contention of
pcp buffer refill", but the cost of the checks is still visible compared
to their removal [1]. In the pcplist draining path free_pcppages_bulk()
the checks are still done under zone->lock.
Thus, remove the checks from pcplist refill and drain paths completely.
Introduce a static key check_pages_enabled to control checks during page
allocation a freeing (whether pcplist is used or bypassed). The static
key is enabled if either is true:
- kernel is built with CONFIG_DEBUG_VM=y (debugging)
- debug_pagealloc or page poisoning is boot-time enabled (debugging)
- init_on_alloc or init_on_free is boot-time enabled (hardening)
The resulting user visible changes:
- no checks when draining/refilling pcplists - less overhead, with
likely no practical reduction of ability to catch bad pages
- no checks when bypassing pcplists in default config (no
debugging/hardening) - less overhead etc. as above
- on typical hardened kernels [2], checks are now performed on each page
allocation/free (previously only when bypassing/draining/refilling
pcplists) - the init_on_alloc/init_on_free enabled should be sufficient
indication for preferring more costly alloc/free operations for
hardening purposes and we shouldn't need to introduce another toggle
- code (various wrappers) removal and simplification
[1] https://lore.kernel.org/all/68ba44d8-6899-c018-dcb3-36f3a96e6bea@sra.uni-hannover.de/
[2] https://lore.kernel.org/all/63ebc499.a70a0220.9ac51.29ea@mx.google.com/
[akpm@linux-foundation.org: coding-style cleanups]
[akpm@linux-foundation.org: make check_pages_enabled static]
Link: https://lkml.kernel.org/r/20230216095131.17336-1-vbabka@suse.cz
Reported-by: Alexander Halbuer <halbuer@sra.uni-hannover.de>
Reported-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Kees Cook <keescook@chromium.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-02-16 10:51:31 +01:00
|
|
|
|
2012-01-11 15:16:11 +01:00
|
|
|
|
2025-03-30 17:28:09 -07:00
|
|
|
|
|
|
|
|
|
2025-02-21 18:44:23 -08:00
|
|
|
|
2025-03-30 17:28:09 -07:00
|
|
|
|
2025-02-21 18:44:23 -08:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
mm, page_alloc: spread allocations across zones before introducing fragmentation
Patch series "Fragmentation avoidance improvements", v5.
It has been noted before that fragmentation avoidance (aka
anti-fragmentation) is not perfect. Given sufficient time or an adverse
workload, memory gets fragmented and the long-term success of high-order
allocations degrades. This series defines an adverse workload, a definition
of external fragmentation events (including serious) ones and a series
that reduces the level of those fragmentation events.
The details of the workload and the consequences are described in more
detail in the changelogs. However, from patch 1, this is a high-level
summary of the adverse workload. The exact details are found in the
mmtests implementation.
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch)
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameterr create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed
3. Warm up a number of fio read-only threads accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll fault back in old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup
Overall the series reduces external fragmentation causing events by over 94%
on 1 and 2 socket machines, which in turn impacts high-order allocation
success rates over the long term. There are differences in latencies and
high-order allocation success rates. Latencies are a mixed bag as they
are vulnerable to exact system state and whether allocations succeeded
so they are treated as a secondary metric.
Patch 1 uses lower zones if they are populated and have free memory
instead of fragmenting a higher zone. It's special cased to
handle a Normal->DMA32 fallback with the reasons explained
in the changelog.
Patch 2-4 boosts watermarks temporarily when an external fragmentation
event occurs. kswapd wakes to reclaim a small amount of old memory
and then wakes kcompactd on completion to recover the system
slightly. This introduces some overhead in the slowpath. The level
of boosting can be tuned or disabled depending on the tolerance
for fragmentation vs allocation latency.
Patch 5 stalls some movable allocation requests to let kswapd from patch 4
make some progress. The duration of the stalls is very low but it
is possible to tune the system to avoid fragmentation events if
larger stalls can be tolerated.
The bulk of the improvement in fragmentation avoidance is from patches
1-4 but patch 5 can deal with a rare corner case and provides the option
of tuning a system for THP allocation success rates in exchange for
some stalls to control fragmentation.
This patch (of 5):
The page allocator zone lists are iterated based on the watermarks of each
zone which does not take anti-fragmentation into account. On x86, node 0
may have multiple zones while other nodes have one zone. A consequence is
that tasks running on node 0 may fragment ZONE_NORMAL even though
ZONE_DMA32 has plenty of free memory. This patch special cases the
allocator fast path such that it'll try an allocation from a lower local
zone before fragmenting a higher zone. In this case, stealing of
pageblocks or orders larger than a pageblock are still allowed in the fast
path as they are uninteresting from a fragmentation point of view.
This was evaluated using a benchmark designed to fragment memory before
attempting THP allocations. It's implemented in mmtests as the following
configurations
configs/config-global-dhp__workload_thpfioscale
configs/config-global-dhp__workload_thpfioscale-defrag
configs/config-global-dhp__workload_thpfioscale-madvhugepage
e.g. from mmtests
./run-mmtests.sh --run-monitor --config configs/config-global-dhp__workload_thpfioscale test-run-1
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch).
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameter create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed.
3. Warm up a number of fio read-only processes accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll refault old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds.
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup the test files.
Note that due to the use of IO and page cache that this benchmark is not
suitable for running on large machines where the time to fragment memory
may be excessive. Also note that while this is one mix that generates
fragmentation that it's not the only mix that generates fragmentation.
Differences in workload that are more slab-intensive or whether SLUB is
used with high-order pages may yield different results.
When the page allocator fragments memory, it records the event using the
mm_page_alloc_extfrag ftrace event. If the fallback_order is smaller than
a pageblock order (order-9 on 64-bit x86) then it's considered to be an
"external fragmentation event" that may cause issues in the future.
Hence, the primary metric here is the number of external fragmentation
events that occur with order < 9. The secondary metric is allocation
latency and huge page allocation success rates but note that differences
in latencies and what the success rate also can affect the number of
external fragmentation event which is why it's a secondary metric.
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 662.92 ( 0.00%) 653.58 * 1.41%*
Amean fault-huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
Fault latencies are slightly reduced while allocation success rates remain
at zero as this configuration does not make any special effort to allocate
THP and fio is heavily active at the time and either filling memory or
keeping pages resident. However, a 49% reduction of serious fragmentation
events reduces the changes of external fragmentation being a problem in
the future.
Vlastimil asked during review for a breakdown of the allocation types
that are falling back.
vanilla
3816 MIGRATE_UNMOVABLE
800845 MIGRATE_MOVABLE
33 MIGRATE_UNRECLAIMABLE
patch
735 MIGRATE_UNMOVABLE
408135 MIGRATE_MOVABLE
42 MIGRATE_UNRECLAIMABLE
The majority of the fallbacks are due to movable allocations and this is
consistent for the workload throughout the series so will not be presented
again as the primary source of fallbacks are movable allocations.
Movable fallbacks are sometimes considered "ok" to fallback because they
can be migrated. The problem is that they can fill an
unmovable/reclaimable pageblock causing those allocations to fallback
later and polluting pageblocks with pages that cannot move. If there is a
movable fallback, it is pretty much guaranteed to affect an
unmovable/reclaimable pageblock and while it might not be enough to
actually cause a unmovable/reclaimable fallback in the future, we cannot
know that in advance so the patch takes the only option available to it.
Hence, it's important to control them. This point is also consistent
throughout the series and will not be repeated.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 1495.14 ( 0.00%) 1467.55 ( 1.85%)
Amean fault-huge-1 1098.48 ( 0.00%) 1127.11 ( -2.61%)
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 78.57 ( 0.00%) 77.64 ( -1.18%)
Fragmentation events were reduced quite a bit although this is known
to be a little variable. The latencies and allocation success rates
are similar but they were already quite high.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 1350.05 ( 0.00%) 1346.45 ( 0.27%)
Amean fault-huge-5 4181.01 ( 0.00%) 3418.60 ( 18.24%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 1.15 ( 0.00%) 0.78 ( -31.88%)
The reduction of external fragmentation events is slight and this is
partially due to the removal of __GFP_THISNODE in commit ac5b2c18911f
("mm: thp: relax __GFP_THISNODE for MADV_HUGEPAGE mappings") as THP
allocations can now spill over to remote nodes instead of fragmenting
local memory.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 6138.97 ( 0.00%) 6217.43 ( -1.28%)
Amean fault-huge-5 2294.28 ( 0.00%) 3163.33 * -37.88%*
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 96.82 ( 0.00%) 95.14 ( -1.74%)
There was a slight reduction in external fragmentation events although the
latencies were higher. The allocation success rate is high enough that
the system is struggling and there is quite a lot of parallel reclaim and
compaction activity. There is also a certain degree of luck on whether
processes start on node 0 or not for this patch but the relevance is
reduced later in the series.
Overall, the patch reduces the number of external fragmentation causing
events so the success of THP over long periods of time would be improved
for this adverse workload.
Link: http://lkml.kernel.org/r/20181123114528.28802-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:41 -08:00
|
|
|
|
mm: page_alloc: speed up fallbacks in rmqueue_bulk()
The test robot identified c2f6ea38fc1b ("mm: page_alloc: don't steal
single pages from biggest buddy") as the root cause of a 56.4% regression
in vm-scalability::lru-file-mmap-read.
Carlos reports an earlier patch, c0cd6f557b90 ("mm: page_alloc: fix
freelist movement during block conversion"), as the root cause for a
regression in worst-case zone->lock+irqoff hold times.
Both of these patches modify the page allocator's fallback path to be less
greedy in an effort to stave off fragmentation. The flip side of this is
that fallbacks are also less productive each time around, which means the
fallback search can run much more frequently.
Carlos' traces point to rmqueue_bulk() specifically, which tries to refill
the percpu cache by allocating a large batch of pages in a loop. It
highlights how once the native freelists are exhausted, the fallback code
first scans orders top-down for whole blocks to claim, then falls back to
a bottom-up search for the smallest buddy to steal. For the next batch
page, it goes through the same thing again.
This can be made more efficient. Since rmqueue_bulk() holds the
zone->lock over the entire batch, the freelists are not subject to outside
changes; when the search for a block to claim has already failed, there is
no point in trying again for the next page.
Modify __rmqueue() to remember the last successful fallback mode, and
restart directly from there on the next rmqueue_bulk() iteration.
Oliver confirms that this improves beyond the regression that the test
robot reported against c2f6ea38fc1b:
commit:
f3b92176f4 ("tools/selftests: add guard region test for /proc/$pid/pagemap")
c2f6ea38fc ("mm: page_alloc: don't steal single pages from biggest buddy")
acc4d5ff0b ("Merge tag 'net-6.15-rc0' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net")
2c847f27c3 ("mm: page_alloc: speed up fallbacks in rmqueue_bulk()") <--- your patch
f3b92176f4f7100f c2f6ea38fc1b640aa7a2e155cc1 acc4d5ff0b61eb1715c498b6536 2c847f27c37da65a93d23c237c5
---------------- --------------------------- --------------------------- ---------------------------
%stddev %change %stddev %change %stddev %change %stddev
\ | \ | \ | \
25525364 ± 3% -56.4% 11135467 -57.8% 10779336 +31.6% 33581409 vm-scalability.throughput
Carlos confirms that worst-case times are almost fully recovered
compared to before the earlier culprit patch:
2dd482ba627d (before freelist hygiene): 1ms
c0cd6f557b90 (after freelist hygiene): 90ms
next-20250319 (steal smallest buddy): 280ms
this patch : 8ms
[jackmanb@google.com: comment updates]
Link: https://lkml.kernel.org/r/D92AC0P9594X.3BML64MUKTF8Z@google.com
[hannes@cmpxchg.org: reset rmqueue_mode in rmqueue_buddy() error loop, per Yunsheng Lin]
Link: https://lkml.kernel.org/r/20250409140023.GA2313@cmpxchg.org
Link: https://lkml.kernel.org/r/20250407180154.63348-1-hannes@cmpxchg.org
Fixes: c0cd6f557b90 ("mm: page_alloc: fix freelist movement during block conversion")
Fixes: c2f6ea38fc1b ("mm: page_alloc: don't steal single pages from biggest buddy")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Brendan Jackman <jackmanb@google.com>
Reported-by: kernel test robot <oliver.sang@intel.com>
Reported-by: Carlos Song <carlos.song@nxp.com>
Tested-by: Carlos Song <carlos.song@nxp.com>
Tested-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202503271547.fc08b188-lkp@intel.com
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Tested-by: Shivank Garg <shivankg@amd.com>
Acked-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: <stable@vger.kernel.org> [6.10+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-07 14:01:53 -04:00
|
|
|
|
2006-01-06 00:11:01 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2007-12-17 16:20:05 -08:00
|
|
|
|
|
|
|
|
|
2017-11-15 17:38:07 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-12-17 16:20:05 -08:00
|
|
|
|
2022-06-24 13:54:17 +01:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2022-11-18 10:17:14 +00:00
|
|
|
|
2023-02-01 17:25:49 +01:00
|
|
|
|
mm, page_alloc: reduce page alloc/free sanity checks
Historically, we have performed sanity checks on all struct pages being
allocated or freed, making sure they have no unexpected page flags or
certain field values. This can detect insufficient cleanup and some cases
of use-after-free, although on its own it can't always identify the
culprit. The result is a warning and the "bad page" being leaked.
The checks do need some cpu cycles, so in 4.7 with commits 479f854a207c
("mm, page_alloc: defer debugging checks of pages allocated from the PCP")
and 4db7548ccbd9 ("mm, page_alloc: defer debugging checks of freed pages
until a PCP drain") they were no longer performed in the hot paths when
allocating and freeing from pcplists, but only when pcplists are bypassed,
refilled or drained. For debugging purposes, with CONFIG_DEBUG_VM enabled
the checks were instead still done in the hot paths and not when refilling
or draining pcplists.
With 4462b32c9285 ("mm, page_alloc: more extensive free page checking with
debug_pagealloc"), enabling debug_pagealloc also moved the sanity checks
back to hot pahs. When both debug_pagealloc and CONFIG_DEBUG_VM are
enabled, the checks are done both in hotpaths and pcplist refill/drain.
Even though the non-debug default today might seem to be a sensible
tradeoff between overhead and ability to detect bad pages, on closer look
it's arguably not. As most allocations go through the pcplists, catching
any bad pages when refilling or draining pcplists has only a small chance,
insufficient for debugging or serious hardening purposes. On the other
hand the cost of the checks is concentrated in the already expensive
drain/refill batching operations, and those are done under the often
contended zone lock. That was recently identified as an issue for page
allocation and the zone lock contention reduced by moving the checks
outside of the locked section with a patch "mm: reduce lock contention of
pcp buffer refill", but the cost of the checks is still visible compared
to their removal [1]. In the pcplist draining path free_pcppages_bulk()
the checks are still done under zone->lock.
Thus, remove the checks from pcplist refill and drain paths completely.
Introduce a static key check_pages_enabled to control checks during page
allocation a freeing (whether pcplist is used or bypassed). The static
key is enabled if either is true:
- kernel is built with CONFIG_DEBUG_VM=y (debugging)
- debug_pagealloc or page poisoning is boot-time enabled (debugging)
- init_on_alloc or init_on_free is boot-time enabled (hardening)
The resulting user visible changes:
- no checks when draining/refilling pcplists - less overhead, with
likely no practical reduction of ability to catch bad pages
- no checks when bypassing pcplists in default config (no
debugging/hardening) - less overhead etc. as above
- on typical hardened kernels [2], checks are now performed on each page
allocation/free (previously only when bypassing/draining/refilling
pcplists) - the init_on_alloc/init_on_free enabled should be sufficient
indication for preferring more costly alloc/free operations for
hardening purposes and we shouldn't need to introduce another toggle
- code (various wrappers) removal and simplification
[1] https://lore.kernel.org/all/68ba44d8-6899-c018-dcb3-36f3a96e6bea@sra.uni-hannover.de/
[2] https://lore.kernel.org/all/63ebc499.a70a0220.9ac51.29ea@mx.google.com/
[akpm@linux-foundation.org: coding-style cleanups]
[akpm@linux-foundation.org: make check_pages_enabled static]
Link: https://lkml.kernel.org/r/20230216095131.17336-1-vbabka@suse.cz
Reported-by: Alexander Halbuer <halbuer@sra.uni-hannover.de>
Reported-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Kees Cook <keescook@chromium.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-02-16 10:51:31 +01:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
mm: tune PCP high automatically
The target to tune PCP high automatically is as follows,
- Minimize allocation/freeing from/to shared zone
- Minimize idle pages in PCP
- Minimize pages in PCP if the system free pages is too few
To reach these target, a tuning algorithm as follows is designed,
- When we refill PCP via allocating from the zone, increase PCP high.
Because if we had larger PCP, we could avoid to allocate from the
zone.
- In periodic vmstat updating kworker (via refresh_cpu_vm_stats()),
decrease PCP high to try to free possible idle PCP pages.
- When page reclaiming is active for the zone, stop increasing PCP
high in allocating path, decrease PCP high and free some pages in
freeing path.
So, the PCP high can be tuned to the page allocating/freeing depth of
workloads eventually.
One issue of the algorithm is that if the number of pages allocated is
much more than that of pages freed on a CPU, the PCP high may become the
maximal value even if the allocating/freeing depth is small. But this
isn't a severe issue, because there are no idle pages in this case.
One alternative choice is to increase PCP high when we drain PCP via
trying to free pages to the zone, but don't increase PCP high during PCP
refilling. This can avoid the issue above. But if the number of pages
allocated is much less than that of pages freed on a CPU, there will be
many idle pages in PCP and it is hard to free these idle pages.
1/8 (>> 3) of PCP high will be decreased periodically. The value 1/8 is
kind of arbitrary. Just to make sure that the idle PCP pages will be
freed eventually.
On a 2-socket Intel server with 224 logical CPU, we run 8 kbuild instances
in parallel (each with `make -j 28`) in 8 cgroup. This simulates the
kbuild server that is used by 0-Day kbuild service. With the patch, the
build time decreases 3.5%. The cycles% of the spinlock contention (mostly
for zone lock) decreases from 11.0% to 0.5%. The number of PCP draining
for high order pages freeing (free_high) decreases 65.6%. The number of
pages allocated from zone (instead of from PCP) decreases 83.9%.
Link: https://lkml.kernel.org/r/20231016053002.756205-8-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Suggested-by: Mel Gorman <mgorman@techsingularity.net>
Suggested-by: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:00 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-06-21 17:14:57 -07:00
|
|
|
|
2006-03-09 17:33:54 -08:00
|
|
|
|
2007-05-09 02:35:14 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-03-09 17:33:54 -08:00
|
|
|
|
2007-05-09 02:35:14 -07:00
|
|
|
|
2005-06-21 17:14:57 -07:00
|
|
|
|
2014-08-06 16:05:15 -07:00
|
|
|
|
2005-06-21 17:14:57 -07:00
|
|
|
|
2015-04-15 16:14:08 -07:00
|
|
|
|
2014-08-06 16:05:15 -07:00
|
|
|
|
2022-06-24 13:54:21 +01:00
|
|
|
|
2022-11-18 10:17:14 +00:00
|
|
|
|
2022-03-22 14:43:36 -07:00
|
|
|
|
2022-11-18 10:17:14 +00:00
|
|
|
|
2022-06-24 13:54:21 +01:00
|
|
|
|
2005-06-21 17:14:57 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-02-04 22:29:11 -08:00
|
|
|
|
2014-12-10 15:43:01 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: page_alloc: control latency caused by zone PCP draining
Patch series "mm/treewide: Remove pXd_huge() API", v2.
In previous work [1], we removed the pXd_large() API, which is arch
specific. This patchset further removes the hugetlb pXd_huge() API.
Hugetlb was never special on creating huge mappings when compared with
other huge mappings. Having a standalone API just to detect such pgtable
entries is more or less redundant, especially after the pXd_leaf() API set
is introduced with/without CONFIG_HUGETLB_PAGE.
When looking at this problem, a few issues are also exposed that we don't
have a clear definition of the *_huge() variance API. This patchset
started by cleaning these issues first, then replace all *_huge() users to
use *_leaf(), then drop all *_huge() code.
On x86/sparc, swap entries will be reported "true" in pXd_huge(), while
for all the rest archs they're reported "false" instead. This part is
done in patch 1-5, in which I suspect patch 1 can be seen as a bug fix,
but I'll leave that to hmm experts to decide.
Besides, there are three archs (arm, arm64, powerpc) that have slightly
different definitions between the *_huge() v.s. *_leaf() variances. I
tackled them separately so that it'll be easier for arch experts to chim
in when necessary. This part is done in patch 6-9.
The final patches 10-14 do the rest on the final removal, since *_leaf()
will be the ultimate API in the future, and we seem to have quite some
confusions on how *_huge() APIs can be defined, provide a rich comment for
*_leaf() API set to define them properly to avoid future misuse, and
hopefully that'll also help new archs to start support huge mappings and
avoid traps (like either swap entries, or PROT_NONE entry checks).
[1] https://lore.kernel.org/r/20240305043750.93762-1-peterx@redhat.com
This patch (of 14):
When the complete PCP is drained a much larger number of pages than the
usual batch size might be freed at once, causing large IRQ and preemption
latency spikes, as they are all freed while holding the pcp and zone
spinlocks.
To avoid those latency spikes, limit the number of pages freed in a single
bulk operation to common batch limits.
Link: https://lkml.kernel.org/r/20240318200404.448346-1-peterx@redhat.com
Link: https://lkml.kernel.org/r/20240318200736.2835502-1-l.stach@pengutronix.de
Signed-off-by: Lucas Stach <l.stach@pengutronix.de>
Signed-off-by: Peter Xu <peterx@redhat.com>
Cc: Christophe Leroy <christophe.leroy@csgroup.eu>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Mike Rapoport (IBM) <rppt@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Alistair Popple <apopple@nvidia.com>
Cc: Andreas Larsson <andreas@gaisler.com>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@kernel.org>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Bjorn Andersson <andersson@kernel.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Fabio Estevam <festevam@denx.de>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Konrad Dybcio <konrad.dybcio@linaro.org>
Cc: Krzysztof Kozlowski <krzysztof.kozlowski@linaro.org>
Cc: Mark Salter <msalter@redhat.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Naoya Horiguchi <nao.horiguchi@gmail.com>
Cc: "Naveen N. Rao" <naveen.n.rao@linux.ibm.com>
Cc: Nicholas Piggin <npiggin@gmail.com>
Cc: Russell King <linux@armlinux.org.uk>
Cc: Shawn Guo <shawnguo@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-18 21:07:36 +01:00
|
|
|
|
2024-07-23 14:44:28 +08:00
|
|
|
|
2014-12-10 15:43:01 -08:00
|
|
|
|
2024-07-23 14:44:28 +08:00
|
|
|
|
2022-11-18 10:17:14 +00:00
|
|
|
|
2024-07-23 14:44:28 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-11-18 10:17:14 +00:00
|
|
|
|
2024-07-23 14:44:28 +08:00
|
|
|
|
2014-12-10 15:43:01 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-02-04 22:29:11 -08:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2009-03-31 15:19:31 -07:00
|
|
|
|
2014-12-10 15:43:01 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-02-04 22:29:11 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-12-10 15:43:01 -08:00
|
|
|
|
2008-02-04 22:29:11 -08:00
|
|
|
|
2014-12-10 15:43:01 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-02-04 22:29:11 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, page_alloc: disable pcplists during memory offline
Memory offlining relies on page isolation to guarantee a forward progress
because pages cannot be reused while they are isolated. But the page
isolation itself doesn't prevent from races while freed pages are stored
on pcp lists and thus can be reused. This can be worked around by
repeated draining of pcplists, as done by commit 968318261221
("mm/memory_hotplug: drain per-cpu pages again during memory offline").
David and Michal would prefer that this race was closed in a way that
callers of page isolation who need stronger guarantees don't need to
repeatedly drain. David suggested disabling pcplists usage completely
during page isolation, instead of repeatedly draining them.
To achieve this without adding special cases in alloc/free fastpath, we
can use the same approach as boot pagesets - when pcp->high is 0, any
pcplist addition will be immediately flushed.
The race can thus be closed by setting pcp->high to 0 and draining
pcplists once, before calling start_isolate_page_range(). The draining
will serialize after processes that already disabled interrupts and read
the old value of pcp->high in free_unref_page_commit(), and processes that
have not yet disabled interrupts, will observe pcp->high == 0 when they
are rescheduled, and skip pcplists. This guarantees no stray pages on
pcplists in zones where isolation happens.
This patch thus adds zone_pcp_disable() and zone_pcp_enable() functions
that page isolation users can call before start_isolate_page_range() and
after unisolating (or offlining) the isolated pages.
Also, drain_all_pages() is optimized to only execute on cpus where
pcplists are not empty. The check can however race with a free to pcplist
that has not yet increased the pcp->count from 0 to 1. Thus make the
drain optionally skip the racy check and drain on all cpus, and use this
option in zone_pcp_disable().
As we have to avoid external updates to high and batch while pcplists are
disabled, we take pcp_batch_high_lock in zone_pcp_disable() and release it
in zone_pcp_enable(). This also synchronizes multiple users of
zone_pcp_disable()/enable().
Currently the only user of this functionality is offline_pages().
[vbabka@suse.cz: add comment, per David]
Link: https://lkml.kernel.org/r/527480ef-ed72-e1c1-52a0-1c5b0113df45@suse.cz
Link: https://lkml.kernel.org/r/20201111092812.11329-8-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Suggested-by: David Hildenbrand <david@redhat.com>
Suggested-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: David Hildenbrand <david@redhat.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-14 19:10:59 -08:00
|
|
|
|
|
|
|
|
|
2012-03-28 14:42:45 -07:00
|
|
|
|
mm, page_alloc: disable pcplists during memory offline
Memory offlining relies on page isolation to guarantee a forward progress
because pages cannot be reused while they are isolated. But the page
isolation itself doesn't prevent from races while freed pages are stored
on pcp lists and thus can be reused. This can be worked around by
repeated draining of pcplists, as done by commit 968318261221
("mm/memory_hotplug: drain per-cpu pages again during memory offline").
David and Michal would prefer that this race was closed in a way that
callers of page isolation who need stronger guarantees don't need to
repeatedly drain. David suggested disabling pcplists usage completely
during page isolation, instead of repeatedly draining them.
To achieve this without adding special cases in alloc/free fastpath, we
can use the same approach as boot pagesets - when pcp->high is 0, any
pcplist addition will be immediately flushed.
The race can thus be closed by setting pcp->high to 0 and draining
pcplists once, before calling start_isolate_page_range(). The draining
will serialize after processes that already disabled interrupts and read
the old value of pcp->high in free_unref_page_commit(), and processes that
have not yet disabled interrupts, will observe pcp->high == 0 when they
are rescheduled, and skip pcplists. This guarantees no stray pages on
pcplists in zones where isolation happens.
This patch thus adds zone_pcp_disable() and zone_pcp_enable() functions
that page isolation users can call before start_isolate_page_range() and
after unisolating (or offlining) the isolated pages.
Also, drain_all_pages() is optimized to only execute on cpus where
pcplists are not empty. The check can however race with a free to pcplist
that has not yet increased the pcp->count from 0 to 1. Thus make the
drain optionally skip the racy check and drain on all cpus, and use this
option in zone_pcp_disable().
As we have to avoid external updates to high and batch while pcplists are
disabled, we take pcp_batch_high_lock in zone_pcp_disable() and release it
in zone_pcp_enable(). This also synchronizes multiple users of
zone_pcp_disable()/enable().
Currently the only user of this functionality is offline_pages().
[vbabka@suse.cz: add comment, per David]
Link: https://lkml.kernel.org/r/527480ef-ed72-e1c1-52a0-1c5b0113df45@suse.cz
Link: https://lkml.kernel.org/r/20201111092812.11329-8-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Suggested-by: David Hildenbrand <david@redhat.com>
Suggested-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: David Hildenbrand <david@redhat.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-14 19:10:59 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-02-04 22:29:11 -08:00
|
|
|
|
2020-12-14 19:11:12 -08:00
|
|
|
|
2008-02-04 22:29:11 -08:00
|
|
|
|
2012-03-28 14:42:45 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-30 18:53:17 -07:00
|
|
|
|
2012-03-28 14:42:45 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-02-24 14:56:56 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-02-24 14:56:32 -08:00
|
|
|
|
2012-03-28 14:42:45 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/page_alloc: split per cpu page lists and zone stats
The PCP (per-cpu page allocator in page_alloc.c) shares locking
requirements with vmstat and the zone lock which is inconvenient and
causes some issues. For example, the PCP list and vmstat share the same
per-cpu space meaning that it's possible that vmstat updates dirty cache
lines holding per-cpu lists across CPUs unless padding is used. Second,
PREEMPT_RT does not want to disable IRQs for too long in the page
allocator.
This series splits the locking requirements and uses locks types more
suitable for PREEMPT_RT, reduces the time when special locking is required
for stats and reduces the time when IRQs need to be disabled on
!PREEMPT_RT kernels.
Why local_lock? PREEMPT_RT considers the following sequence to be unsafe
as documented in Documentation/locking/locktypes.rst
local_irq_disable();
spin_lock(&lock);
The pcp allocator has this sequence for rmqueue_pcplist (local_irq_save)
-> __rmqueue_pcplist -> rmqueue_bulk (spin_lock). While it's possible to
separate this out, it generally means there are points where we enable
IRQs and reenable them again immediately. To prevent a migration and the
per-cpu pointer going stale, migrate_disable is also needed. That is a
custom lock that is similar, but worse, than local_lock. Furthermore, on
PREEMPT_RT, it's undesirable to leave IRQs disabled for too long. By
converting to local_lock which disables migration on PREEMPT_RT, the
locking requirements can be separated and start moving the protections for
PCP, stats and the zone lock to PREEMPT_RT-safe equivalent locking. As a
bonus, local_lock also means that PROVE_LOCKING does something useful.
After that, it's obvious that zone_statistics incurs too much overhead and
leaves IRQs disabled for longer than necessary on !PREEMPT_RT kernels.
zone_statistics uses perfectly accurate counters requiring IRQs be
disabled for parallel RMW sequences when inaccurate ones like vm_events
would do. The series makes the NUMA statistics (NUMA_HIT and friends)
inaccurate counters that then require no special protection on
!PREEMPT_RT.
The bulk page allocator can then do stat updates in bulk with IRQs enabled
which should improve the efficiency. Technically, this could have been
done without the local_lock and vmstat conversion work and the order
simply reflects the timing of when different series were implemented.
Finally, there are places where we conflate IRQs being disabled for the
PCP with the IRQ-safe zone spinlock. The remainder of the series reduces
the scope of what is protected by disabled IRQs on !PREEMPT_RT kernels.
By the end of the series, page_alloc.c does not call local_irq_save so the
locking scope is a bit clearer. The one exception is that modifying
NR_FREE_PAGES still happens in places where it's known the IRQs are
disabled as it's harmless for PREEMPT_RT and would be expensive to split
the locking there.
No performance data is included because despite the overhead of the stats,
it's within the noise for most workloads on !PREEMPT_RT. However, Jesper
Dangaard Brouer ran a page allocation microbenchmark on a E5-1650 v4 @
3.60GHz CPU on the first version of this series. Focusing on the array
variant of the bulk page allocator reveals the following.
(CPU: Intel(R) Xeon(R) CPU E5-1650 v4 @ 3.60GHz)
ARRAY variant: time_bulk_page_alloc_free_array: step=bulk size
Baseline Patched
1 56.383 54.225 (+3.83%)
2 40.047 35.492 (+11.38%)
3 37.339 32.643 (+12.58%)
4 35.578 30.992 (+12.89%)
8 33.592 29.606 (+11.87%)
16 32.362 28.532 (+11.85%)
32 31.476 27.728 (+11.91%)
64 30.633 27.252 (+11.04%)
128 30.596 27.090 (+11.46%)
While this is a positive outcome, the series is more likely to be
interesting to the RT people in terms of getting parts of the PREEMPT_RT
tree into mainline.
This patch (of 9):
The per-cpu page allocator lists and the per-cpu vmstat deltas are stored
in the same struct per_cpu_pages even though vmstats have no direct impact
on the per-cpu page lists. This is inconsistent because the vmstats for a
node are stored on a dedicated structure. The bigger issue is that the
per_cpu_pages structure is not cache-aligned and stat updates either cache
conflict with adjacent per-cpu lists incurring a runtime cost or padding
is required incurring a memory cost.
This patch splits the per-cpu pagelists and the vmstat deltas into
separate structures. It's mostly a mechanical conversion but some
variable renaming is done to clearly distinguish the per-cpu pages
structure (pcp) from the vmstats (pzstats).
Superficially, this appears to increase the size of the per_cpu_pages
structure but the movement of expire fills a structure hole so there is no
impact overall.
[mgorman@techsingularity.net: make it W=1 cleaner]
Link: https://lkml.kernel.org/r/20210514144622.GA3735@techsingularity.net
[mgorman@techsingularity.net: make it W=1 even cleaner]
Link: https://lkml.kernel.org/r/20210516140705.GB3735@techsingularity.net
[lkp@intel.com: check struct per_cpu_zonestat has a non-zero size]
[vbabka@suse.cz: Init zone->per_cpu_zonestats properly]
Link: https://lkml.kernel.org/r/20210512095458.30632-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20210512095458.30632-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-28 19:41:38 -07:00
|
|
|
|
2014-12-10 15:43:01 -08:00
|
|
|
|
2012-03-28 14:42:45 -07:00
|
|
|
|
2014-12-10 15:43:01 -08:00
|
|
|
|
mm, page_alloc: disable pcplists during memory offline
Memory offlining relies on page isolation to guarantee a forward progress
because pages cannot be reused while they are isolated. But the page
isolation itself doesn't prevent from races while freed pages are stored
on pcp lists and thus can be reused. This can be worked around by
repeated draining of pcplists, as done by commit 968318261221
("mm/memory_hotplug: drain per-cpu pages again during memory offline").
David and Michal would prefer that this race was closed in a way that
callers of page isolation who need stronger guarantees don't need to
repeatedly drain. David suggested disabling pcplists usage completely
during page isolation, instead of repeatedly draining them.
To achieve this without adding special cases in alloc/free fastpath, we
can use the same approach as boot pagesets - when pcp->high is 0, any
pcplist addition will be immediately flushed.
The race can thus be closed by setting pcp->high to 0 and draining
pcplists once, before calling start_isolate_page_range(). The draining
will serialize after processes that already disabled interrupts and read
the old value of pcp->high in free_unref_page_commit(), and processes that
have not yet disabled interrupts, will observe pcp->high == 0 when they
are rescheduled, and skip pcplists. This guarantees no stray pages on
pcplists in zones where isolation happens.
This patch thus adds zone_pcp_disable() and zone_pcp_enable() functions
that page isolation users can call before start_isolate_page_range() and
after unisolating (or offlining) the isolated pages.
Also, drain_all_pages() is optimized to only execute on cpus where
pcplists are not empty. The check can however race with a free to pcplist
that has not yet increased the pcp->count from 0 to 1. Thus make the
drain optionally skip the racy check and drain on all cpus, and use this
option in zone_pcp_disable().
As we have to avoid external updates to high and batch while pcplists are
disabled, we take pcp_batch_high_lock in zone_pcp_disable() and release it
in zone_pcp_enable(). This also synchronizes multiple users of
zone_pcp_disable()/enable().
Currently the only user of this functionality is offline_pages().
[vbabka@suse.cz: add comment, per David]
Link: https://lkml.kernel.org/r/527480ef-ed72-e1c1-52a0-1c5b0113df45@suse.cz
Link: https://lkml.kernel.org/r/20201111092812.11329-8-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Suggested-by: David Hildenbrand <david@redhat.com>
Suggested-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: David Hildenbrand <david@redhat.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-14 19:10:59 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/page_alloc: split per cpu page lists and zone stats
The PCP (per-cpu page allocator in page_alloc.c) shares locking
requirements with vmstat and the zone lock which is inconvenient and
causes some issues. For example, the PCP list and vmstat share the same
per-cpu space meaning that it's possible that vmstat updates dirty cache
lines holding per-cpu lists across CPUs unless padding is used. Second,
PREEMPT_RT does not want to disable IRQs for too long in the page
allocator.
This series splits the locking requirements and uses locks types more
suitable for PREEMPT_RT, reduces the time when special locking is required
for stats and reduces the time when IRQs need to be disabled on
!PREEMPT_RT kernels.
Why local_lock? PREEMPT_RT considers the following sequence to be unsafe
as documented in Documentation/locking/locktypes.rst
local_irq_disable();
spin_lock(&lock);
The pcp allocator has this sequence for rmqueue_pcplist (local_irq_save)
-> __rmqueue_pcplist -> rmqueue_bulk (spin_lock). While it's possible to
separate this out, it generally means there are points where we enable
IRQs and reenable them again immediately. To prevent a migration and the
per-cpu pointer going stale, migrate_disable is also needed. That is a
custom lock that is similar, but worse, than local_lock. Furthermore, on
PREEMPT_RT, it's undesirable to leave IRQs disabled for too long. By
converting to local_lock which disables migration on PREEMPT_RT, the
locking requirements can be separated and start moving the protections for
PCP, stats and the zone lock to PREEMPT_RT-safe equivalent locking. As a
bonus, local_lock also means that PROVE_LOCKING does something useful.
After that, it's obvious that zone_statistics incurs too much overhead and
leaves IRQs disabled for longer than necessary on !PREEMPT_RT kernels.
zone_statistics uses perfectly accurate counters requiring IRQs be
disabled for parallel RMW sequences when inaccurate ones like vm_events
would do. The series makes the NUMA statistics (NUMA_HIT and friends)
inaccurate counters that then require no special protection on
!PREEMPT_RT.
The bulk page allocator can then do stat updates in bulk with IRQs enabled
which should improve the efficiency. Technically, this could have been
done without the local_lock and vmstat conversion work and the order
simply reflects the timing of when different series were implemented.
Finally, there are places where we conflate IRQs being disabled for the
PCP with the IRQ-safe zone spinlock. The remainder of the series reduces
the scope of what is protected by disabled IRQs on !PREEMPT_RT kernels.
By the end of the series, page_alloc.c does not call local_irq_save so the
locking scope is a bit clearer. The one exception is that modifying
NR_FREE_PAGES still happens in places where it's known the IRQs are
disabled as it's harmless for PREEMPT_RT and would be expensive to split
the locking there.
No performance data is included because despite the overhead of the stats,
it's within the noise for most workloads on !PREEMPT_RT. However, Jesper
Dangaard Brouer ran a page allocation microbenchmark on a E5-1650 v4 @
3.60GHz CPU on the first version of this series. Focusing on the array
variant of the bulk page allocator reveals the following.
(CPU: Intel(R) Xeon(R) CPU E5-1650 v4 @ 3.60GHz)
ARRAY variant: time_bulk_page_alloc_free_array: step=bulk size
Baseline Patched
1 56.383 54.225 (+3.83%)
2 40.047 35.492 (+11.38%)
3 37.339 32.643 (+12.58%)
4 35.578 30.992 (+12.89%)
8 33.592 29.606 (+11.87%)
16 32.362 28.532 (+11.85%)
32 31.476 27.728 (+11.91%)
64 30.633 27.252 (+11.04%)
128 30.596 27.090 (+11.46%)
While this is a positive outcome, the series is more likely to be
interesting to the RT people in terms of getting parts of the PREEMPT_RT
tree into mainline.
This patch (of 9):
The per-cpu page allocator lists and the per-cpu vmstat deltas are stored
in the same struct per_cpu_pages even though vmstats have no direct impact
on the per-cpu page lists. This is inconsistent because the vmstats for a
node are stored on a dedicated structure. The bigger issue is that the
per_cpu_pages structure is not cache-aligned and stat updates either cache
conflict with adjacent per-cpu lists incurring a runtime cost or padding
is required incurring a memory cost.
This patch splits the per-cpu pagelists and the vmstat deltas into
separate structures. It's mostly a mechanical conversion but some
variable renaming is done to clearly distinguish the per-cpu pages
structure (pcp) from the vmstats (pzstats).
Superficially, this appears to increase the size of the per_cpu_pages
structure but the movement of expire fills a structure hole so there is no
impact overall.
[mgorman@techsingularity.net: make it W=1 cleaner]
Link: https://lkml.kernel.org/r/20210514144622.GA3735@techsingularity.net
[mgorman@techsingularity.net: make it W=1 even cleaner]
Link: https://lkml.kernel.org/r/20210516140705.GB3735@techsingularity.net
[lkp@intel.com: check struct per_cpu_zonestat has a non-zero size]
[vbabka@suse.cz: Init zone->per_cpu_zonestats properly]
Link: https://lkml.kernel.org/r/20210512095458.30632-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20210512095458.30632-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-28 19:41:38 -07:00
|
|
|
|
|
|
|
|
|
2012-03-28 14:42:45 -07:00
|
|
|
|
2014-12-10 15:43:01 -08:00
|
|
|
|
|
|
|
|
|
mm/page_alloc: split per cpu page lists and zone stats
The PCP (per-cpu page allocator in page_alloc.c) shares locking
requirements with vmstat and the zone lock which is inconvenient and
causes some issues. For example, the PCP list and vmstat share the same
per-cpu space meaning that it's possible that vmstat updates dirty cache
lines holding per-cpu lists across CPUs unless padding is used. Second,
PREEMPT_RT does not want to disable IRQs for too long in the page
allocator.
This series splits the locking requirements and uses locks types more
suitable for PREEMPT_RT, reduces the time when special locking is required
for stats and reduces the time when IRQs need to be disabled on
!PREEMPT_RT kernels.
Why local_lock? PREEMPT_RT considers the following sequence to be unsafe
as documented in Documentation/locking/locktypes.rst
local_irq_disable();
spin_lock(&lock);
The pcp allocator has this sequence for rmqueue_pcplist (local_irq_save)
-> __rmqueue_pcplist -> rmqueue_bulk (spin_lock). While it's possible to
separate this out, it generally means there are points where we enable
IRQs and reenable them again immediately. To prevent a migration and the
per-cpu pointer going stale, migrate_disable is also needed. That is a
custom lock that is similar, but worse, than local_lock. Furthermore, on
PREEMPT_RT, it's undesirable to leave IRQs disabled for too long. By
converting to local_lock which disables migration on PREEMPT_RT, the
locking requirements can be separated and start moving the protections for
PCP, stats and the zone lock to PREEMPT_RT-safe equivalent locking. As a
bonus, local_lock also means that PROVE_LOCKING does something useful.
After that, it's obvious that zone_statistics incurs too much overhead and
leaves IRQs disabled for longer than necessary on !PREEMPT_RT kernels.
zone_statistics uses perfectly accurate counters requiring IRQs be
disabled for parallel RMW sequences when inaccurate ones like vm_events
would do. The series makes the NUMA statistics (NUMA_HIT and friends)
inaccurate counters that then require no special protection on
!PREEMPT_RT.
The bulk page allocator can then do stat updates in bulk with IRQs enabled
which should improve the efficiency. Technically, this could have been
done without the local_lock and vmstat conversion work and the order
simply reflects the timing of when different series were implemented.
Finally, there are places where we conflate IRQs being disabled for the
PCP with the IRQ-safe zone spinlock. The remainder of the series reduces
the scope of what is protected by disabled IRQs on !PREEMPT_RT kernels.
By the end of the series, page_alloc.c does not call local_irq_save so the
locking scope is a bit clearer. The one exception is that modifying
NR_FREE_PAGES still happens in places where it's known the IRQs are
disabled as it's harmless for PREEMPT_RT and would be expensive to split
the locking there.
No performance data is included because despite the overhead of the stats,
it's within the noise for most workloads on !PREEMPT_RT. However, Jesper
Dangaard Brouer ran a page allocation microbenchmark on a E5-1650 v4 @
3.60GHz CPU on the first version of this series. Focusing on the array
variant of the bulk page allocator reveals the following.
(CPU: Intel(R) Xeon(R) CPU E5-1650 v4 @ 3.60GHz)
ARRAY variant: time_bulk_page_alloc_free_array: step=bulk size
Baseline Patched
1 56.383 54.225 (+3.83%)
2 40.047 35.492 (+11.38%)
3 37.339 32.643 (+12.58%)
4 35.578 30.992 (+12.89%)
8 33.592 29.606 (+11.87%)
16 32.362 28.532 (+11.85%)
32 31.476 27.728 (+11.91%)
64 30.633 27.252 (+11.04%)
128 30.596 27.090 (+11.46%)
While this is a positive outcome, the series is more likely to be
interesting to the RT people in terms of getting parts of the PREEMPT_RT
tree into mainline.
This patch (of 9):
The per-cpu page allocator lists and the per-cpu vmstat deltas are stored
in the same struct per_cpu_pages even though vmstats have no direct impact
on the per-cpu page lists. This is inconsistent because the vmstats for a
node are stored on a dedicated structure. The bigger issue is that the
per_cpu_pages structure is not cache-aligned and stat updates either cache
conflict with adjacent per-cpu lists incurring a runtime cost or padding
is required incurring a memory cost.
This patch splits the per-cpu pagelists and the vmstat deltas into
separate structures. It's mostly a mechanical conversion but some
variable renaming is done to clearly distinguish the per-cpu pages
structure (pcp) from the vmstats (pzstats).
Superficially, this appears to increase the size of the per_cpu_pages
structure but the movement of expire fills a structure hole so there is no
impact overall.
[mgorman@techsingularity.net: make it W=1 cleaner]
Link: https://lkml.kernel.org/r/20210514144622.GA3735@techsingularity.net
[mgorman@techsingularity.net: make it W=1 even cleaner]
Link: https://lkml.kernel.org/r/20210516140705.GB3735@techsingularity.net
[lkp@intel.com: check struct per_cpu_zonestat has a non-zero size]
[vbabka@suse.cz: Init zone->per_cpu_zonestats properly]
Link: https://lkml.kernel.org/r/20210512095458.30632-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20210512095458.30632-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-28 19:41:38 -07:00
|
|
|
|
|
|
|
|
|
2014-12-10 15:43:01 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-03-28 14:42:45 -07:00
|
|
|
|
|
|
|
|
|
2014-12-10 15:43:01 -08:00
|
|
|
|
2012-03-28 14:42:45 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-02-24 14:56:32 -08:00
|
|
|
|
2017-02-24 14:56:56 -08:00
|
|
|
|
2022-06-24 13:54:22 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-02-24 14:56:32 -08:00
|
|
|
|
2017-02-24 14:56:56 -08:00
|
|
|
|
|
|
|
|
|
2008-02-04 22:29:11 -08:00
|
|
|
|
|
|
|
|
|
mm, page_alloc: disable pcplists during memory offline
Memory offlining relies on page isolation to guarantee a forward progress
because pages cannot be reused while they are isolated. But the page
isolation itself doesn't prevent from races while freed pages are stored
on pcp lists and thus can be reused. This can be worked around by
repeated draining of pcplists, as done by commit 968318261221
("mm/memory_hotplug: drain per-cpu pages again during memory offline").
David and Michal would prefer that this race was closed in a way that
callers of page isolation who need stronger guarantees don't need to
repeatedly drain. David suggested disabling pcplists usage completely
during page isolation, instead of repeatedly draining them.
To achieve this without adding special cases in alloc/free fastpath, we
can use the same approach as boot pagesets - when pcp->high is 0, any
pcplist addition will be immediately flushed.
The race can thus be closed by setting pcp->high to 0 and draining
pcplists once, before calling start_isolate_page_range(). The draining
will serialize after processes that already disabled interrupts and read
the old value of pcp->high in free_unref_page_commit(), and processes that
have not yet disabled interrupts, will observe pcp->high == 0 when they
are rescheduled, and skip pcplists. This guarantees no stray pages on
pcplists in zones where isolation happens.
This patch thus adds zone_pcp_disable() and zone_pcp_enable() functions
that page isolation users can call before start_isolate_page_range() and
after unisolating (or offlining) the isolated pages.
Also, drain_all_pages() is optimized to only execute on cpus where
pcplists are not empty. The check can however race with a free to pcplist
that has not yet increased the pcp->count from 0 to 1. Thus make the
drain optionally skip the racy check and drain on all cpus, and use this
option in zone_pcp_disable().
As we have to avoid external updates to high and batch while pcplists are
disabled, we take pcp_batch_high_lock in zone_pcp_disable() and release it
in zone_pcp_enable(). This also synchronizes multiple users of
zone_pcp_disable()/enable().
Currently the only user of this functionality is offline_pages().
[vbabka@suse.cz: add comment, per David]
Link: https://lkml.kernel.org/r/527480ef-ed72-e1c1-52a0-1c5b0113df45@suse.cz
Link: https://lkml.kernel.org/r/20201111092812.11329-8-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Suggested-by: David Hildenbrand <david@redhat.com>
Suggested-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: David Hildenbrand <david@redhat.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-14 19:10:59 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: tune PCP high automatically
The target to tune PCP high automatically is as follows,
- Minimize allocation/freeing from/to shared zone
- Minimize idle pages in PCP
- Minimize pages in PCP if the system free pages is too few
To reach these target, a tuning algorithm as follows is designed,
- When we refill PCP via allocating from the zone, increase PCP high.
Because if we had larger PCP, we could avoid to allocate from the
zone.
- In periodic vmstat updating kworker (via refresh_cpu_vm_stats()),
decrease PCP high to try to free possible idle PCP pages.
- When page reclaiming is active for the zone, stop increasing PCP
high in allocating path, decrease PCP high and free some pages in
freeing path.
So, the PCP high can be tuned to the page allocating/freeing depth of
workloads eventually.
One issue of the algorithm is that if the number of pages allocated is
much more than that of pages freed on a CPU, the PCP high may become the
maximal value even if the allocating/freeing depth is small. But this
isn't a severe issue, because there are no idle pages in this case.
One alternative choice is to increase PCP high when we drain PCP via
trying to free pages to the zone, but don't increase PCP high during PCP
refilling. This can avoid the issue above. But if the number of pages
allocated is much less than that of pages freed on a CPU, there will be
many idle pages in PCP and it is hard to free these idle pages.
1/8 (>> 3) of PCP high will be decreased periodically. The value 1/8 is
kind of arbitrary. Just to make sure that the idle PCP pages will be
freed eventually.
On a 2-socket Intel server with 224 logical CPU, we run 8 kbuild instances
in parallel (each with `make -j 28`) in 8 cgroup. This simulates the
kbuild server that is used by 0-Day kbuild service. With the patch, the
build time decreases 3.5%. The cycles% of the spinlock contention (mostly
for zone lock) decreases from 11.0% to 0.5%. The number of PCP draining
for high order pages freeing (free_high) decreases 65.6%. The number of
pages allocated from zone (instead of from PCP) decreases 83.9%.
Link: https://lkml.kernel.org/r/20231016053002.756205-8-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Suggested-by: Mel Gorman <mgorman@techsingularity.net>
Suggested-by: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:00 +08:00
|
|
|
|
2021-06-28 19:42:18 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: tune PCP high automatically
The target to tune PCP high automatically is as follows,
- Minimize allocation/freeing from/to shared zone
- Minimize idle pages in PCP
- Minimize pages in PCP if the system free pages is too few
To reach these target, a tuning algorithm as follows is designed,
- When we refill PCP via allocating from the zone, increase PCP high.
Because if we had larger PCP, we could avoid to allocate from the
zone.
- In periodic vmstat updating kworker (via refresh_cpu_vm_stats()),
decrease PCP high to try to free possible idle PCP pages.
- When page reclaiming is active for the zone, stop increasing PCP
high in allocating path, decrease PCP high and free some pages in
freeing path.
So, the PCP high can be tuned to the page allocating/freeing depth of
workloads eventually.
One issue of the algorithm is that if the number of pages allocated is
much more than that of pages freed on a CPU, the PCP high may become the
maximal value even if the allocating/freeing depth is small. But this
isn't a severe issue, because there are no idle pages in this case.
One alternative choice is to increase PCP high when we drain PCP via
trying to free pages to the zone, but don't increase PCP high during PCP
refilling. This can avoid the issue above. But if the number of pages
allocated is much less than that of pages freed on a CPU, there will be
many idle pages in PCP and it is hard to free these idle pages.
1/8 (>> 3) of PCP high will be decreased periodically. The value 1/8 is
kind of arbitrary. Just to make sure that the idle PCP pages will be
freed eventually.
On a 2-socket Intel server with 224 logical CPU, we run 8 kbuild instances
in parallel (each with `make -j 28`) in 8 cgroup. This simulates the
kbuild server that is used by 0-Day kbuild service. With the patch, the
build time decreases 3.5%. The cycles% of the spinlock contention (mostly
for zone lock) decreases from 11.0% to 0.5%. The number of PCP draining
for high order pages freeing (free_high) decreases 65.6%. The number of
pages allocated from zone (instead of from PCP) decreases 83.9%.
Link: https://lkml.kernel.org/r/20231016053002.756205-8-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Suggested-by: Mel Gorman <mgorman@techsingularity.net>
Suggested-by: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:00 +08:00
|
|
|
|
2022-03-22 14:43:45 -07:00
|
|
|
|
mm: tune PCP high automatically
The target to tune PCP high automatically is as follows,
- Minimize allocation/freeing from/to shared zone
- Minimize idle pages in PCP
- Minimize pages in PCP if the system free pages is too few
To reach these target, a tuning algorithm as follows is designed,
- When we refill PCP via allocating from the zone, increase PCP high.
Because if we had larger PCP, we could avoid to allocate from the
zone.
- In periodic vmstat updating kworker (via refresh_cpu_vm_stats()),
decrease PCP high to try to free possible idle PCP pages.
- When page reclaiming is active for the zone, stop increasing PCP
high in allocating path, decrease PCP high and free some pages in
freeing path.
So, the PCP high can be tuned to the page allocating/freeing depth of
workloads eventually.
One issue of the algorithm is that if the number of pages allocated is
much more than that of pages freed on a CPU, the PCP high may become the
maximal value even if the allocating/freeing depth is small. But this
isn't a severe issue, because there are no idle pages in this case.
One alternative choice is to increase PCP high when we drain PCP via
trying to free pages to the zone, but don't increase PCP high during PCP
refilling. This can avoid the issue above. But if the number of pages
allocated is much less than that of pages freed on a CPU, there will be
many idle pages in PCP and it is hard to free these idle pages.
1/8 (>> 3) of PCP high will be decreased periodically. The value 1/8 is
kind of arbitrary. Just to make sure that the idle PCP pages will be
freed eventually.
On a 2-socket Intel server with 224 logical CPU, we run 8 kbuild instances
in parallel (each with `make -j 28`) in 8 cgroup. This simulates the
kbuild server that is used by 0-Day kbuild service. With the patch, the
build time decreases 3.5%. The cycles% of the spinlock contention (mostly
for zone lock) decreases from 11.0% to 0.5%. The number of PCP draining
for high order pages freeing (free_high) decreases 65.6%. The number of
pages allocated from zone (instead of from PCP) decreases 83.9%.
Link: https://lkml.kernel.org/r/20231016053002.756205-8-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Suggested-by: Mel Gorman <mgorman@techsingularity.net>
Suggested-by: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:00 +08:00
|
|
|
|
2022-03-22 14:43:45 -07:00
|
|
|
|
2021-06-28 19:42:18 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, pcp: reduce detecting time of consecutive high order page freeing
In current PCP auto-tuning design, if the number of pages allocated is
much more than that of pages freed on a CPU, the PCP high may become the
maximal value even if the allocating/freeing depth is small, for example,
in the sender of network workloads. If a CPU was used as sender
originally, then it is used as receiver after context switching, we need
to fill the whole PCP with maximal high before triggering PCP draining for
consecutive high order freeing. This will hurt the performance of some
network workloads.
To solve the issue, in this patch, we will track the consecutive page
freeing with a counter in stead of relying on PCP draining. So, we can
detect consecutive page freeing much earlier.
On a 2-socket Intel server with 128 logical CPU, we tested
SCTP_STREAM_MANY test case of netperf test suite with 64-pair processes.
With the patch, the network bandwidth improves 5.0%. This restores the
performance drop caused by PCP auto-tuning.
Link: https://lkml.kernel.org/r/20231016053002.756205-10-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:02 +08:00
|
|
|
|
|
|
|
|
|
2021-06-28 19:42:18 -07:00
|
|
|
|
mm, pcp: reduce detecting time of consecutive high order page freeing
In current PCP auto-tuning design, if the number of pages allocated is
much more than that of pages freed on a CPU, the PCP high may become the
maximal value even if the allocating/freeing depth is small, for example,
in the sender of network workloads. If a CPU was used as sender
originally, then it is used as receiver after context switching, we need
to fill the whole PCP with maximal high before triggering PCP draining for
consecutive high order freeing. This will hurt the performance of some
network workloads.
To solve the issue, in this patch, we will track the consecutive page
freeing with a counter in stead of relying on PCP draining. So, we can
detect consecutive page freeing much earlier.
On a 2-socket Intel server with 128 logical CPU, we tested
SCTP_STREAM_MANY test case of netperf test suite with 64-pair processes.
With the patch, the network bandwidth improves 5.0%. This restores the
performance drop caused by PCP auto-tuning.
Link: https://lkml.kernel.org/r/20231016053002.756205-10-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:02 +08:00
|
|
|
|
2021-06-28 19:42:18 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-03-22 14:43:45 -07:00
|
|
|
|
mm: tune PCP high automatically
The target to tune PCP high automatically is as follows,
- Minimize allocation/freeing from/to shared zone
- Minimize idle pages in PCP
- Minimize pages in PCP if the system free pages is too few
To reach these target, a tuning algorithm as follows is designed,
- When we refill PCP via allocating from the zone, increase PCP high.
Because if we had larger PCP, we could avoid to allocate from the
zone.
- In periodic vmstat updating kworker (via refresh_cpu_vm_stats()),
decrease PCP high to try to free possible idle PCP pages.
- When page reclaiming is active for the zone, stop increasing PCP
high in allocating path, decrease PCP high and free some pages in
freeing path.
So, the PCP high can be tuned to the page allocating/freeing depth of
workloads eventually.
One issue of the algorithm is that if the number of pages allocated is
much more than that of pages freed on a CPU, the PCP high may become the
maximal value even if the allocating/freeing depth is small. But this
isn't a severe issue, because there are no idle pages in this case.
One alternative choice is to increase PCP high when we drain PCP via
trying to free pages to the zone, but don't increase PCP high during PCP
refilling. This can avoid the issue above. But if the number of pages
allocated is much less than that of pages freed on a CPU, there will be
many idle pages in PCP and it is hard to free these idle pages.
1/8 (>> 3) of PCP high will be decreased periodically. The value 1/8 is
kind of arbitrary. Just to make sure that the idle PCP pages will be
freed eventually.
On a 2-socket Intel server with 224 logical CPU, we run 8 kbuild instances
in parallel (each with `make -j 28`) in 8 cgroup. This simulates the
kbuild server that is used by 0-Day kbuild service. With the patch, the
build time decreases 3.5%. The cycles% of the spinlock contention (mostly
for zone lock) decreases from 11.0% to 0.5%. The number of PCP draining
for high order pages freeing (free_high) decreases 65.6%. The number of
pages allocated from zone (instead of from PCP) decreases 83.9%.
Link: https://lkml.kernel.org/r/20231016053002.756205-8-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Suggested-by: Mel Gorman <mgorman@techsingularity.net>
Suggested-by: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:00 +08:00
|
|
|
|
2021-06-28 19:42:21 -07:00
|
|
|
|
mm: tune PCP high automatically
The target to tune PCP high automatically is as follows,
- Minimize allocation/freeing from/to shared zone
- Minimize idle pages in PCP
- Minimize pages in PCP if the system free pages is too few
To reach these target, a tuning algorithm as follows is designed,
- When we refill PCP via allocating from the zone, increase PCP high.
Because if we had larger PCP, we could avoid to allocate from the
zone.
- In periodic vmstat updating kworker (via refresh_cpu_vm_stats()),
decrease PCP high to try to free possible idle PCP pages.
- When page reclaiming is active for the zone, stop increasing PCP
high in allocating path, decrease PCP high and free some pages in
freeing path.
So, the PCP high can be tuned to the page allocating/freeing depth of
workloads eventually.
One issue of the algorithm is that if the number of pages allocated is
much more than that of pages freed on a CPU, the PCP high may become the
maximal value even if the allocating/freeing depth is small. But this
isn't a severe issue, because there are no idle pages in this case.
One alternative choice is to increase PCP high when we drain PCP via
trying to free pages to the zone, but don't increase PCP high during PCP
refilling. This can avoid the issue above. But if the number of pages
allocated is much less than that of pages freed on a CPU, there will be
many idle pages in PCP and it is hard to free these idle pages.
1/8 (>> 3) of PCP high will be decreased periodically. The value 1/8 is
kind of arbitrary. Just to make sure that the idle PCP pages will be
freed eventually.
On a 2-socket Intel server with 224 logical CPU, we run 8 kbuild instances
in parallel (each with `make -j 28`) in 8 cgroup. This simulates the
kbuild server that is used by 0-Day kbuild service. With the patch, the
build time decreases 3.5%. The cycles% of the spinlock contention (mostly
for zone lock) decreases from 11.0% to 0.5%. The number of PCP draining
for high order pages freeing (free_high) decreases 65.6%. The number of
pages allocated from zone (instead of from PCP) decreases 83.9%.
Link: https://lkml.kernel.org/r/20231016053002.756205-8-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Suggested-by: Mel Gorman <mgorman@techsingularity.net>
Suggested-by: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:00 +08:00
|
|
|
|
2021-06-28 19:42:21 -07:00
|
|
|
|
mm: tune PCP high automatically
The target to tune PCP high automatically is as follows,
- Minimize allocation/freeing from/to shared zone
- Minimize idle pages in PCP
- Minimize pages in PCP if the system free pages is too few
To reach these target, a tuning algorithm as follows is designed,
- When we refill PCP via allocating from the zone, increase PCP high.
Because if we had larger PCP, we could avoid to allocate from the
zone.
- In periodic vmstat updating kworker (via refresh_cpu_vm_stats()),
decrease PCP high to try to free possible idle PCP pages.
- When page reclaiming is active for the zone, stop increasing PCP
high in allocating path, decrease PCP high and free some pages in
freeing path.
So, the PCP high can be tuned to the page allocating/freeing depth of
workloads eventually.
One issue of the algorithm is that if the number of pages allocated is
much more than that of pages freed on a CPU, the PCP high may become the
maximal value even if the allocating/freeing depth is small. But this
isn't a severe issue, because there are no idle pages in this case.
One alternative choice is to increase PCP high when we drain PCP via
trying to free pages to the zone, but don't increase PCP high during PCP
refilling. This can avoid the issue above. But if the number of pages
allocated is much less than that of pages freed on a CPU, there will be
many idle pages in PCP and it is hard to free these idle pages.
1/8 (>> 3) of PCP high will be decreased periodically. The value 1/8 is
kind of arbitrary. Just to make sure that the idle PCP pages will be
freed eventually.
On a 2-socket Intel server with 224 logical CPU, we run 8 kbuild instances
in parallel (each with `make -j 28`) in 8 cgroup. This simulates the
kbuild server that is used by 0-Day kbuild service. With the patch, the
build time decreases 3.5%. The cycles% of the spinlock contention (mostly
for zone lock) decreases from 11.0% to 0.5%. The number of PCP draining
for high order pages freeing (free_high) decreases 65.6%. The number of
pages allocated from zone (instead of from PCP) decreases 83.9%.
Link: https://lkml.kernel.org/r/20231016053002.756205-8-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Suggested-by: Mel Gorman <mgorman@techsingularity.net>
Suggested-by: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:00 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-28 19:42:21 -07:00
|
|
|
|
|
|
|
|
|
mm: tune PCP high automatically
The target to tune PCP high automatically is as follows,
- Minimize allocation/freeing from/to shared zone
- Minimize idle pages in PCP
- Minimize pages in PCP if the system free pages is too few
To reach these target, a tuning algorithm as follows is designed,
- When we refill PCP via allocating from the zone, increase PCP high.
Because if we had larger PCP, we could avoid to allocate from the
zone.
- In periodic vmstat updating kworker (via refresh_cpu_vm_stats()),
decrease PCP high to try to free possible idle PCP pages.
- When page reclaiming is active for the zone, stop increasing PCP
high in allocating path, decrease PCP high and free some pages in
freeing path.
So, the PCP high can be tuned to the page allocating/freeing depth of
workloads eventually.
One issue of the algorithm is that if the number of pages allocated is
much more than that of pages freed on a CPU, the PCP high may become the
maximal value even if the allocating/freeing depth is small. But this
isn't a severe issue, because there are no idle pages in this case.
One alternative choice is to increase PCP high when we drain PCP via
trying to free pages to the zone, but don't increase PCP high during PCP
refilling. This can avoid the issue above. But if the number of pages
allocated is much less than that of pages freed on a CPU, there will be
many idle pages in PCP and it is hard to free these idle pages.
1/8 (>> 3) of PCP high will be decreased periodically. The value 1/8 is
kind of arbitrary. Just to make sure that the idle PCP pages will be
freed eventually.
On a 2-socket Intel server with 224 logical CPU, we run 8 kbuild instances
in parallel (each with `make -j 28`) in 8 cgroup. This simulates the
kbuild server that is used by 0-Day kbuild service. With the patch, the
build time decreases 3.5%. The cycles% of the spinlock contention (mostly
for zone lock) decreases from 11.0% to 0.5%. The number of PCP draining
for high order pages freeing (free_high) decreases 65.6%. The number of
pages allocated from zone (instead of from PCP) decreases 83.9%.
Link: https://lkml.kernel.org/r/20231016053002.756205-8-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Suggested-by: Mel Gorman <mgorman@techsingularity.net>
Suggested-by: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:00 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-28 19:42:21 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: tune PCP high automatically
The target to tune PCP high automatically is as follows,
- Minimize allocation/freeing from/to shared zone
- Minimize idle pages in PCP
- Minimize pages in PCP if the system free pages is too few
To reach these target, a tuning algorithm as follows is designed,
- When we refill PCP via allocating from the zone, increase PCP high.
Because if we had larger PCP, we could avoid to allocate from the
zone.
- In periodic vmstat updating kworker (via refresh_cpu_vm_stats()),
decrease PCP high to try to free possible idle PCP pages.
- When page reclaiming is active for the zone, stop increasing PCP
high in allocating path, decrease PCP high and free some pages in
freeing path.
So, the PCP high can be tuned to the page allocating/freeing depth of
workloads eventually.
One issue of the algorithm is that if the number of pages allocated is
much more than that of pages freed on a CPU, the PCP high may become the
maximal value even if the allocating/freeing depth is small. But this
isn't a severe issue, because there are no idle pages in this case.
One alternative choice is to increase PCP high when we drain PCP via
trying to free pages to the zone, but don't increase PCP high during PCP
refilling. This can avoid the issue above. But if the number of pages
allocated is much less than that of pages freed on a CPU, there will be
many idle pages in PCP and it is hard to free these idle pages.
1/8 (>> 3) of PCP high will be decreased periodically. The value 1/8 is
kind of arbitrary. Just to make sure that the idle PCP pages will be
freed eventually.
On a 2-socket Intel server with 224 logical CPU, we run 8 kbuild instances
in parallel (each with `make -j 28`) in 8 cgroup. This simulates the
kbuild server that is used by 0-Day kbuild service. With the patch, the
build time decreases 3.5%. The cycles% of the spinlock contention (mostly
for zone lock) decreases from 11.0% to 0.5%. The number of PCP draining
for high order pages freeing (free_high) decreases 65.6%. The number of
pages allocated from zone (instead of from PCP) decreases 83.9%.
Link: https://lkml.kernel.org/r/20231016053002.756205-8-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Suggested-by: Mel Gorman <mgorman@techsingularity.net>
Suggested-by: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:00 +08:00
|
|
|
|
mm, pcp: reduce detecting time of consecutive high order page freeing
In current PCP auto-tuning design, if the number of pages allocated is
much more than that of pages freed on a CPU, the PCP high may become the
maximal value even if the allocating/freeing depth is small, for example,
in the sender of network workloads. If a CPU was used as sender
originally, then it is used as receiver after context switching, we need
to fill the whole PCP with maximal high before triggering PCP draining for
consecutive high order freeing. This will hurt the performance of some
network workloads.
To solve the issue, in this patch, we will track the consecutive page
freeing with a counter in stead of relying on PCP draining. So, we can
detect consecutive page freeing much earlier.
On a 2-socket Intel server with 128 logical CPU, we tested
SCTP_STREAM_MANY test case of netperf test suite with 64-pair processes.
With the patch, the network bandwidth improves 5.0%. This restores the
performance drop caused by PCP auto-tuning.
Link: https://lkml.kernel.org/r/20231016053002.756205-10-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:02 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: tune PCP high automatically
The target to tune PCP high automatically is as follows,
- Minimize allocation/freeing from/to shared zone
- Minimize idle pages in PCP
- Minimize pages in PCP if the system free pages is too few
To reach these target, a tuning algorithm as follows is designed,
- When we refill PCP via allocating from the zone, increase PCP high.
Because if we had larger PCP, we could avoid to allocate from the
zone.
- In periodic vmstat updating kworker (via refresh_cpu_vm_stats()),
decrease PCP high to try to free possible idle PCP pages.
- When page reclaiming is active for the zone, stop increasing PCP
high in allocating path, decrease PCP high and free some pages in
freeing path.
So, the PCP high can be tuned to the page allocating/freeing depth of
workloads eventually.
One issue of the algorithm is that if the number of pages allocated is
much more than that of pages freed on a CPU, the PCP high may become the
maximal value even if the allocating/freeing depth is small. But this
isn't a severe issue, because there are no idle pages in this case.
One alternative choice is to increase PCP high when we drain PCP via
trying to free pages to the zone, but don't increase PCP high during PCP
refilling. This can avoid the issue above. But if the number of pages
allocated is much less than that of pages freed on a CPU, there will be
many idle pages in PCP and it is hard to free these idle pages.
1/8 (>> 3) of PCP high will be decreased periodically. The value 1/8 is
kind of arbitrary. Just to make sure that the idle PCP pages will be
freed eventually.
On a 2-socket Intel server with 224 logical CPU, we run 8 kbuild instances
in parallel (each with `make -j 28`) in 8 cgroup. This simulates the
kbuild server that is used by 0-Day kbuild service. With the patch, the
build time decreases 3.5%. The cycles% of the spinlock contention (mostly
for zone lock) decreases from 11.0% to 0.5%. The number of PCP draining
for high order pages freeing (free_high) decreases 65.6%. The number of
pages allocated from zone (instead of from PCP) decreases 83.9%.
Link: https://lkml.kernel.org/r/20231016053002.756205-8-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Suggested-by: Mel Gorman <mgorman@techsingularity.net>
Suggested-by: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:00 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, pcp: decrease PCP high if free pages < high watermark
One target of PCP is to minimize pages in PCP if the system free pages is
too few. To reach that target, when page reclaiming is active for the
zone (ZONE_RECLAIM_ACTIVE), we will stop increasing PCP high in allocating
path, decrease PCP high and free some pages in freeing path. But this may
be too late because the background page reclaiming may introduce latency
for some workloads. So, in this patch, during page allocation we will
detect whether the number of free pages of the zone is below high
watermark. If so, we will stop increasing PCP high in allocating path,
decrease PCP high and free some pages in freeing path. With this, we can
reduce the possibility of the premature background page reclaiming caused
by too large PCP.
The high watermark checking is done in allocating path to reduce the
overhead in hotter freeing path.
Link: https://lkml.kernel.org/r/20231016053002.756205-9-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:01 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, pcp: reduce detecting time of consecutive high order page freeing
In current PCP auto-tuning design, if the number of pages allocated is
much more than that of pages freed on a CPU, the PCP high may become the
maximal value even if the allocating/freeing depth is small, for example,
in the sender of network workloads. If a CPU was used as sender
originally, then it is used as receiver after context switching, we need
to fill the whole PCP with maximal high before triggering PCP draining for
consecutive high order freeing. This will hurt the performance of some
network workloads.
To solve the issue, in this patch, we will track the consecutive page
freeing with a counter in stead of relying on PCP draining. So, we can
detect consecutive page freeing much earlier.
On a 2-socket Intel server with 128 logical CPU, we tested
SCTP_STREAM_MANY test case of netperf test suite with 64-pair processes.
With the patch, the network bandwidth improves 5.0%. This restores the
performance drop caused by PCP auto-tuning.
Link: https://lkml.kernel.org/r/20231016053002.756205-10-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:02 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, pcp: decrease PCP high if free pages < high watermark
One target of PCP is to minimize pages in PCP if the system free pages is
too few. To reach that target, when page reclaiming is active for the
zone (ZONE_RECLAIM_ACTIVE), we will stop increasing PCP high in allocating
path, decrease PCP high and free some pages in freeing path. But this may
be too late because the background page reclaiming may introduce latency
for some workloads. So, in this patch, during page allocation we will
detect whether the number of free pages of the zone is below high
watermark. If so, we will stop increasing PCP high in allocating path,
decrease PCP high and free some pages in freeing path. With this, we can
reduce the possibility of the premature background page reclaiming caused
by too large PCP.
The high watermark checking is done in allocating path to reduce the
overhead in hotter freeing path.
Link: https://lkml.kernel.org/r/20231016053002.756205-9-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:01 +08:00
|
|
|
|
|
|
|
|
|
mm, pcp: reduce detecting time of consecutive high order page freeing
In current PCP auto-tuning design, if the number of pages allocated is
much more than that of pages freed on a CPU, the PCP high may become the
maximal value even if the allocating/freeing depth is small, for example,
in the sender of network workloads. If a CPU was used as sender
originally, then it is used as receiver after context switching, we need
to fill the whole PCP with maximal high before triggering PCP draining for
consecutive high order freeing. This will hurt the performance of some
network workloads.
To solve the issue, in this patch, we will track the consecutive page
freeing with a counter in stead of relying on PCP draining. So, we can
detect consecutive page freeing much earlier.
On a 2-socket Intel server with 128 logical CPU, we tested
SCTP_STREAM_MANY test case of netperf test suite with 64-pair processes.
With the patch, the network bandwidth improves 5.0%. This restores the
performance drop caused by PCP auto-tuning.
Link: https://lkml.kernel.org/r/20231016053002.756205-10-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:02 +08:00
|
|
|
|
mm: tune PCP high automatically
The target to tune PCP high automatically is as follows,
- Minimize allocation/freeing from/to shared zone
- Minimize idle pages in PCP
- Minimize pages in PCP if the system free pages is too few
To reach these target, a tuning algorithm as follows is designed,
- When we refill PCP via allocating from the zone, increase PCP high.
Because if we had larger PCP, we could avoid to allocate from the
zone.
- In periodic vmstat updating kworker (via refresh_cpu_vm_stats()),
decrease PCP high to try to free possible idle PCP pages.
- When page reclaiming is active for the zone, stop increasing PCP
high in allocating path, decrease PCP high and free some pages in
freeing path.
So, the PCP high can be tuned to the page allocating/freeing depth of
workloads eventually.
One issue of the algorithm is that if the number of pages allocated is
much more than that of pages freed on a CPU, the PCP high may become the
maximal value even if the allocating/freeing depth is small. But this
isn't a severe issue, because there are no idle pages in this case.
One alternative choice is to increase PCP high when we drain PCP via
trying to free pages to the zone, but don't increase PCP high during PCP
refilling. This can avoid the issue above. But if the number of pages
allocated is much less than that of pages freed on a CPU, there will be
many idle pages in PCP and it is hard to free these idle pages.
1/8 (>> 3) of PCP high will be decreased periodically. The value 1/8 is
kind of arbitrary. Just to make sure that the idle PCP pages will be
freed eventually.
On a 2-socket Intel server with 224 logical CPU, we run 8 kbuild instances
in parallel (each with `make -j 28`) in 8 cgroup. This simulates the
kbuild server that is used by 0-Day kbuild service. With the patch, the
build time decreases 3.5%. The cycles% of the spinlock contention (mostly
for zone lock) decreases from 11.0% to 0.5%. The number of PCP draining
for high order pages freeing (free_high) decreases 65.6%. The number of
pages allocated from zone (instead of from PCP) decreases 83.9%.
Link: https://lkml.kernel.org/r/20231016053002.756205-8-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Suggested-by: Mel Gorman <mgorman@techsingularity.net>
Suggested-by: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:00 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-28 19:42:21 -07:00
|
|
|
|
|
|
|
|
|
2024-11-25 21:01:35 +00:00
|
|
|
|
|
|
|
|
|
2025-02-21 18:44:24 -08:00
|
|
|
|
mm, page_alloc: enable/disable IRQs once when freeing a list of pages
Patch series "Follow-up for speed up page cache truncation", v2.
This series is a follow-on for Jan Kara's series "Speed up page cache
truncation" series. We both ended up looking at the same problem but
saw different problems based on the same data. This series builds upon
his work.
A variety of workloads were compared on four separate machines but each
machine showed gains albeit at different levels. Minimally, some of the
differences are due to NUMA where truncating data from a remote node is
slower than a local node. The workloads checked were
o sparse truncate microbenchmark, tiny
o sparse truncate microbenchmark, large
o reaim-io disk workfile
o dbench4 (modified by mmtests to produce more stable results)
o filebench varmail configuration for small memory size
o bonnie, directory operations, working set size 2*RAM
reaim-io, dbench and filebench all showed minor gains. Truncation does
not dominate those workloads but were tested to ensure no other
regressions. They will not be reported further.
The sparse truncate microbench was written by Jan. It creates a number
of files and then times how long it takes to truncate each one. The
"tiny" configuraiton creates a number of files that easily fits in
memory and times how long it takes to truncate files with page cache.
The large configuration uses enough files to have data that is twice the
size of memory and so timings there include truncating page cache and
working set shadow entries in the radix tree.
Patches 1-4 are the most relevant parts of this series. Patches 5-8 are
optional as they are deleting code that is essentially useless but has a
negligible performance impact.
The changelogs have more information on performance but just for bonnie
delete options, the main comparison is
bonnie
4.14.0-rc5 4.14.0-rc5 4.14.0-rc5
jan-v2 vanilla mel-v2
Hmean SeqCreate ops 76.20 ( 0.00%) 75.80 ( -0.53%) 76.80 ( 0.79%)
Hmean SeqCreate read 85.00 ( 0.00%) 85.00 ( 0.00%) 85.00 ( 0.00%)
Hmean SeqCreate del 13752.31 ( 0.00%) 12090.23 ( -12.09%) 15304.84 ( 11.29%)
Hmean RandCreate ops 76.00 ( 0.00%) 75.60 ( -0.53%) 77.00 ( 1.32%)
Hmean RandCreate read 96.80 ( 0.00%) 96.80 ( 0.00%) 97.00 ( 0.21%)
Hmean RandCreate del 13233.75 ( 0.00%) 11525.35 ( -12.91%) 14446.61 ( 9.16%)
Jan's series is the baseline and the vanilla kernel is 12% slower where
as this series on top gains another 11%. This is from a different
machine than the data in the changelogs but the detailed data was not
collected as there was no substantial change in v2.
This patch (of 8):
Freeing a list of pages current enables/disables IRQs for each page
freed. This patch splits freeing a list of pages into two operations --
preparing the pages for freeing and the actual freeing. This is a
tradeoff - we're taking two passes of the list to free in exchange for
avoiding multiple enable/disable of IRQs.
sparsetruncate (tiny)
4.14.0-rc4 4.14.0-rc4
janbatch-v1r1 oneirq-v1r1
Min Time 149.00 ( 0.00%) 141.00 ( 5.37%)
1st-qrtle Time 150.00 ( 0.00%) 142.00 ( 5.33%)
2nd-qrtle Time 151.00 ( 0.00%) 142.00 ( 5.96%)
3rd-qrtle Time 151.00 ( 0.00%) 143.00 ( 5.30%)
Max-90% Time 153.00 ( 0.00%) 144.00 ( 5.88%)
Max-95% Time 155.00 ( 0.00%) 147.00 ( 5.16%)
Max-99% Time 201.00 ( 0.00%) 195.00 ( 2.99%)
Max Time 236.00 ( 0.00%) 230.00 ( 2.54%)
Amean Time 152.65 ( 0.00%) 144.37 ( 5.43%)
Stddev Time 9.78 ( 0.00%) 10.44 ( -6.72%)
Coeff Time 6.41 ( 0.00%) 7.23 ( -12.84%)
Best99%Amean Time 152.07 ( 0.00%) 143.72 ( 5.50%)
Best95%Amean Time 150.75 ( 0.00%) 142.37 ( 5.56%)
Best90%Amean Time 150.59 ( 0.00%) 142.19 ( 5.58%)
Best75%Amean Time 150.36 ( 0.00%) 141.92 ( 5.61%)
Best50%Amean Time 150.04 ( 0.00%) 141.69 ( 5.56%)
Best25%Amean Time 149.85 ( 0.00%) 141.38 ( 5.65%)
With a tiny number of files, each file truncated has resident page cache
and it shows that time to truncate is roughtly 5-6% with some minor
jitter.
4.14.0-rc4 4.14.0-rc4
janbatch-v1r1 oneirq-v1r1
Hmean SeqCreate ops 65.27 ( 0.00%) 81.86 ( 25.43%)
Hmean SeqCreate read 39.48 ( 0.00%) 47.44 ( 20.16%)
Hmean SeqCreate del 24963.95 ( 0.00%) 26319.99 ( 5.43%)
Hmean RandCreate ops 65.47 ( 0.00%) 82.01 ( 25.26%)
Hmean RandCreate read 42.04 ( 0.00%) 51.75 ( 23.09%)
Hmean RandCreate del 23377.66 ( 0.00%) 23764.79 ( 1.66%)
As expected, there is a small gain for the delete operation.
[mgorman@techsingularity.net: use page_private and set_page_private helpers]
Link: http://lkml.kernel.org/r/20171018101547.mjycw7zreb66jzpa@techsingularity.net
Link: http://lkml.kernel.org/r/20171018075952.10627-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Kara <jack@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-15 17:37:37 -08:00
|
|
|
|
mm: tune PCP high automatically
The target to tune PCP high automatically is as follows,
- Minimize allocation/freeing from/to shared zone
- Minimize idle pages in PCP
- Minimize pages in PCP if the system free pages is too few
To reach these target, a tuning algorithm as follows is designed,
- When we refill PCP via allocating from the zone, increase PCP high.
Because if we had larger PCP, we could avoid to allocate from the
zone.
- In periodic vmstat updating kworker (via refresh_cpu_vm_stats()),
decrease PCP high to try to free possible idle PCP pages.
- When page reclaiming is active for the zone, stop increasing PCP
high in allocating path, decrease PCP high and free some pages in
freeing path.
So, the PCP high can be tuned to the page allocating/freeing depth of
workloads eventually.
One issue of the algorithm is that if the number of pages allocated is
much more than that of pages freed on a CPU, the PCP high may become the
maximal value even if the allocating/freeing depth is small. But this
isn't a severe issue, because there are no idle pages in this case.
One alternative choice is to increase PCP high when we drain PCP via
trying to free pages to the zone, but don't increase PCP high during PCP
refilling. This can avoid the issue above. But if the number of pages
allocated is much less than that of pages freed on a CPU, there will be
many idle pages in PCP and it is hard to free these idle pages.
1/8 (>> 3) of PCP high will be decreased periodically. The value 1/8 is
kind of arbitrary. Just to make sure that the idle PCP pages will be
freed eventually.
On a 2-socket Intel server with 224 logical CPU, we run 8 kbuild instances
in parallel (each with `make -j 28`) in 8 cgroup. This simulates the
kbuild server that is used by 0-Day kbuild service. With the patch, the
build time decreases 3.5%. The cycles% of the spinlock contention (mostly
for zone lock) decreases from 11.0% to 0.5%. The number of PCP draining
for high order pages freeing (free_high) decreases 65.6%. The number of
pages allocated from zone (instead of from PCP) decreases 83.9%.
Link: https://lkml.kernel.org/r/20231016053002.756205-8-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Suggested-by: Mel Gorman <mgorman@techsingularity.net>
Suggested-by: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:00 +08:00
|
|
|
|
2021-06-28 19:43:08 -07:00
|
|
|
|
mm, pcp: avoid to drain PCP when process exit
Patch series "mm: PCP high auto-tuning", v3.
The page allocation performance requirements of different workloads are
often different. So, we need to tune the PCP (Per-CPU Pageset) high on
each CPU automatically to optimize the page allocation performance.
The list of patches in series is as follows,
[1/9] mm, pcp: avoid to drain PCP when process exit
[2/9] cacheinfo: calculate per-CPU data cache size
[3/9] mm, pcp: reduce lock contention for draining high-order pages
[4/9] mm: restrict the pcp batch scale factor to avoid too long latency
[5/9] mm, page_alloc: scale the number of pages that are batch allocated
[6/9] mm: add framework for PCP high auto-tuning
[7/9] mm: tune PCP high automatically
[8/9] mm, pcp: decrease PCP high if free pages < high watermark
[9/9] mm, pcp: reduce detecting time of consecutive high order page freeing
Patch [1/9], [2/9], [3/9] optimize the PCP draining for consecutive
high-order pages freeing.
Patch [4/9], [5/9] optimize batch freeing and allocating.
Patch [6/9], [7/9], [8/9] implement and optimize a PCP high
auto-tuning method.
Patch [9/9] optimize the PCP draining for consecutive high order page
freeing based on PCP high auto-tuning.
The test results for patches with performance impact are as follows,
kbuild
======
On a 2-socket Intel server with 224 logical CPU, we run 8 kbuild instances
in parallel (each with `make -j 28`) in 8 cgroup. This simulates the
kbuild server that is used by 0-Day kbuild service.
build time lock contend% free_high alloc_zone
---------- ---------- --------- ----------
base 100.0 14.0 100.0 100.0
patch1 99.5 12.8 19.5 95.6
patch3 99.4 12.6 7.1 95.6
patch5 98.6 11.0 8.1 97.1
patch7 95.1 0.5 2.8 15.6
patch9 95.0 1.0 8.8 20.0
The PCP draining optimization (patch [1/9], [3/9]) and PCP batch
allocation optimization (patch [5/9]) reduces zone lock contention a
little. The PCP high auto-tuning (patch [7/9], [9/9]) reduces build time
visibly. Where the tuning target: the number of pages allocated from zone
reduces greatly. So, the zone contention cycles% reduces greatly.
With PCP tuning patches (patch [7/9], [9/9]), the average used memory
during test increases up to 18.4% because more pages are cached in PCP.
But at the end of the test, the number of the used memory decreases to the
same level as that of the base patch. That is, the pages cached in PCP
will be released to zone after not being used actively.
netperf SCTP_STREAM_MANY
========================
On a 2-socket Intel server with 128 logical CPU, we tested
SCTP_STREAM_MANY test case of netperf test suite with 64-pair processes.
score lock contend% free_high alloc_zone cache miss rate%
----- ---------- --------- ---------- ----------------
base 100.0 2.1 100.0 100.0 1.3
patch1 99.4 2.1 99.4 99.4 1.3
patch3 106.4 1.3 13.3 106.3 1.3
patch5 106.0 1.2 13.2 105.9 1.3
patch7 103.4 1.9 6.7 90.3 7.6
patch9 108.6 1.3 13.7 108.6 1.3
The PCP draining optimization (patch [1/9]+[3/9]) improves performance.
The PCP high auto-tuning (patch [7/9]) reduces performance a little
because PCP draining cannot be triggered in time sometimes. So, the cache
miss rate% increases. The further PCP draining optimization (patch [9/9])
based on PCP tuning restore the performance.
lmbench3 UNIX (AF_UNIX)
=======================
On a 2-socket Intel server with 128 logical CPU, we tested UNIX
(AF_UNIX socket) test case of lmbench3 test suite with 16-pair
processes.
score lock contend% free_high alloc_zone cache miss rate%
----- ---------- --------- ---------- ----------------
base 100.0 51.4 100.0 100.0 0.2
patch1 116.8 46.1 69.5 104.3 0.2
patch3 199.1 21.3 7.0 104.9 0.2
patch5 200.0 20.8 7.1 106.9 0.3
patch7 191.6 19.9 6.8 103.8 2.8
patch9 193.4 21.7 7.0 104.7 2.1
The PCP draining optimization (patch [1/9], [3/9]) improves performance
much. The PCP tuning (patch [7/9]) reduces performance a little because
PCP draining cannot be triggered in time sometimes. The further PCP
draining optimization (patch [9/9]) based on PCP tuning restores the
performance partly.
The patchset adds several fields in struct per_cpu_pages. The struct
layout before/after the patchset is as follows,
base
====
struct per_cpu_pages {
spinlock_t lock; /* 0 4 */
int count; /* 4 4 */
int high; /* 8 4 */
int batch; /* 12 4 */
short int free_factor; /* 16 2 */
short int expire; /* 18 2 */
/* XXX 4 bytes hole, try to pack */
struct list_head lists[13]; /* 24 208 */
/* size: 256, cachelines: 4, members: 7 */
/* sum members: 228, holes: 1, sum holes: 4 */
/* padding: 24 */
} __attribute__((__aligned__(64)));
patched
=======
struct per_cpu_pages {
spinlock_t lock; /* 0 4 */
int count; /* 4 4 */
int high; /* 8 4 */
int high_min; /* 12 4 */
int high_max; /* 16 4 */
int batch; /* 20 4 */
u8 flags; /* 24 1 */
u8 alloc_factor; /* 25 1 */
u8 expire; /* 26 1 */
/* XXX 1 byte hole, try to pack */
short int free_count; /* 28 2 */
/* XXX 2 bytes hole, try to pack */
struct list_head lists[13]; /* 32 208 */
/* size: 256, cachelines: 4, members: 11 */
/* sum members: 237, holes: 2, sum holes: 3 */
/* padding: 16 */
} __attribute__((__aligned__(64)));
The size of the struct doesn't changed with the patchset.
This patch (of 9):
In commit f26b3fa04611 ("mm/page_alloc: limit number of high-order pages
on PCP during bulk free"), the PCP (Per-CPU Pageset) will be drained when
PCP is mostly used for high-order pages freeing to improve the cache-hot
pages reusing between page allocation and freeing CPUs.
But, the PCP draining mechanism may be triggered unexpectedly when process
exits. With some customized trace point, it was found that PCP draining
(free_high == true) was triggered with the order-1 page freeing with the
following call stack,
=> free_unref_page_commit
=> free_unref_page
=> __mmdrop
=> exit_mm
=> do_exit
=> do_group_exit
=> __x64_sys_exit_group
=> do_syscall_64
Checking the source code, this is the page table PGD freeing
(mm_free_pgd()). It's a order-1 page freeing if
CONFIG_PAGE_TABLE_ISOLATION=y. Which is a common configuration for
security.
Just before that, page freeing with the following call stack was found,
=> free_unref_page_commit
=> free_unref_page_list
=> release_pages
=> tlb_batch_pages_flush
=> tlb_finish_mmu
=> exit_mmap
=> __mmput
=> exit_mm
=> do_exit
=> do_group_exit
=> __x64_sys_exit_group
=> do_syscall_64
So, when a process exits,
- a large number of user pages of the process will be freed without
page allocation, it's highly possible that pcp->free_factor becomes >
0. In fact, this is expected behavior to improve process exit
performance.
- after freeing all user pages, the PGD will be freed, which is a
order-1 page freeing, PCP will be drained.
All in all, when a process exits, it's high possible that the PCP will be
drained. This is an unexpected behavior.
To avoid this, in the patch, the PCP draining will only be triggered for 2
consecutive high-order page freeing.
On a 2-socket Intel server with 224 logical CPU, we run 8 kbuild instances
in parallel (each with `make -j 28`) in 8 cgroup. This simulates the
kbuild server that is used by 0-Day kbuild service. With the patch, the
cycles% of the spinlock contention (mostly for zone lock) decreases from
14.0% to 12.8% (with PCP size == 367). The number of PCP draining for
high order pages freeing (free_high) decreases 80.5%.
This helps network workload too for reduced zone lock contention. On a
2-socket Intel server with 128 logical CPU, with the patch, the network
bandwidth of the UNIX (AF_UNIX) test case of lmbench test suite with
16-pair processes increase 16.8%. The cycles% of the spinlock contention
(mostly for zone lock) decreases from 51.4% to 46.1%. The number of PCP
draining for high order pages freeing (free_high) decreases 30.5%. The
cache miss rate keeps 0.2%.
Link: https://lkml.kernel.org/r/20231016053002.756205-1-ying.huang@intel.com
Link: https://lkml.kernel.org/r/20231016053002.756205-2-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:29:54 +08:00
|
|
|
|
mm, page_alloc: enable/disable IRQs once when freeing a list of pages
Patch series "Follow-up for speed up page cache truncation", v2.
This series is a follow-on for Jan Kara's series "Speed up page cache
truncation" series. We both ended up looking at the same problem but
saw different problems based on the same data. This series builds upon
his work.
A variety of workloads were compared on four separate machines but each
machine showed gains albeit at different levels. Minimally, some of the
differences are due to NUMA where truncating data from a remote node is
slower than a local node. The workloads checked were
o sparse truncate microbenchmark, tiny
o sparse truncate microbenchmark, large
o reaim-io disk workfile
o dbench4 (modified by mmtests to produce more stable results)
o filebench varmail configuration for small memory size
o bonnie, directory operations, working set size 2*RAM
reaim-io, dbench and filebench all showed minor gains. Truncation does
not dominate those workloads but were tested to ensure no other
regressions. They will not be reported further.
The sparse truncate microbench was written by Jan. It creates a number
of files and then times how long it takes to truncate each one. The
"tiny" configuraiton creates a number of files that easily fits in
memory and times how long it takes to truncate files with page cache.
The large configuration uses enough files to have data that is twice the
size of memory and so timings there include truncating page cache and
working set shadow entries in the radix tree.
Patches 1-4 are the most relevant parts of this series. Patches 5-8 are
optional as they are deleting code that is essentially useless but has a
negligible performance impact.
The changelogs have more information on performance but just for bonnie
delete options, the main comparison is
bonnie
4.14.0-rc5 4.14.0-rc5 4.14.0-rc5
jan-v2 vanilla mel-v2
Hmean SeqCreate ops 76.20 ( 0.00%) 75.80 ( -0.53%) 76.80 ( 0.79%)
Hmean SeqCreate read 85.00 ( 0.00%) 85.00 ( 0.00%) 85.00 ( 0.00%)
Hmean SeqCreate del 13752.31 ( 0.00%) 12090.23 ( -12.09%) 15304.84 ( 11.29%)
Hmean RandCreate ops 76.00 ( 0.00%) 75.60 ( -0.53%) 77.00 ( 1.32%)
Hmean RandCreate read 96.80 ( 0.00%) 96.80 ( 0.00%) 97.00 ( 0.21%)
Hmean RandCreate del 13233.75 ( 0.00%) 11525.35 ( -12.91%) 14446.61 ( 9.16%)
Jan's series is the baseline and the vanilla kernel is 12% slower where
as this series on top gains another 11%. This is from a different
machine than the data in the changelogs but the detailed data was not
collected as there was no substantial change in v2.
This patch (of 8):
Freeing a list of pages current enables/disables IRQs for each page
freed. This patch splits freeing a list of pages into two operations --
preparing the pages for freeing and the actual freeing. This is a
tradeoff - we're taking two passes of the list to free in exchange for
avoiding multiple enable/disable of IRQs.
sparsetruncate (tiny)
4.14.0-rc4 4.14.0-rc4
janbatch-v1r1 oneirq-v1r1
Min Time 149.00 ( 0.00%) 141.00 ( 5.37%)
1st-qrtle Time 150.00 ( 0.00%) 142.00 ( 5.33%)
2nd-qrtle Time 151.00 ( 0.00%) 142.00 ( 5.96%)
3rd-qrtle Time 151.00 ( 0.00%) 143.00 ( 5.30%)
Max-90% Time 153.00 ( 0.00%) 144.00 ( 5.88%)
Max-95% Time 155.00 ( 0.00%) 147.00 ( 5.16%)
Max-99% Time 201.00 ( 0.00%) 195.00 ( 2.99%)
Max Time 236.00 ( 0.00%) 230.00 ( 2.54%)
Amean Time 152.65 ( 0.00%) 144.37 ( 5.43%)
Stddev Time 9.78 ( 0.00%) 10.44 ( -6.72%)
Coeff Time 6.41 ( 0.00%) 7.23 ( -12.84%)
Best99%Amean Time 152.07 ( 0.00%) 143.72 ( 5.50%)
Best95%Amean Time 150.75 ( 0.00%) 142.37 ( 5.56%)
Best90%Amean Time 150.59 ( 0.00%) 142.19 ( 5.58%)
Best75%Amean Time 150.36 ( 0.00%) 141.92 ( 5.61%)
Best50%Amean Time 150.04 ( 0.00%) 141.69 ( 5.56%)
Best25%Amean Time 149.85 ( 0.00%) 141.38 ( 5.65%)
With a tiny number of files, each file truncated has resident page cache
and it shows that time to truncate is roughtly 5-6% with some minor
jitter.
4.14.0-rc4 4.14.0-rc4
janbatch-v1r1 oneirq-v1r1
Hmean SeqCreate ops 65.27 ( 0.00%) 81.86 ( 25.43%)
Hmean SeqCreate read 39.48 ( 0.00%) 47.44 ( 20.16%)
Hmean SeqCreate del 24963.95 ( 0.00%) 26319.99 ( 5.43%)
Hmean RandCreate ops 65.47 ( 0.00%) 82.01 ( 25.26%)
Hmean RandCreate read 42.04 ( 0.00%) 51.75 ( 23.09%)
Hmean RandCreate del 23377.66 ( 0.00%) 23764.79 ( 1.66%)
As expected, there is a small gain for the delete operation.
[mgorman@techsingularity.net: use page_private and set_page_private helpers]
Link: http://lkml.kernel.org/r/20171018101547.mjycw7zreb66jzpa@techsingularity.net
Link: http://lkml.kernel.org/r/20171018075952.10627-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Kara <jack@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-15 17:37:37 -08:00
|
|
|
|
2023-10-16 13:29:58 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-10-06 10:15:40 +00:00
|
|
|
|
2021-06-28 19:43:08 -07:00
|
|
|
|
2022-06-24 13:54:17 +01:00
|
|
|
|
2021-06-28 19:43:08 -07:00
|
|
|
|
2022-03-22 14:43:45 -07:00
|
|
|
|
mm: tune PCP high automatically
The target to tune PCP high automatically is as follows,
- Minimize allocation/freeing from/to shared zone
- Minimize idle pages in PCP
- Minimize pages in PCP if the system free pages is too few
To reach these target, a tuning algorithm as follows is designed,
- When we refill PCP via allocating from the zone, increase PCP high.
Because if we had larger PCP, we could avoid to allocate from the
zone.
- In periodic vmstat updating kworker (via refresh_cpu_vm_stats()),
decrease PCP high to try to free possible idle PCP pages.
- When page reclaiming is active for the zone, stop increasing PCP
high in allocating path, decrease PCP high and free some pages in
freeing path.
So, the PCP high can be tuned to the page allocating/freeing depth of
workloads eventually.
One issue of the algorithm is that if the number of pages allocated is
much more than that of pages freed on a CPU, the PCP high may become the
maximal value even if the allocating/freeing depth is small. But this
isn't a severe issue, because there are no idle pages in this case.
One alternative choice is to increase PCP high when we drain PCP via
trying to free pages to the zone, but don't increase PCP high during PCP
refilling. This can avoid the issue above. But if the number of pages
allocated is much less than that of pages freed on a CPU, there will be
many idle pages in PCP and it is hard to free these idle pages.
1/8 (>> 3) of PCP high will be decreased periodically. The value 1/8 is
kind of arbitrary. Just to make sure that the idle PCP pages will be
freed eventually.
On a 2-socket Intel server with 224 logical CPU, we run 8 kbuild instances
in parallel (each with `make -j 28`) in 8 cgroup. This simulates the
kbuild server that is used by 0-Day kbuild service. With the patch, the
build time decreases 3.5%. The cycles% of the spinlock contention (mostly
for zone lock) decreases from 11.0% to 0.5%. The number of PCP draining
for high order pages freeing (free_high) decreases 65.6%. The number of
pages allocated from zone (instead of from PCP) decreases 83.9%.
Link: https://lkml.kernel.org/r/20231016053002.756205-8-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Suggested-by: Mel Gorman <mgorman@techsingularity.net>
Suggested-by: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:00 +08:00
|
|
|
|
2022-03-22 14:43:45 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, pcp: avoid to drain PCP when process exit
Patch series "mm: PCP high auto-tuning", v3.
The page allocation performance requirements of different workloads are
often different. So, we need to tune the PCP (Per-CPU Pageset) high on
each CPU automatically to optimize the page allocation performance.
The list of patches in series is as follows,
[1/9] mm, pcp: avoid to drain PCP when process exit
[2/9] cacheinfo: calculate per-CPU data cache size
[3/9] mm, pcp: reduce lock contention for draining high-order pages
[4/9] mm: restrict the pcp batch scale factor to avoid too long latency
[5/9] mm, page_alloc: scale the number of pages that are batch allocated
[6/9] mm: add framework for PCP high auto-tuning
[7/9] mm: tune PCP high automatically
[8/9] mm, pcp: decrease PCP high if free pages < high watermark
[9/9] mm, pcp: reduce detecting time of consecutive high order page freeing
Patch [1/9], [2/9], [3/9] optimize the PCP draining for consecutive
high-order pages freeing.
Patch [4/9], [5/9] optimize batch freeing and allocating.
Patch [6/9], [7/9], [8/9] implement and optimize a PCP high
auto-tuning method.
Patch [9/9] optimize the PCP draining for consecutive high order page
freeing based on PCP high auto-tuning.
The test results for patches with performance impact are as follows,
kbuild
======
On a 2-socket Intel server with 224 logical CPU, we run 8 kbuild instances
in parallel (each with `make -j 28`) in 8 cgroup. This simulates the
kbuild server that is used by 0-Day kbuild service.
build time lock contend% free_high alloc_zone
---------- ---------- --------- ----------
base 100.0 14.0 100.0 100.0
patch1 99.5 12.8 19.5 95.6
patch3 99.4 12.6 7.1 95.6
patch5 98.6 11.0 8.1 97.1
patch7 95.1 0.5 2.8 15.6
patch9 95.0 1.0 8.8 20.0
The PCP draining optimization (patch [1/9], [3/9]) and PCP batch
allocation optimization (patch [5/9]) reduces zone lock contention a
little. The PCP high auto-tuning (patch [7/9], [9/9]) reduces build time
visibly. Where the tuning target: the number of pages allocated from zone
reduces greatly. So, the zone contention cycles% reduces greatly.
With PCP tuning patches (patch [7/9], [9/9]), the average used memory
during test increases up to 18.4% because more pages are cached in PCP.
But at the end of the test, the number of the used memory decreases to the
same level as that of the base patch. That is, the pages cached in PCP
will be released to zone after not being used actively.
netperf SCTP_STREAM_MANY
========================
On a 2-socket Intel server with 128 logical CPU, we tested
SCTP_STREAM_MANY test case of netperf test suite with 64-pair processes.
score lock contend% free_high alloc_zone cache miss rate%
----- ---------- --------- ---------- ----------------
base 100.0 2.1 100.0 100.0 1.3
patch1 99.4 2.1 99.4 99.4 1.3
patch3 106.4 1.3 13.3 106.3 1.3
patch5 106.0 1.2 13.2 105.9 1.3
patch7 103.4 1.9 6.7 90.3 7.6
patch9 108.6 1.3 13.7 108.6 1.3
The PCP draining optimization (patch [1/9]+[3/9]) improves performance.
The PCP high auto-tuning (patch [7/9]) reduces performance a little
because PCP draining cannot be triggered in time sometimes. So, the cache
miss rate% increases. The further PCP draining optimization (patch [9/9])
based on PCP tuning restore the performance.
lmbench3 UNIX (AF_UNIX)
=======================
On a 2-socket Intel server with 128 logical CPU, we tested UNIX
(AF_UNIX socket) test case of lmbench3 test suite with 16-pair
processes.
score lock contend% free_high alloc_zone cache miss rate%
----- ---------- --------- ---------- ----------------
base 100.0 51.4 100.0 100.0 0.2
patch1 116.8 46.1 69.5 104.3 0.2
patch3 199.1 21.3 7.0 104.9 0.2
patch5 200.0 20.8 7.1 106.9 0.3
patch7 191.6 19.9 6.8 103.8 2.8
patch9 193.4 21.7 7.0 104.7 2.1
The PCP draining optimization (patch [1/9], [3/9]) improves performance
much. The PCP tuning (patch [7/9]) reduces performance a little because
PCP draining cannot be triggered in time sometimes. The further PCP
draining optimization (patch [9/9]) based on PCP tuning restores the
performance partly.
The patchset adds several fields in struct per_cpu_pages. The struct
layout before/after the patchset is as follows,
base
====
struct per_cpu_pages {
spinlock_t lock; /* 0 4 */
int count; /* 4 4 */
int high; /* 8 4 */
int batch; /* 12 4 */
short int free_factor; /* 16 2 */
short int expire; /* 18 2 */
/* XXX 4 bytes hole, try to pack */
struct list_head lists[13]; /* 24 208 */
/* size: 256, cachelines: 4, members: 7 */
/* sum members: 228, holes: 1, sum holes: 4 */
/* padding: 24 */
} __attribute__((__aligned__(64)));
patched
=======
struct per_cpu_pages {
spinlock_t lock; /* 0 4 */
int count; /* 4 4 */
int high; /* 8 4 */
int high_min; /* 12 4 */
int high_max; /* 16 4 */
int batch; /* 20 4 */
u8 flags; /* 24 1 */
u8 alloc_factor; /* 25 1 */
u8 expire; /* 26 1 */
/* XXX 1 byte hole, try to pack */
short int free_count; /* 28 2 */
/* XXX 2 bytes hole, try to pack */
struct list_head lists[13]; /* 32 208 */
/* size: 256, cachelines: 4, members: 11 */
/* sum members: 237, holes: 2, sum holes: 3 */
/* padding: 16 */
} __attribute__((__aligned__(64)));
The size of the struct doesn't changed with the patchset.
This patch (of 9):
In commit f26b3fa04611 ("mm/page_alloc: limit number of high-order pages
on PCP during bulk free"), the PCP (Per-CPU Pageset) will be drained when
PCP is mostly used for high-order pages freeing to improve the cache-hot
pages reusing between page allocation and freeing CPUs.
But, the PCP draining mechanism may be triggered unexpectedly when process
exits. With some customized trace point, it was found that PCP draining
(free_high == true) was triggered with the order-1 page freeing with the
following call stack,
=> free_unref_page_commit
=> free_unref_page
=> __mmdrop
=> exit_mm
=> do_exit
=> do_group_exit
=> __x64_sys_exit_group
=> do_syscall_64
Checking the source code, this is the page table PGD freeing
(mm_free_pgd()). It's a order-1 page freeing if
CONFIG_PAGE_TABLE_ISOLATION=y. Which is a common configuration for
security.
Just before that, page freeing with the following call stack was found,
=> free_unref_page_commit
=> free_unref_page_list
=> release_pages
=> tlb_batch_pages_flush
=> tlb_finish_mmu
=> exit_mmap
=> __mmput
=> exit_mm
=> do_exit
=> do_group_exit
=> __x64_sys_exit_group
=> do_syscall_64
So, when a process exits,
- a large number of user pages of the process will be freed without
page allocation, it's highly possible that pcp->free_factor becomes >
0. In fact, this is expected behavior to improve process exit
performance.
- after freeing all user pages, the PGD will be freed, which is a
order-1 page freeing, PCP will be drained.
All in all, when a process exits, it's high possible that the PCP will be
drained. This is an unexpected behavior.
To avoid this, in the patch, the PCP draining will only be triggered for 2
consecutive high-order page freeing.
On a 2-socket Intel server with 224 logical CPU, we run 8 kbuild instances
in parallel (each with `make -j 28`) in 8 cgroup. This simulates the
kbuild server that is used by 0-Day kbuild service. With the patch, the
cycles% of the spinlock contention (mostly for zone lock) decreases from
14.0% to 12.8% (with PCP size == 367). The number of PCP draining for
high order pages freeing (free_high) decreases 80.5%.
This helps network workload too for reduced zone lock contention. On a
2-socket Intel server with 128 logical CPU, with the patch, the network
bandwidth of the UNIX (AF_UNIX) test case of lmbench test suite with
16-pair processes increase 16.8%. The cycles% of the spinlock contention
(mostly for zone lock) decreases from 51.4% to 46.1%. The number of PCP
draining for high order pages freeing (free_high) decreases 30.5%. The
cache miss rate keeps 0.2%.
Link: https://lkml.kernel.org/r/20231016053002.756205-1-ying.huang@intel.com
Link: https://lkml.kernel.org/r/20231016053002.756205-2-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:29:54 +08:00
|
|
|
|
mm: pcp: increase pcp->free_count threshold to trigger free_high
In old pcp design, pcp->free_factor gets incremented in nr_pcp_free()
which is invoked by free_pcppages_bulk(). So, it used to increase
free_factor by 1 only when we try to reduce the size of pcp list and
free_high used to trigger only for order > 0 and order < costly_order
and pcp->free_factor > 0.
For iperf3 I noticed that with older design in kernel v6.6, pcp list
was drained mostly when pcp->count > high (more often when count goes
above 530). and most of the time pcp->free_factor was 0, triggering
very few high order flushes.
But this is changed in the current design, introduced in commit
6ccdcb6d3a74 ("mm, pcp: reduce detecting time of consecutive high order
page freeing"), where pcp->free_factor is changed to pcp->free_count to
keep track of the number of pages freed contiguously. In this design,
pcp->free_count is incremented on every deallocation, irrespective of
whether pcp list was reduced or not. And logic to trigger free_high is
if pcp->free_count goes above batch (which is 63) and there are two
contiguous page free without any allocation.
With this design, for iperf3, pcp list is getting flushed more
frequently because free_high heuristics is triggered more often now. I
observed that high order pcp list is drained as soon as both count and
free_count goes above 63.
Due to this more aggressive high order flushing, applications doing
contiguous high order allocation will require to go to global list more
frequently.
On a 2-node AMD machine with 384 vCPUs on each node, connected via
Mellonox connectX-7, I am seeing a ~30% performance reduction if we
scale number of iperf3 client/server pairs from 32 to 64.
Though this new design reduced the time to detect high order flushes,
but for application which are allocating high order pages more
frequently it may be flushing the high order list pre-maturely. This
motivates towards tuning on how late or early we should flush high
order lists.
So, in this patch, we increased the pcp->free_count threshold to
trigger free_high from "batch" to "batch + pcp->high_min / 2" as
suggested by Ying [1], In the original pcp->free_factor solution,
free_high is triggered for contiguous freeing with size ranging from
"batch" to "pcp->high + batch". So, the average value is "batch +
pcp->high / 2". While in the pcp->free_count solution, free_high will
be triggered for contiguous freeing with size "batch". So, to restore
the original behavior, we can use the threshold "batch + pcp->high_min
/ 2"
This new threshold keeps high order pages in pcp list for a longer
duration which can help the application doing high order allocations
frequently.
With this patch performace to Iperf3 is restored and score for other
benchmarks on the same machine are as follows:
iperf3 lmbench3 netperf kbuild
(AF_UNIX) (SCTP_STREAM_MANY)
------- --------- ----------------- ------
v6.6 vanilla (base) 100 100 100 100
v6.12 vanilla 69 113 98.5 98.8
v6.12 + this patch 100 110.3 100.2 99.3
netperf-tcp:
6.12 6.12
vanilla this_patch
Hmean 64 732.14 ( 0.00%) 730.45 ( -0.23%)
Hmean 128 1417.46 ( 0.00%) 1419.44 ( 0.14%)
Hmean 256 2679.67 ( 0.00%) 2676.45 ( -0.12%)
Hmean 1024 8328.52 ( 0.00%) 8339.34 ( 0.13%)
Hmean 2048 12716.98 ( 0.00%) 12743.68 ( 0.21%)
Hmean 3312 15787.79 ( 0.00%) 15887.25 ( 0.63%)
Hmean 4096 17311.91 ( 0.00%) 17332.68 ( 0.12%)
Hmean 8192 20310.73 ( 0.00%) 20465.09 ( 0.76%)
Link: https://lore.kernel.org/all/875xjmuiup.fsf@DESKTOP-5N7EMDA/ [1]
Link: https://lkml.kernel.org/r/20250407105219.55351-1-nikhil.dhama@amd.com
Fixes: 6ccdcb6d3a74 ("mm, pcp: reduce detecting time of consecutive high order page freeing")
Signed-off-by: Nikhil Dhama <nikhil.dhama@amd.com>
Suggested-by: Huang Ying <ying.huang@linux.alibaba.com>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Cc: Raghavendra K T <raghavendra.kt@amd.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Bharata B Rao <bharata@amd.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-07 16:22:19 +05:30
|
|
|
|
mm, pcp: reduce lock contention for draining high-order pages
In commit f26b3fa04611 ("mm/page_alloc: limit number of high-order pages
on PCP during bulk free"), the PCP (Per-CPU Pageset) will be drained when
PCP is mostly used for high-order pages freeing to improve the cache-hot
pages reusing between page allocating and freeing CPUs.
On system with small per-CPU data cache slice, pages shouldn't be cached
before draining to guarantee cache-hot. But on a system with large
per-CPU data cache slice, some pages can be cached before draining to
reduce zone lock contention.
So, in this patch, instead of draining without any caching, "pcp->batch"
pages will be cached in PCP before draining if the size of the per-CPU
data cache slice is more than "3 * batch".
In theory, if the size of per-CPU data cache slice is more than "2 *
batch", we can reuse cache-hot pages between CPUs. But considering the
other usage of cache (code, other data accessing, etc.), "3 * batch" is
used.
Note: "3 * batch" is chosen to make sure the optimization works on recent
x86_64 server CPUs. If you want to increase it, please check whether it
breaks the optimization.
On a 2-socket Intel server with 128 logical CPU, with the patch, the
network bandwidth of the UNIX (AF_UNIX) test case of lmbench test suite
with 16-pair processes increase 70.5%. The cycles% of the spinlock
contention (mostly for zone lock) decreases from 46.1% to 21.3%. The
number of PCP draining for high order pages freeing (free_high) decreases
89.9%. The cache miss rate keeps 0.2%.
Link: https://lkml.kernel.org/r/20231016053002.756205-4-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:29:56 +08:00
|
|
|
|
|
|
|
|
|
2025-04-02 03:41:16 -04:00
|
|
|
|
mm, pcp: avoid to drain PCP when process exit
Patch series "mm: PCP high auto-tuning", v3.
The page allocation performance requirements of different workloads are
often different. So, we need to tune the PCP (Per-CPU Pageset) high on
each CPU automatically to optimize the page allocation performance.
The list of patches in series is as follows,
[1/9] mm, pcp: avoid to drain PCP when process exit
[2/9] cacheinfo: calculate per-CPU data cache size
[3/9] mm, pcp: reduce lock contention for draining high-order pages
[4/9] mm: restrict the pcp batch scale factor to avoid too long latency
[5/9] mm, page_alloc: scale the number of pages that are batch allocated
[6/9] mm: add framework for PCP high auto-tuning
[7/9] mm: tune PCP high automatically
[8/9] mm, pcp: decrease PCP high if free pages < high watermark
[9/9] mm, pcp: reduce detecting time of consecutive high order page freeing
Patch [1/9], [2/9], [3/9] optimize the PCP draining for consecutive
high-order pages freeing.
Patch [4/9], [5/9] optimize batch freeing and allocating.
Patch [6/9], [7/9], [8/9] implement and optimize a PCP high
auto-tuning method.
Patch [9/9] optimize the PCP draining for consecutive high order page
freeing based on PCP high auto-tuning.
The test results for patches with performance impact are as follows,
kbuild
======
On a 2-socket Intel server with 224 logical CPU, we run 8 kbuild instances
in parallel (each with `make -j 28`) in 8 cgroup. This simulates the
kbuild server that is used by 0-Day kbuild service.
build time lock contend% free_high alloc_zone
---------- ---------- --------- ----------
base 100.0 14.0 100.0 100.0
patch1 99.5 12.8 19.5 95.6
patch3 99.4 12.6 7.1 95.6
patch5 98.6 11.0 8.1 97.1
patch7 95.1 0.5 2.8 15.6
patch9 95.0 1.0 8.8 20.0
The PCP draining optimization (patch [1/9], [3/9]) and PCP batch
allocation optimization (patch [5/9]) reduces zone lock contention a
little. The PCP high auto-tuning (patch [7/9], [9/9]) reduces build time
visibly. Where the tuning target: the number of pages allocated from zone
reduces greatly. So, the zone contention cycles% reduces greatly.
With PCP tuning patches (patch [7/9], [9/9]), the average used memory
during test increases up to 18.4% because more pages are cached in PCP.
But at the end of the test, the number of the used memory decreases to the
same level as that of the base patch. That is, the pages cached in PCP
will be released to zone after not being used actively.
netperf SCTP_STREAM_MANY
========================
On a 2-socket Intel server with 128 logical CPU, we tested
SCTP_STREAM_MANY test case of netperf test suite with 64-pair processes.
score lock contend% free_high alloc_zone cache miss rate%
----- ---------- --------- ---------- ----------------
base 100.0 2.1 100.0 100.0 1.3
patch1 99.4 2.1 99.4 99.4 1.3
patch3 106.4 1.3 13.3 106.3 1.3
patch5 106.0 1.2 13.2 105.9 1.3
patch7 103.4 1.9 6.7 90.3 7.6
patch9 108.6 1.3 13.7 108.6 1.3
The PCP draining optimization (patch [1/9]+[3/9]) improves performance.
The PCP high auto-tuning (patch [7/9]) reduces performance a little
because PCP draining cannot be triggered in time sometimes. So, the cache
miss rate% increases. The further PCP draining optimization (patch [9/9])
based on PCP tuning restore the performance.
lmbench3 UNIX (AF_UNIX)
=======================
On a 2-socket Intel server with 128 logical CPU, we tested UNIX
(AF_UNIX socket) test case of lmbench3 test suite with 16-pair
processes.
score lock contend% free_high alloc_zone cache miss rate%
----- ---------- --------- ---------- ----------------
base 100.0 51.4 100.0 100.0 0.2
patch1 116.8 46.1 69.5 104.3 0.2
patch3 199.1 21.3 7.0 104.9 0.2
patch5 200.0 20.8 7.1 106.9 0.3
patch7 191.6 19.9 6.8 103.8 2.8
patch9 193.4 21.7 7.0 104.7 2.1
The PCP draining optimization (patch [1/9], [3/9]) improves performance
much. The PCP tuning (patch [7/9]) reduces performance a little because
PCP draining cannot be triggered in time sometimes. The further PCP
draining optimization (patch [9/9]) based on PCP tuning restores the
performance partly.
The patchset adds several fields in struct per_cpu_pages. The struct
layout before/after the patchset is as follows,
base
====
struct per_cpu_pages {
spinlock_t lock; /* 0 4 */
int count; /* 4 4 */
int high; /* 8 4 */
int batch; /* 12 4 */
short int free_factor; /* 16 2 */
short int expire; /* 18 2 */
/* XXX 4 bytes hole, try to pack */
struct list_head lists[13]; /* 24 208 */
/* size: 256, cachelines: 4, members: 7 */
/* sum members: 228, holes: 1, sum holes: 4 */
/* padding: 24 */
} __attribute__((__aligned__(64)));
patched
=======
struct per_cpu_pages {
spinlock_t lock; /* 0 4 */
int count; /* 4 4 */
int high; /* 8 4 */
int high_min; /* 12 4 */
int high_max; /* 16 4 */
int batch; /* 20 4 */
u8 flags; /* 24 1 */
u8 alloc_factor; /* 25 1 */
u8 expire; /* 26 1 */
/* XXX 1 byte hole, try to pack */
short int free_count; /* 28 2 */
/* XXX 2 bytes hole, try to pack */
struct list_head lists[13]; /* 32 208 */
/* size: 256, cachelines: 4, members: 11 */
/* sum members: 237, holes: 2, sum holes: 3 */
/* padding: 16 */
} __attribute__((__aligned__(64)));
The size of the struct doesn't changed with the patchset.
This patch (of 9):
In commit f26b3fa04611 ("mm/page_alloc: limit number of high-order pages
on PCP during bulk free"), the PCP (Per-CPU Pageset) will be drained when
PCP is mostly used for high-order pages freeing to improve the cache-hot
pages reusing between page allocation and freeing CPUs.
But, the PCP draining mechanism may be triggered unexpectedly when process
exits. With some customized trace point, it was found that PCP draining
(free_high == true) was triggered with the order-1 page freeing with the
following call stack,
=> free_unref_page_commit
=> free_unref_page
=> __mmdrop
=> exit_mm
=> do_exit
=> do_group_exit
=> __x64_sys_exit_group
=> do_syscall_64
Checking the source code, this is the page table PGD freeing
(mm_free_pgd()). It's a order-1 page freeing if
CONFIG_PAGE_TABLE_ISOLATION=y. Which is a common configuration for
security.
Just before that, page freeing with the following call stack was found,
=> free_unref_page_commit
=> free_unref_page_list
=> release_pages
=> tlb_batch_pages_flush
=> tlb_finish_mmu
=> exit_mmap
=> __mmput
=> exit_mm
=> do_exit
=> do_group_exit
=> __x64_sys_exit_group
=> do_syscall_64
So, when a process exits,
- a large number of user pages of the process will be freed without
page allocation, it's highly possible that pcp->free_factor becomes >
0. In fact, this is expected behavior to improve process exit
performance.
- after freeing all user pages, the PGD will be freed, which is a
order-1 page freeing, PCP will be drained.
All in all, when a process exits, it's high possible that the PCP will be
drained. This is an unexpected behavior.
To avoid this, in the patch, the PCP draining will only be triggered for 2
consecutive high-order page freeing.
On a 2-socket Intel server with 224 logical CPU, we run 8 kbuild instances
in parallel (each with `make -j 28`) in 8 cgroup. This simulates the
kbuild server that is used by 0-Day kbuild service. With the patch, the
cycles% of the spinlock contention (mostly for zone lock) decreases from
14.0% to 12.8% (with PCP size == 367). The number of PCP draining for
high order pages freeing (free_high) decreases 80.5%.
This helps network workload too for reduced zone lock contention. On a
2-socket Intel server with 128 logical CPU, with the patch, the network
bandwidth of the UNIX (AF_UNIX) test case of lmbench test suite with
16-pair processes increase 16.8%. The cycles% of the spinlock contention
(mostly for zone lock) decreases from 51.4% to 46.1%. The number of PCP
draining for high order pages freeing (free_high) decreases 30.5%. The
cache miss rate keeps 0.2%.
Link: https://lkml.kernel.org/r/20231016053002.756205-1-ying.huang@intel.com
Link: https://lkml.kernel.org/r/20231016053002.756205-2-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:29:54 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, pcp: reduce detecting time of consecutive high order page freeing
In current PCP auto-tuning design, if the number of pages allocated is
much more than that of pages freed on a CPU, the PCP high may become the
maximal value even if the allocating/freeing depth is small, for example,
in the sender of network workloads. If a CPU was used as sender
originally, then it is used as receiver after context switching, we need
to fill the whole PCP with maximal high before triggering PCP draining for
consecutive high order freeing. This will hurt the performance of some
network workloads.
To solve the issue, in this patch, we will track the consecutive page
freeing with a counter in stead of relying on PCP draining. So, we can
detect consecutive page freeing much earlier.
On a 2-socket Intel server with 128 logical CPU, we tested
SCTP_STREAM_MANY test case of netperf test suite with 64-pair processes.
With the patch, the network bandwidth improves 5.0%. This restores the
performance drop caused by PCP auto-tuning.
Link: https://lkml.kernel.org/r/20231016053002.756205-10-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:02 +08:00
|
|
|
|
|
|
|
|
|
2025-02-21 18:44:24 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: re-enable kswapd when memory pressure subsides or demotion is toggled
If kswapd fails to reclaim pages from a node MAX_RECLAIM_RETRIES in a
row, kswapd on that node gets disabled. That is, the system won't wakeup
kswapd for that node until page reclamation is observed at least once.
That reclamation is mostly done by direct reclaim, which in turn enables
kswapd back.
However, on systems with CXL memory nodes, workloads with high anon page
usage can disable kswapd indefinitely, without triggering direct
reclaim. This can be reproduced with following steps:
numa node 0 (32GB memory, 48 CPUs)
numa node 2~5 (512GB CXL memory, 128GB each)
(numa node 1 is disabled)
swap space 8GB
1) Set /sys/kernel/mm/demotion_enabled to 0.
2) Set /proc/sys/kernel/numa_balancing to 0.
3) Run a process that allocates and random accesses 500GB of anon
pages.
4) Let the process exit normally.
During 3), free memory on node 0 gets lower than low watermark, and
kswapd runs and depletes swap space. Then, kswapd fails consecutively
and gets disabled. Allocation afterwards happens on CXL memory, so node
0 never gains more memory pressure to trigger direct reclaim.
After 4), kswapd on node 0 remains disabled, and tasks running on that
node are unable to swap. If you turn on NUMA_BALANCING_MEMORY_TIERING
and demotion now, it won't work properly since kswapd is disabled.
To mitigate this problem, reset kswapd_failures to 0 on following
conditions:
a) ZONE_BELOW_HIGH bit of a zone in hopeless node with a fallback
memory node gets cleared.
b) demotion_enabled is changed from false to true.
Rationale for a):
ZONE_BELOW_HIGH bit being cleared might be a sign that the node may
be reclaimable afterwards. This won't help much if the memory-hungry
process keeps running without freeing anything, but at least the node
will go back to reclaimable state when the process exits.
Rationale for b):
When demotion_enabled is false, kswapd can only reclaim anon pages by
swapping them out to swap space. If demotion_enabled is turned on,
kswapd can demote anon pages to another node for reclaiming. So, the
original failure count for determining reclaimability is no longer
valid.
Since kswapd_failures resets may be missed by ++ operation, it is
changed from int to atomic_t.
[akpm@linux-foundation.org: tweak whitespace]
Link: https://lkml.kernel.org/r/aL6qGi69jWXfPc4D@pcw-MS-7D22
Signed-off-by: Chanwon Park <flyinrm@gmail.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-09-08 19:04:10 +09:00
|
|
|
|
mm: tune PCP high automatically
The target to tune PCP high automatically is as follows,
- Minimize allocation/freeing from/to shared zone
- Minimize idle pages in PCP
- Minimize pages in PCP if the system free pages is too few
To reach these target, a tuning algorithm as follows is designed,
- When we refill PCP via allocating from the zone, increase PCP high.
Because if we had larger PCP, we could avoid to allocate from the
zone.
- In periodic vmstat updating kworker (via refresh_cpu_vm_stats()),
decrease PCP high to try to free possible idle PCP pages.
- When page reclaiming is active for the zone, stop increasing PCP
high in allocating path, decrease PCP high and free some pages in
freeing path.
So, the PCP high can be tuned to the page allocating/freeing depth of
workloads eventually.
One issue of the algorithm is that if the number of pages allocated is
much more than that of pages freed on a CPU, the PCP high may become the
maximal value even if the allocating/freeing depth is small. But this
isn't a severe issue, because there are no idle pages in this case.
One alternative choice is to increase PCP high when we drain PCP via
trying to free pages to the zone, but don't increase PCP high during PCP
refilling. This can avoid the issue above. But if the number of pages
allocated is much less than that of pages freed on a CPU, there will be
many idle pages in PCP and it is hard to free these idle pages.
1/8 (>> 3) of PCP high will be decreased periodically. The value 1/8 is
kind of arbitrary. Just to make sure that the idle PCP pages will be
freed eventually.
On a 2-socket Intel server with 224 logical CPU, we run 8 kbuild instances
in parallel (each with `make -j 28`) in 8 cgroup. This simulates the
kbuild server that is used by 0-Day kbuild service. With the patch, the
build time decreases 3.5%. The cycles% of the spinlock contention (mostly
for zone lock) decreases from 11.0% to 0.5%. The number of PCP draining
for high order pages freeing (free_high) decreases 65.6%. The number of
pages allocated from zone (instead of from PCP) decreases 83.9%.
Link: https://lkml.kernel.org/r/20231016053002.756205-8-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Suggested-by: Mel Gorman <mgorman@techsingularity.net>
Suggested-by: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:00 +08:00
|
|
|
|
mm: re-enable kswapd when memory pressure subsides or demotion is toggled
If kswapd fails to reclaim pages from a node MAX_RECLAIM_RETRIES in a
row, kswapd on that node gets disabled. That is, the system won't wakeup
kswapd for that node until page reclamation is observed at least once.
That reclamation is mostly done by direct reclaim, which in turn enables
kswapd back.
However, on systems with CXL memory nodes, workloads with high anon page
usage can disable kswapd indefinitely, without triggering direct
reclaim. This can be reproduced with following steps:
numa node 0 (32GB memory, 48 CPUs)
numa node 2~5 (512GB CXL memory, 128GB each)
(numa node 1 is disabled)
swap space 8GB
1) Set /sys/kernel/mm/demotion_enabled to 0.
2) Set /proc/sys/kernel/numa_balancing to 0.
3) Run a process that allocates and random accesses 500GB of anon
pages.
4) Let the process exit normally.
During 3), free memory on node 0 gets lower than low watermark, and
kswapd runs and depletes swap space. Then, kswapd fails consecutively
and gets disabled. Allocation afterwards happens on CXL memory, so node
0 never gains more memory pressure to trigger direct reclaim.
After 4), kswapd on node 0 remains disabled, and tasks running on that
node are unable to swap. If you turn on NUMA_BALANCING_MEMORY_TIERING
and demotion now, it won't work properly since kswapd is disabled.
To mitigate this problem, reset kswapd_failures to 0 on following
conditions:
a) ZONE_BELOW_HIGH bit of a zone in hopeless node with a fallback
memory node gets cleared.
b) demotion_enabled is changed from false to true.
Rationale for a):
ZONE_BELOW_HIGH bit being cleared might be a sign that the node may
be reclaimable afterwards. This won't help much if the memory-hungry
process keeps running without freeing anything, but at least the node
will go back to reclaimable state when the process exits.
Rationale for b):
When demotion_enabled is false, kswapd can only reclaim anon pages by
swapping them out to swap space. If demotion_enabled is turned on,
kswapd can demote anon pages to another node for reclaiming. So, the
original failure count for determining reclaimability is no longer
valid.
Since kswapd_failures resets may be missed by ++ operation, it is
changed from int to atomic_t.
[akpm@linux-foundation.org: tweak whitespace]
Link: https://lkml.kernel.org/r/aL6qGi69jWXfPc4D@pcw-MS-7D22
Signed-off-by: Chanwon Park <flyinrm@gmail.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-09-08 19:04:10 +09:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-28 19:42:18 -07:00
|
|
|
|
mm, page_alloc: enable/disable IRQs once when freeing a list of pages
Patch series "Follow-up for speed up page cache truncation", v2.
This series is a follow-on for Jan Kara's series "Speed up page cache
truncation" series. We both ended up looking at the same problem but
saw different problems based on the same data. This series builds upon
his work.
A variety of workloads were compared on four separate machines but each
machine showed gains albeit at different levels. Minimally, some of the
differences are due to NUMA where truncating data from a remote node is
slower than a local node. The workloads checked were
o sparse truncate microbenchmark, tiny
o sparse truncate microbenchmark, large
o reaim-io disk workfile
o dbench4 (modified by mmtests to produce more stable results)
o filebench varmail configuration for small memory size
o bonnie, directory operations, working set size 2*RAM
reaim-io, dbench and filebench all showed minor gains. Truncation does
not dominate those workloads but were tested to ensure no other
regressions. They will not be reported further.
The sparse truncate microbench was written by Jan. It creates a number
of files and then times how long it takes to truncate each one. The
"tiny" configuraiton creates a number of files that easily fits in
memory and times how long it takes to truncate files with page cache.
The large configuration uses enough files to have data that is twice the
size of memory and so timings there include truncating page cache and
working set shadow entries in the radix tree.
Patches 1-4 are the most relevant parts of this series. Patches 5-8 are
optional as they are deleting code that is essentially useless but has a
negligible performance impact.
The changelogs have more information on performance but just for bonnie
delete options, the main comparison is
bonnie
4.14.0-rc5 4.14.0-rc5 4.14.0-rc5
jan-v2 vanilla mel-v2
Hmean SeqCreate ops 76.20 ( 0.00%) 75.80 ( -0.53%) 76.80 ( 0.79%)
Hmean SeqCreate read 85.00 ( 0.00%) 85.00 ( 0.00%) 85.00 ( 0.00%)
Hmean SeqCreate del 13752.31 ( 0.00%) 12090.23 ( -12.09%) 15304.84 ( 11.29%)
Hmean RandCreate ops 76.00 ( 0.00%) 75.60 ( -0.53%) 77.00 ( 1.32%)
Hmean RandCreate read 96.80 ( 0.00%) 96.80 ( 0.00%) 97.00 ( 0.21%)
Hmean RandCreate del 13233.75 ( 0.00%) 11525.35 ( -12.91%) 14446.61 ( 9.16%)
Jan's series is the baseline and the vanilla kernel is 12% slower where
as this series on top gains another 11%. This is from a different
machine than the data in the changelogs but the detailed data was not
collected as there was no substantial change in v2.
This patch (of 8):
Freeing a list of pages current enables/disables IRQs for each page
freed. This patch splits freeing a list of pages into two operations --
preparing the pages for freeing and the actual freeing. This is a
tradeoff - we're taking two passes of the list to free in exchange for
avoiding multiple enable/disable of IRQs.
sparsetruncate (tiny)
4.14.0-rc4 4.14.0-rc4
janbatch-v1r1 oneirq-v1r1
Min Time 149.00 ( 0.00%) 141.00 ( 5.37%)
1st-qrtle Time 150.00 ( 0.00%) 142.00 ( 5.33%)
2nd-qrtle Time 151.00 ( 0.00%) 142.00 ( 5.96%)
3rd-qrtle Time 151.00 ( 0.00%) 143.00 ( 5.30%)
Max-90% Time 153.00 ( 0.00%) 144.00 ( 5.88%)
Max-95% Time 155.00 ( 0.00%) 147.00 ( 5.16%)
Max-99% Time 201.00 ( 0.00%) 195.00 ( 2.99%)
Max Time 236.00 ( 0.00%) 230.00 ( 2.54%)
Amean Time 152.65 ( 0.00%) 144.37 ( 5.43%)
Stddev Time 9.78 ( 0.00%) 10.44 ( -6.72%)
Coeff Time 6.41 ( 0.00%) 7.23 ( -12.84%)
Best99%Amean Time 152.07 ( 0.00%) 143.72 ( 5.50%)
Best95%Amean Time 150.75 ( 0.00%) 142.37 ( 5.56%)
Best90%Amean Time 150.59 ( 0.00%) 142.19 ( 5.58%)
Best75%Amean Time 150.36 ( 0.00%) 141.92 ( 5.61%)
Best50%Amean Time 150.04 ( 0.00%) 141.69 ( 5.56%)
Best25%Amean Time 149.85 ( 0.00%) 141.38 ( 5.65%)
With a tiny number of files, each file truncated has resident page cache
and it shows that time to truncate is roughtly 5-6% with some minor
jitter.
4.14.0-rc4 4.14.0-rc4
janbatch-v1r1 oneirq-v1r1
Hmean SeqCreate ops 65.27 ( 0.00%) 81.86 ( 25.43%)
Hmean SeqCreate read 39.48 ( 0.00%) 47.44 ( 20.16%)
Hmean SeqCreate del 24963.95 ( 0.00%) 26319.99 ( 5.43%)
Hmean RandCreate ops 65.47 ( 0.00%) 82.01 ( 25.26%)
Hmean RandCreate read 42.04 ( 0.00%) 51.75 ( 23.09%)
Hmean RandCreate del 23377.66 ( 0.00%) 23764.79 ( 1.66%)
As expected, there is a small gain for the delete operation.
[mgorman@techsingularity.net: use page_private and set_page_private helpers]
Link: http://lkml.kernel.org/r/20171018101547.mjycw7zreb66jzpa@techsingularity.net
Link: http://lkml.kernel.org/r/20171018075952.10627-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Kara <jack@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-15 17:37:37 -08:00
|
|
|
|
page-allocator: split per-cpu list into one-list-per-migrate-type
The following two patches remove searching in the page allocator fast-path
by maintaining multiple free-lists in the per-cpu structure. At the time
the search was introduced, increasing the per-cpu structures would waste a
lot of memory as per-cpu structures were statically allocated at
compile-time. This is no longer the case.
The patches are as follows. They are based on mmotm-2009-08-27.
Patch 1 adds multiple lists to struct per_cpu_pages, one per
migratetype that can be stored on the PCP lists.
Patch 2 notes that the pcpu drain path check empty lists multiple times. The
patch reduces the number of checks by maintaining a count of free
lists encountered. Lists containing pages will then free multiple
pages in batch
The patches were tested with kernbench, netperf udp/tcp, hackbench and
sysbench. The netperf tests were not bound to any CPU in particular and
were run such that the results should be 99% confidence that the reported
results are within 1% of the estimated mean. sysbench was run with a
postgres background and read-only tests. Similar to netperf, it was run
multiple times so that it's 99% confidence results are within 1%. The
patches were tested on x86, x86-64 and ppc64 as
x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine)
kernbench - No significant difference, variance well within noise
netperf-udp - 1.34% to 2.28% gain
netperf-tcp - 0.45% to 1.22% gain
hackbench - Small variances, very close to noise
sysbench - Very small gains
x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine)
kernbench - No significant difference, variance well within noise
netperf-udp - 1.83% to 10.42% gains
netperf-tcp - No conclusive until buffer >= PAGE_SIZE
4096 +15.83%
8192 + 0.34% (not significant)
16384 + 1%
hackbench - Small gains, very close to noise
sysbench - 0.79% to 1.6% gain
ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation)
kernbench - No significant difference, variance well within noise
netperf-udp - 2-3% gain for almost all buffer sizes tested
netperf-tcp - losses on small buffers, gains on larger buffers
possibly indicates some bad caching effect.
hackbench - No significant difference
sysbench - 2-4% gain
This patch:
Currently the per-cpu page allocator searches the PCP list for pages of
the correct migrate-type to reduce the possibility of pages being
inappropriate placed from a fragmentation perspective. This search is
potentially expensive in a fast-path and undesirable. Splitting the
per-cpu list into multiple lists increases the size of a per-cpu structure
and this was potentially a major problem at the time the search was
introduced. These problem has been mitigated as now only the necessary
number of structures is allocated for the running system.
This patch replaces a list search in the per-cpu allocator with one list
per migrate type. The potential snag with this approach is when bulk
freeing pages. We round-robin free pages based on migrate type which has
little bearing on the cache hotness of the page and potentially checks
empty lists repeatedly in the event the majority of PCP pages are of one
type.
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Acked-by: Nick Piggin <npiggin@suse.de>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-21 17:03:19 -07:00
|
|
|
|
mm, page_alloc: enable/disable IRQs once when freeing a list of pages
Patch series "Follow-up for speed up page cache truncation", v2.
This series is a follow-on for Jan Kara's series "Speed up page cache
truncation" series. We both ended up looking at the same problem but
saw different problems based on the same data. This series builds upon
his work.
A variety of workloads were compared on four separate machines but each
machine showed gains albeit at different levels. Minimally, some of the
differences are due to NUMA where truncating data from a remote node is
slower than a local node. The workloads checked were
o sparse truncate microbenchmark, tiny
o sparse truncate microbenchmark, large
o reaim-io disk workfile
o dbench4 (modified by mmtests to produce more stable results)
o filebench varmail configuration for small memory size
o bonnie, directory operations, working set size 2*RAM
reaim-io, dbench and filebench all showed minor gains. Truncation does
not dominate those workloads but were tested to ensure no other
regressions. They will not be reported further.
The sparse truncate microbench was written by Jan. It creates a number
of files and then times how long it takes to truncate each one. The
"tiny" configuraiton creates a number of files that easily fits in
memory and times how long it takes to truncate files with page cache.
The large configuration uses enough files to have data that is twice the
size of memory and so timings there include truncating page cache and
working set shadow entries in the radix tree.
Patches 1-4 are the most relevant parts of this series. Patches 5-8 are
optional as they are deleting code that is essentially useless but has a
negligible performance impact.
The changelogs have more information on performance but just for bonnie
delete options, the main comparison is
bonnie
4.14.0-rc5 4.14.0-rc5 4.14.0-rc5
jan-v2 vanilla mel-v2
Hmean SeqCreate ops 76.20 ( 0.00%) 75.80 ( -0.53%) 76.80 ( 0.79%)
Hmean SeqCreate read 85.00 ( 0.00%) 85.00 ( 0.00%) 85.00 ( 0.00%)
Hmean SeqCreate del 13752.31 ( 0.00%) 12090.23 ( -12.09%) 15304.84 ( 11.29%)
Hmean RandCreate ops 76.00 ( 0.00%) 75.60 ( -0.53%) 77.00 ( 1.32%)
Hmean RandCreate read 96.80 ( 0.00%) 96.80 ( 0.00%) 97.00 ( 0.21%)
Hmean RandCreate del 13233.75 ( 0.00%) 11525.35 ( -12.91%) 14446.61 ( 9.16%)
Jan's series is the baseline and the vanilla kernel is 12% slower where
as this series on top gains another 11%. This is from a different
machine than the data in the changelogs but the detailed data was not
collected as there was no substantial change in v2.
This patch (of 8):
Freeing a list of pages current enables/disables IRQs for each page
freed. This patch splits freeing a list of pages into two operations --
preparing the pages for freeing and the actual freeing. This is a
tradeoff - we're taking two passes of the list to free in exchange for
avoiding multiple enable/disable of IRQs.
sparsetruncate (tiny)
4.14.0-rc4 4.14.0-rc4
janbatch-v1r1 oneirq-v1r1
Min Time 149.00 ( 0.00%) 141.00 ( 5.37%)
1st-qrtle Time 150.00 ( 0.00%) 142.00 ( 5.33%)
2nd-qrtle Time 151.00 ( 0.00%) 142.00 ( 5.96%)
3rd-qrtle Time 151.00 ( 0.00%) 143.00 ( 5.30%)
Max-90% Time 153.00 ( 0.00%) 144.00 ( 5.88%)
Max-95% Time 155.00 ( 0.00%) 147.00 ( 5.16%)
Max-99% Time 201.00 ( 0.00%) 195.00 ( 2.99%)
Max Time 236.00 ( 0.00%) 230.00 ( 2.54%)
Amean Time 152.65 ( 0.00%) 144.37 ( 5.43%)
Stddev Time 9.78 ( 0.00%) 10.44 ( -6.72%)
Coeff Time 6.41 ( 0.00%) 7.23 ( -12.84%)
Best99%Amean Time 152.07 ( 0.00%) 143.72 ( 5.50%)
Best95%Amean Time 150.75 ( 0.00%) 142.37 ( 5.56%)
Best90%Amean Time 150.59 ( 0.00%) 142.19 ( 5.58%)
Best75%Amean Time 150.36 ( 0.00%) 141.92 ( 5.61%)
Best50%Amean Time 150.04 ( 0.00%) 141.69 ( 5.56%)
Best25%Amean Time 149.85 ( 0.00%) 141.38 ( 5.65%)
With a tiny number of files, each file truncated has resident page cache
and it shows that time to truncate is roughtly 5-6% with some minor
jitter.
4.14.0-rc4 4.14.0-rc4
janbatch-v1r1 oneirq-v1r1
Hmean SeqCreate ops 65.27 ( 0.00%) 81.86 ( 25.43%)
Hmean SeqCreate read 39.48 ( 0.00%) 47.44 ( 20.16%)
Hmean SeqCreate del 24963.95 ( 0.00%) 26319.99 ( 5.43%)
Hmean RandCreate ops 65.47 ( 0.00%) 82.01 ( 25.26%)
Hmean RandCreate read 42.04 ( 0.00%) 51.75 ( 23.09%)
Hmean RandCreate del 23377.66 ( 0.00%) 23764.79 ( 1.66%)
As expected, there is a small gain for the delete operation.
[mgorman@techsingularity.net: use page_private and set_page_private helpers]
Link: http://lkml.kernel.org/r/20171018101547.mjycw7zreb66jzpa@techsingularity.net
Link: http://lkml.kernel.org/r/20171018075952.10627-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Kara <jack@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-15 17:37:37 -08:00
|
|
|
|
2021-06-28 19:43:08 -07:00
|
|
|
|
mm, page_alloc: enable/disable IRQs once when freeing a list of pages
Patch series "Follow-up for speed up page cache truncation", v2.
This series is a follow-on for Jan Kara's series "Speed up page cache
truncation" series. We both ended up looking at the same problem but
saw different problems based on the same data. This series builds upon
his work.
A variety of workloads were compared on four separate machines but each
machine showed gains albeit at different levels. Minimally, some of the
differences are due to NUMA where truncating data from a remote node is
slower than a local node. The workloads checked were
o sparse truncate microbenchmark, tiny
o sparse truncate microbenchmark, large
o reaim-io disk workfile
o dbench4 (modified by mmtests to produce more stable results)
o filebench varmail configuration for small memory size
o bonnie, directory operations, working set size 2*RAM
reaim-io, dbench and filebench all showed minor gains. Truncation does
not dominate those workloads but were tested to ensure no other
regressions. They will not be reported further.
The sparse truncate microbench was written by Jan. It creates a number
of files and then times how long it takes to truncate each one. The
"tiny" configuraiton creates a number of files that easily fits in
memory and times how long it takes to truncate files with page cache.
The large configuration uses enough files to have data that is twice the
size of memory and so timings there include truncating page cache and
working set shadow entries in the radix tree.
Patches 1-4 are the most relevant parts of this series. Patches 5-8 are
optional as they are deleting code that is essentially useless but has a
negligible performance impact.
The changelogs have more information on performance but just for bonnie
delete options, the main comparison is
bonnie
4.14.0-rc5 4.14.0-rc5 4.14.0-rc5
jan-v2 vanilla mel-v2
Hmean SeqCreate ops 76.20 ( 0.00%) 75.80 ( -0.53%) 76.80 ( 0.79%)
Hmean SeqCreate read 85.00 ( 0.00%) 85.00 ( 0.00%) 85.00 ( 0.00%)
Hmean SeqCreate del 13752.31 ( 0.00%) 12090.23 ( -12.09%) 15304.84 ( 11.29%)
Hmean RandCreate ops 76.00 ( 0.00%) 75.60 ( -0.53%) 77.00 ( 1.32%)
Hmean RandCreate read 96.80 ( 0.00%) 96.80 ( 0.00%) 97.00 ( 0.21%)
Hmean RandCreate del 13233.75 ( 0.00%) 11525.35 ( -12.91%) 14446.61 ( 9.16%)
Jan's series is the baseline and the vanilla kernel is 12% slower where
as this series on top gains another 11%. This is from a different
machine than the data in the changelogs but the detailed data was not
collected as there was no substantial change in v2.
This patch (of 8):
Freeing a list of pages current enables/disables IRQs for each page
freed. This patch splits freeing a list of pages into two operations --
preparing the pages for freeing and the actual freeing. This is a
tradeoff - we're taking two passes of the list to free in exchange for
avoiding multiple enable/disable of IRQs.
sparsetruncate (tiny)
4.14.0-rc4 4.14.0-rc4
janbatch-v1r1 oneirq-v1r1
Min Time 149.00 ( 0.00%) 141.00 ( 5.37%)
1st-qrtle Time 150.00 ( 0.00%) 142.00 ( 5.33%)
2nd-qrtle Time 151.00 ( 0.00%) 142.00 ( 5.96%)
3rd-qrtle Time 151.00 ( 0.00%) 143.00 ( 5.30%)
Max-90% Time 153.00 ( 0.00%) 144.00 ( 5.88%)
Max-95% Time 155.00 ( 0.00%) 147.00 ( 5.16%)
Max-99% Time 201.00 ( 0.00%) 195.00 ( 2.99%)
Max Time 236.00 ( 0.00%) 230.00 ( 2.54%)
Amean Time 152.65 ( 0.00%) 144.37 ( 5.43%)
Stddev Time 9.78 ( 0.00%) 10.44 ( -6.72%)
Coeff Time 6.41 ( 0.00%) 7.23 ( -12.84%)
Best99%Amean Time 152.07 ( 0.00%) 143.72 ( 5.50%)
Best95%Amean Time 150.75 ( 0.00%) 142.37 ( 5.56%)
Best90%Amean Time 150.59 ( 0.00%) 142.19 ( 5.58%)
Best75%Amean Time 150.36 ( 0.00%) 141.92 ( 5.61%)
Best50%Amean Time 150.04 ( 0.00%) 141.69 ( 5.56%)
Best25%Amean Time 149.85 ( 0.00%) 141.38 ( 5.65%)
With a tiny number of files, each file truncated has resident page cache
and it shows that time to truncate is roughtly 5-6% with some minor
jitter.
4.14.0-rc4 4.14.0-rc4
janbatch-v1r1 oneirq-v1r1
Hmean SeqCreate ops 65.27 ( 0.00%) 81.86 ( 25.43%)
Hmean SeqCreate read 39.48 ( 0.00%) 47.44 ( 20.16%)
Hmean SeqCreate del 24963.95 ( 0.00%) 26319.99 ( 5.43%)
Hmean RandCreate ops 65.47 ( 0.00%) 82.01 ( 25.26%)
Hmean RandCreate read 42.04 ( 0.00%) 51.75 ( 23.09%)
Hmean RandCreate del 23377.66 ( 0.00%) 23764.79 ( 1.66%)
As expected, there is a small gain for the delete operation.
[mgorman@techsingularity.net: use page_private and set_page_private helpers]
Link: http://lkml.kernel.org/r/20171018101547.mjycw7zreb66jzpa@techsingularity.net
Link: http://lkml.kernel.org/r/20171018075952.10627-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Kara <jack@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-15 17:37:37 -08:00
|
|
|
|
2025-02-21 18:44:24 -08:00
|
|
|
|
|
|
|
|
|
mm, page_alloc: enable/disable IRQs once when freeing a list of pages
Patch series "Follow-up for speed up page cache truncation", v2.
This series is a follow-on for Jan Kara's series "Speed up page cache
truncation" series. We both ended up looking at the same problem but
saw different problems based on the same data. This series builds upon
his work.
A variety of workloads were compared on four separate machines but each
machine showed gains albeit at different levels. Minimally, some of the
differences are due to NUMA where truncating data from a remote node is
slower than a local node. The workloads checked were
o sparse truncate microbenchmark, tiny
o sparse truncate microbenchmark, large
o reaim-io disk workfile
o dbench4 (modified by mmtests to produce more stable results)
o filebench varmail configuration for small memory size
o bonnie, directory operations, working set size 2*RAM
reaim-io, dbench and filebench all showed minor gains. Truncation does
not dominate those workloads but were tested to ensure no other
regressions. They will not be reported further.
The sparse truncate microbench was written by Jan. It creates a number
of files and then times how long it takes to truncate each one. The
"tiny" configuraiton creates a number of files that easily fits in
memory and times how long it takes to truncate files with page cache.
The large configuration uses enough files to have data that is twice the
size of memory and so timings there include truncating page cache and
working set shadow entries in the radix tree.
Patches 1-4 are the most relevant parts of this series. Patches 5-8 are
optional as they are deleting code that is essentially useless but has a
negligible performance impact.
The changelogs have more information on performance but just for bonnie
delete options, the main comparison is
bonnie
4.14.0-rc5 4.14.0-rc5 4.14.0-rc5
jan-v2 vanilla mel-v2
Hmean SeqCreate ops 76.20 ( 0.00%) 75.80 ( -0.53%) 76.80 ( 0.79%)
Hmean SeqCreate read 85.00 ( 0.00%) 85.00 ( 0.00%) 85.00 ( 0.00%)
Hmean SeqCreate del 13752.31 ( 0.00%) 12090.23 ( -12.09%) 15304.84 ( 11.29%)
Hmean RandCreate ops 76.00 ( 0.00%) 75.60 ( -0.53%) 77.00 ( 1.32%)
Hmean RandCreate read 96.80 ( 0.00%) 96.80 ( 0.00%) 97.00 ( 0.21%)
Hmean RandCreate del 13233.75 ( 0.00%) 11525.35 ( -12.91%) 14446.61 ( 9.16%)
Jan's series is the baseline and the vanilla kernel is 12% slower where
as this series on top gains another 11%. This is from a different
machine than the data in the changelogs but the detailed data was not
collected as there was no substantial change in v2.
This patch (of 8):
Freeing a list of pages current enables/disables IRQs for each page
freed. This patch splits freeing a list of pages into two operations --
preparing the pages for freeing and the actual freeing. This is a
tradeoff - we're taking two passes of the list to free in exchange for
avoiding multiple enable/disable of IRQs.
sparsetruncate (tiny)
4.14.0-rc4 4.14.0-rc4
janbatch-v1r1 oneirq-v1r1
Min Time 149.00 ( 0.00%) 141.00 ( 5.37%)
1st-qrtle Time 150.00 ( 0.00%) 142.00 ( 5.33%)
2nd-qrtle Time 151.00 ( 0.00%) 142.00 ( 5.96%)
3rd-qrtle Time 151.00 ( 0.00%) 143.00 ( 5.30%)
Max-90% Time 153.00 ( 0.00%) 144.00 ( 5.88%)
Max-95% Time 155.00 ( 0.00%) 147.00 ( 5.16%)
Max-99% Time 201.00 ( 0.00%) 195.00 ( 2.99%)
Max Time 236.00 ( 0.00%) 230.00 ( 2.54%)
Amean Time 152.65 ( 0.00%) 144.37 ( 5.43%)
Stddev Time 9.78 ( 0.00%) 10.44 ( -6.72%)
Coeff Time 6.41 ( 0.00%) 7.23 ( -12.84%)
Best99%Amean Time 152.07 ( 0.00%) 143.72 ( 5.50%)
Best95%Amean Time 150.75 ( 0.00%) 142.37 ( 5.56%)
Best90%Amean Time 150.59 ( 0.00%) 142.19 ( 5.58%)
Best75%Amean Time 150.36 ( 0.00%) 141.92 ( 5.61%)
Best50%Amean Time 150.04 ( 0.00%) 141.69 ( 5.56%)
Best25%Amean Time 149.85 ( 0.00%) 141.38 ( 5.65%)
With a tiny number of files, each file truncated has resident page cache
and it shows that time to truncate is roughtly 5-6% with some minor
jitter.
4.14.0-rc4 4.14.0-rc4
janbatch-v1r1 oneirq-v1r1
Hmean SeqCreate ops 65.27 ( 0.00%) 81.86 ( 25.43%)
Hmean SeqCreate read 39.48 ( 0.00%) 47.44 ( 20.16%)
Hmean SeqCreate del 24963.95 ( 0.00%) 26319.99 ( 5.43%)
Hmean RandCreate ops 65.47 ( 0.00%) 82.01 ( 25.26%)
Hmean RandCreate read 42.04 ( 0.00%) 51.75 ( 23.09%)
Hmean RandCreate del 23377.66 ( 0.00%) 23764.79 ( 1.66%)
As expected, there is a small gain for the delete operation.
[mgorman@techsingularity.net: use page_private and set_page_private helpers]
Link: http://lkml.kernel.org/r/20171018101547.mjycw7zreb66jzpa@techsingularity.net
Link: http://lkml.kernel.org/r/20171018075952.10627-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Kara <jack@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-15 17:37:37 -08:00
|
|
|
|
2022-06-24 13:54:21 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, page_alloc: enable/disable IRQs once when freeing a list of pages
Patch series "Follow-up for speed up page cache truncation", v2.
This series is a follow-on for Jan Kara's series "Speed up page cache
truncation" series. We both ended up looking at the same problem but
saw different problems based on the same data. This series builds upon
his work.
A variety of workloads were compared on four separate machines but each
machine showed gains albeit at different levels. Minimally, some of the
differences are due to NUMA where truncating data from a remote node is
slower than a local node. The workloads checked were
o sparse truncate microbenchmark, tiny
o sparse truncate microbenchmark, large
o reaim-io disk workfile
o dbench4 (modified by mmtests to produce more stable results)
o filebench varmail configuration for small memory size
o bonnie, directory operations, working set size 2*RAM
reaim-io, dbench and filebench all showed minor gains. Truncation does
not dominate those workloads but were tested to ensure no other
regressions. They will not be reported further.
The sparse truncate microbench was written by Jan. It creates a number
of files and then times how long it takes to truncate each one. The
"tiny" configuraiton creates a number of files that easily fits in
memory and times how long it takes to truncate files with page cache.
The large configuration uses enough files to have data that is twice the
size of memory and so timings there include truncating page cache and
working set shadow entries in the radix tree.
Patches 1-4 are the most relevant parts of this series. Patches 5-8 are
optional as they are deleting code that is essentially useless but has a
negligible performance impact.
The changelogs have more information on performance but just for bonnie
delete options, the main comparison is
bonnie
4.14.0-rc5 4.14.0-rc5 4.14.0-rc5
jan-v2 vanilla mel-v2
Hmean SeqCreate ops 76.20 ( 0.00%) 75.80 ( -0.53%) 76.80 ( 0.79%)
Hmean SeqCreate read 85.00 ( 0.00%) 85.00 ( 0.00%) 85.00 ( 0.00%)
Hmean SeqCreate del 13752.31 ( 0.00%) 12090.23 ( -12.09%) 15304.84 ( 11.29%)
Hmean RandCreate ops 76.00 ( 0.00%) 75.60 ( -0.53%) 77.00 ( 1.32%)
Hmean RandCreate read 96.80 ( 0.00%) 96.80 ( 0.00%) 97.00 ( 0.21%)
Hmean RandCreate del 13233.75 ( 0.00%) 11525.35 ( -12.91%) 14446.61 ( 9.16%)
Jan's series is the baseline and the vanilla kernel is 12% slower where
as this series on top gains another 11%. This is from a different
machine than the data in the changelogs but the detailed data was not
collected as there was no substantial change in v2.
This patch (of 8):
Freeing a list of pages current enables/disables IRQs for each page
freed. This patch splits freeing a list of pages into two operations --
preparing the pages for freeing and the actual freeing. This is a
tradeoff - we're taking two passes of the list to free in exchange for
avoiding multiple enable/disable of IRQs.
sparsetruncate (tiny)
4.14.0-rc4 4.14.0-rc4
janbatch-v1r1 oneirq-v1r1
Min Time 149.00 ( 0.00%) 141.00 ( 5.37%)
1st-qrtle Time 150.00 ( 0.00%) 142.00 ( 5.33%)
2nd-qrtle Time 151.00 ( 0.00%) 142.00 ( 5.96%)
3rd-qrtle Time 151.00 ( 0.00%) 143.00 ( 5.30%)
Max-90% Time 153.00 ( 0.00%) 144.00 ( 5.88%)
Max-95% Time 155.00 ( 0.00%) 147.00 ( 5.16%)
Max-99% Time 201.00 ( 0.00%) 195.00 ( 2.99%)
Max Time 236.00 ( 0.00%) 230.00 ( 2.54%)
Amean Time 152.65 ( 0.00%) 144.37 ( 5.43%)
Stddev Time 9.78 ( 0.00%) 10.44 ( -6.72%)
Coeff Time 6.41 ( 0.00%) 7.23 ( -12.84%)
Best99%Amean Time 152.07 ( 0.00%) 143.72 ( 5.50%)
Best95%Amean Time 150.75 ( 0.00%) 142.37 ( 5.56%)
Best90%Amean Time 150.59 ( 0.00%) 142.19 ( 5.58%)
Best75%Amean Time 150.36 ( 0.00%) 141.92 ( 5.61%)
Best50%Amean Time 150.04 ( 0.00%) 141.69 ( 5.56%)
Best25%Amean Time 149.85 ( 0.00%) 141.38 ( 5.65%)
With a tiny number of files, each file truncated has resident page cache
and it shows that time to truncate is roughtly 5-6% with some minor
jitter.
4.14.0-rc4 4.14.0-rc4
janbatch-v1r1 oneirq-v1r1
Hmean SeqCreate ops 65.27 ( 0.00%) 81.86 ( 25.43%)
Hmean SeqCreate read 39.48 ( 0.00%) 47.44 ( 20.16%)
Hmean SeqCreate del 24963.95 ( 0.00%) 26319.99 ( 5.43%)
Hmean RandCreate ops 65.47 ( 0.00%) 82.01 ( 25.26%)
Hmean RandCreate read 42.04 ( 0.00%) 51.75 ( 23.09%)
Hmean RandCreate del 23377.66 ( 0.00%) 23764.79 ( 1.66%)
As expected, there is a small gain for the delete operation.
[mgorman@techsingularity.net: use page_private and set_page_private helpers]
Link: http://lkml.kernel.org/r/20171018101547.mjycw7zreb66jzpa@techsingularity.net
Link: http://lkml.kernel.org/r/20171018075952.10627-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Kara <jack@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-15 17:37:37 -08:00
|
|
|
|
2024-03-20 14:02:12 -04:00
|
|
|
|
mm, page_alloc: enable/disable IRQs once when freeing a list of pages
Patch series "Follow-up for speed up page cache truncation", v2.
This series is a follow-on for Jan Kara's series "Speed up page cache
truncation" series. We both ended up looking at the same problem but
saw different problems based on the same data. This series builds upon
his work.
A variety of workloads were compared on four separate machines but each
machine showed gains albeit at different levels. Minimally, some of the
differences are due to NUMA where truncating data from a remote node is
slower than a local node. The workloads checked were
o sparse truncate microbenchmark, tiny
o sparse truncate microbenchmark, large
o reaim-io disk workfile
o dbench4 (modified by mmtests to produce more stable results)
o filebench varmail configuration for small memory size
o bonnie, directory operations, working set size 2*RAM
reaim-io, dbench and filebench all showed minor gains. Truncation does
not dominate those workloads but were tested to ensure no other
regressions. They will not be reported further.
The sparse truncate microbench was written by Jan. It creates a number
of files and then times how long it takes to truncate each one. The
"tiny" configuraiton creates a number of files that easily fits in
memory and times how long it takes to truncate files with page cache.
The large configuration uses enough files to have data that is twice the
size of memory and so timings there include truncating page cache and
working set shadow entries in the radix tree.
Patches 1-4 are the most relevant parts of this series. Patches 5-8 are
optional as they are deleting code that is essentially useless but has a
negligible performance impact.
The changelogs have more information on performance but just for bonnie
delete options, the main comparison is
bonnie
4.14.0-rc5 4.14.0-rc5 4.14.0-rc5
jan-v2 vanilla mel-v2
Hmean SeqCreate ops 76.20 ( 0.00%) 75.80 ( -0.53%) 76.80 ( 0.79%)
Hmean SeqCreate read 85.00 ( 0.00%) 85.00 ( 0.00%) 85.00 ( 0.00%)
Hmean SeqCreate del 13752.31 ( 0.00%) 12090.23 ( -12.09%) 15304.84 ( 11.29%)
Hmean RandCreate ops 76.00 ( 0.00%) 75.60 ( -0.53%) 77.00 ( 1.32%)
Hmean RandCreate read 96.80 ( 0.00%) 96.80 ( 0.00%) 97.00 ( 0.21%)
Hmean RandCreate del 13233.75 ( 0.00%) 11525.35 ( -12.91%) 14446.61 ( 9.16%)
Jan's series is the baseline and the vanilla kernel is 12% slower where
as this series on top gains another 11%. This is from a different
machine than the data in the changelogs but the detailed data was not
collected as there was no substantial change in v2.
This patch (of 8):
Freeing a list of pages current enables/disables IRQs for each page
freed. This patch splits freeing a list of pages into two operations --
preparing the pages for freeing and the actual freeing. This is a
tradeoff - we're taking two passes of the list to free in exchange for
avoiding multiple enable/disable of IRQs.
sparsetruncate (tiny)
4.14.0-rc4 4.14.0-rc4
janbatch-v1r1 oneirq-v1r1
Min Time 149.00 ( 0.00%) 141.00 ( 5.37%)
1st-qrtle Time 150.00 ( 0.00%) 142.00 ( 5.33%)
2nd-qrtle Time 151.00 ( 0.00%) 142.00 ( 5.96%)
3rd-qrtle Time 151.00 ( 0.00%) 143.00 ( 5.30%)
Max-90% Time 153.00 ( 0.00%) 144.00 ( 5.88%)
Max-95% Time 155.00 ( 0.00%) 147.00 ( 5.16%)
Max-99% Time 201.00 ( 0.00%) 195.00 ( 2.99%)
Max Time 236.00 ( 0.00%) 230.00 ( 2.54%)
Amean Time 152.65 ( 0.00%) 144.37 ( 5.43%)
Stddev Time 9.78 ( 0.00%) 10.44 ( -6.72%)
Coeff Time 6.41 ( 0.00%) 7.23 ( -12.84%)
Best99%Amean Time 152.07 ( 0.00%) 143.72 ( 5.50%)
Best95%Amean Time 150.75 ( 0.00%) 142.37 ( 5.56%)
Best90%Amean Time 150.59 ( 0.00%) 142.19 ( 5.58%)
Best75%Amean Time 150.36 ( 0.00%) 141.92 ( 5.61%)
Best50%Amean Time 150.04 ( 0.00%) 141.69 ( 5.56%)
Best25%Amean Time 149.85 ( 0.00%) 141.38 ( 5.65%)
With a tiny number of files, each file truncated has resident page cache
and it shows that time to truncate is roughtly 5-6% with some minor
jitter.
4.14.0-rc4 4.14.0-rc4
janbatch-v1r1 oneirq-v1r1
Hmean SeqCreate ops 65.27 ( 0.00%) 81.86 ( 25.43%)
Hmean SeqCreate read 39.48 ( 0.00%) 47.44 ( 20.16%)
Hmean SeqCreate del 24963.95 ( 0.00%) 26319.99 ( 5.43%)
Hmean RandCreate ops 65.47 ( 0.00%) 82.01 ( 25.26%)
Hmean RandCreate read 42.04 ( 0.00%) 51.75 ( 23.09%)
Hmean RandCreate del 23377.66 ( 0.00%) 23764.79 ( 1.66%)
As expected, there is a small gain for the delete operation.
[mgorman@techsingularity.net: use page_private and set_page_private helpers]
Link: http://lkml.kernel.org/r/20171018101547.mjycw7zreb66jzpa@techsingularity.net
Link: http://lkml.kernel.org/r/20171018075952.10627-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Kara <jack@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-15 17:37:37 -08:00
|
|
|
|
2024-04-05 16:32:24 +01:00
|
|
|
|
2025-02-21 18:44:24 -08:00
|
|
|
|
2024-04-05 16:32:24 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: page_alloc: remove pcppage migratetype caching
Patch series "mm: page_alloc: freelist migratetype hygiene", v4.
The page allocator's mobility grouping is intended to keep unmovable pages
separate from reclaimable/compactable ones to allow on-demand
defragmentation for higher-order allocations and huge pages.
Currently, there are several places where accidental type mixing occurs:
an allocation asks for a page of a certain migratetype and receives
another. This ruins pageblocks for compaction, which in turn makes
allocating huge pages more expensive and less reliable.
The series addresses those causes. The last patch adds type checks on all
freelist movements to prevent new violations being introduced.
The benefits can be seen in a mixed workload that stresses the machine
with a memcache-type workload and a kernel build job while periodically
attempting to allocate batches of THP. The following data is aggregated
over 50 consecutive defconfig builds:
VANILLA PATCHED
Hugealloc Time mean 165843.93 ( +0.00%) 113025.88 ( -31.85%)
Hugealloc Time stddev 158957.35 ( +0.00%) 114716.07 ( -27.83%)
Kbuild Real time 310.24 ( +0.00%) 300.73 ( -3.06%)
Kbuild User time 1271.13 ( +0.00%) 1259.42 ( -0.92%)
Kbuild System time 582.02 ( +0.00%) 559.79 ( -3.81%)
THP fault alloc 30585.14 ( +0.00%) 40853.62 ( +33.57%)
THP fault fallback 36626.46 ( +0.00%) 26357.62 ( -28.04%)
THP fault fail rate % 54.49 ( +0.00%) 39.22 ( -27.53%)
Pagealloc fallback 1328.00 ( +0.00%) 1.00 ( -99.85%)
Pagealloc type mismatch 181009.50 ( +0.00%) 0.00 ( -100.00%)
Direct compact stall 434.56 ( +0.00%) 257.66 ( -40.61%)
Direct compact fail 421.70 ( +0.00%) 249.94 ( -40.63%)
Direct compact success 12.86 ( +0.00%) 7.72 ( -37.09%)
Direct compact success rate % 2.86 ( +0.00%) 2.82 ( -0.96%)
Compact daemon scanned migrate 3370059.62 ( +0.00%) 3612054.76 ( +7.18%)
Compact daemon scanned free 7718439.20 ( +0.00%) 5386385.02 ( -30.21%)
Compact direct scanned migrate 309248.62 ( +0.00%) 176721.04 ( -42.85%)
Compact direct scanned free 433582.84 ( +0.00%) 315727.66 ( -27.18%)
Compact migrate scanned daemon % 91.20 ( +0.00%) 94.48 ( +3.56%)
Compact free scanned daemon % 94.58 ( +0.00%) 94.42 ( -0.16%)
Compact total migrate scanned 3679308.24 ( +0.00%) 3788775.80 ( +2.98%)
Compact total free scanned 8152022.04 ( +0.00%) 5702112.68 ( -30.05%)
Alloc stall 872.04 ( +0.00%) 5156.12 ( +490.71%)
Pages kswapd scanned 510645.86 ( +0.00%) 3394.94 ( -99.33%)
Pages kswapd reclaimed 134811.62 ( +0.00%) 2701.26 ( -98.00%)
Pages direct scanned 99546.06 ( +0.00%) 376407.52 ( +278.12%)
Pages direct reclaimed 62123.40 ( +0.00%) 289535.70 ( +366.06%)
Pages total scanned 610191.92 ( +0.00%) 379802.46 ( -37.76%)
Pages scanned kswapd % 76.36 ( +0.00%) 0.10 ( -98.58%)
Swap out 12057.54 ( +0.00%) 15022.98 ( +24.59%)
Swap in 209.16 ( +0.00%) 256.48 ( +22.52%)
File refaults 17701.64 ( +0.00%) 11765.40 ( -33.53%)
Huge page success rate is higher, allocation latencies are shorter and
more predictable.
Stealing (fallback) rate is drastically reduced. Notably, while the
vanilla kernel keeps doing fallbacks on an ongoing basis, the patched
kernel enters a steady state once the distribution of block types is
adequate for the workload. Steals over 50 runs:
VANILLA PATCHED
1504.0 227.0
1557.0 6.0
1391.0 13.0
1080.0 26.0
1057.0 40.0
1156.0 6.0
805.0 46.0
736.0 20.0
1747.0 2.0
1699.0 34.0
1269.0 13.0
1858.0 12.0
907.0 4.0
727.0 2.0
563.0 2.0
3094.0 2.0
10211.0 3.0
2621.0 1.0
5508.0 2.0
1060.0 2.0
538.0 3.0
5773.0 2.0
2199.0 0.0
3781.0 2.0
1387.0 1.0
4977.0 0.0
2865.0 1.0
1814.0 1.0
3739.0 1.0
6857.0 0.0
382.0 0.0
407.0 1.0
3784.0 0.0
297.0 0.0
298.0 0.0
6636.0 0.0
4188.0 0.0
242.0 0.0
9960.0 0.0
5816.0 0.0
354.0 0.0
287.0 0.0
261.0 0.0
140.0 1.0
2065.0 0.0
312.0 0.0
331.0 0.0
164.0 0.0
465.0 1.0
219.0 0.0
Type mismatches are down too. Those count every time an allocation
request asks for one migratetype and gets another. This can still occur
minimally in the patched kernel due to non-stealing fallbacks, but it's
quite rare and follows the pattern of overall fallbacks - once the block
type distribution settles, mismatches cease as well:
VANILLA: PATCHED:
182602.0 268.0
135794.0 20.0
88619.0 19.0
95973.0 0.0
129590.0 0.0
129298.0 0.0
147134.0 0.0
230854.0 0.0
239709.0 0.0
137670.0 0.0
132430.0 0.0
65712.0 0.0
57901.0 0.0
67506.0 0.0
63565.0 4.0
34806.0 0.0
42962.0 0.0
32406.0 0.0
38668.0 0.0
61356.0 0.0
57800.0 0.0
41435.0 0.0
83456.0 0.0
65048.0 0.0
28955.0 0.0
47597.0 0.0
75117.0 0.0
55564.0 0.0
38280.0 0.0
52404.0 0.0
26264.0 0.0
37538.0 0.0
19671.0 0.0
30936.0 0.0
26933.0 0.0
16962.0 0.0
44554.0 0.0
46352.0 0.0
24995.0 0.0
35152.0 0.0
12823.0 0.0
21583.0 0.0
18129.0 0.0
31693.0 0.0
28745.0 0.0
33308.0 0.0
31114.0 0.0
35034.0 0.0
12111.0 0.0
24885.0 0.0
Compaction work is markedly reduced despite much better THP rates.
In the vanilla kernel, reclaim seems to have been driven primarily by
watermark boosting that happens as a result of fallbacks. With those all
but eliminated, watermarks average lower and kswapd does less work. The
uptick in direct reclaim is because THP requests have to fend for
themselves more often - which is intended policy right now. Aggregate
reclaim activity is lowered significantly, though.
This patch (of 10):
The idea behind the cache is to save get_pageblock_migratetype() lookups
during bulk freeing. A microbenchmark suggests this isn't helping,
though. The pcp migratetype can get stale, which means that bulk freeing
has an extra branch to check if the pageblock was isolated while on the
pcp.
While the variance overlaps, the cache write and the branch seem to make
this a net negative. The following test allocates and frees batches of
10,000 pages (~3x the pcp high marks to trigger flushing):
Before:
8,668.48 msec task-clock # 99.735 CPUs utilized ( +- 2.90% )
19 context-switches # 4.341 /sec ( +- 3.24% )
0 cpu-migrations # 0.000 /sec
17,440 page-faults # 3.984 K/sec ( +- 2.90% )
41,758,692,473 cycles # 9.541 GHz ( +- 2.90% )
126,201,294,231 instructions # 5.98 insn per cycle ( +- 2.90% )
25,348,098,335 branches # 5.791 G/sec ( +- 2.90% )
33,436,921 branch-misses # 0.26% of all branches ( +- 2.90% )
0.0869148 +- 0.0000302 seconds time elapsed ( +- 0.03% )
After:
8,444.81 msec task-clock # 99.726 CPUs utilized ( +- 2.90% )
22 context-switches # 5.160 /sec ( +- 3.23% )
0 cpu-migrations # 0.000 /sec
17,443 page-faults # 4.091 K/sec ( +- 2.90% )
40,616,738,355 cycles # 9.527 GHz ( +- 2.90% )
126,383,351,792 instructions # 6.16 insn per cycle ( +- 2.90% )
25,224,985,153 branches # 5.917 G/sec ( +- 2.90% )
32,236,793 branch-misses # 0.25% of all branches ( +- 2.90% )
0.0846799 +- 0.0000412 seconds time elapsed ( +- 0.05% )
A side effect is that this also ensures that pages whose pageblock gets
stolen while on the pcplist end up on the right freelist and we don't
perform potentially type-incompatible buddy merges (or skip merges when we
shouldn't), which is likely beneficial to long-term fragmentation
management, although the effects would be harder to measure. Settle for
simpler and faster code as justification here.
Link: https://lkml.kernel.org/r/20240320180429.678181-1-hannes@cmpxchg.org
Link: https://lkml.kernel.org/r/20240320180429.678181-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Tested-by: "Huang, Ying" <ying.huang@intel.com>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-20 14:02:06 -04:00
|
|
|
|
mm, page_alloc: enable/disable IRQs once when freeing a list of pages
Patch series "Follow-up for speed up page cache truncation", v2.
This series is a follow-on for Jan Kara's series "Speed up page cache
truncation" series. We both ended up looking at the same problem but
saw different problems based on the same data. This series builds upon
his work.
A variety of workloads were compared on four separate machines but each
machine showed gains albeit at different levels. Minimally, some of the
differences are due to NUMA where truncating data from a remote node is
slower than a local node. The workloads checked were
o sparse truncate microbenchmark, tiny
o sparse truncate microbenchmark, large
o reaim-io disk workfile
o dbench4 (modified by mmtests to produce more stable results)
o filebench varmail configuration for small memory size
o bonnie, directory operations, working set size 2*RAM
reaim-io, dbench and filebench all showed minor gains. Truncation does
not dominate those workloads but were tested to ensure no other
regressions. They will not be reported further.
The sparse truncate microbench was written by Jan. It creates a number
of files and then times how long it takes to truncate each one. The
"tiny" configuraiton creates a number of files that easily fits in
memory and times how long it takes to truncate files with page cache.
The large configuration uses enough files to have data that is twice the
size of memory and so timings there include truncating page cache and
working set shadow entries in the radix tree.
Patches 1-4 are the most relevant parts of this series. Patches 5-8 are
optional as they are deleting code that is essentially useless but has a
negligible performance impact.
The changelogs have more information on performance but just for bonnie
delete options, the main comparison is
bonnie
4.14.0-rc5 4.14.0-rc5 4.14.0-rc5
jan-v2 vanilla mel-v2
Hmean SeqCreate ops 76.20 ( 0.00%) 75.80 ( -0.53%) 76.80 ( 0.79%)
Hmean SeqCreate read 85.00 ( 0.00%) 85.00 ( 0.00%) 85.00 ( 0.00%)
Hmean SeqCreate del 13752.31 ( 0.00%) 12090.23 ( -12.09%) 15304.84 ( 11.29%)
Hmean RandCreate ops 76.00 ( 0.00%) 75.60 ( -0.53%) 77.00 ( 1.32%)
Hmean RandCreate read 96.80 ( 0.00%) 96.80 ( 0.00%) 97.00 ( 0.21%)
Hmean RandCreate del 13233.75 ( 0.00%) 11525.35 ( -12.91%) 14446.61 ( 9.16%)
Jan's series is the baseline and the vanilla kernel is 12% slower where
as this series on top gains another 11%. This is from a different
machine than the data in the changelogs but the detailed data was not
collected as there was no substantial change in v2.
This patch (of 8):
Freeing a list of pages current enables/disables IRQs for each page
freed. This patch splits freeing a list of pages into two operations --
preparing the pages for freeing and the actual freeing. This is a
tradeoff - we're taking two passes of the list to free in exchange for
avoiding multiple enable/disable of IRQs.
sparsetruncate (tiny)
4.14.0-rc4 4.14.0-rc4
janbatch-v1r1 oneirq-v1r1
Min Time 149.00 ( 0.00%) 141.00 ( 5.37%)
1st-qrtle Time 150.00 ( 0.00%) 142.00 ( 5.33%)
2nd-qrtle Time 151.00 ( 0.00%) 142.00 ( 5.96%)
3rd-qrtle Time 151.00 ( 0.00%) 143.00 ( 5.30%)
Max-90% Time 153.00 ( 0.00%) 144.00 ( 5.88%)
Max-95% Time 155.00 ( 0.00%) 147.00 ( 5.16%)
Max-99% Time 201.00 ( 0.00%) 195.00 ( 2.99%)
Max Time 236.00 ( 0.00%) 230.00 ( 2.54%)
Amean Time 152.65 ( 0.00%) 144.37 ( 5.43%)
Stddev Time 9.78 ( 0.00%) 10.44 ( -6.72%)
Coeff Time 6.41 ( 0.00%) 7.23 ( -12.84%)
Best99%Amean Time 152.07 ( 0.00%) 143.72 ( 5.50%)
Best95%Amean Time 150.75 ( 0.00%) 142.37 ( 5.56%)
Best90%Amean Time 150.59 ( 0.00%) 142.19 ( 5.58%)
Best75%Amean Time 150.36 ( 0.00%) 141.92 ( 5.61%)
Best50%Amean Time 150.04 ( 0.00%) 141.69 ( 5.56%)
Best25%Amean Time 149.85 ( 0.00%) 141.38 ( 5.65%)
With a tiny number of files, each file truncated has resident page cache
and it shows that time to truncate is roughtly 5-6% with some minor
jitter.
4.14.0-rc4 4.14.0-rc4
janbatch-v1r1 oneirq-v1r1
Hmean SeqCreate ops 65.27 ( 0.00%) 81.86 ( 25.43%)
Hmean SeqCreate read 39.48 ( 0.00%) 47.44 ( 20.16%)
Hmean SeqCreate del 24963.95 ( 0.00%) 26319.99 ( 5.43%)
Hmean RandCreate ops 65.47 ( 0.00%) 82.01 ( 25.26%)
Hmean RandCreate read 42.04 ( 0.00%) 51.75 ( 23.09%)
Hmean RandCreate del 23377.66 ( 0.00%) 23764.79 ( 1.66%)
As expected, there is a small gain for the delete operation.
[mgorman@techsingularity.net: use page_private and set_page_private helpers]
Link: http://lkml.kernel.org/r/20171018101547.mjycw7zreb66jzpa@techsingularity.net
Link: http://lkml.kernel.org/r/20171018075952.10627-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Kara <jack@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-15 17:37:37 -08:00
|
|
|
|
2009-06-16 15:32:08 -07:00
|
|
|
|
page-allocator: split per-cpu list into one-list-per-migrate-type
The following two patches remove searching in the page allocator fast-path
by maintaining multiple free-lists in the per-cpu structure. At the time
the search was introduced, increasing the per-cpu structures would waste a
lot of memory as per-cpu structures were statically allocated at
compile-time. This is no longer the case.
The patches are as follows. They are based on mmotm-2009-08-27.
Patch 1 adds multiple lists to struct per_cpu_pages, one per
migratetype that can be stored on the PCP lists.
Patch 2 notes that the pcpu drain path check empty lists multiple times. The
patch reduces the number of checks by maintaining a count of free
lists encountered. Lists containing pages will then free multiple
pages in batch
The patches were tested with kernbench, netperf udp/tcp, hackbench and
sysbench. The netperf tests were not bound to any CPU in particular and
were run such that the results should be 99% confidence that the reported
results are within 1% of the estimated mean. sysbench was run with a
postgres background and read-only tests. Similar to netperf, it was run
multiple times so that it's 99% confidence results are within 1%. The
patches were tested on x86, x86-64 and ppc64 as
x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine)
kernbench - No significant difference, variance well within noise
netperf-udp - 1.34% to 2.28% gain
netperf-tcp - 0.45% to 1.22% gain
hackbench - Small variances, very close to noise
sysbench - Very small gains
x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine)
kernbench - No significant difference, variance well within noise
netperf-udp - 1.83% to 10.42% gains
netperf-tcp - No conclusive until buffer >= PAGE_SIZE
4096 +15.83%
8192 + 0.34% (not significant)
16384 + 1%
hackbench - Small gains, very close to noise
sysbench - 0.79% to 1.6% gain
ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation)
kernbench - No significant difference, variance well within noise
netperf-udp - 2-3% gain for almost all buffer sizes tested
netperf-tcp - losses on small buffers, gains on larger buffers
possibly indicates some bad caching effect.
hackbench - No significant difference
sysbench - 2-4% gain
This patch:
Currently the per-cpu page allocator searches the PCP list for pages of
the correct migrate-type to reduce the possibility of pages being
inappropriate placed from a fragmentation perspective. This search is
potentially expensive in a fast-path and undesirable. Splitting the
per-cpu list into multiple lists increases the size of a per-cpu structure
and this was potentially a major problem at the time the search was
introduced. These problem has been mitigated as now only the necessary
number of structures is allocated for the running system.
This patch replaces a list search in the per-cpu allocator with one list
per migrate type. The potential snag with this approach is when bulk
freeing pages. We round-robin free pages based on migrate type which has
little bearing on the cache hotness of the page and potentially checks
empty lists repeatedly in the event the majority of PCP pages are of one
type.
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Acked-by: Nick Piggin <npiggin@suse.de>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-21 17:03:19 -07:00
|
|
|
|
|
|
|
|
|
2021-06-28 19:42:00 -07:00
|
|
|
|
2023-09-11 14:11:08 -04:00
|
|
|
|
|
|
|
|
|
page-allocator: split per-cpu list into one-list-per-migrate-type
The following two patches remove searching in the page allocator fast-path
by maintaining multiple free-lists in the per-cpu structure. At the time
the search was introduced, increasing the per-cpu structures would waste a
lot of memory as per-cpu structures were statically allocated at
compile-time. This is no longer the case.
The patches are as follows. They are based on mmotm-2009-08-27.
Patch 1 adds multiple lists to struct per_cpu_pages, one per
migratetype that can be stored on the PCP lists.
Patch 2 notes that the pcpu drain path check empty lists multiple times. The
patch reduces the number of checks by maintaining a count of free
lists encountered. Lists containing pages will then free multiple
pages in batch
The patches were tested with kernbench, netperf udp/tcp, hackbench and
sysbench. The netperf tests were not bound to any CPU in particular and
were run such that the results should be 99% confidence that the reported
results are within 1% of the estimated mean. sysbench was run with a
postgres background and read-only tests. Similar to netperf, it was run
multiple times so that it's 99% confidence results are within 1%. The
patches were tested on x86, x86-64 and ppc64 as
x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine)
kernbench - No significant difference, variance well within noise
netperf-udp - 1.34% to 2.28% gain
netperf-tcp - 0.45% to 1.22% gain
hackbench - Small variances, very close to noise
sysbench - Very small gains
x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine)
kernbench - No significant difference, variance well within noise
netperf-udp - 1.83% to 10.42% gains
netperf-tcp - No conclusive until buffer >= PAGE_SIZE
4096 +15.83%
8192 + 0.34% (not significant)
16384 + 1%
hackbench - Small gains, very close to noise
sysbench - 0.79% to 1.6% gain
ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation)
kernbench - No significant difference, variance well within noise
netperf-udp - 2-3% gain for almost all buffer sizes tested
netperf-tcp - losses on small buffers, gains on larger buffers
possibly indicates some bad caching effect.
hackbench - No significant difference
sysbench - 2-4% gain
This patch:
Currently the per-cpu page allocator searches the PCP list for pages of
the correct migrate-type to reduce the possibility of pages being
inappropriate placed from a fragmentation perspective. This search is
potentially expensive in a fast-path and undesirable. Splitting the
per-cpu list into multiple lists increases the size of a per-cpu structure
and this was potentially a major problem at the time the search was
introduced. These problem has been mitigated as now only the necessary
number of structures is allocated for the running system.
This patch replaces a list search in the per-cpu allocator with one list
per migrate type. The potential snag with this approach is when bulk
freeing pages. We round-robin free pages based on migrate type which has
little bearing on the cache hotness of the page and potentially checks
empty lists repeatedly in the event the majority of PCP pages are of one
type.
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Acked-by: Nick Piggin <npiggin@suse.de>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-21 17:03:19 -07:00
|
|
|
|
|
|
|
|
|
2024-11-25 21:01:33 +00:00
|
|
|
|
2024-03-20 14:02:12 -04:00
|
|
|
|
2021-06-28 19:42:00 -07:00
|
|
|
|
2013-02-22 16:33:58 -08:00
|
|
|
|
2025-02-21 18:44:24 -08:00
|
|
|
|
mm, page_alloc: enable/disable IRQs once when freeing a list of pages
Patch series "Follow-up for speed up page cache truncation", v2.
This series is a follow-on for Jan Kara's series "Speed up page cache
truncation" series. We both ended up looking at the same problem but
saw different problems based on the same data. This series builds upon
his work.
A variety of workloads were compared on four separate machines but each
machine showed gains albeit at different levels. Minimally, some of the
differences are due to NUMA where truncating data from a remote node is
slower than a local node. The workloads checked were
o sparse truncate microbenchmark, tiny
o sparse truncate microbenchmark, large
o reaim-io disk workfile
o dbench4 (modified by mmtests to produce more stable results)
o filebench varmail configuration for small memory size
o bonnie, directory operations, working set size 2*RAM
reaim-io, dbench and filebench all showed minor gains. Truncation does
not dominate those workloads but were tested to ensure no other
regressions. They will not be reported further.
The sparse truncate microbench was written by Jan. It creates a number
of files and then times how long it takes to truncate each one. The
"tiny" configuraiton creates a number of files that easily fits in
memory and times how long it takes to truncate files with page cache.
The large configuration uses enough files to have data that is twice the
size of memory and so timings there include truncating page cache and
working set shadow entries in the radix tree.
Patches 1-4 are the most relevant parts of this series. Patches 5-8 are
optional as they are deleting code that is essentially useless but has a
negligible performance impact.
The changelogs have more information on performance but just for bonnie
delete options, the main comparison is
bonnie
4.14.0-rc5 4.14.0-rc5 4.14.0-rc5
jan-v2 vanilla mel-v2
Hmean SeqCreate ops 76.20 ( 0.00%) 75.80 ( -0.53%) 76.80 ( 0.79%)
Hmean SeqCreate read 85.00 ( 0.00%) 85.00 ( 0.00%) 85.00 ( 0.00%)
Hmean SeqCreate del 13752.31 ( 0.00%) 12090.23 ( -12.09%) 15304.84 ( 11.29%)
Hmean RandCreate ops 76.00 ( 0.00%) 75.60 ( -0.53%) 77.00 ( 1.32%)
Hmean RandCreate read 96.80 ( 0.00%) 96.80 ( 0.00%) 97.00 ( 0.21%)
Hmean RandCreate del 13233.75 ( 0.00%) 11525.35 ( -12.91%) 14446.61 ( 9.16%)
Jan's series is the baseline and the vanilla kernel is 12% slower where
as this series on top gains another 11%. This is from a different
machine than the data in the changelogs but the detailed data was not
collected as there was no substantial change in v2.
This patch (of 8):
Freeing a list of pages current enables/disables IRQs for each page
freed. This patch splits freeing a list of pages into two operations --
preparing the pages for freeing and the actual freeing. This is a
tradeoff - we're taking two passes of the list to free in exchange for
avoiding multiple enable/disable of IRQs.
sparsetruncate (tiny)
4.14.0-rc4 4.14.0-rc4
janbatch-v1r1 oneirq-v1r1
Min Time 149.00 ( 0.00%) 141.00 ( 5.37%)
1st-qrtle Time 150.00 ( 0.00%) 142.00 ( 5.33%)
2nd-qrtle Time 151.00 ( 0.00%) 142.00 ( 5.96%)
3rd-qrtle Time 151.00 ( 0.00%) 143.00 ( 5.30%)
Max-90% Time 153.00 ( 0.00%) 144.00 ( 5.88%)
Max-95% Time 155.00 ( 0.00%) 147.00 ( 5.16%)
Max-99% Time 201.00 ( 0.00%) 195.00 ( 2.99%)
Max Time 236.00 ( 0.00%) 230.00 ( 2.54%)
Amean Time 152.65 ( 0.00%) 144.37 ( 5.43%)
Stddev Time 9.78 ( 0.00%) 10.44 ( -6.72%)
Coeff Time 6.41 ( 0.00%) 7.23 ( -12.84%)
Best99%Amean Time 152.07 ( 0.00%) 143.72 ( 5.50%)
Best95%Amean Time 150.75 ( 0.00%) 142.37 ( 5.56%)
Best90%Amean Time 150.59 ( 0.00%) 142.19 ( 5.58%)
Best75%Amean Time 150.36 ( 0.00%) 141.92 ( 5.61%)
Best50%Amean Time 150.04 ( 0.00%) 141.69 ( 5.56%)
Best25%Amean Time 149.85 ( 0.00%) 141.38 ( 5.65%)
With a tiny number of files, each file truncated has resident page cache
and it shows that time to truncate is roughtly 5-6% with some minor
jitter.
4.14.0-rc4 4.14.0-rc4
janbatch-v1r1 oneirq-v1r1
Hmean SeqCreate ops 65.27 ( 0.00%) 81.86 ( 25.43%)
Hmean SeqCreate read 39.48 ( 0.00%) 47.44 ( 20.16%)
Hmean SeqCreate del 24963.95 ( 0.00%) 26319.99 ( 5.43%)
Hmean RandCreate ops 65.47 ( 0.00%) 82.01 ( 25.26%)
Hmean RandCreate read 42.04 ( 0.00%) 51.75 ( 23.09%)
Hmean RandCreate del 23377.66 ( 0.00%) 23764.79 ( 1.66%)
As expected, there is a small gain for the delete operation.
[mgorman@techsingularity.net: use page_private and set_page_private helpers]
Link: http://lkml.kernel.org/r/20171018101547.mjycw7zreb66jzpa@techsingularity.net
Link: http://lkml.kernel.org/r/20171018075952.10627-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Kara <jack@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-15 17:37:37 -08:00
|
|
|
|
page-allocator: split per-cpu list into one-list-per-migrate-type
The following two patches remove searching in the page allocator fast-path
by maintaining multiple free-lists in the per-cpu structure. At the time
the search was introduced, increasing the per-cpu structures would waste a
lot of memory as per-cpu structures were statically allocated at
compile-time. This is no longer the case.
The patches are as follows. They are based on mmotm-2009-08-27.
Patch 1 adds multiple lists to struct per_cpu_pages, one per
migratetype that can be stored on the PCP lists.
Patch 2 notes that the pcpu drain path check empty lists multiple times. The
patch reduces the number of checks by maintaining a count of free
lists encountered. Lists containing pages will then free multiple
pages in batch
The patches were tested with kernbench, netperf udp/tcp, hackbench and
sysbench. The netperf tests were not bound to any CPU in particular and
were run such that the results should be 99% confidence that the reported
results are within 1% of the estimated mean. sysbench was run with a
postgres background and read-only tests. Similar to netperf, it was run
multiple times so that it's 99% confidence results are within 1%. The
patches were tested on x86, x86-64 and ppc64 as
x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine)
kernbench - No significant difference, variance well within noise
netperf-udp - 1.34% to 2.28% gain
netperf-tcp - 0.45% to 1.22% gain
hackbench - Small variances, very close to noise
sysbench - Very small gains
x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine)
kernbench - No significant difference, variance well within noise
netperf-udp - 1.83% to 10.42% gains
netperf-tcp - No conclusive until buffer >= PAGE_SIZE
4096 +15.83%
8192 + 0.34% (not significant)
16384 + 1%
hackbench - Small gains, very close to noise
sysbench - 0.79% to 1.6% gain
ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation)
kernbench - No significant difference, variance well within noise
netperf-udp - 2-3% gain for almost all buffer sizes tested
netperf-tcp - losses on small buffers, gains on larger buffers
possibly indicates some bad caching effect.
hackbench - No significant difference
sysbench - 2-4% gain
This patch:
Currently the per-cpu page allocator searches the PCP list for pages of
the correct migrate-type to reduce the possibility of pages being
inappropriate placed from a fragmentation perspective. This search is
potentially expensive in a fast-path and undesirable. Splitting the
per-cpu list into multiple lists increases the size of a per-cpu structure
and this was potentially a major problem at the time the search was
introduced. These problem has been mitigated as now only the necessary
number of structures is allocated for the running system.
This patch replaces a list search in the per-cpu allocator with one list
per migrate type. The potential snag with this approach is when bulk
freeing pages. We round-robin free pages based on migrate type which has
little bearing on the cache hotness of the page and potentially checks
empty lists repeatedly in the event the majority of PCP pages are of one
type.
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Acked-by: Nick Piggin <npiggin@suse.de>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-21 17:03:19 -07:00
|
|
|
|
2024-03-20 14:02:12 -04:00
|
|
|
|
page-allocator: split per-cpu list into one-list-per-migrate-type
The following two patches remove searching in the page allocator fast-path
by maintaining multiple free-lists in the per-cpu structure. At the time
the search was introduced, increasing the per-cpu structures would waste a
lot of memory as per-cpu structures were statically allocated at
compile-time. This is no longer the case.
The patches are as follows. They are based on mmotm-2009-08-27.
Patch 1 adds multiple lists to struct per_cpu_pages, one per
migratetype that can be stored on the PCP lists.
Patch 2 notes that the pcpu drain path check empty lists multiple times. The
patch reduces the number of checks by maintaining a count of free
lists encountered. Lists containing pages will then free multiple
pages in batch
The patches were tested with kernbench, netperf udp/tcp, hackbench and
sysbench. The netperf tests were not bound to any CPU in particular and
were run such that the results should be 99% confidence that the reported
results are within 1% of the estimated mean. sysbench was run with a
postgres background and read-only tests. Similar to netperf, it was run
multiple times so that it's 99% confidence results are within 1%. The
patches were tested on x86, x86-64 and ppc64 as
x86: Intel Pentium D 3GHz with 8G RAM (no-brand machine)
kernbench - No significant difference, variance well within noise
netperf-udp - 1.34% to 2.28% gain
netperf-tcp - 0.45% to 1.22% gain
hackbench - Small variances, very close to noise
sysbench - Very small gains
x86-64: AMD Phenom 9950 1.3GHz with 8G RAM (no-brand machine)
kernbench - No significant difference, variance well within noise
netperf-udp - 1.83% to 10.42% gains
netperf-tcp - No conclusive until buffer >= PAGE_SIZE
4096 +15.83%
8192 + 0.34% (not significant)
16384 + 1%
hackbench - Small gains, very close to noise
sysbench - 0.79% to 1.6% gain
ppc64: PPC970MP 2.5GHz with 10GB RAM (it's a terrasoft powerstation)
kernbench - No significant difference, variance well within noise
netperf-udp - 2-3% gain for almost all buffer sizes tested
netperf-tcp - losses on small buffers, gains on larger buffers
possibly indicates some bad caching effect.
hackbench - No significant difference
sysbench - 2-4% gain
This patch:
Currently the per-cpu page allocator searches the PCP list for pages of
the correct migrate-type to reduce the possibility of pages being
inappropriate placed from a fragmentation perspective. This search is
potentially expensive in a fast-path and undesirable. Splitting the
per-cpu list into multiple lists increases the size of a per-cpu structure
and this was potentially a major problem at the time the search was
introduced. These problem has been mitigated as now only the necessary
number of structures is allocated for the running system.
This patch replaces a list search in the per-cpu allocator with one list
per migrate type. The potential snag with this approach is when bulk
freeing pages. We round-robin free pages based on migrate type which has
little bearing on the cache hotness of the page and potentially checks
empty lists repeatedly in the event the majority of PCP pages are of one
type.
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Acked-by: Nick Piggin <npiggin@suse.de>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-09-21 17:03:19 -07:00
|
|
|
|
|
|
|
|
|
2025-02-21 18:44:24 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-06-24 13:54:21 +01:00
|
|
|
|
2022-11-18 10:17:14 +00:00
|
|
|
|
2022-06-24 13:54:23 +01:00
|
|
|
|
2025-02-21 18:44:24 -08:00
|
|
|
|
2022-11-18 10:17:14 +00:00
|
|
|
|
2022-06-24 13:54:21 +01:00
|
|
|
|
2025-02-21 18:44:24 -08:00
|
|
|
|
2022-06-24 13:54:21 +01:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2025-02-21 18:44:24 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-01-10 15:07:04 -08:00
|
|
|
|
2024-02-27 17:42:43 +00:00
|
|
|
|
2012-01-10 15:07:04 -08:00
|
|
|
|
2024-02-27 17:42:37 +00:00
|
|
|
|
2012-01-10 15:07:04 -08:00
|
|
|
|
2022-11-18 10:17:14 +00:00
|
|
|
|
2022-06-24 13:54:21 +01:00
|
|
|
|
|
|
|
|
|
2024-03-20 14:02:07 -04:00
|
|
|
|
2012-01-10 15:07:04 -08:00
|
|
|
|
2024-02-27 17:42:37 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-02-27 17:42:36 +00:00
|
|
|
|
2024-02-27 17:42:43 +00:00
|
|
|
|
|
|
|
|
|
mm: page_alloc: remove pcppage migratetype caching
Patch series "mm: page_alloc: freelist migratetype hygiene", v4.
The page allocator's mobility grouping is intended to keep unmovable pages
separate from reclaimable/compactable ones to allow on-demand
defragmentation for higher-order allocations and huge pages.
Currently, there are several places where accidental type mixing occurs:
an allocation asks for a page of a certain migratetype and receives
another. This ruins pageblocks for compaction, which in turn makes
allocating huge pages more expensive and less reliable.
The series addresses those causes. The last patch adds type checks on all
freelist movements to prevent new violations being introduced.
The benefits can be seen in a mixed workload that stresses the machine
with a memcache-type workload and a kernel build job while periodically
attempting to allocate batches of THP. The following data is aggregated
over 50 consecutive defconfig builds:
VANILLA PATCHED
Hugealloc Time mean 165843.93 ( +0.00%) 113025.88 ( -31.85%)
Hugealloc Time stddev 158957.35 ( +0.00%) 114716.07 ( -27.83%)
Kbuild Real time 310.24 ( +0.00%) 300.73 ( -3.06%)
Kbuild User time 1271.13 ( +0.00%) 1259.42 ( -0.92%)
Kbuild System time 582.02 ( +0.00%) 559.79 ( -3.81%)
THP fault alloc 30585.14 ( +0.00%) 40853.62 ( +33.57%)
THP fault fallback 36626.46 ( +0.00%) 26357.62 ( -28.04%)
THP fault fail rate % 54.49 ( +0.00%) 39.22 ( -27.53%)
Pagealloc fallback 1328.00 ( +0.00%) 1.00 ( -99.85%)
Pagealloc type mismatch 181009.50 ( +0.00%) 0.00 ( -100.00%)
Direct compact stall 434.56 ( +0.00%) 257.66 ( -40.61%)
Direct compact fail 421.70 ( +0.00%) 249.94 ( -40.63%)
Direct compact success 12.86 ( +0.00%) 7.72 ( -37.09%)
Direct compact success rate % 2.86 ( +0.00%) 2.82 ( -0.96%)
Compact daemon scanned migrate 3370059.62 ( +0.00%) 3612054.76 ( +7.18%)
Compact daemon scanned free 7718439.20 ( +0.00%) 5386385.02 ( -30.21%)
Compact direct scanned migrate 309248.62 ( +0.00%) 176721.04 ( -42.85%)
Compact direct scanned free 433582.84 ( +0.00%) 315727.66 ( -27.18%)
Compact migrate scanned daemon % 91.20 ( +0.00%) 94.48 ( +3.56%)
Compact free scanned daemon % 94.58 ( +0.00%) 94.42 ( -0.16%)
Compact total migrate scanned 3679308.24 ( +0.00%) 3788775.80 ( +2.98%)
Compact total free scanned 8152022.04 ( +0.00%) 5702112.68 ( -30.05%)
Alloc stall 872.04 ( +0.00%) 5156.12 ( +490.71%)
Pages kswapd scanned 510645.86 ( +0.00%) 3394.94 ( -99.33%)
Pages kswapd reclaimed 134811.62 ( +0.00%) 2701.26 ( -98.00%)
Pages direct scanned 99546.06 ( +0.00%) 376407.52 ( +278.12%)
Pages direct reclaimed 62123.40 ( +0.00%) 289535.70 ( +366.06%)
Pages total scanned 610191.92 ( +0.00%) 379802.46 ( -37.76%)
Pages scanned kswapd % 76.36 ( +0.00%) 0.10 ( -98.58%)
Swap out 12057.54 ( +0.00%) 15022.98 ( +24.59%)
Swap in 209.16 ( +0.00%) 256.48 ( +22.52%)
File refaults 17701.64 ( +0.00%) 11765.40 ( -33.53%)
Huge page success rate is higher, allocation latencies are shorter and
more predictable.
Stealing (fallback) rate is drastically reduced. Notably, while the
vanilla kernel keeps doing fallbacks on an ongoing basis, the patched
kernel enters a steady state once the distribution of block types is
adequate for the workload. Steals over 50 runs:
VANILLA PATCHED
1504.0 227.0
1557.0 6.0
1391.0 13.0
1080.0 26.0
1057.0 40.0
1156.0 6.0
805.0 46.0
736.0 20.0
1747.0 2.0
1699.0 34.0
1269.0 13.0
1858.0 12.0
907.0 4.0
727.0 2.0
563.0 2.0
3094.0 2.0
10211.0 3.0
2621.0 1.0
5508.0 2.0
1060.0 2.0
538.0 3.0
5773.0 2.0
2199.0 0.0
3781.0 2.0
1387.0 1.0
4977.0 0.0
2865.0 1.0
1814.0 1.0
3739.0 1.0
6857.0 0.0
382.0 0.0
407.0 1.0
3784.0 0.0
297.0 0.0
298.0 0.0
6636.0 0.0
4188.0 0.0
242.0 0.0
9960.0 0.0
5816.0 0.0
354.0 0.0
287.0 0.0
261.0 0.0
140.0 1.0
2065.0 0.0
312.0 0.0
331.0 0.0
164.0 0.0
465.0 1.0
219.0 0.0
Type mismatches are down too. Those count every time an allocation
request asks for one migratetype and gets another. This can still occur
minimally in the patched kernel due to non-stealing fallbacks, but it's
quite rare and follows the pattern of overall fallbacks - once the block
type distribution settles, mismatches cease as well:
VANILLA: PATCHED:
182602.0 268.0
135794.0 20.0
88619.0 19.0
95973.0 0.0
129590.0 0.0
129298.0 0.0
147134.0 0.0
230854.0 0.0
239709.0 0.0
137670.0 0.0
132430.0 0.0
65712.0 0.0
57901.0 0.0
67506.0 0.0
63565.0 4.0
34806.0 0.0
42962.0 0.0
32406.0 0.0
38668.0 0.0
61356.0 0.0
57800.0 0.0
41435.0 0.0
83456.0 0.0
65048.0 0.0
28955.0 0.0
47597.0 0.0
75117.0 0.0
55564.0 0.0
38280.0 0.0
52404.0 0.0
26264.0 0.0
37538.0 0.0
19671.0 0.0
30936.0 0.0
26933.0 0.0
16962.0 0.0
44554.0 0.0
46352.0 0.0
24995.0 0.0
35152.0 0.0
12823.0 0.0
21583.0 0.0
18129.0 0.0
31693.0 0.0
28745.0 0.0
33308.0 0.0
31114.0 0.0
35034.0 0.0
12111.0 0.0
24885.0 0.0
Compaction work is markedly reduced despite much better THP rates.
In the vanilla kernel, reclaim seems to have been driven primarily by
watermark boosting that happens as a result of fallbacks. With those all
but eliminated, watermarks average lower and kswapd does less work. The
uptick in direct reclaim is because THP requests have to fend for
themselves more often - which is intended policy right now. Aggregate
reclaim activity is lowered significantly, though.
This patch (of 10):
The idea behind the cache is to save get_pageblock_migratetype() lookups
during bulk freeing. A microbenchmark suggests this isn't helping,
though. The pcp migratetype can get stale, which means that bulk freeing
has an extra branch to check if the pageblock was isolated while on the
pcp.
While the variance overlaps, the cache write and the branch seem to make
this a net negative. The following test allocates and frees batches of
10,000 pages (~3x the pcp high marks to trigger flushing):
Before:
8,668.48 msec task-clock # 99.735 CPUs utilized ( +- 2.90% )
19 context-switches # 4.341 /sec ( +- 3.24% )
0 cpu-migrations # 0.000 /sec
17,440 page-faults # 3.984 K/sec ( +- 2.90% )
41,758,692,473 cycles # 9.541 GHz ( +- 2.90% )
126,201,294,231 instructions # 5.98 insn per cycle ( +- 2.90% )
25,348,098,335 branches # 5.791 G/sec ( +- 2.90% )
33,436,921 branch-misses # 0.26% of all branches ( +- 2.90% )
0.0869148 +- 0.0000302 seconds time elapsed ( +- 0.03% )
After:
8,444.81 msec task-clock # 99.726 CPUs utilized ( +- 2.90% )
22 context-switches # 5.160 /sec ( +- 3.23% )
0 cpu-migrations # 0.000 /sec
17,443 page-faults # 4.091 K/sec ( +- 2.90% )
40,616,738,355 cycles # 9.527 GHz ( +- 2.90% )
126,383,351,792 instructions # 6.16 insn per cycle ( +- 2.90% )
25,224,985,153 branches # 5.917 G/sec ( +- 2.90% )
32,236,793 branch-misses # 0.25% of all branches ( +- 2.90% )
0.0846799 +- 0.0000412 seconds time elapsed ( +- 0.05% )
A side effect is that this also ensures that pages whose pageblock gets
stolen while on the pcplist end up on the right freelist and we don't
perform potentially type-incompatible buddy merges (or skip merges when we
shouldn't), which is likely beneficial to long-term fragmentation
management, although the effects would be harder to measure. Settle for
simpler and faster code as justification here.
Link: https://lkml.kernel.org/r/20240320180429.678181-1-hannes@cmpxchg.org
Link: https://lkml.kernel.org/r/20240320180429.678181-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Tested-by: "Huang, Ying" <ying.huang@intel.com>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-20 14:02:06 -04:00
|
|
|
|
2021-09-08 18:10:11 -07:00
|
|
|
|
2021-06-28 19:42:00 -07:00
|
|
|
|
2024-03-20 14:02:07 -04:00
|
|
|
|
|
|
|
|
|
2021-06-28 19:42:00 -07:00
|
|
|
|
2024-03-20 14:02:07 -04:00
|
|
|
|
2024-03-20 14:02:12 -04:00
|
|
|
|
|
|
|
|
|
2021-08-19 19:04:12 -07:00
|
|
|
|
2021-06-28 19:42:00 -07:00
|
|
|
|
2024-02-27 17:42:43 +00:00
|
|
|
|
2024-02-27 17:42:37 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-01-10 15:07:04 -08:00
|
|
|
|
2024-02-27 17:42:37 +00:00
|
|
|
|
mm, page_alloc: enable/disable IRQs once when freeing a list of pages
Patch series "Follow-up for speed up page cache truncation", v2.
This series is a follow-on for Jan Kara's series "Speed up page cache
truncation" series. We both ended up looking at the same problem but
saw different problems based on the same data. This series builds upon
his work.
A variety of workloads were compared on four separate machines but each
machine showed gains albeit at different levels. Minimally, some of the
differences are due to NUMA where truncating data from a remote node is
slower than a local node. The workloads checked were
o sparse truncate microbenchmark, tiny
o sparse truncate microbenchmark, large
o reaim-io disk workfile
o dbench4 (modified by mmtests to produce more stable results)
o filebench varmail configuration for small memory size
o bonnie, directory operations, working set size 2*RAM
reaim-io, dbench and filebench all showed minor gains. Truncation does
not dominate those workloads but were tested to ensure no other
regressions. They will not be reported further.
The sparse truncate microbench was written by Jan. It creates a number
of files and then times how long it takes to truncate each one. The
"tiny" configuraiton creates a number of files that easily fits in
memory and times how long it takes to truncate files with page cache.
The large configuration uses enough files to have data that is twice the
size of memory and so timings there include truncating page cache and
working set shadow entries in the radix tree.
Patches 1-4 are the most relevant parts of this series. Patches 5-8 are
optional as they are deleting code that is essentially useless but has a
negligible performance impact.
The changelogs have more information on performance but just for bonnie
delete options, the main comparison is
bonnie
4.14.0-rc5 4.14.0-rc5 4.14.0-rc5
jan-v2 vanilla mel-v2
Hmean SeqCreate ops 76.20 ( 0.00%) 75.80 ( -0.53%) 76.80 ( 0.79%)
Hmean SeqCreate read 85.00 ( 0.00%) 85.00 ( 0.00%) 85.00 ( 0.00%)
Hmean SeqCreate del 13752.31 ( 0.00%) 12090.23 ( -12.09%) 15304.84 ( 11.29%)
Hmean RandCreate ops 76.00 ( 0.00%) 75.60 ( -0.53%) 77.00 ( 1.32%)
Hmean RandCreate read 96.80 ( 0.00%) 96.80 ( 0.00%) 97.00 ( 0.21%)
Hmean RandCreate del 13233.75 ( 0.00%) 11525.35 ( -12.91%) 14446.61 ( 9.16%)
Jan's series is the baseline and the vanilla kernel is 12% slower where
as this series on top gains another 11%. This is from a different
machine than the data in the changelogs but the detailed data was not
collected as there was no substantial change in v2.
This patch (of 8):
Freeing a list of pages current enables/disables IRQs for each page
freed. This patch splits freeing a list of pages into two operations --
preparing the pages for freeing and the actual freeing. This is a
tradeoff - we're taking two passes of the list to free in exchange for
avoiding multiple enable/disable of IRQs.
sparsetruncate (tiny)
4.14.0-rc4 4.14.0-rc4
janbatch-v1r1 oneirq-v1r1
Min Time 149.00 ( 0.00%) 141.00 ( 5.37%)
1st-qrtle Time 150.00 ( 0.00%) 142.00 ( 5.33%)
2nd-qrtle Time 151.00 ( 0.00%) 142.00 ( 5.96%)
3rd-qrtle Time 151.00 ( 0.00%) 143.00 ( 5.30%)
Max-90% Time 153.00 ( 0.00%) 144.00 ( 5.88%)
Max-95% Time 155.00 ( 0.00%) 147.00 ( 5.16%)
Max-99% Time 201.00 ( 0.00%) 195.00 ( 2.99%)
Max Time 236.00 ( 0.00%) 230.00 ( 2.54%)
Amean Time 152.65 ( 0.00%) 144.37 ( 5.43%)
Stddev Time 9.78 ( 0.00%) 10.44 ( -6.72%)
Coeff Time 6.41 ( 0.00%) 7.23 ( -12.84%)
Best99%Amean Time 152.07 ( 0.00%) 143.72 ( 5.50%)
Best95%Amean Time 150.75 ( 0.00%) 142.37 ( 5.56%)
Best90%Amean Time 150.59 ( 0.00%) 142.19 ( 5.58%)
Best75%Amean Time 150.36 ( 0.00%) 141.92 ( 5.61%)
Best50%Amean Time 150.04 ( 0.00%) 141.69 ( 5.56%)
Best25%Amean Time 149.85 ( 0.00%) 141.38 ( 5.65%)
With a tiny number of files, each file truncated has resident page cache
and it shows that time to truncate is roughtly 5-6% with some minor
jitter.
4.14.0-rc4 4.14.0-rc4
janbatch-v1r1 oneirq-v1r1
Hmean SeqCreate ops 65.27 ( 0.00%) 81.86 ( 25.43%)
Hmean SeqCreate read 39.48 ( 0.00%) 47.44 ( 20.16%)
Hmean SeqCreate del 24963.95 ( 0.00%) 26319.99 ( 5.43%)
Hmean RandCreate ops 65.47 ( 0.00%) 82.01 ( 25.26%)
Hmean RandCreate read 42.04 ( 0.00%) 51.75 ( 23.09%)
Hmean RandCreate del 23377.66 ( 0.00%) 23764.79 ( 1.66%)
As expected, there is a small gain for the delete operation.
[mgorman@techsingularity.net: use page_private and set_page_private helpers]
Link: http://lkml.kernel.org/r/20171018101547.mjycw7zreb66jzpa@techsingularity.net
Link: http://lkml.kernel.org/r/20171018075952.10627-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Kara <jack@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-15 17:37:37 -08:00
|
|
|
|
2024-02-27 17:42:37 +00:00
|
|
|
|
|
|
|
|
|
2024-02-27 17:42:36 +00:00
|
|
|
|
mm: page_alloc: remove pcppage migratetype caching
Patch series "mm: page_alloc: freelist migratetype hygiene", v4.
The page allocator's mobility grouping is intended to keep unmovable pages
separate from reclaimable/compactable ones to allow on-demand
defragmentation for higher-order allocations and huge pages.
Currently, there are several places where accidental type mixing occurs:
an allocation asks for a page of a certain migratetype and receives
another. This ruins pageblocks for compaction, which in turn makes
allocating huge pages more expensive and less reliable.
The series addresses those causes. The last patch adds type checks on all
freelist movements to prevent new violations being introduced.
The benefits can be seen in a mixed workload that stresses the machine
with a memcache-type workload and a kernel build job while periodically
attempting to allocate batches of THP. The following data is aggregated
over 50 consecutive defconfig builds:
VANILLA PATCHED
Hugealloc Time mean 165843.93 ( +0.00%) 113025.88 ( -31.85%)
Hugealloc Time stddev 158957.35 ( +0.00%) 114716.07 ( -27.83%)
Kbuild Real time 310.24 ( +0.00%) 300.73 ( -3.06%)
Kbuild User time 1271.13 ( +0.00%) 1259.42 ( -0.92%)
Kbuild System time 582.02 ( +0.00%) 559.79 ( -3.81%)
THP fault alloc 30585.14 ( +0.00%) 40853.62 ( +33.57%)
THP fault fallback 36626.46 ( +0.00%) 26357.62 ( -28.04%)
THP fault fail rate % 54.49 ( +0.00%) 39.22 ( -27.53%)
Pagealloc fallback 1328.00 ( +0.00%) 1.00 ( -99.85%)
Pagealloc type mismatch 181009.50 ( +0.00%) 0.00 ( -100.00%)
Direct compact stall 434.56 ( +0.00%) 257.66 ( -40.61%)
Direct compact fail 421.70 ( +0.00%) 249.94 ( -40.63%)
Direct compact success 12.86 ( +0.00%) 7.72 ( -37.09%)
Direct compact success rate % 2.86 ( +0.00%) 2.82 ( -0.96%)
Compact daemon scanned migrate 3370059.62 ( +0.00%) 3612054.76 ( +7.18%)
Compact daemon scanned free 7718439.20 ( +0.00%) 5386385.02 ( -30.21%)
Compact direct scanned migrate 309248.62 ( +0.00%) 176721.04 ( -42.85%)
Compact direct scanned free 433582.84 ( +0.00%) 315727.66 ( -27.18%)
Compact migrate scanned daemon % 91.20 ( +0.00%) 94.48 ( +3.56%)
Compact free scanned daemon % 94.58 ( +0.00%) 94.42 ( -0.16%)
Compact total migrate scanned 3679308.24 ( +0.00%) 3788775.80 ( +2.98%)
Compact total free scanned 8152022.04 ( +0.00%) 5702112.68 ( -30.05%)
Alloc stall 872.04 ( +0.00%) 5156.12 ( +490.71%)
Pages kswapd scanned 510645.86 ( +0.00%) 3394.94 ( -99.33%)
Pages kswapd reclaimed 134811.62 ( +0.00%) 2701.26 ( -98.00%)
Pages direct scanned 99546.06 ( +0.00%) 376407.52 ( +278.12%)
Pages direct reclaimed 62123.40 ( +0.00%) 289535.70 ( +366.06%)
Pages total scanned 610191.92 ( +0.00%) 379802.46 ( -37.76%)
Pages scanned kswapd % 76.36 ( +0.00%) 0.10 ( -98.58%)
Swap out 12057.54 ( +0.00%) 15022.98 ( +24.59%)
Swap in 209.16 ( +0.00%) 256.48 ( +22.52%)
File refaults 17701.64 ( +0.00%) 11765.40 ( -33.53%)
Huge page success rate is higher, allocation latencies are shorter and
more predictable.
Stealing (fallback) rate is drastically reduced. Notably, while the
vanilla kernel keeps doing fallbacks on an ongoing basis, the patched
kernel enters a steady state once the distribution of block types is
adequate for the workload. Steals over 50 runs:
VANILLA PATCHED
1504.0 227.0
1557.0 6.0
1391.0 13.0
1080.0 26.0
1057.0 40.0
1156.0 6.0
805.0 46.0
736.0 20.0
1747.0 2.0
1699.0 34.0
1269.0 13.0
1858.0 12.0
907.0 4.0
727.0 2.0
563.0 2.0
3094.0 2.0
10211.0 3.0
2621.0 1.0
5508.0 2.0
1060.0 2.0
538.0 3.0
5773.0 2.0
2199.0 0.0
3781.0 2.0
1387.0 1.0
4977.0 0.0
2865.0 1.0
1814.0 1.0
3739.0 1.0
6857.0 0.0
382.0 0.0
407.0 1.0
3784.0 0.0
297.0 0.0
298.0 0.0
6636.0 0.0
4188.0 0.0
242.0 0.0
9960.0 0.0
5816.0 0.0
354.0 0.0
287.0 0.0
261.0 0.0
140.0 1.0
2065.0 0.0
312.0 0.0
331.0 0.0
164.0 0.0
465.0 1.0
219.0 0.0
Type mismatches are down too. Those count every time an allocation
request asks for one migratetype and gets another. This can still occur
minimally in the patched kernel due to non-stealing fallbacks, but it's
quite rare and follows the pattern of overall fallbacks - once the block
type distribution settles, mismatches cease as well:
VANILLA: PATCHED:
182602.0 268.0
135794.0 20.0
88619.0 19.0
95973.0 0.0
129590.0 0.0
129298.0 0.0
147134.0 0.0
230854.0 0.0
239709.0 0.0
137670.0 0.0
132430.0 0.0
65712.0 0.0
57901.0 0.0
67506.0 0.0
63565.0 4.0
34806.0 0.0
42962.0 0.0
32406.0 0.0
38668.0 0.0
61356.0 0.0
57800.0 0.0
41435.0 0.0
83456.0 0.0
65048.0 0.0
28955.0 0.0
47597.0 0.0
75117.0 0.0
55564.0 0.0
38280.0 0.0
52404.0 0.0
26264.0 0.0
37538.0 0.0
19671.0 0.0
30936.0 0.0
26933.0 0.0
16962.0 0.0
44554.0 0.0
46352.0 0.0
24995.0 0.0
35152.0 0.0
12823.0 0.0
21583.0 0.0
18129.0 0.0
31693.0 0.0
28745.0 0.0
33308.0 0.0
31114.0 0.0
35034.0 0.0
12111.0 0.0
24885.0 0.0
Compaction work is markedly reduced despite much better THP rates.
In the vanilla kernel, reclaim seems to have been driven primarily by
watermark boosting that happens as a result of fallbacks. With those all
but eliminated, watermarks average lower and kswapd does less work. The
uptick in direct reclaim is because THP requests have to fend for
themselves more often - which is intended policy right now. Aggregate
reclaim activity is lowered significantly, though.
This patch (of 10):
The idea behind the cache is to save get_pageblock_migratetype() lookups
during bulk freeing. A microbenchmark suggests this isn't helping,
though. The pcp migratetype can get stale, which means that bulk freeing
has an extra branch to check if the pageblock was isolated while on the
pcp.
While the variance overlaps, the cache write and the branch seem to make
this a net negative. The following test allocates and frees batches of
10,000 pages (~3x the pcp high marks to trigger flushing):
Before:
8,668.48 msec task-clock # 99.735 CPUs utilized ( +- 2.90% )
19 context-switches # 4.341 /sec ( +- 3.24% )
0 cpu-migrations # 0.000 /sec
17,440 page-faults # 3.984 K/sec ( +- 2.90% )
41,758,692,473 cycles # 9.541 GHz ( +- 2.90% )
126,201,294,231 instructions # 5.98 insn per cycle ( +- 2.90% )
25,348,098,335 branches # 5.791 G/sec ( +- 2.90% )
33,436,921 branch-misses # 0.26% of all branches ( +- 2.90% )
0.0869148 +- 0.0000302 seconds time elapsed ( +- 0.03% )
After:
8,444.81 msec task-clock # 99.726 CPUs utilized ( +- 2.90% )
22 context-switches # 5.160 /sec ( +- 3.23% )
0 cpu-migrations # 0.000 /sec
17,443 page-faults # 4.091 K/sec ( +- 2.90% )
40,616,738,355 cycles # 9.527 GHz ( +- 2.90% )
126,383,351,792 instructions # 6.16 insn per cycle ( +- 2.90% )
25,224,985,153 branches # 5.917 G/sec ( +- 2.90% )
32,236,793 branch-misses # 0.25% of all branches ( +- 2.90% )
0.0846799 +- 0.0000412 seconds time elapsed ( +- 0.05% )
A side effect is that this also ensures that pages whose pageblock gets
stolen while on the pcplist end up on the right freelist and we don't
perform potentially type-incompatible buddy merges (or skip merges when we
shouldn't), which is likely beneficial to long-term fragmentation
management, although the effects would be harder to measure. Settle for
simpler and faster code as justification here.
Link: https://lkml.kernel.org/r/20240320180429.678181-1-hannes@cmpxchg.org
Link: https://lkml.kernel.org/r/20240320180429.678181-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Tested-by: "Huang, Ying" <ying.huang@intel.com>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-20 14:02:06 -04:00
|
|
|
|
2024-02-27 17:42:43 +00:00
|
|
|
|
2024-03-20 14:02:07 -04:00
|
|
|
|
2022-06-24 13:54:21 +01:00
|
|
|
|
2024-02-27 17:42:43 +00:00
|
|
|
|
mm: page_alloc: remove pcppage migratetype caching
Patch series "mm: page_alloc: freelist migratetype hygiene", v4.
The page allocator's mobility grouping is intended to keep unmovable pages
separate from reclaimable/compactable ones to allow on-demand
defragmentation for higher-order allocations and huge pages.
Currently, there are several places where accidental type mixing occurs:
an allocation asks for a page of a certain migratetype and receives
another. This ruins pageblocks for compaction, which in turn makes
allocating huge pages more expensive and less reliable.
The series addresses those causes. The last patch adds type checks on all
freelist movements to prevent new violations being introduced.
The benefits can be seen in a mixed workload that stresses the machine
with a memcache-type workload and a kernel build job while periodically
attempting to allocate batches of THP. The following data is aggregated
over 50 consecutive defconfig builds:
VANILLA PATCHED
Hugealloc Time mean 165843.93 ( +0.00%) 113025.88 ( -31.85%)
Hugealloc Time stddev 158957.35 ( +0.00%) 114716.07 ( -27.83%)
Kbuild Real time 310.24 ( +0.00%) 300.73 ( -3.06%)
Kbuild User time 1271.13 ( +0.00%) 1259.42 ( -0.92%)
Kbuild System time 582.02 ( +0.00%) 559.79 ( -3.81%)
THP fault alloc 30585.14 ( +0.00%) 40853.62 ( +33.57%)
THP fault fallback 36626.46 ( +0.00%) 26357.62 ( -28.04%)
THP fault fail rate % 54.49 ( +0.00%) 39.22 ( -27.53%)
Pagealloc fallback 1328.00 ( +0.00%) 1.00 ( -99.85%)
Pagealloc type mismatch 181009.50 ( +0.00%) 0.00 ( -100.00%)
Direct compact stall 434.56 ( +0.00%) 257.66 ( -40.61%)
Direct compact fail 421.70 ( +0.00%) 249.94 ( -40.63%)
Direct compact success 12.86 ( +0.00%) 7.72 ( -37.09%)
Direct compact success rate % 2.86 ( +0.00%) 2.82 ( -0.96%)
Compact daemon scanned migrate 3370059.62 ( +0.00%) 3612054.76 ( +7.18%)
Compact daemon scanned free 7718439.20 ( +0.00%) 5386385.02 ( -30.21%)
Compact direct scanned migrate 309248.62 ( +0.00%) 176721.04 ( -42.85%)
Compact direct scanned free 433582.84 ( +0.00%) 315727.66 ( -27.18%)
Compact migrate scanned daemon % 91.20 ( +0.00%) 94.48 ( +3.56%)
Compact free scanned daemon % 94.58 ( +0.00%) 94.42 ( -0.16%)
Compact total migrate scanned 3679308.24 ( +0.00%) 3788775.80 ( +2.98%)
Compact total free scanned 8152022.04 ( +0.00%) 5702112.68 ( -30.05%)
Alloc stall 872.04 ( +0.00%) 5156.12 ( +490.71%)
Pages kswapd scanned 510645.86 ( +0.00%) 3394.94 ( -99.33%)
Pages kswapd reclaimed 134811.62 ( +0.00%) 2701.26 ( -98.00%)
Pages direct scanned 99546.06 ( +0.00%) 376407.52 ( +278.12%)
Pages direct reclaimed 62123.40 ( +0.00%) 289535.70 ( +366.06%)
Pages total scanned 610191.92 ( +0.00%) 379802.46 ( -37.76%)
Pages scanned kswapd % 76.36 ( +0.00%) 0.10 ( -98.58%)
Swap out 12057.54 ( +0.00%) 15022.98 ( +24.59%)
Swap in 209.16 ( +0.00%) 256.48 ( +22.52%)
File refaults 17701.64 ( +0.00%) 11765.40 ( -33.53%)
Huge page success rate is higher, allocation latencies are shorter and
more predictable.
Stealing (fallback) rate is drastically reduced. Notably, while the
vanilla kernel keeps doing fallbacks on an ongoing basis, the patched
kernel enters a steady state once the distribution of block types is
adequate for the workload. Steals over 50 runs:
VANILLA PATCHED
1504.0 227.0
1557.0 6.0
1391.0 13.0
1080.0 26.0
1057.0 40.0
1156.0 6.0
805.0 46.0
736.0 20.0
1747.0 2.0
1699.0 34.0
1269.0 13.0
1858.0 12.0
907.0 4.0
727.0 2.0
563.0 2.0
3094.0 2.0
10211.0 3.0
2621.0 1.0
5508.0 2.0
1060.0 2.0
538.0 3.0
5773.0 2.0
2199.0 0.0
3781.0 2.0
1387.0 1.0
4977.0 0.0
2865.0 1.0
1814.0 1.0
3739.0 1.0
6857.0 0.0
382.0 0.0
407.0 1.0
3784.0 0.0
297.0 0.0
298.0 0.0
6636.0 0.0
4188.0 0.0
242.0 0.0
9960.0 0.0
5816.0 0.0
354.0 0.0
287.0 0.0
261.0 0.0
140.0 1.0
2065.0 0.0
312.0 0.0
331.0 0.0
164.0 0.0
465.0 1.0
219.0 0.0
Type mismatches are down too. Those count every time an allocation
request asks for one migratetype and gets another. This can still occur
minimally in the patched kernel due to non-stealing fallbacks, but it's
quite rare and follows the pattern of overall fallbacks - once the block
type distribution settles, mismatches cease as well:
VANILLA: PATCHED:
182602.0 268.0
135794.0 20.0
88619.0 19.0
95973.0 0.0
129590.0 0.0
129298.0 0.0
147134.0 0.0
230854.0 0.0
239709.0 0.0
137670.0 0.0
132430.0 0.0
65712.0 0.0
57901.0 0.0
67506.0 0.0
63565.0 4.0
34806.0 0.0
42962.0 0.0
32406.0 0.0
38668.0 0.0
61356.0 0.0
57800.0 0.0
41435.0 0.0
83456.0 0.0
65048.0 0.0
28955.0 0.0
47597.0 0.0
75117.0 0.0
55564.0 0.0
38280.0 0.0
52404.0 0.0
26264.0 0.0
37538.0 0.0
19671.0 0.0
30936.0 0.0
26933.0 0.0
16962.0 0.0
44554.0 0.0
46352.0 0.0
24995.0 0.0
35152.0 0.0
12823.0 0.0
21583.0 0.0
18129.0 0.0
31693.0 0.0
28745.0 0.0
33308.0 0.0
31114.0 0.0
35034.0 0.0
12111.0 0.0
24885.0 0.0
Compaction work is markedly reduced despite much better THP rates.
In the vanilla kernel, reclaim seems to have been driven primarily by
watermark boosting that happens as a result of fallbacks. With those all
but eliminated, watermarks average lower and kswapd does less work. The
uptick in direct reclaim is because THP requests have to fend for
themselves more often - which is intended policy right now. Aggregate
reclaim activity is lowered significantly, though.
This patch (of 10):
The idea behind the cache is to save get_pageblock_migratetype() lookups
during bulk freeing. A microbenchmark suggests this isn't helping,
though. The pcp migratetype can get stale, which means that bulk freeing
has an extra branch to check if the pageblock was isolated while on the
pcp.
While the variance overlaps, the cache write and the branch seem to make
this a net negative. The following test allocates and frees batches of
10,000 pages (~3x the pcp high marks to trigger flushing):
Before:
8,668.48 msec task-clock # 99.735 CPUs utilized ( +- 2.90% )
19 context-switches # 4.341 /sec ( +- 3.24% )
0 cpu-migrations # 0.000 /sec
17,440 page-faults # 3.984 K/sec ( +- 2.90% )
41,758,692,473 cycles # 9.541 GHz ( +- 2.90% )
126,201,294,231 instructions # 5.98 insn per cycle ( +- 2.90% )
25,348,098,335 branches # 5.791 G/sec ( +- 2.90% )
33,436,921 branch-misses # 0.26% of all branches ( +- 2.90% )
0.0869148 +- 0.0000302 seconds time elapsed ( +- 0.03% )
After:
8,444.81 msec task-clock # 99.726 CPUs utilized ( +- 2.90% )
22 context-switches # 5.160 /sec ( +- 3.23% )
0 cpu-migrations # 0.000 /sec
17,443 page-faults # 4.091 K/sec ( +- 2.90% )
40,616,738,355 cycles # 9.527 GHz ( +- 2.90% )
126,383,351,792 instructions # 6.16 insn per cycle ( +- 2.90% )
25,224,985,153 branches # 5.917 G/sec ( +- 2.90% )
32,236,793 branch-misses # 0.25% of all branches ( +- 2.90% )
0.0846799 +- 0.0000412 seconds time elapsed ( +- 0.05% )
A side effect is that this also ensures that pages whose pageblock gets
stolen while on the pcplist end up on the right freelist and we don't
perform potentially type-incompatible buddy merges (or skip merges when we
shouldn't), which is likely beneficial to long-term fragmentation
management, although the effects would be harder to measure. Settle for
simpler and faster code as justification here.
Link: https://lkml.kernel.org/r/20240320180429.678181-1-hannes@cmpxchg.org
Link: https://lkml.kernel.org/r/20240320180429.678181-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Tested-by: "Huang, Ying" <ying.huang@intel.com>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-20 14:02:06 -04:00
|
|
|
|
2022-11-18 10:17:13 +00:00
|
|
|
|
2024-02-27 17:42:37 +00:00
|
|
|
|
2024-03-20 14:02:07 -04:00
|
|
|
|
|
|
|
|
|
2022-11-18 10:17:14 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-20 14:02:07 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-11-25 21:01:35 +00:00
|
|
|
|
2024-03-20 14:02:07 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-20 14:02:12 -04:00
|
|
|
|
2024-03-20 14:02:07 -04:00
|
|
|
|
2022-11-18 10:17:14 +00:00
|
|
|
|
2022-06-24 13:54:23 +01:00
|
|
|
|
2022-11-18 10:17:14 +00:00
|
|
|
|
2024-02-27 17:42:36 +00:00
|
|
|
|
2022-11-18 10:17:14 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: page_alloc: remove pcppage migratetype caching
Patch series "mm: page_alloc: freelist migratetype hygiene", v4.
The page allocator's mobility grouping is intended to keep unmovable pages
separate from reclaimable/compactable ones to allow on-demand
defragmentation for higher-order allocations and huge pages.
Currently, there are several places where accidental type mixing occurs:
an allocation asks for a page of a certain migratetype and receives
another. This ruins pageblocks for compaction, which in turn makes
allocating huge pages more expensive and less reliable.
The series addresses those causes. The last patch adds type checks on all
freelist movements to prevent new violations being introduced.
The benefits can be seen in a mixed workload that stresses the machine
with a memcache-type workload and a kernel build job while periodically
attempting to allocate batches of THP. The following data is aggregated
over 50 consecutive defconfig builds:
VANILLA PATCHED
Hugealloc Time mean 165843.93 ( +0.00%) 113025.88 ( -31.85%)
Hugealloc Time stddev 158957.35 ( +0.00%) 114716.07 ( -27.83%)
Kbuild Real time 310.24 ( +0.00%) 300.73 ( -3.06%)
Kbuild User time 1271.13 ( +0.00%) 1259.42 ( -0.92%)
Kbuild System time 582.02 ( +0.00%) 559.79 ( -3.81%)
THP fault alloc 30585.14 ( +0.00%) 40853.62 ( +33.57%)
THP fault fallback 36626.46 ( +0.00%) 26357.62 ( -28.04%)
THP fault fail rate % 54.49 ( +0.00%) 39.22 ( -27.53%)
Pagealloc fallback 1328.00 ( +0.00%) 1.00 ( -99.85%)
Pagealloc type mismatch 181009.50 ( +0.00%) 0.00 ( -100.00%)
Direct compact stall 434.56 ( +0.00%) 257.66 ( -40.61%)
Direct compact fail 421.70 ( +0.00%) 249.94 ( -40.63%)
Direct compact success 12.86 ( +0.00%) 7.72 ( -37.09%)
Direct compact success rate % 2.86 ( +0.00%) 2.82 ( -0.96%)
Compact daemon scanned migrate 3370059.62 ( +0.00%) 3612054.76 ( +7.18%)
Compact daemon scanned free 7718439.20 ( +0.00%) 5386385.02 ( -30.21%)
Compact direct scanned migrate 309248.62 ( +0.00%) 176721.04 ( -42.85%)
Compact direct scanned free 433582.84 ( +0.00%) 315727.66 ( -27.18%)
Compact migrate scanned daemon % 91.20 ( +0.00%) 94.48 ( +3.56%)
Compact free scanned daemon % 94.58 ( +0.00%) 94.42 ( -0.16%)
Compact total migrate scanned 3679308.24 ( +0.00%) 3788775.80 ( +2.98%)
Compact total free scanned 8152022.04 ( +0.00%) 5702112.68 ( -30.05%)
Alloc stall 872.04 ( +0.00%) 5156.12 ( +490.71%)
Pages kswapd scanned 510645.86 ( +0.00%) 3394.94 ( -99.33%)
Pages kswapd reclaimed 134811.62 ( +0.00%) 2701.26 ( -98.00%)
Pages direct scanned 99546.06 ( +0.00%) 376407.52 ( +278.12%)
Pages direct reclaimed 62123.40 ( +0.00%) 289535.70 ( +366.06%)
Pages total scanned 610191.92 ( +0.00%) 379802.46 ( -37.76%)
Pages scanned kswapd % 76.36 ( +0.00%) 0.10 ( -98.58%)
Swap out 12057.54 ( +0.00%) 15022.98 ( +24.59%)
Swap in 209.16 ( +0.00%) 256.48 ( +22.52%)
File refaults 17701.64 ( +0.00%) 11765.40 ( -33.53%)
Huge page success rate is higher, allocation latencies are shorter and
more predictable.
Stealing (fallback) rate is drastically reduced. Notably, while the
vanilla kernel keeps doing fallbacks on an ongoing basis, the patched
kernel enters a steady state once the distribution of block types is
adequate for the workload. Steals over 50 runs:
VANILLA PATCHED
1504.0 227.0
1557.0 6.0
1391.0 13.0
1080.0 26.0
1057.0 40.0
1156.0 6.0
805.0 46.0
736.0 20.0
1747.0 2.0
1699.0 34.0
1269.0 13.0
1858.0 12.0
907.0 4.0
727.0 2.0
563.0 2.0
3094.0 2.0
10211.0 3.0
2621.0 1.0
5508.0 2.0
1060.0 2.0
538.0 3.0
5773.0 2.0
2199.0 0.0
3781.0 2.0
1387.0 1.0
4977.0 0.0
2865.0 1.0
1814.0 1.0
3739.0 1.0
6857.0 0.0
382.0 0.0
407.0 1.0
3784.0 0.0
297.0 0.0
298.0 0.0
6636.0 0.0
4188.0 0.0
242.0 0.0
9960.0 0.0
5816.0 0.0
354.0 0.0
287.0 0.0
261.0 0.0
140.0 1.0
2065.0 0.0
312.0 0.0
331.0 0.0
164.0 0.0
465.0 1.0
219.0 0.0
Type mismatches are down too. Those count every time an allocation
request asks for one migratetype and gets another. This can still occur
minimally in the patched kernel due to non-stealing fallbacks, but it's
quite rare and follows the pattern of overall fallbacks - once the block
type distribution settles, mismatches cease as well:
VANILLA: PATCHED:
182602.0 268.0
135794.0 20.0
88619.0 19.0
95973.0 0.0
129590.0 0.0
129298.0 0.0
147134.0 0.0
230854.0 0.0
239709.0 0.0
137670.0 0.0
132430.0 0.0
65712.0 0.0
57901.0 0.0
67506.0 0.0
63565.0 4.0
34806.0 0.0
42962.0 0.0
32406.0 0.0
38668.0 0.0
61356.0 0.0
57800.0 0.0
41435.0 0.0
83456.0 0.0
65048.0 0.0
28955.0 0.0
47597.0 0.0
75117.0 0.0
55564.0 0.0
38280.0 0.0
52404.0 0.0
26264.0 0.0
37538.0 0.0
19671.0 0.0
30936.0 0.0
26933.0 0.0
16962.0 0.0
44554.0 0.0
46352.0 0.0
24995.0 0.0
35152.0 0.0
12823.0 0.0
21583.0 0.0
18129.0 0.0
31693.0 0.0
28745.0 0.0
33308.0 0.0
31114.0 0.0
35034.0 0.0
12111.0 0.0
24885.0 0.0
Compaction work is markedly reduced despite much better THP rates.
In the vanilla kernel, reclaim seems to have been driven primarily by
watermark boosting that happens as a result of fallbacks. With those all
but eliminated, watermarks average lower and kswapd does less work. The
uptick in direct reclaim is because THP requests have to fend for
themselves more often - which is intended policy right now. Aggregate
reclaim activity is lowered significantly, though.
This patch (of 10):
The idea behind the cache is to save get_pageblock_migratetype() lookups
during bulk freeing. A microbenchmark suggests this isn't helping,
though. The pcp migratetype can get stale, which means that bulk freeing
has an extra branch to check if the pageblock was isolated while on the
pcp.
While the variance overlaps, the cache write and the branch seem to make
this a net negative. The following test allocates and frees batches of
10,000 pages (~3x the pcp high marks to trigger flushing):
Before:
8,668.48 msec task-clock # 99.735 CPUs utilized ( +- 2.90% )
19 context-switches # 4.341 /sec ( +- 3.24% )
0 cpu-migrations # 0.000 /sec
17,440 page-faults # 3.984 K/sec ( +- 2.90% )
41,758,692,473 cycles # 9.541 GHz ( +- 2.90% )
126,201,294,231 instructions # 5.98 insn per cycle ( +- 2.90% )
25,348,098,335 branches # 5.791 G/sec ( +- 2.90% )
33,436,921 branch-misses # 0.26% of all branches ( +- 2.90% )
0.0869148 +- 0.0000302 seconds time elapsed ( +- 0.03% )
After:
8,444.81 msec task-clock # 99.726 CPUs utilized ( +- 2.90% )
22 context-switches # 5.160 /sec ( +- 3.23% )
0 cpu-migrations # 0.000 /sec
17,443 page-faults # 4.091 K/sec ( +- 2.90% )
40,616,738,355 cycles # 9.527 GHz ( +- 2.90% )
126,383,351,792 instructions # 6.16 insn per cycle ( +- 2.90% )
25,224,985,153 branches # 5.917 G/sec ( +- 2.90% )
32,236,793 branch-misses # 0.25% of all branches ( +- 2.90% )
0.0846799 +- 0.0000412 seconds time elapsed ( +- 0.05% )
A side effect is that this also ensures that pages whose pageblock gets
stolen while on the pcplist end up on the right freelist and we don't
perform potentially type-incompatible buddy merges (or skip merges when we
shouldn't), which is likely beneficial to long-term fragmentation
management, although the effects would be harder to measure. Settle for
simpler and faster code as justification here.
Link: https://lkml.kernel.org/r/20240320180429.678181-1-hannes@cmpxchg.org
Link: https://lkml.kernel.org/r/20240320180429.678181-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Tested-by: "Huang, Ying" <ying.huang@intel.com>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-20 14:02:06 -04:00
|
|
|
|
2024-03-20 14:02:12 -04:00
|
|
|
|
2022-11-18 10:17:14 +00:00
|
|
|
|
|
|
|
|
|
2022-06-24 13:54:21 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-08-19 19:04:12 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-02-27 17:42:36 +00:00
|
|
|
|
2024-11-25 21:01:35 +00:00
|
|
|
|
2025-02-21 18:44:24 -08:00
|
|
|
|
mm, page_alloc: enable/disable IRQs once when freeing a list of pages
Patch series "Follow-up for speed up page cache truncation", v2.
This series is a follow-on for Jan Kara's series "Speed up page cache
truncation" series. We both ended up looking at the same problem but
saw different problems based on the same data. This series builds upon
his work.
A variety of workloads were compared on four separate machines but each
machine showed gains albeit at different levels. Minimally, some of the
differences are due to NUMA where truncating data from a remote node is
slower than a local node. The workloads checked were
o sparse truncate microbenchmark, tiny
o sparse truncate microbenchmark, large
o reaim-io disk workfile
o dbench4 (modified by mmtests to produce more stable results)
o filebench varmail configuration for small memory size
o bonnie, directory operations, working set size 2*RAM
reaim-io, dbench and filebench all showed minor gains. Truncation does
not dominate those workloads but were tested to ensure no other
regressions. They will not be reported further.
The sparse truncate microbench was written by Jan. It creates a number
of files and then times how long it takes to truncate each one. The
"tiny" configuraiton creates a number of files that easily fits in
memory and times how long it takes to truncate files with page cache.
The large configuration uses enough files to have data that is twice the
size of memory and so timings there include truncating page cache and
working set shadow entries in the radix tree.
Patches 1-4 are the most relevant parts of this series. Patches 5-8 are
optional as they are deleting code that is essentially useless but has a
negligible performance impact.
The changelogs have more information on performance but just for bonnie
delete options, the main comparison is
bonnie
4.14.0-rc5 4.14.0-rc5 4.14.0-rc5
jan-v2 vanilla mel-v2
Hmean SeqCreate ops 76.20 ( 0.00%) 75.80 ( -0.53%) 76.80 ( 0.79%)
Hmean SeqCreate read 85.00 ( 0.00%) 85.00 ( 0.00%) 85.00 ( 0.00%)
Hmean SeqCreate del 13752.31 ( 0.00%) 12090.23 ( -12.09%) 15304.84 ( 11.29%)
Hmean RandCreate ops 76.00 ( 0.00%) 75.60 ( -0.53%) 77.00 ( 1.32%)
Hmean RandCreate read 96.80 ( 0.00%) 96.80 ( 0.00%) 97.00 ( 0.21%)
Hmean RandCreate del 13233.75 ( 0.00%) 11525.35 ( -12.91%) 14446.61 ( 9.16%)
Jan's series is the baseline and the vanilla kernel is 12% slower where
as this series on top gains another 11%. This is from a different
machine than the data in the changelogs but the detailed data was not
collected as there was no substantial change in v2.
This patch (of 8):
Freeing a list of pages current enables/disables IRQs for each page
freed. This patch splits freeing a list of pages into two operations --
preparing the pages for freeing and the actual freeing. This is a
tradeoff - we're taking two passes of the list to free in exchange for
avoiding multiple enable/disable of IRQs.
sparsetruncate (tiny)
4.14.0-rc4 4.14.0-rc4
janbatch-v1r1 oneirq-v1r1
Min Time 149.00 ( 0.00%) 141.00 ( 5.37%)
1st-qrtle Time 150.00 ( 0.00%) 142.00 ( 5.33%)
2nd-qrtle Time 151.00 ( 0.00%) 142.00 ( 5.96%)
3rd-qrtle Time 151.00 ( 0.00%) 143.00 ( 5.30%)
Max-90% Time 153.00 ( 0.00%) 144.00 ( 5.88%)
Max-95% Time 155.00 ( 0.00%) 147.00 ( 5.16%)
Max-99% Time 201.00 ( 0.00%) 195.00 ( 2.99%)
Max Time 236.00 ( 0.00%) 230.00 ( 2.54%)
Amean Time 152.65 ( 0.00%) 144.37 ( 5.43%)
Stddev Time 9.78 ( 0.00%) 10.44 ( -6.72%)
Coeff Time 6.41 ( 0.00%) 7.23 ( -12.84%)
Best99%Amean Time 152.07 ( 0.00%) 143.72 ( 5.50%)
Best95%Amean Time 150.75 ( 0.00%) 142.37 ( 5.56%)
Best90%Amean Time 150.59 ( 0.00%) 142.19 ( 5.58%)
Best75%Amean Time 150.36 ( 0.00%) 141.92 ( 5.61%)
Best50%Amean Time 150.04 ( 0.00%) 141.69 ( 5.56%)
Best25%Amean Time 149.85 ( 0.00%) 141.38 ( 5.65%)
With a tiny number of files, each file truncated has resident page cache
and it shows that time to truncate is roughtly 5-6% with some minor
jitter.
4.14.0-rc4 4.14.0-rc4
janbatch-v1r1 oneirq-v1r1
Hmean SeqCreate ops 65.27 ( 0.00%) 81.86 ( 25.43%)
Hmean SeqCreate read 39.48 ( 0.00%) 47.44 ( 20.16%)
Hmean SeqCreate del 24963.95 ( 0.00%) 26319.99 ( 5.43%)
Hmean RandCreate ops 65.47 ( 0.00%) 82.01 ( 25.26%)
Hmean RandCreate read 42.04 ( 0.00%) 51.75 ( 23.09%)
Hmean RandCreate del 23377.66 ( 0.00%) 23764.79 ( 1.66%)
As expected, there is a small gain for the delete operation.
[mgorman@techsingularity.net: use page_private and set_page_private helpers]
Link: http://lkml.kernel.org/r/20171018101547.mjycw7zreb66jzpa@techsingularity.net
Link: http://lkml.kernel.org/r/20171018075952.10627-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Kara <jack@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-15 17:37:37 -08:00
|
|
|
|
2022-06-24 13:54:21 +01:00
|
|
|
|
2022-11-18 10:17:14 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-02-27 17:42:37 +00:00
|
|
|
|
2012-01-10 15:07:04 -08:00
|
|
|
|
|
|
|
|
|
2006-03-22 00:08:05 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-01-23 15:52:54 -08:00
|
|
|
|
|
|
|
|
|
2008-11-25 16:55:53 +01:00
|
|
|
|
2016-07-26 15:23:49 -07:00
|
|
|
|
2006-03-22 00:08:40 -08:00
|
|
|
|
2024-02-26 15:55:32 -05:00
|
|
|
|
2024-09-05 22:21:07 -06:00
|
|
|
|
2025-03-14 13:36:12 +00:00
|
|
|
|
2006-03-22 00:08:05 -08:00
|
|
|
|
2013-03-25 15:47:38 -07:00
|
|
|
|
2006-03-22 00:08:05 -08:00
|
|
|
|
2014-11-13 15:19:21 -08:00
|
|
|
|
2010-05-24 14:32:27 -07:00
|
|
|
|
2022-09-01 09:50:43 +08:00
|
|
|
|
|
|
|
|
|
2010-05-24 14:32:27 -07:00
|
|
|
|
2013-02-22 16:33:58 -08:00
|
|
|
|
2022-09-01 09:50:43 +08:00
|
|
|
|
2016-10-07 16:58:00 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-03-05 15:44:50 -08:00
|
|
|
|
2018-05-23 10:18:21 +09:00
|
|
|
|
2012-12-11 16:02:57 -08:00
|
|
|
|
|
|
|
|
|
2010-05-24 14:32:27 -07:00
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
2012-10-08 16:32:00 -07:00
|
|
|
|
2016-07-28 15:45:07 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-05-24 14:32:27 -07:00
|
|
|
|
|
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
|
|
|
|
|
2022-03-22 14:43:05 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-20 14:02:13 -04:00
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
2024-03-20 14:02:13 -04:00
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
2010-05-24 14:32:27 -07:00
|
|
|
|
|
|
|
|
|
2013-01-11 14:32:16 -08:00
|
|
|
|
2012-10-08 16:29:12 -07:00
|
|
|
|
|
|
|
|
|
2020-04-06 20:04:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-04-10 14:32:29 -07:00
|
|
|
|
2020-04-06 20:04:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/page_alloc: convert "report" flag of __free_one_page() to a proper flag
Patch series "mm: place pages to the freelist tail when onlining and undoing isolation", v2.
When adding separate memory blocks via add_memory*() and onlining them
immediately, the metadata (especially the memmap) of the next block will
be placed onto one of the just added+onlined block. This creates a chain
of unmovable allocations: If the last memory block cannot get
offlined+removed() so will all dependent ones. We directly have unmovable
allocations all over the place.
This can be observed quite easily using virtio-mem, however, it can also
be observed when using DIMMs. The freshly onlined pages will usually be
placed to the head of the freelists, meaning they will be allocated next,
turning the just-added memory usually immediately un-removable. The fresh
pages are cold, prefering to allocate others (that might be hot) also
feels to be the natural thing to do.
It also applies to the hyper-v balloon xen-balloon, and ppc64 dlpar: when
adding separate, successive memory blocks, each memory block will have
unmovable allocations on them - for example gigantic pages will fail to
allocate.
While the ZONE_NORMAL doesn't provide any guarantees that memory can get
offlined+removed again (any kind of fragmentation with unmovable
allocations is possible), there are many scenarios (hotplugging a lot of
memory, running workload, hotunplug some memory/as much as possible) where
we can offline+remove quite a lot with this patchset.
a) To visualize the problem, a very simple example:
Start a VM with 4GB and 8GB of virtio-mem memory:
[root@localhost ~]# lsmem
RANGE SIZE STATE REMOVABLE BLOCK
0x0000000000000000-0x00000000bfffffff 3G online yes 0-23
0x0000000100000000-0x000000033fffffff 9G online yes 32-103
Memory block size: 128M
Total online memory: 12G
Total offline memory: 0B
Then try to unplug as much as possible using virtio-mem. Observe which
memory blocks are still around. Without this patch set:
[root@localhost ~]# lsmem
RANGE SIZE STATE REMOVABLE BLOCK
0x0000000000000000-0x00000000bfffffff 3G online yes 0-23
0x0000000100000000-0x000000013fffffff 1G online yes 32-39
0x0000000148000000-0x000000014fffffff 128M online yes 41
0x0000000158000000-0x000000015fffffff 128M online yes 43
0x0000000168000000-0x000000016fffffff 128M online yes 45
0x0000000178000000-0x000000017fffffff 128M online yes 47
0x0000000188000000-0x0000000197ffffff 256M online yes 49-50
0x00000001a0000000-0x00000001a7ffffff 128M online yes 52
0x00000001b0000000-0x00000001b7ffffff 128M online yes 54
0x00000001c0000000-0x00000001c7ffffff 128M online yes 56
0x00000001d0000000-0x00000001d7ffffff 128M online yes 58
0x00000001e0000000-0x00000001e7ffffff 128M online yes 60
0x00000001f0000000-0x00000001f7ffffff 128M online yes 62
0x0000000200000000-0x0000000207ffffff 128M online yes 64
0x0000000210000000-0x0000000217ffffff 128M online yes 66
0x0000000220000000-0x0000000227ffffff 128M online yes 68
0x0000000230000000-0x0000000237ffffff 128M online yes 70
0x0000000240000000-0x0000000247ffffff 128M online yes 72
0x0000000250000000-0x0000000257ffffff 128M online yes 74
0x0000000260000000-0x0000000267ffffff 128M online yes 76
0x0000000270000000-0x0000000277ffffff 128M online yes 78
0x0000000280000000-0x0000000287ffffff 128M online yes 80
0x0000000290000000-0x0000000297ffffff 128M online yes 82
0x00000002a0000000-0x00000002a7ffffff 128M online yes 84
0x00000002b0000000-0x00000002b7ffffff 128M online yes 86
0x00000002c0000000-0x00000002c7ffffff 128M online yes 88
0x00000002d0000000-0x00000002d7ffffff 128M online yes 90
0x00000002e0000000-0x00000002e7ffffff 128M online yes 92
0x00000002f0000000-0x00000002f7ffffff 128M online yes 94
0x0000000300000000-0x0000000307ffffff 128M online yes 96
0x0000000310000000-0x0000000317ffffff 128M online yes 98
0x0000000320000000-0x0000000327ffffff 128M online yes 100
0x0000000330000000-0x000000033fffffff 256M online yes 102-103
Memory block size: 128M
Total online memory: 8.1G
Total offline memory: 0B
With this patch set:
[root@localhost ~]# lsmem
RANGE SIZE STATE REMOVABLE BLOCK
0x0000000000000000-0x00000000bfffffff 3G online yes 0-23
0x0000000100000000-0x000000013fffffff 1G online yes 32-39
Memory block size: 128M
Total online memory: 4G
Total offline memory: 0B
All memory can get unplugged, all memory block can get removed. Of
course, no workload ran and the system was basically idle, but it
highlights the issue - the fairly deterministic chain of unmovable
allocations. When a huge page for the 2MB memmap is needed, a
just-onlined 4MB page will be split. The remaining 2MB page will be used
for the memmap of the next memory block. So one memory block will hold
the memmap of the two following memory blocks. Finally the pages of the
last-onlined memory block will get used for the next bigger allocations -
if any allocation is unmovable, all dependent memory blocks cannot get
unplugged and removed until that allocation is gone.
Note that with bigger memory blocks (e.g., 256MB), *all* memory
blocks are dependent and none can get unplugged again!
b) Experiment with memory intensive workload
I performed an experiment with an older version of this patch set (before
we used undo_isolate_page_range() in online_pages(): Hotplug 56GB to a VM
with an initial 4GB, onlining all memory to ZONE_NORMAL right from the
kernel when adding it. I then run various memory intensive workloads that
consume most system memory for a total of 45 minutes. Once finished, I
try to unplug as much memory as possible.
With this change, I am able to remove via virtio-mem (adding individual
128MB memory blocks) 413 out of 448 added memory blocks. Via individual
(256MB) DIMMs 380 out of 448 added memory blocks. (I don't have any
numbers without this patchset, but looking at the above example, it's at
most half of the 448 memory blocks for virtio-mem, and most probably none
for DIMMs).
Again, there are workloads that might behave very differently due to the
nature of ZONE_NORMAL.
This change also affects (besides memory onlining):
- Other users of undo_isolate_page_range(): Pages are always placed to the
tail.
-- When memory offlining fails
-- When memory isolation fails after having isolated some pageblocks
-- When alloc_contig_range() either succeeds or fails
- Other users of __putback_isolated_page(): Pages are always placed to the
tail.
-- Free page reporting
- Other users of __free_pages_core()
-- AFAIKs, any memory that is getting exposed to the buddy during boot.
IIUC we will now usually allocate memory from lower addresses within
a zone first (especially during boot).
- Other users of generic_online_page()
-- Hyper-V balloon
This patch (of 5):
Let's prepare for additional flags and avoid long parameter lists of
bools. Follow-up patches will also make use of the flags in
__free_pages_ok().
Signed-off-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Alexander Duyck <alexander.h.duyck@linux.intel.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Wei Yang <richard.weiyang@linux.alibaba.com>
Reviewed-by: Pankaj Gupta <pankaj.gupta.linux@gmail.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Haiyang Zhang <haiyangz@microsoft.com>
Cc: "K. Y. Srinivasan" <kys@microsoft.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Scott Cheloha <cheloha@linux.ibm.com>
Cc: Stephen Hemminger <sthemmin@microsoft.com>
Cc: Wei Liu <wei.liu@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Link: https://lkml.kernel.org/r/20201005121534.15649-1-david@redhat.com
Link: https://lkml.kernel.org/r/20201005121534.15649-2-david@redhat.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-10-15 20:09:20 -07:00
|
|
|
|
mm/page_alloc: place pages to tail in __putback_isolated_page()
__putback_isolated_page() already documents that pages will be placed to
the tail of the freelist - this is, however, not the case for "order >=
MAX_ORDER - 2" (see buddy_merge_likely()) - which should be the case for
all existing users.
This change affects two users:
- free page reporting
- page isolation, when undoing the isolation (including memory onlining).
This behavior is desirable for pages that haven't really been touched
lately, so exactly the two users that don't actually read/write page
content, but rather move untouched pages.
The new behavior is especially desirable for memory onlining, where we
allow allocation of newly onlined pages via undo_isolate_page_range() in
online_pages(). Right now, we always place them to the head of the
freelist, resulting in undesireable behavior: Assume we add individual
memory chunks via add_memory() and online them right away to the NORMAL
zone. We create a dependency chain of unmovable allocations e.g., via the
memmap. The memmap of the next chunk will be placed onto previous chunks
- if the last block cannot get offlined+removed, all dependent ones cannot
get offlined+removed. While this can already be observed with individual
DIMMs, it's more of an issue for virtio-mem (and I suspect also ppc
DLPAR).
Document that this should only be used for optimizations, and no code
should rely on this behavior for correction (if the order of the freelists
ever changes).
We won't care about page shuffling: memory onlining already properly
shuffles after onlining. free page reporting doesn't care about
physically contiguous ranges, and there are already cases where page
isolation will simply move (physically close) free pages to (currently)
the head of the freelists via move_freepages_block() instead of shuffling.
If this becomes ever relevant, we should shuffle the whole zone when
undoing isolation of larger ranges, and after free_contig_range().
Signed-off-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Alexander Duyck <alexander.h.duyck@linux.intel.com>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Wei Yang <richard.weiyang@linux.alibaba.com>
Reviewed-by: Pankaj Gupta <pankaj.gupta.linux@gmail.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Scott Cheloha <cheloha@linux.ibm.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Haiyang Zhang <haiyangz@microsoft.com>
Cc: "K. Y. Srinivasan" <kys@microsoft.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Stephen Hemminger <sthemmin@microsoft.com>
Cc: Wei Liu <wei.liu@kernel.org>
Link: https://lkml.kernel.org/r/20201005121534.15649-3-david@redhat.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-10-15 20:09:26 -07:00
|
|
|
|
2020-04-06 20:04:53 -07:00
|
|
|
|
|
|
|
|
|
2016-05-19 17:13:27 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-28 19:41:50 -07:00
|
|
|
|
|
|
|
|
|
2016-05-19 17:13:27 -07:00
|
|
|
|
|
|
|
|
|
mm: change the call sites of numa statistics items
Patch series "Separate NUMA statistics from zone statistics", v2.
Each page allocation updates a set of per-zone statistics with a call to
zone_statistics(). As discussed in 2017 MM summit, these are a
substantial source of overhead in the page allocator and are very rarely
consumed. This significant overhead in cache bouncing caused by zone
counters (NUMA associated counters) update in parallel in multi-threaded
page allocation (pointed out by Dave Hansen).
A link to the MM summit slides:
http://people.netfilter.org/hawk/presentations/MM-summit2017/MM-summit2017-JesperBrouer.pdf
To mitigate this overhead, this patchset separates NUMA statistics from
zone statistics framework, and update NUMA counter threshold to a fixed
size of MAX_U16 - 2, as a small threshold greatly increases the update
frequency of the global counter from local per cpu counter (suggested by
Ying Huang). The rationality is that these statistics counters don't
need to be read often, unlike other VM counters, so it's not a problem
to use a large threshold and make readers more expensive.
With this patchset, we see 31.3% drop of CPU cycles(537-->369, see
below) for per single page allocation and reclaim on Jesper's
page_bench03 benchmark. Meanwhile, this patchset keeps the same style
of virtual memory statistics with little end-user-visible effects (only
move the numa stats to show behind zone page stats, see the first patch
for details).
I did an experiment of single page allocation and reclaim concurrently
using Jesper's page_bench03 benchmark on a 2-Socket Broadwell-based
server (88 processors with 126G memory) with different size of threshold
of pcp counter.
Benchmark provided by Jesper D Brouer(increase loop times to 10000000):
https://github.com/netoptimizer/prototype-kernel/tree/master/kernel/mm/bench
Threshold CPU cycles Throughput(88 threads)
32 799 241760478
64 640 301628829
125 537 358906028 <==> system by default
256 468 412397590
512 428 450550704
4096 399 482520943
20000 394 489009617
30000 395 488017817
65533 369(-31.3%) 521661345(+45.3%) <==> with this patchset
N/A 342(-36.3%) 562900157(+56.8%) <==> disable zone_statistics
This patch (of 3):
In this patch, NUMA statistics is separated from zone statistics
framework, all the call sites of NUMA stats are changed to use
numa-stats-specific functions, it does not have any functionality change
except that the number of NUMA stats is shown behind zone page stats
when users *read* the zone info.
E.g. cat /proc/zoneinfo
***Base*** ***With this patch***
nr_free_pages 3976 nr_free_pages 3976
nr_zone_inactive_anon 0 nr_zone_inactive_anon 0
nr_zone_active_anon 0 nr_zone_active_anon 0
nr_zone_inactive_file 0 nr_zone_inactive_file 0
nr_zone_active_file 0 nr_zone_active_file 0
nr_zone_unevictable 0 nr_zone_unevictable 0
nr_zone_write_pending 0 nr_zone_write_pending 0
nr_mlock 0 nr_mlock 0
nr_page_table_pages 0 nr_page_table_pages 0
nr_kernel_stack 0 nr_kernel_stack 0
nr_bounce 0 nr_bounce 0
nr_zspages 0 nr_zspages 0
numa_hit 0 *nr_free_cma 0*
numa_miss 0 numa_hit 0
numa_foreign 0 numa_miss 0
numa_interleave 0 numa_foreign 0
numa_local 0 numa_interleave 0
numa_other 0 numa_local 0
*nr_free_cma 0* numa_other 0
... ...
vm stats threshold: 10 vm stats threshold: 10
... ...
The next patch updates the numa stats counter size and threshold.
[akpm@linux-foundation.org: coding-style fixes]
Link: http://lkml.kernel.org/r/1503568801-21305-2-git-send-email-kemi.wang@intel.com
Signed-off-by: Kemi Wang <kemi.wang@intel.com>
Reported-by: Jesper Dangaard Brouer <brouer@redhat.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Christopher Lameter <cl@linux.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Andi Kleen <andi.kleen@intel.com>
Cc: Ying Huang <ying.huang@intel.com>
Cc: Aaron Lu <aaron.lu@intel.com>
Cc: Tim Chen <tim.c.chen@intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-08 16:12:48 -07:00
|
|
|
|
2016-05-19 17:13:27 -07:00
|
|
|
|
2017-11-15 17:38:22 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-08-21 21:53:32 -07:00
|
|
|
|
2016-05-19 17:13:27 -07:00
|
|
|
|
|
|
|
|
|
2018-08-21 21:53:32 -07:00
|
|
|
|
2021-06-28 19:41:50 -07:00
|
|
|
|
2017-01-10 16:57:39 -08:00
|
|
|
|
2021-06-28 19:41:50 -07:00
|
|
|
|
|
|
|
|
|
2016-05-19 17:13:27 -07:00
|
|
|
|
2021-06-28 19:41:50 -07:00
|
|
|
|
2016-05-19 17:13:27 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-06-24 13:54:19 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-03-30 17:28:09 -07:00
|
|
|
|
|
|
|
|
|
2025-02-21 18:44:23 -08:00
|
|
|
|
2025-03-30 17:28:09 -07:00
|
|
|
|
2025-02-21 18:44:23 -08:00
|
|
|
|
|
|
|
|
|
2023-01-13 11:12:14 +00:00
|
|
|
|
2022-06-24 13:54:19 +01:00
|
|
|
|
|
|
|
|
|
mm: page_alloc: speed up fallbacks in rmqueue_bulk()
The test robot identified c2f6ea38fc1b ("mm: page_alloc: don't steal
single pages from biggest buddy") as the root cause of a 56.4% regression
in vm-scalability::lru-file-mmap-read.
Carlos reports an earlier patch, c0cd6f557b90 ("mm: page_alloc: fix
freelist movement during block conversion"), as the root cause for a
regression in worst-case zone->lock+irqoff hold times.
Both of these patches modify the page allocator's fallback path to be less
greedy in an effort to stave off fragmentation. The flip side of this is
that fallbacks are also less productive each time around, which means the
fallback search can run much more frequently.
Carlos' traces point to rmqueue_bulk() specifically, which tries to refill
the percpu cache by allocating a large batch of pages in a loop. It
highlights how once the native freelists are exhausted, the fallback code
first scans orders top-down for whole blocks to claim, then falls back to
a bottom-up search for the smallest buddy to steal. For the next batch
page, it goes through the same thing again.
This can be made more efficient. Since rmqueue_bulk() holds the
zone->lock over the entire batch, the freelists are not subject to outside
changes; when the search for a block to claim has already failed, there is
no point in trying again for the next page.
Modify __rmqueue() to remember the last successful fallback mode, and
restart directly from there on the next rmqueue_bulk() iteration.
Oliver confirms that this improves beyond the regression that the test
robot reported against c2f6ea38fc1b:
commit:
f3b92176f4 ("tools/selftests: add guard region test for /proc/$pid/pagemap")
c2f6ea38fc ("mm: page_alloc: don't steal single pages from biggest buddy")
acc4d5ff0b ("Merge tag 'net-6.15-rc0' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net")
2c847f27c3 ("mm: page_alloc: speed up fallbacks in rmqueue_bulk()") <--- your patch
f3b92176f4f7100f c2f6ea38fc1b640aa7a2e155cc1 acc4d5ff0b61eb1715c498b6536 2c847f27c37da65a93d23c237c5
---------------- --------------------------- --------------------------- ---------------------------
%stddev %change %stddev %change %stddev %change %stddev
\ | \ | \ | \
25525364 ± 3% -56.4% 11135467 -57.8% 10779336 +31.6% 33581409 vm-scalability.throughput
Carlos confirms that worst-case times are almost fully recovered
compared to before the earlier culprit patch:
2dd482ba627d (before freelist hygiene): 1ms
c0cd6f557b90 (after freelist hygiene): 90ms
next-20250319 (steal smallest buddy): 280ms
this patch : 8ms
[jackmanb@google.com: comment updates]
Link: https://lkml.kernel.org/r/D92AC0P9594X.3BML64MUKTF8Z@google.com
[hannes@cmpxchg.org: reset rmqueue_mode in rmqueue_buddy() error loop, per Yunsheng Lin]
Link: https://lkml.kernel.org/r/20250409140023.GA2313@cmpxchg.org
Link: https://lkml.kernel.org/r/20250407180154.63348-1-hannes@cmpxchg.org
Fixes: c0cd6f557b90 ("mm: page_alloc: fix freelist movement during block conversion")
Fixes: c2f6ea38fc1b ("mm: page_alloc: don't steal single pages from biggest buddy")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Brendan Jackman <jackmanb@google.com>
Reported-by: kernel test robot <oliver.sang@intel.com>
Reported-by: Carlos Song <carlos.song@nxp.com>
Tested-by: Carlos Song <carlos.song@nxp.com>
Tested-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202503271547.fc08b188-lkp@intel.com
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Tested-by: Shivank Garg <shivankg@amd.com>
Acked-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: <stable@vger.kernel.org> [6.10+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-07 14:01:53 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-01-13 11:12:14 +00:00
|
|
|
|
|
|
|
|
|
2024-10-11 13:07:37 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-01-13 11:12:14 +00:00
|
|
|
|
2024-10-11 13:07:37 +01:00
|
|
|
|
2023-01-13 11:12:14 +00:00
|
|
|
|
|
|
|
|
|
2022-06-24 13:54:19 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: tune PCP high automatically
The target to tune PCP high automatically is as follows,
- Minimize allocation/freeing from/to shared zone
- Minimize idle pages in PCP
- Minimize pages in PCP if the system free pages is too few
To reach these target, a tuning algorithm as follows is designed,
- When we refill PCP via allocating from the zone, increase PCP high.
Because if we had larger PCP, we could avoid to allocate from the
zone.
- In periodic vmstat updating kworker (via refresh_cpu_vm_stats()),
decrease PCP high to try to free possible idle PCP pages.
- When page reclaiming is active for the zone, stop increasing PCP
high in allocating path, decrease PCP high and free some pages in
freeing path.
So, the PCP high can be tuned to the page allocating/freeing depth of
workloads eventually.
One issue of the algorithm is that if the number of pages allocated is
much more than that of pages freed on a CPU, the PCP high may become the
maximal value even if the allocating/freeing depth is small. But this
isn't a severe issue, because there are no idle pages in this case.
One alternative choice is to increase PCP high when we drain PCP via
trying to free pages to the zone, but don't increase PCP high during PCP
refilling. This can avoid the issue above. But if the number of pages
allocated is much less than that of pages freed on a CPU, there will be
many idle pages in PCP and it is hard to free these idle pages.
1/8 (>> 3) of PCP high will be decreased periodically. The value 1/8 is
kind of arbitrary. Just to make sure that the idle PCP pages will be
freed eventually.
On a 2-socket Intel server with 224 logical CPU, we run 8 kbuild instances
in parallel (each with `make -j 28`) in 8 cgroup. This simulates the
kbuild server that is used by 0-Day kbuild service. With the patch, the
build time decreases 3.5%. The cycles% of the spinlock contention (mostly
for zone lock) decreases from 11.0% to 0.5%. The number of PCP draining
for high order pages freeing (free_high) decreases 65.6%. The number of
pages allocated from zone (instead of from PCP) decreases 83.9%.
Link: https://lkml.kernel.org/r/20231016053002.756205-8-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Suggested-by: Mel Gorman <mgorman@techsingularity.net>
Suggested-by: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:00 +08:00
|
|
|
|
2023-10-16 13:29:58 +08:00
|
|
|
|
mm: tune PCP high automatically
The target to tune PCP high automatically is as follows,
- Minimize allocation/freeing from/to shared zone
- Minimize idle pages in PCP
- Minimize pages in PCP if the system free pages is too few
To reach these target, a tuning algorithm as follows is designed,
- When we refill PCP via allocating from the zone, increase PCP high.
Because if we had larger PCP, we could avoid to allocate from the
zone.
- In periodic vmstat updating kworker (via refresh_cpu_vm_stats()),
decrease PCP high to try to free possible idle PCP pages.
- When page reclaiming is active for the zone, stop increasing PCP
high in allocating path, decrease PCP high and free some pages in
freeing path.
So, the PCP high can be tuned to the page allocating/freeing depth of
workloads eventually.
One issue of the algorithm is that if the number of pages allocated is
much more than that of pages freed on a CPU, the PCP high may become the
maximal value even if the allocating/freeing depth is small. But this
isn't a severe issue, because there are no idle pages in this case.
One alternative choice is to increase PCP high when we drain PCP via
trying to free pages to the zone, but don't increase PCP high during PCP
refilling. This can avoid the issue above. But if the number of pages
allocated is much less than that of pages freed on a CPU, there will be
many idle pages in PCP and it is hard to free these idle pages.
1/8 (>> 3) of PCP high will be decreased periodically. The value 1/8 is
kind of arbitrary. Just to make sure that the idle PCP pages will be
freed eventually.
On a 2-socket Intel server with 224 logical CPU, we run 8 kbuild instances
in parallel (each with `make -j 28`) in 8 cgroup. This simulates the
kbuild server that is used by 0-Day kbuild service. With the patch, the
build time decreases 3.5%. The cycles% of the spinlock contention (mostly
for zone lock) decreases from 11.0% to 0.5%. The number of PCP draining
for high order pages freeing (free_high) decreases 65.6%. The number of
pages allocated from zone (instead of from PCP) decreases 83.9%.
Link: https://lkml.kernel.org/r/20231016053002.756205-8-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Suggested-by: Mel Gorman <mgorman@techsingularity.net>
Suggested-by: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:00 +08:00
|
|
|
|
|
|
|
|
|
2023-10-16 13:29:58 +08:00
|
|
|
|
mm: tune PCP high automatically
The target to tune PCP high automatically is as follows,
- Minimize allocation/freeing from/to shared zone
- Minimize idle pages in PCP
- Minimize pages in PCP if the system free pages is too few
To reach these target, a tuning algorithm as follows is designed,
- When we refill PCP via allocating from the zone, increase PCP high.
Because if we had larger PCP, we could avoid to allocate from the
zone.
- In periodic vmstat updating kworker (via refresh_cpu_vm_stats()),
decrease PCP high to try to free possible idle PCP pages.
- When page reclaiming is active for the zone, stop increasing PCP
high in allocating path, decrease PCP high and free some pages in
freeing path.
So, the PCP high can be tuned to the page allocating/freeing depth of
workloads eventually.
One issue of the algorithm is that if the number of pages allocated is
much more than that of pages freed on a CPU, the PCP high may become the
maximal value even if the allocating/freeing depth is small. But this
isn't a severe issue, because there are no idle pages in this case.
One alternative choice is to increase PCP high when we drain PCP via
trying to free pages to the zone, but don't increase PCP high during PCP
refilling. This can avoid the issue above. But if the number of pages
allocated is much less than that of pages freed on a CPU, there will be
many idle pages in PCP and it is hard to free these idle pages.
1/8 (>> 3) of PCP high will be decreased periodically. The value 1/8 is
kind of arbitrary. Just to make sure that the idle PCP pages will be
freed eventually.
On a 2-socket Intel server with 224 logical CPU, we run 8 kbuild instances
in parallel (each with `make -j 28`) in 8 cgroup. This simulates the
kbuild server that is used by 0-Day kbuild service. With the patch, the
build time decreases 3.5%. The cycles% of the spinlock contention (mostly
for zone lock) decreases from 11.0% to 0.5%. The number of PCP draining
for high order pages freeing (free_high) decreases 65.6%. The number of
pages allocated from zone (instead of from PCP) decreases 83.9%.
Link: https://lkml.kernel.org/r/20231016053002.756205-8-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Suggested-by: Mel Gorman <mgorman@techsingularity.net>
Suggested-by: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:00 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-16 13:29:58 +08:00
|
|
|
|
|
|
|
|
|
mm: tune PCP high automatically
The target to tune PCP high automatically is as follows,
- Minimize allocation/freeing from/to shared zone
- Minimize idle pages in PCP
- Minimize pages in PCP if the system free pages is too few
To reach these target, a tuning algorithm as follows is designed,
- When we refill PCP via allocating from the zone, increase PCP high.
Because if we had larger PCP, we could avoid to allocate from the
zone.
- In periodic vmstat updating kworker (via refresh_cpu_vm_stats()),
decrease PCP high to try to free possible idle PCP pages.
- When page reclaiming is active for the zone, stop increasing PCP
high in allocating path, decrease PCP high and free some pages in
freeing path.
So, the PCP high can be tuned to the page allocating/freeing depth of
workloads eventually.
One issue of the algorithm is that if the number of pages allocated is
much more than that of pages freed on a CPU, the PCP high may become the
maximal value even if the allocating/freeing depth is small. But this
isn't a severe issue, because there are no idle pages in this case.
One alternative choice is to increase PCP high when we drain PCP via
trying to free pages to the zone, but don't increase PCP high during PCP
refilling. This can avoid the issue above. But if the number of pages
allocated is much less than that of pages freed on a CPU, there will be
many idle pages in PCP and it is hard to free these idle pages.
1/8 (>> 3) of PCP high will be decreased periodically. The value 1/8 is
kind of arbitrary. Just to make sure that the idle PCP pages will be
freed eventually.
On a 2-socket Intel server with 224 logical CPU, we run 8 kbuild instances
in parallel (each with `make -j 28`) in 8 cgroup. This simulates the
kbuild server that is used by 0-Day kbuild service. With the patch, the
build time decreases 3.5%. The cycles% of the spinlock contention (mostly
for zone lock) decreases from 11.0% to 0.5%. The number of PCP draining
for high order pages freeing (free_high) decreases 65.6%. The number of
pages allocated from zone (instead of from PCP) decreases 83.9%.
Link: https://lkml.kernel.org/r/20231016053002.756205-8-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Suggested-by: Mel Gorman <mgorman@techsingularity.net>
Suggested-by: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:00 +08:00
|
|
|
|
2023-10-16 13:29:58 +08:00
|
|
|
|
|
|
|
|
|
mm: tune PCP high automatically
The target to tune PCP high automatically is as follows,
- Minimize allocation/freeing from/to shared zone
- Minimize idle pages in PCP
- Minimize pages in PCP if the system free pages is too few
To reach these target, a tuning algorithm as follows is designed,
- When we refill PCP via allocating from the zone, increase PCP high.
Because if we had larger PCP, we could avoid to allocate from the
zone.
- In periodic vmstat updating kworker (via refresh_cpu_vm_stats()),
decrease PCP high to try to free possible idle PCP pages.
- When page reclaiming is active for the zone, stop increasing PCP
high in allocating path, decrease PCP high and free some pages in
freeing path.
So, the PCP high can be tuned to the page allocating/freeing depth of
workloads eventually.
One issue of the algorithm is that if the number of pages allocated is
much more than that of pages freed on a CPU, the PCP high may become the
maximal value even if the allocating/freeing depth is small. But this
isn't a severe issue, because there are no idle pages in this case.
One alternative choice is to increase PCP high when we drain PCP via
trying to free pages to the zone, but don't increase PCP high during PCP
refilling. This can avoid the issue above. But if the number of pages
allocated is much less than that of pages freed on a CPU, there will be
many idle pages in PCP and it is hard to free these idle pages.
1/8 (>> 3) of PCP high will be decreased periodically. The value 1/8 is
kind of arbitrary. Just to make sure that the idle PCP pages will be
freed eventually.
On a 2-socket Intel server with 224 logical CPU, we run 8 kbuild instances
in parallel (each with `make -j 28`) in 8 cgroup. This simulates the
kbuild server that is used by 0-Day kbuild service. With the patch, the
build time decreases 3.5%. The cycles% of the spinlock contention (mostly
for zone lock) decreases from 11.0% to 0.5%. The number of PCP draining
for high order pages freeing (free_high) decreases 65.6%. The number of
pages allocated from zone (instead of from PCP) decreases 83.9%.
Link: https://lkml.kernel.org/r/20231016053002.756205-8-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Suggested-by: Mel Gorman <mgorman@techsingularity.net>
Suggested-by: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:00 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-16 13:29:58 +08:00
|
|
|
|
mm: tune PCP high automatically
The target to tune PCP high automatically is as follows,
- Minimize allocation/freeing from/to shared zone
- Minimize idle pages in PCP
- Minimize pages in PCP if the system free pages is too few
To reach these target, a tuning algorithm as follows is designed,
- When we refill PCP via allocating from the zone, increase PCP high.
Because if we had larger PCP, we could avoid to allocate from the
zone.
- In periodic vmstat updating kworker (via refresh_cpu_vm_stats()),
decrease PCP high to try to free possible idle PCP pages.
- When page reclaiming is active for the zone, stop increasing PCP
high in allocating path, decrease PCP high and free some pages in
freeing path.
So, the PCP high can be tuned to the page allocating/freeing depth of
workloads eventually.
One issue of the algorithm is that if the number of pages allocated is
much more than that of pages freed on a CPU, the PCP high may become the
maximal value even if the allocating/freeing depth is small. But this
isn't a severe issue, because there are no idle pages in this case.
One alternative choice is to increase PCP high when we drain PCP via
trying to free pages to the zone, but don't increase PCP high during PCP
refilling. This can avoid the issue above. But if the number of pages
allocated is much less than that of pages freed on a CPU, there will be
many idle pages in PCP and it is hard to free these idle pages.
1/8 (>> 3) of PCP high will be decreased periodically. The value 1/8 is
kind of arbitrary. Just to make sure that the idle PCP pages will be
freed eventually.
On a 2-socket Intel server with 224 logical CPU, we run 8 kbuild instances
in parallel (each with `make -j 28`) in 8 cgroup. This simulates the
kbuild server that is used by 0-Day kbuild service. With the patch, the
build time decreases 3.5%. The cycles% of the spinlock contention (mostly
for zone lock) decreases from 11.0% to 0.5%. The number of PCP draining
for high order pages freeing (free_high) decreases 65.6%. The number of
pages allocated from zone (instead of from PCP) decreases 83.9%.
Link: https://lkml.kernel.org/r/20231016053002.756205-8-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Suggested-by: Mel Gorman <mgorman@techsingularity.net>
Suggested-by: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:00 +08:00
|
|
|
|
|
|
|
|
|
2023-10-16 13:29:58 +08:00
|
|
|
|
mm, pcp: decrease PCP high if free pages < high watermark
One target of PCP is to minimize pages in PCP if the system free pages is
too few. To reach that target, when page reclaiming is active for the
zone (ZONE_RECLAIM_ACTIVE), we will stop increasing PCP high in allocating
path, decrease PCP high and free some pages in freeing path. But this may
be too late because the background page reclaiming may introduce latency
for some workloads. So, in this patch, during page allocation we will
detect whether the number of free pages of the zone is below high
watermark. If so, we will stop increasing PCP high in allocating path,
decrease PCP high and free some pages in freeing path. With this, we can
reduce the possibility of the premature background page reclaiming caused
by too large PCP.
The high watermark checking is done in allocating path to reduce the
overhead in hotter freeing path.
Link: https://lkml.kernel.org/r/20231016053002.756205-9-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:01 +08:00
|
|
|
|
mm: tune PCP high automatically
The target to tune PCP high automatically is as follows,
- Minimize allocation/freeing from/to shared zone
- Minimize idle pages in PCP
- Minimize pages in PCP if the system free pages is too few
To reach these target, a tuning algorithm as follows is designed,
- When we refill PCP via allocating from the zone, increase PCP high.
Because if we had larger PCP, we could avoid to allocate from the
zone.
- In periodic vmstat updating kworker (via refresh_cpu_vm_stats()),
decrease PCP high to try to free possible idle PCP pages.
- When page reclaiming is active for the zone, stop increasing PCP
high in allocating path, decrease PCP high and free some pages in
freeing path.
So, the PCP high can be tuned to the page allocating/freeing depth of
workloads eventually.
One issue of the algorithm is that if the number of pages allocated is
much more than that of pages freed on a CPU, the PCP high may become the
maximal value even if the allocating/freeing depth is small. But this
isn't a severe issue, because there are no idle pages in this case.
One alternative choice is to increase PCP high when we drain PCP via
trying to free pages to the zone, but don't increase PCP high during PCP
refilling. This can avoid the issue above. But if the number of pages
allocated is much less than that of pages freed on a CPU, there will be
many idle pages in PCP and it is hard to free these idle pages.
1/8 (>> 3) of PCP high will be decreased periodically. The value 1/8 is
kind of arbitrary. Just to make sure that the idle PCP pages will be
freed eventually.
On a 2-socket Intel server with 224 logical CPU, we run 8 kbuild instances
in parallel (each with `make -j 28`) in 8 cgroup. This simulates the
kbuild server that is used by 0-Day kbuild service. With the patch, the
build time decreases 3.5%. The cycles% of the spinlock contention (mostly
for zone lock) decreases from 11.0% to 0.5%. The number of PCP draining
for high order pages freeing (free_high) decreases 65.6%. The number of
pages allocated from zone (instead of from PCP) decreases 83.9%.
Link: https://lkml.kernel.org/r/20231016053002.756205-8-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Suggested-by: Mel Gorman <mgorman@techsingularity.net>
Suggested-by: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:00 +08:00
|
|
|
|
|
|
|
|
|
2023-10-16 13:29:58 +08:00
|
|
|
|
mm: tune PCP high automatically
The target to tune PCP high automatically is as follows,
- Minimize allocation/freeing from/to shared zone
- Minimize idle pages in PCP
- Minimize pages in PCP if the system free pages is too few
To reach these target, a tuning algorithm as follows is designed,
- When we refill PCP via allocating from the zone, increase PCP high.
Because if we had larger PCP, we could avoid to allocate from the
zone.
- In periodic vmstat updating kworker (via refresh_cpu_vm_stats()),
decrease PCP high to try to free possible idle PCP pages.
- When page reclaiming is active for the zone, stop increasing PCP
high in allocating path, decrease PCP high and free some pages in
freeing path.
So, the PCP high can be tuned to the page allocating/freeing depth of
workloads eventually.
One issue of the algorithm is that if the number of pages allocated is
much more than that of pages freed on a CPU, the PCP high may become the
maximal value even if the allocating/freeing depth is small. But this
isn't a severe issue, because there are no idle pages in this case.
One alternative choice is to increase PCP high when we drain PCP via
trying to free pages to the zone, but don't increase PCP high during PCP
refilling. This can avoid the issue above. But if the number of pages
allocated is much less than that of pages freed on a CPU, there will be
many idle pages in PCP and it is hard to free these idle pages.
1/8 (>> 3) of PCP high will be decreased periodically. The value 1/8 is
kind of arbitrary. Just to make sure that the idle PCP pages will be
freed eventually.
On a 2-socket Intel server with 224 logical CPU, we run 8 kbuild instances
in parallel (each with `make -j 28`) in 8 cgroup. This simulates the
kbuild server that is used by 0-Day kbuild service. With the patch, the
build time decreases 3.5%. The cycles% of the spinlock contention (mostly
for zone lock) decreases from 11.0% to 0.5%. The number of PCP draining
for high order pages freeing (free_high) decreases 65.6%. The number of
pages allocated from zone (instead of from PCP) decreases 83.9%.
Link: https://lkml.kernel.org/r/20231016053002.756205-8-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Suggested-by: Mel Gorman <mgorman@techsingularity.net>
Suggested-by: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:00 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-16 13:29:58 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-02-24 14:56:26 -08:00
|
|
|
|
2021-04-29 23:01:55 -07:00
|
|
|
|
2021-06-28 19:43:08 -07:00
|
|
|
|
|
|
|
|
|
mm, page_alloc: spread allocations across zones before introducing fragmentation
Patch series "Fragmentation avoidance improvements", v5.
It has been noted before that fragmentation avoidance (aka
anti-fragmentation) is not perfect. Given sufficient time or an adverse
workload, memory gets fragmented and the long-term success of high-order
allocations degrades. This series defines an adverse workload, a definition
of external fragmentation events (including serious) ones and a series
that reduces the level of those fragmentation events.
The details of the workload and the consequences are described in more
detail in the changelogs. However, from patch 1, this is a high-level
summary of the adverse workload. The exact details are found in the
mmtests implementation.
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch)
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameterr create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed
3. Warm up a number of fio read-only threads accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll fault back in old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup
Overall the series reduces external fragmentation causing events by over 94%
on 1 and 2 socket machines, which in turn impacts high-order allocation
success rates over the long term. There are differences in latencies and
high-order allocation success rates. Latencies are a mixed bag as they
are vulnerable to exact system state and whether allocations succeeded
so they are treated as a secondary metric.
Patch 1 uses lower zones if they are populated and have free memory
instead of fragmenting a higher zone. It's special cased to
handle a Normal->DMA32 fallback with the reasons explained
in the changelog.
Patch 2-4 boosts watermarks temporarily when an external fragmentation
event occurs. kswapd wakes to reclaim a small amount of old memory
and then wakes kcompactd on completion to recover the system
slightly. This introduces some overhead in the slowpath. The level
of boosting can be tuned or disabled depending on the tolerance
for fragmentation vs allocation latency.
Patch 5 stalls some movable allocation requests to let kswapd from patch 4
make some progress. The duration of the stalls is very low but it
is possible to tune the system to avoid fragmentation events if
larger stalls can be tolerated.
The bulk of the improvement in fragmentation avoidance is from patches
1-4 but patch 5 can deal with a rare corner case and provides the option
of tuning a system for THP allocation success rates in exchange for
some stalls to control fragmentation.
This patch (of 5):
The page allocator zone lists are iterated based on the watermarks of each
zone which does not take anti-fragmentation into account. On x86, node 0
may have multiple zones while other nodes have one zone. A consequence is
that tasks running on node 0 may fragment ZONE_NORMAL even though
ZONE_DMA32 has plenty of free memory. This patch special cases the
allocator fast path such that it'll try an allocation from a lower local
zone before fragmenting a higher zone. In this case, stealing of
pageblocks or orders larger than a pageblock are still allowed in the fast
path as they are uninteresting from a fragmentation point of view.
This was evaluated using a benchmark designed to fragment memory before
attempting THP allocations. It's implemented in mmtests as the following
configurations
configs/config-global-dhp__workload_thpfioscale
configs/config-global-dhp__workload_thpfioscale-defrag
configs/config-global-dhp__workload_thpfioscale-madvhugepage
e.g. from mmtests
./run-mmtests.sh --run-monitor --config configs/config-global-dhp__workload_thpfioscale test-run-1
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch).
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameter create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed.
3. Warm up a number of fio read-only processes accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll refault old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds.
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup the test files.
Note that due to the use of IO and page cache that this benchmark is not
suitable for running on large machines where the time to fragment memory
may be excessive. Also note that while this is one mix that generates
fragmentation that it's not the only mix that generates fragmentation.
Differences in workload that are more slab-intensive or whether SLUB is
used with high-order pages may yield different results.
When the page allocator fragments memory, it records the event using the
mm_page_alloc_extfrag ftrace event. If the fallback_order is smaller than
a pageblock order (order-9 on 64-bit x86) then it's considered to be an
"external fragmentation event" that may cause issues in the future.
Hence, the primary metric here is the number of external fragmentation
events that occur with order < 9. The secondary metric is allocation
latency and huge page allocation success rates but note that differences
in latencies and what the success rate also can affect the number of
external fragmentation event which is why it's a secondary metric.
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 662.92 ( 0.00%) 653.58 * 1.41%*
Amean fault-huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
Fault latencies are slightly reduced while allocation success rates remain
at zero as this configuration does not make any special effort to allocate
THP and fio is heavily active at the time and either filling memory or
keeping pages resident. However, a 49% reduction of serious fragmentation
events reduces the changes of external fragmentation being a problem in
the future.
Vlastimil asked during review for a breakdown of the allocation types
that are falling back.
vanilla
3816 MIGRATE_UNMOVABLE
800845 MIGRATE_MOVABLE
33 MIGRATE_UNRECLAIMABLE
patch
735 MIGRATE_UNMOVABLE
408135 MIGRATE_MOVABLE
42 MIGRATE_UNRECLAIMABLE
The majority of the fallbacks are due to movable allocations and this is
consistent for the workload throughout the series so will not be presented
again as the primary source of fallbacks are movable allocations.
Movable fallbacks are sometimes considered "ok" to fallback because they
can be migrated. The problem is that they can fill an
unmovable/reclaimable pageblock causing those allocations to fallback
later and polluting pageblocks with pages that cannot move. If there is a
movable fallback, it is pretty much guaranteed to affect an
unmovable/reclaimable pageblock and while it might not be enough to
actually cause a unmovable/reclaimable fallback in the future, we cannot
know that in advance so the patch takes the only option available to it.
Hence, it's important to control them. This point is also consistent
throughout the series and will not be repeated.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 1495.14 ( 0.00%) 1467.55 ( 1.85%)
Amean fault-huge-1 1098.48 ( 0.00%) 1127.11 ( -2.61%)
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 78.57 ( 0.00%) 77.64 ( -1.18%)
Fragmentation events were reduced quite a bit although this is known
to be a little variable. The latencies and allocation success rates
are similar but they were already quite high.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 1350.05 ( 0.00%) 1346.45 ( 0.27%)
Amean fault-huge-5 4181.01 ( 0.00%) 3418.60 ( 18.24%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 1.15 ( 0.00%) 0.78 ( -31.88%)
The reduction of external fragmentation events is slight and this is
partially due to the removal of __GFP_THISNODE in commit ac5b2c18911f
("mm: thp: relax __GFP_THISNODE for MADV_HUGEPAGE mappings") as THP
allocations can now spill over to remote nodes instead of fragmenting
local memory.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 6138.97 ( 0.00%) 6217.43 ( -1.28%)
Amean fault-huge-5 2294.28 ( 0.00%) 3163.33 * -37.88%*
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 96.82 ( 0.00%) 95.14 ( -1.74%)
There was a slight reduction in external fragmentation events although the
latencies were higher. The allocation success rate is high enough that
the system is struggling and there is quite a lot of parallel reclaim and
compaction activity. There is also a certain degree of luck on whether
processes start on node 0 or not for this patch but the relevance is
reduced later in the series.
Overall, the patch reduces the number of external fragmentation causing
events so the success of THP over long periods of time would be improved
for this adverse workload.
Link: http://lkml.kernel.org/r/20181123114528.28802-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:41 -08:00
|
|
|
|
2017-11-15 17:38:03 -08:00
|
|
|
|
2017-02-24 14:56:26 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: tune PCP high automatically
The target to tune PCP high automatically is as follows,
- Minimize allocation/freeing from/to shared zone
- Minimize idle pages in PCP
- Minimize pages in PCP if the system free pages is too few
To reach these target, a tuning algorithm as follows is designed,
- When we refill PCP via allocating from the zone, increase PCP high.
Because if we had larger PCP, we could avoid to allocate from the
zone.
- In periodic vmstat updating kworker (via refresh_cpu_vm_stats()),
decrease PCP high to try to free possible idle PCP pages.
- When page reclaiming is active for the zone, stop increasing PCP
high in allocating path, decrease PCP high and free some pages in
freeing path.
So, the PCP high can be tuned to the page allocating/freeing depth of
workloads eventually.
One issue of the algorithm is that if the number of pages allocated is
much more than that of pages freed on a CPU, the PCP high may become the
maximal value even if the allocating/freeing depth is small. But this
isn't a severe issue, because there are no idle pages in this case.
One alternative choice is to increase PCP high when we drain PCP via
trying to free pages to the zone, but don't increase PCP high during PCP
refilling. This can avoid the issue above. But if the number of pages
allocated is much less than that of pages freed on a CPU, there will be
many idle pages in PCP and it is hard to free these idle pages.
1/8 (>> 3) of PCP high will be decreased periodically. The value 1/8 is
kind of arbitrary. Just to make sure that the idle PCP pages will be
freed eventually.
On a 2-socket Intel server with 224 logical CPU, we run 8 kbuild instances
in parallel (each with `make -j 28`) in 8 cgroup. This simulates the
kbuild server that is used by 0-Day kbuild service. With the patch, the
build time decreases 3.5%. The cycles% of the spinlock contention (mostly
for zone lock) decreases from 11.0% to 0.5%. The number of PCP draining
for high order pages freeing (free_high) decreases 65.6%. The number of
pages allocated from zone (instead of from PCP) decreases 83.9%.
Link: https://lkml.kernel.org/r/20231016053002.756205-8-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Suggested-by: Mel Gorman <mgorman@techsingularity.net>
Suggested-by: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:00 +08:00
|
|
|
|
2021-06-28 19:43:08 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, page_alloc: spread allocations across zones before introducing fragmentation
Patch series "Fragmentation avoidance improvements", v5.
It has been noted before that fragmentation avoidance (aka
anti-fragmentation) is not perfect. Given sufficient time or an adverse
workload, memory gets fragmented and the long-term success of high-order
allocations degrades. This series defines an adverse workload, a definition
of external fragmentation events (including serious) ones and a series
that reduces the level of those fragmentation events.
The details of the workload and the consequences are described in more
detail in the changelogs. However, from patch 1, this is a high-level
summary of the adverse workload. The exact details are found in the
mmtests implementation.
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch)
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameterr create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed
3. Warm up a number of fio read-only threads accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll fault back in old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup
Overall the series reduces external fragmentation causing events by over 94%
on 1 and 2 socket machines, which in turn impacts high-order allocation
success rates over the long term. There are differences in latencies and
high-order allocation success rates. Latencies are a mixed bag as they
are vulnerable to exact system state and whether allocations succeeded
so they are treated as a secondary metric.
Patch 1 uses lower zones if they are populated and have free memory
instead of fragmenting a higher zone. It's special cased to
handle a Normal->DMA32 fallback with the reasons explained
in the changelog.
Patch 2-4 boosts watermarks temporarily when an external fragmentation
event occurs. kswapd wakes to reclaim a small amount of old memory
and then wakes kcompactd on completion to recover the system
slightly. This introduces some overhead in the slowpath. The level
of boosting can be tuned or disabled depending on the tolerance
for fragmentation vs allocation latency.
Patch 5 stalls some movable allocation requests to let kswapd from patch 4
make some progress. The duration of the stalls is very low but it
is possible to tune the system to avoid fragmentation events if
larger stalls can be tolerated.
The bulk of the improvement in fragmentation avoidance is from patches
1-4 but patch 5 can deal with a rare corner case and provides the option
of tuning a system for THP allocation success rates in exchange for
some stalls to control fragmentation.
This patch (of 5):
The page allocator zone lists are iterated based on the watermarks of each
zone which does not take anti-fragmentation into account. On x86, node 0
may have multiple zones while other nodes have one zone. A consequence is
that tasks running on node 0 may fragment ZONE_NORMAL even though
ZONE_DMA32 has plenty of free memory. This patch special cases the
allocator fast path such that it'll try an allocation from a lower local
zone before fragmenting a higher zone. In this case, stealing of
pageblocks or orders larger than a pageblock are still allowed in the fast
path as they are uninteresting from a fragmentation point of view.
This was evaluated using a benchmark designed to fragment memory before
attempting THP allocations. It's implemented in mmtests as the following
configurations
configs/config-global-dhp__workload_thpfioscale
configs/config-global-dhp__workload_thpfioscale-defrag
configs/config-global-dhp__workload_thpfioscale-madvhugepage
e.g. from mmtests
./run-mmtests.sh --run-monitor --config configs/config-global-dhp__workload_thpfioscale test-run-1
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch).
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameter create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed.
3. Warm up a number of fio read-only processes accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll refault old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds.
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup the test files.
Note that due to the use of IO and page cache that this benchmark is not
suitable for running on large machines where the time to fragment memory
may be excessive. Also note that while this is one mix that generates
fragmentation that it's not the only mix that generates fragmentation.
Differences in workload that are more slab-intensive or whether SLUB is
used with high-order pages may yield different results.
When the page allocator fragments memory, it records the event using the
mm_page_alloc_extfrag ftrace event. If the fallback_order is smaller than
a pageblock order (order-9 on 64-bit x86) then it's considered to be an
"external fragmentation event" that may cause issues in the future.
Hence, the primary metric here is the number of external fragmentation
events that occur with order < 9. The secondary metric is allocation
latency and huge page allocation success rates but note that differences
in latencies and what the success rate also can affect the number of
external fragmentation event which is why it's a secondary metric.
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 662.92 ( 0.00%) 653.58 * 1.41%*
Amean fault-huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
Fault latencies are slightly reduced while allocation success rates remain
at zero as this configuration does not make any special effort to allocate
THP and fio is heavily active at the time and either filling memory or
keeping pages resident. However, a 49% reduction of serious fragmentation
events reduces the changes of external fragmentation being a problem in
the future.
Vlastimil asked during review for a breakdown of the allocation types
that are falling back.
vanilla
3816 MIGRATE_UNMOVABLE
800845 MIGRATE_MOVABLE
33 MIGRATE_UNRECLAIMABLE
patch
735 MIGRATE_UNMOVABLE
408135 MIGRATE_MOVABLE
42 MIGRATE_UNRECLAIMABLE
The majority of the fallbacks are due to movable allocations and this is
consistent for the workload throughout the series so will not be presented
again as the primary source of fallbacks are movable allocations.
Movable fallbacks are sometimes considered "ok" to fallback because they
can be migrated. The problem is that they can fill an
unmovable/reclaimable pageblock causing those allocations to fallback
later and polluting pageblocks with pages that cannot move. If there is a
movable fallback, it is pretty much guaranteed to affect an
unmovable/reclaimable pageblock and while it might not be enough to
actually cause a unmovable/reclaimable fallback in the future, we cannot
know that in advance so the patch takes the only option available to it.
Hence, it's important to control them. This point is also consistent
throughout the series and will not be repeated.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 1495.14 ( 0.00%) 1467.55 ( 1.85%)
Amean fault-huge-1 1098.48 ( 0.00%) 1127.11 ( -2.61%)
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 78.57 ( 0.00%) 77.64 ( -1.18%)
Fragmentation events were reduced quite a bit although this is known
to be a little variable. The latencies and allocation success rates
are similar but they were already quite high.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 1350.05 ( 0.00%) 1346.45 ( 0.27%)
Amean fault-huge-5 4181.01 ( 0.00%) 3418.60 ( 18.24%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 1.15 ( 0.00%) 0.78 ( -31.88%)
The reduction of external fragmentation events is slight and this is
partially due to the removal of __GFP_THISNODE in commit ac5b2c18911f
("mm: thp: relax __GFP_THISNODE for MADV_HUGEPAGE mappings") as THP
allocations can now spill over to remote nodes instead of fragmenting
local memory.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 6138.97 ( 0.00%) 6217.43 ( -1.28%)
Amean fault-huge-5 2294.28 ( 0.00%) 3163.33 * -37.88%*
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 96.82 ( 0.00%) 95.14 ( -1.74%)
There was a slight reduction in external fragmentation events although the
latencies were higher. The allocation success rate is high enough that
the system is struggling and there is quite a lot of parallel reclaim and
compaction activity. There is also a certain degree of luck on whether
processes start on node 0 or not for this patch but the relevance is
reduced later in the series.
Overall, the patch reduces the number of external fragmentation causing
events so the success of THP over long periods of time would be improved
for this adverse workload.
Link: http://lkml.kernel.org/r/20181123114528.28802-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:41 -08:00
|
|
|
|
2021-06-28 19:43:08 -07:00
|
|
|
|
|
|
|
|
|
2017-02-24 14:56:26 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-06-24 13:54:17 +01:00
|
|
|
|
|
|
|
|
|
2021-06-28 19:43:08 -07:00
|
|
|
|
mm, page_alloc: reduce page alloc/free sanity checks
Historically, we have performed sanity checks on all struct pages being
allocated or freed, making sure they have no unexpected page flags or
certain field values. This can detect insufficient cleanup and some cases
of use-after-free, although on its own it can't always identify the
culprit. The result is a warning and the "bad page" being leaked.
The checks do need some cpu cycles, so in 4.7 with commits 479f854a207c
("mm, page_alloc: defer debugging checks of pages allocated from the PCP")
and 4db7548ccbd9 ("mm, page_alloc: defer debugging checks of freed pages
until a PCP drain") they were no longer performed in the hot paths when
allocating and freeing from pcplists, but only when pcplists are bypassed,
refilled or drained. For debugging purposes, with CONFIG_DEBUG_VM enabled
the checks were instead still done in the hot paths and not when refilling
or draining pcplists.
With 4462b32c9285 ("mm, page_alloc: more extensive free page checking with
debug_pagealloc"), enabling debug_pagealloc also moved the sanity checks
back to hot pahs. When both debug_pagealloc and CONFIG_DEBUG_VM are
enabled, the checks are done both in hotpaths and pcplist refill/drain.
Even though the non-debug default today might seem to be a sensible
tradeoff between overhead and ability to detect bad pages, on closer look
it's arguably not. As most allocations go through the pcplists, catching
any bad pages when refilling or draining pcplists has only a small chance,
insufficient for debugging or serious hardening purposes. On the other
hand the cost of the checks is concentrated in the already expensive
drain/refill batching operations, and those are done under the often
contended zone lock. That was recently identified as an issue for page
allocation and the zone lock contention reduced by moving the checks
outside of the locked section with a patch "mm: reduce lock contention of
pcp buffer refill", but the cost of the checks is still visible compared
to their removal [1]. In the pcplist draining path free_pcppages_bulk()
the checks are still done under zone->lock.
Thus, remove the checks from pcplist refill and drain paths completely.
Introduce a static key check_pages_enabled to control checks during page
allocation a freeing (whether pcplist is used or bypassed). The static
key is enabled if either is true:
- kernel is built with CONFIG_DEBUG_VM=y (debugging)
- debug_pagealloc or page poisoning is boot-time enabled (debugging)
- init_on_alloc or init_on_free is boot-time enabled (hardening)
The resulting user visible changes:
- no checks when draining/refilling pcplists - less overhead, with
likely no practical reduction of ability to catch bad pages
- no checks when bypassing pcplists in default config (no
debugging/hardening) - less overhead etc. as above
- on typical hardened kernels [2], checks are now performed on each page
allocation/free (previously only when bypassing/draining/refilling
pcplists) - the init_on_alloc/init_on_free enabled should be sufficient
indication for preferring more costly alloc/free operations for
hardening purposes and we shouldn't need to introduce another toggle
- code (various wrappers) removal and simplification
[1] https://lore.kernel.org/all/68ba44d8-6899-c018-dcb3-36f3a96e6bea@sra.uni-hannover.de/
[2] https://lore.kernel.org/all/63ebc499.a70a0220.9ac51.29ea@mx.google.com/
[akpm@linux-foundation.org: coding-style cleanups]
[akpm@linux-foundation.org: make check_pages_enabled static]
Link: https://lkml.kernel.org/r/20230216095131.17336-1-vbabka@suse.cz
Reported-by: Alexander Halbuer <halbuer@sra.uni-hannover.de>
Reported-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Kees Cook <keescook@chromium.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-02-16 10:51:31 +01:00
|
|
|
|
2017-02-24 14:56:26 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-28 19:43:08 -07:00
|
|
|
|
2022-08-31 01:34:04 +00:00
|
|
|
|
2017-02-24 14:56:26 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-06-24 13:54:21 +01:00
|
|
|
|
2017-02-24 14:56:26 -08:00
|
|
|
|
2022-11-18 10:17:14 +00:00
|
|
|
|
2022-06-24 13:54:21 +01:00
|
|
|
|
2022-11-18 10:17:14 +00:00
|
|
|
|
2022-06-24 13:54:23 +01:00
|
|
|
|
2022-06-24 13:54:21 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-28 19:42:18 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, pcp: reduce detecting time of consecutive high order page freeing
In current PCP auto-tuning design, if the number of pages allocated is
much more than that of pages freed on a CPU, the PCP high may become the
maximal value even if the allocating/freeing depth is small, for example,
in the sender of network workloads. If a CPU was used as sender
originally, then it is used as receiver after context switching, we need
to fill the whole PCP with maximal high before triggering PCP draining for
consecutive high order freeing. This will hurt the performance of some
network workloads.
To solve the issue, in this patch, we will track the consecutive page
freeing with a counter in stead of relying on PCP draining. So, we can
detect consecutive page freeing much earlier.
On a 2-socket Intel server with 128 logical CPU, we tested
SCTP_STREAM_MANY test case of netperf test suite with 64-pair processes.
With the patch, the network bandwidth improves 5.0%. This restores the
performance drop caused by PCP auto-tuning.
Link: https://lkml.kernel.org/r/20231016053002.756205-10-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:02 +08:00
|
|
|
|
2021-06-28 19:43:08 -07:00
|
|
|
|
|
|
|
|
|
2022-11-18 10:17:14 +00:00
|
|
|
|
2022-06-24 13:54:21 +01:00
|
|
|
|
2017-02-24 14:56:26 -08:00
|
|
|
|
2022-10-06 10:15:40 +00:00
|
|
|
|
2021-06-28 19:41:50 -07:00
|
|
|
|
2017-02-24 14:56:26 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2022-09-18 02:56:40 +00:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2022-09-15 17:03:48 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/page_alloc: pageblock flags functions clean up
Patch series "Make MIGRATE_ISOLATE a standalone bit", v10.
This patchset moves MIGRATE_ISOLATE to a standalone bit to avoid being
overwritten during pageblock isolation process. Currently,
MIGRATE_ISOLATE is part of enum migratetype (in include/linux/mmzone.h),
thus, setting a pageblock to MIGRATE_ISOLATE overwrites its original
migratetype. This causes pageblock migratetype loss during
alloc_contig_range() and memory offline, especially when the process fails
due to a failed pageblock isolation and the code tries to undo the
finished pageblock isolations.
In terms of performance for changing pageblock types, no performance
change is observed:
1. I used perf to collect stats of offlining and onlining all memory
of a 40GB VM 10 times and see that get_pfnblock_flags_mask() and
set_pfnblock_flags_mask() take about 0.12% and 0.02% of the whole
process respectively with and without this patchset across 3 runs.
2. I used perf to collect stats of dd from /dev/random to a 40GB tmpfs
file and find get_pfnblock_flags_mask() takes about 0.05% of the
process with and without this patchset across 3 runs.
This patch (of 6):
No functional change is intended.
1. Add __NR_PAGEBLOCK_BITS for the number of pageblock flag bits and use
roundup_pow_of_two(__NR_PAGEBLOCK_BITS) as NR_PAGEBLOCK_BITS to take
right amount of bits for pageblock flags.
2. Rename PB_migrate_skip to PB_compact_skip.
3. Add {get,set,clear}_pfnblock_bit() to operate one a standalone bit,
like PB_compact_skip.
3. Make {get,set}_pfnblock_flags_mask() internal functions and use
{get,set}_pfnblock_migratetype() for pageblock migratetype operations.
4. Move pageblock flags common code to get_pfnblock_bitmap_bitidx().
3. Use MIGRATETYPE_MASK to get the migratetype of a pageblock from its
flags.
4. Use PB_migrate_end in the definition of MIGRATETYPE_MASK instead of
PB_migrate_bits.
5. Add a comment on is_migrate_cma_folio() to prevent one from changing it
to use get_pageblock_migratetype() and causing issues.
Link: https://lkml.kernel.org/r/20250617021115.2331563-1-ziy@nvidia.com
Link: https://lkml.kernel.org/r/20250617021115.2331563-2-ziy@nvidia.com
Signed-off-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kirill A. Shuemov <kirill.shutemov@linux.intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Richard Chang <richardycc@google.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-06-16 22:11:09 -04:00
|
|
|
|
2022-09-15 17:03:48 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2009-06-16 15:32:05 -07:00
|
|
|
|
2017-02-24 14:56:26 -08:00
|
|
|
|
2014-06-04 16:10:21 -07:00
|
|
|
|
2016-05-19 17:13:38 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2005-11-21 21:32:20 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2021-06-28 19:43:08 -07:00
|
|
|
|
2023-08-24 11:38:21 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2022-06-24 13:54:19 +01:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2017-02-24 14:56:26 -08:00
|
|
|
|
mm, page_alloc: do not wake kswapd with zone lock held
syzbot reported the following regression in the latest merge window and
it was confirmed by Qian Cai that a similar bug was visible from a
different context.
======================================================
WARNING: possible circular locking dependency detected
4.20.0+ #297 Not tainted
------------------------------------------------------
syz-executor0/8529 is trying to acquire lock:
000000005e7fb829 (&pgdat->kswapd_wait){....}, at:
__wake_up_common_lock+0x19e/0x330 kernel/sched/wait.c:120
but task is already holding lock:
000000009bb7bae0 (&(&zone->lock)->rlock){-.-.}, at: spin_lock
include/linux/spinlock.h:329 [inline]
000000009bb7bae0 (&(&zone->lock)->rlock){-.-.}, at: rmqueue_bulk
mm/page_alloc.c:2548 [inline]
000000009bb7bae0 (&(&zone->lock)->rlock){-.-.}, at: __rmqueue_pcplist
mm/page_alloc.c:3021 [inline]
000000009bb7bae0 (&(&zone->lock)->rlock){-.-.}, at: rmqueue_pcplist
mm/page_alloc.c:3050 [inline]
000000009bb7bae0 (&(&zone->lock)->rlock){-.-.}, at: rmqueue
mm/page_alloc.c:3072 [inline]
000000009bb7bae0 (&(&zone->lock)->rlock){-.-.}, at:
get_page_from_freelist+0x1bae/0x52a0 mm/page_alloc.c:3491
It appears to be a false positive in that the only way the lock ordering
should be inverted is if kswapd is waking itself and the wakeup
allocates debugging objects which should already be allocated if it's
kswapd doing the waking. Nevertheless, the possibility exists and so
it's best to avoid the problem.
This patch flags a zone as needing a kswapd using the, surprisingly,
unused zone flag field. The flag is read without the lock held to do
the wakeup. It's possible that the flag setting context is not the same
as the flag clearing context or for small races to occur. However, each
race possibility is harmless and there is no visible degredation in
fragmentation treatment.
While zone->flag could have continued to be unused, there is potential
for moving some existing fields into the flags field instead.
Particularly read-mostly ones like zone->initialized and
zone->contiguous.
Link: http://lkml.kernel.org/r/20190103225712.GJ31517@techsingularity.net
Fixes: 1c30844d2dfe ("mm: reclaim small amounts of memory when an external fragmentation event occurs")
Reported-by: syzbot+93d94a001cfbce9e60e1@syzkaller.appspotmail.com
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Tested-by: Qian Cai <cai@lca.pw>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-01-08 15:23:39 -08:00
|
|
|
|
2023-05-14 09:28:56 +09:00
|
|
|
|
|
|
|
|
|
mm, page_alloc: do not wake kswapd with zone lock held
syzbot reported the following regression in the latest merge window and
it was confirmed by Qian Cai that a similar bug was visible from a
different context.
======================================================
WARNING: possible circular locking dependency detected
4.20.0+ #297 Not tainted
------------------------------------------------------
syz-executor0/8529 is trying to acquire lock:
000000005e7fb829 (&pgdat->kswapd_wait){....}, at:
__wake_up_common_lock+0x19e/0x330 kernel/sched/wait.c:120
but task is already holding lock:
000000009bb7bae0 (&(&zone->lock)->rlock){-.-.}, at: spin_lock
include/linux/spinlock.h:329 [inline]
000000009bb7bae0 (&(&zone->lock)->rlock){-.-.}, at: rmqueue_bulk
mm/page_alloc.c:2548 [inline]
000000009bb7bae0 (&(&zone->lock)->rlock){-.-.}, at: __rmqueue_pcplist
mm/page_alloc.c:3021 [inline]
000000009bb7bae0 (&(&zone->lock)->rlock){-.-.}, at: rmqueue_pcplist
mm/page_alloc.c:3050 [inline]
000000009bb7bae0 (&(&zone->lock)->rlock){-.-.}, at: rmqueue
mm/page_alloc.c:3072 [inline]
000000009bb7bae0 (&(&zone->lock)->rlock){-.-.}, at:
get_page_from_freelist+0x1bae/0x52a0 mm/page_alloc.c:3491
It appears to be a false positive in that the only way the lock ordering
should be inverted is if kswapd is waking itself and the wakeup
allocates debugging objects which should already be allocated if it's
kswapd doing the waking. Nevertheless, the possibility exists and so
it's best to avoid the problem.
This patch flags a zone as needing a kswapd using the, surprisingly,
unused zone flag field. The flag is read without the lock held to do
the wakeup. It's possible that the flag setting context is not the same
as the flag clearing context or for small races to occur. However, each
race possibility is harmless and there is no visible degredation in
fragmentation treatment.
While zone->flag could have continued to be unused, there is potential
for moving some existing fields into the flags field instead.
Particularly read-mostly ones like zone->initialized and
zone->contiguous.
Link: http://lkml.kernel.org/r/20190103225712.GJ31517@techsingularity.net
Fixes: 1c30844d2dfe ("mm: reclaim small amounts of memory when an external fragmentation event occurs")
Reported-by: syzbot+93d94a001cfbce9e60e1@syzkaller.appspotmail.com
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Tested-by: Qian Cai <cai@lca.pw>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-01-08 15:23:39 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-02-24 14:56:26 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-02-24 19:08:26 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-02-25 18:45:09 +00:00
|
|
|
|
2025-02-24 19:08:26 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-02-25 18:45:09 +00:00
|
|
|
|
|
|
|
|
|
2025-02-24 19:08:26 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
page_alloc: consider highatomic reserve in watermark fast
zone_watermark_fast was introduced by commit 48ee5f3696f6 ("mm,
page_alloc: shortcut watermark checks for order-0 pages"). The commit
simply checks if free pages is bigger than watermark without additional
calculation such like reducing watermark.
It considered free cma pages but it did not consider highatomic reserved.
This may incur exhaustion of free pages except high order atomic free
pages.
Assume that reserved_highatomic pageblock is bigger than watermark min,
and there are only few free pages except high order atomic free. Because
zone_watermark_fast passes the allocation without considering high order
atomic free, normal reclaimable allocation like GFP_HIGHUSER will consume
all the free pages. Then finally order-0 atomic allocation may fail on
allocation.
This means watermark min is not protected against non-atomic allocation.
The order-0 atomic allocation with ALLOC_HARDER unwantedly can be failed.
Additionally the __GFP_MEMALLOC allocation with ALLOC_NO_WATERMARKS also
can be failed.
To avoid the problem, zone_watermark_fast should consider highatomic
reserve. If the actual size of high atomic free is counted accurately
like cma free, we may use it. On this patch just use
nr_reserved_highatomic. Additionally introduce
__zone_watermark_unusable_free to factor out common parts between
zone_watermark_fast and __zone_watermark_ok.
This is an example of ALLOC_HARDER allocation failure using v4.19 based
kernel.
Binder:9343_3: page allocation failure: order:0, mode:0x480020(GFP_ATOMIC), nodemask=(null)
Call trace:
[<ffffff8008f40f8c>] dump_stack+0xb8/0xf0
[<ffffff8008223320>] warn_alloc+0xd8/0x12c
[<ffffff80082245e4>] __alloc_pages_nodemask+0x120c/0x1250
[<ffffff800827f6e8>] new_slab+0x128/0x604
[<ffffff800827b0cc>] ___slab_alloc+0x508/0x670
[<ffffff800827ba00>] __kmalloc+0x2f8/0x310
[<ffffff80084ac3e0>] context_struct_to_string+0x104/0x1cc
[<ffffff80084ad8fc>] security_sid_to_context_core+0x74/0x144
[<ffffff80084ad880>] security_sid_to_context+0x10/0x18
[<ffffff800849bd80>] selinux_secid_to_secctx+0x20/0x28
[<ffffff800849109c>] security_secid_to_secctx+0x3c/0x70
[<ffffff8008bfe118>] binder_transaction+0xe68/0x454c
Mem-Info:
active_anon:102061 inactive_anon:81551 isolated_anon:0
active_file:59102 inactive_file:68924 isolated_file:64
unevictable:611 dirty:63 writeback:0 unstable:0
slab_reclaimable:13324 slab_unreclaimable:44354
mapped:83015 shmem:4858 pagetables:26316 bounce:0
free:2727 free_pcp:1035 free_cma:178
Node 0 active_anon:408244kB inactive_anon:326204kB active_file:236408kB inactive_file:275696kB unevictable:2444kB isolated(anon):0kB isolated(file):256kB mapped:332060kB dirty:252kB writeback:0kB shmem:19432kB writeback_tmp:0kB unstable:0kB all_unreclaimable? no
Normal free:10908kB min:6192kB low:44388kB high:47060kB active_anon:409160kB inactive_anon:325924kB active_file:235820kB inactive_file:276628kB unevictable:2444kB writepending:252kB present:3076096kB managed:2673676kB mlocked:2444kB kernel_stack:62512kB pagetables:105264kB bounce:0kB free_pcp:4140kB local_pcp:40kB free_cma:712kB
lowmem_reserve[]: 0 0
Normal: 505*4kB (H) 357*8kB (H) 201*16kB (H) 65*32kB (H) 1*64kB (H) 0*128kB 0*256kB 0*512kB 0*1024kB 0*2048kB 0*4096kB = 10236kB
138826 total pagecache pages
5460 pages in swap cache
Swap cache stats: add 8273090, delete 8267506, find 1004381/4060142
This is an example of ALLOC_NO_WATERMARKS allocation failure using v4.14
based kernel.
kswapd0: page allocation failure: order:0, mode:0x140000a(GFP_NOIO|__GFP_HIGHMEM|__GFP_MOVABLE), nodemask=(null)
kswapd0 cpuset=/ mems_allowed=0
CPU: 4 PID: 1221 Comm: kswapd0 Not tainted 4.14.113-18770262-userdebug #1
Call trace:
[<0000000000000000>] dump_backtrace+0x0/0x248
[<0000000000000000>] show_stack+0x18/0x20
[<0000000000000000>] __dump_stack+0x20/0x28
[<0000000000000000>] dump_stack+0x68/0x90
[<0000000000000000>] warn_alloc+0x104/0x198
[<0000000000000000>] __alloc_pages_nodemask+0xdc0/0xdf0
[<0000000000000000>] zs_malloc+0x148/0x3d0
[<0000000000000000>] zram_bvec_rw+0x410/0x798
[<0000000000000000>] zram_rw_page+0x88/0xdc
[<0000000000000000>] bdev_write_page+0x70/0xbc
[<0000000000000000>] __swap_writepage+0x58/0x37c
[<0000000000000000>] swap_writepage+0x40/0x4c
[<0000000000000000>] shrink_page_list+0xc30/0xf48
[<0000000000000000>] shrink_inactive_list+0x2b0/0x61c
[<0000000000000000>] shrink_node_memcg+0x23c/0x618
[<0000000000000000>] shrink_node+0x1c8/0x304
[<0000000000000000>] kswapd+0x680/0x7c4
[<0000000000000000>] kthread+0x110/0x120
[<0000000000000000>] ret_from_fork+0x10/0x18
Mem-Info:
active_anon:111826 inactive_anon:65557 isolated_anon:0\x0a active_file:44260 inactive_file:83422 isolated_file:0\x0a unevictable:4158 dirty:117 writeback:0 unstable:0\x0a slab_reclaimable:13943 slab_unreclaimable:43315\x0a mapped:102511 shmem:3299 pagetables:19566 bounce:0\x0a free:3510 free_pcp:553 free_cma:0
Node 0 active_anon:447304kB inactive_anon:262228kB active_file:177040kB inactive_file:333688kB unevictable:16632kB isolated(anon):0kB isolated(file):0kB mapped:410044kB d irty:468kB writeback:0kB shmem:13196kB writeback_tmp:0kB unstable:0kB all_unreclaimable? no
Normal free:14040kB min:7440kB low:94500kB high:98136kB reserved_highatomic:32768KB active_anon:447336kB inactive_anon:261668kB active_file:177572kB inactive_file:333768k B unevictable:16632kB writepending:480kB present:4081664kB managed:3637088kB mlocked:16632kB kernel_stack:47072kB pagetables:78264kB bounce:0kB free_pcp:2280kB local_pcp:720kB free_cma:0kB [ 4738.329607] lowmem_reserve[]: 0 0
Normal: 860*4kB (H) 453*8kB (H) 180*16kB (H) 26*32kB (H) 34*64kB (H) 6*128kB (H) 2*256kB (H) 0*512kB 0*1024kB 0*2048kB 0*4096kB = 14232kB
This is trace log which shows GFP_HIGHUSER consumes free pages right
before ALLOC_NO_WATERMARKS.
<...>-22275 [006] .... 889.213383: mm_page_alloc: page=00000000d2be5665 pfn=970744 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213385: mm_page_alloc: page=000000004b2335c2 pfn=970745 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213387: mm_page_alloc: page=00000000017272e1 pfn=970278 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213389: mm_page_alloc: page=00000000c4be79fb pfn=970279 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213391: mm_page_alloc: page=00000000f8a51d4f pfn=970260 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213393: mm_page_alloc: page=000000006ba8f5ac pfn=970261 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213395: mm_page_alloc: page=00000000819f1cd3 pfn=970196 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213396: mm_page_alloc: page=00000000f6b72a64 pfn=970197 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
kswapd0-1207 [005] ...1 889.213398: mm_page_alloc: page= (null) pfn=0 order=0 migratetype=1 nr_free=3650 gfp_flags=GFP_NOWAIT|__GFP_HIGHMEM|__GFP_NOWARN|__GFP_MOVABLE
[jaewon31.kim@samsung.com: remove redundant code for high-order]
Link: http://lkml.kernel.org/r/20200623035242.27232-1-jaewon31.kim@samsung.com
Reported-by: Yong-Taek Lee <ytk.lee@samsung.com>
Suggested-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Jaewon Kim <jaewon31.kim@samsung.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Baoquan He <bhe@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Yong-Taek Lee <ytk.lee@samsung.com>
Cc: Michal Hocko <mhocko@kernel.org>
Link: http://lkml.kernel.org/r/20200619235958.11283-1-jaewon31.kim@samsung.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-06 23:25:20 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-01-13 11:12:15 +00:00
|
|
|
|
mm/page_alloc: keep track of free highatomic
OOM kills due to vastly overestimated free highatomic reserves were
observed:
... invoked oom-killer: gfp_mask=0x100cca(GFP_HIGHUSER_MOVABLE), order=0 ...
Node 0 Normal free:1482936kB boost:0kB min:410416kB low:739404kB high:1068392kB reserved_highatomic:1073152KB ...
Node 0 Normal: 1292*4kB (ME) 1920*8kB (E) 383*16kB (UE) 220*32kB (ME) 340*64kB (E) 2155*128kB (UE) 3243*256kB (UE) 615*512kB (U) 1*1024kB (M) 0*2048kB 0*4096kB = 1477408kB
The second line above shows that the OOM kill was due to the following
condition:
free (1482936kB) - reserved_highatomic (1073152kB) = 409784KB < min (410416kB)
And the third line shows there were no free pages in any
MIGRATE_HIGHATOMIC pageblocks, which otherwise would show up as type 'H'.
Therefore __zone_watermark_unusable_free() underestimated the usable free
memory by over 1GB, which resulted in the unnecessary OOM kill above.
The comments in __zone_watermark_unusable_free() warns about the potential
risk, i.e.,
If the caller does not have rights to reserves below the min
watermark then subtract the high-atomic reserves. This will
over-estimate the size of the atomic reserve but it avoids a search.
However, it is possible to keep track of free pages in reserved highatomic
pageblocks with a new per-zone counter nr_free_highatomic protected by the
zone lock, to avoid a search when calculating the usable free memory. And
the cost would be minimal, i.e., simple arithmetics in the highatomic
alloc/free/move paths.
Note that since nr_free_highatomic can be relatively small, using a
per-cpu counter might cause too much drift and defeat its purpose, in
addition to the extra memory overhead.
Dependson e0932b6c1f94 ("mm: page_alloc: consolidate free page accounting") - see [1]
[akpm@linux-foundation.org: s/if/else if/, per Johannes, stealth whitespace tweak]
Link: https://lkml.kernel.org/r/20241028182653.3420139-1-yuzhao@google.com
Link: https://lkml.kernel.org/r/0d0ddb33-fcdc-43e2-801f-0c1df2031afb@suse.cz [1]
Fixes: 0aaa29a56e4f ("mm, page_alloc: reserve pageblocks for high-order atomic allocations on demand")
Signed-off-by: Yu Zhao <yuzhao@google.com>
Reported-by: Link Lin <linkl@google.com>
Acked-by: David Rientjes <rientjes@google.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-10-28 12:26:53 -06:00
|
|
|
|
page_alloc: consider highatomic reserve in watermark fast
zone_watermark_fast was introduced by commit 48ee5f3696f6 ("mm,
page_alloc: shortcut watermark checks for order-0 pages"). The commit
simply checks if free pages is bigger than watermark without additional
calculation such like reducing watermark.
It considered free cma pages but it did not consider highatomic reserved.
This may incur exhaustion of free pages except high order atomic free
pages.
Assume that reserved_highatomic pageblock is bigger than watermark min,
and there are only few free pages except high order atomic free. Because
zone_watermark_fast passes the allocation without considering high order
atomic free, normal reclaimable allocation like GFP_HIGHUSER will consume
all the free pages. Then finally order-0 atomic allocation may fail on
allocation.
This means watermark min is not protected against non-atomic allocation.
The order-0 atomic allocation with ALLOC_HARDER unwantedly can be failed.
Additionally the __GFP_MEMALLOC allocation with ALLOC_NO_WATERMARKS also
can be failed.
To avoid the problem, zone_watermark_fast should consider highatomic
reserve. If the actual size of high atomic free is counted accurately
like cma free, we may use it. On this patch just use
nr_reserved_highatomic. Additionally introduce
__zone_watermark_unusable_free to factor out common parts between
zone_watermark_fast and __zone_watermark_ok.
This is an example of ALLOC_HARDER allocation failure using v4.19 based
kernel.
Binder:9343_3: page allocation failure: order:0, mode:0x480020(GFP_ATOMIC), nodemask=(null)
Call trace:
[<ffffff8008f40f8c>] dump_stack+0xb8/0xf0
[<ffffff8008223320>] warn_alloc+0xd8/0x12c
[<ffffff80082245e4>] __alloc_pages_nodemask+0x120c/0x1250
[<ffffff800827f6e8>] new_slab+0x128/0x604
[<ffffff800827b0cc>] ___slab_alloc+0x508/0x670
[<ffffff800827ba00>] __kmalloc+0x2f8/0x310
[<ffffff80084ac3e0>] context_struct_to_string+0x104/0x1cc
[<ffffff80084ad8fc>] security_sid_to_context_core+0x74/0x144
[<ffffff80084ad880>] security_sid_to_context+0x10/0x18
[<ffffff800849bd80>] selinux_secid_to_secctx+0x20/0x28
[<ffffff800849109c>] security_secid_to_secctx+0x3c/0x70
[<ffffff8008bfe118>] binder_transaction+0xe68/0x454c
Mem-Info:
active_anon:102061 inactive_anon:81551 isolated_anon:0
active_file:59102 inactive_file:68924 isolated_file:64
unevictable:611 dirty:63 writeback:0 unstable:0
slab_reclaimable:13324 slab_unreclaimable:44354
mapped:83015 shmem:4858 pagetables:26316 bounce:0
free:2727 free_pcp:1035 free_cma:178
Node 0 active_anon:408244kB inactive_anon:326204kB active_file:236408kB inactive_file:275696kB unevictable:2444kB isolated(anon):0kB isolated(file):256kB mapped:332060kB dirty:252kB writeback:0kB shmem:19432kB writeback_tmp:0kB unstable:0kB all_unreclaimable? no
Normal free:10908kB min:6192kB low:44388kB high:47060kB active_anon:409160kB inactive_anon:325924kB active_file:235820kB inactive_file:276628kB unevictable:2444kB writepending:252kB present:3076096kB managed:2673676kB mlocked:2444kB kernel_stack:62512kB pagetables:105264kB bounce:0kB free_pcp:4140kB local_pcp:40kB free_cma:712kB
lowmem_reserve[]: 0 0
Normal: 505*4kB (H) 357*8kB (H) 201*16kB (H) 65*32kB (H) 1*64kB (H) 0*128kB 0*256kB 0*512kB 0*1024kB 0*2048kB 0*4096kB = 10236kB
138826 total pagecache pages
5460 pages in swap cache
Swap cache stats: add 8273090, delete 8267506, find 1004381/4060142
This is an example of ALLOC_NO_WATERMARKS allocation failure using v4.14
based kernel.
kswapd0: page allocation failure: order:0, mode:0x140000a(GFP_NOIO|__GFP_HIGHMEM|__GFP_MOVABLE), nodemask=(null)
kswapd0 cpuset=/ mems_allowed=0
CPU: 4 PID: 1221 Comm: kswapd0 Not tainted 4.14.113-18770262-userdebug #1
Call trace:
[<0000000000000000>] dump_backtrace+0x0/0x248
[<0000000000000000>] show_stack+0x18/0x20
[<0000000000000000>] __dump_stack+0x20/0x28
[<0000000000000000>] dump_stack+0x68/0x90
[<0000000000000000>] warn_alloc+0x104/0x198
[<0000000000000000>] __alloc_pages_nodemask+0xdc0/0xdf0
[<0000000000000000>] zs_malloc+0x148/0x3d0
[<0000000000000000>] zram_bvec_rw+0x410/0x798
[<0000000000000000>] zram_rw_page+0x88/0xdc
[<0000000000000000>] bdev_write_page+0x70/0xbc
[<0000000000000000>] __swap_writepage+0x58/0x37c
[<0000000000000000>] swap_writepage+0x40/0x4c
[<0000000000000000>] shrink_page_list+0xc30/0xf48
[<0000000000000000>] shrink_inactive_list+0x2b0/0x61c
[<0000000000000000>] shrink_node_memcg+0x23c/0x618
[<0000000000000000>] shrink_node+0x1c8/0x304
[<0000000000000000>] kswapd+0x680/0x7c4
[<0000000000000000>] kthread+0x110/0x120
[<0000000000000000>] ret_from_fork+0x10/0x18
Mem-Info:
active_anon:111826 inactive_anon:65557 isolated_anon:0\x0a active_file:44260 inactive_file:83422 isolated_file:0\x0a unevictable:4158 dirty:117 writeback:0 unstable:0\x0a slab_reclaimable:13943 slab_unreclaimable:43315\x0a mapped:102511 shmem:3299 pagetables:19566 bounce:0\x0a free:3510 free_pcp:553 free_cma:0
Node 0 active_anon:447304kB inactive_anon:262228kB active_file:177040kB inactive_file:333688kB unevictable:16632kB isolated(anon):0kB isolated(file):0kB mapped:410044kB d irty:468kB writeback:0kB shmem:13196kB writeback_tmp:0kB unstable:0kB all_unreclaimable? no
Normal free:14040kB min:7440kB low:94500kB high:98136kB reserved_highatomic:32768KB active_anon:447336kB inactive_anon:261668kB active_file:177572kB inactive_file:333768k B unevictable:16632kB writepending:480kB present:4081664kB managed:3637088kB mlocked:16632kB kernel_stack:47072kB pagetables:78264kB bounce:0kB free_pcp:2280kB local_pcp:720kB free_cma:0kB [ 4738.329607] lowmem_reserve[]: 0 0
Normal: 860*4kB (H) 453*8kB (H) 180*16kB (H) 26*32kB (H) 34*64kB (H) 6*128kB (H) 2*256kB (H) 0*512kB 0*1024kB 0*2048kB 0*4096kB = 14232kB
This is trace log which shows GFP_HIGHUSER consumes free pages right
before ALLOC_NO_WATERMARKS.
<...>-22275 [006] .... 889.213383: mm_page_alloc: page=00000000d2be5665 pfn=970744 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213385: mm_page_alloc: page=000000004b2335c2 pfn=970745 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213387: mm_page_alloc: page=00000000017272e1 pfn=970278 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213389: mm_page_alloc: page=00000000c4be79fb pfn=970279 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213391: mm_page_alloc: page=00000000f8a51d4f pfn=970260 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213393: mm_page_alloc: page=000000006ba8f5ac pfn=970261 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213395: mm_page_alloc: page=00000000819f1cd3 pfn=970196 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213396: mm_page_alloc: page=00000000f6b72a64 pfn=970197 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
kswapd0-1207 [005] ...1 889.213398: mm_page_alloc: page= (null) pfn=0 order=0 migratetype=1 nr_free=3650 gfp_flags=GFP_NOWAIT|__GFP_HIGHMEM|__GFP_NOWARN|__GFP_MOVABLE
[jaewon31.kim@samsung.com: remove redundant code for high-order]
Link: http://lkml.kernel.org/r/20200623035242.27232-1-jaewon31.kim@samsung.com
Reported-by: Yong-Taek Lee <ytk.lee@samsung.com>
Suggested-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Jaewon Kim <jaewon31.kim@samsung.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Baoquan He <bhe@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Yong-Taek Lee <ytk.lee@samsung.com>
Cc: Michal Hocko <mhocko@kernel.org>
Link: http://lkml.kernel.org/r/20200619235958.11283-1-jaewon31.kim@samsung.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-06 23:25:20 -07:00
|
|
|
|
2023-01-13 11:12:15 +00:00
|
|
|
|
mm/page_alloc: keep track of free highatomic
OOM kills due to vastly overestimated free highatomic reserves were
observed:
... invoked oom-killer: gfp_mask=0x100cca(GFP_HIGHUSER_MOVABLE), order=0 ...
Node 0 Normal free:1482936kB boost:0kB min:410416kB low:739404kB high:1068392kB reserved_highatomic:1073152KB ...
Node 0 Normal: 1292*4kB (ME) 1920*8kB (E) 383*16kB (UE) 220*32kB (ME) 340*64kB (E) 2155*128kB (UE) 3243*256kB (UE) 615*512kB (U) 1*1024kB (M) 0*2048kB 0*4096kB = 1477408kB
The second line above shows that the OOM kill was due to the following
condition:
free (1482936kB) - reserved_highatomic (1073152kB) = 409784KB < min (410416kB)
And the third line shows there were no free pages in any
MIGRATE_HIGHATOMIC pageblocks, which otherwise would show up as type 'H'.
Therefore __zone_watermark_unusable_free() underestimated the usable free
memory by over 1GB, which resulted in the unnecessary OOM kill above.
The comments in __zone_watermark_unusable_free() warns about the potential
risk, i.e.,
If the caller does not have rights to reserves below the min
watermark then subtract the high-atomic reserves. This will
over-estimate the size of the atomic reserve but it avoids a search.
However, it is possible to keep track of free pages in reserved highatomic
pageblocks with a new per-zone counter nr_free_highatomic protected by the
zone lock, to avoid a search when calculating the usable free memory. And
the cost would be minimal, i.e., simple arithmetics in the highatomic
alloc/free/move paths.
Note that since nr_free_highatomic can be relatively small, using a
per-cpu counter might cause too much drift and defeat its purpose, in
addition to the extra memory overhead.
Dependson e0932b6c1f94 ("mm: page_alloc: consolidate free page accounting") - see [1]
[akpm@linux-foundation.org: s/if/else if/, per Johannes, stealth whitespace tweak]
Link: https://lkml.kernel.org/r/20241028182653.3420139-1-yuzhao@google.com
Link: https://lkml.kernel.org/r/0d0ddb33-fcdc-43e2-801f-0c1df2031afb@suse.cz [1]
Fixes: 0aaa29a56e4f ("mm, page_alloc: reserve pageblocks for high-order atomic allocations on demand")
Signed-off-by: Yu Zhao <yuzhao@google.com>
Reported-by: Link Lin <linkl@google.com>
Acked-by: David Rientjes <rientjes@google.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-10-28 12:26:53 -06:00
|
|
|
|
page_alloc: consider highatomic reserve in watermark fast
zone_watermark_fast was introduced by commit 48ee5f3696f6 ("mm,
page_alloc: shortcut watermark checks for order-0 pages"). The commit
simply checks if free pages is bigger than watermark without additional
calculation such like reducing watermark.
It considered free cma pages but it did not consider highatomic reserved.
This may incur exhaustion of free pages except high order atomic free
pages.
Assume that reserved_highatomic pageblock is bigger than watermark min,
and there are only few free pages except high order atomic free. Because
zone_watermark_fast passes the allocation without considering high order
atomic free, normal reclaimable allocation like GFP_HIGHUSER will consume
all the free pages. Then finally order-0 atomic allocation may fail on
allocation.
This means watermark min is not protected against non-atomic allocation.
The order-0 atomic allocation with ALLOC_HARDER unwantedly can be failed.
Additionally the __GFP_MEMALLOC allocation with ALLOC_NO_WATERMARKS also
can be failed.
To avoid the problem, zone_watermark_fast should consider highatomic
reserve. If the actual size of high atomic free is counted accurately
like cma free, we may use it. On this patch just use
nr_reserved_highatomic. Additionally introduce
__zone_watermark_unusable_free to factor out common parts between
zone_watermark_fast and __zone_watermark_ok.
This is an example of ALLOC_HARDER allocation failure using v4.19 based
kernel.
Binder:9343_3: page allocation failure: order:0, mode:0x480020(GFP_ATOMIC), nodemask=(null)
Call trace:
[<ffffff8008f40f8c>] dump_stack+0xb8/0xf0
[<ffffff8008223320>] warn_alloc+0xd8/0x12c
[<ffffff80082245e4>] __alloc_pages_nodemask+0x120c/0x1250
[<ffffff800827f6e8>] new_slab+0x128/0x604
[<ffffff800827b0cc>] ___slab_alloc+0x508/0x670
[<ffffff800827ba00>] __kmalloc+0x2f8/0x310
[<ffffff80084ac3e0>] context_struct_to_string+0x104/0x1cc
[<ffffff80084ad8fc>] security_sid_to_context_core+0x74/0x144
[<ffffff80084ad880>] security_sid_to_context+0x10/0x18
[<ffffff800849bd80>] selinux_secid_to_secctx+0x20/0x28
[<ffffff800849109c>] security_secid_to_secctx+0x3c/0x70
[<ffffff8008bfe118>] binder_transaction+0xe68/0x454c
Mem-Info:
active_anon:102061 inactive_anon:81551 isolated_anon:0
active_file:59102 inactive_file:68924 isolated_file:64
unevictable:611 dirty:63 writeback:0 unstable:0
slab_reclaimable:13324 slab_unreclaimable:44354
mapped:83015 shmem:4858 pagetables:26316 bounce:0
free:2727 free_pcp:1035 free_cma:178
Node 0 active_anon:408244kB inactive_anon:326204kB active_file:236408kB inactive_file:275696kB unevictable:2444kB isolated(anon):0kB isolated(file):256kB mapped:332060kB dirty:252kB writeback:0kB shmem:19432kB writeback_tmp:0kB unstable:0kB all_unreclaimable? no
Normal free:10908kB min:6192kB low:44388kB high:47060kB active_anon:409160kB inactive_anon:325924kB active_file:235820kB inactive_file:276628kB unevictable:2444kB writepending:252kB present:3076096kB managed:2673676kB mlocked:2444kB kernel_stack:62512kB pagetables:105264kB bounce:0kB free_pcp:4140kB local_pcp:40kB free_cma:712kB
lowmem_reserve[]: 0 0
Normal: 505*4kB (H) 357*8kB (H) 201*16kB (H) 65*32kB (H) 1*64kB (H) 0*128kB 0*256kB 0*512kB 0*1024kB 0*2048kB 0*4096kB = 10236kB
138826 total pagecache pages
5460 pages in swap cache
Swap cache stats: add 8273090, delete 8267506, find 1004381/4060142
This is an example of ALLOC_NO_WATERMARKS allocation failure using v4.14
based kernel.
kswapd0: page allocation failure: order:0, mode:0x140000a(GFP_NOIO|__GFP_HIGHMEM|__GFP_MOVABLE), nodemask=(null)
kswapd0 cpuset=/ mems_allowed=0
CPU: 4 PID: 1221 Comm: kswapd0 Not tainted 4.14.113-18770262-userdebug #1
Call trace:
[<0000000000000000>] dump_backtrace+0x0/0x248
[<0000000000000000>] show_stack+0x18/0x20
[<0000000000000000>] __dump_stack+0x20/0x28
[<0000000000000000>] dump_stack+0x68/0x90
[<0000000000000000>] warn_alloc+0x104/0x198
[<0000000000000000>] __alloc_pages_nodemask+0xdc0/0xdf0
[<0000000000000000>] zs_malloc+0x148/0x3d0
[<0000000000000000>] zram_bvec_rw+0x410/0x798
[<0000000000000000>] zram_rw_page+0x88/0xdc
[<0000000000000000>] bdev_write_page+0x70/0xbc
[<0000000000000000>] __swap_writepage+0x58/0x37c
[<0000000000000000>] swap_writepage+0x40/0x4c
[<0000000000000000>] shrink_page_list+0xc30/0xf48
[<0000000000000000>] shrink_inactive_list+0x2b0/0x61c
[<0000000000000000>] shrink_node_memcg+0x23c/0x618
[<0000000000000000>] shrink_node+0x1c8/0x304
[<0000000000000000>] kswapd+0x680/0x7c4
[<0000000000000000>] kthread+0x110/0x120
[<0000000000000000>] ret_from_fork+0x10/0x18
Mem-Info:
active_anon:111826 inactive_anon:65557 isolated_anon:0\x0a active_file:44260 inactive_file:83422 isolated_file:0\x0a unevictable:4158 dirty:117 writeback:0 unstable:0\x0a slab_reclaimable:13943 slab_unreclaimable:43315\x0a mapped:102511 shmem:3299 pagetables:19566 bounce:0\x0a free:3510 free_pcp:553 free_cma:0
Node 0 active_anon:447304kB inactive_anon:262228kB active_file:177040kB inactive_file:333688kB unevictable:16632kB isolated(anon):0kB isolated(file):0kB mapped:410044kB d irty:468kB writeback:0kB shmem:13196kB writeback_tmp:0kB unstable:0kB all_unreclaimable? no
Normal free:14040kB min:7440kB low:94500kB high:98136kB reserved_highatomic:32768KB active_anon:447336kB inactive_anon:261668kB active_file:177572kB inactive_file:333768k B unevictable:16632kB writepending:480kB present:4081664kB managed:3637088kB mlocked:16632kB kernel_stack:47072kB pagetables:78264kB bounce:0kB free_pcp:2280kB local_pcp:720kB free_cma:0kB [ 4738.329607] lowmem_reserve[]: 0 0
Normal: 860*4kB (H) 453*8kB (H) 180*16kB (H) 26*32kB (H) 34*64kB (H) 6*128kB (H) 2*256kB (H) 0*512kB 0*1024kB 0*2048kB 0*4096kB = 14232kB
This is trace log which shows GFP_HIGHUSER consumes free pages right
before ALLOC_NO_WATERMARKS.
<...>-22275 [006] .... 889.213383: mm_page_alloc: page=00000000d2be5665 pfn=970744 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213385: mm_page_alloc: page=000000004b2335c2 pfn=970745 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213387: mm_page_alloc: page=00000000017272e1 pfn=970278 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213389: mm_page_alloc: page=00000000c4be79fb pfn=970279 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213391: mm_page_alloc: page=00000000f8a51d4f pfn=970260 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213393: mm_page_alloc: page=000000006ba8f5ac pfn=970261 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213395: mm_page_alloc: page=00000000819f1cd3 pfn=970196 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213396: mm_page_alloc: page=00000000f6b72a64 pfn=970197 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
kswapd0-1207 [005] ...1 889.213398: mm_page_alloc: page= (null) pfn=0 order=0 migratetype=1 nr_free=3650 gfp_flags=GFP_NOWAIT|__GFP_HIGHMEM|__GFP_NOWARN|__GFP_MOVABLE
[jaewon31.kim@samsung.com: remove redundant code for high-order]
Link: http://lkml.kernel.org/r/20200623035242.27232-1-jaewon31.kim@samsung.com
Reported-by: Yong-Taek Lee <ytk.lee@samsung.com>
Suggested-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Jaewon Kim <jaewon31.kim@samsung.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Baoquan He <bhe@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Yong-Taek Lee <ytk.lee@samsung.com>
Cc: Michal Hocko <mhocko@kernel.org>
Link: http://lkml.kernel.org/r/20200619235958.11283-1-jaewon31.kim@samsung.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-06 23:25:20 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2015-11-06 16:28:40 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2016-05-20 16:57:12 -07:00
|
|
|
|
2020-06-03 15:59:01 -07:00
|
|
|
|
2016-05-20 16:57:12 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2007-02-10 01:43:02 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2015-11-06 16:28:37 -08:00
|
|
|
|
page_alloc: consider highatomic reserve in watermark fast
zone_watermark_fast was introduced by commit 48ee5f3696f6 ("mm,
page_alloc: shortcut watermark checks for order-0 pages"). The commit
simply checks if free pages is bigger than watermark without additional
calculation such like reducing watermark.
It considered free cma pages but it did not consider highatomic reserved.
This may incur exhaustion of free pages except high order atomic free
pages.
Assume that reserved_highatomic pageblock is bigger than watermark min,
and there are only few free pages except high order atomic free. Because
zone_watermark_fast passes the allocation without considering high order
atomic free, normal reclaimable allocation like GFP_HIGHUSER will consume
all the free pages. Then finally order-0 atomic allocation may fail on
allocation.
This means watermark min is not protected against non-atomic allocation.
The order-0 atomic allocation with ALLOC_HARDER unwantedly can be failed.
Additionally the __GFP_MEMALLOC allocation with ALLOC_NO_WATERMARKS also
can be failed.
To avoid the problem, zone_watermark_fast should consider highatomic
reserve. If the actual size of high atomic free is counted accurately
like cma free, we may use it. On this patch just use
nr_reserved_highatomic. Additionally introduce
__zone_watermark_unusable_free to factor out common parts between
zone_watermark_fast and __zone_watermark_ok.
This is an example of ALLOC_HARDER allocation failure using v4.19 based
kernel.
Binder:9343_3: page allocation failure: order:0, mode:0x480020(GFP_ATOMIC), nodemask=(null)
Call trace:
[<ffffff8008f40f8c>] dump_stack+0xb8/0xf0
[<ffffff8008223320>] warn_alloc+0xd8/0x12c
[<ffffff80082245e4>] __alloc_pages_nodemask+0x120c/0x1250
[<ffffff800827f6e8>] new_slab+0x128/0x604
[<ffffff800827b0cc>] ___slab_alloc+0x508/0x670
[<ffffff800827ba00>] __kmalloc+0x2f8/0x310
[<ffffff80084ac3e0>] context_struct_to_string+0x104/0x1cc
[<ffffff80084ad8fc>] security_sid_to_context_core+0x74/0x144
[<ffffff80084ad880>] security_sid_to_context+0x10/0x18
[<ffffff800849bd80>] selinux_secid_to_secctx+0x20/0x28
[<ffffff800849109c>] security_secid_to_secctx+0x3c/0x70
[<ffffff8008bfe118>] binder_transaction+0xe68/0x454c
Mem-Info:
active_anon:102061 inactive_anon:81551 isolated_anon:0
active_file:59102 inactive_file:68924 isolated_file:64
unevictable:611 dirty:63 writeback:0 unstable:0
slab_reclaimable:13324 slab_unreclaimable:44354
mapped:83015 shmem:4858 pagetables:26316 bounce:0
free:2727 free_pcp:1035 free_cma:178
Node 0 active_anon:408244kB inactive_anon:326204kB active_file:236408kB inactive_file:275696kB unevictable:2444kB isolated(anon):0kB isolated(file):256kB mapped:332060kB dirty:252kB writeback:0kB shmem:19432kB writeback_tmp:0kB unstable:0kB all_unreclaimable? no
Normal free:10908kB min:6192kB low:44388kB high:47060kB active_anon:409160kB inactive_anon:325924kB active_file:235820kB inactive_file:276628kB unevictable:2444kB writepending:252kB present:3076096kB managed:2673676kB mlocked:2444kB kernel_stack:62512kB pagetables:105264kB bounce:0kB free_pcp:4140kB local_pcp:40kB free_cma:712kB
lowmem_reserve[]: 0 0
Normal: 505*4kB (H) 357*8kB (H) 201*16kB (H) 65*32kB (H) 1*64kB (H) 0*128kB 0*256kB 0*512kB 0*1024kB 0*2048kB 0*4096kB = 10236kB
138826 total pagecache pages
5460 pages in swap cache
Swap cache stats: add 8273090, delete 8267506, find 1004381/4060142
This is an example of ALLOC_NO_WATERMARKS allocation failure using v4.14
based kernel.
kswapd0: page allocation failure: order:0, mode:0x140000a(GFP_NOIO|__GFP_HIGHMEM|__GFP_MOVABLE), nodemask=(null)
kswapd0 cpuset=/ mems_allowed=0
CPU: 4 PID: 1221 Comm: kswapd0 Not tainted 4.14.113-18770262-userdebug #1
Call trace:
[<0000000000000000>] dump_backtrace+0x0/0x248
[<0000000000000000>] show_stack+0x18/0x20
[<0000000000000000>] __dump_stack+0x20/0x28
[<0000000000000000>] dump_stack+0x68/0x90
[<0000000000000000>] warn_alloc+0x104/0x198
[<0000000000000000>] __alloc_pages_nodemask+0xdc0/0xdf0
[<0000000000000000>] zs_malloc+0x148/0x3d0
[<0000000000000000>] zram_bvec_rw+0x410/0x798
[<0000000000000000>] zram_rw_page+0x88/0xdc
[<0000000000000000>] bdev_write_page+0x70/0xbc
[<0000000000000000>] __swap_writepage+0x58/0x37c
[<0000000000000000>] swap_writepage+0x40/0x4c
[<0000000000000000>] shrink_page_list+0xc30/0xf48
[<0000000000000000>] shrink_inactive_list+0x2b0/0x61c
[<0000000000000000>] shrink_node_memcg+0x23c/0x618
[<0000000000000000>] shrink_node+0x1c8/0x304
[<0000000000000000>] kswapd+0x680/0x7c4
[<0000000000000000>] kthread+0x110/0x120
[<0000000000000000>] ret_from_fork+0x10/0x18
Mem-Info:
active_anon:111826 inactive_anon:65557 isolated_anon:0\x0a active_file:44260 inactive_file:83422 isolated_file:0\x0a unevictable:4158 dirty:117 writeback:0 unstable:0\x0a slab_reclaimable:13943 slab_unreclaimable:43315\x0a mapped:102511 shmem:3299 pagetables:19566 bounce:0\x0a free:3510 free_pcp:553 free_cma:0
Node 0 active_anon:447304kB inactive_anon:262228kB active_file:177040kB inactive_file:333688kB unevictable:16632kB isolated(anon):0kB isolated(file):0kB mapped:410044kB d irty:468kB writeback:0kB shmem:13196kB writeback_tmp:0kB unstable:0kB all_unreclaimable? no
Normal free:14040kB min:7440kB low:94500kB high:98136kB reserved_highatomic:32768KB active_anon:447336kB inactive_anon:261668kB active_file:177572kB inactive_file:333768k B unevictable:16632kB writepending:480kB present:4081664kB managed:3637088kB mlocked:16632kB kernel_stack:47072kB pagetables:78264kB bounce:0kB free_pcp:2280kB local_pcp:720kB free_cma:0kB [ 4738.329607] lowmem_reserve[]: 0 0
Normal: 860*4kB (H) 453*8kB (H) 180*16kB (H) 26*32kB (H) 34*64kB (H) 6*128kB (H) 2*256kB (H) 0*512kB 0*1024kB 0*2048kB 0*4096kB = 14232kB
This is trace log which shows GFP_HIGHUSER consumes free pages right
before ALLOC_NO_WATERMARKS.
<...>-22275 [006] .... 889.213383: mm_page_alloc: page=00000000d2be5665 pfn=970744 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213385: mm_page_alloc: page=000000004b2335c2 pfn=970745 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213387: mm_page_alloc: page=00000000017272e1 pfn=970278 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213389: mm_page_alloc: page=00000000c4be79fb pfn=970279 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213391: mm_page_alloc: page=00000000f8a51d4f pfn=970260 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213393: mm_page_alloc: page=000000006ba8f5ac pfn=970261 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213395: mm_page_alloc: page=00000000819f1cd3 pfn=970196 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213396: mm_page_alloc: page=00000000f6b72a64 pfn=970197 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
kswapd0-1207 [005] ...1 889.213398: mm_page_alloc: page= (null) pfn=0 order=0 migratetype=1 nr_free=3650 gfp_flags=GFP_NOWAIT|__GFP_HIGHMEM|__GFP_NOWARN|__GFP_MOVABLE
[jaewon31.kim@samsung.com: remove redundant code for high-order]
Link: http://lkml.kernel.org/r/20200623035242.27232-1-jaewon31.kim@samsung.com
Reported-by: Yong-Taek Lee <ytk.lee@samsung.com>
Suggested-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Jaewon Kim <jaewon31.kim@samsung.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Baoquan He <bhe@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Yong-Taek Lee <ytk.lee@samsung.com>
Cc: Michal Hocko <mhocko@kernel.org>
Link: http://lkml.kernel.org/r/20200619235958.11283-1-jaewon31.kim@samsung.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-06 23:25:20 -07:00
|
|
|
|
2015-11-06 16:28:37 -08:00
|
|
|
|
2023-01-13 11:12:15 +00:00
|
|
|
|
2017-09-06 16:24:50 -07:00
|
|
|
|
2023-01-13 11:12:15 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-01-13 11:12:16 +00:00
|
|
|
|
2023-01-13 11:12:15 +00:00
|
|
|
|
|
|
|
|
|
2023-01-13 11:12:16 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-01-13 11:12:15 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-09-06 16:24:50 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-11-06 16:28:40 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-06-03 15:59:01 -07:00
|
|
|
|
mm: page allocator: adjust the per-cpu counter threshold when memory is low
Commit aa45484 ("calculate a better estimate of NR_FREE_PAGES when memory
is low") noted that watermarks were based on the vmstat NR_FREE_PAGES. To
avoid synchronization overhead, these counters are maintained on a per-cpu
basis and drained both periodically and when a threshold is above a
threshold. On large CPU systems, the difference between the estimate and
real value of NR_FREE_PAGES can be very high. The system can get into a
case where pages are allocated far below the min watermark potentially
causing livelock issues. The commit solved the problem by taking a better
reading of NR_FREE_PAGES when memory was low.
Unfortately, as reported by Shaohua Li this accurate reading can consume a
large amount of CPU time on systems with many sockets due to cache line
bouncing. This patch takes a different approach. For large machines
where counter drift might be unsafe and while kswapd is awake, the per-cpu
thresholds for the target pgdat are reduced to limit the level of drift to
what should be a safe level. This incurs a performance penalty in heavy
memory pressure by a factor that depends on the workload and the machine
but the machine should function correctly without accidentally exhausting
all memory on a node. There is an additional cost when kswapd wakes and
sleeps but the event is not expected to be frequent - in Shaohua's test
case, there was one recorded sleep and wake event at least.
To ensure that kswapd wakes up, a safe version of zone_watermark_ok() is
introduced that takes a more accurate reading of NR_FREE_PAGES when called
from wakeup_kswapd, when deciding whether it is really safe to go back to
sleep in sleeping_prematurely() and when deciding if a zone is really
balanced or not in balance_pgdat(). We are still using an expensive
function but limiting how often it is called.
When the test case is reproduced, the time spent in the watermark
functions is reduced. The following report is on the percentage of time
spent cumulatively spent in the functions zone_nr_free_pages(),
zone_watermark_ok(), __zone_watermark_ok(), zone_watermark_ok_safe(),
zone_page_state_snapshot(), zone_page_state().
vanilla 11.6615%
disable-threshold 0.2584%
David said:
: We had to pull aa454840 "mm: page allocator: calculate a better estimate
: of NR_FREE_PAGES when memory is low and kswapd is awake" from 2.6.36
: internally because tests showed that it would cause the machine to stall
: as the result of heavy kswapd activity. I merged it back with this fix as
: it is pending in the -mm tree and it solves the issue we were seeing, so I
: definitely think this should be pushed to -stable (and I would seriously
: consider it for 2.6.37 inclusion even at this late date).
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reported-by: Shaohua Li <shaohua.li@intel.com>
Reviewed-by: Christoph Lameter <cl@linux.com>
Tested-by: Nicolas Bareil <nico@chdir.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Kyle McMartin <kyle@mcmartin.ca>
Cc: <stable@kernel.org> [2.6.37.1, 2.6.36.x]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-13 15:45:41 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2015-11-06 16:28:40 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2015-11-06 16:28:40 -08:00
|
|
|
|
2023-12-28 17:47:03 +03:00
|
|
|
|
2015-11-06 16:28:40 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-05-14 15:41:32 -07:00
|
|
|
|
2015-11-06 16:28:40 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-05-23 10:18:21 +09:00
|
|
|
|
2019-05-14 15:41:32 -07:00
|
|
|
|
2015-11-06 16:28:40 -08:00
|
|
|
|
2018-05-23 10:18:21 +09:00
|
|
|
|
2015-11-06 16:28:40 -08:00
|
|
|
|
2023-01-13 11:12:14 +00:00
|
|
|
|
|
|
|
|
|
mm, page_alloc: fix potential false positive in __zone_watermark_ok
Since commit 97a16fc82a7c ("mm, page_alloc: only enforce watermarks for
order-0 allocations"), __zone_watermark_ok() check for high-order
allocations will shortcut per-migratetype free list checks for
ALLOC_HARDER allocations, and return true as long as there's free page
of any migratetype. The intention is that ALLOC_HARDER can allocate
from MIGRATE_HIGHATOMIC free lists, while normal allocations can't.
However, as a side effect, the watermark check will then also return
true when there are pages only on the MIGRATE_ISOLATE list, or (prior to
CMA conversion to ZONE_MOVABLE) on the MIGRATE_CMA list. Since the
allocation cannot actually obtain isolated pages, and might not be able
to obtain CMA pages, this can result in a false positive.
The condition should be rare and perhaps the outcome is not a fatal one.
Still, it's better if the watermark check is correct. There also
shouldn't be a performance tradeoff here.
Link: http://lkml.kernel.org/r/20171102125001.23708-1-vbabka@suse.cz
Fixes: 97a16fc82a7c ("mm, page_alloc: only enforce watermarks for order-0 allocations")
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-15 17:38:30 -08:00
|
|
|
|
2023-01-13 11:12:14 +00:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2015-11-06 16:28:40 -08:00
|
|
|
|
mm: page allocator: adjust the per-cpu counter threshold when memory is low
Commit aa45484 ("calculate a better estimate of NR_FREE_PAGES when memory
is low") noted that watermarks were based on the vmstat NR_FREE_PAGES. To
avoid synchronization overhead, these counters are maintained on a per-cpu
basis and drained both periodically and when a threshold is above a
threshold. On large CPU systems, the difference between the estimate and
real value of NR_FREE_PAGES can be very high. The system can get into a
case where pages are allocated far below the min watermark potentially
causing livelock issues. The commit solved the problem by taking a better
reading of NR_FREE_PAGES when memory was low.
Unfortately, as reported by Shaohua Li this accurate reading can consume a
large amount of CPU time on systems with many sockets due to cache line
bouncing. This patch takes a different approach. For large machines
where counter drift might be unsafe and while kswapd is awake, the per-cpu
thresholds for the target pgdat are reduced to limit the level of drift to
what should be a safe level. This incurs a performance penalty in heavy
memory pressure by a factor that depends on the workload and the machine
but the machine should function correctly without accidentally exhausting
all memory on a node. There is an additional cost when kswapd wakes and
sleeps but the event is not expected to be frequent - in Shaohua's test
case, there was one recorded sleep and wake event at least.
To ensure that kswapd wakes up, a safe version of zone_watermark_ok() is
introduced that takes a more accurate reading of NR_FREE_PAGES when called
from wakeup_kswapd, when deciding whether it is really safe to go back to
sleep in sleeping_prematurely() and when deciding if a zone is really
balanced or not in balance_pgdat(). We are still using an expensive
function but limiting how often it is called.
When the test case is reproduced, the time spent in the watermark
functions is reduced. The following report is on the percentage of time
spent cumulatively spent in the functions zone_nr_free_pages(),
zone_watermark_ok(), __zone_watermark_ok(), zone_watermark_ok_safe(),
zone_page_state_snapshot(), zone_page_state().
vanilla 11.6615%
disable-threshold 0.2584%
David said:
: We had to pull aa454840 "mm: page allocator: calculate a better estimate
: of NR_FREE_PAGES when memory is low and kswapd is awake" from 2.6.36
: internally because tests showed that it would cause the machine to stall
: as the result of heavy kswapd activity. I merged it back with this fix as
: it is pending in the -mm tree and it solves the issue we were seeing, so I
: definitely think this should be pushed to -stable (and I would seriously
: consider it for 2.6.37 inclusion even at this late date).
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reported-by: Shaohua Li <shaohua.li@intel.com>
Reviewed-by: Christoph Lameter <cl@linux.com>
Tested-by: Nicolas Bareil <nico@chdir.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Kyle McMartin <kyle@mcmartin.ca>
Cc: <stable@kernel.org> [2.6.37.1, 2.6.36.x]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-13 15:45:41 -08:00
|
|
|
|
|
|
|
|
|
2014-06-04 16:10:21 -07:00
|
|
|
|
2020-06-03 15:59:01 -07:00
|
|
|
|
mm: page allocator: adjust the per-cpu counter threshold when memory is low
Commit aa45484 ("calculate a better estimate of NR_FREE_PAGES when memory
is low") noted that watermarks were based on the vmstat NR_FREE_PAGES. To
avoid synchronization overhead, these counters are maintained on a per-cpu
basis and drained both periodically and when a threshold is above a
threshold. On large CPU systems, the difference between the estimate and
real value of NR_FREE_PAGES can be very high. The system can get into a
case where pages are allocated far below the min watermark potentially
causing livelock issues. The commit solved the problem by taking a better
reading of NR_FREE_PAGES when memory was low.
Unfortately, as reported by Shaohua Li this accurate reading can consume a
large amount of CPU time on systems with many sockets due to cache line
bouncing. This patch takes a different approach. For large machines
where counter drift might be unsafe and while kswapd is awake, the per-cpu
thresholds for the target pgdat are reduced to limit the level of drift to
what should be a safe level. This incurs a performance penalty in heavy
memory pressure by a factor that depends on the workload and the machine
but the machine should function correctly without accidentally exhausting
all memory on a node. There is an additional cost when kswapd wakes and
sleeps but the event is not expected to be frequent - in Shaohua's test
case, there was one recorded sleep and wake event at least.
To ensure that kswapd wakes up, a safe version of zone_watermark_ok() is
introduced that takes a more accurate reading of NR_FREE_PAGES when called
from wakeup_kswapd, when deciding whether it is really safe to go back to
sleep in sleeping_prematurely() and when deciding if a zone is really
balanced or not in balance_pgdat(). We are still using an expensive
function but limiting how often it is called.
When the test case is reproduced, the time spent in the watermark
functions is reduced. The following report is on the percentage of time
spent cumulatively spent in the functions zone_nr_free_pages(),
zone_watermark_ok(), __zone_watermark_ok(), zone_watermark_ok_safe(),
zone_page_state_snapshot(), zone_page_state().
vanilla 11.6615%
disable-threshold 0.2584%
David said:
: We had to pull aa454840 "mm: page allocator: calculate a better estimate
: of NR_FREE_PAGES when memory is low and kswapd is awake" from 2.6.36
: internally because tests showed that it would cause the machine to stall
: as the result of heavy kswapd activity. I merged it back with this fix as
: it is pending in the -mm tree and it solves the issue we were seeing, so I
: definitely think this should be pushed to -stable (and I would seriously
: consider it for 2.6.37 inclusion even at this late date).
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reported-by: Shaohua Li <shaohua.li@intel.com>
Reviewed-by: Christoph Lameter <cl@linux.com>
Tested-by: Nicolas Bareil <nico@chdir.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Kyle McMartin <kyle@mcmartin.ca>
Cc: <stable@kernel.org> [2.6.37.1, 2.6.36.x]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-13 15:45:41 -08:00
|
|
|
|
2020-06-03 15:59:01 -07:00
|
|
|
|
mm: page allocator: adjust the per-cpu counter threshold when memory is low
Commit aa45484 ("calculate a better estimate of NR_FREE_PAGES when memory
is low") noted that watermarks were based on the vmstat NR_FREE_PAGES. To
avoid synchronization overhead, these counters are maintained on a per-cpu
basis and drained both periodically and when a threshold is above a
threshold. On large CPU systems, the difference between the estimate and
real value of NR_FREE_PAGES can be very high. The system can get into a
case where pages are allocated far below the min watermark potentially
causing livelock issues. The commit solved the problem by taking a better
reading of NR_FREE_PAGES when memory was low.
Unfortately, as reported by Shaohua Li this accurate reading can consume a
large amount of CPU time on systems with many sockets due to cache line
bouncing. This patch takes a different approach. For large machines
where counter drift might be unsafe and while kswapd is awake, the per-cpu
thresholds for the target pgdat are reduced to limit the level of drift to
what should be a safe level. This incurs a performance penalty in heavy
memory pressure by a factor that depends on the workload and the machine
but the machine should function correctly without accidentally exhausting
all memory on a node. There is an additional cost when kswapd wakes and
sleeps but the event is not expected to be frequent - in Shaohua's test
case, there was one recorded sleep and wake event at least.
To ensure that kswapd wakes up, a safe version of zone_watermark_ok() is
introduced that takes a more accurate reading of NR_FREE_PAGES when called
from wakeup_kswapd, when deciding whether it is really safe to go back to
sleep in sleeping_prematurely() and when deciding if a zone is really
balanced or not in balance_pgdat(). We are still using an expensive
function but limiting how often it is called.
When the test case is reproduced, the time spent in the watermark
functions is reduced. The following report is on the percentage of time
spent cumulatively spent in the functions zone_nr_free_pages(),
zone_watermark_ok(), __zone_watermark_ok(), zone_watermark_ok_safe(),
zone_page_state_snapshot(), zone_page_state().
vanilla 11.6615%
disable-threshold 0.2584%
David said:
: We had to pull aa454840 "mm: page allocator: calculate a better estimate
: of NR_FREE_PAGES when memory is low and kswapd is awake" from 2.6.36
: internally because tests showed that it would cause the machine to stall
: as the result of heavy kswapd activity. I merged it back with this fix as
: it is pending in the -mm tree and it solves the issue we were seeing, so I
: definitely think this should be pushed to -stable (and I would seriously
: consider it for 2.6.37 inclusion even at this late date).
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reported-by: Shaohua Li <shaohua.li@intel.com>
Reviewed-by: Christoph Lameter <cl@linux.com>
Tested-by: Nicolas Bareil <nico@chdir.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Kyle McMartin <kyle@mcmartin.ca>
Cc: <stable@kernel.org> [2.6.37.1, 2.6.36.x]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-13 15:45:41 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-05-19 17:14:07 -07:00
|
|
|
|
2020-06-03 15:59:01 -07:00
|
|
|
|
mm, page_alloc: skip ->waternark_boost for atomic order-0 allocations
When boosting is enabled, it is observed that rate of atomic order-0
allocation failures are high due to the fact that free levels in the
system are checked with ->watermark_boost offset. This is not a problem
for sleepable allocations but for atomic allocations which looks like
regression.
This problem is seen frequently on system setup of Android kernel running
on Snapdragon hardware with 4GB RAM size. When no extfrag event occurred
in the system, ->watermark_boost factor is zero, thus the watermark
configurations in the system are:
_watermark = (
[WMARK_MIN] = 1272, --> ~5MB
[WMARK_LOW] = 9067, --> ~36MB
[WMARK_HIGH] = 9385), --> ~38MB
watermark_boost = 0
After launching some memory hungry applications in Android which can cause
extfrag events in the system to an extent that ->watermark_boost can be
set to max i.e. default boost factor makes it to 150% of high watermark.
_watermark = (
[WMARK_MIN] = 1272, --> ~5MB
[WMARK_LOW] = 9067, --> ~36MB
[WMARK_HIGH] = 9385), --> ~38MB
watermark_boost = 14077, -->~57MB
With default system configuration, for an atomic order-0 allocation to
succeed, having free memory of ~2MB will suffice. But boosting makes the
min_wmark to ~61MB thus for an atomic order-0 allocation to be successful
system should have minimum of ~23MB of free memory(from calculations of
zone_watermark_ok(), min = 3/4(min/2)). But failures are observed despite
system is having ~20MB of free memory. In the testing, this is
reproducible as early as first 300secs since boot and with furtherlowram
configurations(<2GB) it is observed as early as first 150secs since boot.
These failures can be avoided by excluding the ->watermark_boost in
watermark caluculations for atomic order-0 allocations.
[akpm@linux-foundation.org: fix comment grammar, reflow comment]
[charante@codeaurora.org: fix suggested by Mel Gorman]
Link: http://lkml.kernel.org/r/31556793-57b1-1c21-1a9d-22674d9bd938@codeaurora.org
Signed-off-by: Charan Teja Reddy <charante@codeaurora.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Vinayak Menon <vinmenon@codeaurora.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Link: http://lkml.kernel.org/r/1589882284-21010-1-git-send-email-charante@codeaurora.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-06 23:25:24 -07:00
|
|
|
|
2016-05-19 17:14:07 -07:00
|
|
|
|
page_alloc: consider highatomic reserve in watermark fast
zone_watermark_fast was introduced by commit 48ee5f3696f6 ("mm,
page_alloc: shortcut watermark checks for order-0 pages"). The commit
simply checks if free pages is bigger than watermark without additional
calculation such like reducing watermark.
It considered free cma pages but it did not consider highatomic reserved.
This may incur exhaustion of free pages except high order atomic free
pages.
Assume that reserved_highatomic pageblock is bigger than watermark min,
and there are only few free pages except high order atomic free. Because
zone_watermark_fast passes the allocation without considering high order
atomic free, normal reclaimable allocation like GFP_HIGHUSER will consume
all the free pages. Then finally order-0 atomic allocation may fail on
allocation.
This means watermark min is not protected against non-atomic allocation.
The order-0 atomic allocation with ALLOC_HARDER unwantedly can be failed.
Additionally the __GFP_MEMALLOC allocation with ALLOC_NO_WATERMARKS also
can be failed.
To avoid the problem, zone_watermark_fast should consider highatomic
reserve. If the actual size of high atomic free is counted accurately
like cma free, we may use it. On this patch just use
nr_reserved_highatomic. Additionally introduce
__zone_watermark_unusable_free to factor out common parts between
zone_watermark_fast and __zone_watermark_ok.
This is an example of ALLOC_HARDER allocation failure using v4.19 based
kernel.
Binder:9343_3: page allocation failure: order:0, mode:0x480020(GFP_ATOMIC), nodemask=(null)
Call trace:
[<ffffff8008f40f8c>] dump_stack+0xb8/0xf0
[<ffffff8008223320>] warn_alloc+0xd8/0x12c
[<ffffff80082245e4>] __alloc_pages_nodemask+0x120c/0x1250
[<ffffff800827f6e8>] new_slab+0x128/0x604
[<ffffff800827b0cc>] ___slab_alloc+0x508/0x670
[<ffffff800827ba00>] __kmalloc+0x2f8/0x310
[<ffffff80084ac3e0>] context_struct_to_string+0x104/0x1cc
[<ffffff80084ad8fc>] security_sid_to_context_core+0x74/0x144
[<ffffff80084ad880>] security_sid_to_context+0x10/0x18
[<ffffff800849bd80>] selinux_secid_to_secctx+0x20/0x28
[<ffffff800849109c>] security_secid_to_secctx+0x3c/0x70
[<ffffff8008bfe118>] binder_transaction+0xe68/0x454c
Mem-Info:
active_anon:102061 inactive_anon:81551 isolated_anon:0
active_file:59102 inactive_file:68924 isolated_file:64
unevictable:611 dirty:63 writeback:0 unstable:0
slab_reclaimable:13324 slab_unreclaimable:44354
mapped:83015 shmem:4858 pagetables:26316 bounce:0
free:2727 free_pcp:1035 free_cma:178
Node 0 active_anon:408244kB inactive_anon:326204kB active_file:236408kB inactive_file:275696kB unevictable:2444kB isolated(anon):0kB isolated(file):256kB mapped:332060kB dirty:252kB writeback:0kB shmem:19432kB writeback_tmp:0kB unstable:0kB all_unreclaimable? no
Normal free:10908kB min:6192kB low:44388kB high:47060kB active_anon:409160kB inactive_anon:325924kB active_file:235820kB inactive_file:276628kB unevictable:2444kB writepending:252kB present:3076096kB managed:2673676kB mlocked:2444kB kernel_stack:62512kB pagetables:105264kB bounce:0kB free_pcp:4140kB local_pcp:40kB free_cma:712kB
lowmem_reserve[]: 0 0
Normal: 505*4kB (H) 357*8kB (H) 201*16kB (H) 65*32kB (H) 1*64kB (H) 0*128kB 0*256kB 0*512kB 0*1024kB 0*2048kB 0*4096kB = 10236kB
138826 total pagecache pages
5460 pages in swap cache
Swap cache stats: add 8273090, delete 8267506, find 1004381/4060142
This is an example of ALLOC_NO_WATERMARKS allocation failure using v4.14
based kernel.
kswapd0: page allocation failure: order:0, mode:0x140000a(GFP_NOIO|__GFP_HIGHMEM|__GFP_MOVABLE), nodemask=(null)
kswapd0 cpuset=/ mems_allowed=0
CPU: 4 PID: 1221 Comm: kswapd0 Not tainted 4.14.113-18770262-userdebug #1
Call trace:
[<0000000000000000>] dump_backtrace+0x0/0x248
[<0000000000000000>] show_stack+0x18/0x20
[<0000000000000000>] __dump_stack+0x20/0x28
[<0000000000000000>] dump_stack+0x68/0x90
[<0000000000000000>] warn_alloc+0x104/0x198
[<0000000000000000>] __alloc_pages_nodemask+0xdc0/0xdf0
[<0000000000000000>] zs_malloc+0x148/0x3d0
[<0000000000000000>] zram_bvec_rw+0x410/0x798
[<0000000000000000>] zram_rw_page+0x88/0xdc
[<0000000000000000>] bdev_write_page+0x70/0xbc
[<0000000000000000>] __swap_writepage+0x58/0x37c
[<0000000000000000>] swap_writepage+0x40/0x4c
[<0000000000000000>] shrink_page_list+0xc30/0xf48
[<0000000000000000>] shrink_inactive_list+0x2b0/0x61c
[<0000000000000000>] shrink_node_memcg+0x23c/0x618
[<0000000000000000>] shrink_node+0x1c8/0x304
[<0000000000000000>] kswapd+0x680/0x7c4
[<0000000000000000>] kthread+0x110/0x120
[<0000000000000000>] ret_from_fork+0x10/0x18
Mem-Info:
active_anon:111826 inactive_anon:65557 isolated_anon:0\x0a active_file:44260 inactive_file:83422 isolated_file:0\x0a unevictable:4158 dirty:117 writeback:0 unstable:0\x0a slab_reclaimable:13943 slab_unreclaimable:43315\x0a mapped:102511 shmem:3299 pagetables:19566 bounce:0\x0a free:3510 free_pcp:553 free_cma:0
Node 0 active_anon:447304kB inactive_anon:262228kB active_file:177040kB inactive_file:333688kB unevictable:16632kB isolated(anon):0kB isolated(file):0kB mapped:410044kB d irty:468kB writeback:0kB shmem:13196kB writeback_tmp:0kB unstable:0kB all_unreclaimable? no
Normal free:14040kB min:7440kB low:94500kB high:98136kB reserved_highatomic:32768KB active_anon:447336kB inactive_anon:261668kB active_file:177572kB inactive_file:333768k B unevictable:16632kB writepending:480kB present:4081664kB managed:3637088kB mlocked:16632kB kernel_stack:47072kB pagetables:78264kB bounce:0kB free_pcp:2280kB local_pcp:720kB free_cma:0kB [ 4738.329607] lowmem_reserve[]: 0 0
Normal: 860*4kB (H) 453*8kB (H) 180*16kB (H) 26*32kB (H) 34*64kB (H) 6*128kB (H) 2*256kB (H) 0*512kB 0*1024kB 0*2048kB 0*4096kB = 14232kB
This is trace log which shows GFP_HIGHUSER consumes free pages right
before ALLOC_NO_WATERMARKS.
<...>-22275 [006] .... 889.213383: mm_page_alloc: page=00000000d2be5665 pfn=970744 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213385: mm_page_alloc: page=000000004b2335c2 pfn=970745 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213387: mm_page_alloc: page=00000000017272e1 pfn=970278 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213389: mm_page_alloc: page=00000000c4be79fb pfn=970279 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213391: mm_page_alloc: page=00000000f8a51d4f pfn=970260 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213393: mm_page_alloc: page=000000006ba8f5ac pfn=970261 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213395: mm_page_alloc: page=00000000819f1cd3 pfn=970196 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213396: mm_page_alloc: page=00000000f6b72a64 pfn=970197 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
kswapd0-1207 [005] ...1 889.213398: mm_page_alloc: page= (null) pfn=0 order=0 migratetype=1 nr_free=3650 gfp_flags=GFP_NOWAIT|__GFP_HIGHMEM|__GFP_NOWARN|__GFP_MOVABLE
[jaewon31.kim@samsung.com: remove redundant code for high-order]
Link: http://lkml.kernel.org/r/20200623035242.27232-1-jaewon31.kim@samsung.com
Reported-by: Yong-Taek Lee <ytk.lee@samsung.com>
Suggested-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Jaewon Kim <jaewon31.kim@samsung.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Baoquan He <bhe@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Yong-Taek Lee <ytk.lee@samsung.com>
Cc: Michal Hocko <mhocko@kernel.org>
Link: http://lkml.kernel.org/r/20200619235958.11283-1-jaewon31.kim@samsung.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-06 23:25:20 -07:00
|
|
|
|
2018-05-23 10:18:21 +09:00
|
|
|
|
page_alloc: consider highatomic reserve in watermark fast
zone_watermark_fast was introduced by commit 48ee5f3696f6 ("mm,
page_alloc: shortcut watermark checks for order-0 pages"). The commit
simply checks if free pages is bigger than watermark without additional
calculation such like reducing watermark.
It considered free cma pages but it did not consider highatomic reserved.
This may incur exhaustion of free pages except high order atomic free
pages.
Assume that reserved_highatomic pageblock is bigger than watermark min,
and there are only few free pages except high order atomic free. Because
zone_watermark_fast passes the allocation without considering high order
atomic free, normal reclaimable allocation like GFP_HIGHUSER will consume
all the free pages. Then finally order-0 atomic allocation may fail on
allocation.
This means watermark min is not protected against non-atomic allocation.
The order-0 atomic allocation with ALLOC_HARDER unwantedly can be failed.
Additionally the __GFP_MEMALLOC allocation with ALLOC_NO_WATERMARKS also
can be failed.
To avoid the problem, zone_watermark_fast should consider highatomic
reserve. If the actual size of high atomic free is counted accurately
like cma free, we may use it. On this patch just use
nr_reserved_highatomic. Additionally introduce
__zone_watermark_unusable_free to factor out common parts between
zone_watermark_fast and __zone_watermark_ok.
This is an example of ALLOC_HARDER allocation failure using v4.19 based
kernel.
Binder:9343_3: page allocation failure: order:0, mode:0x480020(GFP_ATOMIC), nodemask=(null)
Call trace:
[<ffffff8008f40f8c>] dump_stack+0xb8/0xf0
[<ffffff8008223320>] warn_alloc+0xd8/0x12c
[<ffffff80082245e4>] __alloc_pages_nodemask+0x120c/0x1250
[<ffffff800827f6e8>] new_slab+0x128/0x604
[<ffffff800827b0cc>] ___slab_alloc+0x508/0x670
[<ffffff800827ba00>] __kmalloc+0x2f8/0x310
[<ffffff80084ac3e0>] context_struct_to_string+0x104/0x1cc
[<ffffff80084ad8fc>] security_sid_to_context_core+0x74/0x144
[<ffffff80084ad880>] security_sid_to_context+0x10/0x18
[<ffffff800849bd80>] selinux_secid_to_secctx+0x20/0x28
[<ffffff800849109c>] security_secid_to_secctx+0x3c/0x70
[<ffffff8008bfe118>] binder_transaction+0xe68/0x454c
Mem-Info:
active_anon:102061 inactive_anon:81551 isolated_anon:0
active_file:59102 inactive_file:68924 isolated_file:64
unevictable:611 dirty:63 writeback:0 unstable:0
slab_reclaimable:13324 slab_unreclaimable:44354
mapped:83015 shmem:4858 pagetables:26316 bounce:0
free:2727 free_pcp:1035 free_cma:178
Node 0 active_anon:408244kB inactive_anon:326204kB active_file:236408kB inactive_file:275696kB unevictable:2444kB isolated(anon):0kB isolated(file):256kB mapped:332060kB dirty:252kB writeback:0kB shmem:19432kB writeback_tmp:0kB unstable:0kB all_unreclaimable? no
Normal free:10908kB min:6192kB low:44388kB high:47060kB active_anon:409160kB inactive_anon:325924kB active_file:235820kB inactive_file:276628kB unevictable:2444kB writepending:252kB present:3076096kB managed:2673676kB mlocked:2444kB kernel_stack:62512kB pagetables:105264kB bounce:0kB free_pcp:4140kB local_pcp:40kB free_cma:712kB
lowmem_reserve[]: 0 0
Normal: 505*4kB (H) 357*8kB (H) 201*16kB (H) 65*32kB (H) 1*64kB (H) 0*128kB 0*256kB 0*512kB 0*1024kB 0*2048kB 0*4096kB = 10236kB
138826 total pagecache pages
5460 pages in swap cache
Swap cache stats: add 8273090, delete 8267506, find 1004381/4060142
This is an example of ALLOC_NO_WATERMARKS allocation failure using v4.14
based kernel.
kswapd0: page allocation failure: order:0, mode:0x140000a(GFP_NOIO|__GFP_HIGHMEM|__GFP_MOVABLE), nodemask=(null)
kswapd0 cpuset=/ mems_allowed=0
CPU: 4 PID: 1221 Comm: kswapd0 Not tainted 4.14.113-18770262-userdebug #1
Call trace:
[<0000000000000000>] dump_backtrace+0x0/0x248
[<0000000000000000>] show_stack+0x18/0x20
[<0000000000000000>] __dump_stack+0x20/0x28
[<0000000000000000>] dump_stack+0x68/0x90
[<0000000000000000>] warn_alloc+0x104/0x198
[<0000000000000000>] __alloc_pages_nodemask+0xdc0/0xdf0
[<0000000000000000>] zs_malloc+0x148/0x3d0
[<0000000000000000>] zram_bvec_rw+0x410/0x798
[<0000000000000000>] zram_rw_page+0x88/0xdc
[<0000000000000000>] bdev_write_page+0x70/0xbc
[<0000000000000000>] __swap_writepage+0x58/0x37c
[<0000000000000000>] swap_writepage+0x40/0x4c
[<0000000000000000>] shrink_page_list+0xc30/0xf48
[<0000000000000000>] shrink_inactive_list+0x2b0/0x61c
[<0000000000000000>] shrink_node_memcg+0x23c/0x618
[<0000000000000000>] shrink_node+0x1c8/0x304
[<0000000000000000>] kswapd+0x680/0x7c4
[<0000000000000000>] kthread+0x110/0x120
[<0000000000000000>] ret_from_fork+0x10/0x18
Mem-Info:
active_anon:111826 inactive_anon:65557 isolated_anon:0\x0a active_file:44260 inactive_file:83422 isolated_file:0\x0a unevictable:4158 dirty:117 writeback:0 unstable:0\x0a slab_reclaimable:13943 slab_unreclaimable:43315\x0a mapped:102511 shmem:3299 pagetables:19566 bounce:0\x0a free:3510 free_pcp:553 free_cma:0
Node 0 active_anon:447304kB inactive_anon:262228kB active_file:177040kB inactive_file:333688kB unevictable:16632kB isolated(anon):0kB isolated(file):0kB mapped:410044kB d irty:468kB writeback:0kB shmem:13196kB writeback_tmp:0kB unstable:0kB all_unreclaimable? no
Normal free:14040kB min:7440kB low:94500kB high:98136kB reserved_highatomic:32768KB active_anon:447336kB inactive_anon:261668kB active_file:177572kB inactive_file:333768k B unevictable:16632kB writepending:480kB present:4081664kB managed:3637088kB mlocked:16632kB kernel_stack:47072kB pagetables:78264kB bounce:0kB free_pcp:2280kB local_pcp:720kB free_cma:0kB [ 4738.329607] lowmem_reserve[]: 0 0
Normal: 860*4kB (H) 453*8kB (H) 180*16kB (H) 26*32kB (H) 34*64kB (H) 6*128kB (H) 2*256kB (H) 0*512kB 0*1024kB 0*2048kB 0*4096kB = 14232kB
This is trace log which shows GFP_HIGHUSER consumes free pages right
before ALLOC_NO_WATERMARKS.
<...>-22275 [006] .... 889.213383: mm_page_alloc: page=00000000d2be5665 pfn=970744 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213385: mm_page_alloc: page=000000004b2335c2 pfn=970745 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213387: mm_page_alloc: page=00000000017272e1 pfn=970278 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213389: mm_page_alloc: page=00000000c4be79fb pfn=970279 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213391: mm_page_alloc: page=00000000f8a51d4f pfn=970260 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213393: mm_page_alloc: page=000000006ba8f5ac pfn=970261 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213395: mm_page_alloc: page=00000000819f1cd3 pfn=970196 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213396: mm_page_alloc: page=00000000f6b72a64 pfn=970197 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
kswapd0-1207 [005] ...1 889.213398: mm_page_alloc: page= (null) pfn=0 order=0 migratetype=1 nr_free=3650 gfp_flags=GFP_NOWAIT|__GFP_HIGHMEM|__GFP_NOWARN|__GFP_MOVABLE
[jaewon31.kim@samsung.com: remove redundant code for high-order]
Link: http://lkml.kernel.org/r/20200623035242.27232-1-jaewon31.kim@samsung.com
Reported-by: Yong-Taek Lee <ytk.lee@samsung.com>
Suggested-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Jaewon Kim <jaewon31.kim@samsung.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Baoquan He <bhe@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Yong-Taek Lee <ytk.lee@samsung.com>
Cc: Michal Hocko <mhocko@kernel.org>
Link: http://lkml.kernel.org/r/20200619235958.11283-1-jaewon31.kim@samsung.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-06 23:25:20 -07:00
|
|
|
|
2016-05-19 17:14:07 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
page_alloc: consider highatomic reserve in watermark fast
zone_watermark_fast was introduced by commit 48ee5f3696f6 ("mm,
page_alloc: shortcut watermark checks for order-0 pages"). The commit
simply checks if free pages is bigger than watermark without additional
calculation such like reducing watermark.
It considered free cma pages but it did not consider highatomic reserved.
This may incur exhaustion of free pages except high order atomic free
pages.
Assume that reserved_highatomic pageblock is bigger than watermark min,
and there are only few free pages except high order atomic free. Because
zone_watermark_fast passes the allocation without considering high order
atomic free, normal reclaimable allocation like GFP_HIGHUSER will consume
all the free pages. Then finally order-0 atomic allocation may fail on
allocation.
This means watermark min is not protected against non-atomic allocation.
The order-0 atomic allocation with ALLOC_HARDER unwantedly can be failed.
Additionally the __GFP_MEMALLOC allocation with ALLOC_NO_WATERMARKS also
can be failed.
To avoid the problem, zone_watermark_fast should consider highatomic
reserve. If the actual size of high atomic free is counted accurately
like cma free, we may use it. On this patch just use
nr_reserved_highatomic. Additionally introduce
__zone_watermark_unusable_free to factor out common parts between
zone_watermark_fast and __zone_watermark_ok.
This is an example of ALLOC_HARDER allocation failure using v4.19 based
kernel.
Binder:9343_3: page allocation failure: order:0, mode:0x480020(GFP_ATOMIC), nodemask=(null)
Call trace:
[<ffffff8008f40f8c>] dump_stack+0xb8/0xf0
[<ffffff8008223320>] warn_alloc+0xd8/0x12c
[<ffffff80082245e4>] __alloc_pages_nodemask+0x120c/0x1250
[<ffffff800827f6e8>] new_slab+0x128/0x604
[<ffffff800827b0cc>] ___slab_alloc+0x508/0x670
[<ffffff800827ba00>] __kmalloc+0x2f8/0x310
[<ffffff80084ac3e0>] context_struct_to_string+0x104/0x1cc
[<ffffff80084ad8fc>] security_sid_to_context_core+0x74/0x144
[<ffffff80084ad880>] security_sid_to_context+0x10/0x18
[<ffffff800849bd80>] selinux_secid_to_secctx+0x20/0x28
[<ffffff800849109c>] security_secid_to_secctx+0x3c/0x70
[<ffffff8008bfe118>] binder_transaction+0xe68/0x454c
Mem-Info:
active_anon:102061 inactive_anon:81551 isolated_anon:0
active_file:59102 inactive_file:68924 isolated_file:64
unevictable:611 dirty:63 writeback:0 unstable:0
slab_reclaimable:13324 slab_unreclaimable:44354
mapped:83015 shmem:4858 pagetables:26316 bounce:0
free:2727 free_pcp:1035 free_cma:178
Node 0 active_anon:408244kB inactive_anon:326204kB active_file:236408kB inactive_file:275696kB unevictable:2444kB isolated(anon):0kB isolated(file):256kB mapped:332060kB dirty:252kB writeback:0kB shmem:19432kB writeback_tmp:0kB unstable:0kB all_unreclaimable? no
Normal free:10908kB min:6192kB low:44388kB high:47060kB active_anon:409160kB inactive_anon:325924kB active_file:235820kB inactive_file:276628kB unevictable:2444kB writepending:252kB present:3076096kB managed:2673676kB mlocked:2444kB kernel_stack:62512kB pagetables:105264kB bounce:0kB free_pcp:4140kB local_pcp:40kB free_cma:712kB
lowmem_reserve[]: 0 0
Normal: 505*4kB (H) 357*8kB (H) 201*16kB (H) 65*32kB (H) 1*64kB (H) 0*128kB 0*256kB 0*512kB 0*1024kB 0*2048kB 0*4096kB = 10236kB
138826 total pagecache pages
5460 pages in swap cache
Swap cache stats: add 8273090, delete 8267506, find 1004381/4060142
This is an example of ALLOC_NO_WATERMARKS allocation failure using v4.14
based kernel.
kswapd0: page allocation failure: order:0, mode:0x140000a(GFP_NOIO|__GFP_HIGHMEM|__GFP_MOVABLE), nodemask=(null)
kswapd0 cpuset=/ mems_allowed=0
CPU: 4 PID: 1221 Comm: kswapd0 Not tainted 4.14.113-18770262-userdebug #1
Call trace:
[<0000000000000000>] dump_backtrace+0x0/0x248
[<0000000000000000>] show_stack+0x18/0x20
[<0000000000000000>] __dump_stack+0x20/0x28
[<0000000000000000>] dump_stack+0x68/0x90
[<0000000000000000>] warn_alloc+0x104/0x198
[<0000000000000000>] __alloc_pages_nodemask+0xdc0/0xdf0
[<0000000000000000>] zs_malloc+0x148/0x3d0
[<0000000000000000>] zram_bvec_rw+0x410/0x798
[<0000000000000000>] zram_rw_page+0x88/0xdc
[<0000000000000000>] bdev_write_page+0x70/0xbc
[<0000000000000000>] __swap_writepage+0x58/0x37c
[<0000000000000000>] swap_writepage+0x40/0x4c
[<0000000000000000>] shrink_page_list+0xc30/0xf48
[<0000000000000000>] shrink_inactive_list+0x2b0/0x61c
[<0000000000000000>] shrink_node_memcg+0x23c/0x618
[<0000000000000000>] shrink_node+0x1c8/0x304
[<0000000000000000>] kswapd+0x680/0x7c4
[<0000000000000000>] kthread+0x110/0x120
[<0000000000000000>] ret_from_fork+0x10/0x18
Mem-Info:
active_anon:111826 inactive_anon:65557 isolated_anon:0\x0a active_file:44260 inactive_file:83422 isolated_file:0\x0a unevictable:4158 dirty:117 writeback:0 unstable:0\x0a slab_reclaimable:13943 slab_unreclaimable:43315\x0a mapped:102511 shmem:3299 pagetables:19566 bounce:0\x0a free:3510 free_pcp:553 free_cma:0
Node 0 active_anon:447304kB inactive_anon:262228kB active_file:177040kB inactive_file:333688kB unevictable:16632kB isolated(anon):0kB isolated(file):0kB mapped:410044kB d irty:468kB writeback:0kB shmem:13196kB writeback_tmp:0kB unstable:0kB all_unreclaimable? no
Normal free:14040kB min:7440kB low:94500kB high:98136kB reserved_highatomic:32768KB active_anon:447336kB inactive_anon:261668kB active_file:177572kB inactive_file:333768k B unevictable:16632kB writepending:480kB present:4081664kB managed:3637088kB mlocked:16632kB kernel_stack:47072kB pagetables:78264kB bounce:0kB free_pcp:2280kB local_pcp:720kB free_cma:0kB [ 4738.329607] lowmem_reserve[]: 0 0
Normal: 860*4kB (H) 453*8kB (H) 180*16kB (H) 26*32kB (H) 34*64kB (H) 6*128kB (H) 2*256kB (H) 0*512kB 0*1024kB 0*2048kB 0*4096kB = 14232kB
This is trace log which shows GFP_HIGHUSER consumes free pages right
before ALLOC_NO_WATERMARKS.
<...>-22275 [006] .... 889.213383: mm_page_alloc: page=00000000d2be5665 pfn=970744 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213385: mm_page_alloc: page=000000004b2335c2 pfn=970745 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213387: mm_page_alloc: page=00000000017272e1 pfn=970278 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213389: mm_page_alloc: page=00000000c4be79fb pfn=970279 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213391: mm_page_alloc: page=00000000f8a51d4f pfn=970260 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213393: mm_page_alloc: page=000000006ba8f5ac pfn=970261 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213395: mm_page_alloc: page=00000000819f1cd3 pfn=970196 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213396: mm_page_alloc: page=00000000f6b72a64 pfn=970197 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
kswapd0-1207 [005] ...1 889.213398: mm_page_alloc: page= (null) pfn=0 order=0 migratetype=1 nr_free=3650 gfp_flags=GFP_NOWAIT|__GFP_HIGHMEM|__GFP_NOWARN|__GFP_MOVABLE
[jaewon31.kim@samsung.com: remove redundant code for high-order]
Link: http://lkml.kernel.org/r/20200623035242.27232-1-jaewon31.kim@samsung.com
Reported-by: Yong-Taek Lee <ytk.lee@samsung.com>
Suggested-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Jaewon Kim <jaewon31.kim@samsung.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Baoquan He <bhe@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Yong-Taek Lee <ytk.lee@samsung.com>
Cc: Michal Hocko <mhocko@kernel.org>
Link: http://lkml.kernel.org/r/20200619235958.11283-1-jaewon31.kim@samsung.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-06 23:25:20 -07:00
|
|
|
|
2016-05-19 17:14:07 -07:00
|
|
|
|
page_alloc: consider highatomic reserve in watermark fast
zone_watermark_fast was introduced by commit 48ee5f3696f6 ("mm,
page_alloc: shortcut watermark checks for order-0 pages"). The commit
simply checks if free pages is bigger than watermark without additional
calculation such like reducing watermark.
It considered free cma pages but it did not consider highatomic reserved.
This may incur exhaustion of free pages except high order atomic free
pages.
Assume that reserved_highatomic pageblock is bigger than watermark min,
and there are only few free pages except high order atomic free. Because
zone_watermark_fast passes the allocation without considering high order
atomic free, normal reclaimable allocation like GFP_HIGHUSER will consume
all the free pages. Then finally order-0 atomic allocation may fail on
allocation.
This means watermark min is not protected against non-atomic allocation.
The order-0 atomic allocation with ALLOC_HARDER unwantedly can be failed.
Additionally the __GFP_MEMALLOC allocation with ALLOC_NO_WATERMARKS also
can be failed.
To avoid the problem, zone_watermark_fast should consider highatomic
reserve. If the actual size of high atomic free is counted accurately
like cma free, we may use it. On this patch just use
nr_reserved_highatomic. Additionally introduce
__zone_watermark_unusable_free to factor out common parts between
zone_watermark_fast and __zone_watermark_ok.
This is an example of ALLOC_HARDER allocation failure using v4.19 based
kernel.
Binder:9343_3: page allocation failure: order:0, mode:0x480020(GFP_ATOMIC), nodemask=(null)
Call trace:
[<ffffff8008f40f8c>] dump_stack+0xb8/0xf0
[<ffffff8008223320>] warn_alloc+0xd8/0x12c
[<ffffff80082245e4>] __alloc_pages_nodemask+0x120c/0x1250
[<ffffff800827f6e8>] new_slab+0x128/0x604
[<ffffff800827b0cc>] ___slab_alloc+0x508/0x670
[<ffffff800827ba00>] __kmalloc+0x2f8/0x310
[<ffffff80084ac3e0>] context_struct_to_string+0x104/0x1cc
[<ffffff80084ad8fc>] security_sid_to_context_core+0x74/0x144
[<ffffff80084ad880>] security_sid_to_context+0x10/0x18
[<ffffff800849bd80>] selinux_secid_to_secctx+0x20/0x28
[<ffffff800849109c>] security_secid_to_secctx+0x3c/0x70
[<ffffff8008bfe118>] binder_transaction+0xe68/0x454c
Mem-Info:
active_anon:102061 inactive_anon:81551 isolated_anon:0
active_file:59102 inactive_file:68924 isolated_file:64
unevictable:611 dirty:63 writeback:0 unstable:0
slab_reclaimable:13324 slab_unreclaimable:44354
mapped:83015 shmem:4858 pagetables:26316 bounce:0
free:2727 free_pcp:1035 free_cma:178
Node 0 active_anon:408244kB inactive_anon:326204kB active_file:236408kB inactive_file:275696kB unevictable:2444kB isolated(anon):0kB isolated(file):256kB mapped:332060kB dirty:252kB writeback:0kB shmem:19432kB writeback_tmp:0kB unstable:0kB all_unreclaimable? no
Normal free:10908kB min:6192kB low:44388kB high:47060kB active_anon:409160kB inactive_anon:325924kB active_file:235820kB inactive_file:276628kB unevictable:2444kB writepending:252kB present:3076096kB managed:2673676kB mlocked:2444kB kernel_stack:62512kB pagetables:105264kB bounce:0kB free_pcp:4140kB local_pcp:40kB free_cma:712kB
lowmem_reserve[]: 0 0
Normal: 505*4kB (H) 357*8kB (H) 201*16kB (H) 65*32kB (H) 1*64kB (H) 0*128kB 0*256kB 0*512kB 0*1024kB 0*2048kB 0*4096kB = 10236kB
138826 total pagecache pages
5460 pages in swap cache
Swap cache stats: add 8273090, delete 8267506, find 1004381/4060142
This is an example of ALLOC_NO_WATERMARKS allocation failure using v4.14
based kernel.
kswapd0: page allocation failure: order:0, mode:0x140000a(GFP_NOIO|__GFP_HIGHMEM|__GFP_MOVABLE), nodemask=(null)
kswapd0 cpuset=/ mems_allowed=0
CPU: 4 PID: 1221 Comm: kswapd0 Not tainted 4.14.113-18770262-userdebug #1
Call trace:
[<0000000000000000>] dump_backtrace+0x0/0x248
[<0000000000000000>] show_stack+0x18/0x20
[<0000000000000000>] __dump_stack+0x20/0x28
[<0000000000000000>] dump_stack+0x68/0x90
[<0000000000000000>] warn_alloc+0x104/0x198
[<0000000000000000>] __alloc_pages_nodemask+0xdc0/0xdf0
[<0000000000000000>] zs_malloc+0x148/0x3d0
[<0000000000000000>] zram_bvec_rw+0x410/0x798
[<0000000000000000>] zram_rw_page+0x88/0xdc
[<0000000000000000>] bdev_write_page+0x70/0xbc
[<0000000000000000>] __swap_writepage+0x58/0x37c
[<0000000000000000>] swap_writepage+0x40/0x4c
[<0000000000000000>] shrink_page_list+0xc30/0xf48
[<0000000000000000>] shrink_inactive_list+0x2b0/0x61c
[<0000000000000000>] shrink_node_memcg+0x23c/0x618
[<0000000000000000>] shrink_node+0x1c8/0x304
[<0000000000000000>] kswapd+0x680/0x7c4
[<0000000000000000>] kthread+0x110/0x120
[<0000000000000000>] ret_from_fork+0x10/0x18
Mem-Info:
active_anon:111826 inactive_anon:65557 isolated_anon:0\x0a active_file:44260 inactive_file:83422 isolated_file:0\x0a unevictable:4158 dirty:117 writeback:0 unstable:0\x0a slab_reclaimable:13943 slab_unreclaimable:43315\x0a mapped:102511 shmem:3299 pagetables:19566 bounce:0\x0a free:3510 free_pcp:553 free_cma:0
Node 0 active_anon:447304kB inactive_anon:262228kB active_file:177040kB inactive_file:333688kB unevictable:16632kB isolated(anon):0kB isolated(file):0kB mapped:410044kB d irty:468kB writeback:0kB shmem:13196kB writeback_tmp:0kB unstable:0kB all_unreclaimable? no
Normal free:14040kB min:7440kB low:94500kB high:98136kB reserved_highatomic:32768KB active_anon:447336kB inactive_anon:261668kB active_file:177572kB inactive_file:333768k B unevictable:16632kB writepending:480kB present:4081664kB managed:3637088kB mlocked:16632kB kernel_stack:47072kB pagetables:78264kB bounce:0kB free_pcp:2280kB local_pcp:720kB free_cma:0kB [ 4738.329607] lowmem_reserve[]: 0 0
Normal: 860*4kB (H) 453*8kB (H) 180*16kB (H) 26*32kB (H) 34*64kB (H) 6*128kB (H) 2*256kB (H) 0*512kB 0*1024kB 0*2048kB 0*4096kB = 14232kB
This is trace log which shows GFP_HIGHUSER consumes free pages right
before ALLOC_NO_WATERMARKS.
<...>-22275 [006] .... 889.213383: mm_page_alloc: page=00000000d2be5665 pfn=970744 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213385: mm_page_alloc: page=000000004b2335c2 pfn=970745 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213387: mm_page_alloc: page=00000000017272e1 pfn=970278 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213389: mm_page_alloc: page=00000000c4be79fb pfn=970279 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213391: mm_page_alloc: page=00000000f8a51d4f pfn=970260 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213393: mm_page_alloc: page=000000006ba8f5ac pfn=970261 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213395: mm_page_alloc: page=00000000819f1cd3 pfn=970196 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213396: mm_page_alloc: page=00000000f6b72a64 pfn=970197 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
kswapd0-1207 [005] ...1 889.213398: mm_page_alloc: page= (null) pfn=0 order=0 migratetype=1 nr_free=3650 gfp_flags=GFP_NOWAIT|__GFP_HIGHMEM|__GFP_NOWARN|__GFP_MOVABLE
[jaewon31.kim@samsung.com: remove redundant code for high-order]
Link: http://lkml.kernel.org/r/20200623035242.27232-1-jaewon31.kim@samsung.com
Reported-by: Yong-Taek Lee <ytk.lee@samsung.com>
Suggested-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Jaewon Kim <jaewon31.kim@samsung.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Baoquan He <bhe@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Yong-Taek Lee <ytk.lee@samsung.com>
Cc: Michal Hocko <mhocko@kernel.org>
Link: http://lkml.kernel.org/r/20200619235958.11283-1-jaewon31.kim@samsung.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-06 23:25:20 -07:00
|
|
|
|
2022-07-25 18:52:12 +09:00
|
|
|
|
|
|
|
|
|
page_alloc: consider highatomic reserve in watermark fast
zone_watermark_fast was introduced by commit 48ee5f3696f6 ("mm,
page_alloc: shortcut watermark checks for order-0 pages"). The commit
simply checks if free pages is bigger than watermark without additional
calculation such like reducing watermark.
It considered free cma pages but it did not consider highatomic reserved.
This may incur exhaustion of free pages except high order atomic free
pages.
Assume that reserved_highatomic pageblock is bigger than watermark min,
and there are only few free pages except high order atomic free. Because
zone_watermark_fast passes the allocation without considering high order
atomic free, normal reclaimable allocation like GFP_HIGHUSER will consume
all the free pages. Then finally order-0 atomic allocation may fail on
allocation.
This means watermark min is not protected against non-atomic allocation.
The order-0 atomic allocation with ALLOC_HARDER unwantedly can be failed.
Additionally the __GFP_MEMALLOC allocation with ALLOC_NO_WATERMARKS also
can be failed.
To avoid the problem, zone_watermark_fast should consider highatomic
reserve. If the actual size of high atomic free is counted accurately
like cma free, we may use it. On this patch just use
nr_reserved_highatomic. Additionally introduce
__zone_watermark_unusable_free to factor out common parts between
zone_watermark_fast and __zone_watermark_ok.
This is an example of ALLOC_HARDER allocation failure using v4.19 based
kernel.
Binder:9343_3: page allocation failure: order:0, mode:0x480020(GFP_ATOMIC), nodemask=(null)
Call trace:
[<ffffff8008f40f8c>] dump_stack+0xb8/0xf0
[<ffffff8008223320>] warn_alloc+0xd8/0x12c
[<ffffff80082245e4>] __alloc_pages_nodemask+0x120c/0x1250
[<ffffff800827f6e8>] new_slab+0x128/0x604
[<ffffff800827b0cc>] ___slab_alloc+0x508/0x670
[<ffffff800827ba00>] __kmalloc+0x2f8/0x310
[<ffffff80084ac3e0>] context_struct_to_string+0x104/0x1cc
[<ffffff80084ad8fc>] security_sid_to_context_core+0x74/0x144
[<ffffff80084ad880>] security_sid_to_context+0x10/0x18
[<ffffff800849bd80>] selinux_secid_to_secctx+0x20/0x28
[<ffffff800849109c>] security_secid_to_secctx+0x3c/0x70
[<ffffff8008bfe118>] binder_transaction+0xe68/0x454c
Mem-Info:
active_anon:102061 inactive_anon:81551 isolated_anon:0
active_file:59102 inactive_file:68924 isolated_file:64
unevictable:611 dirty:63 writeback:0 unstable:0
slab_reclaimable:13324 slab_unreclaimable:44354
mapped:83015 shmem:4858 pagetables:26316 bounce:0
free:2727 free_pcp:1035 free_cma:178
Node 0 active_anon:408244kB inactive_anon:326204kB active_file:236408kB inactive_file:275696kB unevictable:2444kB isolated(anon):0kB isolated(file):256kB mapped:332060kB dirty:252kB writeback:0kB shmem:19432kB writeback_tmp:0kB unstable:0kB all_unreclaimable? no
Normal free:10908kB min:6192kB low:44388kB high:47060kB active_anon:409160kB inactive_anon:325924kB active_file:235820kB inactive_file:276628kB unevictable:2444kB writepending:252kB present:3076096kB managed:2673676kB mlocked:2444kB kernel_stack:62512kB pagetables:105264kB bounce:0kB free_pcp:4140kB local_pcp:40kB free_cma:712kB
lowmem_reserve[]: 0 0
Normal: 505*4kB (H) 357*8kB (H) 201*16kB (H) 65*32kB (H) 1*64kB (H) 0*128kB 0*256kB 0*512kB 0*1024kB 0*2048kB 0*4096kB = 10236kB
138826 total pagecache pages
5460 pages in swap cache
Swap cache stats: add 8273090, delete 8267506, find 1004381/4060142
This is an example of ALLOC_NO_WATERMARKS allocation failure using v4.14
based kernel.
kswapd0: page allocation failure: order:0, mode:0x140000a(GFP_NOIO|__GFP_HIGHMEM|__GFP_MOVABLE), nodemask=(null)
kswapd0 cpuset=/ mems_allowed=0
CPU: 4 PID: 1221 Comm: kswapd0 Not tainted 4.14.113-18770262-userdebug #1
Call trace:
[<0000000000000000>] dump_backtrace+0x0/0x248
[<0000000000000000>] show_stack+0x18/0x20
[<0000000000000000>] __dump_stack+0x20/0x28
[<0000000000000000>] dump_stack+0x68/0x90
[<0000000000000000>] warn_alloc+0x104/0x198
[<0000000000000000>] __alloc_pages_nodemask+0xdc0/0xdf0
[<0000000000000000>] zs_malloc+0x148/0x3d0
[<0000000000000000>] zram_bvec_rw+0x410/0x798
[<0000000000000000>] zram_rw_page+0x88/0xdc
[<0000000000000000>] bdev_write_page+0x70/0xbc
[<0000000000000000>] __swap_writepage+0x58/0x37c
[<0000000000000000>] swap_writepage+0x40/0x4c
[<0000000000000000>] shrink_page_list+0xc30/0xf48
[<0000000000000000>] shrink_inactive_list+0x2b0/0x61c
[<0000000000000000>] shrink_node_memcg+0x23c/0x618
[<0000000000000000>] shrink_node+0x1c8/0x304
[<0000000000000000>] kswapd+0x680/0x7c4
[<0000000000000000>] kthread+0x110/0x120
[<0000000000000000>] ret_from_fork+0x10/0x18
Mem-Info:
active_anon:111826 inactive_anon:65557 isolated_anon:0\x0a active_file:44260 inactive_file:83422 isolated_file:0\x0a unevictable:4158 dirty:117 writeback:0 unstable:0\x0a slab_reclaimable:13943 slab_unreclaimable:43315\x0a mapped:102511 shmem:3299 pagetables:19566 bounce:0\x0a free:3510 free_pcp:553 free_cma:0
Node 0 active_anon:447304kB inactive_anon:262228kB active_file:177040kB inactive_file:333688kB unevictable:16632kB isolated(anon):0kB isolated(file):0kB mapped:410044kB d irty:468kB writeback:0kB shmem:13196kB writeback_tmp:0kB unstable:0kB all_unreclaimable? no
Normal free:14040kB min:7440kB low:94500kB high:98136kB reserved_highatomic:32768KB active_anon:447336kB inactive_anon:261668kB active_file:177572kB inactive_file:333768k B unevictable:16632kB writepending:480kB present:4081664kB managed:3637088kB mlocked:16632kB kernel_stack:47072kB pagetables:78264kB bounce:0kB free_pcp:2280kB local_pcp:720kB free_cma:0kB [ 4738.329607] lowmem_reserve[]: 0 0
Normal: 860*4kB (H) 453*8kB (H) 180*16kB (H) 26*32kB (H) 34*64kB (H) 6*128kB (H) 2*256kB (H) 0*512kB 0*1024kB 0*2048kB 0*4096kB = 14232kB
This is trace log which shows GFP_HIGHUSER consumes free pages right
before ALLOC_NO_WATERMARKS.
<...>-22275 [006] .... 889.213383: mm_page_alloc: page=00000000d2be5665 pfn=970744 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213385: mm_page_alloc: page=000000004b2335c2 pfn=970745 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213387: mm_page_alloc: page=00000000017272e1 pfn=970278 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213389: mm_page_alloc: page=00000000c4be79fb pfn=970279 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213391: mm_page_alloc: page=00000000f8a51d4f pfn=970260 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213393: mm_page_alloc: page=000000006ba8f5ac pfn=970261 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213395: mm_page_alloc: page=00000000819f1cd3 pfn=970196 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213396: mm_page_alloc: page=00000000f6b72a64 pfn=970197 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
kswapd0-1207 [005] ...1 889.213398: mm_page_alloc: page= (null) pfn=0 order=0 migratetype=1 nr_free=3650 gfp_flags=GFP_NOWAIT|__GFP_HIGHMEM|__GFP_NOWARN|__GFP_MOVABLE
[jaewon31.kim@samsung.com: remove redundant code for high-order]
Link: http://lkml.kernel.org/r/20200623035242.27232-1-jaewon31.kim@samsung.com
Reported-by: Yong-Taek Lee <ytk.lee@samsung.com>
Suggested-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Jaewon Kim <jaewon31.kim@samsung.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Baoquan He <bhe@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Yong-Taek Lee <ytk.lee@samsung.com>
Cc: Michal Hocko <mhocko@kernel.org>
Link: http://lkml.kernel.org/r/20200619235958.11283-1-jaewon31.kim@samsung.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-06 23:25:20 -07:00
|
|
|
|
2022-07-25 18:52:12 +09:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
page_alloc: consider highatomic reserve in watermark fast
zone_watermark_fast was introduced by commit 48ee5f3696f6 ("mm,
page_alloc: shortcut watermark checks for order-0 pages"). The commit
simply checks if free pages is bigger than watermark without additional
calculation such like reducing watermark.
It considered free cma pages but it did not consider highatomic reserved.
This may incur exhaustion of free pages except high order atomic free
pages.
Assume that reserved_highatomic pageblock is bigger than watermark min,
and there are only few free pages except high order atomic free. Because
zone_watermark_fast passes the allocation without considering high order
atomic free, normal reclaimable allocation like GFP_HIGHUSER will consume
all the free pages. Then finally order-0 atomic allocation may fail on
allocation.
This means watermark min is not protected against non-atomic allocation.
The order-0 atomic allocation with ALLOC_HARDER unwantedly can be failed.
Additionally the __GFP_MEMALLOC allocation with ALLOC_NO_WATERMARKS also
can be failed.
To avoid the problem, zone_watermark_fast should consider highatomic
reserve. If the actual size of high atomic free is counted accurately
like cma free, we may use it. On this patch just use
nr_reserved_highatomic. Additionally introduce
__zone_watermark_unusable_free to factor out common parts between
zone_watermark_fast and __zone_watermark_ok.
This is an example of ALLOC_HARDER allocation failure using v4.19 based
kernel.
Binder:9343_3: page allocation failure: order:0, mode:0x480020(GFP_ATOMIC), nodemask=(null)
Call trace:
[<ffffff8008f40f8c>] dump_stack+0xb8/0xf0
[<ffffff8008223320>] warn_alloc+0xd8/0x12c
[<ffffff80082245e4>] __alloc_pages_nodemask+0x120c/0x1250
[<ffffff800827f6e8>] new_slab+0x128/0x604
[<ffffff800827b0cc>] ___slab_alloc+0x508/0x670
[<ffffff800827ba00>] __kmalloc+0x2f8/0x310
[<ffffff80084ac3e0>] context_struct_to_string+0x104/0x1cc
[<ffffff80084ad8fc>] security_sid_to_context_core+0x74/0x144
[<ffffff80084ad880>] security_sid_to_context+0x10/0x18
[<ffffff800849bd80>] selinux_secid_to_secctx+0x20/0x28
[<ffffff800849109c>] security_secid_to_secctx+0x3c/0x70
[<ffffff8008bfe118>] binder_transaction+0xe68/0x454c
Mem-Info:
active_anon:102061 inactive_anon:81551 isolated_anon:0
active_file:59102 inactive_file:68924 isolated_file:64
unevictable:611 dirty:63 writeback:0 unstable:0
slab_reclaimable:13324 slab_unreclaimable:44354
mapped:83015 shmem:4858 pagetables:26316 bounce:0
free:2727 free_pcp:1035 free_cma:178
Node 0 active_anon:408244kB inactive_anon:326204kB active_file:236408kB inactive_file:275696kB unevictable:2444kB isolated(anon):0kB isolated(file):256kB mapped:332060kB dirty:252kB writeback:0kB shmem:19432kB writeback_tmp:0kB unstable:0kB all_unreclaimable? no
Normal free:10908kB min:6192kB low:44388kB high:47060kB active_anon:409160kB inactive_anon:325924kB active_file:235820kB inactive_file:276628kB unevictable:2444kB writepending:252kB present:3076096kB managed:2673676kB mlocked:2444kB kernel_stack:62512kB pagetables:105264kB bounce:0kB free_pcp:4140kB local_pcp:40kB free_cma:712kB
lowmem_reserve[]: 0 0
Normal: 505*4kB (H) 357*8kB (H) 201*16kB (H) 65*32kB (H) 1*64kB (H) 0*128kB 0*256kB 0*512kB 0*1024kB 0*2048kB 0*4096kB = 10236kB
138826 total pagecache pages
5460 pages in swap cache
Swap cache stats: add 8273090, delete 8267506, find 1004381/4060142
This is an example of ALLOC_NO_WATERMARKS allocation failure using v4.14
based kernel.
kswapd0: page allocation failure: order:0, mode:0x140000a(GFP_NOIO|__GFP_HIGHMEM|__GFP_MOVABLE), nodemask=(null)
kswapd0 cpuset=/ mems_allowed=0
CPU: 4 PID: 1221 Comm: kswapd0 Not tainted 4.14.113-18770262-userdebug #1
Call trace:
[<0000000000000000>] dump_backtrace+0x0/0x248
[<0000000000000000>] show_stack+0x18/0x20
[<0000000000000000>] __dump_stack+0x20/0x28
[<0000000000000000>] dump_stack+0x68/0x90
[<0000000000000000>] warn_alloc+0x104/0x198
[<0000000000000000>] __alloc_pages_nodemask+0xdc0/0xdf0
[<0000000000000000>] zs_malloc+0x148/0x3d0
[<0000000000000000>] zram_bvec_rw+0x410/0x798
[<0000000000000000>] zram_rw_page+0x88/0xdc
[<0000000000000000>] bdev_write_page+0x70/0xbc
[<0000000000000000>] __swap_writepage+0x58/0x37c
[<0000000000000000>] swap_writepage+0x40/0x4c
[<0000000000000000>] shrink_page_list+0xc30/0xf48
[<0000000000000000>] shrink_inactive_list+0x2b0/0x61c
[<0000000000000000>] shrink_node_memcg+0x23c/0x618
[<0000000000000000>] shrink_node+0x1c8/0x304
[<0000000000000000>] kswapd+0x680/0x7c4
[<0000000000000000>] kthread+0x110/0x120
[<0000000000000000>] ret_from_fork+0x10/0x18
Mem-Info:
active_anon:111826 inactive_anon:65557 isolated_anon:0\x0a active_file:44260 inactive_file:83422 isolated_file:0\x0a unevictable:4158 dirty:117 writeback:0 unstable:0\x0a slab_reclaimable:13943 slab_unreclaimable:43315\x0a mapped:102511 shmem:3299 pagetables:19566 bounce:0\x0a free:3510 free_pcp:553 free_cma:0
Node 0 active_anon:447304kB inactive_anon:262228kB active_file:177040kB inactive_file:333688kB unevictable:16632kB isolated(anon):0kB isolated(file):0kB mapped:410044kB d irty:468kB writeback:0kB shmem:13196kB writeback_tmp:0kB unstable:0kB all_unreclaimable? no
Normal free:14040kB min:7440kB low:94500kB high:98136kB reserved_highatomic:32768KB active_anon:447336kB inactive_anon:261668kB active_file:177572kB inactive_file:333768k B unevictable:16632kB writepending:480kB present:4081664kB managed:3637088kB mlocked:16632kB kernel_stack:47072kB pagetables:78264kB bounce:0kB free_pcp:2280kB local_pcp:720kB free_cma:0kB [ 4738.329607] lowmem_reserve[]: 0 0
Normal: 860*4kB (H) 453*8kB (H) 180*16kB (H) 26*32kB (H) 34*64kB (H) 6*128kB (H) 2*256kB (H) 0*512kB 0*1024kB 0*2048kB 0*4096kB = 14232kB
This is trace log which shows GFP_HIGHUSER consumes free pages right
before ALLOC_NO_WATERMARKS.
<...>-22275 [006] .... 889.213383: mm_page_alloc: page=00000000d2be5665 pfn=970744 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213385: mm_page_alloc: page=000000004b2335c2 pfn=970745 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213387: mm_page_alloc: page=00000000017272e1 pfn=970278 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213389: mm_page_alloc: page=00000000c4be79fb pfn=970279 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213391: mm_page_alloc: page=00000000f8a51d4f pfn=970260 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213393: mm_page_alloc: page=000000006ba8f5ac pfn=970261 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213395: mm_page_alloc: page=00000000819f1cd3 pfn=970196 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
<...>-22275 [006] .... 889.213396: mm_page_alloc: page=00000000f6b72a64 pfn=970197 order=0 migratetype=0 nr_free=3650 gfp_flags=GFP_HIGHUSER|__GFP_ZERO
kswapd0-1207 [005] ...1 889.213398: mm_page_alloc: page= (null) pfn=0 order=0 migratetype=1 nr_free=3650 gfp_flags=GFP_NOWAIT|__GFP_HIGHMEM|__GFP_NOWARN|__GFP_MOVABLE
[jaewon31.kim@samsung.com: remove redundant code for high-order]
Link: http://lkml.kernel.org/r/20200623035242.27232-1-jaewon31.kim@samsung.com
Reported-by: Yong-Taek Lee <ytk.lee@samsung.com>
Suggested-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Jaewon Kim <jaewon31.kim@samsung.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Baoquan He <bhe@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Yong-Taek Lee <ytk.lee@samsung.com>
Cc: Michal Hocko <mhocko@kernel.org>
Link: http://lkml.kernel.org/r/20200619235958.11283-1-jaewon31.kim@samsung.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-06 23:25:20 -07:00
|
|
|
|
|
|
|
|
|
2016-05-19 17:14:07 -07:00
|
|
|
|
mm, page_alloc: skip ->waternark_boost for atomic order-0 allocations
When boosting is enabled, it is observed that rate of atomic order-0
allocation failures are high due to the fact that free levels in the
system are checked with ->watermark_boost offset. This is not a problem
for sleepable allocations but for atomic allocations which looks like
regression.
This problem is seen frequently on system setup of Android kernel running
on Snapdragon hardware with 4GB RAM size. When no extfrag event occurred
in the system, ->watermark_boost factor is zero, thus the watermark
configurations in the system are:
_watermark = (
[WMARK_MIN] = 1272, --> ~5MB
[WMARK_LOW] = 9067, --> ~36MB
[WMARK_HIGH] = 9385), --> ~38MB
watermark_boost = 0
After launching some memory hungry applications in Android which can cause
extfrag events in the system to an extent that ->watermark_boost can be
set to max i.e. default boost factor makes it to 150% of high watermark.
_watermark = (
[WMARK_MIN] = 1272, --> ~5MB
[WMARK_LOW] = 9067, --> ~36MB
[WMARK_HIGH] = 9385), --> ~38MB
watermark_boost = 14077, -->~57MB
With default system configuration, for an atomic order-0 allocation to
succeed, having free memory of ~2MB will suffice. But boosting makes the
min_wmark to ~61MB thus for an atomic order-0 allocation to be successful
system should have minimum of ~23MB of free memory(from calculations of
zone_watermark_ok(), min = 3/4(min/2)). But failures are observed despite
system is having ~20MB of free memory. In the testing, this is
reproducible as early as first 300secs since boot and with furtherlowram
configurations(<2GB) it is observed as early as first 150secs since boot.
These failures can be avoided by excluding the ->watermark_boost in
watermark caluculations for atomic order-0 allocations.
[akpm@linux-foundation.org: fix comment grammar, reflow comment]
[charante@codeaurora.org: fix suggested by Mel Gorman]
Link: http://lkml.kernel.org/r/31556793-57b1-1c21-1a9d-22674d9bd938@codeaurora.org
Signed-off-by: Charan Teja Reddy <charante@codeaurora.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Vinayak Menon <vinmenon@codeaurora.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Link: http://lkml.kernel.org/r/1589882284-21010-1-git-send-email-charante@codeaurora.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-06 23:25:24 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-01-13 11:12:17 +00:00
|
|
|
|
mm, page_alloc: skip ->waternark_boost for atomic order-0 allocations
When boosting is enabled, it is observed that rate of atomic order-0
allocation failures are high due to the fact that free levels in the
system are checked with ->watermark_boost offset. This is not a problem
for sleepable allocations but for atomic allocations which looks like
regression.
This problem is seen frequently on system setup of Android kernel running
on Snapdragon hardware with 4GB RAM size. When no extfrag event occurred
in the system, ->watermark_boost factor is zero, thus the watermark
configurations in the system are:
_watermark = (
[WMARK_MIN] = 1272, --> ~5MB
[WMARK_LOW] = 9067, --> ~36MB
[WMARK_HIGH] = 9385), --> ~38MB
watermark_boost = 0
After launching some memory hungry applications in Android which can cause
extfrag events in the system to an extent that ->watermark_boost can be
set to max i.e. default boost factor makes it to 150% of high watermark.
_watermark = (
[WMARK_MIN] = 1272, --> ~5MB
[WMARK_LOW] = 9067, --> ~36MB
[WMARK_HIGH] = 9385), --> ~38MB
watermark_boost = 14077, -->~57MB
With default system configuration, for an atomic order-0 allocation to
succeed, having free memory of ~2MB will suffice. But boosting makes the
min_wmark to ~61MB thus for an atomic order-0 allocation to be successful
system should have minimum of ~23MB of free memory(from calculations of
zone_watermark_ok(), min = 3/4(min/2)). But failures are observed despite
system is having ~20MB of free memory. In the testing, this is
reproducible as early as first 300secs since boot and with furtherlowram
configurations(<2GB) it is observed as early as first 150secs since boot.
These failures can be avoided by excluding the ->watermark_boost in
watermark caluculations for atomic order-0 allocations.
[akpm@linux-foundation.org: fix comment grammar, reflow comment]
[charante@codeaurora.org: fix suggested by Mel Gorman]
Link: http://lkml.kernel.org/r/31556793-57b1-1c21-1a9d-22674d9bd938@codeaurora.org
Signed-off-by: Charan Teja Reddy <charante@codeaurora.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Vinayak Menon <vinmenon@codeaurora.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Link: http://lkml.kernel.org/r/1589882284-21010-1-git-send-email-charante@codeaurora.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-06 23:25:24 -07:00
|
|
|
|
2023-01-13 11:12:17 +00:00
|
|
|
|
mm, page_alloc: skip ->waternark_boost for atomic order-0 allocations
When boosting is enabled, it is observed that rate of atomic order-0
allocation failures are high due to the fact that free levels in the
system are checked with ->watermark_boost offset. This is not a problem
for sleepable allocations but for atomic allocations which looks like
regression.
This problem is seen frequently on system setup of Android kernel running
on Snapdragon hardware with 4GB RAM size. When no extfrag event occurred
in the system, ->watermark_boost factor is zero, thus the watermark
configurations in the system are:
_watermark = (
[WMARK_MIN] = 1272, --> ~5MB
[WMARK_LOW] = 9067, --> ~36MB
[WMARK_HIGH] = 9385), --> ~38MB
watermark_boost = 0
After launching some memory hungry applications in Android which can cause
extfrag events in the system to an extent that ->watermark_boost can be
set to max i.e. default boost factor makes it to 150% of high watermark.
_watermark = (
[WMARK_MIN] = 1272, --> ~5MB
[WMARK_LOW] = 9067, --> ~36MB
[WMARK_HIGH] = 9385), --> ~38MB
watermark_boost = 14077, -->~57MB
With default system configuration, for an atomic order-0 allocation to
succeed, having free memory of ~2MB will suffice. But boosting makes the
min_wmark to ~61MB thus for an atomic order-0 allocation to be successful
system should have minimum of ~23MB of free memory(from calculations of
zone_watermark_ok(), min = 3/4(min/2)). But failures are observed despite
system is having ~20MB of free memory. In the testing, this is
reproducible as early as first 300secs since boot and with furtherlowram
configurations(<2GB) it is observed as early as first 150secs since boot.
These failures can be avoided by excluding the ->watermark_boost in
watermark caluculations for atomic order-0 allocations.
[akpm@linux-foundation.org: fix comment grammar, reflow comment]
[charante@codeaurora.org: fix suggested by Mel Gorman]
Link: http://lkml.kernel.org/r/31556793-57b1-1c21-1a9d-22674d9bd938@codeaurora.org
Signed-off-by: Charan Teja Reddy <charante@codeaurora.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Vinayak Menon <vinmenon@codeaurora.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Link: http://lkml.kernel.org/r/1589882284-21010-1-git-send-email-charante@codeaurora.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-06 23:25:24 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-01-13 11:12:17 +00:00
|
|
|
|
mm, page_alloc: skip ->waternark_boost for atomic order-0 allocations
When boosting is enabled, it is observed that rate of atomic order-0
allocation failures are high due to the fact that free levels in the
system are checked with ->watermark_boost offset. This is not a problem
for sleepable allocations but for atomic allocations which looks like
regression.
This problem is seen frequently on system setup of Android kernel running
on Snapdragon hardware with 4GB RAM size. When no extfrag event occurred
in the system, ->watermark_boost factor is zero, thus the watermark
configurations in the system are:
_watermark = (
[WMARK_MIN] = 1272, --> ~5MB
[WMARK_LOW] = 9067, --> ~36MB
[WMARK_HIGH] = 9385), --> ~38MB
watermark_boost = 0
After launching some memory hungry applications in Android which can cause
extfrag events in the system to an extent that ->watermark_boost can be
set to max i.e. default boost factor makes it to 150% of high watermark.
_watermark = (
[WMARK_MIN] = 1272, --> ~5MB
[WMARK_LOW] = 9067, --> ~36MB
[WMARK_HIGH] = 9385), --> ~38MB
watermark_boost = 14077, -->~57MB
With default system configuration, for an atomic order-0 allocation to
succeed, having free memory of ~2MB will suffice. But boosting makes the
min_wmark to ~61MB thus for an atomic order-0 allocation to be successful
system should have minimum of ~23MB of free memory(from calculations of
zone_watermark_ok(), min = 3/4(min/2)). But failures are observed despite
system is having ~20MB of free memory. In the testing, this is
reproducible as early as first 300secs since boot and with furtherlowram
configurations(<2GB) it is observed as early as first 150secs since boot.
These failures can be avoided by excluding the ->watermark_boost in
watermark caluculations for atomic order-0 allocations.
[akpm@linux-foundation.org: fix comment grammar, reflow comment]
[charante@codeaurora.org: fix suggested by Mel Gorman]
Link: http://lkml.kernel.org/r/31556793-57b1-1c21-1a9d-22674d9bd938@codeaurora.org
Signed-off-by: Charan Teja Reddy <charante@codeaurora.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Vinayak Menon <vinmenon@codeaurora.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Link: http://lkml.kernel.org/r/1589882284-21010-1-git-send-email-charante@codeaurora.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-06 23:25:24 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-05-19 17:14:07 -07:00
|
|
|
|
|
|
|
|
|
[PATCH] memory page_alloc zonelist caching speedup
Optimize the critical zonelist scanning for free pages in the kernel memory
allocator by caching the zones that were found to be full recently, and
skipping them.
Remembers the zones in a zonelist that were short of free memory in the
last second. And it stashes a zone-to-node table in the zonelist struct,
to optimize that conversion (minimize its cache footprint.)
Recent changes:
This differs in a significant way from a similar patch that I
posted a week ago. Now, instead of having a nodemask_t of
recently full nodes, I have a bitmask of recently full zones.
This solves a problem that last weeks patch had, which on
systems with multiple zones per node (such as DMA zone) would
take seeing any of these zones full as meaning that all zones
on that node were full.
Also I changed names - from "zonelist faster" to "zonelist cache",
as that seemed to better convey what we're doing here - caching
some of the key zonelist state (for faster access.)
See below for some performance benchmark results. After all that
discussion with David on why I didn't need them, I went and got
some ;). I wanted to verify that I had not hurt the normal case
of memory allocation noticeably. At least for my one little
microbenchmark, I found (1) the normal case wasn't affected, and
(2) workloads that forced scanning across multiple nodes for
memory improved up to 10% fewer System CPU cycles and lower
elapsed clock time ('sys' and 'real'). Good. See details, below.
I didn't have the logic in get_page_from_freelist() for various
full nodes and zone reclaim failures correct. That should be
fixed up now - notice the new goto labels zonelist_scan,
this_zone_full, and try_next_zone, in get_page_from_freelist().
There are two reasons I persued this alternative, over some earlier
proposals that would have focused on optimizing the fake numa
emulation case by caching the last useful zone:
1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems)
have seen real customer loads where the cost to scan the zonelist
was a problem, due to many nodes being full of memory before
we got to a node we could use. Or at least, I think we have.
This was related to me by another engineer, based on experiences
from some time past. So this is not guaranteed. Most likely, though.
The following approach should help such real numa systems just as
much as it helps fake numa systems, or any combination thereof.
2) The effort to distinguish fake from real numa, using node_distance,
so that we could cache a fake numa node and optimize choosing
it over equivalent distance fake nodes, while continuing to
properly scan all real nodes in distance order, was going to
require a nasty blob of zonelist and node distance munging.
The following approach has no new dependency on node distances or
zone sorting.
See comment in the patch below for a description of what it actually does.
Technical details of note (or controversy):
- See the use of "zlc_active" and "did_zlc_setup" below, to delay
adding any work for this new mechanism until we've looked at the
first zone in zonelist. I figured the odds of the first zone
having the memory we needed were high enough that we should just
look there, first, then get fancy only if we need to keep looking.
- Some odd hackery was needed to add items to struct zonelist, while
not tripping up the custom zonelists built by the mm/mempolicy.c
code for MPOL_BIND. My usual wordy comments below explain this.
Search for "MPOL_BIND".
- Some per-node data in the struct zonelist is now modified frequently,
with no locking. Multiple CPU cores on a node could hit and mangle
this data. The theory is that this is just performance hint data,
and the memory allocator will work just fine despite any such mangling.
The fields at risk are the struct 'zonelist_cache' fields 'fullzones'
(a bitmask) and 'last_full_zap' (unsigned long jiffies). It should
all be self correcting after at most a one second delay.
- This still does a linear scan of the same lengths as before. All
I've optimized is making the scan faster, not algorithmically
shorter. It is now able to scan a compact array of 'unsigned
short' in the case of many full nodes, so one cache line should
cover quite a few nodes, rather than each node hitting another
one or two new and distinct cache lines.
- If both Andi and Nick don't find this too complicated, I will be
(pleasantly) flabbergasted.
- I removed the comment claiming we only use one cachline's worth of
zonelist. We seem, at least in the fake numa case, to have put the
lie to that claim.
- I pay no attention to the various watermarks and such in this performance
hint. A node could be marked full for one watermark, and then skipped
over when searching for a page using a different watermark. I think
that's actually quite ok, as it will tend to slightly increase the
spreading of memory over other nodes, away from a memory stressed node.
===============
Performance - some benchmark results and analysis:
This benchmark runs a memory hog program that uses multiple
threads to touch alot of memory as quickly as it can.
Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of
the total 96 GBytes on the system, and using 1, 19, 37, or 55
threads (on a 56 CPU system.) System, user and real (elapsed)
timings were recorded for each run, shown in units of seconds,
in the table below.
Two kernels were tested - 2.6.18-mm3 and the same kernel with
this zonelist caching patch added. The table also shows the
percentage improvement the zonelist caching sys time is over
(lower than) the stock *-mm kernel.
number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent
GBs N ------------ -------------- ---------------- systime
mem threads sys user real sys user real sys user real better
12 1 153 24 177 151 24 176 -2 0 -1 1%
12 19 99 22 8 99 22 8 0 0 0 0%
12 37 111 25 6 112 25 6 1 0 0 -0%
12 55 115 25 5 110 23 5 -5 -2 0 4%
38 1 502 74 576 497 73 570 -5 -1 -6 0%
38 19 426 78 48 373 76 39 -53 -2 -9 12%
38 37 544 83 36 547 82 36 3 -1 0 -0%
38 55 501 77 23 511 80 24 10 3 1 -1%
64 1 917 125 1042 890 124 1014 -27 -1 -28 2%
64 19 1118 138 119 965 141 103 -153 3 -16 13%
64 37 1202 151 94 1136 150 81 -66 -1 -13 5%
64 55 1118 141 61 1072 140 58 -46 -1 -3 4%
90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4%
90 19 2392 199 192 2116 189 176 -276 -10 -16 11%
90 37 3313 238 175 2972 225 145 -341 -13 -30 10%
90 55 1948 210 104 1843 213 100 -105 3 -4 5%
Notes:
1) This test ran a memory hog program that started a specified number N of
threads, and had each thread allocate and touch 1/N'th of
the total memory to be used in the test run in a single loop,
writing a constant word to memory, one store every 4096 bytes.
Watching this test during some earlier trial runs, I would see
each of these threads sit down on one CPU and stay there, for
the remainder of the pass, a different CPU for each thread.
2) The 'real' column is not comparable to the 'sys' or 'user' columns.
The 'real' column is seconds wall clock time elapsed, from beginning
to end of that test pass. The 'sys' and 'user' columns are total
CPU seconds spent on that test pass. For a 19 thread test run,
for example, the sum of 'sys' and 'user' could be up to 19 times the
number of 'real' elapsed wall clock seconds.
3) Tests were run on a fresh, single-user boot, to minimize the amount
of memory already in use at the start of the test, and to minimize
the amount of background activity that might interfere.
4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM.
5) Notice that the 'real' time gets large for the single thread runs, even
though the measured 'sys' and 'user' times are modest. I'm not sure what
that means - probably something to do with it being slow for one thread to
be accessing memory along ways away. Perhaps the fake numa system, running
ostensibly the same workload, would not show this substantial degradation
of 'real' time for one thread on many nodes -- lets hope not.
6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs)
ran quite efficiently, as one might expect. Each pair of threads needed
to allocate and touch the memory on the node the two threads shared, a
pleasantly parallizable workload.
7) The intermediate thread count passes, when asking for alot of memory forcing
them to go to a few neighboring nodes, improved the most with this zonelist
caching patch.
Conclusions:
* This zonelist cache patch probably makes little difference one way or the
other for most workloads on real numa hardware, if those workloads avoid
heavy off node allocations.
* For memory intensive workloads requiring substantial off-node allocations
on real numa hardware, this patch improves both kernel and elapsed timings
up to ten per-cent.
* For fake numa systems, I'm optimistic, but will have to leave that up to
Rohit Seth to actually test (once I get him a 2.6.18 backport.)
Signed-off-by: Paul Jackson <pj@sgi.com>
Cc: Rohit Seth <rohitseth@google.com>
Cc: Christoph Lameter <clameter@engr.sgi.com>
Cc: David Rientjes <rientjes@cs.washington.edu>
Cc: Paul Menage <menage@google.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-06 20:31:48 -08:00
|
|
|
|
2021-11-05 13:40:24 -07:00
|
|
|
|
|
|
|
|
|
2012-10-08 16:33:24 -07:00
|
|
|
|
|
|
|
|
|
mm/page_alloc: fix nodes for reclaim in fast path
When @node_reclaim_node isn't 0, the page allocator tries to reclaim
pages if the amount of free memory in the zones are below the low
watermark. On Power platform, none of NUMA nodes are scanned for page
reclaim because no nodes match the condition in zone_allows_reclaim().
On Power platform, RECLAIM_DISTANCE is set to 10 which is the distance
of Node-A to Node-A. So the preferred node even won't be scanned for
page reclaim.
__alloc_pages_nodemask()
get_page_from_freelist()
zone_allows_reclaim()
Anton proposed the test code as below:
# cat alloc.c
:
int main(int argc, char *argv[])
{
void *p;
unsigned long size;
unsigned long start, end;
start = time(NULL);
size = strtoul(argv[1], NULL, 0);
printf("To allocate %ldGB memory\n", size);
size <<= 30;
p = malloc(size);
assert(p);
memset(p, 0, size);
end = time(NULL);
printf("Used time: %ld seconds\n", end - start);
sleep(3600);
return 0;
}
The system I use for testing has two NUMA nodes. Both have 128GB
memory. In below scnario, the page caches on node#0 should be reclaimed
when it encounters pressure to accommodate request of allocation.
# echo 2 > /proc/sys/vm/zone_reclaim_mode; \
sync; \
echo 3 > /proc/sys/vm/drop_caches; \
# taskset -c 0 cat file.32G > /dev/null; \
grep FilePages /sys/devices/system/node/node0/meminfo
Node 0 FilePages: 33619712 kB
# taskset -c 0 ./alloc 128
# grep FilePages /sys/devices/system/node/node0/meminfo
Node 0 FilePages: 33619840 kB
# grep MemFree /sys/devices/system/node/node0/meminfo
Node 0 MemFree: 186816 kB
With the patch applied, the pagecache on node-0 is reclaimed when its
free memory is running out. It's the expected behaviour.
# echo 2 > /proc/sys/vm/zone_reclaim_mode; \
sync; \
echo 3 > /proc/sys/vm/drop_caches
# taskset -c 0 cat file.32G > /dev/null; \
grep FilePages /sys/devices/system/node/node0/meminfo
Node 0 FilePages: 33605568 kB
# taskset -c 0 ./alloc 128
# grep FilePages /sys/devices/system/node/node0/meminfo
Node 0 FilePages: 1379520 kB
# grep MemFree /sys/devices/system/node/node0/meminfo
Node 0 MemFree: 317120 kB
Fixes: 5f7a75acdb24 ("mm: page_alloc: do not cache reclaim distances")
Link: http://lkml.kernel.org/r/1486532455-29613-1-git-send-email-gwshan@linux.vnet.ibm.com
Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Anton Blanchard <anton@samba.org>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: <stable@vger.kernel.org> [3.16+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-24 14:59:33 -08:00
|
|
|
|
2019-08-08 20:53:01 +01:00
|
|
|
|
2012-10-08 16:33:24 -07:00
|
|
|
|
[PATCH] memory page_alloc zonelist caching speedup
Optimize the critical zonelist scanning for free pages in the kernel memory
allocator by caching the zones that were found to be full recently, and
skipping them.
Remembers the zones in a zonelist that were short of free memory in the
last second. And it stashes a zone-to-node table in the zonelist struct,
to optimize that conversion (minimize its cache footprint.)
Recent changes:
This differs in a significant way from a similar patch that I
posted a week ago. Now, instead of having a nodemask_t of
recently full nodes, I have a bitmask of recently full zones.
This solves a problem that last weeks patch had, which on
systems with multiple zones per node (such as DMA zone) would
take seeing any of these zones full as meaning that all zones
on that node were full.
Also I changed names - from "zonelist faster" to "zonelist cache",
as that seemed to better convey what we're doing here - caching
some of the key zonelist state (for faster access.)
See below for some performance benchmark results. After all that
discussion with David on why I didn't need them, I went and got
some ;). I wanted to verify that I had not hurt the normal case
of memory allocation noticeably. At least for my one little
microbenchmark, I found (1) the normal case wasn't affected, and
(2) workloads that forced scanning across multiple nodes for
memory improved up to 10% fewer System CPU cycles and lower
elapsed clock time ('sys' and 'real'). Good. See details, below.
I didn't have the logic in get_page_from_freelist() for various
full nodes and zone reclaim failures correct. That should be
fixed up now - notice the new goto labels zonelist_scan,
this_zone_full, and try_next_zone, in get_page_from_freelist().
There are two reasons I persued this alternative, over some earlier
proposals that would have focused on optimizing the fake numa
emulation case by caching the last useful zone:
1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems)
have seen real customer loads where the cost to scan the zonelist
was a problem, due to many nodes being full of memory before
we got to a node we could use. Or at least, I think we have.
This was related to me by another engineer, based on experiences
from some time past. So this is not guaranteed. Most likely, though.
The following approach should help such real numa systems just as
much as it helps fake numa systems, or any combination thereof.
2) The effort to distinguish fake from real numa, using node_distance,
so that we could cache a fake numa node and optimize choosing
it over equivalent distance fake nodes, while continuing to
properly scan all real nodes in distance order, was going to
require a nasty blob of zonelist and node distance munging.
The following approach has no new dependency on node distances or
zone sorting.
See comment in the patch below for a description of what it actually does.
Technical details of note (or controversy):
- See the use of "zlc_active" and "did_zlc_setup" below, to delay
adding any work for this new mechanism until we've looked at the
first zone in zonelist. I figured the odds of the first zone
having the memory we needed were high enough that we should just
look there, first, then get fancy only if we need to keep looking.
- Some odd hackery was needed to add items to struct zonelist, while
not tripping up the custom zonelists built by the mm/mempolicy.c
code for MPOL_BIND. My usual wordy comments below explain this.
Search for "MPOL_BIND".
- Some per-node data in the struct zonelist is now modified frequently,
with no locking. Multiple CPU cores on a node could hit and mangle
this data. The theory is that this is just performance hint data,
and the memory allocator will work just fine despite any such mangling.
The fields at risk are the struct 'zonelist_cache' fields 'fullzones'
(a bitmask) and 'last_full_zap' (unsigned long jiffies). It should
all be self correcting after at most a one second delay.
- This still does a linear scan of the same lengths as before. All
I've optimized is making the scan faster, not algorithmically
shorter. It is now able to scan a compact array of 'unsigned
short' in the case of many full nodes, so one cache line should
cover quite a few nodes, rather than each node hitting another
one or two new and distinct cache lines.
- If both Andi and Nick don't find this too complicated, I will be
(pleasantly) flabbergasted.
- I removed the comment claiming we only use one cachline's worth of
zonelist. We seem, at least in the fake numa case, to have put the
lie to that claim.
- I pay no attention to the various watermarks and such in this performance
hint. A node could be marked full for one watermark, and then skipped
over when searching for a page using a different watermark. I think
that's actually quite ok, as it will tend to slightly increase the
spreading of memory over other nodes, away from a memory stressed node.
===============
Performance - some benchmark results and analysis:
This benchmark runs a memory hog program that uses multiple
threads to touch alot of memory as quickly as it can.
Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of
the total 96 GBytes on the system, and using 1, 19, 37, or 55
threads (on a 56 CPU system.) System, user and real (elapsed)
timings were recorded for each run, shown in units of seconds,
in the table below.
Two kernels were tested - 2.6.18-mm3 and the same kernel with
this zonelist caching patch added. The table also shows the
percentage improvement the zonelist caching sys time is over
(lower than) the stock *-mm kernel.
number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent
GBs N ------------ -------------- ---------------- systime
mem threads sys user real sys user real sys user real better
12 1 153 24 177 151 24 176 -2 0 -1 1%
12 19 99 22 8 99 22 8 0 0 0 0%
12 37 111 25 6 112 25 6 1 0 0 -0%
12 55 115 25 5 110 23 5 -5 -2 0 4%
38 1 502 74 576 497 73 570 -5 -1 -6 0%
38 19 426 78 48 373 76 39 -53 -2 -9 12%
38 37 544 83 36 547 82 36 3 -1 0 -0%
38 55 501 77 23 511 80 24 10 3 1 -1%
64 1 917 125 1042 890 124 1014 -27 -1 -28 2%
64 19 1118 138 119 965 141 103 -153 3 -16 13%
64 37 1202 151 94 1136 150 81 -66 -1 -13 5%
64 55 1118 141 61 1072 140 58 -46 -1 -3 4%
90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4%
90 19 2392 199 192 2116 189 176 -276 -10 -16 11%
90 37 3313 238 175 2972 225 145 -341 -13 -30 10%
90 55 1948 210 104 1843 213 100 -105 3 -4 5%
Notes:
1) This test ran a memory hog program that started a specified number N of
threads, and had each thread allocate and touch 1/N'th of
the total memory to be used in the test run in a single loop,
writing a constant word to memory, one store every 4096 bytes.
Watching this test during some earlier trial runs, I would see
each of these threads sit down on one CPU and stay there, for
the remainder of the pass, a different CPU for each thread.
2) The 'real' column is not comparable to the 'sys' or 'user' columns.
The 'real' column is seconds wall clock time elapsed, from beginning
to end of that test pass. The 'sys' and 'user' columns are total
CPU seconds spent on that test pass. For a 19 thread test run,
for example, the sum of 'sys' and 'user' could be up to 19 times the
number of 'real' elapsed wall clock seconds.
3) Tests were run on a fresh, single-user boot, to minimize the amount
of memory already in use at the start of the test, and to minimize
the amount of background activity that might interfere.
4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM.
5) Notice that the 'real' time gets large for the single thread runs, even
though the measured 'sys' and 'user' times are modest. I'm not sure what
that means - probably something to do with it being slow for one thread to
be accessing memory along ways away. Perhaps the fake numa system, running
ostensibly the same workload, would not show this substantial degradation
of 'real' time for one thread on many nodes -- lets hope not.
6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs)
ran quite efficiently, as one might expect. Each pair of threads needed
to allocate and touch the memory on the node the two threads shared, a
pleasantly parallizable workload.
7) The intermediate thread count passes, when asking for alot of memory forcing
them to go to a few neighboring nodes, improved the most with this zonelist
caching patch.
Conclusions:
* This zonelist cache patch probably makes little difference one way or the
other for most workloads on real numa hardware, if those workloads avoid
heavy off node allocations.
* For memory intensive workloads requiring substantial off-node allocations
on real numa hardware, this patch improves both kernel and elapsed timings
up to ten per-cent.
* For fake numa systems, I'm optimistic, but will have to leave that up to
Rohit Seth to actually test (once I get him a 2.6.18 backport.)
Signed-off-by: Paul Jackson <pj@sgi.com>
Cc: Rohit Seth <rohitseth@google.com>
Cc: Christoph Lameter <clameter@engr.sgi.com>
Cc: David Rientjes <rientjes@cs.washington.edu>
Cc: Paul Menage <menage@google.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-06 20:31:48 -08:00
|
|
|
|
2012-10-08 16:33:24 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
[PATCH] memory page_alloc zonelist caching speedup
Optimize the critical zonelist scanning for free pages in the kernel memory
allocator by caching the zones that were found to be full recently, and
skipping them.
Remembers the zones in a zonelist that were short of free memory in the
last second. And it stashes a zone-to-node table in the zonelist struct,
to optimize that conversion (minimize its cache footprint.)
Recent changes:
This differs in a significant way from a similar patch that I
posted a week ago. Now, instead of having a nodemask_t of
recently full nodes, I have a bitmask of recently full zones.
This solves a problem that last weeks patch had, which on
systems with multiple zones per node (such as DMA zone) would
take seeing any of these zones full as meaning that all zones
on that node were full.
Also I changed names - from "zonelist faster" to "zonelist cache",
as that seemed to better convey what we're doing here - caching
some of the key zonelist state (for faster access.)
See below for some performance benchmark results. After all that
discussion with David on why I didn't need them, I went and got
some ;). I wanted to verify that I had not hurt the normal case
of memory allocation noticeably. At least for my one little
microbenchmark, I found (1) the normal case wasn't affected, and
(2) workloads that forced scanning across multiple nodes for
memory improved up to 10% fewer System CPU cycles and lower
elapsed clock time ('sys' and 'real'). Good. See details, below.
I didn't have the logic in get_page_from_freelist() for various
full nodes and zone reclaim failures correct. That should be
fixed up now - notice the new goto labels zonelist_scan,
this_zone_full, and try_next_zone, in get_page_from_freelist().
There are two reasons I persued this alternative, over some earlier
proposals that would have focused on optimizing the fake numa
emulation case by caching the last useful zone:
1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems)
have seen real customer loads where the cost to scan the zonelist
was a problem, due to many nodes being full of memory before
we got to a node we could use. Or at least, I think we have.
This was related to me by another engineer, based on experiences
from some time past. So this is not guaranteed. Most likely, though.
The following approach should help such real numa systems just as
much as it helps fake numa systems, or any combination thereof.
2) The effort to distinguish fake from real numa, using node_distance,
so that we could cache a fake numa node and optimize choosing
it over equivalent distance fake nodes, while continuing to
properly scan all real nodes in distance order, was going to
require a nasty blob of zonelist and node distance munging.
The following approach has no new dependency on node distances or
zone sorting.
See comment in the patch below for a description of what it actually does.
Technical details of note (or controversy):
- See the use of "zlc_active" and "did_zlc_setup" below, to delay
adding any work for this new mechanism until we've looked at the
first zone in zonelist. I figured the odds of the first zone
having the memory we needed were high enough that we should just
look there, first, then get fancy only if we need to keep looking.
- Some odd hackery was needed to add items to struct zonelist, while
not tripping up the custom zonelists built by the mm/mempolicy.c
code for MPOL_BIND. My usual wordy comments below explain this.
Search for "MPOL_BIND".
- Some per-node data in the struct zonelist is now modified frequently,
with no locking. Multiple CPU cores on a node could hit and mangle
this data. The theory is that this is just performance hint data,
and the memory allocator will work just fine despite any such mangling.
The fields at risk are the struct 'zonelist_cache' fields 'fullzones'
(a bitmask) and 'last_full_zap' (unsigned long jiffies). It should
all be self correcting after at most a one second delay.
- This still does a linear scan of the same lengths as before. All
I've optimized is making the scan faster, not algorithmically
shorter. It is now able to scan a compact array of 'unsigned
short' in the case of many full nodes, so one cache line should
cover quite a few nodes, rather than each node hitting another
one or two new and distinct cache lines.
- If both Andi and Nick don't find this too complicated, I will be
(pleasantly) flabbergasted.
- I removed the comment claiming we only use one cachline's worth of
zonelist. We seem, at least in the fake numa case, to have put the
lie to that claim.
- I pay no attention to the various watermarks and such in this performance
hint. A node could be marked full for one watermark, and then skipped
over when searching for a page using a different watermark. I think
that's actually quite ok, as it will tend to slightly increase the
spreading of memory over other nodes, away from a memory stressed node.
===============
Performance - some benchmark results and analysis:
This benchmark runs a memory hog program that uses multiple
threads to touch alot of memory as quickly as it can.
Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of
the total 96 GBytes on the system, and using 1, 19, 37, or 55
threads (on a 56 CPU system.) System, user and real (elapsed)
timings were recorded for each run, shown in units of seconds,
in the table below.
Two kernels were tested - 2.6.18-mm3 and the same kernel with
this zonelist caching patch added. The table also shows the
percentage improvement the zonelist caching sys time is over
(lower than) the stock *-mm kernel.
number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent
GBs N ------------ -------------- ---------------- systime
mem threads sys user real sys user real sys user real better
12 1 153 24 177 151 24 176 -2 0 -1 1%
12 19 99 22 8 99 22 8 0 0 0 0%
12 37 111 25 6 112 25 6 1 0 0 -0%
12 55 115 25 5 110 23 5 -5 -2 0 4%
38 1 502 74 576 497 73 570 -5 -1 -6 0%
38 19 426 78 48 373 76 39 -53 -2 -9 12%
38 37 544 83 36 547 82 36 3 -1 0 -0%
38 55 501 77 23 511 80 24 10 3 1 -1%
64 1 917 125 1042 890 124 1014 -27 -1 -28 2%
64 19 1118 138 119 965 141 103 -153 3 -16 13%
64 37 1202 151 94 1136 150 81 -66 -1 -13 5%
64 55 1118 141 61 1072 140 58 -46 -1 -3 4%
90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4%
90 19 2392 199 192 2116 189 176 -276 -10 -16 11%
90 37 3313 238 175 2972 225 145 -341 -13 -30 10%
90 55 1948 210 104 1843 213 100 -105 3 -4 5%
Notes:
1) This test ran a memory hog program that started a specified number N of
threads, and had each thread allocate and touch 1/N'th of
the total memory to be used in the test run in a single loop,
writing a constant word to memory, one store every 4096 bytes.
Watching this test during some earlier trial runs, I would see
each of these threads sit down on one CPU and stay there, for
the remainder of the pass, a different CPU for each thread.
2) The 'real' column is not comparable to the 'sys' or 'user' columns.
The 'real' column is seconds wall clock time elapsed, from beginning
to end of that test pass. The 'sys' and 'user' columns are total
CPU seconds spent on that test pass. For a 19 thread test run,
for example, the sum of 'sys' and 'user' could be up to 19 times the
number of 'real' elapsed wall clock seconds.
3) Tests were run on a fresh, single-user boot, to minimize the amount
of memory already in use at the start of the test, and to minimize
the amount of background activity that might interfere.
4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM.
5) Notice that the 'real' time gets large for the single thread runs, even
though the measured 'sys' and 'user' times are modest. I'm not sure what
that means - probably something to do with it being slow for one thread to
be accessing memory along ways away. Perhaps the fake numa system, running
ostensibly the same workload, would not show this substantial degradation
of 'real' time for one thread on many nodes -- lets hope not.
6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs)
ran quite efficiently, as one might expect. Each pair of threads needed
to allocate and touch the memory on the node the two threads shared, a
pleasantly parallizable workload.
7) The intermediate thread count passes, when asking for alot of memory forcing
them to go to a few neighboring nodes, improved the most with this zonelist
caching patch.
Conclusions:
* This zonelist cache patch probably makes little difference one way or the
other for most workloads on real numa hardware, if those workloads avoid
heavy off node allocations.
* For memory intensive workloads requiring substantial off-node allocations
on real numa hardware, this patch improves both kernel and elapsed timings
up to ten per-cent.
* For fake numa systems, I'm optimistic, but will have to leave that up to
Rohit Seth to actually test (once I get him a 2.6.18 backport.)
Signed-off-by: Paul Jackson <pj@sgi.com>
Cc: Rohit Seth <rohitseth@google.com>
Cc: Christoph Lameter <clameter@engr.sgi.com>
Cc: David Rientjes <rientjes@cs.washington.edu>
Cc: Paul Menage <menage@google.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-06 20:31:48 -08:00
|
|
|
|
|
|
|
|
|
mm, page_alloc: spread allocations across zones before introducing fragmentation
Patch series "Fragmentation avoidance improvements", v5.
It has been noted before that fragmentation avoidance (aka
anti-fragmentation) is not perfect. Given sufficient time or an adverse
workload, memory gets fragmented and the long-term success of high-order
allocations degrades. This series defines an adverse workload, a definition
of external fragmentation events (including serious) ones and a series
that reduces the level of those fragmentation events.
The details of the workload and the consequences are described in more
detail in the changelogs. However, from patch 1, this is a high-level
summary of the adverse workload. The exact details are found in the
mmtests implementation.
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch)
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameterr create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed
3. Warm up a number of fio read-only threads accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll fault back in old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup
Overall the series reduces external fragmentation causing events by over 94%
on 1 and 2 socket machines, which in turn impacts high-order allocation
success rates over the long term. There are differences in latencies and
high-order allocation success rates. Latencies are a mixed bag as they
are vulnerable to exact system state and whether allocations succeeded
so they are treated as a secondary metric.
Patch 1 uses lower zones if they are populated and have free memory
instead of fragmenting a higher zone. It's special cased to
handle a Normal->DMA32 fallback with the reasons explained
in the changelog.
Patch 2-4 boosts watermarks temporarily when an external fragmentation
event occurs. kswapd wakes to reclaim a small amount of old memory
and then wakes kcompactd on completion to recover the system
slightly. This introduces some overhead in the slowpath. The level
of boosting can be tuned or disabled depending on the tolerance
for fragmentation vs allocation latency.
Patch 5 stalls some movable allocation requests to let kswapd from patch 4
make some progress. The duration of the stalls is very low but it
is possible to tune the system to avoid fragmentation events if
larger stalls can be tolerated.
The bulk of the improvement in fragmentation avoidance is from patches
1-4 but patch 5 can deal with a rare corner case and provides the option
of tuning a system for THP allocation success rates in exchange for
some stalls to control fragmentation.
This patch (of 5):
The page allocator zone lists are iterated based on the watermarks of each
zone which does not take anti-fragmentation into account. On x86, node 0
may have multiple zones while other nodes have one zone. A consequence is
that tasks running on node 0 may fragment ZONE_NORMAL even though
ZONE_DMA32 has plenty of free memory. This patch special cases the
allocator fast path such that it'll try an allocation from a lower local
zone before fragmenting a higher zone. In this case, stealing of
pageblocks or orders larger than a pageblock are still allowed in the fast
path as they are uninteresting from a fragmentation point of view.
This was evaluated using a benchmark designed to fragment memory before
attempting THP allocations. It's implemented in mmtests as the following
configurations
configs/config-global-dhp__workload_thpfioscale
configs/config-global-dhp__workload_thpfioscale-defrag
configs/config-global-dhp__workload_thpfioscale-madvhugepage
e.g. from mmtests
./run-mmtests.sh --run-monitor --config configs/config-global-dhp__workload_thpfioscale test-run-1
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch).
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameter create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed.
3. Warm up a number of fio read-only processes accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll refault old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds.
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup the test files.
Note that due to the use of IO and page cache that this benchmark is not
suitable for running on large machines where the time to fragment memory
may be excessive. Also note that while this is one mix that generates
fragmentation that it's not the only mix that generates fragmentation.
Differences in workload that are more slab-intensive or whether SLUB is
used with high-order pages may yield different results.
When the page allocator fragments memory, it records the event using the
mm_page_alloc_extfrag ftrace event. If the fallback_order is smaller than
a pageblock order (order-9 on 64-bit x86) then it's considered to be an
"external fragmentation event" that may cause issues in the future.
Hence, the primary metric here is the number of external fragmentation
events that occur with order < 9. The secondary metric is allocation
latency and huge page allocation success rates but note that differences
in latencies and what the success rate also can affect the number of
external fragmentation event which is why it's a secondary metric.
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 662.92 ( 0.00%) 653.58 * 1.41%*
Amean fault-huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
Fault latencies are slightly reduced while allocation success rates remain
at zero as this configuration does not make any special effort to allocate
THP and fio is heavily active at the time and either filling memory or
keeping pages resident. However, a 49% reduction of serious fragmentation
events reduces the changes of external fragmentation being a problem in
the future.
Vlastimil asked during review for a breakdown of the allocation types
that are falling back.
vanilla
3816 MIGRATE_UNMOVABLE
800845 MIGRATE_MOVABLE
33 MIGRATE_UNRECLAIMABLE
patch
735 MIGRATE_UNMOVABLE
408135 MIGRATE_MOVABLE
42 MIGRATE_UNRECLAIMABLE
The majority of the fallbacks are due to movable allocations and this is
consistent for the workload throughout the series so will not be presented
again as the primary source of fallbacks are movable allocations.
Movable fallbacks are sometimes considered "ok" to fallback because they
can be migrated. The problem is that they can fill an
unmovable/reclaimable pageblock causing those allocations to fallback
later and polluting pageblocks with pages that cannot move. If there is a
movable fallback, it is pretty much guaranteed to affect an
unmovable/reclaimable pageblock and while it might not be enough to
actually cause a unmovable/reclaimable fallback in the future, we cannot
know that in advance so the patch takes the only option available to it.
Hence, it's important to control them. This point is also consistent
throughout the series and will not be repeated.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 1495.14 ( 0.00%) 1467.55 ( 1.85%)
Amean fault-huge-1 1098.48 ( 0.00%) 1127.11 ( -2.61%)
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 78.57 ( 0.00%) 77.64 ( -1.18%)
Fragmentation events were reduced quite a bit although this is known
to be a little variable. The latencies and allocation success rates
are similar but they were already quite high.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 1350.05 ( 0.00%) 1346.45 ( 0.27%)
Amean fault-huge-5 4181.01 ( 0.00%) 3418.60 ( 18.24%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 1.15 ( 0.00%) 0.78 ( -31.88%)
The reduction of external fragmentation events is slight and this is
partially due to the removal of __GFP_THISNODE in commit ac5b2c18911f
("mm: thp: relax __GFP_THISNODE for MADV_HUGEPAGE mappings") as THP
allocations can now spill over to remote nodes instead of fragmenting
local memory.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 6138.97 ( 0.00%) 6217.43 ( -1.28%)
Amean fault-huge-5 2294.28 ( 0.00%) 3163.33 * -37.88%*
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 96.82 ( 0.00%) 95.14 ( -1.74%)
There was a slight reduction in external fragmentation events although the
latencies were higher. The allocation success rate is high enough that
the system is struggling and there is quite a lot of parallel reclaim and
compaction activity. There is also a certain degree of luck on whether
processes start on node 0 or not for this patch but the relevance is
reduced later in the series.
Overall, the patch reduces the number of external fragmentation causing
events so the success of THP over long periods of time would be improved
for this adverse workload.
Link: http://lkml.kernel.org/r/20181123114528.28802-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:41 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-12-28 00:35:48 -08:00
|
|
|
|
mm, page_alloc: spread allocations across zones before introducing fragmentation
Patch series "Fragmentation avoidance improvements", v5.
It has been noted before that fragmentation avoidance (aka
anti-fragmentation) is not perfect. Given sufficient time or an adverse
workload, memory gets fragmented and the long-term success of high-order
allocations degrades. This series defines an adverse workload, a definition
of external fragmentation events (including serious) ones and a series
that reduces the level of those fragmentation events.
The details of the workload and the consequences are described in more
detail in the changelogs. However, from patch 1, this is a high-level
summary of the adverse workload. The exact details are found in the
mmtests implementation.
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch)
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameterr create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed
3. Warm up a number of fio read-only threads accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll fault back in old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup
Overall the series reduces external fragmentation causing events by over 94%
on 1 and 2 socket machines, which in turn impacts high-order allocation
success rates over the long term. There are differences in latencies and
high-order allocation success rates. Latencies are a mixed bag as they
are vulnerable to exact system state and whether allocations succeeded
so they are treated as a secondary metric.
Patch 1 uses lower zones if they are populated and have free memory
instead of fragmenting a higher zone. It's special cased to
handle a Normal->DMA32 fallback with the reasons explained
in the changelog.
Patch 2-4 boosts watermarks temporarily when an external fragmentation
event occurs. kswapd wakes to reclaim a small amount of old memory
and then wakes kcompactd on completion to recover the system
slightly. This introduces some overhead in the slowpath. The level
of boosting can be tuned or disabled depending on the tolerance
for fragmentation vs allocation latency.
Patch 5 stalls some movable allocation requests to let kswapd from patch 4
make some progress. The duration of the stalls is very low but it
is possible to tune the system to avoid fragmentation events if
larger stalls can be tolerated.
The bulk of the improvement in fragmentation avoidance is from patches
1-4 but patch 5 can deal with a rare corner case and provides the option
of tuning a system for THP allocation success rates in exchange for
some stalls to control fragmentation.
This patch (of 5):
The page allocator zone lists are iterated based on the watermarks of each
zone which does not take anti-fragmentation into account. On x86, node 0
may have multiple zones while other nodes have one zone. A consequence is
that tasks running on node 0 may fragment ZONE_NORMAL even though
ZONE_DMA32 has plenty of free memory. This patch special cases the
allocator fast path such that it'll try an allocation from a lower local
zone before fragmenting a higher zone. In this case, stealing of
pageblocks or orders larger than a pageblock are still allowed in the fast
path as they are uninteresting from a fragmentation point of view.
This was evaluated using a benchmark designed to fragment memory before
attempting THP allocations. It's implemented in mmtests as the following
configurations
configs/config-global-dhp__workload_thpfioscale
configs/config-global-dhp__workload_thpfioscale-defrag
configs/config-global-dhp__workload_thpfioscale-madvhugepage
e.g. from mmtests
./run-mmtests.sh --run-monitor --config configs/config-global-dhp__workload_thpfioscale test-run-1
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch).
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameter create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed.
3. Warm up a number of fio read-only processes accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll refault old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds.
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup the test files.
Note that due to the use of IO and page cache that this benchmark is not
suitable for running on large machines where the time to fragment memory
may be excessive. Also note that while this is one mix that generates
fragmentation that it's not the only mix that generates fragmentation.
Differences in workload that are more slab-intensive or whether SLUB is
used with high-order pages may yield different results.
When the page allocator fragments memory, it records the event using the
mm_page_alloc_extfrag ftrace event. If the fallback_order is smaller than
a pageblock order (order-9 on 64-bit x86) then it's considered to be an
"external fragmentation event" that may cause issues in the future.
Hence, the primary metric here is the number of external fragmentation
events that occur with order < 9. The secondary metric is allocation
latency and huge page allocation success rates but note that differences
in latencies and what the success rate also can affect the number of
external fragmentation event which is why it's a secondary metric.
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 662.92 ( 0.00%) 653.58 * 1.41%*
Amean fault-huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
Fault latencies are slightly reduced while allocation success rates remain
at zero as this configuration does not make any special effort to allocate
THP and fio is heavily active at the time and either filling memory or
keeping pages resident. However, a 49% reduction of serious fragmentation
events reduces the changes of external fragmentation being a problem in
the future.
Vlastimil asked during review for a breakdown of the allocation types
that are falling back.
vanilla
3816 MIGRATE_UNMOVABLE
800845 MIGRATE_MOVABLE
33 MIGRATE_UNRECLAIMABLE
patch
735 MIGRATE_UNMOVABLE
408135 MIGRATE_MOVABLE
42 MIGRATE_UNRECLAIMABLE
The majority of the fallbacks are due to movable allocations and this is
consistent for the workload throughout the series so will not be presented
again as the primary source of fallbacks are movable allocations.
Movable fallbacks are sometimes considered "ok" to fallback because they
can be migrated. The problem is that they can fill an
unmovable/reclaimable pageblock causing those allocations to fallback
later and polluting pageblocks with pages that cannot move. If there is a
movable fallback, it is pretty much guaranteed to affect an
unmovable/reclaimable pageblock and while it might not be enough to
actually cause a unmovable/reclaimable fallback in the future, we cannot
know that in advance so the patch takes the only option available to it.
Hence, it's important to control them. This point is also consistent
throughout the series and will not be repeated.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 1495.14 ( 0.00%) 1467.55 ( 1.85%)
Amean fault-huge-1 1098.48 ( 0.00%) 1127.11 ( -2.61%)
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 78.57 ( 0.00%) 77.64 ( -1.18%)
Fragmentation events were reduced quite a bit although this is known
to be a little variable. The latencies and allocation success rates
are similar but they were already quite high.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 1350.05 ( 0.00%) 1346.45 ( 0.27%)
Amean fault-huge-5 4181.01 ( 0.00%) 3418.60 ( 18.24%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 1.15 ( 0.00%) 0.78 ( -31.88%)
The reduction of external fragmentation events is slight and this is
partially due to the removal of __GFP_THISNODE in commit ac5b2c18911f
("mm: thp: relax __GFP_THISNODE for MADV_HUGEPAGE mappings") as THP
allocations can now spill over to remote nodes instead of fragmenting
local memory.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 6138.97 ( 0.00%) 6217.43 ( -1.28%)
Amean fault-huge-5 2294.28 ( 0.00%) 3163.33 * -37.88%*
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 96.82 ( 0.00%) 95.14 ( -1.74%)
There was a slight reduction in external fragmentation events although the
latencies were higher. The allocation success rate is high enough that
the system is struggling and there is quite a lot of parallel reclaim and
compaction activity. There is also a certain degree of luck on whether
processes start on node 0 or not for this patch but the relevance is
reduced later in the series.
Overall, the patch reduces the number of external fragmentation causing
events so the success of THP over long periods of time would be improved
for this adverse workload.
Link: http://lkml.kernel.org/r/20181123114528.28802-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:41 -08:00
|
|
|
|
2020-04-01 21:09:47 -07:00
|
|
|
|
2018-12-28 00:35:48 -08:00
|
|
|
|
2020-04-01 21:09:47 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-12-28 00:35:48 -08:00
|
|
|
|
mm: page_alloc: defrag_mode
The page allocator groups requests by migratetype to stave off
fragmentation. However, in practice this is routinely defeated by the
fact that it gives up *before* invoking reclaim and compaction - which may
well produce suitable pages. As a result, fragmentation of physical
memory is a common ongoing process in many load scenarios.
Fragmentation deteriorates compaction's ability to produce huge pages.
Depending on the lifetime of the fragmenting allocations, those effects
can be long-lasting or even permanent, requiring drastic measures like
forcible idle states or even reboots as the only reliable ways to recover
the address space for THP production.
In a kernel build test with supplemental THP pressure, the THP allocation
rate steadily declines over 15 runs:
thp_fault_alloc
61988
56474
57258
50187
52388
55409
52925
47648
43669
40621
36077
41721
36685
34641
33215
This is a hurdle in adopting THP in any environment where hosts are shared
between multiple overlapping workloads (cloud environments), and rarely
experience true idle periods. To make THP a reliable and predictable
optimization, there needs to be a stronger guarantee to avoid such
fragmentation.
Introduce defrag_mode. When enabled, reclaim/compaction is invoked to its
full extent *before* falling back. Specifically, ALLOC_NOFRAGMENT is
enforced on the allocator fastpath and the reclaiming slowpath.
For now, fallbacks are permitted to avert OOMs. There is a plan to add
defrag_mode=2 to prefer OOMs over fragmentation, but this requires
additional prep work in compaction and the reserve management to make it
ready for all possible allocation contexts.
The following test results are from a kernel build with periodic bursts of
THP allocations, over 15 runs:
vanilla defrag_mode=1
@claimer[unmovable]: 189 103
@claimer[movable]: 92 103
@claimer[reclaimable]: 207 61
@pollute[unmovable from movable]: 25 0
@pollute[unmovable from reclaimable]: 28 0
@pollute[movable from unmovable]: 38835 0
@pollute[movable from reclaimable]: 147136 0
@pollute[reclaimable from unmovable]: 178 0
@pollute[reclaimable from movable]: 33 0
@steal[unmovable from movable]: 11 0
@steal[unmovable from reclaimable]: 5 0
@steal[reclaimable from unmovable]: 107 0
@steal[reclaimable from movable]: 90 0
@steal[movable from reclaimable]: 354 0
@steal[movable from unmovable]: 130 0
Both types of polluting fallbacks are eliminated in this workload.
Interestingly, whole block conversions are reduced as well. This is
because once a block is claimed for a type, its empty space remains
available for future allocations, instead of being padded with fallbacks;
this allows the native type to group up instead of spreading out to new
blocks. The assumption in the allocator has been that pollution from
movable allocations is less harmful than from other types, since they can
be reclaimed or migrated out should the space be needed. However, since
fallbacks occur *before* reclaim/compaction is invoked, movable pollution
will still cause non-movable allocations to spread out and claim more
blocks.
Without fragmentation, THP rates hold steady with defrag_mode=1:
thp_fault_alloc
32478
20725
45045
32130
14018
21711
40791
29134
34458
45381
28305
17265
22584
28454
30850
While the downward trend is eliminated, the keen reader will of course
notice that the baseline rate is much smaller than the vanilla kernel's to
begin with. This is due to deficiencies in how reclaim and compaction are
currently driven: ALLOC_NOFRAGMENT increases the extent to which smaller
allocations are competing with THPs for pageblocks, while making no effort
themselves to reclaim or compact beyond their own request size. This
effect already exists with the current usage of ALLOC_NOFRAGMENT, but is
amplified by defrag_mode insisting on whole block stealing much more
strongly.
Subsequent patches will address defrag_mode reclaim strategy to raise the
THP success baseline above the vanilla kernel.
Link: https://lkml.kernel.org/r/20250313210647.1314586-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-03-13 17:05:34 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-12-28 00:35:48 -08:00
|
|
|
|
2019-04-25 22:23:58 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, page_alloc: spread allocations across zones before introducing fragmentation
Patch series "Fragmentation avoidance improvements", v5.
It has been noted before that fragmentation avoidance (aka
anti-fragmentation) is not perfect. Given sufficient time or an adverse
workload, memory gets fragmented and the long-term success of high-order
allocations degrades. This series defines an adverse workload, a definition
of external fragmentation events (including serious) ones and a series
that reduces the level of those fragmentation events.
The details of the workload and the consequences are described in more
detail in the changelogs. However, from patch 1, this is a high-level
summary of the adverse workload. The exact details are found in the
mmtests implementation.
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch)
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameterr create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed
3. Warm up a number of fio read-only threads accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll fault back in old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup
Overall the series reduces external fragmentation causing events by over 94%
on 1 and 2 socket machines, which in turn impacts high-order allocation
success rates over the long term. There are differences in latencies and
high-order allocation success rates. Latencies are a mixed bag as they
are vulnerable to exact system state and whether allocations succeeded
so they are treated as a secondary metric.
Patch 1 uses lower zones if they are populated and have free memory
instead of fragmenting a higher zone. It's special cased to
handle a Normal->DMA32 fallback with the reasons explained
in the changelog.
Patch 2-4 boosts watermarks temporarily when an external fragmentation
event occurs. kswapd wakes to reclaim a small amount of old memory
and then wakes kcompactd on completion to recover the system
slightly. This introduces some overhead in the slowpath. The level
of boosting can be tuned or disabled depending on the tolerance
for fragmentation vs allocation latency.
Patch 5 stalls some movable allocation requests to let kswapd from patch 4
make some progress. The duration of the stalls is very low but it
is possible to tune the system to avoid fragmentation events if
larger stalls can be tolerated.
The bulk of the improvement in fragmentation avoidance is from patches
1-4 but patch 5 can deal with a rare corner case and provides the option
of tuning a system for THP allocation success rates in exchange for
some stalls to control fragmentation.
This patch (of 5):
The page allocator zone lists are iterated based on the watermarks of each
zone which does not take anti-fragmentation into account. On x86, node 0
may have multiple zones while other nodes have one zone. A consequence is
that tasks running on node 0 may fragment ZONE_NORMAL even though
ZONE_DMA32 has plenty of free memory. This patch special cases the
allocator fast path such that it'll try an allocation from a lower local
zone before fragmenting a higher zone. In this case, stealing of
pageblocks or orders larger than a pageblock are still allowed in the fast
path as they are uninteresting from a fragmentation point of view.
This was evaluated using a benchmark designed to fragment memory before
attempting THP allocations. It's implemented in mmtests as the following
configurations
configs/config-global-dhp__workload_thpfioscale
configs/config-global-dhp__workload_thpfioscale-defrag
configs/config-global-dhp__workload_thpfioscale-madvhugepage
e.g. from mmtests
./run-mmtests.sh --run-monitor --config configs/config-global-dhp__workload_thpfioscale test-run-1
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch).
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameter create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed.
3. Warm up a number of fio read-only processes accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll refault old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds.
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup the test files.
Note that due to the use of IO and page cache that this benchmark is not
suitable for running on large machines where the time to fragment memory
may be excessive. Also note that while this is one mix that generates
fragmentation that it's not the only mix that generates fragmentation.
Differences in workload that are more slab-intensive or whether SLUB is
used with high-order pages may yield different results.
When the page allocator fragments memory, it records the event using the
mm_page_alloc_extfrag ftrace event. If the fallback_order is smaller than
a pageblock order (order-9 on 64-bit x86) then it's considered to be an
"external fragmentation event" that may cause issues in the future.
Hence, the primary metric here is the number of external fragmentation
events that occur with order < 9. The secondary metric is allocation
latency and huge page allocation success rates but note that differences
in latencies and what the success rate also can affect the number of
external fragmentation event which is why it's a secondary metric.
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 662.92 ( 0.00%) 653.58 * 1.41%*
Amean fault-huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
Fault latencies are slightly reduced while allocation success rates remain
at zero as this configuration does not make any special effort to allocate
THP and fio is heavily active at the time and either filling memory or
keeping pages resident. However, a 49% reduction of serious fragmentation
events reduces the changes of external fragmentation being a problem in
the future.
Vlastimil asked during review for a breakdown of the allocation types
that are falling back.
vanilla
3816 MIGRATE_UNMOVABLE
800845 MIGRATE_MOVABLE
33 MIGRATE_UNRECLAIMABLE
patch
735 MIGRATE_UNMOVABLE
408135 MIGRATE_MOVABLE
42 MIGRATE_UNRECLAIMABLE
The majority of the fallbacks are due to movable allocations and this is
consistent for the workload throughout the series so will not be presented
again as the primary source of fallbacks are movable allocations.
Movable fallbacks are sometimes considered "ok" to fallback because they
can be migrated. The problem is that they can fill an
unmovable/reclaimable pageblock causing those allocations to fallback
later and polluting pageblocks with pages that cannot move. If there is a
movable fallback, it is pretty much guaranteed to affect an
unmovable/reclaimable pageblock and while it might not be enough to
actually cause a unmovable/reclaimable fallback in the future, we cannot
know that in advance so the patch takes the only option available to it.
Hence, it's important to control them. This point is also consistent
throughout the series and will not be repeated.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 1495.14 ( 0.00%) 1467.55 ( 1.85%)
Amean fault-huge-1 1098.48 ( 0.00%) 1127.11 ( -2.61%)
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 78.57 ( 0.00%) 77.64 ( -1.18%)
Fragmentation events were reduced quite a bit although this is known
to be a little variable. The latencies and allocation success rates
are similar but they were already quite high.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 1350.05 ( 0.00%) 1346.45 ( 0.27%)
Amean fault-huge-5 4181.01 ( 0.00%) 3418.60 ( 18.24%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 1.15 ( 0.00%) 0.78 ( -31.88%)
The reduction of external fragmentation events is slight and this is
partially due to the removal of __GFP_THISNODE in commit ac5b2c18911f
("mm: thp: relax __GFP_THISNODE for MADV_HUGEPAGE mappings") as THP
allocations can now spill over to remote nodes instead of fragmenting
local memory.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 6138.97 ( 0.00%) 6217.43 ( -1.28%)
Amean fault-huge-5 2294.28 ( 0.00%) 3163.33 * -37.88%*
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 96.82 ( 0.00%) 95.14 ( -1.74%)
There was a slight reduction in external fragmentation events although the
latencies were higher. The allocation success rate is high enough that
the system is struggling and there is quite a lot of parallel reclaim and
compaction activity. There is also a certain degree of luck on whether
processes start on node 0 or not for this patch but the relevance is
reduced later in the series.
Overall, the patch reduces the number of external fragmentation causing
events so the success of THP over long periods of time would be improved
for this adverse workload.
Link: http://lkml.kernel.org/r/20181123114528.28802-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:41 -08:00
|
|
|
|
2019-04-25 22:24:01 -07:00
|
|
|
|
mm, page_alloc: spread allocations across zones before introducing fragmentation
Patch series "Fragmentation avoidance improvements", v5.
It has been noted before that fragmentation avoidance (aka
anti-fragmentation) is not perfect. Given sufficient time or an adverse
workload, memory gets fragmented and the long-term success of high-order
allocations degrades. This series defines an adverse workload, a definition
of external fragmentation events (including serious) ones and a series
that reduces the level of those fragmentation events.
The details of the workload and the consequences are described in more
detail in the changelogs. However, from patch 1, this is a high-level
summary of the adverse workload. The exact details are found in the
mmtests implementation.
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch)
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameterr create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed
3. Warm up a number of fio read-only threads accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll fault back in old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup
Overall the series reduces external fragmentation causing events by over 94%
on 1 and 2 socket machines, which in turn impacts high-order allocation
success rates over the long term. There are differences in latencies and
high-order allocation success rates. Latencies are a mixed bag as they
are vulnerable to exact system state and whether allocations succeeded
so they are treated as a secondary metric.
Patch 1 uses lower zones if they are populated and have free memory
instead of fragmenting a higher zone. It's special cased to
handle a Normal->DMA32 fallback with the reasons explained
in the changelog.
Patch 2-4 boosts watermarks temporarily when an external fragmentation
event occurs. kswapd wakes to reclaim a small amount of old memory
and then wakes kcompactd on completion to recover the system
slightly. This introduces some overhead in the slowpath. The level
of boosting can be tuned or disabled depending on the tolerance
for fragmentation vs allocation latency.
Patch 5 stalls some movable allocation requests to let kswapd from patch 4
make some progress. The duration of the stalls is very low but it
is possible to tune the system to avoid fragmentation events if
larger stalls can be tolerated.
The bulk of the improvement in fragmentation avoidance is from patches
1-4 but patch 5 can deal with a rare corner case and provides the option
of tuning a system for THP allocation success rates in exchange for
some stalls to control fragmentation.
This patch (of 5):
The page allocator zone lists are iterated based on the watermarks of each
zone which does not take anti-fragmentation into account. On x86, node 0
may have multiple zones while other nodes have one zone. A consequence is
that tasks running on node 0 may fragment ZONE_NORMAL even though
ZONE_DMA32 has plenty of free memory. This patch special cases the
allocator fast path such that it'll try an allocation from a lower local
zone before fragmenting a higher zone. In this case, stealing of
pageblocks or orders larger than a pageblock are still allowed in the fast
path as they are uninteresting from a fragmentation point of view.
This was evaluated using a benchmark designed to fragment memory before
attempting THP allocations. It's implemented in mmtests as the following
configurations
configs/config-global-dhp__workload_thpfioscale
configs/config-global-dhp__workload_thpfioscale-defrag
configs/config-global-dhp__workload_thpfioscale-madvhugepage
e.g. from mmtests
./run-mmtests.sh --run-monitor --config configs/config-global-dhp__workload_thpfioscale test-run-1
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch).
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameter create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed.
3. Warm up a number of fio read-only processes accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll refault old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds.
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup the test files.
Note that due to the use of IO and page cache that this benchmark is not
suitable for running on large machines where the time to fragment memory
may be excessive. Also note that while this is one mix that generates
fragmentation that it's not the only mix that generates fragmentation.
Differences in workload that are more slab-intensive or whether SLUB is
used with high-order pages may yield different results.
When the page allocator fragments memory, it records the event using the
mm_page_alloc_extfrag ftrace event. If the fallback_order is smaller than
a pageblock order (order-9 on 64-bit x86) then it's considered to be an
"external fragmentation event" that may cause issues in the future.
Hence, the primary metric here is the number of external fragmentation
events that occur with order < 9. The secondary metric is allocation
latency and huge page allocation success rates but note that differences
in latencies and what the success rate also can affect the number of
external fragmentation event which is why it's a secondary metric.
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 662.92 ( 0.00%) 653.58 * 1.41%*
Amean fault-huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
Fault latencies are slightly reduced while allocation success rates remain
at zero as this configuration does not make any special effort to allocate
THP and fio is heavily active at the time and either filling memory or
keeping pages resident. However, a 49% reduction of serious fragmentation
events reduces the changes of external fragmentation being a problem in
the future.
Vlastimil asked during review for a breakdown of the allocation types
that are falling back.
vanilla
3816 MIGRATE_UNMOVABLE
800845 MIGRATE_MOVABLE
33 MIGRATE_UNRECLAIMABLE
patch
735 MIGRATE_UNMOVABLE
408135 MIGRATE_MOVABLE
42 MIGRATE_UNRECLAIMABLE
The majority of the fallbacks are due to movable allocations and this is
consistent for the workload throughout the series so will not be presented
again as the primary source of fallbacks are movable allocations.
Movable fallbacks are sometimes considered "ok" to fallback because they
can be migrated. The problem is that they can fill an
unmovable/reclaimable pageblock causing those allocations to fallback
later and polluting pageblocks with pages that cannot move. If there is a
movable fallback, it is pretty much guaranteed to affect an
unmovable/reclaimable pageblock and while it might not be enough to
actually cause a unmovable/reclaimable fallback in the future, we cannot
know that in advance so the patch takes the only option available to it.
Hence, it's important to control them. This point is also consistent
throughout the series and will not be repeated.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 1495.14 ( 0.00%) 1467.55 ( 1.85%)
Amean fault-huge-1 1098.48 ( 0.00%) 1127.11 ( -2.61%)
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 78.57 ( 0.00%) 77.64 ( -1.18%)
Fragmentation events were reduced quite a bit although this is known
to be a little variable. The latencies and allocation success rates
are similar but they were already quite high.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 1350.05 ( 0.00%) 1346.45 ( 0.27%)
Amean fault-huge-5 4181.01 ( 0.00%) 3418.60 ( 18.24%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 1.15 ( 0.00%) 0.78 ( -31.88%)
The reduction of external fragmentation events is slight and this is
partially due to the removal of __GFP_THISNODE in commit ac5b2c18911f
("mm: thp: relax __GFP_THISNODE for MADV_HUGEPAGE mappings") as THP
allocations can now spill over to remote nodes instead of fragmenting
local memory.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 6138.97 ( 0.00%) 6217.43 ( -1.28%)
Amean fault-huge-5 2294.28 ( 0.00%) 3163.33 * -37.88%*
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 96.82 ( 0.00%) 95.14 ( -1.74%)
There was a slight reduction in external fragmentation events although the
latencies were higher. The allocation success rate is high enough that
the system is struggling and there is quite a lot of parallel reclaim and
compaction activity. There is also a certain degree of luck on whether
processes start on node 0 or not for this patch but the relevance is
reduced later in the series.
Overall, the patch reduces the number of external fragmentation causing
events so the success of THP over long periods of time would be improved
for this adverse workload.
Link: http://lkml.kernel.org/r/20181123114528.28802-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:41 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-04-25 22:24:01 -07:00
|
|
|
|
mm, page_alloc: spread allocations across zones before introducing fragmentation
Patch series "Fragmentation avoidance improvements", v5.
It has been noted before that fragmentation avoidance (aka
anti-fragmentation) is not perfect. Given sufficient time or an adverse
workload, memory gets fragmented and the long-term success of high-order
allocations degrades. This series defines an adverse workload, a definition
of external fragmentation events (including serious) ones and a series
that reduces the level of those fragmentation events.
The details of the workload and the consequences are described in more
detail in the changelogs. However, from patch 1, this is a high-level
summary of the adverse workload. The exact details are found in the
mmtests implementation.
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch)
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameterr create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed
3. Warm up a number of fio read-only threads accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll fault back in old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup
Overall the series reduces external fragmentation causing events by over 94%
on 1 and 2 socket machines, which in turn impacts high-order allocation
success rates over the long term. There are differences in latencies and
high-order allocation success rates. Latencies are a mixed bag as they
are vulnerable to exact system state and whether allocations succeeded
so they are treated as a secondary metric.
Patch 1 uses lower zones if they are populated and have free memory
instead of fragmenting a higher zone. It's special cased to
handle a Normal->DMA32 fallback with the reasons explained
in the changelog.
Patch 2-4 boosts watermarks temporarily when an external fragmentation
event occurs. kswapd wakes to reclaim a small amount of old memory
and then wakes kcompactd on completion to recover the system
slightly. This introduces some overhead in the slowpath. The level
of boosting can be tuned or disabled depending on the tolerance
for fragmentation vs allocation latency.
Patch 5 stalls some movable allocation requests to let kswapd from patch 4
make some progress. The duration of the stalls is very low but it
is possible to tune the system to avoid fragmentation events if
larger stalls can be tolerated.
The bulk of the improvement in fragmentation avoidance is from patches
1-4 but patch 5 can deal with a rare corner case and provides the option
of tuning a system for THP allocation success rates in exchange for
some stalls to control fragmentation.
This patch (of 5):
The page allocator zone lists are iterated based on the watermarks of each
zone which does not take anti-fragmentation into account. On x86, node 0
may have multiple zones while other nodes have one zone. A consequence is
that tasks running on node 0 may fragment ZONE_NORMAL even though
ZONE_DMA32 has plenty of free memory. This patch special cases the
allocator fast path such that it'll try an allocation from a lower local
zone before fragmenting a higher zone. In this case, stealing of
pageblocks or orders larger than a pageblock are still allowed in the fast
path as they are uninteresting from a fragmentation point of view.
This was evaluated using a benchmark designed to fragment memory before
attempting THP allocations. It's implemented in mmtests as the following
configurations
configs/config-global-dhp__workload_thpfioscale
configs/config-global-dhp__workload_thpfioscale-defrag
configs/config-global-dhp__workload_thpfioscale-madvhugepage
e.g. from mmtests
./run-mmtests.sh --run-monitor --config configs/config-global-dhp__workload_thpfioscale test-run-1
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch).
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameter create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed.
3. Warm up a number of fio read-only processes accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll refault old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds.
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup the test files.
Note that due to the use of IO and page cache that this benchmark is not
suitable for running on large machines where the time to fragment memory
may be excessive. Also note that while this is one mix that generates
fragmentation that it's not the only mix that generates fragmentation.
Differences in workload that are more slab-intensive or whether SLUB is
used with high-order pages may yield different results.
When the page allocator fragments memory, it records the event using the
mm_page_alloc_extfrag ftrace event. If the fallback_order is smaller than
a pageblock order (order-9 on 64-bit x86) then it's considered to be an
"external fragmentation event" that may cause issues in the future.
Hence, the primary metric here is the number of external fragmentation
events that occur with order < 9. The secondary metric is allocation
latency and huge page allocation success rates but note that differences
in latencies and what the success rate also can affect the number of
external fragmentation event which is why it's a secondary metric.
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 662.92 ( 0.00%) 653.58 * 1.41%*
Amean fault-huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
Fault latencies are slightly reduced while allocation success rates remain
at zero as this configuration does not make any special effort to allocate
THP and fio is heavily active at the time and either filling memory or
keeping pages resident. However, a 49% reduction of serious fragmentation
events reduces the changes of external fragmentation being a problem in
the future.
Vlastimil asked during review for a breakdown of the allocation types
that are falling back.
vanilla
3816 MIGRATE_UNMOVABLE
800845 MIGRATE_MOVABLE
33 MIGRATE_UNRECLAIMABLE
patch
735 MIGRATE_UNMOVABLE
408135 MIGRATE_MOVABLE
42 MIGRATE_UNRECLAIMABLE
The majority of the fallbacks are due to movable allocations and this is
consistent for the workload throughout the series so will not be presented
again as the primary source of fallbacks are movable allocations.
Movable fallbacks are sometimes considered "ok" to fallback because they
can be migrated. The problem is that they can fill an
unmovable/reclaimable pageblock causing those allocations to fallback
later and polluting pageblocks with pages that cannot move. If there is a
movable fallback, it is pretty much guaranteed to affect an
unmovable/reclaimable pageblock and while it might not be enough to
actually cause a unmovable/reclaimable fallback in the future, we cannot
know that in advance so the patch takes the only option available to it.
Hence, it's important to control them. This point is also consistent
throughout the series and will not be repeated.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 1495.14 ( 0.00%) 1467.55 ( 1.85%)
Amean fault-huge-1 1098.48 ( 0.00%) 1127.11 ( -2.61%)
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 78.57 ( 0.00%) 77.64 ( -1.18%)
Fragmentation events were reduced quite a bit although this is known
to be a little variable. The latencies and allocation success rates
are similar but they were already quite high.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 1350.05 ( 0.00%) 1346.45 ( 0.27%)
Amean fault-huge-5 4181.01 ( 0.00%) 3418.60 ( 18.24%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 1.15 ( 0.00%) 0.78 ( -31.88%)
The reduction of external fragmentation events is slight and this is
partially due to the removal of __GFP_THISNODE in commit ac5b2c18911f
("mm: thp: relax __GFP_THISNODE for MADV_HUGEPAGE mappings") as THP
allocations can now spill over to remote nodes instead of fragmenting
local memory.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 6138.97 ( 0.00%) 6217.43 ( -1.28%)
Amean fault-huge-5 2294.28 ( 0.00%) 3163.33 * -37.88%*
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 96.82 ( 0.00%) 95.14 ( -1.74%)
There was a slight reduction in external fragmentation events although the
latencies were higher. The allocation success rate is high enough that
the system is struggling and there is quite a lot of parallel reclaim and
compaction activity. There is also a certain degree of luck on whether
processes start on node 0 or not for this patch but the relevance is
reduced later in the series.
Overall, the patch reduces the number of external fragmentation causing
events so the success of THP over long periods of time would be improved
for this adverse workload.
Link: http://lkml.kernel.org/r/20181123114528.28802-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:41 -08:00
|
|
|
|
2019-04-25 22:24:01 -07:00
|
|
|
|
2018-12-28 00:35:48 -08:00
|
|
|
|
|
|
|
|
|
mm, page_alloc: spread allocations across zones before introducing fragmentation
Patch series "Fragmentation avoidance improvements", v5.
It has been noted before that fragmentation avoidance (aka
anti-fragmentation) is not perfect. Given sufficient time or an adverse
workload, memory gets fragmented and the long-term success of high-order
allocations degrades. This series defines an adverse workload, a definition
of external fragmentation events (including serious) ones and a series
that reduces the level of those fragmentation events.
The details of the workload and the consequences are described in more
detail in the changelogs. However, from patch 1, this is a high-level
summary of the adverse workload. The exact details are found in the
mmtests implementation.
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch)
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameterr create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed
3. Warm up a number of fio read-only threads accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll fault back in old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup
Overall the series reduces external fragmentation causing events by over 94%
on 1 and 2 socket machines, which in turn impacts high-order allocation
success rates over the long term. There are differences in latencies and
high-order allocation success rates. Latencies are a mixed bag as they
are vulnerable to exact system state and whether allocations succeeded
so they are treated as a secondary metric.
Patch 1 uses lower zones if they are populated and have free memory
instead of fragmenting a higher zone. It's special cased to
handle a Normal->DMA32 fallback with the reasons explained
in the changelog.
Patch 2-4 boosts watermarks temporarily when an external fragmentation
event occurs. kswapd wakes to reclaim a small amount of old memory
and then wakes kcompactd on completion to recover the system
slightly. This introduces some overhead in the slowpath. The level
of boosting can be tuned or disabled depending on the tolerance
for fragmentation vs allocation latency.
Patch 5 stalls some movable allocation requests to let kswapd from patch 4
make some progress. The duration of the stalls is very low but it
is possible to tune the system to avoid fragmentation events if
larger stalls can be tolerated.
The bulk of the improvement in fragmentation avoidance is from patches
1-4 but patch 5 can deal with a rare corner case and provides the option
of tuning a system for THP allocation success rates in exchange for
some stalls to control fragmentation.
This patch (of 5):
The page allocator zone lists are iterated based on the watermarks of each
zone which does not take anti-fragmentation into account. On x86, node 0
may have multiple zones while other nodes have one zone. A consequence is
that tasks running on node 0 may fragment ZONE_NORMAL even though
ZONE_DMA32 has plenty of free memory. This patch special cases the
allocator fast path such that it'll try an allocation from a lower local
zone before fragmenting a higher zone. In this case, stealing of
pageblocks or orders larger than a pageblock are still allowed in the fast
path as they are uninteresting from a fragmentation point of view.
This was evaluated using a benchmark designed to fragment memory before
attempting THP allocations. It's implemented in mmtests as the following
configurations
configs/config-global-dhp__workload_thpfioscale
configs/config-global-dhp__workload_thpfioscale-defrag
configs/config-global-dhp__workload_thpfioscale-madvhugepage
e.g. from mmtests
./run-mmtests.sh --run-monitor --config configs/config-global-dhp__workload_thpfioscale test-run-1
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch).
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameter create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed.
3. Warm up a number of fio read-only processes accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll refault old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds.
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup the test files.
Note that due to the use of IO and page cache that this benchmark is not
suitable for running on large machines where the time to fragment memory
may be excessive. Also note that while this is one mix that generates
fragmentation that it's not the only mix that generates fragmentation.
Differences in workload that are more slab-intensive or whether SLUB is
used with high-order pages may yield different results.
When the page allocator fragments memory, it records the event using the
mm_page_alloc_extfrag ftrace event. If the fallback_order is smaller than
a pageblock order (order-9 on 64-bit x86) then it's considered to be an
"external fragmentation event" that may cause issues in the future.
Hence, the primary metric here is the number of external fragmentation
events that occur with order < 9. The secondary metric is allocation
latency and huge page allocation success rates but note that differences
in latencies and what the success rate also can affect the number of
external fragmentation event which is why it's a secondary metric.
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 662.92 ( 0.00%) 653.58 * 1.41%*
Amean fault-huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
Fault latencies are slightly reduced while allocation success rates remain
at zero as this configuration does not make any special effort to allocate
THP and fio is heavily active at the time and either filling memory or
keeping pages resident. However, a 49% reduction of serious fragmentation
events reduces the changes of external fragmentation being a problem in
the future.
Vlastimil asked during review for a breakdown of the allocation types
that are falling back.
vanilla
3816 MIGRATE_UNMOVABLE
800845 MIGRATE_MOVABLE
33 MIGRATE_UNRECLAIMABLE
patch
735 MIGRATE_UNMOVABLE
408135 MIGRATE_MOVABLE
42 MIGRATE_UNRECLAIMABLE
The majority of the fallbacks are due to movable allocations and this is
consistent for the workload throughout the series so will not be presented
again as the primary source of fallbacks are movable allocations.
Movable fallbacks are sometimes considered "ok" to fallback because they
can be migrated. The problem is that they can fill an
unmovable/reclaimable pageblock causing those allocations to fallback
later and polluting pageblocks with pages that cannot move. If there is a
movable fallback, it is pretty much guaranteed to affect an
unmovable/reclaimable pageblock and while it might not be enough to
actually cause a unmovable/reclaimable fallback in the future, we cannot
know that in advance so the patch takes the only option available to it.
Hence, it's important to control them. This point is also consistent
throughout the series and will not be repeated.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 1495.14 ( 0.00%) 1467.55 ( 1.85%)
Amean fault-huge-1 1098.48 ( 0.00%) 1127.11 ( -2.61%)
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 78.57 ( 0.00%) 77.64 ( -1.18%)
Fragmentation events were reduced quite a bit although this is known
to be a little variable. The latencies and allocation success rates
are similar but they were already quite high.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 1350.05 ( 0.00%) 1346.45 ( 0.27%)
Amean fault-huge-5 4181.01 ( 0.00%) 3418.60 ( 18.24%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 1.15 ( 0.00%) 0.78 ( -31.88%)
The reduction of external fragmentation events is slight and this is
partially due to the removal of __GFP_THISNODE in commit ac5b2c18911f
("mm: thp: relax __GFP_THISNODE for MADV_HUGEPAGE mappings") as THP
allocations can now spill over to remote nodes instead of fragmenting
local memory.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 6138.97 ( 0.00%) 6217.43 ( -1.28%)
Amean fault-huge-5 2294.28 ( 0.00%) 3163.33 * -37.88%*
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 96.82 ( 0.00%) 95.14 ( -1.74%)
There was a slight reduction in external fragmentation events although the
latencies were higher. The allocation success rate is high enough that
the system is struggling and there is quite a lot of parallel reclaim and
compaction activity. There is also a certain degree of luck on whether
processes start on node 0 or not for this patch but the relevance is
reduced later in the series.
Overall, the patch reduces the number of external fragmentation causing
events so the success of THP over long periods of time would be improved
for this adverse workload.
Link: http://lkml.kernel.org/r/20181123114528.28802-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:41 -08:00
|
|
|
|
|
|
|
|
|
2021-05-04 18:39:00 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-08-06 23:26:04 -07:00
|
|
|
|
|
|
|
|
|
2021-05-04 18:39:00 -07:00
|
|
|
|
2020-08-06 23:26:04 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-11-13 16:06:43 -08:00
|
|
|
|
2006-12-06 20:31:38 -08:00
|
|
|
|
2005-11-13 16:06:43 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-02-11 15:25:41 -08:00
|
|
|
|
|
|
|
|
|
2005-06-21 17:14:41 -07:00
|
|
|
|
mm, page_alloc: spread allocations across zones before introducing fragmentation
Patch series "Fragmentation avoidance improvements", v5.
It has been noted before that fragmentation avoidance (aka
anti-fragmentation) is not perfect. Given sufficient time or an adverse
workload, memory gets fragmented and the long-term success of high-order
allocations degrades. This series defines an adverse workload, a definition
of external fragmentation events (including serious) ones and a series
that reduces the level of those fragmentation events.
The details of the workload and the consequences are described in more
detail in the changelogs. However, from patch 1, this is a high-level
summary of the adverse workload. The exact details are found in the
mmtests implementation.
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch)
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameterr create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed
3. Warm up a number of fio read-only threads accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll fault back in old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup
Overall the series reduces external fragmentation causing events by over 94%
on 1 and 2 socket machines, which in turn impacts high-order allocation
success rates over the long term. There are differences in latencies and
high-order allocation success rates. Latencies are a mixed bag as they
are vulnerable to exact system state and whether allocations succeeded
so they are treated as a secondary metric.
Patch 1 uses lower zones if they are populated and have free memory
instead of fragmenting a higher zone. It's special cased to
handle a Normal->DMA32 fallback with the reasons explained
in the changelog.
Patch 2-4 boosts watermarks temporarily when an external fragmentation
event occurs. kswapd wakes to reclaim a small amount of old memory
and then wakes kcompactd on completion to recover the system
slightly. This introduces some overhead in the slowpath. The level
of boosting can be tuned or disabled depending on the tolerance
for fragmentation vs allocation latency.
Patch 5 stalls some movable allocation requests to let kswapd from patch 4
make some progress. The duration of the stalls is very low but it
is possible to tune the system to avoid fragmentation events if
larger stalls can be tolerated.
The bulk of the improvement in fragmentation avoidance is from patches
1-4 but patch 5 can deal with a rare corner case and provides the option
of tuning a system for THP allocation success rates in exchange for
some stalls to control fragmentation.
This patch (of 5):
The page allocator zone lists are iterated based on the watermarks of each
zone which does not take anti-fragmentation into account. On x86, node 0
may have multiple zones while other nodes have one zone. A consequence is
that tasks running on node 0 may fragment ZONE_NORMAL even though
ZONE_DMA32 has plenty of free memory. This patch special cases the
allocator fast path such that it'll try an allocation from a lower local
zone before fragmenting a higher zone. In this case, stealing of
pageblocks or orders larger than a pageblock are still allowed in the fast
path as they are uninteresting from a fragmentation point of view.
This was evaluated using a benchmark designed to fragment memory before
attempting THP allocations. It's implemented in mmtests as the following
configurations
configs/config-global-dhp__workload_thpfioscale
configs/config-global-dhp__workload_thpfioscale-defrag
configs/config-global-dhp__workload_thpfioscale-madvhugepage
e.g. from mmtests
./run-mmtests.sh --run-monitor --config configs/config-global-dhp__workload_thpfioscale test-run-1
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch).
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameter create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed.
3. Warm up a number of fio read-only processes accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll refault old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds.
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup the test files.
Note that due to the use of IO and page cache that this benchmark is not
suitable for running on large machines where the time to fragment memory
may be excessive. Also note that while this is one mix that generates
fragmentation that it's not the only mix that generates fragmentation.
Differences in workload that are more slab-intensive or whether SLUB is
used with high-order pages may yield different results.
When the page allocator fragments memory, it records the event using the
mm_page_alloc_extfrag ftrace event. If the fallback_order is smaller than
a pageblock order (order-9 on 64-bit x86) then it's considered to be an
"external fragmentation event" that may cause issues in the future.
Hence, the primary metric here is the number of external fragmentation
events that occur with order < 9. The secondary metric is allocation
latency and huge page allocation success rates but note that differences
in latencies and what the success rate also can affect the number of
external fragmentation event which is why it's a secondary metric.
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 662.92 ( 0.00%) 653.58 * 1.41%*
Amean fault-huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
Fault latencies are slightly reduced while allocation success rates remain
at zero as this configuration does not make any special effort to allocate
THP and fio is heavily active at the time and either filling memory or
keeping pages resident. However, a 49% reduction of serious fragmentation
events reduces the changes of external fragmentation being a problem in
the future.
Vlastimil asked during review for a breakdown of the allocation types
that are falling back.
vanilla
3816 MIGRATE_UNMOVABLE
800845 MIGRATE_MOVABLE
33 MIGRATE_UNRECLAIMABLE
patch
735 MIGRATE_UNMOVABLE
408135 MIGRATE_MOVABLE
42 MIGRATE_UNRECLAIMABLE
The majority of the fallbacks are due to movable allocations and this is
consistent for the workload throughout the series so will not be presented
again as the primary source of fallbacks are movable allocations.
Movable fallbacks are sometimes considered "ok" to fallback because they
can be migrated. The problem is that they can fill an
unmovable/reclaimable pageblock causing those allocations to fallback
later and polluting pageblocks with pages that cannot move. If there is a
movable fallback, it is pretty much guaranteed to affect an
unmovable/reclaimable pageblock and while it might not be enough to
actually cause a unmovable/reclaimable fallback in the future, we cannot
know that in advance so the patch takes the only option available to it.
Hence, it's important to control them. This point is also consistent
throughout the series and will not be repeated.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 1495.14 ( 0.00%) 1467.55 ( 1.85%)
Amean fault-huge-1 1098.48 ( 0.00%) 1127.11 ( -2.61%)
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 78.57 ( 0.00%) 77.64 ( -1.18%)
Fragmentation events were reduced quite a bit although this is known
to be a little variable. The latencies and allocation success rates
are similar but they were already quite high.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 1350.05 ( 0.00%) 1346.45 ( 0.27%)
Amean fault-huge-5 4181.01 ( 0.00%) 3418.60 ( 18.24%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 1.15 ( 0.00%) 0.78 ( -31.88%)
The reduction of external fragmentation events is slight and this is
partially due to the removal of __GFP_THISNODE in commit ac5b2c18911f
("mm: thp: relax __GFP_THISNODE for MADV_HUGEPAGE mappings") as THP
allocations can now spill over to remote nodes instead of fragmenting
local memory.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 6138.97 ( 0.00%) 6217.43 ( -1.28%)
Amean fault-huge-5 2294.28 ( 0.00%) 3163.33 * -37.88%*
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 96.82 ( 0.00%) 95.14 ( -1.74%)
There was a slight reduction in external fragmentation events although the
latencies were higher. The allocation success rate is high enough that
the system is struggling and there is quite a lot of parallel reclaim and
compaction activity. There is also a certain degree of luck on whether
processes start on node 0 or not for this patch but the relevance is
reduced later in the series.
Overall, the patch reduces the number of external fragmentation causing
events so the success of THP over long periods of time would be improved
for this adverse workload.
Link: http://lkml.kernel.org/r/20181123114528.28802-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:41 -08:00
|
|
|
|
2009-06-16 15:31:59 -07:00
|
|
|
|
2022-05-12 20:22:51 -07:00
|
|
|
|
|
|
|
|
|
mm, page_alloc: spread allocations across zones before introducing fragmentation
Patch series "Fragmentation avoidance improvements", v5.
It has been noted before that fragmentation avoidance (aka
anti-fragmentation) is not perfect. Given sufficient time or an adverse
workload, memory gets fragmented and the long-term success of high-order
allocations degrades. This series defines an adverse workload, a definition
of external fragmentation events (including serious) ones and a series
that reduces the level of those fragmentation events.
The details of the workload and the consequences are described in more
detail in the changelogs. However, from patch 1, this is a high-level
summary of the adverse workload. The exact details are found in the
mmtests implementation.
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch)
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameterr create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed
3. Warm up a number of fio read-only threads accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll fault back in old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup
Overall the series reduces external fragmentation causing events by over 94%
on 1 and 2 socket machines, which in turn impacts high-order allocation
success rates over the long term. There are differences in latencies and
high-order allocation success rates. Latencies are a mixed bag as they
are vulnerable to exact system state and whether allocations succeeded
so they are treated as a secondary metric.
Patch 1 uses lower zones if they are populated and have free memory
instead of fragmenting a higher zone. It's special cased to
handle a Normal->DMA32 fallback with the reasons explained
in the changelog.
Patch 2-4 boosts watermarks temporarily when an external fragmentation
event occurs. kswapd wakes to reclaim a small amount of old memory
and then wakes kcompactd on completion to recover the system
slightly. This introduces some overhead in the slowpath. The level
of boosting can be tuned or disabled depending on the tolerance
for fragmentation vs allocation latency.
Patch 5 stalls some movable allocation requests to let kswapd from patch 4
make some progress. The duration of the stalls is very low but it
is possible to tune the system to avoid fragmentation events if
larger stalls can be tolerated.
The bulk of the improvement in fragmentation avoidance is from patches
1-4 but patch 5 can deal with a rare corner case and provides the option
of tuning a system for THP allocation success rates in exchange for
some stalls to control fragmentation.
This patch (of 5):
The page allocator zone lists are iterated based on the watermarks of each
zone which does not take anti-fragmentation into account. On x86, node 0
may have multiple zones while other nodes have one zone. A consequence is
that tasks running on node 0 may fragment ZONE_NORMAL even though
ZONE_DMA32 has plenty of free memory. This patch special cases the
allocator fast path such that it'll try an allocation from a lower local
zone before fragmenting a higher zone. In this case, stealing of
pageblocks or orders larger than a pageblock are still allowed in the fast
path as they are uninteresting from a fragmentation point of view.
This was evaluated using a benchmark designed to fragment memory before
attempting THP allocations. It's implemented in mmtests as the following
configurations
configs/config-global-dhp__workload_thpfioscale
configs/config-global-dhp__workload_thpfioscale-defrag
configs/config-global-dhp__workload_thpfioscale-madvhugepage
e.g. from mmtests
./run-mmtests.sh --run-monitor --config configs/config-global-dhp__workload_thpfioscale test-run-1
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch).
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameter create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed.
3. Warm up a number of fio read-only processes accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll refault old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds.
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup the test files.
Note that due to the use of IO and page cache that this benchmark is not
suitable for running on large machines where the time to fragment memory
may be excessive. Also note that while this is one mix that generates
fragmentation that it's not the only mix that generates fragmentation.
Differences in workload that are more slab-intensive or whether SLUB is
used with high-order pages may yield different results.
When the page allocator fragments memory, it records the event using the
mm_page_alloc_extfrag ftrace event. If the fallback_order is smaller than
a pageblock order (order-9 on 64-bit x86) then it's considered to be an
"external fragmentation event" that may cause issues in the future.
Hence, the primary metric here is the number of external fragmentation
events that occur with order < 9. The secondary metric is allocation
latency and huge page allocation success rates but note that differences
in latencies and what the success rate also can affect the number of
external fragmentation event which is why it's a secondary metric.
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 662.92 ( 0.00%) 653.58 * 1.41%*
Amean fault-huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
Fault latencies are slightly reduced while allocation success rates remain
at zero as this configuration does not make any special effort to allocate
THP and fio is heavily active at the time and either filling memory or
keeping pages resident. However, a 49% reduction of serious fragmentation
events reduces the changes of external fragmentation being a problem in
the future.
Vlastimil asked during review for a breakdown of the allocation types
that are falling back.
vanilla
3816 MIGRATE_UNMOVABLE
800845 MIGRATE_MOVABLE
33 MIGRATE_UNRECLAIMABLE
patch
735 MIGRATE_UNMOVABLE
408135 MIGRATE_MOVABLE
42 MIGRATE_UNRECLAIMABLE
The majority of the fallbacks are due to movable allocations and this is
consistent for the workload throughout the series so will not be presented
again as the primary source of fallbacks are movable allocations.
Movable fallbacks are sometimes considered "ok" to fallback because they
can be migrated. The problem is that they can fill an
unmovable/reclaimable pageblock causing those allocations to fallback
later and polluting pageblocks with pages that cannot move. If there is a
movable fallback, it is pretty much guaranteed to affect an
unmovable/reclaimable pageblock and while it might not be enough to
actually cause a unmovable/reclaimable fallback in the future, we cannot
know that in advance so the patch takes the only option available to it.
Hence, it's important to control them. This point is also consistent
throughout the series and will not be repeated.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 1495.14 ( 0.00%) 1467.55 ( 1.85%)
Amean fault-huge-1 1098.48 ( 0.00%) 1127.11 ( -2.61%)
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 78.57 ( 0.00%) 77.64 ( -1.18%)
Fragmentation events were reduced quite a bit although this is known
to be a little variable. The latencies and allocation success rates
are similar but they were already quite high.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 1350.05 ( 0.00%) 1346.45 ( 0.27%)
Amean fault-huge-5 4181.01 ( 0.00%) 3418.60 ( 18.24%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 1.15 ( 0.00%) 0.78 ( -31.88%)
The reduction of external fragmentation events is slight and this is
partially due to the removal of __GFP_THISNODE in commit ac5b2c18911f
("mm: thp: relax __GFP_THISNODE for MADV_HUGEPAGE mappings") as THP
allocations can now spill over to remote nodes instead of fragmenting
local memory.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 6138.97 ( 0.00%) 6217.43 ( -1.28%)
Amean fault-huge-5 2294.28 ( 0.00%) 3163.33 * -37.88%*
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 96.82 ( 0.00%) 95.14 ( -1.74%)
There was a slight reduction in external fragmentation events although the
latencies were higher. The allocation success rate is high enough that
the system is struggling and there is quite a lot of parallel reclaim and
compaction activity. There is also a certain degree of luck on whether
processes start on node 0 or not for this patch but the relevance is
reduced later in the series.
Overall, the patch reduces the number of external fragmentation causing
events so the success of THP over long periods of time would be improved
for this adverse workload.
Link: http://lkml.kernel.org/r/20181123114528.28802-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:41 -08:00
|
|
|
|
mm: page_alloc: avoid kswapd thrashing due to NUMA restrictions
On NUMA systems without bindings, allocations check all nodes for free
space, then wake up the kswapds on all nodes and retry. This ensures
all available space is evenly used before reclaim begins. However,
when one process or certain allocations have node restrictions, they
can cause kswapds on only a subset of nodes to be woken up.
Since kswapd hysteresis targets watermarks that are *higher* than
needed for allocation, even *unrestricted* allocations can now get
suckered onto such nodes that are already pressured. This ends up
concentrating all allocations on them, even when there are idle nodes
available for the unrestricted requests.
This was observed with two numa nodes, where node0 is normal and node1
is ZONE_MOVABLE to facilitate hotplugging: a kernel allocation wakes
kswapd on node0 only (since node1 is not eligible); once kswapd0 is
active, the watermarks hover between low and high, and then even the
movable allocations end up on node0, only to be kicked out again;
meanwhile node1 is empty and idle.
Similar behavior is possible when a process with NUMA bindings is
causing selective kswapd wakeups.
To fix this, on NUMA systems augment the (misleading) watermark test
with a check for whether kswapd is already active during the first
iteration through the zonelist. If this fails to place the request,
kswapd must be running everywhere already, and the watermark test is
good enough to decide placement.
With this patch, unrestricted requests successfully make use of node1,
even while kswapd is reclaiming node0 for restricted allocations.
[gourry@gourry.net: don't retry if no kswapds were active]
Link: https://lkml.kernel.org/r/20250919162134.1098208-1-hannes@cmpxchg.org
Signed-off-by: Gregory Price <gourry@gourry.net>
Tested-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Zi Yan <ziy@nvidia.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: Joshua Hahn <joshua.hahnjy@gmail.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-09-19 12:21:34 -04:00
|
|
|
|
|
|
|
|
|
2016-07-28 15:46:53 -07:00
|
|
|
|
mm, page_alloc: spread allocations across zones before introducing fragmentation
Patch series "Fragmentation avoidance improvements", v5.
It has been noted before that fragmentation avoidance (aka
anti-fragmentation) is not perfect. Given sufficient time or an adverse
workload, memory gets fragmented and the long-term success of high-order
allocations degrades. This series defines an adverse workload, a definition
of external fragmentation events (including serious) ones and a series
that reduces the level of those fragmentation events.
The details of the workload and the consequences are described in more
detail in the changelogs. However, from patch 1, this is a high-level
summary of the adverse workload. The exact details are found in the
mmtests implementation.
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch)
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameterr create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed
3. Warm up a number of fio read-only threads accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll fault back in old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup
Overall the series reduces external fragmentation causing events by over 94%
on 1 and 2 socket machines, which in turn impacts high-order allocation
success rates over the long term. There are differences in latencies and
high-order allocation success rates. Latencies are a mixed bag as they
are vulnerable to exact system state and whether allocations succeeded
so they are treated as a secondary metric.
Patch 1 uses lower zones if they are populated and have free memory
instead of fragmenting a higher zone. It's special cased to
handle a Normal->DMA32 fallback with the reasons explained
in the changelog.
Patch 2-4 boosts watermarks temporarily when an external fragmentation
event occurs. kswapd wakes to reclaim a small amount of old memory
and then wakes kcompactd on completion to recover the system
slightly. This introduces some overhead in the slowpath. The level
of boosting can be tuned or disabled depending on the tolerance
for fragmentation vs allocation latency.
Patch 5 stalls some movable allocation requests to let kswapd from patch 4
make some progress. The duration of the stalls is very low but it
is possible to tune the system to avoid fragmentation events if
larger stalls can be tolerated.
The bulk of the improvement in fragmentation avoidance is from patches
1-4 but patch 5 can deal with a rare corner case and provides the option
of tuning a system for THP allocation success rates in exchange for
some stalls to control fragmentation.
This patch (of 5):
The page allocator zone lists are iterated based on the watermarks of each
zone which does not take anti-fragmentation into account. On x86, node 0
may have multiple zones while other nodes have one zone. A consequence is
that tasks running on node 0 may fragment ZONE_NORMAL even though
ZONE_DMA32 has plenty of free memory. This patch special cases the
allocator fast path such that it'll try an allocation from a lower local
zone before fragmenting a higher zone. In this case, stealing of
pageblocks or orders larger than a pageblock are still allowed in the fast
path as they are uninteresting from a fragmentation point of view.
This was evaluated using a benchmark designed to fragment memory before
attempting THP allocations. It's implemented in mmtests as the following
configurations
configs/config-global-dhp__workload_thpfioscale
configs/config-global-dhp__workload_thpfioscale-defrag
configs/config-global-dhp__workload_thpfioscale-madvhugepage
e.g. from mmtests
./run-mmtests.sh --run-monitor --config configs/config-global-dhp__workload_thpfioscale test-run-1
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch).
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameter create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed.
3. Warm up a number of fio read-only processes accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll refault old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds.
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup the test files.
Note that due to the use of IO and page cache that this benchmark is not
suitable for running on large machines where the time to fragment memory
may be excessive. Also note that while this is one mix that generates
fragmentation that it's not the only mix that generates fragmentation.
Differences in workload that are more slab-intensive or whether SLUB is
used with high-order pages may yield different results.
When the page allocator fragments memory, it records the event using the
mm_page_alloc_extfrag ftrace event. If the fallback_order is smaller than
a pageblock order (order-9 on 64-bit x86) then it's considered to be an
"external fragmentation event" that may cause issues in the future.
Hence, the primary metric here is the number of external fragmentation
events that occur with order < 9. The secondary metric is allocation
latency and huge page allocation success rates but note that differences
in latencies and what the success rate also can affect the number of
external fragmentation event which is why it's a secondary metric.
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 662.92 ( 0.00%) 653.58 * 1.41%*
Amean fault-huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
Fault latencies are slightly reduced while allocation success rates remain
at zero as this configuration does not make any special effort to allocate
THP and fio is heavily active at the time and either filling memory or
keeping pages resident. However, a 49% reduction of serious fragmentation
events reduces the changes of external fragmentation being a problem in
the future.
Vlastimil asked during review for a breakdown of the allocation types
that are falling back.
vanilla
3816 MIGRATE_UNMOVABLE
800845 MIGRATE_MOVABLE
33 MIGRATE_UNRECLAIMABLE
patch
735 MIGRATE_UNMOVABLE
408135 MIGRATE_MOVABLE
42 MIGRATE_UNRECLAIMABLE
The majority of the fallbacks are due to movable allocations and this is
consistent for the workload throughout the series so will not be presented
again as the primary source of fallbacks are movable allocations.
Movable fallbacks are sometimes considered "ok" to fallback because they
can be migrated. The problem is that they can fill an
unmovable/reclaimable pageblock causing those allocations to fallback
later and polluting pageblocks with pages that cannot move. If there is a
movable fallback, it is pretty much guaranteed to affect an
unmovable/reclaimable pageblock and while it might not be enough to
actually cause a unmovable/reclaimable fallback in the future, we cannot
know that in advance so the patch takes the only option available to it.
Hence, it's important to control them. This point is also consistent
throughout the series and will not be repeated.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 1495.14 ( 0.00%) 1467.55 ( 1.85%)
Amean fault-huge-1 1098.48 ( 0.00%) 1127.11 ( -2.61%)
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 78.57 ( 0.00%) 77.64 ( -1.18%)
Fragmentation events were reduced quite a bit although this is known
to be a little variable. The latencies and allocation success rates
are similar but they were already quite high.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 1350.05 ( 0.00%) 1346.45 ( 0.27%)
Amean fault-huge-5 4181.01 ( 0.00%) 3418.60 ( 18.24%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 1.15 ( 0.00%) 0.78 ( -31.88%)
The reduction of external fragmentation events is slight and this is
partially due to the removal of __GFP_THISNODE in commit ac5b2c18911f
("mm: thp: relax __GFP_THISNODE for MADV_HUGEPAGE mappings") as THP
allocations can now spill over to remote nodes instead of fragmenting
local memory.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 6138.97 ( 0.00%) 6217.43 ( -1.28%)
Amean fault-huge-5 2294.28 ( 0.00%) 3163.33 * -37.88%*
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 96.82 ( 0.00%) 95.14 ( -1.74%)
There was a slight reduction in external fragmentation events although the
latencies were higher. The allocation success rate is high enough that
the system is struggling and there is quite a lot of parallel reclaim and
compaction activity. There is also a certain degree of luck on whether
processes start on node 0 or not for this patch but the relevance is
reduced later in the series.
Overall, the patch reduces the number of external fragmentation causing
events so the success of THP over long periods of time would be improved
for this adverse workload.
Link: http://lkml.kernel.org/r/20181123114528.28802-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:41 -08:00
|
|
|
|
2005-11-13 16:06:43 -08:00
|
|
|
|
[PATCH] memory page_alloc zonelist caching speedup
Optimize the critical zonelist scanning for free pages in the kernel memory
allocator by caching the zones that were found to be full recently, and
skipping them.
Remembers the zones in a zonelist that were short of free memory in the
last second. And it stashes a zone-to-node table in the zonelist struct,
to optimize that conversion (minimize its cache footprint.)
Recent changes:
This differs in a significant way from a similar patch that I
posted a week ago. Now, instead of having a nodemask_t of
recently full nodes, I have a bitmask of recently full zones.
This solves a problem that last weeks patch had, which on
systems with multiple zones per node (such as DMA zone) would
take seeing any of these zones full as meaning that all zones
on that node were full.
Also I changed names - from "zonelist faster" to "zonelist cache",
as that seemed to better convey what we're doing here - caching
some of the key zonelist state (for faster access.)
See below for some performance benchmark results. After all that
discussion with David on why I didn't need them, I went and got
some ;). I wanted to verify that I had not hurt the normal case
of memory allocation noticeably. At least for my one little
microbenchmark, I found (1) the normal case wasn't affected, and
(2) workloads that forced scanning across multiple nodes for
memory improved up to 10% fewer System CPU cycles and lower
elapsed clock time ('sys' and 'real'). Good. See details, below.
I didn't have the logic in get_page_from_freelist() for various
full nodes and zone reclaim failures correct. That should be
fixed up now - notice the new goto labels zonelist_scan,
this_zone_full, and try_next_zone, in get_page_from_freelist().
There are two reasons I persued this alternative, over some earlier
proposals that would have focused on optimizing the fake numa
emulation case by caching the last useful zone:
1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems)
have seen real customer loads where the cost to scan the zonelist
was a problem, due to many nodes being full of memory before
we got to a node we could use. Or at least, I think we have.
This was related to me by another engineer, based on experiences
from some time past. So this is not guaranteed. Most likely, though.
The following approach should help such real numa systems just as
much as it helps fake numa systems, or any combination thereof.
2) The effort to distinguish fake from real numa, using node_distance,
so that we could cache a fake numa node and optimize choosing
it over equivalent distance fake nodes, while continuing to
properly scan all real nodes in distance order, was going to
require a nasty blob of zonelist and node distance munging.
The following approach has no new dependency on node distances or
zone sorting.
See comment in the patch below for a description of what it actually does.
Technical details of note (or controversy):
- See the use of "zlc_active" and "did_zlc_setup" below, to delay
adding any work for this new mechanism until we've looked at the
first zone in zonelist. I figured the odds of the first zone
having the memory we needed were high enough that we should just
look there, first, then get fancy only if we need to keep looking.
- Some odd hackery was needed to add items to struct zonelist, while
not tripping up the custom zonelists built by the mm/mempolicy.c
code for MPOL_BIND. My usual wordy comments below explain this.
Search for "MPOL_BIND".
- Some per-node data in the struct zonelist is now modified frequently,
with no locking. Multiple CPU cores on a node could hit and mangle
this data. The theory is that this is just performance hint data,
and the memory allocator will work just fine despite any such mangling.
The fields at risk are the struct 'zonelist_cache' fields 'fullzones'
(a bitmask) and 'last_full_zap' (unsigned long jiffies). It should
all be self correcting after at most a one second delay.
- This still does a linear scan of the same lengths as before. All
I've optimized is making the scan faster, not algorithmically
shorter. It is now able to scan a compact array of 'unsigned
short' in the case of many full nodes, so one cache line should
cover quite a few nodes, rather than each node hitting another
one or two new and distinct cache lines.
- If both Andi and Nick don't find this too complicated, I will be
(pleasantly) flabbergasted.
- I removed the comment claiming we only use one cachline's worth of
zonelist. We seem, at least in the fake numa case, to have put the
lie to that claim.
- I pay no attention to the various watermarks and such in this performance
hint. A node could be marked full for one watermark, and then skipped
over when searching for a page using a different watermark. I think
that's actually quite ok, as it will tend to slightly increase the
spreading of memory over other nodes, away from a memory stressed node.
===============
Performance - some benchmark results and analysis:
This benchmark runs a memory hog program that uses multiple
threads to touch alot of memory as quickly as it can.
Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of
the total 96 GBytes on the system, and using 1, 19, 37, or 55
threads (on a 56 CPU system.) System, user and real (elapsed)
timings were recorded for each run, shown in units of seconds,
in the table below.
Two kernels were tested - 2.6.18-mm3 and the same kernel with
this zonelist caching patch added. The table also shows the
percentage improvement the zonelist caching sys time is over
(lower than) the stock *-mm kernel.
number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent
GBs N ------------ -------------- ---------------- systime
mem threads sys user real sys user real sys user real better
12 1 153 24 177 151 24 176 -2 0 -1 1%
12 19 99 22 8 99 22 8 0 0 0 0%
12 37 111 25 6 112 25 6 1 0 0 -0%
12 55 115 25 5 110 23 5 -5 -2 0 4%
38 1 502 74 576 497 73 570 -5 -1 -6 0%
38 19 426 78 48 373 76 39 -53 -2 -9 12%
38 37 544 83 36 547 82 36 3 -1 0 -0%
38 55 501 77 23 511 80 24 10 3 1 -1%
64 1 917 125 1042 890 124 1014 -27 -1 -28 2%
64 19 1118 138 119 965 141 103 -153 3 -16 13%
64 37 1202 151 94 1136 150 81 -66 -1 -13 5%
64 55 1118 141 61 1072 140 58 -46 -1 -3 4%
90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4%
90 19 2392 199 192 2116 189 176 -276 -10 -16 11%
90 37 3313 238 175 2972 225 145 -341 -13 -30 10%
90 55 1948 210 104 1843 213 100 -105 3 -4 5%
Notes:
1) This test ran a memory hog program that started a specified number N of
threads, and had each thread allocate and touch 1/N'th of
the total memory to be used in the test run in a single loop,
writing a constant word to memory, one store every 4096 bytes.
Watching this test during some earlier trial runs, I would see
each of these threads sit down on one CPU and stay there, for
the remainder of the pass, a different CPU for each thread.
2) The 'real' column is not comparable to the 'sys' or 'user' columns.
The 'real' column is seconds wall clock time elapsed, from beginning
to end of that test pass. The 'sys' and 'user' columns are total
CPU seconds spent on that test pass. For a 19 thread test run,
for example, the sum of 'sys' and 'user' could be up to 19 times the
number of 'real' elapsed wall clock seconds.
3) Tests were run on a fresh, single-user boot, to minimize the amount
of memory already in use at the start of the test, and to minimize
the amount of background activity that might interfere.
4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM.
5) Notice that the 'real' time gets large for the single thread runs, even
though the measured 'sys' and 'user' times are modest. I'm not sure what
that means - probably something to do with it being slow for one thread to
be accessing memory along ways away. Perhaps the fake numa system, running
ostensibly the same workload, would not show this substantial degradation
of 'real' time for one thread on many nodes -- lets hope not.
6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs)
ran quite efficiently, as one might expect. Each pair of threads needed
to allocate and touch the memory on the node the two threads shared, a
pleasantly parallizable workload.
7) The intermediate thread count passes, when asking for alot of memory forcing
them to go to a few neighboring nodes, improved the most with this zonelist
caching patch.
Conclusions:
* This zonelist cache patch probably makes little difference one way or the
other for most workloads on real numa hardware, if those workloads avoid
heavy off node allocations.
* For memory intensive workloads requiring substantial off-node allocations
on real numa hardware, this patch improves both kernel and elapsed timings
up to ten per-cent.
* For fake numa systems, I'm optimistic, but will have to leave that up to
Rohit Seth to actually test (once I get him a 2.6.18 backport.)
Signed-off-by: Paul Jackson <pj@sgi.com>
Cc: Rohit Seth <rohitseth@google.com>
Cc: Christoph Lameter <clameter@engr.sgi.com>
Cc: David Rientjes <rientjes@cs.washington.edu>
Cc: Paul Menage <menage@google.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-06 20:31:48 -08:00
|
|
|
|
2025-04-24 16:28:05 -04:00
|
|
|
|
2005-11-13 16:06:43 -08:00
|
|
|
|
mm, page_alloc: spread allocations across zones before introducing fragmentation
Patch series "Fragmentation avoidance improvements", v5.
It has been noted before that fragmentation avoidance (aka
anti-fragmentation) is not perfect. Given sufficient time or an adverse
workload, memory gets fragmented and the long-term success of high-order
allocations degrades. This series defines an adverse workload, a definition
of external fragmentation events (including serious) ones and a series
that reduces the level of those fragmentation events.
The details of the workload and the consequences are described in more
detail in the changelogs. However, from patch 1, this is a high-level
summary of the adverse workload. The exact details are found in the
mmtests implementation.
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch)
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameterr create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed
3. Warm up a number of fio read-only threads accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll fault back in old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup
Overall the series reduces external fragmentation causing events by over 94%
on 1 and 2 socket machines, which in turn impacts high-order allocation
success rates over the long term. There are differences in latencies and
high-order allocation success rates. Latencies are a mixed bag as they
are vulnerable to exact system state and whether allocations succeeded
so they are treated as a secondary metric.
Patch 1 uses lower zones if they are populated and have free memory
instead of fragmenting a higher zone. It's special cased to
handle a Normal->DMA32 fallback with the reasons explained
in the changelog.
Patch 2-4 boosts watermarks temporarily when an external fragmentation
event occurs. kswapd wakes to reclaim a small amount of old memory
and then wakes kcompactd on completion to recover the system
slightly. This introduces some overhead in the slowpath. The level
of boosting can be tuned or disabled depending on the tolerance
for fragmentation vs allocation latency.
Patch 5 stalls some movable allocation requests to let kswapd from patch 4
make some progress. The duration of the stalls is very low but it
is possible to tune the system to avoid fragmentation events if
larger stalls can be tolerated.
The bulk of the improvement in fragmentation avoidance is from patches
1-4 but patch 5 can deal with a rare corner case and provides the option
of tuning a system for THP allocation success rates in exchange for
some stalls to control fragmentation.
This patch (of 5):
The page allocator zone lists are iterated based on the watermarks of each
zone which does not take anti-fragmentation into account. On x86, node 0
may have multiple zones while other nodes have one zone. A consequence is
that tasks running on node 0 may fragment ZONE_NORMAL even though
ZONE_DMA32 has plenty of free memory. This patch special cases the
allocator fast path such that it'll try an allocation from a lower local
zone before fragmenting a higher zone. In this case, stealing of
pageblocks or orders larger than a pageblock are still allowed in the fast
path as they are uninteresting from a fragmentation point of view.
This was evaluated using a benchmark designed to fragment memory before
attempting THP allocations. It's implemented in mmtests as the following
configurations
configs/config-global-dhp__workload_thpfioscale
configs/config-global-dhp__workload_thpfioscale-defrag
configs/config-global-dhp__workload_thpfioscale-madvhugepage
e.g. from mmtests
./run-mmtests.sh --run-monitor --config configs/config-global-dhp__workload_thpfioscale test-run-1
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch).
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameter create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed.
3. Warm up a number of fio read-only processes accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll refault old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds.
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup the test files.
Note that due to the use of IO and page cache that this benchmark is not
suitable for running on large machines where the time to fragment memory
may be excessive. Also note that while this is one mix that generates
fragmentation that it's not the only mix that generates fragmentation.
Differences in workload that are more slab-intensive or whether SLUB is
used with high-order pages may yield different results.
When the page allocator fragments memory, it records the event using the
mm_page_alloc_extfrag ftrace event. If the fallback_order is smaller than
a pageblock order (order-9 on 64-bit x86) then it's considered to be an
"external fragmentation event" that may cause issues in the future.
Hence, the primary metric here is the number of external fragmentation
events that occur with order < 9. The secondary metric is allocation
latency and huge page allocation success rates but note that differences
in latencies and what the success rate also can affect the number of
external fragmentation event which is why it's a secondary metric.
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 662.92 ( 0.00%) 653.58 * 1.41%*
Amean fault-huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
Fault latencies are slightly reduced while allocation success rates remain
at zero as this configuration does not make any special effort to allocate
THP and fio is heavily active at the time and either filling memory or
keeping pages resident. However, a 49% reduction of serious fragmentation
events reduces the changes of external fragmentation being a problem in
the future.
Vlastimil asked during review for a breakdown of the allocation types
that are falling back.
vanilla
3816 MIGRATE_UNMOVABLE
800845 MIGRATE_MOVABLE
33 MIGRATE_UNRECLAIMABLE
patch
735 MIGRATE_UNMOVABLE
408135 MIGRATE_MOVABLE
42 MIGRATE_UNRECLAIMABLE
The majority of the fallbacks are due to movable allocations and this is
consistent for the workload throughout the series so will not be presented
again as the primary source of fallbacks are movable allocations.
Movable fallbacks are sometimes considered "ok" to fallback because they
can be migrated. The problem is that they can fill an
unmovable/reclaimable pageblock causing those allocations to fallback
later and polluting pageblocks with pages that cannot move. If there is a
movable fallback, it is pretty much guaranteed to affect an
unmovable/reclaimable pageblock and while it might not be enough to
actually cause a unmovable/reclaimable fallback in the future, we cannot
know that in advance so the patch takes the only option available to it.
Hence, it's important to control them. This point is also consistent
throughout the series and will not be repeated.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 1495.14 ( 0.00%) 1467.55 ( 1.85%)
Amean fault-huge-1 1098.48 ( 0.00%) 1127.11 ( -2.61%)
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 78.57 ( 0.00%) 77.64 ( -1.18%)
Fragmentation events were reduced quite a bit although this is known
to be a little variable. The latencies and allocation success rates
are similar but they were already quite high.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 1350.05 ( 0.00%) 1346.45 ( 0.27%)
Amean fault-huge-5 4181.01 ( 0.00%) 3418.60 ( 18.24%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 1.15 ( 0.00%) 0.78 ( -31.88%)
The reduction of external fragmentation events is slight and this is
partially due to the removal of __GFP_THISNODE in commit ac5b2c18911f
("mm: thp: relax __GFP_THISNODE for MADV_HUGEPAGE mappings") as THP
allocations can now spill over to remote nodes instead of fragmenting
local memory.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 6138.97 ( 0.00%) 6217.43 ( -1.28%)
Amean fault-huge-5 2294.28 ( 0.00%) 3163.33 * -37.88%*
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 96.82 ( 0.00%) 95.14 ( -1.74%)
There was a slight reduction in external fragmentation events although the
latencies were higher. The allocation success rate is high enough that
the system is struggling and there is quite a lot of parallel reclaim and
compaction activity. There is also a certain degree of luck on whether
processes start on node 0 or not for this patch but the relevance is
reduced later in the series.
Overall, the patch reduces the number of external fragmentation causing
events so the success of THP over long periods of time would be improved
for this adverse workload.
Link: http://lkml.kernel.org/r/20181123114528.28802-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:41 -08:00
|
|
|
|
|
|
|
|
|
2020-10-13 16:55:57 -07:00
|
|
|
|
|
|
|
|
|
2016-05-19 17:13:47 -07:00
|
|
|
|
2013-09-11 14:20:46 -07:00
|
|
|
|
|
|
|
|
|
2014-06-04 16:10:08 -07:00
|
|
|
|
|
|
|
|
|
2016-05-19 17:14:30 -07:00
|
|
|
|
mm: page allocator: initialise ZLC for first zone eligible for zone_reclaim
There have been a small number of complaints about significant stalls
while copying large amounts of data on NUMA machines reported on a
distribution bugzilla. In these cases, zone_reclaim was enabled by
default due to large NUMA distances. In general, the complaints have not
been about the workload itself unless it was a file server (in which case
the recommendation was disable zone_reclaim).
The stalls are mostly due to significant amounts of time spent scanning
the preferred zone for pages to free. After a failure, it might fallback
to another node (as zonelists are often node-ordered rather than
zone-ordered) but stall quickly again when the next allocation attempt
occurs. In bad cases, each page allocated results in a full scan of the
preferred zone.
Patch 1 checks the preferred zone for recent allocation failure
which is particularly important if zone_reclaim has failed
recently. This avoids rescanning the zone in the near future and
instead falling back to another node. This may hurt node locality
in some cases but a failure to zone_reclaim is more expensive than
a remote access.
Patch 2 clears the zlc information after direct reclaim.
Otherwise, zone_reclaim can mark zones full, direct reclaim can
reclaim enough pages but the zone is still not considered for
allocation.
This was tested on a 24-thread 2-node x86_64 machine. The tests were
focused on large amounts of IO. All tests were bound to the CPUs on
node-0 to avoid disturbances due to processes being scheduled on different
nodes. The kernels tested are
3.0-rc6-vanilla Vanilla 3.0-rc6
zlcfirst Patch 1 applied
zlcreconsider Patches 1+2 applied
FS-Mark
./fs_mark -d /tmp/fsmark-10813 -D 100 -N 5000 -n 208 -L 35 -t 24 -S0 -s 524288
fsmark-3.0-rc6 3.0-rc6 3.0-rc6
vanilla zlcfirs zlcreconsider
Files/s min 54.90 ( 0.00%) 49.80 (-10.24%) 49.10 (-11.81%)
Files/s mean 100.11 ( 0.00%) 135.17 (25.94%) 146.93 (31.87%)
Files/s stddev 57.51 ( 0.00%) 138.97 (58.62%) 158.69 (63.76%)
Files/s max 361.10 ( 0.00%) 834.40 (56.72%) 802.40 (55.00%)
Overhead min 76704.00 ( 0.00%) 76501.00 ( 0.27%) 77784.00 (-1.39%)
Overhead mean 1485356.51 ( 0.00%) 1035797.83 (43.40%) 1594680.26 (-6.86%)
Overhead stddev 1848122.53 ( 0.00%) 881489.88 (109.66%) 1772354.90 ( 4.27%)
Overhead max 7989060.00 ( 0.00%) 3369118.00 (137.13%) 10135324.00 (-21.18%)
MMTests Statistics: duration
User/Sys Time Running Test (seconds) 501.49 493.91 499.93
Total Elapsed Time (seconds) 2451.57 2257.48 2215.92
MMTests Statistics: vmstat
Page Ins 46268 63840 66008
Page Outs 90821596 90671128 88043732
Swap Ins 0 0 0
Swap Outs 0 0 0
Direct pages scanned 13091697 8966863 8971790
Kswapd pages scanned 0 1830011 1831116
Kswapd pages reclaimed 0 1829068 1829930
Direct pages reclaimed 13037777 8956828 8648314
Kswapd efficiency 100% 99% 99%
Kswapd velocity 0.000 810.643 826.346
Direct efficiency 99% 99% 96%
Direct velocity 5340.128 3972.068 4048.788
Percentage direct scans 100% 83% 83%
Page writes by reclaim 0 3 0
Slabs scanned 796672 720640 720256
Direct inode steals 7422667 7160012 7088638
Kswapd inode steals 0 1736840 2021238
Test completes far faster with a large increase in the number of files
created per second. Standard deviation is high as a small number of
iterations were much higher than the mean. The number of pages scanned by
zone_reclaim is reduced and kswapd is used for more work.
LARGE DD
3.0-rc6 3.0-rc6 3.0-rc6
vanilla zlcfirst zlcreconsider
download tar 59 ( 0.00%) 59 ( 0.00%) 55 ( 7.27%)
dd source files 527 ( 0.00%) 296 (78.04%) 320 (64.69%)
delete source 36 ( 0.00%) 19 (89.47%) 20 (80.00%)
MMTests Statistics: duration
User/Sys Time Running Test (seconds) 125.03 118.98 122.01
Total Elapsed Time (seconds) 624.56 375.02 398.06
MMTests Statistics: vmstat
Page Ins 3594216 439368 407032
Page Outs 23380832 23380488 23377444
Swap Ins 0 0 0
Swap Outs 0 436 287
Direct pages scanned 17482342 69315973 82864918
Kswapd pages scanned 0 519123 575425
Kswapd pages reclaimed 0 466501 522487
Direct pages reclaimed 5858054 2732949 2712547
Kswapd efficiency 100% 89% 90%
Kswapd velocity 0.000 1384.254 1445.574
Direct efficiency 33% 3% 3%
Direct velocity 27991.453 184832.737 208171.929
Percentage direct scans 100% 99% 99%
Page writes by reclaim 0 5082 13917
Slabs scanned 17280 29952 35328
Direct inode steals 115257 1431122 332201
Kswapd inode steals 0 0 979532
This test downloads a large tarfile and copies it with dd a number of
times - similar to the most recent bug report I've dealt with. Time to
completion is reduced. The number of pages scanned directly is still
disturbingly high with a low efficiency but this is likely due to the
number of dirty pages encountered. The figures could probably be improved
with more work around how kswapd is used and how dirty pages are handled
but that is separate work and this result is significant on its own.
Streaming Mapped Writer
MMTests Statistics: duration
User/Sys Time Running Test (seconds) 124.47 111.67 112.64
Total Elapsed Time (seconds) 2138.14 1816.30 1867.56
MMTests Statistics: vmstat
Page Ins 90760 89124 89516
Page Outs 121028340 120199524 120736696
Swap Ins 0 86 55
Swap Outs 0 0 0
Direct pages scanned 114989363 96461439 96330619
Kswapd pages scanned 56430948 56965763 57075875
Kswapd pages reclaimed 27743219 27752044 27766606
Direct pages reclaimed 49777 46884 36655
Kswapd efficiency 49% 48% 48%
Kswapd velocity 26392.541 31363.631 30561.736
Direct efficiency 0% 0% 0%
Direct velocity 53780.091 53108.759 51581.004
Percentage direct scans 67% 62% 62%
Page writes by reclaim 385 122 1513
Slabs scanned 43008 39040 42112
Direct inode steals 0 10 8
Kswapd inode steals 733 534 477
This test just creates a large file mapping and writes to it linearly.
Time to completion is again reduced.
The gains are mostly down to two things. In many cases, there is less
scanning as zone_reclaim simply gives up faster due to recent failures.
The second reason is that memory is used more efficiently. Instead of
scanning the preferred zone every time, the allocator falls back to
another zone and uses it instead improving overall memory utilisation.
This patch: initialise ZLC for first zone eligible for zone_reclaim.
The zonelist cache (ZLC) is used among other things to record if
zone_reclaim() failed for a particular zone recently. The intention is to
avoid a high cost scanning extremely long zonelists or scanning within the
zone uselessly.
Currently the zonelist cache is setup only after the first zone has been
considered and zone_reclaim() has been called. The objective was to avoid
a costly setup but zone_reclaim is itself quite expensive. If it is
failing regularly such as the first eligible zone having mostly mapped
pages, the cost in scanning and allocation stalls is far higher than the
ZLC initialisation step.
This patch initialises ZLC before the first eligible zone calls
zone_reclaim(). Once initialised, it is checked whether the zone failed
zone_reclaim recently. If it has, the zone is skipped. As the first zone
is now being checked, additional care has to be taken about zones marked
full. A zone can be marked "full" because it should not have enough
unmapped pages for zone_reclaim but this is excessive as direct reclaim or
kswapd may succeed where zone_reclaim fails. Only mark zones "full" after
zone_reclaim fails if it failed to reclaim enough pages after scanning.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Christoph Lameter <cl@linux.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-07-25 17:12:29 -07:00
|
|
|
|
mm: try to distribute dirty pages fairly across zones
The maximum number of dirty pages that exist in the system at any time is
determined by a number of pages considered dirtyable and a user-configured
percentage of those, or an absolute number in bytes.
This number of dirtyable pages is the sum of memory provided by all the
zones in the system minus their lowmem reserves and high watermarks, so
that the system can retain a healthy number of free pages without having
to reclaim dirty pages.
But there is a flaw in that we have a zoned page allocator which does not
care about the global state but rather the state of individual memory
zones. And right now there is nothing that prevents one zone from filling
up with dirty pages while other zones are spared, which frequently leads
to situations where kswapd, in order to restore the watermark of free
pages, does indeed have to write pages from that zone's LRU list. This
can interfere so badly with IO from the flusher threads that major
filesystems (btrfs, xfs, ext4) mostly ignore write requests from reclaim
already, taking away the VM's only possibility to keep such a zone
balanced, aside from hoping the flushers will soon clean pages from that
zone.
Enter per-zone dirty limits. They are to a zone's dirtyable memory what
the global limit is to the global amount of dirtyable memory, and try to
make sure that no single zone receives more than its fair share of the
globally allowed dirty pages in the first place. As the number of pages
considered dirtyable excludes the zones' lowmem reserves and high
watermarks, the maximum number of dirty pages in a zone is such that the
zone can always be balanced without requiring page cleaning.
As this is a placement decision in the page allocator and pages are
dirtied only after the allocation, this patch allows allocators to pass
__GFP_WRITE when they know in advance that the page will be written to and
become dirty soon. The page allocator will then attempt to allocate from
the first zone of the zonelist - which on NUMA is determined by the task's
NUMA memory policy - that has not exceeded its dirty limit.
At first glance, it would appear that the diversion to lower zones can
increase pressure on them, but this is not the case. With a full high
zone, allocations will be diverted to lower zones eventually, so it is
more of a shift in timing of the lower zone allocations. Workloads that
previously could fit their dirty pages completely in the higher zone may
be forced to allocate from lower zones, but the amount of pages that
"spill over" are limited themselves by the lower zones' dirty constraints,
and thus unlikely to become a problem.
For now, the problem of unfair dirty page distribution remains for NUMA
configurations where the zones allowed for allocation are in sum not big
enough to trigger the global dirty limits, wake up the flusher threads and
remedy the situation. Because of this, an allocation that could not
succeed on any of the considered zones is allowed to ignore the dirty
limits before going into direct reclaim or even failing the allocation,
until a future patch changes the global dirty throttling and flusher
thread activation so that they take individual zone states into account.
Test results
15M DMA + 3246M DMA32 + 504 Normal = 3765M memory
40% dirty ratio
16G USB thumb drive
10 runs of dd if=/dev/zero of=disk/zeroes bs=32k count=$((10 << 15))
seconds nr_vmscan_write
(stddev) min| median| max
xfs
vanilla: 549.747( 3.492) 0.000| 0.000| 0.000
patched: 550.996( 3.802) 0.000| 0.000| 0.000
fuse-ntfs
vanilla: 1183.094(53.178) 54349.000| 59341.000| 65163.000
patched: 558.049(17.914) 0.000| 0.000| 43.000
btrfs
vanilla: 573.679(14.015) 156657.000| 460178.000| 606926.000
patched: 563.365(11.368) 0.000| 0.000| 1362.000
ext4
vanilla: 561.197(15.782) 0.000|2725438.000|4143837.000
patched: 568.806(17.496) 0.000| 0.000| 0.000
Signed-off-by: Johannes Weiner <jweiner@redhat.com>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Michal Hocko <mhocko@suse.cz>
Tested-by: Wu Fengguang <fengguang.wu@intel.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Shaohua Li <shaohua.li@intel.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Chris Mason <chris.mason@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-10 15:07:49 -08:00
|
|
|
|
|
|
|
|
|
2016-07-28 15:46:11 -07:00
|
|
|
|
|
|
|
|
|
mm: try to distribute dirty pages fairly across zones
The maximum number of dirty pages that exist in the system at any time is
determined by a number of pages considered dirtyable and a user-configured
percentage of those, or an absolute number in bytes.
This number of dirtyable pages is the sum of memory provided by all the
zones in the system minus their lowmem reserves and high watermarks, so
that the system can retain a healthy number of free pages without having
to reclaim dirty pages.
But there is a flaw in that we have a zoned page allocator which does not
care about the global state but rather the state of individual memory
zones. And right now there is nothing that prevents one zone from filling
up with dirty pages while other zones are spared, which frequently leads
to situations where kswapd, in order to restore the watermark of free
pages, does indeed have to write pages from that zone's LRU list. This
can interfere so badly with IO from the flusher threads that major
filesystems (btrfs, xfs, ext4) mostly ignore write requests from reclaim
already, taking away the VM's only possibility to keep such a zone
balanced, aside from hoping the flushers will soon clean pages from that
zone.
Enter per-zone dirty limits. They are to a zone's dirtyable memory what
the global limit is to the global amount of dirtyable memory, and try to
make sure that no single zone receives more than its fair share of the
globally allowed dirty pages in the first place. As the number of pages
considered dirtyable excludes the zones' lowmem reserves and high
watermarks, the maximum number of dirty pages in a zone is such that the
zone can always be balanced without requiring page cleaning.
As this is a placement decision in the page allocator and pages are
dirtied only after the allocation, this patch allows allocators to pass
__GFP_WRITE when they know in advance that the page will be written to and
become dirty soon. The page allocator will then attempt to allocate from
the first zone of the zonelist - which on NUMA is determined by the task's
NUMA memory policy - that has not exceeded its dirty limit.
At first glance, it would appear that the diversion to lower zones can
increase pressure on them, but this is not the case. With a full high
zone, allocations will be diverted to lower zones eventually, so it is
more of a shift in timing of the lower zone allocations. Workloads that
previously could fit their dirty pages completely in the higher zone may
be forced to allocate from lower zones, but the amount of pages that
"spill over" are limited themselves by the lower zones' dirty constraints,
and thus unlikely to become a problem.
For now, the problem of unfair dirty page distribution remains for NUMA
configurations where the zones allowed for allocation are in sum not big
enough to trigger the global dirty limits, wake up the flusher threads and
remedy the situation. Because of this, an allocation that could not
succeed on any of the considered zones is allowed to ignore the dirty
limits before going into direct reclaim or even failing the allocation,
until a future patch changes the global dirty throttling and flusher
thread activation so that they take individual zone states into account.
Test results
15M DMA + 3246M DMA32 + 504 Normal = 3765M memory
40% dirty ratio
16G USB thumb drive
10 runs of dd if=/dev/zero of=disk/zeroes bs=32k count=$((10 << 15))
seconds nr_vmscan_write
(stddev) min| median| max
xfs
vanilla: 549.747( 3.492) 0.000| 0.000| 0.000
patched: 550.996( 3.802) 0.000| 0.000| 0.000
fuse-ntfs
vanilla: 1183.094(53.178) 54349.000| 59341.000| 65163.000
patched: 558.049(17.914) 0.000| 0.000| 43.000
btrfs
vanilla: 573.679(14.015) 156657.000| 460178.000| 606926.000
patched: 563.365(11.368) 0.000| 0.000| 1362.000
ext4
vanilla: 561.197(15.782) 0.000|2725438.000|4143837.000
patched: 568.806(17.496) 0.000| 0.000| 0.000
Signed-off-by: Johannes Weiner <jweiner@redhat.com>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Michal Hocko <mhocko@suse.cz>
Tested-by: Wu Fengguang <fengguang.wu@intel.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Shaohua Li <shaohua.li@intel.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Chris Mason <chris.mason@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-10 15:07:49 -08:00
|
|
|
|
2016-07-28 15:46:11 -07:00
|
|
|
|
mm: try to distribute dirty pages fairly across zones
The maximum number of dirty pages that exist in the system at any time is
determined by a number of pages considered dirtyable and a user-configured
percentage of those, or an absolute number in bytes.
This number of dirtyable pages is the sum of memory provided by all the
zones in the system minus their lowmem reserves and high watermarks, so
that the system can retain a healthy number of free pages without having
to reclaim dirty pages.
But there is a flaw in that we have a zoned page allocator which does not
care about the global state but rather the state of individual memory
zones. And right now there is nothing that prevents one zone from filling
up with dirty pages while other zones are spared, which frequently leads
to situations where kswapd, in order to restore the watermark of free
pages, does indeed have to write pages from that zone's LRU list. This
can interfere so badly with IO from the flusher threads that major
filesystems (btrfs, xfs, ext4) mostly ignore write requests from reclaim
already, taking away the VM's only possibility to keep such a zone
balanced, aside from hoping the flushers will soon clean pages from that
zone.
Enter per-zone dirty limits. They are to a zone's dirtyable memory what
the global limit is to the global amount of dirtyable memory, and try to
make sure that no single zone receives more than its fair share of the
globally allowed dirty pages in the first place. As the number of pages
considered dirtyable excludes the zones' lowmem reserves and high
watermarks, the maximum number of dirty pages in a zone is such that the
zone can always be balanced without requiring page cleaning.
As this is a placement decision in the page allocator and pages are
dirtied only after the allocation, this patch allows allocators to pass
__GFP_WRITE when they know in advance that the page will be written to and
become dirty soon. The page allocator will then attempt to allocate from
the first zone of the zonelist - which on NUMA is determined by the task's
NUMA memory policy - that has not exceeded its dirty limit.
At first glance, it would appear that the diversion to lower zones can
increase pressure on them, but this is not the case. With a full high
zone, allocations will be diverted to lower zones eventually, so it is
more of a shift in timing of the lower zone allocations. Workloads that
previously could fit their dirty pages completely in the higher zone may
be forced to allocate from lower zones, but the amount of pages that
"spill over" are limited themselves by the lower zones' dirty constraints,
and thus unlikely to become a problem.
For now, the problem of unfair dirty page distribution remains for NUMA
configurations where the zones allowed for allocation are in sum not big
enough to trigger the global dirty limits, wake up the flusher threads and
remedy the situation. Because of this, an allocation that could not
succeed on any of the considered zones is allowed to ignore the dirty
limits before going into direct reclaim or even failing the allocation,
until a future patch changes the global dirty throttling and flusher
thread activation so that they take individual zone states into account.
Test results
15M DMA + 3246M DMA32 + 504 Normal = 3765M memory
40% dirty ratio
16G USB thumb drive
10 runs of dd if=/dev/zero of=disk/zeroes bs=32k count=$((10 << 15))
seconds nr_vmscan_write
(stddev) min| median| max
xfs
vanilla: 549.747( 3.492) 0.000| 0.000| 0.000
patched: 550.996( 3.802) 0.000| 0.000| 0.000
fuse-ntfs
vanilla: 1183.094(53.178) 54349.000| 59341.000| 65163.000
patched: 558.049(17.914) 0.000| 0.000| 43.000
btrfs
vanilla: 573.679(14.015) 156657.000| 460178.000| 606926.000
patched: 563.365(11.368) 0.000| 0.000| 1362.000
ext4
vanilla: 561.197(15.782) 0.000|2725438.000|4143837.000
patched: 568.806(17.496) 0.000| 0.000| 0.000
Signed-off-by: Johannes Weiner <jweiner@redhat.com>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Michal Hocko <mhocko@suse.cz>
Tested-by: Wu Fengguang <fengguang.wu@intel.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Shaohua Li <shaohua.li@intel.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Chris Mason <chris.mason@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-10 15:07:49 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-07-28 15:46:11 -07:00
|
|
|
|
2015-11-06 16:28:12 -08:00
|
|
|
|
mm: try to distribute dirty pages fairly across zones
The maximum number of dirty pages that exist in the system at any time is
determined by a number of pages considered dirtyable and a user-configured
percentage of those, or an absolute number in bytes.
This number of dirtyable pages is the sum of memory provided by all the
zones in the system minus their lowmem reserves and high watermarks, so
that the system can retain a healthy number of free pages without having
to reclaim dirty pages.
But there is a flaw in that we have a zoned page allocator which does not
care about the global state but rather the state of individual memory
zones. And right now there is nothing that prevents one zone from filling
up with dirty pages while other zones are spared, which frequently leads
to situations where kswapd, in order to restore the watermark of free
pages, does indeed have to write pages from that zone's LRU list. This
can interfere so badly with IO from the flusher threads that major
filesystems (btrfs, xfs, ext4) mostly ignore write requests from reclaim
already, taking away the VM's only possibility to keep such a zone
balanced, aside from hoping the flushers will soon clean pages from that
zone.
Enter per-zone dirty limits. They are to a zone's dirtyable memory what
the global limit is to the global amount of dirtyable memory, and try to
make sure that no single zone receives more than its fair share of the
globally allowed dirty pages in the first place. As the number of pages
considered dirtyable excludes the zones' lowmem reserves and high
watermarks, the maximum number of dirty pages in a zone is such that the
zone can always be balanced without requiring page cleaning.
As this is a placement decision in the page allocator and pages are
dirtied only after the allocation, this patch allows allocators to pass
__GFP_WRITE when they know in advance that the page will be written to and
become dirty soon. The page allocator will then attempt to allocate from
the first zone of the zonelist - which on NUMA is determined by the task's
NUMA memory policy - that has not exceeded its dirty limit.
At first glance, it would appear that the diversion to lower zones can
increase pressure on them, but this is not the case. With a full high
zone, allocations will be diverted to lower zones eventually, so it is
more of a shift in timing of the lower zone allocations. Workloads that
previously could fit their dirty pages completely in the higher zone may
be forced to allocate from lower zones, but the amount of pages that
"spill over" are limited themselves by the lower zones' dirty constraints,
and thus unlikely to become a problem.
For now, the problem of unfair dirty page distribution remains for NUMA
configurations where the zones allowed for allocation are in sum not big
enough to trigger the global dirty limits, wake up the flusher threads and
remedy the situation. Because of this, an allocation that could not
succeed on any of the considered zones is allowed to ignore the dirty
limits before going into direct reclaim or even failing the allocation,
until a future patch changes the global dirty throttling and flusher
thread activation so that they take individual zone states into account.
Test results
15M DMA + 3246M DMA32 + 504 Normal = 3765M memory
40% dirty ratio
16G USB thumb drive
10 runs of dd if=/dev/zero of=disk/zeroes bs=32k count=$((10 << 15))
seconds nr_vmscan_write
(stddev) min| median| max
xfs
vanilla: 549.747( 3.492) 0.000| 0.000| 0.000
patched: 550.996( 3.802) 0.000| 0.000| 0.000
fuse-ntfs
vanilla: 1183.094(53.178) 54349.000| 59341.000| 65163.000
patched: 558.049(17.914) 0.000| 0.000| 43.000
btrfs
vanilla: 573.679(14.015) 156657.000| 460178.000| 606926.000
patched: 563.365(11.368) 0.000| 0.000| 1362.000
ext4
vanilla: 561.197(15.782) 0.000|2725438.000|4143837.000
patched: 568.806(17.496) 0.000| 0.000| 0.000
Signed-off-by: Johannes Weiner <jweiner@redhat.com>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Michal Hocko <mhocko@suse.cz>
Tested-by: Wu Fengguang <fengguang.wu@intel.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Shaohua Li <shaohua.li@intel.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Chris Mason <chris.mason@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-10 15:07:49 -08:00
|
|
|
|
2016-07-28 15:46:11 -07:00
|
|
|
|
mm: try to distribute dirty pages fairly across zones
The maximum number of dirty pages that exist in the system at any time is
determined by a number of pages considered dirtyable and a user-configured
percentage of those, or an absolute number in bytes.
This number of dirtyable pages is the sum of memory provided by all the
zones in the system minus their lowmem reserves and high watermarks, so
that the system can retain a healthy number of free pages without having
to reclaim dirty pages.
But there is a flaw in that we have a zoned page allocator which does not
care about the global state but rather the state of individual memory
zones. And right now there is nothing that prevents one zone from filling
up with dirty pages while other zones are spared, which frequently leads
to situations where kswapd, in order to restore the watermark of free
pages, does indeed have to write pages from that zone's LRU list. This
can interfere so badly with IO from the flusher threads that major
filesystems (btrfs, xfs, ext4) mostly ignore write requests from reclaim
already, taking away the VM's only possibility to keep such a zone
balanced, aside from hoping the flushers will soon clean pages from that
zone.
Enter per-zone dirty limits. They are to a zone's dirtyable memory what
the global limit is to the global amount of dirtyable memory, and try to
make sure that no single zone receives more than its fair share of the
globally allowed dirty pages in the first place. As the number of pages
considered dirtyable excludes the zones' lowmem reserves and high
watermarks, the maximum number of dirty pages in a zone is such that the
zone can always be balanced without requiring page cleaning.
As this is a placement decision in the page allocator and pages are
dirtied only after the allocation, this patch allows allocators to pass
__GFP_WRITE when they know in advance that the page will be written to and
become dirty soon. The page allocator will then attempt to allocate from
the first zone of the zonelist - which on NUMA is determined by the task's
NUMA memory policy - that has not exceeded its dirty limit.
At first glance, it would appear that the diversion to lower zones can
increase pressure on them, but this is not the case. With a full high
zone, allocations will be diverted to lower zones eventually, so it is
more of a shift in timing of the lower zone allocations. Workloads that
previously could fit their dirty pages completely in the higher zone may
be forced to allocate from lower zones, but the amount of pages that
"spill over" are limited themselves by the lower zones' dirty constraints,
and thus unlikely to become a problem.
For now, the problem of unfair dirty page distribution remains for NUMA
configurations where the zones allowed for allocation are in sum not big
enough to trigger the global dirty limits, wake up the flusher threads and
remedy the situation. Because of this, an allocation that could not
succeed on any of the considered zones is allowed to ignore the dirty
limits before going into direct reclaim or even failing the allocation,
until a future patch changes the global dirty throttling and flusher
thread activation so that they take individual zone states into account.
Test results
15M DMA + 3246M DMA32 + 504 Normal = 3765M memory
40% dirty ratio
16G USB thumb drive
10 runs of dd if=/dev/zero of=disk/zeroes bs=32k count=$((10 << 15))
seconds nr_vmscan_write
(stddev) min| median| max
xfs
vanilla: 549.747( 3.492) 0.000| 0.000| 0.000
patched: 550.996( 3.802) 0.000| 0.000| 0.000
fuse-ntfs
vanilla: 1183.094(53.178) 54349.000| 59341.000| 65163.000
patched: 558.049(17.914) 0.000| 0.000| 43.000
btrfs
vanilla: 573.679(14.015) 156657.000| 460178.000| 606926.000
patched: 563.365(11.368) 0.000| 0.000| 1362.000
ext4
vanilla: 561.197(15.782) 0.000|2725438.000|4143837.000
patched: 568.806(17.496) 0.000| 0.000| 0.000
Signed-off-by: Johannes Weiner <jweiner@redhat.com>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Michal Hocko <mhocko@suse.cz>
Tested-by: Wu Fengguang <fengguang.wu@intel.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Shaohua Li <shaohua.li@intel.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Chris Mason <chris.mason@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-10 15:07:49 -08:00
|
|
|
|
2016-07-28 15:46:11 -07:00
|
|
|
|
mm: try to distribute dirty pages fairly across zones
The maximum number of dirty pages that exist in the system at any time is
determined by a number of pages considered dirtyable and a user-configured
percentage of those, or an absolute number in bytes.
This number of dirtyable pages is the sum of memory provided by all the
zones in the system minus their lowmem reserves and high watermarks, so
that the system can retain a healthy number of free pages without having
to reclaim dirty pages.
But there is a flaw in that we have a zoned page allocator which does not
care about the global state but rather the state of individual memory
zones. And right now there is nothing that prevents one zone from filling
up with dirty pages while other zones are spared, which frequently leads
to situations where kswapd, in order to restore the watermark of free
pages, does indeed have to write pages from that zone's LRU list. This
can interfere so badly with IO from the flusher threads that major
filesystems (btrfs, xfs, ext4) mostly ignore write requests from reclaim
already, taking away the VM's only possibility to keep such a zone
balanced, aside from hoping the flushers will soon clean pages from that
zone.
Enter per-zone dirty limits. They are to a zone's dirtyable memory what
the global limit is to the global amount of dirtyable memory, and try to
make sure that no single zone receives more than its fair share of the
globally allowed dirty pages in the first place. As the number of pages
considered dirtyable excludes the zones' lowmem reserves and high
watermarks, the maximum number of dirty pages in a zone is such that the
zone can always be balanced without requiring page cleaning.
As this is a placement decision in the page allocator and pages are
dirtied only after the allocation, this patch allows allocators to pass
__GFP_WRITE when they know in advance that the page will be written to and
become dirty soon. The page allocator will then attempt to allocate from
the first zone of the zonelist - which on NUMA is determined by the task's
NUMA memory policy - that has not exceeded its dirty limit.
At first glance, it would appear that the diversion to lower zones can
increase pressure on them, but this is not the case. With a full high
zone, allocations will be diverted to lower zones eventually, so it is
more of a shift in timing of the lower zone allocations. Workloads that
previously could fit their dirty pages completely in the higher zone may
be forced to allocate from lower zones, but the amount of pages that
"spill over" are limited themselves by the lower zones' dirty constraints,
and thus unlikely to become a problem.
For now, the problem of unfair dirty page distribution remains for NUMA
configurations where the zones allowed for allocation are in sum not big
enough to trigger the global dirty limits, wake up the flusher threads and
remedy the situation. Because of this, an allocation that could not
succeed on any of the considered zones is allowed to ignore the dirty
limits before going into direct reclaim or even failing the allocation,
until a future patch changes the global dirty throttling and flusher
thread activation so that they take individual zone states into account.
Test results
15M DMA + 3246M DMA32 + 504 Normal = 3765M memory
40% dirty ratio
16G USB thumb drive
10 runs of dd if=/dev/zero of=disk/zeroes bs=32k count=$((10 << 15))
seconds nr_vmscan_write
(stddev) min| median| max
xfs
vanilla: 549.747( 3.492) 0.000| 0.000| 0.000
patched: 550.996( 3.802) 0.000| 0.000| 0.000
fuse-ntfs
vanilla: 1183.094(53.178) 54349.000| 59341.000| 65163.000
patched: 558.049(17.914) 0.000| 0.000| 43.000
btrfs
vanilla: 573.679(14.015) 156657.000| 460178.000| 606926.000
patched: 563.365(11.368) 0.000| 0.000| 1362.000
ext4
vanilla: 561.197(15.782) 0.000|2725438.000|4143837.000
patched: 568.806(17.496) 0.000| 0.000| 0.000
Signed-off-by: Johannes Weiner <jweiner@redhat.com>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Michal Hocko <mhocko@suse.cz>
Tested-by: Wu Fengguang <fengguang.wu@intel.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Shaohua Li <shaohua.li@intel.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Chris Mason <chris.mason@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-10 15:07:49 -08:00
|
|
|
|
|
|
|
|
|
2016-07-28 15:46:53 -07:00
|
|
|
|
2022-05-12 20:22:51 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-07-28 15:46:53 -07:00
|
|
|
|
2022-05-12 20:22:51 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-07-28 15:46:53 -07:00
|
|
|
|
2005-11-13 16:06:43 -08:00
|
|
|
|
mm: page_alloc: defrag_mode
The page allocator groups requests by migratetype to stave off
fragmentation. However, in practice this is routinely defeated by the
fact that it gives up *before* invoking reclaim and compaction - which may
well produce suitable pages. As a result, fragmentation of physical
memory is a common ongoing process in many load scenarios.
Fragmentation deteriorates compaction's ability to produce huge pages.
Depending on the lifetime of the fragmenting allocations, those effects
can be long-lasting or even permanent, requiring drastic measures like
forcible idle states or even reboots as the only reliable ways to recover
the address space for THP production.
In a kernel build test with supplemental THP pressure, the THP allocation
rate steadily declines over 15 runs:
thp_fault_alloc
61988
56474
57258
50187
52388
55409
52925
47648
43669
40621
36077
41721
36685
34641
33215
This is a hurdle in adopting THP in any environment where hosts are shared
between multiple overlapping workloads (cloud environments), and rarely
experience true idle periods. To make THP a reliable and predictable
optimization, there needs to be a stronger guarantee to avoid such
fragmentation.
Introduce defrag_mode. When enabled, reclaim/compaction is invoked to its
full extent *before* falling back. Specifically, ALLOC_NOFRAGMENT is
enforced on the allocator fastpath and the reclaiming slowpath.
For now, fallbacks are permitted to avert OOMs. There is a plan to add
defrag_mode=2 to prefer OOMs over fragmentation, but this requires
additional prep work in compaction and the reserve management to make it
ready for all possible allocation contexts.
The following test results are from a kernel build with periodic bursts of
THP allocations, over 15 runs:
vanilla defrag_mode=1
@claimer[unmovable]: 189 103
@claimer[movable]: 92 103
@claimer[reclaimable]: 207 61
@pollute[unmovable from movable]: 25 0
@pollute[unmovable from reclaimable]: 28 0
@pollute[movable from unmovable]: 38835 0
@pollute[movable from reclaimable]: 147136 0
@pollute[reclaimable from unmovable]: 178 0
@pollute[reclaimable from movable]: 33 0
@steal[unmovable from movable]: 11 0
@steal[unmovable from reclaimable]: 5 0
@steal[reclaimable from unmovable]: 107 0
@steal[reclaimable from movable]: 90 0
@steal[movable from reclaimable]: 354 0
@steal[movable from unmovable]: 130 0
Both types of polluting fallbacks are eliminated in this workload.
Interestingly, whole block conversions are reduced as well. This is
because once a block is claimed for a type, its empty space remains
available for future allocations, instead of being padded with fallbacks;
this allows the native type to group up instead of spreading out to new
blocks. The assumption in the allocator has been that pollution from
movable allocations is less harmful than from other types, since they can
be reclaimed or migrated out should the space be needed. However, since
fallbacks occur *before* reclaim/compaction is invoked, movable pollution
will still cause non-movable allocations to spread out and claim more
blocks.
Without fragmentation, THP rates hold steady with defrag_mode=1:
thp_fault_alloc
32478
20725
45045
32130
14018
21711
40791
29134
34458
45381
28305
17265
22584
28454
30850
While the downward trend is eliminated, the keen reader will of course
notice that the baseline rate is much smaller than the vanilla kernel's to
begin with. This is due to deficiencies in how reclaim and compaction are
currently driven: ALLOC_NOFRAGMENT increases the extent to which smaller
allocations are competing with THPs for pageblocks, while making no effort
themselves to reclaim or compact beyond their own request size. This
effect already exists with the current usage of ALLOC_NOFRAGMENT, but is
amplified by defrag_mode insisting on whole block stealing much more
strongly.
Subsequent patches will address defrag_mode reclaim strategy to raise the
THP success baseline above the vanilla kernel.
Link: https://lkml.kernel.org/r/20250313210647.1314586-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-03-13 17:05:34 -04:00
|
|
|
|
2024-07-29 14:47:17 +05:30
|
|
|
|
mm, page_alloc: spread allocations across zones before introducing fragmentation
Patch series "Fragmentation avoidance improvements", v5.
It has been noted before that fragmentation avoidance (aka
anti-fragmentation) is not perfect. Given sufficient time or an adverse
workload, memory gets fragmented and the long-term success of high-order
allocations degrades. This series defines an adverse workload, a definition
of external fragmentation events (including serious) ones and a series
that reduces the level of those fragmentation events.
The details of the workload and the consequences are described in more
detail in the changelogs. However, from patch 1, this is a high-level
summary of the adverse workload. The exact details are found in the
mmtests implementation.
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch)
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameterr create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed
3. Warm up a number of fio read-only threads accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll fault back in old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup
Overall the series reduces external fragmentation causing events by over 94%
on 1 and 2 socket machines, which in turn impacts high-order allocation
success rates over the long term. There are differences in latencies and
high-order allocation success rates. Latencies are a mixed bag as they
are vulnerable to exact system state and whether allocations succeeded
so they are treated as a secondary metric.
Patch 1 uses lower zones if they are populated and have free memory
instead of fragmenting a higher zone. It's special cased to
handle a Normal->DMA32 fallback with the reasons explained
in the changelog.
Patch 2-4 boosts watermarks temporarily when an external fragmentation
event occurs. kswapd wakes to reclaim a small amount of old memory
and then wakes kcompactd on completion to recover the system
slightly. This introduces some overhead in the slowpath. The level
of boosting can be tuned or disabled depending on the tolerance
for fragmentation vs allocation latency.
Patch 5 stalls some movable allocation requests to let kswapd from patch 4
make some progress. The duration of the stalls is very low but it
is possible to tune the system to avoid fragmentation events if
larger stalls can be tolerated.
The bulk of the improvement in fragmentation avoidance is from patches
1-4 but patch 5 can deal with a rare corner case and provides the option
of tuning a system for THP allocation success rates in exchange for
some stalls to control fragmentation.
This patch (of 5):
The page allocator zone lists are iterated based on the watermarks of each
zone which does not take anti-fragmentation into account. On x86, node 0
may have multiple zones while other nodes have one zone. A consequence is
that tasks running on node 0 may fragment ZONE_NORMAL even though
ZONE_DMA32 has plenty of free memory. This patch special cases the
allocator fast path such that it'll try an allocation from a lower local
zone before fragmenting a higher zone. In this case, stealing of
pageblocks or orders larger than a pageblock are still allowed in the fast
path as they are uninteresting from a fragmentation point of view.
This was evaluated using a benchmark designed to fragment memory before
attempting THP allocations. It's implemented in mmtests as the following
configurations
configs/config-global-dhp__workload_thpfioscale
configs/config-global-dhp__workload_thpfioscale-defrag
configs/config-global-dhp__workload_thpfioscale-madvhugepage
e.g. from mmtests
./run-mmtests.sh --run-monitor --config configs/config-global-dhp__workload_thpfioscale test-run-1
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch).
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameter create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed.
3. Warm up a number of fio read-only processes accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll refault old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds.
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup the test files.
Note that due to the use of IO and page cache that this benchmark is not
suitable for running on large machines where the time to fragment memory
may be excessive. Also note that while this is one mix that generates
fragmentation that it's not the only mix that generates fragmentation.
Differences in workload that are more slab-intensive or whether SLUB is
used with high-order pages may yield different results.
When the page allocator fragments memory, it records the event using the
mm_page_alloc_extfrag ftrace event. If the fallback_order is smaller than
a pageblock order (order-9 on 64-bit x86) then it's considered to be an
"external fragmentation event" that may cause issues in the future.
Hence, the primary metric here is the number of external fragmentation
events that occur with order < 9. The secondary metric is allocation
latency and huge page allocation success rates but note that differences
in latencies and what the success rate also can affect the number of
external fragmentation event which is why it's a secondary metric.
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 662.92 ( 0.00%) 653.58 * 1.41%*
Amean fault-huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
Fault latencies are slightly reduced while allocation success rates remain
at zero as this configuration does not make any special effort to allocate
THP and fio is heavily active at the time and either filling memory or
keeping pages resident. However, a 49% reduction of serious fragmentation
events reduces the changes of external fragmentation being a problem in
the future.
Vlastimil asked during review for a breakdown of the allocation types
that are falling back.
vanilla
3816 MIGRATE_UNMOVABLE
800845 MIGRATE_MOVABLE
33 MIGRATE_UNRECLAIMABLE
patch
735 MIGRATE_UNMOVABLE
408135 MIGRATE_MOVABLE
42 MIGRATE_UNRECLAIMABLE
The majority of the fallbacks are due to movable allocations and this is
consistent for the workload throughout the series so will not be presented
again as the primary source of fallbacks are movable allocations.
Movable fallbacks are sometimes considered "ok" to fallback because they
can be migrated. The problem is that they can fill an
unmovable/reclaimable pageblock causing those allocations to fallback
later and polluting pageblocks with pages that cannot move. If there is a
movable fallback, it is pretty much guaranteed to affect an
unmovable/reclaimable pageblock and while it might not be enough to
actually cause a unmovable/reclaimable fallback in the future, we cannot
know that in advance so the patch takes the only option available to it.
Hence, it's important to control them. This point is also consistent
throughout the series and will not be repeated.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 1495.14 ( 0.00%) 1467.55 ( 1.85%)
Amean fault-huge-1 1098.48 ( 0.00%) 1127.11 ( -2.61%)
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 78.57 ( 0.00%) 77.64 ( -1.18%)
Fragmentation events were reduced quite a bit although this is known
to be a little variable. The latencies and allocation success rates
are similar but they were already quite high.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 1350.05 ( 0.00%) 1346.45 ( 0.27%)
Amean fault-huge-5 4181.01 ( 0.00%) 3418.60 ( 18.24%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 1.15 ( 0.00%) 0.78 ( -31.88%)
The reduction of external fragmentation events is slight and this is
partially due to the removal of __GFP_THISNODE in commit ac5b2c18911f
("mm: thp: relax __GFP_THISNODE for MADV_HUGEPAGE mappings") as THP
allocations can now spill over to remote nodes instead of fragmenting
local memory.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 6138.97 ( 0.00%) 6217.43 ( -1.28%)
Amean fault-huge-5 2294.28 ( 0.00%) 3163.33 * -37.88%*
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 96.82 ( 0.00%) 95.14 ( -1.74%)
There was a slight reduction in external fragmentation events although the
latencies were higher. The allocation success rate is high enough that
the system is struggling and there is quite a lot of parallel reclaim and
compaction activity. There is also a certain degree of luck on whether
processes start on node 0 or not for this patch but the relevance is
reduced later in the series.
Overall, the patch reduces the number of external fragmentation causing
events so the success of THP over long periods of time would be improved
for this adverse workload.
Link: http://lkml.kernel.org/r/20181123114528.28802-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:41 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-07-29 14:47:17 +05:30
|
|
|
|
mm, page_alloc: spread allocations across zones before introducing fragmentation
Patch series "Fragmentation avoidance improvements", v5.
It has been noted before that fragmentation avoidance (aka
anti-fragmentation) is not perfect. Given sufficient time or an adverse
workload, memory gets fragmented and the long-term success of high-order
allocations degrades. This series defines an adverse workload, a definition
of external fragmentation events (including serious) ones and a series
that reduces the level of those fragmentation events.
The details of the workload and the consequences are described in more
detail in the changelogs. However, from patch 1, this is a high-level
summary of the adverse workload. The exact details are found in the
mmtests implementation.
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch)
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameterr create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed
3. Warm up a number of fio read-only threads accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll fault back in old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup
Overall the series reduces external fragmentation causing events by over 94%
on 1 and 2 socket machines, which in turn impacts high-order allocation
success rates over the long term. There are differences in latencies and
high-order allocation success rates. Latencies are a mixed bag as they
are vulnerable to exact system state and whether allocations succeeded
so they are treated as a secondary metric.
Patch 1 uses lower zones if they are populated and have free memory
instead of fragmenting a higher zone. It's special cased to
handle a Normal->DMA32 fallback with the reasons explained
in the changelog.
Patch 2-4 boosts watermarks temporarily when an external fragmentation
event occurs. kswapd wakes to reclaim a small amount of old memory
and then wakes kcompactd on completion to recover the system
slightly. This introduces some overhead in the slowpath. The level
of boosting can be tuned or disabled depending on the tolerance
for fragmentation vs allocation latency.
Patch 5 stalls some movable allocation requests to let kswapd from patch 4
make some progress. The duration of the stalls is very low but it
is possible to tune the system to avoid fragmentation events if
larger stalls can be tolerated.
The bulk of the improvement in fragmentation avoidance is from patches
1-4 but patch 5 can deal with a rare corner case and provides the option
of tuning a system for THP allocation success rates in exchange for
some stalls to control fragmentation.
This patch (of 5):
The page allocator zone lists are iterated based on the watermarks of each
zone which does not take anti-fragmentation into account. On x86, node 0
may have multiple zones while other nodes have one zone. A consequence is
that tasks running on node 0 may fragment ZONE_NORMAL even though
ZONE_DMA32 has plenty of free memory. This patch special cases the
allocator fast path such that it'll try an allocation from a lower local
zone before fragmenting a higher zone. In this case, stealing of
pageblocks or orders larger than a pageblock are still allowed in the fast
path as they are uninteresting from a fragmentation point of view.
This was evaluated using a benchmark designed to fragment memory before
attempting THP allocations. It's implemented in mmtests as the following
configurations
configs/config-global-dhp__workload_thpfioscale
configs/config-global-dhp__workload_thpfioscale-defrag
configs/config-global-dhp__workload_thpfioscale-madvhugepage
e.g. from mmtests
./run-mmtests.sh --run-monitor --config configs/config-global-dhp__workload_thpfioscale test-run-1
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch).
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameter create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed.
3. Warm up a number of fio read-only processes accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll refault old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds.
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup the test files.
Note that due to the use of IO and page cache that this benchmark is not
suitable for running on large machines where the time to fragment memory
may be excessive. Also note that while this is one mix that generates
fragmentation that it's not the only mix that generates fragmentation.
Differences in workload that are more slab-intensive or whether SLUB is
used with high-order pages may yield different results.
When the page allocator fragments memory, it records the event using the
mm_page_alloc_extfrag ftrace event. If the fallback_order is smaller than
a pageblock order (order-9 on 64-bit x86) then it's considered to be an
"external fragmentation event" that may cause issues in the future.
Hence, the primary metric here is the number of external fragmentation
events that occur with order < 9. The secondary metric is allocation
latency and huge page allocation success rates but note that differences
in latencies and what the success rate also can affect the number of
external fragmentation event which is why it's a secondary metric.
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 662.92 ( 0.00%) 653.58 * 1.41%*
Amean fault-huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
Fault latencies are slightly reduced while allocation success rates remain
at zero as this configuration does not make any special effort to allocate
THP and fio is heavily active at the time and either filling memory or
keeping pages resident. However, a 49% reduction of serious fragmentation
events reduces the changes of external fragmentation being a problem in
the future.
Vlastimil asked during review for a breakdown of the allocation types
that are falling back.
vanilla
3816 MIGRATE_UNMOVABLE
800845 MIGRATE_MOVABLE
33 MIGRATE_UNRECLAIMABLE
patch
735 MIGRATE_UNMOVABLE
408135 MIGRATE_MOVABLE
42 MIGRATE_UNRECLAIMABLE
The majority of the fallbacks are due to movable allocations and this is
consistent for the workload throughout the series so will not be presented
again as the primary source of fallbacks are movable allocations.
Movable fallbacks are sometimes considered "ok" to fallback because they
can be migrated. The problem is that they can fill an
unmovable/reclaimable pageblock causing those allocations to fallback
later and polluting pageblocks with pages that cannot move. If there is a
movable fallback, it is pretty much guaranteed to affect an
unmovable/reclaimable pageblock and while it might not be enough to
actually cause a unmovable/reclaimable fallback in the future, we cannot
know that in advance so the patch takes the only option available to it.
Hence, it's important to control them. This point is also consistent
throughout the series and will not be repeated.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 1495.14 ( 0.00%) 1467.55 ( 1.85%)
Amean fault-huge-1 1098.48 ( 0.00%) 1127.11 ( -2.61%)
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 78.57 ( 0.00%) 77.64 ( -1.18%)
Fragmentation events were reduced quite a bit although this is known
to be a little variable. The latencies and allocation success rates
are similar but they were already quite high.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 1350.05 ( 0.00%) 1346.45 ( 0.27%)
Amean fault-huge-5 4181.01 ( 0.00%) 3418.60 ( 18.24%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 1.15 ( 0.00%) 0.78 ( -31.88%)
The reduction of external fragmentation events is slight and this is
partially due to the removal of __GFP_THISNODE in commit ac5b2c18911f
("mm: thp: relax __GFP_THISNODE for MADV_HUGEPAGE mappings") as THP
allocations can now spill over to remote nodes instead of fragmenting
local memory.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 6138.97 ( 0.00%) 6217.43 ( -1.28%)
Amean fault-huge-5 2294.28 ( 0.00%) 3163.33 * -37.88%*
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 96.82 ( 0.00%) 95.14 ( -1.74%)
There was a slight reduction in external fragmentation events although the
latencies were higher. The allocation success rate is high enough that
the system is struggling and there is quite a lot of parallel reclaim and
compaction activity. There is also a certain degree of luck on whether
processes start on node 0 or not for this patch but the relevance is
reduced later in the series.
Overall, the patch reduces the number of external fragmentation causing
events so the success of THP over long periods of time would be improved
for this adverse workload.
Link: http://lkml.kernel.org/r/20181123114528.28802-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:41 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: page_alloc: avoid kswapd thrashing due to NUMA restrictions
On NUMA systems without bindings, allocations check all nodes for free
space, then wake up the kswapds on all nodes and retry. This ensures
all available space is evenly used before reclaim begins. However,
when one process or certain allocations have node restrictions, they
can cause kswapds on only a subset of nodes to be woken up.
Since kswapd hysteresis targets watermarks that are *higher* than
needed for allocation, even *unrestricted* allocations can now get
suckered onto such nodes that are already pressured. This ends up
concentrating all allocations on them, even when there are idle nodes
available for the unrestricted requests.
This was observed with two numa nodes, where node0 is normal and node1
is ZONE_MOVABLE to facilitate hotplugging: a kernel allocation wakes
kswapd on node0 only (since node1 is not eligible); once kswapd0 is
active, the watermarks hover between low and high, and then even the
movable allocations end up on node0, only to be kicked out again;
meanwhile node1 is empty and idle.
Similar behavior is possible when a process with NUMA bindings is
causing selective kswapd wakeups.
To fix this, on NUMA systems augment the (misleading) watermark test
with a check for whether kswapd is already active during the first
iteration through the zonelist. If this fails to place the request,
kswapd must be running everywhere already, and the watermark test is
good enough to decide placement.
With this patch, unrestricted requests successfully make use of node1,
even while kswapd is reclaiming node0 for restricted allocations.
[gourry@gourry.net: don't retry if no kswapds were active]
Link: https://lkml.kernel.org/r/20250919162134.1098208-1-hannes@cmpxchg.org
Signed-off-by: Gregory Price <gourry@gourry.net>
Tested-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Zi Yan <ziy@nvidia.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: Joshua Hahn <joshua.hahnjy@gmail.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-09-19 12:21:34 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-05-06 14:25:08 +03:00
|
|
|
|
2024-08-09 14:48:47 +03:00
|
|
|
|
mm, pcp: decrease PCP high if free pages < high watermark
One target of PCP is to minimize pages in PCP if the system free pages is
too few. To reach that target, when page reclaiming is active for the
zone (ZONE_RECLAIM_ACTIVE), we will stop increasing PCP high in allocating
path, decrease PCP high and free some pages in freeing path. But this may
be too late because the background page reclaiming may introduce latency
for some workloads. So, in this patch, during page allocation we will
detect whether the number of free pages of the zone is below high
watermark. If so, we will stop increasing PCP high in allocating path,
decrease PCP high and free some pages in freeing path. With this, we can
reduce the possibility of the premature background page reclaiming caused
by too large PCP.
The high watermark checking is done in allocating path to reduce the
overhead in hotter freeing path.
Link: https://lkml.kernel.org/r/20231016053002.756205-9-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:30:01 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-12-28 00:35:44 -08:00
|
|
|
|
2016-05-19 17:14:07 -07:00
|
|
|
|
mm, page_alloc: skip ->waternark_boost for atomic order-0 allocations
When boosting is enabled, it is observed that rate of atomic order-0
allocation failures are high due to the fact that free levels in the
system are checked with ->watermark_boost offset. This is not a problem
for sleepable allocations but for atomic allocations which looks like
regression.
This problem is seen frequently on system setup of Android kernel running
on Snapdragon hardware with 4GB RAM size. When no extfrag event occurred
in the system, ->watermark_boost factor is zero, thus the watermark
configurations in the system are:
_watermark = (
[WMARK_MIN] = 1272, --> ~5MB
[WMARK_LOW] = 9067, --> ~36MB
[WMARK_HIGH] = 9385), --> ~38MB
watermark_boost = 0
After launching some memory hungry applications in Android which can cause
extfrag events in the system to an extent that ->watermark_boost can be
set to max i.e. default boost factor makes it to 150% of high watermark.
_watermark = (
[WMARK_MIN] = 1272, --> ~5MB
[WMARK_LOW] = 9067, --> ~36MB
[WMARK_HIGH] = 9385), --> ~38MB
watermark_boost = 14077, -->~57MB
With default system configuration, for an atomic order-0 allocation to
succeed, having free memory of ~2MB will suffice. But boosting makes the
min_wmark to ~61MB thus for an atomic order-0 allocation to be successful
system should have minimum of ~23MB of free memory(from calculations of
zone_watermark_ok(), min = 3/4(min/2)). But failures are observed despite
system is having ~20MB of free memory. In the testing, this is
reproducible as early as first 300secs since boot and with furtherlowram
configurations(<2GB) it is observed as early as first 150secs since boot.
These failures can be avoided by excluding the ->watermark_boost in
watermark caluculations for atomic order-0 allocations.
[akpm@linux-foundation.org: fix comment grammar, reflow comment]
[charante@codeaurora.org: fix suggested by Mel Gorman]
Link: http://lkml.kernel.org/r/31556793-57b1-1c21-1a9d-22674d9bd938@codeaurora.org
Signed-off-by: Charan Teja Reddy <charante@codeaurora.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Vinayak Menon <vinmenon@codeaurora.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Link: http://lkml.kernel.org/r/1589882284-21010-1-git-send-email-charante@codeaurora.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-06 23:25:24 -07:00
|
|
|
|
|
|
|
|
|
2009-06-16 15:33:22 -07:00
|
|
|
|
|
|
|
|
|
2025-05-06 14:25:08 +03:00
|
|
|
|
2024-08-09 14:48:47 +03:00
|
|
|
|
2023-06-06 17:26:29 +03:00
|
|
|
|
2018-04-05 16:22:31 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-01-05 13:55:06 +05:30
|
|
|
|
2018-04-05 16:22:31 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-06-04 16:10:14 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-05-04 18:36:04 -07:00
|
|
|
|
2024-07-29 14:47:17 +05:30
|
|
|
|
mm: page allocator: initialise ZLC for first zone eligible for zone_reclaim
There have been a small number of complaints about significant stalls
while copying large amounts of data on NUMA machines reported on a
distribution bugzilla. In these cases, zone_reclaim was enabled by
default due to large NUMA distances. In general, the complaints have not
been about the workload itself unless it was a file server (in which case
the recommendation was disable zone_reclaim).
The stalls are mostly due to significant amounts of time spent scanning
the preferred zone for pages to free. After a failure, it might fallback
to another node (as zonelists are often node-ordered rather than
zone-ordered) but stall quickly again when the next allocation attempt
occurs. In bad cases, each page allocated results in a full scan of the
preferred zone.
Patch 1 checks the preferred zone for recent allocation failure
which is particularly important if zone_reclaim has failed
recently. This avoids rescanning the zone in the near future and
instead falling back to another node. This may hurt node locality
in some cases but a failure to zone_reclaim is more expensive than
a remote access.
Patch 2 clears the zlc information after direct reclaim.
Otherwise, zone_reclaim can mark zones full, direct reclaim can
reclaim enough pages but the zone is still not considered for
allocation.
This was tested on a 24-thread 2-node x86_64 machine. The tests were
focused on large amounts of IO. All tests were bound to the CPUs on
node-0 to avoid disturbances due to processes being scheduled on different
nodes. The kernels tested are
3.0-rc6-vanilla Vanilla 3.0-rc6
zlcfirst Patch 1 applied
zlcreconsider Patches 1+2 applied
FS-Mark
./fs_mark -d /tmp/fsmark-10813 -D 100 -N 5000 -n 208 -L 35 -t 24 -S0 -s 524288
fsmark-3.0-rc6 3.0-rc6 3.0-rc6
vanilla zlcfirs zlcreconsider
Files/s min 54.90 ( 0.00%) 49.80 (-10.24%) 49.10 (-11.81%)
Files/s mean 100.11 ( 0.00%) 135.17 (25.94%) 146.93 (31.87%)
Files/s stddev 57.51 ( 0.00%) 138.97 (58.62%) 158.69 (63.76%)
Files/s max 361.10 ( 0.00%) 834.40 (56.72%) 802.40 (55.00%)
Overhead min 76704.00 ( 0.00%) 76501.00 ( 0.27%) 77784.00 (-1.39%)
Overhead mean 1485356.51 ( 0.00%) 1035797.83 (43.40%) 1594680.26 (-6.86%)
Overhead stddev 1848122.53 ( 0.00%) 881489.88 (109.66%) 1772354.90 ( 4.27%)
Overhead max 7989060.00 ( 0.00%) 3369118.00 (137.13%) 10135324.00 (-21.18%)
MMTests Statistics: duration
User/Sys Time Running Test (seconds) 501.49 493.91 499.93
Total Elapsed Time (seconds) 2451.57 2257.48 2215.92
MMTests Statistics: vmstat
Page Ins 46268 63840 66008
Page Outs 90821596 90671128 88043732
Swap Ins 0 0 0
Swap Outs 0 0 0
Direct pages scanned 13091697 8966863 8971790
Kswapd pages scanned 0 1830011 1831116
Kswapd pages reclaimed 0 1829068 1829930
Direct pages reclaimed 13037777 8956828 8648314
Kswapd efficiency 100% 99% 99%
Kswapd velocity 0.000 810.643 826.346
Direct efficiency 99% 99% 96%
Direct velocity 5340.128 3972.068 4048.788
Percentage direct scans 100% 83% 83%
Page writes by reclaim 0 3 0
Slabs scanned 796672 720640 720256
Direct inode steals 7422667 7160012 7088638
Kswapd inode steals 0 1736840 2021238
Test completes far faster with a large increase in the number of files
created per second. Standard deviation is high as a small number of
iterations were much higher than the mean. The number of pages scanned by
zone_reclaim is reduced and kswapd is used for more work.
LARGE DD
3.0-rc6 3.0-rc6 3.0-rc6
vanilla zlcfirst zlcreconsider
download tar 59 ( 0.00%) 59 ( 0.00%) 55 ( 7.27%)
dd source files 527 ( 0.00%) 296 (78.04%) 320 (64.69%)
delete source 36 ( 0.00%) 19 (89.47%) 20 (80.00%)
MMTests Statistics: duration
User/Sys Time Running Test (seconds) 125.03 118.98 122.01
Total Elapsed Time (seconds) 624.56 375.02 398.06
MMTests Statistics: vmstat
Page Ins 3594216 439368 407032
Page Outs 23380832 23380488 23377444
Swap Ins 0 0 0
Swap Outs 0 436 287
Direct pages scanned 17482342 69315973 82864918
Kswapd pages scanned 0 519123 575425
Kswapd pages reclaimed 0 466501 522487
Direct pages reclaimed 5858054 2732949 2712547
Kswapd efficiency 100% 89% 90%
Kswapd velocity 0.000 1384.254 1445.574
Direct efficiency 33% 3% 3%
Direct velocity 27991.453 184832.737 208171.929
Percentage direct scans 100% 99% 99%
Page writes by reclaim 0 5082 13917
Slabs scanned 17280 29952 35328
Direct inode steals 115257 1431122 332201
Kswapd inode steals 0 0 979532
This test downloads a large tarfile and copies it with dd a number of
times - similar to the most recent bug report I've dealt with. Time to
completion is reduced. The number of pages scanned directly is still
disturbingly high with a low efficiency but this is likely due to the
number of dirty pages encountered. The figures could probably be improved
with more work around how kswapd is used and how dirty pages are handled
but that is separate work and this result is significant on its own.
Streaming Mapped Writer
MMTests Statistics: duration
User/Sys Time Running Test (seconds) 124.47 111.67 112.64
Total Elapsed Time (seconds) 2138.14 1816.30 1867.56
MMTests Statistics: vmstat
Page Ins 90760 89124 89516
Page Outs 121028340 120199524 120736696
Swap Ins 0 86 55
Swap Outs 0 0 0
Direct pages scanned 114989363 96461439 96330619
Kswapd pages scanned 56430948 56965763 57075875
Kswapd pages reclaimed 27743219 27752044 27766606
Direct pages reclaimed 49777 46884 36655
Kswapd efficiency 49% 48% 48%
Kswapd velocity 26392.541 31363.631 30561.736
Direct efficiency 0% 0% 0%
Direct velocity 53780.091 53108.759 51581.004
Percentage direct scans 67% 62% 62%
Page writes by reclaim 385 122 1513
Slabs scanned 43008 39040 42112
Direct inode steals 0 10 8
Kswapd inode steals 733 534 477
This test just creates a large file mapping and writes to it linearly.
Time to completion is again reduced.
The gains are mostly down to two things. In many cases, there is less
scanning as zone_reclaim simply gives up faster due to recent failures.
The second reason is that memory is used more efficiently. Instead of
scanning the preferred zone every time, the allocator falls back to
another zone and uses it instead improving overall memory utilisation.
This patch: initialise ZLC for first zone eligible for zone_reclaim.
The zonelist cache (ZLC) is used among other things to record if
zone_reclaim() failed for a particular zone recently. The intention is to
avoid a high cost scanning extremely long zonelists or scanning within the
zone uselessly.
Currently the zonelist cache is setup only after the first zone has been
considered and zone_reclaim() has been called. The objective was to avoid
a costly setup but zone_reclaim is itself quite expensive. If it is
failing regularly such as the first eligible zone having mostly mapped
pages, the cost in scanning and allocation stalls is far higher than the
ZLC initialisation step.
This patch initialises ZLC before the first eligible zone calls
zone_reclaim(). Once initialised, it is checked whether the zone failed
zone_reclaim recently. If it has, the zone is skipped. As the first zone
is now being checked, additional care has to be taken about zones marked
full. A zone can be marked "full" because it should not have enough
unmapped pages for zone_reclaim but this is excessive as direct reclaim or
kswapd may succeed where zone_reclaim fails. Only mark zones "full" after
zone_reclaim fails if it failed to reclaim enough pages after scanning.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Christoph Lameter <cl@linux.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-07-25 17:12:29 -07:00
|
|
|
|
|
|
|
|
|
2016-07-28 15:46:32 -07:00
|
|
|
|
2009-06-16 15:33:22 -07:00
|
|
|
|
2016-07-28 15:46:32 -07:00
|
|
|
|
2009-06-16 15:33:22 -07:00
|
|
|
|
mm: page allocator: initialise ZLC for first zone eligible for zone_reclaim
There have been a small number of complaints about significant stalls
while copying large amounts of data on NUMA machines reported on a
distribution bugzilla. In these cases, zone_reclaim was enabled by
default due to large NUMA distances. In general, the complaints have not
been about the workload itself unless it was a file server (in which case
the recommendation was disable zone_reclaim).
The stalls are mostly due to significant amounts of time spent scanning
the preferred zone for pages to free. After a failure, it might fallback
to another node (as zonelists are often node-ordered rather than
zone-ordered) but stall quickly again when the next allocation attempt
occurs. In bad cases, each page allocated results in a full scan of the
preferred zone.
Patch 1 checks the preferred zone for recent allocation failure
which is particularly important if zone_reclaim has failed
recently. This avoids rescanning the zone in the near future and
instead falling back to another node. This may hurt node locality
in some cases but a failure to zone_reclaim is more expensive than
a remote access.
Patch 2 clears the zlc information after direct reclaim.
Otherwise, zone_reclaim can mark zones full, direct reclaim can
reclaim enough pages but the zone is still not considered for
allocation.
This was tested on a 24-thread 2-node x86_64 machine. The tests were
focused on large amounts of IO. All tests were bound to the CPUs on
node-0 to avoid disturbances due to processes being scheduled on different
nodes. The kernels tested are
3.0-rc6-vanilla Vanilla 3.0-rc6
zlcfirst Patch 1 applied
zlcreconsider Patches 1+2 applied
FS-Mark
./fs_mark -d /tmp/fsmark-10813 -D 100 -N 5000 -n 208 -L 35 -t 24 -S0 -s 524288
fsmark-3.0-rc6 3.0-rc6 3.0-rc6
vanilla zlcfirs zlcreconsider
Files/s min 54.90 ( 0.00%) 49.80 (-10.24%) 49.10 (-11.81%)
Files/s mean 100.11 ( 0.00%) 135.17 (25.94%) 146.93 (31.87%)
Files/s stddev 57.51 ( 0.00%) 138.97 (58.62%) 158.69 (63.76%)
Files/s max 361.10 ( 0.00%) 834.40 (56.72%) 802.40 (55.00%)
Overhead min 76704.00 ( 0.00%) 76501.00 ( 0.27%) 77784.00 (-1.39%)
Overhead mean 1485356.51 ( 0.00%) 1035797.83 (43.40%) 1594680.26 (-6.86%)
Overhead stddev 1848122.53 ( 0.00%) 881489.88 (109.66%) 1772354.90 ( 4.27%)
Overhead max 7989060.00 ( 0.00%) 3369118.00 (137.13%) 10135324.00 (-21.18%)
MMTests Statistics: duration
User/Sys Time Running Test (seconds) 501.49 493.91 499.93
Total Elapsed Time (seconds) 2451.57 2257.48 2215.92
MMTests Statistics: vmstat
Page Ins 46268 63840 66008
Page Outs 90821596 90671128 88043732
Swap Ins 0 0 0
Swap Outs 0 0 0
Direct pages scanned 13091697 8966863 8971790
Kswapd pages scanned 0 1830011 1831116
Kswapd pages reclaimed 0 1829068 1829930
Direct pages reclaimed 13037777 8956828 8648314
Kswapd efficiency 100% 99% 99%
Kswapd velocity 0.000 810.643 826.346
Direct efficiency 99% 99% 96%
Direct velocity 5340.128 3972.068 4048.788
Percentage direct scans 100% 83% 83%
Page writes by reclaim 0 3 0
Slabs scanned 796672 720640 720256
Direct inode steals 7422667 7160012 7088638
Kswapd inode steals 0 1736840 2021238
Test completes far faster with a large increase in the number of files
created per second. Standard deviation is high as a small number of
iterations were much higher than the mean. The number of pages scanned by
zone_reclaim is reduced and kswapd is used for more work.
LARGE DD
3.0-rc6 3.0-rc6 3.0-rc6
vanilla zlcfirst zlcreconsider
download tar 59 ( 0.00%) 59 ( 0.00%) 55 ( 7.27%)
dd source files 527 ( 0.00%) 296 (78.04%) 320 (64.69%)
delete source 36 ( 0.00%) 19 (89.47%) 20 (80.00%)
MMTests Statistics: duration
User/Sys Time Running Test (seconds) 125.03 118.98 122.01
Total Elapsed Time (seconds) 624.56 375.02 398.06
MMTests Statistics: vmstat
Page Ins 3594216 439368 407032
Page Outs 23380832 23380488 23377444
Swap Ins 0 0 0
Swap Outs 0 436 287
Direct pages scanned 17482342 69315973 82864918
Kswapd pages scanned 0 519123 575425
Kswapd pages reclaimed 0 466501 522487
Direct pages reclaimed 5858054 2732949 2712547
Kswapd efficiency 100% 89% 90%
Kswapd velocity 0.000 1384.254 1445.574
Direct efficiency 33% 3% 3%
Direct velocity 27991.453 184832.737 208171.929
Percentage direct scans 100% 99% 99%
Page writes by reclaim 0 5082 13917
Slabs scanned 17280 29952 35328
Direct inode steals 115257 1431122 332201
Kswapd inode steals 0 0 979532
This test downloads a large tarfile and copies it with dd a number of
times - similar to the most recent bug report I've dealt with. Time to
completion is reduced. The number of pages scanned directly is still
disturbingly high with a low efficiency but this is likely due to the
number of dirty pages encountered. The figures could probably be improved
with more work around how kswapd is used and how dirty pages are handled
but that is separate work and this result is significant on its own.
Streaming Mapped Writer
MMTests Statistics: duration
User/Sys Time Running Test (seconds) 124.47 111.67 112.64
Total Elapsed Time (seconds) 2138.14 1816.30 1867.56
MMTests Statistics: vmstat
Page Ins 90760 89124 89516
Page Outs 121028340 120199524 120736696
Swap Ins 0 86 55
Swap Outs 0 0 0
Direct pages scanned 114989363 96461439 96330619
Kswapd pages scanned 56430948 56965763 57075875
Kswapd pages reclaimed 27743219 27752044 27766606
Direct pages reclaimed 49777 46884 36655
Kswapd efficiency 49% 48% 48%
Kswapd velocity 26392.541 31363.631 30561.736
Direct efficiency 0% 0% 0%
Direct velocity 53780.091 53108.759 51581.004
Percentage direct scans 67% 62% 62%
Page writes by reclaim 385 122 1513
Slabs scanned 43008 39040 42112
Direct inode steals 0 10 8
Kswapd inode steals 733 534 477
This test just creates a large file mapping and writes to it linearly.
Time to completion is again reduced.
The gains are mostly down to two things. In many cases, there is less
scanning as zone_reclaim simply gives up faster due to recent failures.
The second reason is that memory is used more efficiently. Instead of
scanning the preferred zone every time, the allocator falls back to
another zone and uses it instead improving overall memory utilisation.
This patch: initialise ZLC for first zone eligible for zone_reclaim.
The zonelist cache (ZLC) is used among other things to record if
zone_reclaim() failed for a particular zone recently. The intention is to
avoid a high cost scanning extremely long zonelists or scanning within the
zone uselessly.
Currently the zonelist cache is setup only after the first zone has been
considered and zone_reclaim() has been called. The objective was to avoid
a costly setup but zone_reclaim is itself quite expensive. If it is
failing regularly such as the first eligible zone having mostly mapped
pages, the cost in scanning and allocation stalls is far higher than the
ZLC initialisation step.
This patch initialises ZLC before the first eligible zone calls
zone_reclaim(). Once initialised, it is checked whether the zone failed
zone_reclaim recently. If it has, the zone is skipped. As the first zone
is now being checked, additional care has to be taken about zones marked
full. A zone can be marked "full" because it should not have enough
unmapped pages for zone_reclaim but this is excessive as direct reclaim or
kswapd may succeed where zone_reclaim fails. Only mark zones "full" after
zone_reclaim fails if it failed to reclaim enough pages after scanning.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Christoph Lameter <cl@linux.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-07-25 17:12:29 -07:00
|
|
|
|
2016-07-28 15:46:32 -07:00
|
|
|
|
2009-06-16 15:33:22 -07:00
|
|
|
|
mm: page allocator: initialise ZLC for first zone eligible for zone_reclaim
There have been a small number of complaints about significant stalls
while copying large amounts of data on NUMA machines reported on a
distribution bugzilla. In these cases, zone_reclaim was enabled by
default due to large NUMA distances. In general, the complaints have not
been about the workload itself unless it was a file server (in which case
the recommendation was disable zone_reclaim).
The stalls are mostly due to significant amounts of time spent scanning
the preferred zone for pages to free. After a failure, it might fallback
to another node (as zonelists are often node-ordered rather than
zone-ordered) but stall quickly again when the next allocation attempt
occurs. In bad cases, each page allocated results in a full scan of the
preferred zone.
Patch 1 checks the preferred zone for recent allocation failure
which is particularly important if zone_reclaim has failed
recently. This avoids rescanning the zone in the near future and
instead falling back to another node. This may hurt node locality
in some cases but a failure to zone_reclaim is more expensive than
a remote access.
Patch 2 clears the zlc information after direct reclaim.
Otherwise, zone_reclaim can mark zones full, direct reclaim can
reclaim enough pages but the zone is still not considered for
allocation.
This was tested on a 24-thread 2-node x86_64 machine. The tests were
focused on large amounts of IO. All tests were bound to the CPUs on
node-0 to avoid disturbances due to processes being scheduled on different
nodes. The kernels tested are
3.0-rc6-vanilla Vanilla 3.0-rc6
zlcfirst Patch 1 applied
zlcreconsider Patches 1+2 applied
FS-Mark
./fs_mark -d /tmp/fsmark-10813 -D 100 -N 5000 -n 208 -L 35 -t 24 -S0 -s 524288
fsmark-3.0-rc6 3.0-rc6 3.0-rc6
vanilla zlcfirs zlcreconsider
Files/s min 54.90 ( 0.00%) 49.80 (-10.24%) 49.10 (-11.81%)
Files/s mean 100.11 ( 0.00%) 135.17 (25.94%) 146.93 (31.87%)
Files/s stddev 57.51 ( 0.00%) 138.97 (58.62%) 158.69 (63.76%)
Files/s max 361.10 ( 0.00%) 834.40 (56.72%) 802.40 (55.00%)
Overhead min 76704.00 ( 0.00%) 76501.00 ( 0.27%) 77784.00 (-1.39%)
Overhead mean 1485356.51 ( 0.00%) 1035797.83 (43.40%) 1594680.26 (-6.86%)
Overhead stddev 1848122.53 ( 0.00%) 881489.88 (109.66%) 1772354.90 ( 4.27%)
Overhead max 7989060.00 ( 0.00%) 3369118.00 (137.13%) 10135324.00 (-21.18%)
MMTests Statistics: duration
User/Sys Time Running Test (seconds) 501.49 493.91 499.93
Total Elapsed Time (seconds) 2451.57 2257.48 2215.92
MMTests Statistics: vmstat
Page Ins 46268 63840 66008
Page Outs 90821596 90671128 88043732
Swap Ins 0 0 0
Swap Outs 0 0 0
Direct pages scanned 13091697 8966863 8971790
Kswapd pages scanned 0 1830011 1831116
Kswapd pages reclaimed 0 1829068 1829930
Direct pages reclaimed 13037777 8956828 8648314
Kswapd efficiency 100% 99% 99%
Kswapd velocity 0.000 810.643 826.346
Direct efficiency 99% 99% 96%
Direct velocity 5340.128 3972.068 4048.788
Percentage direct scans 100% 83% 83%
Page writes by reclaim 0 3 0
Slabs scanned 796672 720640 720256
Direct inode steals 7422667 7160012 7088638
Kswapd inode steals 0 1736840 2021238
Test completes far faster with a large increase in the number of files
created per second. Standard deviation is high as a small number of
iterations were much higher than the mean. The number of pages scanned by
zone_reclaim is reduced and kswapd is used for more work.
LARGE DD
3.0-rc6 3.0-rc6 3.0-rc6
vanilla zlcfirst zlcreconsider
download tar 59 ( 0.00%) 59 ( 0.00%) 55 ( 7.27%)
dd source files 527 ( 0.00%) 296 (78.04%) 320 (64.69%)
delete source 36 ( 0.00%) 19 (89.47%) 20 (80.00%)
MMTests Statistics: duration
User/Sys Time Running Test (seconds) 125.03 118.98 122.01
Total Elapsed Time (seconds) 624.56 375.02 398.06
MMTests Statistics: vmstat
Page Ins 3594216 439368 407032
Page Outs 23380832 23380488 23377444
Swap Ins 0 0 0
Swap Outs 0 436 287
Direct pages scanned 17482342 69315973 82864918
Kswapd pages scanned 0 519123 575425
Kswapd pages reclaimed 0 466501 522487
Direct pages reclaimed 5858054 2732949 2712547
Kswapd efficiency 100% 89% 90%
Kswapd velocity 0.000 1384.254 1445.574
Direct efficiency 33% 3% 3%
Direct velocity 27991.453 184832.737 208171.929
Percentage direct scans 100% 99% 99%
Page writes by reclaim 0 5082 13917
Slabs scanned 17280 29952 35328
Direct inode steals 115257 1431122 332201
Kswapd inode steals 0 0 979532
This test downloads a large tarfile and copies it with dd a number of
times - similar to the most recent bug report I've dealt with. Time to
completion is reduced. The number of pages scanned directly is still
disturbingly high with a low efficiency but this is likely due to the
number of dirty pages encountered. The figures could probably be improved
with more work around how kswapd is used and how dirty pages are handled
but that is separate work and this result is significant on its own.
Streaming Mapped Writer
MMTests Statistics: duration
User/Sys Time Running Test (seconds) 124.47 111.67 112.64
Total Elapsed Time (seconds) 2138.14 1816.30 1867.56
MMTests Statistics: vmstat
Page Ins 90760 89124 89516
Page Outs 121028340 120199524 120736696
Swap Ins 0 86 55
Swap Outs 0 0 0
Direct pages scanned 114989363 96461439 96330619
Kswapd pages scanned 56430948 56965763 57075875
Kswapd pages reclaimed 27743219 27752044 27766606
Direct pages reclaimed 49777 46884 36655
Kswapd efficiency 49% 48% 48%
Kswapd velocity 26392.541 31363.631 30561.736
Direct efficiency 0% 0% 0%
Direct velocity 53780.091 53108.759 51581.004
Percentage direct scans 67% 62% 62%
Page writes by reclaim 385 122 1513
Slabs scanned 43008 39040 42112
Direct inode steals 0 10 8
Kswapd inode steals 733 534 477
This test just creates a large file mapping and writes to it linearly.
Time to completion is again reduced.
The gains are mostly down to two things. In many cases, there is less
scanning as zone_reclaim simply gives up faster due to recent failures.
The second reason is that memory is used more efficiently. Instead of
scanning the preferred zone every time, the allocator falls back to
another zone and uses it instead improving overall memory utilisation.
This patch: initialise ZLC for first zone eligible for zone_reclaim.
The zonelist cache (ZLC) is used among other things to record if
zone_reclaim() failed for a particular zone recently. The intention is to
avoid a high cost scanning extremely long zonelists or scanning within the
zone uselessly.
Currently the zonelist cache is setup only after the first zone has been
considered and zone_reclaim() has been called. The objective was to avoid
a costly setup but zone_reclaim is itself quite expensive. If it is
failing regularly such as the first eligible zone having mostly mapped
pages, the cost in scanning and allocation stalls is far higher than the
ZLC initialisation step.
This patch initialises ZLC before the first eligible zone calls
zone_reclaim(). Once initialised, it is checked whether the zone failed
zone_reclaim recently. If it has, the zone is skipped. As the first zone
is now being checked, additional care has to be taken about zones marked
full. A zone can be marked "full" because it should not have enough
unmapped pages for zone_reclaim but this is excessive as direct reclaim or
kswapd may succeed where zone_reclaim fails. Only mark zones "full" after
zone_reclaim fails if it failed to reclaim enough pages after scanning.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Christoph Lameter <cl@linux.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-07-25 17:12:29 -07:00
|
|
|
|
2009-06-16 15:33:22 -07:00
|
|
|
|
|
|
|
|
|
2013-04-29 15:07:57 -07:00
|
|
|
|
2020-06-03 15:59:01 -07:00
|
|
|
|
2013-04-29 15:07:57 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-12-06 20:31:38 -08:00
|
|
|
|
2005-11-13 16:06:43 -08:00
|
|
|
|
|
|
|
|
|
2009-06-16 15:33:22 -07:00
|
|
|
|
2024-07-29 14:47:17 +05:30
|
|
|
|
2015-11-06 16:28:37 -08:00
|
|
|
|
mm: set page->pfmemalloc in prep_new_page()
The possibility of replacing the numerous parameters of alloc_pages*
functions with a single structure has been discussed when Minchan proposed
to expand the x86 kernel stack [1]. This series implements the change,
along with few more cleanups/microoptimizations.
The series is based on next-20150108 and I used gcc 4.8.3 20140627 on
openSUSE 13.2 for compiling. Config includess NUMA and COMPACTION.
The core change is the introduction of a new struct alloc_context, which looks
like this:
struct alloc_context {
struct zonelist *zonelist;
nodemask_t *nodemask;
struct zone *preferred_zone;
int classzone_idx;
int migratetype;
enum zone_type high_zoneidx;
};
All the contents is mostly constant, except that __alloc_pages_slowpath()
changes preferred_zone, classzone_idx and potentially zonelist. But
that's not a problem in case control returns to retry_cpuset: in
__alloc_pages_nodemask(), those will be reset to initial values again
(although it's a bit subtle). On the other hand, gfp_flags and alloc_info
mutate so much that it doesn't make sense to put them into alloc_context.
Still, the result is one parameter instead of up to 7. This is all in
Patch 2.
Patch 3 is a step to expand alloc_context usage out of page_alloc.c
itself. The function try_to_compact_pages() can also much benefit from
the parameter reduction, but it means the struct definition has to be
moved to a shared header.
Patch 1 should IMHO be included even if the rest is deemed not useful
enough. It improves maintainability and also has some code/stack
reduction. Patch 4 is OTOH a tiny optimization.
Overall bloat-o-meter results:
add/remove: 0/0 grow/shrink: 0/4 up/down: 0/-460 (-460)
function old new delta
nr_free_zone_pages 129 115 -14
__alloc_pages_direct_compact 329 256 -73
get_page_from_freelist 2670 2576 -94
__alloc_pages_nodemask 2564 2285 -279
try_to_compact_pages 582 579 -3
Overall stack sizes per ./scripts/checkstack.pl:
old new delta
get_page_from_freelist: 184 184 0
__alloc_pages_nodemask 248 200 -48
__alloc_pages_direct_c 40 - -40
try_to_compact_pages 72 72 0
-88
[1] http://marc.info/?l=linux-mm&m=140142462528257&w=2
This patch (of 4):
prep_new_page() sets almost everything in the struct page of the page
being allocated, except page->pfmemalloc. This is not obvious and has at
least once led to a bug where page->pfmemalloc was forgotten to be set
correctly, see commit 8fb74b9fb2b1 ("mm: compaction: partially revert
capture of suitable high-order page").
This patch moves the pfmemalloc setting to prep_new_page(), which means it
needs to gain alloc_flags parameter. The call to prep_new_page is moved
from buffered_rmqueue() to get_page_from_freelist(), which also leads to
simpler code. An obsolete comment for buffered_rmqueue() is replaced.
In addition to better maintainability there is a small reduction of code
and stack usage for get_page_from_freelist(), which inlines the other
functions involved.
add/remove: 0/0 grow/shrink: 0/1 up/down: 0/-145 (-145)
function old new delta
get_page_from_freelist 2670 2525 -145
Stack usage is reduced from 184 to 168 bytes.
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:25:38 -08:00
|
|
|
|
2016-05-19 17:14:35 -07:00
|
|
|
|
2015-11-06 16:28:37 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-01-13 11:12:14 +00:00
|
|
|
|
mm: page_alloc: fix highatomic typing in multi-block buddies
Christoph reports a page allocator splat triggered by xfstests:
generic/176 214s ... [ 1204.507931] run fstests generic/176 at 2024-05-27 12:52:30
XFS (nvme0n1): Mounting V5 Filesystem cd936307-415f-48a3-b99d-a2d52ae1f273
XFS (nvme0n1): Ending clean mount
XFS (nvme1n1): Mounting V5 Filesystem ab3ee1a4-af62-4934-9a6a-6c2fde321850
XFS (nvme1n1): Ending clean mount
XFS (nvme1n1): Unmounting Filesystem ab3ee1a4-af62-4934-9a6a-6c2fde321850
XFS (nvme1n1): Mounting V5 Filesystem 7099b02d-9c58-4d1d-be1d-2cc472d12cd9
XFS (nvme1n1): Ending clean mount
------------[ cut here ]------------
page type is 3, passed migratetype is 1 (nr=512)
WARNING: CPU: 0 PID: 509870 at mm/page_alloc.c:645 expand+0x1c5/0x1f0
Modules linked in: i2c_i801 crc32_pclmul i2c_smbus [last unloaded: scsi_debug]
CPU: 0 PID: 509870 Comm: xfs_io Not tainted 6.10.0-rc1+ #2437
Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.16.3-debian-1.16.3-2 04/01/2014
RIP: 0010:expand+0x1c5/0x1f0
Code: 05 16 70 bf 02 01 e8 ca fc ff ff 8b 54 24 34 44 89 e1 48 c7 c7 80 a2 28 83 48 89 c6 b8 01 00 3
RSP: 0018:ffffc90003b2b968 EFLAGS: 00010082
RAX: 0000000000000000 RBX: ffffffff83fa9480 RCX: 0000000000000000
RDX: 0000000000000005 RSI: 0000000000000027 RDI: 00000000ffffffff
RBP: 00000000001f2600 R08: 00000000fffeffff R09: 0000000000000001
R10: 0000000000000000 R11: ffffffff83676200 R12: 0000000000000009
R13: 0000000000000200 R14: 0000000000000001 R15: ffffea0007c98000
FS: 00007f72ca3d5780(0000) GS:ffff8881f9c00000(0000) knlGS:0000000000000000
CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 00007f72ca1fff38 CR3: 00000001aa0c6002 CR4: 0000000000770ef0
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000ffff07f0 DR7: 0000000000000400
PKRU: 55555554
Call Trace:
<TASK>
? __warn+0x7b/0x120
? expand+0x1c5/0x1f0
? report_bug+0x191/0x1c0
? handle_bug+0x3c/0x80
? exc_invalid_op+0x17/0x70
? asm_exc_invalid_op+0x1a/0x20
? expand+0x1c5/0x1f0
? expand+0x1c5/0x1f0
__rmqueue_pcplist+0x3a9/0x730
get_page_from_freelist+0x7a0/0xf00
__alloc_pages_noprof+0x153/0x2e0
__folio_alloc_noprof+0x10/0xa0
__filemap_get_folio+0x16b/0x370
iomap_write_begin+0x496/0x680
While trying to service a movable allocation (page type 1), the page
allocator runs into a two-pageblock buddy on the movable freelist whose
second block is typed as highatomic (page type 3).
This inconsistency is caused by the highatomic reservation system
operating on single pageblocks, while MAX_ORDER can be bigger than that -
in this configuration, pageblock_order is 9 while MAX_PAGE_ORDER is 10.
The test case is observed to make several adjacent order-3 requests with
__GFP_DIRECT_RECLAIM cleared, which marks the surrounding block as
highatomic. Upon freeing, the blocks merge into an order-10 buddy. When
the highatomic pool is drained later on, this order-10 buddy gets moved
back to the movable list, but only the first pageblock is marked movable
again. A subsequent expand() of this buddy warns about the tail being of
a different type.
This is a long-standing bug that's surfaced by the recent block type
warnings added to the allocator. The consequences seem mostly benign, it
just results in odd behavior: the highatomic tail blocks are not properly
drained, instead they end up on the movable list first, then go back to
the highatomic list after an alloc-free cycle.
To fix this, make the highatomic reservation code aware that
allocations/buddies can be larger than a pageblock.
While it's an old quirk, the recently added type consistency warnings seem
to be the most prominent consequence of it. Set the Fixes: tag
accordingly to highlight this backporting dependency.
Link: https://lkml.kernel.org/r/20240530114203.GA1222079@cmpxchg.org
Fixes: e0932b6c1f94 ("mm: page_alloc: consolidate free page accounting")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reported-by: Christoph Hellwig <hch@infradead.org>
Reviewed-by: Zi Yan <ziy@nvidia.com>
Tested-by: Christoph Hellwig <hch@lst.de>
Cc: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-05-29 18:18:12 -04:00
|
|
|
|
2015-11-06 16:28:37 -08:00
|
|
|
|
mm: set page->pfmemalloc in prep_new_page()
The possibility of replacing the numerous parameters of alloc_pages*
functions with a single structure has been discussed when Minchan proposed
to expand the x86 kernel stack [1]. This series implements the change,
along with few more cleanups/microoptimizations.
The series is based on next-20150108 and I used gcc 4.8.3 20140627 on
openSUSE 13.2 for compiling. Config includess NUMA and COMPACTION.
The core change is the introduction of a new struct alloc_context, which looks
like this:
struct alloc_context {
struct zonelist *zonelist;
nodemask_t *nodemask;
struct zone *preferred_zone;
int classzone_idx;
int migratetype;
enum zone_type high_zoneidx;
};
All the contents is mostly constant, except that __alloc_pages_slowpath()
changes preferred_zone, classzone_idx and potentially zonelist. But
that's not a problem in case control returns to retry_cpuset: in
__alloc_pages_nodemask(), those will be reset to initial values again
(although it's a bit subtle). On the other hand, gfp_flags and alloc_info
mutate so much that it doesn't make sense to put them into alloc_context.
Still, the result is one parameter instead of up to 7. This is all in
Patch 2.
Patch 3 is a step to expand alloc_context usage out of page_alloc.c
itself. The function try_to_compact_pages() can also much benefit from
the parameter reduction, but it means the struct definition has to be
moved to a shared header.
Patch 1 should IMHO be included even if the rest is deemed not useful
enough. It improves maintainability and also has some code/stack
reduction. Patch 4 is OTOH a tiny optimization.
Overall bloat-o-meter results:
add/remove: 0/0 grow/shrink: 0/4 up/down: 0/-460 (-460)
function old new delta
nr_free_zone_pages 129 115 -14
__alloc_pages_direct_compact 329 256 -73
get_page_from_freelist 2670 2576 -94
__alloc_pages_nodemask 2564 2285 -279
try_to_compact_pages 582 579 -3
Overall stack sizes per ./scripts/checkstack.pl:
old new delta
get_page_from_freelist: 184 184 0
__alloc_pages_nodemask 248 200 -48
__alloc_pages_direct_c 40 - -40
try_to_compact_pages 72 72 0
-88
[1] http://marc.info/?l=linux-mm&m=140142462528257&w=2
This patch (of 4):
prep_new_page() sets almost everything in the struct page of the page
being allocated, except page->pfmemalloc. This is not obvious and has at
least once led to a bug where page->pfmemalloc was forgotten to be set
correctly, see commit 8fb74b9fb2b1 ("mm: compaction: partially revert
capture of suitable high-order page").
This patch moves the pfmemalloc setting to prep_new_page(), which means it
needs to gain alloc_flags parameter. The call to prep_new_page is moved
from buffered_rmqueue() to get_page_from_freelist(), which also leads to
simpler code. An obsolete comment for buffered_rmqueue() is replaced.
In addition to better maintainability there is a small reduction of code
and stack usage for get_page_from_freelist(), which inlines the other
functions involved.
add/remove: 0/0 grow/shrink: 0/1 up/down: 0/-145 (-145)
function old new delta
get_page_from_freelist 2670 2525 -145
Stack usage is reduced from 184 to 168 bytes.
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:25:38 -08:00
|
|
|
|
2018-04-05 16:22:31 -07:00
|
|
|
|
2025-05-06 14:25:08 +03:00
|
|
|
|
2024-08-09 14:48:47 +03:00
|
|
|
|
2023-06-06 17:26:29 +03:00
|
|
|
|
2018-04-05 16:22:31 -07:00
|
|
|
|
2023-01-05 13:55:06 +05:30
|
|
|
|
2018-04-05 16:22:31 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: set page->pfmemalloc in prep_new_page()
The possibility of replacing the numerous parameters of alloc_pages*
functions with a single structure has been discussed when Minchan proposed
to expand the x86 kernel stack [1]. This series implements the change,
along with few more cleanups/microoptimizations.
The series is based on next-20150108 and I used gcc 4.8.3 20140627 on
openSUSE 13.2 for compiling. Config includess NUMA and COMPACTION.
The core change is the introduction of a new struct alloc_context, which looks
like this:
struct alloc_context {
struct zonelist *zonelist;
nodemask_t *nodemask;
struct zone *preferred_zone;
int classzone_idx;
int migratetype;
enum zone_type high_zoneidx;
};
All the contents is mostly constant, except that __alloc_pages_slowpath()
changes preferred_zone, classzone_idx and potentially zonelist. But
that's not a problem in case control returns to retry_cpuset: in
__alloc_pages_nodemask(), those will be reset to initial values again
(although it's a bit subtle). On the other hand, gfp_flags and alloc_info
mutate so much that it doesn't make sense to put them into alloc_context.
Still, the result is one parameter instead of up to 7. This is all in
Patch 2.
Patch 3 is a step to expand alloc_context usage out of page_alloc.c
itself. The function try_to_compact_pages() can also much benefit from
the parameter reduction, but it means the struct definition has to be
moved to a shared header.
Patch 1 should IMHO be included even if the rest is deemed not useful
enough. It improves maintainability and also has some code/stack
reduction. Patch 4 is OTOH a tiny optimization.
Overall bloat-o-meter results:
add/remove: 0/0 grow/shrink: 0/4 up/down: 0/-460 (-460)
function old new delta
nr_free_zone_pages 129 115 -14
__alloc_pages_direct_compact 329 256 -73
get_page_from_freelist 2670 2576 -94
__alloc_pages_nodemask 2564 2285 -279
try_to_compact_pages 582 579 -3
Overall stack sizes per ./scripts/checkstack.pl:
old new delta
get_page_from_freelist: 184 184 0
__alloc_pages_nodemask 248 200 -48
__alloc_pages_direct_c 40 - -40
try_to_compact_pages 72 72 0
-88
[1] http://marc.info/?l=linux-mm&m=140142462528257&w=2
This patch (of 4):
prep_new_page() sets almost everything in the struct page of the page
being allocated, except page->pfmemalloc. This is not obvious and has at
least once led to a bug where page->pfmemalloc was forgotten to be set
correctly, see commit 8fb74b9fb2b1 ("mm: compaction: partially revert
capture of suitable high-order page").
This patch moves the pfmemalloc setting to prep_new_page(), which means it
needs to gain alloc_flags parameter. The call to prep_new_page is moved
from buffered_rmqueue() to get_page_from_freelist(), which also leads to
simpler code. An obsolete comment for buffered_rmqueue() is replaced.
In addition to better maintainability there is a small reduction of code
and stack usage for get_page_from_freelist(), which inlines the other
functions involved.
add/remove: 0/0 grow/shrink: 0/1 up/down: 0/-145 (-145)
function old new delta
get_page_from_freelist 2670 2525 -145
Stack usage is reduced from 184 to 168 bytes.
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:25:38 -08:00
|
|
|
|
2008-04-28 02:12:16 -07:00
|
|
|
|
[PATCH] memory page_alloc zonelist caching speedup
Optimize the critical zonelist scanning for free pages in the kernel memory
allocator by caching the zones that were found to be full recently, and
skipping them.
Remembers the zones in a zonelist that were short of free memory in the
last second. And it stashes a zone-to-node table in the zonelist struct,
to optimize that conversion (minimize its cache footprint.)
Recent changes:
This differs in a significant way from a similar patch that I
posted a week ago. Now, instead of having a nodemask_t of
recently full nodes, I have a bitmask of recently full zones.
This solves a problem that last weeks patch had, which on
systems with multiple zones per node (such as DMA zone) would
take seeing any of these zones full as meaning that all zones
on that node were full.
Also I changed names - from "zonelist faster" to "zonelist cache",
as that seemed to better convey what we're doing here - caching
some of the key zonelist state (for faster access.)
See below for some performance benchmark results. After all that
discussion with David on why I didn't need them, I went and got
some ;). I wanted to verify that I had not hurt the normal case
of memory allocation noticeably. At least for my one little
microbenchmark, I found (1) the normal case wasn't affected, and
(2) workloads that forced scanning across multiple nodes for
memory improved up to 10% fewer System CPU cycles and lower
elapsed clock time ('sys' and 'real'). Good. See details, below.
I didn't have the logic in get_page_from_freelist() for various
full nodes and zone reclaim failures correct. That should be
fixed up now - notice the new goto labels zonelist_scan,
this_zone_full, and try_next_zone, in get_page_from_freelist().
There are two reasons I persued this alternative, over some earlier
proposals that would have focused on optimizing the fake numa
emulation case by caching the last useful zone:
1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems)
have seen real customer loads where the cost to scan the zonelist
was a problem, due to many nodes being full of memory before
we got to a node we could use. Or at least, I think we have.
This was related to me by another engineer, based on experiences
from some time past. So this is not guaranteed. Most likely, though.
The following approach should help such real numa systems just as
much as it helps fake numa systems, or any combination thereof.
2) The effort to distinguish fake from real numa, using node_distance,
so that we could cache a fake numa node and optimize choosing
it over equivalent distance fake nodes, while continuing to
properly scan all real nodes in distance order, was going to
require a nasty blob of zonelist and node distance munging.
The following approach has no new dependency on node distances or
zone sorting.
See comment in the patch below for a description of what it actually does.
Technical details of note (or controversy):
- See the use of "zlc_active" and "did_zlc_setup" below, to delay
adding any work for this new mechanism until we've looked at the
first zone in zonelist. I figured the odds of the first zone
having the memory we needed were high enough that we should just
look there, first, then get fancy only if we need to keep looking.
- Some odd hackery was needed to add items to struct zonelist, while
not tripping up the custom zonelists built by the mm/mempolicy.c
code for MPOL_BIND. My usual wordy comments below explain this.
Search for "MPOL_BIND".
- Some per-node data in the struct zonelist is now modified frequently,
with no locking. Multiple CPU cores on a node could hit and mangle
this data. The theory is that this is just performance hint data,
and the memory allocator will work just fine despite any such mangling.
The fields at risk are the struct 'zonelist_cache' fields 'fullzones'
(a bitmask) and 'last_full_zap' (unsigned long jiffies). It should
all be self correcting after at most a one second delay.
- This still does a linear scan of the same lengths as before. All
I've optimized is making the scan faster, not algorithmically
shorter. It is now able to scan a compact array of 'unsigned
short' in the case of many full nodes, so one cache line should
cover quite a few nodes, rather than each node hitting another
one or two new and distinct cache lines.
- If both Andi and Nick don't find this too complicated, I will be
(pleasantly) flabbergasted.
- I removed the comment claiming we only use one cachline's worth of
zonelist. We seem, at least in the fake numa case, to have put the
lie to that claim.
- I pay no attention to the various watermarks and such in this performance
hint. A node could be marked full for one watermark, and then skipped
over when searching for a page using a different watermark. I think
that's actually quite ok, as it will tend to slightly increase the
spreading of memory over other nodes, away from a memory stressed node.
===============
Performance - some benchmark results and analysis:
This benchmark runs a memory hog program that uses multiple
threads to touch alot of memory as quickly as it can.
Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of
the total 96 GBytes on the system, and using 1, 19, 37, or 55
threads (on a 56 CPU system.) System, user and real (elapsed)
timings were recorded for each run, shown in units of seconds,
in the table below.
Two kernels were tested - 2.6.18-mm3 and the same kernel with
this zonelist caching patch added. The table also shows the
percentage improvement the zonelist caching sys time is over
(lower than) the stock *-mm kernel.
number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent
GBs N ------------ -------------- ---------------- systime
mem threads sys user real sys user real sys user real better
12 1 153 24 177 151 24 176 -2 0 -1 1%
12 19 99 22 8 99 22 8 0 0 0 0%
12 37 111 25 6 112 25 6 1 0 0 -0%
12 55 115 25 5 110 23 5 -5 -2 0 4%
38 1 502 74 576 497 73 570 -5 -1 -6 0%
38 19 426 78 48 373 76 39 -53 -2 -9 12%
38 37 544 83 36 547 82 36 3 -1 0 -0%
38 55 501 77 23 511 80 24 10 3 1 -1%
64 1 917 125 1042 890 124 1014 -27 -1 -28 2%
64 19 1118 138 119 965 141 103 -153 3 -16 13%
64 37 1202 151 94 1136 150 81 -66 -1 -13 5%
64 55 1118 141 61 1072 140 58 -46 -1 -3 4%
90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4%
90 19 2392 199 192 2116 189 176 -276 -10 -16 11%
90 37 3313 238 175 2972 225 145 -341 -13 -30 10%
90 55 1948 210 104 1843 213 100 -105 3 -4 5%
Notes:
1) This test ran a memory hog program that started a specified number N of
threads, and had each thread allocate and touch 1/N'th of
the total memory to be used in the test run in a single loop,
writing a constant word to memory, one store every 4096 bytes.
Watching this test during some earlier trial runs, I would see
each of these threads sit down on one CPU and stay there, for
the remainder of the pass, a different CPU for each thread.
2) The 'real' column is not comparable to the 'sys' or 'user' columns.
The 'real' column is seconds wall clock time elapsed, from beginning
to end of that test pass. The 'sys' and 'user' columns are total
CPU seconds spent on that test pass. For a 19 thread test run,
for example, the sum of 'sys' and 'user' could be up to 19 times the
number of 'real' elapsed wall clock seconds.
3) Tests were run on a fresh, single-user boot, to minimize the amount
of memory already in use at the start of the test, and to minimize
the amount of background activity that might interfere.
4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM.
5) Notice that the 'real' time gets large for the single thread runs, even
though the measured 'sys' and 'user' times are modest. I'm not sure what
that means - probably something to do with it being slow for one thread to
be accessing memory along ways away. Perhaps the fake numa system, running
ostensibly the same workload, would not show this substantial degradation
of 'real' time for one thread on many nodes -- lets hope not.
6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs)
ran quite efficiently, as one might expect. Each pair of threads needed
to allocate and touch the memory on the node the two threads shared, a
pleasantly parallizable workload.
7) The intermediate thread count passes, when asking for alot of memory forcing
them to go to a few neighboring nodes, improved the most with this zonelist
caching patch.
Conclusions:
* This zonelist cache patch probably makes little difference one way or the
other for most workloads on real numa hardware, if those workloads avoid
heavy off node allocations.
* For memory intensive workloads requiring substantial off-node allocations
on real numa hardware, this patch improves both kernel and elapsed timings
up to ten per-cent.
* For fake numa systems, I'm optimistic, but will have to leave that up to
Rohit Seth to actually test (once I get him a 2.6.18 backport.)
Signed-off-by: Paul Jackson <pj@sgi.com>
Cc: Rohit Seth <rohitseth@google.com>
Cc: Christoph Lameter <clameter@engr.sgi.com>
Cc: David Rientjes <rientjes@cs.washington.edu>
Cc: Paul Menage <menage@google.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-06 20:31:48 -08:00
|
|
|
|
mm: page_alloc: avoid kswapd thrashing due to NUMA restrictions
On NUMA systems without bindings, allocations check all nodes for free
space, then wake up the kswapds on all nodes and retry. This ensures
all available space is evenly used before reclaim begins. However,
when one process or certain allocations have node restrictions, they
can cause kswapds on only a subset of nodes to be woken up.
Since kswapd hysteresis targets watermarks that are *higher* than
needed for allocation, even *unrestricted* allocations can now get
suckered onto such nodes that are already pressured. This ends up
concentrating all allocations on them, even when there are idle nodes
available for the unrestricted requests.
This was observed with two numa nodes, where node0 is normal and node1
is ZONE_MOVABLE to facilitate hotplugging: a kernel allocation wakes
kswapd on node0 only (since node1 is not eligible); once kswapd0 is
active, the watermarks hover between low and high, and then even the
movable allocations end up on node0, only to be kicked out again;
meanwhile node1 is empty and idle.
Similar behavior is possible when a process with NUMA bindings is
causing selective kswapd wakeups.
To fix this, on NUMA systems augment the (misleading) watermark test
with a check for whether kswapd is already active during the first
iteration through the zonelist. If this fails to place the request,
kswapd must be running everywhere already, and the watermark test is
good enough to decide placement.
With this patch, unrestricted requests successfully make use of node1,
even while kswapd is reclaiming node0 for restricted allocations.
[gourry@gourry.net: don't retry if no kswapds were active]
Link: https://lkml.kernel.org/r/20250919162134.1098208-1-hannes@cmpxchg.org
Signed-off-by: Gregory Price <gourry@gourry.net>
Tested-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Zi Yan <ziy@nvidia.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: Joshua Hahn <joshua.hahnjy@gmail.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-09-19 12:21:34 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, page_alloc: spread allocations across zones before introducing fragmentation
Patch series "Fragmentation avoidance improvements", v5.
It has been noted before that fragmentation avoidance (aka
anti-fragmentation) is not perfect. Given sufficient time or an adverse
workload, memory gets fragmented and the long-term success of high-order
allocations degrades. This series defines an adverse workload, a definition
of external fragmentation events (including serious) ones and a series
that reduces the level of those fragmentation events.
The details of the workload and the consequences are described in more
detail in the changelogs. However, from patch 1, this is a high-level
summary of the adverse workload. The exact details are found in the
mmtests implementation.
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch)
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameterr create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed
3. Warm up a number of fio read-only threads accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll fault back in old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup
Overall the series reduces external fragmentation causing events by over 94%
on 1 and 2 socket machines, which in turn impacts high-order allocation
success rates over the long term. There are differences in latencies and
high-order allocation success rates. Latencies are a mixed bag as they
are vulnerable to exact system state and whether allocations succeeded
so they are treated as a secondary metric.
Patch 1 uses lower zones if they are populated and have free memory
instead of fragmenting a higher zone. It's special cased to
handle a Normal->DMA32 fallback with the reasons explained
in the changelog.
Patch 2-4 boosts watermarks temporarily when an external fragmentation
event occurs. kswapd wakes to reclaim a small amount of old memory
and then wakes kcompactd on completion to recover the system
slightly. This introduces some overhead in the slowpath. The level
of boosting can be tuned or disabled depending on the tolerance
for fragmentation vs allocation latency.
Patch 5 stalls some movable allocation requests to let kswapd from patch 4
make some progress. The duration of the stalls is very low but it
is possible to tune the system to avoid fragmentation events if
larger stalls can be tolerated.
The bulk of the improvement in fragmentation avoidance is from patches
1-4 but patch 5 can deal with a rare corner case and provides the option
of tuning a system for THP allocation success rates in exchange for
some stalls to control fragmentation.
This patch (of 5):
The page allocator zone lists are iterated based on the watermarks of each
zone which does not take anti-fragmentation into account. On x86, node 0
may have multiple zones while other nodes have one zone. A consequence is
that tasks running on node 0 may fragment ZONE_NORMAL even though
ZONE_DMA32 has plenty of free memory. This patch special cases the
allocator fast path such that it'll try an allocation from a lower local
zone before fragmenting a higher zone. In this case, stealing of
pageblocks or orders larger than a pageblock are still allowed in the fast
path as they are uninteresting from a fragmentation point of view.
This was evaluated using a benchmark designed to fragment memory before
attempting THP allocations. It's implemented in mmtests as the following
configurations
configs/config-global-dhp__workload_thpfioscale
configs/config-global-dhp__workload_thpfioscale-defrag
configs/config-global-dhp__workload_thpfioscale-madvhugepage
e.g. from mmtests
./run-mmtests.sh --run-monitor --config configs/config-global-dhp__workload_thpfioscale test-run-1
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch).
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameter create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed.
3. Warm up a number of fio read-only processes accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll refault old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds.
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup the test files.
Note that due to the use of IO and page cache that this benchmark is not
suitable for running on large machines where the time to fragment memory
may be excessive. Also note that while this is one mix that generates
fragmentation that it's not the only mix that generates fragmentation.
Differences in workload that are more slab-intensive or whether SLUB is
used with high-order pages may yield different results.
When the page allocator fragments memory, it records the event using the
mm_page_alloc_extfrag ftrace event. If the fallback_order is smaller than
a pageblock order (order-9 on 64-bit x86) then it's considered to be an
"external fragmentation event" that may cause issues in the future.
Hence, the primary metric here is the number of external fragmentation
events that occur with order < 9. The secondary metric is allocation
latency and huge page allocation success rates but note that differences
in latencies and what the success rate also can affect the number of
external fragmentation event which is why it's a secondary metric.
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 662.92 ( 0.00%) 653.58 * 1.41%*
Amean fault-huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
Fault latencies are slightly reduced while allocation success rates remain
at zero as this configuration does not make any special effort to allocate
THP and fio is heavily active at the time and either filling memory or
keeping pages resident. However, a 49% reduction of serious fragmentation
events reduces the changes of external fragmentation being a problem in
the future.
Vlastimil asked during review for a breakdown of the allocation types
that are falling back.
vanilla
3816 MIGRATE_UNMOVABLE
800845 MIGRATE_MOVABLE
33 MIGRATE_UNRECLAIMABLE
patch
735 MIGRATE_UNMOVABLE
408135 MIGRATE_MOVABLE
42 MIGRATE_UNRECLAIMABLE
The majority of the fallbacks are due to movable allocations and this is
consistent for the workload throughout the series so will not be presented
again as the primary source of fallbacks are movable allocations.
Movable fallbacks are sometimes considered "ok" to fallback because they
can be migrated. The problem is that they can fill an
unmovable/reclaimable pageblock causing those allocations to fallback
later and polluting pageblocks with pages that cannot move. If there is a
movable fallback, it is pretty much guaranteed to affect an
unmovable/reclaimable pageblock and while it might not be enough to
actually cause a unmovable/reclaimable fallback in the future, we cannot
know that in advance so the patch takes the only option available to it.
Hence, it's important to control them. This point is also consistent
throughout the series and will not be repeated.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 1495.14 ( 0.00%) 1467.55 ( 1.85%)
Amean fault-huge-1 1098.48 ( 0.00%) 1127.11 ( -2.61%)
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 78.57 ( 0.00%) 77.64 ( -1.18%)
Fragmentation events were reduced quite a bit although this is known
to be a little variable. The latencies and allocation success rates
are similar but they were already quite high.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 1350.05 ( 0.00%) 1346.45 ( 0.27%)
Amean fault-huge-5 4181.01 ( 0.00%) 3418.60 ( 18.24%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 1.15 ( 0.00%) 0.78 ( -31.88%)
The reduction of external fragmentation events is slight and this is
partially due to the removal of __GFP_THISNODE in commit ac5b2c18911f
("mm: thp: relax __GFP_THISNODE for MADV_HUGEPAGE mappings") as THP
allocations can now spill over to remote nodes instead of fragmenting
local memory.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 6138.97 ( 0.00%) 6217.43 ( -1.28%)
Amean fault-huge-5 2294.28 ( 0.00%) 3163.33 * -37.88%*
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 96.82 ( 0.00%) 95.14 ( -1.74%)
There was a slight reduction in external fragmentation events although the
latencies were higher. The allocation success rate is high enough that
the system is struggling and there is quite a lot of parallel reclaim and
compaction activity. There is also a certain degree of luck on whether
processes start on node 0 or not for this patch but the relevance is
reduced later in the series.
Overall, the patch reduces the number of external fragmentation causing
events so the success of THP over long periods of time would be improved
for this adverse workload.
Link: http://lkml.kernel.org/r/20181123114528.28802-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:41 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: page_alloc: defrag_mode
The page allocator groups requests by migratetype to stave off
fragmentation. However, in practice this is routinely defeated by the
fact that it gives up *before* invoking reclaim and compaction - which may
well produce suitable pages. As a result, fragmentation of physical
memory is a common ongoing process in many load scenarios.
Fragmentation deteriorates compaction's ability to produce huge pages.
Depending on the lifetime of the fragmenting allocations, those effects
can be long-lasting or even permanent, requiring drastic measures like
forcible idle states or even reboots as the only reliable ways to recover
the address space for THP production.
In a kernel build test with supplemental THP pressure, the THP allocation
rate steadily declines over 15 runs:
thp_fault_alloc
61988
56474
57258
50187
52388
55409
52925
47648
43669
40621
36077
41721
36685
34641
33215
This is a hurdle in adopting THP in any environment where hosts are shared
between multiple overlapping workloads (cloud environments), and rarely
experience true idle periods. To make THP a reliable and predictable
optimization, there needs to be a stronger guarantee to avoid such
fragmentation.
Introduce defrag_mode. When enabled, reclaim/compaction is invoked to its
full extent *before* falling back. Specifically, ALLOC_NOFRAGMENT is
enforced on the allocator fastpath and the reclaiming slowpath.
For now, fallbacks are permitted to avert OOMs. There is a plan to add
defrag_mode=2 to prefer OOMs over fragmentation, but this requires
additional prep work in compaction and the reserve management to make it
ready for all possible allocation contexts.
The following test results are from a kernel build with periodic bursts of
THP allocations, over 15 runs:
vanilla defrag_mode=1
@claimer[unmovable]: 189 103
@claimer[movable]: 92 103
@claimer[reclaimable]: 207 61
@pollute[unmovable from movable]: 25 0
@pollute[unmovable from reclaimable]: 28 0
@pollute[movable from unmovable]: 38835 0
@pollute[movable from reclaimable]: 147136 0
@pollute[reclaimable from unmovable]: 178 0
@pollute[reclaimable from movable]: 33 0
@steal[unmovable from movable]: 11 0
@steal[unmovable from reclaimable]: 5 0
@steal[reclaimable from unmovable]: 107 0
@steal[reclaimable from movable]: 90 0
@steal[movable from reclaimable]: 354 0
@steal[movable from unmovable]: 130 0
Both types of polluting fallbacks are eliminated in this workload.
Interestingly, whole block conversions are reduced as well. This is
because once a block is claimed for a type, its empty space remains
available for future allocations, instead of being padded with fallbacks;
this allows the native type to group up instead of spreading out to new
blocks. The assumption in the allocator has been that pollution from
movable allocations is less harmful than from other types, since they can
be reclaimed or migrated out should the space be needed. However, since
fallbacks occur *before* reclaim/compaction is invoked, movable pollution
will still cause non-movable allocations to spread out and claim more
blocks.
Without fragmentation, THP rates hold steady with defrag_mode=1:
thp_fault_alloc
32478
20725
45045
32130
14018
21711
40791
29134
34458
45381
28305
17265
22584
28454
30850
While the downward trend is eliminated, the keen reader will of course
notice that the baseline rate is much smaller than the vanilla kernel's to
begin with. This is due to deficiencies in how reclaim and compaction are
currently driven: ALLOC_NOFRAGMENT increases the extent to which smaller
allocations are competing with THPs for pageblocks, while making no effort
themselves to reclaim or compact beyond their own request size. This
effect already exists with the current usage of ALLOC_NOFRAGMENT, but is
amplified by defrag_mode insisting on whole block stealing much more
strongly.
Subsequent patches will address defrag_mode reclaim strategy to raise the
THP success baseline above the vanilla kernel.
Link: https://lkml.kernel.org/r/20250313210647.1314586-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-03-13 17:05:34 -04:00
|
|
|
|
mm, page_alloc: spread allocations across zones before introducing fragmentation
Patch series "Fragmentation avoidance improvements", v5.
It has been noted before that fragmentation avoidance (aka
anti-fragmentation) is not perfect. Given sufficient time or an adverse
workload, memory gets fragmented and the long-term success of high-order
allocations degrades. This series defines an adverse workload, a definition
of external fragmentation events (including serious) ones and a series
that reduces the level of those fragmentation events.
The details of the workload and the consequences are described in more
detail in the changelogs. However, from patch 1, this is a high-level
summary of the adverse workload. The exact details are found in the
mmtests implementation.
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch)
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameterr create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed
3. Warm up a number of fio read-only threads accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll fault back in old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup
Overall the series reduces external fragmentation causing events by over 94%
on 1 and 2 socket machines, which in turn impacts high-order allocation
success rates over the long term. There are differences in latencies and
high-order allocation success rates. Latencies are a mixed bag as they
are vulnerable to exact system state and whether allocations succeeded
so they are treated as a secondary metric.
Patch 1 uses lower zones if they are populated and have free memory
instead of fragmenting a higher zone. It's special cased to
handle a Normal->DMA32 fallback with the reasons explained
in the changelog.
Patch 2-4 boosts watermarks temporarily when an external fragmentation
event occurs. kswapd wakes to reclaim a small amount of old memory
and then wakes kcompactd on completion to recover the system
slightly. This introduces some overhead in the slowpath. The level
of boosting can be tuned or disabled depending on the tolerance
for fragmentation vs allocation latency.
Patch 5 stalls some movable allocation requests to let kswapd from patch 4
make some progress. The duration of the stalls is very low but it
is possible to tune the system to avoid fragmentation events if
larger stalls can be tolerated.
The bulk of the improvement in fragmentation avoidance is from patches
1-4 but patch 5 can deal with a rare corner case and provides the option
of tuning a system for THP allocation success rates in exchange for
some stalls to control fragmentation.
This patch (of 5):
The page allocator zone lists are iterated based on the watermarks of each
zone which does not take anti-fragmentation into account. On x86, node 0
may have multiple zones while other nodes have one zone. A consequence is
that tasks running on node 0 may fragment ZONE_NORMAL even though
ZONE_DMA32 has plenty of free memory. This patch special cases the
allocator fast path such that it'll try an allocation from a lower local
zone before fragmenting a higher zone. In this case, stealing of
pageblocks or orders larger than a pageblock are still allowed in the fast
path as they are uninteresting from a fragmentation point of view.
This was evaluated using a benchmark designed to fragment memory before
attempting THP allocations. It's implemented in mmtests as the following
configurations
configs/config-global-dhp__workload_thpfioscale
configs/config-global-dhp__workload_thpfioscale-defrag
configs/config-global-dhp__workload_thpfioscale-madvhugepage
e.g. from mmtests
./run-mmtests.sh --run-monitor --config configs/config-global-dhp__workload_thpfioscale test-run-1
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch).
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameter create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed.
3. Warm up a number of fio read-only processes accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll refault old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds.
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup the test files.
Note that due to the use of IO and page cache that this benchmark is not
suitable for running on large machines where the time to fragment memory
may be excessive. Also note that while this is one mix that generates
fragmentation that it's not the only mix that generates fragmentation.
Differences in workload that are more slab-intensive or whether SLUB is
used with high-order pages may yield different results.
When the page allocator fragments memory, it records the event using the
mm_page_alloc_extfrag ftrace event. If the fallback_order is smaller than
a pageblock order (order-9 on 64-bit x86) then it's considered to be an
"external fragmentation event" that may cause issues in the future.
Hence, the primary metric here is the number of external fragmentation
events that occur with order < 9. The secondary metric is allocation
latency and huge page allocation success rates but note that differences
in latencies and what the success rate also can affect the number of
external fragmentation event which is why it's a secondary metric.
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 662.92 ( 0.00%) 653.58 * 1.41%*
Amean fault-huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
Fault latencies are slightly reduced while allocation success rates remain
at zero as this configuration does not make any special effort to allocate
THP and fio is heavily active at the time and either filling memory or
keeping pages resident. However, a 49% reduction of serious fragmentation
events reduces the changes of external fragmentation being a problem in
the future.
Vlastimil asked during review for a breakdown of the allocation types
that are falling back.
vanilla
3816 MIGRATE_UNMOVABLE
800845 MIGRATE_MOVABLE
33 MIGRATE_UNRECLAIMABLE
patch
735 MIGRATE_UNMOVABLE
408135 MIGRATE_MOVABLE
42 MIGRATE_UNRECLAIMABLE
The majority of the fallbacks are due to movable allocations and this is
consistent for the workload throughout the series so will not be presented
again as the primary source of fallbacks are movable allocations.
Movable fallbacks are sometimes considered "ok" to fallback because they
can be migrated. The problem is that they can fill an
unmovable/reclaimable pageblock causing those allocations to fallback
later and polluting pageblocks with pages that cannot move. If there is a
movable fallback, it is pretty much guaranteed to affect an
unmovable/reclaimable pageblock and while it might not be enough to
actually cause a unmovable/reclaimable fallback in the future, we cannot
know that in advance so the patch takes the only option available to it.
Hence, it's important to control them. This point is also consistent
throughout the series and will not be repeated.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 1495.14 ( 0.00%) 1467.55 ( 1.85%)
Amean fault-huge-1 1098.48 ( 0.00%) 1127.11 ( -2.61%)
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 78.57 ( 0.00%) 77.64 ( -1.18%)
Fragmentation events were reduced quite a bit although this is known
to be a little variable. The latencies and allocation success rates
are similar but they were already quite high.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 1350.05 ( 0.00%) 1346.45 ( 0.27%)
Amean fault-huge-5 4181.01 ( 0.00%) 3418.60 ( 18.24%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 1.15 ( 0.00%) 0.78 ( -31.88%)
The reduction of external fragmentation events is slight and this is
partially due to the removal of __GFP_THISNODE in commit ac5b2c18911f
("mm: thp: relax __GFP_THISNODE for MADV_HUGEPAGE mappings") as THP
allocations can now spill over to remote nodes instead of fragmenting
local memory.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 6138.97 ( 0.00%) 6217.43 ( -1.28%)
Amean fault-huge-5 2294.28 ( 0.00%) 3163.33 * -37.88%*
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 96.82 ( 0.00%) 95.14 ( -1.74%)
There was a slight reduction in external fragmentation events although the
latencies were higher. The allocation success rate is high enough that
the system is struggling and there is quite a lot of parallel reclaim and
compaction activity. There is also a certain degree of luck on whether
processes start on node 0 or not for this patch but the relevance is
reduced later in the series.
Overall, the patch reduces the number of external fragmentation causing
events so the success of THP over long periods of time would be improved
for this adverse workload.
Link: http://lkml.kernel.org/r/20181123114528.28802-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:41 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-08-06 16:07:22 -07:00
|
|
|
|
2005-06-21 17:14:41 -07:00
|
|
|
|
|
|
|
|
|
2017-02-22 15:46:16 -08:00
|
|
|
|
2011-05-24 17:12:16 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-09-06 16:24:50 -07:00
|
|
|
|
2011-05-24 17:12:16 -07:00
|
|
|
|
|
|
|
|
|
2021-09-02 14:58:13 -07:00
|
|
|
|
2011-05-24 17:12:16 -07:00
|
|
|
|
|
|
|
|
|
2022-08-23 11:22:30 +02:00
|
|
|
|
2017-02-22 15:41:45 -08:00
|
|
|
|
|
|
|
|
|
2017-02-22 15:46:10 -08:00
|
|
|
|
2017-02-22 15:41:45 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-11-05 21:16:51 -08:00
|
|
|
|
2017-02-22 15:41:45 -08:00
|
|
|
|
mm/page_alloc.c: do not warn allocation failure on zone DMA if no managed pages
In kdump kernel of x86_64, page allocation failure is observed:
kworker/u2:2: page allocation failure: order:0, mode:0xcc1(GFP_KERNEL|GFP_DMA), nodemask=(null),cpuset=/,mems_allowed=0
CPU: 0 PID: 55 Comm: kworker/u2:2 Not tainted 5.16.0-rc4+ #5
Hardware name: AMD Dinar/Dinar, BIOS RDN1505B 06/05/2013
Workqueue: events_unbound async_run_entry_fn
Call Trace:
<TASK>
dump_stack_lvl+0x48/0x5e
warn_alloc.cold+0x72/0xd6
__alloc_pages_slowpath.constprop.0+0xc69/0xcd0
__alloc_pages+0x1df/0x210
new_slab+0x389/0x4d0
___slab_alloc+0x58f/0x770
__slab_alloc.constprop.0+0x4a/0x80
kmem_cache_alloc_trace+0x24b/0x2c0
sr_probe+0x1db/0x620
......
device_add+0x405/0x920
......
__scsi_add_device+0xe5/0x100
ata_scsi_scan_host+0x97/0x1d0
async_run_entry_fn+0x30/0x130
process_one_work+0x1e8/0x3c0
worker_thread+0x50/0x3b0
? rescuer_thread+0x350/0x350
kthread+0x16b/0x190
? set_kthread_struct+0x40/0x40
ret_from_fork+0x22/0x30
</TASK>
Mem-Info:
......
The above failure happened when calling kmalloc() to allocate buffer with
GFP_DMA. It requests to allocate slab page from DMA zone while no managed
pages at all in there.
sr_probe()
--> get_capabilities()
--> buffer = kmalloc(512, GFP_KERNEL | GFP_DMA);
Because in the current kernel, dma-kmalloc will be created as long as
CONFIG_ZONE_DMA is enabled. However, kdump kernel of x86_64 doesn't have
managed pages on DMA zone since commit 6f599d84231f ("x86/kdump: Always
reserve the low 1M when the crashkernel option is specified"). The
failure can be always reproduced.
For now, let's mute the warning of allocation failure if requesting pages
from DMA zone while no managed pages.
[akpm@linux-foundation.org: fix warning]
Link: https://lkml.kernel.org/r/20211223094435.248523-4-bhe@redhat.com
Fixes: 6f599d84231f ("x86/kdump: Always reserve the low 1M when the crashkernel option is specified")
Signed-off-by: Baoquan He <bhe@redhat.com>
Acked-by: John Donnelly <john.p.donnelly@oracle.com>
Reviewed-by: Hyeonggon Yoo <42.hyeyoo@gmail.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Christoph Hellwig <hch@lst.de>
Cc: David Hildenbrand <david@redhat.com>
Cc: David Laight <David.Laight@ACULAB.COM>
Cc: Marek Szyprowski <m.szyprowski@samsung.com>
Cc: Robin Murphy <robin.murphy@arm.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-14 14:07:44 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-02-22 15:41:45 -08:00
|
|
|
|
|
|
|
|
|
2016-10-07 17:01:55 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, oom: reorganize the oom report in dump_header
OOM report contains several sections. The first one is the allocation
context that has triggered the OOM. Then we have cpuset context followed
by the stack trace of the OOM path. The tird one is the OOM memory
information. Followed by the current memory state of all system tasks.
At last, we will show oom eligible tasks and the information about the
chosen oom victim.
One thing that makes parsing more awkward than necessary is that we do not
have a single and easily parsable line about the oom context. This patch
is reorganizing the oom report to
1) who invoked oom and what was the allocation request
[ 515.902945] tuned invoked oom-killer: gfp_mask=0x6200ca(GFP_HIGHUSER_MOVABLE), order=0, oom_score_adj=0
2) OOM stack trace
[ 515.904273] CPU: 24 PID: 1809 Comm: tuned Not tainted 4.20.0-rc3+ #3
[ 515.905518] Hardware name: Inspur SA5212M4/YZMB-00370-107, BIOS 4.1.10 11/14/2016
[ 515.906821] Call Trace:
[ 515.908062] dump_stack+0x5a/0x73
[ 515.909311] dump_header+0x55/0x28c
[ 515.914260] oom_kill_process+0x2d8/0x300
[ 515.916708] out_of_memory+0x145/0x4a0
[ 515.917932] __alloc_pages_slowpath+0x7d2/0xa16
[ 515.919157] __alloc_pages_nodemask+0x277/0x290
[ 515.920367] filemap_fault+0x3d0/0x6c0
[ 515.921529] ? filemap_map_pages+0x2b8/0x420
[ 515.922709] ext4_filemap_fault+0x2c/0x40 [ext4]
[ 515.923884] __do_fault+0x20/0x80
[ 515.925032] __handle_mm_fault+0xbc0/0xe80
[ 515.926195] handle_mm_fault+0xfa/0x210
[ 515.927357] __do_page_fault+0x233/0x4c0
[ 515.928506] do_page_fault+0x32/0x140
[ 515.929646] ? page_fault+0x8/0x30
[ 515.930770] page_fault+0x1e/0x30
3) OOM memory information
[ 515.958093] Mem-Info:
[ 515.959647] active_anon:26501758 inactive_anon:1179809 isolated_anon:0
active_file:4402672 inactive_file:483963 isolated_file:1344
unevictable:0 dirty:4886753 writeback:0 unstable:0
slab_reclaimable:148442 slab_unreclaimable:18741
mapped:1347 shmem:1347 pagetables:58669 bounce:0
free:88663 free_pcp:0 free_cma:0
...
4) current memory state of all system tasks
[ 516.079544] [ 744] 0 744 9211 1345 114688 82 0 systemd-journal
[ 516.082034] [ 787] 0 787 31764 0 143360 92 0 lvmetad
[ 516.084465] [ 792] 0 792 10930 1 110592 208 -1000 systemd-udevd
[ 516.086865] [ 1199] 0 1199 13866 0 131072 112 -1000 auditd
[ 516.089190] [ 1222] 0 1222 31990 1 110592 157 0 smartd
[ 516.091477] [ 1225] 0 1225 4864 85 81920 43 0 irqbalance
[ 516.093712] [ 1226] 0 1226 52612 0 258048 426 0 abrtd
[ 516.112128] [ 1280] 0 1280 109774 55 299008 400 0 NetworkManager
[ 516.113998] [ 1295] 0 1295 28817 37 69632 24 0 ksmtuned
[ 516.144596] [ 10718] 0 10718 2622484 1721372 15998976 267219 0 panic
[ 516.145792] [ 10719] 0 10719 2622484 1164767 9818112 53576 0 panic
[ 516.146977] [ 10720] 0 10720 2622484 1174361 9904128 53709 0 panic
[ 516.148163] [ 10721] 0 10721 2622484 1209070 10194944 54824 0 panic
[ 516.149329] [ 10722] 0 10722 2622484 1745799 14774272 91138 0 panic
5) oom context (contrains and the chosen victim).
oom-kill:constraint=CONSTRAINT_NONE,nodemask=(null),cpuset=/,mems_allowed=0-1,task=panic,pid=10737,uid=0
An admin can easily get the full oom context at a single line which
makes parsing much easier.
Link: http://lkml.kernel.org/r/1542799799-36184-1-git-send-email-ufo19890607@gmail.com
Signed-off-by: yuzhoujian <yuzhoujian@didichuxing.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Tetsuo Handa <penguin-kernel@i-love.sakura.ne.jp>
Cc: Yang Shi <yang.s@alibaba-inc.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:36:07 -08:00
|
|
|
|
2017-11-15 17:39:14 -08:00
|
|
|
|
|
|
|
|
|
2016-10-07 17:01:55 -07:00
|
|
|
|
2011-10-31 17:08:35 -07:00
|
|
|
|
2017-02-22 15:46:10 -08:00
|
|
|
|
mm, oom: reorganize the oom report in dump_header
OOM report contains several sections. The first one is the allocation
context that has triggered the OOM. Then we have cpuset context followed
by the stack trace of the OOM path. The tird one is the OOM memory
information. Followed by the current memory state of all system tasks.
At last, we will show oom eligible tasks and the information about the
chosen oom victim.
One thing that makes parsing more awkward than necessary is that we do not
have a single and easily parsable line about the oom context. This patch
is reorganizing the oom report to
1) who invoked oom and what was the allocation request
[ 515.902945] tuned invoked oom-killer: gfp_mask=0x6200ca(GFP_HIGHUSER_MOVABLE), order=0, oom_score_adj=0
2) OOM stack trace
[ 515.904273] CPU: 24 PID: 1809 Comm: tuned Not tainted 4.20.0-rc3+ #3
[ 515.905518] Hardware name: Inspur SA5212M4/YZMB-00370-107, BIOS 4.1.10 11/14/2016
[ 515.906821] Call Trace:
[ 515.908062] dump_stack+0x5a/0x73
[ 515.909311] dump_header+0x55/0x28c
[ 515.914260] oom_kill_process+0x2d8/0x300
[ 515.916708] out_of_memory+0x145/0x4a0
[ 515.917932] __alloc_pages_slowpath+0x7d2/0xa16
[ 515.919157] __alloc_pages_nodemask+0x277/0x290
[ 515.920367] filemap_fault+0x3d0/0x6c0
[ 515.921529] ? filemap_map_pages+0x2b8/0x420
[ 515.922709] ext4_filemap_fault+0x2c/0x40 [ext4]
[ 515.923884] __do_fault+0x20/0x80
[ 515.925032] __handle_mm_fault+0xbc0/0xe80
[ 515.926195] handle_mm_fault+0xfa/0x210
[ 515.927357] __do_page_fault+0x233/0x4c0
[ 515.928506] do_page_fault+0x32/0x140
[ 515.929646] ? page_fault+0x8/0x30
[ 515.930770] page_fault+0x1e/0x30
3) OOM memory information
[ 515.958093] Mem-Info:
[ 515.959647] active_anon:26501758 inactive_anon:1179809 isolated_anon:0
active_file:4402672 inactive_file:483963 isolated_file:1344
unevictable:0 dirty:4886753 writeback:0 unstable:0
slab_reclaimable:148442 slab_unreclaimable:18741
mapped:1347 shmem:1347 pagetables:58669 bounce:0
free:88663 free_pcp:0 free_cma:0
...
4) current memory state of all system tasks
[ 516.079544] [ 744] 0 744 9211 1345 114688 82 0 systemd-journal
[ 516.082034] [ 787] 0 787 31764 0 143360 92 0 lvmetad
[ 516.084465] [ 792] 0 792 10930 1 110592 208 -1000 systemd-udevd
[ 516.086865] [ 1199] 0 1199 13866 0 131072 112 -1000 auditd
[ 516.089190] [ 1222] 0 1222 31990 1 110592 157 0 smartd
[ 516.091477] [ 1225] 0 1225 4864 85 81920 43 0 irqbalance
[ 516.093712] [ 1226] 0 1226 52612 0 258048 426 0 abrtd
[ 516.112128] [ 1280] 0 1280 109774 55 299008 400 0 NetworkManager
[ 516.113998] [ 1295] 0 1295 28817 37 69632 24 0 ksmtuned
[ 516.144596] [ 10718] 0 10718 2622484 1721372 15998976 267219 0 panic
[ 516.145792] [ 10719] 0 10719 2622484 1164767 9818112 53576 0 panic
[ 516.146977] [ 10720] 0 10720 2622484 1174361 9904128 53709 0 panic
[ 516.148163] [ 10721] 0 10721 2622484 1209070 10194944 54824 0 panic
[ 516.149329] [ 10722] 0 10722 2622484 1745799 14774272 91138 0 panic
5) oom context (contrains and the chosen victim).
oom-kill:constraint=CONSTRAINT_NONE,nodemask=(null),cpuset=/,mems_allowed=0-1,task=panic,pid=10737,uid=0
An admin can easily get the full oom context at a single line which
makes parsing much easier.
Link: http://lkml.kernel.org/r/1542799799-36184-1-git-send-email-ufo19890607@gmail.com
Signed-off-by: yuzhoujian <yuzhoujian@didichuxing.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Tetsuo Handa <penguin-kernel@i-love.sakura.ne.jp>
Cc: Yang Shi <yang.s@alibaba-inc.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:36:07 -08:00
|
|
|
|
2011-05-24 17:12:16 -07:00
|
|
|
|
2017-02-22 15:46:28 -08:00
|
|
|
|
2011-05-24 17:12:16 -07:00
|
|
|
|
|
|
|
|
|
2017-02-22 15:46:25 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2009-06-16 15:31:57 -07:00
|
|
|
|
|
|
|
|
|
2015-02-11 15:25:41 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2015-09-08 15:00:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-07-26 15:22:33 -07:00
|
|
|
|
2015-09-08 15:00:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2015-01-26 12:58:32 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-06-24 16:57:19 -07:00
|
|
|
|
|
|
|
|
|
2015-01-26 12:58:32 -08:00
|
|
|
|
2015-06-24 16:57:19 -07:00
|
|
|
|
2015-01-26 12:58:32 -08:00
|
|
|
|
2009-06-16 15:31:57 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2005-11-17 21:35:02 +01:00
|
|
|
|
2009-06-16 15:31:57 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-08-31 16:15:20 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2009-06-16 15:31:57 -07:00
|
|
|
|
2017-08-31 16:15:20 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-11-13 16:06:43 -08:00
|
|
|
|
2009-06-16 15:31:57 -07:00
|
|
|
|
|
|
|
|
|
2017-02-22 15:46:22 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-07-12 14:36:45 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-10-13 16:55:45 -07:00
|
|
|
|
|
|
|
|
|
2017-07-12 14:36:45 -07:00
|
|
|
|
2020-10-13 16:55:45 -07:00
|
|
|
|
2017-07-12 14:36:45 -07:00
|
|
|
|
2017-02-22 15:46:22 -08:00
|
|
|
|
2020-06-03 15:59:01 -07:00
|
|
|
|
2017-02-22 15:46:22 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-01-31 16:20:07 -08:00
|
|
|
|
2022-05-19 14:08:55 -07:00
|
|
|
|
|
|
|
|
|
2015-02-11 15:26:24 -08:00
|
|
|
|
2016-01-14 15:20:36 -08:00
|
|
|
|
2017-02-22 15:46:25 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-01-14 15:20:36 -08:00
|
|
|
|
|
|
|
|
|
2009-06-16 15:31:57 -07:00
|
|
|
|
2015-06-24 16:57:19 -07:00
|
|
|
|
2009-06-16 15:31:57 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-05-20 16:57:06 -07:00
|
|
|
|
2021-05-06 18:06:50 -07:00
|
|
|
|
2016-05-20 16:57:06 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-05-24 14:32:30 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-05-19 17:13:38 -07:00
|
|
|
|
2016-07-28 15:49:28 -07:00
|
|
|
|
2010-05-24 14:32:30 -07:00
|
|
|
|
2019-03-05 15:45:41 -08:00
|
|
|
|
psi: pressure stall information for CPU, memory, and IO
When systems are overcommitted and resources become contended, it's hard
to tell exactly the impact this has on workload productivity, or how close
the system is to lockups and OOM kills. In particular, when machines work
multiple jobs concurrently, the impact of overcommit in terms of latency
and throughput on the individual job can be enormous.
In order to maximize hardware utilization without sacrificing individual
job health or risk complete machine lockups, this patch implements a way
to quantify resource pressure in the system.
A kernel built with CONFIG_PSI=y creates files in /proc/pressure/ that
expose the percentage of time the system is stalled on CPU, memory, or IO,
respectively. Stall states are aggregate versions of the per-task delay
accounting delays:
cpu: some tasks are runnable but not executing on a CPU
memory: tasks are reclaiming, or waiting for swapin or thrashing cache
io: tasks are waiting for io completions
These percentages of walltime can be thought of as pressure percentages,
and they give a general sense of system health and productivity loss
incurred by resource overcommit. They can also indicate when the system
is approaching lockup scenarios and OOMs.
To do this, psi keeps track of the task states associated with each CPU
and samples the time they spend in stall states. Every 2 seconds, the
samples are averaged across CPUs - weighted by the CPUs' non-idle time to
eliminate artifacts from unused CPUs - and translated into percentages of
walltime. A running average of those percentages is maintained over 10s,
1m, and 5m periods (similar to the loadaverage).
[hannes@cmpxchg.org: doc fixlet, per Randy]
Link: http://lkml.kernel.org/r/20180828205625.GA14030@cmpxchg.org
[hannes@cmpxchg.org: code optimization]
Link: http://lkml.kernel.org/r/20180907175015.GA8479@cmpxchg.org
[hannes@cmpxchg.org: rename psi_clock() to psi_update_work(), per Peter]
Link: http://lkml.kernel.org/r/20180907145404.GB11088@cmpxchg.org
[hannes@cmpxchg.org: fix build]
Link: http://lkml.kernel.org/r/20180913014222.GA2370@cmpxchg.org
Link: http://lkml.kernel.org/r/20180828172258.3185-9-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Daniel Drake <drake@endlessm.com>
Tested-by: Suren Baghdasaryan <surenb@google.com>
Cc: Christopher Lameter <cl@linux.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Johannes Weiner <jweiner@fb.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Enderborg <peter.enderborg@sony.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vinayak Menon <vinmenon@codeaurora.org>
Cc: Randy Dunlap <rdunlap@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-26 15:06:27 -07:00
|
|
|
|
2017-05-08 15:59:50 -07:00
|
|
|
|
mm, compaction: defer each zone individually instead of preferred zone
When direct sync compaction is often unsuccessful, it may become deferred
for some time to avoid further useless attempts, both sync and async.
Successful high-order allocations un-defer compaction, while further
unsuccessful compaction attempts prolong the compaction deferred period.
Currently the checking and setting deferred status is performed only on
the preferred zone of the allocation that invoked direct compaction. But
compaction itself is attempted on all eligible zones in the zonelist, so
the behavior is suboptimal and may lead both to scenarios where 1)
compaction is attempted uselessly, or 2) where it's not attempted despite
good chances of succeeding, as shown on the examples below:
1) A direct compaction with Normal preferred zone failed and set
deferred compaction for the Normal zone. Another unrelated direct
compaction with DMA32 as preferred zone will attempt to compact DMA32
zone even though the first compaction attempt also included DMA32 zone.
In another scenario, compaction with Normal preferred zone failed to
compact Normal zone, but succeeded in the DMA32 zone, so it will not
defer compaction. In the next attempt, it will try Normal zone which
will fail again, instead of skipping Normal zone and trying DMA32
directly.
2) Kswapd will balance DMA32 zone and reset defer status based on
watermarks looking good. A direct compaction with preferred Normal
zone will skip compaction of all zones including DMA32 because Normal
was still deferred. The allocation might have succeeded in DMA32, but
won't.
This patch makes compaction deferring work on individual zone basis
instead of preferred zone. For each zone, it checks compaction_deferred()
to decide if the zone should be skipped. If watermarks fail after
compacting the zone, defer_compaction() is called. The zone where
watermarks passed can still be deferred when the allocation attempt is
unsuccessful. When allocation is successful, compaction_defer_reset() is
called for the zone containing the allocated page. This approach should
approximate calling defer_compaction() only on zones where compaction was
attempted and did not yield allocated page. There might be corner cases
but that is inevitable as long as the decision to stop compacting dues not
guarantee that a page will be allocated.
Due to a new COMPACT_DEFERRED return value, some functions relying
implicitly on COMPACT_SKIPPED = 0 had to be updated, with comments made
more accurate. The did_some_progress output parameter of
__alloc_pages_direct_compact() is removed completely, as the caller
actually does not use it after compaction sets it - it is only considered
when direct reclaim sets it.
During testing on a two-node machine with a single very small Normal zone
on node 1, this patch has improved success rates in stress-highalloc
mmtests benchmark. The success here were previously made worse by commit
3a025760fc15 ("mm: page_alloc: spill to remote nodes before waking
kswapd") as kswapd was no longer resetting often enough the deferred
compaction for the Normal zone, and DMA32 zones on both nodes were thus
not considered for compaction. On different machine, success rates were
improved with __GFP_NO_KSWAPD allocations.
[akpm@linux-foundation.org: fix CONFIG_COMPACTION=n build]
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Minchan Kim <minchan@kernel.org>
Reviewed-by: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Michal Nazarewicz <mina86@mina86.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-10-09 15:27:02 -07:00
|
|
|
|
2012-01-12 17:19:41 -08:00
|
|
|
|
2010-05-24 14:32:30 -07:00
|
|
|
|
|
|
|
|
|
psi: pressure stall information for CPU, memory, and IO
When systems are overcommitted and resources become contended, it's hard
to tell exactly the impact this has on workload productivity, or how close
the system is to lockups and OOM kills. In particular, when machines work
multiple jobs concurrently, the impact of overcommit in terms of latency
and throughput on the individual job can be enormous.
In order to maximize hardware utilization without sacrificing individual
job health or risk complete machine lockups, this patch implements a way
to quantify resource pressure in the system.
A kernel built with CONFIG_PSI=y creates files in /proc/pressure/ that
expose the percentage of time the system is stalled on CPU, memory, or IO,
respectively. Stall states are aggregate versions of the per-task delay
accounting delays:
cpu: some tasks are runnable but not executing on a CPU
memory: tasks are reclaiming, or waiting for swapin or thrashing cache
io: tasks are waiting for io completions
These percentages of walltime can be thought of as pressure percentages,
and they give a general sense of system health and productivity loss
incurred by resource overcommit. They can also indicate when the system
is approaching lockup scenarios and OOMs.
To do this, psi keeps track of the task states associated with each CPU
and samples the time they spend in stall states. Every 2 seconds, the
samples are averaged across CPUs - weighted by the CPUs' non-idle time to
eliminate artifacts from unused CPUs - and translated into percentages of
walltime. A running average of those percentages is maintained over 10s,
1m, and 5m periods (similar to the loadaverage).
[hannes@cmpxchg.org: doc fixlet, per Randy]
Link: http://lkml.kernel.org/r/20180828205625.GA14030@cmpxchg.org
[hannes@cmpxchg.org: code optimization]
Link: http://lkml.kernel.org/r/20180907175015.GA8479@cmpxchg.org
[hannes@cmpxchg.org: rename psi_clock() to psi_update_work(), per Peter]
Link: http://lkml.kernel.org/r/20180907145404.GB11088@cmpxchg.org
[hannes@cmpxchg.org: fix build]
Link: http://lkml.kernel.org/r/20180913014222.GA2370@cmpxchg.org
Link: http://lkml.kernel.org/r/20180828172258.3185-9-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Daniel Drake <drake@endlessm.com>
Tested-by: Suren Baghdasaryan <surenb@google.com>
Cc: Christopher Lameter <cl@linux.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Johannes Weiner <jweiner@fb.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Enderborg <peter.enderborg@sony.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vinayak Menon <vinmenon@codeaurora.org>
Cc: Randy Dunlap <rdunlap@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-26 15:06:27 -07:00
|
|
|
|
2022-01-19 18:10:15 -08:00
|
|
|
|
2017-05-08 15:59:50 -07:00
|
|
|
|
psi: pressure stall information for CPU, memory, and IO
When systems are overcommitted and resources become contended, it's hard
to tell exactly the impact this has on workload productivity, or how close
the system is to lockups and OOM kills. In particular, when machines work
multiple jobs concurrently, the impact of overcommit in terms of latency
and throughput on the individual job can be enormous.
In order to maximize hardware utilization without sacrificing individual
job health or risk complete machine lockups, this patch implements a way
to quantify resource pressure in the system.
A kernel built with CONFIG_PSI=y creates files in /proc/pressure/ that
expose the percentage of time the system is stalled on CPU, memory, or IO,
respectively. Stall states are aggregate versions of the per-task delay
accounting delays:
cpu: some tasks are runnable but not executing on a CPU
memory: tasks are reclaiming, or waiting for swapin or thrashing cache
io: tasks are waiting for io completions
These percentages of walltime can be thought of as pressure percentages,
and they give a general sense of system health and productivity loss
incurred by resource overcommit. They can also indicate when the system
is approaching lockup scenarios and OOMs.
To do this, psi keeps track of the task states associated with each CPU
and samples the time they spend in stall states. Every 2 seconds, the
samples are averaged across CPUs - weighted by the CPUs' non-idle time to
eliminate artifacts from unused CPUs - and translated into percentages of
walltime. A running average of those percentages is maintained over 10s,
1m, and 5m periods (similar to the loadaverage).
[hannes@cmpxchg.org: doc fixlet, per Randy]
Link: http://lkml.kernel.org/r/20180828205625.GA14030@cmpxchg.org
[hannes@cmpxchg.org: code optimization]
Link: http://lkml.kernel.org/r/20180907175015.GA8479@cmpxchg.org
[hannes@cmpxchg.org: rename psi_clock() to psi_update_work(), per Peter]
Link: http://lkml.kernel.org/r/20180907145404.GB11088@cmpxchg.org
[hannes@cmpxchg.org: fix build]
Link: http://lkml.kernel.org/r/20180913014222.GA2370@cmpxchg.org
Link: http://lkml.kernel.org/r/20180828172258.3185-9-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Daniel Drake <drake@endlessm.com>
Tested-by: Suren Baghdasaryan <surenb@google.com>
Cc: Christopher Lameter <cl@linux.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Johannes Weiner <jweiner@fb.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Enderborg <peter.enderborg@sony.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vinayak Menon <vinmenon@codeaurora.org>
Cc: Randy Dunlap <rdunlap@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-26 15:06:27 -07:00
|
|
|
|
2016-05-20 16:56:53 -07:00
|
|
|
|
2019-03-05 15:45:41 -08:00
|
|
|
|
psi: pressure stall information for CPU, memory, and IO
When systems are overcommitted and resources become contended, it's hard
to tell exactly the impact this has on workload productivity, or how close
the system is to lockups and OOM kills. In particular, when machines work
multiple jobs concurrently, the impact of overcommit in terms of latency
and throughput on the individual job can be enormous.
In order to maximize hardware utilization without sacrificing individual
job health or risk complete machine lockups, this patch implements a way
to quantify resource pressure in the system.
A kernel built with CONFIG_PSI=y creates files in /proc/pressure/ that
expose the percentage of time the system is stalled on CPU, memory, or IO,
respectively. Stall states are aggregate versions of the per-task delay
accounting delays:
cpu: some tasks are runnable but not executing on a CPU
memory: tasks are reclaiming, or waiting for swapin or thrashing cache
io: tasks are waiting for io completions
These percentages of walltime can be thought of as pressure percentages,
and they give a general sense of system health and productivity loss
incurred by resource overcommit. They can also indicate when the system
is approaching lockup scenarios and OOMs.
To do this, psi keeps track of the task states associated with each CPU
and samples the time they spend in stall states. Every 2 seconds, the
samples are averaged across CPUs - weighted by the CPUs' non-idle time to
eliminate artifacts from unused CPUs - and translated into percentages of
walltime. A running average of those percentages is maintained over 10s,
1m, and 5m periods (similar to the loadaverage).
[hannes@cmpxchg.org: doc fixlet, per Randy]
Link: http://lkml.kernel.org/r/20180828205625.GA14030@cmpxchg.org
[hannes@cmpxchg.org: code optimization]
Link: http://lkml.kernel.org/r/20180907175015.GA8479@cmpxchg.org
[hannes@cmpxchg.org: rename psi_clock() to psi_update_work(), per Peter]
Link: http://lkml.kernel.org/r/20180907145404.GB11088@cmpxchg.org
[hannes@cmpxchg.org: fix build]
Link: http://lkml.kernel.org/r/20180913014222.GA2370@cmpxchg.org
Link: http://lkml.kernel.org/r/20180828172258.3185-9-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Daniel Drake <drake@endlessm.com>
Tested-by: Suren Baghdasaryan <surenb@google.com>
Cc: Christopher Lameter <cl@linux.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Johannes Weiner <jweiner@fb.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Enderborg <peter.enderborg@sony.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vinayak Menon <vinmenon@codeaurora.org>
Cc: Randy Dunlap <rdunlap@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-26 15:06:27 -07:00
|
|
|
|
2017-05-08 15:59:50 -07:00
|
|
|
|
psi: pressure stall information for CPU, memory, and IO
When systems are overcommitted and resources become contended, it's hard
to tell exactly the impact this has on workload productivity, or how close
the system is to lockups and OOM kills. In particular, when machines work
multiple jobs concurrently, the impact of overcommit in terms of latency
and throughput on the individual job can be enormous.
In order to maximize hardware utilization without sacrificing individual
job health or risk complete machine lockups, this patch implements a way
to quantify resource pressure in the system.
A kernel built with CONFIG_PSI=y creates files in /proc/pressure/ that
expose the percentage of time the system is stalled on CPU, memory, or IO,
respectively. Stall states are aggregate versions of the per-task delay
accounting delays:
cpu: some tasks are runnable but not executing on a CPU
memory: tasks are reclaiming, or waiting for swapin or thrashing cache
io: tasks are waiting for io completions
These percentages of walltime can be thought of as pressure percentages,
and they give a general sense of system health and productivity loss
incurred by resource overcommit. They can also indicate when the system
is approaching lockup scenarios and OOMs.
To do this, psi keeps track of the task states associated with each CPU
and samples the time they spend in stall states. Every 2 seconds, the
samples are averaged across CPUs - weighted by the CPUs' non-idle time to
eliminate artifacts from unused CPUs - and translated into percentages of
walltime. A running average of those percentages is maintained over 10s,
1m, and 5m periods (similar to the loadaverage).
[hannes@cmpxchg.org: doc fixlet, per Randy]
Link: http://lkml.kernel.org/r/20180828205625.GA14030@cmpxchg.org
[hannes@cmpxchg.org: code optimization]
Link: http://lkml.kernel.org/r/20180907175015.GA8479@cmpxchg.org
[hannes@cmpxchg.org: rename psi_clock() to psi_update_work(), per Peter]
Link: http://lkml.kernel.org/r/20180907145404.GB11088@cmpxchg.org
[hannes@cmpxchg.org: fix build]
Link: http://lkml.kernel.org/r/20180913014222.GA2370@cmpxchg.org
Link: http://lkml.kernel.org/r/20180828172258.3185-9-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Daniel Drake <drake@endlessm.com>
Tested-by: Suren Baghdasaryan <surenb@google.com>
Cc: Christopher Lameter <cl@linux.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Johannes Weiner <jweiner@fb.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Enderborg <peter.enderborg@sony.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vinayak Menon <vinmenon@codeaurora.org>
Cc: Randy Dunlap <rdunlap@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-26 15:06:27 -07:00
|
|
|
|
2022-01-19 18:10:15 -08:00
|
|
|
|
2010-05-24 14:32:30 -07:00
|
|
|
|
2021-05-04 18:36:51 -07:00
|
|
|
|
|
|
|
|
|
2014-10-09 15:27:04 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-03-05 15:45:41 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-10-09 15:27:04 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-05-24 14:32:30 -07:00
|
|
|
|
|
|
|
|
|
2016-05-20 16:57:06 -07:00
|
|
|
|
2016-10-07 17:00:28 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-10-07 17:00:31 -07:00
|
|
|
|
2016-10-07 17:00:28 -07:00
|
|
|
|
|
|
|
|
|
2016-10-07 17:00:34 -07:00
|
|
|
|
2017-02-22 15:42:03 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-10-07 17:00:28 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/page_alloc: bail out on fatal signal during reclaim/compaction retry attempt
A customer experienced a low-memory situation and decided to issue a
SIGKILL (i.e. a fatal signal). Instead of promptly terminating as one
would expect, the aforementioned task remained unresponsive.
Further investigation indicated that the task was "stuck" in the
reclaim/compaction retry loop. Now, it does not make sense to retry
compaction when a fatal signal is pending.
In the context of try_to_compact_pages(), indeed COMPACT_SKIPPED can be
returned; albeit, not every zone, on the zone list, would be considered in
the case a fatal signal is found to be pending. Yet, in
should_compact_retry(), given the last known compaction result, each zone,
on the zone list, can be considered/or checked (see
compaction_zonelist_suitable()). For example, if a zone was found to
succeed, then reclaim/compaction would be tried again (notwithstanding the
above).
This patch ensures that compaction is not needlessly retried irrespective
of the last known compaction result e.g. if it was skipped, in the
unlikely case a fatal signal is found pending. So, OOM is at least
attempted.
Link: https://lkml.kernel.org/r/20210520142901.3371299-1-atomlin@redhat.com
Signed-off-by: Aaron Tomlin <atomlin@redhat.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-28 19:41:10 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-10-07 17:00:28 -07:00
|
|
|
|
2023-05-19 14:39:55 +02:00
|
|
|
|
|
|
|
|
|
2016-10-07 17:00:28 -07:00
|
|
|
|
2023-05-19 14:39:55 +02:00
|
|
|
|
2017-02-22 15:42:03 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-10-07 17:00:28 -07:00
|
|
|
|
mm, compaction: raise compaction priority after it withdrawns
Mike Kravetz reports that "hugetlb allocations could stall for minutes or
hours when should_compact_retry() would return true more often then it
should. Specifically, this was in the case where compact_result was
COMPACT_DEFERRED and COMPACT_PARTIAL_SKIPPED and no progress was being
made."
The problem is that the compaction_withdrawn() test in
should_compact_retry() includes compaction outcomes that are only possible
on low compaction priority, and results in a retry without increasing the
priority. This may result in furter reclaim, and more incomplete
compaction attempts.
With this patch, compaction priority is raised when possible, or
should_compact_retry() returns false.
The COMPACT_SKIPPED result doesn't really fit together with the other
outcomes in compaction_withdrawn(), as that's a result caused by
insufficient order-0 pages, not due to low compaction priority. With this
patch, it is moved to a new compaction_needs_reclaim() function, and for
that outcome we keep the current logic of retrying if it looks like
reclaim will be able to help.
Link: http://lkml.kernel.org/r/20190806014744.15446-4-mike.kravetz@oracle.com
Reported-by: Mike Kravetz <mike.kravetz@oracle.com>
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Mike Kravetz <mike.kravetz@oracle.com>
Tested-by: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-09-23 15:37:32 -07:00
|
|
|
|
2023-05-19 14:39:56 +02:00
|
|
|
|
|
|
|
|
|
mm, compaction: raise compaction priority after it withdrawns
Mike Kravetz reports that "hugetlb allocations could stall for minutes or
hours when should_compact_retry() would return true more often then it
should. Specifically, this was in the case where compact_result was
COMPACT_DEFERRED and COMPACT_PARTIAL_SKIPPED and no progress was being
made."
The problem is that the compaction_withdrawn() test in
should_compact_retry() includes compaction outcomes that are only possible
on low compaction priority, and results in a retry without increasing the
priority. This may result in furter reclaim, and more incomplete
compaction attempts.
With this patch, compaction priority is raised when possible, or
should_compact_retry() returns false.
The COMPACT_SKIPPED result doesn't really fit together with the other
outcomes in compaction_withdrawn(), as that's a result caused by
insufficient order-0 pages, not due to low compaction priority. With this
patch, it is moved to a new compaction_needs_reclaim() function, and for
that outcome we keep the current logic of retrying if it looks like
reclaim will be able to help.
Link: http://lkml.kernel.org/r/20190806014744.15446-4-mike.kravetz@oracle.com
Reported-by: Mike Kravetz <mike.kravetz@oracle.com>
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Mike Kravetz <mike.kravetz@oracle.com>
Tested-by: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-09-23 15:37:32 -07:00
|
|
|
|
2023-05-19 14:39:56 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, compaction: raise compaction priority after it withdrawns
Mike Kravetz reports that "hugetlb allocations could stall for minutes or
hours when should_compact_retry() would return true more often then it
should. Specifically, this was in the case where compact_result was
COMPACT_DEFERRED and COMPACT_PARTIAL_SKIPPED and no progress was being
made."
The problem is that the compaction_withdrawn() test in
should_compact_retry() includes compaction outcomes that are only possible
on low compaction priority, and results in a retry without increasing the
priority. This may result in furter reclaim, and more incomplete
compaction attempts.
With this patch, compaction priority is raised when possible, or
should_compact_retry() returns false.
The COMPACT_SKIPPED result doesn't really fit together with the other
outcomes in compaction_withdrawn(), as that's a result caused by
insufficient order-0 pages, not due to low compaction priority. With this
patch, it is moved to a new compaction_needs_reclaim() function, and for
that outcome we keep the current logic of retrying if it looks like
reclaim will be able to help.
Link: http://lkml.kernel.org/r/20190806014744.15446-4-mike.kravetz@oracle.com
Reported-by: Mike Kravetz <mike.kravetz@oracle.com>
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Mike Kravetz <mike.kravetz@oracle.com>
Tested-by: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-09-23 15:37:32 -07:00
|
|
|
|
2023-05-19 14:39:56 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, compaction: raise compaction priority after it withdrawns
Mike Kravetz reports that "hugetlb allocations could stall for minutes or
hours when should_compact_retry() would return true more often then it
should. Specifically, this was in the case where compact_result was
COMPACT_DEFERRED and COMPACT_PARTIAL_SKIPPED and no progress was being
made."
The problem is that the compaction_withdrawn() test in
should_compact_retry() includes compaction outcomes that are only possible
on low compaction priority, and results in a retry without increasing the
priority. This may result in furter reclaim, and more incomplete
compaction attempts.
With this patch, compaction priority is raised when possible, or
should_compact_retry() returns false.
The COMPACT_SKIPPED result doesn't really fit together with the other
outcomes in compaction_withdrawn(), as that's a result caused by
insufficient order-0 pages, not due to low compaction priority. With this
patch, it is moved to a new compaction_needs_reclaim() function, and for
that outcome we keep the current logic of retrying if it looks like
reclaim will be able to help.
Link: http://lkml.kernel.org/r/20190806014744.15446-4-mike.kravetz@oracle.com
Reported-by: Mike Kravetz <mike.kravetz@oracle.com>
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Mike Kravetz <mike.kravetz@oracle.com>
Tested-by: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-09-23 15:37:32 -07:00
|
|
|
|
|
|
|
|
|
2016-10-07 17:00:28 -07:00
|
|
|
|
2023-05-19 14:39:56 +02:00
|
|
|
|
2016-10-07 17:00:28 -07:00
|
|
|
|
2016-10-07 17:00:34 -07:00
|
|
|
|
|
|
|
|
|
2017-02-22 15:42:03 -08:00
|
|
|
|
2016-10-07 17:00:34 -07:00
|
|
|
|
2016-10-07 17:00:31 -07:00
|
|
|
|
|
|
|
|
|
2017-02-22 15:42:03 -08:00
|
|
|
|
2016-10-07 17:00:31 -07:00
|
|
|
|
2017-02-22 15:42:03 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-10-07 17:00:28 -07:00
|
|
|
|
2010-05-24 14:32:30 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-05-19 17:13:38 -07:00
|
|
|
|
2016-07-28 15:49:28 -07:00
|
|
|
|
2010-05-24 14:32:30 -07:00
|
|
|
|
2016-05-20 16:57:06 -07:00
|
|
|
|
2010-05-24 14:32:30 -07:00
|
|
|
|
|
|
|
|
|
2016-05-20 16:57:06 -07:00
|
|
|
|
|
|
|
|
|
2025-08-26 14:06:54 +00:00
|
|
|
|
2016-05-20 16:57:12 -07:00
|
|
|
|
2016-07-28 15:49:28 -07:00
|
|
|
|
2016-10-07 17:00:31 -07:00
|
|
|
|
2016-05-20 16:57:06 -07:00
|
|
|
|
2016-05-20 16:57:15 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-06-03 15:59:01 -07:00
|
|
|
|
|
|
|
|
|
2016-05-20 16:57:15 -07:00
|
|
|
|
2020-06-03 15:59:01 -07:00
|
|
|
|
2016-05-20 16:57:15 -07:00
|
|
|
|
|
|
|
|
|
2016-05-20 16:57:06 -07:00
|
|
|
|
|
|
|
|
|
2016-10-07 17:00:28 -07:00
|
|
|
|
2010-05-24 14:32:30 -07:00
|
|
|
|
2017-03-03 10:13:38 +01:00
|
|
|
|
lockdep: fix fs_reclaim annotation
While revisiting my Btrfs swapfile series [1], I introduced a situation
in which reclaim would lock i_rwsem, and even though the swapon() path
clearly made GFP_KERNEL allocations while holding i_rwsem, I got no
complaints from lockdep. It turns out that the rework of the fs_reclaim
annotation was broken: if the current task has PF_MEMALLOC set, we don't
acquire the dummy fs_reclaim lock, but when reclaiming we always check
this _after_ we've just set the PF_MEMALLOC flag. In most cases, we can
fix this by moving the fs_reclaim_{acquire,release}() outside of the
memalloc_noreclaim_{save,restore}(), althought kswapd is slightly
different. After applying this, I got the expected lockdep splats.
1: https://lwn.net/Articles/625412/
Link: http://lkml.kernel.org/r/9f8aa70652a98e98d7c4de0fc96a4addcee13efe.1523778026.git.osandov@fb.com
Fixes: d92a8cfcb37e ("locking/lockdep: Rework FS_RECLAIM annotation")
Signed-off-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-06-07 17:07:02 -07:00
|
|
|
|
2017-03-03 10:13:38 +01:00
|
|
|
|
|
|
|
|
|
2020-12-14 19:08:30 -08:00
|
|
|
|
2017-03-03 10:13:38 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-03-22 16:17:10 -07:00
|
|
|
|
2017-03-03 10:13:38 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-09-02 14:52:58 -07:00
|
|
|
|
lockdep: fix fs_reclaim annotation
While revisiting my Btrfs swapfile series [1], I introduced a situation
in which reclaim would lock i_rwsem, and even though the swapon() path
clearly made GFP_KERNEL allocations while holding i_rwsem, I got no
complaints from lockdep. It turns out that the rework of the fs_reclaim
annotation was broken: if the current task has PF_MEMALLOC set, we don't
acquire the dummy fs_reclaim lock, but when reclaiming we always check
this _after_ we've just set the PF_MEMALLOC flag. In most cases, we can
fix this by moving the fs_reclaim_{acquire,release}() outside of the
memalloc_noreclaim_{save,restore}(), althought kswapd is slightly
different. After applying this, I got the expected lockdep splats.
1: https://lwn.net/Articles/625412/
Link: http://lkml.kernel.org/r/9f8aa70652a98e98d7c4de0fc96a4addcee13efe.1523778026.git.osandov@fb.com
Fixes: d92a8cfcb37e ("locking/lockdep: Rework FS_RECLAIM annotation")
Signed-off-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-06-07 17:07:02 -07:00
|
|
|
|
2021-09-02 14:52:58 -07:00
|
|
|
|
lockdep: fix fs_reclaim annotation
While revisiting my Btrfs swapfile series [1], I introduced a situation
in which reclaim would lock i_rwsem, and even though the swapon() path
clearly made GFP_KERNEL allocations while holding i_rwsem, I got no
complaints from lockdep. It turns out that the rework of the fs_reclaim
annotation was broken: if the current task has PF_MEMALLOC set, we don't
acquire the dummy fs_reclaim lock, but when reclaiming we always check
this _after_ we've just set the PF_MEMALLOC flag. In most cases, we can
fix this by moving the fs_reclaim_{acquire,release}() outside of the
memalloc_noreclaim_{save,restore}(), althought kswapd is slightly
different. After applying this, I got the expected lockdep splats.
1: https://lwn.net/Articles/625412/
Link: http://lkml.kernel.org/r/9f8aa70652a98e98d7c4de0fc96a4addcee13efe.1523778026.git.osandov@fb.com
Fixes: d92a8cfcb37e ("locking/lockdep: Rework FS_RECLAIM annotation")
Signed-off-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-06-07 17:07:02 -07:00
|
|
|
|
|
|
|
|
|
2021-09-02 14:52:58 -07:00
|
|
|
|
lockdep: fix fs_reclaim annotation
While revisiting my Btrfs swapfile series [1], I introduced a situation
in which reclaim would lock i_rwsem, and even though the swapon() path
clearly made GFP_KERNEL allocations while holding i_rwsem, I got no
complaints from lockdep. It turns out that the rework of the fs_reclaim
annotation was broken: if the current task has PF_MEMALLOC set, we don't
acquire the dummy fs_reclaim lock, but when reclaiming we always check
this _after_ we've just set the PF_MEMALLOC flag. In most cases, we can
fix this by moving the fs_reclaim_{acquire,release}() outside of the
memalloc_noreclaim_{save,restore}(), althought kswapd is slightly
different. After applying this, I got the expected lockdep splats.
1: https://lwn.net/Articles/625412/
Link: http://lkml.kernel.org/r/9f8aa70652a98e98d7c4de0fc96a4addcee13efe.1523778026.git.osandov@fb.com
Fixes: d92a8cfcb37e ("locking/lockdep: Rework FS_RECLAIM annotation")
Signed-off-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-06-07 17:07:02 -07:00
|
|
|
|
2021-09-02 14:52:58 -07:00
|
|
|
|
lockdep: fix fs_reclaim annotation
While revisiting my Btrfs swapfile series [1], I introduced a situation
in which reclaim would lock i_rwsem, and even though the swapon() path
clearly made GFP_KERNEL allocations while holding i_rwsem, I got no
complaints from lockdep. It turns out that the rework of the fs_reclaim
annotation was broken: if the current task has PF_MEMALLOC set, we don't
acquire the dummy fs_reclaim lock, but when reclaiming we always check
this _after_ we've just set the PF_MEMALLOC flag. In most cases, we can
fix this by moving the fs_reclaim_{acquire,release}() outside of the
memalloc_noreclaim_{save,restore}(), althought kswapd is slightly
different. After applying this, I got the expected lockdep splats.
1: https://lwn.net/Articles/625412/
Link: http://lkml.kernel.org/r/9f8aa70652a98e98d7c4de0fc96a4addcee13efe.1523778026.git.osandov@fb.com
Fixes: d92a8cfcb37e ("locking/lockdep: Rework FS_RECLAIM annotation")
Signed-off-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-06-07 17:07:02 -07:00
|
|
|
|
|
|
|
|
|
2017-03-03 10:13:38 +01:00
|
|
|
|
|
|
|
|
|
2020-12-14 19:08:30 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-09-02 14:52:58 -07:00
|
|
|
|
2020-12-14 19:08:30 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-03-03 10:13:38 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-12-14 19:08:30 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-09-02 14:52:58 -07:00
|
|
|
|
2020-12-14 19:08:30 -08:00
|
|
|
|
2017-03-03 10:13:38 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/page_alloc: fix race condition between build_all_zonelists and page allocation
Patrick Daly reported the following problem;
NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK] - before offline operation
[0] - ZONE_MOVABLE
[1] - ZONE_NORMAL
[2] - NULL
For a GFP_KERNEL allocation, alloc_pages_slowpath() will save the
offset of ZONE_NORMAL in ac->preferred_zoneref. If a concurrent
memory_offline operation removes the last page from ZONE_MOVABLE,
build_all_zonelists() & build_zonerefs_node() will update
node_zonelists as shown below. Only populated zones are added.
NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK] - after offline operation
[0] - ZONE_NORMAL
[1] - NULL
[2] - NULL
The race is simple -- page allocation could be in progress when a memory
hot-remove operation triggers a zonelist rebuild that removes zones. The
allocation request will still have a valid ac->preferred_zoneref that is
now pointing to NULL and triggers an OOM kill.
This problem probably always existed but may be slightly easier to trigger
due to 6aa303defb74 ("mm, vmscan: only allocate and reclaim from zones
with pages managed by the buddy allocator") which distinguishes between
zones that are completely unpopulated versus zones that have valid pages
not managed by the buddy allocator (e.g. reserved, memblock, ballooning
etc). Memory hotplug had multiple stages with timing considerations
around managed/present page updates, the zonelist rebuild and the zone
span updates. As David Hildenbrand puts it
memory offlining adjusts managed+present pages of the zone
essentially in one go. If after the adjustments, the zone is no
longer populated (present==0), we rebuild the zone lists.
Once that's done, we try shrinking the zone (start+spanned
pages) -- which results in zone_start_pfn == 0 if there are no
more pages. That happens *after* rebuilding the zonelists via
remove_pfn_range_from_zone().
The only requirement to fix the race is that a page allocation request
identifies when a zonelist rebuild has happened since the allocation
request started and no page has yet been allocated. Use a seqlock_t to
track zonelist updates with a lockless read-side of the zonelist and
protecting the rebuild and update of the counter with a spinlock.
[akpm@linux-foundation.org: make zonelist_update_seq static]
Link: https://lkml.kernel.org/r/20220824110900.vh674ltxmzb3proq@techsingularity.net
Fixes: 6aa303defb74 ("mm, vmscan: only allocate and reclaim from zones with pages managed by the buddy allocator")
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Reported-by: Patrick Daly <quic_pdaly@quicinc.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: David Hildenbrand <david@redhat.com>
Cc: <stable@vger.kernel.org> [4.9+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-08-24 12:14:50 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-01-25 12:09:52 +01:00
|
|
|
|
2020-10-13 16:55:54 -07:00
|
|
|
|
2015-02-11 15:25:41 -08:00
|
|
|
|
|
|
|
|
|
2009-06-16 15:31:57 -07:00
|
|
|
|
2017-05-08 15:59:50 -07:00
|
|
|
|
2022-03-22 14:43:54 -07:00
|
|
|
|
2009-06-16 15:31:57 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-03-03 10:13:38 +01:00
|
|
|
|
lockdep: fix fs_reclaim annotation
While revisiting my Btrfs swapfile series [1], I introduced a situation
in which reclaim would lock i_rwsem, and even though the swapon() path
clearly made GFP_KERNEL allocations while holding i_rwsem, I got no
complaints from lockdep. It turns out that the rework of the fs_reclaim
annotation was broken: if the current task has PF_MEMALLOC set, we don't
acquire the dummy fs_reclaim lock, but when reclaiming we always check
this _after_ we've just set the PF_MEMALLOC flag. In most cases, we can
fix this by moving the fs_reclaim_{acquire,release}() outside of the
memalloc_noreclaim_{save,restore}(), althought kswapd is slightly
different. After applying this, I got the expected lockdep splats.
1: https://lwn.net/Articles/625412/
Link: http://lkml.kernel.org/r/9f8aa70652a98e98d7c4de0fc96a4addcee13efe.1523778026.git.osandov@fb.com
Fixes: d92a8cfcb37e ("locking/lockdep: Rework FS_RECLAIM annotation")
Signed-off-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-06-07 17:07:02 -07:00
|
|
|
|
2009-06-16 15:31:57 -07:00
|
|
|
|
2015-02-11 15:25:41 -08:00
|
|
|
|
|
|
|
|
|
2009-06-16 15:31:57 -07:00
|
|
|
|
2017-05-08 15:59:50 -07:00
|
|
|
|
lockdep: fix fs_reclaim annotation
While revisiting my Btrfs swapfile series [1], I introduced a situation
in which reclaim would lock i_rwsem, and even though the swapon() path
clearly made GFP_KERNEL allocations while holding i_rwsem, I got no
complaints from lockdep. It turns out that the rework of the fs_reclaim
annotation was broken: if the current task has PF_MEMALLOC set, we don't
acquire the dummy fs_reclaim lock, but when reclaiming we always check
this _after_ we've just set the PF_MEMALLOC flag. In most cases, we can
fix this by moving the fs_reclaim_{acquire,release}() outside of the
memalloc_noreclaim_{save,restore}(), althought kswapd is slightly
different. After applying this, I got the expected lockdep splats.
1: https://lwn.net/Articles/625412/
Link: http://lkml.kernel.org/r/9f8aa70652a98e98d7c4de0fc96a4addcee13efe.1523778026.git.osandov@fb.com
Fixes: d92a8cfcb37e ("locking/lockdep: Rework FS_RECLAIM annotation")
Signed-off-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-06-07 17:07:02 -07:00
|
|
|
|
2009-06-16 15:31:57 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-01-25 12:09:52 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-05-19 17:13:38 -07:00
|
|
|
|
2015-02-11 15:25:41 -08:00
|
|
|
|
2012-01-25 12:09:52 +01:00
|
|
|
|
|
|
|
|
|
2022-03-22 14:43:54 -07:00
|
|
|
|
2012-01-25 12:09:52 +01:00
|
|
|
|
|
|
|
|
|
2022-03-22 14:43:54 -07:00
|
|
|
|
2015-02-11 15:25:41 -08:00
|
|
|
|
2010-09-09 16:38:18 -07:00
|
|
|
|
2022-03-22 14:43:54 -07:00
|
|
|
|
2009-06-16 15:31:57 -07:00
|
|
|
|
2010-09-09 16:38:18 -07:00
|
|
|
|
2016-07-28 15:49:13 -07:00
|
|
|
|
2010-09-09 16:38:18 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-11-06 16:28:37 -08:00
|
|
|
|
2020-08-11 18:33:14 -07:00
|
|
|
|
2010-09-09 16:38:18 -07:00
|
|
|
|
|
|
|
|
|
2016-12-12 16:42:14 -08:00
|
|
|
|
2014-12-10 15:43:01 -08:00
|
|
|
|
2010-09-09 16:38:18 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-03-22 14:43:54 -07:00
|
|
|
|
|
|
|
|
|
2010-09-09 16:38:18 -07:00
|
|
|
|
2009-06-16 15:31:57 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-04-05 16:25:16 -07:00
|
|
|
|
|
|
|
|
|
2009-06-16 15:31:57 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-07-28 15:46:26 -07:00
|
|
|
|
2020-06-03 15:59:01 -07:00
|
|
|
|
2025-03-13 17:05:35 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2009-06-16 15:31:57 -07:00
|
|
|
|
2020-06-03 15:59:01 -07:00
|
|
|
|
2018-04-05 16:25:16 -07:00
|
|
|
|
2022-04-28 23:16:03 -07:00
|
|
|
|
|
|
|
|
|
2025-03-13 17:05:35 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-07-28 15:46:26 -07:00
|
|
|
|
2009-06-16 15:31:57 -07:00
|
|
|
|
|
|
|
|
|
2016-05-19 17:13:38 -07:00
|
|
|
|
2023-01-13 11:12:14 +00:00
|
|
|
|
2009-06-16 15:32:02 -07:00
|
|
|
|
2016-05-19 17:13:38 -07:00
|
|
|
|
2009-06-16 15:32:02 -07:00
|
|
|
|
2020-04-01 21:09:47 -07:00
|
|
|
|
2023-01-13 11:12:12 +00:00
|
|
|
|
2020-04-01 21:09:47 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-01-13 11:12:12 +00:00
|
|
|
|
2020-04-01 21:09:47 -07:00
|
|
|
|
2009-06-16 15:32:02 -07:00
|
|
|
|
2009-06-16 15:32:02 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-01-13 11:12:16 +00:00
|
|
|
|
2009-06-16 15:32:02 -07:00
|
|
|
|
2020-04-01 21:09:47 -07:00
|
|
|
|
|
|
|
|
|
2009-06-16 15:32:02 -07:00
|
|
|
|
2023-01-13 11:12:16 +00:00
|
|
|
|
2011-01-13 15:46:49 -08:00
|
|
|
|
2014-07-30 16:08:24 -07:00
|
|
|
|
|
|
|
|
|
2011-01-13 15:46:49 -08:00
|
|
|
|
2023-01-13 11:12:14 +00:00
|
|
|
|
2023-01-13 11:12:16 +00:00
|
|
|
|
2023-01-13 11:12:14 +00:00
|
|
|
|
mm/page_alloc: only set ALLOC_HIGHATOMIC for __GPF_HIGH allocations
Commit 524c48072e56 ("mm/page_alloc: rename ALLOC_HIGH to
ALLOC_MIN_RESERVE") is the start of a series that explains how __GFP_HIGH,
which implies ALLOC_MIN_RESERVE, is going to be used instead of
__GFP_ATOMIC for high atomic reserves.
Commit eb2e2b425c69 ("mm/page_alloc: explicitly record high-order atomic
allocations in alloc_flags") introduced ALLOC_HIGHATOMIC for such
allocations of order higher than 0. It still used __GFP_ATOMIC, though.
Then, commit 1ebbb21811b7 ("mm/page_alloc: explicitly define how
__GFP_HIGH non-blocking allocations accesses reserves") just turned that
check for !__GFP_DIRECT_RECLAIM, ignoring that high atomic reserves were
expected to test for __GFP_HIGH.
This leads to high atomic reserves being added for high-order GFP_NOWAIT
allocations and others that clear __GFP_DIRECT_RECLAIM, which is
unexpected. Later, those reserves lead to 0-order allocations going to
the slow path and starting reclaim.
From /proc/pagetypeinfo, without the patch:
Node 0, zone DMA, type HighAtomic 0 0 0 0 0 0 0 0 0 0 0
Node 0, zone DMA32, type HighAtomic 1 8 10 9 7 3 0 0 0 0 0
Node 0, zone Normal, type HighAtomic 64 20 12 5 0 0 0 0 0 0 0
With the patch:
Node 0, zone DMA, type HighAtomic 0 0 0 0 0 0 0 0 0 0 0
Node 0, zone DMA32, type HighAtomic 0 0 0 0 0 0 0 0 0 0 0
Node 0, zone Normal, type HighAtomic 0 0 0 0 0 0 0 0 0 0 0
Link: https://lkml.kernel.org/r/20250814172245.1259625-1-cascardo@igalia.com
Fixes: 1ebbb21811b7 ("mm/page_alloc: explicitly define how __GFP_HIGH non-blocking allocations accesses reserves")
Signed-off-by: Thadeu Lima de Souza Cascardo <cascardo@igalia.com>
Tested-by: Helen Koike <koike@igalia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Tested-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: NeilBrown <neilb@suse.de>
Cc: Thierry Reding <thierry.reding@gmail.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Zi Yan <ziy@nvidia.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-08-14 14:22:45 -03:00
|
|
|
|
2023-01-13 11:12:14 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2009-06-16 15:32:02 -07:00
|
|
|
|
2023-01-13 11:12:16 +00:00
|
|
|
|
|
|
|
|
|
2025-04-24 16:28:05 -04:00
|
|
|
|
2009-06-16 15:32:02 -07:00
|
|
|
|
2023-01-13 11:12:16 +00:00
|
|
|
|
|
|
|
|
|
2024-06-10 20:20:18 +01:00
|
|
|
|
2023-01-13 11:12:13 +00:00
|
|
|
|
2009-06-16 15:32:02 -07:00
|
|
|
|
2021-05-04 18:39:00 -07:00
|
|
|
|
2020-08-06 23:26:04 -07:00
|
|
|
|
mm: page_alloc: defrag_mode
The page allocator groups requests by migratetype to stave off
fragmentation. However, in practice this is routinely defeated by the
fact that it gives up *before* invoking reclaim and compaction - which may
well produce suitable pages. As a result, fragmentation of physical
memory is a common ongoing process in many load scenarios.
Fragmentation deteriorates compaction's ability to produce huge pages.
Depending on the lifetime of the fragmenting allocations, those effects
can be long-lasting or even permanent, requiring drastic measures like
forcible idle states or even reboots as the only reliable ways to recover
the address space for THP production.
In a kernel build test with supplemental THP pressure, the THP allocation
rate steadily declines over 15 runs:
thp_fault_alloc
61988
56474
57258
50187
52388
55409
52925
47648
43669
40621
36077
41721
36685
34641
33215
This is a hurdle in adopting THP in any environment where hosts are shared
between multiple overlapping workloads (cloud environments), and rarely
experience true idle periods. To make THP a reliable and predictable
optimization, there needs to be a stronger guarantee to avoid such
fragmentation.
Introduce defrag_mode. When enabled, reclaim/compaction is invoked to its
full extent *before* falling back. Specifically, ALLOC_NOFRAGMENT is
enforced on the allocator fastpath and the reclaiming slowpath.
For now, fallbacks are permitted to avert OOMs. There is a plan to add
defrag_mode=2 to prefer OOMs over fragmentation, but this requires
additional prep work in compaction and the reserve management to make it
ready for all possible allocation contexts.
The following test results are from a kernel build with periodic bursts of
THP allocations, over 15 runs:
vanilla defrag_mode=1
@claimer[unmovable]: 189 103
@claimer[movable]: 92 103
@claimer[reclaimable]: 207 61
@pollute[unmovable from movable]: 25 0
@pollute[unmovable from reclaimable]: 28 0
@pollute[movable from unmovable]: 38835 0
@pollute[movable from reclaimable]: 147136 0
@pollute[reclaimable from unmovable]: 178 0
@pollute[reclaimable from movable]: 33 0
@steal[unmovable from movable]: 11 0
@steal[unmovable from reclaimable]: 5 0
@steal[reclaimable from unmovable]: 107 0
@steal[reclaimable from movable]: 90 0
@steal[movable from reclaimable]: 354 0
@steal[movable from unmovable]: 130 0
Both types of polluting fallbacks are eliminated in this workload.
Interestingly, whole block conversions are reduced as well. This is
because once a block is claimed for a type, its empty space remains
available for future allocations, instead of being padded with fallbacks;
this allows the native type to group up instead of spreading out to new
blocks. The assumption in the allocator has been that pollution from
movable allocations is less harmful than from other types, since they can
be reclaimed or migrated out should the space be needed. However, since
fallbacks occur *before* reclaim/compaction is invoked, movable pollution
will still cause non-movable allocations to spread out and claim more
blocks.
Without fragmentation, THP rates hold steady with defrag_mode=1:
thp_fault_alloc
32478
20725
45045
32130
14018
21711
40791
29134
34458
45381
28305
17265
22584
28454
30850
While the downward trend is eliminated, the keen reader will of course
notice that the baseline rate is much smaller than the vanilla kernel's to
begin with. This is due to deficiencies in how reclaim and compaction are
currently driven: ALLOC_NOFRAGMENT increases the extent to which smaller
allocations are competing with THPs for pageblocks, while making no effort
themselves to reclaim or compact beyond their own request size. This
effect already exists with the current usage of ALLOC_NOFRAGMENT, but is
amplified by defrag_mode insisting on whole block stealing much more
strongly.
Subsequent patches will address defrag_mode reclaim strategy to raise the
THP success baseline above the vanilla kernel.
Link: https://lkml.kernel.org/r/20250313210647.1314586-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-03-13 17:05:34 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2009-06-16 15:32:02 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-09-06 16:24:50 -07:00
|
|
|
|
mm: sl[au]b: add knowledge of PFMEMALLOC reserve pages
When a user or administrator requires swap for their application, they
create a swap partition and file, format it with mkswap and activate it
with swapon. Swap over the network is considered as an option in diskless
systems. The two likely scenarios are when blade servers are used as part
of a cluster where the form factor or maintenance costs do not allow the
use of disks and thin clients.
The Linux Terminal Server Project recommends the use of the Network Block
Device (NBD) for swap according to the manual at
https://sourceforge.net/projects/ltsp/files/Docs-Admin-Guide/LTSPManual.pdf/download
There is also documentation and tutorials on how to setup swap over NBD at
places like https://help.ubuntu.com/community/UbuntuLTSP/EnableNBDSWAP The
nbd-client also documents the use of NBD as swap. Despite this, the fact
is that a machine using NBD for swap can deadlock within minutes if swap
is used intensively. This patch series addresses the problem.
The core issue is that network block devices do not use mempools like
normal block devices do. As the host cannot control where they receive
packets from, they cannot reliably work out in advance how much memory
they might need. Some years ago, Peter Zijlstra developed a series of
patches that supported swap over an NFS that at least one distribution is
carrying within their kernels. This patch series borrows very heavily
from Peter's work to support swapping over NBD as a pre-requisite to
supporting swap-over-NFS. The bulk of the complexity is concerned with
preserving memory that is allocated from the PFMEMALLOC reserves for use
by the network layer which is needed for both NBD and NFS.
Patch 1 adds knowledge of the PFMEMALLOC reserves to SLAB and SLUB to
preserve access to pages allocated under low memory situations
to callers that are freeing memory.
Patch 2 optimises the SLUB fast path to avoid pfmemalloc checks
Patch 3 introduces __GFP_MEMALLOC to allow access to the PFMEMALLOC
reserves without setting PFMEMALLOC.
Patch 4 opens the possibility for softirqs to use PFMEMALLOC reserves
for later use by network packet processing.
Patch 5 only sets page->pfmemalloc when ALLOC_NO_WATERMARKS was required
Patch 6 ignores memory policies when ALLOC_NO_WATERMARKS is set.
Patches 7-12 allows network processing to use PFMEMALLOC reserves when
the socket has been marked as being used by the VM to clean pages. If
packets are received and stored in pages that were allocated under
low-memory situations and are unrelated to the VM, the packets
are dropped.
Patch 11 reintroduces __skb_alloc_page which the networking
folk may object to but is needed in some cases to propogate
pfmemalloc from a newly allocated page to an skb. If there is a
strong objection, this patch can be dropped with the impact being
that swap-over-network will be slower in some cases but it should
not fail.
Patch 13 is a micro-optimisation to avoid a function call in the
common case.
Patch 14 tags NBD sockets as being SOCK_MEMALLOC so they can use
PFMEMALLOC if necessary.
Patch 15 notes that it is still possible for the PFMEMALLOC reserve
to be depleted. To prevent this, direct reclaimers get throttled on
a waitqueue if 50% of the PFMEMALLOC reserves are depleted. It is
expected that kswapd and the direct reclaimers already running
will clean enough pages for the low watermark to be reached and
the throttled processes are woken up.
Patch 16 adds a statistic to track how often processes get throttled
Some basic performance testing was run using kernel builds, netperf on
loopback for UDP and TCP, hackbench (pipes and sockets), iozone and
sysbench. Each of them were expected to use the sl*b allocators
reasonably heavily but there did not appear to be significant performance
variances.
For testing swap-over-NBD, a machine was booted with 2G of RAM with a
swapfile backed by NBD. 8*NUM_CPU processes were started that create
anonymous memory mappings and read them linearly in a loop. The total
size of the mappings were 4*PHYSICAL_MEMORY to use swap heavily under
memory pressure.
Without the patches and using SLUB, the machine locks up within minutes
and runs to completion with them applied. With SLAB, the story is
different as an unpatched kernel run to completion. However, the patched
kernel completed the test 45% faster.
MICRO
3.5.0-rc2 3.5.0-rc2
vanilla swapnbd
Unrecognised test vmscan-anon-mmap-write
MMTests Statistics: duration
Sys Time Running Test (seconds) 197.80 173.07
User+Sys Time Running Test (seconds) 206.96 182.03
Total Elapsed Time (seconds) 3240.70 1762.09
This patch: mm: sl[au]b: add knowledge of PFMEMALLOC reserve pages
Allocations of pages below the min watermark run a risk of the machine
hanging due to a lack of memory. To prevent this, only callers who have
PF_MEMALLOC or TIF_MEMDIE set and are not processing an interrupt are
allowed to allocate with ALLOC_NO_WATERMARKS. Once they are allocated to
a slab though, nothing prevents other callers consuming free objects
within those slabs. This patch limits access to slab pages that were
alloced from the PFMEMALLOC reserves.
When this patch is applied, pages allocated from below the low watermark
are returned with page->pfmemalloc set and it is up to the caller to
determine how the page should be protected. SLAB restricts access to any
page with page->pfmemalloc set to callers which are known to able to
access the PFMEMALLOC reserve. If one is not available, an attempt is
made to allocate a new page rather than use a reserve. SLUB is a bit more
relaxed in that it only records if the current per-CPU page was allocated
from PFMEMALLOC reserve and uses another partial slab if the caller does
not have the necessary GFP or process flags. This was found to be
sufficient in tests to avoid hangs due to SLUB generally maintaining
smaller lists than SLAB.
In low-memory conditions it does mean that !PFMEMALLOC allocators can fail
a slab allocation even though free objects are available because they are
being preserved for callers that are freeing pages.
[a.p.zijlstra@chello.nl: Original implementation]
[sebastian@breakpoint.cc: Correct order of page flag clearing]
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: David Miller <davem@davemloft.net>
Cc: Neil Brown <neilb@suse.de>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Christie <michaelc@cs.wisc.edu>
Cc: Eric B Munson <emunson@mgebm.net>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Sebastian Andrzej Siewior <sebastian@breakpoint.cc>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Christoph Lameter <cl@linux.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-07-31 16:43:58 -07:00
|
|
|
|
2017-09-06 16:24:50 -07:00
|
|
|
|
2016-07-28 15:49:13 -07:00
|
|
|
|
|
|
|
|
|
2017-09-06 16:24:50 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-07-28 15:49:13 -07:00
|
|
|
|
2017-09-06 16:24:50 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: sl[au]b: add knowledge of PFMEMALLOC reserve pages
When a user or administrator requires swap for their application, they
create a swap partition and file, format it with mkswap and activate it
with swapon. Swap over the network is considered as an option in diskless
systems. The two likely scenarios are when blade servers are used as part
of a cluster where the form factor or maintenance costs do not allow the
use of disks and thin clients.
The Linux Terminal Server Project recommends the use of the Network Block
Device (NBD) for swap according to the manual at
https://sourceforge.net/projects/ltsp/files/Docs-Admin-Guide/LTSPManual.pdf/download
There is also documentation and tutorials on how to setup swap over NBD at
places like https://help.ubuntu.com/community/UbuntuLTSP/EnableNBDSWAP The
nbd-client also documents the use of NBD as swap. Despite this, the fact
is that a machine using NBD for swap can deadlock within minutes if swap
is used intensively. This patch series addresses the problem.
The core issue is that network block devices do not use mempools like
normal block devices do. As the host cannot control where they receive
packets from, they cannot reliably work out in advance how much memory
they might need. Some years ago, Peter Zijlstra developed a series of
patches that supported swap over an NFS that at least one distribution is
carrying within their kernels. This patch series borrows very heavily
from Peter's work to support swapping over NBD as a pre-requisite to
supporting swap-over-NFS. The bulk of the complexity is concerned with
preserving memory that is allocated from the PFMEMALLOC reserves for use
by the network layer which is needed for both NBD and NFS.
Patch 1 adds knowledge of the PFMEMALLOC reserves to SLAB and SLUB to
preserve access to pages allocated under low memory situations
to callers that are freeing memory.
Patch 2 optimises the SLUB fast path to avoid pfmemalloc checks
Patch 3 introduces __GFP_MEMALLOC to allow access to the PFMEMALLOC
reserves without setting PFMEMALLOC.
Patch 4 opens the possibility for softirqs to use PFMEMALLOC reserves
for later use by network packet processing.
Patch 5 only sets page->pfmemalloc when ALLOC_NO_WATERMARKS was required
Patch 6 ignores memory policies when ALLOC_NO_WATERMARKS is set.
Patches 7-12 allows network processing to use PFMEMALLOC reserves when
the socket has been marked as being used by the VM to clean pages. If
packets are received and stored in pages that were allocated under
low-memory situations and are unrelated to the VM, the packets
are dropped.
Patch 11 reintroduces __skb_alloc_page which the networking
folk may object to but is needed in some cases to propogate
pfmemalloc from a newly allocated page to an skb. If there is a
strong objection, this patch can be dropped with the impact being
that swap-over-network will be slower in some cases but it should
not fail.
Patch 13 is a micro-optimisation to avoid a function call in the
common case.
Patch 14 tags NBD sockets as being SOCK_MEMALLOC so they can use
PFMEMALLOC if necessary.
Patch 15 notes that it is still possible for the PFMEMALLOC reserve
to be depleted. To prevent this, direct reclaimers get throttled on
a waitqueue if 50% of the PFMEMALLOC reserves are depleted. It is
expected that kswapd and the direct reclaimers already running
will clean enough pages for the low watermark to be reached and
the throttled processes are woken up.
Patch 16 adds a statistic to track how often processes get throttled
Some basic performance testing was run using kernel builds, netperf on
loopback for UDP and TCP, hackbench (pipes and sockets), iozone and
sysbench. Each of them were expected to use the sl*b allocators
reasonably heavily but there did not appear to be significant performance
variances.
For testing swap-over-NBD, a machine was booted with 2G of RAM with a
swapfile backed by NBD. 8*NUM_CPU processes were started that create
anonymous memory mappings and read them linearly in a loop. The total
size of the mappings were 4*PHYSICAL_MEMORY to use swap heavily under
memory pressure.
Without the patches and using SLUB, the machine locks up within minutes
and runs to completion with them applied. With SLAB, the story is
different as an unpatched kernel run to completion. However, the patched
kernel completed the test 45% faster.
MICRO
3.5.0-rc2 3.5.0-rc2
vanilla swapnbd
Unrecognised test vmscan-anon-mmap-write
MMTests Statistics: duration
Sys Time Running Test (seconds) 197.80 173.07
User+Sys Time Running Test (seconds) 206.96 182.03
Total Elapsed Time (seconds) 3240.70 1762.09
This patch: mm: sl[au]b: add knowledge of PFMEMALLOC reserve pages
Allocations of pages below the min watermark run a risk of the machine
hanging due to a lack of memory. To prevent this, only callers who have
PF_MEMALLOC or TIF_MEMDIE set and are not processing an interrupt are
allowed to allocate with ALLOC_NO_WATERMARKS. Once they are allocated to
a slab though, nothing prevents other callers consuming free objects
within those slabs. This patch limits access to slab pages that were
alloced from the PFMEMALLOC reserves.
When this patch is applied, pages allocated from below the low watermark
are returned with page->pfmemalloc set and it is up to the caller to
determine how the page should be protected. SLAB restricts access to any
page with page->pfmemalloc set to callers which are known to able to
access the PFMEMALLOC reserve. If one is not available, an attempt is
made to allocate a new page rather than use a reserve. SLUB is a bit more
relaxed in that it only records if the current per-CPU page was allocated
from PFMEMALLOC reserve and uses another partial slab if the caller does
not have the necessary GFP or process flags. This was found to be
sufficient in tests to avoid hangs due to SLUB generally maintaining
smaller lists than SLAB.
In low-memory conditions it does mean that !PFMEMALLOC allocators can fail
a slab allocation even though free objects are available because they are
being preserved for callers that are freeing pages.
[a.p.zijlstra@chello.nl: Original implementation]
[sebastian@breakpoint.cc: Correct order of page flag clearing]
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: David Miller <davem@davemloft.net>
Cc: Neil Brown <neilb@suse.de>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Christie <michaelc@cs.wisc.edu>
Cc: Eric B Munson <emunson@mgebm.net>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Sebastian Andrzej Siewior <sebastian@breakpoint.cc>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Christoph Lameter <cl@linux.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-07-31 16:43:58 -07:00
|
|
|
|
|
|
|
|
|
2016-05-20 16:57:00 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-05-03 14:52:16 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-05-20 16:57:00 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-10-07 17:00:40 -07:00
|
|
|
|
2016-05-20 16:57:00 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-10-26 15:03:31 -07:00
|
|
|
|
2016-05-20 16:57:00 -07:00
|
|
|
|
2016-10-07 17:00:40 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-11-24 16:27:25 +05:30
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-05-20 16:57:00 -07:00
|
|
|
|
2016-07-28 15:47:05 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-05-20 16:57:00 -07:00
|
|
|
|
2020-06-03 15:59:01 -07:00
|
|
|
|
|
|
|
|
|
2016-05-20 16:57:00 -07:00
|
|
|
|
2016-05-20 16:57:03 -07:00
|
|
|
|
2017-02-22 15:42:00 -08:00
|
|
|
|
|
|
|
|
|
2016-05-20 16:57:00 -07:00
|
|
|
|
2024-08-22 17:26:12 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-07-28 15:47:31 -07:00
|
|
|
|
|
|
|
|
|
2016-05-20 16:57:00 -07:00
|
|
|
|
|
|
|
|
|
2017-05-03 14:52:16 -07:00
|
|
|
|
|
|
|
|
|
2016-05-20 16:57:00 -07:00
|
|
|
|
2017-02-22 15:42:00 -08:00
|
|
|
|
2020-06-03 15:59:01 -07:00
|
|
|
|
2017-02-22 15:42:00 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-10-26 15:03:31 -07:00
|
|
|
|
2021-11-05 13:42:38 -07:00
|
|
|
|
2016-05-20 16:57:00 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-10-26 15:03:31 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-11-24 16:27:25 +05:30
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-10-26 15:03:31 -07:00
|
|
|
|
2016-05-20 16:57:00 -07:00
|
|
|
|
|
|
|
|
|
mm, page_alloc: fix more premature OOM due to race with cpuset update
I would like to stress that this patchset aims to fix issues and cleanup
the code *within the existing documented semantics*, i.e. patch 1
ignores mempolicy restrictions if the set of allowed nodes has no
intersection with set of nodes allowed by cpuset. I believe discussing
potential changes of the semantics can be better done once we have a
baseline with no known bugs of the current semantics.
I've recently summarized the cpuset/mempolicy issues in a LSF/MM
proposal [1] and the discussion itself [2]. I've been trying to rewrite
the handling as proposed, with the idea that changing semantics to make
all mempolicies static wrt cpuset updates (and discarding the relative
and default modes) can be tried on top, as there's a high risk of being
rejected/reverted because somebody might still care about the removed
modes.
However I haven't yet figured out how to properly:
1) make mempolicies swappable instead of rebinding in place. I thought
mbind() already works that way and uses refcounting to avoid
use-after-free of the old policy by a parallel allocation, but turns
out true refcounting is only done for shared (shmem) mempolicies, and
the actual protection for mbind() comes from mmap_sem. Extending the
refcounting means more overhead in allocator hot path. Also swapping
whole mempolicies means that we have to allocate the new ones, which
can fail, and reverting of the partially done work also means
allocating (note that mbind() doesn't care and will just leave part
of the range updated and part not updated when returning -ENOMEM...).
2) make cpuset's task->mems_allowed also swappable (after converting it
from nodemask to zonelist, which is the easy part) for mostly the
same reasons.
The good news is that while trying to do the above, I've at least
figured out how to hopefully close the remaining premature OOM's, and do
a buch of cleanups on top, removing quite some of the code that was also
supposed to prevent the cpuset update races, but doesn't work anymore
nowadays. This should fix the most pressing concerns with this topic
and give us a better baseline before either proceeding with the original
proposal, or pushing a change of semantics that removes the problem 1)
above. I'd be then fine with trying to change the semantic first and
rewrite later.
Patchset has been tested with the LTP cpuset01 stress test.
[1] https://lkml.kernel.org/r/4c44a589-5fd8-08d0-892c-e893bb525b71@suse.cz
[2] https://lwn.net/Articles/717797/
[3] https://marc.info/?l=linux-mm&m=149191957922828&w=2
This patch (of 6):
Commit e47483bca2cc ("mm, page_alloc: fix premature OOM when racing with
cpuset mems update") has fixed known recent regressions found by LTP's
cpuset01 testcase. I have however found that by modifying the testcase
to use per-vma mempolicies via bind(2) instead of per-task mempolicies
via set_mempolicy(2), the premature OOM still happens and the issue is
much older.
The root of the problem is that the cpuset's mems_allowed and
mempolicy's nodemask can temporarily have no intersection, thus
get_page_from_freelist() cannot find any usable zone. The current
semantic for empty intersection is to ignore mempolicy's nodemask and
honour cpuset restrictions. This is checked in node_zonelist(), but the
racy update can happen after we already passed the check. Such races
should be protected by the seqlock task->mems_allowed_seq, but it
doesn't work here, because 1) mpol_rebind_mm() does not happen under
seqlock for write, and doing so would lead to deadlock, as it takes
mmap_sem for write, while the allocation can have mmap_sem for read when
it's taking the seqlock for read. And 2) the seqlock cookie of callers
of node_zonelist() (alloc_pages_vma() and alloc_pages_current()) is
different than the one of __alloc_pages_slowpath(), so there's still a
potential race window.
This patch fixes the issue by having __alloc_pages_slowpath() check for
empty intersection of cpuset and ac->nodemask before OOM or allocation
failure. If it's indeed empty, the nodemask is ignored and allocation
retried, which mimics node_zonelist(). This works fine, because almost
all callers of __alloc_pages_nodemask are obtaining the nodemask via
node_zonelist(). The only exception is new_node_page() from hotplug,
where the potential violation of nodemask isn't an issue, as there's
already a fallback allocation attempt without any nodemask. If there's
a future caller that needs to have its specific nodemask honoured over
task's cpuset restrictions, we'll have to e.g. add a gfp flag for that.
Link: http://lkml.kernel.org/r/20170517081140.30654-2-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Li Zefan <lizefan@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: David Rientjes <rientjes@google.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Anshuman Khandual <khandual@linux.vnet.ibm.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Dimitri Sivanich <sivanich@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 15:39:56 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2009-06-16 15:31:57 -07:00
|
|
|
|
|
|
|
|
|
2015-02-11 15:25:41 -08:00
|
|
|
|
2009-06-16 15:31:57 -07:00
|
|
|
|
2015-11-06 16:28:21 -08:00
|
|
|
|
mm, vmscan: prevent infinite loop for costly GFP_NOIO | __GFP_RETRY_MAYFAIL allocations
Sven reports an infinite loop in __alloc_pages_slowpath() for costly order
__GFP_RETRY_MAYFAIL allocations that are also GFP_NOIO. Such combination
can happen in a suspend/resume context where a GFP_KERNEL allocation can
have __GFP_IO masked out via gfp_allowed_mask.
Quoting Sven:
1. try to do a "costly" allocation (order > PAGE_ALLOC_COSTLY_ORDER)
with __GFP_RETRY_MAYFAIL set.
2. page alloc's __alloc_pages_slowpath tries to get a page from the
freelist. This fails because there is nothing free of that costly
order.
3. page alloc tries to reclaim by calling __alloc_pages_direct_reclaim,
which bails out because a zone is ready to be compacted; it pretends
to have made a single page of progress.
4. page alloc tries to compact, but this always bails out early because
__GFP_IO is not set (it's not passed by the snd allocator, and even
if it were, we are suspending so the __GFP_IO flag would be cleared
anyway).
5. page alloc believes reclaim progress was made (because of the
pretense in item 3) and so it checks whether it should retry
compaction. The compaction retry logic thinks it should try again,
because:
a) reclaim is needed because of the early bail-out in item 4
b) a zonelist is suitable for compaction
6. goto 2. indefinite stall.
(end quote)
The immediate root cause is confusing the COMPACT_SKIPPED returned from
__alloc_pages_direct_compact() (step 4) due to lack of __GFP_IO to be
indicating a lack of order-0 pages, and in step 5 evaluating that in
should_compact_retry() as a reason to retry, before incrementing and
limiting the number of retries. There are however other places that
wrongly assume that compaction can happen while we lack __GFP_IO.
To fix this, introduce gfp_compaction_allowed() to abstract the __GFP_IO
evaluation and switch the open-coded test in try_to_compact_pages() to use
it.
Also use the new helper in:
- compaction_ready(), which will make reclaim not bail out in step 3, so
there's at least one attempt to actually reclaim, even if chances are
small for a costly order
- in_reclaim_compaction() which will make should_continue_reclaim()
return false and we don't over-reclaim unnecessarily
- in __alloc_pages_slowpath() to set a local variable can_compact,
which is then used to avoid retrying reclaim/compaction for costly
allocations (step 5) if we can't compact and also to skip the early
compaction attempt that we do in some cases
Link: https://lkml.kernel.org/r/20240221114357.13655-2-vbabka@suse.cz
Fixes: 3250845d0526 ("Revert "mm, oom: prevent premature OOM killer invocation for high order request"")
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Reported-by: Sven van Ashbrook <svenva@chromium.org>
Closes: https://lore.kernel.org/all/CAG-rBihs_xMKb3wrMO1%2B-%2Bp4fowP9oy1pa_OTkfxBzPUVOZF%2Bg@mail.gmail.com/
Tested-by: Karthikeyan Ramasubramanian <kramasub@chromium.org>
Cc: Brian Geffon <bgeffon@google.com>
Cc: Curtis Malainey <cujomalainey@chromium.org>
Cc: Jaroslav Kysela <perex@perex.cz>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Takashi Iwai <tiwai@suse.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-21 12:43:58 +01:00
|
|
|
|
2024-08-31 08:28:23 +12:00
|
|
|
|
mm, compaction: restrict async compaction to pageblocks of same migratetype
The migrate scanner in async compaction is currently limited to
MIGRATE_MOVABLE pageblocks. This is a heuristic intended to reduce
latency, based on the assumption that non-MOVABLE pageblocks are
unlikely to contain movable pages.
However, with the exception of THP's, most high-order allocations are
not movable. Should the async compaction succeed, this increases the
chance that the non-MOVABLE allocations will fallback to a MOVABLE
pageblock, making the long-term fragmentation worse.
This patch attempts to help the situation by changing async direct
compaction so that the migrate scanner only scans the pageblocks of the
requested migratetype. If it's a non-MOVABLE type and there are such
pageblocks that do contain movable pages, chances are that the
allocation can succeed within one of such pageblocks, removing the need
for a fallback. If that fails, the subsequent sync attempt will ignore
this restriction.
In testing based on 4.9 kernel with stress-highalloc from mmtests
configured for order-4 GFP_KERNEL allocations, this patch has reduced
the number of unmovable allocations falling back to movable pageblocks
by 30%. The number of movable allocations falling back is reduced by
12%.
Link: http://lkml.kernel.org/r/20170307131545.28577-8-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-08 15:54:49 -07:00
|
|
|
|
2009-06-16 15:31:57 -07:00
|
|
|
|
2016-05-19 17:13:38 -07:00
|
|
|
|
2009-06-16 15:31:57 -07:00
|
|
|
|
2017-01-24 15:18:38 -08:00
|
|
|
|
2016-05-20 16:56:53 -07:00
|
|
|
|
2017-01-24 15:18:38 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/page_alloc: fix race condition between build_all_zonelists and page allocation
Patrick Daly reported the following problem;
NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK] - before offline operation
[0] - ZONE_MOVABLE
[1] - ZONE_NORMAL
[2] - NULL
For a GFP_KERNEL allocation, alloc_pages_slowpath() will save the
offset of ZONE_NORMAL in ac->preferred_zoneref. If a concurrent
memory_offline operation removes the last page from ZONE_MOVABLE,
build_all_zonelists() & build_zonerefs_node() will update
node_zonelists as shown below. Only populated zones are added.
NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK] - after offline operation
[0] - ZONE_NORMAL
[1] - NULL
[2] - NULL
The race is simple -- page allocation could be in progress when a memory
hot-remove operation triggers a zonelist rebuild that removes zones. The
allocation request will still have a valid ac->preferred_zoneref that is
now pointing to NULL and triggers an OOM kill.
This problem probably always existed but may be slightly easier to trigger
due to 6aa303defb74 ("mm, vmscan: only allocate and reclaim from zones
with pages managed by the buddy allocator") which distinguishes between
zones that are completely unpopulated versus zones that have valid pages
not managed by the buddy allocator (e.g. reserved, memblock, ballooning
etc). Memory hotplug had multiple stages with timing considerations
around managed/present page updates, the zonelist rebuild and the zone
span updates. As David Hildenbrand puts it
memory offlining adjusts managed+present pages of the zone
essentially in one go. If after the adjustments, the zone is no
longer populated (present==0), we rebuild the zone lists.
Once that's done, we try shrinking the zone (start+spanned
pages) -- which results in zone_start_pfn == 0 if there are no
more pages. That happens *after* rebuilding the zonelists via
remove_pfn_range_from_zone().
The only requirement to fix the race is that a page allocation request
identifies when a zonelist rebuild has happened since the allocation
request started and no page has yet been allocated. Use a seqlock_t to
track zonelist updates with a lockless read-side of the zonelist and
protecting the rebuild and update of the counter with a spinlock.
[akpm@linux-foundation.org: make zonelist_update_seq static]
Link: https://lkml.kernel.org/r/20220824110900.vh674ltxmzb3proq@techsingularity.net
Fixes: 6aa303defb74 ("mm, vmscan: only allocate and reclaim from zones with pages managed by the buddy allocator")
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Reported-by: Patrick Daly <quic_pdaly@quicinc.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: David Hildenbrand <david@redhat.com>
Cc: <stable@vger.kernel.org> [4.9+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-08-24 12:14:50 +01:00
|
|
|
|
2017-09-06 16:24:50 -07:00
|
|
|
|
2005-11-13 16:06:43 -08:00
|
|
|
|
2024-08-31 08:28:23 +12:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/page_alloc: fix race condition between build_all_zonelists and page allocation
Patrick Daly reported the following problem;
NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK] - before offline operation
[0] - ZONE_MOVABLE
[1] - ZONE_NORMAL
[2] - NULL
For a GFP_KERNEL allocation, alloc_pages_slowpath() will save the
offset of ZONE_NORMAL in ac->preferred_zoneref. If a concurrent
memory_offline operation removes the last page from ZONE_MOVABLE,
build_all_zonelists() & build_zonerefs_node() will update
node_zonelists as shown below. Only populated zones are added.
NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK] - after offline operation
[0] - ZONE_NORMAL
[1] - NULL
[2] - NULL
The race is simple -- page allocation could be in progress when a memory
hot-remove operation triggers a zonelist rebuild that removes zones. The
allocation request will still have a valid ac->preferred_zoneref that is
now pointing to NULL and triggers an OOM kill.
This problem probably always existed but may be slightly easier to trigger
due to 6aa303defb74 ("mm, vmscan: only allocate and reclaim from zones
with pages managed by the buddy allocator") which distinguishes between
zones that are completely unpopulated versus zones that have valid pages
not managed by the buddy allocator (e.g. reserved, memblock, ballooning
etc). Memory hotplug had multiple stages with timing considerations
around managed/present page updates, the zonelist rebuild and the zone
span updates. As David Hildenbrand puts it
memory offlining adjusts managed+present pages of the zone
essentially in one go. If after the adjustments, the zone is no
longer populated (present==0), we rebuild the zone lists.
Once that's done, we try shrinking the zone (start+spanned
pages) -- which results in zone_start_pfn == 0 if there are no
more pages. That happens *after* rebuilding the zonelists via
remove_pfn_range_from_zone().
The only requirement to fix the race is that a page allocation request
identifies when a zonelist rebuild has happened since the allocation
request started and no page has yet been allocated. Use a seqlock_t to
track zonelist updates with a lockless read-side of the zonelist and
protecting the rebuild and update of the counter with a spinlock.
[akpm@linux-foundation.org: make zonelist_update_seq static]
Link: https://lkml.kernel.org/r/20220824110900.vh674ltxmzb3proq@techsingularity.net
Fixes: 6aa303defb74 ("mm, vmscan: only allocate and reclaim from zones with pages managed by the buddy allocator")
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Reported-by: Patrick Daly <quic_pdaly@quicinc.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: David Hildenbrand <david@redhat.com>
Cc: <stable@vger.kernel.org> [4.9+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-08-24 12:14:50 +01:00
|
|
|
|
2017-01-24 15:18:38 -08:00
|
|
|
|
|
|
|
|
|
2025-02-27 11:41:29 +08:00
|
|
|
|
2017-01-24 15:18:38 -08:00
|
|
|
|
|
|
|
|
|
mm/page_alloc: fix race condition between build_all_zonelists and page allocation
Patrick Daly reported the following problem;
NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK] - before offline operation
[0] - ZONE_MOVABLE
[1] - ZONE_NORMAL
[2] - NULL
For a GFP_KERNEL allocation, alloc_pages_slowpath() will save the
offset of ZONE_NORMAL in ac->preferred_zoneref. If a concurrent
memory_offline operation removes the last page from ZONE_MOVABLE,
build_all_zonelists() & build_zonerefs_node() will update
node_zonelists as shown below. Only populated zones are added.
NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK] - after offline operation
[0] - ZONE_NORMAL
[1] - NULL
[2] - NULL
The race is simple -- page allocation could be in progress when a memory
hot-remove operation triggers a zonelist rebuild that removes zones. The
allocation request will still have a valid ac->preferred_zoneref that is
now pointing to NULL and triggers an OOM kill.
This problem probably always existed but may be slightly easier to trigger
due to 6aa303defb74 ("mm, vmscan: only allocate and reclaim from zones
with pages managed by the buddy allocator") which distinguishes between
zones that are completely unpopulated versus zones that have valid pages
not managed by the buddy allocator (e.g. reserved, memblock, ballooning
etc). Memory hotplug had multiple stages with timing considerations
around managed/present page updates, the zonelist rebuild and the zone
span updates. As David Hildenbrand puts it
memory offlining adjusts managed+present pages of the zone
essentially in one go. If after the adjustments, the zone is no
longer populated (present==0), we rebuild the zone lists.
Once that's done, we try shrinking the zone (start+spanned
pages) -- which results in zone_start_pfn == 0 if there are no
more pages. That happens *after* rebuilding the zonelists via
remove_pfn_range_from_zone().
The only requirement to fix the race is that a page allocation request
identifies when a zonelist rebuild has happened since the allocation
request started and no page has yet been allocated. Use a seqlock_t to
track zonelist updates with a lockless read-side of the zonelist and
protecting the rebuild and update of the counter with a spinlock.
[akpm@linux-foundation.org: make zonelist_update_seq static]
Link: https://lkml.kernel.org/r/20220824110900.vh674ltxmzb3proq@techsingularity.net
Fixes: 6aa303defb74 ("mm, vmscan: only allocate and reclaim from zones with pages managed by the buddy allocator")
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Reported-by: Patrick Daly <quic_pdaly@quicinc.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: David Hildenbrand <david@redhat.com>
Cc: <stable@vger.kernel.org> [4.9+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-08-24 12:14:50 +01:00
|
|
|
|
2017-02-22 15:46:19 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-01-13 11:12:14 +00:00
|
|
|
|
2017-02-22 15:46:19 -08:00
|
|
|
|
mm, page_alloc: fix premature OOM when racing with cpuset mems update
Ganapatrao Kulkarni reported that the LTP test cpuset01 in stress mode
triggers OOM killer in few seconds, despite lots of free memory. The
test attempts to repeatedly fault in memory in one process in a cpuset,
while changing allowed nodes of the cpuset between 0 and 1 in another
process.
The problem comes from insufficient protection against cpuset changes,
which can cause get_page_from_freelist() to consider all zones as
non-eligible due to nodemask and/or current->mems_allowed. This was
masked in the past by sufficient retries, but since commit 682a3385e773
("mm, page_alloc: inline the fast path of the zonelist iterator") we fix
the preferred_zoneref once, and don't iterate over the whole zonelist in
further attempts, thus the only eligible zones might be placed in the
zonelist before our starting point and we always miss them.
A previous patch fixed this problem for current->mems_allowed. However,
cpuset changes also update the task's mempolicy nodemask. The fix has
two parts. We have to repeat the preferred_zoneref search when we
detect cpuset update by way of seqcount, and we have to check the
seqcount before considering OOM.
[akpm@linux-foundation.org: fix typo in comment]
Link: http://lkml.kernel.org/r/20170120103843.24587-5-vbabka@suse.cz
Fixes: c33d6c06f60f ("mm, page_alloc: avoid looking up the first zone in a zonelist twice")
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Reported-by: Ganapatrao Kulkarni <gpkulkarni@gmail.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-01-24 15:18:41 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-06-03 15:59:01 -07:00
|
|
|
|
2024-07-29 14:47:17 +05:30
|
|
|
|
mm, page_alloc: fix premature OOM when racing with cpuset mems update
Ganapatrao Kulkarni reported that the LTP test cpuset01 in stress mode
triggers OOM killer in few seconds, despite lots of free memory. The
test attempts to repeatedly fault in memory in one process in a cpuset,
while changing allowed nodes of the cpuset between 0 and 1 in another
process.
The problem comes from insufficient protection against cpuset changes,
which can cause get_page_from_freelist() to consider all zones as
non-eligible due to nodemask and/or current->mems_allowed. This was
masked in the past by sufficient retries, but since commit 682a3385e773
("mm, page_alloc: inline the fast path of the zonelist iterator") we fix
the preferred_zoneref once, and don't iterate over the whole zonelist in
further attempts, thus the only eligible zones might be placed in the
zonelist before our starting point and we always miss them.
A previous patch fixed this problem for current->mems_allowed. However,
cpuset changes also update the task's mempolicy nodemask. The fix has
two parts. We have to repeat the preferred_zoneref search when we
detect cpuset update by way of seqcount, and we have to check the
seqcount before considering OOM.
[akpm@linux-foundation.org: fix typo in comment]
Link: http://lkml.kernel.org/r/20170120103843.24587-5-vbabka@suse.cz
Fixes: c33d6c06f60f ("mm, page_alloc: avoid looking up the first zone in a zonelist twice")
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Reported-by: Ganapatrao Kulkarni <gpkulkarni@gmail.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-01-24 15:18:41 -08:00
|
|
|
|
|
|
|
|
|
mm/page_alloc: detect allocation forbidden by cpuset and bail out early
There was a report that starting an Ubuntu in docker while using cpuset
to bind it to movable nodes (a node only has movable zone, like a node
for hotplug or a Persistent Memory node in normal usage) will fail due
to memory allocation failure, and then OOM is involved and many other
innocent processes got killed.
It can be reproduced with command:
$ docker run -it --rm --cpuset-mems 4 ubuntu:latest bash -c "grep Mems_allowed /proc/self/status"
(where node 4 is a movable node)
runc:[2:INIT] invoked oom-killer: gfp_mask=0x500cc2(GFP_HIGHUSER|__GFP_ACCOUNT), order=0, oom_score_adj=0
CPU: 8 PID: 8291 Comm: runc:[2:INIT] Tainted: G W I E 5.8.2-0.g71b519a-default #1 openSUSE Tumbleweed (unreleased)
Hardware name: Dell Inc. PowerEdge R640/0PHYDR, BIOS 2.6.4 04/09/2020
Call Trace:
dump_stack+0x6b/0x88
dump_header+0x4a/0x1e2
oom_kill_process.cold+0xb/0x10
out_of_memory.part.0+0xaf/0x230
out_of_memory+0x3d/0x80
__alloc_pages_slowpath.constprop.0+0x954/0xa20
__alloc_pages_nodemask+0x2d3/0x300
pipe_write+0x322/0x590
new_sync_write+0x196/0x1b0
vfs_write+0x1c3/0x1f0
ksys_write+0xa7/0xe0
do_syscall_64+0x52/0xd0
entry_SYSCALL_64_after_hwframe+0x44/0xa9
Mem-Info:
active_anon:392832 inactive_anon:182 isolated_anon:0
active_file:68130 inactive_file:151527 isolated_file:0
unevictable:2701 dirty:0 writeback:7
slab_reclaimable:51418 slab_unreclaimable:116300
mapped:45825 shmem:735 pagetables:2540 bounce:0
free:159849484 free_pcp:73 free_cma:0
Node 4 active_anon:1448kB inactive_anon:0kB active_file:0kB inactive_file:0kB unevictable:0kB isolated(anon):0kB isolated(file):0kB mapped:0kB dirty:0kB writeback:0kB shmem:0kB shmem_thp: 0kB shmem_pmdmapped: 0kB anon_thp: 0kB writeback_tmp:0kB all_unreclaimable? no
Node 4 Movable free:130021408kB min:9140kB low:139160kB high:269180kB reserved_highatomic:0KB active_anon:1448kB inactive_anon:0kB active_file:0kB inactive_file:0kB unevictable:0kB writepending:0kB present:130023424kB managed:130023424kB mlocked:0kB kernel_stack:0kB pagetables:0kB bounce:0kB free_pcp:292kB local_pcp:84kB free_cma:0kB
lowmem_reserve[]: 0 0 0 0 0
Node 4 Movable: 1*4kB (M) 0*8kB 0*16kB 1*32kB (M) 0*64kB 0*128kB 1*256kB (M) 1*512kB (M) 1*1024kB (M) 0*2048kB 31743*4096kB (M) = 130021156kB
oom-kill:constraint=CONSTRAINT_CPUSET,nodemask=(null),cpuset=docker-9976a269caec812c134fa317f27487ee36e1129beba7278a463dd53e5fb9997b.scope,mems_allowed=4,global_oom,task_memcg=/system.slice/containerd.service,task=containerd,pid=4100,uid=0
Out of memory: Killed process 4100 (containerd) total-vm:4077036kB, anon-rss:51184kB, file-rss:26016kB, shmem-rss:0kB, UID:0 pgtables:676kB oom_score_adj:0
oom_reaper: reaped process 8248 (docker), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
oom_reaper: reaped process 2054 (node_exporter), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
oom_reaper: reaped process 1452 (systemd-journal), now anon-rss:0kB, file-rss:8564kB, shmem-rss:4kB
oom_reaper: reaped process 2146 (munin-node), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
oom_reaper: reaped process 8291 (runc:[2:INIT]), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
The reason is that in this case, the target cpuset nodes only have
movable zone, while the creation of an OS in docker sometimes needs to
allocate memory in non-movable zones (dma/dma32/normal) like
GFP_HIGHUSER, and the cpuset limit forbids the allocation, then
out-of-memory killing is involved even when normal nodes and movable
nodes both have many free memory.
The OOM killer cannot help to resolve the situation as there is no
usable memory for the request in the cpuset scope. The only reasonable
measure to take is to fail the allocation right away and have the caller
to deal with it.
So add a check for cases like this in the slowpath of allocation, and
bail out early returning NULL for the allocation.
As page allocation is one of the hottest path in kernel, this check will
hurt all users with sane cpuset configuration, add a static branch check
and detect the abnormal config in cpuset memory binding setup so that
the extra check cost in page allocation is not paid by everyone.
[thanks to Micho Hocko and David Rientjes for suggesting not handling
it inside OOM code, adding cpuset check, refining comments]
Link: https://lkml.kernel.org/r/1632481657-68112-1-git-send-email-feng.tang@intel.com
Signed-off-by: Feng Tang <feng.tang@intel.com>
Suggested-by: Michal Hocko <mhocko@suse.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Zefan Li <lizefan.x@bytedance.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-11-05 13:40:34 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-07-29 14:47:17 +05:30
|
|
|
|
mm/page_alloc: detect allocation forbidden by cpuset and bail out early
There was a report that starting an Ubuntu in docker while using cpuset
to bind it to movable nodes (a node only has movable zone, like a node
for hotplug or a Persistent Memory node in normal usage) will fail due
to memory allocation failure, and then OOM is involved and many other
innocent processes got killed.
It can be reproduced with command:
$ docker run -it --rm --cpuset-mems 4 ubuntu:latest bash -c "grep Mems_allowed /proc/self/status"
(where node 4 is a movable node)
runc:[2:INIT] invoked oom-killer: gfp_mask=0x500cc2(GFP_HIGHUSER|__GFP_ACCOUNT), order=0, oom_score_adj=0
CPU: 8 PID: 8291 Comm: runc:[2:INIT] Tainted: G W I E 5.8.2-0.g71b519a-default #1 openSUSE Tumbleweed (unreleased)
Hardware name: Dell Inc. PowerEdge R640/0PHYDR, BIOS 2.6.4 04/09/2020
Call Trace:
dump_stack+0x6b/0x88
dump_header+0x4a/0x1e2
oom_kill_process.cold+0xb/0x10
out_of_memory.part.0+0xaf/0x230
out_of_memory+0x3d/0x80
__alloc_pages_slowpath.constprop.0+0x954/0xa20
__alloc_pages_nodemask+0x2d3/0x300
pipe_write+0x322/0x590
new_sync_write+0x196/0x1b0
vfs_write+0x1c3/0x1f0
ksys_write+0xa7/0xe0
do_syscall_64+0x52/0xd0
entry_SYSCALL_64_after_hwframe+0x44/0xa9
Mem-Info:
active_anon:392832 inactive_anon:182 isolated_anon:0
active_file:68130 inactive_file:151527 isolated_file:0
unevictable:2701 dirty:0 writeback:7
slab_reclaimable:51418 slab_unreclaimable:116300
mapped:45825 shmem:735 pagetables:2540 bounce:0
free:159849484 free_pcp:73 free_cma:0
Node 4 active_anon:1448kB inactive_anon:0kB active_file:0kB inactive_file:0kB unevictable:0kB isolated(anon):0kB isolated(file):0kB mapped:0kB dirty:0kB writeback:0kB shmem:0kB shmem_thp: 0kB shmem_pmdmapped: 0kB anon_thp: 0kB writeback_tmp:0kB all_unreclaimable? no
Node 4 Movable free:130021408kB min:9140kB low:139160kB high:269180kB reserved_highatomic:0KB active_anon:1448kB inactive_anon:0kB active_file:0kB inactive_file:0kB unevictable:0kB writepending:0kB present:130023424kB managed:130023424kB mlocked:0kB kernel_stack:0kB pagetables:0kB bounce:0kB free_pcp:292kB local_pcp:84kB free_cma:0kB
lowmem_reserve[]: 0 0 0 0 0
Node 4 Movable: 1*4kB (M) 0*8kB 0*16kB 1*32kB (M) 0*64kB 0*128kB 1*256kB (M) 1*512kB (M) 1*1024kB (M) 0*2048kB 31743*4096kB (M) = 130021156kB
oom-kill:constraint=CONSTRAINT_CPUSET,nodemask=(null),cpuset=docker-9976a269caec812c134fa317f27487ee36e1129beba7278a463dd53e5fb9997b.scope,mems_allowed=4,global_oom,task_memcg=/system.slice/containerd.service,task=containerd,pid=4100,uid=0
Out of memory: Killed process 4100 (containerd) total-vm:4077036kB, anon-rss:51184kB, file-rss:26016kB, shmem-rss:0kB, UID:0 pgtables:676kB oom_score_adj:0
oom_reaper: reaped process 8248 (docker), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
oom_reaper: reaped process 2054 (node_exporter), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
oom_reaper: reaped process 1452 (systemd-journal), now anon-rss:0kB, file-rss:8564kB, shmem-rss:4kB
oom_reaper: reaped process 2146 (munin-node), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
oom_reaper: reaped process 8291 (runc:[2:INIT]), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
The reason is that in this case, the target cpuset nodes only have
movable zone, while the creation of an OS in docker sometimes needs to
allocate memory in non-movable zones (dma/dma32/normal) like
GFP_HIGHUSER, and the cpuset limit forbids the allocation, then
out-of-memory killing is involved even when normal nodes and movable
nodes both have many free memory.
The OOM killer cannot help to resolve the situation as there is no
usable memory for the request in the cpuset scope. The only reasonable
measure to take is to fail the allocation right away and have the caller
to deal with it.
So add a check for cases like this in the slowpath of allocation, and
bail out early returning NULL for the allocation.
As page allocation is one of the hottest path in kernel, this check will
hurt all users with sane cpuset configuration, add a static branch check
and detect the abnormal config in cpuset memory binding setup so that
the extra check cost in page allocation is not paid by everyone.
[thanks to Micho Hocko and David Rientjes for suggesting not handling
it inside OOM code, adding cpuset check, refining comments]
Link: https://lkml.kernel.org/r/1632481657-68112-1-git-send-email-feng.tang@intel.com
Signed-off-by: Feng Tang <feng.tang@intel.com>
Suggested-by: Michal Hocko <mhocko@suse.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Zefan Li <lizefan.x@bytedance.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-11-05 13:40:34 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-12-28 00:35:48 -08:00
|
|
|
|
2018-04-05 16:25:16 -07:00
|
|
|
|
2005-11-13 16:06:43 -08:00
|
|
|
|
2016-07-28 15:49:16 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, page_alloc: restructure direct compaction handling in slowpath
The retry loop in __alloc_pages_slowpath is supposed to keep trying
reclaim and compaction (and OOM), until either the allocation succeeds,
or returns with failure. Success here is more probable when reclaim
precedes compaction, as certain watermarks have to be met for compaction
to even try, and more free pages increase the probability of compaction
success. On the other hand, starting with light async compaction (if
the watermarks allow it), can be more efficient, especially for smaller
orders, if there's enough free memory which is just fragmented.
Thus, the current code starts with compaction before reclaim, and to
make sure that the last reclaim is always followed by a final
compaction, there's another direct compaction call at the end of the
loop. This makes the code hard to follow and adds some duplicated
handling of migration_mode decisions. It's also somewhat inefficient
that even if reclaim or compaction decides not to retry, the final
compaction is still attempted. Some gfp flags combination also shortcut
these retry decisions by "goto noretry;", making it even harder to
follow.
This patch attempts to restructure the code with only minimal functional
changes. The call to the first compaction and THP-specific checks are
now placed above the retry loop, and the "noretry" direct compaction is
removed.
The initial compaction is additionally restricted only to costly orders,
as we can expect smaller orders to be held back by watermarks, and only
larger orders to suffer primarily from fragmentation. This better
matches the checks in reclaim's shrink_zones().
There are two other smaller functional changes. One is that the upgrade
from async migration to light sync migration will always occur after the
initial compaction. This is how it has been until recent patch "mm,
oom: protect !costly allocations some more", which introduced upgrading
the mode based on COMPACT_COMPLETE result, but kept the final compaction
always upgraded, which made it even more special. It's better to return
to the simpler handling for now, as migration modes will be further
modified later in the series.
The second change is that once both reclaim and compaction declare it's
not worth to retry the reclaim/compact loop, there is no final
compaction attempt. As argued above, this is intentional. If that
final compaction were to succeed, it would be due to a wrong retry
decision, or simply a race with somebody else freeing memory for us.
The main outcome of this patch should be simpler code. Logically, the
initial compaction without reclaim is the exceptional case to the
reclaim/compaction scheme, but prior to the patch, it was the last loop
iteration that was exceptional. Now the code matches the logic better.
The change also enable the following patches.
Link: http://lkml.kernel.org/r/20160721073614.24395-5-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-07-28 15:49:19 -07:00
|
|
|
|
|
|
|
|
|
mm, compaction: restrict async compaction to pageblocks of same migratetype
The migrate scanner in async compaction is currently limited to
MIGRATE_MOVABLE pageblocks. This is a heuristic intended to reduce
latency, based on the assumption that non-MOVABLE pageblocks are
unlikely to contain movable pages.
However, with the exception of THP's, most high-order allocations are
not movable. Should the async compaction succeed, this increases the
chance that the non-MOVABLE allocations will fallback to a MOVABLE
pageblock, making the long-term fragmentation worse.
This patch attempts to help the situation by changing async direct
compaction so that the migrate scanner only scans the pageblocks of the
requested migratetype. If it's a non-MOVABLE type and there are such
pageblocks that do contain movable pages, chances are that the
allocation can succeed within one of such pageblocks, removing the need
for a fallback. If that fails, the subsequent sync attempt will ignore
this restriction.
In testing based on 4.9 kernel with stress-highalloc from mmtests
configured for order-4 GFP_KERNEL allocations, this patch has reduced
the number of unmovable allocations falling back to movable pageblocks
by 30%. The number of movable allocations falling back is reduced by
12%.
Link: http://lkml.kernel.org/r/20170307131545.28577-8-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-08 15:54:49 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, page_alloc: restructure direct compaction handling in slowpath
The retry loop in __alloc_pages_slowpath is supposed to keep trying
reclaim and compaction (and OOM), until either the allocation succeeds,
or returns with failure. Success here is more probable when reclaim
precedes compaction, as certain watermarks have to be met for compaction
to even try, and more free pages increase the probability of compaction
success. On the other hand, starting with light async compaction (if
the watermarks allow it), can be more efficient, especially for smaller
orders, if there's enough free memory which is just fragmented.
Thus, the current code starts with compaction before reclaim, and to
make sure that the last reclaim is always followed by a final
compaction, there's another direct compaction call at the end of the
loop. This makes the code hard to follow and adds some duplicated
handling of migration_mode decisions. It's also somewhat inefficient
that even if reclaim or compaction decides not to retry, the final
compaction is still attempted. Some gfp flags combination also shortcut
these retry decisions by "goto noretry;", making it even harder to
follow.
This patch attempts to restructure the code with only minimal functional
changes. The call to the first compaction and THP-specific checks are
now placed above the retry loop, and the "noretry" direct compaction is
removed.
The initial compaction is additionally restricted only to costly orders,
as we can expect smaller orders to be held back by watermarks, and only
larger orders to suffer primarily from fragmentation. This better
matches the checks in reclaim's shrink_zones().
There are two other smaller functional changes. One is that the upgrade
from async migration to light sync migration will always occur after the
initial compaction. This is how it has been until recent patch "mm,
oom: protect !costly allocations some more", which introduced upgrading
the mode based on COMPACT_COMPLETE result, but kept the final compaction
always upgraded, which made it even more special. It's better to return
to the simpler handling for now, as migration modes will be further
modified later in the series.
The second change is that once both reclaim and compaction declare it's
not worth to retry the reclaim/compact loop, there is no final
compaction attempt. As argued above, this is intentional. If that
final compaction were to succeed, it would be due to a wrong retry
decision, or simply a race with somebody else freeing memory for us.
The main outcome of this patch should be simpler code. Logically, the
initial compaction without reclaim is the exceptional case to the
reclaim/compaction scheme, but prior to the patch, it was the last loop
iteration that was exceptional. Now the code matches the logic better.
The change also enable the following patches.
Link: http://lkml.kernel.org/r/20160721073614.24395-5-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-07-28 15:49:19 -07:00
|
|
|
|
mm, vmscan: prevent infinite loop for costly GFP_NOIO | __GFP_RETRY_MAYFAIL allocations
Sven reports an infinite loop in __alloc_pages_slowpath() for costly order
__GFP_RETRY_MAYFAIL allocations that are also GFP_NOIO. Such combination
can happen in a suspend/resume context where a GFP_KERNEL allocation can
have __GFP_IO masked out via gfp_allowed_mask.
Quoting Sven:
1. try to do a "costly" allocation (order > PAGE_ALLOC_COSTLY_ORDER)
with __GFP_RETRY_MAYFAIL set.
2. page alloc's __alloc_pages_slowpath tries to get a page from the
freelist. This fails because there is nothing free of that costly
order.
3. page alloc tries to reclaim by calling __alloc_pages_direct_reclaim,
which bails out because a zone is ready to be compacted; it pretends
to have made a single page of progress.
4. page alloc tries to compact, but this always bails out early because
__GFP_IO is not set (it's not passed by the snd allocator, and even
if it were, we are suspending so the __GFP_IO flag would be cleared
anyway).
5. page alloc believes reclaim progress was made (because of the
pretense in item 3) and so it checks whether it should retry
compaction. The compaction retry logic thinks it should try again,
because:
a) reclaim is needed because of the early bail-out in item 4
b) a zonelist is suitable for compaction
6. goto 2. indefinite stall.
(end quote)
The immediate root cause is confusing the COMPACT_SKIPPED returned from
__alloc_pages_direct_compact() (step 4) due to lack of __GFP_IO to be
indicating a lack of order-0 pages, and in step 5 evaluating that in
should_compact_retry() as a reason to retry, before incrementing and
limiting the number of retries. There are however other places that
wrongly assume that compaction can happen while we lack __GFP_IO.
To fix this, introduce gfp_compaction_allowed() to abstract the __GFP_IO
evaluation and switch the open-coded test in try_to_compact_pages() to use
it.
Also use the new helper in:
- compaction_ready(), which will make reclaim not bail out in step 3, so
there's at least one attempt to actually reclaim, even if chances are
small for a costly order
- in_reclaim_compaction() which will make should_continue_reclaim()
return false and we don't over-reclaim unnecessarily
- in __alloc_pages_slowpath() to set a local variable can_compact,
which is then used to avoid retrying reclaim/compaction for costly
allocations (step 5) if we can't compact and also to skip the early
compaction attempt that we do in some cases
Link: https://lkml.kernel.org/r/20240221114357.13655-2-vbabka@suse.cz
Fixes: 3250845d0526 ("Revert "mm, oom: prevent premature OOM killer invocation for high order request"")
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Reported-by: Sven van Ashbrook <svenva@chromium.org>
Closes: https://lore.kernel.org/all/CAG-rBihs_xMKb3wrMO1%2B-%2Bp4fowP9oy1pa_OTkfxBzPUVOZF%2Bg@mail.gmail.com/
Tested-by: Karthikeyan Ramasubramanian <kramasub@chromium.org>
Cc: Brian Geffon <bgeffon@google.com>
Cc: Curtis Malainey <cujomalainey@chromium.org>
Cc: Jaroslav Kysela <perex@perex.cz>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Takashi Iwai <tiwai@suse.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-21 12:43:58 +01:00
|
|
|
|
mm, compaction: restrict async compaction to pageblocks of same migratetype
The migrate scanner in async compaction is currently limited to
MIGRATE_MOVABLE pageblocks. This is a heuristic intended to reduce
latency, based on the assumption that non-MOVABLE pageblocks are
unlikely to contain movable pages.
However, with the exception of THP's, most high-order allocations are
not movable. Should the async compaction succeed, this increases the
chance that the non-MOVABLE allocations will fallback to a MOVABLE
pageblock, making the long-term fragmentation worse.
This patch attempts to help the situation by changing async direct
compaction so that the migrate scanner only scans the pageblocks of the
requested migratetype. If it's a non-MOVABLE type and there are such
pageblocks that do contain movable pages, chances are that the
allocation can succeed within one of such pageblocks, removing the need
for a fallback. If that fails, the subsequent sync attempt will ignore
this restriction.
In testing based on 4.9 kernel with stress-highalloc from mmtests
configured for order-4 GFP_KERNEL allocations, this patch has reduced
the number of unmovable allocations falling back to movable pageblocks
by 30%. The number of movable allocations falling back is reduced by
12%.
Link: http://lkml.kernel.org/r/20170307131545.28577-8-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-08 15:54:49 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, page_alloc: restructure direct compaction handling in slowpath
The retry loop in __alloc_pages_slowpath is supposed to keep trying
reclaim and compaction (and OOM), until either the allocation succeeds,
or returns with failure. Success here is more probable when reclaim
precedes compaction, as certain watermarks have to be met for compaction
to even try, and more free pages increase the probability of compaction
success. On the other hand, starting with light async compaction (if
the watermarks allow it), can be more efficient, especially for smaller
orders, if there's enough free memory which is just fragmented.
Thus, the current code starts with compaction before reclaim, and to
make sure that the last reclaim is always followed by a final
compaction, there's another direct compaction call at the end of the
loop. This makes the code hard to follow and adds some duplicated
handling of migration_mode decisions. It's also somewhat inefficient
that even if reclaim or compaction decides not to retry, the final
compaction is still attempted. Some gfp flags combination also shortcut
these retry decisions by "goto noretry;", making it even harder to
follow.
This patch attempts to restructure the code with only minimal functional
changes. The call to the first compaction and THP-specific checks are
now placed above the retry loop, and the "noretry" direct compaction is
removed.
The initial compaction is additionally restricted only to costly orders,
as we can expect smaller orders to be held back by watermarks, and only
larger orders to suffer primarily from fragmentation. This better
matches the checks in reclaim's shrink_zones().
There are two other smaller functional changes. One is that the upgrade
from async migration to light sync migration will always occur after the
initial compaction. This is how it has been until recent patch "mm,
oom: protect !costly allocations some more", which introduced upgrading
the mode based on COMPACT_COMPLETE result, but kept the final compaction
always upgraded, which made it even more special. It's better to return
to the simpler handling for now, as migration modes will be further
modified later in the series.
The second change is that once both reclaim and compaction declare it's
not worth to retry the reclaim/compact loop, there is no final
compaction attempt. As argued above, this is intentional. If that
final compaction were to succeed, it would be due to a wrong retry
decision, or simply a race with somebody else freeing memory for us.
The main outcome of this patch should be simpler code. Logically, the
initial compaction without reclaim is the exceptional case to the
reclaim/compaction scheme, but prior to the patch, it was the last loop
iteration that was exceptional. Now the code matches the logic better.
The change also enable the following patches.
Link: http://lkml.kernel.org/r/20160721073614.24395-5-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-07-28 15:49:19 -07:00
|
|
|
|
|
|
|
|
|
2016-07-28 15:49:28 -07:00
|
|
|
|
mm, page_alloc: restructure direct compaction handling in slowpath
The retry loop in __alloc_pages_slowpath is supposed to keep trying
reclaim and compaction (and OOM), until either the allocation succeeds,
or returns with failure. Success here is more probable when reclaim
precedes compaction, as certain watermarks have to be met for compaction
to even try, and more free pages increase the probability of compaction
success. On the other hand, starting with light async compaction (if
the watermarks allow it), can be more efficient, especially for smaller
orders, if there's enough free memory which is just fragmented.
Thus, the current code starts with compaction before reclaim, and to
make sure that the last reclaim is always followed by a final
compaction, there's another direct compaction call at the end of the
loop. This makes the code hard to follow and adds some duplicated
handling of migration_mode decisions. It's also somewhat inefficient
that even if reclaim or compaction decides not to retry, the final
compaction is still attempted. Some gfp flags combination also shortcut
these retry decisions by "goto noretry;", making it even harder to
follow.
This patch attempts to restructure the code with only minimal functional
changes. The call to the first compaction and THP-specific checks are
now placed above the retry loop, and the "noretry" direct compaction is
removed.
The initial compaction is additionally restricted only to costly orders,
as we can expect smaller orders to be held back by watermarks, and only
larger orders to suffer primarily from fragmentation. This better
matches the checks in reclaim's shrink_zones().
There are two other smaller functional changes. One is that the upgrade
from async migration to light sync migration will always occur after the
initial compaction. This is how it has been until recent patch "mm,
oom: protect !costly allocations some more", which introduced upgrading
the mode based on COMPACT_COMPLETE result, but kept the final compaction
always upgraded, which made it even more special. It's better to return
to the simpler handling for now, as migration modes will be further
modified later in the series.
The second change is that once both reclaim and compaction declare it's
not worth to retry the reclaim/compact loop, there is no final
compaction attempt. As argued above, this is intentional. If that
final compaction were to succeed, it would be due to a wrong retry
decision, or simply a race with somebody else freeing memory for us.
The main outcome of this patch should be simpler code. Logically, the
initial compaction without reclaim is the exceptional case to the
reclaim/compaction scheme, but prior to the patch, it was the last loop
iteration that was exceptional. Now the code matches the logic better.
The change also enable the following patches.
Link: http://lkml.kernel.org/r/20160721073614.24395-5-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-07-28 15:49:19 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, thp: tweak reclaim/compaction effort of local-only and all-node allocations
THP page faults now attempt a __GFP_THISNODE allocation first, which
should only compact existing free memory, followed by another attempt
that can allocate from any node using reclaim/compaction effort
specified by global defrag setting and madvise.
This patch makes the following changes to the scheme:
- Before the patch, the first allocation relies on a check for
pageblock order and __GFP_IO to prevent excessive reclaim. This
however affects also the second attempt, which is not limited to
single node.
Instead of that, reuse the existing check for costly order
__GFP_NORETRY allocations, and make sure the first THP attempt uses
__GFP_NORETRY. As a side-effect, all costly order __GFP_NORETRY
allocations will bail out if compaction needs reclaim, while
previously they only bailed out when compaction was deferred due to
previous failures.
This should be still acceptable within the __GFP_NORETRY semantics.
- Before the patch, the second allocation attempt (on all nodes) was
passing __GFP_NORETRY. This is redundant as the check for pageblock
order (discussed above) was stronger. It's also contrary to
madvise(MADV_HUGEPAGE) which means some effort to allocate THP is
requested.
After this patch, the second attempt doesn't pass __GFP_THISNODE nor
__GFP_NORETRY.
To sum up, THP page faults now try the following attempts:
1. local node only THP allocation with no reclaim, just compaction.
2. for madvised VMA's or when synchronous compaction is enabled always - THP
allocation from any node with effort determined by global defrag setting
and VMA madvise
3. fallback to base pages on any node
Link: http://lkml.kernel.org/r/08a3f4dd-c3ce-0009-86c5-9ee51aba8557@suse.cz
Fixes: b39d0ee2632d ("mm, page_alloc: avoid expensive reclaim when compaction may not succeed")
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-01-13 16:29:04 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-09-04 12:54:22 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, hugetlb: allow hugepage allocations to reclaim as needed
Commit b39d0ee2632d ("mm, page_alloc: avoid expensive reclaim when
compaction may not succeed") has chnaged the allocator to bail out from
the allocator early to prevent from a potentially excessive memory
reclaim. __GFP_RETRY_MAYFAIL is designed to retry the allocation,
reclaim and compaction loop as long as there is a reasonable chance to
make forward progress. Neither COMPACT_SKIPPED nor COMPACT_DEFERRED at
the INIT_COMPACT_PRIORITY compaction attempt gives this feedback.
The most obvious affected subsystem is hugetlbfs which allocates huge
pages based on an admin request (or via admin configured overcommit). I
have done a simple test which tries to allocate half of the memory for
hugetlb pages while the memory is full of a clean page cache. This is
not an unusual situation because we try to cache as much of the memory
as possible and sysctl/sysfs interface to allocate huge pages is there
for flexibility to allocate hugetlb pages at any time.
System has 1GB of RAM and we are requesting 515MB worth of hugetlb pages
after the memory is prefilled by a clean page cache:
root@test1:~# cat hugetlb_test.sh
set -x
echo 0 > /proc/sys/vm/nr_hugepages
echo 3 > /proc/sys/vm/drop_caches
echo 1 > /proc/sys/vm/compact_memory
dd if=/mnt/data/file-1G of=/dev/null bs=$((4<<10))
TS=$(date +%s)
echo 256 > /proc/sys/vm/nr_hugepages
cat /proc/sys/vm/nr_hugepages
The results for 2 consecutive runs on clean 5.3
root@test1:~# sh hugetlb_test.sh
+ echo 0
+ echo 3
+ echo 1
+ dd if=/mnt/data/file-1G of=/dev/null bs=4096
262144+0 records in
262144+0 records out
1073741824 bytes (1.1 GB) copied, 21.0694 s, 51.0 MB/s
+ date +%s
+ TS=1569905284
+ echo 256
+ cat /proc/sys/vm/nr_hugepages
256
root@test1:~# sh hugetlb_test.sh
+ echo 0
+ echo 3
+ echo 1
+ dd if=/mnt/data/file-1G of=/dev/null bs=4096
262144+0 records in
262144+0 records out
1073741824 bytes (1.1 GB) copied, 21.7548 s, 49.4 MB/s
+ date +%s
+ TS=1569905311
+ echo 256
+ cat /proc/sys/vm/nr_hugepages
256
Now with b39d0ee2632d applied
root@test1:~# sh hugetlb_test.sh
+ echo 0
+ echo 3
+ echo 1
+ dd if=/mnt/data/file-1G of=/dev/null bs=4096
262144+0 records in
262144+0 records out
1073741824 bytes (1.1 GB) copied, 20.1815 s, 53.2 MB/s
+ date +%s
+ TS=1569905516
+ echo 256
+ cat /proc/sys/vm/nr_hugepages
11
root@test1:~# sh hugetlb_test.sh
+ echo 0
+ echo 3
+ echo 1
+ dd if=/mnt/data/file-1G of=/dev/null bs=4096
262144+0 records in
262144+0 records out
1073741824 bytes (1.1 GB) copied, 21.9485 s, 48.9 MB/s
+ date +%s
+ TS=1569905541
+ echo 256
+ cat /proc/sys/vm/nr_hugepages
12
The success rate went down by factor of 20!
Although hugetlb allocation requests might fail and it is reasonable to
expect them to under extremely fragmented memory or when the memory is
under a heavy pressure but the above situation is not that case.
Fix the regression by reverting back to the previous behavior for
__GFP_RETRY_MAYFAIL requests and disable the beail out heuristic for
those requests.
Mike said:
: hugetlbfs allocations are commonly done via sysctl/sysfs shortly after
: boot where this may not be as much of an issue. However, I am aware of at
: least three use cases where allocations are made after the system has been
: up and running for quite some time:
:
: - DB reconfiguration. If sysctl/sysfs fails to get required number of
: huge pages, system is rebooted to perform allocation after boot.
:
: - VM provisioning. If unable get required number of huge pages, fall
: back to base pages.
:
: - An application that does not preallocate pool, but rather allocates
: pages at fault time for optimal NUMA locality.
:
: In all cases, I would expect b39d0ee2632d to cause regressions and
: noticable behavior changes.
:
: My quick/limited testing in
: https://lkml.kernel.org/r/3468b605-a3a9-6978-9699-57c52a90bd7e@oracle.com
: was insufficient. It was also mentioned that if something like
: b39d0ee2632d went forward, I would like exemptions for __GFP_RETRY_MAYFAIL
: requests as in this patch.
[mhocko@suse.com: reworded changelog]
Link: http://lkml.kernel.org/r/20191007075548.12456-1-mhocko@kernel.org
Fixes: b39d0ee2632d ("mm, page_alloc: avoid expensive reclaim when compaction may not succeed")
Signed-off-by: David Rientjes <rientjes@google.com>
Signed-off-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Mel Gorman <mgorman@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-10-14 14:12:04 -07:00
|
|
|
|
|
|
|
|
|
2019-09-04 12:54:22 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, page_alloc: restructure direct compaction handling in slowpath
The retry loop in __alloc_pages_slowpath is supposed to keep trying
reclaim and compaction (and OOM), until either the allocation succeeds,
or returns with failure. Success here is more probable when reclaim
precedes compaction, as certain watermarks have to be met for compaction
to even try, and more free pages increase the probability of compaction
success. On the other hand, starting with light async compaction (if
the watermarks allow it), can be more efficient, especially for smaller
orders, if there's enough free memory which is just fragmented.
Thus, the current code starts with compaction before reclaim, and to
make sure that the last reclaim is always followed by a final
compaction, there's another direct compaction call at the end of the
loop. This makes the code hard to follow and adds some duplicated
handling of migration_mode decisions. It's also somewhat inefficient
that even if reclaim or compaction decides not to retry, the final
compaction is still attempted. Some gfp flags combination also shortcut
these retry decisions by "goto noretry;", making it even harder to
follow.
This patch attempts to restructure the code with only minimal functional
changes. The call to the first compaction and THP-specific checks are
now placed above the retry loop, and the "noretry" direct compaction is
removed.
The initial compaction is additionally restricted only to costly orders,
as we can expect smaller orders to be held back by watermarks, and only
larger orders to suffer primarily from fragmentation. This better
matches the checks in reclaim's shrink_zones().
There are two other smaller functional changes. One is that the upgrade
from async migration to light sync migration will always occur after the
initial compaction. This is how it has been until recent patch "mm,
oom: protect !costly allocations some more", which introduced upgrading
the mode based on COMPACT_COMPLETE result, but kept the final compaction
always upgraded, which made it even more special. It's better to return
to the simpler handling for now, as migration modes will be further
modified later in the series.
The second change is that once both reclaim and compaction declare it's
not worth to retry the reclaim/compact loop, there is no final
compaction attempt. As argued above, this is intentional. If that
final compaction were to succeed, it would be due to a wrong retry
decision, or simply a race with somebody else freeing memory for us.
The main outcome of this patch should be simpler code. Logically, the
initial compaction without reclaim is the exceptional case to the
reclaim/compaction scheme, but prior to the patch, it was the last loop
iteration that was exceptional. Now the code matches the logic better.
The change also enable the following patches.
Link: http://lkml.kernel.org/r/20160721073614.24395-5-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-07-28 15:49:19 -07:00
|
|
|
|
|
|
|
|
|
mm, page_alloc: make THP-specific decisions more generic
Since THP allocations during page faults can be costly, extra decisions
are employed for them to avoid excessive reclaim and compaction, if the
initial compaction doesn't look promising. The detection has never been
perfect as there is no gfp flag specific to THP allocations. At this
moment it checks the whole combination of flags that makes up
GFP_TRANSHUGE, and hopes that no other users of such combination exist,
or would mind being treated the same way. Extra care is also taken to
separate allocations from khugepaged, where latency doesn't matter that
much.
It is however possible to distinguish these allocations in a simpler and
more reliable way. The key observation is that after the initial
compaction followed by the first iteration of "standard"
reclaim/compaction, both __GFP_NORETRY allocations and costly
allocations without __GFP_REPEAT are declared as failures:
/* Do not loop if specifically requested */
if (gfp_mask & __GFP_NORETRY)
goto nopage;
/*
* Do not retry costly high order allocations unless they are
* __GFP_REPEAT
*/
if (order > PAGE_ALLOC_COSTLY_ORDER && !(gfp_mask & __GFP_REPEAT))
goto nopage;
This means we can further distinguish allocations that are costly order
*and* additionally include the __GFP_NORETRY flag. As it happens,
GFP_TRANSHUGE allocations do already fall into this category. This will
also allow other costly allocations with similar high-order benefit vs
latency considerations to use this semantic. Furthermore, we can
distinguish THP allocations that should try a bit harder (such as from
khugepageed) by removing __GFP_NORETRY, as will be done in the next
patch.
Link: http://lkml.kernel.org/r/20160721073614.24395-6-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-07-28 15:49:22 -07:00
|
|
|
|
|
|
|
|
|
mm, thp: remove __GFP_NORETRY from khugepaged and madvised allocations
After the previous patch, we can distinguish costly allocations that
should be really lightweight, such as THP page faults, with
__GFP_NORETRY. This means we don't need to recognize khugepaged
allocations via PF_KTHREAD anymore. We can also change THP page faults
in areas where madvise(MADV_HUGEPAGE) was used to try as hard as
khugepaged, as the process has indicated that it benefits from THP's and
is willing to pay some initial latency costs.
We can also make the flags handling less cryptic by distinguishing
GFP_TRANSHUGE_LIGHT (no reclaim at all, default mode in page fault) from
GFP_TRANSHUGE (only direct reclaim, khugepaged default). Adding
__GFP_NORETRY or __GFP_KSWAPD_RECLAIM is done where needed.
The patch effectively changes the current GFP_TRANSHUGE users as
follows:
* get_huge_zero_page() - the zero page lifetime should be relatively
long and it's shared by multiple users, so it's worth spending some
effort on it. We use GFP_TRANSHUGE, and __GFP_NORETRY is not added.
This also restores direct reclaim to this allocation, which was
unintentionally removed by commit e4a49efe4e7e ("mm: thp: set THP defrag
by default to madvise and add a stall-free defrag option")
* alloc_hugepage_khugepaged_gfpmask() - this is khugepaged, so latency
is not an issue. So if khugepaged "defrag" is enabled (the default), do
reclaim via GFP_TRANSHUGE without __GFP_NORETRY. We can remove the
PF_KTHREAD check from page alloc.
As a side-effect, khugepaged will now no longer check if the initial
compaction was deferred or contended. This is OK, as khugepaged sleep
times between collapsion attempts are long enough to prevent noticeable
disruption, so we should allow it to spend some effort.
* migrate_misplaced_transhuge_page() - already was masking out
__GFP_RECLAIM, so just convert to GFP_TRANSHUGE_LIGHT which is
equivalent.
* alloc_hugepage_direct_gfpmask() - vma's with VM_HUGEPAGE (via madvise)
are now allocating without __GFP_NORETRY. Other vma's keep using
__GFP_NORETRY if direct reclaim/compaction is at all allowed (by default
it's allowed only for madvised vma's). The rest is conversion to
GFP_TRANSHUGE(_LIGHT).
[mhocko@suse.com: suggested GFP_TRANSHUGE_LIGHT]
Link: http://lkml.kernel.org/r/20160721073614.24395-7-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-07-28 15:49:25 -07:00
|
|
|
|
mm, page_alloc: restructure direct compaction handling in slowpath
The retry loop in __alloc_pages_slowpath is supposed to keep trying
reclaim and compaction (and OOM), until either the allocation succeeds,
or returns with failure. Success here is more probable when reclaim
precedes compaction, as certain watermarks have to be met for compaction
to even try, and more free pages increase the probability of compaction
success. On the other hand, starting with light async compaction (if
the watermarks allow it), can be more efficient, especially for smaller
orders, if there's enough free memory which is just fragmented.
Thus, the current code starts with compaction before reclaim, and to
make sure that the last reclaim is always followed by a final
compaction, there's another direct compaction call at the end of the
loop. This makes the code hard to follow and adds some duplicated
handling of migration_mode decisions. It's also somewhat inefficient
that even if reclaim or compaction decides not to retry, the final
compaction is still attempted. Some gfp flags combination also shortcut
these retry decisions by "goto noretry;", making it even harder to
follow.
This patch attempts to restructure the code with only minimal functional
changes. The call to the first compaction and THP-specific checks are
now placed above the retry loop, and the "noretry" direct compaction is
removed.
The initial compaction is additionally restricted only to costly orders,
as we can expect smaller orders to be held back by watermarks, and only
larger orders to suffer primarily from fragmentation. This better
matches the checks in reclaim's shrink_zones().
There are two other smaller functional changes. One is that the upgrade
from async migration to light sync migration will always occur after the
initial compaction. This is how it has been until recent patch "mm,
oom: protect !costly allocations some more", which introduced upgrading
the mode based on COMPACT_COMPLETE result, but kept the final compaction
always upgraded, which made it even more special. It's better to return
to the simpler handling for now, as migration modes will be further
modified later in the series.
The second change is that once both reclaim and compaction declare it's
not worth to retry the reclaim/compact loop, there is no final
compaction attempt. As argued above, this is intentional. If that
final compaction were to succeed, it would be due to a wrong retry
decision, or simply a race with somebody else freeing memory for us.
The main outcome of this patch should be simpler code. Logically, the
initial compaction without reclaim is the exceptional case to the
reclaim/compaction scheme, but prior to the patch, it was the last loop
iteration that was exceptional. Now the code matches the logic better.
The change also enable the following patches.
Link: http://lkml.kernel.org/r/20160721073614.24395-5-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-07-28 15:49:19 -07:00
|
|
|
|
2016-07-28 15:49:28 -07:00
|
|
|
|
mm, page_alloc: restructure direct compaction handling in slowpath
The retry loop in __alloc_pages_slowpath is supposed to keep trying
reclaim and compaction (and OOM), until either the allocation succeeds,
or returns with failure. Success here is more probable when reclaim
precedes compaction, as certain watermarks have to be met for compaction
to even try, and more free pages increase the probability of compaction
success. On the other hand, starting with light async compaction (if
the watermarks allow it), can be more efficient, especially for smaller
orders, if there's enough free memory which is just fragmented.
Thus, the current code starts with compaction before reclaim, and to
make sure that the last reclaim is always followed by a final
compaction, there's another direct compaction call at the end of the
loop. This makes the code hard to follow and adds some duplicated
handling of migration_mode decisions. It's also somewhat inefficient
that even if reclaim or compaction decides not to retry, the final
compaction is still attempted. Some gfp flags combination also shortcut
these retry decisions by "goto noretry;", making it even harder to
follow.
This patch attempts to restructure the code with only minimal functional
changes. The call to the first compaction and THP-specific checks are
now placed above the retry loop, and the "noretry" direct compaction is
removed.
The initial compaction is additionally restricted only to costly orders,
as we can expect smaller orders to be held back by watermarks, and only
larger orders to suffer primarily from fragmentation. This better
matches the checks in reclaim's shrink_zones().
There are two other smaller functional changes. One is that the upgrade
from async migration to light sync migration will always occur after the
initial compaction. This is how it has been until recent patch "mm,
oom: protect !costly allocations some more", which introduced upgrading
the mode based on COMPACT_COMPLETE result, but kept the final compaction
always upgraded, which made it even more special. It's better to return
to the simpler handling for now, as migration modes will be further
modified later in the series.
The second change is that once both reclaim and compaction declare it's
not worth to retry the reclaim/compact loop, there is no final
compaction attempt. As argued above, this is intentional. If that
final compaction were to succeed, it would be due to a wrong retry
decision, or simply a race with somebody else freeing memory for us.
The main outcome of this patch should be simpler code. Logically, the
initial compaction without reclaim is the exceptional case to the
reclaim/compaction scheme, but prior to the patch, it was the last loop
iteration that was exceptional. Now the code matches the logic better.
The change also enable the following patches.
Link: http://lkml.kernel.org/r/20160721073614.24395-5-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-07-28 15:49:19 -07:00
|
|
|
|
|
|
|
|
|
2016-07-28 15:49:16 -07:00
|
|
|
|
|
|
|
|
|
mm/page_alloc.c: avoid infinite retries caused by cpuset race
__alloc_pages_slowpath has no change detection for ac->nodemask in the
part of retry path, while cpuset can modify it in parallel. For some
processes that set mempolicy as MPOL_BIND, this results ac->nodemask
changes, and then the should_reclaim_retry will judge based on the latest
nodemask and jump to retry, while the get_page_from_freelist only
traverses the zonelist from ac->preferred_zoneref, which selected by a
expired nodemask and may cause infinite retries in some cases
cpu 64:
__alloc_pages_slowpath {
/* ..... */
retry:
/* ac->nodemask = 0x1, ac->preferred->zone->nid = 1 */
if (alloc_flags & ALLOC_KSWAPD)
wake_all_kswapds(order, gfp_mask, ac);
/* cpu 1:
cpuset_write_resmask
update_nodemask
update_nodemasks_hier
update_tasks_nodemask
mpol_rebind_task
mpol_rebind_policy
mpol_rebind_nodemask
// mempolicy->nodes has been modified,
// which ac->nodemask point to
*/
/* ac->nodemask = 0x3, ac->preferred->zone->nid = 1 */
if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
did_some_progress > 0, &no_progress_loops))
goto retry;
}
Simultaneously starting multiple cpuset01 from LTP can quickly reproduce
this issue on a multi node server when the maximum memory pressure is
reached and the swap is enabled
Link: https://lkml.kernel.org/r/20250416082405.20988-1-zhangtianyang@loongson.cn
Fixes: c33d6c06f60f ("mm, page_alloc: avoid looking up the first zone in a zonelist twice")
Signed-off-by: Tianyang Zhang <zhangtianyang@loongson.cn>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Zi Yan <ziy@nvidia.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-16 16:24:05 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-07-28 15:49:16 -07:00
|
|
|
|
2018-12-28 00:35:48 -08:00
|
|
|
|
2018-04-05 16:25:16 -07:00
|
|
|
|
2016-07-28 15:49:16 -07:00
|
|
|
|
2017-09-06 16:24:50 -07:00
|
|
|
|
|
|
|
|
|
2022-09-16 15:22:42 +08:00
|
|
|
|
|
|
|
|
|
2016-07-28 15:49:16 -07:00
|
|
|
|
2016-06-03 14:56:01 -07:00
|
|
|
|
mm, page_alloc: actually ignore mempolicies for high priority allocations
__alloc_pages_slowpath() has for a long time contained code to ignore
node restrictions from memory policies for high priority allocations.
The current code that resets the zonelist iterator however does
effectively nothing after commit 7810e6781e0f ("mm, page_alloc: do not
break __GFP_THISNODE by zonelist reset") removed a buggy zonelist reset.
Even before that commit, mempolicy restrictions were still not ignored,
as they are passed in ac->nodemask which is untouched by the code.
We can either remove the code, or make it work as intended. Since
ac->nodemask can be set from task's mempolicy via alloc_pages_current()
and thus also alloc_pages(), it may indeed affect kernel allocations,
and it makes sense to ignore it to allow progress for high priority
allocations.
Thus, this patch resets ac->nodemask to NULL in such cases. This
assumes all callers can handle it (i.e. there are no guarantees as in
the case of __GFP_THISNODE) which seems to be the case. The same
assumption is already present in check_retry_cpuset() for some time.
The expected effect is that high priority kernel allocations in the
context of userspace tasks (e.g. OOM victims) restricted by mempolicies
will have higher chance to succeed if they are restricted to nodes with
depleted memory, while there are other nodes with free memory left.
It's not a new intention, but for the first time the code will match the
intention, AFAICS. It was intended by commit 183f6371aac2 ("mm: ignore
mempolicies when using ALLOC_NO_WATERMARK") in v3.6 but I think it never
really worked, as mempolicy restriction was already encoded in nodemask,
not zonelist, at that time.
So originally that was for ALLOC_NO_WATERMARK only. Then it was
adjusted by e46e7b77c909 ("mm, page_alloc: recalculate the preferred
zoneref if the context can ignore memory policies") and cd04ae1e2dc8
("mm, oom: do not rely on TIF_MEMDIE for memory reserves access") to the
current state. So even GFP_ATOMIC would now ignore mempolicies after
the initial attempts fail - if the code worked as people thought it
does.
Link: http://lkml.kernel.org/r/20180612122624.8045-1-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-17 15:45:05 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-06-03 14:56:01 -07:00
|
|
|
|
2017-09-06 16:24:50 -07:00
|
|
|
|
mm, page_alloc: actually ignore mempolicies for high priority allocations
__alloc_pages_slowpath() has for a long time contained code to ignore
node restrictions from memory policies for high priority allocations.
The current code that resets the zonelist iterator however does
effectively nothing after commit 7810e6781e0f ("mm, page_alloc: do not
break __GFP_THISNODE by zonelist reset") removed a buggy zonelist reset.
Even before that commit, mempolicy restrictions were still not ignored,
as they are passed in ac->nodemask which is untouched by the code.
We can either remove the code, or make it work as intended. Since
ac->nodemask can be set from task's mempolicy via alloc_pages_current()
and thus also alloc_pages(), it may indeed affect kernel allocations,
and it makes sense to ignore it to allow progress for high priority
allocations.
Thus, this patch resets ac->nodemask to NULL in such cases. This
assumes all callers can handle it (i.e. there are no guarantees as in
the case of __GFP_THISNODE) which seems to be the case. The same
assumption is already present in check_retry_cpuset() for some time.
The expected effect is that high priority kernel allocations in the
context of userspace tasks (e.g. OOM victims) restricted by mempolicies
will have higher chance to succeed if they are restricted to nodes with
depleted memory, while there are other nodes with free memory left.
It's not a new intention, but for the first time the code will match the
intention, AFAICS. It was intended by commit 183f6371aac2 ("mm: ignore
mempolicies when using ALLOC_NO_WATERMARK") in v3.6 but I think it never
really worked, as mempolicy restriction was already encoded in nodemask,
not zonelist, at that time.
So originally that was for ALLOC_NO_WATERMARK only. Then it was
adjusted by e46e7b77c909 ("mm, page_alloc: recalculate the preferred
zoneref if the context can ignore memory policies") and cd04ae1e2dc8
("mm, oom: do not rely on TIF_MEMDIE for memory reserves access") to the
current state. So even GFP_ATOMIC would now ignore mempolicies after
the initial attempts fail - if the code worked as people thought it
does.
Link: http://lkml.kernel.org/r/20180612122624.8045-1-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-17 15:45:05 -07:00
|
|
|
|
2016-06-03 14:56:01 -07:00
|
|
|
|
2020-06-03 15:59:01 -07:00
|
|
|
|
2016-06-03 14:56:01 -07:00
|
|
|
|
|
|
|
|
|
2016-07-28 15:49:16 -07:00
|
|
|
|
2016-07-28 15:49:13 -07:00
|
|
|
|
2005-11-13 16:06:43 -08:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2015-11-06 16:28:21 -08:00
|
|
|
|
2017-02-22 15:46:19 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2017-02-22 15:46:19 -08:00
|
|
|
|
2009-06-16 15:32:02 -07:00
|
|
|
|
2017-02-22 15:46:19 -08:00
|
|
|
|
2009-06-16 15:32:02 -07:00
|
|
|
|
2014-08-06 16:07:54 -07:00
|
|
|
|
2009-06-16 15:31:57 -07:00
|
|
|
|
2015-02-11 15:25:41 -08:00
|
|
|
|
|
|
|
|
|
2009-06-16 15:31:57 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
mm, page_alloc: restructure direct compaction handling in slowpath
The retry loop in __alloc_pages_slowpath is supposed to keep trying
reclaim and compaction (and OOM), until either the allocation succeeds,
or returns with failure. Success here is more probable when reclaim
precedes compaction, as certain watermarks have to be met for compaction
to even try, and more free pages increase the probability of compaction
success. On the other hand, starting with light async compaction (if
the watermarks allow it), can be more efficient, especially for smaller
orders, if there's enough free memory which is just fragmented.
Thus, the current code starts with compaction before reclaim, and to
make sure that the last reclaim is always followed by a final
compaction, there's another direct compaction call at the end of the
loop. This makes the code hard to follow and adds some duplicated
handling of migration_mode decisions. It's also somewhat inefficient
that even if reclaim or compaction decides not to retry, the final
compaction is still attempted. Some gfp flags combination also shortcut
these retry decisions by "goto noretry;", making it even harder to
follow.
This patch attempts to restructure the code with only minimal functional
changes. The call to the first compaction and THP-specific checks are
now placed above the retry loop, and the "noretry" direct compaction is
removed.
The initial compaction is additionally restricted only to costly orders,
as we can expect smaller orders to be held back by watermarks, and only
larger orders to suffer primarily from fragmentation. This better
matches the checks in reclaim's shrink_zones().
There are two other smaller functional changes. One is that the upgrade
from async migration to light sync migration will always occur after the
initial compaction. This is how it has been until recent patch "mm,
oom: protect !costly allocations some more", which introduced upgrading
the mode based on COMPACT_COMPLETE result, but kept the final compaction
always upgraded, which made it even more special. It's better to return
to the simpler handling for now, as migration modes will be further
modified later in the series.
The second change is that once both reclaim and compaction declare it's
not worth to retry the reclaim/compact loop, there is no final
compaction attempt. As argued above, this is intentional. If that
final compaction were to succeed, it would be due to a wrong retry
decision, or simply a race with somebody else freeing memory for us.
The main outcome of this patch should be simpler code. Logically, the
initial compaction without reclaim is the exceptional case to the
reclaim/compaction scheme, but prior to the patch, it was the last loop
iteration that was exceptional. Now the code matches the logic better.
The change also enable the following patches.
Link: http://lkml.kernel.org/r/20160721073614.24395-5-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-07-28 15:49:19 -07:00
|
|
|
|
|
|
|
|
|
2016-07-28 15:49:28 -07:00
|
|
|
|
mm, page_alloc: restructure direct compaction handling in slowpath
The retry loop in __alloc_pages_slowpath is supposed to keep trying
reclaim and compaction (and OOM), until either the allocation succeeds,
or returns with failure. Success here is more probable when reclaim
precedes compaction, as certain watermarks have to be met for compaction
to even try, and more free pages increase the probability of compaction
success. On the other hand, starting with light async compaction (if
the watermarks allow it), can be more efficient, especially for smaller
orders, if there's enough free memory which is just fragmented.
Thus, the current code starts with compaction before reclaim, and to
make sure that the last reclaim is always followed by a final
compaction, there's another direct compaction call at the end of the
loop. This makes the code hard to follow and adds some duplicated
handling of migration_mode decisions. It's also somewhat inefficient
that even if reclaim or compaction decides not to retry, the final
compaction is still attempted. Some gfp flags combination also shortcut
these retry decisions by "goto noretry;", making it even harder to
follow.
This patch attempts to restructure the code with only minimal functional
changes. The call to the first compaction and THP-specific checks are
now placed above the retry loop, and the "noretry" direct compaction is
removed.
The initial compaction is additionally restricted only to costly orders,
as we can expect smaller orders to be held back by watermarks, and only
larger orders to suffer primarily from fragmentation. This better
matches the checks in reclaim's shrink_zones().
There are two other smaller functional changes. One is that the upgrade
from async migration to light sync migration will always occur after the
initial compaction. This is how it has been until recent patch "mm,
oom: protect !costly allocations some more", which introduced upgrading
the mode based on COMPACT_COMPLETE result, but kept the final compaction
always upgraded, which made it even more special. It's better to return
to the simpler handling for now, as migration modes will be further
modified later in the series.
The second change is that once both reclaim and compaction declare it's
not worth to retry the reclaim/compact loop, there is no final
compaction attempt. As argued above, this is intentional. If that
final compaction were to succeed, it would be due to a wrong retry
decision, or simply a race with somebody else freeing memory for us.
The main outcome of this patch should be simpler code. Logically, the
initial compaction without reclaim is the exceptional case to the
reclaim/compaction scheme, but prior to the patch, it was the last loop
iteration that was exceptional. Now the code matches the logic better.
The change also enable the following patches.
Link: http://lkml.kernel.org/r/20160721073614.24395-5-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-07-28 15:49:19 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-06-24 16:57:21 -07:00
|
|
|
|
|
|
|
|
|
mm, page_alloc: restructure direct compaction handling in slowpath
The retry loop in __alloc_pages_slowpath is supposed to keep trying
reclaim and compaction (and OOM), until either the allocation succeeds,
or returns with failure. Success here is more probable when reclaim
precedes compaction, as certain watermarks have to be met for compaction
to even try, and more free pages increase the probability of compaction
success. On the other hand, starting with light async compaction (if
the watermarks allow it), can be more efficient, especially for smaller
orders, if there's enough free memory which is just fragmented.
Thus, the current code starts with compaction before reclaim, and to
make sure that the last reclaim is always followed by a final
compaction, there's another direct compaction call at the end of the
loop. This makes the code hard to follow and adds some duplicated
handling of migration_mode decisions. It's also somewhat inefficient
that even if reclaim or compaction decides not to retry, the final
compaction is still attempted. Some gfp flags combination also shortcut
these retry decisions by "goto noretry;", making it even harder to
follow.
This patch attempts to restructure the code with only minimal functional
changes. The call to the first compaction and THP-specific checks are
now placed above the retry loop, and the "noretry" direct compaction is
removed.
The initial compaction is additionally restricted only to costly orders,
as we can expect smaller orders to be held back by watermarks, and only
larger orders to suffer primarily from fragmentation. This better
matches the checks in reclaim's shrink_zones().
There are two other smaller functional changes. One is that the upgrade
from async migration to light sync migration will always occur after the
initial compaction. This is how it has been until recent patch "mm,
oom: protect !costly allocations some more", which introduced upgrading
the mode based on COMPACT_COMPLETE result, but kept the final compaction
always upgraded, which made it even more special. It's better to return
to the simpler handling for now, as migration modes will be further
modified later in the series.
The second change is that once both reclaim and compaction declare it's
not worth to retry the reclaim/compact loop, there is no final
compaction attempt. As argued above, this is intentional. If that
final compaction were to succeed, it would be due to a wrong retry
decision, or simply a race with somebody else freeing memory for us.
The main outcome of this patch should be simpler code. Logically, the
initial compaction without reclaim is the exceptional case to the
reclaim/compaction scheme, but prior to the patch, it was the last loop
iteration that was exceptional. Now the code matches the logic better.
The change also enable the following patches.
Link: http://lkml.kernel.org/r/20160721073614.24395-5-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-07-28 15:49:19 -07:00
|
|
|
|
2015-06-24 16:57:21 -07:00
|
|
|
|
2016-05-20 16:57:00 -07:00
|
|
|
|
|
|
|
|
|
mm, vmscan: prevent infinite loop for costly GFP_NOIO | __GFP_RETRY_MAYFAIL allocations
Sven reports an infinite loop in __alloc_pages_slowpath() for costly order
__GFP_RETRY_MAYFAIL allocations that are also GFP_NOIO. Such combination
can happen in a suspend/resume context where a GFP_KERNEL allocation can
have __GFP_IO masked out via gfp_allowed_mask.
Quoting Sven:
1. try to do a "costly" allocation (order > PAGE_ALLOC_COSTLY_ORDER)
with __GFP_RETRY_MAYFAIL set.
2. page alloc's __alloc_pages_slowpath tries to get a page from the
freelist. This fails because there is nothing free of that costly
order.
3. page alloc tries to reclaim by calling __alloc_pages_direct_reclaim,
which bails out because a zone is ready to be compacted; it pretends
to have made a single page of progress.
4. page alloc tries to compact, but this always bails out early because
__GFP_IO is not set (it's not passed by the snd allocator, and even
if it were, we are suspending so the __GFP_IO flag would be cleared
anyway).
5. page alloc believes reclaim progress was made (because of the
pretense in item 3) and so it checks whether it should retry
compaction. The compaction retry logic thinks it should try again,
because:
a) reclaim is needed because of the early bail-out in item 4
b) a zonelist is suitable for compaction
6. goto 2. indefinite stall.
(end quote)
The immediate root cause is confusing the COMPACT_SKIPPED returned from
__alloc_pages_direct_compact() (step 4) due to lack of __GFP_IO to be
indicating a lack of order-0 pages, and in step 5 evaluating that in
should_compact_retry() as a reason to retry, before incrementing and
limiting the number of retries. There are however other places that
wrongly assume that compaction can happen while we lack __GFP_IO.
To fix this, introduce gfp_compaction_allowed() to abstract the __GFP_IO
evaluation and switch the open-coded test in try_to_compact_pages() to use
it.
Also use the new helper in:
- compaction_ready(), which will make reclaim not bail out in step 3, so
there's at least one attempt to actually reclaim, even if chances are
small for a costly order
- in_reclaim_compaction() which will make should_continue_reclaim()
return false and we don't over-reclaim unnecessarily
- in __alloc_pages_slowpath() to set a local variable can_compact,
which is then used to avoid retrying reclaim/compaction for costly
allocations (step 5) if we can't compact and also to skip the early
compaction attempt that we do in some cases
Link: https://lkml.kernel.org/r/20240221114357.13655-2-vbabka@suse.cz
Fixes: 3250845d0526 ("Revert "mm, oom: prevent premature OOM killer invocation for high order request"")
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Reported-by: Sven van Ashbrook <svenva@chromium.org>
Closes: https://lore.kernel.org/all/CAG-rBihs_xMKb3wrMO1%2B-%2Bp4fowP9oy1pa_OTkfxBzPUVOZF%2Bg@mail.gmail.com/
Tested-by: Karthikeyan Ramasubramanian <kramasub@chromium.org>
Cc: Brian Geffon <bgeffon@google.com>
Cc: Curtis Malainey <cujomalainey@chromium.org>
Cc: Jaroslav Kysela <perex@perex.cz>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Takashi Iwai <tiwai@suse.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-21 12:43:58 +01:00
|
|
|
|
2016-05-20 16:57:00 -07:00
|
|
|
|
mm, vmscan: prevent infinite loop for costly GFP_NOIO | __GFP_RETRY_MAYFAIL allocations
Sven reports an infinite loop in __alloc_pages_slowpath() for costly order
__GFP_RETRY_MAYFAIL allocations that are also GFP_NOIO. Such combination
can happen in a suspend/resume context where a GFP_KERNEL allocation can
have __GFP_IO masked out via gfp_allowed_mask.
Quoting Sven:
1. try to do a "costly" allocation (order > PAGE_ALLOC_COSTLY_ORDER)
with __GFP_RETRY_MAYFAIL set.
2. page alloc's __alloc_pages_slowpath tries to get a page from the
freelist. This fails because there is nothing free of that costly
order.
3. page alloc tries to reclaim by calling __alloc_pages_direct_reclaim,
which bails out because a zone is ready to be compacted; it pretends
to have made a single page of progress.
4. page alloc tries to compact, but this always bails out early because
__GFP_IO is not set (it's not passed by the snd allocator, and even
if it were, we are suspending so the __GFP_IO flag would be cleared
anyway).
5. page alloc believes reclaim progress was made (because of the
pretense in item 3) and so it checks whether it should retry
compaction. The compaction retry logic thinks it should try again,
because:
a) reclaim is needed because of the early bail-out in item 4
b) a zonelist is suitable for compaction
6. goto 2. indefinite stall.
(end quote)
The immediate root cause is confusing the COMPACT_SKIPPED returned from
__alloc_pages_direct_compact() (step 4) due to lack of __GFP_IO to be
indicating a lack of order-0 pages, and in step 5 evaluating that in
should_compact_retry() as a reason to retry, before incrementing and
limiting the number of retries. There are however other places that
wrongly assume that compaction can happen while we lack __GFP_IO.
To fix this, introduce gfp_compaction_allowed() to abstract the __GFP_IO
evaluation and switch the open-coded test in try_to_compact_pages() to use
it.
Also use the new helper in:
- compaction_ready(), which will make reclaim not bail out in step 3, so
there's at least one attempt to actually reclaim, even if chances are
small for a costly order
- in_reclaim_compaction() which will make should_continue_reclaim()
return false and we don't over-reclaim unnecessarily
- in __alloc_pages_slowpath() to set a local variable can_compact,
which is then used to avoid retrying reclaim/compaction for costly
allocations (step 5) if we can't compact and also to skip the early
compaction attempt that we do in some cases
Link: https://lkml.kernel.org/r/20240221114357.13655-2-vbabka@suse.cz
Fixes: 3250845d0526 ("Revert "mm, oom: prevent premature OOM killer invocation for high order request"")
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Reported-by: Sven van Ashbrook <svenva@chromium.org>
Closes: https://lore.kernel.org/all/CAG-rBihs_xMKb3wrMO1%2B-%2Bp4fowP9oy1pa_OTkfxBzPUVOZF%2Bg@mail.gmail.com/
Tested-by: Karthikeyan Ramasubramanian <kramasub@chromium.org>
Cc: Brian Geffon <bgeffon@google.com>
Cc: Curtis Malainey <cujomalainey@chromium.org>
Cc: Jaroslav Kysela <perex@perex.cz>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Takashi Iwai <tiwai@suse.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-21 12:43:58 +01:00
|
|
|
|
|
|
|
|
|
mm, page_alloc: restructure direct compaction handling in slowpath
The retry loop in __alloc_pages_slowpath is supposed to keep trying
reclaim and compaction (and OOM), until either the allocation succeeds,
or returns with failure. Success here is more probable when reclaim
precedes compaction, as certain watermarks have to be met for compaction
to even try, and more free pages increase the probability of compaction
success. On the other hand, starting with light async compaction (if
the watermarks allow it), can be more efficient, especially for smaller
orders, if there's enough free memory which is just fragmented.
Thus, the current code starts with compaction before reclaim, and to
make sure that the last reclaim is always followed by a final
compaction, there's another direct compaction call at the end of the
loop. This makes the code hard to follow and adds some duplicated
handling of migration_mode decisions. It's also somewhat inefficient
that even if reclaim or compaction decides not to retry, the final
compaction is still attempted. Some gfp flags combination also shortcut
these retry decisions by "goto noretry;", making it even harder to
follow.
This patch attempts to restructure the code with only minimal functional
changes. The call to the first compaction and THP-specific checks are
now placed above the retry loop, and the "noretry" direct compaction is
removed.
The initial compaction is additionally restricted only to costly orders,
as we can expect smaller orders to be held back by watermarks, and only
larger orders to suffer primarily from fragmentation. This better
matches the checks in reclaim's shrink_zones().
There are two other smaller functional changes. One is that the upgrade
from async migration to light sync migration will always occur after the
initial compaction. This is how it has been until recent patch "mm,
oom: protect !costly allocations some more", which introduced upgrading
the mode based on COMPACT_COMPLETE result, but kept the final compaction
always upgraded, which made it even more special. It's better to return
to the simpler handling for now, as migration modes will be further
modified later in the series.
The second change is that once both reclaim and compaction declare it's
not worth to retry the reclaim/compact loop, there is no final
compaction attempt. As argued above, this is intentional. If that
final compaction were to succeed, it would be due to a wrong retry
decision, or simply a race with somebody else freeing memory for us.
The main outcome of this patch should be simpler code. Logically, the
initial compaction without reclaim is the exceptional case to the
reclaim/compaction scheme, but prior to the patch, it was the last loop
iteration that was exceptional. Now the code matches the logic better.
The change also enable the following patches.
Link: http://lkml.kernel.org/r/20160721073614.24395-5-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-07-28 15:49:19 -07:00
|
|
|
|
2016-05-20 16:57:00 -07:00
|
|
|
|
|
|
|
|
|
2016-10-07 17:00:40 -07:00
|
|
|
|
2016-05-20 16:57:00 -07:00
|
|
|
|
|
|
|
|
|
2016-05-20 16:57:06 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, vmscan: prevent infinite loop for costly GFP_NOIO | __GFP_RETRY_MAYFAIL allocations
Sven reports an infinite loop in __alloc_pages_slowpath() for costly order
__GFP_RETRY_MAYFAIL allocations that are also GFP_NOIO. Such combination
can happen in a suspend/resume context where a GFP_KERNEL allocation can
have __GFP_IO masked out via gfp_allowed_mask.
Quoting Sven:
1. try to do a "costly" allocation (order > PAGE_ALLOC_COSTLY_ORDER)
with __GFP_RETRY_MAYFAIL set.
2. page alloc's __alloc_pages_slowpath tries to get a page from the
freelist. This fails because there is nothing free of that costly
order.
3. page alloc tries to reclaim by calling __alloc_pages_direct_reclaim,
which bails out because a zone is ready to be compacted; it pretends
to have made a single page of progress.
4. page alloc tries to compact, but this always bails out early because
__GFP_IO is not set (it's not passed by the snd allocator, and even
if it were, we are suspending so the __GFP_IO flag would be cleared
anyway).
5. page alloc believes reclaim progress was made (because of the
pretense in item 3) and so it checks whether it should retry
compaction. The compaction retry logic thinks it should try again,
because:
a) reclaim is needed because of the early bail-out in item 4
b) a zonelist is suitable for compaction
6. goto 2. indefinite stall.
(end quote)
The immediate root cause is confusing the COMPACT_SKIPPED returned from
__alloc_pages_direct_compact() (step 4) due to lack of __GFP_IO to be
indicating a lack of order-0 pages, and in step 5 evaluating that in
should_compact_retry() as a reason to retry, before incrementing and
limiting the number of retries. There are however other places that
wrongly assume that compaction can happen while we lack __GFP_IO.
To fix this, introduce gfp_compaction_allowed() to abstract the __GFP_IO
evaluation and switch the open-coded test in try_to_compact_pages() to use
it.
Also use the new helper in:
- compaction_ready(), which will make reclaim not bail out in step 3, so
there's at least one attempt to actually reclaim, even if chances are
small for a costly order
- in_reclaim_compaction() which will make should_continue_reclaim()
return false and we don't over-reclaim unnecessarily
- in __alloc_pages_slowpath() to set a local variable can_compact,
which is then used to avoid retrying reclaim/compaction for costly
allocations (step 5) if we can't compact and also to skip the early
compaction attempt that we do in some cases
Link: https://lkml.kernel.org/r/20240221114357.13655-2-vbabka@suse.cz
Fixes: 3250845d0526 ("Revert "mm, oom: prevent premature OOM killer invocation for high order request"")
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Reported-by: Sven van Ashbrook <svenva@chromium.org>
Closes: https://lore.kernel.org/all/CAG-rBihs_xMKb3wrMO1%2B-%2Bp4fowP9oy1pa_OTkfxBzPUVOZF%2Bg@mail.gmail.com/
Tested-by: Karthikeyan Ramasubramanian <kramasub@chromium.org>
Cc: Brian Geffon <bgeffon@google.com>
Cc: Curtis Malainey <cujomalainey@chromium.org>
Cc: Jaroslav Kysela <perex@perex.cz>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Takashi Iwai <tiwai@suse.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-21 12:43:58 +01:00
|
|
|
|
2016-05-20 16:57:12 -07:00
|
|
|
|
2016-07-28 15:49:28 -07:00
|
|
|
|
2016-10-07 17:00:31 -07:00
|
|
|
|
2016-05-20 16:57:06 -07:00
|
|
|
|
|
|
|
|
|
mm: page_alloc: defrag_mode
The page allocator groups requests by migratetype to stave off
fragmentation. However, in practice this is routinely defeated by the
fact that it gives up *before* invoking reclaim and compaction - which may
well produce suitable pages. As a result, fragmentation of physical
memory is a common ongoing process in many load scenarios.
Fragmentation deteriorates compaction's ability to produce huge pages.
Depending on the lifetime of the fragmenting allocations, those effects
can be long-lasting or even permanent, requiring drastic measures like
forcible idle states or even reboots as the only reliable ways to recover
the address space for THP production.
In a kernel build test with supplemental THP pressure, the THP allocation
rate steadily declines over 15 runs:
thp_fault_alloc
61988
56474
57258
50187
52388
55409
52925
47648
43669
40621
36077
41721
36685
34641
33215
This is a hurdle in adopting THP in any environment where hosts are shared
between multiple overlapping workloads (cloud environments), and rarely
experience true idle periods. To make THP a reliable and predictable
optimization, there needs to be a stronger guarantee to avoid such
fragmentation.
Introduce defrag_mode. When enabled, reclaim/compaction is invoked to its
full extent *before* falling back. Specifically, ALLOC_NOFRAGMENT is
enforced on the allocator fastpath and the reclaiming slowpath.
For now, fallbacks are permitted to avert OOMs. There is a plan to add
defrag_mode=2 to prefer OOMs over fragmentation, but this requires
additional prep work in compaction and the reserve management to make it
ready for all possible allocation contexts.
The following test results are from a kernel build with periodic bursts of
THP allocations, over 15 runs:
vanilla defrag_mode=1
@claimer[unmovable]: 189 103
@claimer[movable]: 92 103
@claimer[reclaimable]: 207 61
@pollute[unmovable from movable]: 25 0
@pollute[unmovable from reclaimable]: 28 0
@pollute[movable from unmovable]: 38835 0
@pollute[movable from reclaimable]: 147136 0
@pollute[reclaimable from unmovable]: 178 0
@pollute[reclaimable from movable]: 33 0
@steal[unmovable from movable]: 11 0
@steal[unmovable from reclaimable]: 5 0
@steal[reclaimable from unmovable]: 107 0
@steal[reclaimable from movable]: 90 0
@steal[movable from reclaimable]: 354 0
@steal[movable from unmovable]: 130 0
Both types of polluting fallbacks are eliminated in this workload.
Interestingly, whole block conversions are reduced as well. This is
because once a block is claimed for a type, its empty space remains
available for future allocations, instead of being padded with fallbacks;
this allows the native type to group up instead of spreading out to new
blocks. The assumption in the allocator has been that pollution from
movable allocations is less harmful than from other types, since they can
be reclaimed or migrated out should the space be needed. However, since
fallbacks occur *before* reclaim/compaction is invoked, movable pollution
will still cause non-movable allocations to spread out and claim more
blocks.
Without fragmentation, THP rates hold steady with defrag_mode=1:
thp_fault_alloc
32478
20725
45045
32130
14018
21711
40791
29134
34458
45381
28305
17265
22584
28454
30850
While the downward trend is eliminated, the keen reader will of course
notice that the baseline rate is much smaller than the vanilla kernel's to
begin with. This is due to deficiencies in how reclaim and compaction are
currently driven: ALLOC_NOFRAGMENT increases the extent to which smaller
allocations are competing with THPs for pageblocks, while making no effort
themselves to reclaim or compact beyond their own request size. This
effect already exists with the current usage of ALLOC_NOFRAGMENT, but is
amplified by defrag_mode insisting on whole block stealing much more
strongly.
Subsequent patches will address defrag_mode reclaim strategy to raise the
THP success baseline above the vanilla kernel.
Link: https://lkml.kernel.org/r/20250313210647.1314586-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-03-13 17:05:34 -04:00
|
|
|
|
2025-03-22 19:21:45 -04:00
|
|
|
|
|
|
|
|
|
mm: page_alloc: defrag_mode
The page allocator groups requests by migratetype to stave off
fragmentation. However, in practice this is routinely defeated by the
fact that it gives up *before* invoking reclaim and compaction - which may
well produce suitable pages. As a result, fragmentation of physical
memory is a common ongoing process in many load scenarios.
Fragmentation deteriorates compaction's ability to produce huge pages.
Depending on the lifetime of the fragmenting allocations, those effects
can be long-lasting or even permanent, requiring drastic measures like
forcible idle states or even reboots as the only reliable ways to recover
the address space for THP production.
In a kernel build test with supplemental THP pressure, the THP allocation
rate steadily declines over 15 runs:
thp_fault_alloc
61988
56474
57258
50187
52388
55409
52925
47648
43669
40621
36077
41721
36685
34641
33215
This is a hurdle in adopting THP in any environment where hosts are shared
between multiple overlapping workloads (cloud environments), and rarely
experience true idle periods. To make THP a reliable and predictable
optimization, there needs to be a stronger guarantee to avoid such
fragmentation.
Introduce defrag_mode. When enabled, reclaim/compaction is invoked to its
full extent *before* falling back. Specifically, ALLOC_NOFRAGMENT is
enforced on the allocator fastpath and the reclaiming slowpath.
For now, fallbacks are permitted to avert OOMs. There is a plan to add
defrag_mode=2 to prefer OOMs over fragmentation, but this requires
additional prep work in compaction and the reserve management to make it
ready for all possible allocation contexts.
The following test results are from a kernel build with periodic bursts of
THP allocations, over 15 runs:
vanilla defrag_mode=1
@claimer[unmovable]: 189 103
@claimer[movable]: 92 103
@claimer[reclaimable]: 207 61
@pollute[unmovable from movable]: 25 0
@pollute[unmovable from reclaimable]: 28 0
@pollute[movable from unmovable]: 38835 0
@pollute[movable from reclaimable]: 147136 0
@pollute[reclaimable from unmovable]: 178 0
@pollute[reclaimable from movable]: 33 0
@steal[unmovable from movable]: 11 0
@steal[unmovable from reclaimable]: 5 0
@steal[reclaimable from unmovable]: 107 0
@steal[reclaimable from movable]: 90 0
@steal[movable from reclaimable]: 354 0
@steal[movable from unmovable]: 130 0
Both types of polluting fallbacks are eliminated in this workload.
Interestingly, whole block conversions are reduced as well. This is
because once a block is claimed for a type, its empty space remains
available for future allocations, instead of being padded with fallbacks;
this allows the native type to group up instead of spreading out to new
blocks. The assumption in the allocator has been that pollution from
movable allocations is less harmful than from other types, since they can
be reclaimed or migrated out should the space be needed. However, since
fallbacks occur *before* reclaim/compaction is invoked, movable pollution
will still cause non-movable allocations to spread out and claim more
blocks.
Without fragmentation, THP rates hold steady with defrag_mode=1:
thp_fault_alloc
32478
20725
45045
32130
14018
21711
40791
29134
34458
45381
28305
17265
22584
28454
30850
While the downward trend is eliminated, the keen reader will of course
notice that the baseline rate is much smaller than the vanilla kernel's to
begin with. This is due to deficiencies in how reclaim and compaction are
currently driven: ALLOC_NOFRAGMENT increases the extent to which smaller
allocations are competing with THPs for pageblocks, while making no effort
themselves to reclaim or compact beyond their own request size. This
effect already exists with the current usage of ALLOC_NOFRAGMENT, but is
amplified by defrag_mode insisting on whole block stealing much more
strongly.
Subsequent patches will address defrag_mode reclaim strategy to raise the
THP success baseline above the vanilla kernel.
Link: https://lkml.kernel.org/r/20250313210647.1314586-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-03-13 17:05:34 -04:00
|
|
|
|
|
|
|
|
|
mm, page_alloc: fix more premature OOM due to race with cpuset update
I would like to stress that this patchset aims to fix issues and cleanup
the code *within the existing documented semantics*, i.e. patch 1
ignores mempolicy restrictions if the set of allowed nodes has no
intersection with set of nodes allowed by cpuset. I believe discussing
potential changes of the semantics can be better done once we have a
baseline with no known bugs of the current semantics.
I've recently summarized the cpuset/mempolicy issues in a LSF/MM
proposal [1] and the discussion itself [2]. I've been trying to rewrite
the handling as proposed, with the idea that changing semantics to make
all mempolicies static wrt cpuset updates (and discarding the relative
and default modes) can be tried on top, as there's a high risk of being
rejected/reverted because somebody might still care about the removed
modes.
However I haven't yet figured out how to properly:
1) make mempolicies swappable instead of rebinding in place. I thought
mbind() already works that way and uses refcounting to avoid
use-after-free of the old policy by a parallel allocation, but turns
out true refcounting is only done for shared (shmem) mempolicies, and
the actual protection for mbind() comes from mmap_sem. Extending the
refcounting means more overhead in allocator hot path. Also swapping
whole mempolicies means that we have to allocate the new ones, which
can fail, and reverting of the partially done work also means
allocating (note that mbind() doesn't care and will just leave part
of the range updated and part not updated when returning -ENOMEM...).
2) make cpuset's task->mems_allowed also swappable (after converting it
from nodemask to zonelist, which is the easy part) for mostly the
same reasons.
The good news is that while trying to do the above, I've at least
figured out how to hopefully close the remaining premature OOM's, and do
a buch of cleanups on top, removing quite some of the code that was also
supposed to prevent the cpuset update races, but doesn't work anymore
nowadays. This should fix the most pressing concerns with this topic
and give us a better baseline before either proceeding with the original
proposal, or pushing a change of semantics that removes the problem 1)
above. I'd be then fine with trying to change the semantic first and
rewrite later.
Patchset has been tested with the LTP cpuset01 stress test.
[1] https://lkml.kernel.org/r/4c44a589-5fd8-08d0-892c-e893bb525b71@suse.cz
[2] https://lwn.net/Articles/717797/
[3] https://marc.info/?l=linux-mm&m=149191957922828&w=2
This patch (of 6):
Commit e47483bca2cc ("mm, page_alloc: fix premature OOM when racing with
cpuset mems update") has fixed known recent regressions found by LTP's
cpuset01 testcase. I have however found that by modifying the testcase
to use per-vma mempolicies via bind(2) instead of per-task mempolicies
via set_mempolicy(2), the premature OOM still happens and the issue is
much older.
The root of the problem is that the cpuset's mems_allowed and
mempolicy's nodemask can temporarily have no intersection, thus
get_page_from_freelist() cannot find any usable zone. The current
semantic for empty intersection is to ignore mempolicy's nodemask and
honour cpuset restrictions. This is checked in node_zonelist(), but the
racy update can happen after we already passed the check. Such races
should be protected by the seqlock task->mems_allowed_seq, but it
doesn't work here, because 1) mpol_rebind_mm() does not happen under
seqlock for write, and doing so would lead to deadlock, as it takes
mmap_sem for write, while the allocation can have mmap_sem for read when
it's taking the seqlock for read. And 2) the seqlock cookie of callers
of node_zonelist() (alloc_pages_vma() and alloc_pages_current()) is
different than the one of __alloc_pages_slowpath(), so there's still a
potential race window.
This patch fixes the issue by having __alloc_pages_slowpath() check for
empty intersection of cpuset and ac->nodemask before OOM or allocation
failure. If it's indeed empty, the nodemask is ignored and allocation
retried, which mimics node_zonelist(). This works fine, because almost
all callers of __alloc_pages_nodemask are obtaining the nodemask via
node_zonelist(). The only exception is new_node_page() from hotplug,
where the potential violation of nodemask isn't an issue, as there's
already a fallback allocation attempt without any nodemask. If there's
a future caller that needs to have its specific nodemask honoured over
task's cpuset restrictions, we'll have to e.g. add a gfp flag for that.
Link: http://lkml.kernel.org/r/20170517081140.30654-2-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Li Zefan <lizefan@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: David Rientjes <rientjes@google.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Anshuman Khandual <khandual@linux.vnet.ibm.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Dimitri Sivanich <sivanich@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 15:39:56 -07:00
|
|
|
|
mm/page_alloc: fix race condition between build_all_zonelists and page allocation
Patrick Daly reported the following problem;
NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK] - before offline operation
[0] - ZONE_MOVABLE
[1] - ZONE_NORMAL
[2] - NULL
For a GFP_KERNEL allocation, alloc_pages_slowpath() will save the
offset of ZONE_NORMAL in ac->preferred_zoneref. If a concurrent
memory_offline operation removes the last page from ZONE_MOVABLE,
build_all_zonelists() & build_zonerefs_node() will update
node_zonelists as shown below. Only populated zones are added.
NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK] - after offline operation
[0] - ZONE_NORMAL
[1] - NULL
[2] - NULL
The race is simple -- page allocation could be in progress when a memory
hot-remove operation triggers a zonelist rebuild that removes zones. The
allocation request will still have a valid ac->preferred_zoneref that is
now pointing to NULL and triggers an OOM kill.
This problem probably always existed but may be slightly easier to trigger
due to 6aa303defb74 ("mm, vmscan: only allocate and reclaim from zones
with pages managed by the buddy allocator") which distinguishes between
zones that are completely unpopulated versus zones that have valid pages
not managed by the buddy allocator (e.g. reserved, memblock, ballooning
etc). Memory hotplug had multiple stages with timing considerations
around managed/present page updates, the zonelist rebuild and the zone
span updates. As David Hildenbrand puts it
memory offlining adjusts managed+present pages of the zone
essentially in one go. If after the adjustments, the zone is no
longer populated (present==0), we rebuild the zone lists.
Once that's done, we try shrinking the zone (start+spanned
pages) -- which results in zone_start_pfn == 0 if there are no
more pages. That happens *after* rebuilding the zonelists via
remove_pfn_range_from_zone().
The only requirement to fix the race is that a page allocation request
identifies when a zonelist rebuild has happened since the allocation
request started and no page has yet been allocated. Use a seqlock_t to
track zonelist updates with a lockless read-side of the zonelist and
protecting the rebuild and update of the counter with a spinlock.
[akpm@linux-foundation.org: make zonelist_update_seq static]
Link: https://lkml.kernel.org/r/20220824110900.vh674ltxmzb3proq@techsingularity.net
Fixes: 6aa303defb74 ("mm, vmscan: only allocate and reclaim from zones with pages managed by the buddy allocator")
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Reported-by: Patrick Daly <quic_pdaly@quicinc.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: David Hildenbrand <david@redhat.com>
Cc: <stable@vger.kernel.org> [4.9+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-08-24 12:14:50 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, page_alloc: fix premature OOM when racing with cpuset mems update
Ganapatrao Kulkarni reported that the LTP test cpuset01 in stress mode
triggers OOM killer in few seconds, despite lots of free memory. The
test attempts to repeatedly fault in memory in one process in a cpuset,
while changing allowed nodes of the cpuset between 0 and 1 in another
process.
The problem comes from insufficient protection against cpuset changes,
which can cause get_page_from_freelist() to consider all zones as
non-eligible due to nodemask and/or current->mems_allowed. This was
masked in the past by sufficient retries, but since commit 682a3385e773
("mm, page_alloc: inline the fast path of the zonelist iterator") we fix
the preferred_zoneref once, and don't iterate over the whole zonelist in
further attempts, thus the only eligible zones might be placed in the
zonelist before our starting point and we always miss them.
A previous patch fixed this problem for current->mems_allowed. However,
cpuset changes also update the task's mempolicy nodemask. The fix has
two parts. We have to repeat the preferred_zoneref search when we
detect cpuset update by way of seqcount, and we have to check the
seqcount before considering OOM.
[akpm@linux-foundation.org: fix typo in comment]
Link: http://lkml.kernel.org/r/20170120103843.24587-5-vbabka@suse.cz
Fixes: c33d6c06f60f ("mm, page_alloc: avoid looking up the first zone in a zonelist twice")
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Reported-by: Ganapatrao Kulkarni <gpkulkarni@gmail.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-01-24 15:18:41 -08:00
|
|
|
|
2015-06-24 16:57:21 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-02-22 15:46:19 -08:00
|
|
|
|
2017-09-06 16:24:50 -07:00
|
|
|
|
2020-08-06 23:26:04 -07:00
|
|
|
|
mm/page_alloc.c: make sure OOM victim can try allocations with no watermarks once
Roman Gushchin has reported that the OOM killer can trivially selects
next OOM victim when a thread doing memory allocation from page fault
path was selected as first OOM victim.
allocate invoked oom-killer: gfp_mask=0x14280ca(GFP_HIGHUSER_MOVABLE|__GFP_ZERO), nodemask=(null), order=0, oom_score_adj=0
allocate cpuset=/ mems_allowed=0
CPU: 1 PID: 492 Comm: allocate Not tainted 4.12.0-rc1-mm1+ #181
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Ubuntu-1.8.2-1ubuntu1 04/01/2014
Call Trace:
oom_kill_process+0x219/0x3e0
out_of_memory+0x11d/0x480
__alloc_pages_slowpath+0xc84/0xd40
__alloc_pages_nodemask+0x245/0x260
alloc_pages_vma+0xa2/0x270
__handle_mm_fault+0xca9/0x10c0
handle_mm_fault+0xf3/0x210
__do_page_fault+0x240/0x4e0
trace_do_page_fault+0x37/0xe0
do_async_page_fault+0x19/0x70
async_page_fault+0x28/0x30
...
Out of memory: Kill process 492 (allocate) score 899 or sacrifice child
Killed process 492 (allocate) total-vm:2052368kB, anon-rss:1894576kB, file-rss:4kB, shmem-rss:0kB
allocate: page allocation failure: order:0, mode:0x14280ca(GFP_HIGHUSER_MOVABLE|__GFP_ZERO), nodemask=(null)
allocate cpuset=/ mems_allowed=0
CPU: 1 PID: 492 Comm: allocate Not tainted 4.12.0-rc1-mm1+ #181
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Ubuntu-1.8.2-1ubuntu1 04/01/2014
Call Trace:
__alloc_pages_slowpath+0xd32/0xd40
__alloc_pages_nodemask+0x245/0x260
alloc_pages_vma+0xa2/0x270
__handle_mm_fault+0xca9/0x10c0
handle_mm_fault+0xf3/0x210
__do_page_fault+0x240/0x4e0
trace_do_page_fault+0x37/0xe0
do_async_page_fault+0x19/0x70
async_page_fault+0x28/0x30
...
oom_reaper: reaped process 492 (allocate), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
...
allocate invoked oom-killer: gfp_mask=0x0(), nodemask=(null), order=0, oom_score_adj=0
allocate cpuset=/ mems_allowed=0
CPU: 1 PID: 492 Comm: allocate Not tainted 4.12.0-rc1-mm1+ #181
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Ubuntu-1.8.2-1ubuntu1 04/01/2014
Call Trace:
oom_kill_process+0x219/0x3e0
out_of_memory+0x11d/0x480
pagefault_out_of_memory+0x68/0x80
mm_fault_error+0x8f/0x190
? handle_mm_fault+0xf3/0x210
__do_page_fault+0x4b2/0x4e0
trace_do_page_fault+0x37/0xe0
do_async_page_fault+0x19/0x70
async_page_fault+0x28/0x30
...
Out of memory: Kill process 233 (firewalld) score 10 or sacrifice child
Killed process 233 (firewalld) total-vm:246076kB, anon-rss:20956kB, file-rss:0kB, shmem-rss:0kB
There is a race window that the OOM reaper completes reclaiming the
first victim's memory while nothing but mutex_trylock() prevents the
first victim from calling out_of_memory() from pagefault_out_of_memory()
after memory allocation for page fault path failed due to being selected
as an OOM victim.
This is a side effect of commit 9a67f6488eca926f ("mm: consolidate
GFP_NOFAIL checks in the allocator slowpath") because that commit
silently changed the behavior from
/* Avoid allocations with no watermarks from looping endlessly */
to
/*
* Give up allocations without trying memory reserves if selected
* as an OOM victim
*/
in __alloc_pages_slowpath() by moving the location to check TIF_MEMDIE
flag. I have noticed this change but I didn't post a patch because I
thought it is an acceptable change other than noise by warn_alloc()
because !__GFP_NOFAIL allocations are allowed to fail. But we
overlooked that failing memory allocation from page fault path makes
difference due to the race window explained above.
While it might be possible to add a check to pagefault_out_of_memory()
that prevents the first victim from calling out_of_memory() or remove
out_of_memory() from pagefault_out_of_memory(), changing
pagefault_out_of_memory() does not suppress noise by warn_alloc() when
allocating thread was selected as an OOM victim. There is little point
with printing similar backtraces and memory information from both
out_of_memory() and warn_alloc().
Instead, if we guarantee that current thread can try allocations with no
watermarks once when current thread looping inside
__alloc_pages_slowpath() was selected as an OOM victim, we can follow "who
can use memory reserves" rules and suppress noise by warn_alloc() and
prevent memory allocations from page fault path from calling
pagefault_out_of_memory().
If we take the comment literally, this patch would do
- if (test_thread_flag(TIF_MEMDIE))
- goto nopage;
+ if (alloc_flags == ALLOC_NO_WATERMARKS || (gfp_mask & __GFP_NOMEMALLOC))
+ goto nopage;
because gfp_pfmemalloc_allowed() returns false if __GFP_NOMEMALLOC is
given. But if I recall correctly (I couldn't find the message), the
condition is meant to apply to only OOM victims despite the comment.
Therefore, this patch preserves TIF_MEMDIE check.
Fixes: 9a67f6488eca926f ("mm: consolidate GFP_NOFAIL checks in the allocator slowpath")
Link: http://lkml.kernel.org/r/201705192112.IAF69238.OQOHSJLFOFFMtV@I-love.SAKURA.ne.jp
Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Reported-by: Roman Gushchin <guro@fb.com>
Tested-by: Roman Gushchin <guro@fb.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: <stable@vger.kernel.org> [4.11]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-02 14:46:31 -07:00
|
|
|
|
2017-02-22 15:46:19 -08:00
|
|
|
|
|
|
|
|
|
2015-06-24 16:57:21 -07:00
|
|
|
|
2016-05-20 16:57:00 -07:00
|
|
|
|
|
|
|
|
|
2015-06-24 16:57:21 -07:00
|
|
|
|
2016-05-20 16:57:00 -07:00
|
|
|
|
2015-06-24 16:57:21 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
mm/page_alloc: fix race condition between build_all_zonelists and page allocation
Patrick Daly reported the following problem;
NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK] - before offline operation
[0] - ZONE_MOVABLE
[1] - ZONE_NORMAL
[2] - NULL
For a GFP_KERNEL allocation, alloc_pages_slowpath() will save the
offset of ZONE_NORMAL in ac->preferred_zoneref. If a concurrent
memory_offline operation removes the last page from ZONE_MOVABLE,
build_all_zonelists() & build_zonerefs_node() will update
node_zonelists as shown below. Only populated zones are added.
NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK] - after offline operation
[0] - ZONE_NORMAL
[1] - NULL
[2] - NULL
The race is simple -- page allocation could be in progress when a memory
hot-remove operation triggers a zonelist rebuild that removes zones. The
allocation request will still have a valid ac->preferred_zoneref that is
now pointing to NULL and triggers an OOM kill.
This problem probably always existed but may be slightly easier to trigger
due to 6aa303defb74 ("mm, vmscan: only allocate and reclaim from zones
with pages managed by the buddy allocator") which distinguishes between
zones that are completely unpopulated versus zones that have valid pages
not managed by the buddy allocator (e.g. reserved, memblock, ballooning
etc). Memory hotplug had multiple stages with timing considerations
around managed/present page updates, the zonelist rebuild and the zone
span updates. As David Hildenbrand puts it
memory offlining adjusts managed+present pages of the zone
essentially in one go. If after the adjustments, the zone is no
longer populated (present==0), we rebuild the zone lists.
Once that's done, we try shrinking the zone (start+spanned
pages) -- which results in zone_start_pfn == 0 if there are no
more pages. That happens *after* rebuilding the zonelists via
remove_pfn_range_from_zone().
The only requirement to fix the race is that a page allocation request
identifies when a zonelist rebuild has happened since the allocation
request started and no page has yet been allocated. Use a seqlock_t to
track zonelist updates with a lockless read-side of the zonelist and
protecting the rebuild and update of the counter with a spinlock.
[akpm@linux-foundation.org: make zonelist_update_seq static]
Link: https://lkml.kernel.org/r/20220824110900.vh674ltxmzb3proq@techsingularity.net
Fixes: 6aa303defb74 ("mm, vmscan: only allocate and reclaim from zones with pages managed by the buddy allocator")
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Reported-by: Patrick Daly <quic_pdaly@quicinc.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: David Hildenbrand <david@redhat.com>
Cc: <stable@vger.kernel.org> [4.9+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-08-24 12:14:50 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-01-24 15:18:38 -08:00
|
|
|
|
2017-02-22 15:46:19 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-08-31 08:28:23 +12:00
|
|
|
|
2017-02-22 15:46:19 -08:00
|
|
|
|
2024-08-31 08:28:23 +12:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-02-22 15:46:19 -08:00
|
|
|
|
2024-08-31 08:28:23 +12:00
|
|
|
|
2017-02-22 15:46:19 -08:00
|
|
|
|
|
|
|
|
|
2017-02-22 15:46:25 -08:00
|
|
|
|
2023-01-13 11:12:16 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-02-22 15:46:25 -08:00
|
|
|
|
2023-01-13 11:12:16 +00:00
|
|
|
|
2017-02-22 15:46:25 -08:00
|
|
|
|
2023-01-13 11:12:16 +00:00
|
|
|
|
2017-02-22 15:46:25 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-02-22 15:46:19 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-02-22 15:46:10 -08:00
|
|
|
|
2016-10-07 17:01:55 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
mm: sl[au]b: add knowledge of PFMEMALLOC reserve pages
When a user or administrator requires swap for their application, they
create a swap partition and file, format it with mkswap and activate it
with swapon. Swap over the network is considered as an option in diskless
systems. The two likely scenarios are when blade servers are used as part
of a cluster where the form factor or maintenance costs do not allow the
use of disks and thin clients.
The Linux Terminal Server Project recommends the use of the Network Block
Device (NBD) for swap according to the manual at
https://sourceforge.net/projects/ltsp/files/Docs-Admin-Guide/LTSPManual.pdf/download
There is also documentation and tutorials on how to setup swap over NBD at
places like https://help.ubuntu.com/community/UbuntuLTSP/EnableNBDSWAP The
nbd-client also documents the use of NBD as swap. Despite this, the fact
is that a machine using NBD for swap can deadlock within minutes if swap
is used intensively. This patch series addresses the problem.
The core issue is that network block devices do not use mempools like
normal block devices do. As the host cannot control where they receive
packets from, they cannot reliably work out in advance how much memory
they might need. Some years ago, Peter Zijlstra developed a series of
patches that supported swap over an NFS that at least one distribution is
carrying within their kernels. This patch series borrows very heavily
from Peter's work to support swapping over NBD as a pre-requisite to
supporting swap-over-NFS. The bulk of the complexity is concerned with
preserving memory that is allocated from the PFMEMALLOC reserves for use
by the network layer which is needed for both NBD and NFS.
Patch 1 adds knowledge of the PFMEMALLOC reserves to SLAB and SLUB to
preserve access to pages allocated under low memory situations
to callers that are freeing memory.
Patch 2 optimises the SLUB fast path to avoid pfmemalloc checks
Patch 3 introduces __GFP_MEMALLOC to allow access to the PFMEMALLOC
reserves without setting PFMEMALLOC.
Patch 4 opens the possibility for softirqs to use PFMEMALLOC reserves
for later use by network packet processing.
Patch 5 only sets page->pfmemalloc when ALLOC_NO_WATERMARKS was required
Patch 6 ignores memory policies when ALLOC_NO_WATERMARKS is set.
Patches 7-12 allows network processing to use PFMEMALLOC reserves when
the socket has been marked as being used by the VM to clean pages. If
packets are received and stored in pages that were allocated under
low-memory situations and are unrelated to the VM, the packets
are dropped.
Patch 11 reintroduces __skb_alloc_page which the networking
folk may object to but is needed in some cases to propogate
pfmemalloc from a newly allocated page to an skb. If there is a
strong objection, this patch can be dropped with the impact being
that swap-over-network will be slower in some cases but it should
not fail.
Patch 13 is a micro-optimisation to avoid a function call in the
common case.
Patch 14 tags NBD sockets as being SOCK_MEMALLOC so they can use
PFMEMALLOC if necessary.
Patch 15 notes that it is still possible for the PFMEMALLOC reserve
to be depleted. To prevent this, direct reclaimers get throttled on
a waitqueue if 50% of the PFMEMALLOC reserves are depleted. It is
expected that kswapd and the direct reclaimers already running
will clean enough pages for the low watermark to be reached and
the throttled processes are woken up.
Patch 16 adds a statistic to track how often processes get throttled
Some basic performance testing was run using kernel builds, netperf on
loopback for UDP and TCP, hackbench (pipes and sockets), iozone and
sysbench. Each of them were expected to use the sl*b allocators
reasonably heavily but there did not appear to be significant performance
variances.
For testing swap-over-NBD, a machine was booted with 2G of RAM with a
swapfile backed by NBD. 8*NUM_CPU processes were started that create
anonymous memory mappings and read them linearly in a loop. The total
size of the mappings were 4*PHYSICAL_MEMORY to use swap heavily under
memory pressure.
Without the patches and using SLUB, the machine locks up within minutes
and runs to completion with them applied. With SLAB, the story is
different as an unpatched kernel run to completion. However, the patched
kernel completed the test 45% faster.
MICRO
3.5.0-rc2 3.5.0-rc2
vanilla swapnbd
Unrecognised test vmscan-anon-mmap-write
MMTests Statistics: duration
Sys Time Running Test (seconds) 197.80 173.07
User+Sys Time Running Test (seconds) 206.96 182.03
Total Elapsed Time (seconds) 3240.70 1762.09
This patch: mm: sl[au]b: add knowledge of PFMEMALLOC reserve pages
Allocations of pages below the min watermark run a risk of the machine
hanging due to a lack of memory. To prevent this, only callers who have
PF_MEMALLOC or TIF_MEMDIE set and are not processing an interrupt are
allowed to allocate with ALLOC_NO_WATERMARKS. Once they are allocated to
a slab though, nothing prevents other callers consuming free objects
within those slabs. This patch limits access to slab pages that were
alloced from the PFMEMALLOC reserves.
When this patch is applied, pages allocated from below the low watermark
are returned with page->pfmemalloc set and it is up to the caller to
determine how the page should be protected. SLAB restricts access to any
page with page->pfmemalloc set to callers which are known to able to
access the PFMEMALLOC reserve. If one is not available, an attempt is
made to allocate a new page rather than use a reserve. SLUB is a bit more
relaxed in that it only records if the current per-CPU page was allocated
from PFMEMALLOC reserve and uses another partial slab if the caller does
not have the necessary GFP or process flags. This was found to be
sufficient in tests to avoid hangs due to SLUB generally maintaining
smaller lists than SLAB.
In low-memory conditions it does mean that !PFMEMALLOC allocators can fail
a slab allocation even though free objects are available because they are
being preserved for callers that are freeing pages.
[a.p.zijlstra@chello.nl: Original implementation]
[sebastian@breakpoint.cc: Correct order of page flag clearing]
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: David Miller <davem@davemloft.net>
Cc: Neil Brown <neilb@suse.de>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Christie <michaelc@cs.wisc.edu>
Cc: Eric B Munson <emunson@mgebm.net>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Sebastian Andrzej Siewior <sebastian@breakpoint.cc>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Christoph Lameter <cl@linux.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-07-31 16:43:58 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2009-06-16 15:31:57 -07:00
|
|
|
|
2017-02-24 14:56:29 -08:00
|
|
|
|
2017-07-06 15:40:03 -07:00
|
|
|
|
2021-04-29 23:01:10 -07:00
|
|
|
|
2017-02-24 14:56:29 -08:00
|
|
|
|
|
|
|
|
|
2020-06-03 15:59:01 -07:00
|
|
|
|
2017-07-06 15:40:03 -07:00
|
|
|
|
2017-02-24 14:56:29 -08:00
|
|
|
|
2020-06-03 15:59:08 -07:00
|
|
|
|
2017-02-24 14:56:29 -08:00
|
|
|
|
|
|
|
|
|
2021-04-29 23:01:10 -07:00
|
|
|
|
2020-08-06 23:26:01 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-09-02 14:58:13 -07:00
|
|
|
|
2017-02-24 14:56:29 -08:00
|
|
|
|
2017-02-24 14:56:53 -08:00
|
|
|
|
|
|
|
|
|
2017-02-24 14:56:29 -08:00
|
|
|
|
|
|
|
|
|
2022-06-05 17:25:37 +02:00
|
|
|
|
2017-02-24 14:56:29 -08:00
|
|
|
|
2025-02-21 18:44:23 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-02-24 14:56:29 -08:00
|
|
|
|
|
|
|
|
|
2021-05-04 18:39:00 -07:00
|
|
|
|
2018-05-23 10:18:21 +09:00
|
|
|
|
2017-02-24 14:56:29 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-06-03 15:59:01 -07:00
|
|
|
|
2020-10-13 16:55:51 -07:00
|
|
|
|
|
|
|
|
|
2017-02-24 14:56:29 -08:00
|
|
|
|
|
|
|
|
|
2021-04-29 23:01:45 -07:00
|
|
|
|
2024-12-23 17:00:37 -05:00
|
|
|
|
2021-04-29 23:01:45 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-12-23 17:00:37 -05:00
|
|
|
|
|
|
|
|
|
2021-04-29 23:01:45 -07:00
|
|
|
|
|
|
|
|
|
2024-12-23 17:00:37 -05:00
|
|
|
|
2021-04-29 23:01:45 -07:00
|
|
|
|
2024-12-23 17:00:37 -05:00
|
|
|
|
2021-04-29 23:01:48 -07:00
|
|
|
|
|
|
|
|
|
2024-12-23 17:00:37 -05:00
|
|
|
|
2021-04-29 23:01:45 -07:00
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
2021-04-29 23:01:45 -07:00
|
|
|
|
2021-04-29 23:01:48 -07:00
|
|
|
|
2021-04-29 23:01:45 -07:00
|
|
|
|
|
|
|
|
|
2022-06-24 13:54:21 +01:00
|
|
|
|
2021-04-29 23:01:45 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-28 19:41:50 -07:00
|
|
|
|
2021-04-29 23:01:45 -07:00
|
|
|
|
2021-04-29 23:01:48 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-12-23 17:00:37 -05:00
|
|
|
|
2021-04-29 23:01:48 -07:00
|
|
|
|
|
|
|
|
|
2021-07-14 21:26:52 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-24 18:40:07 -07:00
|
|
|
|
2024-12-23 17:00:37 -05:00
|
|
|
|
2021-07-14 21:26:52 -07:00
|
|
|
|
2021-06-24 18:40:07 -07:00
|
|
|
|
2021-10-28 14:36:04 -07:00
|
|
|
|
2023-02-13 11:29:22 -08:00
|
|
|
|
2021-10-28 14:36:04 -07:00
|
|
|
|
|
|
|
|
|
2021-04-29 23:01:45 -07:00
|
|
|
|
2021-04-29 23:01:48 -07:00
|
|
|
|
2021-04-29 23:01:45 -07:00
|
|
|
|
|
|
|
|
|
2021-07-14 21:26:46 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-04-29 23:01:45 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-07-14 21:26:52 -07:00
|
|
|
|
2021-04-29 23:01:45 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: fix NULL pointer dereference in alloc_pages_bulk_noprof
We triggered a NULL pointer dereference for ac.preferred_zoneref->zone in
alloc_pages_bulk_noprof() when the task is migrated between cpusets.
When cpuset is enabled, in prepare_alloc_pages(), ac->nodemask may be
¤t->mems_allowed. when first_zones_zonelist() is called to find
preferred_zoneref, the ac->nodemask may be modified concurrently if the
task is migrated between different cpusets. Assuming we have 2 NUMA Node,
when traversing Node1 in ac->zonelist, the nodemask is 2, and when
traversing Node2 in ac->zonelist, the nodemask is 1. As a result, the
ac->preferred_zoneref points to NULL zone.
In alloc_pages_bulk_noprof(), for_each_zone_zonelist_nodemask() finds a
allowable zone and calls zonelist_node_idx(ac.preferred_zoneref), leading
to NULL pointer dereference.
__alloc_pages_noprof() fixes this issue by checking NULL pointer in commit
ea57485af8f4 ("mm, page_alloc: fix check for NULL preferred_zone") and
commit df76cee6bbeb ("mm, page_alloc: remove redundant checks from alloc
fastpath").
To fix it, check NULL pointer for preferred_zoneref->zone.
Link: https://lkml.kernel.org/r/20241113083235.166798-1-tujinjiang@huawei.com
Fixes: 387ba26fb1cb ("mm/page_alloc: add a bulk page allocator")
Signed-off-by: Jinjiang Tu <tujinjiang@huawei.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Alexander Lobakin <alobakin@pm.me>
Cc: David Hildenbrand <david@redhat.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Nanyong Sun <sunnanyong@huawei.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-11-13 16:32:35 +08:00
|
|
|
|
|
|
|
|
|
2021-04-29 23:01:45 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-07-29 14:47:17 +05:30
|
|
|
|
|
|
|
|
|
2021-04-29 23:01:45 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-05-06 14:25:08 +03:00
|
|
|
|
2024-08-09 14:48:49 +03:00
|
|
|
|
2021-04-29 23:01:45 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-08-09 14:48:49 +03:00
|
|
|
|
2025-05-06 14:25:08 +03:00
|
|
|
|
2024-08-09 14:48:49 +03:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-04-29 23:01:45 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-04-29 23:01:51 -07:00
|
|
|
|
2021-04-29 23:01:45 -07:00
|
|
|
|
|
|
|
|
|
2022-11-18 10:17:14 +00:00
|
|
|
|
2022-06-24 13:54:21 +01:00
|
|
|
|
2022-11-18 10:17:14 +00:00
|
|
|
|
2022-06-24 13:54:23 +01:00
|
|
|
|
2022-06-24 13:54:21 +01:00
|
|
|
|
2021-04-29 23:01:45 -07:00
|
|
|
|
2022-06-24 13:54:21 +01:00
|
|
|
|
|
|
|
|
|
2021-04-29 23:01:48 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-12-23 17:00:37 -05:00
|
|
|
|
2021-04-29 23:01:48 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-28 19:43:08 -07:00
|
|
|
|
2021-04-29 23:01:45 -07:00
|
|
|
|
2021-04-29 23:01:51 -07:00
|
|
|
|
2022-05-26 10:12:10 +01:00
|
|
|
|
2022-06-24 13:54:21 +01:00
|
|
|
|
2022-11-18 10:17:14 +00:00
|
|
|
|
2021-04-29 23:01:45 -07:00
|
|
|
|
2022-06-24 13:54:21 +01:00
|
|
|
|
2021-04-29 23:01:45 -07:00
|
|
|
|
|
|
|
|
|
2021-06-28 19:41:50 -07:00
|
|
|
|
2021-04-29 23:01:45 -07:00
|
|
|
|
|
|
|
|
|
2024-11-25 21:01:37 +00:00
|
|
|
|
2024-12-23 17:00:37 -05:00
|
|
|
|
2021-04-29 23:01:45 -07:00
|
|
|
|
|
|
|
|
|
2022-11-18 10:17:14 +00:00
|
|
|
|
2022-06-24 13:54:21 +01:00
|
|
|
|
2021-06-28 19:41:54 -07:00
|
|
|
|
2021-06-28 19:41:50 -07:00
|
|
|
|
2024-07-29 14:47:17 +05:30
|
|
|
|
2021-04-29 23:01:45 -07:00
|
|
|
|
2021-07-14 21:26:52 -07:00
|
|
|
|
2021-04-29 23:01:48 -07:00
|
|
|
|
2021-04-29 23:01:45 -07:00
|
|
|
|
|
|
|
|
|
2022-06-24 13:54:21 +01:00
|
|
|
|
2021-04-29 23:01:45 -07:00
|
|
|
|
|
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
2024-12-23 17:00:37 -05:00
|
|
|
|
|
|
|
|
|
2021-07-14 21:26:52 -07:00
|
|
|
|
2021-04-29 23:01:45 -07:00
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
2021-04-29 23:01:45 -07:00
|
|
|
|
2009-06-16 15:31:57 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-11-25 21:01:45 +00:00
|
|
|
|
|
|
|
|
|
2009-06-16 15:31:57 -07:00
|
|
|
|
2016-05-19 17:13:50 -07:00
|
|
|
|
2016-07-28 15:46:50 -07:00
|
|
|
|
2021-04-29 23:01:10 -07:00
|
|
|
|
2017-02-24 14:56:29 -08:00
|
|
|
|
2016-05-19 17:13:30 -07:00
|
|
|
|
2018-11-16 15:08:53 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-12-28 17:47:04 +03:00
|
|
|
|
2018-11-16 15:08:53 -08:00
|
|
|
|
|
|
|
|
|
2021-04-29 23:01:13 -07:00
|
|
|
|
2021-05-04 18:38:57 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-05-04 18:39:00 -07:00
|
|
|
|
|
|
|
|
|
2021-05-04 18:38:57 -07:00
|
|
|
|
|
|
|
|
|
2021-04-29 23:01:13 -07:00
|
|
|
|
|
|
|
|
|
2021-04-29 23:01:10 -07:00
|
|
|
|
2009-06-16 15:31:57 -07:00
|
|
|
|
|
|
|
|
|
mm, page_alloc: spread allocations across zones before introducing fragmentation
Patch series "Fragmentation avoidance improvements", v5.
It has been noted before that fragmentation avoidance (aka
anti-fragmentation) is not perfect. Given sufficient time or an adverse
workload, memory gets fragmented and the long-term success of high-order
allocations degrades. This series defines an adverse workload, a definition
of external fragmentation events (including serious) ones and a series
that reduces the level of those fragmentation events.
The details of the workload and the consequences are described in more
detail in the changelogs. However, from patch 1, this is a high-level
summary of the adverse workload. The exact details are found in the
mmtests implementation.
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch)
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameterr create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed
3. Warm up a number of fio read-only threads accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll fault back in old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup
Overall the series reduces external fragmentation causing events by over 94%
on 1 and 2 socket machines, which in turn impacts high-order allocation
success rates over the long term. There are differences in latencies and
high-order allocation success rates. Latencies are a mixed bag as they
are vulnerable to exact system state and whether allocations succeeded
so they are treated as a secondary metric.
Patch 1 uses lower zones if they are populated and have free memory
instead of fragmenting a higher zone. It's special cased to
handle a Normal->DMA32 fallback with the reasons explained
in the changelog.
Patch 2-4 boosts watermarks temporarily when an external fragmentation
event occurs. kswapd wakes to reclaim a small amount of old memory
and then wakes kcompactd on completion to recover the system
slightly. This introduces some overhead in the slowpath. The level
of boosting can be tuned or disabled depending on the tolerance
for fragmentation vs allocation latency.
Patch 5 stalls some movable allocation requests to let kswapd from patch 4
make some progress. The duration of the stalls is very low but it
is possible to tune the system to avoid fragmentation events if
larger stalls can be tolerated.
The bulk of the improvement in fragmentation avoidance is from patches
1-4 but patch 5 can deal with a rare corner case and provides the option
of tuning a system for THP allocation success rates in exchange for
some stalls to control fragmentation.
This patch (of 5):
The page allocator zone lists are iterated based on the watermarks of each
zone which does not take anti-fragmentation into account. On x86, node 0
may have multiple zones while other nodes have one zone. A consequence is
that tasks running on node 0 may fragment ZONE_NORMAL even though
ZONE_DMA32 has plenty of free memory. This patch special cases the
allocator fast path such that it'll try an allocation from a lower local
zone before fragmenting a higher zone. In this case, stealing of
pageblocks or orders larger than a pageblock are still allowed in the fast
path as they are uninteresting from a fragmentation point of view.
This was evaluated using a benchmark designed to fragment memory before
attempting THP allocations. It's implemented in mmtests as the following
configurations
configs/config-global-dhp__workload_thpfioscale
configs/config-global-dhp__workload_thpfioscale-defrag
configs/config-global-dhp__workload_thpfioscale-madvhugepage
e.g. from mmtests
./run-mmtests.sh --run-monitor --config configs/config-global-dhp__workload_thpfioscale test-run-1
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch).
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameter create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed.
3. Warm up a number of fio read-only processes accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll refault old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds.
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup the test files.
Note that due to the use of IO and page cache that this benchmark is not
suitable for running on large machines where the time to fragment memory
may be excessive. Also note that while this is one mix that generates
fragmentation that it's not the only mix that generates fragmentation.
Differences in workload that are more slab-intensive or whether SLUB is
used with high-order pages may yield different results.
When the page allocator fragments memory, it records the event using the
mm_page_alloc_extfrag ftrace event. If the fallback_order is smaller than
a pageblock order (order-9 on 64-bit x86) then it's considered to be an
"external fragmentation event" that may cause issues in the future.
Hence, the primary metric here is the number of external fragmentation
events that occur with order < 9. The secondary metric is allocation
latency and huge page allocation success rates but note that differences
in latencies and what the success rate also can affect the number of
external fragmentation event which is why it's a secondary metric.
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 662.92 ( 0.00%) 653.58 * 1.41%*
Amean fault-huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
Fault latencies are slightly reduced while allocation success rates remain
at zero as this configuration does not make any special effort to allocate
THP and fio is heavily active at the time and either filling memory or
keeping pages resident. However, a 49% reduction of serious fragmentation
events reduces the changes of external fragmentation being a problem in
the future.
Vlastimil asked during review for a breakdown of the allocation types
that are falling back.
vanilla
3816 MIGRATE_UNMOVABLE
800845 MIGRATE_MOVABLE
33 MIGRATE_UNRECLAIMABLE
patch
735 MIGRATE_UNMOVABLE
408135 MIGRATE_MOVABLE
42 MIGRATE_UNRECLAIMABLE
The majority of the fallbacks are due to movable allocations and this is
consistent for the workload throughout the series so will not be presented
again as the primary source of fallbacks are movable allocations.
Movable fallbacks are sometimes considered "ok" to fallback because they
can be migrated. The problem is that they can fill an
unmovable/reclaimable pageblock causing those allocations to fallback
later and polluting pageblocks with pages that cannot move. If there is a
movable fallback, it is pretty much guaranteed to affect an
unmovable/reclaimable pageblock and while it might not be enough to
actually cause a unmovable/reclaimable fallback in the future, we cannot
know that in advance so the patch takes the only option available to it.
Hence, it's important to control them. This point is also consistent
throughout the series and will not be repeated.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 1495.14 ( 0.00%) 1467.55 ( 1.85%)
Amean fault-huge-1 1098.48 ( 0.00%) 1127.11 ( -2.61%)
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 78.57 ( 0.00%) 77.64 ( -1.18%)
Fragmentation events were reduced quite a bit although this is known
to be a little variable. The latencies and allocation success rates
are similar but they were already quite high.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 1350.05 ( 0.00%) 1346.45 ( 0.27%)
Amean fault-huge-5 4181.01 ( 0.00%) 3418.60 ( 18.24%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 1.15 ( 0.00%) 0.78 ( -31.88%)
The reduction of external fragmentation events is slight and this is
partially due to the removal of __GFP_THISNODE in commit ac5b2c18911f
("mm: thp: relax __GFP_THISNODE for MADV_HUGEPAGE mappings") as THP
allocations can now spill over to remote nodes instead of fragmenting
local memory.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 6138.97 ( 0.00%) 6217.43 ( -1.28%)
Amean fault-huge-5 2294.28 ( 0.00%) 3163.33 * -37.88%*
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 96.82 ( 0.00%) 95.14 ( -1.74%)
There was a slight reduction in external fragmentation events although the
latencies were higher. The allocation success rate is high enough that
the system is struggling and there is quite a lot of parallel reclaim and
compaction activity. There is also a certain degree of luck on whether
processes start on node 0 or not for this patch but the relevance is
reduced later in the series.
Overall, the patch reduces the number of external fragmentation causing
events so the success of THP over long periods of time would be improved
for this adverse workload.
Link: http://lkml.kernel.org/r/20181123114528.28802-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:41 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-07-29 14:47:17 +05:30
|
|
|
|
mm, page_alloc: spread allocations across zones before introducing fragmentation
Patch series "Fragmentation avoidance improvements", v5.
It has been noted before that fragmentation avoidance (aka
anti-fragmentation) is not perfect. Given sufficient time or an adverse
workload, memory gets fragmented and the long-term success of high-order
allocations degrades. This series defines an adverse workload, a definition
of external fragmentation events (including serious) ones and a series
that reduces the level of those fragmentation events.
The details of the workload and the consequences are described in more
detail in the changelogs. However, from patch 1, this is a high-level
summary of the adverse workload. The exact details are found in the
mmtests implementation.
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch)
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameterr create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed
3. Warm up a number of fio read-only threads accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll fault back in old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup
Overall the series reduces external fragmentation causing events by over 94%
on 1 and 2 socket machines, which in turn impacts high-order allocation
success rates over the long term. There are differences in latencies and
high-order allocation success rates. Latencies are a mixed bag as they
are vulnerable to exact system state and whether allocations succeeded
so they are treated as a secondary metric.
Patch 1 uses lower zones if they are populated and have free memory
instead of fragmenting a higher zone. It's special cased to
handle a Normal->DMA32 fallback with the reasons explained
in the changelog.
Patch 2-4 boosts watermarks temporarily when an external fragmentation
event occurs. kswapd wakes to reclaim a small amount of old memory
and then wakes kcompactd on completion to recover the system
slightly. This introduces some overhead in the slowpath. The level
of boosting can be tuned or disabled depending on the tolerance
for fragmentation vs allocation latency.
Patch 5 stalls some movable allocation requests to let kswapd from patch 4
make some progress. The duration of the stalls is very low but it
is possible to tune the system to avoid fragmentation events if
larger stalls can be tolerated.
The bulk of the improvement in fragmentation avoidance is from patches
1-4 but patch 5 can deal with a rare corner case and provides the option
of tuning a system for THP allocation success rates in exchange for
some stalls to control fragmentation.
This patch (of 5):
The page allocator zone lists are iterated based on the watermarks of each
zone which does not take anti-fragmentation into account. On x86, node 0
may have multiple zones while other nodes have one zone. A consequence is
that tasks running on node 0 may fragment ZONE_NORMAL even though
ZONE_DMA32 has plenty of free memory. This patch special cases the
allocator fast path such that it'll try an allocation from a lower local
zone before fragmenting a higher zone. In this case, stealing of
pageblocks or orders larger than a pageblock are still allowed in the fast
path as they are uninteresting from a fragmentation point of view.
This was evaluated using a benchmark designed to fragment memory before
attempting THP allocations. It's implemented in mmtests as the following
configurations
configs/config-global-dhp__workload_thpfioscale
configs/config-global-dhp__workload_thpfioscale-defrag
configs/config-global-dhp__workload_thpfioscale-madvhugepage
e.g. from mmtests
./run-mmtests.sh --run-monitor --config configs/config-global-dhp__workload_thpfioscale test-run-1
The broad details of the workload are as follows;
1. Create an XFS filesystem (not specified in the configuration but done
as part of the testing for this patch).
2. Start 4 fio threads that write a number of 64K files inefficiently.
Inefficiently means that files are created on first access and not
created in advance (fio parameter create_on_open=1) and fallocate
is not used (fallocate=none). With multiple IO issuers this creates
a mix of slab and page cache allocations over time. The total size
of the files is 150% physical memory so that the slabs and page cache
pages get mixed.
3. Warm up a number of fio read-only processes accessing the same files
created in step 2. This part runs for the same length of time it
took to create the files. It'll refault old data and further
interleave slab and page cache allocations. As it's now low on
memory due to step 2, fragmentation occurs as pageblocks get
stolen.
4. While step 3 is still running, start a process that tries to allocate
75% of memory as huge pages with a number of threads. The number of
threads is based on a (NR_CPUS_SOCKET - NR_FIO_THREADS)/4 to avoid THP
threads contending with fio, any other threads or forcing cross-NUMA
scheduling. Note that the test has not been used on a machine with less
than 8 cores. The benchmark records whether huge pages were allocated
and what the fault latency was in microseconds.
5. Measure the number of events potentially causing external fragmentation,
the fault latency and the huge page allocation success rate.
6. Cleanup the test files.
Note that due to the use of IO and page cache that this benchmark is not
suitable for running on large machines where the time to fragment memory
may be excessive. Also note that while this is one mix that generates
fragmentation that it's not the only mix that generates fragmentation.
Differences in workload that are more slab-intensive or whether SLUB is
used with high-order pages may yield different results.
When the page allocator fragments memory, it records the event using the
mm_page_alloc_extfrag ftrace event. If the fallback_order is smaller than
a pageblock order (order-9 on 64-bit x86) then it's considered to be an
"external fragmentation event" that may cause issues in the future.
Hence, the primary metric here is the number of external fragmentation
events that occur with order < 9. The secondary metric is allocation
latency and huge page allocation success rates but note that differences
in latencies and what the success rate also can affect the number of
external fragmentation event which is why it's a secondary metric.
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 662.92 ( 0.00%) 653.58 * 1.41%*
Amean fault-huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 0.00 ( 0.00%) 0.00 ( 0.00%)
Fault latencies are slightly reduced while allocation success rates remain
at zero as this configuration does not make any special effort to allocate
THP and fio is heavily active at the time and either filling memory or
keeping pages resident. However, a 49% reduction of serious fragmentation
events reduces the changes of external fragmentation being a problem in
the future.
Vlastimil asked during review for a breakdown of the allocation types
that are falling back.
vanilla
3816 MIGRATE_UNMOVABLE
800845 MIGRATE_MOVABLE
33 MIGRATE_UNRECLAIMABLE
patch
735 MIGRATE_UNMOVABLE
408135 MIGRATE_MOVABLE
42 MIGRATE_UNRECLAIMABLE
The majority of the fallbacks are due to movable allocations and this is
consistent for the workload throughout the series so will not be presented
again as the primary source of fallbacks are movable allocations.
Movable fallbacks are sometimes considered "ok" to fallback because they
can be migrated. The problem is that they can fill an
unmovable/reclaimable pageblock causing those allocations to fallback
later and polluting pageblocks with pages that cannot move. If there is a
movable fallback, it is pretty much guaranteed to affect an
unmovable/reclaimable pageblock and while it might not be enough to
actually cause a unmovable/reclaimable fallback in the future, we cannot
know that in advance so the patch takes the only option available to it.
Hence, it's important to control them. This point is also consistent
throughout the series and will not be repeated.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-1 1495.14 ( 0.00%) 1467.55 ( 1.85%)
Amean fault-huge-1 1098.48 ( 0.00%) 1127.11 ( -2.61%)
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-1 78.57 ( 0.00%) 77.64 ( -1.18%)
Fragmentation events were reduced quite a bit although this is known
to be a little variable. The latencies and allocation success rates
are similar but they were already quite high.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 1350.05 ( 0.00%) 1346.45 ( 0.27%)
Amean fault-huge-5 4181.01 ( 0.00%) 3418.60 ( 18.24%)
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 1.15 ( 0.00%) 0.78 ( -31.88%)
The reduction of external fragmentation events is slight and this is
partially due to the removal of __GFP_THISNODE in commit ac5b2c18911f
("mm: thp: relax __GFP_THISNODE for MADV_HUGEPAGE mappings") as THP
allocations can now spill over to remote nodes instead of fragmenting
local memory.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Amean fault-base-5 6138.97 ( 0.00%) 6217.43 ( -1.28%)
Amean fault-huge-5 2294.28 ( 0.00%) 3163.33 * -37.88%*
thpfioscale Percentage Faults Huge
4.20.0-rc3 4.20.0-rc3
vanilla lowzone-v5r8
Percentage huge-5 96.82 ( 0.00%) 95.14 ( -1.74%)
There was a slight reduction in external fragmentation events although the
latencies were higher. The allocation success rate is high enough that
the system is struggling and there is quite a lot of parallel reclaim and
compaction activity. There is also a certain degree of luck on whether
processes start on node 0 or not for this patch but the relevance is
reduced later in the series.
Overall, the patch reduces the number of external fragmentation causing
events so the success of THP over long periods of time would be improved
for this adverse workload.
Link: http://lkml.kernel.org/r/20181123114528.28802-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 00:35:41 -08:00
|
|
|
|
2009-06-16 15:31:59 -07:00
|
|
|
|
2021-04-29 23:01:10 -07:00
|
|
|
|
2016-05-19 17:14:01 -07:00
|
|
|
|
|
|
|
|
|
2015-02-11 15:25:04 -08:00
|
|
|
|
2021-05-04 18:38:57 -07:00
|
|
|
|
2016-05-19 17:14:01 -07:00
|
|
|
|
2009-06-16 15:31:57 -07:00
|
|
|
|
2016-05-19 17:14:44 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-04-01 21:09:53 -07:00
|
|
|
|
mm, page_alloc: fix fast-path race with cpuset update or removal
Ganapatrao Kulkarni reported that the LTP test cpuset01 in stress mode
triggers OOM killer in few seconds, despite lots of free memory. The
test attempts to repeatedly fault in memory in one process in a cpuset,
while changing allowed nodes of the cpuset between 0 and 1 in another
process.
One possible cause is that in the fast path we find the preferred
zoneref according to current mems_allowed, so that it points to the
middle of the zonelist, skipping e.g. zones of node 1 completely. If
the mems_allowed is updated to contain only node 1, we never reach it in
the zonelist, and trigger OOM before checking the cpuset_mems_cookie.
This patch fixes the particular case by redoing the preferred zoneref
search if we switch back to the original nodemask. The condition is
also slightly changed so that when the last non-root cpuset is removed,
we don't miss it.
Note that this is not a full fix, and more patches will follow.
Link: http://lkml.kernel.org/r/20170120103843.24587-3-vbabka@suse.cz
Fixes: 682a3385e773 ("mm, page_alloc: inline the fast path of the zonelist iterator")
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Reported-by: Ganapatrao Kulkarni <gpkulkarni@gmail.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-01-24 15:18:35 -08:00
|
|
|
|
2021-04-29 23:01:10 -07:00
|
|
|
|
2015-02-11 15:25:07 -08:00
|
|
|
|
2016-05-19 17:14:01 -07:00
|
|
|
|
2023-02-13 11:29:22 -08:00
|
|
|
|
2021-04-29 23:01:13 -07:00
|
|
|
|
2024-11-25 21:01:44 +00:00
|
|
|
|
mm: memcontrol: only mark charged pages with PageKmemcg
To distinguish non-slab pages charged to kmemcg we mark them PageKmemcg,
which sets page->_mapcount to -512. Currently, we set/clear PageKmemcg
in __alloc_pages_nodemask()/free_pages_prepare() for any page allocated
with __GFP_ACCOUNT, including those that aren't actually charged to any
cgroup, i.e. allocated from the root cgroup context. To avoid overhead
in case cgroups are not used, we only do that if memcg_kmem_enabled() is
true. The latter is set iff there are kmem-enabled memory cgroups
(online or offline). The root cgroup is not considered kmem-enabled.
As a result, if a page is allocated with __GFP_ACCOUNT for the root
cgroup when there are kmem-enabled memory cgroups and is freed after all
kmem-enabled memory cgroups were removed, e.g.
# no memory cgroups has been created yet, create one
mkdir /sys/fs/cgroup/memory/test
# run something allocating pages with __GFP_ACCOUNT, e.g.
# a program using pipe
dmesg | tail
# remove the memory cgroup
rmdir /sys/fs/cgroup/memory/test
we'll get bad page state bug complaining about page->_mapcount != -1:
BUG: Bad page state in process swapper/0 pfn:1fd945c
page:ffffea007f651700 count:0 mapcount:-511 mapping: (null) index:0x0
flags: 0x1000000000000000()
To avoid that, let's mark with PageKmemcg only those pages that are
actually charged to and hence pin a non-root memory cgroup.
Fixes: 4949148ad433 ("mm: charge/uncharge kmemcg from generic page allocator paths")
Reported-and-tested-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: Vladimir Davydov <vdavydov@virtuozzo.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-08-08 23:03:12 +03:00
|
|
|
|
mm: charge/uncharge kmemcg from generic page allocator paths
Currently, to charge a non-slab allocation to kmemcg one has to use
alloc_kmem_pages helper with __GFP_ACCOUNT flag. A page allocated with
this helper should finally be freed using free_kmem_pages, otherwise it
won't be uncharged.
This API suits its current users fine, but it turns out to be impossible
to use along with page reference counting, i.e. when an allocation is
supposed to be freed with put_page, as it is the case with pipe or unix
socket buffers.
To overcome this limitation, this patch moves charging/uncharging to
generic page allocator paths, i.e. to __alloc_pages_nodemask and
free_pages_prepare, and zaps alloc/free_kmem_pages helpers. This way,
one can use any of the available page allocation functions to get the
allocated page charged to kmemcg - it's enough to pass __GFP_ACCOUNT,
just like in case of kmalloc and friends. A charged page will be
automatically uncharged on free.
To make it possible, we need to mark pages charged to kmemcg somehow.
To avoid introducing a new page flag, we make use of page->_mapcount for
marking such pages. Since pages charged to kmemcg are not supposed to
be mapped to userspace, it should work just fine. There are other
(ab)users of page->_mapcount - buddy and balloon pages - but we don't
conflict with them.
In case kmemcg is compiled out or not used at runtime, this patch
introduces no overhead to generic page allocator paths. If kmemcg is
used, it will be plus one gfp flags check on alloc and plus one
page->_mapcount check on free, which shouldn't hurt performance, because
the data accessed are hot.
Link: http://lkml.kernel.org/r/a9736d856f895bcb465d9f257b54efe32eda6f99.1464079538.git.vdavydov@virtuozzo.com
Signed-off-by: Vladimir Davydov <vdavydov@virtuozzo.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-07-26 15:24:24 -07:00
|
|
|
|
|
|
|
|
|
2021-04-29 23:01:10 -07:00
|
|
|
|
2022-09-15 17:03:48 +02:00
|
|
|
|
2016-05-19 17:14:01 -07:00
|
|
|
|
2009-06-16 15:31:57 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2024-11-25 21:01:45 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
2020-12-15 22:55:54 -05:00
|
|
|
|
|
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
2023-10-03 02:25:33 -07:00
|
|
|
|
|
|
|
|
|
2020-12-15 22:55:54 -05:00
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
2020-12-15 22:55:54 -05:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2018-08-17 15:46:01 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2009-09-21 17:01:47 -07:00
|
|
|
|
|
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2025-02-21 18:44:24 -08:00
|
|
|
|
|
|
|
|
|
2018-12-28 00:35:22 -08:00
|
|
|
|
2023-02-09 17:48:28 +00:00
|
|
|
|
|
|
|
|
|
2025-05-06 03:30:34 +08:00
|
|
|
|
|
|
|
|
|
2023-02-09 17:48:28 +00:00
|
|
|
|
2018-12-28 00:35:22 -08:00
|
|
|
|
2025-02-21 18:44:24 -08:00
|
|
|
|
2024-03-21 09:36:42 -07:00
|
|
|
|
2025-05-06 03:30:34 +08:00
|
|
|
|
2025-09-15 14:27:56 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-02-21 18:44:24 -08:00
|
|
|
|
|
|
|
|
|
2025-09-15 14:27:56 -07:00
|
|
|
|
2024-03-21 09:36:42 -07:00
|
|
|
|
2018-12-28 00:35:22 -08:00
|
|
|
|
2025-06-04 20:03:26 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-02-21 18:44:24 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2025-02-21 18:44:24 -08:00
|
|
|
|
|
|
|
|
|
2025-05-16 17:34:46 -07:00
|
|
|
|
2025-02-21 18:44:24 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-09-03 11:59:15 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-02-04 22:29:26 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2006-09-25 23:30:55 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-11-06 16:29:57 -08:00
|
|
|
|
|
|
|
|
|
2011-05-11 15:13:34 -07:00
|
|
|
|
|
|
|
|
|
2022-05-31 09:20:51 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-05-11 15:13:34 -07:00
|
|
|
|
2024-02-26 15:55:32 -05:00
|
|
|
|
2024-09-05 22:21:07 -06:00
|
|
|
|
2025-03-14 13:36:12 +00:00
|
|
|
|
2022-05-31 09:20:51 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-05-11 15:13:34 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-07-23 21:28:11 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, page_alloc: disallow __GFP_COMP in alloc_pages_exact()
alloc_pages_exact*() allocates a page of sufficient order and then splits
it to return only the number of pages requested. That makes it
incompatible with __GFP_COMP, because compound pages cannot be split.
As shown by [1] things may silently work until the requested size
(possibly depending on user) stops being power of two. Then for
CONFIG_DEBUG_VM, BUG_ON() triggers in split_page(). Without
CONFIG_DEBUG_VM, consequences are unclear.
There are several options here, none of them great:
1) Don't do the splitting when __GFP_COMP is passed, and return the
whole compound page. However if caller then returns it via
free_pages_exact(), that will be unexpected and the freeing actions
there will be wrong.
2) Warn and remove __GFP_COMP from the flags. But the caller may have
really wanted it, so things may break later somewhere.
3) Warn and return NULL. However NULL may be unexpected, especially
for small sizes.
This patch picks option 2, because as Michal Hocko put it: "callers wanted
it" is much less probable than "caller is simply confused and more gfp
flags is surely better than fewer".
[1] https://lore.kernel.org/lkml/20181126002805.GI18977@shao2-debian/T/#u
Link: http://lkml.kernel.org/r/0c6393eb-b28d-4607-c386-862a71f09de6@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Takashi Iwai <tiwai@suse.de>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-13 17:16:47 -07:00
|
|
|
|
2008-07-23 21:28:11 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-12-28 17:47:04 +03:00
|
|
|
|
2008-07-23 21:28:11 -07:00
|
|
|
|
|
|
|
|
|
2019-03-05 15:48:42 -08:00
|
|
|
|
|
|
|
|
|
2008-07-23 21:28:11 -07:00
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
2008-07-23 21:28:11 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-11-05 13:40:15 -07:00
|
|
|
|
|
|
|
|
|
mm, page_alloc: disallow __GFP_COMP in alloc_pages_exact()
alloc_pages_exact*() allocates a page of sufficient order and then splits
it to return only the number of pages requested. That makes it
incompatible with __GFP_COMP, because compound pages cannot be split.
As shown by [1] things may silently work until the requested size
(possibly depending on user) stops being power of two. Then for
CONFIG_DEBUG_VM, BUG_ON() triggers in split_page(). Without
CONFIG_DEBUG_VM, consequences are unclear.
There are several options here, none of them great:
1) Don't do the splitting when __GFP_COMP is passed, and return the
whole compound page. However if caller then returns it via
free_pages_exact(), that will be unexpected and the freeing actions
there will be wrong.
2) Warn and remove __GFP_COMP from the flags. But the caller may have
really wanted it, so things may break later somewhere.
3) Warn and return NULL. However NULL may be unexpected, especially
for small sizes.
This patch picks option 2, because as Michal Hocko put it: "callers wanted
it" is much less probable than "caller is simply confused and more gfp
flags is surely better than fewer".
[1] https://lore.kernel.org/lkml/20181126002805.GI18977@shao2-debian/T/#u
Link: http://lkml.kernel.org/r/0c6393eb-b28d-4607-c386-862a71f09de6@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Takashi Iwai <tiwai@suse.de>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-13 17:16:47 -07:00
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
2011-05-11 15:13:34 -07:00
|
|
|
|
2008-07-23 21:28:11 -07:00
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
2008-07-23 21:28:11 -07:00
|
|
|
|
2011-05-11 15:13:34 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-05-16 13:16:54 -07:00
|
|
|
|
2011-05-11 15:13:34 -07:00
|
|
|
|
mm, page_alloc: disallow __GFP_COMP in alloc_pages_exact()
alloc_pages_exact*() allocates a page of sufficient order and then splits
it to return only the number of pages requested. That makes it
incompatible with __GFP_COMP, because compound pages cannot be split.
As shown by [1] things may silently work until the requested size
(possibly depending on user) stops being power of two. Then for
CONFIG_DEBUG_VM, BUG_ON() triggers in split_page(). Without
CONFIG_DEBUG_VM, consequences are unclear.
There are several options here, none of them great:
1) Don't do the splitting when __GFP_COMP is passed, and return the
whole compound page. However if caller then returns it via
free_pages_exact(), that will be unexpected and the freeing actions
there will be wrong.
2) Warn and remove __GFP_COMP from the flags. But the caller may have
really wanted it, so things may break later somewhere.
3) Warn and return NULL. However NULL may be unexpected, especially
for small sizes.
This patch picks option 2, because as Michal Hocko put it: "callers wanted
it" is much less probable than "caller is simply confused and more gfp
flags is surely better than fewer".
[1] https://lore.kernel.org/lkml/20181126002805.GI18977@shao2-debian/T/#u
Link: http://lkml.kernel.org/r/0c6393eb-b28d-4607-c386-862a71f09de6@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Takashi Iwai <tiwai@suse.de>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-13 17:16:47 -07:00
|
|
|
|
2011-05-11 15:13:34 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-03-05 15:48:42 -08:00
|
|
|
|
|
|
|
|
|
2011-05-11 15:13:34 -07:00
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
2011-05-11 15:13:34 -07:00
|
|
|
|
2015-11-06 16:29:57 -08:00
|
|
|
|
mm, page_alloc: disallow __GFP_COMP in alloc_pages_exact()
alloc_pages_exact*() allocates a page of sufficient order and then splits
it to return only the number of pages requested. That makes it
incompatible with __GFP_COMP, because compound pages cannot be split.
As shown by [1] things may silently work until the requested size
(possibly depending on user) stops being power of two. Then for
CONFIG_DEBUG_VM, BUG_ON() triggers in split_page(). Without
CONFIG_DEBUG_VM, consequences are unclear.
There are several options here, none of them great:
1) Don't do the splitting when __GFP_COMP is passed, and return the
whole compound page. However if caller then returns it via
free_pages_exact(), that will be unexpected and the freeing actions
there will be wrong.
2) Warn and remove __GFP_COMP from the flags. But the caller may have
really wanted it, so things may break later somewhere.
3) Warn and return NULL. However NULL may be unexpected, especially
for small sizes.
This patch picks option 2, because as Michal Hocko put it: "callers wanted
it" is much less probable than "caller is simply confused and more gfp
flags is surely better than fewer".
[1] https://lore.kernel.org/lkml/20181126002805.GI18977@shao2-debian/T/#u
Link: http://lkml.kernel.org/r/0c6393eb-b28d-4607-c386-862a71f09de6@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Takashi Iwai <tiwai@suse.de>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-13 17:16:47 -07:00
|
|
|
|
|
|
|
|
|
2021-11-05 13:40:15 -07:00
|
|
|
|
|
|
|
|
|
mm, page_alloc: disallow __GFP_COMP in alloc_pages_exact()
alloc_pages_exact*() allocates a page of sufficient order and then splits
it to return only the number of pages requested. That makes it
incompatible with __GFP_COMP, because compound pages cannot be split.
As shown by [1] things may silently work until the requested size
(possibly depending on user) stops being power of two. Then for
CONFIG_DEBUG_VM, BUG_ON() triggers in split_page(). Without
CONFIG_DEBUG_VM, consequences are unclear.
There are several options here, none of them great:
1) Don't do the splitting when __GFP_COMP is passed, and return the
whole compound page. However if caller then returns it via
free_pages_exact(), that will be unexpected and the freeing actions
there will be wrong.
2) Warn and remove __GFP_COMP from the flags. But the caller may have
really wanted it, so things may break later somewhere.
3) Warn and return NULL. However NULL may be unexpected, especially
for small sizes.
This patch picks option 2, because as Michal Hocko put it: "callers wanted
it" is much less probable than "caller is simply confused and more gfp
flags is surely better than fewer".
[1] https://lore.kernel.org/lkml/20181126002805.GI18977@shao2-debian/T/#u
Link: http://lkml.kernel.org/r/0c6393eb-b28d-4607-c386-862a71f09de6@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Takashi Iwai <tiwai@suse.de>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-13 17:16:47 -07:00
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
2011-05-11 15:13:34 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-07-23 21:28:11 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-02-22 16:35:54 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-03-05 15:48:42 -08:00
|
|
|
|
2013-02-22 16:35:54 -08:00
|
|
|
|
|
|
|
|
|
2017-03-30 17:11:36 -03:00
|
|
|
|
|
|
|
|
|
2019-03-05 15:48:42 -08:00
|
|
|
|
|
|
|
|
|
2013-02-22 16:35:54 -08:00
|
|
|
|
2013-02-22 16:35:43 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2008-04-28 02:12:17 -07:00
|
|
|
|
2008-04-28 02:12:16 -07:00
|
|
|
|
|
|
|
|
|
2005-07-29 22:59:18 -07:00
|
|
|
|
2013-02-22 16:35:43 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2008-04-28 02:12:14 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2008-04-28 02:12:16 -07:00
|
|
|
|
2018-12-28 00:34:24 -08:00
|
|
|
|
2009-06-16 15:32:12 -07:00
|
|
|
|
2005-07-29 22:59:18 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-02-22 16:35:54 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-03-05 15:48:42 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2013-02-22 16:35:43 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2005-10-21 02:55:38 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2007-07-17 04:04:39 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2008-04-28 02:12:18 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2006-01-06 00:11:16 -08:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2017-09-06 16:20:30 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2006-01-06 00:11:16 -08:00
|
|
|
|
2013-07-08 16:00:06 -07:00
|
|
|
|
2017-09-06 16:20:30 -07:00
|
|
|
|
2006-01-06 00:11:18 -08:00
|
|
|
|
|
|
|
|
|
2006-09-25 23:31:18 -07:00
|
|
|
|
2006-01-06 00:11:19 -08:00
|
|
|
|
2022-04-14 19:13:43 -07:00
|
|
|
|
2017-09-06 16:20:30 -07:00
|
|
|
|
2006-01-06 00:11:19 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2006-09-25 23:31:18 -07:00
|
|
|
|
2013-07-08 16:00:06 -07:00
|
|
|
|
2006-01-06 00:11:19 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-07-15 23:38:01 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-09-06 16:20:13 -07:00
|
|
|
|
2021-05-06 18:06:47 -07:00
|
|
|
|
2017-09-06 16:20:13 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-07-15 23:38:01 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-05-16 14:38:20 +08:00
|
|
|
|
|
|
|
|
|
2007-07-15 23:38:01 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
sysctl: treewide: constify the ctl_table argument of proc_handlers
const qualify the struct ctl_table argument in the proc_handler function
signatures. This is a prerequisite to moving the static ctl_table
structs into .rodata data which will ensure that proc_handler function
pointers cannot be modified.
This patch has been generated by the following coccinelle script:
```
virtual patch
@r1@
identifier ctl, write, buffer, lenp, ppos;
identifier func !~ "appldata_(timer|interval)_handler|sched_(rt|rr)_handler|rds_tcp_skbuf_handler|proc_sctp_do_(hmac_alg|rto_min|rto_max|udp_port|alpha_beta|auth|probe_interval)";
@@
int func(
- struct ctl_table *ctl
+ const struct ctl_table *ctl
,int write, void *buffer, size_t *lenp, loff_t *ppos);
@r2@
identifier func, ctl, write, buffer, lenp, ppos;
@@
int func(
- struct ctl_table *ctl
+ const struct ctl_table *ctl
,int write, void *buffer, size_t *lenp, loff_t *ppos)
{ ... }
@r3@
identifier func;
@@
int func(
- struct ctl_table *
+ const struct ctl_table *
,int , void *, size_t *, loff_t *);
@r4@
identifier func, ctl;
@@
int func(
- struct ctl_table *ctl
+ const struct ctl_table *ctl
,int , void *, size_t *, loff_t *);
@r5@
identifier func, write, buffer, lenp, ppos;
@@
int func(
- struct ctl_table *
+ const struct ctl_table *
,int write, void *buffer, size_t *lenp, loff_t *ppos);
```
* Code formatting was adjusted in xfs_sysctl.c to comply with code
conventions. The xfs_stats_clear_proc_handler,
xfs_panic_mask_proc_handler and xfs_deprecated_dointvec_minmax where
adjusted.
* The ctl_table argument in proc_watchdog_common was const qualified.
This is called from a proc_handler itself and is calling back into
another proc_handler, making it necessary to change it as part of the
proc_handler migration.
Co-developed-by: Thomas Weißschuh <linux@weissschuh.net>
Signed-off-by: Thomas Weißschuh <linux@weissschuh.net>
Co-developed-by: Joel Granados <j.granados@samsung.com>
Signed-off-by: Joel Granados <j.granados@samsung.com>
2024-07-24 20:59:29 +02:00
|
|
|
|
2020-04-24 08:43:38 +02:00
|
|
|
|
2007-07-15 23:38:01 -07:00
|
|
|
|
2020-04-24 08:43:38 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-07-15 23:38:01 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2005-05-01 08:59:25 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-03-05 15:48:42 -08:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
mm/numa: automatically generate node migration order
Patch series "Migrate Pages in lieu of discard", v11.
We're starting to see systems with more and more kinds of memory such as
Intel's implementation of persistent memory.
Let's say you have a system with some DRAM and some persistent memory.
Today, once DRAM fills up, reclaim will start and some of the DRAM
contents will be thrown out. Allocations will, at some point, start
falling over to the slower persistent memory.
That has two nasty properties. First, the newer allocations can end up in
the slower persistent memory. Second, reclaimed data in DRAM are just
discarded even if there are gobs of space in persistent memory that could
be used.
This patchset implements a solution to these problems. At the end of the
reclaim process in shrink_page_list() just before the last page refcount
is dropped, the page is migrated to persistent memory instead of being
dropped.
While I've talked about a DRAM/PMEM pairing, this approach would function
in any environment where memory tiers exist.
This is not perfect. It "strands" pages in slower memory and never brings
them back to fast DRAM. Huang Ying has follow-on work which repurposes
NUMA balancing to promote hot pages back to DRAM.
This is also all based on an upstream mechanism that allows persistent
memory to be onlined and used as if it were volatile:
http://lkml.kernel.org/r/20190124231441.37A4A305@viggo.jf.intel.com
With that, the DRAM and PMEM in each socket will be represented as 2
separate NUMA nodes, with the CPUs sit in the DRAM node. So the
general inter-NUMA demotion mechanism introduced in the patchset can
migrate the cold DRAM pages to the PMEM node.
We have tested the patchset with the postgresql and pgbench. On a
2-socket server machine with DRAM and PMEM, the kernel with the patchset
can improve the score of pgbench up to 22.1% compared with that of the
DRAM only + disk case. This comes from the reduced disk read throughput
(which reduces up to 70.8%).
== Open Issues ==
* Memory policies and cpusets that, for instance, restrict allocations
to DRAM can be demoted to PMEM whenever they opt in to this
new mechanism. A cgroup-level API to opt-in or opt-out of
these migrations will likely be required as a follow-on.
* Could be more aggressive about where anon LRU scanning occurs
since it no longer necessarily involves I/O. get_scan_count()
for instance says: "If we have no swap space, do not bother
scanning anon pages"
This patch (of 9):
Prepare for the kernel to auto-migrate pages to other memory nodes with a
node migration table. This allows creating single migration target for
each NUMA node to enable the kernel to do NUMA page migrations instead of
simply discarding colder pages. A node with no target is a "terminal
node", so reclaim acts normally there. The migration target does not
fundamentally _need_ to be a single node, but this implementation starts
there to limit complexity.
When memory fills up on a node, memory contents can be automatically
migrated to another node. The biggest problems are knowing when to
migrate and to where the migration should be targeted.
The most straightforward way to generate the "to where" list would be to
follow the page allocator fallback lists. Those lists already tell us if
memory is full where to look next. It would also be logical to move
memory in that order.
But, the allocator fallback lists have a fatal flaw: most nodes appear in
all the lists. This would potentially lead to migration cycles (A->B,
B->A, A->B, ...).
Instead of using the allocator fallback lists directly, keep a separate
node migration ordering. But, reuse the same data used to generate page
allocator fallback in the first place: find_next_best_node().
This means that the firmware data used to populate node distances
essentially dictates the ordering for now. It should also be
architecture-neutral since all NUMA architectures have a working
find_next_best_node().
RCU is used to allow lock-less read of node_demotion[] and prevent
demotion cycles been observed. If multiple reads of node_demotion[] are
performed, a single rcu_read_lock() must be held over all reads to ensure
no cycles are observed. Details are as follows.
=== What does RCU provide? ===
Imagine a simple loop which walks down the demotion path looking
for the last node:
terminal_node = start_node;
while (node_demotion[terminal_node] != NUMA_NO_NODE) {
terminal_node = node_demotion[terminal_node];
}
The initial values are:
node_demotion[0] = 1;
node_demotion[1] = NUMA_NO_NODE;
and are updated to:
node_demotion[0] = NUMA_NO_NODE;
node_demotion[1] = 0;
What guarantees that the cycle is not observed:
node_demotion[0] = 1;
node_demotion[1] = 0;
and would loop forever?
With RCU, a rcu_read_lock/unlock() can be placed around the loop. Since
the write side does a synchronize_rcu(), the loop that observed the old
contents is known to be complete before the synchronize_rcu() has
completed.
RCU, combined with disable_all_migrate_targets(), ensures that the old
migration state is not visible by the time __set_migration_target_nodes()
is called.
=== What does READ_ONCE() provide? ===
READ_ONCE() forbids the compiler from merging or reordering successive
reads of node_demotion[]. This ensures that any updates are *eventually*
observed.
Consider the above loop again. The compiler could theoretically read the
entirety of node_demotion[] into local storage (registers) and never go
back to memory, and *permanently* observe bad values for node_demotion[].
Note: RCU does not provide any universal compiler-ordering
guarantees:
https://lore.kernel.org/lkml/20150921204327.GH4029@linux.vnet.ibm.com/
This code is unused for now. It will be called later in the
series.
Link: https://lkml.kernel.org/r/20210721063926.3024591-1-ying.huang@intel.com
Link: https://lkml.kernel.org/r/20210715055145.195411-1-ying.huang@intel.com
Link: https://lkml.kernel.org/r/20210715055145.195411-2-ying.huang@intel.com
Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com>
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Reviewed-by: Yang Shi <shy828301@gmail.com>
Reviewed-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Wei Xu <weixugc@google.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Keith Busch <kbusch@kernel.org>
Cc: Yang Shi <yang.shi@linux.alibaba.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-09-02 14:59:06 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2006-02-17 20:38:21 +01:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2013-02-22 16:35:36 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
mm: page_alloc: skip memoryless nodes entirely
Patch series "handle memoryless nodes more appropriately", v3.
Currently, in the process of initialization or offline memory, memoryless
nodes will still be built into the fallback list of itself or other nodes.
This is not what we expected, so this patch series removes memoryless
nodes from the fallback list entirely.
This patch (of 2):
In find_next_best_node(), we skipped the memoryless nodes when building
the zonelists of other normal nodes (N_NORMAL), but did not skip the
memoryless node itself when building the zonelist. This will cause it to
be traversed at runtime.
For example, say we have node0 and node1, node0 is memoryless
node, then the fallback order of node0 and node1 as follows:
[ 0.153005] Fallback order for Node 0: 0 1
[ 0.153564] Fallback order for Node 1: 1
After this patch, we skip memoryless node0 entirely, then
the fallback order of node0 and node1 as follows:
[ 0.155236] Fallback order for Node 0: 1
[ 0.155806] Fallback order for Node 1: 1
So it becomes completely invisible, which will reduce runtime
overhead.
And in this way, we will not try to allocate pages from memoryless node0,
then the panic mentioned in [1] will also be fixed. Even though this
problem has been solved by dropping the NODE_MIN_SIZE constrain in x86
[2], it would be better to fix it in core MM as well.
[1]. https://lore.kernel.org/all/20230212110305.93670-1-zhengqi.arch@bytedance.com/
[2]. https://lore.kernel.org/all/20231017062215.171670-1-rppt@kernel.org/
[zhengqi.arch@bytedance.com: update comment, per Ingo]
Link: https://lkml.kernel.org/r/7300fc00a057eefeb9a68c8ad28171c3f0ce66ce.1697799303.git.zhengqi.arch@bytedance.com
Link: https://lkml.kernel.org/r/cover.1697799303.git.zhengqi.arch@bytedance.com
Link: https://lkml.kernel.org/r/cover.1697711415.git.zhengqi.arch@bytedance.com
Link: https://lkml.kernel.org/r/157013e978468241de4a4c05d5337a44638ecb0e.1697711415.git.zhengqi.arch@bytedance.com
Signed-off-by: Qi Zheng <zhengqi.arch@bytedance.com>
Acked-by: David Hildenbrand <david@redhat.com>
Acked-by: Ingo Molnar <mingo@kernel.org>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 18:43:54 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-02-17 20:38:21 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2012-12-12 13:51:46 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-02-17 20:38:21 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2020-10-13 16:55:42 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-04-28 23:16:01 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-07-15 23:38:01 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-09-06 16:20:30 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2017-09-06 16:20:30 -07:00
|
|
|
|
|
|
|
|
|
2007-07-15 23:38:01 -07:00
|
|
|
|
2017-09-06 16:20:30 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-07-15 23:38:01 -07:00
|
|
|
|
|
|
|
|
|
2007-10-16 01:25:37 -07:00
|
|
|
|
2024-06-19 01:06:11 +00:00
|
|
|
|
2007-10-16 01:25:37 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-09-06 16:20:30 -07:00
|
|
|
|
|
|
|
|
|
2007-10-16 01:25:37 -07:00
|
|
|
|
2017-09-06 16:20:30 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-10-16 01:25:37 -07:00
|
|
|
|
|
|
|
|
|
2007-07-15 23:38:01 -07:00
|
|
|
|
|
|
|
|
|
2017-09-06 16:20:30 -07:00
|
|
|
|
2022-04-28 23:16:01 -07:00
|
|
|
|
2020-06-03 15:59:05 -07:00
|
|
|
|
2007-07-15 23:38:01 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-07-15 23:38:01 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-10-08 16:33:24 -07:00
|
|
|
|
|
|
|
|
|
2022-04-28 23:16:01 -07:00
|
|
|
|
2007-07-15 23:38:01 -07:00
|
|
|
|
2017-09-06 16:20:30 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2007-10-16 01:25:37 -07:00
|
|
|
|
2017-09-06 16:20:30 -07:00
|
|
|
|
2007-10-16 01:25:37 -07:00
|
|
|
|
2021-11-05 13:40:18 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2010-05-26 14:45:00 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-05-19 17:14:10 -07:00
|
|
|
|
2010-05-26 14:45:00 -07:00
|
|
|
|
2016-05-19 17:14:10 -07:00
|
|
|
|
2010-05-26 14:45:00 -07:00
|
|
|
|
2016-05-19 17:14:10 -07:00
|
|
|
|
2024-07-29 14:47:17 +05:30
|
|
|
|
2010-05-26 14:45:00 -07:00
|
|
|
|
|
|
|
|
|
2007-07-15 23:38:01 -07:00
|
|
|
|
2016-08-10 16:27:49 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2007-07-15 23:38:01 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2017-09-06 16:20:30 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2017-09-06 16:20:30 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2017-09-06 16:20:30 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-01-05 15:34:51 +09:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/page_alloc: split per cpu page lists and zone stats
The PCP (per-cpu page allocator in page_alloc.c) shares locking
requirements with vmstat and the zone lock which is inconvenient and
causes some issues. For example, the PCP list and vmstat share the same
per-cpu space meaning that it's possible that vmstat updates dirty cache
lines holding per-cpu lists across CPUs unless padding is used. Second,
PREEMPT_RT does not want to disable IRQs for too long in the page
allocator.
This series splits the locking requirements and uses locks types more
suitable for PREEMPT_RT, reduces the time when special locking is required
for stats and reduces the time when IRQs need to be disabled on
!PREEMPT_RT kernels.
Why local_lock? PREEMPT_RT considers the following sequence to be unsafe
as documented in Documentation/locking/locktypes.rst
local_irq_disable();
spin_lock(&lock);
The pcp allocator has this sequence for rmqueue_pcplist (local_irq_save)
-> __rmqueue_pcplist -> rmqueue_bulk (spin_lock). While it's possible to
separate this out, it generally means there are points where we enable
IRQs and reenable them again immediately. To prevent a migration and the
per-cpu pointer going stale, migrate_disable is also needed. That is a
custom lock that is similar, but worse, than local_lock. Furthermore, on
PREEMPT_RT, it's undesirable to leave IRQs disabled for too long. By
converting to local_lock which disables migration on PREEMPT_RT, the
locking requirements can be separated and start moving the protections for
PCP, stats and the zone lock to PREEMPT_RT-safe equivalent locking. As a
bonus, local_lock also means that PROVE_LOCKING does something useful.
After that, it's obvious that zone_statistics incurs too much overhead and
leaves IRQs disabled for longer than necessary on !PREEMPT_RT kernels.
zone_statistics uses perfectly accurate counters requiring IRQs be
disabled for parallel RMW sequences when inaccurate ones like vm_events
would do. The series makes the NUMA statistics (NUMA_HIT and friends)
inaccurate counters that then require no special protection on
!PREEMPT_RT.
The bulk page allocator can then do stat updates in bulk with IRQs enabled
which should improve the efficiency. Technically, this could have been
done without the local_lock and vmstat conversion work and the order
simply reflects the timing of when different series were implemented.
Finally, there are places where we conflate IRQs being disabled for the
PCP with the IRQ-safe zone spinlock. The remainder of the series reduces
the scope of what is protected by disabled IRQs on !PREEMPT_RT kernels.
By the end of the series, page_alloc.c does not call local_irq_save so the
locking scope is a bit clearer. The one exception is that modifying
NR_FREE_PAGES still happens in places where it's known the IRQs are
disabled as it's harmless for PREEMPT_RT and would be expensive to split
the locking there.
No performance data is included because despite the overhead of the stats,
it's within the noise for most workloads on !PREEMPT_RT. However, Jesper
Dangaard Brouer ran a page allocation microbenchmark on a E5-1650 v4 @
3.60GHz CPU on the first version of this series. Focusing on the array
variant of the bulk page allocator reveals the following.
(CPU: Intel(R) Xeon(R) CPU E5-1650 v4 @ 3.60GHz)
ARRAY variant: time_bulk_page_alloc_free_array: step=bulk size
Baseline Patched
1 56.383 54.225 (+3.83%)
2 40.047 35.492 (+11.38%)
3 37.339 32.643 (+12.58%)
4 35.578 30.992 (+12.89%)
8 33.592 29.606 (+11.87%)
16 32.362 28.532 (+11.85%)
32 31.476 27.728 (+11.91%)
64 30.633 27.252 (+11.04%)
128 30.596 27.090 (+11.46%)
While this is a positive outcome, the series is more likely to be
interesting to the RT people in terms of getting parts of the PREEMPT_RT
tree into mainline.
This patch (of 9):
The per-cpu page allocator lists and the per-cpu vmstat deltas are stored
in the same struct per_cpu_pages even though vmstats have no direct impact
on the per-cpu page lists. This is inconsistent because the vmstats for a
node are stored on a dedicated structure. The bigger issue is that the
per_cpu_pages structure is not cache-aligned and stat updates either cache
conflict with adjacent per-cpu lists incurring a runtime cost or padding
is required incurring a memory cost.
This patch splits the per-cpu pagelists and the vmstat deltas into
separate structures. It's mostly a mechanical conversion but some
variable renaming is done to clearly distinguish the per-cpu pages
structure (pcp) from the vmstats (pzstats).
Superficially, this appears to increase the size of the per_cpu_pages
structure but the movement of expire fills a structure hole so there is no
impact overall.
[mgorman@techsingularity.net: make it W=1 cleaner]
Link: https://lkml.kernel.org/r/20210514144622.GA3735@techsingularity.net
[mgorman@techsingularity.net: make it W=1 even cleaner]
Link: https://lkml.kernel.org/r/20210516140705.GB3735@techsingularity.net
[lkp@intel.com: check struct per_cpu_zonestat has a non-zero size]
[vbabka@suse.cz: Init zone->per_cpu_zonestats properly]
Link: https://lkml.kernel.org/r/20210512095458.30632-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20210512095458.30632-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-28 19:41:38 -07:00
|
|
|
|
2020-12-14 19:10:53 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/page_alloc: split per cpu page lists and zone stats
The PCP (per-cpu page allocator in page_alloc.c) shares locking
requirements with vmstat and the zone lock which is inconvenient and
causes some issues. For example, the PCP list and vmstat share the same
per-cpu space meaning that it's possible that vmstat updates dirty cache
lines holding per-cpu lists across CPUs unless padding is used. Second,
PREEMPT_RT does not want to disable IRQs for too long in the page
allocator.
This series splits the locking requirements and uses locks types more
suitable for PREEMPT_RT, reduces the time when special locking is required
for stats and reduces the time when IRQs need to be disabled on
!PREEMPT_RT kernels.
Why local_lock? PREEMPT_RT considers the following sequence to be unsafe
as documented in Documentation/locking/locktypes.rst
local_irq_disable();
spin_lock(&lock);
The pcp allocator has this sequence for rmqueue_pcplist (local_irq_save)
-> __rmqueue_pcplist -> rmqueue_bulk (spin_lock). While it's possible to
separate this out, it generally means there are points where we enable
IRQs and reenable them again immediately. To prevent a migration and the
per-cpu pointer going stale, migrate_disable is also needed. That is a
custom lock that is similar, but worse, than local_lock. Furthermore, on
PREEMPT_RT, it's undesirable to leave IRQs disabled for too long. By
converting to local_lock which disables migration on PREEMPT_RT, the
locking requirements can be separated and start moving the protections for
PCP, stats and the zone lock to PREEMPT_RT-safe equivalent locking. As a
bonus, local_lock also means that PROVE_LOCKING does something useful.
After that, it's obvious that zone_statistics incurs too much overhead and
leaves IRQs disabled for longer than necessary on !PREEMPT_RT kernels.
zone_statistics uses perfectly accurate counters requiring IRQs be
disabled for parallel RMW sequences when inaccurate ones like vm_events
would do. The series makes the NUMA statistics (NUMA_HIT and friends)
inaccurate counters that then require no special protection on
!PREEMPT_RT.
The bulk page allocator can then do stat updates in bulk with IRQs enabled
which should improve the efficiency. Technically, this could have been
done without the local_lock and vmstat conversion work and the order
simply reflects the timing of when different series were implemented.
Finally, there are places where we conflate IRQs being disabled for the
PCP with the IRQ-safe zone spinlock. The remainder of the series reduces
the scope of what is protected by disabled IRQs on !PREEMPT_RT kernels.
By the end of the series, page_alloc.c does not call local_irq_save so the
locking scope is a bit clearer. The one exception is that modifying
NR_FREE_PAGES still happens in places where it's known the IRQs are
disabled as it's harmless for PREEMPT_RT and would be expensive to split
the locking there.
No performance data is included because despite the overhead of the stats,
it's within the noise for most workloads on !PREEMPT_RT. However, Jesper
Dangaard Brouer ran a page allocation microbenchmark on a E5-1650 v4 @
3.60GHz CPU on the first version of this series. Focusing on the array
variant of the bulk page allocator reveals the following.
(CPU: Intel(R) Xeon(R) CPU E5-1650 v4 @ 3.60GHz)
ARRAY variant: time_bulk_page_alloc_free_array: step=bulk size
Baseline Patched
1 56.383 54.225 (+3.83%)
2 40.047 35.492 (+11.38%)
3 37.339 32.643 (+12.58%)
4 35.578 30.992 (+12.89%)
8 33.592 29.606 (+11.87%)
16 32.362 28.532 (+11.85%)
32 31.476 27.728 (+11.91%)
64 30.633 27.252 (+11.04%)
128 30.596 27.090 (+11.46%)
While this is a positive outcome, the series is more likely to be
interesting to the RT people in terms of getting parts of the PREEMPT_RT
tree into mainline.
This patch (of 9):
The per-cpu page allocator lists and the per-cpu vmstat deltas are stored
in the same struct per_cpu_pages even though vmstats have no direct impact
on the per-cpu page lists. This is inconsistent because the vmstats for a
node are stored on a dedicated structure. The bigger issue is that the
per_cpu_pages structure is not cache-aligned and stat updates either cache
conflict with adjacent per-cpu lists incurring a runtime cost or padding
is required incurring a memory cost.
This patch splits the per-cpu pagelists and the vmstat deltas into
separate structures. It's mostly a mechanical conversion but some
variable renaming is done to clearly distinguish the per-cpu pages
structure (pcp) from the vmstats (pzstats).
Superficially, this appears to increase the size of the per_cpu_pages
structure but the movement of expire fills a structure hole so there is no
impact overall.
[mgorman@techsingularity.net: make it W=1 cleaner]
Link: https://lkml.kernel.org/r/20210514144622.GA3735@techsingularity.net
[mgorman@techsingularity.net: make it W=1 even cleaner]
Link: https://lkml.kernel.org/r/20210516140705.GB3735@techsingularity.net
[lkp@intel.com: check struct per_cpu_zonestat has a non-zero size]
[vbabka@suse.cz: Init zone->per_cpu_zonestats properly]
Link: https://lkml.kernel.org/r/20210512095458.30632-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20210512095458.30632-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-28 19:41:38 -07:00
|
|
|
|
|
|
|
|
|
2010-01-05 15:34:51 +09:00
|
|
|
|
mm, page_alloc: remove stop_machine from build_all_zonelists
build_all_zonelists has been (ab)using stop_machine to make sure that
zonelists do not change while somebody is looking at them. This is is
just a gross hack because a) it complicates the context from which we
can call build_all_zonelists (see 3f906ba23689 ("mm/memory-hotplug:
switch locking to a percpu rwsem")) and b) is is not really necessary
especially after "mm, page_alloc: simplify zonelist initialization" and
c) it doesn't really provide the protection it claims (see below).
Updates of the zonelists happen very seldom, basically only when a zone
becomes populated during memory online or when it loses all the memory
during offline. A racing iteration over zonelists could either miss a
zone or try to work on one zone twice. Both of these are something we
can live with occasionally because there will always be at least one
zone visible so we are not likely to fail allocation too easily for
example.
Please note that the original stop_machine approach doesn't really
provide a better exclusion because the iteration might be interrupted
half way (unless the whole iteration is preempt disabled which is not
the case in most cases) so the some zones could still be seen twice or a
zone missed.
I have run the pathological online/offline of the single memblock in the
movable zone while stressing the same small node with some memory
pressure.
Node 1, zone DMA
pages free 0
min 0
low 0
high 0
spanned 0
present 0
managed 0
protection: (0, 943, 943, 943)
Node 1, zone DMA32
pages free 227310
min 8294
low 10367
high 12440
spanned 262112
present 262112
managed 241436
protection: (0, 0, 0, 0)
Node 1, zone Normal
pages free 0
min 0
low 0
high 0
spanned 0
present 0
managed 0
protection: (0, 0, 0, 1024)
Node 1, zone Movable
pages free 32722
min 85
low 117
high 149
spanned 32768
present 32768
managed 32768
protection: (0, 0, 0, 0)
root@test1:/sys/devices/system/node/node1# while true
do
echo offline > memory34/state
echo online_movable > memory34/state
done
root@test1:/mnt/data/test/linux-3.7-rc5# numactl --preferred=1 make -j4
and it survived without any unexpected behavior. While this is not
really a great testing coverage it should exercise the allocation path
quite a lot.
Link: http://lkml.kernel.org/r/20170721143915.14161-8-mhocko@kernel.org
Signed-off-by: Michal Hocko <mhocko@suse.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Joonsoo Kim <js1304@gmail.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Shaohua Li <shaohua.li@intel.com>
Cc: Toshi Kani <toshi.kani@hpe.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 16:20:34 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2006-06-23 02:03:11 -07:00
|
|
|
|
2017-09-06 16:20:17 -07:00
|
|
|
|
2012-07-31 16:43:28 -07:00
|
|
|
|
mm/page_alloc: fix potential deadlock on zonelist_update_seq seqlock
syzbot is reporting circular locking dependency which involves
zonelist_update_seq seqlock [1], for this lock is checked by memory
allocation requests which do not need to be retried.
One deadlock scenario is kmalloc(GFP_ATOMIC) from an interrupt handler.
CPU0
----
__build_all_zonelists() {
write_seqlock(&zonelist_update_seq); // makes zonelist_update_seq.seqcount odd
// e.g. timer interrupt handler runs at this moment
some_timer_func() {
kmalloc(GFP_ATOMIC) {
__alloc_pages_slowpath() {
read_seqbegin(&zonelist_update_seq) {
// spins forever because zonelist_update_seq.seqcount is odd
}
}
}
}
// e.g. timer interrupt handler finishes
write_sequnlock(&zonelist_update_seq); // makes zonelist_update_seq.seqcount even
}
This deadlock scenario can be easily eliminated by not calling
read_seqbegin(&zonelist_update_seq) from !__GFP_DIRECT_RECLAIM allocation
requests, for retry is applicable to only __GFP_DIRECT_RECLAIM allocation
requests. But Michal Hocko does not know whether we should go with this
approach.
Another deadlock scenario which syzbot is reporting is a race between
kmalloc(GFP_ATOMIC) from tty_insert_flip_string_and_push_buffer() with
port->lock held and printk() from __build_all_zonelists() with
zonelist_update_seq held.
CPU0 CPU1
---- ----
pty_write() {
tty_insert_flip_string_and_push_buffer() {
__build_all_zonelists() {
write_seqlock(&zonelist_update_seq);
build_zonelists() {
printk() {
vprintk() {
vprintk_default() {
vprintk_emit() {
console_unlock() {
console_flush_all() {
console_emit_next_record() {
con->write() = serial8250_console_write() {
spin_lock_irqsave(&port->lock, flags);
tty_insert_flip_string() {
tty_insert_flip_string_fixed_flag() {
__tty_buffer_request_room() {
tty_buffer_alloc() {
kmalloc(GFP_ATOMIC | __GFP_NOWARN) {
__alloc_pages_slowpath() {
zonelist_iter_begin() {
read_seqbegin(&zonelist_update_seq); // spins forever because zonelist_update_seq.seqcount is odd
spin_lock_irqsave(&port->lock, flags); // spins forever because port->lock is held
}
}
}
}
}
}
}
}
spin_unlock_irqrestore(&port->lock, flags);
// message is printed to console
spin_unlock_irqrestore(&port->lock, flags);
}
}
}
}
}
}
}
}
}
write_sequnlock(&zonelist_update_seq);
}
}
}
This deadlock scenario can be eliminated by
preventing interrupt context from calling kmalloc(GFP_ATOMIC)
and
preventing printk() from calling console_flush_all()
while zonelist_update_seq.seqcount is odd.
Since Petr Mladek thinks that __build_all_zonelists() can become a
candidate for deferring printk() [2], let's address this problem by
disabling local interrupts in order to avoid kmalloc(GFP_ATOMIC)
and
disabling synchronous printk() in order to avoid console_flush_all()
.
As a side effect of minimizing duration of zonelist_update_seq.seqcount
being odd by disabling synchronous printk(), latency at
read_seqbegin(&zonelist_update_seq) for both !__GFP_DIRECT_RECLAIM and
__GFP_DIRECT_RECLAIM allocation requests will be reduced. Although, from
lockdep perspective, not calling read_seqbegin(&zonelist_update_seq) (i.e.
do not record unnecessary locking dependency) from interrupt context is
still preferable, even if we don't allow calling kmalloc(GFP_ATOMIC)
inside
write_seqlock(&zonelist_update_seq)/write_sequnlock(&zonelist_update_seq)
section...
Link: https://lkml.kernel.org/r/8796b95c-3da3-5885-fddd-6ef55f30e4d3@I-love.SAKURA.ne.jp
Fixes: 3d36424b3b58 ("mm/page_alloc: fix race condition between build_all_zonelists and page allocation")
Link: https://lkml.kernel.org/r/ZCrs+1cDqPWTDFNM@alley [2]
Reported-by: syzbot <syzbot+223c7461c58c58a4cb10@syzkaller.appspotmail.com>
Link: https://syzkaller.appspot.com/bug?extid=223c7461c58c58a4cb10 [1]
Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Petr Mladek <pmladek@suse.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Ilpo Järvinen <ilpo.jarvinen@linux.intel.com>
Cc: John Ogness <john.ogness@linutronix.de>
Cc: Patrick Daly <quic_pdaly@quicinc.com>
Cc: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-04-04 23:31:58 +09:00
|
|
|
|
2017-09-06 16:20:37 -07:00
|
|
|
|
mm/page_alloc: fix potential deadlock on zonelist_update_seq seqlock
syzbot is reporting circular locking dependency which involves
zonelist_update_seq seqlock [1], for this lock is checked by memory
allocation requests which do not need to be retried.
One deadlock scenario is kmalloc(GFP_ATOMIC) from an interrupt handler.
CPU0
----
__build_all_zonelists() {
write_seqlock(&zonelist_update_seq); // makes zonelist_update_seq.seqcount odd
// e.g. timer interrupt handler runs at this moment
some_timer_func() {
kmalloc(GFP_ATOMIC) {
__alloc_pages_slowpath() {
read_seqbegin(&zonelist_update_seq) {
// spins forever because zonelist_update_seq.seqcount is odd
}
}
}
}
// e.g. timer interrupt handler finishes
write_sequnlock(&zonelist_update_seq); // makes zonelist_update_seq.seqcount even
}
This deadlock scenario can be easily eliminated by not calling
read_seqbegin(&zonelist_update_seq) from !__GFP_DIRECT_RECLAIM allocation
requests, for retry is applicable to only __GFP_DIRECT_RECLAIM allocation
requests. But Michal Hocko does not know whether we should go with this
approach.
Another deadlock scenario which syzbot is reporting is a race between
kmalloc(GFP_ATOMIC) from tty_insert_flip_string_and_push_buffer() with
port->lock held and printk() from __build_all_zonelists() with
zonelist_update_seq held.
CPU0 CPU1
---- ----
pty_write() {
tty_insert_flip_string_and_push_buffer() {
__build_all_zonelists() {
write_seqlock(&zonelist_update_seq);
build_zonelists() {
printk() {
vprintk() {
vprintk_default() {
vprintk_emit() {
console_unlock() {
console_flush_all() {
console_emit_next_record() {
con->write() = serial8250_console_write() {
spin_lock_irqsave(&port->lock, flags);
tty_insert_flip_string() {
tty_insert_flip_string_fixed_flag() {
__tty_buffer_request_room() {
tty_buffer_alloc() {
kmalloc(GFP_ATOMIC | __GFP_NOWARN) {
__alloc_pages_slowpath() {
zonelist_iter_begin() {
read_seqbegin(&zonelist_update_seq); // spins forever because zonelist_update_seq.seqcount is odd
spin_lock_irqsave(&port->lock, flags); // spins forever because port->lock is held
}
}
}
}
}
}
}
}
spin_unlock_irqrestore(&port->lock, flags);
// message is printed to console
spin_unlock_irqrestore(&port->lock, flags);
}
}
}
}
}
}
}
}
}
write_sequnlock(&zonelist_update_seq);
}
}
}
This deadlock scenario can be eliminated by
preventing interrupt context from calling kmalloc(GFP_ATOMIC)
and
preventing printk() from calling console_flush_all()
while zonelist_update_seq.seqcount is odd.
Since Petr Mladek thinks that __build_all_zonelists() can become a
candidate for deferring printk() [2], let's address this problem by
disabling local interrupts in order to avoid kmalloc(GFP_ATOMIC)
and
disabling synchronous printk() in order to avoid console_flush_all()
.
As a side effect of minimizing duration of zonelist_update_seq.seqcount
being odd by disabling synchronous printk(), latency at
read_seqbegin(&zonelist_update_seq) for both !__GFP_DIRECT_RECLAIM and
__GFP_DIRECT_RECLAIM allocation requests will be reduced. Although, from
lockdep perspective, not calling read_seqbegin(&zonelist_update_seq) (i.e.
do not record unnecessary locking dependency) from interrupt context is
still preferable, even if we don't allow calling kmalloc(GFP_ATOMIC)
inside
write_seqlock(&zonelist_update_seq)/write_sequnlock(&zonelist_update_seq)
section...
Link: https://lkml.kernel.org/r/8796b95c-3da3-5885-fddd-6ef55f30e4d3@I-love.SAKURA.ne.jp
Fixes: 3d36424b3b58 ("mm/page_alloc: fix race condition between build_all_zonelists and page allocation")
Link: https://lkml.kernel.org/r/ZCrs+1cDqPWTDFNM@alley [2]
Reported-by: syzbot <syzbot+223c7461c58c58a4cb10@syzkaller.appspotmail.com>
Link: https://syzkaller.appspot.com/bug?extid=223c7461c58c58a4cb10 [1]
Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Petr Mladek <pmladek@suse.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Ilpo Järvinen <ilpo.jarvinen@linux.intel.com>
Cc: John Ogness <john.ogness@linutronix.de>
Cc: Patrick Daly <quic_pdaly@quicinc.com>
Cc: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-04-04 23:31:58 +09:00
|
|
|
|
2023-06-23 22:15:17 +02:00
|
|
|
|
|
|
|
|
|
mm/page_alloc: fix potential deadlock on zonelist_update_seq seqlock
syzbot is reporting circular locking dependency which involves
zonelist_update_seq seqlock [1], for this lock is checked by memory
allocation requests which do not need to be retried.
One deadlock scenario is kmalloc(GFP_ATOMIC) from an interrupt handler.
CPU0
----
__build_all_zonelists() {
write_seqlock(&zonelist_update_seq); // makes zonelist_update_seq.seqcount odd
// e.g. timer interrupt handler runs at this moment
some_timer_func() {
kmalloc(GFP_ATOMIC) {
__alloc_pages_slowpath() {
read_seqbegin(&zonelist_update_seq) {
// spins forever because zonelist_update_seq.seqcount is odd
}
}
}
}
// e.g. timer interrupt handler finishes
write_sequnlock(&zonelist_update_seq); // makes zonelist_update_seq.seqcount even
}
This deadlock scenario can be easily eliminated by not calling
read_seqbegin(&zonelist_update_seq) from !__GFP_DIRECT_RECLAIM allocation
requests, for retry is applicable to only __GFP_DIRECT_RECLAIM allocation
requests. But Michal Hocko does not know whether we should go with this
approach.
Another deadlock scenario which syzbot is reporting is a race between
kmalloc(GFP_ATOMIC) from tty_insert_flip_string_and_push_buffer() with
port->lock held and printk() from __build_all_zonelists() with
zonelist_update_seq held.
CPU0 CPU1
---- ----
pty_write() {
tty_insert_flip_string_and_push_buffer() {
__build_all_zonelists() {
write_seqlock(&zonelist_update_seq);
build_zonelists() {
printk() {
vprintk() {
vprintk_default() {
vprintk_emit() {
console_unlock() {
console_flush_all() {
console_emit_next_record() {
con->write() = serial8250_console_write() {
spin_lock_irqsave(&port->lock, flags);
tty_insert_flip_string() {
tty_insert_flip_string_fixed_flag() {
__tty_buffer_request_room() {
tty_buffer_alloc() {
kmalloc(GFP_ATOMIC | __GFP_NOWARN) {
__alloc_pages_slowpath() {
zonelist_iter_begin() {
read_seqbegin(&zonelist_update_seq); // spins forever because zonelist_update_seq.seqcount is odd
spin_lock_irqsave(&port->lock, flags); // spins forever because port->lock is held
}
}
}
}
}
}
}
}
spin_unlock_irqrestore(&port->lock, flags);
// message is printed to console
spin_unlock_irqrestore(&port->lock, flags);
}
}
}
}
}
}
}
}
}
write_sequnlock(&zonelist_update_seq);
}
}
}
This deadlock scenario can be eliminated by
preventing interrupt context from calling kmalloc(GFP_ATOMIC)
and
preventing printk() from calling console_flush_all()
while zonelist_update_seq.seqcount is odd.
Since Petr Mladek thinks that __build_all_zonelists() can become a
candidate for deferring printk() [2], let's address this problem by
disabling local interrupts in order to avoid kmalloc(GFP_ATOMIC)
and
disabling synchronous printk() in order to avoid console_flush_all()
.
As a side effect of minimizing duration of zonelist_update_seq.seqcount
being odd by disabling synchronous printk(), latency at
read_seqbegin(&zonelist_update_seq) for both !__GFP_DIRECT_RECLAIM and
__GFP_DIRECT_RECLAIM allocation requests will be reduced. Although, from
lockdep perspective, not calling read_seqbegin(&zonelist_update_seq) (i.e.
do not record unnecessary locking dependency) from interrupt context is
still preferable, even if we don't allow calling kmalloc(GFP_ATOMIC)
inside
write_seqlock(&zonelist_update_seq)/write_sequnlock(&zonelist_update_seq)
section...
Link: https://lkml.kernel.org/r/8796b95c-3da3-5885-fddd-6ef55f30e4d3@I-love.SAKURA.ne.jp
Fixes: 3d36424b3b58 ("mm/page_alloc: fix race condition between build_all_zonelists and page allocation")
Link: https://lkml.kernel.org/r/ZCrs+1cDqPWTDFNM@alley [2]
Reported-by: syzbot <syzbot+223c7461c58c58a4cb10@syzkaller.appspotmail.com>
Link: https://syzkaller.appspot.com/bug?extid=223c7461c58c58a4cb10 [1]
Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Petr Mladek <pmladek@suse.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Ilpo Järvinen <ilpo.jarvinen@linux.intel.com>
Cc: John Ogness <john.ogness@linutronix.de>
Cc: Patrick Daly <quic_pdaly@quicinc.com>
Cc: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-04-04 23:31:58 +09:00
|
|
|
|
2023-06-23 22:15:17 +02:00
|
|
|
|
mm/page_alloc: fix potential deadlock on zonelist_update_seq seqlock
syzbot is reporting circular locking dependency which involves
zonelist_update_seq seqlock [1], for this lock is checked by memory
allocation requests which do not need to be retried.
One deadlock scenario is kmalloc(GFP_ATOMIC) from an interrupt handler.
CPU0
----
__build_all_zonelists() {
write_seqlock(&zonelist_update_seq); // makes zonelist_update_seq.seqcount odd
// e.g. timer interrupt handler runs at this moment
some_timer_func() {
kmalloc(GFP_ATOMIC) {
__alloc_pages_slowpath() {
read_seqbegin(&zonelist_update_seq) {
// spins forever because zonelist_update_seq.seqcount is odd
}
}
}
}
// e.g. timer interrupt handler finishes
write_sequnlock(&zonelist_update_seq); // makes zonelist_update_seq.seqcount even
}
This deadlock scenario can be easily eliminated by not calling
read_seqbegin(&zonelist_update_seq) from !__GFP_DIRECT_RECLAIM allocation
requests, for retry is applicable to only __GFP_DIRECT_RECLAIM allocation
requests. But Michal Hocko does not know whether we should go with this
approach.
Another deadlock scenario which syzbot is reporting is a race between
kmalloc(GFP_ATOMIC) from tty_insert_flip_string_and_push_buffer() with
port->lock held and printk() from __build_all_zonelists() with
zonelist_update_seq held.
CPU0 CPU1
---- ----
pty_write() {
tty_insert_flip_string_and_push_buffer() {
__build_all_zonelists() {
write_seqlock(&zonelist_update_seq);
build_zonelists() {
printk() {
vprintk() {
vprintk_default() {
vprintk_emit() {
console_unlock() {
console_flush_all() {
console_emit_next_record() {
con->write() = serial8250_console_write() {
spin_lock_irqsave(&port->lock, flags);
tty_insert_flip_string() {
tty_insert_flip_string_fixed_flag() {
__tty_buffer_request_room() {
tty_buffer_alloc() {
kmalloc(GFP_ATOMIC | __GFP_NOWARN) {
__alloc_pages_slowpath() {
zonelist_iter_begin() {
read_seqbegin(&zonelist_update_seq); // spins forever because zonelist_update_seq.seqcount is odd
spin_lock_irqsave(&port->lock, flags); // spins forever because port->lock is held
}
}
}
}
}
}
}
}
spin_unlock_irqrestore(&port->lock, flags);
// message is printed to console
spin_unlock_irqrestore(&port->lock, flags);
}
}
}
}
}
}
}
}
}
write_sequnlock(&zonelist_update_seq);
}
}
}
This deadlock scenario can be eliminated by
preventing interrupt context from calling kmalloc(GFP_ATOMIC)
and
preventing printk() from calling console_flush_all()
while zonelist_update_seq.seqcount is odd.
Since Petr Mladek thinks that __build_all_zonelists() can become a
candidate for deferring printk() [2], let's address this problem by
disabling local interrupts in order to avoid kmalloc(GFP_ATOMIC)
and
disabling synchronous printk() in order to avoid console_flush_all()
.
As a side effect of minimizing duration of zonelist_update_seq.seqcount
being odd by disabling synchronous printk(), latency at
read_seqbegin(&zonelist_update_seq) for both !__GFP_DIRECT_RECLAIM and
__GFP_DIRECT_RECLAIM allocation requests will be reduced. Although, from
lockdep perspective, not calling read_seqbegin(&zonelist_update_seq) (i.e.
do not record unnecessary locking dependency) from interrupt context is
still preferable, even if we don't allow calling kmalloc(GFP_ATOMIC)
inside
write_seqlock(&zonelist_update_seq)/write_sequnlock(&zonelist_update_seq)
section...
Link: https://lkml.kernel.org/r/8796b95c-3da3-5885-fddd-6ef55f30e4d3@I-love.SAKURA.ne.jp
Fixes: 3d36424b3b58 ("mm/page_alloc: fix race condition between build_all_zonelists and page allocation")
Link: https://lkml.kernel.org/r/ZCrs+1cDqPWTDFNM@alley [2]
Reported-by: syzbot <syzbot+223c7461c58c58a4cb10@syzkaller.appspotmail.com>
Link: https://syzkaller.appspot.com/bug?extid=223c7461c58c58a4cb10 [1]
Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Petr Mladek <pmladek@suse.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Ilpo Järvinen <ilpo.jarvinen@linux.intel.com>
Cc: John Ogness <john.ogness@linutronix.de>
Cc: Patrick Daly <quic_pdaly@quicinc.com>
Cc: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-04-04 23:31:58 +09:00
|
|
|
|
2023-06-23 22:15:17 +02:00
|
|
|
|
|
|
|
|
|
mm/page_alloc: fix potential deadlock on zonelist_update_seq seqlock
syzbot is reporting circular locking dependency which involves
zonelist_update_seq seqlock [1], for this lock is checked by memory
allocation requests which do not need to be retried.
One deadlock scenario is kmalloc(GFP_ATOMIC) from an interrupt handler.
CPU0
----
__build_all_zonelists() {
write_seqlock(&zonelist_update_seq); // makes zonelist_update_seq.seqcount odd
// e.g. timer interrupt handler runs at this moment
some_timer_func() {
kmalloc(GFP_ATOMIC) {
__alloc_pages_slowpath() {
read_seqbegin(&zonelist_update_seq) {
// spins forever because zonelist_update_seq.seqcount is odd
}
}
}
}
// e.g. timer interrupt handler finishes
write_sequnlock(&zonelist_update_seq); // makes zonelist_update_seq.seqcount even
}
This deadlock scenario can be easily eliminated by not calling
read_seqbegin(&zonelist_update_seq) from !__GFP_DIRECT_RECLAIM allocation
requests, for retry is applicable to only __GFP_DIRECT_RECLAIM allocation
requests. But Michal Hocko does not know whether we should go with this
approach.
Another deadlock scenario which syzbot is reporting is a race between
kmalloc(GFP_ATOMIC) from tty_insert_flip_string_and_push_buffer() with
port->lock held and printk() from __build_all_zonelists() with
zonelist_update_seq held.
CPU0 CPU1
---- ----
pty_write() {
tty_insert_flip_string_and_push_buffer() {
__build_all_zonelists() {
write_seqlock(&zonelist_update_seq);
build_zonelists() {
printk() {
vprintk() {
vprintk_default() {
vprintk_emit() {
console_unlock() {
console_flush_all() {
console_emit_next_record() {
con->write() = serial8250_console_write() {
spin_lock_irqsave(&port->lock, flags);
tty_insert_flip_string() {
tty_insert_flip_string_fixed_flag() {
__tty_buffer_request_room() {
tty_buffer_alloc() {
kmalloc(GFP_ATOMIC | __GFP_NOWARN) {
__alloc_pages_slowpath() {
zonelist_iter_begin() {
read_seqbegin(&zonelist_update_seq); // spins forever because zonelist_update_seq.seqcount is odd
spin_lock_irqsave(&port->lock, flags); // spins forever because port->lock is held
}
}
}
}
}
}
}
}
spin_unlock_irqrestore(&port->lock, flags);
// message is printed to console
spin_unlock_irqrestore(&port->lock, flags);
}
}
}
}
}
}
}
}
}
write_sequnlock(&zonelist_update_seq);
}
}
}
This deadlock scenario can be eliminated by
preventing interrupt context from calling kmalloc(GFP_ATOMIC)
and
preventing printk() from calling console_flush_all()
while zonelist_update_seq.seqcount is odd.
Since Petr Mladek thinks that __build_all_zonelists() can become a
candidate for deferring printk() [2], let's address this problem by
disabling local interrupts in order to avoid kmalloc(GFP_ATOMIC)
and
disabling synchronous printk() in order to avoid console_flush_all()
.
As a side effect of minimizing duration of zonelist_update_seq.seqcount
being odd by disabling synchronous printk(), latency at
read_seqbegin(&zonelist_update_seq) for both !__GFP_DIRECT_RECLAIM and
__GFP_DIRECT_RECLAIM allocation requests will be reduced. Although, from
lockdep perspective, not calling read_seqbegin(&zonelist_update_seq) (i.e.
do not record unnecessary locking dependency) from interrupt context is
still preferable, even if we don't allow calling kmalloc(GFP_ATOMIC)
inside
write_seqlock(&zonelist_update_seq)/write_sequnlock(&zonelist_update_seq)
section...
Link: https://lkml.kernel.org/r/8796b95c-3da3-5885-fddd-6ef55f30e4d3@I-love.SAKURA.ne.jp
Fixes: 3d36424b3b58 ("mm/page_alloc: fix race condition between build_all_zonelists and page allocation")
Link: https://lkml.kernel.org/r/ZCrs+1cDqPWTDFNM@alley [2]
Reported-by: syzbot <syzbot+223c7461c58c58a4cb10@syzkaller.appspotmail.com>
Link: https://syzkaller.appspot.com/bug?extid=223c7461c58c58a4cb10 [1]
Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Petr Mladek <pmladek@suse.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Ilpo Järvinen <ilpo.jarvinen@linux.intel.com>
Cc: John Ogness <john.ogness@linutronix.de>
Cc: Patrick Daly <quic_pdaly@quicinc.com>
Cc: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-04-04 23:31:58 +09:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
[PATCH] memory page_alloc zonelist caching speedup
Optimize the critical zonelist scanning for free pages in the kernel memory
allocator by caching the zones that were found to be full recently, and
skipping them.
Remembers the zones in a zonelist that were short of free memory in the
last second. And it stashes a zone-to-node table in the zonelist struct,
to optimize that conversion (minimize its cache footprint.)
Recent changes:
This differs in a significant way from a similar patch that I
posted a week ago. Now, instead of having a nodemask_t of
recently full nodes, I have a bitmask of recently full zones.
This solves a problem that last weeks patch had, which on
systems with multiple zones per node (such as DMA zone) would
take seeing any of these zones full as meaning that all zones
on that node were full.
Also I changed names - from "zonelist faster" to "zonelist cache",
as that seemed to better convey what we're doing here - caching
some of the key zonelist state (for faster access.)
See below for some performance benchmark results. After all that
discussion with David on why I didn't need them, I went and got
some ;). I wanted to verify that I had not hurt the normal case
of memory allocation noticeably. At least for my one little
microbenchmark, I found (1) the normal case wasn't affected, and
(2) workloads that forced scanning across multiple nodes for
memory improved up to 10% fewer System CPU cycles and lower
elapsed clock time ('sys' and 'real'). Good. See details, below.
I didn't have the logic in get_page_from_freelist() for various
full nodes and zone reclaim failures correct. That should be
fixed up now - notice the new goto labels zonelist_scan,
this_zone_full, and try_next_zone, in get_page_from_freelist().
There are two reasons I persued this alternative, over some earlier
proposals that would have focused on optimizing the fake numa
emulation case by caching the last useful zone:
1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems)
have seen real customer loads where the cost to scan the zonelist
was a problem, due to many nodes being full of memory before
we got to a node we could use. Or at least, I think we have.
This was related to me by another engineer, based on experiences
from some time past. So this is not guaranteed. Most likely, though.
The following approach should help such real numa systems just as
much as it helps fake numa systems, or any combination thereof.
2) The effort to distinguish fake from real numa, using node_distance,
so that we could cache a fake numa node and optimize choosing
it over equivalent distance fake nodes, while continuing to
properly scan all real nodes in distance order, was going to
require a nasty blob of zonelist and node distance munging.
The following approach has no new dependency on node distances or
zone sorting.
See comment in the patch below for a description of what it actually does.
Technical details of note (or controversy):
- See the use of "zlc_active" and "did_zlc_setup" below, to delay
adding any work for this new mechanism until we've looked at the
first zone in zonelist. I figured the odds of the first zone
having the memory we needed were high enough that we should just
look there, first, then get fancy only if we need to keep looking.
- Some odd hackery was needed to add items to struct zonelist, while
not tripping up the custom zonelists built by the mm/mempolicy.c
code for MPOL_BIND. My usual wordy comments below explain this.
Search for "MPOL_BIND".
- Some per-node data in the struct zonelist is now modified frequently,
with no locking. Multiple CPU cores on a node could hit and mangle
this data. The theory is that this is just performance hint data,
and the memory allocator will work just fine despite any such mangling.
The fields at risk are the struct 'zonelist_cache' fields 'fullzones'
(a bitmask) and 'last_full_zap' (unsigned long jiffies). It should
all be self correcting after at most a one second delay.
- This still does a linear scan of the same lengths as before. All
I've optimized is making the scan faster, not algorithmically
shorter. It is now able to scan a compact array of 'unsigned
short' in the case of many full nodes, so one cache line should
cover quite a few nodes, rather than each node hitting another
one or two new and distinct cache lines.
- If both Andi and Nick don't find this too complicated, I will be
(pleasantly) flabbergasted.
- I removed the comment claiming we only use one cachline's worth of
zonelist. We seem, at least in the fake numa case, to have put the
lie to that claim.
- I pay no attention to the various watermarks and such in this performance
hint. A node could be marked full for one watermark, and then skipped
over when searching for a page using a different watermark. I think
that's actually quite ok, as it will tend to slightly increase the
spreading of memory over other nodes, away from a memory stressed node.
===============
Performance - some benchmark results and analysis:
This benchmark runs a memory hog program that uses multiple
threads to touch alot of memory as quickly as it can.
Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of
the total 96 GBytes on the system, and using 1, 19, 37, or 55
threads (on a 56 CPU system.) System, user and real (elapsed)
timings were recorded for each run, shown in units of seconds,
in the table below.
Two kernels were tested - 2.6.18-mm3 and the same kernel with
this zonelist caching patch added. The table also shows the
percentage improvement the zonelist caching sys time is over
(lower than) the stock *-mm kernel.
number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent
GBs N ------------ -------------- ---------------- systime
mem threads sys user real sys user real sys user real better
12 1 153 24 177 151 24 176 -2 0 -1 1%
12 19 99 22 8 99 22 8 0 0 0 0%
12 37 111 25 6 112 25 6 1 0 0 -0%
12 55 115 25 5 110 23 5 -5 -2 0 4%
38 1 502 74 576 497 73 570 -5 -1 -6 0%
38 19 426 78 48 373 76 39 -53 -2 -9 12%
38 37 544 83 36 547 82 36 3 -1 0 -0%
38 55 501 77 23 511 80 24 10 3 1 -1%
64 1 917 125 1042 890 124 1014 -27 -1 -28 2%
64 19 1118 138 119 965 141 103 -153 3 -16 13%
64 37 1202 151 94 1136 150 81 -66 -1 -13 5%
64 55 1118 141 61 1072 140 58 -46 -1 -3 4%
90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4%
90 19 2392 199 192 2116 189 176 -276 -10 -16 11%
90 37 3313 238 175 2972 225 145 -341 -13 -30 10%
90 55 1948 210 104 1843 213 100 -105 3 -4 5%
Notes:
1) This test ran a memory hog program that started a specified number N of
threads, and had each thread allocate and touch 1/N'th of
the total memory to be used in the test run in a single loop,
writing a constant word to memory, one store every 4096 bytes.
Watching this test during some earlier trial runs, I would see
each of these threads sit down on one CPU and stay there, for
the remainder of the pass, a different CPU for each thread.
2) The 'real' column is not comparable to the 'sys' or 'user' columns.
The 'real' column is seconds wall clock time elapsed, from beginning
to end of that test pass. The 'sys' and 'user' columns are total
CPU seconds spent on that test pass. For a 19 thread test run,
for example, the sum of 'sys' and 'user' could be up to 19 times the
number of 'real' elapsed wall clock seconds.
3) Tests were run on a fresh, single-user boot, to minimize the amount
of memory already in use at the start of the test, and to minimize
the amount of background activity that might interfere.
4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM.
5) Notice that the 'real' time gets large for the single thread runs, even
though the measured 'sys' and 'user' times are modest. I'm not sure what
that means - probably something to do with it being slow for one thread to
be accessing memory along ways away. Perhaps the fake numa system, running
ostensibly the same workload, would not show this substantial degradation
of 'real' time for one thread on many nodes -- lets hope not.
6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs)
ran quite efficiently, as one might expect. Each pair of threads needed
to allocate and touch the memory on the node the two threads shared, a
pleasantly parallizable workload.
7) The intermediate thread count passes, when asking for alot of memory forcing
them to go to a few neighboring nodes, improved the most with this zonelist
caching patch.
Conclusions:
* This zonelist cache patch probably makes little difference one way or the
other for most workloads on real numa hardware, if those workloads avoid
heavy off node allocations.
* For memory intensive workloads requiring substantial off-node allocations
on real numa hardware, this patch improves both kernel and elapsed timings
up to ten per-cent.
* For fake numa systems, I'm optimistic, but will have to leave that up to
Rohit Seth to actually test (once I get him a 2.6.18 backport.)
Signed-off-by: Paul Jackson <pj@sgi.com>
Cc: Rohit Seth <rohitseth@google.com>
Cc: Christoph Lameter <clameter@engr.sgi.com>
Cc: David Rientjes <rientjes@cs.washington.edu>
Cc: Paul Menage <menage@google.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-06 20:31:48 -08:00
|
|
|
|
2009-08-18 14:11:19 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-07-31 16:43:28 -07:00
|
|
|
|
2017-09-06 16:19:33 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-07-31 16:43:28 -07:00
|
|
|
|
|
|
|
|
|
2017-09-06 16:19:33 -07:00
|
|
|
|
2022-03-22 14:46:54 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-09-06 16:19:33 -07:00
|
|
|
|
2012-07-31 16:43:28 -07:00
|
|
|
|
2017-09-06 16:19:33 -07:00
|
|
|
|
|
|
|
|
|
2010-01-05 15:34:51 +09:00
|
|
|
|
2017-09-06 16:20:17 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-09-06 16:20:20 -07:00
|
|
|
|
2017-09-06 16:20:17 -07:00
|
|
|
|
|
|
|
|
|
2017-09-06 16:20:20 -07:00
|
|
|
|
2017-09-06 16:20:37 -07:00
|
|
|
|
mm/page_alloc: fix potential deadlock on zonelist_update_seq seqlock
syzbot is reporting circular locking dependency which involves
zonelist_update_seq seqlock [1], for this lock is checked by memory
allocation requests which do not need to be retried.
One deadlock scenario is kmalloc(GFP_ATOMIC) from an interrupt handler.
CPU0
----
__build_all_zonelists() {
write_seqlock(&zonelist_update_seq); // makes zonelist_update_seq.seqcount odd
// e.g. timer interrupt handler runs at this moment
some_timer_func() {
kmalloc(GFP_ATOMIC) {
__alloc_pages_slowpath() {
read_seqbegin(&zonelist_update_seq) {
// spins forever because zonelist_update_seq.seqcount is odd
}
}
}
}
// e.g. timer interrupt handler finishes
write_sequnlock(&zonelist_update_seq); // makes zonelist_update_seq.seqcount even
}
This deadlock scenario can be easily eliminated by not calling
read_seqbegin(&zonelist_update_seq) from !__GFP_DIRECT_RECLAIM allocation
requests, for retry is applicable to only __GFP_DIRECT_RECLAIM allocation
requests. But Michal Hocko does not know whether we should go with this
approach.
Another deadlock scenario which syzbot is reporting is a race between
kmalloc(GFP_ATOMIC) from tty_insert_flip_string_and_push_buffer() with
port->lock held and printk() from __build_all_zonelists() with
zonelist_update_seq held.
CPU0 CPU1
---- ----
pty_write() {
tty_insert_flip_string_and_push_buffer() {
__build_all_zonelists() {
write_seqlock(&zonelist_update_seq);
build_zonelists() {
printk() {
vprintk() {
vprintk_default() {
vprintk_emit() {
console_unlock() {
console_flush_all() {
console_emit_next_record() {
con->write() = serial8250_console_write() {
spin_lock_irqsave(&port->lock, flags);
tty_insert_flip_string() {
tty_insert_flip_string_fixed_flag() {
__tty_buffer_request_room() {
tty_buffer_alloc() {
kmalloc(GFP_ATOMIC | __GFP_NOWARN) {
__alloc_pages_slowpath() {
zonelist_iter_begin() {
read_seqbegin(&zonelist_update_seq); // spins forever because zonelist_update_seq.seqcount is odd
spin_lock_irqsave(&port->lock, flags); // spins forever because port->lock is held
}
}
}
}
}
}
}
}
spin_unlock_irqrestore(&port->lock, flags);
// message is printed to console
spin_unlock_irqrestore(&port->lock, flags);
}
}
}
}
}
}
}
}
}
write_sequnlock(&zonelist_update_seq);
}
}
}
This deadlock scenario can be eliminated by
preventing interrupt context from calling kmalloc(GFP_ATOMIC)
and
preventing printk() from calling console_flush_all()
while zonelist_update_seq.seqcount is odd.
Since Petr Mladek thinks that __build_all_zonelists() can become a
candidate for deferring printk() [2], let's address this problem by
disabling local interrupts in order to avoid kmalloc(GFP_ATOMIC)
and
disabling synchronous printk() in order to avoid console_flush_all()
.
As a side effect of minimizing duration of zonelist_update_seq.seqcount
being odd by disabling synchronous printk(), latency at
read_seqbegin(&zonelist_update_seq) for both !__GFP_DIRECT_RECLAIM and
__GFP_DIRECT_RECLAIM allocation requests will be reduced. Although, from
lockdep perspective, not calling read_seqbegin(&zonelist_update_seq) (i.e.
do not record unnecessary locking dependency) from interrupt context is
still preferable, even if we don't allow calling kmalloc(GFP_ATOMIC)
inside
write_seqlock(&zonelist_update_seq)/write_sequnlock(&zonelist_update_seq)
section...
Link: https://lkml.kernel.org/r/8796b95c-3da3-5885-fddd-6ef55f30e4d3@I-love.SAKURA.ne.jp
Fixes: 3d36424b3b58 ("mm/page_alloc: fix race condition between build_all_zonelists and page allocation")
Link: https://lkml.kernel.org/r/ZCrs+1cDqPWTDFNM@alley [2]
Reported-by: syzbot <syzbot+223c7461c58c58a4cb10@syzkaller.appspotmail.com>
Link: https://syzkaller.appspot.com/bug?extid=223c7461c58c58a4cb10 [1]
Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Petr Mladek <pmladek@suse.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Ilpo Järvinen <ilpo.jarvinen@linux.intel.com>
Cc: John Ogness <john.ogness@linutronix.de>
Cc: Patrick Daly <quic_pdaly@quicinc.com>
Cc: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-04-04 23:31:58 +09:00
|
|
|
|
2023-06-23 22:15:17 +02:00
|
|
|
|
2017-09-06 16:20:17 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-01-05 15:34:51 +09:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-09-06 16:20:17 -07:00
|
|
|
|
mm/page_alloc: split per cpu page lists and zone stats
The PCP (per-cpu page allocator in page_alloc.c) shares locking
requirements with vmstat and the zone lock which is inconvenient and
causes some issues. For example, the PCP list and vmstat share the same
per-cpu space meaning that it's possible that vmstat updates dirty cache
lines holding per-cpu lists across CPUs unless padding is used. Second,
PREEMPT_RT does not want to disable IRQs for too long in the page
allocator.
This series splits the locking requirements and uses locks types more
suitable for PREEMPT_RT, reduces the time when special locking is required
for stats and reduces the time when IRQs need to be disabled on
!PREEMPT_RT kernels.
Why local_lock? PREEMPT_RT considers the following sequence to be unsafe
as documented in Documentation/locking/locktypes.rst
local_irq_disable();
spin_lock(&lock);
The pcp allocator has this sequence for rmqueue_pcplist (local_irq_save)
-> __rmqueue_pcplist -> rmqueue_bulk (spin_lock). While it's possible to
separate this out, it generally means there are points where we enable
IRQs and reenable them again immediately. To prevent a migration and the
per-cpu pointer going stale, migrate_disable is also needed. That is a
custom lock that is similar, but worse, than local_lock. Furthermore, on
PREEMPT_RT, it's undesirable to leave IRQs disabled for too long. By
converting to local_lock which disables migration on PREEMPT_RT, the
locking requirements can be separated and start moving the protections for
PCP, stats and the zone lock to PREEMPT_RT-safe equivalent locking. As a
bonus, local_lock also means that PROVE_LOCKING does something useful.
After that, it's obvious that zone_statistics incurs too much overhead and
leaves IRQs disabled for longer than necessary on !PREEMPT_RT kernels.
zone_statistics uses perfectly accurate counters requiring IRQs be
disabled for parallel RMW sequences when inaccurate ones like vm_events
would do. The series makes the NUMA statistics (NUMA_HIT and friends)
inaccurate counters that then require no special protection on
!PREEMPT_RT.
The bulk page allocator can then do stat updates in bulk with IRQs enabled
which should improve the efficiency. Technically, this could have been
done without the local_lock and vmstat conversion work and the order
simply reflects the timing of when different series were implemented.
Finally, there are places where we conflate IRQs being disabled for the
PCP with the IRQ-safe zone spinlock. The remainder of the series reduces
the scope of what is protected by disabled IRQs on !PREEMPT_RT kernels.
By the end of the series, page_alloc.c does not call local_irq_save so the
locking scope is a bit clearer. The one exception is that modifying
NR_FREE_PAGES still happens in places where it's known the IRQs are
disabled as it's harmless for PREEMPT_RT and would be expensive to split
the locking there.
No performance data is included because despite the overhead of the stats,
it's within the noise for most workloads on !PREEMPT_RT. However, Jesper
Dangaard Brouer ran a page allocation microbenchmark on a E5-1650 v4 @
3.60GHz CPU on the first version of this series. Focusing on the array
variant of the bulk page allocator reveals the following.
(CPU: Intel(R) Xeon(R) CPU E5-1650 v4 @ 3.60GHz)
ARRAY variant: time_bulk_page_alloc_free_array: step=bulk size
Baseline Patched
1 56.383 54.225 (+3.83%)
2 40.047 35.492 (+11.38%)
3 37.339 32.643 (+12.58%)
4 35.578 30.992 (+12.89%)
8 33.592 29.606 (+11.87%)
16 32.362 28.532 (+11.85%)
32 31.476 27.728 (+11.91%)
64 30.633 27.252 (+11.04%)
128 30.596 27.090 (+11.46%)
While this is a positive outcome, the series is more likely to be
interesting to the RT people in terms of getting parts of the PREEMPT_RT
tree into mainline.
This patch (of 9):
The per-cpu page allocator lists and the per-cpu vmstat deltas are stored
in the same struct per_cpu_pages even though vmstats have no direct impact
on the per-cpu page lists. This is inconsistent because the vmstats for a
node are stored on a dedicated structure. The bigger issue is that the
per_cpu_pages structure is not cache-aligned and stat updates either cache
conflict with adjacent per-cpu lists incurring a runtime cost or padding
is required incurring a memory cost.
This patch splits the per-cpu pagelists and the vmstat deltas into
separate structures. It's mostly a mechanical conversion but some
variable renaming is done to clearly distinguish the per-cpu pages
structure (pcp) from the vmstats (pzstats).
Superficially, this appears to increase the size of the per_cpu_pages
structure but the movement of expire fills a structure hole so there is no
impact overall.
[mgorman@techsingularity.net: make it W=1 cleaner]
Link: https://lkml.kernel.org/r/20210514144622.GA3735@techsingularity.net
[mgorman@techsingularity.net: make it W=1 even cleaner]
Link: https://lkml.kernel.org/r/20210516140705.GB3735@techsingularity.net
[lkp@intel.com: check struct per_cpu_zonestat has a non-zero size]
[vbabka@suse.cz: Init zone->per_cpu_zonestats properly]
Link: https://lkml.kernel.org/r/20210512095458.30632-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20210512095458.30632-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-28 19:41:38 -07:00
|
|
|
|
2010-01-05 15:34:51 +09:00
|
|
|
|
2015-02-12 15:00:06 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-05-24 14:32:52 -07:00
|
|
|
|
|
|
|
|
|
2015-02-12 15:00:06 -08:00
|
|
|
|
2017-09-06 16:20:24 -07:00
|
|
|
|
2015-02-12 15:00:06 -08:00
|
|
|
|
2010-05-24 14:32:52 -07:00
|
|
|
|
2017-09-06 16:20:24 -07:00
|
|
|
|
2006-06-23 02:03:11 -07:00
|
|
|
|
2020-08-06 23:25:27 -07:00
|
|
|
|
|
|
|
|
|
2006-06-23 02:03:11 -07:00
|
|
|
|
2015-02-12 15:00:06 -08:00
|
|
|
|
2006-06-23 02:03:11 -07:00
|
|
|
|
mm, page_alloc: remove stop_machine from build_all_zonelists
build_all_zonelists has been (ab)using stop_machine to make sure that
zonelists do not change while somebody is looking at them. This is is
just a gross hack because a) it complicates the context from which we
can call build_all_zonelists (see 3f906ba23689 ("mm/memory-hotplug:
switch locking to a percpu rwsem")) and b) is is not really necessary
especially after "mm, page_alloc: simplify zonelist initialization" and
c) it doesn't really provide the protection it claims (see below).
Updates of the zonelists happen very seldom, basically only when a zone
becomes populated during memory online or when it loses all the memory
during offline. A racing iteration over zonelists could either miss a
zone or try to work on one zone twice. Both of these are something we
can live with occasionally because there will always be at least one
zone visible so we are not likely to fail allocation too easily for
example.
Please note that the original stop_machine approach doesn't really
provide a better exclusion because the iteration might be interrupted
half way (unless the whole iteration is preempt disabled which is not
the case in most cases) so the some zones could still be seen twice or a
zone missed.
I have run the pathological online/offline of the single memblock in the
movable zone while stressing the same small node with some memory
pressure.
Node 1, zone DMA
pages free 0
min 0
low 0
high 0
spanned 0
present 0
managed 0
protection: (0, 943, 943, 943)
Node 1, zone DMA32
pages free 227310
min 8294
low 10367
high 12440
spanned 262112
present 262112
managed 241436
protection: (0, 0, 0, 0)
Node 1, zone Normal
pages free 0
min 0
low 0
high 0
spanned 0
present 0
managed 0
protection: (0, 0, 0, 1024)
Node 1, zone Movable
pages free 32722
min 85
low 117
high 149
spanned 32768
present 32768
managed 32768
protection: (0, 0, 0, 0)
root@test1:/sys/devices/system/node/node1# while true
do
echo offline > memory34/state
echo online_movable > memory34/state
done
root@test1:/mnt/data/test/linux-3.7-rc5# numactl --preferred=1 make -j4
and it survived without any unexpected behavior. While this is not
really a great testing coverage it should exercise the allocation path
quite a lot.
Link: http://lkml.kernel.org/r/20170721143915.14161-8-mhocko@kernel.org
Signed-off-by: Michal Hocko <mhocko@suse.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Joonsoo Kim <js1304@gmail.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Shaohua Li <shaohua.li@intel.com>
Cc: Toshi Kani <toshi.kani@hpe.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 16:20:34 -07:00
|
|
|
|
2006-06-23 02:03:11 -07:00
|
|
|
|
|
|
|
|
|
2020-08-06 23:25:30 -07:00
|
|
|
|
|
|
|
|
|
2007-10-16 01:25:54 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-10-16 01:26:01 -07:00
|
|
|
|
2007-10-16 01:25:54 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-03-05 15:48:29 -08:00
|
|
|
|
2016-03-17 14:19:47 -07:00
|
|
|
|
2024-10-21 11:13:40 +02:00
|
|
|
|
2016-03-17 14:19:47 -07:00
|
|
|
|
2007-07-15 23:38:01 -07:00
|
|
|
|
2014-12-10 15:42:53 -08:00
|
|
|
|
2007-07-15 23:38:01 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2014-06-23 13:22:04 -07:00
|
|
|
|
2005-06-21 17:14:47 -07:00
|
|
|
|
nommu: clamp zone_batchsize() to 0 under NOMMU conditions
Clamp zone_batchsize() to 0 under NOMMU conditions to stop
free_hot_cold_page() from queueing and batching frees.
The problem is that under NOMMU conditions it is really important to be
able to allocate large contiguous chunks of memory, but when munmap() or
exit_mmap() releases big stretches of memory, return of these to the buddy
allocator can be deferred, and when it does finally happen, it can be in
small chunks.
Whilst the fragmentation this incurs isn't so much of a problem under MMU
conditions as userspace VM is glued together from individual pages with
the aid of the MMU, it is a real problem if there isn't an MMU.
By clamping the page freeing queue size to 0, pages are returned to the
allocator immediately, and the buddy detector is more likely to be able to
glue them together into large chunks immediately, and fragmentation is
less likely to occur.
By disabling batching of frees, and by turning off the trimming of excess
space during boot, Coldfire can manage to boot.
Reported-by: Lanttor Guo <lanttor.guo@freescale.com>
Signed-off-by: David Howells <dhowells@redhat.com>
Tested-by: Lanttor Guo <lanttor.guo@freescale.com>
Cc: Greg Ungerer <gerg@snapgear.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-05-06 16:03:03 -07:00
|
|
|
|
2005-06-21 17:14:47 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-28 19:42:12 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-06-21 17:14:47 -07:00
|
|
|
|
2022-09-16 15:22:53 +08:00
|
|
|
|
2005-06-21 17:14:47 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-12-04 13:55:25 +11:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-06-21 17:14:47 -07:00
|
|
|
|
2005-12-04 13:55:25 +11:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-06-21 17:14:47 -07:00
|
|
|
|
2009-05-06 16:03:02 -07:00
|
|
|
|
2005-10-29 18:15:47 -07:00
|
|
|
|
2005-06-21 17:14:47 -07:00
|
|
|
|
nommu: clamp zone_batchsize() to 0 under NOMMU conditions
Clamp zone_batchsize() to 0 under NOMMU conditions to stop
free_hot_cold_page() from queueing and batching frees.
The problem is that under NOMMU conditions it is really important to be
able to allocate large contiguous chunks of memory, but when munmap() or
exit_mmap() releases big stretches of memory, return of these to the buddy
allocator can be deferred, and when it does finally happen, it can be in
small chunks.
Whilst the fragmentation this incurs isn't so much of a problem under MMU
conditions as userspace VM is glued together from individual pages with
the aid of the MMU, it is a real problem if there isn't an MMU.
By clamping the page freeing queue size to 0, pages are returned to the
allocator immediately, and the buddy detector is more likely to be able to
glue them together into large chunks immediately, and fragmentation is
less likely to occur.
By disabling batching of frees, and by turning off the trimming of excess
space during boot, Coldfire can manage to boot.
Reported-by: Lanttor Guo <lanttor.guo@freescale.com>
Signed-off-by: David Howells <dhowells@redhat.com>
Tested-by: Lanttor Guo <lanttor.guo@freescale.com>
Cc: Greg Ungerer <gerg@snapgear.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-05-06 16:03:03 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-06-21 17:14:47 -07:00
|
|
|
|
|
|
|
|
|
2023-05-16 14:38:20 +08:00
|
|
|
|
2023-10-16 13:29:59 +08:00
|
|
|
|
|
|
|
|
|
2021-06-28 19:42:12 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-28 19:43:11 -07:00
|
|
|
|
2021-06-28 19:42:24 -07:00
|
|
|
|
|
|
|
|
|
2023-10-16 13:29:59 +08:00
|
|
|
|
2021-06-28 19:42:24 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-16 13:29:59 +08:00
|
|
|
|
2021-06-28 19:42:24 -07:00
|
|
|
|
2021-06-28 19:42:12 -07:00
|
|
|
|
|
|
|
|
|
2021-06-28 19:42:24 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-16 13:29:59 +08:00
|
|
|
|
|
|
|
|
|
2021-06-28 19:43:11 -07:00
|
|
|
|
2021-06-28 19:42:12 -07:00
|
|
|
|
2021-06-28 19:43:11 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-28 19:42:12 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-07-03 15:01:31 -07:00
|
|
|
|
2020-12-14 19:10:50 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-07-03 15:01:31 -07:00
|
|
|
|
2020-12-14 19:10:50 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-16 13:29:59 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-07-03 15:01:31 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-16 13:29:59 +08:00
|
|
|
|
|
|
|
|
|
2013-07-03 15:01:31 -07:00
|
|
|
|
2020-12-14 19:10:50 -08:00
|
|
|
|
2023-10-16 13:29:59 +08:00
|
|
|
|
|
|
|
|
|
2013-07-03 15:01:31 -07:00
|
|
|
|
|
|
|
|
|
mm/page_alloc: split per cpu page lists and zone stats
The PCP (per-cpu page allocator in page_alloc.c) shares locking
requirements with vmstat and the zone lock which is inconvenient and
causes some issues. For example, the PCP list and vmstat share the same
per-cpu space meaning that it's possible that vmstat updates dirty cache
lines holding per-cpu lists across CPUs unless padding is used. Second,
PREEMPT_RT does not want to disable IRQs for too long in the page
allocator.
This series splits the locking requirements and uses locks types more
suitable for PREEMPT_RT, reduces the time when special locking is required
for stats and reduces the time when IRQs need to be disabled on
!PREEMPT_RT kernels.
Why local_lock? PREEMPT_RT considers the following sequence to be unsafe
as documented in Documentation/locking/locktypes.rst
local_irq_disable();
spin_lock(&lock);
The pcp allocator has this sequence for rmqueue_pcplist (local_irq_save)
-> __rmqueue_pcplist -> rmqueue_bulk (spin_lock). While it's possible to
separate this out, it generally means there are points where we enable
IRQs and reenable them again immediately. To prevent a migration and the
per-cpu pointer going stale, migrate_disable is also needed. That is a
custom lock that is similar, but worse, than local_lock. Furthermore, on
PREEMPT_RT, it's undesirable to leave IRQs disabled for too long. By
converting to local_lock which disables migration on PREEMPT_RT, the
locking requirements can be separated and start moving the protections for
PCP, stats and the zone lock to PREEMPT_RT-safe equivalent locking. As a
bonus, local_lock also means that PROVE_LOCKING does something useful.
After that, it's obvious that zone_statistics incurs too much overhead and
leaves IRQs disabled for longer than necessary on !PREEMPT_RT kernels.
zone_statistics uses perfectly accurate counters requiring IRQs be
disabled for parallel RMW sequences when inaccurate ones like vm_events
would do. The series makes the NUMA statistics (NUMA_HIT and friends)
inaccurate counters that then require no special protection on
!PREEMPT_RT.
The bulk page allocator can then do stat updates in bulk with IRQs enabled
which should improve the efficiency. Technically, this could have been
done without the local_lock and vmstat conversion work and the order
simply reflects the timing of when different series were implemented.
Finally, there are places where we conflate IRQs being disabled for the
PCP with the IRQ-safe zone spinlock. The remainder of the series reduces
the scope of what is protected by disabled IRQs on !PREEMPT_RT kernels.
By the end of the series, page_alloc.c does not call local_irq_save so the
locking scope is a bit clearer. The one exception is that modifying
NR_FREE_PAGES still happens in places where it's known the IRQs are
disabled as it's harmless for PREEMPT_RT and would be expensive to split
the locking there.
No performance data is included because despite the overhead of the stats,
it's within the noise for most workloads on !PREEMPT_RT. However, Jesper
Dangaard Brouer ran a page allocation microbenchmark on a E5-1650 v4 @
3.60GHz CPU on the first version of this series. Focusing on the array
variant of the bulk page allocator reveals the following.
(CPU: Intel(R) Xeon(R) CPU E5-1650 v4 @ 3.60GHz)
ARRAY variant: time_bulk_page_alloc_free_array: step=bulk size
Baseline Patched
1 56.383 54.225 (+3.83%)
2 40.047 35.492 (+11.38%)
3 37.339 32.643 (+12.58%)
4 35.578 30.992 (+12.89%)
8 33.592 29.606 (+11.87%)
16 32.362 28.532 (+11.85%)
32 31.476 27.728 (+11.91%)
64 30.633 27.252 (+11.04%)
128 30.596 27.090 (+11.46%)
While this is a positive outcome, the series is more likely to be
interesting to the RT people in terms of getting parts of the PREEMPT_RT
tree into mainline.
This patch (of 9):
The per-cpu page allocator lists and the per-cpu vmstat deltas are stored
in the same struct per_cpu_pages even though vmstats have no direct impact
on the per-cpu page lists. This is inconsistent because the vmstats for a
node are stored on a dedicated structure. The bigger issue is that the
per_cpu_pages structure is not cache-aligned and stat updates either cache
conflict with adjacent per-cpu lists incurring a runtime cost or padding
is required incurring a memory cost.
This patch splits the per-cpu pagelists and the vmstat deltas into
separate structures. It's mostly a mechanical conversion but some
variable renaming is done to clearly distinguish the per-cpu pages
structure (pcp) from the vmstats (pzstats).
Superficially, this appears to increase the size of the per_cpu_pages
structure but the movement of expire fills a structure hole so there is no
impact overall.
[mgorman@techsingularity.net: make it W=1 cleaner]
Link: https://lkml.kernel.org/r/20210514144622.GA3735@techsingularity.net
[mgorman@techsingularity.net: make it W=1 even cleaner]
Link: https://lkml.kernel.org/r/20210516140705.GB3735@techsingularity.net
[lkp@intel.com: check struct per_cpu_zonestat has a non-zero size]
[vbabka@suse.cz: Init zone->per_cpu_zonestats properly]
Link: https://lkml.kernel.org/r/20210512095458.30632-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20210512095458.30632-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-28 19:41:38 -07:00
|
|
|
|
2005-06-21 17:15:00 -07:00
|
|
|
|
2021-06-28 19:43:08 -07:00
|
|
|
|
2005-06-21 17:15:00 -07:00
|
|
|
|
mm/page_alloc: split per cpu page lists and zone stats
The PCP (per-cpu page allocator in page_alloc.c) shares locking
requirements with vmstat and the zone lock which is inconvenient and
causes some issues. For example, the PCP list and vmstat share the same
per-cpu space meaning that it's possible that vmstat updates dirty cache
lines holding per-cpu lists across CPUs unless padding is used. Second,
PREEMPT_RT does not want to disable IRQs for too long in the page
allocator.
This series splits the locking requirements and uses locks types more
suitable for PREEMPT_RT, reduces the time when special locking is required
for stats and reduces the time when IRQs need to be disabled on
!PREEMPT_RT kernels.
Why local_lock? PREEMPT_RT considers the following sequence to be unsafe
as documented in Documentation/locking/locktypes.rst
local_irq_disable();
spin_lock(&lock);
The pcp allocator has this sequence for rmqueue_pcplist (local_irq_save)
-> __rmqueue_pcplist -> rmqueue_bulk (spin_lock). While it's possible to
separate this out, it generally means there are points where we enable
IRQs and reenable them again immediately. To prevent a migration and the
per-cpu pointer going stale, migrate_disable is also needed. That is a
custom lock that is similar, but worse, than local_lock. Furthermore, on
PREEMPT_RT, it's undesirable to leave IRQs disabled for too long. By
converting to local_lock which disables migration on PREEMPT_RT, the
locking requirements can be separated and start moving the protections for
PCP, stats and the zone lock to PREEMPT_RT-safe equivalent locking. As a
bonus, local_lock also means that PROVE_LOCKING does something useful.
After that, it's obvious that zone_statistics incurs too much overhead and
leaves IRQs disabled for longer than necessary on !PREEMPT_RT kernels.
zone_statistics uses perfectly accurate counters requiring IRQs be
disabled for parallel RMW sequences when inaccurate ones like vm_events
would do. The series makes the NUMA statistics (NUMA_HIT and friends)
inaccurate counters that then require no special protection on
!PREEMPT_RT.
The bulk page allocator can then do stat updates in bulk with IRQs enabled
which should improve the efficiency. Technically, this could have been
done without the local_lock and vmstat conversion work and the order
simply reflects the timing of when different series were implemented.
Finally, there are places where we conflate IRQs being disabled for the
PCP with the IRQ-safe zone spinlock. The remainder of the series reduces
the scope of what is protected by disabled IRQs on !PREEMPT_RT kernels.
By the end of the series, page_alloc.c does not call local_irq_save so the
locking scope is a bit clearer. The one exception is that modifying
NR_FREE_PAGES still happens in places where it's known the IRQs are
disabled as it's harmless for PREEMPT_RT and would be expensive to split
the locking there.
No performance data is included because despite the overhead of the stats,
it's within the noise for most workloads on !PREEMPT_RT. However, Jesper
Dangaard Brouer ran a page allocation microbenchmark on a E5-1650 v4 @
3.60GHz CPU on the first version of this series. Focusing on the array
variant of the bulk page allocator reveals the following.
(CPU: Intel(R) Xeon(R) CPU E5-1650 v4 @ 3.60GHz)
ARRAY variant: time_bulk_page_alloc_free_array: step=bulk size
Baseline Patched
1 56.383 54.225 (+3.83%)
2 40.047 35.492 (+11.38%)
3 37.339 32.643 (+12.58%)
4 35.578 30.992 (+12.89%)
8 33.592 29.606 (+11.87%)
16 32.362 28.532 (+11.85%)
32 31.476 27.728 (+11.91%)
64 30.633 27.252 (+11.04%)
128 30.596 27.090 (+11.46%)
While this is a positive outcome, the series is more likely to be
interesting to the RT people in terms of getting parts of the PREEMPT_RT
tree into mainline.
This patch (of 9):
The per-cpu page allocator lists and the per-cpu vmstat deltas are stored
in the same struct per_cpu_pages even though vmstats have no direct impact
on the per-cpu page lists. This is inconsistent because the vmstats for a
node are stored on a dedicated structure. The bigger issue is that the
per_cpu_pages structure is not cache-aligned and stat updates either cache
conflict with adjacent per-cpu lists incurring a runtime cost or padding
is required incurring a memory cost.
This patch splits the per-cpu pagelists and the vmstat deltas into
separate structures. It's mostly a mechanical conversion but some
variable renaming is done to clearly distinguish the per-cpu pages
structure (pcp) from the vmstats (pzstats).
Superficially, this appears to increase the size of the per_cpu_pages
structure but the movement of expire fills a structure hole so there is no
impact overall.
[mgorman@techsingularity.net: make it W=1 cleaner]
Link: https://lkml.kernel.org/r/20210514144622.GA3735@techsingularity.net
[mgorman@techsingularity.net: make it W=1 even cleaner]
Link: https://lkml.kernel.org/r/20210516140705.GB3735@techsingularity.net
[lkp@intel.com: check struct per_cpu_zonestat has a non-zero size]
[vbabka@suse.cz: Init zone->per_cpu_zonestats properly]
Link: https://lkml.kernel.org/r/20210512095458.30632-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20210512095458.30632-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-28 19:41:38 -07:00
|
|
|
|
|
|
|
|
|
2005-10-26 01:58:59 -07:00
|
|
|
|
2022-06-24 13:54:21 +01:00
|
|
|
|
2021-06-28 19:43:08 -07:00
|
|
|
|
|
|
|
|
|
2005-06-21 17:15:00 -07:00
|
|
|
|
2020-12-14 19:10:47 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-16 13:29:59 +08:00
|
|
|
|
|
|
|
|
|
2020-12-14 19:10:53 -08:00
|
|
|
|
2013-07-03 15:01:35 -07:00
|
|
|
|
|
|
|
|
|
2023-10-16 13:29:59 +08:00
|
|
|
|
|
|
|
|
|
mm, page_alloc: disable pcplists during memory offline
Memory offlining relies on page isolation to guarantee a forward progress
because pages cannot be reused while they are isolated. But the page
isolation itself doesn't prevent from races while freed pages are stored
on pcp lists and thus can be reused. This can be worked around by
repeated draining of pcplists, as done by commit 968318261221
("mm/memory_hotplug: drain per-cpu pages again during memory offline").
David and Michal would prefer that this race was closed in a way that
callers of page isolation who need stronger guarantees don't need to
repeatedly drain. David suggested disabling pcplists usage completely
during page isolation, instead of repeatedly draining them.
To achieve this without adding special cases in alloc/free fastpath, we
can use the same approach as boot pagesets - when pcp->high is 0, any
pcplist addition will be immediately flushed.
The race can thus be closed by setting pcp->high to 0 and draining
pcplists once, before calling start_isolate_page_range(). The draining
will serialize after processes that already disabled interrupts and read
the old value of pcp->high in free_unref_page_commit(), and processes that
have not yet disabled interrupts, will observe pcp->high == 0 when they
are rescheduled, and skip pcplists. This guarantees no stray pages on
pcplists in zones where isolation happens.
This patch thus adds zone_pcp_disable() and zone_pcp_enable() functions
that page isolation users can call before start_isolate_page_range() and
after unisolating (or offlining) the isolated pages.
Also, drain_all_pages() is optimized to only execute on cpus where
pcplists are not empty. The check can however race with a free to pcplist
that has not yet increased the pcp->count from 0 to 1. Thus make the
drain optionally skip the racy check and drain on all cpus, and use this
option in zone_pcp_disable().
As we have to avoid external updates to high and batch while pcplists are
disabled, we take pcp_batch_high_lock in zone_pcp_disable() and release it
in zone_pcp_enable(). This also synchronizes multiple users of
zone_pcp_disable()/enable().
Currently the only user of this functionality is offline_pages().
[vbabka@suse.cz: add comment, per David]
Link: https://lkml.kernel.org/r/527480ef-ed72-e1c1-52a0-1c5b0113df45@suse.cz
Link: https://lkml.kernel.org/r/20201111092812.11329-8-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Suggested-by: David Hildenbrand <david@redhat.com>
Suggested-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: David Hildenbrand <david@redhat.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-14 19:10:59 -08:00
|
|
|
|
mm/page_alloc: split per cpu page lists and zone stats
The PCP (per-cpu page allocator in page_alloc.c) shares locking
requirements with vmstat and the zone lock which is inconvenient and
causes some issues. For example, the PCP list and vmstat share the same
per-cpu space meaning that it's possible that vmstat updates dirty cache
lines holding per-cpu lists across CPUs unless padding is used. Second,
PREEMPT_RT does not want to disable IRQs for too long in the page
allocator.
This series splits the locking requirements and uses locks types more
suitable for PREEMPT_RT, reduces the time when special locking is required
for stats and reduces the time when IRQs need to be disabled on
!PREEMPT_RT kernels.
Why local_lock? PREEMPT_RT considers the following sequence to be unsafe
as documented in Documentation/locking/locktypes.rst
local_irq_disable();
spin_lock(&lock);
The pcp allocator has this sequence for rmqueue_pcplist (local_irq_save)
-> __rmqueue_pcplist -> rmqueue_bulk (spin_lock). While it's possible to
separate this out, it generally means there are points where we enable
IRQs and reenable them again immediately. To prevent a migration and the
per-cpu pointer going stale, migrate_disable is also needed. That is a
custom lock that is similar, but worse, than local_lock. Furthermore, on
PREEMPT_RT, it's undesirable to leave IRQs disabled for too long. By
converting to local_lock which disables migration on PREEMPT_RT, the
locking requirements can be separated and start moving the protections for
PCP, stats and the zone lock to PREEMPT_RT-safe equivalent locking. As a
bonus, local_lock also means that PROVE_LOCKING does something useful.
After that, it's obvious that zone_statistics incurs too much overhead and
leaves IRQs disabled for longer than necessary on !PREEMPT_RT kernels.
zone_statistics uses perfectly accurate counters requiring IRQs be
disabled for parallel RMW sequences when inaccurate ones like vm_events
would do. The series makes the NUMA statistics (NUMA_HIT and friends)
inaccurate counters that then require no special protection on
!PREEMPT_RT.
The bulk page allocator can then do stat updates in bulk with IRQs enabled
which should improve the efficiency. Technically, this could have been
done without the local_lock and vmstat conversion work and the order
simply reflects the timing of when different series were implemented.
Finally, there are places where we conflate IRQs being disabled for the
PCP with the IRQ-safe zone spinlock. The remainder of the series reduces
the scope of what is protected by disabled IRQs on !PREEMPT_RT kernels.
By the end of the series, page_alloc.c does not call local_irq_save so the
locking scope is a bit clearer. The one exception is that modifying
NR_FREE_PAGES still happens in places where it's known the IRQs are
disabled as it's harmless for PREEMPT_RT and would be expensive to split
the locking there.
No performance data is included because despite the overhead of the stats,
it's within the noise for most workloads on !PREEMPT_RT. However, Jesper
Dangaard Brouer ran a page allocation microbenchmark on a E5-1650 v4 @
3.60GHz CPU on the first version of this series. Focusing on the array
variant of the bulk page allocator reveals the following.
(CPU: Intel(R) Xeon(R) CPU E5-1650 v4 @ 3.60GHz)
ARRAY variant: time_bulk_page_alloc_free_array: step=bulk size
Baseline Patched
1 56.383 54.225 (+3.83%)
2 40.047 35.492 (+11.38%)
3 37.339 32.643 (+12.58%)
4 35.578 30.992 (+12.89%)
8 33.592 29.606 (+11.87%)
16 32.362 28.532 (+11.85%)
32 31.476 27.728 (+11.91%)
64 30.633 27.252 (+11.04%)
128 30.596 27.090 (+11.46%)
While this is a positive outcome, the series is more likely to be
interesting to the RT people in terms of getting parts of the PREEMPT_RT
tree into mainline.
This patch (of 9):
The per-cpu page allocator lists and the per-cpu vmstat deltas are stored
in the same struct per_cpu_pages even though vmstats have no direct impact
on the per-cpu page lists. This is inconsistent because the vmstats for a
node are stored on a dedicated structure. The bigger issue is that the
per_cpu_pages structure is not cache-aligned and stat updates either cache
conflict with adjacent per-cpu lists incurring a runtime cost or padding
is required incurring a memory cost.
This patch splits the per-cpu pagelists and the vmstat deltas into
separate structures. It's mostly a mechanical conversion but some
variable renaming is done to clearly distinguish the per-cpu pages
structure (pcp) from the vmstats (pzstats).
Superficially, this appears to increase the size of the per_cpu_pages
structure but the movement of expire fills a structure hole so there is no
impact overall.
[mgorman@techsingularity.net: make it W=1 cleaner]
Link: https://lkml.kernel.org/r/20210514144622.GA3735@techsingularity.net
[mgorman@techsingularity.net: make it W=1 even cleaner]
Link: https://lkml.kernel.org/r/20210516140705.GB3735@techsingularity.net
[lkp@intel.com: check struct per_cpu_zonestat has a non-zero size]
[vbabka@suse.cz: Init zone->per_cpu_zonestats properly]
Link: https://lkml.kernel.org/r/20210512095458.30632-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20210512095458.30632-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-28 19:41:38 -07:00
|
|
|
|
mm, page_alloc: disable pcplists during memory offline
Memory offlining relies on page isolation to guarantee a forward progress
because pages cannot be reused while they are isolated. But the page
isolation itself doesn't prevent from races while freed pages are stored
on pcp lists and thus can be reused. This can be worked around by
repeated draining of pcplists, as done by commit 968318261221
("mm/memory_hotplug: drain per-cpu pages again during memory offline").
David and Michal would prefer that this race was closed in a way that
callers of page isolation who need stronger guarantees don't need to
repeatedly drain. David suggested disabling pcplists usage completely
during page isolation, instead of repeatedly draining them.
To achieve this without adding special cases in alloc/free fastpath, we
can use the same approach as boot pagesets - when pcp->high is 0, any
pcplist addition will be immediately flushed.
The race can thus be closed by setting pcp->high to 0 and draining
pcplists once, before calling start_isolate_page_range(). The draining
will serialize after processes that already disabled interrupts and read
the old value of pcp->high in free_unref_page_commit(), and processes that
have not yet disabled interrupts, will observe pcp->high == 0 when they
are rescheduled, and skip pcplists. This guarantees no stray pages on
pcplists in zones where isolation happens.
This patch thus adds zone_pcp_disable() and zone_pcp_enable() functions
that page isolation users can call before start_isolate_page_range() and
after unisolating (or offlining) the isolated pages.
Also, drain_all_pages() is optimized to only execute on cpus where
pcplists are not empty. The check can however race with a free to pcplist
that has not yet increased the pcp->count from 0 to 1. Thus make the
drain optionally skip the racy check and drain on all cpus, and use this
option in zone_pcp_disable().
As we have to avoid external updates to high and batch while pcplists are
disabled, we take pcp_batch_high_lock in zone_pcp_disable() and release it
in zone_pcp_enable(). This also synchronizes multiple users of
zone_pcp_disable()/enable().
Currently the only user of this functionality is offline_pages().
[vbabka@suse.cz: add comment, per David]
Link: https://lkml.kernel.org/r/527480ef-ed72-e1c1-52a0-1c5b0113df45@suse.cz
Link: https://lkml.kernel.org/r/20201111092812.11329-8-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Suggested-by: David Hildenbrand <david@redhat.com>
Suggested-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: David Hildenbrand <david@redhat.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-14 19:10:59 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/page_alloc: split per cpu page lists and zone stats
The PCP (per-cpu page allocator in page_alloc.c) shares locking
requirements with vmstat and the zone lock which is inconvenient and
causes some issues. For example, the PCP list and vmstat share the same
per-cpu space meaning that it's possible that vmstat updates dirty cache
lines holding per-cpu lists across CPUs unless padding is used. Second,
PREEMPT_RT does not want to disable IRQs for too long in the page
allocator.
This series splits the locking requirements and uses locks types more
suitable for PREEMPT_RT, reduces the time when special locking is required
for stats and reduces the time when IRQs need to be disabled on
!PREEMPT_RT kernels.
Why local_lock? PREEMPT_RT considers the following sequence to be unsafe
as documented in Documentation/locking/locktypes.rst
local_irq_disable();
spin_lock(&lock);
The pcp allocator has this sequence for rmqueue_pcplist (local_irq_save)
-> __rmqueue_pcplist -> rmqueue_bulk (spin_lock). While it's possible to
separate this out, it generally means there are points where we enable
IRQs and reenable them again immediately. To prevent a migration and the
per-cpu pointer going stale, migrate_disable is also needed. That is a
custom lock that is similar, but worse, than local_lock. Furthermore, on
PREEMPT_RT, it's undesirable to leave IRQs disabled for too long. By
converting to local_lock which disables migration on PREEMPT_RT, the
locking requirements can be separated and start moving the protections for
PCP, stats and the zone lock to PREEMPT_RT-safe equivalent locking. As a
bonus, local_lock also means that PROVE_LOCKING does something useful.
After that, it's obvious that zone_statistics incurs too much overhead and
leaves IRQs disabled for longer than necessary on !PREEMPT_RT kernels.
zone_statistics uses perfectly accurate counters requiring IRQs be
disabled for parallel RMW sequences when inaccurate ones like vm_events
would do. The series makes the NUMA statistics (NUMA_HIT and friends)
inaccurate counters that then require no special protection on
!PREEMPT_RT.
The bulk page allocator can then do stat updates in bulk with IRQs enabled
which should improve the efficiency. Technically, this could have been
done without the local_lock and vmstat conversion work and the order
simply reflects the timing of when different series were implemented.
Finally, there are places where we conflate IRQs being disabled for the
PCP with the IRQ-safe zone spinlock. The remainder of the series reduces
the scope of what is protected by disabled IRQs on !PREEMPT_RT kernels.
By the end of the series, page_alloc.c does not call local_irq_save so the
locking scope is a bit clearer. The one exception is that modifying
NR_FREE_PAGES still happens in places where it's known the IRQs are
disabled as it's harmless for PREEMPT_RT and would be expensive to split
the locking there.
No performance data is included because despite the overhead of the stats,
it's within the noise for most workloads on !PREEMPT_RT. However, Jesper
Dangaard Brouer ran a page allocation microbenchmark on a E5-1650 v4 @
3.60GHz CPU on the first version of this series. Focusing on the array
variant of the bulk page allocator reveals the following.
(CPU: Intel(R) Xeon(R) CPU E5-1650 v4 @ 3.60GHz)
ARRAY variant: time_bulk_page_alloc_free_array: step=bulk size
Baseline Patched
1 56.383 54.225 (+3.83%)
2 40.047 35.492 (+11.38%)
3 37.339 32.643 (+12.58%)
4 35.578 30.992 (+12.89%)
8 33.592 29.606 (+11.87%)
16 32.362 28.532 (+11.85%)
32 31.476 27.728 (+11.91%)
64 30.633 27.252 (+11.04%)
128 30.596 27.090 (+11.46%)
While this is a positive outcome, the series is more likely to be
interesting to the RT people in terms of getting parts of the PREEMPT_RT
tree into mainline.
This patch (of 9):
The per-cpu page allocator lists and the per-cpu vmstat deltas are stored
in the same struct per_cpu_pages even though vmstats have no direct impact
on the per-cpu page lists. This is inconsistent because the vmstats for a
node are stored on a dedicated structure. The bigger issue is that the
per_cpu_pages structure is not cache-aligned and stat updates either cache
conflict with adjacent per-cpu lists incurring a runtime cost or padding
is required incurring a memory cost.
This patch splits the per-cpu pagelists and the vmstat deltas into
separate structures. It's mostly a mechanical conversion but some
variable renaming is done to clearly distinguish the per-cpu pages
structure (pcp) from the vmstats (pzstats).
Superficially, this appears to increase the size of the per_cpu_pages
structure but the movement of expire fills a structure hole so there is no
impact overall.
[mgorman@techsingularity.net: make it W=1 cleaner]
Link: https://lkml.kernel.org/r/20210514144622.GA3735@techsingularity.net
[mgorman@techsingularity.net: make it W=1 even cleaner]
Link: https://lkml.kernel.org/r/20210516140705.GB3735@techsingularity.net
[lkp@intel.com: check struct per_cpu_zonestat has a non-zero size]
[vbabka@suse.cz: Init zone->per_cpu_zonestats properly]
Link: https://lkml.kernel.org/r/20210512095458.30632-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20210512095458.30632-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-28 19:41:38 -07:00
|
|
|
|
2023-10-16 13:29:59 +08:00
|
|
|
|
mm, page_alloc: disable pcplists during memory offline
Memory offlining relies on page isolation to guarantee a forward progress
because pages cannot be reused while they are isolated. But the page
isolation itself doesn't prevent from races while freed pages are stored
on pcp lists and thus can be reused. This can be worked around by
repeated draining of pcplists, as done by commit 968318261221
("mm/memory_hotplug: drain per-cpu pages again during memory offline").
David and Michal would prefer that this race was closed in a way that
callers of page isolation who need stronger guarantees don't need to
repeatedly drain. David suggested disabling pcplists usage completely
during page isolation, instead of repeatedly draining them.
To achieve this without adding special cases in alloc/free fastpath, we
can use the same approach as boot pagesets - when pcp->high is 0, any
pcplist addition will be immediately flushed.
The race can thus be closed by setting pcp->high to 0 and draining
pcplists once, before calling start_isolate_page_range(). The draining
will serialize after processes that already disabled interrupts and read
the old value of pcp->high in free_unref_page_commit(), and processes that
have not yet disabled interrupts, will observe pcp->high == 0 when they
are rescheduled, and skip pcplists. This guarantees no stray pages on
pcplists in zones where isolation happens.
This patch thus adds zone_pcp_disable() and zone_pcp_enable() functions
that page isolation users can call before start_isolate_page_range() and
after unisolating (or offlining) the isolated pages.
Also, drain_all_pages() is optimized to only execute on cpus where
pcplists are not empty. The check can however race with a free to pcplist
that has not yet increased the pcp->count from 0 to 1. Thus make the
drain optionally skip the racy check and drain on all cpus, and use this
option in zone_pcp_disable().
As we have to avoid external updates to high and batch while pcplists are
disabled, we take pcp_batch_high_lock in zone_pcp_disable() and release it
in zone_pcp_enable(). This also synchronizes multiple users of
zone_pcp_disable()/enable().
Currently the only user of this functionality is offline_pages().
[vbabka@suse.cz: add comment, per David]
Link: https://lkml.kernel.org/r/527480ef-ed72-e1c1-52a0-1c5b0113df45@suse.cz
Link: https://lkml.kernel.org/r/20201111092812.11329-8-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Suggested-by: David Hildenbrand <david@redhat.com>
Suggested-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: David Hildenbrand <david@redhat.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-14 19:10:59 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-01-08 01:00:40 -08:00
|
|
|
|
2020-12-14 19:10:43 -08:00
|
|
|
|
2021-06-28 19:42:09 -07:00
|
|
|
|
2006-01-08 01:00:40 -08:00
|
|
|
|
2021-06-28 19:42:15 -07:00
|
|
|
|
2013-07-03 15:01:38 -07:00
|
|
|
|
2023-10-16 13:29:59 +08:00
|
|
|
|
mm, page_alloc: clean up pageset high and batch update
Patch series "disable pcplists during memory offline", v3.
As per the discussions [1] [2] this is an attempt to implement David's
suggestion that page isolation should disable pcplists to avoid races with
page freeing in progress. This is done without extra checks in fast
paths, as explained in Patch 9. The repeated draining done by [2] is then
no longer needed. Previous version (RFC) is at [3].
The RFC tried to hide pcplists disabling/enabling into page isolation, but
it wasn't completely possible, as memory offline does not unisolation.
Michal suggested an explicit API in [4] so that's the current
implementation and it seems indeed nicer.
Once we accept that page isolation users need to do explicit actions
around it depending on the needed guarantees, we can also IMHO accept that
the current pcplist draining can be also done by the callers, which is
more effective. After all, there are only two users of page isolation.
So patch 6 does effectively the same thing as Pavel proposed in [5], and
patch 7 implement stronger guarantees only for memory offline. If CMA
decides to opt-in to the stronger guarantee, it can be added later.
Patches 1-5 are preparatory cleanups for pcplist disabling.
Patchset was briefly tested in QEMU so that memory online/offline works,
but I haven't done a stress test that would prove the race fixed by [2] is
eliminated.
Note that patch 7 could be avoided if we instead adjusted page freeing in
shown in [6], but I believe the current implementation of disabling
pcplists is not too much complex, so I would prefer this instead of adding
new checks and longer irq-disabled section into page freeing hotpaths.
[1] https://lore.kernel.org/linux-mm/20200901124615.137200-1-pasha.tatashin@soleen.com/
[2] https://lore.kernel.org/linux-mm/20200903140032.380431-1-pasha.tatashin@soleen.com/
[3] https://lore.kernel.org/linux-mm/20200907163628.26495-1-vbabka@suse.cz/
[4] https://lore.kernel.org/linux-mm/20200909113647.GG7348@dhcp22.suse.cz/
[5] https://lore.kernel.org/linux-mm/20200904151448.100489-3-pasha.tatashin@soleen.com/
[6] https://lore.kernel.org/linux-mm/3d3b53db-aeaa-ff24-260b-36427fac9b1c@suse.cz/
[7] https://lore.kernel.org/linux-mm/20200922143712.12048-1-vbabka@suse.cz/
[8] https://lore.kernel.org/linux-mm/20201008114201.18824-1-vbabka@suse.cz/
This patch (of 7):
The updates to pcplists' high and batch values are handled by multiple
functions that make the calculations hard to follow. Consolidate
everything to pageset_set_high_and_batch() and remove pageset_set_batch()
and pageset_set_high() wrappers.
The only special case using one of the removed wrappers was:
build_all_zonelists_init()
setup_pageset()
pageset_set_batch()
which was hardcoding batch as 0, so we can just open-code a call to
pageset_update() with constant parameters instead.
No functional change.
Link: https://lkml.kernel.org/r/20201111092812.11329-1-vbabka@suse.cz
Link: https://lkml.kernel.org/r/20201111092812.11329-2-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: David Hildenbrand <david@redhat.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Pankaj Gupta <pankaj.gupta@cloud.ionos.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-14 19:10:40 -08:00
|
|
|
|
2021-06-28 19:42:12 -07:00
|
|
|
|
2023-10-16 13:29:59 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-07-03 15:01:38 -07:00
|
|
|
|
2023-10-16 13:29:59 +08:00
|
|
|
|
|
|
|
|
|
2020-12-14 19:10:53 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-16 13:29:59 +08:00
|
|
|
|
|
|
|
|
|
2020-12-14 19:10:53 -08:00
|
|
|
|
|
|
|
|
|
2023-10-16 13:29:59 +08:00
|
|
|
|
|
|
|
|
|
2013-07-03 15:01:41 -07:00
|
|
|
|
|
|
|
|
|
2017-09-06 16:20:24 -07:00
|
|
|
|
2010-05-24 14:32:49 -07:00
|
|
|
|
|
|
|
|
|
2020-12-14 19:10:43 -08:00
|
|
|
|
mm/page_alloc: split per cpu page lists and zone stats
The PCP (per-cpu page allocator in page_alloc.c) shares locking
requirements with vmstat and the zone lock which is inconvenient and
causes some issues. For example, the PCP list and vmstat share the same
per-cpu space meaning that it's possible that vmstat updates dirty cache
lines holding per-cpu lists across CPUs unless padding is used. Second,
PREEMPT_RT does not want to disable IRQs for too long in the page
allocator.
This series splits the locking requirements and uses locks types more
suitable for PREEMPT_RT, reduces the time when special locking is required
for stats and reduces the time when IRQs need to be disabled on
!PREEMPT_RT kernels.
Why local_lock? PREEMPT_RT considers the following sequence to be unsafe
as documented in Documentation/locking/locktypes.rst
local_irq_disable();
spin_lock(&lock);
The pcp allocator has this sequence for rmqueue_pcplist (local_irq_save)
-> __rmqueue_pcplist -> rmqueue_bulk (spin_lock). While it's possible to
separate this out, it generally means there are points where we enable
IRQs and reenable them again immediately. To prevent a migration and the
per-cpu pointer going stale, migrate_disable is also needed. That is a
custom lock that is similar, but worse, than local_lock. Furthermore, on
PREEMPT_RT, it's undesirable to leave IRQs disabled for too long. By
converting to local_lock which disables migration on PREEMPT_RT, the
locking requirements can be separated and start moving the protections for
PCP, stats and the zone lock to PREEMPT_RT-safe equivalent locking. As a
bonus, local_lock also means that PROVE_LOCKING does something useful.
After that, it's obvious that zone_statistics incurs too much overhead and
leaves IRQs disabled for longer than necessary on !PREEMPT_RT kernels.
zone_statistics uses perfectly accurate counters requiring IRQs be
disabled for parallel RMW sequences when inaccurate ones like vm_events
would do. The series makes the NUMA statistics (NUMA_HIT and friends)
inaccurate counters that then require no special protection on
!PREEMPT_RT.
The bulk page allocator can then do stat updates in bulk with IRQs enabled
which should improve the efficiency. Technically, this could have been
done without the local_lock and vmstat conversion work and the order
simply reflects the timing of when different series were implemented.
Finally, there are places where we conflate IRQs being disabled for the
PCP with the IRQ-safe zone spinlock. The remainder of the series reduces
the scope of what is protected by disabled IRQs on !PREEMPT_RT kernels.
By the end of the series, page_alloc.c does not call local_irq_save so the
locking scope is a bit clearer. The one exception is that modifying
NR_FREE_PAGES still happens in places where it's known the IRQs are
disabled as it's harmless for PREEMPT_RT and would be expensive to split
the locking there.
No performance data is included because despite the overhead of the stats,
it's within the noise for most workloads on !PREEMPT_RT. However, Jesper
Dangaard Brouer ran a page allocation microbenchmark on a E5-1650 v4 @
3.60GHz CPU on the first version of this series. Focusing on the array
variant of the bulk page allocator reveals the following.
(CPU: Intel(R) Xeon(R) CPU E5-1650 v4 @ 3.60GHz)
ARRAY variant: time_bulk_page_alloc_free_array: step=bulk size
Baseline Patched
1 56.383 54.225 (+3.83%)
2 40.047 35.492 (+11.38%)
3 37.339 32.643 (+12.58%)
4 35.578 30.992 (+12.89%)
8 33.592 29.606 (+11.87%)
16 32.362 28.532 (+11.85%)
32 31.476 27.728 (+11.91%)
64 30.633 27.252 (+11.04%)
128 30.596 27.090 (+11.46%)
While this is a positive outcome, the series is more likely to be
interesting to the RT people in terms of getting parts of the PREEMPT_RT
tree into mainline.
This patch (of 9):
The per-cpu page allocator lists and the per-cpu vmstat deltas are stored
in the same struct per_cpu_pages even though vmstats have no direct impact
on the per-cpu page lists. This is inconsistent because the vmstats for a
node are stored on a dedicated structure. The bigger issue is that the
per_cpu_pages structure is not cache-aligned and stat updates either cache
conflict with adjacent per-cpu lists incurring a runtime cost or padding
is required incurring a memory cost.
This patch splits the per-cpu pagelists and the vmstat deltas into
separate structures. It's mostly a mechanical conversion but some
variable renaming is done to clearly distinguish the per-cpu pages
structure (pcp) from the vmstats (pzstats).
Superficially, this appears to increase the size of the per_cpu_pages
structure but the movement of expire fills a structure hole so there is no
impact overall.
[mgorman@techsingularity.net: make it W=1 cleaner]
Link: https://lkml.kernel.org/r/20210514144622.GA3735@techsingularity.net
[mgorman@techsingularity.net: make it W=1 even cleaner]
Link: https://lkml.kernel.org/r/20210516140705.GB3735@techsingularity.net
[lkp@intel.com: check struct per_cpu_zonestat has a non-zero size]
[vbabka@suse.cz: Init zone->per_cpu_zonestats properly]
Link: https://lkml.kernel.org/r/20210512095458.30632-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20210512095458.30632-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-28 19:41:38 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-12-14 19:10:43 -08:00
|
|
|
|
mm/page_alloc: split per cpu page lists and zone stats
The PCP (per-cpu page allocator in page_alloc.c) shares locking
requirements with vmstat and the zone lock which is inconvenient and
causes some issues. For example, the PCP list and vmstat share the same
per-cpu space meaning that it's possible that vmstat updates dirty cache
lines holding per-cpu lists across CPUs unless padding is used. Second,
PREEMPT_RT does not want to disable IRQs for too long in the page
allocator.
This series splits the locking requirements and uses locks types more
suitable for PREEMPT_RT, reduces the time when special locking is required
for stats and reduces the time when IRQs need to be disabled on
!PREEMPT_RT kernels.
Why local_lock? PREEMPT_RT considers the following sequence to be unsafe
as documented in Documentation/locking/locktypes.rst
local_irq_disable();
spin_lock(&lock);
The pcp allocator has this sequence for rmqueue_pcplist (local_irq_save)
-> __rmqueue_pcplist -> rmqueue_bulk (spin_lock). While it's possible to
separate this out, it generally means there are points where we enable
IRQs and reenable them again immediately. To prevent a migration and the
per-cpu pointer going stale, migrate_disable is also needed. That is a
custom lock that is similar, but worse, than local_lock. Furthermore, on
PREEMPT_RT, it's undesirable to leave IRQs disabled for too long. By
converting to local_lock which disables migration on PREEMPT_RT, the
locking requirements can be separated and start moving the protections for
PCP, stats and the zone lock to PREEMPT_RT-safe equivalent locking. As a
bonus, local_lock also means that PROVE_LOCKING does something useful.
After that, it's obvious that zone_statistics incurs too much overhead and
leaves IRQs disabled for longer than necessary on !PREEMPT_RT kernels.
zone_statistics uses perfectly accurate counters requiring IRQs be
disabled for parallel RMW sequences when inaccurate ones like vm_events
would do. The series makes the NUMA statistics (NUMA_HIT and friends)
inaccurate counters that then require no special protection on
!PREEMPT_RT.
The bulk page allocator can then do stat updates in bulk with IRQs enabled
which should improve the efficiency. Technically, this could have been
done without the local_lock and vmstat conversion work and the order
simply reflects the timing of when different series were implemented.
Finally, there are places where we conflate IRQs being disabled for the
PCP with the IRQ-safe zone spinlock. The remainder of the series reduces
the scope of what is protected by disabled IRQs on !PREEMPT_RT kernels.
By the end of the series, page_alloc.c does not call local_irq_save so the
locking scope is a bit clearer. The one exception is that modifying
NR_FREE_PAGES still happens in places where it's known the IRQs are
disabled as it's harmless for PREEMPT_RT and would be expensive to split
the locking there.
No performance data is included because despite the overhead of the stats,
it's within the noise for most workloads on !PREEMPT_RT. However, Jesper
Dangaard Brouer ran a page allocation microbenchmark on a E5-1650 v4 @
3.60GHz CPU on the first version of this series. Focusing on the array
variant of the bulk page allocator reveals the following.
(CPU: Intel(R) Xeon(R) CPU E5-1650 v4 @ 3.60GHz)
ARRAY variant: time_bulk_page_alloc_free_array: step=bulk size
Baseline Patched
1 56.383 54.225 (+3.83%)
2 40.047 35.492 (+11.38%)
3 37.339 32.643 (+12.58%)
4 35.578 30.992 (+12.89%)
8 33.592 29.606 (+11.87%)
16 32.362 28.532 (+11.85%)
32 31.476 27.728 (+11.91%)
64 30.633 27.252 (+11.04%)
128 30.596 27.090 (+11.46%)
While this is a positive outcome, the series is more likely to be
interesting to the RT people in terms of getting parts of the PREEMPT_RT
tree into mainline.
This patch (of 9):
The per-cpu page allocator lists and the per-cpu vmstat deltas are stored
in the same struct per_cpu_pages even though vmstats have no direct impact
on the per-cpu page lists. This is inconsistent because the vmstats for a
node are stored on a dedicated structure. The bigger issue is that the
per_cpu_pages structure is not cache-aligned and stat updates either cache
conflict with adjacent per-cpu lists incurring a runtime cost or padding
is required incurring a memory cost.
This patch splits the per-cpu pagelists and the vmstat deltas into
separate structures. It's mostly a mechanical conversion but some
variable renaming is done to clearly distinguish the per-cpu pages
structure (pcp) from the vmstats (pzstats).
Superficially, this appears to increase the size of the per_cpu_pages
structure but the movement of expire fills a structure hole so there is no
impact overall.
[mgorman@techsingularity.net: make it W=1 cleaner]
Link: https://lkml.kernel.org/r/20210514144622.GA3735@techsingularity.net
[mgorman@techsingularity.net: make it W=1 even cleaner]
Link: https://lkml.kernel.org/r/20210516140705.GB3735@techsingularity.net
[lkp@intel.com: check struct per_cpu_zonestat has a non-zero size]
[vbabka@suse.cz: Init zone->per_cpu_zonestats properly]
Link: https://lkml.kernel.org/r/20210512095458.30632-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20210512095458.30632-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-28 19:41:38 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-12-14 19:10:43 -08:00
|
|
|
|
|
|
|
|
|
2021-06-28 19:42:15 -07:00
|
|
|
|
2010-05-24 14:32:49 -07:00
|
|
|
|
|
|
|
|
|
2022-09-16 15:22:43 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-01-26 16:19:44 +08:00
|
|
|
|
mm, pcp: reduce lock contention for draining high-order pages
In commit f26b3fa04611 ("mm/page_alloc: limit number of high-order pages
on PCP during bulk free"), the PCP (Per-CPU Pageset) will be drained when
PCP is mostly used for high-order pages freeing to improve the cache-hot
pages reusing between page allocating and freeing CPUs.
On system with small per-CPU data cache slice, pages shouldn't be cached
before draining to guarantee cache-hot. But on a system with large
per-CPU data cache slice, some pages can be cached before draining to
reduce zone lock contention.
So, in this patch, instead of draining without any caching, "pcp->batch"
pages will be cached in PCP before draining if the size of the per-CPU
data cache slice is more than "3 * batch".
In theory, if the size of per-CPU data cache slice is more than "2 *
batch", we can reuse cache-hot pages between CPUs. But considering the
other usage of cache (code, other data accessing, etc.), "3 * batch" is
used.
Note: "3 * batch" is chosen to make sure the optimization works on recent
x86_64 server CPUs. If you want to increase it, please check whether it
breaks the optimization.
On a 2-socket Intel server with 128 logical CPU, with the patch, the
network bandwidth of the UNIX (AF_UNIX) test case of lmbench test suite
with 16-pair processes increase 70.5%. The cycles% of the spinlock
contention (mostly for zone lock) decreases from 46.1% to 21.3%. The
number of PCP draining for high order pages freeing (free_high) decreases
89.9%. The cache miss rate keeps 0.2%.
Link: https://lkml.kernel.org/r/20231016053002.756205-4-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:29:56 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-01-26 16:19:44 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, pcp: reduce lock contention for draining high-order pages
In commit f26b3fa04611 ("mm/page_alloc: limit number of high-order pages
on PCP during bulk free"), the PCP (Per-CPU Pageset) will be drained when
PCP is mostly used for high-order pages freeing to improve the cache-hot
pages reusing between page allocating and freeing CPUs.
On system with small per-CPU data cache slice, pages shouldn't be cached
before draining to guarantee cache-hot. But on a system with large
per-CPU data cache slice, some pages can be cached before draining to
reduce zone lock contention.
So, in this patch, instead of draining without any caching, "pcp->batch"
pages will be cached in PCP before draining if the size of the per-CPU
data cache slice is more than "3 * batch".
In theory, if the size of per-CPU data cache slice is more than "2 *
batch", we can reuse cache-hot pages between CPUs. But considering the
other usage of cache (code, other data accessing, etc.), "3 * batch" is
used.
Note: "3 * batch" is chosen to make sure the optimization works on recent
x86_64 server CPUs. If you want to increase it, please check whether it
breaks the optimization.
On a 2-socket Intel server with 128 logical CPU, with the patch, the
network bandwidth of the UNIX (AF_UNIX) test case of lmbench test suite
with 16-pair processes increase 70.5%. The cycles% of the spinlock
contention (mostly for zone lock) decreases from 46.1% to 21.3%. The
number of PCP draining for high order pages freeing (free_high) decreases
89.9%. The cache miss rate keeps 0.2%.
Link: https://lkml.kernel.org/r/20231016053002.756205-4-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:29:56 +08:00
|
|
|
|
|
|
|
|
|
2024-01-26 16:19:44 +08:00
|
|
|
|
mm, pcp: reduce lock contention for draining high-order pages
In commit f26b3fa04611 ("mm/page_alloc: limit number of high-order pages
on PCP during bulk free"), the PCP (Per-CPU Pageset) will be drained when
PCP is mostly used for high-order pages freeing to improve the cache-hot
pages reusing between page allocating and freeing CPUs.
On system with small per-CPU data cache slice, pages shouldn't be cached
before draining to guarantee cache-hot. But on a system with large
per-CPU data cache slice, some pages can be cached before draining to
reduce zone lock contention.
So, in this patch, instead of draining without any caching, "pcp->batch"
pages will be cached in PCP before draining if the size of the per-CPU
data cache slice is more than "3 * batch".
In theory, if the size of per-CPU data cache slice is more than "2 *
batch", we can reuse cache-hot pages between CPUs. But considering the
other usage of cache (code, other data accessing, etc.), "3 * batch" is
used.
Note: "3 * batch" is chosen to make sure the optimization works on recent
x86_64 server CPUs. If you want to increase it, please check whether it
breaks the optimization.
On a 2-socket Intel server with 128 logical CPU, with the patch, the
network bandwidth of the UNIX (AF_UNIX) test case of lmbench test suite
with 16-pair processes increase 70.5%. The cycles% of the spinlock
contention (mostly for zone lock) decreases from 46.1% to 21.3%. The
number of PCP draining for high order pages freeing (free_high) decreases
89.9%. The cache miss rate keeps 0.2%.
Link: https://lkml.kernel.org/r/20231016053002.756205-4-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:29:56 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-01-26 16:19:44 +08:00
|
|
|
|
mm, pcp: reduce lock contention for draining high-order pages
In commit f26b3fa04611 ("mm/page_alloc: limit number of high-order pages
on PCP during bulk free"), the PCP (Per-CPU Pageset) will be drained when
PCP is mostly used for high-order pages freeing to improve the cache-hot
pages reusing between page allocating and freeing CPUs.
On system with small per-CPU data cache slice, pages shouldn't be cached
before draining to guarantee cache-hot. But on a system with large
per-CPU data cache slice, some pages can be cached before draining to
reduce zone lock contention.
So, in this patch, instead of draining without any caching, "pcp->batch"
pages will be cached in PCP before draining if the size of the per-CPU
data cache slice is more than "3 * batch".
In theory, if the size of per-CPU data cache slice is more than "2 *
batch", we can reuse cache-hot pages between CPUs. But considering the
other usage of cache (code, other data accessing, etc.), "3 * batch" is
used.
Note: "3 * batch" is chosen to make sure the optimization works on recent
x86_64 server CPUs. If you want to increase it, please check whether it
breaks the optimization.
On a 2-socket Intel server with 128 logical CPU, with the patch, the
network bandwidth of the UNIX (AF_UNIX) test case of lmbench test suite
with 16-pair processes increase 70.5%. The cycles% of the spinlock
contention (mostly for zone lock) decreases from 46.1% to 21.3%. The
number of PCP draining for high order pages freeing (free_high) decreases
89.9%. The cache miss rate keeps 0.2%.
Link: https://lkml.kernel.org/r/20231016053002.756205-4-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:29:56 +08:00
|
|
|
|
|
|
|
|
|
2005-06-21 17:14:47 -07:00
|
|
|
|
2010-01-05 15:34:51 +09:00
|
|
|
|
|
|
|
|
|
2005-06-21 17:15:00 -07:00
|
|
|
|
2005-12-15 09:18:25 +00:00
|
|
|
|
2005-06-21 17:14:47 -07:00
|
|
|
|
mm: initialise per_cpu_nodestats for all online pgdats at boot
Paul Mackerras and Reza Arbab reported that machines with memoryless
nodes fail when vmstats are refreshed. Paul reported an oops as follows
Unable to handle kernel paging request for data at address 0xff7a10000
Faulting instruction address: 0xc000000000270cd0
Oops: Kernel access of bad area, sig: 11 [#1]
SMP NR_CPUS=2048 NUMA PowerNV
Modules linked in:
CPU: 0 PID: 1 Comm: swapper/0 Not tainted 4.7.0-kvm+ #118
task: c000000ff0680010 task.stack: c000000ff0704000
NIP: c000000000270cd0 LR: c000000000270ce8 CTR: 0000000000000000
REGS: c000000ff0707900 TRAP: 0300 Not tainted (4.7.0-kvm+)
MSR: 9000000102009033 <SF,HV,VEC,EE,ME,IR,DR,RI,LE,TM[E]> CR: 846b6824 XER: 20000000
CFAR: c000000000008768 DAR: 0000000ff7a10000 DSISR: 42000000 SOFTE: 1
NIP refresh_zone_stat_thresholds+0x80/0x240
LR refresh_zone_stat_thresholds+0x98/0x240
Call Trace:
refresh_zone_stat_thresholds+0xb8/0x240 (unreliable)
Both supplied potential fixes but one potentially misses checks and
another had redundant initialisations. This version initialises
per_cpu_nodestats on a per-pgdat basis instead of on a per-zone basis.
Link: http://lkml.kernel.org/r/20160804092404.GI2799@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Reported-by: Paul Mackerras <paulus@ozlabs.org>
Reported-by: Reza Arbab <arbab@linux.vnet.ibm.com>
Tested-by: Reza Arbab <arbab@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-08-04 15:31:49 -07:00
|
|
|
|
2010-01-05 15:34:51 +09:00
|
|
|
|
2020-06-03 15:59:11 -07:00
|
|
|
|
2005-06-21 17:14:47 -07:00
|
|
|
|
2010-05-24 14:32:49 -07:00
|
|
|
|
|
|
|
|
|
mm: initialise per_cpu_nodestats for all online pgdats at boot
Paul Mackerras and Reza Arbab reported that machines with memoryless
nodes fail when vmstats are refreshed. Paul reported an oops as follows
Unable to handle kernel paging request for data at address 0xff7a10000
Faulting instruction address: 0xc000000000270cd0
Oops: Kernel access of bad area, sig: 11 [#1]
SMP NR_CPUS=2048 NUMA PowerNV
Modules linked in:
CPU: 0 PID: 1 Comm: swapper/0 Not tainted 4.7.0-kvm+ #118
task: c000000ff0680010 task.stack: c000000ff0704000
NIP: c000000000270cd0 LR: c000000000270ce8 CTR: 0000000000000000
REGS: c000000ff0707900 TRAP: 0300 Not tainted (4.7.0-kvm+)
MSR: 9000000102009033 <SF,HV,VEC,EE,ME,IR,DR,RI,LE,TM[E]> CR: 846b6824 XER: 20000000
CFAR: c000000000008768 DAR: 0000000ff7a10000 DSISR: 42000000 SOFTE: 1
NIP refresh_zone_stat_thresholds+0x80/0x240
LR refresh_zone_stat_thresholds+0x98/0x240
Call Trace:
refresh_zone_stat_thresholds+0xb8/0x240 (unreliable)
Both supplied potential fixes but one potentially misses checks and
another had redundant initialisations. This version initialises
per_cpu_nodestats on a per-pgdat basis instead of on a per-zone basis.
Link: http://lkml.kernel.org/r/20160804092404.GI2799@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Reported-by: Paul Mackerras <paulus@ozlabs.org>
Reported-by: Reza Arbab <arbab@linux.vnet.ibm.com>
Tested-by: Reza Arbab <arbab@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-08-04 15:31:49 -07:00
|
|
|
|
2020-06-03 15:59:11 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/page_alloc: split per cpu page lists and zone stats
The PCP (per-cpu page allocator in page_alloc.c) shares locking
requirements with vmstat and the zone lock which is inconvenient and
causes some issues. For example, the PCP list and vmstat share the same
per-cpu space meaning that it's possible that vmstat updates dirty cache
lines holding per-cpu lists across CPUs unless padding is used. Second,
PREEMPT_RT does not want to disable IRQs for too long in the page
allocator.
This series splits the locking requirements and uses locks types more
suitable for PREEMPT_RT, reduces the time when special locking is required
for stats and reduces the time when IRQs need to be disabled on
!PREEMPT_RT kernels.
Why local_lock? PREEMPT_RT considers the following sequence to be unsafe
as documented in Documentation/locking/locktypes.rst
local_irq_disable();
spin_lock(&lock);
The pcp allocator has this sequence for rmqueue_pcplist (local_irq_save)
-> __rmqueue_pcplist -> rmqueue_bulk (spin_lock). While it's possible to
separate this out, it generally means there are points where we enable
IRQs and reenable them again immediately. To prevent a migration and the
per-cpu pointer going stale, migrate_disable is also needed. That is a
custom lock that is similar, but worse, than local_lock. Furthermore, on
PREEMPT_RT, it's undesirable to leave IRQs disabled for too long. By
converting to local_lock which disables migration on PREEMPT_RT, the
locking requirements can be separated and start moving the protections for
PCP, stats and the zone lock to PREEMPT_RT-safe equivalent locking. As a
bonus, local_lock also means that PROVE_LOCKING does something useful.
After that, it's obvious that zone_statistics incurs too much overhead and
leaves IRQs disabled for longer than necessary on !PREEMPT_RT kernels.
zone_statistics uses perfectly accurate counters requiring IRQs be
disabled for parallel RMW sequences when inaccurate ones like vm_events
would do. The series makes the NUMA statistics (NUMA_HIT and friends)
inaccurate counters that then require no special protection on
!PREEMPT_RT.
The bulk page allocator can then do stat updates in bulk with IRQs enabled
which should improve the efficiency. Technically, this could have been
done without the local_lock and vmstat conversion work and the order
simply reflects the timing of when different series were implemented.
Finally, there are places where we conflate IRQs being disabled for the
PCP with the IRQ-safe zone spinlock. The remainder of the series reduces
the scope of what is protected by disabled IRQs on !PREEMPT_RT kernels.
By the end of the series, page_alloc.c does not call local_irq_save so the
locking scope is a bit clearer. The one exception is that modifying
NR_FREE_PAGES still happens in places where it's known the IRQs are
disabled as it's harmless for PREEMPT_RT and would be expensive to split
the locking there.
No performance data is included because despite the overhead of the stats,
it's within the noise for most workloads on !PREEMPT_RT. However, Jesper
Dangaard Brouer ran a page allocation microbenchmark on a E5-1650 v4 @
3.60GHz CPU on the first version of this series. Focusing on the array
variant of the bulk page allocator reveals the following.
(CPU: Intel(R) Xeon(R) CPU E5-1650 v4 @ 3.60GHz)
ARRAY variant: time_bulk_page_alloc_free_array: step=bulk size
Baseline Patched
1 56.383 54.225 (+3.83%)
2 40.047 35.492 (+11.38%)
3 37.339 32.643 (+12.58%)
4 35.578 30.992 (+12.89%)
8 33.592 29.606 (+11.87%)
16 32.362 28.532 (+11.85%)
32 31.476 27.728 (+11.91%)
64 30.633 27.252 (+11.04%)
128 30.596 27.090 (+11.46%)
While this is a positive outcome, the series is more likely to be
interesting to the RT people in terms of getting parts of the PREEMPT_RT
tree into mainline.
This patch (of 9):
The per-cpu page allocator lists and the per-cpu vmstat deltas are stored
in the same struct per_cpu_pages even though vmstats have no direct impact
on the per-cpu page lists. This is inconsistent because the vmstats for a
node are stored on a dedicated structure. The bigger issue is that the
per_cpu_pages structure is not cache-aligned and stat updates either cache
conflict with adjacent per-cpu lists incurring a runtime cost or padding
is required incurring a memory cost.
This patch splits the per-cpu pagelists and the vmstat deltas into
separate structures. It's mostly a mechanical conversion but some
variable renaming is done to clearly distinguish the per-cpu pages
structure (pcp) from the vmstats (pzstats).
Superficially, this appears to increase the size of the per_cpu_pages
structure but the movement of expire fills a structure hole so there is no
impact overall.
[mgorman@techsingularity.net: make it W=1 cleaner]
Link: https://lkml.kernel.org/r/20210514144622.GA3735@techsingularity.net
[mgorman@techsingularity.net: make it W=1 even cleaner]
Link: https://lkml.kernel.org/r/20210516140705.GB3735@techsingularity.net
[lkp@intel.com: check struct per_cpu_zonestat has a non-zero size]
[vbabka@suse.cz: Init zone->per_cpu_zonestats properly]
Link: https://lkml.kernel.org/r/20210512095458.30632-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20210512095458.30632-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-28 19:41:38 -07:00
|
|
|
|
2021-06-28 19:41:44 -07:00
|
|
|
|
|
|
|
|
|
2020-06-03 15:59:11 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: initialise per_cpu_nodestats for all online pgdats at boot
Paul Mackerras and Reza Arbab reported that machines with memoryless
nodes fail when vmstats are refreshed. Paul reported an oops as follows
Unable to handle kernel paging request for data at address 0xff7a10000
Faulting instruction address: 0xc000000000270cd0
Oops: Kernel access of bad area, sig: 11 [#1]
SMP NR_CPUS=2048 NUMA PowerNV
Modules linked in:
CPU: 0 PID: 1 Comm: swapper/0 Not tainted 4.7.0-kvm+ #118
task: c000000ff0680010 task.stack: c000000ff0704000
NIP: c000000000270cd0 LR: c000000000270ce8 CTR: 0000000000000000
REGS: c000000ff0707900 TRAP: 0300 Not tainted (4.7.0-kvm+)
MSR: 9000000102009033 <SF,HV,VEC,EE,ME,IR,DR,RI,LE,TM[E]> CR: 846b6824 XER: 20000000
CFAR: c000000000008768 DAR: 0000000ff7a10000 DSISR: 42000000 SOFTE: 1
NIP refresh_zone_stat_thresholds+0x80/0x240
LR refresh_zone_stat_thresholds+0x98/0x240
Call Trace:
refresh_zone_stat_thresholds+0xb8/0x240 (unreliable)
Both supplied potential fixes but one potentially misses checks and
another had redundant initialisations. This version initialises
per_cpu_nodestats on a per-pgdat basis instead of on a per-zone basis.
Link: http://lkml.kernel.org/r/20160804092404.GI2799@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Reported-by: Paul Mackerras <paulus@ozlabs.org>
Reported-by: Reza Arbab <arbab@linux.vnet.ibm.com>
Tested-by: Reza Arbab <arbab@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-08-04 15:31:49 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-06-21 17:14:47 -07:00
|
|
|
|
|
|
|
|
|
2023-03-21 19:05:02 +02:00
|
|
|
|
2005-10-29 18:16:50 -07:00
|
|
|
|
2010-01-05 15:34:51 +09:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/page_alloc: split per cpu page lists and zone stats
The PCP (per-cpu page allocator in page_alloc.c) shares locking
requirements with vmstat and the zone lock which is inconvenient and
causes some issues. For example, the PCP list and vmstat share the same
per-cpu space meaning that it's possible that vmstat updates dirty cache
lines holding per-cpu lists across CPUs unless padding is used. Second,
PREEMPT_RT does not want to disable IRQs for too long in the page
allocator.
This series splits the locking requirements and uses locks types more
suitable for PREEMPT_RT, reduces the time when special locking is required
for stats and reduces the time when IRQs need to be disabled on
!PREEMPT_RT kernels.
Why local_lock? PREEMPT_RT considers the following sequence to be unsafe
as documented in Documentation/locking/locktypes.rst
local_irq_disable();
spin_lock(&lock);
The pcp allocator has this sequence for rmqueue_pcplist (local_irq_save)
-> __rmqueue_pcplist -> rmqueue_bulk (spin_lock). While it's possible to
separate this out, it generally means there are points where we enable
IRQs and reenable them again immediately. To prevent a migration and the
per-cpu pointer going stale, migrate_disable is also needed. That is a
custom lock that is similar, but worse, than local_lock. Furthermore, on
PREEMPT_RT, it's undesirable to leave IRQs disabled for too long. By
converting to local_lock which disables migration on PREEMPT_RT, the
locking requirements can be separated and start moving the protections for
PCP, stats and the zone lock to PREEMPT_RT-safe equivalent locking. As a
bonus, local_lock also means that PROVE_LOCKING does something useful.
After that, it's obvious that zone_statistics incurs too much overhead and
leaves IRQs disabled for longer than necessary on !PREEMPT_RT kernels.
zone_statistics uses perfectly accurate counters requiring IRQs be
disabled for parallel RMW sequences when inaccurate ones like vm_events
would do. The series makes the NUMA statistics (NUMA_HIT and friends)
inaccurate counters that then require no special protection on
!PREEMPT_RT.
The bulk page allocator can then do stat updates in bulk with IRQs enabled
which should improve the efficiency. Technically, this could have been
done without the local_lock and vmstat conversion work and the order
simply reflects the timing of when different series were implemented.
Finally, there are places where we conflate IRQs being disabled for the
PCP with the IRQ-safe zone spinlock. The remainder of the series reduces
the scope of what is protected by disabled IRQs on !PREEMPT_RT kernels.
By the end of the series, page_alloc.c does not call local_irq_save so the
locking scope is a bit clearer. The one exception is that modifying
NR_FREE_PAGES still happens in places where it's known the IRQs are
disabled as it's harmless for PREEMPT_RT and would be expensive to split
the locking there.
No performance data is included because despite the overhead of the stats,
it's within the noise for most workloads on !PREEMPT_RT. However, Jesper
Dangaard Brouer ran a page allocation microbenchmark on a E5-1650 v4 @
3.60GHz CPU on the first version of this series. Focusing on the array
variant of the bulk page allocator reveals the following.
(CPU: Intel(R) Xeon(R) CPU E5-1650 v4 @ 3.60GHz)
ARRAY variant: time_bulk_page_alloc_free_array: step=bulk size
Baseline Patched
1 56.383 54.225 (+3.83%)
2 40.047 35.492 (+11.38%)
3 37.339 32.643 (+12.58%)
4 35.578 30.992 (+12.89%)
8 33.592 29.606 (+11.87%)
16 32.362 28.532 (+11.85%)
32 31.476 27.728 (+11.91%)
64 30.633 27.252 (+11.04%)
128 30.596 27.090 (+11.46%)
While this is a positive outcome, the series is more likely to be
interesting to the RT people in terms of getting parts of the PREEMPT_RT
tree into mainline.
This patch (of 9):
The per-cpu page allocator lists and the per-cpu vmstat deltas are stored
in the same struct per_cpu_pages even though vmstats have no direct impact
on the per-cpu page lists. This is inconsistent because the vmstats for a
node are stored on a dedicated structure. The bigger issue is that the
per_cpu_pages structure is not cache-aligned and stat updates either cache
conflict with adjacent per-cpu lists incurring a runtime cost or padding
is required incurring a memory cost.
This patch splits the per-cpu pagelists and the vmstat deltas into
separate structures. It's mostly a mechanical conversion but some
variable renaming is done to clearly distinguish the per-cpu pages
structure (pcp) from the vmstats (pzstats).
Superficially, this appears to increase the size of the per_cpu_pages
structure but the movement of expire fills a structure hole so there is no
impact overall.
[mgorman@techsingularity.net: make it W=1 cleaner]
Link: https://lkml.kernel.org/r/20210514144622.GA3735@techsingularity.net
[mgorman@techsingularity.net: make it W=1 even cleaner]
Link: https://lkml.kernel.org/r/20210516140705.GB3735@techsingularity.net
[lkp@intel.com: check struct per_cpu_zonestat has a non-zero size]
[vbabka@suse.cz: Init zone->per_cpu_zonestats properly]
Link: https://lkml.kernel.org/r/20210512095458.30632-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20210512095458.30632-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-28 19:41:38 -07:00
|
|
|
|
|
|
|
|
|
2023-10-16 13:29:59 +08:00
|
|
|
|
|
|
|
|
|
2020-12-14 19:10:53 -08:00
|
|
|
|
2005-10-29 18:16:50 -07:00
|
|
|
|
2013-11-12 15:07:20 -08:00
|
|
|
|
2021-06-28 19:41:31 -07:00
|
|
|
|
|
|
|
|
|
2005-10-29 18:16:50 -07:00
|
|
|
|
|
|
|
|
|
mm: page_alloc: fix missed updates of lowmem_reserve in adjust_managed_page_count
In the kernel, the zone's lowmem_reserve and _watermark, and the global
variable 'totalreserve_pages' depend on the value of managed_pages, but
after running adjust_managed_page_count, these values aren't updated,
which causes some problems.
For example, in a system with six 1GB large pages, we found that the value
of protection in zoneinfo (zone->lowmem_reserve), is not right. Its value
seems to be calculated from the initial managed_pages, but after the
managed_pages changed, was not updated. Only after reading the file
/proc/sys/vm/lowmem_reserve_ratio, updates happen.
read file /proc/sys/vm/lowmem_reserve_ratio:
lowmem_reserve_ratio_sysctl_handler
----setup_per_zone_lowmem_reserve
--------calculate_totalreserve_pages
protection changed after reading file:
[root@test ~]# cat /proc/zoneinfo | grep protection
protection: (0, 2719, 57360, 0)
protection: (0, 0, 54640, 0)
protection: (0, 0, 0, 0)
protection: (0, 0, 0, 0)
[root@test ~]# cat /proc/sys/vm/lowmem_reserve_ratio
256 256 32 0
[root@test ~]# cat /proc/zoneinfo | grep protection
protection: (0, 2735, 63524, 0)
protection: (0, 0, 60788, 0)
protection: (0, 0, 0, 0)
protection: (0, 0, 0, 0)
lowmem_reserve increased also makes the totalreserve_pages increased,
which causes a decrease in available memory. The one above is just a test
machine, and the increase is not significant. On our online machine, the
reserved memory will increase by several GB due to reading this file. It
is clearly unreasonable to cause a sharp drop in available memory just by
reading a file.
In this patch, we update reserve memory when update managed_pages, The
size of reserved memory becomes stable. But it seems that the _watermark
should also be updated along with the managed_pages. We have not done it
because we are unsure if it is reasonable to set the watermark through the
initial managed_pages. If it is not reasonable, we will propose new
patch.
Link: https://lkml.kernel.org/r/20241225021034.45693-1-15645113830zzh@gmail.com
Signed-off-by: zihan zhou <15645113830zzh@gmail.com>
Signed-off-by: yaowenchao <yaowenchao@jd.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-12-25 10:10:35 +08:00
|
|
|
|
|
|
|
|
|
2013-07-03 15:03:14 -07:00
|
|
|
|
|
|
|
|
|
2018-12-28 00:34:24 -08:00
|
|
|
|
2018-12-28 00:34:29 -08:00
|
|
|
|
mm: page_alloc: fix missed updates of lowmem_reserve in adjust_managed_page_count
In the kernel, the zone's lowmem_reserve and _watermark, and the global
variable 'totalreserve_pages' depend on the value of managed_pages, but
after running adjust_managed_page_count, these values aren't updated,
which causes some problems.
For example, in a system with six 1GB large pages, we found that the value
of protection in zoneinfo (zone->lowmem_reserve), is not right. Its value
seems to be calculated from the initial managed_pages, but after the
managed_pages changed, was not updated. Only after reading the file
/proc/sys/vm/lowmem_reserve_ratio, updates happen.
read file /proc/sys/vm/lowmem_reserve_ratio:
lowmem_reserve_ratio_sysctl_handler
----setup_per_zone_lowmem_reserve
--------calculate_totalreserve_pages
protection changed after reading file:
[root@test ~]# cat /proc/zoneinfo | grep protection
protection: (0, 2719, 57360, 0)
protection: (0, 0, 54640, 0)
protection: (0, 0, 0, 0)
protection: (0, 0, 0, 0)
[root@test ~]# cat /proc/sys/vm/lowmem_reserve_ratio
256 256 32 0
[root@test ~]# cat /proc/zoneinfo | grep protection
protection: (0, 2735, 63524, 0)
protection: (0, 0, 60788, 0)
protection: (0, 0, 0, 0)
protection: (0, 0, 0, 0)
lowmem_reserve increased also makes the totalreserve_pages increased,
which causes a decrease in available memory. The one above is just a test
machine, and the increase is not significant. On our online machine, the
reserved memory will increase by several GB due to reading this file. It
is clearly unreasonable to cause a sharp drop in available memory just by
reading a file.
In this patch, we update reserve memory when update managed_pages, The
size of reserved memory becomes stable. But it seems that the _watermark
should also be updated along with the managed_pages. We have not done it
because we are unsure if it is reasonable to set the watermark through the
initial managed_pages. If it is not reasonable, we will propose new
patch.
Link: https://lkml.kernel.org/r/20241225021034.45693-1-15645113830zzh@gmail.com
Signed-off-by: zihan zhou <15645113830zzh@gmail.com>
Signed-off-by: yaowenchao <yaowenchao@jd.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-12-25 10:10:35 +08:00
|
|
|
|
2013-07-03 15:03:14 -07:00
|
|
|
|
2013-07-03 15:03:21 -07:00
|
|
|
|
2013-07-03 15:03:14 -07:00
|
|
|
|
2018-12-28 00:36:03 -08:00
|
|
|
|
2013-04-29 15:06:21 -07:00
|
|
|
|
mm: change signature of free_reserved_area() to fix building warnings
Change signature of free_reserved_area() according to Russell King's
suggestion to fix following build warnings:
arch/arm/mm/init.c: In function 'mem_init':
arch/arm/mm/init.c:603:2: warning: passing argument 1 of 'free_reserved_area' makes integer from pointer without a cast [enabled by default]
free_reserved_area(__va(PHYS_PFN_OFFSET), swapper_pg_dir, 0, NULL);
^
In file included from include/linux/mman.h:4:0,
from arch/arm/mm/init.c:15:
include/linux/mm.h:1301:22: note: expected 'long unsigned int' but argument is of type 'void *'
extern unsigned long free_reserved_area(unsigned long start, unsigned long end,
mm/page_alloc.c: In function 'free_reserved_area':
>> mm/page_alloc.c:5134:3: warning: passing argument 1 of 'virt_to_phys' makes pointer from integer without a cast [enabled by default]
In file included from arch/mips/include/asm/page.h:49:0,
from include/linux/mmzone.h:20,
from include/linux/gfp.h:4,
from include/linux/mm.h:8,
from mm/page_alloc.c:18:
arch/mips/include/asm/io.h:119:29: note: expected 'const volatile void *' but argument is of type 'long unsigned int'
mm/page_alloc.c: In function 'free_area_init_nodes':
mm/page_alloc.c:5030:34: warning: array subscript is below array bounds [-Warray-bounds]
Also address some minor code review comments.
Signed-off-by: Jiang Liu <jiang.liu@huawei.com>
Reported-by: Arnd Bergmann <arnd@arndb.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: "Michael S. Tsirkin" <mst@redhat.com>
Cc: <sworddragon2@aol.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Chris Metcalf <cmetcalf@tilera.com>
Cc: David Howells <dhowells@redhat.com>
Cc: Geert Uytterhoeven <geert@linux-m68k.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Jianguo Wu <wujianguo@huawei.com>
Cc: Joonsoo Kim <js1304@gmail.com>
Cc: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Cc: Marek Szyprowski <m.szyprowski@samsung.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Michel Lespinasse <walken@google.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Cc: Tang Chen <tangchen@cn.fujitsu.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Wen Congyang <wency@cn.fujitsu.com>
Cc: Will Deacon <will.deacon@arm.com>
Cc: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Cc: Yinghai Lu <yinghai@kernel.org>
Cc: Russell King <rmk@arm.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-07-03 15:02:48 -07:00
|
|
|
|
|
|
|
|
|
2013-04-29 15:06:21 -07:00
|
|
|
|
mm: change signature of free_reserved_area() to fix building warnings
Change signature of free_reserved_area() according to Russell King's
suggestion to fix following build warnings:
arch/arm/mm/init.c: In function 'mem_init':
arch/arm/mm/init.c:603:2: warning: passing argument 1 of 'free_reserved_area' makes integer from pointer without a cast [enabled by default]
free_reserved_area(__va(PHYS_PFN_OFFSET), swapper_pg_dir, 0, NULL);
^
In file included from include/linux/mman.h:4:0,
from arch/arm/mm/init.c:15:
include/linux/mm.h:1301:22: note: expected 'long unsigned int' but argument is of type 'void *'
extern unsigned long free_reserved_area(unsigned long start, unsigned long end,
mm/page_alloc.c: In function 'free_reserved_area':
>> mm/page_alloc.c:5134:3: warning: passing argument 1 of 'virt_to_phys' makes pointer from integer without a cast [enabled by default]
In file included from arch/mips/include/asm/page.h:49:0,
from include/linux/mmzone.h:20,
from include/linux/gfp.h:4,
from include/linux/mm.h:8,
from mm/page_alloc.c:18:
arch/mips/include/asm/io.h:119:29: note: expected 'const volatile void *' but argument is of type 'long unsigned int'
mm/page_alloc.c: In function 'free_area_init_nodes':
mm/page_alloc.c:5030:34: warning: array subscript is below array bounds [-Warray-bounds]
Also address some minor code review comments.
Signed-off-by: Jiang Liu <jiang.liu@huawei.com>
Reported-by: Arnd Bergmann <arnd@arndb.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: "Michael S. Tsirkin" <mst@redhat.com>
Cc: <sworddragon2@aol.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Chris Metcalf <cmetcalf@tilera.com>
Cc: David Howells <dhowells@redhat.com>
Cc: Geert Uytterhoeven <geert@linux-m68k.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Jianguo Wu <wujianguo@huawei.com>
Cc: Joonsoo Kim <js1304@gmail.com>
Cc: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Cc: Marek Szyprowski <m.szyprowski@samsung.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Michel Lespinasse <walken@google.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Cc: Tang Chen <tangchen@cn.fujitsu.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Wen Congyang <wency@cn.fujitsu.com>
Cc: Will Deacon <will.deacon@arm.com>
Cc: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Cc: Yinghai Lu <yinghai@kernel.org>
Cc: Russell King <rmk@arm.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-07-03 15:02:48 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-08-02 15:58:26 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-12-22 12:01:49 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-07-03 15:02:51 -07:00
|
|
|
|
2018-08-02 15:58:26 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-04-29 15:06:21 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-11-05 13:40:05 -07:00
|
|
|
|
2013-04-29 15:06:21 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-07-17 14:28:44 -07:00
|
|
|
|
|
|
|
|
|
2024-08-13 08:07:56 -07:00
|
|
|
|
2024-07-17 14:28:44 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-11-03 15:50:02 +01:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2021-06-28 19:42:15 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2016-11-03 15:50:02 +01:00
|
|
|
|
2023-01-12 12:39:31 +00:00
|
|
|
|
2016-11-03 15:50:02 +01:00
|
|
|
|
2008-02-04 22:29:11 -08:00
|
|
|
|
2016-11-03 15:50:02 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-02-04 22:29:11 -08:00
|
|
|
|
2016-11-03 15:50:02 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-28 19:42:15 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-11-03 15:50:02 +01:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2023-03-21 19:05:04 +02:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2016-11-03 15:50:02 +01:00
|
|
|
|
|
|
|
|
|
2021-06-28 19:42:15 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-11-03 15:50:02 +01:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2006-04-10 22:52:59 -07:00
|
|
|
|
2015-09-08 15:04:13 -07:00
|
|
|
|
2006-04-10 22:52:59 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-09-25 23:31:18 -07:00
|
|
|
|
2006-04-10 22:52:59 -07:00
|
|
|
|
|
|
|
|
|
2016-07-28 15:46:11 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-04-10 22:52:59 -07:00
|
|
|
|
|
|
|
|
|
2014-08-06 16:07:14 -07:00
|
|
|
|
2018-12-28 00:34:24 -08:00
|
|
|
|
2006-04-10 22:52:59 -07:00
|
|
|
|
|
|
|
|
|
2025-08-14 17:00:52 +08:00
|
|
|
|
|
|
|
|
|
2006-04-10 22:52:59 -07:00
|
|
|
|
2009-06-16 15:32:12 -07:00
|
|
|
|
|
|
|
|
|
2006-04-10 22:52:59 -07:00
|
|
|
|
2025-08-14 17:00:52 +08:00
|
|
|
|
2016-01-14 15:20:15 -08:00
|
|
|
|
2016-07-28 15:46:11 -07:00
|
|
|
|
2016-01-14 15:20:15 -08:00
|
|
|
|
2006-04-10 22:52:59 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-03-08 03:46:02 +00:00
|
|
|
|
2006-04-10 22:52:59 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2015-09-08 15:04:13 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-12-14 19:11:22 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2006-03-27 01:15:59 -08:00
|
|
|
|
2020-12-14 19:11:22 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2020-12-14 19:11:22 -08:00
|
|
|
|
2021-06-28 19:42:33 -07:00
|
|
|
|
2020-12-14 19:11:22 -08:00
|
|
|
|
2021-06-28 19:42:33 -07:00
|
|
|
|
|
|
|
|
|
Revert "mm/page_alloc.c: don't show protection in zone's ->lowmem_reserve[] for empty zone"
Commit 96a5c186efff ("mm/page_alloc.c: don't show protection in zone's
->lowmem_reserve[] for empty zone") removes the protection of lower zones
from allocations targeting memory-less high zones. This had an unintended
impact on the pattern of reclaims because it makes the high-zone-targeted
allocation more likely to succeed in lower zones, which adds pressure to
said zones. I.e, the following corresponding checks in
zone_watermark_ok/zone_watermark_fast are less likely to trigger:
if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
return false;
As a result, we are observing an increase in reclaim and kswapd scans, due
to the increased pressure. This was initially observed as increased
latency in filesystem operations when benchmarking with fio on a machine
with some memory-less zones, but it has since been associated with
increased contention in locks related to memory reclaim. By reverting
this patch, the original performance was recovered on that machine.
The original commit was introduced as a clarification of the
/proc/zoneinfo output, so it doesn't seem there are usecases depending on
it, making the revert a simple solution.
For reference, I collected vmstat with and without this patch on a freshly
booted system running intensive randread io from an nvme for 5 minutes. I
got:
rpm-6.12.0-slfo.1.2 -> pgscan_kswapd 5629543865
Patched -> pgscan_kswapd 33580844
33M scans is similar to what we had in kernels predating this patch.
These numbers is fairly representative of the workload on this machine, as
measured in several runs. So we are talking about a 2-order of magnitude
increase.
Link: https://lkml.kernel.org/r/20250226032258.234099-1-krisman@suse.de
Fixes: 96a5c186efff ("mm/page_alloc.c: don't show protection in zone's ->lowmem_reserve[] for empty zone")
Signed-off-by: Gabriel Krisman Bertazi <krisman@suse.de>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Baoquan He <bhe@redhat.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-02-25 22:22:58 -05:00
|
|
|
|
2021-06-28 19:42:33 -07:00
|
|
|
|
|
|
|
|
|
2020-12-14 19:11:22 -08:00
|
|
|
|
2025-03-08 03:46:01 +00:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-04-10 22:52:59 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2011-04-25 21:36:42 +00:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-06-25 11:16:56 +08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2023-06-25 11:16:56 +08:00
|
|
|
|
2018-12-28 00:34:24 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-05-15 09:43:59 -07:00
|
|
|
|
|
|
|
|
|
2008-10-18 20:27:11 -07:00
|
|
|
|
2018-12-28 00:34:24 -08:00
|
|
|
|
2024-02-28 23:49:12 +01:00
|
|
|
|
2023-06-25 11:16:56 +08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2005-11-13 16:06:45 -08:00
|
|
|
|
2023-06-25 11:16:56 +08:00
|
|
|
|
|
|
|
|
|
2005-11-13 16:06:45 -08:00
|
|
|
|
2009-06-16 15:32:12 -07:00
|
|
|
|
2019-03-05 15:46:22 -08:00
|
|
|
|
2023-06-25 11:16:56 +08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2013-02-22 16:32:22 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2018-12-28 00:34:24 -08:00
|
|
|
|
2013-02-22 16:32:22 -08:00
|
|
|
|
2018-12-28 00:35:44 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2005-11-13 16:06:45 -08:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2018-12-28 00:35:44 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2016-03-17 14:19:14 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-12-28 00:34:24 -08:00
|
|
|
|
2016-03-17 14:19:14 -07:00
|
|
|
|
|
|
|
|
|
mm, page_alloc: reset the zone->watermark_boost early
Updating the zone watermarks by any means, like min_free_kbytes,
water_mark_scale_factor etc, when ->watermark_boost is set will result in
higher low and high watermarks than the user asked.
Below are the steps to reproduce the problem on system setup of Android
kernel running on Snapdragon hardware.
1) Default settings of the system are as below:
#cat /proc/sys/vm/min_free_kbytes = 5162
#cat /proc/zoneinfo | grep -e boost -e low -e "high " -e min -e Node
Node 0, zone Normal
min 797
low 8340
high 8539
2) Monitor the zone->watermark_boost(by adding a debug print in the
kernel) and whenever it is greater than zero value, write the same
value of min_free_kbytes obtained from step 1.
#echo 5162 > /proc/sys/vm/min_free_kbytes
3) Then read the zone watermarks in the system while the
->watermark_boost is zero. This should show the same values of
watermarks as step 1 but shown a higher values than asked.
#cat /proc/zoneinfo | grep -e boost -e low -e "high " -e min -e Node
Node 0, zone Normal
min 797
low 21148
high 21347
These higher values are because of updating the zone watermarks using the
macro min_wmark_pages(zone) which also adds the zone->watermark_boost.
#define min_wmark_pages(z) (z->_watermark[WMARK_MIN] +
z->watermark_boost)
So the steps that lead to the issue are:
1) On the extfrag event, watermarks are boosted by storing the required
value in ->watermark_boost.
2) User tries to update the zone watermarks level in the system through
min_free_kbytes or watermark_scale_factor.
3) Later, when kswapd woke up, it resets the zone->watermark_boost to
zero.
In step 2), we use the min_wmark_pages() macro to store the watermarks
in the zone structure thus the values are always offsetted by
->watermark_boost value. This can be avoided by resetting the
->watermark_boost to zero before it is used.
Signed-off-by: Charan Teja Reddy <charante@codeaurora.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Baoquan He <bhe@redhat.com>
Cc: Vinayak Menon <vinmenon@codeaurora.org>
Link: http://lkml.kernel.org/r/1589457511-4255-1-git-send-email-charante@codeaurora.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-03 15:59:14 -07:00
|
|
|
|
2018-12-28 00:35:44 -08:00
|
|
|
|
NUMA balancing: optimize page placement for memory tiering system
With the advent of various new memory types, some machines will have
multiple types of memory, e.g. DRAM and PMEM (persistent memory). The
memory subsystem of these machines can be called memory tiering system,
because the performance of the different types of memory are usually
different.
In such system, because of the memory accessing pattern changing etc,
some pages in the slow memory may become hot globally. So in this
patch, the NUMA balancing mechanism is enhanced to optimize the page
placement among the different memory types according to hot/cold
dynamically.
In a typical memory tiering system, there are CPUs, fast memory and slow
memory in each physical NUMA node. The CPUs and the fast memory will be
put in one logical node (called fast memory node), while the slow memory
will be put in another (faked) logical node (called slow memory node).
That is, the fast memory is regarded as local while the slow memory is
regarded as remote. So it's possible for the recently accessed pages in
the slow memory node to be promoted to the fast memory node via the
existing NUMA balancing mechanism.
The original NUMA balancing mechanism will stop to migrate pages if the
free memory of the target node becomes below the high watermark. This
is a reasonable policy if there's only one memory type. But this makes
the original NUMA balancing mechanism almost do not work to optimize
page placement among different memory types. Details are as follows.
It's the common cases that the working-set size of the workload is
larger than the size of the fast memory nodes. Otherwise, it's
unnecessary to use the slow memory at all. So, there are almost always
no enough free pages in the fast memory nodes, so that the globally hot
pages in the slow memory node cannot be promoted to the fast memory
node. To solve the issue, we have 2 choices as follows,
a. Ignore the free pages watermark checking when promoting hot pages
from the slow memory node to the fast memory node. This will
create some memory pressure in the fast memory node, thus trigger
the memory reclaiming. So that, the cold pages in the fast memory
node will be demoted to the slow memory node.
b. Define a new watermark called wmark_promo which is higher than
wmark_high, and have kswapd reclaiming pages until free pages reach
such watermark. The scenario is as follows: when we want to promote
hot-pages from a slow memory to a fast memory, but fast memory's free
pages would go lower than high watermark with such promotion, we wake
up kswapd with wmark_promo watermark in order to demote cold pages and
free us up some space. So, next time we want to promote hot-pages we
might have a chance of doing so.
The choice "a" may create high memory pressure in the fast memory node.
If the memory pressure of the workload is high, the memory pressure
may become so high that the memory allocation latency of the workload
is influenced, e.g. the direct reclaiming may be triggered.
The choice "b" works much better at this aspect. If the memory
pressure of the workload is high, the hot pages promotion will stop
earlier because its allocation watermark is higher than that of the
normal memory allocation. So in this patch, choice "b" is implemented.
A new zone watermark (WMARK_PROMO) is added. Which is larger than the
high watermark and can be controlled via watermark_scale_factor.
In addition to the original page placement optimization among sockets,
the NUMA balancing mechanism is extended to be used to optimize page
placement according to hot/cold among different memory types. So the
sysctl user space interface (numa_balancing) is extended in a backward
compatible way as follow, so that the users can enable/disable these
functionality individually.
The sysctl is converted from a Boolean value to a bits field. The
definition of the flags is,
- 0: NUMA_BALANCING_DISABLED
- 1: NUMA_BALANCING_NORMAL
- 2: NUMA_BALANCING_MEMORY_TIERING
We have tested the patch with the pmbench memory accessing benchmark
with the 80:20 read/write ratio and the Gauss access address
distribution on a 2 socket Intel server with Optane DC Persistent
Memory Model. The test results shows that the pmbench score can
improve up to 95.9%.
Thanks Andrew Morton to help fix the document format error.
Link: https://lkml.kernel.org/r/20220221084529.1052339-3-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Reviewed-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Yang Shi <shy828301@gmail.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Zi Yan <ziy@nvidia.com>
Cc: Wei Xu <weixugc@google.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: zhongjiang-ali <zhongjiang-ali@linux.alibaba.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Feng Tang <feng.tang@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-03-22 14:46:23 -07:00
|
|
|
|
|
|
|
|
|
2025-03-08 03:46:00 +00:00
|
|
|
|
2012-01-25 12:49:24 +01:00
|
|
|
|
2008-10-18 20:27:11 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2006-04-10 22:52:59 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2011-04-25 21:36:42 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-28 19:42:12 -07:00
|
|
|
|
2017-09-06 16:20:37 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-04-25 21:36:42 +00:00
|
|
|
|
2017-09-06 16:20:37 -07:00
|
|
|
|
2021-06-28 19:42:12 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-28 19:42:15 -07:00
|
|
|
|
2011-04-25 21:36:42 +00:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-07-03 15:15:30 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2013-09-11 14:20:34 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-11-05 13:41:36 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2013-07-08 16:00:40 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2013-07-08 16:00:40 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2021-11-05 13:40:55 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-07-08 16:00:40 -07:00
|
|
|
|
|
|
|
|
|
2021-11-05 13:40:55 -07:00
|
|
|
|
2021-11-05 13:41:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2009-06-16 15:32:48 -07:00
|
|
|
|
2011-05-24 17:11:33 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2016-08-10 16:27:49 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-10-10 23:16:40 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2020-08-20 17:42:24 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2013-09-11 14:20:34 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
sysctl: treewide: constify the ctl_table argument of proc_handlers
const qualify the struct ctl_table argument in the proc_handler function
signatures. This is a prerequisite to moving the static ctl_table
structs into .rodata data which will ensure that proc_handler function
pointers cannot be modified.
This patch has been generated by the following coccinelle script:
```
virtual patch
@r1@
identifier ctl, write, buffer, lenp, ppos;
identifier func !~ "appldata_(timer|interval)_handler|sched_(rt|rr)_handler|rds_tcp_skbuf_handler|proc_sctp_do_(hmac_alg|rto_min|rto_max|udp_port|alpha_beta|auth|probe_interval)";
@@
int func(
- struct ctl_table *ctl
+ const struct ctl_table *ctl
,int write, void *buffer, size_t *lenp, loff_t *ppos);
@r2@
identifier func, ctl, write, buffer, lenp, ppos;
@@
int func(
- struct ctl_table *ctl
+ const struct ctl_table *ctl
,int write, void *buffer, size_t *lenp, loff_t *ppos)
{ ... }
@r3@
identifier func;
@@
int func(
- struct ctl_table *
+ const struct ctl_table *
,int , void *, size_t *, loff_t *);
@r4@
identifier func, ctl;
@@
int func(
- struct ctl_table *ctl
+ const struct ctl_table *ctl
,int , void *, size_t *, loff_t *);
@r5@
identifier func, write, buffer, lenp, ppos;
@@
int func(
- struct ctl_table *
+ const struct ctl_table *
,int write, void *buffer, size_t *lenp, loff_t *ppos);
```
* Code formatting was adjusted in xfs_sysctl.c to comply with code
conventions. The xfs_stats_clear_proc_handler,
xfs_panic_mask_proc_handler and xfs_deprecated_dointvec_minmax where
adjusted.
* The ctl_table argument in proc_watchdog_common was const qualified.
This is called from a proc_handler itself and is calling back into
another proc_handler, making it necessary to change it as part of the
proc_handler migration.
Co-developed-by: Thomas Weißschuh <linux@weissschuh.net>
Signed-off-by: Thomas Weißschuh <linux@weissschuh.net>
Co-developed-by: Joel Granados <j.granados@samsung.com>
Signed-off-by: Joel Granados <j.granados@samsung.com>
2024-07-24 20:59:29 +02:00
|
|
|
|
2020-04-24 08:43:38 +02:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2014-01-23 15:53:17 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-07-08 16:00:40 -07:00
|
|
|
|
|
|
|
|
|
2009-06-16 15:32:48 -07:00
|
|
|
|
2013-07-08 16:00:40 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
sysctl: treewide: constify the ctl_table argument of proc_handlers
const qualify the struct ctl_table argument in the proc_handler function
signatures. This is a prerequisite to moving the static ctl_table
structs into .rodata data which will ensure that proc_handler function
pointers cannot be modified.
This patch has been generated by the following coccinelle script:
```
virtual patch
@r1@
identifier ctl, write, buffer, lenp, ppos;
identifier func !~ "appldata_(timer|interval)_handler|sched_(rt|rr)_handler|rds_tcp_skbuf_handler|proc_sctp_do_(hmac_alg|rto_min|rto_max|udp_port|alpha_beta|auth|probe_interval)";
@@
int func(
- struct ctl_table *ctl
+ const struct ctl_table *ctl
,int write, void *buffer, size_t *lenp, loff_t *ppos);
@r2@
identifier func, ctl, write, buffer, lenp, ppos;
@@
int func(
- struct ctl_table *ctl
+ const struct ctl_table *ctl
,int write, void *buffer, size_t *lenp, loff_t *ppos)
{ ... }
@r3@
identifier func;
@@
int func(
- struct ctl_table *
+ const struct ctl_table *
,int , void *, size_t *, loff_t *);
@r4@
identifier func, ctl;
@@
int func(
- struct ctl_table *ctl
+ const struct ctl_table *ctl
,int , void *, size_t *, loff_t *);
@r5@
identifier func, write, buffer, lenp, ppos;
@@
int func(
- struct ctl_table *
+ const struct ctl_table *
,int write, void *buffer, size_t *lenp, loff_t *ppos);
```
* Code formatting was adjusted in xfs_sysctl.c to comply with code
conventions. The xfs_stats_clear_proc_handler,
xfs_panic_mask_proc_handler and xfs_deprecated_dointvec_minmax where
adjusted.
* The ctl_table argument in proc_watchdog_common was const qualified.
This is called from a proc_handler itself and is calling back into
another proc_handler, making it necessary to change it as part of the
proc_handler migration.
Co-developed-by: Thomas Weißschuh <linux@weissschuh.net>
Signed-off-by: Thomas Weißschuh <linux@weissschuh.net>
Co-developed-by: Joel Granados <j.granados@samsung.com>
Signed-off-by: Joel Granados <j.granados@samsung.com>
2024-07-24 20:59:29 +02:00
|
|
|
|
2020-04-24 08:43:38 +02:00
|
|
|
|
2016-03-17 14:19:14 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-07-03 00:24:13 -07:00
|
|
|
|
2016-08-10 16:27:49 -07:00
|
|
|
|
2006-07-03 00:24:13 -07:00
|
|
|
|
2016-08-10 16:27:49 -07:00
|
|
|
|
2006-07-03 00:24:13 -07:00
|
|
|
|
|
|
|
|
|
2016-07-28 15:46:32 -07:00
|
|
|
|
2016-08-10 16:27:46 -07:00
|
|
|
|
2016-07-28 15:46:32 -07:00
|
|
|
|
2006-07-03 00:24:13 -07:00
|
|
|
|
2018-12-28 00:34:24 -08:00
|
|
|
|
|
|
|
|
|
2006-07-03 00:24:13 -07:00
|
|
|
|
2006-09-25 23:31:52 -07:00
|
|
|
|
2016-08-10 16:27:49 -07:00
|
|
|
|
sysctl: treewide: constify the ctl_table argument of proc_handlers
const qualify the struct ctl_table argument in the proc_handler function
signatures. This is a prerequisite to moving the static ctl_table
structs into .rodata data which will ensure that proc_handler function
pointers cannot be modified.
This patch has been generated by the following coccinelle script:
```
virtual patch
@r1@
identifier ctl, write, buffer, lenp, ppos;
identifier func !~ "appldata_(timer|interval)_handler|sched_(rt|rr)_handler|rds_tcp_skbuf_handler|proc_sctp_do_(hmac_alg|rto_min|rto_max|udp_port|alpha_beta|auth|probe_interval)";
@@
int func(
- struct ctl_table *ctl
+ const struct ctl_table *ctl
,int write, void *buffer, size_t *lenp, loff_t *ppos);
@r2@
identifier func, ctl, write, buffer, lenp, ppos;
@@
int func(
- struct ctl_table *ctl
+ const struct ctl_table *ctl
,int write, void *buffer, size_t *lenp, loff_t *ppos)
{ ... }
@r3@
identifier func;
@@
int func(
- struct ctl_table *
+ const struct ctl_table *
,int , void *, size_t *, loff_t *);
@r4@
identifier func, ctl;
@@
int func(
- struct ctl_table *ctl
+ const struct ctl_table *ctl
,int , void *, size_t *, loff_t *);
@r5@
identifier func, write, buffer, lenp, ppos;
@@
int func(
- struct ctl_table *
+ const struct ctl_table *
,int write, void *buffer, size_t *lenp, loff_t *ppos);
```
* Code formatting was adjusted in xfs_sysctl.c to comply with code
conventions. The xfs_stats_clear_proc_handler,
xfs_panic_mask_proc_handler and xfs_deprecated_dointvec_minmax where
adjusted.
* The ctl_table argument in proc_watchdog_common was const qualified.
This is called from a proc_handler itself and is calling back into
another proc_handler, making it necessary to change it as part of the
proc_handler migration.
Co-developed-by: Thomas Weißschuh <linux@weissschuh.net>
Signed-off-by: Thomas Weißschuh <linux@weissschuh.net>
Co-developed-by: Joel Granados <j.granados@samsung.com>
Signed-off-by: Joel Granados <j.granados@samsung.com>
2024-07-24 20:59:29 +02:00
|
|
|
|
2020-04-24 08:43:38 +02:00
|
|
|
|
2006-09-25 23:31:52 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2009-09-23 15:57:19 -07:00
|
|
|
|
2006-09-25 23:31:52 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-08-10 16:27:49 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-07-28 15:46:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-09-25 23:31:52 -07:00
|
|
|
|
2018-12-28 00:34:24 -08:00
|
|
|
|
|
|
|
|
|
2016-08-10 16:27:49 -07:00
|
|
|
|
|
|
|
|
|
sysctl: treewide: constify the ctl_table argument of proc_handlers
const qualify the struct ctl_table argument in the proc_handler function
signatures. This is a prerequisite to moving the static ctl_table
structs into .rodata data which will ensure that proc_handler function
pointers cannot be modified.
This patch has been generated by the following coccinelle script:
```
virtual patch
@r1@
identifier ctl, write, buffer, lenp, ppos;
identifier func !~ "appldata_(timer|interval)_handler|sched_(rt|rr)_handler|rds_tcp_skbuf_handler|proc_sctp_do_(hmac_alg|rto_min|rto_max|udp_port|alpha_beta|auth|probe_interval)";
@@
int func(
- struct ctl_table *ctl
+ const struct ctl_table *ctl
,int write, void *buffer, size_t *lenp, loff_t *ppos);
@r2@
identifier func, ctl, write, buffer, lenp, ppos;
@@
int func(
- struct ctl_table *ctl
+ const struct ctl_table *ctl
,int write, void *buffer, size_t *lenp, loff_t *ppos)
{ ... }
@r3@
identifier func;
@@
int func(
- struct ctl_table *
+ const struct ctl_table *
,int , void *, size_t *, loff_t *);
@r4@
identifier func, ctl;
@@
int func(
- struct ctl_table *ctl
+ const struct ctl_table *ctl
,int , void *, size_t *, loff_t *);
@r5@
identifier func, write, buffer, lenp, ppos;
@@
int func(
- struct ctl_table *
+ const struct ctl_table *
,int write, void *buffer, size_t *lenp, loff_t *ppos);
```
* Code formatting was adjusted in xfs_sysctl.c to comply with code
conventions. The xfs_stats_clear_proc_handler,
xfs_panic_mask_proc_handler and xfs_deprecated_dointvec_minmax where
adjusted.
* The ctl_table argument in proc_watchdog_common was const qualified.
This is called from a proc_handler itself and is calling back into
another proc_handler, making it necessary to change it as part of the
proc_handler migration.
Co-developed-by: Thomas Weißschuh <linux@weissschuh.net>
Signed-off-by: Thomas Weißschuh <linux@weissschuh.net>
Co-developed-by: Joel Granados <j.granados@samsung.com>
Signed-off-by: Joel Granados <j.granados@samsung.com>
2024-07-24 20:59:29 +02:00
|
|
|
|
2020-04-24 08:43:38 +02:00
|
|
|
|
2016-08-10 16:27:49 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-09-25 23:31:52 -07:00
|
|
|
|
|
|
|
|
|
2006-07-03 00:24:13 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2009-06-16 15:32:12 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
sysctl: treewide: constify the ctl_table argument of proc_handlers
const qualify the struct ctl_table argument in the proc_handler function
signatures. This is a prerequisite to moving the static ctl_table
structs into .rodata data which will ensure that proc_handler function
pointers cannot be modified.
This patch has been generated by the following coccinelle script:
```
virtual patch
@r1@
identifier ctl, write, buffer, lenp, ppos;
identifier func !~ "appldata_(timer|interval)_handler|sched_(rt|rr)_handler|rds_tcp_skbuf_handler|proc_sctp_do_(hmac_alg|rto_min|rto_max|udp_port|alpha_beta|auth|probe_interval)";
@@
int func(
- struct ctl_table *ctl
+ const struct ctl_table *ctl
,int write, void *buffer, size_t *lenp, loff_t *ppos);
@r2@
identifier func, ctl, write, buffer, lenp, ppos;
@@
int func(
- struct ctl_table *ctl
+ const struct ctl_table *ctl
,int write, void *buffer, size_t *lenp, loff_t *ppos)
{ ... }
@r3@
identifier func;
@@
int func(
- struct ctl_table *
+ const struct ctl_table *
,int , void *, size_t *, loff_t *);
@r4@
identifier func, ctl;
@@
int func(
- struct ctl_table *ctl
+ const struct ctl_table *ctl
,int , void *, size_t *, loff_t *);
@r5@
identifier func, write, buffer, lenp, ppos;
@@
int func(
- struct ctl_table *
+ const struct ctl_table *
,int write, void *buffer, size_t *lenp, loff_t *ppos);
```
* Code formatting was adjusted in xfs_sysctl.c to comply with code
conventions. The xfs_stats_clear_proc_handler,
xfs_panic_mask_proc_handler and xfs_deprecated_dointvec_minmax where
adjusted.
* The ctl_table argument in proc_watchdog_common was const qualified.
This is called from a proc_handler itself and is calling back into
another proc_handler, making it necessary to change it as part of the
proc_handler migration.
Co-developed-by: Thomas Weißschuh <linux@weissschuh.net>
Signed-off-by: Thomas Weißschuh <linux@weissschuh.net>
Co-developed-by: Joel Granados <j.granados@samsung.com>
Signed-off-by: Joel Granados <j.granados@samsung.com>
2024-07-24 20:59:29 +02:00
|
|
|
|
2023-05-16 14:38:20 +08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2020-06-03 15:58:48 -07:00
|
|
|
|
|
|
|
|
|
2009-09-23 15:57:19 -07:00
|
|
|
|
2020-06-03 15:58:48 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-01-08 01:00:40 -08:00
|
|
|
|
2021-06-28 19:42:24 -07:00
|
|
|
|
|
|
|
|
|
2013-09-11 14:20:34 -07:00
|
|
|
|
2006-01-08 01:00:40 -08:00
|
|
|
|
sysctl: treewide: constify the ctl_table argument of proc_handlers
const qualify the struct ctl_table argument in the proc_handler function
signatures. This is a prerequisite to moving the static ctl_table
structs into .rodata data which will ensure that proc_handler function
pointers cannot be modified.
This patch has been generated by the following coccinelle script:
```
virtual patch
@r1@
identifier ctl, write, buffer, lenp, ppos;
identifier func !~ "appldata_(timer|interval)_handler|sched_(rt|rr)_handler|rds_tcp_skbuf_handler|proc_sctp_do_(hmac_alg|rto_min|rto_max|udp_port|alpha_beta|auth|probe_interval)";
@@
int func(
- struct ctl_table *ctl
+ const struct ctl_table *ctl
,int write, void *buffer, size_t *lenp, loff_t *ppos);
@r2@
identifier func, ctl, write, buffer, lenp, ppos;
@@
int func(
- struct ctl_table *ctl
+ const struct ctl_table *ctl
,int write, void *buffer, size_t *lenp, loff_t *ppos)
{ ... }
@r3@
identifier func;
@@
int func(
- struct ctl_table *
+ const struct ctl_table *
,int , void *, size_t *, loff_t *);
@r4@
identifier func, ctl;
@@
int func(
- struct ctl_table *ctl
+ const struct ctl_table *ctl
,int , void *, size_t *, loff_t *);
@r5@
identifier func, write, buffer, lenp, ppos;
@@
int func(
- struct ctl_table *
+ const struct ctl_table *
,int write, void *buffer, size_t *lenp, loff_t *ppos);
```
* Code formatting was adjusted in xfs_sysctl.c to comply with code
conventions. The xfs_stats_clear_proc_handler,
xfs_panic_mask_proc_handler and xfs_deprecated_dointvec_minmax where
adjusted.
* The ctl_table argument in proc_watchdog_common was const qualified.
This is called from a proc_handler itself and is calling back into
another proc_handler, making it necessary to change it as part of the
proc_handler migration.
Co-developed-by: Thomas Weißschuh <linux@weissschuh.net>
Signed-off-by: Thomas Weißschuh <linux@weissschuh.net>
Co-developed-by: Joel Granados <j.granados@samsung.com>
Signed-off-by: Joel Granados <j.granados@samsung.com>
2024-07-24 20:59:29 +02:00
|
|
|
|
2021-06-28 19:42:24 -07:00
|
|
|
|
2006-01-08 01:00:40 -08:00
|
|
|
|
|
|
|
|
|
2021-06-28 19:42:24 -07:00
|
|
|
|
2006-01-08 01:00:40 -08:00
|
|
|
|
|
|
|
|
|
2013-07-03 15:01:29 -07:00
|
|
|
|
2021-06-28 19:42:24 -07:00
|
|
|
|
2014-06-23 13:22:04 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-28 19:42:24 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-06-23 13:22:04 -07:00
|
|
|
|
|
|
|
|
|
2006-01-08 01:00:40 -08:00
|
|
|
|
2014-06-23 13:22:04 -07:00
|
|
|
|
|
|
|
|
|
2021-06-28 19:42:24 -07:00
|
|
|
|
2014-06-23 13:22:04 -07:00
|
|
|
|
|
|
|
|
|
2019-11-30 17:55:11 -08:00
|
|
|
|
2021-06-28 19:42:24 -07:00
|
|
|
|
2014-06-23 13:22:04 -07:00
|
|
|
|
2013-07-03 15:01:29 -07:00
|
|
|
|
2014-06-23 13:22:04 -07:00
|
|
|
|
2006-01-08 01:00:40 -08:00
|
|
|
|
|
|
|
|
|
2025-01-28 13:48:37 +01:00
|
|
|
|
2023-05-16 14:38:20 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: page_alloc: defrag_mode
The page allocator groups requests by migratetype to stave off
fragmentation. However, in practice this is routinely defeated by the
fact that it gives up *before* invoking reclaim and compaction - which may
well produce suitable pages. As a result, fragmentation of physical
memory is a common ongoing process in many load scenarios.
Fragmentation deteriorates compaction's ability to produce huge pages.
Depending on the lifetime of the fragmenting allocations, those effects
can be long-lasting or even permanent, requiring drastic measures like
forcible idle states or even reboots as the only reliable ways to recover
the address space for THP production.
In a kernel build test with supplemental THP pressure, the THP allocation
rate steadily declines over 15 runs:
thp_fault_alloc
61988
56474
57258
50187
52388
55409
52925
47648
43669
40621
36077
41721
36685
34641
33215
This is a hurdle in adopting THP in any environment where hosts are shared
between multiple overlapping workloads (cloud environments), and rarely
experience true idle periods. To make THP a reliable and predictable
optimization, there needs to be a stronger guarantee to avoid such
fragmentation.
Introduce defrag_mode. When enabled, reclaim/compaction is invoked to its
full extent *before* falling back. Specifically, ALLOC_NOFRAGMENT is
enforced on the allocator fastpath and the reclaiming slowpath.
For now, fallbacks are permitted to avert OOMs. There is a plan to add
defrag_mode=2 to prefer OOMs over fragmentation, but this requires
additional prep work in compaction and the reserve management to make it
ready for all possible allocation contexts.
The following test results are from a kernel build with periodic bursts of
THP allocations, over 15 runs:
vanilla defrag_mode=1
@claimer[unmovable]: 189 103
@claimer[movable]: 92 103
@claimer[reclaimable]: 207 61
@pollute[unmovable from movable]: 25 0
@pollute[unmovable from reclaimable]: 28 0
@pollute[movable from unmovable]: 38835 0
@pollute[movable from reclaimable]: 147136 0
@pollute[reclaimable from unmovable]: 178 0
@pollute[reclaimable from movable]: 33 0
@steal[unmovable from movable]: 11 0
@steal[unmovable from reclaimable]: 5 0
@steal[reclaimable from unmovable]: 107 0
@steal[reclaimable from movable]: 90 0
@steal[movable from reclaimable]: 354 0
@steal[movable from unmovable]: 130 0
Both types of polluting fallbacks are eliminated in this workload.
Interestingly, whole block conversions are reduced as well. This is
because once a block is claimed for a type, its empty space remains
available for future allocations, instead of being padded with fallbacks;
this allows the native type to group up instead of spreading out to new
blocks. The assumption in the allocator has been that pollution from
movable allocations is less harmful than from other types, since they can
be reclaimed or migrated out should the space be needed. However, since
fallbacks occur *before* reclaim/compaction is invoked, movable pollution
will still cause non-movable allocations to spread out and claim more
blocks.
Without fragmentation, THP rates hold steady with defrag_mode=1:
thp_fault_alloc
32478
20725
45045
32130
14018
21711
40791
29134
34458
45381
28305
17265
22584
28454
30850
While the downward trend is eliminated, the keen reader will of course
notice that the baseline rate is much smaller than the vanilla kernel's to
begin with. This is due to deficiencies in how reclaim and compaction are
currently driven: ALLOC_NOFRAGMENT increases the extent to which smaller
allocations are competing with THPs for pageblocks, while making no effort
themselves to reclaim or compact beyond their own request size. This
effect already exists with the current usage of ALLOC_NOFRAGMENT, but is
amplified by defrag_mode insisting on whole block stealing much more
strongly.
Subsequent patches will address defrag_mode reclaim strategy to raise the
THP success baseline above the vanilla kernel.
Link: https://lkml.kernel.org/r/20250313210647.1314586-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-03-13 17:05:34 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-05-16 14:38:20 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-05-13 17:19:00 -07:00
|
|
|
|
2021-04-29 23:01:30 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-07-22 15:46:49 -04:00
|
|
|
|
2024-12-03 10:47:29 +01:00
|
|
|
|
2025-07-22 15:46:49 -04:00
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
|
|
|
|
|
2020-06-03 16:01:18 -07:00
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-08-11 18:37:31 -07:00
|
|
|
|
|
|
|
|
|
2024-12-03 10:47:30 +01:00
|
|
|
|
2024-03-06 18:13:26 +08:00
|
|
|
|
2020-08-11 18:37:31 -07:00
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
2021-05-04 18:36:57 -07:00
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
2012-10-08 16:32:41 -07:00
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-10-08 16:32:41 -07:00
|
|
|
|
|
|
|
|
|
2021-05-04 18:35:17 -07:00
|
|
|
|
|
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
2021-05-04 18:35:17 -07:00
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
|
|
|
|
|
mm,page_alloc: bail out earlier on -ENOMEM in alloc_contig_migrate_range
Patch series "Make alloc_contig_range handle Hugetlb pages", v10.
alloc_contig_range lacks the ability to handle HugeTLB pages. This can
be problematic for some users, e.g: CMA and virtio-mem, where those
users will fail the call if alloc_contig_range ever sees a HugeTLB page,
even when those pages lay in ZONE_MOVABLE and are free. That problem
can be easily solved by replacing the page in the free hugepage pool.
In-use HugeTLB are no exception though, as those can be isolated and
migrated as any other LRU or Movable page.
This aims to improve alloc_contig_range->isolate_migratepages_block, so
that HugeTLB pages can be recognized and handled.
Since we also need to start reporting errors down the chain (e.g:
-ENOMEM due to not be able to allocate a new hugetlb page),
isolate_migratepages_{range,block} interfaces need to change to start
reporting error codes instead of the pfn == 0 vs pfn != 0 scheme it is
using right now. From now on, isolate_migratepages_block will not
return the next pfn to be scanned anymore, but -EINTR, -ENOMEM or 0, so
we the next pfn to be scanned will be recorded in cc->migrate_pfn field
(as it is already done in isolate_migratepages_range()).
Below is an insight from David (thanks), where the problem can clearly be
seen:
"Start a VM with 4G. Hotplug 1G via virtio-mem and online it to
ZONE_MOVABLE. Allocate 512 huge pages.
[root@localhost ~]# cat /proc/meminfo
MemTotal: 5061512 kB
MemFree: 3319396 kB
MemAvailable: 3457144 kB
...
HugePages_Total: 512
HugePages_Free: 512
HugePages_Rsvd: 0
HugePages_Surp: 0
Hugepagesize: 2048 kB
The huge pages get partially allocate from ZONE_MOVABLE. Try unplugging
1G via virtio-mem (remember, all ZONE_MOVABLE). Inside the guest:
[ 180.058992] alloc_contig_range: [1b8000, 1c0000) PFNs busy
[ 180.060531] alloc_contig_range: [1b8000, 1c0000) PFNs busy
[ 180.061972] alloc_contig_range: [1b8000, 1c0000) PFNs busy
[ 180.063413] alloc_contig_range: [1b8000, 1c0000) PFNs busy
[ 180.064838] alloc_contig_range: [1b8000, 1c0000) PFNs busy
[ 180.065848] alloc_contig_range: [1bfc00, 1c0000) PFNs busy
[ 180.066794] alloc_contig_range: [1bfc00, 1c0000) PFNs busy
[ 180.067738] alloc_contig_range: [1bfc00, 1c0000) PFNs busy
[ 180.068669] alloc_contig_range: [1bfc00, 1c0000) PFNs busy
[ 180.069598] alloc_contig_range: [1bfc00, 1c0000) PFNs busy"
And then with this patchset running:
"Same experiment with ZONE_MOVABLE:
a) Free huge pages: all memory can get unplugged again.
b) Allocated/populated but idle huge pages: all memory can get unplugged
again.
c) Allocated/populated but all 512 huge pages are read/written in a
loop: all memory can get unplugged again, but I get a single
[ 121.192345] alloc_contig_range: [180000, 188000) PFNs busy
Most probably because it happened to try migrating a huge page
while it was busy. As virtio-mem retries on ZONE_MOVABLE a couple of
times, it can deal with this temporary failure.
Last but not least, I did something extreme:
# cat /proc/meminfo
MemTotal: 5061568 kB
MemFree: 186560 kB
MemAvailable: 354524 kB
...
HugePages_Total: 2048
HugePages_Free: 2048
HugePages_Rsvd: 0
HugePages_Surp: 0
Triggering unplug would require to dissolve+alloc - which now fails
when trying to allocate an additional ~512 huge pages (1G).
As expected, I can properly see memory unplug not fully succeeding. +
I get a fairly continuous stream of
[ 226.611584] alloc_contig_range: [19f400, 19f800) PFNs busy
...
But more importantly, the hugepage count remains stable, as configured
by the admin (me):
HugePages_Total: 2048
HugePages_Free: 2048
HugePages_Rsvd: 0
HugePages_Surp: 0"
This patch (of 7):
Currently, __alloc_contig_migrate_range can generate -EINTR, -ENOMEM or
-EBUSY, and report them down the chain. The problem is that when
migrate_pages() reports -ENOMEM, we keep going till we exhaust all the
try-attempts (5 at the moment) instead of bailing out.
migrate_pages() bails out right away on -ENOMEM because it is considered a
fatal error. Do the same here instead of keep going and retrying. Note
that this is not fixing a real issue, just a cosmetic change. Although we
can save some cycles by backing off ealier
Link: https://lkml.kernel.org/r/20210419075413.1064-1-osalvador@suse.de
Link: https://lkml.kernel.org/r/20210419075413.1064-2-osalvador@suse.de
Signed-off-by: Oscar Salvador <osalvador@suse.de>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: David Hildenbrand <david@redhat.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Muchun Song <songmuchun@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-05-04 18:35:14 -07:00
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-10-08 16:33:51 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-10-08 16:31:55 -07:00
|
|
|
|
2020-08-11 18:37:31 -07:00
|
|
|
|
2021-09-02 14:59:13 -07:00
|
|
|
|
mm,page_alloc: bail out earlier on -ENOMEM in alloc_contig_migrate_range
Patch series "Make alloc_contig_range handle Hugetlb pages", v10.
alloc_contig_range lacks the ability to handle HugeTLB pages. This can
be problematic for some users, e.g: CMA and virtio-mem, where those
users will fail the call if alloc_contig_range ever sees a HugeTLB page,
even when those pages lay in ZONE_MOVABLE and are free. That problem
can be easily solved by replacing the page in the free hugepage pool.
In-use HugeTLB are no exception though, as those can be isolated and
migrated as any other LRU or Movable page.
This aims to improve alloc_contig_range->isolate_migratepages_block, so
that HugeTLB pages can be recognized and handled.
Since we also need to start reporting errors down the chain (e.g:
-ENOMEM due to not be able to allocate a new hugetlb page),
isolate_migratepages_{range,block} interfaces need to change to start
reporting error codes instead of the pfn == 0 vs pfn != 0 scheme it is
using right now. From now on, isolate_migratepages_block will not
return the next pfn to be scanned anymore, but -EINTR, -ENOMEM or 0, so
we the next pfn to be scanned will be recorded in cc->migrate_pfn field
(as it is already done in isolate_migratepages_range()).
Below is an insight from David (thanks), where the problem can clearly be
seen:
"Start a VM with 4G. Hotplug 1G via virtio-mem and online it to
ZONE_MOVABLE. Allocate 512 huge pages.
[root@localhost ~]# cat /proc/meminfo
MemTotal: 5061512 kB
MemFree: 3319396 kB
MemAvailable: 3457144 kB
...
HugePages_Total: 512
HugePages_Free: 512
HugePages_Rsvd: 0
HugePages_Surp: 0
Hugepagesize: 2048 kB
The huge pages get partially allocate from ZONE_MOVABLE. Try unplugging
1G via virtio-mem (remember, all ZONE_MOVABLE). Inside the guest:
[ 180.058992] alloc_contig_range: [1b8000, 1c0000) PFNs busy
[ 180.060531] alloc_contig_range: [1b8000, 1c0000) PFNs busy
[ 180.061972] alloc_contig_range: [1b8000, 1c0000) PFNs busy
[ 180.063413] alloc_contig_range: [1b8000, 1c0000) PFNs busy
[ 180.064838] alloc_contig_range: [1b8000, 1c0000) PFNs busy
[ 180.065848] alloc_contig_range: [1bfc00, 1c0000) PFNs busy
[ 180.066794] alloc_contig_range: [1bfc00, 1c0000) PFNs busy
[ 180.067738] alloc_contig_range: [1bfc00, 1c0000) PFNs busy
[ 180.068669] alloc_contig_range: [1bfc00, 1c0000) PFNs busy
[ 180.069598] alloc_contig_range: [1bfc00, 1c0000) PFNs busy"
And then with this patchset running:
"Same experiment with ZONE_MOVABLE:
a) Free huge pages: all memory can get unplugged again.
b) Allocated/populated but idle huge pages: all memory can get unplugged
again.
c) Allocated/populated but all 512 huge pages are read/written in a
loop: all memory can get unplugged again, but I get a single
[ 121.192345] alloc_contig_range: [180000, 188000) PFNs busy
Most probably because it happened to try migrating a huge page
while it was busy. As virtio-mem retries on ZONE_MOVABLE a couple of
times, it can deal with this temporary failure.
Last but not least, I did something extreme:
# cat /proc/meminfo
MemTotal: 5061568 kB
MemFree: 186560 kB
MemAvailable: 354524 kB
...
HugePages_Total: 2048
HugePages_Free: 2048
HugePages_Rsvd: 0
HugePages_Surp: 0
Triggering unplug would require to dissolve+alloc - which now fails
when trying to allocate an additional ~512 huge pages (1G).
As expected, I can properly see memory unplug not fully succeeding. +
I get a fairly continuous stream of
[ 226.611584] alloc_contig_range: [19f400, 19f800) PFNs busy
...
But more importantly, the hugepage count remains stable, as configured
by the admin (me):
HugePages_Total: 2048
HugePages_Free: 2048
HugePages_Rsvd: 0
HugePages_Surp: 0"
This patch (of 7):
Currently, __alloc_contig_migrate_range can generate -EINTR, -ENOMEM or
-EBUSY, and report them down the chain. The problem is that when
migrate_pages() reports -ENOMEM, we keep going till we exhaust all the
try-attempts (5 at the moment) instead of bailing out.
migrate_pages() bails out right away on -ENOMEM because it is considered a
fatal error. Do the same here instead of keep going and retrying. Note
that this is not fixing a real issue, just a cosmetic change. Although we
can save some cycles by backing off ealier
Link: https://lkml.kernel.org/r/20210419075413.1064-1-osalvador@suse.de
Link: https://lkml.kernel.org/r/20210419075413.1064-2-osalvador@suse.de
Signed-off-by: Oscar Salvador <osalvador@suse.de>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: David Hildenbrand <david@redhat.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Muchun Song <songmuchun@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-05-04 18:35:14 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
mm: disable LRU pagevec during the migration temporarily
LRU pagevec holds refcount of pages until the pagevec are drained. It
could prevent migration since the refcount of the page is greater than
the expection in migration logic. To mitigate the issue, callers of
migrate_pages drains LRU pagevec via migrate_prep or lru_add_drain_all
before migrate_pages call.
However, it's not enough because pages coming into pagevec after the
draining call still could stay at the pagevec so it could keep
preventing page migration. Since some callers of migrate_pages have
retrial logic with LRU draining, the page would migrate at next trail
but it is still fragile in that it doesn't close the fundamental race
between upcoming LRU pages into pagvec and migration so the migration
failure could cause contiguous memory allocation failure in the end.
To close the race, this patch disables lru caches(i.e, pagevec) during
ongoing migration until migrate is done.
Since it's really hard to reproduce, I measured how many times
migrate_pages retried with force mode(it is about a fallback to a sync
migration) with below debug code.
int migrate_pages(struct list_head *from, new_page_t get_new_page,
..
..
if (rc && reason == MR_CONTIG_RANGE && pass > 2) {
printk(KERN_ERR, "pfn 0x%lx reason %d", page_to_pfn(page), rc);
dump_page(page, "fail to migrate");
}
The test was repeating android apps launching with cma allocation in
background every five seconds. Total cma allocation count was about 500
during the testing. With this patch, the dump_page count was reduced
from 400 to 30.
The new interface is also useful for memory hotplug which currently
drains lru pcp caches after each migration failure. This is rather
suboptimal as it has to disrupt others running during the operation.
With the new interface the operation happens only once. This is also in
line with pcp allocator cache which are disabled for the offlining as
well.
Link: https://lkml.kernel.org/r/20210319175127.886124-1-minchan@kernel.org
Signed-off-by: Minchan Kim <minchan@kernel.org>
Reviewed-by: Chris Goldsworthy <cgoldswo@codeaurora.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: John Dias <joaodias@google.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: David Hildenbrand <david@redhat.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Oliver Sang <oliver.sang@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-05-04 18:36:54 -07:00
|
|
|
|
2021-05-04 18:36:57 -07:00
|
|
|
|
2013-02-22 16:32:09 -08:00
|
|
|
|
2022-05-19 14:08:55 -07:00
|
|
|
|
2021-06-28 19:42:06 -07:00
|
|
|
|
2013-02-22 16:32:09 -08:00
|
|
|
|
|
|
|
|
|
2024-02-28 05:11:17 +00:00
|
|
|
|
|
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
|
|
|
|
|
2024-12-03 10:47:31 +01:00
|
|
|
|
2024-08-13 21:54:49 -06:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-12-03 10:47:31 +01:00
|
|
|
|
2024-11-25 21:01:36 +00:00
|
|
|
|
2024-08-13 21:54:49 -06:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-12-03 10:47:30 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-12-03 10:47:31 +01:00
|
|
|
|
|
|
|
|
|
2024-12-03 10:47:30 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-12-05 10:05:07 +01:00
|
|
|
|
|
|
|
|
|
2024-12-03 10:47:30 +01:00
|
|
|
|
|
|
|
|
|
2024-12-05 10:05:07 +01:00
|
|
|
|
2024-12-03 10:47:30 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-06-16 22:11:14 -04:00
|
|
|
|
2024-12-03 10:47:30 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
2022-05-12 20:22:58 -07:00
|
|
|
|
|
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
2018-04-05 16:25:26 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
2019-03-05 15:48:42 -08:00
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
2025-06-16 22:11:14 -04:00
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
2025-09-01 17:03:27 +02:00
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
2015-11-06 16:29:57 -08:00
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
2012-10-08 16:32:41 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-06-04 16:08:28 -07:00
|
|
|
|
2012-10-08 16:32:41 -07:00
|
|
|
|
2017-11-17 15:26:38 -08:00
|
|
|
|
2020-04-01 21:10:28 -07:00
|
|
|
|
2012-10-08 16:32:41 -07:00
|
|
|
|
|
|
|
|
|
2025-06-16 22:11:14 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-10-08 16:32:41 -07:00
|
|
|
|
2025-09-01 17:03:27 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-12-03 10:47:30 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-05-12 20:22:58 -07:00
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-05-12 20:22:58 -07:00
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-06-16 22:11:14 -04:00
|
|
|
|
2020-10-15 20:08:07 -07:00
|
|
|
|
2022-05-12 20:22:58 -07:00
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
2020-12-14 19:10:56 -08:00
|
|
|
|
|
|
|
|
|
mm/cma: always check which page caused allocation failure
Now, we have tracepoint in test_pages_isolated() to notify pfn which
cannot be isolated. But, in alloc_contig_range(), some error path
doesn't call test_pages_isolated() so it's still hard to know exact pfn
that causes allocation failure.
This patch change this situation by calling test_pages_isolated() in
almost error path. In allocation failure case, some overhead is added
by this change, but, allocation failure is really rare event so it would
not matter.
In fatal signal pending case, we don't call test_pages_isolated()
because this failure is intentional one.
There was a bogus outer_start problem due to unchecked buddy order and
this patch also fix it. Before this patch, it didn't matter, because
end result is same thing. But, after this patch, tracepoint will report
failed pfn so it should be accurate.
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Nazarewicz <mina86@mina86.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-14 15:18:45 -08:00
|
|
|
|
|
|
|
|
|
2017-11-29 16:10:01 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/cma: always check which page caused allocation failure
Now, we have tracepoint in test_pages_isolated() to notify pfn which
cannot be isolated. But, in alloc_contig_range(), some error path
doesn't call test_pages_isolated() so it's still hard to know exact pfn
that causes allocation failure.
This patch change this situation by calling test_pages_isolated() in
almost error path. In allocation failure case, some overhead is added
by this change, but, allocation failure is really rare event so it would
not matter.
In fatal signal pending case, we don't call test_pages_isolated()
because this failure is intentional one.
There was a bogus outer_start problem due to unchecked buddy order and
this patch also fix it. Before this patch, it didn't matter, because
end result is same thing. But, after this patch, tracepoint will report
failed pfn so it should be accurate.
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Nazarewicz <mina86@mina86.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-14 15:18:45 -08:00
|
|
|
|
2025-07-22 15:46:49 -04:00
|
|
|
|
mm/cma: always check which page caused allocation failure
Now, we have tracepoint in test_pages_isolated() to notify pfn which
cannot be isolated. But, in alloc_contig_range(), some error path
doesn't call test_pages_isolated() so it's still hard to know exact pfn
that causes allocation failure.
This patch change this situation by calling test_pages_isolated() in
almost error path. In allocation failure case, some overhead is added
by this change, but, allocation failure is really rare event so it would
not matter.
In fatal signal pending case, we don't call test_pages_isolated()
because this failure is intentional one.
There was a bogus outer_start problem due to unchecked buddy order and
this patch also fix it. Before this patch, it didn't matter, because
end result is same thing. But, after this patch, tracepoint will report
failed pfn so it should be accurate.
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Nazarewicz <mina86@mina86.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-14 15:18:45 -08:00
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
mm: replace free hugepage folios after migration
My machine has 4 NUMA nodes, each equipped with 32GB of memory. I have
configured each NUMA node with 16GB of CMA and 16GB of in-use hugetlb
pages. The allocation of contiguous memory via cma_alloc() can fail
probabilistically.
When there are free hugetlb folios in the hugetlb pool, during the
migration of in-use hugetlb folios, new folios are allocated from the free
hugetlb pool. After the migration is completed, the old folios are
released back to the free hugetlb pool instead of being returned to the
buddy system. This can cause test_pages_isolated() check to fail,
ultimately leading to the failure of cma_alloc().
Call trace:
cma_alloc()
__alloc_contig_migrate_range() // migrate in-use hugepage
test_pages_isolated()
__test_page_isolated_in_pageblock()
PageBuddy(page) // check if the page is in buddy
To address this issue, we introduce a function named
replace_free_hugepage_folios(). This function will replace the hugepage
in the free hugepage pool with a new one and release the old one to the
buddy system. After the migration of in-use hugetlb pages is completed,
we will invoke replace_free_hugepage_folios() to ensure that these
hugepages are properly released to the buddy system. Following this step,
when test_pages_isolated() is executed for inspection, it will
successfully pass.
Additionally, when alloc_contig_range() is used to migrate multiple in-use
hugetlb pages, it can result in some in-use hugetlb pages being released
back to the free hugetlb pool and subsequently being reallocated and used
again. For example:
[huge 0] [huge 1]
To migrate huge 0, we obtain huge x from the pool. After the migration is
completed, we return the now-freed huge 0 back to the pool. When it's
time to migrate huge 1, we can simply reuse the now-freed huge 0 from the
pool. As a result, when replace_free_hugepage_folios() is executed, it
cannot release huge 0 back to the buddy system. To address this issue, we
should prevent the reuse of isolated free hugepages during the migration
process.
Link: https://lkml.kernel.org/r/1734503588-16254-1-git-send-email-yangge1116@126.com
Link: https://lkml.kernel.org/r/1736582300-11364-1-git-send-email-yangge1116@126.com
Signed-off-by: yangge <yangge1116@126.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Barry Song <21cnbao@gmail.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: SeongJae Park <sj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-01-11 15:58:20 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
|
|
|
|
|
2022-05-12 20:22:58 -07:00
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-20 14:02:14 -04:00
|
|
|
|
mm/cma: always check which page caused allocation failure
Now, we have tracepoint in test_pages_isolated() to notify pfn which
cannot be isolated. But, in alloc_contig_range(), some error path
doesn't call test_pages_isolated() so it's still hard to know exact pfn
that causes allocation failure.
This patch change this situation by calling test_pages_isolated() in
almost error path. In allocation failure case, some overhead is added
by this change, but, allocation failure is really rare event so it would
not matter.
In fatal signal pending case, we don't call test_pages_isolated()
because this failure is intentional one.
There was a bogus outer_start problem due to unchecked buddy order and
this patch also fix it. Before this patch, it didn't matter, because
end result is same thing. But, after this patch, tracepoint will report
failed pfn so it should be accurate.
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Nazarewicz <mina86@mina86.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-14 15:18:45 -08:00
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
2025-06-16 22:11:14 -04:00
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-01-25 12:49:24 +01:00
|
|
|
|
2012-10-08 16:32:41 -07:00
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-08-13 21:54:49 -06:00
|
|
|
|
2024-12-03 10:47:31 +01:00
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
2024-08-13 21:54:49 -06:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-11-25 21:01:37 +00:00
|
|
|
|
2024-08-13 21:54:49 -06:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
2025-06-16 22:11:13 -04:00
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
|
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
2019-11-30 17:55:06 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-06-16 22:11:14 -04:00
|
|
|
|
|
|
|
|
|
2019-11-30 17:55:06 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-04-14 15:14:29 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-11-30 17:55:06 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-12-03 10:47:30 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-11-30 17:55:06 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-01-14 14:07:33 -08:00
|
|
|
|
|
|
|
|
|
2019-11-30 17:55:06 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
|
|
|
|
|
2019-11-30 17:55:06 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-05-13 17:19:04 -07:00
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
2021-05-04 18:37:34 -07:00
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
2021-05-04 18:37:34 -07:00
|
|
|
|
2024-08-13 21:54:49 -06:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-12-20 15:05:18 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-05-04 18:37:34 -07:00
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
virtio-mem: Paravirtualized memory hotunplug part 2
We also want to unplug online memory (contained in online memory blocks
and, therefore, managed by the buddy), and eventually replug it later.
When requested to unplug memory, we use alloc_contig_range() to allocate
subblocks in online memory blocks (so we are the owner) and send them to
our hypervisor. When requested to plug memory, we can replug such memory
using free_contig_range() after asking our hypervisor.
We also want to mark all allocated pages PG_offline, so nobody will
touch them. To differentiate pages that were never onlined when
onlining the memory block from pages allocated via alloc_contig_range(), we
use PageDirty(). Based on this flag, virtio_mem_fake_online() can either
online the pages for the first time or use free_contig_range().
It is worth noting that there are no guarantees on how much memory can
actually get unplugged again. All device memory might completely be
fragmented with unmovable data, such that no subblock can get unplugged.
We are not touching the ZONE_MOVABLE. If memory is onlined to the
ZONE_MOVABLE, it can only get unplugged after that memory was offlined
manually by user space. In normal operation, virtio-mem memory is
suggested to be onlined to ZONE_NORMAL. In the future, we will try to
make unplug more likely to succeed.
Add a module parameter to control if online memory shall be touched.
As we want to access alloc_contig_range()/free_contig_range() from
kernel module context, export the symbols.
Note: Whenever virtio-mem uses alloc_contig_range(), all affected pages
are on the same node, in the same zone, and contain no holes.
Acked-by: Michal Hocko <mhocko@suse.com> # to export contig range allocator API
Tested-by: Pankaj Gupta <pankaj.gupta.linux@gmail.com>
Cc: "Michael S. Tsirkin" <mst@redhat.com>
Cc: Jason Wang <jasowang@redhat.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Igor Mammedov <imammedo@redhat.com>
Cc: Dave Young <dyoung@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Stefan Hajnoczi <stefanha@redhat.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Alexander Duyck <alexander.h.duyck@linux.intel.com>
Cc: Alexander Potapenko <glider@google.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Link: https://lore.kernel.org/r/20200507140139.17083-6-david@redhat.com
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
2020-05-07 16:01:29 +02:00
|
|
|
|
2011-12-29 13:09:50 +01:00
|
|
|
|
mm, page_alloc: disable pcplists during memory offline
Memory offlining relies on page isolation to guarantee a forward progress
because pages cannot be reused while they are isolated. But the page
isolation itself doesn't prevent from races while freed pages are stored
on pcp lists and thus can be reused. This can be worked around by
repeated draining of pcplists, as done by commit 968318261221
("mm/memory_hotplug: drain per-cpu pages again during memory offline").
David and Michal would prefer that this race was closed in a way that
callers of page isolation who need stronger guarantees don't need to
repeatedly drain. David suggested disabling pcplists usage completely
during page isolation, instead of repeatedly draining them.
To achieve this without adding special cases in alloc/free fastpath, we
can use the same approach as boot pagesets - when pcp->high is 0, any
pcplist addition will be immediately flushed.
The race can thus be closed by setting pcp->high to 0 and draining
pcplists once, before calling start_isolate_page_range(). The draining
will serialize after processes that already disabled interrupts and read
the old value of pcp->high in free_unref_page_commit(), and processes that
have not yet disabled interrupts, will observe pcp->high == 0 when they
are rescheduled, and skip pcplists. This guarantees no stray pages on
pcplists in zones where isolation happens.
This patch thus adds zone_pcp_disable() and zone_pcp_enable() functions
that page isolation users can call before start_isolate_page_range() and
after unisolating (or offlining) the isolated pages.
Also, drain_all_pages() is optimized to only execute on cpus where
pcplists are not empty. The check can however race with a free to pcplist
that has not yet increased the pcp->count from 0 to 1. Thus make the
drain optionally skip the racy check and drain on all cpus, and use this
option in zone_pcp_disable().
As we have to avoid external updates to high and batch while pcplists are
disabled, we take pcp_batch_high_lock in zone_pcp_disable() and release it
in zone_pcp_enable(). This also synchronizes multiple users of
zone_pcp_disable()/enable().
Currently the only user of this functionality is offline_pages().
[vbabka@suse.cz: add comment, per David]
Link: https://lkml.kernel.org/r/527480ef-ed72-e1c1-52a0-1c5b0113df45@suse.cz
Link: https://lkml.kernel.org/r/20201111092812.11329-8-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Suggested-by: David Hildenbrand <david@redhat.com>
Suggested-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: David Hildenbrand <david@redhat.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-14 19:10:59 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-16 13:29:59 +08:00
|
|
|
|
mm, page_alloc: disable pcplists during memory offline
Memory offlining relies on page isolation to guarantee a forward progress
because pages cannot be reused while they are isolated. But the page
isolation itself doesn't prevent from races while freed pages are stored
on pcp lists and thus can be reused. This can be worked around by
repeated draining of pcplists, as done by commit 968318261221
("mm/memory_hotplug: drain per-cpu pages again during memory offline").
David and Michal would prefer that this race was closed in a way that
callers of page isolation who need stronger guarantees don't need to
repeatedly drain. David suggested disabling pcplists usage completely
during page isolation, instead of repeatedly draining them.
To achieve this without adding special cases in alloc/free fastpath, we
can use the same approach as boot pagesets - when pcp->high is 0, any
pcplist addition will be immediately flushed.
The race can thus be closed by setting pcp->high to 0 and draining
pcplists once, before calling start_isolate_page_range(). The draining
will serialize after processes that already disabled interrupts and read
the old value of pcp->high in free_unref_page_commit(), and processes that
have not yet disabled interrupts, will observe pcp->high == 0 when they
are rescheduled, and skip pcplists. This guarantees no stray pages on
pcplists in zones where isolation happens.
This patch thus adds zone_pcp_disable() and zone_pcp_enable() functions
that page isolation users can call before start_isolate_page_range() and
after unisolating (or offlining) the isolated pages.
Also, drain_all_pages() is optimized to only execute on cpus where
pcplists are not empty. The check can however race with a free to pcplist
that has not yet increased the pcp->count from 0 to 1. Thus make the
drain optionally skip the racy check and drain on all cpus, and use this
option in zone_pcp_disable().
As we have to avoid external updates to high and batch while pcplists are
disabled, we take pcp_batch_high_lock in zone_pcp_disable() and release it
in zone_pcp_enable(). This also synchronizes multiple users of
zone_pcp_disable()/enable().
Currently the only user of this functionality is offline_pages().
[vbabka@suse.cz: add comment, per David]
Link: https://lkml.kernel.org/r/527480ef-ed72-e1c1-52a0-1c5b0113df45@suse.cz
Link: https://lkml.kernel.org/r/20201111092812.11329-8-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Suggested-by: David Hildenbrand <david@redhat.com>
Suggested-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: David Hildenbrand <david@redhat.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-14 19:10:59 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-16 13:29:59 +08:00
|
|
|
|
|
|
|
|
|
mm, page_alloc: disable pcplists during memory offline
Memory offlining relies on page isolation to guarantee a forward progress
because pages cannot be reused while they are isolated. But the page
isolation itself doesn't prevent from races while freed pages are stored
on pcp lists and thus can be reused. This can be worked around by
repeated draining of pcplists, as done by commit 968318261221
("mm/memory_hotplug: drain per-cpu pages again during memory offline").
David and Michal would prefer that this race was closed in a way that
callers of page isolation who need stronger guarantees don't need to
repeatedly drain. David suggested disabling pcplists usage completely
during page isolation, instead of repeatedly draining them.
To achieve this without adding special cases in alloc/free fastpath, we
can use the same approach as boot pagesets - when pcp->high is 0, any
pcplist addition will be immediately flushed.
The race can thus be closed by setting pcp->high to 0 and draining
pcplists once, before calling start_isolate_page_range(). The draining
will serialize after processes that already disabled interrupts and read
the old value of pcp->high in free_unref_page_commit(), and processes that
have not yet disabled interrupts, will observe pcp->high == 0 when they
are rescheduled, and skip pcplists. This guarantees no stray pages on
pcplists in zones where isolation happens.
This patch thus adds zone_pcp_disable() and zone_pcp_enable() functions
that page isolation users can call before start_isolate_page_range() and
after unisolating (or offlining) the isolated pages.
Also, drain_all_pages() is optimized to only execute on cpus where
pcplists are not empty. The check can however race with a free to pcplist
that has not yet increased the pcp->count from 0 to 1. Thus make the
drain optionally skip the racy check and drain on all cpus, and use this
option in zone_pcp_disable().
As we have to avoid external updates to high and batch while pcplists are
disabled, we take pcp_batch_high_lock in zone_pcp_disable() and release it
in zone_pcp_enable(). This also synchronizes multiple users of
zone_pcp_disable()/enable().
Currently the only user of this functionality is offline_pages().
[vbabka@suse.cz: add comment, per David]
Link: https://lkml.kernel.org/r/527480ef-ed72-e1c1-52a0-1c5b0113df45@suse.cz
Link: https://lkml.kernel.org/r/20201111092812.11329-8-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Suggested-by: David Hildenbrand <david@redhat.com>
Suggested-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: David Hildenbrand <david@redhat.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-14 19:10:59 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-07-31 16:43:32 -07:00
|
|
|
|
|
|
|
|
|
2012-10-08 16:33:39 -07:00
|
|
|
|
mm/page_alloc: split per cpu page lists and zone stats
The PCP (per-cpu page allocator in page_alloc.c) shares locking
requirements with vmstat and the zone lock which is inconvenient and
causes some issues. For example, the PCP list and vmstat share the same
per-cpu space meaning that it's possible that vmstat updates dirty cache
lines holding per-cpu lists across CPUs unless padding is used. Second,
PREEMPT_RT does not want to disable IRQs for too long in the page
allocator.
This series splits the locking requirements and uses locks types more
suitable for PREEMPT_RT, reduces the time when special locking is required
for stats and reduces the time when IRQs need to be disabled on
!PREEMPT_RT kernels.
Why local_lock? PREEMPT_RT considers the following sequence to be unsafe
as documented in Documentation/locking/locktypes.rst
local_irq_disable();
spin_lock(&lock);
The pcp allocator has this sequence for rmqueue_pcplist (local_irq_save)
-> __rmqueue_pcplist -> rmqueue_bulk (spin_lock). While it's possible to
separate this out, it generally means there are points where we enable
IRQs and reenable them again immediately. To prevent a migration and the
per-cpu pointer going stale, migrate_disable is also needed. That is a
custom lock that is similar, but worse, than local_lock. Furthermore, on
PREEMPT_RT, it's undesirable to leave IRQs disabled for too long. By
converting to local_lock which disables migration on PREEMPT_RT, the
locking requirements can be separated and start moving the protections for
PCP, stats and the zone lock to PREEMPT_RT-safe equivalent locking. As a
bonus, local_lock also means that PROVE_LOCKING does something useful.
After that, it's obvious that zone_statistics incurs too much overhead and
leaves IRQs disabled for longer than necessary on !PREEMPT_RT kernels.
zone_statistics uses perfectly accurate counters requiring IRQs be
disabled for parallel RMW sequences when inaccurate ones like vm_events
would do. The series makes the NUMA statistics (NUMA_HIT and friends)
inaccurate counters that then require no special protection on
!PREEMPT_RT.
The bulk page allocator can then do stat updates in bulk with IRQs enabled
which should improve the efficiency. Technically, this could have been
done without the local_lock and vmstat conversion work and the order
simply reflects the timing of when different series were implemented.
Finally, there are places where we conflate IRQs being disabled for the
PCP with the IRQ-safe zone spinlock. The remainder of the series reduces
the scope of what is protected by disabled IRQs on !PREEMPT_RT kernels.
By the end of the series, page_alloc.c does not call local_irq_save so the
locking scope is a bit clearer. The one exception is that modifying
NR_FREE_PAGES still happens in places where it's known the IRQs are
disabled as it's harmless for PREEMPT_RT and would be expensive to split
the locking there.
No performance data is included because despite the overhead of the stats,
it's within the noise for most workloads on !PREEMPT_RT. However, Jesper
Dangaard Brouer ran a page allocation microbenchmark on a E5-1650 v4 @
3.60GHz CPU on the first version of this series. Focusing on the array
variant of the bulk page allocator reveals the following.
(CPU: Intel(R) Xeon(R) CPU E5-1650 v4 @ 3.60GHz)
ARRAY variant: time_bulk_page_alloc_free_array: step=bulk size
Baseline Patched
1 56.383 54.225 (+3.83%)
2 40.047 35.492 (+11.38%)
3 37.339 32.643 (+12.58%)
4 35.578 30.992 (+12.89%)
8 33.592 29.606 (+11.87%)
16 32.362 28.532 (+11.85%)
32 31.476 27.728 (+11.91%)
64 30.633 27.252 (+11.04%)
128 30.596 27.090 (+11.46%)
While this is a positive outcome, the series is more likely to be
interesting to the RT people in terms of getting parts of the PREEMPT_RT
tree into mainline.
This patch (of 9):
The per-cpu page allocator lists and the per-cpu vmstat deltas are stored
in the same struct per_cpu_pages even though vmstats have no direct impact
on the per-cpu page lists. This is inconsistent because the vmstats for a
node are stored on a dedicated structure. The bigger issue is that the
per_cpu_pages structure is not cache-aligned and stat updates either cache
conflict with adjacent per-cpu lists incurring a runtime cost or padding
is required incurring a memory cost.
This patch splits the per-cpu pagelists and the vmstat deltas into
separate structures. It's mostly a mechanical conversion but some
variable renaming is done to clearly distinguish the per-cpu pages
structure (pcp) from the vmstats (pzstats).
Superficially, this appears to increase the size of the per_cpu_pages
structure but the movement of expire fills a structure hole so there is no
impact overall.
[mgorman@techsingularity.net: make it W=1 cleaner]
Link: https://lkml.kernel.org/r/20210514144622.GA3735@techsingularity.net
[mgorman@techsingularity.net: make it W=1 even cleaner]
Link: https://lkml.kernel.org/r/20210516140705.GB3735@techsingularity.net
[lkp@intel.com: check struct per_cpu_zonestat has a non-zero size]
[vbabka@suse.cz: Init zone->per_cpu_zonestats properly]
Link: https://lkml.kernel.org/r/20210512095458.30632-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20210512095458.30632-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-28 19:41:38 -07:00
|
|
|
|
2012-07-31 16:43:32 -07:00
|
|
|
|
mm/page_alloc: split per cpu page lists and zone stats
The PCP (per-cpu page allocator in page_alloc.c) shares locking
requirements with vmstat and the zone lock which is inconvenient and
causes some issues. For example, the PCP list and vmstat share the same
per-cpu space meaning that it's possible that vmstat updates dirty cache
lines holding per-cpu lists across CPUs unless padding is used. Second,
PREEMPT_RT does not want to disable IRQs for too long in the page
allocator.
This series splits the locking requirements and uses locks types more
suitable for PREEMPT_RT, reduces the time when special locking is required
for stats and reduces the time when IRQs need to be disabled on
!PREEMPT_RT kernels.
Why local_lock? PREEMPT_RT considers the following sequence to be unsafe
as documented in Documentation/locking/locktypes.rst
local_irq_disable();
spin_lock(&lock);
The pcp allocator has this sequence for rmqueue_pcplist (local_irq_save)
-> __rmqueue_pcplist -> rmqueue_bulk (spin_lock). While it's possible to
separate this out, it generally means there are points where we enable
IRQs and reenable them again immediately. To prevent a migration and the
per-cpu pointer going stale, migrate_disable is also needed. That is a
custom lock that is similar, but worse, than local_lock. Furthermore, on
PREEMPT_RT, it's undesirable to leave IRQs disabled for too long. By
converting to local_lock which disables migration on PREEMPT_RT, the
locking requirements can be separated and start moving the protections for
PCP, stats and the zone lock to PREEMPT_RT-safe equivalent locking. As a
bonus, local_lock also means that PROVE_LOCKING does something useful.
After that, it's obvious that zone_statistics incurs too much overhead and
leaves IRQs disabled for longer than necessary on !PREEMPT_RT kernels.
zone_statistics uses perfectly accurate counters requiring IRQs be
disabled for parallel RMW sequences when inaccurate ones like vm_events
would do. The series makes the NUMA statistics (NUMA_HIT and friends)
inaccurate counters that then require no special protection on
!PREEMPT_RT.
The bulk page allocator can then do stat updates in bulk with IRQs enabled
which should improve the efficiency. Technically, this could have been
done without the local_lock and vmstat conversion work and the order
simply reflects the timing of when different series were implemented.
Finally, there are places where we conflate IRQs being disabled for the
PCP with the IRQ-safe zone spinlock. The remainder of the series reduces
the scope of what is protected by disabled IRQs on !PREEMPT_RT kernels.
By the end of the series, page_alloc.c does not call local_irq_save so the
locking scope is a bit clearer. The one exception is that modifying
NR_FREE_PAGES still happens in places where it's known the IRQs are
disabled as it's harmless for PREEMPT_RT and would be expensive to split
the locking there.
No performance data is included because despite the overhead of the stats,
it's within the noise for most workloads on !PREEMPT_RT. However, Jesper
Dangaard Brouer ran a page allocation microbenchmark on a E5-1650 v4 @
3.60GHz CPU on the first version of this series. Focusing on the array
variant of the bulk page allocator reveals the following.
(CPU: Intel(R) Xeon(R) CPU E5-1650 v4 @ 3.60GHz)
ARRAY variant: time_bulk_page_alloc_free_array: step=bulk size
Baseline Patched
1 56.383 54.225 (+3.83%)
2 40.047 35.492 (+11.38%)
3 37.339 32.643 (+12.58%)
4 35.578 30.992 (+12.89%)
8 33.592 29.606 (+11.87%)
16 32.362 28.532 (+11.85%)
32 31.476 27.728 (+11.91%)
64 30.633 27.252 (+11.04%)
128 30.596 27.090 (+11.46%)
While this is a positive outcome, the series is more likely to be
interesting to the RT people in terms of getting parts of the PREEMPT_RT
tree into mainline.
This patch (of 9):
The per-cpu page allocator lists and the per-cpu vmstat deltas are stored
in the same struct per_cpu_pages even though vmstats have no direct impact
on the per-cpu page lists. This is inconsistent because the vmstats for a
node are stored on a dedicated structure. The bigger issue is that the
per_cpu_pages structure is not cache-aligned and stat updates either cache
conflict with adjacent per-cpu lists incurring a runtime cost or padding
is required incurring a memory cost.
This patch splits the per-cpu pagelists and the vmstat deltas into
separate structures. It's mostly a mechanical conversion but some
variable renaming is done to clearly distinguish the per-cpu pages
structure (pcp) from the vmstats (pzstats).
Superficially, this appears to increase the size of the per_cpu_pages
structure but the movement of expire fills a structure hole so there is no
impact overall.
[mgorman@techsingularity.net: make it W=1 cleaner]
Link: https://lkml.kernel.org/r/20210514144622.GA3735@techsingularity.net
[mgorman@techsingularity.net: make it W=1 even cleaner]
Link: https://lkml.kernel.org/r/20210516140705.GB3735@techsingularity.net
[lkp@intel.com: check struct per_cpu_zonestat has a non-zero size]
[vbabka@suse.cz: Init zone->per_cpu_zonestats properly]
Link: https://lkml.kernel.org/r/20210512095458.30632-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20210512095458.30632-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-28 19:41:38 -07:00
|
|
|
|
2012-10-08 16:33:39 -07:00
|
|
|
|
mm/page_alloc: split per cpu page lists and zone stats
The PCP (per-cpu page allocator in page_alloc.c) shares locking
requirements with vmstat and the zone lock which is inconvenient and
causes some issues. For example, the PCP list and vmstat share the same
per-cpu space meaning that it's possible that vmstat updates dirty cache
lines holding per-cpu lists across CPUs unless padding is used. Second,
PREEMPT_RT does not want to disable IRQs for too long in the page
allocator.
This series splits the locking requirements and uses locks types more
suitable for PREEMPT_RT, reduces the time when special locking is required
for stats and reduces the time when IRQs need to be disabled on
!PREEMPT_RT kernels.
Why local_lock? PREEMPT_RT considers the following sequence to be unsafe
as documented in Documentation/locking/locktypes.rst
local_irq_disable();
spin_lock(&lock);
The pcp allocator has this sequence for rmqueue_pcplist (local_irq_save)
-> __rmqueue_pcplist -> rmqueue_bulk (spin_lock). While it's possible to
separate this out, it generally means there are points where we enable
IRQs and reenable them again immediately. To prevent a migration and the
per-cpu pointer going stale, migrate_disable is also needed. That is a
custom lock that is similar, but worse, than local_lock. Furthermore, on
PREEMPT_RT, it's undesirable to leave IRQs disabled for too long. By
converting to local_lock which disables migration on PREEMPT_RT, the
locking requirements can be separated and start moving the protections for
PCP, stats and the zone lock to PREEMPT_RT-safe equivalent locking. As a
bonus, local_lock also means that PROVE_LOCKING does something useful.
After that, it's obvious that zone_statistics incurs too much overhead and
leaves IRQs disabled for longer than necessary on !PREEMPT_RT kernels.
zone_statistics uses perfectly accurate counters requiring IRQs be
disabled for parallel RMW sequences when inaccurate ones like vm_events
would do. The series makes the NUMA statistics (NUMA_HIT and friends)
inaccurate counters that then require no special protection on
!PREEMPT_RT.
The bulk page allocator can then do stat updates in bulk with IRQs enabled
which should improve the efficiency. Technically, this could have been
done without the local_lock and vmstat conversion work and the order
simply reflects the timing of when different series were implemented.
Finally, there are places where we conflate IRQs being disabled for the
PCP with the IRQ-safe zone spinlock. The remainder of the series reduces
the scope of what is protected by disabled IRQs on !PREEMPT_RT kernels.
By the end of the series, page_alloc.c does not call local_irq_save so the
locking scope is a bit clearer. The one exception is that modifying
NR_FREE_PAGES still happens in places where it's known the IRQs are
disabled as it's harmless for PREEMPT_RT and would be expensive to split
the locking there.
No performance data is included because despite the overhead of the stats,
it's within the noise for most workloads on !PREEMPT_RT. However, Jesper
Dangaard Brouer ran a page allocation microbenchmark on a E5-1650 v4 @
3.60GHz CPU on the first version of this series. Focusing on the array
variant of the bulk page allocator reveals the following.
(CPU: Intel(R) Xeon(R) CPU E5-1650 v4 @ 3.60GHz)
ARRAY variant: time_bulk_page_alloc_free_array: step=bulk size
Baseline Patched
1 56.383 54.225 (+3.83%)
2 40.047 35.492 (+11.38%)
3 37.339 32.643 (+12.58%)
4 35.578 30.992 (+12.89%)
8 33.592 29.606 (+11.87%)
16 32.362 28.532 (+11.85%)
32 31.476 27.728 (+11.91%)
64 30.633 27.252 (+11.04%)
128 30.596 27.090 (+11.46%)
While this is a positive outcome, the series is more likely to be
interesting to the RT people in terms of getting parts of the PREEMPT_RT
tree into mainline.
This patch (of 9):
The per-cpu page allocator lists and the per-cpu vmstat deltas are stored
in the same struct per_cpu_pages even though vmstats have no direct impact
on the per-cpu page lists. This is inconsistent because the vmstats for a
node are stored on a dedicated structure. The bigger issue is that the
per_cpu_pages structure is not cache-aligned and stat updates either cache
conflict with adjacent per-cpu lists incurring a runtime cost or padding
is required incurring a memory cost.
This patch splits the per-cpu pagelists and the vmstat deltas into
separate structures. It's mostly a mechanical conversion but some
variable renaming is done to clearly distinguish the per-cpu pages
structure (pcp) from the vmstats (pzstats).
Superficially, this appears to increase the size of the per_cpu_pages
structure but the movement of expire fills a structure hole so there is no
impact overall.
[mgorman@techsingularity.net: make it W=1 cleaner]
Link: https://lkml.kernel.org/r/20210514144622.GA3735@techsingularity.net
[mgorman@techsingularity.net: make it W=1 even cleaner]
Link: https://lkml.kernel.org/r/20210516140705.GB3735@techsingularity.net
[lkp@intel.com: check struct per_cpu_zonestat has a non-zero size]
[vbabka@suse.cz: Init zone->per_cpu_zonestats properly]
Link: https://lkml.kernel.org/r/20210512095458.30632-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20210512095458.30632-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-28 19:41:38 -07:00
|
|
|
|
|
|
|
|
|
2012-10-08 16:33:39 -07:00
|
|
|
|
mm/page_alloc: split per cpu page lists and zone stats
The PCP (per-cpu page allocator in page_alloc.c) shares locking
requirements with vmstat and the zone lock which is inconvenient and
causes some issues. For example, the PCP list and vmstat share the same
per-cpu space meaning that it's possible that vmstat updates dirty cache
lines holding per-cpu lists across CPUs unless padding is used. Second,
PREEMPT_RT does not want to disable IRQs for too long in the page
allocator.
This series splits the locking requirements and uses locks types more
suitable for PREEMPT_RT, reduces the time when special locking is required
for stats and reduces the time when IRQs need to be disabled on
!PREEMPT_RT kernels.
Why local_lock? PREEMPT_RT considers the following sequence to be unsafe
as documented in Documentation/locking/locktypes.rst
local_irq_disable();
spin_lock(&lock);
The pcp allocator has this sequence for rmqueue_pcplist (local_irq_save)
-> __rmqueue_pcplist -> rmqueue_bulk (spin_lock). While it's possible to
separate this out, it generally means there are points where we enable
IRQs and reenable them again immediately. To prevent a migration and the
per-cpu pointer going stale, migrate_disable is also needed. That is a
custom lock that is similar, but worse, than local_lock. Furthermore, on
PREEMPT_RT, it's undesirable to leave IRQs disabled for too long. By
converting to local_lock which disables migration on PREEMPT_RT, the
locking requirements can be separated and start moving the protections for
PCP, stats and the zone lock to PREEMPT_RT-safe equivalent locking. As a
bonus, local_lock also means that PROVE_LOCKING does something useful.
After that, it's obvious that zone_statistics incurs too much overhead and
leaves IRQs disabled for longer than necessary on !PREEMPT_RT kernels.
zone_statistics uses perfectly accurate counters requiring IRQs be
disabled for parallel RMW sequences when inaccurate ones like vm_events
would do. The series makes the NUMA statistics (NUMA_HIT and friends)
inaccurate counters that then require no special protection on
!PREEMPT_RT.
The bulk page allocator can then do stat updates in bulk with IRQs enabled
which should improve the efficiency. Technically, this could have been
done without the local_lock and vmstat conversion work and the order
simply reflects the timing of when different series were implemented.
Finally, there are places where we conflate IRQs being disabled for the
PCP with the IRQ-safe zone spinlock. The remainder of the series reduces
the scope of what is protected by disabled IRQs on !PREEMPT_RT kernels.
By the end of the series, page_alloc.c does not call local_irq_save so the
locking scope is a bit clearer. The one exception is that modifying
NR_FREE_PAGES still happens in places where it's known the IRQs are
disabled as it's harmless for PREEMPT_RT and would be expensive to split
the locking there.
No performance data is included because despite the overhead of the stats,
it's within the noise for most workloads on !PREEMPT_RT. However, Jesper
Dangaard Brouer ran a page allocation microbenchmark on a E5-1650 v4 @
3.60GHz CPU on the first version of this series. Focusing on the array
variant of the bulk page allocator reveals the following.
(CPU: Intel(R) Xeon(R) CPU E5-1650 v4 @ 3.60GHz)
ARRAY variant: time_bulk_page_alloc_free_array: step=bulk size
Baseline Patched
1 56.383 54.225 (+3.83%)
2 40.047 35.492 (+11.38%)
3 37.339 32.643 (+12.58%)
4 35.578 30.992 (+12.89%)
8 33.592 29.606 (+11.87%)
16 32.362 28.532 (+11.85%)
32 31.476 27.728 (+11.91%)
64 30.633 27.252 (+11.04%)
128 30.596 27.090 (+11.46%)
While this is a positive outcome, the series is more likely to be
interesting to the RT people in terms of getting parts of the PREEMPT_RT
tree into mainline.
This patch (of 9):
The per-cpu page allocator lists and the per-cpu vmstat deltas are stored
in the same struct per_cpu_pages even though vmstats have no direct impact
on the per-cpu page lists. This is inconsistent because the vmstats for a
node are stored on a dedicated structure. The bigger issue is that the
per_cpu_pages structure is not cache-aligned and stat updates either cache
conflict with adjacent per-cpu lists incurring a runtime cost or padding
is required incurring a memory cost.
This patch splits the per-cpu pagelists and the vmstat deltas into
separate structures. It's mostly a mechanical conversion but some
variable renaming is done to clearly distinguish the per-cpu pages
structure (pcp) from the vmstats (pzstats).
Superficially, this appears to increase the size of the per_cpu_pages
structure but the movement of expire fills a structure hole so there is no
impact overall.
[mgorman@techsingularity.net: make it W=1 cleaner]
Link: https://lkml.kernel.org/r/20210514144622.GA3735@techsingularity.net
[mgorman@techsingularity.net: make it W=1 even cleaner]
Link: https://lkml.kernel.org/r/20210516140705.GB3735@techsingularity.net
[lkp@intel.com: check struct per_cpu_zonestat has a non-zero size]
[vbabka@suse.cz: Init zone->per_cpu_zonestats properly]
Link: https://lkml.kernel.org/r/20210512095458.30632-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20210512095458.30632-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-28 19:41:38 -07:00
|
|
|
|
|
|
|
|
|
2022-09-16 15:22:47 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-07-31 16:43:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-12-11 16:01:01 -08:00
|
|
|
|
2007-10-16 01:26:12 -07:00
|
|
|
|
2020-10-15 20:07:59 -07:00
|
|
|
|
|
|
|
|
|
2024-06-07 11:09:38 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-10-16 01:26:12 -07:00
|
|
|
|
2024-06-07 11:09:38 +02:00
|
|
|
|
|
|
|
|
|
2007-10-16 01:26:12 -07:00
|
|
|
|
2024-06-07 11:09:38 +02:00
|
|
|
|
2020-10-15 20:07:59 -07:00
|
|
|
|
2007-10-16 01:26:12 -07:00
|
|
|
|
|
|
|
|
|
2019-11-30 17:54:03 -08:00
|
|
|
|
2019-05-13 17:21:24 -07:00
|
|
|
|
2017-07-06 15:37:56 -07:00
|
|
|
|
2007-10-16 01:26:12 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-12-11 16:00:45 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: Allow to offline unmovable PageOffline() pages via MEM_GOING_OFFLINE
virtio-mem wants to allow to offline memory blocks of which some parts
were unplugged (allocated via alloc_contig_range()), especially, to later
offline and remove completely unplugged memory blocks. The important part
is that PageOffline() has to remain set until the section is offline, so
these pages will never get accessed (e.g., when dumping). The pages should
not be handed back to the buddy (which would require clearing PageOffline()
and result in issues if offlining fails and the pages are suddenly in the
buddy).
Let's allow to do that by allowing to isolate any PageOffline() page
when offlining. This way, we can reach the memory hotplug notifier
MEM_GOING_OFFLINE, where the driver can signal that he is fine with
offlining this page by dropping its reference count. PageOffline() pages
with a reference count of 0 can then be skipped when offlining the
pages (like if they were free, however they are not in the buddy).
Anybody who uses PageOffline() pages and does not agree to offline them
(e.g., Hyper-V balloon, XEN balloon, VMWare balloon for 2MB pages) will not
decrement the reference count and make offlining fail when trying to
migrate such an unmovable page. So there should be no observable change.
Same applies to balloon compaction users (movable PageOffline() pages), the
pages will simply be migrated.
Note 1: If offlining fails, a driver has to increment the reference
count again in MEM_CANCEL_OFFLINE.
Note 2: A driver that makes use of this has to be aware that re-onlining
the memory block has to be handled by hooking into onlining code
(online_page_callback_t), resetting the page PageOffline() and
not giving them to the buddy.
Reviewed-by: Alexander Duyck <alexander.h.duyck@linux.intel.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Pankaj Gupta <pankaj.gupta.linux@gmail.com>
Acked-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Juergen Gross <jgross@suse.com>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Cc: Pavel Tatashin <pavel.tatashin@microsoft.com>
Cc: Alexander Duyck <alexander.h.duyck@linux.intel.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Anthony Yznaga <anthony.yznaga@oracle.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Anshuman Khandual <anshuman.khandual@arm.com>
Cc: Qian Cai <cai@lca.pw>
Cc: Pingfan Liu <kernelfans@gmail.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Link: https://lore.kernel.org/r/20200507140139.17083-7-david@redhat.com
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
2020-05-07 16:01:30 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-06-07 11:09:38 +02:00
|
|
|
|
mm: Allow to offline unmovable PageOffline() pages via MEM_GOING_OFFLINE
virtio-mem wants to allow to offline memory blocks of which some parts
were unplugged (allocated via alloc_contig_range()), especially, to later
offline and remove completely unplugged memory blocks. The important part
is that PageOffline() has to remain set until the section is offline, so
these pages will never get accessed (e.g., when dumping). The pages should
not be handed back to the buddy (which would require clearing PageOffline()
and result in issues if offlining fails and the pages are suddenly in the
buddy).
Let's allow to do that by allowing to isolate any PageOffline() page
when offlining. This way, we can reach the memory hotplug notifier
MEM_GOING_OFFLINE, where the driver can signal that he is fine with
offlining this page by dropping its reference count. PageOffline() pages
with a reference count of 0 can then be skipped when offlining the
pages (like if they were free, however they are not in the buddy).
Anybody who uses PageOffline() pages and does not agree to offline them
(e.g., Hyper-V balloon, XEN balloon, VMWare balloon for 2MB pages) will not
decrement the reference count and make offlining fail when trying to
migrate such an unmovable page. So there should be no observable change.
Same applies to balloon compaction users (movable PageOffline() pages), the
pages will simply be migrated.
Note 1: If offlining fails, a driver has to increment the reference
count again in MEM_CANCEL_OFFLINE.
Note 2: A driver that makes use of this has to be aware that re-onlining
the memory block has to be handled by hooking into onlining code
(online_page_callback_t), resetting the page PageOffline() and
not giving them to the buddy.
Reviewed-by: Alexander Duyck <alexander.h.duyck@linux.intel.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Pankaj Gupta <pankaj.gupta.linux@gmail.com>
Acked-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Juergen Gross <jgross@suse.com>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Cc: Pavel Tatashin <pavel.tatashin@microsoft.com>
Cc: Alexander Duyck <alexander.h.duyck@linux.intel.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Anthony Yznaga <anthony.yznaga@oracle.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Anshuman Khandual <anshuman.khandual@arm.com>
Cc: Qian Cai <cai@lca.pw>
Cc: Pingfan Liu <kernelfans@gmail.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Link: https://lore.kernel.org/r/20200507140139.17083-7-david@redhat.com
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
2020-05-07 16:01:30 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-12-11 16:00:45 -08:00
|
|
|
|
2007-10-16 01:26:12 -07:00
|
|
|
|
|
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
2020-10-15 20:10:15 -07:00
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
2007-10-16 01:26:12 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-06-07 11:09:38 +02:00
|
|
|
|
|
|
|
|
|
2007-10-16 01:26:12 -07:00
|
|
|
|
|
|
|
|
|
2009-12-16 12:19:58 +01:00
|
|
|
|
2021-11-05 13:40:31 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-26 17:10:27 +00:00
|
|
|
|
2009-12-16 12:19:58 +01:00
|
|
|
|
|
|
|
|
|
2014-06-04 16:10:21 -07:00
|
|
|
|
2009-12-16 12:19:58 +01:00
|
|
|
|
2023-12-28 17:47:03 +03:00
|
|
|
|
2024-03-26 17:10:27 +00:00
|
|
|
|
2009-12-16 12:19:58 +01:00
|
|
|
|
2024-03-26 17:10:27 +00:00
|
|
|
|
|
|
|
|
|
2009-12-16 12:19:58 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-12-28 17:47:04 +03:00
|
|
|
|
2009-12-16 12:19:58 +01:00
|
|
|
|
2022-03-22 14:44:35 -07:00
|
|
|
|
2018-08-23 17:00:42 -07:00
|
|
|
|
|
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-08-23 17:00:42 -07:00
|
|
|
|
2020-10-15 20:07:05 -07:00
|
|
|
|
|
|
|
|
|
2018-08-23 17:00:42 -07:00
|
|
|
|
2020-10-15 20:07:05 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-09-27 18:35:14 +08:00
|
|
|
|
2020-10-15 20:07:05 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-09-27 18:35:14 +08:00
|
|
|
|
2020-10-15 20:07:05 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
2020-10-15 20:07:05 -07:00
|
|
|
|
|
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
2023-09-27 18:35:13 +08:00
|
|
|
|
2020-10-15 20:07:05 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-08-23 17:00:42 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-10-15 20:07:05 -07:00
|
|
|
|
2018-08-23 17:00:42 -07:00
|
|
|
|
|
|
|
|
|
2023-12-28 17:47:03 +03:00
|
|
|
|
2018-08-23 17:00:42 -07:00
|
|
|
|
2020-10-15 20:10:15 -07:00
|
|
|
|
2018-08-23 17:00:42 -07:00
|
|
|
|
2020-10-15 20:10:15 -07:00
|
|
|
|
2020-10-15 20:07:05 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
|
|
|
|
|
2020-10-15 20:07:05 -07:00
|
|
|
|
2020-10-15 20:10:15 -07:00
|
|
|
|
mm/hwpoison: fix unpoison_memory()
After recent soft-offline rework, error pages can be taken off from
buddy allocator, but the existing unpoison_memory() does not properly
undo the operation. Moreover, due to the recent change on
__get_hwpoison_page(), get_page_unless_zero() is hardly called for
hwpoisoned pages. So __get_hwpoison_page() highly likely returns -EBUSY
(meaning to fail to grab page refcount) and unpoison just clears
PG_hwpoison without releasing a refcount. That does not lead to a
critical issue like kernel panic, but unpoisoned pages never get back to
buddy (leaked permanently), which is not good.
To (partially) fix this, we need to identify "taken off" pages from
other types of hwpoisoned pages. We can't use refcount or page flags
for this purpose, so a pseudo flag is defined by hacking ->private
field. Someone might think that put_page() is enough to cancel
taken-off pages, but the normal free path contains some operations not
suitable for the current purpose, and can fire VM_BUG_ON().
Note that unpoison_memory() is now supposed to be cancel hwpoison events
injected only by madvise() or
/sys/devices/system/memory/{hard,soft}_offline_page, not by MCE
injection, so please don't try to use unpoison when testing with MCE
injection.
[lkp@intel.com: report build failure for ARCH=i386]
Link: https://lkml.kernel.org/r/20211115084006.3728254-4-naoya.horiguchi@linux.dev
Signed-off-by: Naoya Horiguchi <naoya.horiguchi@nec.com>
Reviewed-by: Yang Shi <shy828301@gmail.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Ding Hui <dinghui@sangfor.com.cn>
Cc: Tony Luck <tony.luck@intel.com>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Peter Xu <peterx@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-14 14:09:09 -08:00
|
|
|
|
2020-10-15 20:07:05 -07:00
|
|
|
|
2018-08-23 17:00:42 -07:00
|
|
|
|
|
|
|
|
|
2020-10-15 20:07:05 -07:00
|
|
|
|
|
|
|
|
|
2018-08-23 17:00:42 -07:00
|
|
|
|
|
|
|
|
|
2020-10-15 20:07:05 -07:00
|
|
|
|
2018-08-23 17:00:42 -07:00
|
|
|
|
mm/hwpoison: fix unpoison_memory()
After recent soft-offline rework, error pages can be taken off from
buddy allocator, but the existing unpoison_memory() does not properly
undo the operation. Moreover, due to the recent change on
__get_hwpoison_page(), get_page_unless_zero() is hardly called for
hwpoisoned pages. So __get_hwpoison_page() highly likely returns -EBUSY
(meaning to fail to grab page refcount) and unpoison just clears
PG_hwpoison without releasing a refcount. That does not lead to a
critical issue like kernel panic, but unpoisoned pages never get back to
buddy (leaked permanently), which is not good.
To (partially) fix this, we need to identify "taken off" pages from
other types of hwpoisoned pages. We can't use refcount or page flags
for this purpose, so a pseudo flag is defined by hacking ->private
field. Someone might think that put_page() is enough to cancel
taken-off pages, but the normal free path contains some operations not
suitable for the current purpose, and can fire VM_BUG_ON().
Note that unpoison_memory() is now supposed to be cancel hwpoison events
injected only by madvise() or
/sys/devices/system/memory/{hard,soft}_offline_page, not by MCE
injection, so please don't try to use unpoison when testing with MCE
injection.
[lkp@intel.com: report build failure for ARCH=i386]
Link: https://lkml.kernel.org/r/20211115084006.3728254-4-naoya.horiguchi@linux.dev
Signed-off-by: Naoya Horiguchi <naoya.horiguchi@nec.com>
Reviewed-by: Yang Shi <shy828301@gmail.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Ding Hui <dinghui@sangfor.com.cn>
Cc: Tony Luck <tony.luck@intel.com>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Peter Xu <peterx@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-14 14:09:09 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-20 14:02:12 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/hwpoison: fix unpoison_memory()
After recent soft-offline rework, error pages can be taken off from
buddy allocator, but the existing unpoison_memory() does not properly
undo the operation. Moreover, due to the recent change on
__get_hwpoison_page(), get_page_unless_zero() is hardly called for
hwpoisoned pages. So __get_hwpoison_page() highly likely returns -EBUSY
(meaning to fail to grab page refcount) and unpoison just clears
PG_hwpoison without releasing a refcount. That does not lead to a
critical issue like kernel panic, but unpoisoned pages never get back to
buddy (leaked permanently), which is not good.
To (partially) fix this, we need to identify "taken off" pages from
other types of hwpoisoned pages. We can't use refcount or page flags
for this purpose, so a pseudo flag is defined by hacking ->private
field. Someone might think that put_page() is enough to cancel
taken-off pages, but the normal free path contains some operations not
suitable for the current purpose, and can fire VM_BUG_ON().
Note that unpoison_memory() is now supposed to be cancel hwpoison events
injected only by madvise() or
/sys/devices/system/memory/{hard,soft}_offline_page, not by MCE
injection, so please don't try to use unpoison when testing with MCE
injection.
[lkp@intel.com: report build failure for ARCH=i386]
Link: https://lkml.kernel.org/r/20211115084006.3728254-4-naoya.horiguchi@linux.dev
Signed-off-by: Naoya Horiguchi <naoya.horiguchi@nec.com>
Reviewed-by: Yang Shi <shy828301@gmail.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Ding Hui <dinghui@sangfor.com.cn>
Cc: Tony Luck <tony.luck@intel.com>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Peter Xu <peterx@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-14 14:09:09 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-08-23 17:00:42 -07:00
|
|
|
|
mm_zone: add function to check if managed dma zone exists
Patch series "Handle warning of allocation failure on DMA zone w/o
managed pages", v4.
**Problem observed:
On x86_64, when crash is triggered and entering into kdump kernel, page
allocation failure can always be seen.
---------------------------------
DMA: preallocated 128 KiB GFP_KERNEL pool for atomic allocations
swapper/0: page allocation failure: order:5, mode:0xcc1(GFP_KERNEL|GFP_DMA), nodemask=(null),cpuset=/,mems_allowed=0
CPU: 0 PID: 1 Comm: swapper/0
Call Trace:
dump_stack+0x7f/0xa1
warn_alloc.cold+0x72/0xd6
......
__alloc_pages+0x24d/0x2c0
......
dma_atomic_pool_init+0xdb/0x176
do_one_initcall+0x67/0x320
? rcu_read_lock_sched_held+0x3f/0x80
kernel_init_freeable+0x290/0x2dc
? rest_init+0x24f/0x24f
kernel_init+0xa/0x111
ret_from_fork+0x22/0x30
Mem-Info:
------------------------------------
***Root cause:
In the current kernel, it assumes that DMA zone must have managed pages
and try to request pages if CONFIG_ZONE_DMA is enabled. While this is not
always true. E.g in kdump kernel of x86_64, only low 1M is presented and
locked down at very early stage of boot, so that this low 1M won't be
added into buddy allocator to become managed pages of DMA zone. This
exception will always cause page allocation failure if page is requested
from DMA zone.
***Investigation:
This failure happens since below commit merged into linus's tree.
1a6a9044b967 x86/setup: Remove CONFIG_X86_RESERVE_LOW and reservelow= options
23721c8e92f7 x86/crash: Remove crash_reserve_low_1M()
f1d4d47c5851 x86/setup: Always reserve the first 1M of RAM
7c321eb2b843 x86/kdump: Remove the backup region handling
6f599d84231f x86/kdump: Always reserve the low 1M when the crashkernel option is specified
Before them, on x86_64, the low 640K area will be reused by kdump kernel.
So in kdump kernel, the content of low 640K area is copied into a backup
region for dumping before jumping into kdump. Then except of those firmware
reserved region in [0, 640K], the left area will be added into buddy
allocator to become available managed pages of DMA zone.
However, after above commits applied, in kdump kernel of x86_64, the low
1M is reserved by memblock, but not released to buddy allocator. So any
later page allocation requested from DMA zone will fail.
At the beginning, if crashkernel is reserved, the low 1M need be locked
down because AMD SME encrypts memory making the old backup region
mechanims impossible when switching into kdump kernel.
Later, it was also observed that there are BIOSes corrupting memory
under 1M. To solve this, in commit f1d4d47c5851, the entire region of
low 1M is always reserved after the real mode trampoline is allocated.
Besides, recently, Intel engineer mentioned their TDX (Trusted domain
extensions) which is under development in kernel also needs to lock down
the low 1M. So we can't simply revert above commits to fix the page allocation
failure from DMA zone as someone suggested.
***Solution:
Currently, only DMA atomic pool and dma-kmalloc will initialize and
request page allocation with GFP_DMA during bootup.
So only initializ DMA atomic pool when DMA zone has available managed
pages, otherwise just skip the initialization.
For dma-kmalloc(), for the time being, let's mute the warning of
allocation failure if requesting pages from DMA zone while no manged
pages. Meanwhile, change code to use dma_alloc_xx/dma_map_xx API to
replace kmalloc(GFP_DMA), or do not use GFP_DMA when calling kmalloc() if
not necessary. Christoph is posting patches to fix those under
drivers/scsi/. Finally, we can remove the need of dma-kmalloc() as people
suggested.
This patch (of 3):
In some places of the current kernel, it assumes that dma zone must have
managed pages if CONFIG_ZONE_DMA is enabled. While this is not always
true. E.g in kdump kernel of x86_64, only low 1M is presented and locked
down at very early stage of boot, so that there's no managed pages at all
in DMA zone. This exception will always cause page allocation failure if
page is requested from DMA zone.
Here add function has_managed_dma() and the relevant helper functions to
check if there's DMA zone with managed pages. It will be used in later
patches.
Link: https://lkml.kernel.org/r/20211223094435.248523-1-bhe@redhat.com
Link: https://lkml.kernel.org/r/20211223094435.248523-2-bhe@redhat.com
Fixes: 6f599d84231f ("x86/kdump: Always reserve the low 1M when the crashkernel option is specified")
Signed-off-by: Baoquan He <bhe@redhat.com>
Reviewed-by: David Hildenbrand <david@redhat.com>
Acked-by: John Donnelly <john.p.donnelly@oracle.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Christoph Lameter <cl@linux.com>
Cc: Hyeonggon Yoo <42.hyeyoo@gmail.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Laight <David.Laight@ACULAB.COM>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Marek Szyprowski <m.szyprowski@samsung.com>
Cc: Robin Murphy <robin.murphy@arm.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-14 14:07:37 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-06-06 17:26:29 +03:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-08-09 14:48:51 +03:00
|
|
|
|
2023-06-06 17:26:29 +03:00
|
|
|
|
|
|
|
|
|
2024-08-09 14:48:52 +03:00
|
|
|
|
|
|
|
|
|
2023-06-06 17:26:29 +03:00
|
|
|
|
2024-08-09 14:48:52 +03:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-06-06 17:26:29 +03:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-08-09 14:48:52 +03:00
|
|
|
|
|
|
|
|
|
2023-06-06 17:26:29 +03:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-05-06 14:25:08 +03:00
|
|
|
|
|
|
|
|
|
2023-06-06 17:26:29 +03:00
|
|
|
|
2025-03-10 10:28:55 +02:00
|
|
|
|
2024-08-09 14:48:47 +03:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-06-06 17:26:29 +03:00
|
|
|
|
2025-05-06 14:25:08 +03:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-03-10 10:28:55 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-08-09 14:48:54 +03:00
|
|
|
|
2025-03-10 10:28:55 +02:00
|
|
|
|
2023-06-06 17:26:29 +03:00
|
|
|
|
2024-08-09 14:48:47 +03:00
|
|
|
|
|
|
|
|
|
2023-06-06 17:26:29 +03:00
|
|
|
|
2024-08-09 14:48:47 +03:00
|
|
|
|
2023-06-06 17:26:29 +03:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-08-09 14:48:47 +03:00
|
|
|
|
2023-06-06 17:26:29 +03:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-20 14:02:15 -04:00
|
|
|
|
2023-06-06 17:26:29 +03:00
|
|
|
|
2024-08-09 14:48:50 +03:00
|
|
|
|
2023-06-06 17:26:29 +03:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-05-06 14:25:08 +03:00
|
|
|
|
|
|
|
|
|
2023-06-06 17:26:29 +03:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-02-21 18:44:23 -08:00
|
|
|
|
2025-09-08 18:00:04 -07:00
|
|
|
|
2025-02-21 18:44:23 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-05-16 17:34:46 -07:00
|
|
|
|
2025-02-21 18:44:23 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-05-16 17:34:46 -07:00
|
|
|
|
2025-02-21 18:44:23 -08:00
|
|
|
|
|
|
|
|
|
2025-09-08 18:00:03 -07:00
|
|
|
|
|
|
|
|
|
2025-02-21 18:44:23 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-09-08 18:00:03 -07:00
|
|
|
|
2025-02-21 18:44:23 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-09-08 18:00:03 -07:00
|
|
|
|
2025-02-21 18:44:26 -08:00
|
|
|
|
2025-09-08 18:00:04 -07:00
|
|
|
|
2025-02-21 18:44:26 -08:00
|
|
|
|
|
|
|
|
|
2025-02-21 18:44:23 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-09-08 18:00:04 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
