2019-05-27 08:55:06 +02:00
|
|
|
|
2008-02-07 00:13:50 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-02-07 00:13:51 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-03-10 15:22:24 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-12-18 14:21:56 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-04-14 15:44:51 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-12-15 12:34:29 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-02-07 00:13:50 -08:00
|
|
|
|
|
|
|
|
|
mm, memcg: cg2 memory{.swap,}.peak write handlers
Patch series "mm, memcg: cg2 memory{.swap,}.peak write handlers", v7.
This patch (of 2):
Other mechanisms for querying the peak memory usage of either a process or
v1 memory cgroup allow for resetting the high watermark. Restore parity
with those mechanisms, but with a less racy API.
For example:
- Any write to memory.max_usage_in_bytes in a cgroup v1 mount resets
the high watermark.
- writing "5" to the clear_refs pseudo-file in a processes's proc
directory resets the peak RSS.
This change is an evolution of a previous patch, which mostly copied the
cgroup v1 behavior, however, there were concerns about races/ownership
issues with a global reset, so instead this change makes the reset
filedescriptor-local.
Writing any non-empty string to the memory.peak and memory.swap.peak
pseudo-files reset the high watermark to the current usage for subsequent
reads through that same FD.
Notably, following Johannes's suggestion, this implementation moves the
O(FDs that have written) behavior onto the FD write(2) path. Instead, on
the page-allocation path, we simply add one additional watermark to
conditionally bump per-hierarchy level in the page-counter.
Additionally, this takes Longman's suggestion of nesting the
page-charging-path checks for the two watermarks to reduce the number of
common-case comparisons.
This behavior is particularly useful for work scheduling systems that need
to track memory usage of worker processes/cgroups per-work-item. Since
memory can't be squeezed like CPU can (the OOM-killer has opinions), these
systems need to track the peak memory usage to compute system/container
fullness when binpacking workitems.
Most notably, Vimeo's use-case involves a system that's doing global
binpacking across many Kubernetes pods/containers, and while we can use
PSI for some local decisions about overload, we strive to avoid packing
workloads too tightly in the first place. To facilitate this, we track
the peak memory usage. However, since we run with long-lived workers (to
amortize startup costs) we need a way to track the high watermark while a
work-item is executing. Polling runs the risk of missing short spikes
that last for timescales below the polling interval, and peak memory
tracking at the cgroup level is otherwise perfect for this use-case.
As this data is used to ensure that binpacked work ends up with sufficient
headroom, this use-case mostly avoids the inaccuracies surrounding
reclaimable memory.
Link: https://lkml.kernel.org/r/20240730231304.761942-1-davidf@vimeo.com
Link: https://lkml.kernel.org/r/20240729143743.34236-1-davidf@vimeo.com
Link: https://lkml.kernel.org/r/20240729143743.34236-2-davidf@vimeo.com
Signed-off-by: David Finkel <davidf@vimeo.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Waiman Long <longman@redhat.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-07-29 10:37:42 -04:00
|
|
|
|
mm: memcontrol: lockless page counters
Memory is internally accounted in bytes, using spinlock-protected 64-bit
counters, even though the smallest accounting delta is a page. The
counter interface is also convoluted and does too many things.
Introduce a new lockless word-sized page counter API, then change all
memory accounting over to it. The translation from and to bytes then only
happens when interfacing with userspace.
The removed locking overhead is noticable when scaling beyond the per-cpu
charge caches - on a 4-socket machine with 144-threads, the following test
shows the performance differences of 288 memcgs concurrently running a
page fault benchmark:
vanilla:
18631648.500498 task-clock (msec) # 140.643 CPUs utilized ( +- 0.33% )
1,380,638 context-switches # 0.074 K/sec ( +- 0.75% )
24,390 cpu-migrations # 0.001 K/sec ( +- 8.44% )
1,843,305,768 page-faults # 0.099 M/sec ( +- 0.00% )
50,134,994,088,218 cycles # 2.691 GHz ( +- 0.33% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
8,049,712,224,651 instructions # 0.16 insns per cycle ( +- 0.04% )
1,586,970,584,979 branches # 85.176 M/sec ( +- 0.05% )
1,724,989,949 branch-misses # 0.11% of all branches ( +- 0.48% )
132.474343877 seconds time elapsed ( +- 0.21% )
lockless:
12195979.037525 task-clock (msec) # 133.480 CPUs utilized ( +- 0.18% )
832,850 context-switches # 0.068 K/sec ( +- 0.54% )
15,624 cpu-migrations # 0.001 K/sec ( +- 10.17% )
1,843,304,774 page-faults # 0.151 M/sec ( +- 0.00% )
32,811,216,801,141 cycles # 2.690 GHz ( +- 0.18% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
9,999,265,091,727 instructions # 0.30 insns per cycle ( +- 0.10% )
2,076,759,325,203 branches # 170.282 M/sec ( +- 0.12% )
1,656,917,214 branch-misses # 0.08% of all branches ( +- 0.55% )
91.369330729 seconds time elapsed ( +- 0.45% )
On top of improved scalability, this also gets rid of the icky long long
types in the very heart of memcg, which is great for 32 bit and also makes
the code a lot more readable.
Notable differences between the old and new API:
- res_counter_charge() and res_counter_charge_nofail() become
page_counter_try_charge() and page_counter_charge() resp. to match
the more common kernel naming scheme of try_do()/do()
- res_counter_uncharge_until() is only ever used to cancel a local
counter and never to uncharge bigger segments of a hierarchy, so
it's replaced by the simpler page_counter_cancel()
- res_counter_set_limit() is replaced by page_counter_limit(), which
expects its callers to serialize against themselves
- res_counter_memparse_write_strategy() is replaced by
page_counter_limit(), which rounds down to the nearest page size -
rather than up. This is more reasonable for explicitely requested
hard upper limits.
- to keep charging light-weight, page_counter_try_charge() charges
speculatively, only to roll back if the result exceeds the limit.
Because of this, a failing bigger charge can temporarily lock out
smaller charges that would otherwise succeed. The error is bounded
to the difference between the smallest and the biggest possible
charge size, so for memcg, this means that a failing THP charge can
send base page charges into reclaim upto 2MB (4MB) before the limit
would have been reached. This should be acceptable.
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE and memparse]
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE, memparse, strncmp, and PAGE_SIZE]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Vladimir Davydov <vdavydov@parallels.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-12-10 15:42:31 -08:00
|
|
|
|
2008-02-07 00:13:50 -08:00
|
|
|
|
|
|
|
|
|
2025-04-24 16:28:06 -04:00
|
|
|
|
2017-02-08 18:51:29 +01:00
|
|
|
|
2017-02-24 14:59:36 -08:00
|
|
|
|
2010-03-10 15:22:14 -08:00
|
|
|
|
2009-01-07 18:07:56 -08:00
|
|
|
|
2024-02-27 17:42:39 +00:00
|
|
|
|
2019-03-05 15:48:09 -08:00
|
|
|
|
2008-02-07 00:14:24 -08:00
|
|
|
|
2008-02-07 00:13:53 -08:00
|
|
|
|
2008-02-07 00:13:56 -08:00
|
|
|
|
2008-02-07 00:13:53 -08:00
|
|
|
|
|
|
|
|
|
2009-04-02 16:57:39 -07:00
|
|
|
|
2011-05-26 16:00:52 -04:00
|
|
|
|
mm, memcg: cg2 memory{.swap,}.peak write handlers
Patch series "mm, memcg: cg2 memory{.swap,}.peak write handlers", v7.
This patch (of 2):
Other mechanisms for querying the peak memory usage of either a process or
v1 memory cgroup allow for resetting the high watermark. Restore parity
with those mechanisms, but with a less racy API.
For example:
- Any write to memory.max_usage_in_bytes in a cgroup v1 mount resets
the high watermark.
- writing "5" to the clear_refs pseudo-file in a processes's proc
directory resets the peak RSS.
This change is an evolution of a previous patch, which mostly copied the
cgroup v1 behavior, however, there were concerns about races/ownership
issues with a global reset, so instead this change makes the reset
filedescriptor-local.
Writing any non-empty string to the memory.peak and memory.swap.peak
pseudo-files reset the high watermark to the current usage for subsequent
reads through that same FD.
Notably, following Johannes's suggestion, this implementation moves the
O(FDs that have written) behavior onto the FD write(2) path. Instead, on
the page-allocation path, we simply add one additional watermark to
conditionally bump per-hierarchy level in the page-counter.
Additionally, this takes Longman's suggestion of nesting the
page-charging-path checks for the two watermarks to reduce the number of
common-case comparisons.
This behavior is particularly useful for work scheduling systems that need
to track memory usage of worker processes/cgroups per-work-item. Since
memory can't be squeezed like CPU can (the OOM-killer has opinions), these
systems need to track the peak memory usage to compute system/container
fullness when binpacking workitems.
Most notably, Vimeo's use-case involves a system that's doing global
binpacking across many Kubernetes pods/containers, and while we can use
PSI for some local decisions about overload, we strive to avoid packing
workloads too tightly in the first place. To facilitate this, we track
the peak memory usage. However, since we run with long-lived workers (to
amortize startup costs) we need a way to track the high watermark while a
work-item is executing. Polling runs the risk of missing short spikes
that last for timescales below the polling interval, and peak memory
tracking at the cgroup level is otherwise perfect for this use-case.
As this data is used to ensure that binpacked work ends up with sufficient
headroom, this use-case mostly avoids the inaccuracies surrounding
reclaimable memory.
Link: https://lkml.kernel.org/r/20240730231304.761942-1-davidf@vimeo.com
Link: https://lkml.kernel.org/r/20240729143743.34236-1-davidf@vimeo.com
Link: https://lkml.kernel.org/r/20240729143743.34236-2-davidf@vimeo.com
Signed-off-by: David Finkel <davidf@vimeo.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Waiman Long <longman@redhat.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-07-29 10:37:42 -04:00
|
|
|
|
2009-01-07 18:08:00 -08:00
|
|
|
|
2013-09-24 15:27:40 -07:00
|
|
|
|
2008-04-29 01:00:19 -07:00
|
|
|
|
2010-03-10 15:22:17 -08:00
|
|
|
|
2008-02-07 00:13:56 -08:00
|
|
|
|
|
|
|
|
|
2008-02-07 00:14:25 -08:00
|
|
|
|
memcg: add memory.pressure_level events
With this patch userland applications that want to maintain the
interactivity/memory allocation cost can use the pressure level
notifications. The levels are defined like this:
The "low" level means that the system is reclaiming memory for new
allocations. Monitoring this reclaiming activity might be useful for
maintaining cache level. Upon notification, the program (typically
"Activity Manager") might analyze vmstat and act in advance (i.e.
prematurely shutdown unimportant services).
The "medium" level means that the system is experiencing medium memory
pressure, the system might be making swap, paging out active file
caches, etc. Upon this event applications may decide to further analyze
vmstat/zoneinfo/memcg or internal memory usage statistics and free any
resources that can be easily reconstructed or re-read from a disk.
The "critical" level means that the system is actively thrashing, it is
about to out of memory (OOM) or even the in-kernel OOM killer is on its
way to trigger. Applications should do whatever they can to help the
system. It might be too late to consult with vmstat or any other
statistics, so it's advisable to take an immediate action.
The events are propagated upward until the event is handled, i.e. the
events are not pass-through. Here is what this means: for example you
have three cgroups: A->B->C. Now you set up an event listener on
cgroups A, B and C, and suppose group C experiences some pressure. In
this situation, only group C will receive the notification, i.e. groups
A and B will not receive it. This is done to avoid excessive
"broadcasting" of messages, which disturbs the system and which is
especially bad if we are low on memory or thrashing. So, organize the
cgroups wisely, or propagate the events manually (or, ask us to
implement the pass-through events, explaining why would you need them.)
Performance wise, the memory pressure notifications feature itself is
lightweight and does not require much of bookkeeping, in contrast to the
rest of memcg features. Unfortunately, as of current memcg
implementation, pages accounting is an inseparable part and cannot be
turned off. The good news is that there are some efforts[1] to improve
the situation; plus, implementing the same, fully API-compatible[2]
interface for CONFIG_MEMCG=n case (e.g. embedded) is also a viable
option, so it will not require any changes on the userland side.
[1] http://permalink.gmane.org/gmane.linux.kernel.cgroups/6291
[2] http://lkml.org/lkml/2013/2/21/454
[akpm@linux-foundation.org: coding-style fixes]
[akpm@linux-foundation.org: fix CONFIG_CGROPUPS=n warnings]
Signed-off-by: Anton Vorontsov <anton.vorontsov@linaro.org>
Acked-by: Kirill A. Shutemov <kirill@shutemov.name>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Glauber Costa <glommer@parallels.com>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Luiz Capitulino <lcapitulino@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Leonid Moiseichuk <leonid.moiseichuk@nokia.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com>
Cc: John Stultz <john.stultz@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-04-29 15:08:31 -07:00
|
|
|
|
2022-02-16 15:31:36 +11:00
|
|
|
|
2008-10-18 20:26:14 -07:00
|
|
|
|
2014-12-10 15:44:55 -08:00
|
|
|
|
2009-12-15 16:47:08 -08:00
|
|
|
|
2010-08-10 18:03:00 -07:00
|
|
|
|
2013-10-31 16:34:14 -07:00
|
|
|
|
2022-02-09 12:20:45 -06:00
|
|
|
|
mm, memcg: throttle allocators when failing reclaim over memory.high
We're trying to use memory.high to limit workloads, but have found that
containment can frequently fail completely and cause OOM situations
outside of the cgroup. This happens especially with swap space -- either
when none is configured, or swap is full. These failures often also don't
have enough warning to allow one to react, whether for a human or for a
daemon monitoring PSI.
Here is output from a simple program showing how long it takes in usec
(column 2) to allocate a megabyte of anonymous memory (column 1) when a
cgroup is already beyond its memory high setting, and no swap is
available:
[root@ktst ~]# systemd-run -p MemoryHigh=100M -p MemorySwapMax=1 \
> --wait -t timeout 300 /root/mdf
[...]
95 1035
96 1038
97 1000
98 1036
99 1048
100 1590
101 1968
102 1776
103 1863
104 1757
105 1921
106 1893
107 1760
108 1748
109 1843
110 1716
111 1924
112 1776
113 1831
114 1766
115 1836
116 1588
117 1912
118 1802
119 1857
120 1731
[...]
[System OOM in 2-3 seconds]
The delay does go up extremely marginally past the 100MB memory.high
threshold, as now we spend time scanning before returning to usermode, but
it's nowhere near enough to contain growth. It also doesn't get worse the
more pages you have, since it only considers nr_pages.
The current situation goes against both the expectations of users of
memory.high, and our intentions as cgroup v2 developers. In
cgroup-v2.txt, we claim that we will throttle and only under "extreme
conditions" will memory.high protection be breached. Likewise, cgroup v2
users generally also expect that memory.high should throttle workloads as
they exceed their high threshold. However, as seen above, this isn't
always how it works in practice -- even on banal setups like those with no
swap, or where swap has become exhausted, we can end up with memory.high
being breached and us having no weapons left in our arsenal to combat
runaway growth with, since reclaim is futile.
It's also hard for system monitoring software or users to tell how bad the
situation is, as "high" events for the memcg may in some cases be benign,
and in others be catastrophic. The current status quo is that we fail
containment in a way that doesn't provide any advance warning that things
are about to go horribly wrong (for example, we are about to invoke the
kernel OOM killer).
This patch introduces explicit throttling when reclaim is failing to keep
memcg size contained at the memory.high setting. It does so by applying
an exponential delay curve derived from the memcg's overage compared to
memory.high. In the normal case where the memcg is either below or only
marginally over its memory.high setting, no throttling will be performed.
This composes well with system health monitoring and remediation, as these
allocator delays are factored into PSI's memory pressure calculations.
This both creates a mechanism system administrators or applications
consuming the PSI interface to trivially see that the memcg in question is
struggling and use that to make more reasonable decisions, and permits
them enough time to act. Either of these can act with significantly more
nuance than that we can provide using the system OOM killer.
This is a similar idea to memory.oom_control in cgroup v1 which would put
the cgroup to sleep if the threshold was violated, but it's also
significantly improved as it results in visible memory pressure, and also
doesn't schedule indefinitely, which previously made tracing and other
introspection difficult (ie. it's clamped at 2*HZ per allocation through
MEMCG_MAX_HIGH_DELAY_JIFFIES).
Contrast the previous results with a kernel with this patch:
[root@ktst ~]# systemd-run -p MemoryHigh=100M -p MemorySwapMax=1 \
> --wait -t timeout 300 /root/mdf
[...]
95 1002
96 1000
97 1002
98 1003
99 1000
100 1043
101 84724
102 330628
103 610511
104 1016265
105 1503969
106 2391692
107 2872061
108 3248003
109 4791904
110 5759832
111 6912509
112 8127818
113 9472203
114 12287622
115 12480079
116 14144008
117 15808029
118 16384500
119 16383242
120 16384979
[...]
As you can see, in the normal case, memory allocation takes around 1000
usec. However, as we exceed our memory.high, things start to increase
exponentially, but fairly leniently at first. Our first megabyte over
memory.high takes us 0.16 seconds, then the next is 0.46 seconds, then the
next is almost an entire second. This gets worse until we reach our
eventual 2*HZ clamp per batch, resulting in 16 seconds per megabyte.
However, this is still making forward progress, so permits tracing or
further analysis with programs like GDB.
We use an exponential curve for our delay penalty for a few reasons:
1. We run mem_cgroup_handle_over_high to potentially do reclaim after
we've already performed allocations, which means that temporarily
going over memory.high by a small amount may be perfectly legitimate,
even for compliant workloads. We don't want to unduly penalise such
cases.
2. An exponential curve (as opposed to a static or linear delay) allows
ramping up memory pressure stats more gradually, which can be useful
to work out that you have set memory.high too low, without destroying
application performance entirely.
This patch expands on earlier work by Johannes Weiner. Thanks!
[akpm@linux-foundation.org: fix max() warning]
[akpm@linux-foundation.org: fix __udivdi3 ref on 32-bit]
[akpm@linux-foundation.org: fix it even more]
[chris@chrisdown.name: fix 64-bit divide even more]
Link: http://lkml.kernel.org/r/20190723180700.GA29459@chrisdown.name
Signed-off-by: Chris Down <chris@chrisdown.name>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Nathan Chancellor <natechancellor@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-09-23 15:34:55 -07:00
|
|
|
|
mm: memcontrol: dump memory.stat during cgroup OOM
The current cgroup OOM memory info dump doesn't include all the memory
we are tracking, nor does it give insight into what the VM tried to do
leading up to the OOM. All that useful info is in memory.stat.
Furthermore, the recursive printing for every child cgroup can
generate absurd amounts of data on the console for larger cgroup
trees, and it's not like we provide a per-cgroup breakdown during
global OOM kills.
When an OOM kill is triggered, print one set of recursive memory.stat
items at the level whose limit triggered the OOM condition.
Example output:
stress invoked oom-killer: gfp_mask=0x100cca(GFP_HIGHUSER_MOVABLE), order=0, oom_score_adj=0
CPU: 2 PID: 210 Comm: stress Not tainted 5.2.0-rc2-mm1-00247-g47d49835983c #135
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.12.0-20181126_142135-anatol 04/01/2014
Call Trace:
dump_stack+0x46/0x60
dump_header+0x4c/0x2d0
oom_kill_process.cold.10+0xb/0x10
out_of_memory+0x200/0x270
? try_to_free_mem_cgroup_pages+0xdf/0x130
mem_cgroup_out_of_memory+0xb7/0xc0
try_charge+0x680/0x6f0
mem_cgroup_try_charge+0xb5/0x160
__add_to_page_cache_locked+0xc6/0x300
? list_lru_destroy+0x80/0x80
add_to_page_cache_lru+0x45/0xc0
pagecache_get_page+0x11b/0x290
filemap_fault+0x458/0x6d0
ext4_filemap_fault+0x27/0x36
__do_fault+0x2f/0xb0
__handle_mm_fault+0x9c5/0x1140
? apic_timer_interrupt+0xa/0x20
handle_mm_fault+0xc5/0x180
__do_page_fault+0x1ab/0x440
? page_fault+0x8/0x30
page_fault+0x1e/0x30
RIP: 0033:0x55c32167fc10
Code: Bad RIP value.
RSP: 002b:00007fff1d031c50 EFLAGS: 00010206
RAX: 000000000dc00000 RBX: 00007fd2db000010 RCX: 00007fd2db000010
RDX: 0000000000000000 RSI: 0000000010001000 RDI: 0000000000000000
RBP: 000055c321680a54 R08: 00000000ffffffff R09: 0000000000000000
R10: 0000000000000022 R11: 0000000000000246 R12: ffffffffffffffff
R13: 0000000000000002 R14: 0000000000001000 R15: 0000000010000000
memory: usage 1024kB, limit 1024kB, failcnt 75131
swap: usage 0kB, limit 9007199254740988kB, failcnt 0
Memory cgroup stats for /foo:
anon 0
file 0
kernel_stack 36864
slab 274432
sock 0
shmem 0
file_mapped 0
file_dirty 0
file_writeback 0
anon_thp 0
inactive_anon 126976
active_anon 0
inactive_file 0
active_file 0
unevictable 0
slab_reclaimable 0
slab_unreclaimable 274432
pgfault 59466
pgmajfault 1617
workingset_refault 2145
workingset_activate 0
workingset_nodereclaim 0
pgrefill 98952
pgscan 200060
pgsteal 59340
pgactivate 40095
pgdeactivate 96787
pglazyfree 0
pglazyfreed 0
thp_fault_alloc 0
thp_collapse_alloc 0
Tasks state (memory values in pages):
[ pid ] uid tgid total_vm rss pgtables_bytes swapents oom_score_adj name
[ 200] 0 200 1121 884 53248 29 0 bash
[ 209] 0 209 905 246 45056 19 0 stress
[ 210] 0 210 66442 56 499712 56349 0 stress
oom-kill:constraint=CONSTRAINT_NONE,nodemask=(null),oom_memcg=/foo,task_memcg=/foo,task=stress,pid=210,uid=0
Memory cgroup out of memory: Killed process 210 (stress) total-vm:265768kB, anon-rss:0kB, file-rss:224kB, shmem-rss:0kB
oom_reaper: reaped process 210 (stress), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
[hannes@cmpxchg.org: s/kvmalloc/kmalloc/ per Michal]
Link: http://lkml.kernel.org/r/20190605161133.GA12453@cmpxchg.org
Link: http://lkml.kernel.org/r/20190604210509.9744-1-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-11 20:55:59 -07:00
|
|
|
|
2023-03-17 14:44:48 +01:00
|
|
|
|
2023-10-03 11:57:45 +02:00
|
|
|
|
memcg: synchronized LRU
A big patch for changing memcg's LRU semantics.
Now,
- page_cgroup is linked to mem_cgroup's its own LRU (per zone).
- LRU of page_cgroup is not synchronous with global LRU.
- page and page_cgroup is one-to-one and statically allocated.
- To find page_cgroup is on what LRU, you have to check pc->mem_cgroup as
- lru = page_cgroup_zoneinfo(pc, nid_of_pc, zid_of_pc);
- SwapCache is handled.
And, when we handle LRU list of page_cgroup, we do following.
pc = lookup_page_cgroup(page);
lock_page_cgroup(pc); .....................(1)
mz = page_cgroup_zoneinfo(pc);
spin_lock(&mz->lru_lock);
.....add to LRU
spin_unlock(&mz->lru_lock);
unlock_page_cgroup(pc);
But (1) is spin_lock and we have to be afraid of dead-lock with zone->lru_lock.
So, trylock() is used at (1), now. Without (1), we can't trust "mz" is correct.
This is a trial to remove this dirty nesting of locks.
This patch changes mz->lru_lock to be zone->lru_lock.
Then, above sequence will be written as
spin_lock(&zone->lru_lock); # in vmscan.c or swap.c via global LRU
mem_cgroup_add/remove/etc_lru() {
pc = lookup_page_cgroup(page);
mz = page_cgroup_zoneinfo(pc);
if (PageCgroupUsed(pc)) {
....add to LRU
}
spin_lock(&zone->lru_lock); # in vmscan.c or swap.c via global LRU
This is much simpler.
(*) We're safe even if we don't take lock_page_cgroup(pc). Because..
1. When pc->mem_cgroup can be modified.
- at charge.
- at account_move().
2. at charge
the PCG_USED bit is not set before pc->mem_cgroup is fixed.
3. at account_move()
the page is isolated and not on LRU.
Pros.
- easy for maintenance.
- memcg can make use of laziness of pagevec.
- we don't have to duplicated LRU/Active/Unevictable bit in page_cgroup.
- LRU status of memcg will be synchronized with global LRU's one.
- # of locks are reduced.
- account_move() is simplified very much.
Cons.
- may increase cost of LRU rotation.
(no impact if memcg is not configured.)
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Li Zefan <lizf@cn.fujitsu.com>
Cc: Balbir Singh <balbir@in.ibm.com>
Cc: Pavel Emelyanov <xemul@openvz.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-01-07 18:08:01 -08:00
|
|
|
|
2011-12-11 21:47:04 +00:00
|
|
|
|
2012-10-08 16:33:10 -07:00
|
|
|
|
2013-11-12 15:08:22 -08:00
|
|
|
|
2024-06-24 17:58:54 -07:00
|
|
|
|
2008-02-07 00:13:50 -08:00
|
|
|
|
2016-12-24 11:46:01 -08:00
|
|
|
|
2008-02-07 00:13:59 -08:00
|
|
|
|
2024-10-09 17:35:50 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-08-09 17:19:57 -07:00
|
|
|
|
|
|
|
|
|
2014-02-08 10:36:58 -05:00
|
|
|
|
|
|
|
|
|
2012-12-12 13:51:57 -08:00
|
|
|
|
2016-01-14 15:20:56 -08:00
|
|
|
|
|
|
|
|
|
mm: kmem: prepare remote memcg charging infra for interrupt contexts
Remote memcg charging API uses current->active_memcg to store the
currently active memory cgroup, which overwrites the memory cgroup of the
current process. It works well for normal contexts, but doesn't work for
interrupt contexts: indeed, if an interrupt occurs during the execution of
a section with an active memcg set, all allocations inside the interrupt
will be charged to the active memcg set (given that we'll enable
accounting for allocations from an interrupt context). But because the
interrupt might have no relation to the active memcg set outside, it's
obviously wrong from the accounting prospective.
To resolve this problem, let's add a global percpu int_active_memcg
variable, which will be used to store an active memory cgroup which will
be used from interrupt contexts. set_active_memcg() will transparently
use current->active_memcg or int_active_memcg depending on the context.
To make the read part simple and transparent for the caller, let's
introduce two new functions:
- struct mem_cgroup *active_memcg(void),
- struct mem_cgroup *get_active_memcg(void).
They are returning the active memcg if it's set, hiding all implementation
details: where to get it depending on the current context.
Signed-off-by: Roman Gushchin <guro@fb.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Link: http://lkml.kernel.org/r/20200827225843.1270629-4-guro@fb.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-10-17 16:13:50 -07:00
|
|
|
|
|
|
|
|
|
2021-06-28 19:38:21 -07:00
|
|
|
|
mm: kmem: prepare remote memcg charging infra for interrupt contexts
Remote memcg charging API uses current->active_memcg to store the
currently active memory cgroup, which overwrites the memory cgroup of the
current process. It works well for normal contexts, but doesn't work for
interrupt contexts: indeed, if an interrupt occurs during the execution of
a section with an active memcg set, all allocations inside the interrupt
will be charged to the active memcg set (given that we'll enable
accounting for allocations from an interrupt context). But because the
interrupt might have no relation to the active memcg set outside, it's
obviously wrong from the accounting prospective.
To resolve this problem, let's add a global percpu int_active_memcg
variable, which will be used to store an active memory cgroup which will
be used from interrupt contexts. set_active_memcg() will transparently
use current->active_memcg or int_active_memcg depending on the context.
To make the read part simple and transparent for the caller, let's
introduce two new functions:
- struct mem_cgroup *active_memcg(void),
- struct mem_cgroup *get_active_memcg(void).
They are returning the active memcg if it's set, hiding all implementation
details: where to get it depending on the current context.
Signed-off-by: Roman Gushchin <guro@fb.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Link: http://lkml.kernel.org/r/20200827225843.1270629-4-guro@fb.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-10-17 16:13:50 -07:00
|
|
|
|
2016-01-14 15:21:29 -08:00
|
|
|
|
2021-06-02 18:09:29 -07:00
|
|
|
|
2016-01-14 15:21:29 -08:00
|
|
|
|
2016-01-20 15:02:38 -08:00
|
|
|
|
2022-01-14 14:05:29 -08:00
|
|
|
|
2016-01-20 15:02:38 -08:00
|
|
|
|
2023-02-10 15:47:31 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-04-25 11:19:24 +08:00
|
|
|
|
2025-04-25 11:19:25 +08:00
|
|
|
|
2025-04-25 11:19:24 +08:00
|
|
|
|
writeback, memcg: Implement foreign dirty flushing
There's an inherent mismatch between memcg and writeback. The former
trackes ownership per-page while the latter per-inode. This was a
deliberate design decision because honoring per-page ownership in the
writeback path is complicated, may lead to higher CPU and IO overheads
and deemed unnecessary given that write-sharing an inode across
different cgroups isn't a common use-case.
Combined with inode majority-writer ownership switching, this works
well enough in most cases but there are some pathological cases. For
example, let's say there are two cgroups A and B which keep writing to
different but confined parts of the same inode. B owns the inode and
A's memory is limited far below B's. A's dirty ratio can rise enough
to trigger balance_dirty_pages() sleeps but B's can be low enough to
avoid triggering background writeback. A will be slowed down without
a way to make writeback of the dirty pages happen.
This patch implements foreign dirty recording and foreign mechanism so
that when a memcg encounters a condition as above it can trigger
flushes on bdi_writebacks which can clean its pages. Please see the
comment on top of mem_cgroup_track_foreign_dirty_slowpath() for
details.
A reproducer follows.
write-range.c::
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <fcntl.h>
#include <sys/types.h>
static const char *usage = "write-range FILE START SIZE\n";
int main(int argc, char **argv)
{
int fd;
unsigned long start, size, end, pos;
char *endp;
char buf[4096];
if (argc < 4) {
fprintf(stderr, usage);
return 1;
}
fd = open(argv[1], O_WRONLY);
if (fd < 0) {
perror("open");
return 1;
}
start = strtoul(argv[2], &endp, 0);
if (*endp != '\0') {
fprintf(stderr, usage);
return 1;
}
size = strtoul(argv[3], &endp, 0);
if (*endp != '\0') {
fprintf(stderr, usage);
return 1;
}
end = start + size;
while (1) {
for (pos = start; pos < end; ) {
long bread, bwritten = 0;
if (lseek(fd, pos, SEEK_SET) < 0) {
perror("lseek");
return 1;
}
bread = read(0, buf, sizeof(buf) < end - pos ?
sizeof(buf) : end - pos);
if (bread < 0) {
perror("read");
return 1;
}
if (bread == 0)
return 0;
while (bwritten < bread) {
long this;
this = write(fd, buf + bwritten,
bread - bwritten);
if (this < 0) {
perror("write");
return 1;
}
bwritten += this;
pos += bwritten;
}
}
}
}
repro.sh::
#!/bin/bash
set -e
set -x
sysctl -w vm.dirty_expire_centisecs=300000
sysctl -w vm.dirty_writeback_centisecs=300000
sysctl -w vm.dirtytime_expire_seconds=300000
echo 3 > /proc/sys/vm/drop_caches
TEST=/sys/fs/cgroup/test
A=$TEST/A
B=$TEST/B
mkdir -p $A $B
echo "+memory +io" > $TEST/cgroup.subtree_control
echo $((1<<30)) > $A/memory.high
echo $((32<<30)) > $B/memory.high
rm -f testfile
touch testfile
fallocate -l 4G testfile
echo "Starting B"
(echo $BASHPID > $B/cgroup.procs
pv -q --rate-limit 70M < /dev/urandom | ./write-range testfile $((2<<30)) $((2<<30))) &
echo "Waiting 10s to ensure B claims the testfile inode"
sleep 5
sync
sleep 5
sync
echo "Starting A"
(echo $BASHPID > $A/cgroup.procs
pv < /dev/urandom | ./write-range testfile 0 $((2<<30)))
v2: Added comments explaining why the specific intervals are being used.
v3: Use 0 @nr when calling cgroup_writeback_by_id() to use best-effort
flushing while avoding possible livelocks.
v4: Use get_jiffies_64() and time_before/after64() instead of raw
jiffies_64 and arthimetic comparisons as suggested by Jan.
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-08-26 09:06:56 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-11-05 13:38:09 -07:00
|
|
|
|
2019-03-05 15:46:47 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
memcg: add memory.pressure_level events
With this patch userland applications that want to maintain the
interactivity/memory allocation cost can use the pressure level
notifications. The levels are defined like this:
The "low" level means that the system is reclaiming memory for new
allocations. Monitoring this reclaiming activity might be useful for
maintaining cache level. Upon notification, the program (typically
"Activity Manager") might analyze vmstat and act in advance (i.e.
prematurely shutdown unimportant services).
The "medium" level means that the system is experiencing medium memory
pressure, the system might be making swap, paging out active file
caches, etc. Upon this event applications may decide to further analyze
vmstat/zoneinfo/memcg or internal memory usage statistics and free any
resources that can be easily reconstructed or re-read from a disk.
The "critical" level means that the system is actively thrashing, it is
about to out of memory (OOM) or even the in-kernel OOM killer is on its
way to trigger. Applications should do whatever they can to help the
system. It might be too late to consult with vmstat or any other
statistics, so it's advisable to take an immediate action.
The events are propagated upward until the event is handled, i.e. the
events are not pass-through. Here is what this means: for example you
have three cgroups: A->B->C. Now you set up an event listener on
cgroups A, B and C, and suppose group C experiences some pressure. In
this situation, only group C will receive the notification, i.e. groups
A and B will not receive it. This is done to avoid excessive
"broadcasting" of messages, which disturbs the system and which is
especially bad if we are low on memory or thrashing. So, organize the
cgroups wisely, or propagate the events manually (or, ask us to
implement the pass-through events, explaining why would you need them.)
Performance wise, the memory pressure notifications feature itself is
lightweight and does not require much of bookkeeping, in contrast to the
rest of memcg features. Unfortunately, as of current memcg
implementation, pages accounting is an inseparable part and cannot be
turned off. The good news is that there are some efforts[1] to improve
the situation; plus, implementing the same, fully API-compatible[2]
interface for CONFIG_MEMCG=n case (e.g. embedded) is also a viable
option, so it will not require any changes on the userland side.
[1] http://permalink.gmane.org/gmane.linux.kernel.cgroups/6291
[2] http://lkml.org/lkml/2013/2/21/454
[akpm@linux-foundation.org: coding-style fixes]
[akpm@linux-foundation.org: fix CONFIG_CGROPUPS=n warnings]
Signed-off-by: Anton Vorontsov <anton.vorontsov@linaro.org>
Acked-by: Kirill A. Shutemov <kirill@shutemov.name>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Glauber Costa <glommer@parallels.com>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Luiz Capitulino <lcapitulino@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Leonid Moiseichuk <leonid.moiseichuk@nokia.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com>
Cc: John Stultz <john.stultz@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-04-29 15:08:31 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-09-02 14:59:36 -07:00
|
|
|
|
memcg: add memory.pressure_level events
With this patch userland applications that want to maintain the
interactivity/memory allocation cost can use the pressure level
notifications. The levels are defined like this:
The "low" level means that the system is reclaiming memory for new
allocations. Monitoring this reclaiming activity might be useful for
maintaining cache level. Upon notification, the program (typically
"Activity Manager") might analyze vmstat and act in advance (i.e.
prematurely shutdown unimportant services).
The "medium" level means that the system is experiencing medium memory
pressure, the system might be making swap, paging out active file
caches, etc. Upon this event applications may decide to further analyze
vmstat/zoneinfo/memcg or internal memory usage statistics and free any
resources that can be easily reconstructed or re-read from a disk.
The "critical" level means that the system is actively thrashing, it is
about to out of memory (OOM) or even the in-kernel OOM killer is on its
way to trigger. Applications should do whatever they can to help the
system. It might be too late to consult with vmstat or any other
statistics, so it's advisable to take an immediate action.
The events are propagated upward until the event is handled, i.e. the
events are not pass-through. Here is what this means: for example you
have three cgroups: A->B->C. Now you set up an event listener on
cgroups A, B and C, and suppose group C experiences some pressure. In
this situation, only group C will receive the notification, i.e. groups
A and B will not receive it. This is done to avoid excessive
"broadcasting" of messages, which disturbs the system and which is
especially bad if we are low on memory or thrashing. So, organize the
cgroups wisely, or propagate the events manually (or, ask us to
implement the pass-through events, explaining why would you need them.)
Performance wise, the memory pressure notifications feature itself is
lightweight and does not require much of bookkeeping, in contrast to the
rest of memcg features. Unfortunately, as of current memcg
implementation, pages accounting is an inseparable part and cannot be
turned off. The good news is that there are some efforts[1] to improve
the situation; plus, implementing the same, fully API-compatible[2]
interface for CONFIG_MEMCG=n case (e.g. embedded) is also a viable
option, so it will not require any changes on the userland side.
[1] http://permalink.gmane.org/gmane.linux.kernel.cgroups/6291
[2] http://lkml.org/lkml/2013/2/21/454
[akpm@linux-foundation.org: coding-style fixes]
[akpm@linux-foundation.org: fix CONFIG_CGROPUPS=n warnings]
Signed-off-by: Anton Vorontsov <anton.vorontsov@linaro.org>
Acked-by: Kirill A. Shutemov <kirill@shutemov.name>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Glauber Costa <glommer@parallels.com>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Luiz Capitulino <lcapitulino@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Leonid Moiseichuk <leonid.moiseichuk@nokia.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com>
Cc: John Stultz <john.stultz@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-04-29 15:08:31 -07:00
|
|
|
|
2021-09-02 14:59:36 -07:00
|
|
|
|
memcg: add memory.pressure_level events
With this patch userland applications that want to maintain the
interactivity/memory allocation cost can use the pressure level
notifications. The levels are defined like this:
The "low" level means that the system is reclaiming memory for new
allocations. Monitoring this reclaiming activity might be useful for
maintaining cache level. Upon notification, the program (typically
"Activity Manager") might analyze vmstat and act in advance (i.e.
prematurely shutdown unimportant services).
The "medium" level means that the system is experiencing medium memory
pressure, the system might be making swap, paging out active file
caches, etc. Upon this event applications may decide to further analyze
vmstat/zoneinfo/memcg or internal memory usage statistics and free any
resources that can be easily reconstructed or re-read from a disk.
The "critical" level means that the system is actively thrashing, it is
about to out of memory (OOM) or even the in-kernel OOM killer is on its
way to trigger. Applications should do whatever they can to help the
system. It might be too late to consult with vmstat or any other
statistics, so it's advisable to take an immediate action.
The events are propagated upward until the event is handled, i.e. the
events are not pass-through. Here is what this means: for example you
have three cgroups: A->B->C. Now you set up an event listener on
cgroups A, B and C, and suppose group C experiences some pressure. In
this situation, only group C will receive the notification, i.e. groups
A and B will not receive it. This is done to avoid excessive
"broadcasting" of messages, which disturbs the system and which is
especially bad if we are low on memory or thrashing. So, organize the
cgroups wisely, or propagate the events manually (or, ask us to
implement the pass-through events, explaining why would you need them.)
Performance wise, the memory pressure notifications feature itself is
lightweight and does not require much of bookkeeping, in contrast to the
rest of memcg features. Unfortunately, as of current memcg
implementation, pages accounting is an inseparable part and cannot be
turned off. The good news is that there are some efforts[1] to improve
the situation; plus, implementing the same, fully API-compatible[2]
interface for CONFIG_MEMCG=n case (e.g. embedded) is also a viable
option, so it will not require any changes on the userland side.
[1] http://permalink.gmane.org/gmane.linux.kernel.cgroups/6291
[2] http://lkml.org/lkml/2013/2/21/454
[akpm@linux-foundation.org: coding-style fixes]
[akpm@linux-foundation.org: fix CONFIG_CGROPUPS=n warnings]
Signed-off-by: Anton Vorontsov <anton.vorontsov@linaro.org>
Acked-by: Kirill A. Shutemov <kirill@shutemov.name>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Glauber Costa <glommer@parallels.com>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Luiz Capitulino <lcapitulino@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Leonid Moiseichuk <leonid.moiseichuk@nokia.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com>
Cc: John Stultz <john.stultz@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-04-29 15:08:31 -07:00
|
|
|
|
|
|
|
|
|
2024-10-21 14:00:26 +01:00
|
|
|
|
2023-10-19 15:53:42 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-02-11 16:32:32 -08:00
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
2021-06-02 18:09:30 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-04-03 18:39:09 -07:00
|
|
|
|
2021-02-24 12:03:58 -08:00
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-04-03 18:39:09 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-28 19:38:06 -07:00
|
|
|
|
2022-02-11 16:32:32 -08:00
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
2022-02-11 16:32:32 -08:00
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-02-11 16:32:32 -08:00
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
2021-06-28 19:38:03 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
|
|
|
|
|
2022-02-11 16:32:32 -08:00
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-12-18 14:22:40 -08:00
|
|
|
|
|
|
|
|
|
2024-03-26 11:37:38 +01:00
|
|
|
|
2012-12-18 14:22:40 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-02-13 11:29:22 -08:00
|
|
|
|
|
|
|
|
|
2023-02-10 15:47:31 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-02-22 15:41:36 -08:00
|
|
|
|
2015-05-27 20:00:02 -04:00
|
|
|
|
2023-01-16 19:25:07 +00:00
|
|
|
|
|
|
|
|
|
2015-05-27 20:00:02 -04:00
|
|
|
|
|
|
|
|
|
2023-01-16 19:25:07 +00:00
|
|
|
|
2015-05-27 20:00:02 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-01-16 19:25:07 +00:00
|
|
|
|
2015-05-27 20:00:02 -04:00
|
|
|
|
2023-01-16 19:25:07 +00:00
|
|
|
|
2015-05-27 20:00:02 -04:00
|
|
|
|
2015-09-18 11:56:28 -04:00
|
|
|
|
2015-05-27 20:00:02 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
memcg: add page_cgroup_ino helper
This patchset introduces a new user API for tracking user memory pages
that have not been used for a given period of time. The purpose of this
is to provide the userspace with the means of tracking a workload's
working set, i.e. the set of pages that are actively used by the
workload. Knowing the working set size can be useful for partitioning the
system more efficiently, e.g. by tuning memory cgroup limits
appropriately, or for job placement within a compute cluster.
==== USE CASES ====
The unified cgroup hierarchy has memory.low and memory.high knobs, which
are defined as the low and high boundaries for the workload working set
size. However, the working set size of a workload may be unknown or
change in time. With this patch set, one can periodically estimate the
amount of memory unused by each cgroup and tune their memory.low and
memory.high parameters accordingly, therefore optimizing the overall
memory utilization.
Another use case is balancing workloads within a compute cluster. Knowing
how much memory is not really used by a workload unit may help take a more
optimal decision when considering migrating the unit to another node
within the cluster.
Also, as noted by Minchan, this would be useful for per-process reclaim
(https://lwn.net/Articles/545668/). With idle tracking, we could reclaim idle
pages only by smart user memory manager.
==== USER API ====
The user API consists of two new files:
* /sys/kernel/mm/page_idle/bitmap. This file implements a bitmap where each
bit corresponds to a page, indexed by PFN. When the bit is set, the
corresponding page is idle. A page is considered idle if it has not been
accessed since it was marked idle. To mark a page idle one should set the
bit corresponding to the page by writing to the file. A value written to the
file is OR-ed with the current bitmap value. Only user memory pages can be
marked idle, for other page types input is silently ignored. Writing to this
file beyond max PFN results in the ENXIO error. Only available when
CONFIG_IDLE_PAGE_TRACKING is set.
This file can be used to estimate the amount of pages that are not
used by a particular workload as follows:
1. mark all pages of interest idle by setting corresponding bits in the
/sys/kernel/mm/page_idle/bitmap
2. wait until the workload accesses its working set
3. read /sys/kernel/mm/page_idle/bitmap and count the number of bits set
* /proc/kpagecgroup. This file contains a 64-bit inode number of the
memory cgroup each page is charged to, indexed by PFN. Only available when
CONFIG_MEMCG is set.
This file can be used to find all pages (including unmapped file pages)
accounted to a particular cgroup. Using /sys/kernel/mm/page_idle/bitmap, one
can then estimate the cgroup working set size.
For an example of using these files for estimating the amount of unused
memory pages per each memory cgroup, please see the script attached
below.
==== REASONING ====
The reason to introduce the new user API instead of using
/proc/PID/{clear_refs,smaps} is that the latter has two serious
drawbacks:
- it does not count unmapped file pages
- it affects the reclaimer logic
The new API attempts to overcome them both. For more details on how it
is achieved, please see the comment to patch 6.
==== PATCHSET STRUCTURE ====
The patch set is organized as follows:
- patch 1 adds page_cgroup_ino() helper for the sake of
/proc/kpagecgroup and patches 2-3 do related cleanup
- patch 4 adds /proc/kpagecgroup, which reports cgroup ino each page is
charged to
- patch 5 introduces a new mmu notifier callback, clear_young, which is
a lightweight version of clear_flush_young; it is used in patch 6
- patch 6 implements the idle page tracking feature, including the
userspace API, /sys/kernel/mm/page_idle/bitmap
- patch 7 exports idle flag via /proc/kpageflags
==== SIMILAR WORKS ====
Originally, the patch for tracking idle memory was proposed back in 2011
by Michel Lespinasse (see http://lwn.net/Articles/459269/). The main
difference between Michel's patch and this one is that Michel implemented
a kernel space daemon for estimating idle memory size per cgroup while
this patch only provides the userspace with the minimal API for doing the
job, leaving the rest up to the userspace. However, they both share the
same idea of Idle/Young page flags to avoid affecting the reclaimer logic.
==== PERFORMANCE EVALUATION ====
SPECjvm2008 (https://www.spec.org/jvm2008/) was used to evaluate the
performance impact introduced by this patch set. Three runs were carried
out:
- base: kernel without the patch
- patched: patched kernel, the feature is not used
- patched-active: patched kernel, 1 minute-period daemon is used for
tracking idle memory
For tracking idle memory, idlememstat utility was used:
https://github.com/locker/idlememstat
testcase base patched patched-active
compiler 537.40 ( 0.00)% 532.26 (-0.96)% 538.31 ( 0.17)%
compress 305.47 ( 0.00)% 301.08 (-1.44)% 300.71 (-1.56)%
crypto 284.32 ( 0.00)% 282.21 (-0.74)% 284.87 ( 0.19)%
derby 411.05 ( 0.00)% 413.44 ( 0.58)% 412.07 ( 0.25)%
mpegaudio 189.96 ( 0.00)% 190.87 ( 0.48)% 189.42 (-0.28)%
scimark.large 46.85 ( 0.00)% 46.41 (-0.94)% 47.83 ( 2.09)%
scimark.small 412.91 ( 0.00)% 415.41 ( 0.61)% 421.17 ( 2.00)%
serial 204.23 ( 0.00)% 213.46 ( 4.52)% 203.17 (-0.52)%
startup 36.76 ( 0.00)% 35.49 (-3.45)% 35.64 (-3.05)%
sunflow 115.34 ( 0.00)% 115.08 (-0.23)% 117.37 ( 1.76)%
xml 620.55 ( 0.00)% 619.95 (-0.10)% 620.39 (-0.03)%
composite 211.50 ( 0.00)% 211.15 (-0.17)% 211.67 ( 0.08)%
time idlememstat:
17.20user 65.16system 2:15:23elapsed 1%CPU (0avgtext+0avgdata 8476maxresident)k
448inputs+40outputs (1major+36052minor)pagefaults 0swaps
==== SCRIPT FOR COUNTING IDLE PAGES PER CGROUP ====
#! /usr/bin/python
#
import os
import stat
import errno
import struct
CGROUP_MOUNT = "/sys/fs/cgroup/memory"
BUFSIZE = 8 * 1024 # must be multiple of 8
def get_hugepage_size():
with open("/proc/meminfo", "r") as f:
for s in f:
k, v = s.split(":")
if k == "Hugepagesize":
return int(v.split()[0]) * 1024
PAGE_SIZE = os.sysconf("SC_PAGE_SIZE")
HUGEPAGE_SIZE = get_hugepage_size()
def set_idle():
f = open("/sys/kernel/mm/page_idle/bitmap", "wb", BUFSIZE)
while True:
try:
f.write(struct.pack("Q", pow(2, 64) - 1))
except IOError as err:
if err.errno == errno.ENXIO:
break
raise
f.close()
def count_idle():
f_flags = open("/proc/kpageflags", "rb", BUFSIZE)
f_cgroup = open("/proc/kpagecgroup", "rb", BUFSIZE)
with open("/sys/kernel/mm/page_idle/bitmap", "rb", BUFSIZE) as f:
while f.read(BUFSIZE): pass # update idle flag
idlememsz = {}
while True:
s1, s2 = f_flags.read(8), f_cgroup.read(8)
if not s1 or not s2:
break
flags, = struct.unpack('Q', s1)
cgino, = struct.unpack('Q', s2)
unevictable = (flags >> 18) & 1
huge = (flags >> 22) & 1
idle = (flags >> 25) & 1
if idle and not unevictable:
idlememsz[cgino] = idlememsz.get(cgino, 0) + \
(HUGEPAGE_SIZE if huge else PAGE_SIZE)
f_flags.close()
f_cgroup.close()
return idlememsz
if __name__ == "__main__":
print "Setting the idle flag for each page..."
set_idle()
raw_input("Wait until the workload accesses its working set, "
"then press Enter")
print "Counting idle pages..."
idlememsz = count_idle()
for dir, subdirs, files in os.walk(CGROUP_MOUNT):
ino = os.stat(dir)[stat.ST_INO]
print dir + ": " + str(idlememsz.get(ino, 0) / 1024) + " kB"
==== END SCRIPT ====
This patch (of 8):
Add page_cgroup_ino() helper to memcg.
This function returns the inode number of the closest online ancestor of
the memory cgroup a page is charged to. It is required for exporting
information about which page is charged to which cgroup to userspace,
which will be introduced by a following patch.
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Reviewed-by: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Greg Thelen <gthelen@google.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-09 15:35:28 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
memcg: page_cgroup_ino() get memcg from the page's folio
In a kernel with added WARN_ON_ONCE(PageTail) in page_memcg_check(), we
observed a warning from page_cgroup_ino() when reading /proc/kpagecgroup.
This warning was added to catch fragile reads of a page memcg. Make
page_cgroup_ino() get memcg from the page's folio using
folio_memcg_check(): that gives it the correct memcg for each page of a
folio, so is the right fix.
Note that page_folio() is racy, the page's folio can change from under us,
but the entire function is racy and documented as such.
I dithered between the right fix and the safer "fix": it's unlikely but
conceivable that some userspace has learnt that /proc/kpagecgroup gives no
memcg on tail pages, and compensates for that in some (racy) way: so
continuing to give no memcg on tails, without warning, might be safer.
But hwpoison_filter_task(), the only other user of page_cgroup_ino(),
persuaded me. It looks as if it currently leaves out tail pages of the
selected memcg, by mistake: whereas hwpoison_inject() uses compound_head()
and expects the tails to be included. So hwpoison testing coverage has
probably been restricted by the wrong output from page_cgroup_ino() (if
that memcg filter is used at all): in the short term, it might be safer
not to enable wider coverage there, but long term we would regret that.
This is based on a patch originally written by Hugh Dickins and retains
most of the original commit log [1]
The patch was changed to use folio_memcg_check(page_folio(page)) instead
of page_memcg_check(compound_head(page)) based on discussions with Matthew
Wilcox; where he stated that callers of page_memcg_check() should stop
using it due to the ambiguity around tail pages -- instead they should use
folio_memcg_check() and handle tail pages themselves.
Link: https://lkml.kernel.org/r/20230412003451.4018887-1-yosryahmed@google.com
Link: https://lore.kernel.org/linux-mm/20230313083452.1319968-1-yosryahmed@google.com/ [1]
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Naoya Horiguchi <naoya.horiguchi@nec.com>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-04-12 00:34:51 +00:00
|
|
|
|
|
|
|
|
|
2020-08-06 23:20:52 -07:00
|
|
|
|
memcg: add page_cgroup_ino helper
This patchset introduces a new user API for tracking user memory pages
that have not been used for a given period of time. The purpose of this
is to provide the userspace with the means of tracking a workload's
working set, i.e. the set of pages that are actively used by the
workload. Knowing the working set size can be useful for partitioning the
system more efficiently, e.g. by tuning memory cgroup limits
appropriately, or for job placement within a compute cluster.
==== USE CASES ====
The unified cgroup hierarchy has memory.low and memory.high knobs, which
are defined as the low and high boundaries for the workload working set
size. However, the working set size of a workload may be unknown or
change in time. With this patch set, one can periodically estimate the
amount of memory unused by each cgroup and tune their memory.low and
memory.high parameters accordingly, therefore optimizing the overall
memory utilization.
Another use case is balancing workloads within a compute cluster. Knowing
how much memory is not really used by a workload unit may help take a more
optimal decision when considering migrating the unit to another node
within the cluster.
Also, as noted by Minchan, this would be useful for per-process reclaim
(https://lwn.net/Articles/545668/). With idle tracking, we could reclaim idle
pages only by smart user memory manager.
==== USER API ====
The user API consists of two new files:
* /sys/kernel/mm/page_idle/bitmap. This file implements a bitmap where each
bit corresponds to a page, indexed by PFN. When the bit is set, the
corresponding page is idle. A page is considered idle if it has not been
accessed since it was marked idle. To mark a page idle one should set the
bit corresponding to the page by writing to the file. A value written to the
file is OR-ed with the current bitmap value. Only user memory pages can be
marked idle, for other page types input is silently ignored. Writing to this
file beyond max PFN results in the ENXIO error. Only available when
CONFIG_IDLE_PAGE_TRACKING is set.
This file can be used to estimate the amount of pages that are not
used by a particular workload as follows:
1. mark all pages of interest idle by setting corresponding bits in the
/sys/kernel/mm/page_idle/bitmap
2. wait until the workload accesses its working set
3. read /sys/kernel/mm/page_idle/bitmap and count the number of bits set
* /proc/kpagecgroup. This file contains a 64-bit inode number of the
memory cgroup each page is charged to, indexed by PFN. Only available when
CONFIG_MEMCG is set.
This file can be used to find all pages (including unmapped file pages)
accounted to a particular cgroup. Using /sys/kernel/mm/page_idle/bitmap, one
can then estimate the cgroup working set size.
For an example of using these files for estimating the amount of unused
memory pages per each memory cgroup, please see the script attached
below.
==== REASONING ====
The reason to introduce the new user API instead of using
/proc/PID/{clear_refs,smaps} is that the latter has two serious
drawbacks:
- it does not count unmapped file pages
- it affects the reclaimer logic
The new API attempts to overcome them both. For more details on how it
is achieved, please see the comment to patch 6.
==== PATCHSET STRUCTURE ====
The patch set is organized as follows:
- patch 1 adds page_cgroup_ino() helper for the sake of
/proc/kpagecgroup and patches 2-3 do related cleanup
- patch 4 adds /proc/kpagecgroup, which reports cgroup ino each page is
charged to
- patch 5 introduces a new mmu notifier callback, clear_young, which is
a lightweight version of clear_flush_young; it is used in patch 6
- patch 6 implements the idle page tracking feature, including the
userspace API, /sys/kernel/mm/page_idle/bitmap
- patch 7 exports idle flag via /proc/kpageflags
==== SIMILAR WORKS ====
Originally, the patch for tracking idle memory was proposed back in 2011
by Michel Lespinasse (see http://lwn.net/Articles/459269/). The main
difference between Michel's patch and this one is that Michel implemented
a kernel space daemon for estimating idle memory size per cgroup while
this patch only provides the userspace with the minimal API for doing the
job, leaving the rest up to the userspace. However, they both share the
same idea of Idle/Young page flags to avoid affecting the reclaimer logic.
==== PERFORMANCE EVALUATION ====
SPECjvm2008 (https://www.spec.org/jvm2008/) was used to evaluate the
performance impact introduced by this patch set. Three runs were carried
out:
- base: kernel without the patch
- patched: patched kernel, the feature is not used
- patched-active: patched kernel, 1 minute-period daemon is used for
tracking idle memory
For tracking idle memory, idlememstat utility was used:
https://github.com/locker/idlememstat
testcase base patched patched-active
compiler 537.40 ( 0.00)% 532.26 (-0.96)% 538.31 ( 0.17)%
compress 305.47 ( 0.00)% 301.08 (-1.44)% 300.71 (-1.56)%
crypto 284.32 ( 0.00)% 282.21 (-0.74)% 284.87 ( 0.19)%
derby 411.05 ( 0.00)% 413.44 ( 0.58)% 412.07 ( 0.25)%
mpegaudio 189.96 ( 0.00)% 190.87 ( 0.48)% 189.42 (-0.28)%
scimark.large 46.85 ( 0.00)% 46.41 (-0.94)% 47.83 ( 2.09)%
scimark.small 412.91 ( 0.00)% 415.41 ( 0.61)% 421.17 ( 2.00)%
serial 204.23 ( 0.00)% 213.46 ( 4.52)% 203.17 (-0.52)%
startup 36.76 ( 0.00)% 35.49 (-3.45)% 35.64 (-3.05)%
sunflow 115.34 ( 0.00)% 115.08 (-0.23)% 117.37 ( 1.76)%
xml 620.55 ( 0.00)% 619.95 (-0.10)% 620.39 (-0.03)%
composite 211.50 ( 0.00)% 211.15 (-0.17)% 211.67 ( 0.08)%
time idlememstat:
17.20user 65.16system 2:15:23elapsed 1%CPU (0avgtext+0avgdata 8476maxresident)k
448inputs+40outputs (1major+36052minor)pagefaults 0swaps
==== SCRIPT FOR COUNTING IDLE PAGES PER CGROUP ====
#! /usr/bin/python
#
import os
import stat
import errno
import struct
CGROUP_MOUNT = "/sys/fs/cgroup/memory"
BUFSIZE = 8 * 1024 # must be multiple of 8
def get_hugepage_size():
with open("/proc/meminfo", "r") as f:
for s in f:
k, v = s.split(":")
if k == "Hugepagesize":
return int(v.split()[0]) * 1024
PAGE_SIZE = os.sysconf("SC_PAGE_SIZE")
HUGEPAGE_SIZE = get_hugepage_size()
def set_idle():
f = open("/sys/kernel/mm/page_idle/bitmap", "wb", BUFSIZE)
while True:
try:
f.write(struct.pack("Q", pow(2, 64) - 1))
except IOError as err:
if err.errno == errno.ENXIO:
break
raise
f.close()
def count_idle():
f_flags = open("/proc/kpageflags", "rb", BUFSIZE)
f_cgroup = open("/proc/kpagecgroup", "rb", BUFSIZE)
with open("/sys/kernel/mm/page_idle/bitmap", "rb", BUFSIZE) as f:
while f.read(BUFSIZE): pass # update idle flag
idlememsz = {}
while True:
s1, s2 = f_flags.read(8), f_cgroup.read(8)
if not s1 or not s2:
break
flags, = struct.unpack('Q', s1)
cgino, = struct.unpack('Q', s2)
unevictable = (flags >> 18) & 1
huge = (flags >> 22) & 1
idle = (flags >> 25) & 1
if idle and not unevictable:
idlememsz[cgino] = idlememsz.get(cgino, 0) + \
(HUGEPAGE_SIZE if huge else PAGE_SIZE)
f_flags.close()
f_cgroup.close()
return idlememsz
if __name__ == "__main__":
print "Setting the idle flag for each page..."
set_idle()
raw_input("Wait until the workload accesses its working set, "
"then press Enter")
print "Counting idle pages..."
idlememsz = count_idle()
for dir, subdirs, files in os.walk(CGROUP_MOUNT):
ino = os.stat(dir)[stat.ST_INO]
print dir + ": " + str(idlememsz.get(ino, 0) / 1024) + " kB"
==== END SCRIPT ====
This patch (of 8):
Add page_cgroup_ino() helper to memcg.
This function returns the inode number of the closest online ancestor of
the memory cgroup a page is charged to. It is required for exporting
information about which page is charged to which cgroup to userspace,
which will be introduced by a following patch.
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Reviewed-by: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Greg Thelen <gthelen@google.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-09 15:35:28 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-09-04 14:22:58 +08:00
|
|
|
|
memcg: add page_cgroup_ino helper
This patchset introduces a new user API for tracking user memory pages
that have not been used for a given period of time. The purpose of this
is to provide the userspace with the means of tracking a workload's
working set, i.e. the set of pages that are actively used by the
workload. Knowing the working set size can be useful for partitioning the
system more efficiently, e.g. by tuning memory cgroup limits
appropriately, or for job placement within a compute cluster.
==== USE CASES ====
The unified cgroup hierarchy has memory.low and memory.high knobs, which
are defined as the low and high boundaries for the workload working set
size. However, the working set size of a workload may be unknown or
change in time. With this patch set, one can periodically estimate the
amount of memory unused by each cgroup and tune their memory.low and
memory.high parameters accordingly, therefore optimizing the overall
memory utilization.
Another use case is balancing workloads within a compute cluster. Knowing
how much memory is not really used by a workload unit may help take a more
optimal decision when considering migrating the unit to another node
within the cluster.
Also, as noted by Minchan, this would be useful for per-process reclaim
(https://lwn.net/Articles/545668/). With idle tracking, we could reclaim idle
pages only by smart user memory manager.
==== USER API ====
The user API consists of two new files:
* /sys/kernel/mm/page_idle/bitmap. This file implements a bitmap where each
bit corresponds to a page, indexed by PFN. When the bit is set, the
corresponding page is idle. A page is considered idle if it has not been
accessed since it was marked idle. To mark a page idle one should set the
bit corresponding to the page by writing to the file. A value written to the
file is OR-ed with the current bitmap value. Only user memory pages can be
marked idle, for other page types input is silently ignored. Writing to this
file beyond max PFN results in the ENXIO error. Only available when
CONFIG_IDLE_PAGE_TRACKING is set.
This file can be used to estimate the amount of pages that are not
used by a particular workload as follows:
1. mark all pages of interest idle by setting corresponding bits in the
/sys/kernel/mm/page_idle/bitmap
2. wait until the workload accesses its working set
3. read /sys/kernel/mm/page_idle/bitmap and count the number of bits set
* /proc/kpagecgroup. This file contains a 64-bit inode number of the
memory cgroup each page is charged to, indexed by PFN. Only available when
CONFIG_MEMCG is set.
This file can be used to find all pages (including unmapped file pages)
accounted to a particular cgroup. Using /sys/kernel/mm/page_idle/bitmap, one
can then estimate the cgroup working set size.
For an example of using these files for estimating the amount of unused
memory pages per each memory cgroup, please see the script attached
below.
==== REASONING ====
The reason to introduce the new user API instead of using
/proc/PID/{clear_refs,smaps} is that the latter has two serious
drawbacks:
- it does not count unmapped file pages
- it affects the reclaimer logic
The new API attempts to overcome them both. For more details on how it
is achieved, please see the comment to patch 6.
==== PATCHSET STRUCTURE ====
The patch set is organized as follows:
- patch 1 adds page_cgroup_ino() helper for the sake of
/proc/kpagecgroup and patches 2-3 do related cleanup
- patch 4 adds /proc/kpagecgroup, which reports cgroup ino each page is
charged to
- patch 5 introduces a new mmu notifier callback, clear_young, which is
a lightweight version of clear_flush_young; it is used in patch 6
- patch 6 implements the idle page tracking feature, including the
userspace API, /sys/kernel/mm/page_idle/bitmap
- patch 7 exports idle flag via /proc/kpageflags
==== SIMILAR WORKS ====
Originally, the patch for tracking idle memory was proposed back in 2011
by Michel Lespinasse (see http://lwn.net/Articles/459269/). The main
difference between Michel's patch and this one is that Michel implemented
a kernel space daemon for estimating idle memory size per cgroup while
this patch only provides the userspace with the minimal API for doing the
job, leaving the rest up to the userspace. However, they both share the
same idea of Idle/Young page flags to avoid affecting the reclaimer logic.
==== PERFORMANCE EVALUATION ====
SPECjvm2008 (https://www.spec.org/jvm2008/) was used to evaluate the
performance impact introduced by this patch set. Three runs were carried
out:
- base: kernel without the patch
- patched: patched kernel, the feature is not used
- patched-active: patched kernel, 1 minute-period daemon is used for
tracking idle memory
For tracking idle memory, idlememstat utility was used:
https://github.com/locker/idlememstat
testcase base patched patched-active
compiler 537.40 ( 0.00)% 532.26 (-0.96)% 538.31 ( 0.17)%
compress 305.47 ( 0.00)% 301.08 (-1.44)% 300.71 (-1.56)%
crypto 284.32 ( 0.00)% 282.21 (-0.74)% 284.87 ( 0.19)%
derby 411.05 ( 0.00)% 413.44 ( 0.58)% 412.07 ( 0.25)%
mpegaudio 189.96 ( 0.00)% 190.87 ( 0.48)% 189.42 (-0.28)%
scimark.large 46.85 ( 0.00)% 46.41 (-0.94)% 47.83 ( 2.09)%
scimark.small 412.91 ( 0.00)% 415.41 ( 0.61)% 421.17 ( 2.00)%
serial 204.23 ( 0.00)% 213.46 ( 4.52)% 203.17 (-0.52)%
startup 36.76 ( 0.00)% 35.49 (-3.45)% 35.64 (-3.05)%
sunflow 115.34 ( 0.00)% 115.08 (-0.23)% 117.37 ( 1.76)%
xml 620.55 ( 0.00)% 619.95 (-0.10)% 620.39 (-0.03)%
composite 211.50 ( 0.00)% 211.15 (-0.17)% 211.67 ( 0.08)%
time idlememstat:
17.20user 65.16system 2:15:23elapsed 1%CPU (0avgtext+0avgdata 8476maxresident)k
448inputs+40outputs (1major+36052minor)pagefaults 0swaps
==== SCRIPT FOR COUNTING IDLE PAGES PER CGROUP ====
#! /usr/bin/python
#
import os
import stat
import errno
import struct
CGROUP_MOUNT = "/sys/fs/cgroup/memory"
BUFSIZE = 8 * 1024 # must be multiple of 8
def get_hugepage_size():
with open("/proc/meminfo", "r") as f:
for s in f:
k, v = s.split(":")
if k == "Hugepagesize":
return int(v.split()[0]) * 1024
PAGE_SIZE = os.sysconf("SC_PAGE_SIZE")
HUGEPAGE_SIZE = get_hugepage_size()
def set_idle():
f = open("/sys/kernel/mm/page_idle/bitmap", "wb", BUFSIZE)
while True:
try:
f.write(struct.pack("Q", pow(2, 64) - 1))
except IOError as err:
if err.errno == errno.ENXIO:
break
raise
f.close()
def count_idle():
f_flags = open("/proc/kpageflags", "rb", BUFSIZE)
f_cgroup = open("/proc/kpagecgroup", "rb", BUFSIZE)
with open("/sys/kernel/mm/page_idle/bitmap", "rb", BUFSIZE) as f:
while f.read(BUFSIZE): pass # update idle flag
idlememsz = {}
while True:
s1, s2 = f_flags.read(8), f_cgroup.read(8)
if not s1 or not s2:
break
flags, = struct.unpack('Q', s1)
cgino, = struct.unpack('Q', s2)
unevictable = (flags >> 18) & 1
huge = (flags >> 22) & 1
idle = (flags >> 25) & 1
if idle and not unevictable:
idlememsz[cgino] = idlememsz.get(cgino, 0) + \
(HUGEPAGE_SIZE if huge else PAGE_SIZE)
f_flags.close()
f_cgroup.close()
return idlememsz
if __name__ == "__main__":
print "Setting the idle flag for each page..."
set_idle()
raw_input("Wait until the workload accesses its working set, "
"then press Enter")
print "Counting idle pages..."
idlememsz = count_idle()
for dir, subdirs, files in os.walk(CGROUP_MOUNT):
ino = os.stat(dir)[stat.ST_INO]
print dir + ": " + str(idlememsz.get(ino, 0) / 1024) + " kB"
==== END SCRIPT ====
This patch (of 8):
Add page_cgroup_ino() helper to memcg.
This function returns the inode number of the closest online ancestor of
the memory cgroup a page is charged to. It is required for exporting
information about which page is charged to which cgroup to userspace,
which will be introduced by a following patch.
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Reviewed-by: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Greg Thelen <gthelen@google.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-09 15:35:28 -07:00
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm,memcg: provide per-cgroup counters for NUMA balancing operations
The ability to observe the demotion and promotion decisions made by the
kernel on a per-cgroup basis is important for monitoring and tuning
containerized workloads on machines equipped with tiered memory.
Different containers in the system may experience drastically different
memory tiering actions that cannot be distinguished from the global
counters alone.
For example, a container running a workload that has a much hotter memory
accesses will likely see more promotions and fewer demotions, potentially
depriving a colocated container of top tier memory to such an extent that
its performance degrades unacceptably.
For another example, some containers may exhibit longer periods between
data reuse, causing much more numa_hint_faults than numa_pages_migrated.
In this case, tuning hot_threshold_ms may be appropriate, but the signal
can easily be lost if only global counters are available.
In the long term, we hope to introduce per-cgroup control of promotion and
demotion actions to implement memory placement policies in tiering.
This patch set adds seven counters to memory.stat in a cgroup:
numa_pages_migrated, numa_pte_updates, numa_hint_faults, pgdemote_kswapd,
pgdemote_khugepaged, pgdemote_direct and pgpromote_success. pgdemote_*
and pgpromote_success are also available in memory.numa_stat.
count_memcg_events_mm() is added to count multiple event occurrences at
once, and get_mem_cgroup_from_folio() is added because we need to get a
reference to the memcg of a folio before it's migrated to track
numa_pages_migrated. The accounting of PGDEMOTE_* is moved to
shrink_inactive_list() before being changed to per-cgroup.
[kaiyang2@cs.cmu.edu: add documentation of the memcg counters in cgroup-v2.rst]
Link: https://lkml.kernel.org/r/20240814235122.252309-1-kaiyang2@cs.cmu.edu
Link: https://lkml.kernel.org/r/20240814174227.30639-1-kaiyang2@cs.cmu.edu
Signed-off-by: Kaiyang Zhao <kaiyang2@cs.cmu.edu>
Cc: David Rientjes <rientjes@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Wei Xu <weixugc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-08-14 17:42:27 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-03-18 15:58:32 +08:00
|
|
|
|
memcg/hugetlb: add hugeTLB counters to memcg
This patch introduces a new counter to memory.stat that tracks hugeTLB
usage, only if hugeTLB accounting is done to memory.current. This feature
is enabled the same way hugeTLB accounting is enabled, via the
memory_hugetlb_accounting mount flag for cgroupsv2.
1. Why is this patch necessary?
Currently, memcg hugeTLB accounting is an opt-in feature [1] that adds
hugeTLB usage to memory.current. However, the metric is not reported in
memory.stat. Given that users often interpret memory.stat as a breakdown
of the value reported in memory.current, the disparity between the two
reports can be confusing. This patch solves this problem by including the
metric in memory.stat as well, but only if it is also reported in
memory.current (it would also be confusing if the value was reported in
memory.stat, but not in memory.current)
Aside from the consistency between the two files, we also see benefits in
observability. Userspace might be interested in the hugeTLB footprint of
cgroups for many reasons. For instance, system admins might want to
verify that hugeTLB usage is distributed as expected across tasks: i.e.
memory-intensive tasks are using more hugeTLB pages than tasks that don't
consume a lot of memory, or are seen to fault frequently. Note that this
is separate from wanting to inspect the distribution for limiting purposes
(in which case, hugeTLB controller makes more sense).
2. We already have a hugeTLB controller. Why not use that?
It is true that hugeTLB tracks the exact value that we want. In fact, by
enabling the hugeTLB controller, we get all of the observability benefits
that I mentioned above, and users can check the total hugeTLB usage,
verify if it is distributed as expected, etc.
With this said, there are 2 problems:
(a) They are still not reported in memory.stat, which means the
disparity between the memcg reports are still there.
(b) We cannot reasonably expect users to enable the hugeTLB controller
just for the sake of hugeTLB usage reporting, especially since
they don't have any use for hugeTLB usage enforcing [2].
3. Implementation Details:
In the alloc / free hugetlb functions, we call lruvec_stat_mod_folio
regardless of whether memcg accounts hugetlb. mem_cgroup_commit_charge
which is called from alloc_hugetlb_folio will set memcg for the folio only
if the CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING cgroup mount option is used, so
lruvec_stat_mod_folio accounts per-memcg hugetlb counters only if the
feature is enabled. Regardless of whether memcg accounts for hugetlb, the
newly added global counter is updated and shown in /proc/vmstat.
The global counter is added because vmstats is the preferred framework for
cgroup stats. It makes stat items consistent between global and cgroups.
It also provides a per-node breakdown, which is useful. Because it does
not use cgroup-specific hooks, we also keep generic MM code separate from
memcg code.
[1] https://lore.kernel.org/all/20231006184629.155543-1-nphamcs@gmail.com/
[2] Of course, we can't make a new patch for every feature that can be
duplicated. However, since the existing solution of enabling the
hugeTLB controller is an imperfect solution that still leaves a
discrepancy between memory.stat and memory.curent, I think that it
is reasonable to isolate the feature in this case.
Link: https://lkml.kernel.org/r/20241101204402.1885383-1-joshua.hahnjy@gmail.com
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Nhat Pham <nphamcs@gmail.com>
Suggested-by: Shakeel Butt <shakeel.butt@linux.dev>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Shakeel Butt <shakeel.butt@linux.dev>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Chris Down <chris@chrisdown.name>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Roman Gushchin <roman.gushchin@linux.dev>
Reviewed-by: Nhat Pham <nphamcs@gmail.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Koutný <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-11-01 13:44:02 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-07-24 20:33:20 +00:00
|
|
|
|
|
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-07-24 20:33:20 +00:00
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
2024-07-24 20:33:20 +00:00
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
2024-07-24 20:33:20 +00:00
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
2024-07-24 20:33:20 +00:00
|
|
|
|
|
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
2024-07-24 20:33:20 +00:00
|
|
|
|
|
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-07-24 20:33:20 +00:00
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
|
|
|
|
|
2024-05-01 10:26:11 -07:00
|
|
|
|
|
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
2024-05-01 10:26:11 -07:00
|
|
|
|
|
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
2024-05-01 10:26:11 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
2024-05-01 10:26:11 -07:00
|
|
|
|
|
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
2024-05-01 10:26:11 -07:00
|
|
|
|
|
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
2024-05-01 10:26:11 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
2024-05-01 10:26:11 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
2024-07-24 20:33:20 +00:00
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
|
|
|
|
|
2024-05-01 10:26:11 -07:00
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
2024-05-01 10:26:11 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-05-01 10:26:16 -07:00
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
2024-05-01 10:26:11 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
2024-07-24 20:33:20 +00:00
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
|
|
|
|
|
2024-05-01 10:26:11 -07:00
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
2024-05-01 10:26:11 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-09-07 04:35:36 +00:00
|
|
|
|
|
|
|
|
|
2024-08-14 22:04:53 -07:00
|
|
|
|
2022-09-07 04:35:37 +00:00
|
|
|
|
|
|
|
|
|
2024-08-14 22:04:53 -07:00
|
|
|
|
2024-08-30 16:22:44 +08:00
|
|
|
|
|
|
|
|
|
2022-09-07 04:35:36 +00:00
|
|
|
|
|
|
|
|
|
2022-10-26 14:01:33 -04:00
|
|
|
|
2025-03-18 15:58:32 +08:00
|
|
|
|
2022-09-07 04:35:36 +00:00
|
|
|
|
|
|
|
|
|
2022-10-26 14:01:33 -04:00
|
|
|
|
2025-03-18 15:58:32 +08:00
|
|
|
|
2022-09-07 04:35:36 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: count zeromap read and set for swapout and swapin
When the proportion of folios from the zeromap is small, missing their
accounting may not significantly impact profiling. However, it's easy to
construct a scenario where this becomes an issue—for example, allocating
1 GB of memory, writing zeros from userspace, followed by MADV_PAGEOUT,
and then swapping it back in. In this case, the swap-out and swap-in
counts seem to vanish into a black hole, potentially causing semantic
ambiguity.
On the other hand, Usama reported that zero-filled pages can exceed 10% in
workloads utilizing zswap, while Hailong noted that some app in Android
have more than 6% zero-filled pages. Before commit 0ca0c24e3211 ("mm:
store zero pages to be swapped out in a bitmap"), both zswap and zRAM
implemented similar optimizations, leading to these optimized-out pages
being counted in either zswap or zRAM counters (with pswpin/pswpout also
increasing for zRAM). With zeromap functioning prior to both zswap and
zRAM, userspace will no longer detect these swap-out and swap-in actions.
We have three ways to address this:
1. Introduce a dedicated counter specifically for the zeromap.
2. Use pswpin/pswpout accounting, treating the zero map as a standard
backend. This approach aligns with zRAM's current handling of
same-page fills at the device level. However, it would mean losing the
optimized-out page counters previously available in zRAM and would not
align with systems using zswap. Additionally, as noted by Nhat Pham,
pswpin/pswpout counters apply only to I/O done directly to the backend
device.
3. Count zeromap pages under zswap, aligning with system behavior when
zswap is enabled. However, this would not be consistent with zRAM, nor
would it align with systems lacking both zswap and zRAM.
Given the complications with options 2 and 3, this patch selects
option 1.
We can find these counters from /proc/vmstat (counters for the whole
system) and memcg's memory.stat (counters for the interested memcg).
For example:
$ grep -E 'swpin_zero|swpout_zero' /proc/vmstat
swpin_zero 1648
swpout_zero 33536
$ grep -E 'swpin_zero|swpout_zero' /sys/fs/cgroup/system.slice/memory.stat
swpin_zero 3905
swpout_zero 3985
This patch does not address any specific zeromap bug, but the missing
swpout and swpin counts for zero-filled pages can be highly confusing and
may mislead user-space agents that rely on changes in these counters as
indicators. Therefore, we add a Fixes tag to encourage the inclusion of
this counter in any kernel versions with zeromap.
Many thanks to Kanchana for the contribution of changing
count_objcg_event() to count_objcg_events() to support large folios[1],
which has now been incorporated into this patch.
[1] https://lkml.kernel.org/r/20241001053222.6944-5-kanchana.p.sridhar@intel.com
Link: https://lkml.kernel.org/r/20241107011246.59137-1-21cnbao@gmail.com
Fixes: 0ca0c24e3211 ("mm: store zero pages to be swapped out in a bitmap")
Co-developed-by: Kanchana P Sridhar <kanchana.p.sridhar@intel.com>
Signed-off-by: Barry Song <v-songbaohua@oppo.com>
Reviewed-by: Nhat Pham <nphamcs@gmail.com>
Reviewed-by: Chengming Zhou <chengming.zhou@linux.dev>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Usama Arif <usamaarif642@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Hailong Liu <hailong.liu@oppo.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Kairui Song <kasong@tencent.com>
Cc: Ryan Roberts <ryan.roberts@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-11-07 14:12:46 +13:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-07-01 11:31:15 -04:00
|
|
|
|
2022-09-07 04:35:36 +00:00
|
|
|
|
|
|
|
|
|
2023-11-29 03:21:50 +00:00
|
|
|
|
2022-09-07 04:35:36 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcg: add THP swap out info for anonymous reclaim
At present, we support per-memcg reclaim strategy, however we do not know
the number of transparent huge pages being reclaimed, as we know the
transparent huge pages need to be splited before reclaim them, and they
will bring some performance bottleneck effect. for example, when two
memcg (A & B) are doing reclaim for anonymous pages at same time, and 'A'
memcg is reclaiming a large number of transparent huge pages, we can
better analyze that the performance bottleneck will be caused by 'A'
memcg. therefore, in order to better analyze such problems, there add THP
swap out info for per-memcg.
[akpm@linux-foundation.orgL fix swap_writepage_fs(), per Johannes]
Link: https://lkml.kernel.org/r/20230913213343.GB48476@cmpxchg.org
Link: https://lkml.kernel.org/r/20230913164938.16918-1-vernhao@tencent.com
Signed-off-by: Xin Hao <vernhao@tencent.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Muchun Song <songmuchun@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-09-14 00:49:37 +08:00
|
|
|
|
|
|
|
|
|
2022-09-07 04:35:36 +00:00
|
|
|
|
mm,memcg: provide per-cgroup counters for NUMA balancing operations
The ability to observe the demotion and promotion decisions made by the
kernel on a per-cgroup basis is important for monitoring and tuning
containerized workloads on machines equipped with tiered memory.
Different containers in the system may experience drastically different
memory tiering actions that cannot be distinguished from the global
counters alone.
For example, a container running a workload that has a much hotter memory
accesses will likely see more promotions and fewer demotions, potentially
depriving a colocated container of top tier memory to such an extent that
its performance degrades unacceptably.
For another example, some containers may exhibit longer periods between
data reuse, causing much more numa_hint_faults than numa_pages_migrated.
In this case, tuning hot_threshold_ms may be appropriate, but the signal
can easily be lost if only global counters are available.
In the long term, we hope to introduce per-cgroup control of promotion and
demotion actions to implement memory placement policies in tiering.
This patch set adds seven counters to memory.stat in a cgroup:
numa_pages_migrated, numa_pte_updates, numa_hint_faults, pgdemote_kswapd,
pgdemote_khugepaged, pgdemote_direct and pgpromote_success. pgdemote_*
and pgpromote_success are also available in memory.numa_stat.
count_memcg_events_mm() is added to count multiple event occurrences at
once, and get_mem_cgroup_from_folio() is added because we need to get a
reference to the memcg of a folio before it's migrated to track
numa_pages_migrated. The accounting of PGDEMOTE_* is moved to
shrink_inactive_list() before being changed to per-cgroup.
[kaiyang2@cs.cmu.edu: add documentation of the memcg counters in cgroup-v2.rst]
Link: https://lkml.kernel.org/r/20240814235122.252309-1-kaiyang2@cs.cmu.edu
Link: https://lkml.kernel.org/r/20240814174227.30639-1-kaiyang2@cs.cmu.edu
Signed-off-by: Kaiyang Zhao <kaiyang2@cs.cmu.edu>
Cc: David Rientjes <rientjes@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Wei Xu <weixugc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-08-14 17:42:27 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-09-07 04:35:36 +00:00
|
|
|
|
|
|
|
|
|
2022-09-07 04:35:37 +00:00
|
|
|
|
2024-07-24 20:33:20 +00:00
|
|
|
|
2022-09-07 04:35:37 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-07-24 20:33:20 +00:00
|
|
|
|
2024-05-01 10:26:10 -07:00
|
|
|
|
2024-07-24 20:33:20 +00:00
|
|
|
|
2024-05-01 10:26:10 -07:00
|
|
|
|
2024-07-24 20:33:20 +00:00
|
|
|
|
|
|
|
|
|
2022-09-07 04:35:37 +00:00
|
|
|
|
|
|
|
|
|
2024-07-24 20:33:20 +00:00
|
|
|
|
2022-09-07 04:35:37 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-07-24 20:33:20 +00:00
|
|
|
|
2022-09-07 04:35:37 +00:00
|
|
|
|
|
|
|
|
|
2022-09-07 04:35:35 +00:00
|
|
|
|
mm: memcg: optimize parent iteration in memcg_rstat_updated()
In memcg_rstat_updated(), we iterate the memcg being updated and its
parents to update memcg->vmstats_percpu->stats_updates in the fast path
(i.e. no atomic updates). According to my math, this is 3 memory loads
(and potentially 3 cache misses) per memcg:
- Load the address of memcg->vmstats_percpu.
- Load vmstats_percpu->stats_updates (based on some percpu calculation).
- Load the address of the parent memcg.
Avoid most of the cache misses by caching a pointer from each struct
memcg_vmstats_percpu to its parent on the corresponding CPU. In this
case, for the first memcg we have 2 memory loads (same as above):
- Load the address of memcg->vmstats_percpu.
- Load vmstats_percpu->stats_updates (based on some percpu calculation).
Then for each additional memcg, we need a single load to get the
parent's stats_updates directly. This reduces the number of loads from
O(3N) to O(2+N) -- where N is the number of memcgs we need to iterate.
Additionally, stash a pointer to memcg->vmstats in each struct
memcg_vmstats_percpu such that we can access the atomic counter that all
CPUs fold into, memcg->vmstats->stats_updates.
memcg_should_flush_stats() is changed to memcg_vmstats_needs_flush() to
accept a struct memcg_vmstats pointer accordingly.
In struct memcg_vmstats_percpu, make sure both pointers together with
stats_updates live on the same cacheline. Finally, update
mem_cgroup_alloc() to take in a parent pointer and initialize the new
cache pointers on each CPU. The percpu loop in mem_cgroup_alloc() may
look concerning, but there are multiple similar loops in the cgroup
creation path (e.g. cgroup_rstat_init()), most of which are hidden
within alloc_percpu().
According to Oliver's testing [1], this fixes multiple 30-38%
regressions in vm-scalability, will-it-scale-tlb_flush2, and
will-it-scale-fallocate1. This comes at a cost of 2 more pointers per
CPU (<2KB on a machine with 128 CPUs).
[1] https://lore.kernel.org/lkml/ZbDJsfsZt2ITyo61@xsang-OptiPlex-9020/
[yosryahmed@google.com: fix struct memcg_vmstats_percpu size and alignment]
Link: https://lkml.kernel.org/r/20240203044612.1234216-1-yosryahmed@google.com
Link: https://lkml.kernel.org/r/20240124100023.660032-1-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Fixes: 8d59d2214c23 ("mm: memcg: make stats flushing threshold per-memcg")
Tested-by: kernel test robot <oliver.sang@intel.com>
Reported-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202401221624.cb53a8ca-oliver.sang@intel.com
Acked-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-01-24 10:00:22 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-05-14 11:41:52 -07:00
|
|
|
|
|
|
|
|
|
mm: memcg: optimize parent iteration in memcg_rstat_updated()
In memcg_rstat_updated(), we iterate the memcg being updated and its
parents to update memcg->vmstats_percpu->stats_updates in the fast path
(i.e. no atomic updates). According to my math, this is 3 memory loads
(and potentially 3 cache misses) per memcg:
- Load the address of memcg->vmstats_percpu.
- Load vmstats_percpu->stats_updates (based on some percpu calculation).
- Load the address of the parent memcg.
Avoid most of the cache misses by caching a pointer from each struct
memcg_vmstats_percpu to its parent on the corresponding CPU. In this
case, for the first memcg we have 2 memory loads (same as above):
- Load the address of memcg->vmstats_percpu.
- Load vmstats_percpu->stats_updates (based on some percpu calculation).
Then for each additional memcg, we need a single load to get the
parent's stats_updates directly. This reduces the number of loads from
O(3N) to O(2+N) -- where N is the number of memcgs we need to iterate.
Additionally, stash a pointer to memcg->vmstats in each struct
memcg_vmstats_percpu such that we can access the atomic counter that all
CPUs fold into, memcg->vmstats->stats_updates.
memcg_should_flush_stats() is changed to memcg_vmstats_needs_flush() to
accept a struct memcg_vmstats pointer accordingly.
In struct memcg_vmstats_percpu, make sure both pointers together with
stats_updates live on the same cacheline. Finally, update
mem_cgroup_alloc() to take in a parent pointer and initialize the new
cache pointers on each CPU. The percpu loop in mem_cgroup_alloc() may
look concerning, but there are multiple similar loops in the cgroup
creation path (e.g. cgroup_rstat_init()), most of which are hidden
within alloc_percpu().
According to Oliver's testing [1], this fixes multiple 30-38%
regressions in vm-scalability, will-it-scale-tlb_flush2, and
will-it-scale-fallocate1. This comes at a cost of 2 more pointers per
CPU (<2KB on a machine with 128 CPUs).
[1] https://lore.kernel.org/lkml/ZbDJsfsZt2ITyo61@xsang-OptiPlex-9020/
[yosryahmed@google.com: fix struct memcg_vmstats_percpu size and alignment]
Link: https://lkml.kernel.org/r/20240203044612.1234216-1-yosryahmed@google.com
Link: https://lkml.kernel.org/r/20240124100023.660032-1-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Fixes: 8d59d2214c23 ("mm: memcg: make stats flushing threshold per-memcg")
Tested-by: kernel test robot <oliver.sang@intel.com>
Reported-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202401221624.cb53a8ca-oliver.sang@intel.com
Acked-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-01-24 10:00:22 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-09-07 04:35:35 +00:00
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
2022-09-07 04:35:37 +00:00
|
|
|
|
2022-09-07 04:35:35 +00:00
|
|
|
|
|
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
2022-09-07 04:35:37 +00:00
|
|
|
|
mm: memcg: optimize parent iteration in memcg_rstat_updated()
In memcg_rstat_updated(), we iterate the memcg being updated and its
parents to update memcg->vmstats_percpu->stats_updates in the fast path
(i.e. no atomic updates). According to my math, this is 3 memory loads
(and potentially 3 cache misses) per memcg:
- Load the address of memcg->vmstats_percpu.
- Load vmstats_percpu->stats_updates (based on some percpu calculation).
- Load the address of the parent memcg.
Avoid most of the cache misses by caching a pointer from each struct
memcg_vmstats_percpu to its parent on the corresponding CPU. In this
case, for the first memcg we have 2 memory loads (same as above):
- Load the address of memcg->vmstats_percpu.
- Load vmstats_percpu->stats_updates (based on some percpu calculation).
Then for each additional memcg, we need a single load to get the
parent's stats_updates directly. This reduces the number of loads from
O(3N) to O(2+N) -- where N is the number of memcgs we need to iterate.
Additionally, stash a pointer to memcg->vmstats in each struct
memcg_vmstats_percpu such that we can access the atomic counter that all
CPUs fold into, memcg->vmstats->stats_updates.
memcg_should_flush_stats() is changed to memcg_vmstats_needs_flush() to
accept a struct memcg_vmstats pointer accordingly.
In struct memcg_vmstats_percpu, make sure both pointers together with
stats_updates live on the same cacheline. Finally, update
mem_cgroup_alloc() to take in a parent pointer and initialize the new
cache pointers on each CPU. The percpu loop in mem_cgroup_alloc() may
look concerning, but there are multiple similar loops in the cgroup
creation path (e.g. cgroup_rstat_init()), most of which are hidden
within alloc_percpu().
According to Oliver's testing [1], this fixes multiple 30-38%
regressions in vm-scalability, will-it-scale-tlb_flush2, and
will-it-scale-fallocate1. This comes at a cost of 2 more pointers per
CPU (<2KB on a machine with 128 CPUs).
[1] https://lore.kernel.org/lkml/ZbDJsfsZt2ITyo61@xsang-OptiPlex-9020/
[yosryahmed@google.com: fix struct memcg_vmstats_percpu size and alignment]
Link: https://lkml.kernel.org/r/20240203044612.1234216-1-yosryahmed@google.com
Link: https://lkml.kernel.org/r/20240124100023.660032-1-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Fixes: 8d59d2214c23 ("mm: memcg: make stats flushing threshold per-memcg")
Tested-by: kernel test robot <oliver.sang@intel.com>
Reported-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202401221624.cb53a8ca-oliver.sang@intel.com
Acked-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-01-24 10:00:22 +00:00
|
|
|
|
2022-09-07 04:35:35 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
2022-09-07 04:35:37 +00:00
|
|
|
|
2022-09-07 04:35:35 +00:00
|
|
|
|
mm: memcg: use rstat for non-hierarchical stats
Currently, memcg uses rstat to maintain aggregated hierarchical stats.
Counters are maintained for hierarchical stats at each memcg. Rstat
tracks which cgroups have updates on which cpus to keep those counters
fresh on the read-side.
Non-hierarchical stats are currently not covered by rstat. Their per-cpu
counters are summed up on every read, which is expensive. The original
implementation did the same. At some point before rstat, non-hierarchical
aggregated counters were introduced by commit a983b5ebee57 ("mm:
memcontrol: fix excessive complexity in memory.stat reporting"). However,
those counters were updated on the performance critical write-side, which
caused regressions, so they were later removed by commit 815744d75152
("mm: memcontrol: don't batch updates of local VM stats and events"). See
[1] for more detailed history.
Kernel versions in between a983b5ebee57 & 815744d75152 (a year and a half)
enjoyed cheap reads of non-hierarchical stats, specifically on cgroup v1.
When moving to more recent kernels, a performance regression for reading
non-hierarchical stats is observed.
Now that we have rstat, we know exactly which percpu counters have updates
for each stat. We can maintain non-hierarchical counters again, making
reads much more efficient, without affecting the performance critical
write-side. Hence, add non-hierarchical (i.e local) counters for the
stats, and extend rstat flushing to keep those up-to-date.
A caveat is that we now need a stats flush before reading
local/non-hierarchical stats through {memcg/lruvec}_page_state_local() or
memcg_events_local(), where we previously only needed a flush to read
hierarchical stats. Most contexts reading non-hierarchical stats are
already doing a flush, add a flush to the only missing context in
count_shadow_nodes().
With this patch, reading memory.stat from 1000 memcgs is 3x faster on a
machine with 256 cpus on cgroup v1:
# for i in $(seq 1000); do mkdir /sys/fs/cgroup/memory/cg$i; done
# time cat /sys/fs/cgroup/memory/cg*/memory.stat > /dev/null
real 0m0.125s
user 0m0.005s
sys 0m0.120s
After:
real 0m0.032s
user 0m0.005s
sys 0m0.027s
To make sure there are no regressions on cgroup v2, I ran an artificial
reclaim/refault stress test [2] that creates (NR_CPUS * 2) cgroups,
assigns them limits, runs a worker process in each cgroup that allocates
tmpfs memory equal to quadruple the limit (to invoke reclaim
continuously), and then reads back the entire file (to invoke refaults).
All workers are run in parallel, and zram is used as a swapping backend.
Both reclaim and refault have conditional stats flushing. I ran this on a
machine with 112 cpus, once on mm-unstable, and once on mm-unstable with
this patch reverted.
(1) A few runs without this patch:
# time ./stress_reclaim_refault.sh
real 0m9.949s
user 0m0.496s
sys 14m44.974s
# time ./stress_reclaim_refault.sh
real 0m10.049s
user 0m0.486s
sys 14m55.791s
# time ./stress_reclaim_refault.sh
real 0m9.984s
user 0m0.481s
sys 14m53.841s
(2) A few runs with this patch:
# time ./stress_reclaim_refault.sh
real 0m9.885s
user 0m0.486s
sys 14m48.753s
# time ./stress_reclaim_refault.sh
real 0m9.903s
user 0m0.495s
sys 14m48.339s
# time ./stress_reclaim_refault.sh
real 0m9.861s
user 0m0.507s
sys 14m49.317s
No regressions are observed with this patch. There is actually a very
slight improvement. If I have to guess, maybe it's because we avoid
the percpu loop in count_shadow_nodes() when calling
lruvec_page_state_local(), but I could not prove this using perf, it's
probably in the noise.
[1] https://lore.kernel.org/lkml/20230725201811.GA1231514@cmpxchg.org/
[2] https://lore.kernel.org/lkml/CAJD7tkb17x=qwoO37uxyYXLEUVp15BQKR+Xfh7Sg9Hx-wTQ_=w@mail.gmail.com/
Link: https://lkml.kernel.org/r/20230803185046.1385770-1-yosryahmed@google.com
Link: https://lkml.kernel.org/r/20230726153223.821757-2-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Roman Gushchin <roman.gushchin@linux.dev>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Shakeel Butt <shakeelb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-07-26 15:32:23 +00:00
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
mm: memcg: use rstat for non-hierarchical stats
Currently, memcg uses rstat to maintain aggregated hierarchical stats.
Counters are maintained for hierarchical stats at each memcg. Rstat
tracks which cgroups have updates on which cpus to keep those counters
fresh on the read-side.
Non-hierarchical stats are currently not covered by rstat. Their per-cpu
counters are summed up on every read, which is expensive. The original
implementation did the same. At some point before rstat, non-hierarchical
aggregated counters were introduced by commit a983b5ebee57 ("mm:
memcontrol: fix excessive complexity in memory.stat reporting"). However,
those counters were updated on the performance critical write-side, which
caused regressions, so they were later removed by commit 815744d75152
("mm: memcontrol: don't batch updates of local VM stats and events"). See
[1] for more detailed history.
Kernel versions in between a983b5ebee57 & 815744d75152 (a year and a half)
enjoyed cheap reads of non-hierarchical stats, specifically on cgroup v1.
When moving to more recent kernels, a performance regression for reading
non-hierarchical stats is observed.
Now that we have rstat, we know exactly which percpu counters have updates
for each stat. We can maintain non-hierarchical counters again, making
reads much more efficient, without affecting the performance critical
write-side. Hence, add non-hierarchical (i.e local) counters for the
stats, and extend rstat flushing to keep those up-to-date.
A caveat is that we now need a stats flush before reading
local/non-hierarchical stats through {memcg/lruvec}_page_state_local() or
memcg_events_local(), where we previously only needed a flush to read
hierarchical stats. Most contexts reading non-hierarchical stats are
already doing a flush, add a flush to the only missing context in
count_shadow_nodes().
With this patch, reading memory.stat from 1000 memcgs is 3x faster on a
machine with 256 cpus on cgroup v1:
# for i in $(seq 1000); do mkdir /sys/fs/cgroup/memory/cg$i; done
# time cat /sys/fs/cgroup/memory/cg*/memory.stat > /dev/null
real 0m0.125s
user 0m0.005s
sys 0m0.120s
After:
real 0m0.032s
user 0m0.005s
sys 0m0.027s
To make sure there are no regressions on cgroup v2, I ran an artificial
reclaim/refault stress test [2] that creates (NR_CPUS * 2) cgroups,
assigns them limits, runs a worker process in each cgroup that allocates
tmpfs memory equal to quadruple the limit (to invoke reclaim
continuously), and then reads back the entire file (to invoke refaults).
All workers are run in parallel, and zram is used as a swapping backend.
Both reclaim and refault have conditional stats flushing. I ran this on a
machine with 112 cpus, once on mm-unstable, and once on mm-unstable with
this patch reverted.
(1) A few runs without this patch:
# time ./stress_reclaim_refault.sh
real 0m9.949s
user 0m0.496s
sys 14m44.974s
# time ./stress_reclaim_refault.sh
real 0m10.049s
user 0m0.486s
sys 14m55.791s
# time ./stress_reclaim_refault.sh
real 0m9.984s
user 0m0.481s
sys 14m53.841s
(2) A few runs with this patch:
# time ./stress_reclaim_refault.sh
real 0m9.885s
user 0m0.486s
sys 14m48.753s
# time ./stress_reclaim_refault.sh
real 0m9.903s
user 0m0.495s
sys 14m48.339s
# time ./stress_reclaim_refault.sh
real 0m9.861s
user 0m0.507s
sys 14m49.317s
No regressions are observed with this patch. There is actually a very
slight improvement. If I have to guess, maybe it's because we avoid
the percpu loop in count_shadow_nodes() when calling
lruvec_page_state_local(), but I could not prove this using perf, it's
probably in the noise.
[1] https://lore.kernel.org/lkml/20230725201811.GA1231514@cmpxchg.org/
[2] https://lore.kernel.org/lkml/CAJD7tkb17x=qwoO37uxyYXLEUVp15BQKR+Xfh7Sg9Hx-wTQ_=w@mail.gmail.com/
Link: https://lkml.kernel.org/r/20230803185046.1385770-1-yosryahmed@google.com
Link: https://lkml.kernel.org/r/20230726153223.821757-2-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Roman Gushchin <roman.gushchin@linux.dev>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Shakeel Butt <shakeelb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-07-26 15:32:23 +00:00
|
|
|
|
|
|
|
|
|
2022-09-07 04:35:35 +00:00
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
2022-09-07 04:35:37 +00:00
|
|
|
|
mm: memcg: make stats flushing threshold per-memcg
A global counter for the magnitude of memcg stats update is maintained on
the memcg side to avoid invoking rstat flushes when the pending updates
are not significant. This avoids unnecessary flushes, which are not very
cheap even if there isn't a lot of stats to flush. It also avoids
unnecessary lock contention on the underlying global rstat lock.
Make this threshold per-memcg. The scheme is followed where percpu (now
also per-memcg) counters are incremented in the update path, and only
propagated to per-memcg atomics when they exceed a certain threshold.
This provides two benefits: (a) On large machines with a lot of memcgs,
the global threshold can be reached relatively fast, so guarding the
underlying lock becomes less effective. Making the threshold per-memcg
avoids this.
(b) Having a global threshold makes it hard to do subtree flushes, as we
cannot reset the global counter except for a full flush. Per-memcg
counters removes this as a blocker from doing subtree flushes, which helps
avoid unnecessary work when the stats of a small subtree are needed.
Nothing is free, of course. This comes at a cost: (a) A new per-cpu
counter per memcg, consuming NR_CPUS * NR_MEMCGS * 4 bytes. The extra
memory usage is insigificant.
(b) More work on the update side, although in the common case it will only
be percpu counter updates. The amount of work scales with the number of
ancestors (i.e. tree depth). This is not a new concept, adding a cgroup
to the rstat tree involves a parent loop, so is charging. Testing results
below show no significant regressions.
(c) The error margin in the stats for the system as a whole increases from
NR_CPUS * MEMCG_CHARGE_BATCH to NR_CPUS * MEMCG_CHARGE_BATCH * NR_MEMCGS.
This is probably fine because we have a similar per-memcg error in charges
coming from percpu stocks, and we have a periodic flusher that makes sure
we always flush all the stats every 2s anyway.
This patch was tested to make sure no significant regressions are
introduced on the update path as follows. The following benchmarks were
ran in a cgroup that is 2 levels deep (/sys/fs/cgroup/a/b/):
(1) Running 22 instances of netperf on a 44 cpu machine with
hyperthreading disabled. All instances are run in a level 2 cgroup, as
well as netserver:
# netserver -6
# netperf -6 -H ::1 -l 60 -t TCP_SENDFILE -- -m 10K
Averaging 20 runs, the numbers are as follows:
Base: 40198.0 mbps
Patched: 38629.7 mbps (-3.9%)
The regression is minimal, especially for 22 instances in the same
cgroup sharing all ancestors (so updating the same atomics).
(2) will-it-scale page_fault tests. These tests (specifically
per_process_ops in page_fault3 test) detected a 25.9% regression before
for a change in the stats update path [1]. These are the
numbers from 10 runs (+ is good) on a machine with 256 cpus:
LABEL | MEAN | MEDIAN | STDDEV |
------------------------------+-------------+-------------+-------------
page_fault1_per_process_ops | | | |
(A) base | 270249.164 | 265437.000 | 13451.836 |
(B) patched | 261368.709 | 255725.000 | 13394.767 |
| -3.29% | -3.66% | |
page_fault1_per_thread_ops | | | |
(A) base | 242111.345 | 239737.000 | 10026.031 |
(B) patched | 237057.109 | 235305.000 | 9769.687 |
| -2.09% | -1.85% | |
page_fault1_scalability | | |
(A) base | 0.034387 | 0.035168 | 0.0018283 |
(B) patched | 0.033988 | 0.034573 | 0.0018056 |
| -1.16% | -1.69% | |
page_fault2_per_process_ops | | |
(A) base | 203561.836 | 203301.000 | 2550.764 |
(B) patched | 197195.945 | 197746.000 | 2264.263 |
| -3.13% | -2.73% | |
page_fault2_per_thread_ops | | |
(A) base | 171046.473 | 170776.000 | 1509.679 |
(B) patched | 166626.327 | 166406.000 | 768.753 |
| -2.58% | -2.56% | |
page_fault2_scalability | | |
(A) base | 0.054026 | 0.053821 | 0.00062121 |
(B) patched | 0.053329 | 0.05306 | 0.00048394 |
| -1.29% | -1.41% | |
page_fault3_per_process_ops | | |
(A) base | 1295807.782 | 1297550.000 | 5907.585 |
(B) patched | 1275579.873 | 1273359.000 | 8759.160 |
| -1.56% | -1.86% | |
page_fault3_per_thread_ops | | |
(A) base | 391234.164 | 390860.000 | 1760.720 |
(B) patched | 377231.273 | 376369.000 | 1874.971 |
| -3.58% | -3.71% | |
page_fault3_scalability | | |
(A) base | 0.60369 | 0.60072 | 0.0083029 |
(B) patched | 0.61733 | 0.61544 | 0.009855 |
| +2.26% | +2.45% | |
All regressions seem to be minimal, and within the normal variance for the
benchmark. The fix for [1] assumes that 3% is noise -- and there were no
further practical complaints), so hopefully this means that such
variations in these microbenchmarks do not reflect on practical workloads.
(3) I also ran stress-ng in a nested cgroup and did not observe any
obvious regressions.
[1]https://lore.kernel.org/all/20190520063534.GB19312@shao2-debian/
Link: https://lkml.kernel.org/r/20231129032154.3710765-4-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Tested-by: Domenico Cerasuolo <cerasuolodomenico@gmail.com>
Acked-by: Shakeel Butt <shakeelb@google.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: Greg Thelen <gthelen@google.com>
Cc: Ivan Babrou <ivan@cloudflare.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Michal Koutny <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Tejun Heo <tj@kernel.org>
Cc: Waiman Long <longman@redhat.com>
Cc: Wei Xu <weixugc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-11-29 03:21:51 +00:00
|
|
|
|
|
|
|
|
|
2025-05-18 23:31:42 -07:00
|
|
|
|
2022-09-07 04:35:35 +00:00
|
|
|
|
|
|
|
|
|
2021-11-05 13:37:31 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-11-29 03:21:49 +00:00
|
|
|
|
2022-04-21 16:35:40 -07:00
|
|
|
|
|
|
|
|
|
2021-11-05 13:37:31 -07:00
|
|
|
|
mm: memcg: optimize parent iteration in memcg_rstat_updated()
In memcg_rstat_updated(), we iterate the memcg being updated and its
parents to update memcg->vmstats_percpu->stats_updates in the fast path
(i.e. no atomic updates). According to my math, this is 3 memory loads
(and potentially 3 cache misses) per memcg:
- Load the address of memcg->vmstats_percpu.
- Load vmstats_percpu->stats_updates (based on some percpu calculation).
- Load the address of the parent memcg.
Avoid most of the cache misses by caching a pointer from each struct
memcg_vmstats_percpu to its parent on the corresponding CPU. In this
case, for the first memcg we have 2 memory loads (same as above):
- Load the address of memcg->vmstats_percpu.
- Load vmstats_percpu->stats_updates (based on some percpu calculation).
Then for each additional memcg, we need a single load to get the
parent's stats_updates directly. This reduces the number of loads from
O(3N) to O(2+N) -- where N is the number of memcgs we need to iterate.
Additionally, stash a pointer to memcg->vmstats in each struct
memcg_vmstats_percpu such that we can access the atomic counter that all
CPUs fold into, memcg->vmstats->stats_updates.
memcg_should_flush_stats() is changed to memcg_vmstats_needs_flush() to
accept a struct memcg_vmstats pointer accordingly.
In struct memcg_vmstats_percpu, make sure both pointers together with
stats_updates live on the same cacheline. Finally, update
mem_cgroup_alloc() to take in a parent pointer and initialize the new
cache pointers on each CPU. The percpu loop in mem_cgroup_alloc() may
look concerning, but there are multiple similar loops in the cgroup
creation path (e.g. cgroup_rstat_init()), most of which are hidden
within alloc_percpu().
According to Oliver's testing [1], this fixes multiple 30-38%
regressions in vm-scalability, will-it-scale-tlb_flush2, and
will-it-scale-fallocate1. This comes at a cost of 2 more pointers per
CPU (<2KB on a machine with 128 CPUs).
[1] https://lore.kernel.org/lkml/ZbDJsfsZt2ITyo61@xsang-OptiPlex-9020/
[yosryahmed@google.com: fix struct memcg_vmstats_percpu size and alignment]
Link: https://lkml.kernel.org/r/20240203044612.1234216-1-yosryahmed@google.com
Link: https://lkml.kernel.org/r/20240124100023.660032-1-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Fixes: 8d59d2214c23 ("mm: memcg: make stats flushing threshold per-memcg")
Tested-by: kernel test robot <oliver.sang@intel.com>
Reported-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202401221624.cb53a8ca-oliver.sang@intel.com
Acked-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-01-24 10:00:22 +00:00
|
|
|
|
mm: memcg: make stats flushing threshold per-memcg
A global counter for the magnitude of memcg stats update is maintained on
the memcg side to avoid invoking rstat flushes when the pending updates
are not significant. This avoids unnecessary flushes, which are not very
cheap even if there isn't a lot of stats to flush. It also avoids
unnecessary lock contention on the underlying global rstat lock.
Make this threshold per-memcg. The scheme is followed where percpu (now
also per-memcg) counters are incremented in the update path, and only
propagated to per-memcg atomics when they exceed a certain threshold.
This provides two benefits: (a) On large machines with a lot of memcgs,
the global threshold can be reached relatively fast, so guarding the
underlying lock becomes less effective. Making the threshold per-memcg
avoids this.
(b) Having a global threshold makes it hard to do subtree flushes, as we
cannot reset the global counter except for a full flush. Per-memcg
counters removes this as a blocker from doing subtree flushes, which helps
avoid unnecessary work when the stats of a small subtree are needed.
Nothing is free, of course. This comes at a cost: (a) A new per-cpu
counter per memcg, consuming NR_CPUS * NR_MEMCGS * 4 bytes. The extra
memory usage is insigificant.
(b) More work on the update side, although in the common case it will only
be percpu counter updates. The amount of work scales with the number of
ancestors (i.e. tree depth). This is not a new concept, adding a cgroup
to the rstat tree involves a parent loop, so is charging. Testing results
below show no significant regressions.
(c) The error margin in the stats for the system as a whole increases from
NR_CPUS * MEMCG_CHARGE_BATCH to NR_CPUS * MEMCG_CHARGE_BATCH * NR_MEMCGS.
This is probably fine because we have a similar per-memcg error in charges
coming from percpu stocks, and we have a periodic flusher that makes sure
we always flush all the stats every 2s anyway.
This patch was tested to make sure no significant regressions are
introduced on the update path as follows. The following benchmarks were
ran in a cgroup that is 2 levels deep (/sys/fs/cgroup/a/b/):
(1) Running 22 instances of netperf on a 44 cpu machine with
hyperthreading disabled. All instances are run in a level 2 cgroup, as
well as netserver:
# netserver -6
# netperf -6 -H ::1 -l 60 -t TCP_SENDFILE -- -m 10K
Averaging 20 runs, the numbers are as follows:
Base: 40198.0 mbps
Patched: 38629.7 mbps (-3.9%)
The regression is minimal, especially for 22 instances in the same
cgroup sharing all ancestors (so updating the same atomics).
(2) will-it-scale page_fault tests. These tests (specifically
per_process_ops in page_fault3 test) detected a 25.9% regression before
for a change in the stats update path [1]. These are the
numbers from 10 runs (+ is good) on a machine with 256 cpus:
LABEL | MEAN | MEDIAN | STDDEV |
------------------------------+-------------+-------------+-------------
page_fault1_per_process_ops | | | |
(A) base | 270249.164 | 265437.000 | 13451.836 |
(B) patched | 261368.709 | 255725.000 | 13394.767 |
| -3.29% | -3.66% | |
page_fault1_per_thread_ops | | | |
(A) base | 242111.345 | 239737.000 | 10026.031 |
(B) patched | 237057.109 | 235305.000 | 9769.687 |
| -2.09% | -1.85% | |
page_fault1_scalability | | |
(A) base | 0.034387 | 0.035168 | 0.0018283 |
(B) patched | 0.033988 | 0.034573 | 0.0018056 |
| -1.16% | -1.69% | |
page_fault2_per_process_ops | | |
(A) base | 203561.836 | 203301.000 | 2550.764 |
(B) patched | 197195.945 | 197746.000 | 2264.263 |
| -3.13% | -2.73% | |
page_fault2_per_thread_ops | | |
(A) base | 171046.473 | 170776.000 | 1509.679 |
(B) patched | 166626.327 | 166406.000 | 768.753 |
| -2.58% | -2.56% | |
page_fault2_scalability | | |
(A) base | 0.054026 | 0.053821 | 0.00062121 |
(B) patched | 0.053329 | 0.05306 | 0.00048394 |
| -1.29% | -1.41% | |
page_fault3_per_process_ops | | |
(A) base | 1295807.782 | 1297550.000 | 5907.585 |
(B) patched | 1275579.873 | 1273359.000 | 8759.160 |
| -1.56% | -1.86% | |
page_fault3_per_thread_ops | | |
(A) base | 391234.164 | 390860.000 | 1760.720 |
(B) patched | 377231.273 | 376369.000 | 1874.971 |
| -3.58% | -3.71% | |
page_fault3_scalability | | |
(A) base | 0.60369 | 0.60072 | 0.0083029 |
(B) patched | 0.61733 | 0.61544 | 0.009855 |
| +2.26% | +2.45% | |
All regressions seem to be minimal, and within the normal variance for the
benchmark. The fix for [1] assumes that 3% is noise -- and there were no
further practical complaints), so hopefully this means that such
variations in these microbenchmarks do not reflect on practical workloads.
(3) I also ran stress-ng in a nested cgroup and did not observe any
obvious regressions.
[1]https://lore.kernel.org/all/20190520063534.GB19312@shao2-debian/
Link: https://lkml.kernel.org/r/20231129032154.3710765-4-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Tested-by: Domenico Cerasuolo <cerasuolodomenico@gmail.com>
Acked-by: Shakeel Butt <shakeelb@google.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: Greg Thelen <gthelen@google.com>
Cc: Ivan Babrou <ivan@cloudflare.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Michal Koutny <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Tejun Heo <tj@kernel.org>
Cc: Waiman Long <longman@redhat.com>
Cc: Wei Xu <weixugc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-11-29 03:21:51 +00:00
|
|
|
|
2025-05-18 23:31:42 -07:00
|
|
|
|
mm: memcg: make stats flushing threshold per-memcg
A global counter for the magnitude of memcg stats update is maintained on
the memcg side to avoid invoking rstat flushes when the pending updates
are not significant. This avoids unnecessary flushes, which are not very
cheap even if there isn't a lot of stats to flush. It also avoids
unnecessary lock contention on the underlying global rstat lock.
Make this threshold per-memcg. The scheme is followed where percpu (now
also per-memcg) counters are incremented in the update path, and only
propagated to per-memcg atomics when they exceed a certain threshold.
This provides two benefits: (a) On large machines with a lot of memcgs,
the global threshold can be reached relatively fast, so guarding the
underlying lock becomes less effective. Making the threshold per-memcg
avoids this.
(b) Having a global threshold makes it hard to do subtree flushes, as we
cannot reset the global counter except for a full flush. Per-memcg
counters removes this as a blocker from doing subtree flushes, which helps
avoid unnecessary work when the stats of a small subtree are needed.
Nothing is free, of course. This comes at a cost: (a) A new per-cpu
counter per memcg, consuming NR_CPUS * NR_MEMCGS * 4 bytes. The extra
memory usage is insigificant.
(b) More work on the update side, although in the common case it will only
be percpu counter updates. The amount of work scales with the number of
ancestors (i.e. tree depth). This is not a new concept, adding a cgroup
to the rstat tree involves a parent loop, so is charging. Testing results
below show no significant regressions.
(c) The error margin in the stats for the system as a whole increases from
NR_CPUS * MEMCG_CHARGE_BATCH to NR_CPUS * MEMCG_CHARGE_BATCH * NR_MEMCGS.
This is probably fine because we have a similar per-memcg error in charges
coming from percpu stocks, and we have a periodic flusher that makes sure
we always flush all the stats every 2s anyway.
This patch was tested to make sure no significant regressions are
introduced on the update path as follows. The following benchmarks were
ran in a cgroup that is 2 levels deep (/sys/fs/cgroup/a/b/):
(1) Running 22 instances of netperf on a 44 cpu machine with
hyperthreading disabled. All instances are run in a level 2 cgroup, as
well as netserver:
# netserver -6
# netperf -6 -H ::1 -l 60 -t TCP_SENDFILE -- -m 10K
Averaging 20 runs, the numbers are as follows:
Base: 40198.0 mbps
Patched: 38629.7 mbps (-3.9%)
The regression is minimal, especially for 22 instances in the same
cgroup sharing all ancestors (so updating the same atomics).
(2) will-it-scale page_fault tests. These tests (specifically
per_process_ops in page_fault3 test) detected a 25.9% regression before
for a change in the stats update path [1]. These are the
numbers from 10 runs (+ is good) on a machine with 256 cpus:
LABEL | MEAN | MEDIAN | STDDEV |
------------------------------+-------------+-------------+-------------
page_fault1_per_process_ops | | | |
(A) base | 270249.164 | 265437.000 | 13451.836 |
(B) patched | 261368.709 | 255725.000 | 13394.767 |
| -3.29% | -3.66% | |
page_fault1_per_thread_ops | | | |
(A) base | 242111.345 | 239737.000 | 10026.031 |
(B) patched | 237057.109 | 235305.000 | 9769.687 |
| -2.09% | -1.85% | |
page_fault1_scalability | | |
(A) base | 0.034387 | 0.035168 | 0.0018283 |
(B) patched | 0.033988 | 0.034573 | 0.0018056 |
| -1.16% | -1.69% | |
page_fault2_per_process_ops | | |
(A) base | 203561.836 | 203301.000 | 2550.764 |
(B) patched | 197195.945 | 197746.000 | 2264.263 |
| -3.13% | -2.73% | |
page_fault2_per_thread_ops | | |
(A) base | 171046.473 | 170776.000 | 1509.679 |
(B) patched | 166626.327 | 166406.000 | 768.753 |
| -2.58% | -2.56% | |
page_fault2_scalability | | |
(A) base | 0.054026 | 0.053821 | 0.00062121 |
(B) patched | 0.053329 | 0.05306 | 0.00048394 |
| -1.29% | -1.41% | |
page_fault3_per_process_ops | | |
(A) base | 1295807.782 | 1297550.000 | 5907.585 |
(B) patched | 1275579.873 | 1273359.000 | 8759.160 |
| -1.56% | -1.86% | |
page_fault3_per_thread_ops | | |
(A) base | 391234.164 | 390860.000 | 1760.720 |
(B) patched | 377231.273 | 376369.000 | 1874.971 |
| -3.58% | -3.71% | |
page_fault3_scalability | | |
(A) base | 0.60369 | 0.60072 | 0.0083029 |
(B) patched | 0.61733 | 0.61544 | 0.009855 |
| +2.26% | +2.45% | |
All regressions seem to be minimal, and within the normal variance for the
benchmark. The fix for [1] assumes that 3% is noise -- and there were no
further practical complaints), so hopefully this means that such
variations in these microbenchmarks do not reflect on practical workloads.
(3) I also ran stress-ng in a nested cgroup and did not observe any
obvious regressions.
[1]https://lore.kernel.org/all/20190520063534.GB19312@shao2-debian/
Link: https://lkml.kernel.org/r/20231129032154.3710765-4-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Tested-by: Domenico Cerasuolo <cerasuolodomenico@gmail.com>
Acked-by: Shakeel Butt <shakeelb@google.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: Greg Thelen <gthelen@google.com>
Cc: Ivan Babrou <ivan@cloudflare.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Michal Koutny <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Tejun Heo <tj@kernel.org>
Cc: Waiman Long <longman@redhat.com>
Cc: Wei Xu <weixugc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-11-29 03:21:51 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-05-14 11:41:53 -07:00
|
|
|
|
|
|
|
|
|
2021-11-05 13:37:31 -07:00
|
|
|
|
2025-05-14 11:41:52 -07:00
|
|
|
|
mm: memcg: optimize parent iteration in memcg_rstat_updated()
In memcg_rstat_updated(), we iterate the memcg being updated and its
parents to update memcg->vmstats_percpu->stats_updates in the fast path
(i.e. no atomic updates). According to my math, this is 3 memory loads
(and potentially 3 cache misses) per memcg:
- Load the address of memcg->vmstats_percpu.
- Load vmstats_percpu->stats_updates (based on some percpu calculation).
- Load the address of the parent memcg.
Avoid most of the cache misses by caching a pointer from each struct
memcg_vmstats_percpu to its parent on the corresponding CPU. In this
case, for the first memcg we have 2 memory loads (same as above):
- Load the address of memcg->vmstats_percpu.
- Load vmstats_percpu->stats_updates (based on some percpu calculation).
Then for each additional memcg, we need a single load to get the
parent's stats_updates directly. This reduces the number of loads from
O(3N) to O(2+N) -- where N is the number of memcgs we need to iterate.
Additionally, stash a pointer to memcg->vmstats in each struct
memcg_vmstats_percpu such that we can access the atomic counter that all
CPUs fold into, memcg->vmstats->stats_updates.
memcg_should_flush_stats() is changed to memcg_vmstats_needs_flush() to
accept a struct memcg_vmstats pointer accordingly.
In struct memcg_vmstats_percpu, make sure both pointers together with
stats_updates live on the same cacheline. Finally, update
mem_cgroup_alloc() to take in a parent pointer and initialize the new
cache pointers on each CPU. The percpu loop in mem_cgroup_alloc() may
look concerning, but there are multiple similar loops in the cgroup
creation path (e.g. cgroup_rstat_init()), most of which are hidden
within alloc_percpu().
According to Oliver's testing [1], this fixes multiple 30-38%
regressions in vm-scalability, will-it-scale-tlb_flush2, and
will-it-scale-fallocate1. This comes at a cost of 2 more pointers per
CPU (<2KB on a machine with 128 CPUs).
[1] https://lore.kernel.org/lkml/ZbDJsfsZt2ITyo61@xsang-OptiPlex-9020/
[yosryahmed@google.com: fix struct memcg_vmstats_percpu size and alignment]
Link: https://lkml.kernel.org/r/20240203044612.1234216-1-yosryahmed@google.com
Link: https://lkml.kernel.org/r/20240124100023.660032-1-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Fixes: 8d59d2214c23 ("mm: memcg: make stats flushing threshold per-memcg")
Tested-by: kernel test robot <oliver.sang@intel.com>
Reported-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202401221624.cb53a8ca-oliver.sang@intel.com
Acked-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-01-24 10:00:22 +00:00
|
|
|
|
2024-04-24 05:59:39 -07:00
|
|
|
|
2022-01-14 14:05:39 -08:00
|
|
|
|
2023-03-30 19:18:01 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-06-17 12:57:25 -07:00
|
|
|
|
2025-05-14 11:41:52 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-04-09 19:57:52 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-05-14 11:41:52 -07:00
|
|
|
|
|
|
|
|
|
2024-04-24 05:59:39 -07:00
|
|
|
|
mm: memcg: make stats flushing threshold per-memcg
A global counter for the magnitude of memcg stats update is maintained on
the memcg side to avoid invoking rstat flushes when the pending updates
are not significant. This avoids unnecessary flushes, which are not very
cheap even if there isn't a lot of stats to flush. It also avoids
unnecessary lock contention on the underlying global rstat lock.
Make this threshold per-memcg. The scheme is followed where percpu (now
also per-memcg) counters are incremented in the update path, and only
propagated to per-memcg atomics when they exceed a certain threshold.
This provides two benefits: (a) On large machines with a lot of memcgs,
the global threshold can be reached relatively fast, so guarding the
underlying lock becomes less effective. Making the threshold per-memcg
avoids this.
(b) Having a global threshold makes it hard to do subtree flushes, as we
cannot reset the global counter except for a full flush. Per-memcg
counters removes this as a blocker from doing subtree flushes, which helps
avoid unnecessary work when the stats of a small subtree are needed.
Nothing is free, of course. This comes at a cost: (a) A new per-cpu
counter per memcg, consuming NR_CPUS * NR_MEMCGS * 4 bytes. The extra
memory usage is insigificant.
(b) More work on the update side, although in the common case it will only
be percpu counter updates. The amount of work scales with the number of
ancestors (i.e. tree depth). This is not a new concept, adding a cgroup
to the rstat tree involves a parent loop, so is charging. Testing results
below show no significant regressions.
(c) The error margin in the stats for the system as a whole increases from
NR_CPUS * MEMCG_CHARGE_BATCH to NR_CPUS * MEMCG_CHARGE_BATCH * NR_MEMCGS.
This is probably fine because we have a similar per-memcg error in charges
coming from percpu stocks, and we have a periodic flusher that makes sure
we always flush all the stats every 2s anyway.
This patch was tested to make sure no significant regressions are
introduced on the update path as follows. The following benchmarks were
ran in a cgroup that is 2 levels deep (/sys/fs/cgroup/a/b/):
(1) Running 22 instances of netperf on a 44 cpu machine with
hyperthreading disabled. All instances are run in a level 2 cgroup, as
well as netserver:
# netserver -6
# netperf -6 -H ::1 -l 60 -t TCP_SENDFILE -- -m 10K
Averaging 20 runs, the numbers are as follows:
Base: 40198.0 mbps
Patched: 38629.7 mbps (-3.9%)
The regression is minimal, especially for 22 instances in the same
cgroup sharing all ancestors (so updating the same atomics).
(2) will-it-scale page_fault tests. These tests (specifically
per_process_ops in page_fault3 test) detected a 25.9% regression before
for a change in the stats update path [1]. These are the
numbers from 10 runs (+ is good) on a machine with 256 cpus:
LABEL | MEAN | MEDIAN | STDDEV |
------------------------------+-------------+-------------+-------------
page_fault1_per_process_ops | | | |
(A) base | 270249.164 | 265437.000 | 13451.836 |
(B) patched | 261368.709 | 255725.000 | 13394.767 |
| -3.29% | -3.66% | |
page_fault1_per_thread_ops | | | |
(A) base | 242111.345 | 239737.000 | 10026.031 |
(B) patched | 237057.109 | 235305.000 | 9769.687 |
| -2.09% | -1.85% | |
page_fault1_scalability | | |
(A) base | 0.034387 | 0.035168 | 0.0018283 |
(B) patched | 0.033988 | 0.034573 | 0.0018056 |
| -1.16% | -1.69% | |
page_fault2_per_process_ops | | |
(A) base | 203561.836 | 203301.000 | 2550.764 |
(B) patched | 197195.945 | 197746.000 | 2264.263 |
| -3.13% | -2.73% | |
page_fault2_per_thread_ops | | |
(A) base | 171046.473 | 170776.000 | 1509.679 |
(B) patched | 166626.327 | 166406.000 | 768.753 |
| -2.58% | -2.56% | |
page_fault2_scalability | | |
(A) base | 0.054026 | 0.053821 | 0.00062121 |
(B) patched | 0.053329 | 0.05306 | 0.00048394 |
| -1.29% | -1.41% | |
page_fault3_per_process_ops | | |
(A) base | 1295807.782 | 1297550.000 | 5907.585 |
(B) patched | 1275579.873 | 1273359.000 | 8759.160 |
| -1.56% | -1.86% | |
page_fault3_per_thread_ops | | |
(A) base | 391234.164 | 390860.000 | 1760.720 |
(B) patched | 377231.273 | 376369.000 | 1874.971 |
| -3.58% | -3.71% | |
page_fault3_scalability | | |
(A) base | 0.60369 | 0.60072 | 0.0083029 |
(B) patched | 0.61733 | 0.61544 | 0.009855 |
| +2.26% | +2.45% | |
All regressions seem to be minimal, and within the normal variance for the
benchmark. The fix for [1] assumes that 3% is noise -- and there were no
further practical complaints), so hopefully this means that such
variations in these microbenchmarks do not reflect on practical workloads.
(3) I also ran stress-ng in a nested cgroup and did not observe any
obvious regressions.
[1]https://lore.kernel.org/all/20190520063534.GB19312@shao2-debian/
Link: https://lkml.kernel.org/r/20231129032154.3710765-4-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Tested-by: Domenico Cerasuolo <cerasuolodomenico@gmail.com>
Acked-by: Shakeel Butt <shakeelb@google.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: Greg Thelen <gthelen@google.com>
Cc: Ivan Babrou <ivan@cloudflare.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Michal Koutny <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Tejun Heo <tj@kernel.org>
Cc: Waiman Long <longman@redhat.com>
Cc: Wei Xu <weixugc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-11-29 03:21:51 +00:00
|
|
|
|
2022-01-14 14:05:39 -08:00
|
|
|
|
2025-05-14 11:41:52 -07:00
|
|
|
|
2025-05-18 23:31:42 -07:00
|
|
|
|
2022-01-14 14:05:39 -08:00
|
|
|
|
2021-11-05 13:37:31 -07:00
|
|
|
|
|
|
|
|
|
2024-10-28 19:11:05 -07:00
|
|
|
|
2021-11-05 13:37:31 -07:00
|
|
|
|
2024-10-28 19:11:06 -07:00
|
|
|
|
|
|
|
|
|
2025-05-18 23:31:42 -07:00
|
|
|
|
2024-10-28 19:11:06 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-10-28 19:11:05 -07:00
|
|
|
|
|
|
|
|
|
mm: memcg: restore subtree stats flushing
Stats flushing for memcg currently follows the following rules:
- Always flush the entire memcg hierarchy (i.e. flush the root).
- Only one flusher is allowed at a time. If someone else tries to flush
concurrently, they skip and return immediately.
- A periodic flusher flushes all the stats every 2 seconds.
The reason this approach is followed is because all flushes are serialized
by a global rstat spinlock. On the memcg side, flushing is invoked from
userspace reads as well as in-kernel flushers (e.g. reclaim, refault,
etc). This approach aims to avoid serializing all flushers on the global
lock, which can cause a significant performance hit under high
concurrency.
This approach has the following problems:
- Occasionally a userspace read of the stats of a non-root cgroup will
be too expensive as it has to flush the entire hierarchy [1].
- Sometimes the stats accuracy are compromised if there is an ongoing
flush, and we skip and return before the subtree of interest is
actually flushed, yielding stale stats (by up to 2s due to periodic
flushing). This is more visible when reading stats from userspace,
but can also affect in-kernel flushers.
The latter problem is particulary a concern when userspace reads stats
after an event occurs, but gets stats from before the event. Examples:
- When memory usage / pressure spikes, a userspace OOM handler may look
at the stats of different memcgs to select a victim based on various
heuristics (e.g. how much private memory will be freed by killing
this). Reading stale stats from before the usage spike in this case
may cause a wrongful OOM kill.
- A proactive reclaimer may read the stats after writing to
memory.reclaim to measure the success of the reclaim operation. Stale
stats from before reclaim may give a false negative.
- Reading the stats of a parent and a child memcg may be inconsistent
(child larger than parent), if the flush doesn't happen when the
parent is read, but happens when the child is read.
As for in-kernel flushers, they will occasionally get stale stats. No
regressions are currently known from this, but if there are regressions,
they would be very difficult to debug and link to the source of the
problem.
This patch aims to fix these problems by restoring subtree flushing, and
removing the unified/coalesced flushing logic that skips flushing if there
is an ongoing flush. This change would introduce a significant regression
with global stats flushing thresholds. With per-memcg stats flushing
thresholds, this seems to perform really well. The thresholds protect the
underlying lock from unnecessary contention.
This patch was tested in two ways to ensure the latency of flushing is
up to par, on a machine with 384 cpus:
- A synthetic test with 5000 concurrent workers in 500 cgroups doing
allocations and reclaim, as well as 1000 readers for memory.stat
(variation of [2]). No regressions were noticed in the total runtime.
Note that significant regressions in this test are observed with
global stats thresholds, but not with per-memcg thresholds.
- A synthetic stress test for concurrently reading memcg stats while
memory allocation/freeing workers are running in the background,
provided by Wei Xu [3]. With 250k threads reading the stats every
100ms in 50k cgroups, 99.9% of reads take <= 50us. Less than 0.01%
of reads take more than 1ms, and no reads take more than 100ms.
[1] https://lore.kernel.org/lkml/CABWYdi0c6__rh-K7dcM_pkf9BJdTRtAU08M43KO9ME4-dsgfoQ@mail.gmail.com/
[2] https://lore.kernel.org/lkml/CAJD7tka13M-zVZTyQJYL1iUAYvuQ1fcHbCjcOBZcz6POYTV-4g@mail.gmail.com/
[3] https://lore.kernel.org/lkml/CAAPL-u9D2b=iF5Lf_cRnKxUfkiEe0AMDTu6yhrUAzX0b6a6rDg@mail.gmail.com/
[akpm@linux-foundation.org: fix mm/zswap.c]
[yosryahmed@google.com: remove stats flushing mutex]
Link: https://lkml.kernel.org/r/CAJD7tkZgP3m-VVPn+fF_YuvXeQYK=tZZjJHj=dzD=CcSSpp2qg@mail.gmail.com
Link: https://lkml.kernel.org/r/20231129032154.3710765-6-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Tested-by: Domenico Cerasuolo <cerasuolodomenico@gmail.com>
Acked-by: Shakeel Butt <shakeelb@google.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: Greg Thelen <gthelen@google.com>
Cc: Ivan Babrou <ivan@cloudflare.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Michal Koutny <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Tejun Heo <tj@kernel.org>
Cc: Waiman Long <longman@redhat.com>
Cc: Wei Xu <weixugc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-11-29 03:21:53 +00:00
|
|
|
|
|
|
|
|
|
2023-03-30 19:17:58 +00:00
|
|
|
|
2025-04-03 18:10:48 -07:00
|
|
|
|
2021-11-05 13:37:31 -07:00
|
|
|
|
|
|
|
|
|
mm: memcg: restore subtree stats flushing
Stats flushing for memcg currently follows the following rules:
- Always flush the entire memcg hierarchy (i.e. flush the root).
- Only one flusher is allowed at a time. If someone else tries to flush
concurrently, they skip and return immediately.
- A periodic flusher flushes all the stats every 2 seconds.
The reason this approach is followed is because all flushes are serialized
by a global rstat spinlock. On the memcg side, flushing is invoked from
userspace reads as well as in-kernel flushers (e.g. reclaim, refault,
etc). This approach aims to avoid serializing all flushers on the global
lock, which can cause a significant performance hit under high
concurrency.
This approach has the following problems:
- Occasionally a userspace read of the stats of a non-root cgroup will
be too expensive as it has to flush the entire hierarchy [1].
- Sometimes the stats accuracy are compromised if there is an ongoing
flush, and we skip and return before the subtree of interest is
actually flushed, yielding stale stats (by up to 2s due to periodic
flushing). This is more visible when reading stats from userspace,
but can also affect in-kernel flushers.
The latter problem is particulary a concern when userspace reads stats
after an event occurs, but gets stats from before the event. Examples:
- When memory usage / pressure spikes, a userspace OOM handler may look
at the stats of different memcgs to select a victim based on various
heuristics (e.g. how much private memory will be freed by killing
this). Reading stale stats from before the usage spike in this case
may cause a wrongful OOM kill.
- A proactive reclaimer may read the stats after writing to
memory.reclaim to measure the success of the reclaim operation. Stale
stats from before reclaim may give a false negative.
- Reading the stats of a parent and a child memcg may be inconsistent
(child larger than parent), if the flush doesn't happen when the
parent is read, but happens when the child is read.
As for in-kernel flushers, they will occasionally get stale stats. No
regressions are currently known from this, but if there are regressions,
they would be very difficult to debug and link to the source of the
problem.
This patch aims to fix these problems by restoring subtree flushing, and
removing the unified/coalesced flushing logic that skips flushing if there
is an ongoing flush. This change would introduce a significant regression
with global stats flushing thresholds. With per-memcg stats flushing
thresholds, this seems to perform really well. The thresholds protect the
underlying lock from unnecessary contention.
This patch was tested in two ways to ensure the latency of flushing is
up to par, on a machine with 384 cpus:
- A synthetic test with 5000 concurrent workers in 500 cgroups doing
allocations and reclaim, as well as 1000 readers for memory.stat
(variation of [2]). No regressions were noticed in the total runtime.
Note that significant regressions in this test are observed with
global stats thresholds, but not with per-memcg thresholds.
- A synthetic stress test for concurrently reading memcg stats while
memory allocation/freeing workers are running in the background,
provided by Wei Xu [3]. With 250k threads reading the stats every
100ms in 50k cgroups, 99.9% of reads take <= 50us. Less than 0.01%
of reads take more than 1ms, and no reads take more than 100ms.
[1] https://lore.kernel.org/lkml/CABWYdi0c6__rh-K7dcM_pkf9BJdTRtAU08M43KO9ME4-dsgfoQ@mail.gmail.com/
[2] https://lore.kernel.org/lkml/CAJD7tka13M-zVZTyQJYL1iUAYvuQ1fcHbCjcOBZcz6POYTV-4g@mail.gmail.com/
[3] https://lore.kernel.org/lkml/CAAPL-u9D2b=iF5Lf_cRnKxUfkiEe0AMDTu6yhrUAzX0b6a6rDg@mail.gmail.com/
[akpm@linux-foundation.org: fix mm/zswap.c]
[yosryahmed@google.com: remove stats flushing mutex]
Link: https://lkml.kernel.org/r/CAJD7tkZgP3m-VVPn+fF_YuvXeQYK=tZZjJHj=dzD=CcSSpp2qg@mail.gmail.com
Link: https://lkml.kernel.org/r/20231129032154.3710765-6-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Tested-by: Domenico Cerasuolo <cerasuolodomenico@gmail.com>
Acked-by: Shakeel Butt <shakeelb@google.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: Greg Thelen <gthelen@google.com>
Cc: Ivan Babrou <ivan@cloudflare.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Michal Koutny <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Tejun Heo <tj@kernel.org>
Cc: Waiman Long <longman@redhat.com>
Cc: Wei Xu <weixugc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-11-29 03:21:53 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-11-05 13:37:31 -07:00
|
|
|
|
mm: memcg: restore subtree stats flushing
Stats flushing for memcg currently follows the following rules:
- Always flush the entire memcg hierarchy (i.e. flush the root).
- Only one flusher is allowed at a time. If someone else tries to flush
concurrently, they skip and return immediately.
- A periodic flusher flushes all the stats every 2 seconds.
The reason this approach is followed is because all flushes are serialized
by a global rstat spinlock. On the memcg side, flushing is invoked from
userspace reads as well as in-kernel flushers (e.g. reclaim, refault,
etc). This approach aims to avoid serializing all flushers on the global
lock, which can cause a significant performance hit under high
concurrency.
This approach has the following problems:
- Occasionally a userspace read of the stats of a non-root cgroup will
be too expensive as it has to flush the entire hierarchy [1].
- Sometimes the stats accuracy are compromised if there is an ongoing
flush, and we skip and return before the subtree of interest is
actually flushed, yielding stale stats (by up to 2s due to periodic
flushing). This is more visible when reading stats from userspace,
but can also affect in-kernel flushers.
The latter problem is particulary a concern when userspace reads stats
after an event occurs, but gets stats from before the event. Examples:
- When memory usage / pressure spikes, a userspace OOM handler may look
at the stats of different memcgs to select a victim based on various
heuristics (e.g. how much private memory will be freed by killing
this). Reading stale stats from before the usage spike in this case
may cause a wrongful OOM kill.
- A proactive reclaimer may read the stats after writing to
memory.reclaim to measure the success of the reclaim operation. Stale
stats from before reclaim may give a false negative.
- Reading the stats of a parent and a child memcg may be inconsistent
(child larger than parent), if the flush doesn't happen when the
parent is read, but happens when the child is read.
As for in-kernel flushers, they will occasionally get stale stats. No
regressions are currently known from this, but if there are regressions,
they would be very difficult to debug and link to the source of the
problem.
This patch aims to fix these problems by restoring subtree flushing, and
removing the unified/coalesced flushing logic that skips flushing if there
is an ongoing flush. This change would introduce a significant regression
with global stats flushing thresholds. With per-memcg stats flushing
thresholds, this seems to perform really well. The thresholds protect the
underlying lock from unnecessary contention.
This patch was tested in two ways to ensure the latency of flushing is
up to par, on a machine with 384 cpus:
- A synthetic test with 5000 concurrent workers in 500 cgroups doing
allocations and reclaim, as well as 1000 readers for memory.stat
(variation of [2]). No regressions were noticed in the total runtime.
Note that significant regressions in this test are observed with
global stats thresholds, but not with per-memcg thresholds.
- A synthetic stress test for concurrently reading memcg stats while
memory allocation/freeing workers are running in the background,
provided by Wei Xu [3]. With 250k threads reading the stats every
100ms in 50k cgroups, 99.9% of reads take <= 50us. Less than 0.01%
of reads take more than 1ms, and no reads take more than 100ms.
[1] https://lore.kernel.org/lkml/CABWYdi0c6__rh-K7dcM_pkf9BJdTRtAU08M43KO9ME4-dsgfoQ@mail.gmail.com/
[2] https://lore.kernel.org/lkml/CAJD7tka13M-zVZTyQJYL1iUAYvuQ1fcHbCjcOBZcz6POYTV-4g@mail.gmail.com/
[3] https://lore.kernel.org/lkml/CAAPL-u9D2b=iF5Lf_cRnKxUfkiEe0AMDTu6yhrUAzX0b6a6rDg@mail.gmail.com/
[akpm@linux-foundation.org: fix mm/zswap.c]
[yosryahmed@google.com: remove stats flushing mutex]
Link: https://lkml.kernel.org/r/CAJD7tkZgP3m-VVPn+fF_YuvXeQYK=tZZjJHj=dzD=CcSSpp2qg@mail.gmail.com
Link: https://lkml.kernel.org/r/20231129032154.3710765-6-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Tested-by: Domenico Cerasuolo <cerasuolodomenico@gmail.com>
Acked-by: Shakeel Butt <shakeelb@google.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: Greg Thelen <gthelen@google.com>
Cc: Ivan Babrou <ivan@cloudflare.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Michal Koutny <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Tejun Heo <tj@kernel.org>
Cc: Waiman Long <longman@redhat.com>
Cc: Wei Xu <weixugc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-11-29 03:21:53 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-10-28 19:11:05 -07:00
|
|
|
|
2023-03-30 19:17:58 +00:00
|
|
|
|
|
|
|
|
|
mm: memcg: restore subtree stats flushing
Stats flushing for memcg currently follows the following rules:
- Always flush the entire memcg hierarchy (i.e. flush the root).
- Only one flusher is allowed at a time. If someone else tries to flush
concurrently, they skip and return immediately.
- A periodic flusher flushes all the stats every 2 seconds.
The reason this approach is followed is because all flushes are serialized
by a global rstat spinlock. On the memcg side, flushing is invoked from
userspace reads as well as in-kernel flushers (e.g. reclaim, refault,
etc). This approach aims to avoid serializing all flushers on the global
lock, which can cause a significant performance hit under high
concurrency.
This approach has the following problems:
- Occasionally a userspace read of the stats of a non-root cgroup will
be too expensive as it has to flush the entire hierarchy [1].
- Sometimes the stats accuracy are compromised if there is an ongoing
flush, and we skip and return before the subtree of interest is
actually flushed, yielding stale stats (by up to 2s due to periodic
flushing). This is more visible when reading stats from userspace,
but can also affect in-kernel flushers.
The latter problem is particulary a concern when userspace reads stats
after an event occurs, but gets stats from before the event. Examples:
- When memory usage / pressure spikes, a userspace OOM handler may look
at the stats of different memcgs to select a victim based on various
heuristics (e.g. how much private memory will be freed by killing
this). Reading stale stats from before the usage spike in this case
may cause a wrongful OOM kill.
- A proactive reclaimer may read the stats after writing to
memory.reclaim to measure the success of the reclaim operation. Stale
stats from before reclaim may give a false negative.
- Reading the stats of a parent and a child memcg may be inconsistent
(child larger than parent), if the flush doesn't happen when the
parent is read, but happens when the child is read.
As for in-kernel flushers, they will occasionally get stale stats. No
regressions are currently known from this, but if there are regressions,
they would be very difficult to debug and link to the source of the
problem.
This patch aims to fix these problems by restoring subtree flushing, and
removing the unified/coalesced flushing logic that skips flushing if there
is an ongoing flush. This change would introduce a significant regression
with global stats flushing thresholds. With per-memcg stats flushing
thresholds, this seems to perform really well. The thresholds protect the
underlying lock from unnecessary contention.
This patch was tested in two ways to ensure the latency of flushing is
up to par, on a machine with 384 cpus:
- A synthetic test with 5000 concurrent workers in 500 cgroups doing
allocations and reclaim, as well as 1000 readers for memory.stat
(variation of [2]). No regressions were noticed in the total runtime.
Note that significant regressions in this test are observed with
global stats thresholds, but not with per-memcg thresholds.
- A synthetic stress test for concurrently reading memcg stats while
memory allocation/freeing workers are running in the background,
provided by Wei Xu [3]. With 250k threads reading the stats every
100ms in 50k cgroups, 99.9% of reads take <= 50us. Less than 0.01%
of reads take more than 1ms, and no reads take more than 100ms.
[1] https://lore.kernel.org/lkml/CABWYdi0c6__rh-K7dcM_pkf9BJdTRtAU08M43KO9ME4-dsgfoQ@mail.gmail.com/
[2] https://lore.kernel.org/lkml/CAJD7tka13M-zVZTyQJYL1iUAYvuQ1fcHbCjcOBZcz6POYTV-4g@mail.gmail.com/
[3] https://lore.kernel.org/lkml/CAAPL-u9D2b=iF5Lf_cRnKxUfkiEe0AMDTu6yhrUAzX0b6a6rDg@mail.gmail.com/
[akpm@linux-foundation.org: fix mm/zswap.c]
[yosryahmed@google.com: remove stats flushing mutex]
Link: https://lkml.kernel.org/r/CAJD7tkZgP3m-VVPn+fF_YuvXeQYK=tZZjJHj=dzD=CcSSpp2qg@mail.gmail.com
Link: https://lkml.kernel.org/r/20231129032154.3710765-6-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Tested-by: Domenico Cerasuolo <cerasuolodomenico@gmail.com>
Acked-by: Shakeel Butt <shakeelb@google.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: Greg Thelen <gthelen@google.com>
Cc: Ivan Babrou <ivan@cloudflare.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Michal Koutny <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Tejun Heo <tj@kernel.org>
Cc: Waiman Long <longman@redhat.com>
Cc: Wei Xu <weixugc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-11-29 03:21:53 +00:00
|
|
|
|
2022-04-21 16:35:40 -07:00
|
|
|
|
2023-11-29 03:21:49 +00:00
|
|
|
|
|
|
|
|
|
mm: memcg: restore subtree stats flushing
Stats flushing for memcg currently follows the following rules:
- Always flush the entire memcg hierarchy (i.e. flush the root).
- Only one flusher is allowed at a time. If someone else tries to flush
concurrently, they skip and return immediately.
- A periodic flusher flushes all the stats every 2 seconds.
The reason this approach is followed is because all flushes are serialized
by a global rstat spinlock. On the memcg side, flushing is invoked from
userspace reads as well as in-kernel flushers (e.g. reclaim, refault,
etc). This approach aims to avoid serializing all flushers on the global
lock, which can cause a significant performance hit under high
concurrency.
This approach has the following problems:
- Occasionally a userspace read of the stats of a non-root cgroup will
be too expensive as it has to flush the entire hierarchy [1].
- Sometimes the stats accuracy are compromised if there is an ongoing
flush, and we skip and return before the subtree of interest is
actually flushed, yielding stale stats (by up to 2s due to periodic
flushing). This is more visible when reading stats from userspace,
but can also affect in-kernel flushers.
The latter problem is particulary a concern when userspace reads stats
after an event occurs, but gets stats from before the event. Examples:
- When memory usage / pressure spikes, a userspace OOM handler may look
at the stats of different memcgs to select a victim based on various
heuristics (e.g. how much private memory will be freed by killing
this). Reading stale stats from before the usage spike in this case
may cause a wrongful OOM kill.
- A proactive reclaimer may read the stats after writing to
memory.reclaim to measure the success of the reclaim operation. Stale
stats from before reclaim may give a false negative.
- Reading the stats of a parent and a child memcg may be inconsistent
(child larger than parent), if the flush doesn't happen when the
parent is read, but happens when the child is read.
As for in-kernel flushers, they will occasionally get stale stats. No
regressions are currently known from this, but if there are regressions,
they would be very difficult to debug and link to the source of the
problem.
This patch aims to fix these problems by restoring subtree flushing, and
removing the unified/coalesced flushing logic that skips flushing if there
is an ongoing flush. This change would introduce a significant regression
with global stats flushing thresholds. With per-memcg stats flushing
thresholds, this seems to perform really well. The thresholds protect the
underlying lock from unnecessary contention.
This patch was tested in two ways to ensure the latency of flushing is
up to par, on a machine with 384 cpus:
- A synthetic test with 5000 concurrent workers in 500 cgroups doing
allocations and reclaim, as well as 1000 readers for memory.stat
(variation of [2]). No regressions were noticed in the total runtime.
Note that significant regressions in this test are observed with
global stats thresholds, but not with per-memcg thresholds.
- A synthetic stress test for concurrently reading memcg stats while
memory allocation/freeing workers are running in the background,
provided by Wei Xu [3]. With 250k threads reading the stats every
100ms in 50k cgroups, 99.9% of reads take <= 50us. Less than 0.01%
of reads take more than 1ms, and no reads take more than 100ms.
[1] https://lore.kernel.org/lkml/CABWYdi0c6__rh-K7dcM_pkf9BJdTRtAU08M43KO9ME4-dsgfoQ@mail.gmail.com/
[2] https://lore.kernel.org/lkml/CAJD7tka13M-zVZTyQJYL1iUAYvuQ1fcHbCjcOBZcz6POYTV-4g@mail.gmail.com/
[3] https://lore.kernel.org/lkml/CAAPL-u9D2b=iF5Lf_cRnKxUfkiEe0AMDTu6yhrUAzX0b6a6rDg@mail.gmail.com/
[akpm@linux-foundation.org: fix mm/zswap.c]
[yosryahmed@google.com: remove stats flushing mutex]
Link: https://lkml.kernel.org/r/CAJD7tkZgP3m-VVPn+fF_YuvXeQYK=tZZjJHj=dzD=CcSSpp2qg@mail.gmail.com
Link: https://lkml.kernel.org/r/20231129032154.3710765-6-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Tested-by: Domenico Cerasuolo <cerasuolodomenico@gmail.com>
Acked-by: Shakeel Butt <shakeelb@google.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: Greg Thelen <gthelen@google.com>
Cc: Ivan Babrou <ivan@cloudflare.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Michal Koutny <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Tejun Heo <tj@kernel.org>
Cc: Waiman Long <longman@redhat.com>
Cc: Wei Xu <weixugc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-11-29 03:21:53 +00:00
|
|
|
|
2022-04-21 16:35:40 -07:00
|
|
|
|
|
|
|
|
|
2021-11-05 13:37:31 -07:00
|
|
|
|
|
|
|
|
|
2023-03-30 19:17:58 +00:00
|
|
|
|
mm: memcg: optimize parent iteration in memcg_rstat_updated()
In memcg_rstat_updated(), we iterate the memcg being updated and its
parents to update memcg->vmstats_percpu->stats_updates in the fast path
(i.e. no atomic updates). According to my math, this is 3 memory loads
(and potentially 3 cache misses) per memcg:
- Load the address of memcg->vmstats_percpu.
- Load vmstats_percpu->stats_updates (based on some percpu calculation).
- Load the address of the parent memcg.
Avoid most of the cache misses by caching a pointer from each struct
memcg_vmstats_percpu to its parent on the corresponding CPU. In this
case, for the first memcg we have 2 memory loads (same as above):
- Load the address of memcg->vmstats_percpu.
- Load vmstats_percpu->stats_updates (based on some percpu calculation).
Then for each additional memcg, we need a single load to get the
parent's stats_updates directly. This reduces the number of loads from
O(3N) to O(2+N) -- where N is the number of memcgs we need to iterate.
Additionally, stash a pointer to memcg->vmstats in each struct
memcg_vmstats_percpu such that we can access the atomic counter that all
CPUs fold into, memcg->vmstats->stats_updates.
memcg_should_flush_stats() is changed to memcg_vmstats_needs_flush() to
accept a struct memcg_vmstats pointer accordingly.
In struct memcg_vmstats_percpu, make sure both pointers together with
stats_updates live on the same cacheline. Finally, update
mem_cgroup_alloc() to take in a parent pointer and initialize the new
cache pointers on each CPU. The percpu loop in mem_cgroup_alloc() may
look concerning, but there are multiple similar loops in the cgroup
creation path (e.g. cgroup_rstat_init()), most of which are hidden
within alloc_percpu().
According to Oliver's testing [1], this fixes multiple 30-38%
regressions in vm-scalability, will-it-scale-tlb_flush2, and
will-it-scale-fallocate1. This comes at a cost of 2 more pointers per
CPU (<2KB on a machine with 128 CPUs).
[1] https://lore.kernel.org/lkml/ZbDJsfsZt2ITyo61@xsang-OptiPlex-9020/
[yosryahmed@google.com: fix struct memcg_vmstats_percpu size and alignment]
Link: https://lkml.kernel.org/r/20240203044612.1234216-1-yosryahmed@google.com
Link: https://lkml.kernel.org/r/20240124100023.660032-1-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Fixes: 8d59d2214c23 ("mm: memcg: make stats flushing threshold per-memcg")
Tested-by: kernel test robot <oliver.sang@intel.com>
Reported-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202401221624.cb53a8ca-oliver.sang@intel.com
Acked-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-01-24 10:00:22 +00:00
|
|
|
|
mm: memcg: make stats flushing threshold per-memcg
A global counter for the magnitude of memcg stats update is maintained on
the memcg side to avoid invoking rstat flushes when the pending updates
are not significant. This avoids unnecessary flushes, which are not very
cheap even if there isn't a lot of stats to flush. It also avoids
unnecessary lock contention on the underlying global rstat lock.
Make this threshold per-memcg. The scheme is followed where percpu (now
also per-memcg) counters are incremented in the update path, and only
propagated to per-memcg atomics when they exceed a certain threshold.
This provides two benefits: (a) On large machines with a lot of memcgs,
the global threshold can be reached relatively fast, so guarding the
underlying lock becomes less effective. Making the threshold per-memcg
avoids this.
(b) Having a global threshold makes it hard to do subtree flushes, as we
cannot reset the global counter except for a full flush. Per-memcg
counters removes this as a blocker from doing subtree flushes, which helps
avoid unnecessary work when the stats of a small subtree are needed.
Nothing is free, of course. This comes at a cost: (a) A new per-cpu
counter per memcg, consuming NR_CPUS * NR_MEMCGS * 4 bytes. The extra
memory usage is insigificant.
(b) More work on the update side, although in the common case it will only
be percpu counter updates. The amount of work scales with the number of
ancestors (i.e. tree depth). This is not a new concept, adding a cgroup
to the rstat tree involves a parent loop, so is charging. Testing results
below show no significant regressions.
(c) The error margin in the stats for the system as a whole increases from
NR_CPUS * MEMCG_CHARGE_BATCH to NR_CPUS * MEMCG_CHARGE_BATCH * NR_MEMCGS.
This is probably fine because we have a similar per-memcg error in charges
coming from percpu stocks, and we have a periodic flusher that makes sure
we always flush all the stats every 2s anyway.
This patch was tested to make sure no significant regressions are
introduced on the update path as follows. The following benchmarks were
ran in a cgroup that is 2 levels deep (/sys/fs/cgroup/a/b/):
(1) Running 22 instances of netperf on a 44 cpu machine with
hyperthreading disabled. All instances are run in a level 2 cgroup, as
well as netserver:
# netserver -6
# netperf -6 -H ::1 -l 60 -t TCP_SENDFILE -- -m 10K
Averaging 20 runs, the numbers are as follows:
Base: 40198.0 mbps
Patched: 38629.7 mbps (-3.9%)
The regression is minimal, especially for 22 instances in the same
cgroup sharing all ancestors (so updating the same atomics).
(2) will-it-scale page_fault tests. These tests (specifically
per_process_ops in page_fault3 test) detected a 25.9% regression before
for a change in the stats update path [1]. These are the
numbers from 10 runs (+ is good) on a machine with 256 cpus:
LABEL | MEAN | MEDIAN | STDDEV |
------------------------------+-------------+-------------+-------------
page_fault1_per_process_ops | | | |
(A) base | 270249.164 | 265437.000 | 13451.836 |
(B) patched | 261368.709 | 255725.000 | 13394.767 |
| -3.29% | -3.66% | |
page_fault1_per_thread_ops | | | |
(A) base | 242111.345 | 239737.000 | 10026.031 |
(B) patched | 237057.109 | 235305.000 | 9769.687 |
| -2.09% | -1.85% | |
page_fault1_scalability | | |
(A) base | 0.034387 | 0.035168 | 0.0018283 |
(B) patched | 0.033988 | 0.034573 | 0.0018056 |
| -1.16% | -1.69% | |
page_fault2_per_process_ops | | |
(A) base | 203561.836 | 203301.000 | 2550.764 |
(B) patched | 197195.945 | 197746.000 | 2264.263 |
| -3.13% | -2.73% | |
page_fault2_per_thread_ops | | |
(A) base | 171046.473 | 170776.000 | 1509.679 |
(B) patched | 166626.327 | 166406.000 | 768.753 |
| -2.58% | -2.56% | |
page_fault2_scalability | | |
(A) base | 0.054026 | 0.053821 | 0.00062121 |
(B) patched | 0.053329 | 0.05306 | 0.00048394 |
| -1.29% | -1.41% | |
page_fault3_per_process_ops | | |
(A) base | 1295807.782 | 1297550.000 | 5907.585 |
(B) patched | 1275579.873 | 1273359.000 | 8759.160 |
| -1.56% | -1.86% | |
page_fault3_per_thread_ops | | |
(A) base | 391234.164 | 390860.000 | 1760.720 |
(B) patched | 377231.273 | 376369.000 | 1874.971 |
| -3.58% | -3.71% | |
page_fault3_scalability | | |
(A) base | 0.60369 | 0.60072 | 0.0083029 |
(B) patched | 0.61733 | 0.61544 | 0.009855 |
| +2.26% | +2.45% | |
All regressions seem to be minimal, and within the normal variance for the
benchmark. The fix for [1] assumes that 3% is noise -- and there were no
further practical complaints), so hopefully this means that such
variations in these microbenchmarks do not reflect on practical workloads.
(3) I also ran stress-ng in a nested cgroup and did not observe any
obvious regressions.
[1]https://lore.kernel.org/all/20190520063534.GB19312@shao2-debian/
Link: https://lkml.kernel.org/r/20231129032154.3710765-4-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Tested-by: Domenico Cerasuolo <cerasuolodomenico@gmail.com>
Acked-by: Shakeel Butt <shakeelb@google.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: Greg Thelen <gthelen@google.com>
Cc: Ivan Babrou <ivan@cloudflare.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Michal Koutny <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Tejun Heo <tj@kernel.org>
Cc: Waiman Long <longman@redhat.com>
Cc: Wei Xu <weixugc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-11-29 03:21:51 +00:00
|
|
|
|
2023-03-30 19:17:58 +00:00
|
|
|
|
2024-10-28 19:11:05 -07:00
|
|
|
|
2022-04-21 16:35:40 -07:00
|
|
|
|
2021-11-05 13:37:31 -07:00
|
|
|
|
|
|
|
|
|
2022-09-07 04:35:35 +00:00
|
|
|
|
|
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-07-24 20:33:20 +00:00
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-09-07 04:35:35 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-09-22 17:57:40 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-05-14 15:47:09 -07:00
|
|
|
|
2025-05-14 11:41:54 -07:00
|
|
|
|
2019-05-14 15:47:09 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-05-14 11:41:54 -07:00
|
|
|
|
2024-05-01 10:26:17 -07:00
|
|
|
|
2019-05-14 15:47:09 -07:00
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
2025-05-14 11:41:53 -07:00
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
2024-05-01 10:26:16 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-07-24 20:33:20 +00:00
|
|
|
|
2019-05-14 15:47:09 -07:00
|
|
|
|
|
|
|
|
|
2025-05-14 11:41:53 -07:00
|
|
|
|
|
|
|
|
|
2025-05-14 11:41:54 -07:00
|
|
|
|
2024-10-09 17:35:50 -07:00
|
|
|
|
2025-05-14 11:41:53 -07:00
|
|
|
|
2024-10-09 17:35:50 -07:00
|
|
|
|
2025-05-14 11:41:53 -07:00
|
|
|
|
|
|
|
|
|
2019-05-14 15:47:09 -07:00
|
|
|
|
|
|
|
|
|
2025-01-24 07:35:14 +00:00
|
|
|
|
mm: memcontrol: switch to rstat
Replace the memory controller's custom hierarchical stats code with the
generic rstat infrastructure provided by the cgroup core.
The current implementation does batched upward propagation from the
write side (i.e. as stats change). The per-cpu batches introduce an
error, which is multiplied by the number of subgroups in a tree. In
systems with many CPUs and sizable cgroup trees, the error can be large
enough to confuse users (e.g. 32 batch pages * 32 CPUs * 32 subgroups
results in an error of up to 128M per stat item). This can entirely
swallow allocation bursts inside a workload that the user is expecting
to see reflected in the statistics.
In the past, we've done read-side aggregation, where a memory.stat read
would have to walk the entire subtree and add up per-cpu counts. This
became problematic with lazily-freed cgroups: we could have large
subtrees where most cgroups were entirely idle. Hence the switch to
change-driven upward propagation. Unfortunately, it needed to trade
accuracy for speed due to the write side being so hot.
Rstat combines the best of both worlds: from the write side, it cheaply
maintains a queue of cgroups that have pending changes, so that the read
side can do selective tree aggregation. This way the reported stats
will always be precise and recent as can be, while the aggregation can
skip over potentially large numbers of idle cgroups.
The way rstat works is that it implements a tree for tracking cgroups
with pending local changes, as well as a flush function that walks the
tree upwards. The controller then drives this by 1) telling rstat when
a local cgroup stat changes (e.g. mod_memcg_state) and 2) when a flush
is required to get uptodate hierarchy stats for a given subtree (e.g.
when memory.stat is read). The controller also provides a flush
callback that is called during the rstat flush walk for each cgroup and
aggregates its local per-cpu counters and propagates them upwards.
This adds a second vmstats to struct mem_cgroup (MEMCG_NR_STAT +
NR_VM_EVENT_ITEMS) to track pending subtree deltas during upward
aggregation. It removes 3 words from the per-cpu data. It eliminates
memcg_exact_page_state(), since memcg_page_state() is now exact.
[akpm@linux-foundation.org: merge fix]
[hannes@cmpxchg.org: fix a sleep in atomic section problem]
Link: https://lkml.kernel.org/r/20210315234100.64307-1-hannes@cmpxchg.org
Link: https://lkml.kernel.org/r/20210209163304.77088-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Roman Gushchin <guro@fb.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Balbir Singh <bsingharora@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:26 -07:00
|
|
|
|
2024-06-24 17:59:02 -07:00
|
|
|
|
2021-04-29 22:56:17 -07:00
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-07-24 20:33:20 +00:00
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
2021-04-29 22:56:17 -07:00
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
2021-04-29 22:56:17 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-01-24 07:35:14 +00:00
|
|
|
|
2021-04-29 22:56:17 -07:00
|
|
|
|
2025-05-14 11:41:56 -07:00
|
|
|
|
2024-04-20 16:25:05 -07:00
|
|
|
|
|
|
|
|
|
2019-05-14 15:47:09 -07:00
|
|
|
|
|
|
|
|
|
mm: memcontrol: fix recursive statistics correctness & scalabilty
Right now, when somebody needs to know the recursive memory statistics
and events of a cgroup subtree, they need to walk the entire subtree and
sum up the counters manually.
There are two issues with this:
1. When a cgroup gets deleted, its stats are lost. The state counters
should all be 0 at that point, of course, but the events are not.
When this happens, the event counters, which are supposed to be
monotonic, can go backwards in the parent cgroups.
2. During regular operation, we always have a certain number of lazily
freed cgroups sitting around that have been deleted, have no tasks,
but have a few cache pages remaining. These groups' statistics do not
change until we eventually hit memory pressure, but somebody
watching, say, memory.stat on an ancestor has to iterate those every
time.
This patch addresses both issues by introducing recursive counters at
each level that are propagated from the write side when stats change.
Upward propagation happens when the per-cpu caches spill over into the
local atomic counter. This is the same thing we do during charge and
uncharge, except that the latter uses atomic RMWs, which are more
expensive; stat changes happen at around the same rate. In a sparse
file test (page faults and reclaim at maximum CPU speed) with 5 cgroup
nesting levels, perf shows __mod_memcg_page state at ~1%.
Link: http://lkml.kernel.org/r/20190412151507.2769-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 15:47:12 -07:00
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
2025-05-14 11:41:53 -07:00
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
2024-07-24 20:33:20 +00:00
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
2019-05-14 15:47:09 -07:00
|
|
|
|
|
|
|
|
|
mm: memcontrol: fix recursive statistics correctness & scalabilty
Right now, when somebody needs to know the recursive memory statistics
and events of a cgroup subtree, they need to walk the entire subtree and
sum up the counters manually.
There are two issues with this:
1. When a cgroup gets deleted, its stats are lost. The state counters
should all be 0 at that point, of course, but the events are not.
When this happens, the event counters, which are supposed to be
monotonic, can go backwards in the parent cgroups.
2. During regular operation, we always have a certain number of lazily
freed cgroups sitting around that have been deleted, have no tasks,
but have a few cache pages remaining. These groups' statistics do not
change until we eventually hit memory pressure, but somebody
watching, say, memory.stat on an ancestor has to iterate those every
time.
This patch addresses both issues by introducing recursive counters at
each level that are propagated from the write side when stats change.
Upward propagation happens when the per-cpu caches spill over into the
local atomic counter. This is the same thing we do during charge and
uncharge, except that the latter uses atomic RMWs, which are more
expensive; stat changes happen at around the same rate. In a sparse
file test (page faults and reclaim at maximum CPU speed) with 5 cgroup
nesting levels, perf shows __mod_memcg_page state at ~1%.
Link: http://lkml.kernel.org/r/20190412151507.2769-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 15:47:12 -07:00
|
|
|
|
2019-05-14 15:47:09 -07:00
|
|
|
|
2025-05-14 11:41:53 -07:00
|
|
|
|
2022-03-22 14:40:41 -07:00
|
|
|
|
2019-05-14 15:47:09 -07:00
|
|
|
|
2025-05-14 11:41:56 -07:00
|
|
|
|
2019-05-14 15:47:09 -07:00
|
|
|
|
mm, memcg: partially revert "mm/memcontrol.c: keep local VM counters in sync with the hierarchical ones"
Commit 766a4c19d880 ("mm/memcontrol.c: keep local VM counters in sync
with the hierarchical ones") effectively decreased the precision of
per-memcg vmstats_local and per-memcg-per-node lruvec percpu counters.
That's good for displaying in memory.stat, but brings a serious
regression into the reclaim process.
One issue I've discovered and debugged is the following:
lruvec_lru_size() can return 0 instead of the actual number of pages in
the lru list, preventing the kernel to reclaim last remaining pages.
Result is yet another dying memory cgroups flooding. The opposite is
also happening: scanning an empty lru list is the waste of cpu time.
Also, inactive_list_is_low() can return incorrect values, preventing the
active lru from being scanned and freed. It can fail both because the
size of active and inactive lists are inaccurate, and because the number
of workingset refaults isn't precise. In other words, the result is
pretty random.
I'm not sure, if using the approximate number of slab pages in
count_shadow_number() is acceptable, but issues described above are
enough to partially revert the patch.
Let's keep per-memcg vmstat_local batched (they are only used for
displaying stats to the userspace), but keep lruvec stats precise. This
change fixes the dead memcg flooding on my setup.
Link: http://lkml.kernel.org/r/20190817004726.2530670-1-guro@fb.com
Fixes: 766a4c19d880 ("mm/memcontrol.c: keep local VM counters in sync with the hierarchical ones")
Signed-off-by: Roman Gushchin <guro@fb.com>
Acked-by: Yafang Shao <laoar.shao@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-08-30 16:04:39 -07:00
|
|
|
|
2025-05-14 11:41:56 -07:00
|
|
|
|
2021-11-05 13:37:31 -07:00
|
|
|
|
2024-10-09 17:35:50 -07:00
|
|
|
|
2025-05-14 11:41:53 -07:00
|
|
|
|
2024-10-09 17:35:50 -07:00
|
|
|
|
2025-05-14 11:41:53 -07:00
|
|
|
|
|
|
|
|
|
2019-05-14 15:47:09 -07:00
|
|
|
|
|
|
|
|
|
mm: memcg: factor out memcg- and lruvec-level changes out of __mod_lruvec_state()
Patch series "The new cgroup slab memory controller", v7.
The patchset moves the accounting from the page level to the object level.
It allows to share slab pages between memory cgroups. This leads to a
significant win in the slab utilization (up to 45%) and the corresponding
drop in the total kernel memory footprint. The reduced number of
unmovable slab pages should also have a positive effect on the memory
fragmentation.
The patchset makes the slab accounting code simpler: there is no more need
in the complicated dynamic creation and destruction of per-cgroup slab
caches, all memory cgroups use a global set of shared slab caches. The
lifetime of slab caches is not more connected to the lifetime of memory
cgroups.
The more precise accounting does require more CPU, however in practice the
difference seems to be negligible. We've been using the new slab
controller in Facebook production for several months with different
workloads and haven't seen any noticeable regressions. What we've seen
were memory savings in order of 1 GB per host (it varied heavily depending
on the actual workload, size of RAM, number of CPUs, memory pressure,
etc).
The third version of the patchset added yet another step towards the
simplification of the code: sharing of slab caches between accounted and
non-accounted allocations. It comes with significant upsides (most
noticeable, a complete elimination of dynamic slab caches creation) but
not without some regression risks, so this change sits on top of the
patchset and is not completely merged in. So in the unlikely event of a
noticeable performance regression it can be reverted separately.
The slab memory accounting works in exactly the same way for SLAB and
SLUB. With both allocators the new controller shows significant memory
savings, with SLUB the difference is bigger. On my 16-core desktop
machine running Fedora 32 the size of the slab memory measured after the
start of the system was lower by 58% and 38% with SLUB and SLAB
correspondingly.
As an estimation of a potential CPU overhead, below are results of
slab_bulk_test01 test, kindly provided by Jesper D. Brouer. He also
helped with the evaluation of results.
The test can be found here: https://github.com/netoptimizer/prototype-kernel/
The smallest number in each row should be picked for a comparison.
SLUB-patched - bulk-API
- SLUB-patched : bulk_quick_reuse objects=1 : 187 - 90 - 224 cycles(tsc)
- SLUB-patched : bulk_quick_reuse objects=2 : 110 - 53 - 133 cycles(tsc)
- SLUB-patched : bulk_quick_reuse objects=3 : 88 - 95 - 42 cycles(tsc)
- SLUB-patched : bulk_quick_reuse objects=4 : 91 - 85 - 36 cycles(tsc)
- SLUB-patched : bulk_quick_reuse objects=8 : 32 - 66 - 32 cycles(tsc)
SLUB-original - bulk-API
- SLUB-original: bulk_quick_reuse objects=1 : 87 - 87 - 142 cycles(tsc)
- SLUB-original: bulk_quick_reuse objects=2 : 52 - 53 - 53 cycles(tsc)
- SLUB-original: bulk_quick_reuse objects=3 : 42 - 42 - 91 cycles(tsc)
- SLUB-original: bulk_quick_reuse objects=4 : 91 - 37 - 37 cycles(tsc)
- SLUB-original: bulk_quick_reuse objects=8 : 31 - 79 - 76 cycles(tsc)
SLAB-patched - bulk-API
- SLAB-patched : bulk_quick_reuse objects=1 : 67 - 67 - 140 cycles(tsc)
- SLAB-patched : bulk_quick_reuse objects=2 : 55 - 46 - 46 cycles(tsc)
- SLAB-patched : bulk_quick_reuse objects=3 : 93 - 94 - 39 cycles(tsc)
- SLAB-patched : bulk_quick_reuse objects=4 : 35 - 88 - 85 cycles(tsc)
- SLAB-patched : bulk_quick_reuse objects=8 : 30 - 30 - 30 cycles(tsc)
SLAB-original- bulk-API
- SLAB-original: bulk_quick_reuse objects=1 : 143 - 136 - 67 cycles(tsc)
- SLAB-original: bulk_quick_reuse objects=2 : 45 - 46 - 46 cycles(tsc)
- SLAB-original: bulk_quick_reuse objects=3 : 38 - 39 - 39 cycles(tsc)
- SLAB-original: bulk_quick_reuse objects=4 : 35 - 87 - 87 cycles(tsc)
- SLAB-original: bulk_quick_reuse objects=8 : 29 - 66 - 30 cycles(tsc)
This patch (of 19):
To convert memcg and lruvec slab counters to bytes there must be a way to
change these counters without touching node counters. Factor out
__mod_memcg_lruvec_state() out of __mod_lruvec_state().
Signed-off-by: Roman Gushchin <guro@fb.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Link: http://lkml.kernel.org/r/20200623174037.3951353-1-guro@fb.com
Link: http://lkml.kernel.org/r/20200623174037.3951353-2-guro@fb.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-06 23:20:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-05-14 11:41:56 -07:00
|
|
|
|
mm: memcg: factor out memcg- and lruvec-level changes out of __mod_lruvec_state()
Patch series "The new cgroup slab memory controller", v7.
The patchset moves the accounting from the page level to the object level.
It allows to share slab pages between memory cgroups. This leads to a
significant win in the slab utilization (up to 45%) and the corresponding
drop in the total kernel memory footprint. The reduced number of
unmovable slab pages should also have a positive effect on the memory
fragmentation.
The patchset makes the slab accounting code simpler: there is no more need
in the complicated dynamic creation and destruction of per-cgroup slab
caches, all memory cgroups use a global set of shared slab caches. The
lifetime of slab caches is not more connected to the lifetime of memory
cgroups.
The more precise accounting does require more CPU, however in practice the
difference seems to be negligible. We've been using the new slab
controller in Facebook production for several months with different
workloads and haven't seen any noticeable regressions. What we've seen
were memory savings in order of 1 GB per host (it varied heavily depending
on the actual workload, size of RAM, number of CPUs, memory pressure,
etc).
The third version of the patchset added yet another step towards the
simplification of the code: sharing of slab caches between accounted and
non-accounted allocations. It comes with significant upsides (most
noticeable, a complete elimination of dynamic slab caches creation) but
not without some regression risks, so this change sits on top of the
patchset and is not completely merged in. So in the unlikely event of a
noticeable performance regression it can be reverted separately.
The slab memory accounting works in exactly the same way for SLAB and
SLUB. With both allocators the new controller shows significant memory
savings, with SLUB the difference is bigger. On my 16-core desktop
machine running Fedora 32 the size of the slab memory measured after the
start of the system was lower by 58% and 38% with SLUB and SLAB
correspondingly.
As an estimation of a potential CPU overhead, below are results of
slab_bulk_test01 test, kindly provided by Jesper D. Brouer. He also
helped with the evaluation of results.
The test can be found here: https://github.com/netoptimizer/prototype-kernel/
The smallest number in each row should be picked for a comparison.
SLUB-patched - bulk-API
- SLUB-patched : bulk_quick_reuse objects=1 : 187 - 90 - 224 cycles(tsc)
- SLUB-patched : bulk_quick_reuse objects=2 : 110 - 53 - 133 cycles(tsc)
- SLUB-patched : bulk_quick_reuse objects=3 : 88 - 95 - 42 cycles(tsc)
- SLUB-patched : bulk_quick_reuse objects=4 : 91 - 85 - 36 cycles(tsc)
- SLUB-patched : bulk_quick_reuse objects=8 : 32 - 66 - 32 cycles(tsc)
SLUB-original - bulk-API
- SLUB-original: bulk_quick_reuse objects=1 : 87 - 87 - 142 cycles(tsc)
- SLUB-original: bulk_quick_reuse objects=2 : 52 - 53 - 53 cycles(tsc)
- SLUB-original: bulk_quick_reuse objects=3 : 42 - 42 - 91 cycles(tsc)
- SLUB-original: bulk_quick_reuse objects=4 : 91 - 37 - 37 cycles(tsc)
- SLUB-original: bulk_quick_reuse objects=8 : 31 - 79 - 76 cycles(tsc)
SLAB-patched - bulk-API
- SLAB-patched : bulk_quick_reuse objects=1 : 67 - 67 - 140 cycles(tsc)
- SLAB-patched : bulk_quick_reuse objects=2 : 55 - 46 - 46 cycles(tsc)
- SLAB-patched : bulk_quick_reuse objects=3 : 93 - 94 - 39 cycles(tsc)
- SLAB-patched : bulk_quick_reuse objects=4 : 35 - 88 - 85 cycles(tsc)
- SLAB-patched : bulk_quick_reuse objects=8 : 30 - 30 - 30 cycles(tsc)
SLAB-original- bulk-API
- SLAB-original: bulk_quick_reuse objects=1 : 143 - 136 - 67 cycles(tsc)
- SLAB-original: bulk_quick_reuse objects=2 : 45 - 46 - 46 cycles(tsc)
- SLAB-original: bulk_quick_reuse objects=3 : 38 - 39 - 39 cycles(tsc)
- SLAB-original: bulk_quick_reuse objects=4 : 35 - 87 - 87 cycles(tsc)
- SLAB-original: bulk_quick_reuse objects=8 : 29 - 66 - 30 cycles(tsc)
This patch (of 19):
To convert memcg and lruvec slab counters to bytes there must be a way to
change these counters without touching node counters. Factor out
__mod_memcg_lruvec_state() out of __mod_lruvec_state().
Signed-off-by: Roman Gushchin <guro@fb.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Link: http://lkml.kernel.org/r/20200623174037.3951353-1-guro@fb.com
Link: http://lkml.kernel.org/r/20200623174037.3951353-2-guro@fb.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-06 23:20:32 -07:00
|
|
|
|
|
|
|
|
|
2023-12-28 08:57:48 +00:00
|
|
|
|
2020-12-14 19:07:14 -08:00
|
|
|
|
|
|
|
|
|
mm: memcontrol: use obj_cgroup APIs to charge kmem pages
Since Roman's series "The new cgroup slab memory controller" applied.
All slab objects are charged via the new APIs of obj_cgroup. The new
APIs introduce a struct obj_cgroup to charge slab objects. It prevents
long-living objects from pinning the original memory cgroup in the
memory. But there are still some corner objects (e.g. allocations
larger than order-1 page on SLUB) which are not charged via the new
APIs. Those objects (include the pages which are allocated from buddy
allocator directly) are charged as kmem pages which still hold a
reference to the memory cgroup.
We want to reuse the obj_cgroup APIs to charge the kmem pages. If we do
that, we should store an object cgroup pointer to page->memcg_data for
the kmem pages.
Finally, page->memcg_data will have 3 different meanings.
1) For the slab pages, page->memcg_data points to an object cgroups
vector.
2) For the kmem pages (exclude the slab pages), page->memcg_data
points to an object cgroup.
3) For the user pages (e.g. the LRU pages), page->memcg_data points
to a memory cgroup.
We do not change the behavior of page_memcg() and page_memcg_rcu(). They
are also suitable for LRU pages and kmem pages. Why?
Because memory allocations pinning memcgs for a long time - it exists at a
larger scale and is causing recurring problems in the real world: page
cache doesn't get reclaimed for a long time, or is used by the second,
third, fourth, ... instance of the same job that was restarted into a new
cgroup every time. Unreclaimable dying cgroups pile up, waste memory, and
make page reclaim very inefficient.
We can convert LRU pages and most other raw memcg pins to the objcg
direction to fix this problem, and then the page->memcg will always point
to an object cgroup pointer. At that time, LRU pages and kmem pages will
be treated the same. The implementation of page_memcg() will remove the
kmem page check.
This patch aims to charge the kmem pages by using the new APIs of
obj_cgroup. Finally, the page->memcg_data of the kmem page points to an
object cgroup. We can use the __page_objcg() to get the object cgroup
associated with a kmem page. Or we can use page_memcg() to get the memory
cgroup associated with a kmem page, but caller must ensure that the
returned memcg won't be released (e.g. acquire the rcu_read_lock or
css_set_lock).
Link: https://lkml.kernel.org/r/20210401030141.37061-1-songmuchun@bytedance.com
Link: https://lkml.kernel.org/r/20210319163821.20704-6-songmuchun@bytedance.com
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Roman Gushchin <guro@fb.com>
Reviewed-by: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Xiongchun Duan <duanxiongchun@bytedance.com>
Cc: Christian Borntraeger <borntraeger@de.ibm.com>
[songmuchun@bytedance.com: fix forget to obtain the ref to objcg in split_page_memcg]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:52 -07:00
|
|
|
|
2023-12-28 08:57:48 +00:00
|
|
|
|
2020-12-14 19:07:14 -08:00
|
|
|
|
|
|
|
|
|
mm: memcontrol: use obj_cgroup APIs to charge kmem pages
Since Roman's series "The new cgroup slab memory controller" applied.
All slab objects are charged via the new APIs of obj_cgroup. The new
APIs introduce a struct obj_cgroup to charge slab objects. It prevents
long-living objects from pinning the original memory cgroup in the
memory. But there are still some corner objects (e.g. allocations
larger than order-1 page on SLUB) which are not charged via the new
APIs. Those objects (include the pages which are allocated from buddy
allocator directly) are charged as kmem pages which still hold a
reference to the memory cgroup.
We want to reuse the obj_cgroup APIs to charge the kmem pages. If we do
that, we should store an object cgroup pointer to page->memcg_data for
the kmem pages.
Finally, page->memcg_data will have 3 different meanings.
1) For the slab pages, page->memcg_data points to an object cgroups
vector.
2) For the kmem pages (exclude the slab pages), page->memcg_data
points to an object cgroup.
3) For the user pages (e.g. the LRU pages), page->memcg_data points
to a memory cgroup.
We do not change the behavior of page_memcg() and page_memcg_rcu(). They
are also suitable for LRU pages and kmem pages. Why?
Because memory allocations pinning memcgs for a long time - it exists at a
larger scale and is causing recurring problems in the real world: page
cache doesn't get reclaimed for a long time, or is used by the second,
third, fourth, ... instance of the same job that was restarted into a new
cgroup every time. Unreclaimable dying cgroups pile up, waste memory, and
make page reclaim very inefficient.
We can convert LRU pages and most other raw memcg pins to the objcg
direction to fix this problem, and then the page->memcg will always point
to an object cgroup pointer. At that time, LRU pages and kmem pages will
be treated the same. The implementation of page_memcg() will remove the
kmem page check.
This patch aims to charge the kmem pages by using the new APIs of
obj_cgroup. Finally, the page->memcg_data of the kmem page points to an
object cgroup. We can use the __page_objcg() to get the object cgroup
associated with a kmem page. Or we can use page_memcg() to get the memory
cgroup associated with a kmem page, but caller must ensure that the
returned memcg won't be released (e.g. acquire the rcu_read_lock or
css_set_lock).
Link: https://lkml.kernel.org/r/20210401030141.37061-1-songmuchun@bytedance.com
Link: https://lkml.kernel.org/r/20210319163821.20704-6-songmuchun@bytedance.com
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Roman Gushchin <guro@fb.com>
Reviewed-by: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Xiongchun Duan <duanxiongchun@bytedance.com>
Cc: Christian Borntraeger <borntraeger@de.ibm.com>
[songmuchun@bytedance.com: fix forget to obtain the ref to objcg in split_page_memcg]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:52 -07:00
|
|
|
|
2023-12-28 08:57:48 +00:00
|
|
|
|
2020-12-14 19:07:14 -08:00
|
|
|
|
2020-12-15 13:22:29 -08:00
|
|
|
|
mm: memcontrol: use obj_cgroup APIs to charge kmem pages
Since Roman's series "The new cgroup slab memory controller" applied.
All slab objects are charged via the new APIs of obj_cgroup. The new
APIs introduce a struct obj_cgroup to charge slab objects. It prevents
long-living objects from pinning the original memory cgroup in the
memory. But there are still some corner objects (e.g. allocations
larger than order-1 page on SLUB) which are not charged via the new
APIs. Those objects (include the pages which are allocated from buddy
allocator directly) are charged as kmem pages which still hold a
reference to the memory cgroup.
We want to reuse the obj_cgroup APIs to charge the kmem pages. If we do
that, we should store an object cgroup pointer to page->memcg_data for
the kmem pages.
Finally, page->memcg_data will have 3 different meanings.
1) For the slab pages, page->memcg_data points to an object cgroups
vector.
2) For the kmem pages (exclude the slab pages), page->memcg_data
points to an object cgroup.
3) For the user pages (e.g. the LRU pages), page->memcg_data points
to a memory cgroup.
We do not change the behavior of page_memcg() and page_memcg_rcu(). They
are also suitable for LRU pages and kmem pages. Why?
Because memory allocations pinning memcgs for a long time - it exists at a
larger scale and is causing recurring problems in the real world: page
cache doesn't get reclaimed for a long time, or is used by the second,
third, fourth, ... instance of the same job that was restarted into a new
cgroup every time. Unreclaimable dying cgroups pile up, waste memory, and
make page reclaim very inefficient.
We can convert LRU pages and most other raw memcg pins to the objcg
direction to fix this problem, and then the page->memcg will always point
to an object cgroup pointer. At that time, LRU pages and kmem pages will
be treated the same. The implementation of page_memcg() will remove the
kmem page check.
This patch aims to charge the kmem pages by using the new APIs of
obj_cgroup. Finally, the page->memcg_data of the kmem page points to an
object cgroup. We can use the __page_objcg() to get the object cgroup
associated with a kmem page. Or we can use page_memcg() to get the memory
cgroup associated with a kmem page, but caller must ensure that the
returned memcg won't be released (e.g. acquire the rcu_read_lock or
css_set_lock).
Link: https://lkml.kernel.org/r/20210401030141.37061-1-songmuchun@bytedance.com
Link: https://lkml.kernel.org/r/20210319163821.20704-6-songmuchun@bytedance.com
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Roman Gushchin <guro@fb.com>
Reviewed-by: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Xiongchun Duan <duanxiongchun@bytedance.com>
Cc: Christian Borntraeger <borntraeger@de.ibm.com>
[songmuchun@bytedance.com: fix forget to obtain the ref to objcg in split_page_memcg]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:52 -07:00
|
|
|
|
2020-12-14 19:07:14 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-12-15 13:22:29 -08:00
|
|
|
|
2020-12-14 19:07:14 -08:00
|
|
|
|
mm: memcontrol: use obj_cgroup APIs to charge kmem pages
Since Roman's series "The new cgroup slab memory controller" applied.
All slab objects are charged via the new APIs of obj_cgroup. The new
APIs introduce a struct obj_cgroup to charge slab objects. It prevents
long-living objects from pinning the original memory cgroup in the
memory. But there are still some corner objects (e.g. allocations
larger than order-1 page on SLUB) which are not charged via the new
APIs. Those objects (include the pages which are allocated from buddy
allocator directly) are charged as kmem pages which still hold a
reference to the memory cgroup.
We want to reuse the obj_cgroup APIs to charge the kmem pages. If we do
that, we should store an object cgroup pointer to page->memcg_data for
the kmem pages.
Finally, page->memcg_data will have 3 different meanings.
1) For the slab pages, page->memcg_data points to an object cgroups
vector.
2) For the kmem pages (exclude the slab pages), page->memcg_data
points to an object cgroup.
3) For the user pages (e.g. the LRU pages), page->memcg_data points
to a memory cgroup.
We do not change the behavior of page_memcg() and page_memcg_rcu(). They
are also suitable for LRU pages and kmem pages. Why?
Because memory allocations pinning memcgs for a long time - it exists at a
larger scale and is causing recurring problems in the real world: page
cache doesn't get reclaimed for a long time, or is used by the second,
third, fourth, ... instance of the same job that was restarted into a new
cgroup every time. Unreclaimable dying cgroups pile up, waste memory, and
make page reclaim very inefficient.
We can convert LRU pages and most other raw memcg pins to the objcg
direction to fix this problem, and then the page->memcg will always point
to an object cgroup pointer. At that time, LRU pages and kmem pages will
be treated the same. The implementation of page_memcg() will remove the
kmem page check.
This patch aims to charge the kmem pages by using the new APIs of
obj_cgroup. Finally, the page->memcg_data of the kmem page points to an
object cgroup. We can use the __page_objcg() to get the object cgroup
associated with a kmem page. Or we can use page_memcg() to get the memory
cgroup associated with a kmem page, but caller must ensure that the
returned memcg won't be released (e.g. acquire the rcu_read_lock or
css_set_lock).
Link: https://lkml.kernel.org/r/20210401030141.37061-1-songmuchun@bytedance.com
Link: https://lkml.kernel.org/r/20210319163821.20704-6-songmuchun@bytedance.com
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Roman Gushchin <guro@fb.com>
Reviewed-by: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Xiongchun Duan <duanxiongchun@bytedance.com>
Cc: Christian Borntraeger <borntraeger@de.ibm.com>
[songmuchun@bytedance.com: fix forget to obtain the ref to objcg in split_page_memcg]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:52 -07:00
|
|
|
|
2020-12-14 19:07:14 -08:00
|
|
|
|
2023-12-28 08:57:48 +00:00
|
|
|
|
2020-12-14 19:07:14 -08:00
|
|
|
|
2020-12-14 19:07:04 -08:00
|
|
|
|
2019-08-13 15:37:41 -07:00
|
|
|
|
2020-04-01 21:06:36 -07:00
|
|
|
|
2019-08-13 15:37:41 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-06-10 11:03:10 -07:00
|
|
|
|
2019-08-13 15:37:41 -07:00
|
|
|
|
2020-11-21 22:17:12 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-08-13 15:37:41 -07:00
|
|
|
|
|
|
|
|
|
2019-11-30 17:55:34 -08:00
|
|
|
|
2019-08-13 15:37:41 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-05-14 15:47:09 -07:00
|
|
|
|
2025-05-14 11:41:55 -07:00
|
|
|
|
2019-05-14 15:47:09 -07:00
|
|
|
|
|
|
|
|
|
2021-05-06 18:06:47 -07:00
|
|
|
|
2019-05-14 15:47:09 -07:00
|
|
|
|
2025-05-14 11:41:55 -07:00
|
|
|
|
2019-05-14 15:47:09 -07:00
|
|
|
|
|
|
|
|
|
2024-05-01 10:26:16 -07:00
|
|
|
|
2025-05-14 11:41:53 -07:00
|
|
|
|
2022-09-07 04:35:37 +00:00
|
|
|
|
2024-05-01 10:26:16 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-07-24 20:33:20 +00:00
|
|
|
|
2019-05-14 15:47:09 -07:00
|
|
|
|
|
|
|
|
|
2025-05-14 11:41:53 -07:00
|
|
|
|
|
|
|
|
|
2025-05-14 11:41:55 -07:00
|
|
|
|
2025-05-14 11:41:53 -07:00
|
|
|
|
2024-10-09 17:35:50 -07:00
|
|
|
|
2025-05-14 11:41:53 -07:00
|
|
|
|
|
|
|
|
|
2019-05-14 15:47:09 -07:00
|
|
|
|
|
|
|
|
|
2024-06-24 17:59:02 -07:00
|
|
|
|
2011-03-23 16:42:37 -07:00
|
|
|
|
2024-05-01 10:26:16 -07:00
|
|
|
|
2022-09-07 04:35:37 +00:00
|
|
|
|
2024-07-24 20:33:20 +00:00
|
|
|
|
2022-09-07 04:35:37 +00:00
|
|
|
|
2024-05-01 10:26:16 -07:00
|
|
|
|
|
|
|
|
|
2011-03-23 16:42:37 -07:00
|
|
|
|
|
|
|
|
|
2025-01-24 07:35:14 +00:00
|
|
|
|
2024-06-24 17:59:02 -07:00
|
|
|
|
mm: memcontrol: fix recursive statistics correctness & scalabilty
Right now, when somebody needs to know the recursive memory statistics
and events of a cgroup subtree, they need to walk the entire subtree and
sum up the counters manually.
There are two issues with this:
1. When a cgroup gets deleted, its stats are lost. The state counters
should all be 0 at that point, of course, but the events are not.
When this happens, the event counters, which are supposed to be
monotonic, can go backwards in the parent cgroups.
2. During regular operation, we always have a certain number of lazily
freed cgroups sitting around that have been deleted, have no tasks,
but have a few cache pages remaining. These groups' statistics do not
change until we eventually hit memory pressure, but somebody
watching, say, memory.stat on an ancestor has to iterate those every
time.
This patch addresses both issues by introducing recursive counters at
each level that are propagated from the write side when stats change.
Upward propagation happens when the per-cpu caches spill over into the
local atomic counter. This is the same thing we do during charge and
uncharge, except that the latter uses atomic RMWs, which are more
expensive; stat changes happen at around the same rate. In a sparse
file test (page faults and reclaim at maximum CPU speed) with 5 cgroup
nesting levels, perf shows __mod_memcg_page state at ~1%.
Link: http://lkml.kernel.org/r/20190412151507.2769-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 15:47:12 -07:00
|
|
|
|
2024-05-01 10:26:16 -07:00
|
|
|
|
2022-09-07 04:35:37 +00:00
|
|
|
|
2024-07-24 20:33:20 +00:00
|
|
|
|
2022-09-07 04:35:37 +00:00
|
|
|
|
mm: memcontrol: don't batch updates of local VM stats and events
The kernel test robot noticed a 26% will-it-scale pagefault regression
from commit 42a300353577 ("mm: memcontrol: fix recursive statistics
correctness & scalabilty"). This appears to be caused by bouncing the
additional cachelines from the new hierarchical statistics counters.
We can fix this by getting rid of the batched local counters instead.
Originally, there were *only* group-local counters, and they were fully
maintained per cpu. A reader of a stats file high up in the cgroup tree
would have to walk the entire subtree and collect each level's per-cpu
counters to get the recursive view. This was prohibitively expensive,
and so we switched to per-cpu batched updates of the local counters
during a983b5ebee57 ("mm: memcontrol: fix excessive complexity in
memory.stat reporting"), reducing the complexity from nr_subgroups *
nr_cpus to nr_subgroups.
With growing machines and cgroup trees, the tree walk itself became too
expensive for monitoring top-level groups, and this is when the culprit
patch added hierarchy counters on each cgroup level. When the per-cpu
batch size would be reached, both the local and the hierarchy counters
would get batch-updated from the per-cpu delta simultaneously.
This makes local and hierarchical counter reads blazingly fast, but it
unfortunately makes the write-side too cache line intense.
Since local counter reads were never a problem - we only centralized
them to accelerate the hierarchy walk - and use of the local counters
are becoming rarer due to replacement with hierarchical views (ongoing
rework in the page reclaim and workingset code), we can make those local
counters unbatched per-cpu counters again.
The scheme will then be as such:
when a memcg statistic changes, the writer will:
- update the local counter (per-cpu)
- update the batch counter (per-cpu). If the batch is full:
- spill the batch into the group's atomic_t
- spill the batch into all ancestors' atomic_ts
- empty out the batch counter (per-cpu)
when a local memcg counter is read, the reader will:
- collect the local counter from all cpus
when a hiearchy memcg counter is read, the reader will:
- read the atomic_t
We might be able to simplify this further and make the recursive
counters unbatched per-cpu counters as well (batch upward propagation,
but leave per-cpu collection to the readers), but that will require a
more in-depth analysis and testing of all the callsites. Deal with the
immediate regression for now.
Link: http://lkml.kernel.org/r/20190521151647.GB2870@cmpxchg.org
Fixes: 42a300353577 ("mm: memcontrol: fix recursive statistics correctness & scalabilty")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reported-by: kernel test robot <rong.a.chen@intel.com>
Tested-by: kernel test robot <rong.a.chen@intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Roman Gushchin <guro@fb.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-06-13 15:55:46 -07:00
|
|
|
|
2024-05-01 10:26:16 -07:00
|
|
|
|
mm: memcontrol: fix recursive statistics correctness & scalabilty
Right now, when somebody needs to know the recursive memory statistics
and events of a cgroup subtree, they need to walk the entire subtree and
sum up the counters manually.
There are two issues with this:
1. When a cgroup gets deleted, its stats are lost. The state counters
should all be 0 at that point, of course, but the events are not.
When this happens, the event counters, which are supposed to be
monotonic, can go backwards in the parent cgroups.
2. During regular operation, we always have a certain number of lazily
freed cgroups sitting around that have been deleted, have no tasks,
but have a few cache pages remaining. These groups' statistics do not
change until we eventually hit memory pressure, but somebody
watching, say, memory.stat on an ancestor has to iterate those every
time.
This patch addresses both issues by introducing recursive counters at
each level that are propagated from the write side when stats change.
Upward propagation happens when the per-cpu caches spill over into the
local atomic counter. This is the same thing we do during charge and
uncharge, except that the latter uses atomic RMWs, which are more
expensive; stat changes happen at around the same rate. In a sparse
file test (page faults and reclaim at maximum CPU speed) with 5 cgroup
nesting levels, perf shows __mod_memcg_page state at ~1%.
Link: http://lkml.kernel.org/r/20190412151507.2769-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 15:47:12 -07:00
|
|
|
|
2025-01-24 07:35:14 +00:00
|
|
|
|
mm: memcontrol: fix recursive statistics correctness & scalabilty
Right now, when somebody needs to know the recursive memory statistics
and events of a cgroup subtree, they need to walk the entire subtree and
sum up the counters manually.
There are two issues with this:
1. When a cgroup gets deleted, its stats are lost. The state counters
should all be 0 at that point, of course, but the events are not.
When this happens, the event counters, which are supposed to be
monotonic, can go backwards in the parent cgroups.
2. During regular operation, we always have a certain number of lazily
freed cgroups sitting around that have been deleted, have no tasks,
but have a few cache pages remaining. These groups' statistics do not
change until we eventually hit memory pressure, but somebody
watching, say, memory.stat on an ancestor has to iterate those every
time.
This patch addresses both issues by introducing recursive counters at
each level that are propagated from the write side when stats change.
Upward propagation happens when the per-cpu caches spill over into the
local atomic counter. This is the same thing we do during charge and
uncharge, except that the latter uses atomic RMWs, which are more
expensive; stat changes happen at around the same rate. In a sparse
file test (page faults and reclaim at maximum CPU speed) with 5 cgroup
nesting levels, perf shows __mod_memcg_page state at ~1%.
Link: http://lkml.kernel.org/r/20190412151507.2769-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 15:47:12 -07:00
|
|
|
|
cgroups: add an owner to the mm_struct
Remove the mem_cgroup member from mm_struct and instead adds an owner.
This approach was suggested by Paul Menage. The advantage of this approach
is that, once the mm->owner is known, using the subsystem id, the cgroup
can be determined. It also allows several control groups that are
virtually grouped by mm_struct, to exist independent of the memory
controller i.e., without adding mem_cgroup's for each controller, to
mm_struct.
A new config option CONFIG_MM_OWNER is added and the memory resource
controller selects this config option.
This patch also adds cgroup callbacks to notify subsystems when mm->owner
changes. The mm_cgroup_changed callback is called with the task_lock() of
the new task held and is called just prior to changing the mm->owner.
I am indebted to Paul Menage for the several reviews of this patchset and
helping me make it lighter and simpler.
This patch was tested on a powerpc box, it was compiled with both the
MM_OWNER config turned on and off.
After the thread group leader exits, it's moved to init_css_state by
cgroup_exit(), thus all future charges from runnings threads would be
redirected to the init_css_set's subsystem.
Signed-off-by: Balbir Singh <balbir@linux.vnet.ibm.com>
Cc: Pavel Emelianov <xemul@openvz.org>
Cc: Hugh Dickins <hugh@veritas.com>
Cc: Sudhir Kumar <skumar@linux.vnet.ibm.com>
Cc: YAMAMOTO Takashi <yamamoto@valinux.co.jp>
Cc: Hirokazu Takahashi <taka@valinux.co.jp>
Cc: David Rientjes <rientjes@google.com>,
Cc: Balbir Singh <balbir@linux.vnet.ibm.com>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Acked-by: Pekka Enberg <penberg@cs.helsinki.fi>
Reviewed-by: Paul Menage <menage@google.com>
Cc: Oleg Nesterov <oleg@tv-sign.ru>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-29 01:00:16 -07:00
|
|
|
|
2008-02-07 00:13:51 -08:00
|
|
|
|
mm owner: fix race between swapoff and exit
There's a race between mm->owner assignment and swapoff, more easily
seen when task slab poisoning is turned on. The condition occurs when
try_to_unuse() runs in parallel with an exiting task. A similar race
can occur with callers of get_task_mm(), such as /proc/<pid>/<mmstats>
or ptrace or page migration.
CPU0 CPU1
try_to_unuse
looks at mm = task0->mm
increments mm->mm_users
task 0 exits
mm->owner needs to be updated, but no
new owner is found (mm_users > 1, but
no other task has task->mm = task0->mm)
mm_update_next_owner() leaves
mmput(mm) decrements mm->mm_users
task0 freed
dereferencing mm->owner fails
The fix is to notify the subsystem via mm_owner_changed callback(),
if no new owner is found, by specifying the new task as NULL.
Jiri Slaby:
mm->owner was set to NULL prior to calling cgroup_mm_owner_callbacks(), but
must be set after that, so as not to pass NULL as old owner causing oops.
Daisuke Nishimura:
mm_update_next_owner() may set mm->owner to NULL, but mem_cgroup_from_task()
and its callers need to take account of this situation to avoid oops.
Hugh Dickins:
Lockdep warning and hang below exec_mmap() when testing these patches.
exit_mm() up_reads mmap_sem before calling mm_update_next_owner(),
so exec_mmap() now needs to do the same. And with that repositioning,
there's now no point in mm_need_new_owner() allowing for NULL mm.
Reported-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Balbir Singh <balbir@linux.vnet.ibm.com>
Signed-off-by: Jiri Slaby <jirislaby@gmail.com>
Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Paul Menage <menage@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-09-28 23:09:31 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-02-08 10:36:58 -05:00
|
|
|
|
2008-02-07 00:13:51 -08:00
|
|
|
|
2015-09-08 15:01:02 -07:00
|
|
|
|
2008-02-07 00:13:51 -08:00
|
|
|
|
2021-06-28 19:38:18 -07:00
|
|
|
|
|
|
|
|
|
2021-09-02 14:55:49 -07:00
|
|
|
|
2021-06-28 19:38:18 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
fs: fsnotify: account fsnotify metadata to kmemcg
Patch series "Directed kmem charging", v8.
The Linux kernel's memory cgroup allows limiting the memory usage of the
jobs running on the system to provide isolation between the jobs. All
the kernel memory allocated in the context of the job and marked with
__GFP_ACCOUNT will also be included in the memory usage and be limited
by the job's limit.
The kernel memory can only be charged to the memcg of the process in
whose context kernel memory was allocated. However there are cases
where the allocated kernel memory should be charged to the memcg
different from the current processes's memcg. This patch series
contains two such concrete use-cases i.e. fsnotify and buffer_head.
The fsnotify event objects can consume a lot of system memory for large
or unlimited queues if there is either no or slow listener. The events
are allocated in the context of the event producer. However they should
be charged to the event consumer. Similarly the buffer_head objects can
be allocated in a memcg different from the memcg of the page for which
buffer_head objects are being allocated.
To solve this issue, this patch series introduces mechanism to charge
kernel memory to a given memcg. In case of fsnotify events, the memcg
of the consumer can be used for charging and for buffer_head, the memcg
of the page can be charged. For directed charging, the caller can use
the scope API memalloc_[un]use_memcg() to specify the memcg to charge
for all the __GFP_ACCOUNT allocations within the scope.
This patch (of 2):
A lot of memory can be consumed by the events generated for the huge or
unlimited queues if there is either no or slow listener. This can cause
system level memory pressure or OOMs. So, it's better to account the
fsnotify kmem caches to the memcg of the listener.
However the listener can be in a different memcg than the memcg of the
producer and these allocations happen in the context of the event
producer. This patch introduces remote memcg charging API which the
producer can use to charge the allocations to the memcg of the listener.
There are seven fsnotify kmem caches and among them allocations from
dnotify_struct_cache, dnotify_mark_cache, fanotify_mark_cache and
inotify_inode_mark_cachep happens in the context of syscall from the
listener. So, SLAB_ACCOUNT is enough for these caches.
The objects from fsnotify_mark_connector_cachep are not accounted as
they are small compared to the notification mark or events and it is
unclear whom to account connector to since it is shared by all events
attached to the inode.
The allocations from the event caches happen in the context of the event
producer. For such caches we will need to remote charge the allocations
to the listener's memcg. Thus we save the memcg reference in the
fsnotify_group structure of the listener.
This patch has also moved the members of fsnotify_group to keep the size
same, at least for 64 bit build, even with additional member by filling
the holes.
[shakeelb@google.com: use GFP_KERNEL_ACCOUNT rather than open-coding it]
Link: http://lkml.kernel.org/r/20180702215439.211597-1-shakeelb@google.com
Link: http://lkml.kernel.org/r/20180627191250.209150-2-shakeelb@google.com
Signed-off-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Jan Kara <jack@suse.cz>
Cc: Amir Goldstein <amir73il@gmail.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-17 15:46:39 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-28 19:38:18 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
fs: fsnotify: account fsnotify metadata to kmemcg
Patch series "Directed kmem charging", v8.
The Linux kernel's memory cgroup allows limiting the memory usage of the
jobs running on the system to provide isolation between the jobs. All
the kernel memory allocated in the context of the job and marked with
__GFP_ACCOUNT will also be included in the memory usage and be limited
by the job's limit.
The kernel memory can only be charged to the memcg of the process in
whose context kernel memory was allocated. However there are cases
where the allocated kernel memory should be charged to the memcg
different from the current processes's memcg. This patch series
contains two such concrete use-cases i.e. fsnotify and buffer_head.
The fsnotify event objects can consume a lot of system memory for large
or unlimited queues if there is either no or slow listener. The events
are allocated in the context of the event producer. However they should
be charged to the event consumer. Similarly the buffer_head objects can
be allocated in a memcg different from the memcg of the page for which
buffer_head objects are being allocated.
To solve this issue, this patch series introduces mechanism to charge
kernel memory to a given memcg. In case of fsnotify events, the memcg
of the consumer can be used for charging and for buffer_head, the memcg
of the page can be charged. For directed charging, the caller can use
the scope API memalloc_[un]use_memcg() to specify the memcg to charge
for all the __GFP_ACCOUNT allocations within the scope.
This patch (of 2):
A lot of memory can be consumed by the events generated for the huge or
unlimited queues if there is either no or slow listener. This can cause
system level memory pressure or OOMs. So, it's better to account the
fsnotify kmem caches to the memcg of the listener.
However the listener can be in a different memcg than the memcg of the
producer and these allocations happen in the context of the event
producer. This patch introduces remote memcg charging API which the
producer can use to charge the allocations to the memcg of the listener.
There are seven fsnotify kmem caches and among them allocations from
dnotify_struct_cache, dnotify_mark_cache, fanotify_mark_cache and
inotify_inode_mark_cachep happens in the context of syscall from the
listener. So, SLAB_ACCOUNT is enough for these caches.
The objects from fsnotify_mark_connector_cachep are not accounted as
they are small compared to the notification mark or events and it is
unclear whom to account connector to since it is shared by all events
attached to the inode.
The allocations from the event caches happen in the context of the event
producer. For such caches we will need to remote charge the allocations
to the listener's memcg. Thus we save the memcg reference in the
fsnotify_group structure of the listener.
This patch has also moved the members of fsnotify_group to keep the size
same, at least for 64 bit build, even with additional member by filling
the holes.
[shakeelb@google.com: use GFP_KERNEL_ACCOUNT rather than open-coding it]
Link: http://lkml.kernel.org/r/20180702215439.211597-1-shakeelb@google.com
Link: http://lkml.kernel.org/r/20180627191250.209150-2-shakeelb@google.com
Signed-off-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Jan Kara <jack@suse.cz>
Cc: Amir Goldstein <amir73il@gmail.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-17 15:46:39 -07:00
|
|
|
|
|
|
|
|
|
2009-01-07 18:08:33 -08:00
|
|
|
|
fs: fsnotify: account fsnotify metadata to kmemcg
Patch series "Directed kmem charging", v8.
The Linux kernel's memory cgroup allows limiting the memory usage of the
jobs running on the system to provide isolation between the jobs. All
the kernel memory allocated in the context of the job and marked with
__GFP_ACCOUNT will also be included in the memory usage and be limited
by the job's limit.
The kernel memory can only be charged to the memcg of the process in
whose context kernel memory was allocated. However there are cases
where the allocated kernel memory should be charged to the memcg
different from the current processes's memcg. This patch series
contains two such concrete use-cases i.e. fsnotify and buffer_head.
The fsnotify event objects can consume a lot of system memory for large
or unlimited queues if there is either no or slow listener. The events
are allocated in the context of the event producer. However they should
be charged to the event consumer. Similarly the buffer_head objects can
be allocated in a memcg different from the memcg of the page for which
buffer_head objects are being allocated.
To solve this issue, this patch series introduces mechanism to charge
kernel memory to a given memcg. In case of fsnotify events, the memcg
of the consumer can be used for charging and for buffer_head, the memcg
of the page can be charged. For directed charging, the caller can use
the scope API memalloc_[un]use_memcg() to specify the memcg to charge
for all the __GFP_ACCOUNT allocations within the scope.
This patch (of 2):
A lot of memory can be consumed by the events generated for the huge or
unlimited queues if there is either no or slow listener. This can cause
system level memory pressure or OOMs. So, it's better to account the
fsnotify kmem caches to the memcg of the listener.
However the listener can be in a different memcg than the memcg of the
producer and these allocations happen in the context of the event
producer. This patch introduces remote memcg charging API which the
producer can use to charge the allocations to the memcg of the listener.
There are seven fsnotify kmem caches and among them allocations from
dnotify_struct_cache, dnotify_mark_cache, fanotify_mark_cache and
inotify_inode_mark_cachep happens in the context of syscall from the
listener. So, SLAB_ACCOUNT is enough for these caches.
The objects from fsnotify_mark_connector_cachep are not accounted as
they are small compared to the notification mark or events and it is
unclear whom to account connector to since it is shared by all events
attached to the inode.
The allocations from the event caches happen in the context of the event
producer. For such caches we will need to remote charge the allocations
to the listener's memcg. Thus we save the memcg reference in the
fsnotify_group structure of the listener.
This patch has also moved the members of fsnotify_group to keep the size
same, at least for 64 bit build, even with additional member by filling
the holes.
[shakeelb@google.com: use GFP_KERNEL_ACCOUNT rather than open-coding it]
Link: http://lkml.kernel.org/r/20180702215439.211597-1-shakeelb@google.com
Link: http://lkml.kernel.org/r/20180627191250.209150-2-shakeelb@google.com
Signed-off-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Jan Kara <jack@suse.cz>
Cc: Amir Goldstein <amir73il@gmail.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-17 15:46:39 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
memcg: fix OOM killer under memcg
This patch tries to fix OOM Killer problems caused by hierarchy.
Now, memcg itself has OOM KILL function (in oom_kill.c) and tries to
kill a task in memcg.
But, when hierarchy is used, it's broken and correct task cannot
be killed. For example, in following cgroup
/groupA/ hierarchy=1, limit=1G,
01 nolimit
02 nolimit
All tasks' memory usage under /groupA, /groupA/01, groupA/02 is limited to
groupA's 1Gbytes but OOM Killer just kills tasks in groupA.
This patch provides makes the bad process be selected from all tasks
under hierarchy. BTW, currently, oom_jiffies is updated against groupA
in above case. oom_jiffies of tree should be updated.
To see how oom_jiffies is used, please check mem_cgroup_oom_called()
callers.
[akpm@linux-foundation.org: build fix]
[akpm@linux-foundation.org: const fix]
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Paul Menage <menage@google.com>
Cc: Li Zefan <lizf@cn.fujitsu.com>
Cc: Balbir Singh <balbir@in.ibm.com>
Cc: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-04-02 16:57:38 -07:00
|
|
|
|
2021-06-28 19:37:50 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-28 19:38:18 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-28 19:37:50 -07:00
|
|
|
|
2009-01-07 18:08:33 -08:00
|
|
|
|
|
|
|
|
|
2021-06-28 19:37:50 -07:00
|
|
|
|
|
|
|
|
|
2014-04-07 15:37:43 -07:00
|
|
|
|
2019-11-15 17:34:43 -08:00
|
|
|
|
2009-01-07 18:08:33 -08:00
|
|
|
|
2011-11-02 13:38:15 -07:00
|
|
|
|
2009-01-07 18:08:33 -08:00
|
|
|
|
fs: fsnotify: account fsnotify metadata to kmemcg
Patch series "Directed kmem charging", v8.
The Linux kernel's memory cgroup allows limiting the memory usage of the
jobs running on the system to provide isolation between the jobs. All
the kernel memory allocated in the context of the job and marked with
__GFP_ACCOUNT will also be included in the memory usage and be limited
by the job's limit.
The kernel memory can only be charged to the memcg of the process in
whose context kernel memory was allocated. However there are cases
where the allocated kernel memory should be charged to the memcg
different from the current processes's memcg. This patch series
contains two such concrete use-cases i.e. fsnotify and buffer_head.
The fsnotify event objects can consume a lot of system memory for large
or unlimited queues if there is either no or slow listener. The events
are allocated in the context of the event producer. However they should
be charged to the event consumer. Similarly the buffer_head objects can
be allocated in a memcg different from the memcg of the page for which
buffer_head objects are being allocated.
To solve this issue, this patch series introduces mechanism to charge
kernel memory to a given memcg. In case of fsnotify events, the memcg
of the consumer can be used for charging and for buffer_head, the memcg
of the page can be charged. For directed charging, the caller can use
the scope API memalloc_[un]use_memcg() to specify the memcg to charge
for all the __GFP_ACCOUNT allocations within the scope.
This patch (of 2):
A lot of memory can be consumed by the events generated for the huge or
unlimited queues if there is either no or slow listener. This can cause
system level memory pressure or OOMs. So, it's better to account the
fsnotify kmem caches to the memcg of the listener.
However the listener can be in a different memcg than the memcg of the
producer and these allocations happen in the context of the event
producer. This patch introduces remote memcg charging API which the
producer can use to charge the allocations to the memcg of the listener.
There are seven fsnotify kmem caches and among them allocations from
dnotify_struct_cache, dnotify_mark_cache, fanotify_mark_cache and
inotify_inode_mark_cachep happens in the context of syscall from the
listener. So, SLAB_ACCOUNT is enough for these caches.
The objects from fsnotify_mark_connector_cachep are not accounted as
they are small compared to the notification mark or events and it is
unclear whom to account connector to since it is shared by all events
attached to the inode.
The allocations from the event caches happen in the context of the event
producer. For such caches we will need to remote charge the allocations
to the listener's memcg. Thus we save the memcg reference in the
fsnotify_group structure of the listener.
This patch has also moved the members of fsnotify_group to keep the size
same, at least for 64 bit build, even with additional member by filling
the holes.
[shakeelb@google.com: use GFP_KERNEL_ACCOUNT rather than open-coding it]
Link: http://lkml.kernel.org/r/20180702215439.211597-1-shakeelb@google.com
Link: http://lkml.kernel.org/r/20180627191250.209150-2-shakeelb@google.com
Signed-off-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Jan Kara <jack@suse.cz>
Cc: Amir Goldstein <amir73il@gmail.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-17 15:46:39 -07:00
|
|
|
|
|
|
|
|
|
memcontrol: add helpers for hugetlb memcg accounting
Patch series "hugetlb memcg accounting", v4.
Currently, hugetlb memory usage is not acounted for in the memory
controller, which could lead to memory overprotection for cgroups with
hugetlb-backed memory. This has been observed in our production system.
For instance, here is one of our usecases: suppose there are two 32G
containers. The machine is booted with hugetlb_cma=6G, and each container
may or may not use up to 3 gigantic page, depending on the workload within
it. The rest is anon, cache, slab, etc. We can set the hugetlb cgroup
limit of each cgroup to 3G to enforce hugetlb fairness. But it is very
difficult to configure memory.max to keep overall consumption, including
anon, cache, slab etcetera fair.
What we have had to resort to is to constantly poll hugetlb usage and
readjust memory.max. Similar procedure is done to other memory limits
(memory.low for e.g). However, this is rather cumbersome and buggy.
Furthermore, when there is a delay in memory limits correction, (for e.g
when hugetlb usage changes within consecutive runs of the userspace
agent), the system could be in an over/underprotected state.
This patch series rectifies this issue by charging the memcg when the
hugetlb folio is allocated, and uncharging when the folio is freed. In
addition, a new selftest is added to demonstrate and verify this new
behavior.
This patch (of 4):
This patch exposes charge committing and cancelling as parts of the memory
controller interface. These functionalities are useful when the
try_charge() and commit_charge() stages have to be separated by other
actions in between (which can fail). One such example is the new hugetlb
accounting behavior in the following patch.
The patch also adds a helper function to obtain a reference to the
current task's memcg.
Link: https://lkml.kernel.org/r/20231006184629.155543-1-nphamcs@gmail.com
Link: https://lkml.kernel.org/r/20231006184629.155543-2-nphamcs@gmail.com
Signed-off-by: Nhat Pham <nphamcs@gmail.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Frank van der Linden <fvdl@google.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Rik van Riel <riel@surriel.com>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Tejun heo <tj@kernel.org>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-06 11:46:26 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm,memcg: provide per-cgroup counters for NUMA balancing operations
The ability to observe the demotion and promotion decisions made by the
kernel on a per-cgroup basis is important for monitoring and tuning
containerized workloads on machines equipped with tiered memory.
Different containers in the system may experience drastically different
memory tiering actions that cannot be distinguished from the global
counters alone.
For example, a container running a workload that has a much hotter memory
accesses will likely see more promotions and fewer demotions, potentially
depriving a colocated container of top tier memory to such an extent that
its performance degrades unacceptably.
For another example, some containers may exhibit longer periods between
data reuse, causing much more numa_hint_faults than numa_pages_migrated.
In this case, tuning hot_threshold_ms may be appropriate, but the signal
can easily be lost if only global counters are available.
In the long term, we hope to introduce per-cgroup control of promotion and
demotion actions to implement memory placement policies in tiering.
This patch set adds seven counters to memory.stat in a cgroup:
numa_pages_migrated, numa_pte_updates, numa_hint_faults, pgdemote_kswapd,
pgdemote_khugepaged, pgdemote_direct and pgpromote_success. pgdemote_*
and pgpromote_success are also available in memory.numa_stat.
count_memcg_events_mm() is added to count multiple event occurrences at
once, and get_mem_cgroup_from_folio() is added because we need to get a
reference to the memcg of a folio before it's migrated to track
numa_pages_migrated. The accounting of PGDEMOTE_* is moved to
shrink_inactive_list() before being changed to per-cgroup.
[kaiyang2@cs.cmu.edu: add documentation of the memcg counters in cgroup-v2.rst]
Link: https://lkml.kernel.org/r/20240814235122.252309-1-kaiyang2@cs.cmu.edu
Link: https://lkml.kernel.org/r/20240814174227.30639-1-kaiyang2@cs.cmu.edu
Signed-off-by: Kaiyang Zhao <kaiyang2@cs.cmu.edu>
Cc: David Rientjes <rientjes@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Wei Xu <weixugc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-08-14 17:42:27 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-01-12 17:17:59 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-10-13 16:52:45 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-01-12 17:17:59 -08:00
|
|
|
|
2013-09-24 15:27:37 -07:00
|
|
|
|
2012-01-12 17:17:59 -08:00
|
|
|
|
2013-09-24 15:27:37 -07:00
|
|
|
|
2009-04-02 16:57:35 -07:00
|
|
|
|
treewide: Remove uninitialized_var() usage
Using uninitialized_var() is dangerous as it papers over real bugs[1]
(or can in the future), and suppresses unrelated compiler warnings
(e.g. "unused variable"). If the compiler thinks it is uninitialized,
either simply initialize the variable or make compiler changes.
In preparation for removing[2] the[3] macro[4], remove all remaining
needless uses with the following script:
git grep '\buninitialized_var\b' | cut -d: -f1 | sort -u | \
xargs perl -pi -e \
's/\buninitialized_var\(([^\)]+)\)/\1/g;
s:\s*/\* (GCC be quiet|to make compiler happy) \*/$::g;'
drivers/video/fbdev/riva/riva_hw.c was manually tweaked to avoid
pathological white-space.
No outstanding warnings were found building allmodconfig with GCC 9.3.0
for x86_64, i386, arm64, arm, powerpc, powerpc64le, s390x, mips, sparc64,
alpha, and m68k.
[1] https://lore.kernel.org/lkml/20200603174714.192027-1-glider@google.com/
[2] https://lore.kernel.org/lkml/CA+55aFw+Vbj0i=1TGqCR5vQkCzWJ0QxK6CernOU6eedsudAixw@mail.gmail.com/
[3] https://lore.kernel.org/lkml/CA+55aFwgbgqhbp1fkxvRKEpzyR5J8n1vKT1VZdz9knmPuXhOeg@mail.gmail.com/
[4] https://lore.kernel.org/lkml/CA+55aFz2500WfbKXAx8s67wrm9=yVJu65TpLgN_ybYNv0VEOKA@mail.gmail.com/
Reviewed-by: Leon Romanovsky <leonro@mellanox.com> # drivers/infiniband and mlx4/mlx5
Acked-by: Jason Gunthorpe <jgg@mellanox.com> # IB
Acked-by: Kalle Valo <kvalo@codeaurora.org> # wireless drivers
Reviewed-by: Chao Yu <yuchao0@huawei.com> # erofs
Signed-off-by: Kees Cook <keescook@chromium.org>
2020-06-03 13:09:38 -07:00
|
|
|
|
2024-09-05 00:30:53 +00:00
|
|
|
|
2024-09-05 00:30:54 +00:00
|
|
|
|
|
|
|
|
|
2010-10-27 15:33:42 -07:00
|
|
|
|
2013-09-24 15:27:37 -07:00
|
|
|
|
|
|
|
|
|
2012-01-12 17:17:59 -08:00
|
|
|
|
mm: memcg: consolidate hierarchy iteration primitives
The memcg naturalization series:
Memory control groups are currently bolted onto the side of
traditional memory management in places where better integration would
be preferrable. To reclaim memory, for example, memory control groups
maintain their own LRU list and reclaim strategy aside from the global
per-zone LRU list reclaim. But an extra list head for each existing
page frame is expensive and maintaining it requires additional code.
This patchset disables the global per-zone LRU lists on memory cgroup
configurations and converts all its users to operate on the per-memory
cgroup lists instead. As LRU pages are then exclusively on one list,
this saves two list pointers for each page frame in the system:
page_cgroup array size with 4G physical memory
vanilla: allocated 31457280 bytes of page_cgroup
patched: allocated 15728640 bytes of page_cgroup
At the same time, system performance for various workloads is
unaffected:
100G sparse file cat, 4G physical memory, 10 runs, to test for code
bloat in the traditional LRU handling and kswapd & direct reclaim
paths, without/with the memory controller configured in
vanilla: 71.603(0.207) seconds
patched: 71.640(0.156) seconds
vanilla: 79.558(0.288) seconds
patched: 77.233(0.147) seconds
100G sparse file cat in 1G memory cgroup, 10 runs, to test for code
bloat in the traditional memory cgroup LRU handling and reclaim path
vanilla: 96.844(0.281) seconds
patched: 94.454(0.311) seconds
4 unlimited memcgs running kbuild -j32 each, 4G physical memory, 500M
swap on SSD, 10 runs, to test for regressions in kswapd & direct
reclaim using per-memcg LRU lists with multiple memcgs and multiple
allocators within each memcg
vanilla: 717.722(1.440) seconds [ 69720.100(11600.835) majfaults ]
patched: 714.106(2.313) seconds [ 71109.300(14886.186) majfaults ]
16 unlimited memcgs running kbuild, 1900M hierarchical limit, 500M
swap on SSD, 10 runs, to test for regressions in hierarchical memcg
setups
vanilla: 2742.058(1.992) seconds [ 26479.600(1736.737) majfaults ]
patched: 2743.267(1.214) seconds [ 27240.700(1076.063) majfaults ]
This patch:
There are currently two different implementations of iterating over a
memory cgroup hierarchy tree.
Consolidate them into one worker function and base the convenience
looping-macros on top of it.
Signed-off-by: Johannes Weiner <jweiner@redhat.com>
Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Reviewed-by: Michal Hocko <mhocko@suse.cz>
Reviewed-by: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Cc: Balbir Singh <bsingharora@gmail.com>
Cc: Ying Han <yinghan@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-12 17:17:48 -08:00
|
|
|
|
|
|
|
|
|
2010-10-27 15:33:41 -07:00
|
|
|
|
2013-04-29 15:07:15 -07:00
|
|
|
|
2024-09-05 00:30:52 +00:00
|
|
|
|
2024-09-05 00:30:54 +00:00
|
|
|
|
2013-04-29 15:07:17 -07:00
|
|
|
|
2014-12-10 15:42:39 -08:00
|
|
|
|
2024-09-05 00:30:53 +00:00
|
|
|
|
2024-09-05 00:30:54 +00:00
|
|
|
|
2014-12-10 15:42:39 -08:00
|
|
|
|
2024-09-05 00:30:54 +00:00
|
|
|
|
2024-09-05 00:30:53 +00:00
|
|
|
|
2014-12-10 15:42:39 -08:00
|
|
|
|
2022-04-28 23:15:59 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-09-05 00:30:53 +00:00
|
|
|
|
|
|
|
|
|
2014-12-10 15:42:39 -08:00
|
|
|
|
|
|
|
|
|
2024-09-05 00:30:51 +00:00
|
|
|
|
2024-09-05 00:30:54 +00:00
|
|
|
|
2022-04-28 23:15:58 -07:00
|
|
|
|
2014-12-10 15:42:39 -08:00
|
|
|
|
2024-09-05 00:30:53 +00:00
|
|
|
|
2010-10-27 15:33:41 -07:00
|
|
|
|
2024-09-05 00:30:54 +00:00
|
|
|
|
2014-12-10 15:42:39 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-09-05 00:30:54 +00:00
|
|
|
|
2016-01-20 15:02:53 -08:00
|
|
|
|
mm: memcg: consolidate hierarchy iteration primitives
The memcg naturalization series:
Memory control groups are currently bolted onto the side of
traditional memory management in places where better integration would
be preferrable. To reclaim memory, for example, memory control groups
maintain their own LRU list and reclaim strategy aside from the global
per-zone LRU list reclaim. But an extra list head for each existing
page frame is expensive and maintaining it requires additional code.
This patchset disables the global per-zone LRU lists on memory cgroup
configurations and converts all its users to operate on the per-memory
cgroup lists instead. As LRU pages are then exclusively on one list,
this saves two list pointers for each page frame in the system:
page_cgroup array size with 4G physical memory
vanilla: allocated 31457280 bytes of page_cgroup
patched: allocated 15728640 bytes of page_cgroup
At the same time, system performance for various workloads is
unaffected:
100G sparse file cat, 4G physical memory, 10 runs, to test for code
bloat in the traditional LRU handling and kswapd & direct reclaim
paths, without/with the memory controller configured in
vanilla: 71.603(0.207) seconds
patched: 71.640(0.156) seconds
vanilla: 79.558(0.288) seconds
patched: 77.233(0.147) seconds
100G sparse file cat in 1G memory cgroup, 10 runs, to test for code
bloat in the traditional memory cgroup LRU handling and reclaim path
vanilla: 96.844(0.281) seconds
patched: 94.454(0.311) seconds
4 unlimited memcgs running kbuild -j32 each, 4G physical memory, 500M
swap on SSD, 10 runs, to test for regressions in kswapd & direct
reclaim using per-memcg LRU lists with multiple memcgs and multiple
allocators within each memcg
vanilla: 717.722(1.440) seconds [ 69720.100(11600.835) majfaults ]
patched: 714.106(2.313) seconds [ 71109.300(14886.186) majfaults ]
16 unlimited memcgs running kbuild, 1900M hierarchical limit, 500M
swap on SSD, 10 runs, to test for regressions in hierarchical memcg
setups
vanilla: 2742.058(1.992) seconds [ 26479.600(1736.737) majfaults ]
patched: 2743.267(1.214) seconds [ 27240.700(1076.063) majfaults ]
This patch:
There are currently two different implementations of iterating over a
memory cgroup hierarchy tree.
Consolidate them into one worker function and base the convenience
looping-macros on top of it.
Signed-off-by: Johannes Weiner <jweiner@redhat.com>
Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Reviewed-by: Michal Hocko <mhocko@suse.cz>
Reviewed-by: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Cc: Balbir Singh <bsingharora@gmail.com>
Cc: Ying Han <yinghan@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-12 17:17:48 -08:00
|
|
|
|
2014-12-10 15:42:39 -08:00
|
|
|
|
2024-09-05 00:30:54 +00:00
|
|
|
|
2024-09-05 00:30:53 +00:00
|
|
|
|
2014-12-10 15:42:39 -08:00
|
|
|
|
|
|
|
|
|
2015-12-29 14:54:10 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-12-10 15:42:39 -08:00
|
|
|
|
2024-09-05 00:30:54 +00:00
|
|
|
|
2024-09-05 00:30:53 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-12-29 14:54:10 -08:00
|
|
|
|
2024-09-05 00:30:54 +00:00
|
|
|
|
2024-09-05 00:30:53 +00:00
|
|
|
|
2014-12-10 15:42:39 -08:00
|
|
|
|
2024-09-05 00:30:52 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcg: consolidate hierarchy iteration primitives
The memcg naturalization series:
Memory control groups are currently bolted onto the side of
traditional memory management in places where better integration would
be preferrable. To reclaim memory, for example, memory control groups
maintain their own LRU list and reclaim strategy aside from the global
per-zone LRU list reclaim. But an extra list head for each existing
page frame is expensive and maintaining it requires additional code.
This patchset disables the global per-zone LRU lists on memory cgroup
configurations and converts all its users to operate on the per-memory
cgroup lists instead. As LRU pages are then exclusively on one list,
this saves two list pointers for each page frame in the system:
page_cgroup array size with 4G physical memory
vanilla: allocated 31457280 bytes of page_cgroup
patched: allocated 15728640 bytes of page_cgroup
At the same time, system performance for various workloads is
unaffected:
100G sparse file cat, 4G physical memory, 10 runs, to test for code
bloat in the traditional LRU handling and kswapd & direct reclaim
paths, without/with the memory controller configured in
vanilla: 71.603(0.207) seconds
patched: 71.640(0.156) seconds
vanilla: 79.558(0.288) seconds
patched: 77.233(0.147) seconds
100G sparse file cat in 1G memory cgroup, 10 runs, to test for code
bloat in the traditional memory cgroup LRU handling and reclaim path
vanilla: 96.844(0.281) seconds
patched: 94.454(0.311) seconds
4 unlimited memcgs running kbuild -j32 each, 4G physical memory, 500M
swap on SSD, 10 runs, to test for regressions in kswapd & direct
reclaim using per-memcg LRU lists with multiple memcgs and multiple
allocators within each memcg
vanilla: 717.722(1.440) seconds [ 69720.100(11600.835) majfaults ]
patched: 714.106(2.313) seconds [ 71109.300(14886.186) majfaults ]
16 unlimited memcgs running kbuild, 1900M hierarchical limit, 500M
swap on SSD, 10 runs, to test for regressions in hierarchical memcg
setups
vanilla: 2742.058(1.992) seconds [ 26479.600(1736.737) majfaults ]
patched: 2743.267(1.214) seconds [ 27240.700(1076.063) majfaults ]
This patch:
There are currently two different implementations of iterating over a
memory cgroup hierarchy tree.
Consolidate them into one worker function and base the convenience
looping-macros on top of it.
Signed-off-by: Johannes Weiner <jweiner@redhat.com>
Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Reviewed-by: Michal Hocko <mhocko@suse.cz>
Reviewed-by: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Cc: Balbir Singh <bsingharora@gmail.com>
Cc: Ying Han <yinghan@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-12 17:17:48 -08:00
|
|
|
|
2014-12-10 15:42:39 -08:00
|
|
|
|
2013-04-29 15:07:15 -07:00
|
|
|
|
|
|
|
|
|
2013-04-29 15:07:14 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-09-05 00:30:54 +00:00
|
|
|
|
2009-04-02 16:57:35 -07:00
|
|
|
|
2010-10-27 15:33:41 -07:00
|
|
|
|
2012-01-12 17:17:59 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcg: consolidate hierarchy iteration primitives
The memcg naturalization series:
Memory control groups are currently bolted onto the side of
traditional memory management in places where better integration would
be preferrable. To reclaim memory, for example, memory control groups
maintain their own LRU list and reclaim strategy aside from the global
per-zone LRU list reclaim. But an extra list head for each existing
page frame is expensive and maintaining it requires additional code.
This patchset disables the global per-zone LRU lists on memory cgroup
configurations and converts all its users to operate on the per-memory
cgroup lists instead. As LRU pages are then exclusively on one list,
this saves two list pointers for each page frame in the system:
page_cgroup array size with 4G physical memory
vanilla: allocated 31457280 bytes of page_cgroup
patched: allocated 15728640 bytes of page_cgroup
At the same time, system performance for various workloads is
unaffected:
100G sparse file cat, 4G physical memory, 10 runs, to test for code
bloat in the traditional LRU handling and kswapd & direct reclaim
paths, without/with the memory controller configured in
vanilla: 71.603(0.207) seconds
patched: 71.640(0.156) seconds
vanilla: 79.558(0.288) seconds
patched: 77.233(0.147) seconds
100G sparse file cat in 1G memory cgroup, 10 runs, to test for code
bloat in the traditional memory cgroup LRU handling and reclaim path
vanilla: 96.844(0.281) seconds
patched: 94.454(0.311) seconds
4 unlimited memcgs running kbuild -j32 each, 4G physical memory, 500M
swap on SSD, 10 runs, to test for regressions in kswapd & direct
reclaim using per-memcg LRU lists with multiple memcgs and multiple
allocators within each memcg
vanilla: 717.722(1.440) seconds [ 69720.100(11600.835) majfaults ]
patched: 714.106(2.313) seconds [ 71109.300(14886.186) majfaults ]
16 unlimited memcgs running kbuild, 1900M hierarchical limit, 500M
swap on SSD, 10 runs, to test for regressions in hierarchical memcg
setups
vanilla: 2742.058(1.992) seconds [ 26479.600(1736.737) majfaults ]
patched: 2743.267(1.214) seconds [ 27240.700(1076.063) majfaults ]
This patch:
There are currently two different implementations of iterating over a
memory cgroup hierarchy tree.
Consolidate them into one worker function and base the convenience
looping-macros on top of it.
Signed-off-by: Johannes Weiner <jweiner@redhat.com>
Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Reviewed-by: Michal Hocko <mhocko@suse.cz>
Reviewed-by: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Cc: Balbir Singh <bsingharora@gmail.com>
Cc: Ying Han <yinghan@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-12 17:17:48 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-10-27 15:33:41 -07:00
|
|
|
|
mm/memcontrol.c: fix use after free in mem_cgroup_iter()
This patch is sent to report an use after free in mem_cgroup_iter()
after merging commit be2657752e9e ("mm: memcg: fix use after free in
mem_cgroup_iter()").
I work with android kernel tree (4.9 & 4.14), and commit be2657752e9e
("mm: memcg: fix use after free in mem_cgroup_iter()") has been merged
to the trees. However, I can still observe use after free issues
addressed in the commit be2657752e9e. (on low-end devices, a few times
this month)
backtrace:
css_tryget <- crash here
mem_cgroup_iter
shrink_node
shrink_zones
do_try_to_free_pages
try_to_free_pages
__perform_reclaim
__alloc_pages_direct_reclaim
__alloc_pages_slowpath
__alloc_pages_nodemask
To debug, I poisoned mem_cgroup before freeing it:
static void __mem_cgroup_free(struct mem_cgroup *memcg)
for_each_node(node)
free_mem_cgroup_per_node_info(memcg, node);
free_percpu(memcg->stat);
+ /* poison memcg before freeing it */
+ memset(memcg, 0x78, sizeof(struct mem_cgroup));
kfree(memcg);
}
The coredump shows the position=0xdbbc2a00 is freed.
(gdb) p/x ((struct mem_cgroup_per_node *)0xe5009e00)->iter[8]
$13 = {position = 0xdbbc2a00, generation = 0x2efd}
0xdbbc2a00: 0xdbbc2e00 0x00000000 0xdbbc2800 0x00000100
0xdbbc2a10: 0x00000200 0x78787878 0x00026218 0x00000000
0xdbbc2a20: 0xdcad6000 0x00000001 0x78787800 0x00000000
0xdbbc2a30: 0x78780000 0x00000000 0x0068fb84 0x78787878
0xdbbc2a40: 0x78787878 0x78787878 0x78787878 0xe3fa5cc0
0xdbbc2a50: 0x78787878 0x78787878 0x00000000 0x00000000
0xdbbc2a60: 0x00000000 0x00000000 0x00000000 0x00000000
0xdbbc2a70: 0x00000000 0x00000000 0x00000000 0x00000000
0xdbbc2a80: 0x00000000 0x00000000 0x00000000 0x00000000
0xdbbc2a90: 0x00000001 0x00000000 0x00000000 0x00100000
0xdbbc2aa0: 0x00000001 0xdbbc2ac8 0x00000000 0x00000000
0xdbbc2ab0: 0x00000000 0x00000000 0x00000000 0x00000000
0xdbbc2ac0: 0x00000000 0x00000000 0xe5b02618 0x00001000
0xdbbc2ad0: 0x00000000 0x78787878 0x78787878 0x78787878
0xdbbc2ae0: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2af0: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b00: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b10: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b20: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b30: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b40: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b50: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b60: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b70: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b80: 0x78787878 0x78787878 0x00000000 0x78787878
0xdbbc2b90: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2ba0: 0x78787878 0x78787878 0x78787878 0x78787878
In the reclaim path, try_to_free_pages() does not setup
sc.target_mem_cgroup and sc is passed to do_try_to_free_pages(), ...,
shrink_node().
In mem_cgroup_iter(), root is set to root_mem_cgroup because
sc->target_mem_cgroup is NULL. It is possible to assign a memcg to
root_mem_cgroup.nodeinfo.iter in mem_cgroup_iter().
try_to_free_pages
struct scan_control sc = {...}, target_mem_cgroup is 0x0;
do_try_to_free_pages
shrink_zones
shrink_node
mem_cgroup *root = sc->target_mem_cgroup;
memcg = mem_cgroup_iter(root, NULL, &reclaim);
mem_cgroup_iter()
if (!root)
root = root_mem_cgroup;
...
css = css_next_descendant_pre(css, &root->css);
memcg = mem_cgroup_from_css(css);
cmpxchg(&iter->position, pos, memcg);
My device uses memcg non-hierarchical mode. When we release a memcg:
invalidate_reclaim_iterators() reaches only dead_memcg and its parents.
If non-hierarchical mode is used, invalidate_reclaim_iterators() never
reaches root_mem_cgroup.
static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
{
struct mem_cgroup *memcg = dead_memcg;
for (; memcg; memcg = parent_mem_cgroup(memcg)
...
}
So the use after free scenario looks like:
CPU1 CPU2
try_to_free_pages
do_try_to_free_pages
shrink_zones
shrink_node
mem_cgroup_iter()
if (!root)
root = root_mem_cgroup;
...
css = css_next_descendant_pre(css, &root->css);
memcg = mem_cgroup_from_css(css);
cmpxchg(&iter->position, pos, memcg);
invalidate_reclaim_iterators(memcg);
...
__mem_cgroup_free()
kfree(memcg);
try_to_free_pages
do_try_to_free_pages
shrink_zones
shrink_node
mem_cgroup_iter()
if (!root)
root = root_mem_cgroup;
...
mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
iter = &mz->iter[reclaim->priority];
pos = READ_ONCE(iter->position);
css_tryget(&pos->css) <- use after free
To avoid this, we should also invalidate root_mem_cgroup.nodeinfo.iter
in invalidate_reclaim_iterators().
[cai@lca.pw: fix -Wparentheses compilation warning]
Link: http://lkml.kernel.org/r/1564580753-17531-1-git-send-email-cai@lca.pw
Link: http://lkml.kernel.org/r/20190730015729.4406-1-miles.chen@mediatek.com
Fixes: 5ac8fb31ad2e ("mm: memcontrol: convert reclaim iterator to simple css refcounting")
Signed-off-by: Miles Chen <miles.chen@mediatek.com>
Signed-off-by: Qian Cai <cai@lca.pw>
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>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-08-13 15:37:28 -07:00
|
|
|
|
|
|
|
|
|
2015-12-29 14:54:10 -08:00
|
|
|
|
|
|
|
|
|
2016-07-28 15:46:05 -07:00
|
|
|
|
|
|
|
|
|
2015-12-29 14:54:10 -08:00
|
|
|
|
mm/memcontrol.c: fix use after free in mem_cgroup_iter()
This patch is sent to report an use after free in mem_cgroup_iter()
after merging commit be2657752e9e ("mm: memcg: fix use after free in
mem_cgroup_iter()").
I work with android kernel tree (4.9 & 4.14), and commit be2657752e9e
("mm: memcg: fix use after free in mem_cgroup_iter()") has been merged
to the trees. However, I can still observe use after free issues
addressed in the commit be2657752e9e. (on low-end devices, a few times
this month)
backtrace:
css_tryget <- crash here
mem_cgroup_iter
shrink_node
shrink_zones
do_try_to_free_pages
try_to_free_pages
__perform_reclaim
__alloc_pages_direct_reclaim
__alloc_pages_slowpath
__alloc_pages_nodemask
To debug, I poisoned mem_cgroup before freeing it:
static void __mem_cgroup_free(struct mem_cgroup *memcg)
for_each_node(node)
free_mem_cgroup_per_node_info(memcg, node);
free_percpu(memcg->stat);
+ /* poison memcg before freeing it */
+ memset(memcg, 0x78, sizeof(struct mem_cgroup));
kfree(memcg);
}
The coredump shows the position=0xdbbc2a00 is freed.
(gdb) p/x ((struct mem_cgroup_per_node *)0xe5009e00)->iter[8]
$13 = {position = 0xdbbc2a00, generation = 0x2efd}
0xdbbc2a00: 0xdbbc2e00 0x00000000 0xdbbc2800 0x00000100
0xdbbc2a10: 0x00000200 0x78787878 0x00026218 0x00000000
0xdbbc2a20: 0xdcad6000 0x00000001 0x78787800 0x00000000
0xdbbc2a30: 0x78780000 0x00000000 0x0068fb84 0x78787878
0xdbbc2a40: 0x78787878 0x78787878 0x78787878 0xe3fa5cc0
0xdbbc2a50: 0x78787878 0x78787878 0x00000000 0x00000000
0xdbbc2a60: 0x00000000 0x00000000 0x00000000 0x00000000
0xdbbc2a70: 0x00000000 0x00000000 0x00000000 0x00000000
0xdbbc2a80: 0x00000000 0x00000000 0x00000000 0x00000000
0xdbbc2a90: 0x00000001 0x00000000 0x00000000 0x00100000
0xdbbc2aa0: 0x00000001 0xdbbc2ac8 0x00000000 0x00000000
0xdbbc2ab0: 0x00000000 0x00000000 0x00000000 0x00000000
0xdbbc2ac0: 0x00000000 0x00000000 0xe5b02618 0x00001000
0xdbbc2ad0: 0x00000000 0x78787878 0x78787878 0x78787878
0xdbbc2ae0: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2af0: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b00: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b10: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b20: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b30: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b40: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b50: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b60: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b70: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b80: 0x78787878 0x78787878 0x00000000 0x78787878
0xdbbc2b90: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2ba0: 0x78787878 0x78787878 0x78787878 0x78787878
In the reclaim path, try_to_free_pages() does not setup
sc.target_mem_cgroup and sc is passed to do_try_to_free_pages(), ...,
shrink_node().
In mem_cgroup_iter(), root is set to root_mem_cgroup because
sc->target_mem_cgroup is NULL. It is possible to assign a memcg to
root_mem_cgroup.nodeinfo.iter in mem_cgroup_iter().
try_to_free_pages
struct scan_control sc = {...}, target_mem_cgroup is 0x0;
do_try_to_free_pages
shrink_zones
shrink_node
mem_cgroup *root = sc->target_mem_cgroup;
memcg = mem_cgroup_iter(root, NULL, &reclaim);
mem_cgroup_iter()
if (!root)
root = root_mem_cgroup;
...
css = css_next_descendant_pre(css, &root->css);
memcg = mem_cgroup_from_css(css);
cmpxchg(&iter->position, pos, memcg);
My device uses memcg non-hierarchical mode. When we release a memcg:
invalidate_reclaim_iterators() reaches only dead_memcg and its parents.
If non-hierarchical mode is used, invalidate_reclaim_iterators() never
reaches root_mem_cgroup.
static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
{
struct mem_cgroup *memcg = dead_memcg;
for (; memcg; memcg = parent_mem_cgroup(memcg)
...
}
So the use after free scenario looks like:
CPU1 CPU2
try_to_free_pages
do_try_to_free_pages
shrink_zones
shrink_node
mem_cgroup_iter()
if (!root)
root = root_mem_cgroup;
...
css = css_next_descendant_pre(css, &root->css);
memcg = mem_cgroup_from_css(css);
cmpxchg(&iter->position, pos, memcg);
invalidate_reclaim_iterators(memcg);
...
__mem_cgroup_free()
kfree(memcg);
try_to_free_pages
do_try_to_free_pages
shrink_zones
shrink_node
mem_cgroup_iter()
if (!root)
root = root_mem_cgroup;
...
mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
iter = &mz->iter[reclaim->priority];
pos = READ_ONCE(iter->position);
css_tryget(&pos->css) <- use after free
To avoid this, we should also invalidate root_mem_cgroup.nodeinfo.iter
in invalidate_reclaim_iterators().
[cai@lca.pw: fix -Wparentheses compilation warning]
Link: http://lkml.kernel.org/r/1564580753-17531-1-git-send-email-cai@lca.pw
Link: http://lkml.kernel.org/r/20190730015729.4406-1-miles.chen@mediatek.com
Fixes: 5ac8fb31ad2e ("mm: memcontrol: convert reclaim iterator to simple css refcounting")
Signed-off-by: Miles Chen <miles.chen@mediatek.com>
Signed-off-by: Qian Cai <cai@lca.pw>
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>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-08-13 15:37:28 -07:00
|
|
|
|
2021-04-29 22:56:14 -07:00
|
|
|
|
2019-11-30 17:50:03 -08:00
|
|
|
|
|
|
|
|
|
2015-12-29 14:54:10 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/memcontrol.c: fix use after free in mem_cgroup_iter()
This patch is sent to report an use after free in mem_cgroup_iter()
after merging commit be2657752e9e ("mm: memcg: fix use after free in
mem_cgroup_iter()").
I work with android kernel tree (4.9 & 4.14), and commit be2657752e9e
("mm: memcg: fix use after free in mem_cgroup_iter()") has been merged
to the trees. However, I can still observe use after free issues
addressed in the commit be2657752e9e. (on low-end devices, a few times
this month)
backtrace:
css_tryget <- crash here
mem_cgroup_iter
shrink_node
shrink_zones
do_try_to_free_pages
try_to_free_pages
__perform_reclaim
__alloc_pages_direct_reclaim
__alloc_pages_slowpath
__alloc_pages_nodemask
To debug, I poisoned mem_cgroup before freeing it:
static void __mem_cgroup_free(struct mem_cgroup *memcg)
for_each_node(node)
free_mem_cgroup_per_node_info(memcg, node);
free_percpu(memcg->stat);
+ /* poison memcg before freeing it */
+ memset(memcg, 0x78, sizeof(struct mem_cgroup));
kfree(memcg);
}
The coredump shows the position=0xdbbc2a00 is freed.
(gdb) p/x ((struct mem_cgroup_per_node *)0xe5009e00)->iter[8]
$13 = {position = 0xdbbc2a00, generation = 0x2efd}
0xdbbc2a00: 0xdbbc2e00 0x00000000 0xdbbc2800 0x00000100
0xdbbc2a10: 0x00000200 0x78787878 0x00026218 0x00000000
0xdbbc2a20: 0xdcad6000 0x00000001 0x78787800 0x00000000
0xdbbc2a30: 0x78780000 0x00000000 0x0068fb84 0x78787878
0xdbbc2a40: 0x78787878 0x78787878 0x78787878 0xe3fa5cc0
0xdbbc2a50: 0x78787878 0x78787878 0x00000000 0x00000000
0xdbbc2a60: 0x00000000 0x00000000 0x00000000 0x00000000
0xdbbc2a70: 0x00000000 0x00000000 0x00000000 0x00000000
0xdbbc2a80: 0x00000000 0x00000000 0x00000000 0x00000000
0xdbbc2a90: 0x00000001 0x00000000 0x00000000 0x00100000
0xdbbc2aa0: 0x00000001 0xdbbc2ac8 0x00000000 0x00000000
0xdbbc2ab0: 0x00000000 0x00000000 0x00000000 0x00000000
0xdbbc2ac0: 0x00000000 0x00000000 0xe5b02618 0x00001000
0xdbbc2ad0: 0x00000000 0x78787878 0x78787878 0x78787878
0xdbbc2ae0: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2af0: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b00: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b10: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b20: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b30: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b40: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b50: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b60: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b70: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b80: 0x78787878 0x78787878 0x00000000 0x78787878
0xdbbc2b90: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2ba0: 0x78787878 0x78787878 0x78787878 0x78787878
In the reclaim path, try_to_free_pages() does not setup
sc.target_mem_cgroup and sc is passed to do_try_to_free_pages(), ...,
shrink_node().
In mem_cgroup_iter(), root is set to root_mem_cgroup because
sc->target_mem_cgroup is NULL. It is possible to assign a memcg to
root_mem_cgroup.nodeinfo.iter in mem_cgroup_iter().
try_to_free_pages
struct scan_control sc = {...}, target_mem_cgroup is 0x0;
do_try_to_free_pages
shrink_zones
shrink_node
mem_cgroup *root = sc->target_mem_cgroup;
memcg = mem_cgroup_iter(root, NULL, &reclaim);
mem_cgroup_iter()
if (!root)
root = root_mem_cgroup;
...
css = css_next_descendant_pre(css, &root->css);
memcg = mem_cgroup_from_css(css);
cmpxchg(&iter->position, pos, memcg);
My device uses memcg non-hierarchical mode. When we release a memcg:
invalidate_reclaim_iterators() reaches only dead_memcg and its parents.
If non-hierarchical mode is used, invalidate_reclaim_iterators() never
reaches root_mem_cgroup.
static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
{
struct mem_cgroup *memcg = dead_memcg;
for (; memcg; memcg = parent_mem_cgroup(memcg)
...
}
So the use after free scenario looks like:
CPU1 CPU2
try_to_free_pages
do_try_to_free_pages
shrink_zones
shrink_node
mem_cgroup_iter()
if (!root)
root = root_mem_cgroup;
...
css = css_next_descendant_pre(css, &root->css);
memcg = mem_cgroup_from_css(css);
cmpxchg(&iter->position, pos, memcg);
invalidate_reclaim_iterators(memcg);
...
__mem_cgroup_free()
kfree(memcg);
try_to_free_pages
do_try_to_free_pages
shrink_zones
shrink_node
mem_cgroup_iter()
if (!root)
root = root_mem_cgroup;
...
mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
iter = &mz->iter[reclaim->priority];
pos = READ_ONCE(iter->position);
css_tryget(&pos->css) <- use after free
To avoid this, we should also invalidate root_mem_cgroup.nodeinfo.iter
in invalidate_reclaim_iterators().
[cai@lca.pw: fix -Wparentheses compilation warning]
Link: http://lkml.kernel.org/r/1564580753-17531-1-git-send-email-cai@lca.pw
Link: http://lkml.kernel.org/r/20190730015729.4406-1-miles.chen@mediatek.com
Fixes: 5ac8fb31ad2e ("mm: memcontrol: convert reclaim iterator to simple css refcounting")
Signed-off-by: Miles Chen <miles.chen@mediatek.com>
Signed-off-by: Qian Cai <cai@lca.pw>
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>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-08-13 15:37:28 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-08-15 06:51:04 +00:00
|
|
|
|
mm/memcontrol.c: fix use after free in mem_cgroup_iter()
This patch is sent to report an use after free in mem_cgroup_iter()
after merging commit be2657752e9e ("mm: memcg: fix use after free in
mem_cgroup_iter()").
I work with android kernel tree (4.9 & 4.14), and commit be2657752e9e
("mm: memcg: fix use after free in mem_cgroup_iter()") has been merged
to the trees. However, I can still observe use after free issues
addressed in the commit be2657752e9e. (on low-end devices, a few times
this month)
backtrace:
css_tryget <- crash here
mem_cgroup_iter
shrink_node
shrink_zones
do_try_to_free_pages
try_to_free_pages
__perform_reclaim
__alloc_pages_direct_reclaim
__alloc_pages_slowpath
__alloc_pages_nodemask
To debug, I poisoned mem_cgroup before freeing it:
static void __mem_cgroup_free(struct mem_cgroup *memcg)
for_each_node(node)
free_mem_cgroup_per_node_info(memcg, node);
free_percpu(memcg->stat);
+ /* poison memcg before freeing it */
+ memset(memcg, 0x78, sizeof(struct mem_cgroup));
kfree(memcg);
}
The coredump shows the position=0xdbbc2a00 is freed.
(gdb) p/x ((struct mem_cgroup_per_node *)0xe5009e00)->iter[8]
$13 = {position = 0xdbbc2a00, generation = 0x2efd}
0xdbbc2a00: 0xdbbc2e00 0x00000000 0xdbbc2800 0x00000100
0xdbbc2a10: 0x00000200 0x78787878 0x00026218 0x00000000
0xdbbc2a20: 0xdcad6000 0x00000001 0x78787800 0x00000000
0xdbbc2a30: 0x78780000 0x00000000 0x0068fb84 0x78787878
0xdbbc2a40: 0x78787878 0x78787878 0x78787878 0xe3fa5cc0
0xdbbc2a50: 0x78787878 0x78787878 0x00000000 0x00000000
0xdbbc2a60: 0x00000000 0x00000000 0x00000000 0x00000000
0xdbbc2a70: 0x00000000 0x00000000 0x00000000 0x00000000
0xdbbc2a80: 0x00000000 0x00000000 0x00000000 0x00000000
0xdbbc2a90: 0x00000001 0x00000000 0x00000000 0x00100000
0xdbbc2aa0: 0x00000001 0xdbbc2ac8 0x00000000 0x00000000
0xdbbc2ab0: 0x00000000 0x00000000 0x00000000 0x00000000
0xdbbc2ac0: 0x00000000 0x00000000 0xe5b02618 0x00001000
0xdbbc2ad0: 0x00000000 0x78787878 0x78787878 0x78787878
0xdbbc2ae0: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2af0: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b00: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b10: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b20: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b30: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b40: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b50: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b60: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b70: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b80: 0x78787878 0x78787878 0x00000000 0x78787878
0xdbbc2b90: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2ba0: 0x78787878 0x78787878 0x78787878 0x78787878
In the reclaim path, try_to_free_pages() does not setup
sc.target_mem_cgroup and sc is passed to do_try_to_free_pages(), ...,
shrink_node().
In mem_cgroup_iter(), root is set to root_mem_cgroup because
sc->target_mem_cgroup is NULL. It is possible to assign a memcg to
root_mem_cgroup.nodeinfo.iter in mem_cgroup_iter().
try_to_free_pages
struct scan_control sc = {...}, target_mem_cgroup is 0x0;
do_try_to_free_pages
shrink_zones
shrink_node
mem_cgroup *root = sc->target_mem_cgroup;
memcg = mem_cgroup_iter(root, NULL, &reclaim);
mem_cgroup_iter()
if (!root)
root = root_mem_cgroup;
...
css = css_next_descendant_pre(css, &root->css);
memcg = mem_cgroup_from_css(css);
cmpxchg(&iter->position, pos, memcg);
My device uses memcg non-hierarchical mode. When we release a memcg:
invalidate_reclaim_iterators() reaches only dead_memcg and its parents.
If non-hierarchical mode is used, invalidate_reclaim_iterators() never
reaches root_mem_cgroup.
static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
{
struct mem_cgroup *memcg = dead_memcg;
for (; memcg; memcg = parent_mem_cgroup(memcg)
...
}
So the use after free scenario looks like:
CPU1 CPU2
try_to_free_pages
do_try_to_free_pages
shrink_zones
shrink_node
mem_cgroup_iter()
if (!root)
root = root_mem_cgroup;
...
css = css_next_descendant_pre(css, &root->css);
memcg = mem_cgroup_from_css(css);
cmpxchg(&iter->position, pos, memcg);
invalidate_reclaim_iterators(memcg);
...
__mem_cgroup_free()
kfree(memcg);
try_to_free_pages
do_try_to_free_pages
shrink_zones
shrink_node
mem_cgroup_iter()
if (!root)
root = root_mem_cgroup;
...
mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
iter = &mz->iter[reclaim->priority];
pos = READ_ONCE(iter->position);
css_tryget(&pos->css) <- use after free
To avoid this, we should also invalidate root_mem_cgroup.nodeinfo.iter
in invalidate_reclaim_iterators().
[cai@lca.pw: fix -Wparentheses compilation warning]
Link: http://lkml.kernel.org/r/1564580753-17531-1-git-send-email-cai@lca.pw
Link: http://lkml.kernel.org/r/20190730015729.4406-1-miles.chen@mediatek.com
Fixes: 5ac8fb31ad2e ("mm: memcontrol: convert reclaim iterator to simple css refcounting")
Signed-off-by: Miles Chen <miles.chen@mediatek.com>
Signed-off-by: Qian Cai <cai@lca.pw>
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>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-08-13 15:37:28 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-09-30 19:14:33 +05:30
|
|
|
|
mm/memcontrol.c: fix use after free in mem_cgroup_iter()
This patch is sent to report an use after free in mem_cgroup_iter()
after merging commit be2657752e9e ("mm: memcg: fix use after free in
mem_cgroup_iter()").
I work with android kernel tree (4.9 & 4.14), and commit be2657752e9e
("mm: memcg: fix use after free in mem_cgroup_iter()") has been merged
to the trees. However, I can still observe use after free issues
addressed in the commit be2657752e9e. (on low-end devices, a few times
this month)
backtrace:
css_tryget <- crash here
mem_cgroup_iter
shrink_node
shrink_zones
do_try_to_free_pages
try_to_free_pages
__perform_reclaim
__alloc_pages_direct_reclaim
__alloc_pages_slowpath
__alloc_pages_nodemask
To debug, I poisoned mem_cgroup before freeing it:
static void __mem_cgroup_free(struct mem_cgroup *memcg)
for_each_node(node)
free_mem_cgroup_per_node_info(memcg, node);
free_percpu(memcg->stat);
+ /* poison memcg before freeing it */
+ memset(memcg, 0x78, sizeof(struct mem_cgroup));
kfree(memcg);
}
The coredump shows the position=0xdbbc2a00 is freed.
(gdb) p/x ((struct mem_cgroup_per_node *)0xe5009e00)->iter[8]
$13 = {position = 0xdbbc2a00, generation = 0x2efd}
0xdbbc2a00: 0xdbbc2e00 0x00000000 0xdbbc2800 0x00000100
0xdbbc2a10: 0x00000200 0x78787878 0x00026218 0x00000000
0xdbbc2a20: 0xdcad6000 0x00000001 0x78787800 0x00000000
0xdbbc2a30: 0x78780000 0x00000000 0x0068fb84 0x78787878
0xdbbc2a40: 0x78787878 0x78787878 0x78787878 0xe3fa5cc0
0xdbbc2a50: 0x78787878 0x78787878 0x00000000 0x00000000
0xdbbc2a60: 0x00000000 0x00000000 0x00000000 0x00000000
0xdbbc2a70: 0x00000000 0x00000000 0x00000000 0x00000000
0xdbbc2a80: 0x00000000 0x00000000 0x00000000 0x00000000
0xdbbc2a90: 0x00000001 0x00000000 0x00000000 0x00100000
0xdbbc2aa0: 0x00000001 0xdbbc2ac8 0x00000000 0x00000000
0xdbbc2ab0: 0x00000000 0x00000000 0x00000000 0x00000000
0xdbbc2ac0: 0x00000000 0x00000000 0xe5b02618 0x00001000
0xdbbc2ad0: 0x00000000 0x78787878 0x78787878 0x78787878
0xdbbc2ae0: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2af0: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b00: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b10: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b20: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b30: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b40: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b50: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b60: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b70: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2b80: 0x78787878 0x78787878 0x00000000 0x78787878
0xdbbc2b90: 0x78787878 0x78787878 0x78787878 0x78787878
0xdbbc2ba0: 0x78787878 0x78787878 0x78787878 0x78787878
In the reclaim path, try_to_free_pages() does not setup
sc.target_mem_cgroup and sc is passed to do_try_to_free_pages(), ...,
shrink_node().
In mem_cgroup_iter(), root is set to root_mem_cgroup because
sc->target_mem_cgroup is NULL. It is possible to assign a memcg to
root_mem_cgroup.nodeinfo.iter in mem_cgroup_iter().
try_to_free_pages
struct scan_control sc = {...}, target_mem_cgroup is 0x0;
do_try_to_free_pages
shrink_zones
shrink_node
mem_cgroup *root = sc->target_mem_cgroup;
memcg = mem_cgroup_iter(root, NULL, &reclaim);
mem_cgroup_iter()
if (!root)
root = root_mem_cgroup;
...
css = css_next_descendant_pre(css, &root->css);
memcg = mem_cgroup_from_css(css);
cmpxchg(&iter->position, pos, memcg);
My device uses memcg non-hierarchical mode. When we release a memcg:
invalidate_reclaim_iterators() reaches only dead_memcg and its parents.
If non-hierarchical mode is used, invalidate_reclaim_iterators() never
reaches root_mem_cgroup.
static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
{
struct mem_cgroup *memcg = dead_memcg;
for (; memcg; memcg = parent_mem_cgroup(memcg)
...
}
So the use after free scenario looks like:
CPU1 CPU2
try_to_free_pages
do_try_to_free_pages
shrink_zones
shrink_node
mem_cgroup_iter()
if (!root)
root = root_mem_cgroup;
...
css = css_next_descendant_pre(css, &root->css);
memcg = mem_cgroup_from_css(css);
cmpxchg(&iter->position, pos, memcg);
invalidate_reclaim_iterators(memcg);
...
__mem_cgroup_free()
kfree(memcg);
try_to_free_pages
do_try_to_free_pages
shrink_zones
shrink_node
mem_cgroup_iter()
if (!root)
root = root_mem_cgroup;
...
mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
iter = &mz->iter[reclaim->priority];
pos = READ_ONCE(iter->position);
css_tryget(&pos->css) <- use after free
To avoid this, we should also invalidate root_mem_cgroup.nodeinfo.iter
in invalidate_reclaim_iterators().
[cai@lca.pw: fix -Wparentheses compilation warning]
Link: http://lkml.kernel.org/r/1564580753-17531-1-git-send-email-cai@lca.pw
Link: http://lkml.kernel.org/r/20190730015729.4406-1-miles.chen@mediatek.com
Fixes: 5ac8fb31ad2e ("mm: memcontrol: convert reclaim iterator to simple css refcounting")
Signed-off-by: Miles Chen <miles.chen@mediatek.com>
Signed-off-by: Qian Cai <cai@lca.pw>
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>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-08-13 15:37:28 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-10-07 16:57:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-06-16 14:30:30 +08:00
|
|
|
|
|
|
|
|
|
2016-10-07 16:57:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-06-16 14:30:30 +08:00
|
|
|
|
|
|
|
|
|
2016-10-07 16:57:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-09-30 19:14:33 +05:30
|
|
|
|
2016-10-07 16:57:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-07-11 21:00:20 -07:00
|
|
|
|
2024-12-24 02:52:38 +00:00
|
|
|
|
2016-10-07 16:57:23 -07:00
|
|
|
|
2025-05-23 10:21:06 -07:00
|
|
|
|
|
|
|
|
|
2024-12-24 02:52:38 +00:00
|
|
|
|
2016-10-07 16:57:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-12-15 12:34:29 -08:00
|
|
|
|
2021-06-28 21:59:47 -04:00
|
|
|
|
2020-12-15 12:34:29 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-28 21:59:47 -04:00
|
|
|
|
2020-12-15 12:34:29 -08:00
|
|
|
|
|
|
|
|
|
2022-09-30 19:14:33 +05:30
|
|
|
|
2020-12-15 12:34:29 -08:00
|
|
|
|
2021-06-28 21:59:47 -04:00
|
|
|
|
2020-12-15 12:34:29 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-28 21:59:47 -04:00
|
|
|
|
|
|
|
|
|
2020-12-15 12:34:29 -08:00
|
|
|
|
2021-02-24 12:03:47 -08:00
|
|
|
|
2021-06-28 21:59:47 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-12-15 12:34:29 -08:00
|
|
|
|
2021-06-28 21:59:47 -04:00
|
|
|
|
2020-12-15 12:34:29 -08:00
|
|
|
|
2021-06-28 21:59:47 -04:00
|
|
|
|
2020-12-15 12:34:29 -08:00
|
|
|
|
|
|
|
|
|
2021-06-28 21:59:47 -04:00
|
|
|
|
2020-12-15 12:34:29 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-28 21:59:47 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-12-15 12:34:29 -08:00
|
|
|
|
2021-06-28 21:59:47 -04:00
|
|
|
|
2020-12-15 12:34:29 -08:00
|
|
|
|
|
|
|
|
|
2021-06-28 21:59:47 -04:00
|
|
|
|
2020-12-15 12:34:29 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-28 21:59:47 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-12-15 12:34:29 -08:00
|
|
|
|
2021-06-28 21:59:47 -04:00
|
|
|
|
2020-12-15 12:34:29 -08:00
|
|
|
|
|
|
|
|
|
2021-06-28 21:59:47 -04:00
|
|
|
|
2020-12-15 12:34:29 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-01-12 17:18:15 -08:00
|
|
|
|
2012-05-29 15:07:09 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-01-10 16:58:04 -08:00
|
|
|
|
2012-05-29 15:07:09 -07:00
|
|
|
|
2012-01-12 17:18:15 -08:00
|
|
|
|
2016-05-19 17:12:35 -07:00
|
|
|
|
2022-02-14 18:29:54 -08:00
|
|
|
|
2011-03-22 16:32:53 -07:00
|
|
|
|
2012-05-29 15:07:09 -07:00
|
|
|
|
2017-01-10 16:58:04 -08:00
|
|
|
|
2011-03-22 16:32:53 -07:00
|
|
|
|
2016-07-28 15:46:05 -07:00
|
|
|
|
2012-05-29 15:07:09 -07:00
|
|
|
|
2016-05-19 17:12:35 -07:00
|
|
|
|
2011-03-22 16:32:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-07-28 15:46:05 -07:00
|
|
|
|
2017-01-10 16:58:04 -08:00
|
|
|
|
2016-05-19 17:12:35 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-01-10 16:58:04 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-05-19 17:12:35 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
memcg: synchronized LRU
A big patch for changing memcg's LRU semantics.
Now,
- page_cgroup is linked to mem_cgroup's its own LRU (per zone).
- LRU of page_cgroup is not synchronous with global LRU.
- page and page_cgroup is one-to-one and statically allocated.
- To find page_cgroup is on what LRU, you have to check pc->mem_cgroup as
- lru = page_cgroup_zoneinfo(pc, nid_of_pc, zid_of_pc);
- SwapCache is handled.
And, when we handle LRU list of page_cgroup, we do following.
pc = lookup_page_cgroup(page);
lock_page_cgroup(pc); .....................(1)
mz = page_cgroup_zoneinfo(pc);
spin_lock(&mz->lru_lock);
.....add to LRU
spin_unlock(&mz->lru_lock);
unlock_page_cgroup(pc);
But (1) is spin_lock and we have to be afraid of dead-lock with zone->lru_lock.
So, trylock() is used at (1), now. Without (1), we can't trust "mz" is correct.
This is a trial to remove this dirty nesting of locks.
This patch changes mz->lru_lock to be zone->lru_lock.
Then, above sequence will be written as
spin_lock(&zone->lru_lock); # in vmscan.c or swap.c via global LRU
mem_cgroup_add/remove/etc_lru() {
pc = lookup_page_cgroup(page);
mz = page_cgroup_zoneinfo(pc);
if (PageCgroupUsed(pc)) {
....add to LRU
}
spin_lock(&zone->lru_lock); # in vmscan.c or swap.c via global LRU
This is much simpler.
(*) We're safe even if we don't take lock_page_cgroup(pc). Because..
1. When pc->mem_cgroup can be modified.
- at charge.
- at account_move().
2. at charge
the PCG_USED bit is not set before pc->mem_cgroup is fixed.
3. at account_move()
the page is isolated and not on LRU.
Pros.
- easy for maintenance.
- memcg can make use of laziness of pagevec.
- we don't have to duplicated LRU/Active/Unevictable bit in page_cgroup.
- LRU status of memcg will be synchronized with global LRU's one.
- # of locks are reduced.
- account_move() is simplified very much.
Cons.
- may increase cost of LRU rotation.
(no impact if memcg is not configured.)
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Li Zefan <lizf@cn.fujitsu.com>
Cc: Balbir Singh <balbir@in.ibm.com>
Cc: Pavel Emelyanov <xemul@openvz.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-01-07 18:08:01 -08:00
|
|
|
|
2009-01-07 18:08:34 -08:00
|
|
|
|
2011-02-01 15:52:43 -08:00
|
|
|
|
2011-03-23 16:42:21 -07:00
|
|
|
|
2012-06-20 12:53:01 -07:00
|
|
|
|
2011-02-01 15:52:43 -08:00
|
|
|
|
2011-03-23 16:42:21 -07:00
|
|
|
|
2011-03-23 16:42:36 -07:00
|
|
|
|
2011-02-01 15:52:43 -08:00
|
|
|
|
2011-11-02 13:38:15 -07:00
|
|
|
|
2011-02-01 15:52:43 -08:00
|
|
|
|
mm: memcontrol: lockless page counters
Memory is internally accounted in bytes, using spinlock-protected 64-bit
counters, even though the smallest accounting delta is a page. The
counter interface is also convoluted and does too many things.
Introduce a new lockless word-sized page counter API, then change all
memory accounting over to it. The translation from and to bytes then only
happens when interfacing with userspace.
The removed locking overhead is noticable when scaling beyond the per-cpu
charge caches - on a 4-socket machine with 144-threads, the following test
shows the performance differences of 288 memcgs concurrently running a
page fault benchmark:
vanilla:
18631648.500498 task-clock (msec) # 140.643 CPUs utilized ( +- 0.33% )
1,380,638 context-switches # 0.074 K/sec ( +- 0.75% )
24,390 cpu-migrations # 0.001 K/sec ( +- 8.44% )
1,843,305,768 page-faults # 0.099 M/sec ( +- 0.00% )
50,134,994,088,218 cycles # 2.691 GHz ( +- 0.33% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
8,049,712,224,651 instructions # 0.16 insns per cycle ( +- 0.04% )
1,586,970,584,979 branches # 85.176 M/sec ( +- 0.05% )
1,724,989,949 branch-misses # 0.11% of all branches ( +- 0.48% )
132.474343877 seconds time elapsed ( +- 0.21% )
lockless:
12195979.037525 task-clock (msec) # 133.480 CPUs utilized ( +- 0.18% )
832,850 context-switches # 0.068 K/sec ( +- 0.54% )
15,624 cpu-migrations # 0.001 K/sec ( +- 10.17% )
1,843,304,774 page-faults # 0.151 M/sec ( +- 0.00% )
32,811,216,801,141 cycles # 2.690 GHz ( +- 0.18% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
9,999,265,091,727 instructions # 0.30 insns per cycle ( +- 0.10% )
2,076,759,325,203 branches # 170.282 M/sec ( +- 0.12% )
1,656,917,214 branch-misses # 0.08% of all branches ( +- 0.55% )
91.369330729 seconds time elapsed ( +- 0.45% )
On top of improved scalability, this also gets rid of the icky long long
types in the very heart of memcg, which is great for 32 bit and also makes
the code a lot more readable.
Notable differences between the old and new API:
- res_counter_charge() and res_counter_charge_nofail() become
page_counter_try_charge() and page_counter_charge() resp. to match
the more common kernel naming scheme of try_do()/do()
- res_counter_uncharge_until() is only ever used to cancel a local
counter and never to uncharge bigger segments of a hierarchy, so
it's replaced by the simpler page_counter_cancel()
- res_counter_set_limit() is replaced by page_counter_limit(), which
expects its callers to serialize against themselves
- res_counter_memparse_write_strategy() is replaced by
page_counter_limit(), which rounds down to the nearest page size -
rather than up. This is more reasonable for explicitely requested
hard upper limits.
- to keep charging light-weight, page_counter_try_charge() charges
speculatively, only to roll back if the result exceeds the limit.
Because of this, a failing bigger charge can temporarily lock out
smaller charges that would otherwise succeed. The error is bounded
to the difference between the smallest and the biggest possible
charge size, so for memcg, this means that a failing THP charge can
send base page charges into reclaim upto 2MB (4MB) before the limit
would have been reached. This should be acceptable.
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE and memparse]
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE, memparse, strncmp, and PAGE_SIZE]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Vladimir Davydov <vdavydov@parallels.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-12-10 15:42:31 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-03-23 16:42:21 -07:00
|
|
|
|
mm: memcontrol: lockless page counters
Memory is internally accounted in bytes, using spinlock-protected 64-bit
counters, even though the smallest accounting delta is a page. The
counter interface is also convoluted and does too many things.
Introduce a new lockless word-sized page counter API, then change all
memory accounting over to it. The translation from and to bytes then only
happens when interfacing with userspace.
The removed locking overhead is noticable when scaling beyond the per-cpu
charge caches - on a 4-socket machine with 144-threads, the following test
shows the performance differences of 288 memcgs concurrently running a
page fault benchmark:
vanilla:
18631648.500498 task-clock (msec) # 140.643 CPUs utilized ( +- 0.33% )
1,380,638 context-switches # 0.074 K/sec ( +- 0.75% )
24,390 cpu-migrations # 0.001 K/sec ( +- 8.44% )
1,843,305,768 page-faults # 0.099 M/sec ( +- 0.00% )
50,134,994,088,218 cycles # 2.691 GHz ( +- 0.33% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
8,049,712,224,651 instructions # 0.16 insns per cycle ( +- 0.04% )
1,586,970,584,979 branches # 85.176 M/sec ( +- 0.05% )
1,724,989,949 branch-misses # 0.11% of all branches ( +- 0.48% )
132.474343877 seconds time elapsed ( +- 0.21% )
lockless:
12195979.037525 task-clock (msec) # 133.480 CPUs utilized ( +- 0.18% )
832,850 context-switches # 0.068 K/sec ( +- 0.54% )
15,624 cpu-migrations # 0.001 K/sec ( +- 10.17% )
1,843,304,774 page-faults # 0.151 M/sec ( +- 0.00% )
32,811,216,801,141 cycles # 2.690 GHz ( +- 0.18% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
9,999,265,091,727 instructions # 0.30 insns per cycle ( +- 0.10% )
2,076,759,325,203 branches # 170.282 M/sec ( +- 0.12% )
1,656,917,214 branch-misses # 0.08% of all branches ( +- 0.55% )
91.369330729 seconds time elapsed ( +- 0.45% )
On top of improved scalability, this also gets rid of the icky long long
types in the very heart of memcg, which is great for 32 bit and also makes
the code a lot more readable.
Notable differences between the old and new API:
- res_counter_charge() and res_counter_charge_nofail() become
page_counter_try_charge() and page_counter_charge() resp. to match
the more common kernel naming scheme of try_do()/do()
- res_counter_uncharge_until() is only ever used to cancel a local
counter and never to uncharge bigger segments of a hierarchy, so
it's replaced by the simpler page_counter_cancel()
- res_counter_set_limit() is replaced by page_counter_limit(), which
expects its callers to serialize against themselves
- res_counter_memparse_write_strategy() is replaced by
page_counter_limit(), which rounds down to the nearest page size -
rather than up. This is more reasonable for explicitely requested
hard upper limits.
- to keep charging light-weight, page_counter_try_charge() charges
speculatively, only to roll back if the result exceeds the limit.
Because of this, a failing bigger charge can temporarily lock out
smaller charges that would otherwise succeed. The error is bounded
to the difference between the smallest and the biggest possible
charge size, so for memcg, this means that a failing THP charge can
send base page charges into reclaim upto 2MB (4MB) before the limit
would have been reached. This should be acceptable.
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE and memparse]
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE, memparse, strncmp, and PAGE_SIZE]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Vladimir Davydov <vdavydov@parallels.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-12-10 15:42:31 -08:00
|
|
|
|
2018-06-07 17:06:18 -07:00
|
|
|
|
mm: memcontrol: lockless page counters
Memory is internally accounted in bytes, using spinlock-protected 64-bit
counters, even though the smallest accounting delta is a page. The
counter interface is also convoluted and does too many things.
Introduce a new lockless word-sized page counter API, then change all
memory accounting over to it. The translation from and to bytes then only
happens when interfacing with userspace.
The removed locking overhead is noticable when scaling beyond the per-cpu
charge caches - on a 4-socket machine with 144-threads, the following test
shows the performance differences of 288 memcgs concurrently running a
page fault benchmark:
vanilla:
18631648.500498 task-clock (msec) # 140.643 CPUs utilized ( +- 0.33% )
1,380,638 context-switches # 0.074 K/sec ( +- 0.75% )
24,390 cpu-migrations # 0.001 K/sec ( +- 8.44% )
1,843,305,768 page-faults # 0.099 M/sec ( +- 0.00% )
50,134,994,088,218 cycles # 2.691 GHz ( +- 0.33% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
8,049,712,224,651 instructions # 0.16 insns per cycle ( +- 0.04% )
1,586,970,584,979 branches # 85.176 M/sec ( +- 0.05% )
1,724,989,949 branch-misses # 0.11% of all branches ( +- 0.48% )
132.474343877 seconds time elapsed ( +- 0.21% )
lockless:
12195979.037525 task-clock (msec) # 133.480 CPUs utilized ( +- 0.18% )
832,850 context-switches # 0.068 K/sec ( +- 0.54% )
15,624 cpu-migrations # 0.001 K/sec ( +- 10.17% )
1,843,304,774 page-faults # 0.151 M/sec ( +- 0.00% )
32,811,216,801,141 cycles # 2.690 GHz ( +- 0.18% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
9,999,265,091,727 instructions # 0.30 insns per cycle ( +- 0.10% )
2,076,759,325,203 branches # 170.282 M/sec ( +- 0.12% )
1,656,917,214 branch-misses # 0.08% of all branches ( +- 0.55% )
91.369330729 seconds time elapsed ( +- 0.45% )
On top of improved scalability, this also gets rid of the icky long long
types in the very heart of memcg, which is great for 32 bit and also makes
the code a lot more readable.
Notable differences between the old and new API:
- res_counter_charge() and res_counter_charge_nofail() become
page_counter_try_charge() and page_counter_charge() resp. to match
the more common kernel naming scheme of try_do()/do()
- res_counter_uncharge_until() is only ever used to cancel a local
counter and never to uncharge bigger segments of a hierarchy, so
it's replaced by the simpler page_counter_cancel()
- res_counter_set_limit() is replaced by page_counter_limit(), which
expects its callers to serialize against themselves
- res_counter_memparse_write_strategy() is replaced by
page_counter_limit(), which rounds down to the nearest page size -
rather than up. This is more reasonable for explicitely requested
hard upper limits.
- to keep charging light-weight, page_counter_try_charge() charges
speculatively, only to roll back if the result exceeds the limit.
Because of this, a failing bigger charge can temporarily lock out
smaller charges that would otherwise succeed. The error is bounded
to the difference between the smallest and the biggest possible
charge size, so for memcg, this means that a failing THP charge can
send base page charges into reclaim upto 2MB (4MB) before the limit
would have been reached. This should be acceptable.
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE and memparse]
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE, memparse, strncmp, and PAGE_SIZE]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Vladimir Davydov <vdavydov@parallels.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-12-10 15:42:31 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-01-14 15:21:23 -08:00
|
|
|
|
mm: memcontrol: lockless page counters
Memory is internally accounted in bytes, using spinlock-protected 64-bit
counters, even though the smallest accounting delta is a page. The
counter interface is also convoluted and does too many things.
Introduce a new lockless word-sized page counter API, then change all
memory accounting over to it. The translation from and to bytes then only
happens when interfacing with userspace.
The removed locking overhead is noticable when scaling beyond the per-cpu
charge caches - on a 4-socket machine with 144-threads, the following test
shows the performance differences of 288 memcgs concurrently running a
page fault benchmark:
vanilla:
18631648.500498 task-clock (msec) # 140.643 CPUs utilized ( +- 0.33% )
1,380,638 context-switches # 0.074 K/sec ( +- 0.75% )
24,390 cpu-migrations # 0.001 K/sec ( +- 8.44% )
1,843,305,768 page-faults # 0.099 M/sec ( +- 0.00% )
50,134,994,088,218 cycles # 2.691 GHz ( +- 0.33% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
8,049,712,224,651 instructions # 0.16 insns per cycle ( +- 0.04% )
1,586,970,584,979 branches # 85.176 M/sec ( +- 0.05% )
1,724,989,949 branch-misses # 0.11% of all branches ( +- 0.48% )
132.474343877 seconds time elapsed ( +- 0.21% )
lockless:
12195979.037525 task-clock (msec) # 133.480 CPUs utilized ( +- 0.18% )
832,850 context-switches # 0.068 K/sec ( +- 0.54% )
15,624 cpu-migrations # 0.001 K/sec ( +- 10.17% )
1,843,304,774 page-faults # 0.151 M/sec ( +- 0.00% )
32,811,216,801,141 cycles # 2.690 GHz ( +- 0.18% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
9,999,265,091,727 instructions # 0.30 insns per cycle ( +- 0.10% )
2,076,759,325,203 branches # 170.282 M/sec ( +- 0.12% )
1,656,917,214 branch-misses # 0.08% of all branches ( +- 0.55% )
91.369330729 seconds time elapsed ( +- 0.45% )
On top of improved scalability, this also gets rid of the icky long long
types in the very heart of memcg, which is great for 32 bit and also makes
the code a lot more readable.
Notable differences between the old and new API:
- res_counter_charge() and res_counter_charge_nofail() become
page_counter_try_charge() and page_counter_charge() resp. to match
the more common kernel naming scheme of try_do()/do()
- res_counter_uncharge_until() is only ever used to cancel a local
counter and never to uncharge bigger segments of a hierarchy, so
it's replaced by the simpler page_counter_cancel()
- res_counter_set_limit() is replaced by page_counter_limit(), which
expects its callers to serialize against themselves
- res_counter_memparse_write_strategy() is replaced by
page_counter_limit(), which rounds down to the nearest page size -
rather than up. This is more reasonable for explicitely requested
hard upper limits.
- to keep charging light-weight, page_counter_try_charge() charges
speculatively, only to roll back if the result exceeds the limit.
Because of this, a failing bigger charge can temporarily lock out
smaller charges that would otherwise succeed. The error is bounded
to the difference between the smallest and the biggest possible
charge size, so for memcg, this means that a failing THP charge can
send base page charges into reclaim upto 2MB (4MB) before the limit
would have been reached. This should be acceptable.
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE and memparse]
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE, memparse, strncmp, and PAGE_SIZE]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Vladimir Davydov <vdavydov@parallels.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-12-10 15:42:31 -08:00
|
|
|
|
2018-06-07 17:06:18 -07:00
|
|
|
|
2020-06-01 21:49:36 -07:00
|
|
|
|
mm: memcontrol: lockless page counters
Memory is internally accounted in bytes, using spinlock-protected 64-bit
counters, even though the smallest accounting delta is a page. The
counter interface is also convoluted and does too many things.
Introduce a new lockless word-sized page counter API, then change all
memory accounting over to it. The translation from and to bytes then only
happens when interfacing with userspace.
The removed locking overhead is noticable when scaling beyond the per-cpu
charge caches - on a 4-socket machine with 144-threads, the following test
shows the performance differences of 288 memcgs concurrently running a
page fault benchmark:
vanilla:
18631648.500498 task-clock (msec) # 140.643 CPUs utilized ( +- 0.33% )
1,380,638 context-switches # 0.074 K/sec ( +- 0.75% )
24,390 cpu-migrations # 0.001 K/sec ( +- 8.44% )
1,843,305,768 page-faults # 0.099 M/sec ( +- 0.00% )
50,134,994,088,218 cycles # 2.691 GHz ( +- 0.33% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
8,049,712,224,651 instructions # 0.16 insns per cycle ( +- 0.04% )
1,586,970,584,979 branches # 85.176 M/sec ( +- 0.05% )
1,724,989,949 branch-misses # 0.11% of all branches ( +- 0.48% )
132.474343877 seconds time elapsed ( +- 0.21% )
lockless:
12195979.037525 task-clock (msec) # 133.480 CPUs utilized ( +- 0.18% )
832,850 context-switches # 0.068 K/sec ( +- 0.54% )
15,624 cpu-migrations # 0.001 K/sec ( +- 10.17% )
1,843,304,774 page-faults # 0.151 M/sec ( +- 0.00% )
32,811,216,801,141 cycles # 2.690 GHz ( +- 0.18% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
9,999,265,091,727 instructions # 0.30 insns per cycle ( +- 0.10% )
2,076,759,325,203 branches # 170.282 M/sec ( +- 0.12% )
1,656,917,214 branch-misses # 0.08% of all branches ( +- 0.55% )
91.369330729 seconds time elapsed ( +- 0.45% )
On top of improved scalability, this also gets rid of the icky long long
types in the very heart of memcg, which is great for 32 bit and also makes
the code a lot more readable.
Notable differences between the old and new API:
- res_counter_charge() and res_counter_charge_nofail() become
page_counter_try_charge() and page_counter_charge() resp. to match
the more common kernel naming scheme of try_do()/do()
- res_counter_uncharge_until() is only ever used to cancel a local
counter and never to uncharge bigger segments of a hierarchy, so
it's replaced by the simpler page_counter_cancel()
- res_counter_set_limit() is replaced by page_counter_limit(), which
expects its callers to serialize against themselves
- res_counter_memparse_write_strategy() is replaced by
page_counter_limit(), which rounds down to the nearest page size -
rather than up. This is more reasonable for explicitely requested
hard upper limits.
- to keep charging light-weight, page_counter_try_charge() charges
speculatively, only to roll back if the result exceeds the limit.
Because of this, a failing bigger charge can temporarily lock out
smaller charges that would otherwise succeed. The error is bounded
to the difference between the smallest and the biggest possible
charge size, so for memcg, this means that a failing THP charge can
send base page charges into reclaim upto 2MB (4MB) before the limit
would have been reached. This should be acceptable.
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE and memparse]
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE, memparse, strncmp, and PAGE_SIZE]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Vladimir Davydov <vdavydov@parallels.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-12-10 15:42:31 -08:00
|
|
|
|
2016-05-27 14:27:43 -07:00
|
|
|
|
|
|
|
|
|
mm: memcontrol: lockless page counters
Memory is internally accounted in bytes, using spinlock-protected 64-bit
counters, even though the smallest accounting delta is a page. The
counter interface is also convoluted and does too many things.
Introduce a new lockless word-sized page counter API, then change all
memory accounting over to it. The translation from and to bytes then only
happens when interfacing with userspace.
The removed locking overhead is noticable when scaling beyond the per-cpu
charge caches - on a 4-socket machine with 144-threads, the following test
shows the performance differences of 288 memcgs concurrently running a
page fault benchmark:
vanilla:
18631648.500498 task-clock (msec) # 140.643 CPUs utilized ( +- 0.33% )
1,380,638 context-switches # 0.074 K/sec ( +- 0.75% )
24,390 cpu-migrations # 0.001 K/sec ( +- 8.44% )
1,843,305,768 page-faults # 0.099 M/sec ( +- 0.00% )
50,134,994,088,218 cycles # 2.691 GHz ( +- 0.33% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
8,049,712,224,651 instructions # 0.16 insns per cycle ( +- 0.04% )
1,586,970,584,979 branches # 85.176 M/sec ( +- 0.05% )
1,724,989,949 branch-misses # 0.11% of all branches ( +- 0.48% )
132.474343877 seconds time elapsed ( +- 0.21% )
lockless:
12195979.037525 task-clock (msec) # 133.480 CPUs utilized ( +- 0.18% )
832,850 context-switches # 0.068 K/sec ( +- 0.54% )
15,624 cpu-migrations # 0.001 K/sec ( +- 10.17% )
1,843,304,774 page-faults # 0.151 M/sec ( +- 0.00% )
32,811,216,801,141 cycles # 2.690 GHz ( +- 0.18% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
9,999,265,091,727 instructions # 0.30 insns per cycle ( +- 0.10% )
2,076,759,325,203 branches # 170.282 M/sec ( +- 0.12% )
1,656,917,214 branch-misses # 0.08% of all branches ( +- 0.55% )
91.369330729 seconds time elapsed ( +- 0.45% )
On top of improved scalability, this also gets rid of the icky long long
types in the very heart of memcg, which is great for 32 bit and also makes
the code a lot more readable.
Notable differences between the old and new API:
- res_counter_charge() and res_counter_charge_nofail() become
page_counter_try_charge() and page_counter_charge() resp. to match
the more common kernel naming scheme of try_do()/do()
- res_counter_uncharge_until() is only ever used to cancel a local
counter and never to uncharge bigger segments of a hierarchy, so
it's replaced by the simpler page_counter_cancel()
- res_counter_set_limit() is replaced by page_counter_limit(), which
expects its callers to serialize against themselves
- res_counter_memparse_write_strategy() is replaced by
page_counter_limit(), which rounds down to the nearest page size -
rather than up. This is more reasonable for explicitely requested
hard upper limits.
- to keep charging light-weight, page_counter_try_charge() charges
speculatively, only to roll back if the result exceeds the limit.
Because of this, a failing bigger charge can temporarily lock out
smaller charges that would otherwise succeed. The error is bounded
to the difference between the smallest and the biggest possible
charge size, so for memcg, this means that a failing THP charge can
send base page charges into reclaim upto 2MB (4MB) before the limit
would have been reached. This should be acceptable.
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE and memparse]
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE, memparse, strncmp, and PAGE_SIZE]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Vladimir Davydov <vdavydov@parallels.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-12-10 15:42:31 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-02-01 15:52:43 -08:00
|
|
|
|
|
|
|
|
|
2020-10-13 16:52:59 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-02-24 12:03:31 -08:00
|
|
|
|
2021-02-24 12:03:43 -08:00
|
|
|
|
|
|
|
|
|
memcg: add per-memcg total kernel memory stat
Currently memcg stats show several types of kernel memory: kernel stack,
page tables, sock, vmalloc, and slab. However, there are other
allocations with __GFP_ACCOUNT (or supersets such as GFP_KERNEL_ACCOUNT)
that are not accounted in any of those stats, a few examples are:
- various kvm allocations (e.g. allocated pages to create vcpus)
- io_uring
- tmp_page in pipes during pipe_write()
- bpf ringbuffers
- unix sockets
Keeping track of the total kernel memory is essential for the ease of
migration from cgroup v1 to v2 as there are large discrepancies between
v1's kmem.usage_in_bytes and the sum of the available kernel memory
stats in v2. Adding separate memcg stats for all __GFP_ACCOUNT kernel
allocations is an impractical maintenance burden as there a lot of those
all over the kernel code, with more use cases likely to show up in the
future.
Therefore, add a "kernel" memcg stat that is analogous to kmem page
counter, with added benefits such as using rstat infrastructure which
aggregates stats more efficiently. Additionally, this provides a
lighter alternative in case the legacy kmem is deprecated in the future
[yosryahmed@google.com: v2]
Link: https://lkml.kernel.org/r/20220203193856.972500-1-yosryahmed@google.com
Link: https://lkml.kernel.org/r/20220201200823.3283171-1-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Acked-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <songmuchun@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-03-22 14:40:10 -07:00
|
|
|
|
2021-02-24 12:03:43 -08:00
|
|
|
|
|
|
|
|
|
mm: add NR_SECONDARY_PAGETABLE to count secondary page table uses.
We keep track of several kernel memory stats (total kernel memory, page
tables, stack, vmalloc, etc) on multiple levels (global, per-node,
per-memcg, etc). These stats give insights to users to how much memory
is used by the kernel and for what purposes.
Currently, memory used by KVM mmu is not accounted in any of those
kernel memory stats. This patch series accounts the memory pages
used by KVM for page tables in those stats in a new
NR_SECONDARY_PAGETABLE stat. This stat can be later extended to account
for other types of secondary pages tables (e.g. iommu page tables).
KVM has a decent number of large allocations that aren't for page
tables, but for most of them, the number/size of those allocations
scales linearly with either the number of vCPUs or the amount of memory
assigned to the VM. KVM's secondary page table allocations do not scale
linearly, especially when nested virtualization is in use.
From a KVM perspective, NR_SECONDARY_PAGETABLE will scale with KVM's
per-VM pages_{4k,2m,1g} stats unless the guest is doing something
bizarre (e.g. accessing only 4kb chunks of 2mb pages so that KVM is
forced to allocate a large number of page tables even though the guest
isn't accessing that much memory). However, someone would need to either
understand how KVM works to make that connection, or know (or be told) to
go look at KVM's stats if they're running VMs to better decipher the stats.
Furthermore, having NR_PAGETABLE side-by-side with NR_SECONDARY_PAGETABLE
is informative. For example, when backing a VM with THP vs. HugeTLB,
NR_SECONDARY_PAGETABLE is roughly the same, but NR_PAGETABLE is an order
of magnitude higher with THP. So having this stat will at the very least
prove to be useful for understanding tradeoffs between VM backing types,
and likely even steer folks towards potential optimizations.
The original discussion with more details about the rationale:
https://lore.kernel.org/all/87ilqoi77b.wl-maz@kernel.org
This stat will be used by subsequent patches to count KVM mmu
memory usage.
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Acked-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20220823004639.2387269-2-yosryahmed@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
2022-08-23 00:46:36 +00:00
|
|
|
|
2021-02-24 12:03:43 -08:00
|
|
|
|
|
|
|
|
|
2022-01-14 14:05:45 -08:00
|
|
|
|
2021-02-24 12:03:43 -08:00
|
|
|
|
2024-07-01 11:31:15 -04:00
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-02-24 12:03:43 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-02-24 12:03:55 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-10-13 16:52:59 -07:00
|
|
|
|
2021-02-24 12:03:43 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-10-13 16:52:59 -07:00
|
|
|
|
2021-02-24 12:03:43 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
memcg/hugetlb: add hugeTLB counters to memcg
This patch introduces a new counter to memory.stat that tracks hugeTLB
usage, only if hugeTLB accounting is done to memory.current. This feature
is enabled the same way hugeTLB accounting is enabled, via the
memory_hugetlb_accounting mount flag for cgroupsv2.
1. Why is this patch necessary?
Currently, memcg hugeTLB accounting is an opt-in feature [1] that adds
hugeTLB usage to memory.current. However, the metric is not reported in
memory.stat. Given that users often interpret memory.stat as a breakdown
of the value reported in memory.current, the disparity between the two
reports can be confusing. This patch solves this problem by including the
metric in memory.stat as well, but only if it is also reported in
memory.current (it would also be confusing if the value was reported in
memory.stat, but not in memory.current)
Aside from the consistency between the two files, we also see benefits in
observability. Userspace might be interested in the hugeTLB footprint of
cgroups for many reasons. For instance, system admins might want to
verify that hugeTLB usage is distributed as expected across tasks: i.e.
memory-intensive tasks are using more hugeTLB pages than tasks that don't
consume a lot of memory, or are seen to fault frequently. Note that this
is separate from wanting to inspect the distribution for limiting purposes
(in which case, hugeTLB controller makes more sense).
2. We already have a hugeTLB controller. Why not use that?
It is true that hugeTLB tracks the exact value that we want. In fact, by
enabling the hugeTLB controller, we get all of the observability benefits
that I mentioned above, and users can check the total hugeTLB usage,
verify if it is distributed as expected, etc.
With this said, there are 2 problems:
(a) They are still not reported in memory.stat, which means the
disparity between the memcg reports are still there.
(b) We cannot reasonably expect users to enable the hugeTLB controller
just for the sake of hugeTLB usage reporting, especially since
they don't have any use for hugeTLB usage enforcing [2].
3. Implementation Details:
In the alloc / free hugetlb functions, we call lruvec_stat_mod_folio
regardless of whether memcg accounts hugetlb. mem_cgroup_commit_charge
which is called from alloc_hugetlb_folio will set memcg for the folio only
if the CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING cgroup mount option is used, so
lruvec_stat_mod_folio accounts per-memcg hugetlb counters only if the
feature is enabled. Regardless of whether memcg accounts for hugetlb, the
newly added global counter is updated and shown in /proc/vmstat.
The global counter is added because vmstats is the preferred framework for
cgroup stats. It makes stat items consistent between global and cgroups.
It also provides a per-node breakdown, which is useful. Because it does
not use cgroup-specific hooks, we also keep generic MM code separate from
memcg code.
[1] https://lore.kernel.org/all/20231006184629.155543-1-nphamcs@gmail.com/
[2] Of course, we can't make a new patch for every feature that can be
duplicated. However, since the existing solution of enabling the
hugeTLB controller is an imperfect solution that still leaves a
discrepancy between memory.stat and memory.curent, I think that it
is reasonable to isolate the feature in this case.
Link: https://lkml.kernel.org/r/20241101204402.1885383-1-joshua.hahnjy@gmail.com
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Nhat Pham <nphamcs@gmail.com>
Suggested-by: Shakeel Butt <shakeel.butt@linux.dev>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Shakeel Butt <shakeel.butt@linux.dev>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Chris Down <chris@chrisdown.name>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Roman Gushchin <roman.gushchin@linux.dev>
Reviewed-by: Nhat Pham <nphamcs@gmail.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Koutný <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-11-01 13:44:02 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-10-13 16:52:59 -07:00
|
|
|
|
|
|
|
|
|
2021-02-24 12:03:43 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm,memcg: provide per-cgroup counters for NUMA balancing operations
The ability to observe the demotion and promotion decisions made by the
kernel on a per-cgroup basis is important for monitoring and tuning
containerized workloads on machines equipped with tiered memory.
Different containers in the system may experience drastically different
memory tiering actions that cannot be distinguished from the global
counters alone.
For example, a container running a workload that has a much hotter memory
accesses will likely see more promotions and fewer demotions, potentially
depriving a colocated container of top tier memory to such an extent that
its performance degrades unacceptably.
For another example, some containers may exhibit longer periods between
data reuse, causing much more numa_hint_faults than numa_pages_migrated.
In this case, tuning hot_threshold_ms may be appropriate, but the signal
can easily be lost if only global counters are available.
In the long term, we hope to introduce per-cgroup control of promotion and
demotion actions to implement memory placement policies in tiering.
This patch set adds seven counters to memory.stat in a cgroup:
numa_pages_migrated, numa_pte_updates, numa_hint_faults, pgdemote_kswapd,
pgdemote_khugepaged, pgdemote_direct and pgpromote_success. pgdemote_*
and pgpromote_success are also available in memory.numa_stat.
count_memcg_events_mm() is added to count multiple event occurrences at
once, and get_mem_cgroup_from_folio() is added because we need to get a
reference to the memcg of a folio before it's migrated to track
numa_pages_migrated. The accounting of PGDEMOTE_* is moved to
shrink_inactive_list() before being changed to per-cgroup.
[kaiyang2@cs.cmu.edu: add documentation of the memcg counters in cgroup-v2.rst]
Link: https://lkml.kernel.org/r/20240814235122.252309-1-kaiyang2@cs.cmu.edu
Link: https://lkml.kernel.org/r/20240814174227.30639-1-kaiyang2@cs.cmu.edu
Signed-off-by: Kaiyang Zhao <kaiyang2@cs.cmu.edu>
Cc: David Rientjes <rientjes@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Wei Xu <weixugc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-08-14 17:42:27 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-03-18 15:58:32 +08:00
|
|
|
|
mm,memcg: provide per-cgroup counters for NUMA balancing operations
The ability to observe the demotion and promotion decisions made by the
kernel on a per-cgroup basis is important for monitoring and tuning
containerized workloads on machines equipped with tiered memory.
Different containers in the system may experience drastically different
memory tiering actions that cannot be distinguished from the global
counters alone.
For example, a container running a workload that has a much hotter memory
accesses will likely see more promotions and fewer demotions, potentially
depriving a colocated container of top tier memory to such an extent that
its performance degrades unacceptably.
For another example, some containers may exhibit longer periods between
data reuse, causing much more numa_hint_faults than numa_pages_migrated.
In this case, tuning hot_threshold_ms may be appropriate, but the signal
can easily be lost if only global counters are available.
In the long term, we hope to introduce per-cgroup control of promotion and
demotion actions to implement memory placement policies in tiering.
This patch set adds seven counters to memory.stat in a cgroup:
numa_pages_migrated, numa_pte_updates, numa_hint_faults, pgdemote_kswapd,
pgdemote_khugepaged, pgdemote_direct and pgpromote_success. pgdemote_*
and pgpromote_success are also available in memory.numa_stat.
count_memcg_events_mm() is added to count multiple event occurrences at
once, and get_mem_cgroup_from_folio() is added because we need to get a
reference to the memcg of a folio before it's migrated to track
numa_pages_migrated. The accounting of PGDEMOTE_* is moved to
shrink_inactive_list() before being changed to per-cgroup.
[kaiyang2@cs.cmu.edu: add documentation of the memcg counters in cgroup-v2.rst]
Link: https://lkml.kernel.org/r/20240814235122.252309-1-kaiyang2@cs.cmu.edu
Link: https://lkml.kernel.org/r/20240814174227.30639-1-kaiyang2@cs.cmu.edu
Signed-off-by: Kaiyang Zhao <kaiyang2@cs.cmu.edu>
Cc: David Rientjes <rientjes@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Wei Xu <weixugc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-08-14 17:42:27 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-10-13 16:52:59 -07:00
|
|
|
|
|
|
|
|
|
2023-09-22 17:57:39 +00:00
|
|
|
|
2021-02-24 12:03:43 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
2021-02-24 12:03:43 -08:00
|
|
|
|
|
|
|
|
|
2023-09-22 17:57:39 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm,memcg: provide per-cgroup counters for NUMA balancing operations
The ability to observe the demotion and promotion decisions made by the
kernel on a per-cgroup basis is important for monitoring and tuning
containerized workloads on machines equipped with tiered memory.
Different containers in the system may experience drastically different
memory tiering actions that cannot be distinguished from the global
counters alone.
For example, a container running a workload that has a much hotter memory
accesses will likely see more promotions and fewer demotions, potentially
depriving a colocated container of top tier memory to such an extent that
its performance degrades unacceptably.
For another example, some containers may exhibit longer periods between
data reuse, causing much more numa_hint_faults than numa_pages_migrated.
In this case, tuning hot_threshold_ms may be appropriate, but the signal
can easily be lost if only global counters are available.
In the long term, we hope to introduce per-cgroup control of promotion and
demotion actions to implement memory placement policies in tiering.
This patch set adds seven counters to memory.stat in a cgroup:
numa_pages_migrated, numa_pte_updates, numa_hint_faults, pgdemote_kswapd,
pgdemote_khugepaged, pgdemote_direct and pgpromote_success. pgdemote_*
and pgpromote_success are also available in memory.numa_stat.
count_memcg_events_mm() is added to count multiple event occurrences at
once, and get_mem_cgroup_from_folio() is added because we need to get a
reference to the memcg of a folio before it's migrated to track
numa_pages_migrated. The accounting of PGDEMOTE_* is moved to
shrink_inactive_list() before being changed to per-cgroup.
[kaiyang2@cs.cmu.edu: add documentation of the memcg counters in cgroup-v2.rst]
Link: https://lkml.kernel.org/r/20240814235122.252309-1-kaiyang2@cs.cmu.edu
Link: https://lkml.kernel.org/r/20240814174227.30639-1-kaiyang2@cs.cmu.edu
Signed-off-by: Kaiyang Zhao <kaiyang2@cs.cmu.edu>
Cc: David Rientjes <rientjes@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Wei Xu <weixugc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-08-14 17:42:27 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-09-22 17:57:39 +00:00
|
|
|
|
|
|
|
|
|
2021-02-24 12:03:43 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm,memcg: provide per-cgroup counters for NUMA balancing operations
The ability to observe the demotion and promotion decisions made by the
kernel on a per-cgroup basis is important for monitoring and tuning
containerized workloads on machines equipped with tiered memory.
Different containers in the system may experience drastically different
memory tiering actions that cannot be distinguished from the global
counters alone.
For example, a container running a workload that has a much hotter memory
accesses will likely see more promotions and fewer demotions, potentially
depriving a colocated container of top tier memory to such an extent that
its performance degrades unacceptably.
For another example, some containers may exhibit longer periods between
data reuse, causing much more numa_hint_faults than numa_pages_migrated.
In this case, tuning hot_threshold_ms may be appropriate, but the signal
can easily be lost if only global counters are available.
In the long term, we hope to introduce per-cgroup control of promotion and
demotion actions to implement memory placement policies in tiering.
This patch set adds seven counters to memory.stat in a cgroup:
numa_pages_migrated, numa_pte_updates, numa_hint_faults, pgdemote_kswapd,
pgdemote_khugepaged, pgdemote_direct and pgpromote_success. pgdemote_*
and pgpromote_success are also available in memory.numa_stat.
count_memcg_events_mm() is added to count multiple event occurrences at
once, and get_mem_cgroup_from_folio() is added because we need to get a
reference to the memcg of a folio before it's migrated to track
numa_pages_migrated. The accounting of PGDEMOTE_* is moved to
shrink_inactive_list() before being changed to per-cgroup.
[kaiyang2@cs.cmu.edu: add documentation of the memcg counters in cgroup-v2.rst]
Link: https://lkml.kernel.org/r/20240814235122.252309-1-kaiyang2@cs.cmu.edu
Link: https://lkml.kernel.org/r/20240814174227.30639-1-kaiyang2@cs.cmu.edu
Signed-off-by: Kaiyang Zhao <kaiyang2@cs.cmu.edu>
Cc: David Rientjes <rientjes@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Wei Xu <weixugc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-08-14 17:42:27 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-03-18 15:58:32 +08:00
|
|
|
|
mm,memcg: provide per-cgroup counters for NUMA balancing operations
The ability to observe the demotion and promotion decisions made by the
kernel on a per-cgroup basis is important for monitoring and tuning
containerized workloads on machines equipped with tiered memory.
Different containers in the system may experience drastically different
memory tiering actions that cannot be distinguished from the global
counters alone.
For example, a container running a workload that has a much hotter memory
accesses will likely see more promotions and fewer demotions, potentially
depriving a colocated container of top tier memory to such an extent that
its performance degrades unacceptably.
For another example, some containers may exhibit longer periods between
data reuse, causing much more numa_hint_faults than numa_pages_migrated.
In this case, tuning hot_threshold_ms may be appropriate, but the signal
can easily be lost if only global counters are available.
In the long term, we hope to introduce per-cgroup control of promotion and
demotion actions to implement memory placement policies in tiering.
This patch set adds seven counters to memory.stat in a cgroup:
numa_pages_migrated, numa_pte_updates, numa_hint_faults, pgdemote_kswapd,
pgdemote_khugepaged, pgdemote_direct and pgpromote_success. pgdemote_*
and pgpromote_success are also available in memory.numa_stat.
count_memcg_events_mm() is added to count multiple event occurrences at
once, and get_mem_cgroup_from_folio() is added because we need to get a
reference to the memcg of a folio before it's migrated to track
numa_pages_migrated. The accounting of PGDEMOTE_* is moved to
shrink_inactive_list() before being changed to per-cgroup.
[kaiyang2@cs.cmu.edu: add documentation of the memcg counters in cgroup-v2.rst]
Link: https://lkml.kernel.org/r/20240814235122.252309-1-kaiyang2@cs.cmu.edu
Link: https://lkml.kernel.org/r/20240814174227.30639-1-kaiyang2@cs.cmu.edu
Signed-off-by: Kaiyang Zhao <kaiyang2@cs.cmu.edu>
Cc: David Rientjes <rientjes@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Wei Xu <weixugc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-08-14 17:42:27 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-02-24 12:03:43 -08:00
|
|
|
|
|
|
|
|
|
2023-09-22 17:57:39 +00:00
|
|
|
|
2021-02-24 12:03:43 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-06-24 17:59:02 -07:00
|
|
|
|
2021-02-24 12:03:43 -08:00
|
|
|
|
2023-09-22 17:57:39 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-01-24 07:35:14 +00:00
|
|
|
|
2024-06-24 17:59:02 -07:00
|
|
|
|
2023-09-22 17:57:39 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-02-24 12:03:43 -08:00
|
|
|
|
2025-01-24 07:35:14 +00:00
|
|
|
|
2021-02-24 12:03:43 -08:00
|
|
|
|
2024-12-11 12:39:49 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
memcg: dump memory.stat during cgroup OOM for v1
Patch series "memcg: OOM log improvements", v2.
This short patch series brings back some cgroup v1 stats in OOM logs
that were unnecessarily changed before. It also makes memcg OOM logs
less reliant on printk() internals.
This patch (of 2):
Commit c8713d0b2312 ("mm: memcontrol: dump memory.stat during cgroup OOM")
made sure we dump all the stats in memory.stat during a cgroup OOM, but it
also introduced a slight behavioral change. The code used to print the
non-hierarchical v1 cgroup stats for the entire cgroup subtree, now it
only prints the v2 cgroup stats for the cgroup under OOM.
For cgroup v1 users, this introduces a few problems:
(a) The non-hierarchical stats of the memcg under OOM are no longer
shown.
(b) A couple of v1-only stats (e.g. pgpgin, pgpgout) are no longer
shown.
(c) We show the list of cgroup v2 stats, even in cgroup v1. This list
of stats is not tracked with v1 in mind. While most of the stats seem
to be working on v1, there may be some stats that are not fully or
correctly tracked.
Although OOM log is not set in stone, we should not change it for no
reason. When upgrading the kernel version to a version including commit
c8713d0b2312 ("mm: memcontrol: dump memory.stat during cgroup OOM"), these
behavioral changes are noticed in cgroup v1.
The fix is simple. Commit c8713d0b2312 ("mm: memcontrol: dump memory.stat
during cgroup OOM") separated stats formatting from stats display for v2,
to reuse the stats formatting in the OOM logs. Do the same for v1.
Move the v2 specific formatting from memory_stat_format() to
memcg_stat_format(), add memcg1_stat_format() for v1, and make
memory_stat_format() select between them based on cgroup version. Since
memory_stat_show() now works for both v1 & v2, drop memcg_stat_show().
Link: https://lkml.kernel.org/r/20230428132406.2540811-1-yosryahmed@google.com
Link: https://lkml.kernel.org/r/20230428132406.2540811-3-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Acked-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Michal Hocko <mhocko@kernel.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Petr Mladek <pmladek@suse.com>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Steven Rostedt (Google) <rostedt@goodmis.org>
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-04-28 13:24:06 +00:00
|
|
|
|
mm: memcontrol: dump memory.stat during cgroup OOM
The current cgroup OOM memory info dump doesn't include all the memory
we are tracking, nor does it give insight into what the VM tried to do
leading up to the OOM. All that useful info is in memory.stat.
Furthermore, the recursive printing for every child cgroup can
generate absurd amounts of data on the console for larger cgroup
trees, and it's not like we provide a per-cgroup breakdown during
global OOM kills.
When an OOM kill is triggered, print one set of recursive memory.stat
items at the level whose limit triggered the OOM condition.
Example output:
stress invoked oom-killer: gfp_mask=0x100cca(GFP_HIGHUSER_MOVABLE), order=0, oom_score_adj=0
CPU: 2 PID: 210 Comm: stress Not tainted 5.2.0-rc2-mm1-00247-g47d49835983c #135
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.12.0-20181126_142135-anatol 04/01/2014
Call Trace:
dump_stack+0x46/0x60
dump_header+0x4c/0x2d0
oom_kill_process.cold.10+0xb/0x10
out_of_memory+0x200/0x270
? try_to_free_mem_cgroup_pages+0xdf/0x130
mem_cgroup_out_of_memory+0xb7/0xc0
try_charge+0x680/0x6f0
mem_cgroup_try_charge+0xb5/0x160
__add_to_page_cache_locked+0xc6/0x300
? list_lru_destroy+0x80/0x80
add_to_page_cache_lru+0x45/0xc0
pagecache_get_page+0x11b/0x290
filemap_fault+0x458/0x6d0
ext4_filemap_fault+0x27/0x36
__do_fault+0x2f/0xb0
__handle_mm_fault+0x9c5/0x1140
? apic_timer_interrupt+0xa/0x20
handle_mm_fault+0xc5/0x180
__do_page_fault+0x1ab/0x440
? page_fault+0x8/0x30
page_fault+0x1e/0x30
RIP: 0033:0x55c32167fc10
Code: Bad RIP value.
RSP: 002b:00007fff1d031c50 EFLAGS: 00010206
RAX: 000000000dc00000 RBX: 00007fd2db000010 RCX: 00007fd2db000010
RDX: 0000000000000000 RSI: 0000000010001000 RDI: 0000000000000000
RBP: 000055c321680a54 R08: 00000000ffffffff R09: 0000000000000000
R10: 0000000000000022 R11: 0000000000000246 R12: ffffffffffffffff
R13: 0000000000000002 R14: 0000000000001000 R15: 0000000010000000
memory: usage 1024kB, limit 1024kB, failcnt 75131
swap: usage 0kB, limit 9007199254740988kB, failcnt 0
Memory cgroup stats for /foo:
anon 0
file 0
kernel_stack 36864
slab 274432
sock 0
shmem 0
file_mapped 0
file_dirty 0
file_writeback 0
anon_thp 0
inactive_anon 126976
active_anon 0
inactive_file 0
active_file 0
unevictable 0
slab_reclaimable 0
slab_unreclaimable 274432
pgfault 59466
pgmajfault 1617
workingset_refault 2145
workingset_activate 0
workingset_nodereclaim 0
pgrefill 98952
pgscan 200060
pgsteal 59340
pgactivate 40095
pgdeactivate 96787
pglazyfree 0
pglazyfreed 0
thp_fault_alloc 0
thp_collapse_alloc 0
Tasks state (memory values in pages):
[ pid ] uid tgid total_vm rss pgtables_bytes swapents oom_score_adj name
[ 200] 0 200 1121 884 53248 29 0 bash
[ 209] 0 209 905 246 45056 19 0 stress
[ 210] 0 210 66442 56 499712 56349 0 stress
oom-kill:constraint=CONSTRAINT_NONE,nodemask=(null),oom_memcg=/foo,task_memcg=/foo,task=stress,pid=210,uid=0
Memory cgroup out of memory: Killed process 210 (stress) total-vm:265768kB, anon-rss:0kB, file-rss:224kB, shmem-rss:0kB
oom_reaper: reaped process 210 (stress), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
[hannes@cmpxchg.org: s/kvmalloc/kmalloc/ per Michal]
Link: http://lkml.kernel.org/r/20190605161133.GA12453@cmpxchg.org
Link: http://lkml.kernel.org/r/20190604210509.9744-1-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-11 20:55:59 -07:00
|
|
|
|
|
|
|
|
|
2017-05-03 14:55:13 -07:00
|
|
|
|
mm: memcontrol: dump memory.stat during cgroup OOM
The current cgroup OOM memory info dump doesn't include all the memory
we are tracking, nor does it give insight into what the VM tried to do
leading up to the OOM. All that useful info is in memory.stat.
Furthermore, the recursive printing for every child cgroup can
generate absurd amounts of data on the console for larger cgroup
trees, and it's not like we provide a per-cgroup breakdown during
global OOM kills.
When an OOM kill is triggered, print one set of recursive memory.stat
items at the level whose limit triggered the OOM condition.
Example output:
stress invoked oom-killer: gfp_mask=0x100cca(GFP_HIGHUSER_MOVABLE), order=0, oom_score_adj=0
CPU: 2 PID: 210 Comm: stress Not tainted 5.2.0-rc2-mm1-00247-g47d49835983c #135
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.12.0-20181126_142135-anatol 04/01/2014
Call Trace:
dump_stack+0x46/0x60
dump_header+0x4c/0x2d0
oom_kill_process.cold.10+0xb/0x10
out_of_memory+0x200/0x270
? try_to_free_mem_cgroup_pages+0xdf/0x130
mem_cgroup_out_of_memory+0xb7/0xc0
try_charge+0x680/0x6f0
mem_cgroup_try_charge+0xb5/0x160
__add_to_page_cache_locked+0xc6/0x300
? list_lru_destroy+0x80/0x80
add_to_page_cache_lru+0x45/0xc0
pagecache_get_page+0x11b/0x290
filemap_fault+0x458/0x6d0
ext4_filemap_fault+0x27/0x36
__do_fault+0x2f/0xb0
__handle_mm_fault+0x9c5/0x1140
? apic_timer_interrupt+0xa/0x20
handle_mm_fault+0xc5/0x180
__do_page_fault+0x1ab/0x440
? page_fault+0x8/0x30
page_fault+0x1e/0x30
RIP: 0033:0x55c32167fc10
Code: Bad RIP value.
RSP: 002b:00007fff1d031c50 EFLAGS: 00010206
RAX: 000000000dc00000 RBX: 00007fd2db000010 RCX: 00007fd2db000010
RDX: 0000000000000000 RSI: 0000000010001000 RDI: 0000000000000000
RBP: 000055c321680a54 R08: 00000000ffffffff R09: 0000000000000000
R10: 0000000000000022 R11: 0000000000000246 R12: ffffffffffffffff
R13: 0000000000000002 R14: 0000000000001000 R15: 0000000010000000
memory: usage 1024kB, limit 1024kB, failcnt 75131
swap: usage 0kB, limit 9007199254740988kB, failcnt 0
Memory cgroup stats for /foo:
anon 0
file 0
kernel_stack 36864
slab 274432
sock 0
shmem 0
file_mapped 0
file_dirty 0
file_writeback 0
anon_thp 0
inactive_anon 126976
active_anon 0
inactive_file 0
active_file 0
unevictable 0
slab_reclaimable 0
slab_unreclaimable 274432
pgfault 59466
pgmajfault 1617
workingset_refault 2145
workingset_activate 0
workingset_nodereclaim 0
pgrefill 98952
pgscan 200060
pgsteal 59340
pgactivate 40095
pgdeactivate 96787
pglazyfree 0
pglazyfreed 0
thp_fault_alloc 0
thp_collapse_alloc 0
Tasks state (memory values in pages):
[ pid ] uid tgid total_vm rss pgtables_bytes swapents oom_score_adj name
[ 200] 0 200 1121 884 53248 29 0 bash
[ 209] 0 209 905 246 45056 19 0 stress
[ 210] 0 210 66442 56 499712 56349 0 stress
oom-kill:constraint=CONSTRAINT_NONE,nodemask=(null),oom_memcg=/foo,task_memcg=/foo,task=stress,pid=210,uid=0
Memory cgroup out of memory: Killed process 210 (stress) total-vm:265768kB, anon-rss:0kB, file-rss:224kB, shmem-rss:0kB
oom_reaper: reaped process 210 (stress), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
[hannes@cmpxchg.org: s/kvmalloc/kmalloc/ per Michal]
Link: http://lkml.kernel.org/r/20190605161133.GA12453@cmpxchg.org
Link: http://lkml.kernel.org/r/20190604210509.9744-1-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-11 20:55:59 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcg: restore subtree stats flushing
Stats flushing for memcg currently follows the following rules:
- Always flush the entire memcg hierarchy (i.e. flush the root).
- Only one flusher is allowed at a time. If someone else tries to flush
concurrently, they skip and return immediately.
- A periodic flusher flushes all the stats every 2 seconds.
The reason this approach is followed is because all flushes are serialized
by a global rstat spinlock. On the memcg side, flushing is invoked from
userspace reads as well as in-kernel flushers (e.g. reclaim, refault,
etc). This approach aims to avoid serializing all flushers on the global
lock, which can cause a significant performance hit under high
concurrency.
This approach has the following problems:
- Occasionally a userspace read of the stats of a non-root cgroup will
be too expensive as it has to flush the entire hierarchy [1].
- Sometimes the stats accuracy are compromised if there is an ongoing
flush, and we skip and return before the subtree of interest is
actually flushed, yielding stale stats (by up to 2s due to periodic
flushing). This is more visible when reading stats from userspace,
but can also affect in-kernel flushers.
The latter problem is particulary a concern when userspace reads stats
after an event occurs, but gets stats from before the event. Examples:
- When memory usage / pressure spikes, a userspace OOM handler may look
at the stats of different memcgs to select a victim based on various
heuristics (e.g. how much private memory will be freed by killing
this). Reading stale stats from before the usage spike in this case
may cause a wrongful OOM kill.
- A proactive reclaimer may read the stats after writing to
memory.reclaim to measure the success of the reclaim operation. Stale
stats from before reclaim may give a false negative.
- Reading the stats of a parent and a child memcg may be inconsistent
(child larger than parent), if the flush doesn't happen when the
parent is read, but happens when the child is read.
As for in-kernel flushers, they will occasionally get stale stats. No
regressions are currently known from this, but if there are regressions,
they would be very difficult to debug and link to the source of the
problem.
This patch aims to fix these problems by restoring subtree flushing, and
removing the unified/coalesced flushing logic that skips flushing if there
is an ongoing flush. This change would introduce a significant regression
with global stats flushing thresholds. With per-memcg stats flushing
thresholds, this seems to perform really well. The thresholds protect the
underlying lock from unnecessary contention.
This patch was tested in two ways to ensure the latency of flushing is
up to par, on a machine with 384 cpus:
- A synthetic test with 5000 concurrent workers in 500 cgroups doing
allocations and reclaim, as well as 1000 readers for memory.stat
(variation of [2]). No regressions were noticed in the total runtime.
Note that significant regressions in this test are observed with
global stats thresholds, but not with per-memcg thresholds.
- A synthetic stress test for concurrently reading memcg stats while
memory allocation/freeing workers are running in the background,
provided by Wei Xu [3]. With 250k threads reading the stats every
100ms in 50k cgroups, 99.9% of reads take <= 50us. Less than 0.01%
of reads take more than 1ms, and no reads take more than 100ms.
[1] https://lore.kernel.org/lkml/CABWYdi0c6__rh-K7dcM_pkf9BJdTRtAU08M43KO9ME4-dsgfoQ@mail.gmail.com/
[2] https://lore.kernel.org/lkml/CAJD7tka13M-zVZTyQJYL1iUAYvuQ1fcHbCjcOBZcz6POYTV-4g@mail.gmail.com/
[3] https://lore.kernel.org/lkml/CAAPL-u9D2b=iF5Lf_cRnKxUfkiEe0AMDTu6yhrUAzX0b6a6rDg@mail.gmail.com/
[akpm@linux-foundation.org: fix mm/zswap.c]
[yosryahmed@google.com: remove stats flushing mutex]
Link: https://lkml.kernel.org/r/CAJD7tkZgP3m-VVPn+fF_YuvXeQYK=tZZjJHj=dzD=CcSSpp2qg@mail.gmail.com
Link: https://lkml.kernel.org/r/20231129032154.3710765-6-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Tested-by: Domenico Cerasuolo <cerasuolodomenico@gmail.com>
Acked-by: Shakeel Butt <shakeelb@google.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: Greg Thelen <gthelen@google.com>
Cc: Ivan Babrou <ivan@cloudflare.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Michal Koutny <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Tejun Heo <tj@kernel.org>
Cc: Waiman Long <longman@redhat.com>
Cc: Wei Xu <weixugc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-11-29 03:21:53 +00:00
|
|
|
|
mm: memcontrol: dump memory.stat during cgroup OOM
The current cgroup OOM memory info dump doesn't include all the memory
we are tracking, nor does it give insight into what the VM tried to do
leading up to the OOM. All that useful info is in memory.stat.
Furthermore, the recursive printing for every child cgroup can
generate absurd amounts of data on the console for larger cgroup
trees, and it's not like we provide a per-cgroup breakdown during
global OOM kills.
When an OOM kill is triggered, print one set of recursive memory.stat
items at the level whose limit triggered the OOM condition.
Example output:
stress invoked oom-killer: gfp_mask=0x100cca(GFP_HIGHUSER_MOVABLE), order=0, oom_score_adj=0
CPU: 2 PID: 210 Comm: stress Not tainted 5.2.0-rc2-mm1-00247-g47d49835983c #135
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.12.0-20181126_142135-anatol 04/01/2014
Call Trace:
dump_stack+0x46/0x60
dump_header+0x4c/0x2d0
oom_kill_process.cold.10+0xb/0x10
out_of_memory+0x200/0x270
? try_to_free_mem_cgroup_pages+0xdf/0x130
mem_cgroup_out_of_memory+0xb7/0xc0
try_charge+0x680/0x6f0
mem_cgroup_try_charge+0xb5/0x160
__add_to_page_cache_locked+0xc6/0x300
? list_lru_destroy+0x80/0x80
add_to_page_cache_lru+0x45/0xc0
pagecache_get_page+0x11b/0x290
filemap_fault+0x458/0x6d0
ext4_filemap_fault+0x27/0x36
__do_fault+0x2f/0xb0
__handle_mm_fault+0x9c5/0x1140
? apic_timer_interrupt+0xa/0x20
handle_mm_fault+0xc5/0x180
__do_page_fault+0x1ab/0x440
? page_fault+0x8/0x30
page_fault+0x1e/0x30
RIP: 0033:0x55c32167fc10
Code: Bad RIP value.
RSP: 002b:00007fff1d031c50 EFLAGS: 00010206
RAX: 000000000dc00000 RBX: 00007fd2db000010 RCX: 00007fd2db000010
RDX: 0000000000000000 RSI: 0000000010001000 RDI: 0000000000000000
RBP: 000055c321680a54 R08: 00000000ffffffff R09: 0000000000000000
R10: 0000000000000022 R11: 0000000000000246 R12: ffffffffffffffff
R13: 0000000000000002 R14: 0000000000001000 R15: 0000000010000000
memory: usage 1024kB, limit 1024kB, failcnt 75131
swap: usage 0kB, limit 9007199254740988kB, failcnt 0
Memory cgroup stats for /foo:
anon 0
file 0
kernel_stack 36864
slab 274432
sock 0
shmem 0
file_mapped 0
file_dirty 0
file_writeback 0
anon_thp 0
inactive_anon 126976
active_anon 0
inactive_file 0
active_file 0
unevictable 0
slab_reclaimable 0
slab_unreclaimable 274432
pgfault 59466
pgmajfault 1617
workingset_refault 2145
workingset_activate 0
workingset_nodereclaim 0
pgrefill 98952
pgscan 200060
pgsteal 59340
pgactivate 40095
pgdeactivate 96787
pglazyfree 0
pglazyfreed 0
thp_fault_alloc 0
thp_collapse_alloc 0
Tasks state (memory values in pages):
[ pid ] uid tgid total_vm rss pgtables_bytes swapents oom_score_adj name
[ 200] 0 200 1121 884 53248 29 0 bash
[ 209] 0 209 905 246 45056 19 0 stress
[ 210] 0 210 66442 56 499712 56349 0 stress
oom-kill:constraint=CONSTRAINT_NONE,nodemask=(null),oom_memcg=/foo,task_memcg=/foo,task=stress,pid=210,uid=0
Memory cgroup out of memory: Killed process 210 (stress) total-vm:265768kB, anon-rss:0kB, file-rss:224kB, shmem-rss:0kB
oom_reaper: reaped process 210 (stress), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
[hannes@cmpxchg.org: s/kvmalloc/kmalloc/ per Michal]
Link: http://lkml.kernel.org/r/20190605161133.GA12453@cmpxchg.org
Link: http://lkml.kernel.org/r/20190604210509.9744-1-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-11 20:55:59 -07:00
|
|
|
|
2020-10-13 16:52:59 -07:00
|
|
|
|
|
|
|
|
|
mm: memcontrol: dump memory.stat during cgroup OOM
The current cgroup OOM memory info dump doesn't include all the memory
we are tracking, nor does it give insight into what the VM tried to do
leading up to the OOM. All that useful info is in memory.stat.
Furthermore, the recursive printing for every child cgroup can
generate absurd amounts of data on the console for larger cgroup
trees, and it's not like we provide a per-cgroup breakdown during
global OOM kills.
When an OOM kill is triggered, print one set of recursive memory.stat
items at the level whose limit triggered the OOM condition.
Example output:
stress invoked oom-killer: gfp_mask=0x100cca(GFP_HIGHUSER_MOVABLE), order=0, oom_score_adj=0
CPU: 2 PID: 210 Comm: stress Not tainted 5.2.0-rc2-mm1-00247-g47d49835983c #135
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.12.0-20181126_142135-anatol 04/01/2014
Call Trace:
dump_stack+0x46/0x60
dump_header+0x4c/0x2d0
oom_kill_process.cold.10+0xb/0x10
out_of_memory+0x200/0x270
? try_to_free_mem_cgroup_pages+0xdf/0x130
mem_cgroup_out_of_memory+0xb7/0xc0
try_charge+0x680/0x6f0
mem_cgroup_try_charge+0xb5/0x160
__add_to_page_cache_locked+0xc6/0x300
? list_lru_destroy+0x80/0x80
add_to_page_cache_lru+0x45/0xc0
pagecache_get_page+0x11b/0x290
filemap_fault+0x458/0x6d0
ext4_filemap_fault+0x27/0x36
__do_fault+0x2f/0xb0
__handle_mm_fault+0x9c5/0x1140
? apic_timer_interrupt+0xa/0x20
handle_mm_fault+0xc5/0x180
__do_page_fault+0x1ab/0x440
? page_fault+0x8/0x30
page_fault+0x1e/0x30
RIP: 0033:0x55c32167fc10
Code: Bad RIP value.
RSP: 002b:00007fff1d031c50 EFLAGS: 00010206
RAX: 000000000dc00000 RBX: 00007fd2db000010 RCX: 00007fd2db000010
RDX: 0000000000000000 RSI: 0000000010001000 RDI: 0000000000000000
RBP: 000055c321680a54 R08: 00000000ffffffff R09: 0000000000000000
R10: 0000000000000022 R11: 0000000000000246 R12: ffffffffffffffff
R13: 0000000000000002 R14: 0000000000001000 R15: 0000000010000000
memory: usage 1024kB, limit 1024kB, failcnt 75131
swap: usage 0kB, limit 9007199254740988kB, failcnt 0
Memory cgroup stats for /foo:
anon 0
file 0
kernel_stack 36864
slab 274432
sock 0
shmem 0
file_mapped 0
file_dirty 0
file_writeback 0
anon_thp 0
inactive_anon 126976
active_anon 0
inactive_file 0
active_file 0
unevictable 0
slab_reclaimable 0
slab_unreclaimable 274432
pgfault 59466
pgmajfault 1617
workingset_refault 2145
workingset_activate 0
workingset_nodereclaim 0
pgrefill 98952
pgscan 200060
pgsteal 59340
pgactivate 40095
pgdeactivate 96787
pglazyfree 0
pglazyfreed 0
thp_fault_alloc 0
thp_collapse_alloc 0
Tasks state (memory values in pages):
[ pid ] uid tgid total_vm rss pgtables_bytes swapents oom_score_adj name
[ 200] 0 200 1121 884 53248 29 0 bash
[ 209] 0 209 905 246 45056 19 0 stress
[ 210] 0 210 66442 56 499712 56349 0 stress
oom-kill:constraint=CONSTRAINT_NONE,nodemask=(null),oom_memcg=/foo,task_memcg=/foo,task=stress,pid=210,uid=0
Memory cgroup out of memory: Killed process 210 (stress) total-vm:265768kB, anon-rss:0kB, file-rss:224kB, shmem-rss:0kB
oom_reaper: reaped process 210 (stress), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
[hannes@cmpxchg.org: s/kvmalloc/kmalloc/ per Michal]
Link: http://lkml.kernel.org/r/20190605161133.GA12453@cmpxchg.org
Link: http://lkml.kernel.org/r/20190604210509.9744-1-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-11 20:55:59 -07:00
|
|
|
|
memcg/hugetlb: add hugeTLB counters to memcg
This patch introduces a new counter to memory.stat that tracks hugeTLB
usage, only if hugeTLB accounting is done to memory.current. This feature
is enabled the same way hugeTLB accounting is enabled, via the
memory_hugetlb_accounting mount flag for cgroupsv2.
1. Why is this patch necessary?
Currently, memcg hugeTLB accounting is an opt-in feature [1] that adds
hugeTLB usage to memory.current. However, the metric is not reported in
memory.stat. Given that users often interpret memory.stat as a breakdown
of the value reported in memory.current, the disparity between the two
reports can be confusing. This patch solves this problem by including the
metric in memory.stat as well, but only if it is also reported in
memory.current (it would also be confusing if the value was reported in
memory.stat, but not in memory.current)
Aside from the consistency between the two files, we also see benefits in
observability. Userspace might be interested in the hugeTLB footprint of
cgroups for many reasons. For instance, system admins might want to
verify that hugeTLB usage is distributed as expected across tasks: i.e.
memory-intensive tasks are using more hugeTLB pages than tasks that don't
consume a lot of memory, or are seen to fault frequently. Note that this
is separate from wanting to inspect the distribution for limiting purposes
(in which case, hugeTLB controller makes more sense).
2. We already have a hugeTLB controller. Why not use that?
It is true that hugeTLB tracks the exact value that we want. In fact, by
enabling the hugeTLB controller, we get all of the observability benefits
that I mentioned above, and users can check the total hugeTLB usage,
verify if it is distributed as expected, etc.
With this said, there are 2 problems:
(a) They are still not reported in memory.stat, which means the
disparity between the memcg reports are still there.
(b) We cannot reasonably expect users to enable the hugeTLB controller
just for the sake of hugeTLB usage reporting, especially since
they don't have any use for hugeTLB usage enforcing [2].
3. Implementation Details:
In the alloc / free hugetlb functions, we call lruvec_stat_mod_folio
regardless of whether memcg accounts hugetlb. mem_cgroup_commit_charge
which is called from alloc_hugetlb_folio will set memcg for the folio only
if the CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING cgroup mount option is used, so
lruvec_stat_mod_folio accounts per-memcg hugetlb counters only if the
feature is enabled. Regardless of whether memcg accounts for hugetlb, the
newly added global counter is updated and shown in /proc/vmstat.
The global counter is added because vmstats is the preferred framework for
cgroup stats. It makes stat items consistent between global and cgroups.
It also provides a per-node breakdown, which is useful. Because it does
not use cgroup-specific hooks, we also keep generic MM code separate from
memcg code.
[1] https://lore.kernel.org/all/20231006184629.155543-1-nphamcs@gmail.com/
[2] Of course, we can't make a new patch for every feature that can be
duplicated. However, since the existing solution of enabling the
hugeTLB controller is an imperfect solution that still leaves a
discrepancy between memory.stat and memory.curent, I think that it
is reasonable to isolate the feature in this case.
Link: https://lkml.kernel.org/r/20241101204402.1885383-1-joshua.hahnjy@gmail.com
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Nhat Pham <nphamcs@gmail.com>
Suggested-by: Shakeel Butt <shakeel.butt@linux.dev>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Shakeel Butt <shakeel.butt@linux.dev>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Chris Down <chris@chrisdown.name>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Roman Gushchin <roman.gushchin@linux.dev>
Reviewed-by: Nhat Pham <nphamcs@gmail.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Koutný <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-11-01 13:44:02 -07:00
|
|
|
|
|
|
|
|
|
2024-12-11 12:39:49 -08:00
|
|
|
|
memcg/hugetlb: add hugeTLB counters to memcg
This patch introduces a new counter to memory.stat that tracks hugeTLB
usage, only if hugeTLB accounting is done to memory.current. This feature
is enabled the same way hugeTLB accounting is enabled, via the
memory_hugetlb_accounting mount flag for cgroupsv2.
1. Why is this patch necessary?
Currently, memcg hugeTLB accounting is an opt-in feature [1] that adds
hugeTLB usage to memory.current. However, the metric is not reported in
memory.stat. Given that users often interpret memory.stat as a breakdown
of the value reported in memory.current, the disparity between the two
reports can be confusing. This patch solves this problem by including the
metric in memory.stat as well, but only if it is also reported in
memory.current (it would also be confusing if the value was reported in
memory.stat, but not in memory.current)
Aside from the consistency between the two files, we also see benefits in
observability. Userspace might be interested in the hugeTLB footprint of
cgroups for many reasons. For instance, system admins might want to
verify that hugeTLB usage is distributed as expected across tasks: i.e.
memory-intensive tasks are using more hugeTLB pages than tasks that don't
consume a lot of memory, or are seen to fault frequently. Note that this
is separate from wanting to inspect the distribution for limiting purposes
(in which case, hugeTLB controller makes more sense).
2. We already have a hugeTLB controller. Why not use that?
It is true that hugeTLB tracks the exact value that we want. In fact, by
enabling the hugeTLB controller, we get all of the observability benefits
that I mentioned above, and users can check the total hugeTLB usage,
verify if it is distributed as expected, etc.
With this said, there are 2 problems:
(a) They are still not reported in memory.stat, which means the
disparity between the memcg reports are still there.
(b) We cannot reasonably expect users to enable the hugeTLB controller
just for the sake of hugeTLB usage reporting, especially since
they don't have any use for hugeTLB usage enforcing [2].
3. Implementation Details:
In the alloc / free hugetlb functions, we call lruvec_stat_mod_folio
regardless of whether memcg accounts hugetlb. mem_cgroup_commit_charge
which is called from alloc_hugetlb_folio will set memcg for the folio only
if the CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING cgroup mount option is used, so
lruvec_stat_mod_folio accounts per-memcg hugetlb counters only if the
feature is enabled. Regardless of whether memcg accounts for hugetlb, the
newly added global counter is updated and shown in /proc/vmstat.
The global counter is added because vmstats is the preferred framework for
cgroup stats. It makes stat items consistent between global and cgroups.
It also provides a per-node breakdown, which is useful. Because it does
not use cgroup-specific hooks, we also keep generic MM code separate from
memcg code.
[1] https://lore.kernel.org/all/20231006184629.155543-1-nphamcs@gmail.com/
[2] Of course, we can't make a new patch for every feature that can be
duplicated. However, since the existing solution of enabling the
hugeTLB controller is an imperfect solution that still leaves a
discrepancy between memory.stat and memory.curent, I think that it
is reasonable to isolate the feature in this case.
Link: https://lkml.kernel.org/r/20241101204402.1885383-1-joshua.hahnjy@gmail.com
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Nhat Pham <nphamcs@gmail.com>
Suggested-by: Shakeel Butt <shakeel.butt@linux.dev>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Shakeel Butt <shakeel.butt@linux.dev>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Chris Down <chris@chrisdown.name>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Roman Gushchin <roman.gushchin@linux.dev>
Reviewed-by: Nhat Pham <nphamcs@gmail.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Koutný <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-11-01 13:44:02 -07:00
|
|
|
|
|
|
|
|
|
2021-02-24 12:03:43 -08:00
|
|
|
|
2023-04-28 13:24:05 +00:00
|
|
|
|
mm: memcontrol: dump memory.stat during cgroup OOM
The current cgroup OOM memory info dump doesn't include all the memory
we are tracking, nor does it give insight into what the VM tried to do
leading up to the OOM. All that useful info is in memory.stat.
Furthermore, the recursive printing for every child cgroup can
generate absurd amounts of data on the console for larger cgroup
trees, and it's not like we provide a per-cgroup breakdown during
global OOM kills.
When an OOM kill is triggered, print one set of recursive memory.stat
items at the level whose limit triggered the OOM condition.
Example output:
stress invoked oom-killer: gfp_mask=0x100cca(GFP_HIGHUSER_MOVABLE), order=0, oom_score_adj=0
CPU: 2 PID: 210 Comm: stress Not tainted 5.2.0-rc2-mm1-00247-g47d49835983c #135
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.12.0-20181126_142135-anatol 04/01/2014
Call Trace:
dump_stack+0x46/0x60
dump_header+0x4c/0x2d0
oom_kill_process.cold.10+0xb/0x10
out_of_memory+0x200/0x270
? try_to_free_mem_cgroup_pages+0xdf/0x130
mem_cgroup_out_of_memory+0xb7/0xc0
try_charge+0x680/0x6f0
mem_cgroup_try_charge+0xb5/0x160
__add_to_page_cache_locked+0xc6/0x300
? list_lru_destroy+0x80/0x80
add_to_page_cache_lru+0x45/0xc0
pagecache_get_page+0x11b/0x290
filemap_fault+0x458/0x6d0
ext4_filemap_fault+0x27/0x36
__do_fault+0x2f/0xb0
__handle_mm_fault+0x9c5/0x1140
? apic_timer_interrupt+0xa/0x20
handle_mm_fault+0xc5/0x180
__do_page_fault+0x1ab/0x440
? page_fault+0x8/0x30
page_fault+0x1e/0x30
RIP: 0033:0x55c32167fc10
Code: Bad RIP value.
RSP: 002b:00007fff1d031c50 EFLAGS: 00010206
RAX: 000000000dc00000 RBX: 00007fd2db000010 RCX: 00007fd2db000010
RDX: 0000000000000000 RSI: 0000000010001000 RDI: 0000000000000000
RBP: 000055c321680a54 R08: 00000000ffffffff R09: 0000000000000000
R10: 0000000000000022 R11: 0000000000000246 R12: ffffffffffffffff
R13: 0000000000000002 R14: 0000000000001000 R15: 0000000010000000
memory: usage 1024kB, limit 1024kB, failcnt 75131
swap: usage 0kB, limit 9007199254740988kB, failcnt 0
Memory cgroup stats for /foo:
anon 0
file 0
kernel_stack 36864
slab 274432
sock 0
shmem 0
file_mapped 0
file_dirty 0
file_writeback 0
anon_thp 0
inactive_anon 126976
active_anon 0
inactive_file 0
active_file 0
unevictable 0
slab_reclaimable 0
slab_unreclaimable 274432
pgfault 59466
pgmajfault 1617
workingset_refault 2145
workingset_activate 0
workingset_nodereclaim 0
pgrefill 98952
pgscan 200060
pgsteal 59340
pgactivate 40095
pgdeactivate 96787
pglazyfree 0
pglazyfreed 0
thp_fault_alloc 0
thp_collapse_alloc 0
Tasks state (memory values in pages):
[ pid ] uid tgid total_vm rss pgtables_bytes swapents oom_score_adj name
[ 200] 0 200 1121 884 53248 29 0 bash
[ 209] 0 209 905 246 45056 19 0 stress
[ 210] 0 210 66442 56 499712 56349 0 stress
oom-kill:constraint=CONSTRAINT_NONE,nodemask=(null),oom_memcg=/foo,task_memcg=/foo,task=stress,pid=210,uid=0
Memory cgroup out of memory: Killed process 210 (stress) total-vm:265768kB, anon-rss:0kB, file-rss:224kB, shmem-rss:0kB
oom_reaper: reaped process 210 (stress), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
[hannes@cmpxchg.org: s/kvmalloc/kmalloc/ per Michal]
Link: http://lkml.kernel.org/r/20190605161133.GA12453@cmpxchg.org
Link: http://lkml.kernel.org/r/20190604210509.9744-1-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-11 20:55:59 -07:00
|
|
|
|
2020-10-13 16:52:59 -07:00
|
|
|
|
2021-02-24 12:03:43 -08:00
|
|
|
|
|
|
|
|
|
2023-04-28 13:24:05 +00:00
|
|
|
|
2020-10-13 16:52:59 -07:00
|
|
|
|
|
|
|
|
|
mm: memcontrol: dump memory.stat during cgroup OOM
The current cgroup OOM memory info dump doesn't include all the memory
we are tracking, nor does it give insight into what the VM tried to do
leading up to the OOM. All that useful info is in memory.stat.
Furthermore, the recursive printing for every child cgroup can
generate absurd amounts of data on the console for larger cgroup
trees, and it's not like we provide a per-cgroup breakdown during
global OOM kills.
When an OOM kill is triggered, print one set of recursive memory.stat
items at the level whose limit triggered the OOM condition.
Example output:
stress invoked oom-killer: gfp_mask=0x100cca(GFP_HIGHUSER_MOVABLE), order=0, oom_score_adj=0
CPU: 2 PID: 210 Comm: stress Not tainted 5.2.0-rc2-mm1-00247-g47d49835983c #135
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.12.0-20181126_142135-anatol 04/01/2014
Call Trace:
dump_stack+0x46/0x60
dump_header+0x4c/0x2d0
oom_kill_process.cold.10+0xb/0x10
out_of_memory+0x200/0x270
? try_to_free_mem_cgroup_pages+0xdf/0x130
mem_cgroup_out_of_memory+0xb7/0xc0
try_charge+0x680/0x6f0
mem_cgroup_try_charge+0xb5/0x160
__add_to_page_cache_locked+0xc6/0x300
? list_lru_destroy+0x80/0x80
add_to_page_cache_lru+0x45/0xc0
pagecache_get_page+0x11b/0x290
filemap_fault+0x458/0x6d0
ext4_filemap_fault+0x27/0x36
__do_fault+0x2f/0xb0
__handle_mm_fault+0x9c5/0x1140
? apic_timer_interrupt+0xa/0x20
handle_mm_fault+0xc5/0x180
__do_page_fault+0x1ab/0x440
? page_fault+0x8/0x30
page_fault+0x1e/0x30
RIP: 0033:0x55c32167fc10
Code: Bad RIP value.
RSP: 002b:00007fff1d031c50 EFLAGS: 00010206
RAX: 000000000dc00000 RBX: 00007fd2db000010 RCX: 00007fd2db000010
RDX: 0000000000000000 RSI: 0000000010001000 RDI: 0000000000000000
RBP: 000055c321680a54 R08: 00000000ffffffff R09: 0000000000000000
R10: 0000000000000022 R11: 0000000000000246 R12: ffffffffffffffff
R13: 0000000000000002 R14: 0000000000001000 R15: 0000000010000000
memory: usage 1024kB, limit 1024kB, failcnt 75131
swap: usage 0kB, limit 9007199254740988kB, failcnt 0
Memory cgroup stats for /foo:
anon 0
file 0
kernel_stack 36864
slab 274432
sock 0
shmem 0
file_mapped 0
file_dirty 0
file_writeback 0
anon_thp 0
inactive_anon 126976
active_anon 0
inactive_file 0
active_file 0
unevictable 0
slab_reclaimable 0
slab_unreclaimable 274432
pgfault 59466
pgmajfault 1617
workingset_refault 2145
workingset_activate 0
workingset_nodereclaim 0
pgrefill 98952
pgscan 200060
pgsteal 59340
pgactivate 40095
pgdeactivate 96787
pglazyfree 0
pglazyfreed 0
thp_fault_alloc 0
thp_collapse_alloc 0
Tasks state (memory values in pages):
[ pid ] uid tgid total_vm rss pgtables_bytes swapents oom_score_adj name
[ 200] 0 200 1121 884 53248 29 0 bash
[ 209] 0 209 905 246 45056 19 0 stress
[ 210] 0 210 66442 56 499712 56349 0 stress
oom-kill:constraint=CONSTRAINT_NONE,nodemask=(null),oom_memcg=/foo,task_memcg=/foo,task=stress,pid=210,uid=0
Memory cgroup out of memory: Killed process 210 (stress) total-vm:265768kB, anon-rss:0kB, file-rss:224kB, shmem-rss:0kB
oom_reaper: reaped process 210 (stress), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
[hannes@cmpxchg.org: s/kvmalloc/kmalloc/ per Michal]
Link: http://lkml.kernel.org/r/20190605161133.GA12453@cmpxchg.org
Link: http://lkml.kernel.org/r/20190604210509.9744-1-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-11 20:55:59 -07:00
|
|
|
|
|
|
|
|
|
2023-04-28 13:24:05 +00:00
|
|
|
|
mm: memcontrol: dump memory.stat during cgroup OOM
The current cgroup OOM memory info dump doesn't include all the memory
we are tracking, nor does it give insight into what the VM tried to do
leading up to the OOM. All that useful info is in memory.stat.
Furthermore, the recursive printing for every child cgroup can
generate absurd amounts of data on the console for larger cgroup
trees, and it's not like we provide a per-cgroup breakdown during
global OOM kills.
When an OOM kill is triggered, print one set of recursive memory.stat
items at the level whose limit triggered the OOM condition.
Example output:
stress invoked oom-killer: gfp_mask=0x100cca(GFP_HIGHUSER_MOVABLE), order=0, oom_score_adj=0
CPU: 2 PID: 210 Comm: stress Not tainted 5.2.0-rc2-mm1-00247-g47d49835983c #135
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.12.0-20181126_142135-anatol 04/01/2014
Call Trace:
dump_stack+0x46/0x60
dump_header+0x4c/0x2d0
oom_kill_process.cold.10+0xb/0x10
out_of_memory+0x200/0x270
? try_to_free_mem_cgroup_pages+0xdf/0x130
mem_cgroup_out_of_memory+0xb7/0xc0
try_charge+0x680/0x6f0
mem_cgroup_try_charge+0xb5/0x160
__add_to_page_cache_locked+0xc6/0x300
? list_lru_destroy+0x80/0x80
add_to_page_cache_lru+0x45/0xc0
pagecache_get_page+0x11b/0x290
filemap_fault+0x458/0x6d0
ext4_filemap_fault+0x27/0x36
__do_fault+0x2f/0xb0
__handle_mm_fault+0x9c5/0x1140
? apic_timer_interrupt+0xa/0x20
handle_mm_fault+0xc5/0x180
__do_page_fault+0x1ab/0x440
? page_fault+0x8/0x30
page_fault+0x1e/0x30
RIP: 0033:0x55c32167fc10
Code: Bad RIP value.
RSP: 002b:00007fff1d031c50 EFLAGS: 00010206
RAX: 000000000dc00000 RBX: 00007fd2db000010 RCX: 00007fd2db000010
RDX: 0000000000000000 RSI: 0000000010001000 RDI: 0000000000000000
RBP: 000055c321680a54 R08: 00000000ffffffff R09: 0000000000000000
R10: 0000000000000022 R11: 0000000000000246 R12: ffffffffffffffff
R13: 0000000000000002 R14: 0000000000001000 R15: 0000000010000000
memory: usage 1024kB, limit 1024kB, failcnt 75131
swap: usage 0kB, limit 9007199254740988kB, failcnt 0
Memory cgroup stats for /foo:
anon 0
file 0
kernel_stack 36864
slab 274432
sock 0
shmem 0
file_mapped 0
file_dirty 0
file_writeback 0
anon_thp 0
inactive_anon 126976
active_anon 0
inactive_file 0
active_file 0
unevictable 0
slab_reclaimable 0
slab_unreclaimable 274432
pgfault 59466
pgmajfault 1617
workingset_refault 2145
workingset_activate 0
workingset_nodereclaim 0
pgrefill 98952
pgscan 200060
pgsteal 59340
pgactivate 40095
pgdeactivate 96787
pglazyfree 0
pglazyfreed 0
thp_fault_alloc 0
thp_collapse_alloc 0
Tasks state (memory values in pages):
[ pid ] uid tgid total_vm rss pgtables_bytes swapents oom_score_adj name
[ 200] 0 200 1121 884 53248 29 0 bash
[ 209] 0 209 905 246 45056 19 0 stress
[ 210] 0 210 66442 56 499712 56349 0 stress
oom-kill:constraint=CONSTRAINT_NONE,nodemask=(null),oom_memcg=/foo,task_memcg=/foo,task=stress,pid=210,uid=0
Memory cgroup out of memory: Killed process 210 (stress) total-vm:265768kB, anon-rss:0kB, file-rss:224kB, shmem-rss:0kB
oom_reaper: reaped process 210 (stress), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
[hannes@cmpxchg.org: s/kvmalloc/kmalloc/ per Michal]
Link: http://lkml.kernel.org/r/20190605161133.GA12453@cmpxchg.org
Link: http://lkml.kernel.org/r/20190604210509.9744-1-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-11 20:55:59 -07:00
|
|
|
|
2022-10-26 14:01:33 -04:00
|
|
|
|
2025-03-18 15:58:32 +08:00
|
|
|
|
2022-10-26 14:01:33 -04:00
|
|
|
|
2023-04-28 13:24:05 +00:00
|
|
|
|
mm: memcontrol: dump memory.stat during cgroup OOM
The current cgroup OOM memory info dump doesn't include all the memory
we are tracking, nor does it give insight into what the VM tried to do
leading up to the OOM. All that useful info is in memory.stat.
Furthermore, the recursive printing for every child cgroup can
generate absurd amounts of data on the console for larger cgroup
trees, and it's not like we provide a per-cgroup breakdown during
global OOM kills.
When an OOM kill is triggered, print one set of recursive memory.stat
items at the level whose limit triggered the OOM condition.
Example output:
stress invoked oom-killer: gfp_mask=0x100cca(GFP_HIGHUSER_MOVABLE), order=0, oom_score_adj=0
CPU: 2 PID: 210 Comm: stress Not tainted 5.2.0-rc2-mm1-00247-g47d49835983c #135
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.12.0-20181126_142135-anatol 04/01/2014
Call Trace:
dump_stack+0x46/0x60
dump_header+0x4c/0x2d0
oom_kill_process.cold.10+0xb/0x10
out_of_memory+0x200/0x270
? try_to_free_mem_cgroup_pages+0xdf/0x130
mem_cgroup_out_of_memory+0xb7/0xc0
try_charge+0x680/0x6f0
mem_cgroup_try_charge+0xb5/0x160
__add_to_page_cache_locked+0xc6/0x300
? list_lru_destroy+0x80/0x80
add_to_page_cache_lru+0x45/0xc0
pagecache_get_page+0x11b/0x290
filemap_fault+0x458/0x6d0
ext4_filemap_fault+0x27/0x36
__do_fault+0x2f/0xb0
__handle_mm_fault+0x9c5/0x1140
? apic_timer_interrupt+0xa/0x20
handle_mm_fault+0xc5/0x180
__do_page_fault+0x1ab/0x440
? page_fault+0x8/0x30
page_fault+0x1e/0x30
RIP: 0033:0x55c32167fc10
Code: Bad RIP value.
RSP: 002b:00007fff1d031c50 EFLAGS: 00010206
RAX: 000000000dc00000 RBX: 00007fd2db000010 RCX: 00007fd2db000010
RDX: 0000000000000000 RSI: 0000000010001000 RDI: 0000000000000000
RBP: 000055c321680a54 R08: 00000000ffffffff R09: 0000000000000000
R10: 0000000000000022 R11: 0000000000000246 R12: ffffffffffffffff
R13: 0000000000000002 R14: 0000000000001000 R15: 0000000010000000
memory: usage 1024kB, limit 1024kB, failcnt 75131
swap: usage 0kB, limit 9007199254740988kB, failcnt 0
Memory cgroup stats for /foo:
anon 0
file 0
kernel_stack 36864
slab 274432
sock 0
shmem 0
file_mapped 0
file_dirty 0
file_writeback 0
anon_thp 0
inactive_anon 126976
active_anon 0
inactive_file 0
active_file 0
unevictable 0
slab_reclaimable 0
slab_unreclaimable 274432
pgfault 59466
pgmajfault 1617
workingset_refault 2145
workingset_activate 0
workingset_nodereclaim 0
pgrefill 98952
pgscan 200060
pgsteal 59340
pgactivate 40095
pgdeactivate 96787
pglazyfree 0
pglazyfreed 0
thp_fault_alloc 0
thp_collapse_alloc 0
Tasks state (memory values in pages):
[ pid ] uid tgid total_vm rss pgtables_bytes swapents oom_score_adj name
[ 200] 0 200 1121 884 53248 29 0 bash
[ 209] 0 209 905 246 45056 19 0 stress
[ 210] 0 210 66442 56 499712 56349 0 stress
oom-kill:constraint=CONSTRAINT_NONE,nodemask=(null),oom_memcg=/foo,task_memcg=/foo,task=stress,pid=210,uid=0
Memory cgroup out of memory: Killed process 210 (stress) total-vm:265768kB, anon-rss:0kB, file-rss:224kB, shmem-rss:0kB
oom_reaper: reaped process 210 (stress), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
[hannes@cmpxchg.org: s/kvmalloc/kmalloc/ per Michal]
Link: http://lkml.kernel.org/r/20190605161133.GA12453@cmpxchg.org
Link: http://lkml.kernel.org/r/20190604210509.9744-1-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-11 20:55:59 -07:00
|
|
|
|
2022-10-26 14:01:33 -04:00
|
|
|
|
2025-03-18 15:58:32 +08:00
|
|
|
|
2022-10-26 14:01:33 -04:00
|
|
|
|
mm: memcontrol: dump memory.stat during cgroup OOM
The current cgroup OOM memory info dump doesn't include all the memory
we are tracking, nor does it give insight into what the VM tried to do
leading up to the OOM. All that useful info is in memory.stat.
Furthermore, the recursive printing for every child cgroup can
generate absurd amounts of data on the console for larger cgroup
trees, and it's not like we provide a per-cgroup breakdown during
global OOM kills.
When an OOM kill is triggered, print one set of recursive memory.stat
items at the level whose limit triggered the OOM condition.
Example output:
stress invoked oom-killer: gfp_mask=0x100cca(GFP_HIGHUSER_MOVABLE), order=0, oom_score_adj=0
CPU: 2 PID: 210 Comm: stress Not tainted 5.2.0-rc2-mm1-00247-g47d49835983c #135
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.12.0-20181126_142135-anatol 04/01/2014
Call Trace:
dump_stack+0x46/0x60
dump_header+0x4c/0x2d0
oom_kill_process.cold.10+0xb/0x10
out_of_memory+0x200/0x270
? try_to_free_mem_cgroup_pages+0xdf/0x130
mem_cgroup_out_of_memory+0xb7/0xc0
try_charge+0x680/0x6f0
mem_cgroup_try_charge+0xb5/0x160
__add_to_page_cache_locked+0xc6/0x300
? list_lru_destroy+0x80/0x80
add_to_page_cache_lru+0x45/0xc0
pagecache_get_page+0x11b/0x290
filemap_fault+0x458/0x6d0
ext4_filemap_fault+0x27/0x36
__do_fault+0x2f/0xb0
__handle_mm_fault+0x9c5/0x1140
? apic_timer_interrupt+0xa/0x20
handle_mm_fault+0xc5/0x180
__do_page_fault+0x1ab/0x440
? page_fault+0x8/0x30
page_fault+0x1e/0x30
RIP: 0033:0x55c32167fc10
Code: Bad RIP value.
RSP: 002b:00007fff1d031c50 EFLAGS: 00010206
RAX: 000000000dc00000 RBX: 00007fd2db000010 RCX: 00007fd2db000010
RDX: 0000000000000000 RSI: 0000000010001000 RDI: 0000000000000000
RBP: 000055c321680a54 R08: 00000000ffffffff R09: 0000000000000000
R10: 0000000000000022 R11: 0000000000000246 R12: ffffffffffffffff
R13: 0000000000000002 R14: 0000000000001000 R15: 0000000010000000
memory: usage 1024kB, limit 1024kB, failcnt 75131
swap: usage 0kB, limit 9007199254740988kB, failcnt 0
Memory cgroup stats for /foo:
anon 0
file 0
kernel_stack 36864
slab 274432
sock 0
shmem 0
file_mapped 0
file_dirty 0
file_writeback 0
anon_thp 0
inactive_anon 126976
active_anon 0
inactive_file 0
active_file 0
unevictable 0
slab_reclaimable 0
slab_unreclaimable 274432
pgfault 59466
pgmajfault 1617
workingset_refault 2145
workingset_activate 0
workingset_nodereclaim 0
pgrefill 98952
pgscan 200060
pgsteal 59340
pgactivate 40095
pgdeactivate 96787
pglazyfree 0
pglazyfreed 0
thp_fault_alloc 0
thp_collapse_alloc 0
Tasks state (memory values in pages):
[ pid ] uid tgid total_vm rss pgtables_bytes swapents oom_score_adj name
[ 200] 0 200 1121 884 53248 29 0 bash
[ 209] 0 209 905 246 45056 19 0 stress
[ 210] 0 210 66442 56 499712 56349 0 stress
oom-kill:constraint=CONSTRAINT_NONE,nodemask=(null),oom_memcg=/foo,task_memcg=/foo,task=stress,pid=210,uid=0
Memory cgroup out of memory: Killed process 210 (stress) total-vm:265768kB, anon-rss:0kB, file-rss:224kB, shmem-rss:0kB
oom_reaper: reaped process 210 (stress), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
[hannes@cmpxchg.org: s/kvmalloc/kmalloc/ per Michal]
Link: http://lkml.kernel.org/r/20190605161133.GA12453@cmpxchg.org
Link: http://lkml.kernel.org/r/20190604210509.9744-1-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-11 20:55:59 -07:00
|
|
|
|
2022-09-07 04:35:37 +00:00
|
|
|
|
2024-08-14 22:04:53 -07:00
|
|
|
|
2022-09-07 04:35:37 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-08-14 22:04:53 -07:00
|
|
|
|
2023-04-28 13:24:05 +00:00
|
|
|
|
mm: memcontrol: add {pgscan,pgsteal}_{kswapd,direct} items in memory.stat of cgroup v2
There are already statistics of {pgscan,pgsteal}_kswapd and
{pgscan,pgsteal}_direct of memcg event here, but now only the sum of the
two is displayed in memory.stat of cgroup v2.
In order to obtain more accurate information during monitoring and
debugging, and to align with the display in /proc/vmstat, it better to
display {pgscan,pgsteal}_kswapd and {pgscan,pgsteal}_direct separately.
Also, for forward compatibility, we still display pgscan and pgsteal items
so that it won't break existing applications.
[zhengqi.arch@bytedance.com: add comment for memcg_vm_event_stat (suggested by Michal)]
Link: https://lkml.kernel.org/r/20220606154028.55030-1-zhengqi.arch@bytedance.com
[zhengqi.arch@bytedance.com: fix the doc, thanks to Johannes]
Link: https://lkml.kernel.org/r/20220607064803.79363-1-zhengqi.arch@bytedance.com
Link: https://lkml.kernel.org/r/20220604082209.55174-1-zhengqi.arch@bytedance.com
Signed-off-by: Qi Zheng <zhengqi.arch@bytedance.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Roman Gushchin <roman.gushchin@linux.dev>
Acked-by: Muchun Song <songmuchun@bytedance.com>
Acked-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Muchun Song <songmuchun@bytedance.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-06-04 16:22:09 +08:00
|
|
|
|
|
|
|
|
|
2022-09-07 04:35:37 +00:00
|
|
|
|
mm: memcontrol: dump memory.stat during cgroup OOM
The current cgroup OOM memory info dump doesn't include all the memory
we are tracking, nor does it give insight into what the VM tried to do
leading up to the OOM. All that useful info is in memory.stat.
Furthermore, the recursive printing for every child cgroup can
generate absurd amounts of data on the console for larger cgroup
trees, and it's not like we provide a per-cgroup breakdown during
global OOM kills.
When an OOM kill is triggered, print one set of recursive memory.stat
items at the level whose limit triggered the OOM condition.
Example output:
stress invoked oom-killer: gfp_mask=0x100cca(GFP_HIGHUSER_MOVABLE), order=0, oom_score_adj=0
CPU: 2 PID: 210 Comm: stress Not tainted 5.2.0-rc2-mm1-00247-g47d49835983c #135
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.12.0-20181126_142135-anatol 04/01/2014
Call Trace:
dump_stack+0x46/0x60
dump_header+0x4c/0x2d0
oom_kill_process.cold.10+0xb/0x10
out_of_memory+0x200/0x270
? try_to_free_mem_cgroup_pages+0xdf/0x130
mem_cgroup_out_of_memory+0xb7/0xc0
try_charge+0x680/0x6f0
mem_cgroup_try_charge+0xb5/0x160
__add_to_page_cache_locked+0xc6/0x300
? list_lru_destroy+0x80/0x80
add_to_page_cache_lru+0x45/0xc0
pagecache_get_page+0x11b/0x290
filemap_fault+0x458/0x6d0
ext4_filemap_fault+0x27/0x36
__do_fault+0x2f/0xb0
__handle_mm_fault+0x9c5/0x1140
? apic_timer_interrupt+0xa/0x20
handle_mm_fault+0xc5/0x180
__do_page_fault+0x1ab/0x440
? page_fault+0x8/0x30
page_fault+0x1e/0x30
RIP: 0033:0x55c32167fc10
Code: Bad RIP value.
RSP: 002b:00007fff1d031c50 EFLAGS: 00010206
RAX: 000000000dc00000 RBX: 00007fd2db000010 RCX: 00007fd2db000010
RDX: 0000000000000000 RSI: 0000000010001000 RDI: 0000000000000000
RBP: 000055c321680a54 R08: 00000000ffffffff R09: 0000000000000000
R10: 0000000000000022 R11: 0000000000000246 R12: ffffffffffffffff
R13: 0000000000000002 R14: 0000000000001000 R15: 0000000010000000
memory: usage 1024kB, limit 1024kB, failcnt 75131
swap: usage 0kB, limit 9007199254740988kB, failcnt 0
Memory cgroup stats for /foo:
anon 0
file 0
kernel_stack 36864
slab 274432
sock 0
shmem 0
file_mapped 0
file_dirty 0
file_writeback 0
anon_thp 0
inactive_anon 126976
active_anon 0
inactive_file 0
active_file 0
unevictable 0
slab_reclaimable 0
slab_unreclaimable 274432
pgfault 59466
pgmajfault 1617
workingset_refault 2145
workingset_activate 0
workingset_nodereclaim 0
pgrefill 98952
pgscan 200060
pgsteal 59340
pgactivate 40095
pgdeactivate 96787
pglazyfree 0
pglazyfreed 0
thp_fault_alloc 0
thp_collapse_alloc 0
Tasks state (memory values in pages):
[ pid ] uid tgid total_vm rss pgtables_bytes swapents oom_score_adj name
[ 200] 0 200 1121 884 53248 29 0 bash
[ 209] 0 209 905 246 45056 19 0 stress
[ 210] 0 210 66442 56 499712 56349 0 stress
oom-kill:constraint=CONSTRAINT_NONE,nodemask=(null),oom_memcg=/foo,task_memcg=/foo,task=stress,pid=210,uid=0
Memory cgroup out of memory: Killed process 210 (stress) total-vm:265768kB, anon-rss:0kB, file-rss:224kB, shmem-rss:0kB
oom_reaper: reaped process 210 (stress), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
[hannes@cmpxchg.org: s/kvmalloc/kmalloc/ per Michal]
Link: http://lkml.kernel.org/r/20190605161133.GA12453@cmpxchg.org
Link: http://lkml.kernel.org/r/20190604210509.9744-1-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-11 20:55:59 -07:00
|
|
|
|
2017-05-03 14:55:13 -07:00
|
|
|
|
memcg: dump memory.stat during cgroup OOM for v1
Patch series "memcg: OOM log improvements", v2.
This short patch series brings back some cgroup v1 stats in OOM logs
that were unnecessarily changed before. It also makes memcg OOM logs
less reliant on printk() internals.
This patch (of 2):
Commit c8713d0b2312 ("mm: memcontrol: dump memory.stat during cgroup OOM")
made sure we dump all the stats in memory.stat during a cgroup OOM, but it
also introduced a slight behavioral change. The code used to print the
non-hierarchical v1 cgroup stats for the entire cgroup subtree, now it
only prints the v2 cgroup stats for the cgroup under OOM.
For cgroup v1 users, this introduces a few problems:
(a) The non-hierarchical stats of the memcg under OOM are no longer
shown.
(b) A couple of v1-only stats (e.g. pgpgin, pgpgout) are no longer
shown.
(c) We show the list of cgroup v2 stats, even in cgroup v1. This list
of stats is not tracked with v1 in mind. While most of the stats seem
to be working on v1, there may be some stats that are not fully or
correctly tracked.
Although OOM log is not set in stone, we should not change it for no
reason. When upgrading the kernel version to a version including commit
c8713d0b2312 ("mm: memcontrol: dump memory.stat during cgroup OOM"), these
behavioral changes are noticed in cgroup v1.
The fix is simple. Commit c8713d0b2312 ("mm: memcontrol: dump memory.stat
during cgroup OOM") separated stats formatting from stats display for v2,
to reuse the stats formatting in the OOM logs. Do the same for v1.
Move the v2 specific formatting from memory_stat_format() to
memcg_stat_format(), add memcg1_stat_format() for v1, and make
memory_stat_format() select between them based on cgroup version. Since
memory_stat_show() now works for both v1 & v2, drop memcg_stat_show().
Link: https://lkml.kernel.org/r/20230428132406.2540811-1-yosryahmed@google.com
Link: https://lkml.kernel.org/r/20230428132406.2540811-3-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Acked-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Michal Hocko <mhocko@kernel.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Petr Mladek <pmladek@suse.com>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Steven Rostedt (Google) <rostedt@goodmis.org>
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-04-28 13:24:06 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-06-28 07:23:33 +00:00
|
|
|
|
|
|
|
|
|
memcg: dump memory.stat during cgroup OOM for v1
Patch series "memcg: OOM log improvements", v2.
This short patch series brings back some cgroup v1 stats in OOM logs
that were unnecessarily changed before. It also makes memcg OOM logs
less reliant on printk() internals.
This patch (of 2):
Commit c8713d0b2312 ("mm: memcontrol: dump memory.stat during cgroup OOM")
made sure we dump all the stats in memory.stat during a cgroup OOM, but it
also introduced a slight behavioral change. The code used to print the
non-hierarchical v1 cgroup stats for the entire cgroup subtree, now it
only prints the v2 cgroup stats for the cgroup under OOM.
For cgroup v1 users, this introduces a few problems:
(a) The non-hierarchical stats of the memcg under OOM are no longer
shown.
(b) A couple of v1-only stats (e.g. pgpgin, pgpgout) are no longer
shown.
(c) We show the list of cgroup v2 stats, even in cgroup v1. This list
of stats is not tracked with v1 in mind. While most of the stats seem
to be working on v1, there may be some stats that are not fully or
correctly tracked.
Although OOM log is not set in stone, we should not change it for no
reason. When upgrading the kernel version to a version including commit
c8713d0b2312 ("mm: memcontrol: dump memory.stat during cgroup OOM"), these
behavioral changes are noticed in cgroup v1.
The fix is simple. Commit c8713d0b2312 ("mm: memcontrol: dump memory.stat
during cgroup OOM") separated stats formatting from stats display for v2,
to reuse the stats formatting in the OOM logs. Do the same for v1.
Move the v2 specific formatting from memory_stat_format() to
memcg_stat_format(), add memcg1_stat_format() for v1, and make
memory_stat_format() select between them based on cgroup version. Since
memory_stat_show() now works for both v1 & v2, drop memcg_stat_show().
Link: https://lkml.kernel.org/r/20230428132406.2540811-1-yosryahmed@google.com
Link: https://lkml.kernel.org/r/20230428132406.2540811-3-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Acked-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Michal Hocko <mhocko@kernel.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Petr Mladek <pmladek@suse.com>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Steven Rostedt (Google) <rostedt@goodmis.org>
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-04-28 13:24:06 +00:00
|
|
|
|
|
|
|
|
|
2009-04-02 16:57:39 -07:00
|
|
|
|
mm, oom: add oom victim's memcg to the oom context information
The current oom report doesn't display victim's memcg context during the
global OOM situation. While this information is not strictly needed, it
can be really helpful for containerized environments to locate which
container has lost a process. Now that we have a single line for the oom
context, we can trivially add both the oom memcg (this can be either
global_oom or a specific memcg which hits its hard limits) and task_memcg
which is the victim's memcg.
Below is the single line output in the oom report after this patch.
- global oom context information:
oom-kill:constraint=<constraint>,nodemask=<nodemask>,cpuset=<cpuset>,mems_allowed=<mems_allowed>,global_oom,task_memcg=<memcg>,task=<comm>,pid=<pid>,uid=<uid>
- memcg oom context information:
oom-kill:constraint=<constraint>,nodemask=<nodemask>,cpuset=<cpuset>,mems_allowed=<mems_allowed>,oom_memcg=<memcg>,task_memcg=<memcg>,task=<comm>,pid=<pid>,uid=<uid>
[penguin-kernel@I-love.SAKURA.ne.jp: use pr_cont() in mem_cgroup_print_oom_context()]
Link: http://lkml.kernel.org/r/201812190723.wBJ7NdkN032628@www262.sakura.ne.jp
Link: http://lkml.kernel.org/r/1542799799-36184-2-git-send-email-ufo19890607@gmail.com
Signed-off-by: yuzhoujian <yuzhoujian@didichuxing.com>
Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: David Rientjes <rientjes@google.com>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Tetsuo Handa <penguin-kernel@i-love.sakura.ne.jp>
Cc: Roman Gushchin <guro@fb.com>
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 00:36:10 -08:00
|
|
|
|
|
|
|
|
|
2009-04-02 16:57:39 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, oom: add oom victim's memcg to the oom context information
The current oom report doesn't display victim's memcg context during the
global OOM situation. While this information is not strictly needed, it
can be really helpful for containerized environments to locate which
container has lost a process. Now that we have a single line for the oom
context, we can trivially add both the oom memcg (this can be either
global_oom or a specific memcg which hits its hard limits) and task_memcg
which is the victim's memcg.
Below is the single line output in the oom report after this patch.
- global oom context information:
oom-kill:constraint=<constraint>,nodemask=<nodemask>,cpuset=<cpuset>,mems_allowed=<mems_allowed>,global_oom,task_memcg=<memcg>,task=<comm>,pid=<pid>,uid=<uid>
- memcg oom context information:
oom-kill:constraint=<constraint>,nodemask=<nodemask>,cpuset=<cpuset>,mems_allowed=<mems_allowed>,oom_memcg=<memcg>,task_memcg=<memcg>,task=<comm>,pid=<pid>,uid=<uid>
[penguin-kernel@I-love.SAKURA.ne.jp: use pr_cont() in mem_cgroup_print_oom_context()]
Link: http://lkml.kernel.org/r/201812190723.wBJ7NdkN032628@www262.sakura.ne.jp
Link: http://lkml.kernel.org/r/1542799799-36184-2-git-send-email-ufo19890607@gmail.com
Signed-off-by: yuzhoujian <yuzhoujian@didichuxing.com>
Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: David Rientjes <rientjes@google.com>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Tetsuo Handa <penguin-kernel@i-love.sakura.ne.jp>
Cc: Roman Gushchin <guro@fb.com>
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 00:36:10 -08:00
|
|
|
|
2009-04-02 16:57:39 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, oom: add oom victim's memcg to the oom context information
The current oom report doesn't display victim's memcg context during the
global OOM situation. While this information is not strictly needed, it
can be really helpful for containerized environments to locate which
container has lost a process. Now that we have a single line for the oom
context, we can trivially add both the oom memcg (this can be either
global_oom or a specific memcg which hits its hard limits) and task_memcg
which is the victim's memcg.
Below is the single line output in the oom report after this patch.
- global oom context information:
oom-kill:constraint=<constraint>,nodemask=<nodemask>,cpuset=<cpuset>,mems_allowed=<mems_allowed>,global_oom,task_memcg=<memcg>,task=<comm>,pid=<pid>,uid=<uid>
- memcg oom context information:
oom-kill:constraint=<constraint>,nodemask=<nodemask>,cpuset=<cpuset>,mems_allowed=<mems_allowed>,oom_memcg=<memcg>,task_memcg=<memcg>,task=<comm>,pid=<pid>,uid=<uid>
[penguin-kernel@I-love.SAKURA.ne.jp: use pr_cont() in mem_cgroup_print_oom_context()]
Link: http://lkml.kernel.org/r/201812190723.wBJ7NdkN032628@www262.sakura.ne.jp
Link: http://lkml.kernel.org/r/1542799799-36184-2-git-send-email-ufo19890607@gmail.com
Signed-off-by: yuzhoujian <yuzhoujian@didichuxing.com>
Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: David Rientjes <rientjes@google.com>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Tetsuo Handa <penguin-kernel@i-love.sakura.ne.jp>
Cc: Roman Gushchin <guro@fb.com>
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 00:36:10 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-04-14 15:48:18 -07:00
|
|
|
|
mm, oom: add oom victim's memcg to the oom context information
The current oom report doesn't display victim's memcg context during the
global OOM situation. While this information is not strictly needed, it
can be really helpful for containerized environments to locate which
container has lost a process. Now that we have a single line for the oom
context, we can trivially add both the oom memcg (this can be either
global_oom or a specific memcg which hits its hard limits) and task_memcg
which is the victim's memcg.
Below is the single line output in the oom report after this patch.
- global oom context information:
oom-kill:constraint=<constraint>,nodemask=<nodemask>,cpuset=<cpuset>,mems_allowed=<mems_allowed>,global_oom,task_memcg=<memcg>,task=<comm>,pid=<pid>,uid=<uid>
- memcg oom context information:
oom-kill:constraint=<constraint>,nodemask=<nodemask>,cpuset=<cpuset>,mems_allowed=<mems_allowed>,oom_memcg=<memcg>,task_memcg=<memcg>,task=<comm>,pid=<pid>,uid=<uid>
[penguin-kernel@I-love.SAKURA.ne.jp: use pr_cont() in mem_cgroup_print_oom_context()]
Link: http://lkml.kernel.org/r/201812190723.wBJ7NdkN032628@www262.sakura.ne.jp
Link: http://lkml.kernel.org/r/1542799799-36184-2-git-send-email-ufo19890607@gmail.com
Signed-off-by: yuzhoujian <yuzhoujian@didichuxing.com>
Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: David Rientjes <rientjes@google.com>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Tetsuo Handa <penguin-kernel@i-love.sakura.ne.jp>
Cc: Roman Gushchin <guro@fb.com>
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 00:36:10 -08:00
|
|
|
|
2015-04-14 15:48:18 -07:00
|
|
|
|
|
|
|
|
|
2009-04-02 16:57:39 -07:00
|
|
|
|
mm, oom: add oom victim's memcg to the oom context information
The current oom report doesn't display victim's memcg context during the
global OOM situation. While this information is not strictly needed, it
can be really helpful for containerized environments to locate which
container has lost a process. Now that we have a single line for the oom
context, we can trivially add both the oom memcg (this can be either
global_oom or a specific memcg which hits its hard limits) and task_memcg
which is the victim's memcg.
Below is the single line output in the oom report after this patch.
- global oom context information:
oom-kill:constraint=<constraint>,nodemask=<nodemask>,cpuset=<cpuset>,mems_allowed=<mems_allowed>,global_oom,task_memcg=<memcg>,task=<comm>,pid=<pid>,uid=<uid>
- memcg oom context information:
oom-kill:constraint=<constraint>,nodemask=<nodemask>,cpuset=<cpuset>,mems_allowed=<mems_allowed>,oom_memcg=<memcg>,task_memcg=<memcg>,task=<comm>,pid=<pid>,uid=<uid>
[penguin-kernel@I-love.SAKURA.ne.jp: use pr_cont() in mem_cgroup_print_oom_context()]
Link: http://lkml.kernel.org/r/201812190723.wBJ7NdkN032628@www262.sakura.ne.jp
Link: http://lkml.kernel.org/r/1542799799-36184-2-git-send-email-ufo19890607@gmail.com
Signed-off-by: yuzhoujian <yuzhoujian@didichuxing.com>
Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: David Rientjes <rientjes@google.com>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Tetsuo Handa <penguin-kernel@i-love.sakura.ne.jp>
Cc: Roman Gushchin <guro@fb.com>
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 00:36:10 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-07-22 19:45:39 +09:00
|
|
|
|
2024-10-21 14:00:26 +01:00
|
|
|
|
2023-04-28 13:24:05 +00:00
|
|
|
|
2025-02-27 23:58:07 -08:00
|
|
|
|
2022-07-22 19:45:39 +09:00
|
|
|
|
|
|
|
|
|
2009-04-02 16:57:39 -07:00
|
|
|
|
2025-02-27 23:58:07 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcontrol: lockless page counters
Memory is internally accounted in bytes, using spinlock-protected 64-bit
counters, even though the smallest accounting delta is a page. The
counter interface is also convoluted and does too many things.
Introduce a new lockless word-sized page counter API, then change all
memory accounting over to it. The translation from and to bytes then only
happens when interfacing with userspace.
The removed locking overhead is noticable when scaling beyond the per-cpu
charge caches - on a 4-socket machine with 144-threads, the following test
shows the performance differences of 288 memcgs concurrently running a
page fault benchmark:
vanilla:
18631648.500498 task-clock (msec) # 140.643 CPUs utilized ( +- 0.33% )
1,380,638 context-switches # 0.074 K/sec ( +- 0.75% )
24,390 cpu-migrations # 0.001 K/sec ( +- 8.44% )
1,843,305,768 page-faults # 0.099 M/sec ( +- 0.00% )
50,134,994,088,218 cycles # 2.691 GHz ( +- 0.33% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
8,049,712,224,651 instructions # 0.16 insns per cycle ( +- 0.04% )
1,586,970,584,979 branches # 85.176 M/sec ( +- 0.05% )
1,724,989,949 branch-misses # 0.11% of all branches ( +- 0.48% )
132.474343877 seconds time elapsed ( +- 0.21% )
lockless:
12195979.037525 task-clock (msec) # 133.480 CPUs utilized ( +- 0.18% )
832,850 context-switches # 0.068 K/sec ( +- 0.54% )
15,624 cpu-migrations # 0.001 K/sec ( +- 10.17% )
1,843,304,774 page-faults # 0.151 M/sec ( +- 0.00% )
32,811,216,801,141 cycles # 2.690 GHz ( +- 0.18% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
9,999,265,091,727 instructions # 0.30 insns per cycle ( +- 0.10% )
2,076,759,325,203 branches # 170.282 M/sec ( +- 0.12% )
1,656,917,214 branch-misses # 0.08% of all branches ( +- 0.55% )
91.369330729 seconds time elapsed ( +- 0.45% )
On top of improved scalability, this also gets rid of the icky long long
types in the very heart of memcg, which is great for 32 bit and also makes
the code a lot more readable.
Notable differences between the old and new API:
- res_counter_charge() and res_counter_charge_nofail() become
page_counter_try_charge() and page_counter_charge() resp. to match
the more common kernel naming scheme of try_do()/do()
- res_counter_uncharge_until() is only ever used to cancel a local
counter and never to uncharge bigger segments of a hierarchy, so
it's replaced by the simpler page_counter_cancel()
- res_counter_set_limit() is replaced by page_counter_limit(), which
expects its callers to serialize against themselves
- res_counter_memparse_write_strategy() is replaced by
page_counter_limit(), which rounds down to the nearest page size -
rather than up. This is more reasonable for explicitely requested
hard upper limits.
- to keep charging light-weight, page_counter_try_charge() charges
speculatively, only to roll back if the result exceeds the limit.
Because of this, a failing bigger charge can temporarily lock out
smaller charges that would otherwise succeed. The error is bounded
to the difference between the smallest and the biggest possible
charge size, so for memcg, this means that a failing THP charge can
send base page charges into reclaim upto 2MB (4MB) before the limit
would have been reached. This should be acceptable.
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE and memparse]
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE, memparse, strncmp, and PAGE_SIZE]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Vladimir Davydov <vdavydov@parallels.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-12-10 15:42:31 -08:00
|
|
|
|
|
|
|
|
|
2025-02-27 23:58:07 -08:00
|
|
|
|
mm: memcontrol: dump memory.stat during cgroup OOM
The current cgroup OOM memory info dump doesn't include all the memory
we are tracking, nor does it give insight into what the VM tried to do
leading up to the OOM. All that useful info is in memory.stat.
Furthermore, the recursive printing for every child cgroup can
generate absurd amounts of data on the console for larger cgroup
trees, and it's not like we provide a per-cgroup breakdown during
global OOM kills.
When an OOM kill is triggered, print one set of recursive memory.stat
items at the level whose limit triggered the OOM condition.
Example output:
stress invoked oom-killer: gfp_mask=0x100cca(GFP_HIGHUSER_MOVABLE), order=0, oom_score_adj=0
CPU: 2 PID: 210 Comm: stress Not tainted 5.2.0-rc2-mm1-00247-g47d49835983c #135
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.12.0-20181126_142135-anatol 04/01/2014
Call Trace:
dump_stack+0x46/0x60
dump_header+0x4c/0x2d0
oom_kill_process.cold.10+0xb/0x10
out_of_memory+0x200/0x270
? try_to_free_mem_cgroup_pages+0xdf/0x130
mem_cgroup_out_of_memory+0xb7/0xc0
try_charge+0x680/0x6f0
mem_cgroup_try_charge+0xb5/0x160
__add_to_page_cache_locked+0xc6/0x300
? list_lru_destroy+0x80/0x80
add_to_page_cache_lru+0x45/0xc0
pagecache_get_page+0x11b/0x290
filemap_fault+0x458/0x6d0
ext4_filemap_fault+0x27/0x36
__do_fault+0x2f/0xb0
__handle_mm_fault+0x9c5/0x1140
? apic_timer_interrupt+0xa/0x20
handle_mm_fault+0xc5/0x180
__do_page_fault+0x1ab/0x440
? page_fault+0x8/0x30
page_fault+0x1e/0x30
RIP: 0033:0x55c32167fc10
Code: Bad RIP value.
RSP: 002b:00007fff1d031c50 EFLAGS: 00010206
RAX: 000000000dc00000 RBX: 00007fd2db000010 RCX: 00007fd2db000010
RDX: 0000000000000000 RSI: 0000000010001000 RDI: 0000000000000000
RBP: 000055c321680a54 R08: 00000000ffffffff R09: 0000000000000000
R10: 0000000000000022 R11: 0000000000000246 R12: ffffffffffffffff
R13: 0000000000000002 R14: 0000000000001000 R15: 0000000010000000
memory: usage 1024kB, limit 1024kB, failcnt 75131
swap: usage 0kB, limit 9007199254740988kB, failcnt 0
Memory cgroup stats for /foo:
anon 0
file 0
kernel_stack 36864
slab 274432
sock 0
shmem 0
file_mapped 0
file_dirty 0
file_writeback 0
anon_thp 0
inactive_anon 126976
active_anon 0
inactive_file 0
active_file 0
unevictable 0
slab_reclaimable 0
slab_unreclaimable 274432
pgfault 59466
pgmajfault 1617
workingset_refault 2145
workingset_activate 0
workingset_nodereclaim 0
pgrefill 98952
pgscan 200060
pgsteal 59340
pgactivate 40095
pgdeactivate 96787
pglazyfree 0
pglazyfreed 0
thp_fault_alloc 0
thp_collapse_alloc 0
Tasks state (memory values in pages):
[ pid ] uid tgid total_vm rss pgtables_bytes swapents oom_score_adj name
[ 200] 0 200 1121 884 53248 29 0 bash
[ 209] 0 209 905 246 45056 19 0 stress
[ 210] 0 210 66442 56 499712 56349 0 stress
oom-kill:constraint=CONSTRAINT_NONE,nodemask=(null),oom_memcg=/foo,task_memcg=/foo,task=stress,pid=210,uid=0
Memory cgroup out of memory: Killed process 210 (stress) total-vm:265768kB, anon-rss:0kB, file-rss:224kB, shmem-rss:0kB
oom_reaper: reaped process 210 (stress), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
[hannes@cmpxchg.org: s/kvmalloc/kmalloc/ per Michal]
Link: http://lkml.kernel.org/r/20190605161133.GA12453@cmpxchg.org
Link: http://lkml.kernel.org/r/20190604210509.9744-1-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-11 20:55:59 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-02-27 23:58:07 -08:00
|
|
|
|
|
|
|
|
|
2024-06-28 21:03:11 +00:00
|
|
|
|
mm: memcontrol: dump memory.stat during cgroup OOM
The current cgroup OOM memory info dump doesn't include all the memory
we are tracking, nor does it give insight into what the VM tried to do
leading up to the OOM. All that useful info is in memory.stat.
Furthermore, the recursive printing for every child cgroup can
generate absurd amounts of data on the console for larger cgroup
trees, and it's not like we provide a per-cgroup breakdown during
global OOM kills.
When an OOM kill is triggered, print one set of recursive memory.stat
items at the level whose limit triggered the OOM condition.
Example output:
stress invoked oom-killer: gfp_mask=0x100cca(GFP_HIGHUSER_MOVABLE), order=0, oom_score_adj=0
CPU: 2 PID: 210 Comm: stress Not tainted 5.2.0-rc2-mm1-00247-g47d49835983c #135
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.12.0-20181126_142135-anatol 04/01/2014
Call Trace:
dump_stack+0x46/0x60
dump_header+0x4c/0x2d0
oom_kill_process.cold.10+0xb/0x10
out_of_memory+0x200/0x270
? try_to_free_mem_cgroup_pages+0xdf/0x130
mem_cgroup_out_of_memory+0xb7/0xc0
try_charge+0x680/0x6f0
mem_cgroup_try_charge+0xb5/0x160
__add_to_page_cache_locked+0xc6/0x300
? list_lru_destroy+0x80/0x80
add_to_page_cache_lru+0x45/0xc0
pagecache_get_page+0x11b/0x290
filemap_fault+0x458/0x6d0
ext4_filemap_fault+0x27/0x36
__do_fault+0x2f/0xb0
__handle_mm_fault+0x9c5/0x1140
? apic_timer_interrupt+0xa/0x20
handle_mm_fault+0xc5/0x180
__do_page_fault+0x1ab/0x440
? page_fault+0x8/0x30
page_fault+0x1e/0x30
RIP: 0033:0x55c32167fc10
Code: Bad RIP value.
RSP: 002b:00007fff1d031c50 EFLAGS: 00010206
RAX: 000000000dc00000 RBX: 00007fd2db000010 RCX: 00007fd2db000010
RDX: 0000000000000000 RSI: 0000000010001000 RDI: 0000000000000000
RBP: 000055c321680a54 R08: 00000000ffffffff R09: 0000000000000000
R10: 0000000000000022 R11: 0000000000000246 R12: ffffffffffffffff
R13: 0000000000000002 R14: 0000000000001000 R15: 0000000010000000
memory: usage 1024kB, limit 1024kB, failcnt 75131
swap: usage 0kB, limit 9007199254740988kB, failcnt 0
Memory cgroup stats for /foo:
anon 0
file 0
kernel_stack 36864
slab 274432
sock 0
shmem 0
file_mapped 0
file_dirty 0
file_writeback 0
anon_thp 0
inactive_anon 126976
active_anon 0
inactive_file 0
active_file 0
unevictable 0
slab_reclaimable 0
slab_unreclaimable 274432
pgfault 59466
pgmajfault 1617
workingset_refault 2145
workingset_activate 0
workingset_nodereclaim 0
pgrefill 98952
pgscan 200060
pgsteal 59340
pgactivate 40095
pgdeactivate 96787
pglazyfree 0
pglazyfreed 0
thp_fault_alloc 0
thp_collapse_alloc 0
Tasks state (memory values in pages):
[ pid ] uid tgid total_vm rss pgtables_bytes swapents oom_score_adj name
[ 200] 0 200 1121 884 53248 29 0 bash
[ 209] 0 209 905 246 45056 19 0 stress
[ 210] 0 210 66442 56 499712 56349 0 stress
oom-kill:constraint=CONSTRAINT_NONE,nodemask=(null),oom_memcg=/foo,task_memcg=/foo,task=stress,pid=210,uid=0
Memory cgroup out of memory: Killed process 210 (stress) total-vm:265768kB, anon-rss:0kB, file-rss:224kB, shmem-rss:0kB
oom_reaper: reaped process 210 (stress), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
[hannes@cmpxchg.org: s/kvmalloc/kmalloc/ per Michal]
Link: http://lkml.kernel.org/r/20190605161133.GA12453@cmpxchg.org
Link: http://lkml.kernel.org/r/20190604210509.9744-1-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-11 20:55:59 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
memcg, oom: provide more precise dump info while memcg oom happening
Currently when a memcg oom is happening the oom dump messages is still
global state and provides few useful info for users. This patch prints
more pointed memcg page statistics for memcg-oom and take hierarchy into
consideration:
Based on Michal's advice, we take hierarchy into consideration: supppose
we trigger an OOM on A's limit
root_memcg
|
A (use_hierachy=1)
/ \
B C
|
D
then the printed info will be:
Memory cgroup stats for /A:...
Memory cgroup stats for /A/B:...
Memory cgroup stats for /A/C:...
Memory cgroup stats for /A/B/D:...
Following are samples of oom output:
(1) Before change:
mal-80 invoked oom-killer:gfp_mask=0xd0, order=0, oom_score_adj=0
mal-80 cpuset=/ mems_allowed=0
Pid: 2976, comm: mal-80 Not tainted 3.7.0+ #10
Call Trace:
[<ffffffff8167fbfb>] dump_header+0x83/0x1ca
..... (call trace)
[<ffffffff8168a818>] page_fault+0x28/0x30
<<<<<<<<<<<<<<<<<<<<< memcg specific information
Task in /A/B/D killed as a result of limit of /A
memory: usage 101376kB, limit 101376kB, failcnt 57
memory+swap: usage 101376kB, limit 101376kB, failcnt 0
kmem: usage 0kB, limit 9007199254740991kB, failcnt 0
<<<<<<<<<<<<<<<<<<<<< print per cpu pageset stat
Mem-Info:
Node 0 DMA per-cpu:
CPU 0: hi: 0, btch: 1 usd: 0
......
CPU 3: hi: 0, btch: 1 usd: 0
Node 0 DMA32 per-cpu:
CPU 0: hi: 186, btch: 31 usd: 173
......
CPU 3: hi: 186, btch: 31 usd: 130
<<<<<<<<<<<<<<<<<<<<< print global page state
active_anon:92963 inactive_anon:40777 isolated_anon:0
active_file:33027 inactive_file:51718 isolated_file:0
unevictable:0 dirty:3 writeback:0 unstable:0
free:729995 slab_reclaimable:6897 slab_unreclaimable:6263
mapped:20278 shmem:35971 pagetables:5885 bounce:0
free_cma:0
<<<<<<<<<<<<<<<<<<<<< print per zone page state
Node 0 DMA free:15836kB ... all_unreclaimable? no
lowmem_reserve[]: 0 3175 3899 3899
Node 0 DMA32 free:2888564kB ... all_unrelaimable? no
lowmem_reserve[]: 0 0 724 724
lowmem_reserve[]: 0 0 0 0
Node 0 DMA: 1*4kB (U) ... 3*4096kB (M) = 15836kB
Node 0 DMA32: 41*4kB (UM) ... 702*4096kB (MR) = 2888316kB
120710 total pagecache pages
0 pages in swap cache
<<<<<<<<<<<<<<<<<<<<< print global swap cache stat
Swap cache stats: add 0, delete 0, find 0/0
Free swap = 499708kB
Total swap = 499708kB
1040368 pages RAM
58678 pages reserved
169065 pages shared
173632 pages non-shared
[ pid ] uid tgid total_vm rss nr_ptes swapents oom_score_adj name
[ 2693] 0 2693 6005 1324 17 0 0 god
[ 2754] 0 2754 6003 1320 16 0 0 god
[ 2811] 0 2811 5992 1304 18 0 0 god
[ 2874] 0 2874 6005 1323 18 0 0 god
[ 2935] 0 2935 8720 7742 21 0 0 mal-30
[ 2976] 0 2976 21520 17577 42 0 0 mal-80
Memory cgroup out of memory: Kill process 2976 (mal-80) score 665 or sacrifice child
Killed process 2976 (mal-80) total-vm:86080kB, anon-rss:69964kB, file-rss:344kB
We can see that messages dumped by show_free_areas() are longsome and can
provide so limited info for memcg that just happen oom.
(2) After change
mal-80 invoked oom-killer: gfp_mask=0xd0, order=0, oom_score_adj=0
mal-80 cpuset=/ mems_allowed=0
Pid: 2704, comm: mal-80 Not tainted 3.7.0+ #10
Call Trace:
[<ffffffff8167fd0b>] dump_header+0x83/0x1d1
.......(call trace)
[<ffffffff8168a918>] page_fault+0x28/0x30
Task in /A/B/D killed as a result of limit of /A
<<<<<<<<<<<<<<<<<<<<< memcg specific information
memory: usage 102400kB, limit 102400kB, failcnt 140
memory+swap: usage 102400kB, limit 102400kB, failcnt 0
kmem: usage 0kB, limit 9007199254740991kB, failcnt 0
Memory cgroup stats for /A: cache:32KB rss:30984KB mapped_file:0KB swap:0KB inactive_anon:6912KB active_anon:24072KB inactive_file:32KB active_file:0KB unevictable:0KB
Memory cgroup stats for /A/B: cache:0KB rss:0KB mapped_file:0KB swap:0KB inactive_anon:0KB active_anon:0KB inactive_file:0KB active_file:0KB unevictable:0KB
Memory cgroup stats for /A/C: cache:0KB rss:0KB mapped_file:0KB swap:0KB inactive_anon:0KB active_anon:0KB inactive_file:0KB active_file:0KB unevictable:0KB
Memory cgroup stats for /A/B/D: cache:32KB rss:71352KB mapped_file:0KB swap:0KB inactive_anon:6656KB active_anon:64696KB inactive_file:16KB active_file:16KB unevictable:0KB
[ pid ] uid tgid total_vm rss nr_ptes swapents oom_score_adj name
[ 2260] 0 2260 6006 1325 18 0 0 god
[ 2383] 0 2383 6003 1319 17 0 0 god
[ 2503] 0 2503 6004 1321 18 0 0 god
[ 2622] 0 2622 6004 1321 16 0 0 god
[ 2695] 0 2695 8720 7741 22 0 0 mal-30
[ 2704] 0 2704 21520 17839 43 0 0 mal-80
Memory cgroup out of memory: Kill process 2704 (mal-80) score 669 or sacrifice child
Killed process 2704 (mal-80) total-vm:86080kB, anon-rss:71016kB, file-rss:340kB
This version provides more pointed info for memcg in "Memory cgroup stats
for XXX" section.
Signed-off-by: Sha Zhengju <handai.szj@taobao.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Acked-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-22 16:32:05 -08:00
|
|
|
|
2024-06-28 21:03:11 +00:00
|
|
|
|
mm: memcontrol: dump memory.stat during cgroup OOM
The current cgroup OOM memory info dump doesn't include all the memory
we are tracking, nor does it give insight into what the VM tried to do
leading up to the OOM. All that useful info is in memory.stat.
Furthermore, the recursive printing for every child cgroup can
generate absurd amounts of data on the console for larger cgroup
trees, and it's not like we provide a per-cgroup breakdown during
global OOM kills.
When an OOM kill is triggered, print one set of recursive memory.stat
items at the level whose limit triggered the OOM condition.
Example output:
stress invoked oom-killer: gfp_mask=0x100cca(GFP_HIGHUSER_MOVABLE), order=0, oom_score_adj=0
CPU: 2 PID: 210 Comm: stress Not tainted 5.2.0-rc2-mm1-00247-g47d49835983c #135
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.12.0-20181126_142135-anatol 04/01/2014
Call Trace:
dump_stack+0x46/0x60
dump_header+0x4c/0x2d0
oom_kill_process.cold.10+0xb/0x10
out_of_memory+0x200/0x270
? try_to_free_mem_cgroup_pages+0xdf/0x130
mem_cgroup_out_of_memory+0xb7/0xc0
try_charge+0x680/0x6f0
mem_cgroup_try_charge+0xb5/0x160
__add_to_page_cache_locked+0xc6/0x300
? list_lru_destroy+0x80/0x80
add_to_page_cache_lru+0x45/0xc0
pagecache_get_page+0x11b/0x290
filemap_fault+0x458/0x6d0
ext4_filemap_fault+0x27/0x36
__do_fault+0x2f/0xb0
__handle_mm_fault+0x9c5/0x1140
? apic_timer_interrupt+0xa/0x20
handle_mm_fault+0xc5/0x180
__do_page_fault+0x1ab/0x440
? page_fault+0x8/0x30
page_fault+0x1e/0x30
RIP: 0033:0x55c32167fc10
Code: Bad RIP value.
RSP: 002b:00007fff1d031c50 EFLAGS: 00010206
RAX: 000000000dc00000 RBX: 00007fd2db000010 RCX: 00007fd2db000010
RDX: 0000000000000000 RSI: 0000000010001000 RDI: 0000000000000000
RBP: 000055c321680a54 R08: 00000000ffffffff R09: 0000000000000000
R10: 0000000000000022 R11: 0000000000000246 R12: ffffffffffffffff
R13: 0000000000000002 R14: 0000000000001000 R15: 0000000010000000
memory: usage 1024kB, limit 1024kB, failcnt 75131
swap: usage 0kB, limit 9007199254740988kB, failcnt 0
Memory cgroup stats for /foo:
anon 0
file 0
kernel_stack 36864
slab 274432
sock 0
shmem 0
file_mapped 0
file_dirty 0
file_writeback 0
anon_thp 0
inactive_anon 126976
active_anon 0
inactive_file 0
active_file 0
unevictable 0
slab_reclaimable 0
slab_unreclaimable 274432
pgfault 59466
pgmajfault 1617
workingset_refault 2145
workingset_activate 0
workingset_nodereclaim 0
pgrefill 98952
pgscan 200060
pgsteal 59340
pgactivate 40095
pgdeactivate 96787
pglazyfree 0
pglazyfreed 0
thp_fault_alloc 0
thp_collapse_alloc 0
Tasks state (memory values in pages):
[ pid ] uid tgid total_vm rss pgtables_bytes swapents oom_score_adj name
[ 200] 0 200 1121 884 53248 29 0 bash
[ 209] 0 209 905 246 45056 19 0 stress
[ 210] 0 210 66442 56 499712 56349 0 stress
oom-kill:constraint=CONSTRAINT_NONE,nodemask=(null),oom_memcg=/foo,task_memcg=/foo,task=stress,pid=210,uid=0
Memory cgroup out of memory: Killed process 210 (stress) total-vm:265768kB, anon-rss:0kB, file-rss:224kB, shmem-rss:0kB
oom_reaper: reaped process 210 (stress), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
[hannes@cmpxchg.org: s/kvmalloc/kmalloc/ per Michal]
Link: http://lkml.kernel.org/r/20190605161133.GA12453@cmpxchg.org
Link: http://lkml.kernel.org/r/20190604210509.9744-1-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-11 20:55:59 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-10-21 14:00:26 +01:00
|
|
|
|
2023-04-28 13:24:05 +00:00
|
|
|
|
|
|
|
|
|
2009-04-02 16:57:39 -07:00
|
|
|
|
|
|
|
|
|
oom: badness heuristic rewrite
This a complete rewrite of the oom killer's badness() heuristic which is
used to determine which task to kill in oom conditions. The goal is to
make it as simple and predictable as possible so the results are better
understood and we end up killing the task which will lead to the most
memory freeing while still respecting the fine-tuning from userspace.
Instead of basing the heuristic on mm->total_vm for each task, the task's
rss and swap space is used instead. This is a better indication of the
amount of memory that will be freeable if the oom killed task is chosen
and subsequently exits. This helps specifically in cases where KDE or
GNOME is chosen for oom kill on desktop systems instead of a memory
hogging task.
The baseline for the heuristic is a proportion of memory that each task is
currently using in memory plus swap compared to the amount of "allowable"
memory. "Allowable," in this sense, means the system-wide resources for
unconstrained oom conditions, the set of mempolicy nodes, the mems
attached to current's cpuset, or a memory controller's limit. The
proportion is given on a scale of 0 (never kill) to 1000 (always kill),
roughly meaning that if a task has a badness() score of 500 that the task
consumes approximately 50% of allowable memory resident in RAM or in swap
space.
The proportion is always relative to the amount of "allowable" memory and
not the total amount of RAM systemwide so that mempolicies and cpusets may
operate in isolation; they shall not need to know the true size of the
machine on which they are running if they are bound to a specific set of
nodes or mems, respectively.
Root tasks are given 3% extra memory just like __vm_enough_memory()
provides in LSMs. In the event of two tasks consuming similar amounts of
memory, it is generally better to save root's task.
Because of the change in the badness() heuristic's baseline, it is also
necessary to introduce a new user interface to tune it. It's not possible
to redefine the meaning of /proc/pid/oom_adj with a new scale since the
ABI cannot be changed for backward compatability. Instead, a new tunable,
/proc/pid/oom_score_adj, is added that ranges from -1000 to +1000. It may
be used to polarize the heuristic such that certain tasks are never
considered for oom kill while others may always be considered. The value
is added directly into the badness() score so a value of -500, for
example, means to discount 50% of its memory consumption in comparison to
other tasks either on the system, bound to the mempolicy, in the cpuset,
or sharing the same memory controller.
/proc/pid/oom_adj is changed so that its meaning is rescaled into the
units used by /proc/pid/oom_score_adj, and vice versa. Changing one of
these per-task tunables will rescale the value of the other to an
equivalent meaning. Although /proc/pid/oom_adj was originally defined as
a bitshift on the badness score, it now shares the same linear growth as
/proc/pid/oom_score_adj but with different granularity. This is required
so the ABI is not broken with userspace applications and allows oom_adj to
be deprecated for future removal.
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: Nick Piggin <npiggin@suse.de>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Balbir Singh <balbir@in.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-08-09 17:19:46 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-06-07 17:06:18 -07:00
|
|
|
|
oom: badness heuristic rewrite
This a complete rewrite of the oom killer's badness() heuristic which is
used to determine which task to kill in oom conditions. The goal is to
make it as simple and predictable as possible so the results are better
understood and we end up killing the task which will lead to the most
memory freeing while still respecting the fine-tuning from userspace.
Instead of basing the heuristic on mm->total_vm for each task, the task's
rss and swap space is used instead. This is a better indication of the
amount of memory that will be freeable if the oom killed task is chosen
and subsequently exits. This helps specifically in cases where KDE or
GNOME is chosen for oom kill on desktop systems instead of a memory
hogging task.
The baseline for the heuristic is a proportion of memory that each task is
currently using in memory plus swap compared to the amount of "allowable"
memory. "Allowable," in this sense, means the system-wide resources for
unconstrained oom conditions, the set of mempolicy nodes, the mems
attached to current's cpuset, or a memory controller's limit. The
proportion is given on a scale of 0 (never kill) to 1000 (always kill),
roughly meaning that if a task has a badness() score of 500 that the task
consumes approximately 50% of allowable memory resident in RAM or in swap
space.
The proportion is always relative to the amount of "allowable" memory and
not the total amount of RAM systemwide so that mempolicies and cpusets may
operate in isolation; they shall not need to know the true size of the
machine on which they are running if they are bound to a specific set of
nodes or mems, respectively.
Root tasks are given 3% extra memory just like __vm_enough_memory()
provides in LSMs. In the event of two tasks consuming similar amounts of
memory, it is generally better to save root's task.
Because of the change in the badness() heuristic's baseline, it is also
necessary to introduce a new user interface to tune it. It's not possible
to redefine the meaning of /proc/pid/oom_adj with a new scale since the
ABI cannot be changed for backward compatability. Instead, a new tunable,
/proc/pid/oom_score_adj, is added that ranges from -1000 to +1000. It may
be used to polarize the heuristic such that certain tasks are never
considered for oom kill while others may always be considered. The value
is added directly into the badness() score so a value of -500, for
example, means to discount 50% of its memory consumption in comparison to
other tasks either on the system, bound to the mempolicy, in the cpuset,
or sharing the same memory controller.
/proc/pid/oom_adj is changed so that its meaning is rescaled into the
units used by /proc/pid/oom_score_adj, and vice versa. Changing one of
these per-task tunables will rescale the value of the other to an
equivalent meaning. Although /proc/pid/oom_adj was originally defined as
a bitshift on the badness score, it now shares the same linear growth as
/proc/pid/oom_score_adj but with different granularity. This is required
so the ABI is not broken with userspace applications and allows oom_adj to
be deprecated for future removal.
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: Nick Piggin <npiggin@suse.de>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Balbir Singh <balbir@in.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-08-09 17:19:46 -07:00
|
|
|
|
2020-10-13 16:52:52 -07:00
|
|
|
|
|
|
|
|
|
2022-09-26 09:57:03 -04:00
|
|
|
|
2020-10-13 16:52:52 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-09-26 09:57:03 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-11-16 14:14:49 -08:00
|
|
|
|
2018-06-07 17:06:18 -07:00
|
|
|
|
oom: badness heuristic rewrite
This a complete rewrite of the oom killer's badness() heuristic which is
used to determine which task to kill in oom conditions. The goal is to
make it as simple and predictable as possible so the results are better
understood and we end up killing the task which will lead to the most
memory freeing while still respecting the fine-tuning from userspace.
Instead of basing the heuristic on mm->total_vm for each task, the task's
rss and swap space is used instead. This is a better indication of the
amount of memory that will be freeable if the oom killed task is chosen
and subsequently exits. This helps specifically in cases where KDE or
GNOME is chosen for oom kill on desktop systems instead of a memory
hogging task.
The baseline for the heuristic is a proportion of memory that each task is
currently using in memory plus swap compared to the amount of "allowable"
memory. "Allowable," in this sense, means the system-wide resources for
unconstrained oom conditions, the set of mempolicy nodes, the mems
attached to current's cpuset, or a memory controller's limit. The
proportion is given on a scale of 0 (never kill) to 1000 (always kill),
roughly meaning that if a task has a badness() score of 500 that the task
consumes approximately 50% of allowable memory resident in RAM or in swap
space.
The proportion is always relative to the amount of "allowable" memory and
not the total amount of RAM systemwide so that mempolicies and cpusets may
operate in isolation; they shall not need to know the true size of the
machine on which they are running if they are bound to a specific set of
nodes or mems, respectively.
Root tasks are given 3% extra memory just like __vm_enough_memory()
provides in LSMs. In the event of two tasks consuming similar amounts of
memory, it is generally better to save root's task.
Because of the change in the badness() heuristic's baseline, it is also
necessary to introduce a new user interface to tune it. It's not possible
to redefine the meaning of /proc/pid/oom_adj with a new scale since the
ABI cannot be changed for backward compatability. Instead, a new tunable,
/proc/pid/oom_score_adj, is added that ranges from -1000 to +1000. It may
be used to polarize the heuristic such that certain tasks are never
considered for oom kill while others may always be considered. The value
is added directly into the badness() score so a value of -500, for
example, means to discount 50% of its memory consumption in comparison to
other tasks either on the system, bound to the mempolicy, in the cpuset,
or sharing the same memory controller.
/proc/pid/oom_adj is changed so that its meaning is rescaled into the
units used by /proc/pid/oom_score_adj, and vice versa. Changing one of
these per-task tunables will rescale the value of the other to an
equivalent meaning. Although /proc/pid/oom_adj was originally defined as
a bitshift on the badness score, it now shares the same linear growth as
/proc/pid/oom_score_adj but with different granularity. This is required
so the ABI is not broken with userspace applications and allows oom_adj to
be deprecated for future removal.
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: Nick Piggin <npiggin@suse.de>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Balbir Singh <balbir@in.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-08-09 17:19:46 -07:00
|
|
|
|
|
|
|
|
|
mm, memcg: proportional memory.{low,min} reclaim
cgroup v2 introduces two memory protection thresholds: memory.low
(best-effort) and memory.min (hard protection). While they generally do
what they say on the tin, there is a limitation in their implementation
that makes them difficult to use effectively: that cliff behaviour often
manifests when they become eligible for reclaim. This patch implements
more intuitive and usable behaviour, where we gradually mount more
reclaim pressure as cgroups further and further exceed their protection
thresholds.
This cliff edge behaviour happens because we only choose whether or not
to reclaim based on whether the memcg is within its protection limits
(see the use of mem_cgroup_protected in shrink_node), but we don't vary
our reclaim behaviour based on this information. Imagine the following
timeline, with the numbers the lruvec size in this zone:
1. memory.low=1000000, memory.current=999999. 0 pages may be scanned.
2. memory.low=1000000, memory.current=1000000. 0 pages may be scanned.
3. memory.low=1000000, memory.current=1000001. 1000001* pages may be
scanned. (?!)
* Of course, we won't usually scan all available pages in the zone even
without this patch because of scan control priority, over-reclaim
protection, etc. However, as shown by the tests at the end, these
techniques don't sufficiently throttle such an extreme change in input,
so cliff-like behaviour isn't really averted by their existence alone.
Here's an example of how this plays out in practice. At Facebook, we are
trying to protect various workloads from "system" software, like
configuration management tools, metric collectors, etc (see this[0] case
study). In order to find a suitable memory.low value, we start by
determining the expected memory range within which the workload will be
comfortable operating. This isn't an exact science -- memory usage deemed
"comfortable" will vary over time due to user behaviour, differences in
composition of work, etc, etc. As such we need to ballpark memory.low,
but doing this is currently problematic:
1. If we end up setting it too low for the workload, it won't have
*any* effect (see discussion above). The group will receive the full
weight of reclaim and won't have any priority while competing with the
less important system software, as if we had no memory.low configured
at all.
2. Because of this behaviour, we end up erring on the side of setting
it too high, such that the comfort range is reliably covered. However,
protected memory is completely unavailable to the rest of the system,
so we might cause undue memory and IO pressure there when we *know* we
have some elasticity in the workload.
3. Even if we get the value totally right, smack in the middle of the
comfort zone, we get extreme jumps between no pressure and full
pressure that cause unpredictable pressure spikes in the workload due
to the current binary reclaim behaviour.
With this patch, we can set it to our ballpark estimation without too much
worry. Any undesirable behaviour, such as too much or too little reclaim
pressure on the workload or system will be proportional to how far our
estimation is off. This means we can set memory.low much more
conservatively and thus waste less resources *without* the risk of the
workload falling off a cliff if we overshoot.
As a more abstract technical description, this unintuitive behaviour
results in having to give high-priority workloads a large protection
buffer on top of their expected usage to function reliably, as otherwise
we have abrupt periods of dramatically increased memory pressure which
hamper performance. Having to set these thresholds so high wastes
resources and generally works against the principle of work conservation.
In addition, having proportional memory reclaim behaviour has other
benefits. Most notably, before this patch it's basically mandatory to set
memory.low to a higher than desirable value because otherwise as soon as
you exceed memory.low, all protection is lost, and all pages are eligible
to scan again. By contrast, having a gradual ramp in reclaim pressure
means that you now still get some protection when thresholds are exceeded,
which means that one can now be more comfortable setting memory.low to
lower values without worrying that all protection will be lost. This is
important because workingset size is really hard to know exactly,
especially with variable workloads, so at least getting *some* protection
if your workingset size grows larger than you expect increases user
confidence in setting memory.low without a huge buffer on top being
needed.
Thanks a lot to Johannes Weiner and Tejun Heo for their advice and
assistance in thinking about how to make this work better.
In testing these changes, I intended to verify that:
1. Changes in page scanning become gradual and proportional instead of
binary.
To test this, I experimented stepping further and further down
memory.low protection on a workload that floats around 19G workingset
when under memory.low protection, watching page scan rates for the
workload cgroup:
+------------+-----------------+--------------------+--------------+
| memory.low | test (pgscan/s) | control (pgscan/s) | % of control |
+------------+-----------------+--------------------+--------------+
| 21G | 0 | 0 | N/A |
| 17G | 867 | 3799 | 23% |
| 12G | 1203 | 3543 | 34% |
| 8G | 2534 | 3979 | 64% |
| 4G | 3980 | 4147 | 96% |
| 0 | 3799 | 3980 | 95% |
+------------+-----------------+--------------------+--------------+
As you can see, the test kernel (with a kernel containing this
patch) ramps up page scanning significantly more gradually than the
control kernel (without this patch).
2. More gradual ramp up in reclaim aggression doesn't result in
premature OOMs.
To test this, I wrote a script that slowly increments the number of
pages held by stress(1)'s --vm-keep mode until a production system
entered severe overall memory contention. This script runs in a highly
protected slice taking up the majority of available system memory.
Watching vmstat revealed that page scanning continued essentially
nominally between test and control, without causing forward reclaim
progress to become arrested.
[0]: https://facebookmicrosites.github.io/cgroup2/docs/overview.html#case-study-the-fbtax2-project
[akpm@linux-foundation.org: reflow block comments to fit in 80 cols]
[chris@chrisdown.name: handle cgroup_disable=memory when getting memcg protection]
Link: http://lkml.kernel.org/r/20190201045711.GA18302@chrisdown.name
Link: http://lkml.kernel.org/r/20190124014455.GA6396@chrisdown.name
Signed-off-by: Chris Down <chris@chrisdown.name>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Dennis Zhou <dennis@kernel.org>
Cc: Tetsuo Handa <penguin-kernel@i-love.sakura.ne.jp>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-10-06 17:58:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-03-17 14:20:28 -07:00
|
|
|
|
2012-12-11 16:00:26 -08:00
|
|
|
|
mm, memcg: introduce own oom handler to iterate only over its own threads
The global oom killer is serialized by the per-zonelist
try_set_zonelist_oom() which is used in the page allocator. Concurrent
oom kills are thus a rare event and only occur in systems using
mempolicies and with a large number of nodes.
Memory controller oom kills, however, can frequently be concurrent since
there is no serialization once the oom killer is called for oom conditions
in several different memcgs in parallel.
This creates a massive contention on tasklist_lock since the oom killer
requires the readside for the tasklist iteration. If several memcgs are
calling the oom killer, this lock can be held for a substantial amount of
time, especially if threads continue to enter it as other threads are
exiting.
Since the exit path grabs the writeside of the lock with irqs disabled in
a few different places, this can cause a soft lockup on cpus as a result
of tasklist_lock starvation.
The kernel lacks unfair writelocks, and successful calls to the oom killer
usually result in at least one thread entering the exit path, so an
alternative solution is needed.
This patch introduces a seperate oom handler for memcgs so that they do
not require tasklist_lock for as much time. Instead, it iterates only
over the threads attached to the oom memcg and grabs a reference to the
selected thread before calling oom_kill_process() to ensure it doesn't
prematurely exit.
This still requires tasklist_lock for the tasklist dump, iterating
children of the selected process, and killing all other threads on the
system sharing the same memory as the selected victim. So while this
isn't a complete solution to tasklist_lock starvation, it significantly
reduces the amount of time that it is held.
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: Sha Zhengju <handai.szj@taobao.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-07-31 16:43:44 -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
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-08-06 23:22:08 -07:00
|
|
|
|
mm, memcg: introduce own oom handler to iterate only over its own threads
The global oom killer is serialized by the per-zonelist
try_set_zonelist_oom() which is used in the page allocator. Concurrent
oom kills are thus a rare event and only occur in systems using
mempolicies and with a large number of nodes.
Memory controller oom kills, however, can frequently be concurrent since
there is no serialization once the oom killer is called for oom conditions
in several different memcgs in parallel.
This creates a massive contention on tasklist_lock since the oom killer
requires the readside for the tasklist iteration. If several memcgs are
calling the oom killer, this lock can be held for a substantial amount of
time, especially if threads continue to enter it as other threads are
exiting.
Since the exit path grabs the writeside of the lock with irqs disabled in
a few different places, this can cause a soft lockup on cpus as a result
of tasklist_lock starvation.
The kernel lacks unfair writelocks, and successful calls to the oom killer
usually result in at least one thread entering the exit path, so an
alternative solution is needed.
This patch introduces a seperate oom handler for memcgs so that they do
not require tasklist_lock for as much time. Instead, it iterates only
over the threads attached to the oom memcg and grabs a reference to the
selected thread before calling oom_kill_process() to ensure it doesn't
prematurely exit.
This still requires tasklist_lock for the tasklist dump, iterating
children of the selected process, and killing all other threads on the
system sharing the same memory as the selected victim. So while this
isn't a complete solution to tasklist_lock starvation, it significantly
reduces the amount of time that it is held.
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: Sha Zhengju <handai.szj@taobao.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-07-31 16:43:44 -07:00
|
|
|
|
2019-03-05 15:46:47 -08:00
|
|
|
|
|
|
|
|
|
2020-08-06 23:22:08 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-03-05 15:46:47 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-04-02 11:01:17 +02:00
|
|
|
|
2020-08-06 23:22:08 -07:00
|
|
|
|
|
|
|
|
|
2015-06-24 16:57:19 -07:00
|
|
|
|
2016-10-07 16:57:23 -07:00
|
|
|
|
mm, memcg: introduce own oom handler to iterate only over its own threads
The global oom killer is serialized by the per-zonelist
try_set_zonelist_oom() which is used in the page allocator. Concurrent
oom kills are thus a rare event and only occur in systems using
mempolicies and with a large number of nodes.
Memory controller oom kills, however, can frequently be concurrent since
there is no serialization once the oom killer is called for oom conditions
in several different memcgs in parallel.
This creates a massive contention on tasklist_lock since the oom killer
requires the readside for the tasklist iteration. If several memcgs are
calling the oom killer, this lock can be held for a substantial amount of
time, especially if threads continue to enter it as other threads are
exiting.
Since the exit path grabs the writeside of the lock with irqs disabled in
a few different places, this can cause a soft lockup on cpus as a result
of tasklist_lock starvation.
The kernel lacks unfair writelocks, and successful calls to the oom killer
usually result in at least one thread entering the exit path, so an
alternative solution is needed.
This patch introduces a seperate oom handler for memcgs so that they do
not require tasklist_lock for as much time. Instead, it iterates only
over the threads attached to the oom memcg and grabs a reference to the
selected thread before calling oom_kill_process() to ensure it doesn't
prematurely exit.
This still requires tasklist_lock for the tasklist dump, iterating
children of the selected process, and killing all other threads on the
system sharing the same memory as the selected victim. So while this
isn't a complete solution to tasklist_lock starvation, it significantly
reduces the amount of time that it is held.
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: Sha Zhengju <handai.szj@taobao.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-07-31 16:43:44 -07:00
|
|
|
|
|
|
|
|
|
2022-03-22 14:40:19 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
memcg: fix OOM killer under memcg
This patch tries to fix OOM Killer problems caused by hierarchy.
Now, memcg itself has OOM KILL function (in oom_kill.c) and tries to
kill a task in memcg.
But, when hierarchy is used, it's broken and correct task cannot
be killed. For example, in following cgroup
/groupA/ hierarchy=1, limit=1G,
01 nolimit
02 nolimit
All tasks' memory usage under /groupA, /groupA/01, groupA/02 is limited to
groupA's 1Gbytes but OOM Killer just kills tasks in groupA.
This patch provides makes the bad process be selected from all tasks
under hierarchy. BTW, currently, oom_jiffies is updated against groupA
in above case. oom_jiffies of tree should be updated.
To see how oom_jiffies is used, please check mem_cgroup_oom_called()
callers.
[akpm@linux-foundation.org: build fix]
[akpm@linux-foundation.org: const fix]
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Paul Menage <menage@google.com>
Cc: Li Zefan <lizf@cn.fujitsu.com>
Cc: Balbir Singh <balbir@in.ibm.com>
Cc: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-04-02 16:57:38 -07:00
|
|
|
|
2022-03-22 14:40:19 -07:00
|
|
|
|
2018-12-28 00:39:57 -08:00
|
|
|
|
2018-08-17 15:47:11 -07:00
|
|
|
|
2022-03-22 14:40:19 -07:00
|
|
|
|
2018-08-17 15:47:11 -07:00
|
|
|
|
mm: don't raise MEMCG_OOM event due to failed high-order allocation
It was reported that on some of our machines containers were restarted
with OOM symptoms without an obvious reason. Despite there were almost no
memory pressure and plenty of page cache, MEMCG_OOM event was raised
occasionally, causing the container management software to think, that OOM
has happened. However, no tasks have been killed.
The following investigation showed that the problem is caused by a failing
attempt to charge a high-order page. In such case, the OOM killer is
never invoked. As shown below, it can happen under conditions, which are
very far from a real OOM: e.g. there is plenty of clean page cache and no
memory pressure.
There is no sense in raising an OOM event in this case, as it might
confuse a user and lead to wrong and excessive actions (e.g. restart the
workload, as in my case).
Let's look at the charging path in try_charge(). If the memory usage is
about memory.max, which is absolutely natural for most memory cgroups, we
try to reclaim some pages. Even if we were able to reclaim enough memory
for the allocation, the following check can fail due to a race with
another concurrent allocation:
if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
goto retry;
For regular pages the following condition will save us from triggering
the OOM:
if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
goto retry;
But for high-order allocation this condition will intentionally fail. The
reason behind is that we'll likely fall to regular pages anyway, so it's
ok and even preferred to return ENOMEM.
In this case the idea of raising MEMCG_OOM looks dubious.
Fix this by moving MEMCG_OOM raising to mem_cgroup_oom() after allocation
order check, so that the event won't be raised for high order allocations.
This change doesn't affect regular pages allocation and charging.
Link: http://lkml.kernel.org/r/20181004214050.7417-1-guro@fb.com
Signed-off-by: Roman Gushchin <guro@fb.com>
Acked-by: David Rientjes <rientjes@google.com>
Acked-by: Michal Hocko <mhocko@kernel.org>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-26 15:09:48 -07:00
|
|
|
|
|
|
|
|
|
2024-06-24 17:59:00 -07:00
|
|
|
|
2022-03-22 14:40:19 -07:00
|
|
|
|
2018-08-17 15:47:11 -07:00
|
|
|
|
2022-03-22 14:40:19 -07:00
|
|
|
|
2018-12-28 00:39:57 -08:00
|
|
|
|
2024-06-24 17:59:00 -07:00
|
|
|
|
2018-08-17 15:47:11 -07:00
|
|
|
|
2018-12-28 00:39:57 -08:00
|
|
|
|
mm: memcg: do not trap chargers with full callstack on OOM
The memcg OOM handling is incredibly fragile and can deadlock. When a
task fails to charge memory, it invokes the OOM killer and loops right
there in the charge code until it succeeds. Comparably, any other task
that enters the charge path at this point will go to a waitqueue right
then and there and sleep until the OOM situation is resolved. The problem
is that these tasks may hold filesystem locks and the mmap_sem; locks that
the selected OOM victim may need to exit.
For example, in one reported case, the task invoking the OOM killer was
about to charge a page cache page during a write(), which holds the
i_mutex. The OOM killer selected a task that was just entering truncate()
and trying to acquire the i_mutex:
OOM invoking task:
mem_cgroup_handle_oom+0x241/0x3b0
mem_cgroup_cache_charge+0xbe/0xe0
add_to_page_cache_locked+0x4c/0x140
add_to_page_cache_lru+0x22/0x50
grab_cache_page_write_begin+0x8b/0xe0
ext3_write_begin+0x88/0x270
generic_file_buffered_write+0x116/0x290
__generic_file_aio_write+0x27c/0x480
generic_file_aio_write+0x76/0xf0 # takes ->i_mutex
do_sync_write+0xea/0x130
vfs_write+0xf3/0x1f0
sys_write+0x51/0x90
system_call_fastpath+0x18/0x1d
OOM kill victim:
do_truncate+0x58/0xa0 # takes i_mutex
do_last+0x250/0xa30
path_openat+0xd7/0x440
do_filp_open+0x49/0xa0
do_sys_open+0x106/0x240
sys_open+0x20/0x30
system_call_fastpath+0x18/0x1d
The OOM handling task will retry the charge indefinitely while the OOM
killed task is not releasing any resources.
A similar scenario can happen when the kernel OOM killer for a memcg is
disabled and a userspace task is in charge of resolving OOM situations.
In this case, ALL tasks that enter the OOM path will be made to sleep on
the OOM waitqueue and wait for userspace to free resources or increase
the group's limit. But a userspace OOM handler is prone to deadlock
itself on the locks held by the waiting tasks. For example one of the
sleeping tasks may be stuck in a brk() call with the mmap_sem held for
writing but the userspace handler, in order to pick an optimal victim,
may need to read files from /proc/<pid>, which tries to acquire the same
mmap_sem for reading and deadlocks.
This patch changes the way tasks behave after detecting a memcg OOM and
makes sure nobody loops or sleeps with locks held:
1. When OOMing in a user fault, invoke the OOM killer and restart the
fault instead of looping on the charge attempt. This way, the OOM
victim can not get stuck on locks the looping task may hold.
2. When OOMing in a user fault but somebody else is handling it
(either the kernel OOM killer or a userspace handler), don't go to
sleep in the charge context. Instead, remember the OOMing memcg in
the task struct and then fully unwind the page fault stack with
-ENOMEM. pagefault_out_of_memory() will then call back into the
memcg code to check if the -ENOMEM came from the memcg, and then
either put the task to sleep on the memcg's OOM waitqueue or just
restart the fault. The OOM victim can no longer get stuck on any
lock a sleeping task may hold.
Debugged by Michal Hocko.
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reported-by: azurIt <azurit@pobox.sk>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-12 15:13:44 -07:00
|
|
|
|
|
|
|
|
|
2018-08-21 21:53:54 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-09-30 19:14:33 +05:30
|
|
|
|
2018-08-21 21:53:54 -07:00
|
|
|
|
|
|
|
|
|
2020-04-01 21:07:39 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-08-21 21:53:54 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-03-06 23:41:35 +08:00
|
|
|
|
2018-08-21 21:53:54 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
memcg: multi-memcg percpu charge cache
Memory cgroup accounting is expensive and to reduce the cost, the kernel
maintains per-cpu charge cache for a single memcg. So, if a charge
request comes for a different memcg, the kernel will flush the old memcg's
charge cache and then charge the newer memcg a fixed amount (64 pages),
subtracts the charge request amount and stores the remaining in the
per-cpu charge cache for the newer memcg.
This mechanism is based on the assumption that the kernel, for locality,
keep a process on a CPU for long period of time and most of the charge
requests from that process will be served by that CPU's local charge
cache.
However this assumption breaks down for incoming network traffic in a
multi-tenant machine. We are in the process of running multiple workloads
on a single machine and if such workloads are network heavy, we are seeing
very high network memory accounting cost. We have observed multiple CPUs
spending almost 100% of their time in net_rx_action and almost all of that
time is spent in memcg accounting of the network traffic.
More precisely, net_rx_action is serving packets from multiple workloads
and is observing/serving mix of packets of these workloads. The memcg
switch of per-cpu cache is very expensive and we are observing a lot of
memcg switches on the machine. Almost all the time is being spent on
charging new memcg and flushing older memcg cache. So, definitely we need
per-cpu cache that support multiple memcgs for this scenario.
This patch implements a simple (and dumb) multiple memcg percpu charge
cache. Actually we started with more sophisticated LRU based approach but
the dumb one was always better than the sophisticated one by 1% to 3%, so
going with the simple approach.
Some of the design choices are:
1. Fit all caches memcgs in a single cacheline.
2. The cache array can be mix of empty slots or memcg charged slots, so
the kernel has to traverse the full array.
3. The cache drain from the reclaim will drain all cached memcgs to keep
things simple.
To evaluate the impact of this optimization, on a 72 CPUs machine, we ran
the following workload where each netperf client runs in a different
cgroup. The next-20250415 kernel is used as base.
$ netserver -6
$ netperf -6 -H ::1 -l 60 -t TCP_SENDFILE -- -m 10K
number of clients | Without patch | With patch
6 | 42584.1 Mbps | 48603.4 Mbps (14.13% improvement)
12 | 30617.1 Mbps | 47919.7 Mbps (56.51% improvement)
18 | 25305.2 Mbps | 45497.3 Mbps (79.79% improvement)
24 | 20104.1 Mbps | 37907.7 Mbps (88.55% improvement)
30 | 14702.4 Mbps | 30746.5 Mbps (109.12% improvement)
36 | 10801.5 Mbps | 26476.3 Mbps (145.11% improvement)
The results show drastic improvement for network intensive workloads.
[shakeel.butt@linux.dev: add BUILD_BUG_ON() for MEMCG_CHARGE_BATCH]
Link: https://lkml.kernel.org/r/rlsgeosg3j7v5nihhbxxxbv3xfy4ejvigihj7lkkbt3n6imyne@2apxx2jm2e57
[shakeel.butt@linux.dev: simplify refill_stock]
Link: https://lkml.kernel.org/r/as5cdsm4lraxupg3t6onep2ixql72za25hvd4x334dsoyo4apr@zyzl4vkuevuv
[hughd@google.com: it's better to stock nr_pages than the uninitialized stock_pages]
Link: https://lkml.kernel.org/r/d542d18f-1caa-6fea-e2c3-3555c87bcf64@google.com
[shakeel.butt@linux.dev: add comment per Michal and use DEFINE_PER_CPU_ALIGNED instead of DEFINE_PER_CPU per Vlastimil]
Link: https://lkml.kernel.org/r/dieeei3squ2gcnqxdjayvxbvzldr266rhnvtl3vjzsqevxkevf@ckui5vjzl2qg
Link: https://lkml.kernel.org/r/20250416180229.2902751-1-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Jakub Kicinski <kuba@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Eric Dumaze <edumazet@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-16 11:02:29 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
2022-03-22 14:40:35 -07:00
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
memcg: multi-memcg percpu charge cache
Memory cgroup accounting is expensive and to reduce the cost, the kernel
maintains per-cpu charge cache for a single memcg. So, if a charge
request comes for a different memcg, the kernel will flush the old memcg's
charge cache and then charge the newer memcg a fixed amount (64 pages),
subtracts the charge request amount and stores the remaining in the
per-cpu charge cache for the newer memcg.
This mechanism is based on the assumption that the kernel, for locality,
keep a process on a CPU for long period of time and most of the charge
requests from that process will be served by that CPU's local charge
cache.
However this assumption breaks down for incoming network traffic in a
multi-tenant machine. We are in the process of running multiple workloads
on a single machine and if such workloads are network heavy, we are seeing
very high network memory accounting cost. We have observed multiple CPUs
spending almost 100% of their time in net_rx_action and almost all of that
time is spent in memcg accounting of the network traffic.
More precisely, net_rx_action is serving packets from multiple workloads
and is observing/serving mix of packets of these workloads. The memcg
switch of per-cpu cache is very expensive and we are observing a lot of
memcg switches on the machine. Almost all the time is being spent on
charging new memcg and flushing older memcg cache. So, definitely we need
per-cpu cache that support multiple memcgs for this scenario.
This patch implements a simple (and dumb) multiple memcg percpu charge
cache. Actually we started with more sophisticated LRU based approach but
the dumb one was always better than the sophisticated one by 1% to 3%, so
going with the simple approach.
Some of the design choices are:
1. Fit all caches memcgs in a single cacheline.
2. The cache array can be mix of empty slots or memcg charged slots, so
the kernel has to traverse the full array.
3. The cache drain from the reclaim will drain all cached memcgs to keep
things simple.
To evaluate the impact of this optimization, on a 72 CPUs machine, we ran
the following workload where each netperf client runs in a different
cgroup. The next-20250415 kernel is used as base.
$ netserver -6
$ netperf -6 -H ::1 -l 60 -t TCP_SENDFILE -- -m 10K
number of clients | Without patch | With patch
6 | 42584.1 Mbps | 48603.4 Mbps (14.13% improvement)
12 | 30617.1 Mbps | 47919.7 Mbps (56.51% improvement)
18 | 25305.2 Mbps | 45497.3 Mbps (79.79% improvement)
24 | 20104.1 Mbps | 37907.7 Mbps (88.55% improvement)
30 | 14702.4 Mbps | 30746.5 Mbps (109.12% improvement)
36 | 10801.5 Mbps | 26476.3 Mbps (145.11% improvement)
The results show drastic improvement for network intensive workloads.
[shakeel.butt@linux.dev: add BUILD_BUG_ON() for MEMCG_CHARGE_BATCH]
Link: https://lkml.kernel.org/r/rlsgeosg3j7v5nihhbxxxbv3xfy4ejvigihj7lkkbt3n6imyne@2apxx2jm2e57
[shakeel.butt@linux.dev: simplify refill_stock]
Link: https://lkml.kernel.org/r/as5cdsm4lraxupg3t6onep2ixql72za25hvd4x334dsoyo4apr@zyzl4vkuevuv
[hughd@google.com: it's better to stock nr_pages than the uninitialized stock_pages]
Link: https://lkml.kernel.org/r/d542d18f-1caa-6fea-e2c3-3555c87bcf64@google.com
[shakeel.butt@linux.dev: add comment per Michal and use DEFINE_PER_CPU_ALIGNED instead of DEFINE_PER_CPU per Vlastimil]
Link: https://lkml.kernel.org/r/dieeei3squ2gcnqxdjayvxbvzldr266rhnvtl3vjzsqevxkevf@ckui5vjzl2qg
Link: https://lkml.kernel.org/r/20250416180229.2902751-1-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Jakub Kicinski <kuba@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Eric Dumaze <edumazet@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-16 11:02:29 -07:00
|
|
|
|
|
|
|
|
|
2022-03-22 14:40:35 -07:00
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-05-06 15:55:31 -07:00
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
2021-06-28 19:37:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
2009-12-15 16:47:08 -08:00
|
|
|
|
2011-06-15 15:08:45 -07:00
|
|
|
|
2009-12-15 16:47:08 -08:00
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-03-22 14:40:47 -07:00
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
2011-08-09 11:56:26 +02:00
|
|
|
|
2009-12-15 16:47:08 -08:00
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
|
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
|
|
|
|
|
2012-12-18 14:21:36 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-05-06 15:55:30 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-12-18 14:21:36 -08:00
|
|
|
|
|
|
|
|
|
2009-12-15 16:47:08 -08:00
|
|
|
|
2025-05-06 15:55:30 -07:00
|
|
|
|
2009-12-15 16:47:08 -08:00
|
|
|
|
|
|
|
|
|
memcg: multi-memcg percpu charge cache
Memory cgroup accounting is expensive and to reduce the cost, the kernel
maintains per-cpu charge cache for a single memcg. So, if a charge
request comes for a different memcg, the kernel will flush the old memcg's
charge cache and then charge the newer memcg a fixed amount (64 pages),
subtracts the charge request amount and stores the remaining in the
per-cpu charge cache for the newer memcg.
This mechanism is based on the assumption that the kernel, for locality,
keep a process on a CPU for long period of time and most of the charge
requests from that process will be served by that CPU's local charge
cache.
However this assumption breaks down for incoming network traffic in a
multi-tenant machine. We are in the process of running multiple workloads
on a single machine and if such workloads are network heavy, we are seeing
very high network memory accounting cost. We have observed multiple CPUs
spending almost 100% of their time in net_rx_action and almost all of that
time is spent in memcg accounting of the network traffic.
More precisely, net_rx_action is serving packets from multiple workloads
and is observing/serving mix of packets of these workloads. The memcg
switch of per-cpu cache is very expensive and we are observing a lot of
memcg switches on the machine. Almost all the time is being spent on
charging new memcg and flushing older memcg cache. So, definitely we need
per-cpu cache that support multiple memcgs for this scenario.
This patch implements a simple (and dumb) multiple memcg percpu charge
cache. Actually we started with more sophisticated LRU based approach but
the dumb one was always better than the sophisticated one by 1% to 3%, so
going with the simple approach.
Some of the design choices are:
1. Fit all caches memcgs in a single cacheline.
2. The cache array can be mix of empty slots or memcg charged slots, so
the kernel has to traverse the full array.
3. The cache drain from the reclaim will drain all cached memcgs to keep
things simple.
To evaluate the impact of this optimization, on a 72 CPUs machine, we ran
the following workload where each netperf client runs in a different
cgroup. The next-20250415 kernel is used as base.
$ netserver -6
$ netperf -6 -H ::1 -l 60 -t TCP_SENDFILE -- -m 10K
number of clients | Without patch | With patch
6 | 42584.1 Mbps | 48603.4 Mbps (14.13% improvement)
12 | 30617.1 Mbps | 47919.7 Mbps (56.51% improvement)
18 | 25305.2 Mbps | 45497.3 Mbps (79.79% improvement)
24 | 20104.1 Mbps | 37907.7 Mbps (88.55% improvement)
30 | 14702.4 Mbps | 30746.5 Mbps (109.12% improvement)
36 | 10801.5 Mbps | 26476.3 Mbps (145.11% improvement)
The results show drastic improvement for network intensive workloads.
[shakeel.butt@linux.dev: add BUILD_BUG_ON() for MEMCG_CHARGE_BATCH]
Link: https://lkml.kernel.org/r/rlsgeosg3j7v5nihhbxxxbv3xfy4ejvigihj7lkkbt3n6imyne@2apxx2jm2e57
[shakeel.butt@linux.dev: simplify refill_stock]
Link: https://lkml.kernel.org/r/as5cdsm4lraxupg3t6onep2ixql72za25hvd4x334dsoyo4apr@zyzl4vkuevuv
[hughd@google.com: it's better to stock nr_pages than the uninitialized stock_pages]
Link: https://lkml.kernel.org/r/d542d18f-1caa-6fea-e2c3-3555c87bcf64@google.com
[shakeel.butt@linux.dev: add comment per Michal and use DEFINE_PER_CPU_ALIGNED instead of DEFINE_PER_CPU per Vlastimil]
Link: https://lkml.kernel.org/r/dieeei3squ2gcnqxdjayvxbvzldr266rhnvtl3vjzsqevxkevf@ckui5vjzl2qg
Link: https://lkml.kernel.org/r/20250416180229.2902751-1-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Jakub Kicinski <kuba@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Eric Dumaze <edumazet@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-16 11:02:29 -07:00
|
|
|
|
mm: memcontrol: lockless page counters
Memory is internally accounted in bytes, using spinlock-protected 64-bit
counters, even though the smallest accounting delta is a page. The
counter interface is also convoluted and does too many things.
Introduce a new lockless word-sized page counter API, then change all
memory accounting over to it. The translation from and to bytes then only
happens when interfacing with userspace.
The removed locking overhead is noticable when scaling beyond the per-cpu
charge caches - on a 4-socket machine with 144-threads, the following test
shows the performance differences of 288 memcgs concurrently running a
page fault benchmark:
vanilla:
18631648.500498 task-clock (msec) # 140.643 CPUs utilized ( +- 0.33% )
1,380,638 context-switches # 0.074 K/sec ( +- 0.75% )
24,390 cpu-migrations # 0.001 K/sec ( +- 8.44% )
1,843,305,768 page-faults # 0.099 M/sec ( +- 0.00% )
50,134,994,088,218 cycles # 2.691 GHz ( +- 0.33% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
8,049,712,224,651 instructions # 0.16 insns per cycle ( +- 0.04% )
1,586,970,584,979 branches # 85.176 M/sec ( +- 0.05% )
1,724,989,949 branch-misses # 0.11% of all branches ( +- 0.48% )
132.474343877 seconds time elapsed ( +- 0.21% )
lockless:
12195979.037525 task-clock (msec) # 133.480 CPUs utilized ( +- 0.18% )
832,850 context-switches # 0.068 K/sec ( +- 0.54% )
15,624 cpu-migrations # 0.001 K/sec ( +- 10.17% )
1,843,304,774 page-faults # 0.151 M/sec ( +- 0.00% )
32,811,216,801,141 cycles # 2.690 GHz ( +- 0.18% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
9,999,265,091,727 instructions # 0.30 insns per cycle ( +- 0.10% )
2,076,759,325,203 branches # 170.282 M/sec ( +- 0.12% )
1,656,917,214 branch-misses # 0.08% of all branches ( +- 0.55% )
91.369330729 seconds time elapsed ( +- 0.45% )
On top of improved scalability, this also gets rid of the icky long long
types in the very heart of memcg, which is great for 32 bit and also makes
the code a lot more readable.
Notable differences between the old and new API:
- res_counter_charge() and res_counter_charge_nofail() become
page_counter_try_charge() and page_counter_charge() resp. to match
the more common kernel naming scheme of try_do()/do()
- res_counter_uncharge_until() is only ever used to cancel a local
counter and never to uncharge bigger segments of a hierarchy, so
it's replaced by the simpler page_counter_cancel()
- res_counter_set_limit() is replaced by page_counter_limit(), which
expects its callers to serialize against themselves
- res_counter_memparse_write_strategy() is replaced by
page_counter_limit(), which rounds down to the nearest page size -
rather than up. This is more reasonable for explicitely requested
hard upper limits.
- to keep charging light-weight, page_counter_try_charge() charges
speculatively, only to roll back if the result exceeds the limit.
Because of this, a failing bigger charge can temporarily lock out
smaller charges that would otherwise succeed. The error is bounded
to the difference between the smallest and the biggest possible
charge size, so for memcg, this means that a failing THP charge can
send base page charges into reclaim upto 2MB (4MB) before the limit
would have been reached. This should be acceptable.
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE and memparse]
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE, memparse, strncmp, and PAGE_SIZE]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Vladimir Davydov <vdavydov@parallels.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-12-10 15:42:31 -08:00
|
|
|
|
memcg: multi-memcg percpu charge cache
Memory cgroup accounting is expensive and to reduce the cost, the kernel
maintains per-cpu charge cache for a single memcg. So, if a charge
request comes for a different memcg, the kernel will flush the old memcg's
charge cache and then charge the newer memcg a fixed amount (64 pages),
subtracts the charge request amount and stores the remaining in the
per-cpu charge cache for the newer memcg.
This mechanism is based on the assumption that the kernel, for locality,
keep a process on a CPU for long period of time and most of the charge
requests from that process will be served by that CPU's local charge
cache.
However this assumption breaks down for incoming network traffic in a
multi-tenant machine. We are in the process of running multiple workloads
on a single machine and if such workloads are network heavy, we are seeing
very high network memory accounting cost. We have observed multiple CPUs
spending almost 100% of their time in net_rx_action and almost all of that
time is spent in memcg accounting of the network traffic.
More precisely, net_rx_action is serving packets from multiple workloads
and is observing/serving mix of packets of these workloads. The memcg
switch of per-cpu cache is very expensive and we are observing a lot of
memcg switches on the machine. Almost all the time is being spent on
charging new memcg and flushing older memcg cache. So, definitely we need
per-cpu cache that support multiple memcgs for this scenario.
This patch implements a simple (and dumb) multiple memcg percpu charge
cache. Actually we started with more sophisticated LRU based approach but
the dumb one was always better than the sophisticated one by 1% to 3%, so
going with the simple approach.
Some of the design choices are:
1. Fit all caches memcgs in a single cacheline.
2. The cache array can be mix of empty slots or memcg charged slots, so
the kernel has to traverse the full array.
3. The cache drain from the reclaim will drain all cached memcgs to keep
things simple.
To evaluate the impact of this optimization, on a 72 CPUs machine, we ran
the following workload where each netperf client runs in a different
cgroup. The next-20250415 kernel is used as base.
$ netserver -6
$ netperf -6 -H ::1 -l 60 -t TCP_SENDFILE -- -m 10K
number of clients | Without patch | With patch
6 | 42584.1 Mbps | 48603.4 Mbps (14.13% improvement)
12 | 30617.1 Mbps | 47919.7 Mbps (56.51% improvement)
18 | 25305.2 Mbps | 45497.3 Mbps (79.79% improvement)
24 | 20104.1 Mbps | 37907.7 Mbps (88.55% improvement)
30 | 14702.4 Mbps | 30746.5 Mbps (109.12% improvement)
36 | 10801.5 Mbps | 26476.3 Mbps (145.11% improvement)
The results show drastic improvement for network intensive workloads.
[shakeel.butt@linux.dev: add BUILD_BUG_ON() for MEMCG_CHARGE_BATCH]
Link: https://lkml.kernel.org/r/rlsgeosg3j7v5nihhbxxxbv3xfy4ejvigihj7lkkbt3n6imyne@2apxx2jm2e57
[shakeel.butt@linux.dev: simplify refill_stock]
Link: https://lkml.kernel.org/r/as5cdsm4lraxupg3t6onep2ixql72za25hvd4x334dsoyo4apr@zyzl4vkuevuv
[hughd@google.com: it's better to stock nr_pages than the uninitialized stock_pages]
Link: https://lkml.kernel.org/r/d542d18f-1caa-6fea-e2c3-3555c87bcf64@google.com
[shakeel.butt@linux.dev: add comment per Michal and use DEFINE_PER_CPU_ALIGNED instead of DEFINE_PER_CPU per Vlastimil]
Link: https://lkml.kernel.org/r/dieeei3squ2gcnqxdjayvxbvzldr266rhnvtl3vjzsqevxkevf@ckui5vjzl2qg
Link: https://lkml.kernel.org/r/20250416180229.2902751-1-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Jakub Kicinski <kuba@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Eric Dumaze <edumazet@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-16 11:02:29 -07:00
|
|
|
|
2009-12-15 16:47:08 -08:00
|
|
|
|
2025-05-06 15:55:30 -07:00
|
|
|
|
2025-05-06 15:55:33 -07:00
|
|
|
|
locking/local_lock, mm: replace localtry_ helpers with local_trylock_t type
Partially revert commit 0aaddfb06882 ("locking/local_lock: Introduce
localtry_lock_t"). Remove localtry_*() helpers, since localtry_lock()
name might be misinterpreted as "try lock".
Introduce local_trylock[_irqsave]() helpers that only work with newly
introduced local_trylock_t type. Note that attempt to use
local_trylock[_irqsave]() with local_lock_t will cause compilation
failure.
Usage and behavior in !PREEMPT_RT:
local_lock_t lock; // sizeof(lock) == 0
local_lock(&lock); // preempt disable
local_lock_irqsave(&lock, ...); // irq save
if (local_trylock_irqsave(&lock, ...)) // compilation error
local_trylock_t lock; // sizeof(lock) == 4
local_lock(&lock); // preempt disable, acquired = 1
local_lock_irqsave(&lock, ...); // irq save, acquired = 1
if (local_trylock(&lock)) // if (!acquired) preempt disable, acquired = 1
if (local_trylock_irqsave(&lock, ...)) // if (!acquired) irq save, acquired = 1
The existing local_lock_*() macros can be used either with local_lock_t or
local_trylock_t. With local_trylock_t they set acquired = 1 while
local_unlock_*() clears it.
In !PREEMPT_RT local_lock_irqsave(local_lock_t *) disables interrupts to
protect critical section, but it doesn't prevent NMI, so the fully
reentrant code cannot use local_lock_irqsave(local_lock_t *) for exclusive
access.
The local_lock_irqsave(local_trylock_t *) helper disables interrupts and
sets acquired=1, so local_trylock_irqsave(local_trylock_t *) from NMI
attempting to acquire the same lock will return false.
In PREEMPT_RT local_lock_irqsave() maps to preemptible spin_lock(). Map
local_trylock_irqsave() to preemptible spin_trylock(). When in hard IRQ
or NMI return false right away, since spin_trylock() is not safe due to
explicit locking in the underneath rt_spin_trylock() implementation.
Removing this explicit locking and attempting only "trylock" is undesired
due to PI implications.
The local_trylock() without _irqsave can be used to avoid the cost of
disabling/enabling interrupts by only disabling preemption, so
local_trylock() in an interrupt attempting to acquire the same lock will
return false.
Note there is no need to use local_inc for acquired variable, since it's a
percpu variable with strict nesting scopes.
Note that guard(local_lock)(&lock) works only for "local_lock_t lock".
The patch also makes sure that local_lock_release(l) is called before
WRITE_ONCE(l->acquired, 0). Though IRQs are disabled at this point the
local_trylock() from NMI will succeed and local_lock_acquire(l) will warn.
Link: https://lkml.kernel.org/r/20250403025514.41186-1-alexei.starovoitov@gmail.com
Fixes: 0aaddfb06882 ("locking/local_lock: Introduce localtry_lock_t")
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Reviewed-by: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Daniel Borkman <daniel@iogearbox.net>
Cc: Linus Torvalds <torvalds@linuxfoundation.org>
Cc: Martin KaFai Lau <martin.lau@kernel.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-02 19:55:14 -07:00
|
|
|
|
2016-09-19 14:44:36 -07:00
|
|
|
|
|
|
|
|
|
memcg: multi-memcg percpu charge cache
Memory cgroup accounting is expensive and to reduce the cost, the kernel
maintains per-cpu charge cache for a single memcg. So, if a charge
request comes for a different memcg, the kernel will flush the old memcg's
charge cache and then charge the newer memcg a fixed amount (64 pages),
subtracts the charge request amount and stores the remaining in the
per-cpu charge cache for the newer memcg.
This mechanism is based on the assumption that the kernel, for locality,
keep a process on a CPU for long period of time and most of the charge
requests from that process will be served by that CPU's local charge
cache.
However this assumption breaks down for incoming network traffic in a
multi-tenant machine. We are in the process of running multiple workloads
on a single machine and if such workloads are network heavy, we are seeing
very high network memory accounting cost. We have observed multiple CPUs
spending almost 100% of their time in net_rx_action and almost all of that
time is spent in memcg accounting of the network traffic.
More precisely, net_rx_action is serving packets from multiple workloads
and is observing/serving mix of packets of these workloads. The memcg
switch of per-cpu cache is very expensive and we are observing a lot of
memcg switches on the machine. Almost all the time is being spent on
charging new memcg and flushing older memcg cache. So, definitely we need
per-cpu cache that support multiple memcgs for this scenario.
This patch implements a simple (and dumb) multiple memcg percpu charge
cache. Actually we started with more sophisticated LRU based approach but
the dumb one was always better than the sophisticated one by 1% to 3%, so
going with the simple approach.
Some of the design choices are:
1. Fit all caches memcgs in a single cacheline.
2. The cache array can be mix of empty slots or memcg charged slots, so
the kernel has to traverse the full array.
3. The cache drain from the reclaim will drain all cached memcgs to keep
things simple.
To evaluate the impact of this optimization, on a 72 CPUs machine, we ran
the following workload where each netperf client runs in a different
cgroup. The next-20250415 kernel is used as base.
$ netserver -6
$ netperf -6 -H ::1 -l 60 -t TCP_SENDFILE -- -m 10K
number of clients | Without patch | With patch
6 | 42584.1 Mbps | 48603.4 Mbps (14.13% improvement)
12 | 30617.1 Mbps | 47919.7 Mbps (56.51% improvement)
18 | 25305.2 Mbps | 45497.3 Mbps (79.79% improvement)
24 | 20104.1 Mbps | 37907.7 Mbps (88.55% improvement)
30 | 14702.4 Mbps | 30746.5 Mbps (109.12% improvement)
36 | 10801.5 Mbps | 26476.3 Mbps (145.11% improvement)
The results show drastic improvement for network intensive workloads.
[shakeel.butt@linux.dev: add BUILD_BUG_ON() for MEMCG_CHARGE_BATCH]
Link: https://lkml.kernel.org/r/rlsgeosg3j7v5nihhbxxxbv3xfy4ejvigihj7lkkbt3n6imyne@2apxx2jm2e57
[shakeel.butt@linux.dev: simplify refill_stock]
Link: https://lkml.kernel.org/r/as5cdsm4lraxupg3t6onep2ixql72za25hvd4x334dsoyo4apr@zyzl4vkuevuv
[hughd@google.com: it's better to stock nr_pages than the uninitialized stock_pages]
Link: https://lkml.kernel.org/r/d542d18f-1caa-6fea-e2c3-3555c87bcf64@google.com
[shakeel.butt@linux.dev: add comment per Michal and use DEFINE_PER_CPU_ALIGNED instead of DEFINE_PER_CPU per Vlastimil]
Link: https://lkml.kernel.org/r/dieeei3squ2gcnqxdjayvxbvzldr266rhnvtl3vjzsqevxkevf@ckui5vjzl2qg
Link: https://lkml.kernel.org/r/20250416180229.2902751-1-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Jakub Kicinski <kuba@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Eric Dumaze <edumazet@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-16 11:02:29 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcontrol: lockless page counters
Memory is internally accounted in bytes, using spinlock-protected 64-bit
counters, even though the smallest accounting delta is a page. The
counter interface is also convoluted and does too many things.
Introduce a new lockless word-sized page counter API, then change all
memory accounting over to it. The translation from and to bytes then only
happens when interfacing with userspace.
The removed locking overhead is noticable when scaling beyond the per-cpu
charge caches - on a 4-socket machine with 144-threads, the following test
shows the performance differences of 288 memcgs concurrently running a
page fault benchmark:
vanilla:
18631648.500498 task-clock (msec) # 140.643 CPUs utilized ( +- 0.33% )
1,380,638 context-switches # 0.074 K/sec ( +- 0.75% )
24,390 cpu-migrations # 0.001 K/sec ( +- 8.44% )
1,843,305,768 page-faults # 0.099 M/sec ( +- 0.00% )
50,134,994,088,218 cycles # 2.691 GHz ( +- 0.33% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
8,049,712,224,651 instructions # 0.16 insns per cycle ( +- 0.04% )
1,586,970,584,979 branches # 85.176 M/sec ( +- 0.05% )
1,724,989,949 branch-misses # 0.11% of all branches ( +- 0.48% )
132.474343877 seconds time elapsed ( +- 0.21% )
lockless:
12195979.037525 task-clock (msec) # 133.480 CPUs utilized ( +- 0.18% )
832,850 context-switches # 0.068 K/sec ( +- 0.54% )
15,624 cpu-migrations # 0.001 K/sec ( +- 10.17% )
1,843,304,774 page-faults # 0.151 M/sec ( +- 0.00% )
32,811,216,801,141 cycles # 2.690 GHz ( +- 0.18% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
9,999,265,091,727 instructions # 0.30 insns per cycle ( +- 0.10% )
2,076,759,325,203 branches # 170.282 M/sec ( +- 0.12% )
1,656,917,214 branch-misses # 0.08% of all branches ( +- 0.55% )
91.369330729 seconds time elapsed ( +- 0.45% )
On top of improved scalability, this also gets rid of the icky long long
types in the very heart of memcg, which is great for 32 bit and also makes
the code a lot more readable.
Notable differences between the old and new API:
- res_counter_charge() and res_counter_charge_nofail() become
page_counter_try_charge() and page_counter_charge() resp. to match
the more common kernel naming scheme of try_do()/do()
- res_counter_uncharge_until() is only ever used to cancel a local
counter and never to uncharge bigger segments of a hierarchy, so
it's replaced by the simpler page_counter_cancel()
- res_counter_set_limit() is replaced by page_counter_limit(), which
expects its callers to serialize against themselves
- res_counter_memparse_write_strategy() is replaced by
page_counter_limit(), which rounds down to the nearest page size -
rather than up. This is more reasonable for explicitely requested
hard upper limits.
- to keep charging light-weight, page_counter_try_charge() charges
speculatively, only to roll back if the result exceeds the limit.
Because of this, a failing bigger charge can temporarily lock out
smaller charges that would otherwise succeed. The error is bounded
to the difference between the smallest and the biggest possible
charge size, so for memcg, this means that a failing THP charge can
send base page charges into reclaim upto 2MB (4MB) before the limit
would have been reached. This should be acceptable.
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE and memparse]
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE, memparse, strncmp, and PAGE_SIZE]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Vladimir Davydov <vdavydov@parallels.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-12-10 15:42:31 -08:00
|
|
|
|
2016-09-19 14:44:36 -07:00
|
|
|
|
2025-05-06 15:55:33 -07:00
|
|
|
|
2016-09-19 14:44:36 -07:00
|
|
|
|
2009-12-15 16:47:08 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-04-03 18:39:07 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2009-12-15 16:47:08 -08:00
|
|
|
|
mm: memcontrol: lockless page counters
Memory is internally accounted in bytes, using spinlock-protected 64-bit
counters, even though the smallest accounting delta is a page. The
counter interface is also convoluted and does too many things.
Introduce a new lockless word-sized page counter API, then change all
memory accounting over to it. The translation from and to bytes then only
happens when interfacing with userspace.
The removed locking overhead is noticable when scaling beyond the per-cpu
charge caches - on a 4-socket machine with 144-threads, the following test
shows the performance differences of 288 memcgs concurrently running a
page fault benchmark:
vanilla:
18631648.500498 task-clock (msec) # 140.643 CPUs utilized ( +- 0.33% )
1,380,638 context-switches # 0.074 K/sec ( +- 0.75% )
24,390 cpu-migrations # 0.001 K/sec ( +- 8.44% )
1,843,305,768 page-faults # 0.099 M/sec ( +- 0.00% )
50,134,994,088,218 cycles # 2.691 GHz ( +- 0.33% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
8,049,712,224,651 instructions # 0.16 insns per cycle ( +- 0.04% )
1,586,970,584,979 branches # 85.176 M/sec ( +- 0.05% )
1,724,989,949 branch-misses # 0.11% of all branches ( +- 0.48% )
132.474343877 seconds time elapsed ( +- 0.21% )
lockless:
12195979.037525 task-clock (msec) # 133.480 CPUs utilized ( +- 0.18% )
832,850 context-switches # 0.068 K/sec ( +- 0.54% )
15,624 cpu-migrations # 0.001 K/sec ( +- 10.17% )
1,843,304,774 page-faults # 0.151 M/sec ( +- 0.00% )
32,811,216,801,141 cycles # 2.690 GHz ( +- 0.18% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
9,999,265,091,727 instructions # 0.30 insns per cycle ( +- 0.10% )
2,076,759,325,203 branches # 170.282 M/sec ( +- 0.12% )
1,656,917,214 branch-misses # 0.08% of all branches ( +- 0.55% )
91.369330729 seconds time elapsed ( +- 0.45% )
On top of improved scalability, this also gets rid of the icky long long
types in the very heart of memcg, which is great for 32 bit and also makes
the code a lot more readable.
Notable differences between the old and new API:
- res_counter_charge() and res_counter_charge_nofail() become
page_counter_try_charge() and page_counter_charge() resp. to match
the more common kernel naming scheme of try_do()/do()
- res_counter_uncharge_until() is only ever used to cancel a local
counter and never to uncharge bigger segments of a hierarchy, so
it's replaced by the simpler page_counter_cancel()
- res_counter_set_limit() is replaced by page_counter_limit(), which
expects its callers to serialize against themselves
- res_counter_memparse_write_strategy() is replaced by
page_counter_limit(), which rounds down to the nearest page size -
rather than up. This is more reasonable for explicitely requested
hard upper limits.
- to keep charging light-weight, page_counter_try_charge() charges
speculatively, only to roll back if the result exceeds the limit.
Because of this, a failing bigger charge can temporarily lock out
smaller charges that would otherwise succeed. The error is bounded
to the difference between the smallest and the biggest possible
charge size, so for memcg, this means that a failing THP charge can
send base page charges into reclaim upto 2MB (4MB) before the limit
would have been reached. This should be acceptable.
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE and memparse]
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE, memparse, strncmp, and PAGE_SIZE]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Vladimir Davydov <vdavydov@parallels.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-12-10 15:42:31 -08:00
|
|
|
|
2009-12-15 16:47:08 -08:00
|
|
|
|
memcg: multi-memcg percpu charge cache
Memory cgroup accounting is expensive and to reduce the cost, the kernel
maintains per-cpu charge cache for a single memcg. So, if a charge
request comes for a different memcg, the kernel will flush the old memcg's
charge cache and then charge the newer memcg a fixed amount (64 pages),
subtracts the charge request amount and stores the remaining in the
per-cpu charge cache for the newer memcg.
This mechanism is based on the assumption that the kernel, for locality,
keep a process on a CPU for long period of time and most of the charge
requests from that process will be served by that CPU's local charge
cache.
However this assumption breaks down for incoming network traffic in a
multi-tenant machine. We are in the process of running multiple workloads
on a single machine and if such workloads are network heavy, we are seeing
very high network memory accounting cost. We have observed multiple CPUs
spending almost 100% of their time in net_rx_action and almost all of that
time is spent in memcg accounting of the network traffic.
More precisely, net_rx_action is serving packets from multiple workloads
and is observing/serving mix of packets of these workloads. The memcg
switch of per-cpu cache is very expensive and we are observing a lot of
memcg switches on the machine. Almost all the time is being spent on
charging new memcg and flushing older memcg cache. So, definitely we need
per-cpu cache that support multiple memcgs for this scenario.
This patch implements a simple (and dumb) multiple memcg percpu charge
cache. Actually we started with more sophisticated LRU based approach but
the dumb one was always better than the sophisticated one by 1% to 3%, so
going with the simple approach.
Some of the design choices are:
1. Fit all caches memcgs in a single cacheline.
2. The cache array can be mix of empty slots or memcg charged slots, so
the kernel has to traverse the full array.
3. The cache drain from the reclaim will drain all cached memcgs to keep
things simple.
To evaluate the impact of this optimization, on a 72 CPUs machine, we ran
the following workload where each netperf client runs in a different
cgroup. The next-20250415 kernel is used as base.
$ netserver -6
$ netperf -6 -H ::1 -l 60 -t TCP_SENDFILE -- -m 10K
number of clients | Without patch | With patch
6 | 42584.1 Mbps | 48603.4 Mbps (14.13% improvement)
12 | 30617.1 Mbps | 47919.7 Mbps (56.51% improvement)
18 | 25305.2 Mbps | 45497.3 Mbps (79.79% improvement)
24 | 20104.1 Mbps | 37907.7 Mbps (88.55% improvement)
30 | 14702.4 Mbps | 30746.5 Mbps (109.12% improvement)
36 | 10801.5 Mbps | 26476.3 Mbps (145.11% improvement)
The results show drastic improvement for network intensive workloads.
[shakeel.butt@linux.dev: add BUILD_BUG_ON() for MEMCG_CHARGE_BATCH]
Link: https://lkml.kernel.org/r/rlsgeosg3j7v5nihhbxxxbv3xfy4ejvigihj7lkkbt3n6imyne@2apxx2jm2e57
[shakeel.butt@linux.dev: simplify refill_stock]
Link: https://lkml.kernel.org/r/as5cdsm4lraxupg3t6onep2ixql72za25hvd4x334dsoyo4apr@zyzl4vkuevuv
[hughd@google.com: it's better to stock nr_pages than the uninitialized stock_pages]
Link: https://lkml.kernel.org/r/d542d18f-1caa-6fea-e2c3-3555c87bcf64@google.com
[shakeel.butt@linux.dev: add comment per Michal and use DEFINE_PER_CPU_ALIGNED instead of DEFINE_PER_CPU per Vlastimil]
Link: https://lkml.kernel.org/r/dieeei3squ2gcnqxdjayvxbvzldr266rhnvtl3vjzsqevxkevf@ckui5vjzl2qg
Link: https://lkml.kernel.org/r/20250416180229.2902751-1-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Jakub Kicinski <kuba@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Eric Dumaze <edumazet@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-16 11:02:29 -07:00
|
|
|
|
2009-12-15 16:47:08 -08:00
|
|
|
|
memcg: multi-memcg percpu charge cache
Memory cgroup accounting is expensive and to reduce the cost, the kernel
maintains per-cpu charge cache for a single memcg. So, if a charge
request comes for a different memcg, the kernel will flush the old memcg's
charge cache and then charge the newer memcg a fixed amount (64 pages),
subtracts the charge request amount and stores the remaining in the
per-cpu charge cache for the newer memcg.
This mechanism is based on the assumption that the kernel, for locality,
keep a process on a CPU for long period of time and most of the charge
requests from that process will be served by that CPU's local charge
cache.
However this assumption breaks down for incoming network traffic in a
multi-tenant machine. We are in the process of running multiple workloads
on a single machine and if such workloads are network heavy, we are seeing
very high network memory accounting cost. We have observed multiple CPUs
spending almost 100% of their time in net_rx_action and almost all of that
time is spent in memcg accounting of the network traffic.
More precisely, net_rx_action is serving packets from multiple workloads
and is observing/serving mix of packets of these workloads. The memcg
switch of per-cpu cache is very expensive and we are observing a lot of
memcg switches on the machine. Almost all the time is being spent on
charging new memcg and flushing older memcg cache. So, definitely we need
per-cpu cache that support multiple memcgs for this scenario.
This patch implements a simple (and dumb) multiple memcg percpu charge
cache. Actually we started with more sophisticated LRU based approach but
the dumb one was always better than the sophisticated one by 1% to 3%, so
going with the simple approach.
Some of the design choices are:
1. Fit all caches memcgs in a single cacheline.
2. The cache array can be mix of empty slots or memcg charged slots, so
the kernel has to traverse the full array.
3. The cache drain from the reclaim will drain all cached memcgs to keep
things simple.
To evaluate the impact of this optimization, on a 72 CPUs machine, we ran
the following workload where each netperf client runs in a different
cgroup. The next-20250415 kernel is used as base.
$ netserver -6
$ netperf -6 -H ::1 -l 60 -t TCP_SENDFILE -- -m 10K
number of clients | Without patch | With patch
6 | 42584.1 Mbps | 48603.4 Mbps (14.13% improvement)
12 | 30617.1 Mbps | 47919.7 Mbps (56.51% improvement)
18 | 25305.2 Mbps | 45497.3 Mbps (79.79% improvement)
24 | 20104.1 Mbps | 37907.7 Mbps (88.55% improvement)
30 | 14702.4 Mbps | 30746.5 Mbps (109.12% improvement)
36 | 10801.5 Mbps | 26476.3 Mbps (145.11% improvement)
The results show drastic improvement for network intensive workloads.
[shakeel.butt@linux.dev: add BUILD_BUG_ON() for MEMCG_CHARGE_BATCH]
Link: https://lkml.kernel.org/r/rlsgeosg3j7v5nihhbxxxbv3xfy4ejvigihj7lkkbt3n6imyne@2apxx2jm2e57
[shakeel.butt@linux.dev: simplify refill_stock]
Link: https://lkml.kernel.org/r/as5cdsm4lraxupg3t6onep2ixql72za25hvd4x334dsoyo4apr@zyzl4vkuevuv
[hughd@google.com: it's better to stock nr_pages than the uninitialized stock_pages]
Link: https://lkml.kernel.org/r/d542d18f-1caa-6fea-e2c3-3555c87bcf64@google.com
[shakeel.butt@linux.dev: add comment per Michal and use DEFINE_PER_CPU_ALIGNED instead of DEFINE_PER_CPU per Vlastimil]
Link: https://lkml.kernel.org/r/dieeei3squ2gcnqxdjayvxbvzldr266rhnvtl3vjzsqevxkevf@ckui5vjzl2qg
Link: https://lkml.kernel.org/r/20250416180229.2902751-1-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Jakub Kicinski <kuba@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Eric Dumaze <edumazet@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-16 11:02:29 -07:00
|
|
|
|
|
|
|
|
|
2009-12-15 16:47:08 -08:00
|
|
|
|
2020-08-06 23:20:45 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
memcg: multi-memcg percpu charge cache
Memory cgroup accounting is expensive and to reduce the cost, the kernel
maintains per-cpu charge cache for a single memcg. So, if a charge
request comes for a different memcg, the kernel will flush the old memcg's
charge cache and then charge the newer memcg a fixed amount (64 pages),
subtracts the charge request amount and stores the remaining in the
per-cpu charge cache for the newer memcg.
This mechanism is based on the assumption that the kernel, for locality,
keep a process on a CPU for long period of time and most of the charge
requests from that process will be served by that CPU's local charge
cache.
However this assumption breaks down for incoming network traffic in a
multi-tenant machine. We are in the process of running multiple workloads
on a single machine and if such workloads are network heavy, we are seeing
very high network memory accounting cost. We have observed multiple CPUs
spending almost 100% of their time in net_rx_action and almost all of that
time is spent in memcg accounting of the network traffic.
More precisely, net_rx_action is serving packets from multiple workloads
and is observing/serving mix of packets of these workloads. The memcg
switch of per-cpu cache is very expensive and we are observing a lot of
memcg switches on the machine. Almost all the time is being spent on
charging new memcg and flushing older memcg cache. So, definitely we need
per-cpu cache that support multiple memcgs for this scenario.
This patch implements a simple (and dumb) multiple memcg percpu charge
cache. Actually we started with more sophisticated LRU based approach but
the dumb one was always better than the sophisticated one by 1% to 3%, so
going with the simple approach.
Some of the design choices are:
1. Fit all caches memcgs in a single cacheline.
2. The cache array can be mix of empty slots or memcg charged slots, so
the kernel has to traverse the full array.
3. The cache drain from the reclaim will drain all cached memcgs to keep
things simple.
To evaluate the impact of this optimization, on a 72 CPUs machine, we ran
the following workload where each netperf client runs in a different
cgroup. The next-20250415 kernel is used as base.
$ netserver -6
$ netperf -6 -H ::1 -l 60 -t TCP_SENDFILE -- -m 10K
number of clients | Without patch | With patch
6 | 42584.1 Mbps | 48603.4 Mbps (14.13% improvement)
12 | 30617.1 Mbps | 47919.7 Mbps (56.51% improvement)
18 | 25305.2 Mbps | 45497.3 Mbps (79.79% improvement)
24 | 20104.1 Mbps | 37907.7 Mbps (88.55% improvement)
30 | 14702.4 Mbps | 30746.5 Mbps (109.12% improvement)
36 | 10801.5 Mbps | 26476.3 Mbps (145.11% improvement)
The results show drastic improvement for network intensive workloads.
[shakeel.butt@linux.dev: add BUILD_BUG_ON() for MEMCG_CHARGE_BATCH]
Link: https://lkml.kernel.org/r/rlsgeosg3j7v5nihhbxxxbv3xfy4ejvigihj7lkkbt3n6imyne@2apxx2jm2e57
[shakeel.butt@linux.dev: simplify refill_stock]
Link: https://lkml.kernel.org/r/as5cdsm4lraxupg3t6onep2ixql72za25hvd4x334dsoyo4apr@zyzl4vkuevuv
[hughd@google.com: it's better to stock nr_pages than the uninitialized stock_pages]
Link: https://lkml.kernel.org/r/d542d18f-1caa-6fea-e2c3-3555c87bcf64@google.com
[shakeel.butt@linux.dev: add comment per Michal and use DEFINE_PER_CPU_ALIGNED instead of DEFINE_PER_CPU per Vlastimil]
Link: https://lkml.kernel.org/r/dieeei3squ2gcnqxdjayvxbvzldr266rhnvtl3vjzsqevxkevf@ckui5vjzl2qg
Link: https://lkml.kernel.org/r/20250416180229.2902751-1-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Jakub Kicinski <kuba@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Eric Dumaze <edumazet@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-16 11:02:29 -07:00
|
|
|
|
2024-05-01 02:54:20 -07:00
|
|
|
|
2025-04-03 18:39:07 -07:00
|
|
|
|
memcg: multi-memcg percpu charge cache
Memory cgroup accounting is expensive and to reduce the cost, the kernel
maintains per-cpu charge cache for a single memcg. So, if a charge
request comes for a different memcg, the kernel will flush the old memcg's
charge cache and then charge the newer memcg a fixed amount (64 pages),
subtracts the charge request amount and stores the remaining in the
per-cpu charge cache for the newer memcg.
This mechanism is based on the assumption that the kernel, for locality,
keep a process on a CPU for long period of time and most of the charge
requests from that process will be served by that CPU's local charge
cache.
However this assumption breaks down for incoming network traffic in a
multi-tenant machine. We are in the process of running multiple workloads
on a single machine and if such workloads are network heavy, we are seeing
very high network memory accounting cost. We have observed multiple CPUs
spending almost 100% of their time in net_rx_action and almost all of that
time is spent in memcg accounting of the network traffic.
More precisely, net_rx_action is serving packets from multiple workloads
and is observing/serving mix of packets of these workloads. The memcg
switch of per-cpu cache is very expensive and we are observing a lot of
memcg switches on the machine. Almost all the time is being spent on
charging new memcg and flushing older memcg cache. So, definitely we need
per-cpu cache that support multiple memcgs for this scenario.
This patch implements a simple (and dumb) multiple memcg percpu charge
cache. Actually we started with more sophisticated LRU based approach but
the dumb one was always better than the sophisticated one by 1% to 3%, so
going with the simple approach.
Some of the design choices are:
1. Fit all caches memcgs in a single cacheline.
2. The cache array can be mix of empty slots or memcg charged slots, so
the kernel has to traverse the full array.
3. The cache drain from the reclaim will drain all cached memcgs to keep
things simple.
To evaluate the impact of this optimization, on a 72 CPUs machine, we ran
the following workload where each netperf client runs in a different
cgroup. The next-20250415 kernel is used as base.
$ netserver -6
$ netperf -6 -H ::1 -l 60 -t TCP_SENDFILE -- -m 10K
number of clients | Without patch | With patch
6 | 42584.1 Mbps | 48603.4 Mbps (14.13% improvement)
12 | 30617.1 Mbps | 47919.7 Mbps (56.51% improvement)
18 | 25305.2 Mbps | 45497.3 Mbps (79.79% improvement)
24 | 20104.1 Mbps | 37907.7 Mbps (88.55% improvement)
30 | 14702.4 Mbps | 30746.5 Mbps (109.12% improvement)
36 | 10801.5 Mbps | 26476.3 Mbps (145.11% improvement)
The results show drastic improvement for network intensive workloads.
[shakeel.butt@linux.dev: add BUILD_BUG_ON() for MEMCG_CHARGE_BATCH]
Link: https://lkml.kernel.org/r/rlsgeosg3j7v5nihhbxxxbv3xfy4ejvigihj7lkkbt3n6imyne@2apxx2jm2e57
[shakeel.butt@linux.dev: simplify refill_stock]
Link: https://lkml.kernel.org/r/as5cdsm4lraxupg3t6onep2ixql72za25hvd4x334dsoyo4apr@zyzl4vkuevuv
[hughd@google.com: it's better to stock nr_pages than the uninitialized stock_pages]
Link: https://lkml.kernel.org/r/d542d18f-1caa-6fea-e2c3-3555c87bcf64@google.com
[shakeel.butt@linux.dev: add comment per Michal and use DEFINE_PER_CPU_ALIGNED instead of DEFINE_PER_CPU per Vlastimil]
Link: https://lkml.kernel.org/r/dieeei3squ2gcnqxdjayvxbvzldr266rhnvtl3vjzsqevxkevf@ckui5vjzl2qg
Link: https://lkml.kernel.org/r/20250416180229.2902751-1-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Jakub Kicinski <kuba@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Eric Dumaze <edumazet@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-16 11:02:29 -07:00
|
|
|
|
2009-12-15 16:47:08 -08:00
|
|
|
|
2020-08-06 23:20:45 -07:00
|
|
|
|
|
|
|
|
|
memcg: multi-memcg percpu charge cache
Memory cgroup accounting is expensive and to reduce the cost, the kernel
maintains per-cpu charge cache for a single memcg. So, if a charge
request comes for a different memcg, the kernel will flush the old memcg's
charge cache and then charge the newer memcg a fixed amount (64 pages),
subtracts the charge request amount and stores the remaining in the
per-cpu charge cache for the newer memcg.
This mechanism is based on the assumption that the kernel, for locality,
keep a process on a CPU for long period of time and most of the charge
requests from that process will be served by that CPU's local charge
cache.
However this assumption breaks down for incoming network traffic in a
multi-tenant machine. We are in the process of running multiple workloads
on a single machine and if such workloads are network heavy, we are seeing
very high network memory accounting cost. We have observed multiple CPUs
spending almost 100% of their time in net_rx_action and almost all of that
time is spent in memcg accounting of the network traffic.
More precisely, net_rx_action is serving packets from multiple workloads
and is observing/serving mix of packets of these workloads. The memcg
switch of per-cpu cache is very expensive and we are observing a lot of
memcg switches on the machine. Almost all the time is being spent on
charging new memcg and flushing older memcg cache. So, definitely we need
per-cpu cache that support multiple memcgs for this scenario.
This patch implements a simple (and dumb) multiple memcg percpu charge
cache. Actually we started with more sophisticated LRU based approach but
the dumb one was always better than the sophisticated one by 1% to 3%, so
going with the simple approach.
Some of the design choices are:
1. Fit all caches memcgs in a single cacheline.
2. The cache array can be mix of empty slots or memcg charged slots, so
the kernel has to traverse the full array.
3. The cache drain from the reclaim will drain all cached memcgs to keep
things simple.
To evaluate the impact of this optimization, on a 72 CPUs machine, we ran
the following workload where each netperf client runs in a different
cgroup. The next-20250415 kernel is used as base.
$ netserver -6
$ netperf -6 -H ::1 -l 60 -t TCP_SENDFILE -- -m 10K
number of clients | Without patch | With patch
6 | 42584.1 Mbps | 48603.4 Mbps (14.13% improvement)
12 | 30617.1 Mbps | 47919.7 Mbps (56.51% improvement)
18 | 25305.2 Mbps | 45497.3 Mbps (79.79% improvement)
24 | 20104.1 Mbps | 37907.7 Mbps (88.55% improvement)
30 | 14702.4 Mbps | 30746.5 Mbps (109.12% improvement)
36 | 10801.5 Mbps | 26476.3 Mbps (145.11% improvement)
The results show drastic improvement for network intensive workloads.
[shakeel.butt@linux.dev: add BUILD_BUG_ON() for MEMCG_CHARGE_BATCH]
Link: https://lkml.kernel.org/r/rlsgeosg3j7v5nihhbxxxbv3xfy4ejvigihj7lkkbt3n6imyne@2apxx2jm2e57
[shakeel.butt@linux.dev: simplify refill_stock]
Link: https://lkml.kernel.org/r/as5cdsm4lraxupg3t6onep2ixql72za25hvd4x334dsoyo4apr@zyzl4vkuevuv
[hughd@google.com: it's better to stock nr_pages than the uninitialized stock_pages]
Link: https://lkml.kernel.org/r/d542d18f-1caa-6fea-e2c3-3555c87bcf64@google.com
[shakeel.butt@linux.dev: add comment per Michal and use DEFINE_PER_CPU_ALIGNED instead of DEFINE_PER_CPU per Vlastimil]
Link: https://lkml.kernel.org/r/dieeei3squ2gcnqxdjayvxbvzldr266rhnvtl3vjzsqevxkevf@ckui5vjzl2qg
Link: https://lkml.kernel.org/r/20250416180229.2902751-1-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Jakub Kicinski <kuba@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Eric Dumaze <edumazet@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-16 11:02:29 -07:00
|
|
|
|
2009-12-15 16:47:08 -08:00
|
|
|
|
|
|
|
|
|
memcg: multi-memcg percpu charge cache
Memory cgroup accounting is expensive and to reduce the cost, the kernel
maintains per-cpu charge cache for a single memcg. So, if a charge
request comes for a different memcg, the kernel will flush the old memcg's
charge cache and then charge the newer memcg a fixed amount (64 pages),
subtracts the charge request amount and stores the remaining in the
per-cpu charge cache for the newer memcg.
This mechanism is based on the assumption that the kernel, for locality,
keep a process on a CPU for long period of time and most of the charge
requests from that process will be served by that CPU's local charge
cache.
However this assumption breaks down for incoming network traffic in a
multi-tenant machine. We are in the process of running multiple workloads
on a single machine and if such workloads are network heavy, we are seeing
very high network memory accounting cost. We have observed multiple CPUs
spending almost 100% of their time in net_rx_action and almost all of that
time is spent in memcg accounting of the network traffic.
More precisely, net_rx_action is serving packets from multiple workloads
and is observing/serving mix of packets of these workloads. The memcg
switch of per-cpu cache is very expensive and we are observing a lot of
memcg switches on the machine. Almost all the time is being spent on
charging new memcg and flushing older memcg cache. So, definitely we need
per-cpu cache that support multiple memcgs for this scenario.
This patch implements a simple (and dumb) multiple memcg percpu charge
cache. Actually we started with more sophisticated LRU based approach but
the dumb one was always better than the sophisticated one by 1% to 3%, so
going with the simple approach.
Some of the design choices are:
1. Fit all caches memcgs in a single cacheline.
2. The cache array can be mix of empty slots or memcg charged slots, so
the kernel has to traverse the full array.
3. The cache drain from the reclaim will drain all cached memcgs to keep
things simple.
To evaluate the impact of this optimization, on a 72 CPUs machine, we ran
the following workload where each netperf client runs in a different
cgroup. The next-20250415 kernel is used as base.
$ netserver -6
$ netperf -6 -H ::1 -l 60 -t TCP_SENDFILE -- -m 10K
number of clients | Without patch | With patch
6 | 42584.1 Mbps | 48603.4 Mbps (14.13% improvement)
12 | 30617.1 Mbps | 47919.7 Mbps (56.51% improvement)
18 | 25305.2 Mbps | 45497.3 Mbps (79.79% improvement)
24 | 20104.1 Mbps | 37907.7 Mbps (88.55% improvement)
30 | 14702.4 Mbps | 30746.5 Mbps (109.12% improvement)
36 | 10801.5 Mbps | 26476.3 Mbps (145.11% improvement)
The results show drastic improvement for network intensive workloads.
[shakeel.butt@linux.dev: add BUILD_BUG_ON() for MEMCG_CHARGE_BATCH]
Link: https://lkml.kernel.org/r/rlsgeosg3j7v5nihhbxxxbv3xfy4ejvigihj7lkkbt3n6imyne@2apxx2jm2e57
[shakeel.butt@linux.dev: simplify refill_stock]
Link: https://lkml.kernel.org/r/as5cdsm4lraxupg3t6onep2ixql72za25hvd4x334dsoyo4apr@zyzl4vkuevuv
[hughd@google.com: it's better to stock nr_pages than the uninitialized stock_pages]
Link: https://lkml.kernel.org/r/d542d18f-1caa-6fea-e2c3-3555c87bcf64@google.com
[shakeel.butt@linux.dev: add comment per Michal and use DEFINE_PER_CPU_ALIGNED instead of DEFINE_PER_CPU per Vlastimil]
Link: https://lkml.kernel.org/r/dieeei3squ2gcnqxdjayvxbvzldr266rhnvtl3vjzsqevxkevf@ckui5vjzl2qg
Link: https://lkml.kernel.org/r/20250416180229.2902751-1-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Jakub Kicinski <kuba@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Eric Dumaze <edumazet@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-16 11:02:29 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2009-12-15 16:47:08 -08:00
|
|
|
|
|
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
2009-12-15 16:47:08 -08:00
|
|
|
|
2016-09-19 14:44:36 -07:00
|
|
|
|
|
|
|
|
|
2025-05-06 15:55:31 -07:00
|
|
|
|
|
|
|
|
|
2016-09-19 14:44:36 -07:00
|
|
|
|
2025-05-06 15:55:33 -07:00
|
|
|
|
2016-09-19 14:44:36 -07:00
|
|
|
|
|
|
|
|
|
memcg: multi-memcg percpu charge cache
Memory cgroup accounting is expensive and to reduce the cost, the kernel
maintains per-cpu charge cache for a single memcg. So, if a charge
request comes for a different memcg, the kernel will flush the old memcg's
charge cache and then charge the newer memcg a fixed amount (64 pages),
subtracts the charge request amount and stores the remaining in the
per-cpu charge cache for the newer memcg.
This mechanism is based on the assumption that the kernel, for locality,
keep a process on a CPU for long period of time and most of the charge
requests from that process will be served by that CPU's local charge
cache.
However this assumption breaks down for incoming network traffic in a
multi-tenant machine. We are in the process of running multiple workloads
on a single machine and if such workloads are network heavy, we are seeing
very high network memory accounting cost. We have observed multiple CPUs
spending almost 100% of their time in net_rx_action and almost all of that
time is spent in memcg accounting of the network traffic.
More precisely, net_rx_action is serving packets from multiple workloads
and is observing/serving mix of packets of these workloads. The memcg
switch of per-cpu cache is very expensive and we are observing a lot of
memcg switches on the machine. Almost all the time is being spent on
charging new memcg and flushing older memcg cache. So, definitely we need
per-cpu cache that support multiple memcgs for this scenario.
This patch implements a simple (and dumb) multiple memcg percpu charge
cache. Actually we started with more sophisticated LRU based approach but
the dumb one was always better than the sophisticated one by 1% to 3%, so
going with the simple approach.
Some of the design choices are:
1. Fit all caches memcgs in a single cacheline.
2. The cache array can be mix of empty slots or memcg charged slots, so
the kernel has to traverse the full array.
3. The cache drain from the reclaim will drain all cached memcgs to keep
things simple.
To evaluate the impact of this optimization, on a 72 CPUs machine, we ran
the following workload where each netperf client runs in a different
cgroup. The next-20250415 kernel is used as base.
$ netserver -6
$ netperf -6 -H ::1 -l 60 -t TCP_SENDFILE -- -m 10K
number of clients | Without patch | With patch
6 | 42584.1 Mbps | 48603.4 Mbps (14.13% improvement)
12 | 30617.1 Mbps | 47919.7 Mbps (56.51% improvement)
18 | 25305.2 Mbps | 45497.3 Mbps (79.79% improvement)
24 | 20104.1 Mbps | 37907.7 Mbps (88.55% improvement)
30 | 14702.4 Mbps | 30746.5 Mbps (109.12% improvement)
36 | 10801.5 Mbps | 26476.3 Mbps (145.11% improvement)
The results show drastic improvement for network intensive workloads.
[shakeel.butt@linux.dev: add BUILD_BUG_ON() for MEMCG_CHARGE_BATCH]
Link: https://lkml.kernel.org/r/rlsgeosg3j7v5nihhbxxxbv3xfy4ejvigihj7lkkbt3n6imyne@2apxx2jm2e57
[shakeel.butt@linux.dev: simplify refill_stock]
Link: https://lkml.kernel.org/r/as5cdsm4lraxupg3t6onep2ixql72za25hvd4x334dsoyo4apr@zyzl4vkuevuv
[hughd@google.com: it's better to stock nr_pages than the uninitialized stock_pages]
Link: https://lkml.kernel.org/r/d542d18f-1caa-6fea-e2c3-3555c87bcf64@google.com
[shakeel.butt@linux.dev: add comment per Michal and use DEFINE_PER_CPU_ALIGNED instead of DEFINE_PER_CPU per Vlastimil]
Link: https://lkml.kernel.org/r/dieeei3squ2gcnqxdjayvxbvzldr266rhnvtl3vjzsqevxkevf@ckui5vjzl2qg
Link: https://lkml.kernel.org/r/20250416180229.2902751-1-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Jakub Kicinski <kuba@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Eric Dumaze <edumazet@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-16 11:02:29 -07:00
|
|
|
|
2011-06-15 15:08:45 -07:00
|
|
|
|
2016-09-19 14:44:36 -07:00
|
|
|
|
2025-05-06 15:55:33 -07:00
|
|
|
|
2009-12-15 16:47:08 -08:00
|
|
|
|
|
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
2009-12-15 16:47:08 -08:00
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
2009-12-15 16:47:08 -08:00
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-05-14 11:41:58 -07:00
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-09-19 14:44:36 -07:00
|
|
|
|
2025-05-14 11:41:58 -07:00
|
|
|
|
2022-03-22 14:40:44 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-09-19 14:44:36 -07:00
|
|
|
|
memcg: multi-memcg percpu charge cache
Memory cgroup accounting is expensive and to reduce the cost, the kernel
maintains per-cpu charge cache for a single memcg. So, if a charge
request comes for a different memcg, the kernel will flush the old memcg's
charge cache and then charge the newer memcg a fixed amount (64 pages),
subtracts the charge request amount and stores the remaining in the
per-cpu charge cache for the newer memcg.
This mechanism is based on the assumption that the kernel, for locality,
keep a process on a CPU for long period of time and most of the charge
requests from that process will be served by that CPU's local charge
cache.
However this assumption breaks down for incoming network traffic in a
multi-tenant machine. We are in the process of running multiple workloads
on a single machine and if such workloads are network heavy, we are seeing
very high network memory accounting cost. We have observed multiple CPUs
spending almost 100% of their time in net_rx_action and almost all of that
time is spent in memcg accounting of the network traffic.
More precisely, net_rx_action is serving packets from multiple workloads
and is observing/serving mix of packets of these workloads. The memcg
switch of per-cpu cache is very expensive and we are observing a lot of
memcg switches on the machine. Almost all the time is being spent on
charging new memcg and flushing older memcg cache. So, definitely we need
per-cpu cache that support multiple memcgs for this scenario.
This patch implements a simple (and dumb) multiple memcg percpu charge
cache. Actually we started with more sophisticated LRU based approach but
the dumb one was always better than the sophisticated one by 1% to 3%, so
going with the simple approach.
Some of the design choices are:
1. Fit all caches memcgs in a single cacheline.
2. The cache array can be mix of empty slots or memcg charged slots, so
the kernel has to traverse the full array.
3. The cache drain from the reclaim will drain all cached memcgs to keep
things simple.
To evaluate the impact of this optimization, on a 72 CPUs machine, we ran
the following workload where each netperf client runs in a different
cgroup. The next-20250415 kernel is used as base.
$ netserver -6
$ netperf -6 -H ::1 -l 60 -t TCP_SENDFILE -- -m 10K
number of clients | Without patch | With patch
6 | 42584.1 Mbps | 48603.4 Mbps (14.13% improvement)
12 | 30617.1 Mbps | 47919.7 Mbps (56.51% improvement)
18 | 25305.2 Mbps | 45497.3 Mbps (79.79% improvement)
24 | 20104.1 Mbps | 37907.7 Mbps (88.55% improvement)
30 | 14702.4 Mbps | 30746.5 Mbps (109.12% improvement)
36 | 10801.5 Mbps | 26476.3 Mbps (145.11% improvement)
The results show drastic improvement for network intensive workloads.
[shakeel.butt@linux.dev: add BUILD_BUG_ON() for MEMCG_CHARGE_BATCH]
Link: https://lkml.kernel.org/r/rlsgeosg3j7v5nihhbxxxbv3xfy4ejvigihj7lkkbt3n6imyne@2apxx2jm2e57
[shakeel.butt@linux.dev: simplify refill_stock]
Link: https://lkml.kernel.org/r/as5cdsm4lraxupg3t6onep2ixql72za25hvd4x334dsoyo4apr@zyzl4vkuevuv
[hughd@google.com: it's better to stock nr_pages than the uninitialized stock_pages]
Link: https://lkml.kernel.org/r/d542d18f-1caa-6fea-e2c3-3555c87bcf64@google.com
[shakeel.butt@linux.dev: add comment per Michal and use DEFINE_PER_CPU_ALIGNED instead of DEFINE_PER_CPU per Vlastimil]
Link: https://lkml.kernel.org/r/dieeei3squ2gcnqxdjayvxbvzldr266rhnvtl3vjzsqevxkevf@ckui5vjzl2qg
Link: https://lkml.kernel.org/r/20250416180229.2902751-1-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Jakub Kicinski <kuba@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Eric Dumaze <edumazet@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-16 11:02:29 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-09-08 16:13:09 -07:00
|
|
|
|
2025-04-03 18:39:08 -07:00
|
|
|
|
|
|
|
|
|
memcg: multi-memcg percpu charge cache
Memory cgroup accounting is expensive and to reduce the cost, the kernel
maintains per-cpu charge cache for a single memcg. So, if a charge
request comes for a different memcg, the kernel will flush the old memcg's
charge cache and then charge the newer memcg a fixed amount (64 pages),
subtracts the charge request amount and stores the remaining in the
per-cpu charge cache for the newer memcg.
This mechanism is based on the assumption that the kernel, for locality,
keep a process on a CPU for long period of time and most of the charge
requests from that process will be served by that CPU's local charge
cache.
However this assumption breaks down for incoming network traffic in a
multi-tenant machine. We are in the process of running multiple workloads
on a single machine and if such workloads are network heavy, we are seeing
very high network memory accounting cost. We have observed multiple CPUs
spending almost 100% of their time in net_rx_action and almost all of that
time is spent in memcg accounting of the network traffic.
More precisely, net_rx_action is serving packets from multiple workloads
and is observing/serving mix of packets of these workloads. The memcg
switch of per-cpu cache is very expensive and we are observing a lot of
memcg switches on the machine. Almost all the time is being spent on
charging new memcg and flushing older memcg cache. So, definitely we need
per-cpu cache that support multiple memcgs for this scenario.
This patch implements a simple (and dumb) multiple memcg percpu charge
cache. Actually we started with more sophisticated LRU based approach but
the dumb one was always better than the sophisticated one by 1% to 3%, so
going with the simple approach.
Some of the design choices are:
1. Fit all caches memcgs in a single cacheline.
2. The cache array can be mix of empty slots or memcg charged slots, so
the kernel has to traverse the full array.
3. The cache drain from the reclaim will drain all cached memcgs to keep
things simple.
To evaluate the impact of this optimization, on a 72 CPUs machine, we ran
the following workload where each netperf client runs in a different
cgroup. The next-20250415 kernel is used as base.
$ netserver -6
$ netperf -6 -H ::1 -l 60 -t TCP_SENDFILE -- -m 10K
number of clients | Without patch | With patch
6 | 42584.1 Mbps | 48603.4 Mbps (14.13% improvement)
12 | 30617.1 Mbps | 47919.7 Mbps (56.51% improvement)
18 | 25305.2 Mbps | 45497.3 Mbps (79.79% improvement)
24 | 20104.1 Mbps | 37907.7 Mbps (88.55% improvement)
30 | 14702.4 Mbps | 30746.5 Mbps (109.12% improvement)
36 | 10801.5 Mbps | 26476.3 Mbps (145.11% improvement)
The results show drastic improvement for network intensive workloads.
[shakeel.butt@linux.dev: add BUILD_BUG_ON() for MEMCG_CHARGE_BATCH]
Link: https://lkml.kernel.org/r/rlsgeosg3j7v5nihhbxxxbv3xfy4ejvigihj7lkkbt3n6imyne@2apxx2jm2e57
[shakeel.butt@linux.dev: simplify refill_stock]
Link: https://lkml.kernel.org/r/as5cdsm4lraxupg3t6onep2ixql72za25hvd4x334dsoyo4apr@zyzl4vkuevuv
[hughd@google.com: it's better to stock nr_pages than the uninitialized stock_pages]
Link: https://lkml.kernel.org/r/d542d18f-1caa-6fea-e2c3-3555c87bcf64@google.com
[shakeel.butt@linux.dev: add comment per Michal and use DEFINE_PER_CPU_ALIGNED instead of DEFINE_PER_CPU per Vlastimil]
Link: https://lkml.kernel.org/r/dieeei3squ2gcnqxdjayvxbvzldr266rhnvtl3vjzsqevxkevf@ckui5vjzl2qg
Link: https://lkml.kernel.org/r/20250416180229.2902751-1-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Jakub Kicinski <kuba@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Eric Dumaze <edumazet@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-16 11:02:29 -07:00
|
|
|
|
2025-05-06 15:55:33 -07:00
|
|
|
|
2025-02-21 18:44:25 -08:00
|
|
|
|
memcg: multi-memcg percpu charge cache
Memory cgroup accounting is expensive and to reduce the cost, the kernel
maintains per-cpu charge cache for a single memcg. So, if a charge
request comes for a different memcg, the kernel will flush the old memcg's
charge cache and then charge the newer memcg a fixed amount (64 pages),
subtracts the charge request amount and stores the remaining in the
per-cpu charge cache for the newer memcg.
This mechanism is based on the assumption that the kernel, for locality,
keep a process on a CPU for long period of time and most of the charge
requests from that process will be served by that CPU's local charge
cache.
However this assumption breaks down for incoming network traffic in a
multi-tenant machine. We are in the process of running multiple workloads
on a single machine and if such workloads are network heavy, we are seeing
very high network memory accounting cost. We have observed multiple CPUs
spending almost 100% of their time in net_rx_action and almost all of that
time is spent in memcg accounting of the network traffic.
More precisely, net_rx_action is serving packets from multiple workloads
and is observing/serving mix of packets of these workloads. The memcg
switch of per-cpu cache is very expensive and we are observing a lot of
memcg switches on the machine. Almost all the time is being spent on
charging new memcg and flushing older memcg cache. So, definitely we need
per-cpu cache that support multiple memcgs for this scenario.
This patch implements a simple (and dumb) multiple memcg percpu charge
cache. Actually we started with more sophisticated LRU based approach but
the dumb one was always better than the sophisticated one by 1% to 3%, so
going with the simple approach.
Some of the design choices are:
1. Fit all caches memcgs in a single cacheline.
2. The cache array can be mix of empty slots or memcg charged slots, so
the kernel has to traverse the full array.
3. The cache drain from the reclaim will drain all cached memcgs to keep
things simple.
To evaluate the impact of this optimization, on a 72 CPUs machine, we ran
the following workload where each netperf client runs in a different
cgroup. The next-20250415 kernel is used as base.
$ netserver -6
$ netperf -6 -H ::1 -l 60 -t TCP_SENDFILE -- -m 10K
number of clients | Without patch | With patch
6 | 42584.1 Mbps | 48603.4 Mbps (14.13% improvement)
12 | 30617.1 Mbps | 47919.7 Mbps (56.51% improvement)
18 | 25305.2 Mbps | 45497.3 Mbps (79.79% improvement)
24 | 20104.1 Mbps | 37907.7 Mbps (88.55% improvement)
30 | 14702.4 Mbps | 30746.5 Mbps (109.12% improvement)
36 | 10801.5 Mbps | 26476.3 Mbps (145.11% improvement)
The results show drastic improvement for network intensive workloads.
[shakeel.butt@linux.dev: add BUILD_BUG_ON() for MEMCG_CHARGE_BATCH]
Link: https://lkml.kernel.org/r/rlsgeosg3j7v5nihhbxxxbv3xfy4ejvigihj7lkkbt3n6imyne@2apxx2jm2e57
[shakeel.butt@linux.dev: simplify refill_stock]
Link: https://lkml.kernel.org/r/as5cdsm4lraxupg3t6onep2ixql72za25hvd4x334dsoyo4apr@zyzl4vkuevuv
[hughd@google.com: it's better to stock nr_pages than the uninitialized stock_pages]
Link: https://lkml.kernel.org/r/d542d18f-1caa-6fea-e2c3-3555c87bcf64@google.com
[shakeel.butt@linux.dev: add comment per Michal and use DEFINE_PER_CPU_ALIGNED instead of DEFINE_PER_CPU per Vlastimil]
Link: https://lkml.kernel.org/r/dieeei3squ2gcnqxdjayvxbvzldr266rhnvtl3vjzsqevxkevf@ckui5vjzl2qg
Link: https://lkml.kernel.org/r/20250416180229.2902751-1-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Jakub Kicinski <kuba@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Eric Dumaze <edumazet@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-16 11:02:29 -07:00
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
2025-02-21 18:44:25 -08:00
|
|
|
|
2025-04-03 18:39:08 -07:00
|
|
|
|
2025-02-21 18:44:25 -08:00
|
|
|
|
|
|
|
|
|
2009-12-15 16:47:08 -08:00
|
|
|
|
2016-09-19 14:44:36 -07:00
|
|
|
|
memcg: multi-memcg percpu charge cache
Memory cgroup accounting is expensive and to reduce the cost, the kernel
maintains per-cpu charge cache for a single memcg. So, if a charge
request comes for a different memcg, the kernel will flush the old memcg's
charge cache and then charge the newer memcg a fixed amount (64 pages),
subtracts the charge request amount and stores the remaining in the
per-cpu charge cache for the newer memcg.
This mechanism is based on the assumption that the kernel, for locality,
keep a process on a CPU for long period of time and most of the charge
requests from that process will be served by that CPU's local charge
cache.
However this assumption breaks down for incoming network traffic in a
multi-tenant machine. We are in the process of running multiple workloads
on a single machine and if such workloads are network heavy, we are seeing
very high network memory accounting cost. We have observed multiple CPUs
spending almost 100% of their time in net_rx_action and almost all of that
time is spent in memcg accounting of the network traffic.
More precisely, net_rx_action is serving packets from multiple workloads
and is observing/serving mix of packets of these workloads. The memcg
switch of per-cpu cache is very expensive and we are observing a lot of
memcg switches on the machine. Almost all the time is being spent on
charging new memcg and flushing older memcg cache. So, definitely we need
per-cpu cache that support multiple memcgs for this scenario.
This patch implements a simple (and dumb) multiple memcg percpu charge
cache. Actually we started with more sophisticated LRU based approach but
the dumb one was always better than the sophisticated one by 1% to 3%, so
going with the simple approach.
Some of the design choices are:
1. Fit all caches memcgs in a single cacheline.
2. The cache array can be mix of empty slots or memcg charged slots, so
the kernel has to traverse the full array.
3. The cache drain from the reclaim will drain all cached memcgs to keep
things simple.
To evaluate the impact of this optimization, on a 72 CPUs machine, we ran
the following workload where each netperf client runs in a different
cgroup. The next-20250415 kernel is used as base.
$ netserver -6
$ netperf -6 -H ::1 -l 60 -t TCP_SENDFILE -- -m 10K
number of clients | Without patch | With patch
6 | 42584.1 Mbps | 48603.4 Mbps (14.13% improvement)
12 | 30617.1 Mbps | 47919.7 Mbps (56.51% improvement)
18 | 25305.2 Mbps | 45497.3 Mbps (79.79% improvement)
24 | 20104.1 Mbps | 37907.7 Mbps (88.55% improvement)
30 | 14702.4 Mbps | 30746.5 Mbps (109.12% improvement)
36 | 10801.5 Mbps | 26476.3 Mbps (145.11% improvement)
The results show drastic improvement for network intensive workloads.
[shakeel.butt@linux.dev: add BUILD_BUG_ON() for MEMCG_CHARGE_BATCH]
Link: https://lkml.kernel.org/r/rlsgeosg3j7v5nihhbxxxbv3xfy4ejvigihj7lkkbt3n6imyne@2apxx2jm2e57
[shakeel.butt@linux.dev: simplify refill_stock]
Link: https://lkml.kernel.org/r/as5cdsm4lraxupg3t6onep2ixql72za25hvd4x334dsoyo4apr@zyzl4vkuevuv
[hughd@google.com: it's better to stock nr_pages than the uninitialized stock_pages]
Link: https://lkml.kernel.org/r/d542d18f-1caa-6fea-e2c3-3555c87bcf64@google.com
[shakeel.butt@linux.dev: add comment per Michal and use DEFINE_PER_CPU_ALIGNED instead of DEFINE_PER_CPU per Vlastimil]
Link: https://lkml.kernel.org/r/dieeei3squ2gcnqxdjayvxbvzldr266rhnvtl3vjzsqevxkevf@ckui5vjzl2qg
Link: https://lkml.kernel.org/r/20250416180229.2902751-1-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Jakub Kicinski <kuba@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Eric Dumaze <edumazet@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-16 11:02:29 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2009-12-15 16:47:08 -08:00
|
|
|
|
2016-09-19 14:44:36 -07:00
|
|
|
|
memcg: multi-memcg percpu charge cache
Memory cgroup accounting is expensive and to reduce the cost, the kernel
maintains per-cpu charge cache for a single memcg. So, if a charge
request comes for a different memcg, the kernel will flush the old memcg's
charge cache and then charge the newer memcg a fixed amount (64 pages),
subtracts the charge request amount and stores the remaining in the
per-cpu charge cache for the newer memcg.
This mechanism is based on the assumption that the kernel, for locality,
keep a process on a CPU for long period of time and most of the charge
requests from that process will be served by that CPU's local charge
cache.
However this assumption breaks down for incoming network traffic in a
multi-tenant machine. We are in the process of running multiple workloads
on a single machine and if such workloads are network heavy, we are seeing
very high network memory accounting cost. We have observed multiple CPUs
spending almost 100% of their time in net_rx_action and almost all of that
time is spent in memcg accounting of the network traffic.
More precisely, net_rx_action is serving packets from multiple workloads
and is observing/serving mix of packets of these workloads. The memcg
switch of per-cpu cache is very expensive and we are observing a lot of
memcg switches on the machine. Almost all the time is being spent on
charging new memcg and flushing older memcg cache. So, definitely we need
per-cpu cache that support multiple memcgs for this scenario.
This patch implements a simple (and dumb) multiple memcg percpu charge
cache. Actually we started with more sophisticated LRU based approach but
the dumb one was always better than the sophisticated one by 1% to 3%, so
going with the simple approach.
Some of the design choices are:
1. Fit all caches memcgs in a single cacheline.
2. The cache array can be mix of empty slots or memcg charged slots, so
the kernel has to traverse the full array.
3. The cache drain from the reclaim will drain all cached memcgs to keep
things simple.
To evaluate the impact of this optimization, on a 72 CPUs machine, we ran
the following workload where each netperf client runs in a different
cgroup. The next-20250415 kernel is used as base.
$ netserver -6
$ netperf -6 -H ::1 -l 60 -t TCP_SENDFILE -- -m 10K
number of clients | Without patch | With patch
6 | 42584.1 Mbps | 48603.4 Mbps (14.13% improvement)
12 | 30617.1 Mbps | 47919.7 Mbps (56.51% improvement)
18 | 25305.2 Mbps | 45497.3 Mbps (79.79% improvement)
24 | 20104.1 Mbps | 37907.7 Mbps (88.55% improvement)
30 | 14702.4 Mbps | 30746.5 Mbps (109.12% improvement)
36 | 10801.5 Mbps | 26476.3 Mbps (145.11% improvement)
The results show drastic improvement for network intensive workloads.
[shakeel.butt@linux.dev: add BUILD_BUG_ON() for MEMCG_CHARGE_BATCH]
Link: https://lkml.kernel.org/r/rlsgeosg3j7v5nihhbxxxbv3xfy4ejvigihj7lkkbt3n6imyne@2apxx2jm2e57
[shakeel.butt@linux.dev: simplify refill_stock]
Link: https://lkml.kernel.org/r/as5cdsm4lraxupg3t6onep2ixql72za25hvd4x334dsoyo4apr@zyzl4vkuevuv
[hughd@google.com: it's better to stock nr_pages than the uninitialized stock_pages]
Link: https://lkml.kernel.org/r/d542d18f-1caa-6fea-e2c3-3555c87bcf64@google.com
[shakeel.butt@linux.dev: add comment per Michal and use DEFINE_PER_CPU_ALIGNED instead of DEFINE_PER_CPU per Vlastimil]
Link: https://lkml.kernel.org/r/dieeei3squ2gcnqxdjayvxbvzldr266rhnvtl3vjzsqevxkevf@ckui5vjzl2qg
Link: https://lkml.kernel.org/r/20250416180229.2902751-1-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Jakub Kicinski <kuba@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Eric Dumaze <edumazet@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-16 11:02:29 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-04-03 18:39:05 -07:00
|
|
|
|
2025-05-06 15:55:33 -07:00
|
|
|
|
2009-12-15 16:47:08 -08:00
|
|
|
|
|
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
|
|
|
|
|
memcg: multi-memcg percpu charge cache
Memory cgroup accounting is expensive and to reduce the cost, the kernel
maintains per-cpu charge cache for a single memcg. So, if a charge
request comes for a different memcg, the kernel will flush the old memcg's
charge cache and then charge the newer memcg a fixed amount (64 pages),
subtracts the charge request amount and stores the remaining in the
per-cpu charge cache for the newer memcg.
This mechanism is based on the assumption that the kernel, for locality,
keep a process on a CPU for long period of time and most of the charge
requests from that process will be served by that CPU's local charge
cache.
However this assumption breaks down for incoming network traffic in a
multi-tenant machine. We are in the process of running multiple workloads
on a single machine and if such workloads are network heavy, we are seeing
very high network memory accounting cost. We have observed multiple CPUs
spending almost 100% of their time in net_rx_action and almost all of that
time is spent in memcg accounting of the network traffic.
More precisely, net_rx_action is serving packets from multiple workloads
and is observing/serving mix of packets of these workloads. The memcg
switch of per-cpu cache is very expensive and we are observing a lot of
memcg switches on the machine. Almost all the time is being spent on
charging new memcg and flushing older memcg cache. So, definitely we need
per-cpu cache that support multiple memcgs for this scenario.
This patch implements a simple (and dumb) multiple memcg percpu charge
cache. Actually we started with more sophisticated LRU based approach but
the dumb one was always better than the sophisticated one by 1% to 3%, so
going with the simple approach.
Some of the design choices are:
1. Fit all caches memcgs in a single cacheline.
2. The cache array can be mix of empty slots or memcg charged slots, so
the kernel has to traverse the full array.
3. The cache drain from the reclaim will drain all cached memcgs to keep
things simple.
To evaluate the impact of this optimization, on a 72 CPUs machine, we ran
the following workload where each netperf client runs in a different
cgroup. The next-20250415 kernel is used as base.
$ netserver -6
$ netperf -6 -H ::1 -l 60 -t TCP_SENDFILE -- -m 10K
number of clients | Without patch | With patch
6 | 42584.1 Mbps | 48603.4 Mbps (14.13% improvement)
12 | 30617.1 Mbps | 47919.7 Mbps (56.51% improvement)
18 | 25305.2 Mbps | 45497.3 Mbps (79.79% improvement)
24 | 20104.1 Mbps | 37907.7 Mbps (88.55% improvement)
30 | 14702.4 Mbps | 30746.5 Mbps (109.12% improvement)
36 | 10801.5 Mbps | 26476.3 Mbps (145.11% improvement)
The results show drastic improvement for network intensive workloads.
[shakeel.butt@linux.dev: add BUILD_BUG_ON() for MEMCG_CHARGE_BATCH]
Link: https://lkml.kernel.org/r/rlsgeosg3j7v5nihhbxxxbv3xfy4ejvigihj7lkkbt3n6imyne@2apxx2jm2e57
[shakeel.butt@linux.dev: simplify refill_stock]
Link: https://lkml.kernel.org/r/as5cdsm4lraxupg3t6onep2ixql72za25hvd4x334dsoyo4apr@zyzl4vkuevuv
[hughd@google.com: it's better to stock nr_pages than the uninitialized stock_pages]
Link: https://lkml.kernel.org/r/d542d18f-1caa-6fea-e2c3-3555c87bcf64@google.com
[shakeel.butt@linux.dev: add comment per Michal and use DEFINE_PER_CPU_ALIGNED instead of DEFINE_PER_CPU per Vlastimil]
Link: https://lkml.kernel.org/r/dieeei3squ2gcnqxdjayvxbvzldr266rhnvtl3vjzsqevxkevf@ckui5vjzl2qg
Link: https://lkml.kernel.org/r/20250416180229.2902751-1-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Jakub Kicinski <kuba@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Eric Dumaze <edumazet@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-16 11:02:29 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2009-12-15 16:47:08 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-11-02 13:38:15 -07:00
|
|
|
|
2014-12-10 15:42:50 -08:00
|
|
|
|
2009-12-15 16:47:08 -08:00
|
|
|
|
2024-06-24 17:59:02 -07:00
|
|
|
|
2009-12-15 16:47:08 -08:00
|
|
|
|
2011-06-15 15:08:45 -07:00
|
|
|
|
2011-07-26 16:08:28 -07:00
|
|
|
|
2014-12-10 15:42:50 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, memcg: remove hotplug locking from try_charge
The following lockdep splat has been noticed during LTP testing
======================================================
WARNING: possible circular locking dependency detected
4.13.0-rc3-next-20170807 #12 Not tainted
------------------------------------------------------
a.out/4771 is trying to acquire lock:
(cpu_hotplug_lock.rw_sem){++++++}, at: [<ffffffff812b4668>] drain_all_stock.part.35+0x18/0x140
but task is already holding lock:
(&mm->mmap_sem){++++++}, at: [<ffffffff8106eb35>] __do_page_fault+0x175/0x530
which lock already depends on the new lock.
the existing dependency chain (in reverse order) is:
-> #3 (&mm->mmap_sem){++++++}:
lock_acquire+0xc9/0x230
__might_fault+0x70/0xa0
_copy_to_user+0x23/0x70
filldir+0xa7/0x110
xfs_dir2_sf_getdents.isra.10+0x20c/0x2c0 [xfs]
xfs_readdir+0x1fa/0x2c0 [xfs]
xfs_file_readdir+0x30/0x40 [xfs]
iterate_dir+0x17a/0x1a0
SyS_getdents+0xb0/0x160
entry_SYSCALL_64_fastpath+0x1f/0xbe
-> #2 (&type->i_mutex_dir_key#3){++++++}:
lock_acquire+0xc9/0x230
down_read+0x51/0xb0
lookup_slow+0xde/0x210
walk_component+0x160/0x250
link_path_walk+0x1a6/0x610
path_openat+0xe4/0xd50
do_filp_open+0x91/0x100
file_open_name+0xf5/0x130
filp_open+0x33/0x50
kernel_read_file_from_path+0x39/0x80
_request_firmware+0x39f/0x880
request_firmware_direct+0x37/0x50
request_microcode_fw+0x64/0xe0
reload_store+0xf7/0x180
dev_attr_store+0x18/0x30
sysfs_kf_write+0x44/0x60
kernfs_fop_write+0x113/0x1a0
__vfs_write+0x37/0x170
vfs_write+0xc7/0x1c0
SyS_write+0x58/0xc0
do_syscall_64+0x6c/0x1f0
return_from_SYSCALL_64+0x0/0x7a
-> #1 (microcode_mutex){+.+.+.}:
lock_acquire+0xc9/0x230
__mutex_lock+0x88/0x960
mutex_lock_nested+0x1b/0x20
microcode_init+0xbb/0x208
do_one_initcall+0x51/0x1a9
kernel_init_freeable+0x208/0x2a7
kernel_init+0xe/0x104
ret_from_fork+0x2a/0x40
-> #0 (cpu_hotplug_lock.rw_sem){++++++}:
__lock_acquire+0x153c/0x1550
lock_acquire+0xc9/0x230
cpus_read_lock+0x4b/0x90
drain_all_stock.part.35+0x18/0x140
try_charge+0x3ab/0x6e0
mem_cgroup_try_charge+0x7f/0x2c0
shmem_getpage_gfp+0x25f/0x1050
shmem_fault+0x96/0x200
__do_fault+0x1e/0xa0
__handle_mm_fault+0x9c3/0xe00
handle_mm_fault+0x16e/0x380
__do_page_fault+0x24a/0x530
do_page_fault+0x30/0x80
page_fault+0x28/0x30
other info that might help us debug this:
Chain exists of:
cpu_hotplug_lock.rw_sem --> &type->i_mutex_dir_key#3 --> &mm->mmap_sem
Possible unsafe locking scenario:
CPU0 CPU1
---- ----
lock(&mm->mmap_sem);
lock(&type->i_mutex_dir_key#3);
lock(&mm->mmap_sem);
lock(cpu_hotplug_lock.rw_sem);
*** DEADLOCK ***
2 locks held by a.out/4771:
#0: (&mm->mmap_sem){++++++}, at: [<ffffffff8106eb35>] __do_page_fault+0x175/0x530
#1: (percpu_charge_mutex){+.+...}, at: [<ffffffff812b4c97>] try_charge+0x397/0x6e0
The problem is very similar to the one fixed by commit a459eeb7b852
("mm, page_alloc: do not depend on cpu hotplug locks inside the
allocator"). We are taking hotplug locks while we can be sitting on top
of basically arbitrary locks. This just calls for problems.
We can get rid of {get,put}_online_cpus, fortunately. We do not have to
be worried about races with memory hotplug because drain_local_stock,
which is called from both the WQ draining and the memory hotplug
contexts, is always operating on the local cpu stock with IRQs disabled.
The only thing to be careful about is that the target memcg doesn't
vanish while we are still in drain_all_stock so take a reference on it.
Link: http://lkml.kernel.org/r/20170913090023.28322-1-mhocko@kernel.org
Signed-off-by: Michal Hocko <mhocko@suse.com>
Reported-by: Artem Savkov <asavkov@redhat.com>
Tested-by: Artem Savkov <asavkov@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-10-03 16:14:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-03-22 14:40:50 -07:00
|
|
|
|
|
|
|
|
|
2009-12-15 16:47:08 -08:00
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
|
|
|
|
|
2011-06-15 15:08:45 -07:00
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-07-26 16:08:27 -07:00
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
2023-03-17 14:44:48 +01:00
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
|
|
|
|
|
mm: memcontrol: switch to rcu protection in drain_all_stock()
Commit 72f0184c8a00 ("mm, memcg: remove hotplug locking from try_charge")
introduced css_tryget()/css_put() calls in drain_all_stock(), which are
supposed to protect the target memory cgroup from being released during
the mem_cgroup_is_descendant() call.
However, it's not completely safe. In theory, memcg can go away between
reading stock->cached pointer and calling css_tryget().
This can happen if drain_all_stock() races with drain_local_stock()
performed on the remote cpu as a result of a work, scheduled by the
previous invocation of drain_all_stock().
The race is a bit theoretical and there are few chances to trigger it, but
the current code looks a bit confusing, so it makes sense to fix it
anyway. The code looks like as if css_tryget() and css_put() are used to
protect stocks drainage. It's not necessary because stocked pages are
holding references to the cached cgroup. And it obviously won't work for
works, scheduled on other cpus.
So, let's read the stock->cached pointer and evaluate the memory cgroup
inside a rcu read section, and get rid of css_tryget()/css_put() calls.
Link: http://lkml.kernel.org/r/20190802192241.3253165-1-guro@fb.com
Signed-off-by: Roman Gushchin <guro@fb.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-09-23 15:34:58 -07:00
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-07-26 16:08:27 -07:00
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
2023-03-17 14:44:48 +01:00
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
2011-07-26 16:08:27 -07:00
|
|
|
|
2009-12-15 16:47:08 -08:00
|
|
|
|
2022-03-22 14:40:50 -07:00
|
|
|
|
2011-08-09 11:56:26 +02:00
|
|
|
|
2009-12-15 16:47:08 -08:00
|
|
|
|
|
|
|
|
|
2021-04-29 22:56:29 -07:00
|
|
|
|
|
|
|
|
|
2025-05-06 15:55:31 -07:00
|
|
|
|
2025-05-14 11:41:57 -07:00
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
mm: memcontrol: fix cpuhotplug statistics flushing
Patch series "mm: memcontrol: switch to rstat", v3.
This series converts memcg stats tracking to the streamlined rstat
infrastructure provided by the cgroup core code. rstat is already used by
the CPU controller and the IO controller. This change is motivated by
recent accuracy problems in memcg's custom stats code, as well as the
benefits of sharing common infra with other controllers.
The current memcg implementation does batched tree aggregation on the
write side: local stat changes are cached in per-cpu counters, which are
then propagated upward in batches when a threshold (32 pages) is exceeded.
This is cheap, but the error introduced by the lazy upward propagation
adds up: 32 pages times CPUs times cgroups in the subtree. We've had
complaints from service owners that the stats do not reliably track and
react to allocation behavior as expected, sometimes swallowing the results
of entire test applications.
The original memcg stat implementation used to do tree aggregation
exclusively on the read side: local stats would only ever be tracked in
per-cpu counters, and a memory.stat read would iterate the entire subtree
and sum those counters up. This didn't keep up with the times:
- Cgroup trees are much bigger now. We switched to lazily-freed
cgroups, where deleted groups would hang around until their remaining
page cache has been reclaimed. This can result in large subtrees that
are expensive to walk, while most of the groups are idle and their
statistics don't change much anymore.
- Automated monitoring increased. With the proliferation of userspace
oom killing, proactive reclaim, and higher-resolution logging of
workload trends in general, top-level stat files are polled at least
once a second in many deployments.
- The lifetime of cgroups got shorter. Where most cgroup setups in the
past would have a few large policy-oriented cgroups for everything
running on the system, newer cgroup deployments tend to create one
group per application - which gets deleted again as the processes
exit. An aggregation scheme that doesn't retain child data inside the
parents loses event history of the subtree.
Rstat addresses all three of those concerns through intelligent,
persistent read-side aggregation. As statistics change at the local
level, rstat tracks - on a per-cpu basis - only those parts of a subtree
that have changes pending and require aggregation. The actual
aggregation occurs on the colder read side - which can now skip over
(potentially large) numbers of recently idle cgroups.
===
The test_kmem cgroup selftest is currently failing due to excessive
cumulative vmstat drift from 100 subgroups:
ok 1 test_kmem_basic
memory.current = 8810496
slab + anon + file + kernel_stack = 17074568
slab = 6101384
anon = 946176
file = 0
kernel_stack = 10027008
not ok 2 test_kmem_memcg_deletion
ok 3 test_kmem_proc_kpagecgroup
ok 4 test_kmem_kernel_stacks
ok 5 test_kmem_dead_cgroups
ok 6 test_percpu_basic
As you can see, memory.stat items far exceed memory.current. The kernel
stack alone is bigger than all of charged memory. That's because the
memory of the test has been uncharged from memory.current, but the
negative vmstat deltas are still sitting in the percpu caches.
The test at this time isn't even counting percpu, pagetables etc. yet,
which would further contribute to the error. The last patch in the series
updates the test to include them - as well as reduces the vmstat
tolerances in general to only expect page_counter batching.
With all patches applied, the (now more stringent) test succeeds:
ok 1 test_kmem_basic
ok 2 test_kmem_memcg_deletion
ok 3 test_kmem_proc_kpagecgroup
ok 4 test_kmem_kernel_stacks
ok 5 test_kmem_dead_cgroups
ok 6 test_percpu_basic
===
A kernel build test confirms that overhead is comparable. Two kernels are
built simultaneously in a nested tree with several idle siblings:
root - kernelbuild - one - two - three - four - build-a (defconfig, make -j16)
`- build-b (defconfig, make -j16)
`- idle-1
`- ...
`- idle-9
During the builds, kernelbuild/memory.stat is read once a second.
A perf diff shows that the changes in cycle distribution is
minimal. Top 10 kernel symbols:
0.09% +0.08% [kernel.kallsyms] [k] __mod_memcg_lruvec_state
0.00% +0.06% [kernel.kallsyms] [k] cgroup_rstat_updated
0.08% -0.05% [kernel.kallsyms] [k] __mod_memcg_state.part.0
0.16% -0.04% [kernel.kallsyms] [k] release_pages
0.00% +0.03% [kernel.kallsyms] [k] __count_memcg_events
0.01% +0.03% [kernel.kallsyms] [k] mem_cgroup_charge_statistics.constprop.0
0.10% -0.02% [kernel.kallsyms] [k] get_mem_cgroup_from_mm
0.05% -0.02% [kernel.kallsyms] [k] mem_cgroup_update_lru_size
0.57% +0.01% [kernel.kallsyms] [k] asm_exc_page_fault
===
The on-demand aggregated stats are now fully accurate:
$ grep -e nr_inactive_file /proc/vmstat | awk '{print($1,$2*4096)}'; \
grep -e inactive_file /sys/fs/cgroup/memory.stat
vanilla: patched:
nr_inactive_file 1574105088 nr_inactive_file 1027801088
inactive_file 1577410560 inactive_file 1027801088
===
This patch (of 8):
The memcg hotunplug callback erroneously flushes counts on the local CPU,
not the counts of the CPU going away; those counts will be lost.
Flush the CPU that is actually going away.
Also simplify the code a bit by using mod_memcg_state() and
count_memcg_events() instead of open-coding the upward flush - this is
comparable to how vmstat.c handles hotunplug flushing.
Link: https://lkml.kernel.org/r/20210209163304.77088-1-hannes@cmpxchg.org
Link: https://lkml.kernel.org/r/20210209163304.77088-2-hannes@cmpxchg.org
Fixes: a983b5ebee572 ("mm: memcontrol: fix excessive complexity in memory.stat reporting")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Roman Gushchin <guro@fb.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:11 -07:00
|
|
|
|
2016-11-03 15:49:59 +01:00
|
|
|
|
2009-12-15 16:47:08 -08:00
|
|
|
|
|
|
|
|
|
mm, memcg: reclaim more aggressively before high allocator throttling
Patch series "mm, memcg: reclaim harder before high throttling", v2.
This patch (of 2):
In Facebook production, we've seen cases where cgroups have been put into
allocator throttling even when they appear to have a lot of slack file
caches which should be trivially reclaimable.
Looking more closely, the problem is that we only try a single cgroup
reclaim walk for each return to usermode before calculating whether or not
we should throttle. This single attempt doesn't produce enough pressure
to shrink for cgroups with a rapidly growing amount of file caches prior
to entering allocator throttling.
As an example, we see that threads in an affected cgroup are stuck in
allocator throttling:
# for i in $(cat cgroup.threads); do
> grep over_high "/proc/$i/stack"
> done
[<0>] mem_cgroup_handle_over_high+0x10b/0x150
[<0>] mem_cgroup_handle_over_high+0x10b/0x150
[<0>] mem_cgroup_handle_over_high+0x10b/0x150
...however, there is no I/O pressure reported by PSI, despite a lot of
slack file pages:
# cat memory.pressure
some avg10=78.50 avg60=84.99 avg300=84.53 total=5702440903
full avg10=78.50 avg60=84.99 avg300=84.53 total=5702116959
# cat io.pressure
some avg10=0.00 avg60=0.00 avg300=0.00 total=78051391
full avg10=0.00 avg60=0.00 avg300=0.00 total=78049640
# grep _file memory.stat
inactive_file 1370939392
active_file 661635072
This patch changes the behaviour to retry reclaim either until the current
task goes below the 10ms grace period, or we are making no reclaim
progress at all. In the latter case, we enter reclaim throttling as
before.
To a user, there's no intuitive reason for the reclaim behaviour to differ
from hitting memory.high as part of a new allocation, as opposed to
hitting memory.high because someone lowered its value. As such this also
brings an added benefit: it unifies the reclaim behaviour between the two.
There's precedent for this behaviour: we already do reclaim retries when
writing to memory.{high,max}, in max reclaim, and in the page allocator
itself.
Signed-off-by: Chris Down <chris@chrisdown.name>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Roman Gushchin <guro@fb.com>
Link: http://lkml.kernel.org/r/cover.1594640214.git.chris@chrisdown.name
Link: http://lkml.kernel.org/r/a4e23b59e9ef499b575ae73a8120ee089b7d3373.1594640214.git.chris@chrisdown.name
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-06 23:21:54 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-01-14 15:21:29 -08:00
|
|
|
|
mm, memcg: reclaim more aggressively before high allocator throttling
Patch series "mm, memcg: reclaim harder before high throttling", v2.
This patch (of 2):
In Facebook production, we've seen cases where cgroups have been put into
allocator throttling even when they appear to have a lot of slack file
caches which should be trivially reclaimable.
Looking more closely, the problem is that we only try a single cgroup
reclaim walk for each return to usermode before calculating whether or not
we should throttle. This single attempt doesn't produce enough pressure
to shrink for cgroups with a rapidly growing amount of file caches prior
to entering allocator throttling.
As an example, we see that threads in an affected cgroup are stuck in
allocator throttling:
# for i in $(cat cgroup.threads); do
> grep over_high "/proc/$i/stack"
> done
[<0>] mem_cgroup_handle_over_high+0x10b/0x150
[<0>] mem_cgroup_handle_over_high+0x10b/0x150
[<0>] mem_cgroup_handle_over_high+0x10b/0x150
...however, there is no I/O pressure reported by PSI, despite a lot of
slack file pages:
# cat memory.pressure
some avg10=78.50 avg60=84.99 avg300=84.53 total=5702440903
full avg10=78.50 avg60=84.99 avg300=84.53 total=5702116959
# cat io.pressure
some avg10=0.00 avg60=0.00 avg300=0.00 total=78051391
full avg10=0.00 avg60=0.00 avg300=0.00 total=78049640
# grep _file memory.stat
inactive_file 1370939392
active_file 661635072
This patch changes the behaviour to retry reclaim either until the current
task goes below the 10ms grace period, or we are making no reclaim
progress at all. In the latter case, we enter reclaim throttling as
before.
To a user, there's no intuitive reason for the reclaim behaviour to differ
from hitting memory.high as part of a new allocation, as opposed to
hitting memory.high because someone lowered its value. As such this also
brings an added benefit: it unifies the reclaim behaviour between the two.
There's precedent for this behaviour: we already do reclaim retries when
writing to memory.{high,max}, in max reclaim, and in the page allocator
itself.
Signed-off-by: Chris Down <chris@chrisdown.name>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Roman Gushchin <guro@fb.com>
Link: http://lkml.kernel.org/r/cover.1594640214.git.chris@chrisdown.name
Link: http://lkml.kernel.org/r/a4e23b59e9ef499b575ae73a8120ee089b7d3373.1594640214.git.chris@chrisdown.name
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-06 23:21:54 -07:00
|
|
|
|
|
|
|
|
|
2016-01-14 15:21:29 -08:00
|
|
|
|
2020-08-06 23:22:15 -07:00
|
|
|
|
|
|
|
|
|
2020-06-01 21:49:49 -07:00
|
|
|
|
|
|
|
|
|
2016-01-14 15:21:29 -08:00
|
|
|
|
2020-08-06 23:22:15 -07:00
|
|
|
|
2018-04-10 16:29:45 -07:00
|
|
|
|
2020-08-06 23:22:15 -07:00
|
|
|
|
|
|
|
|
|
mm, memcg: reclaim more aggressively before high allocator throttling
Patch series "mm, memcg: reclaim harder before high throttling", v2.
This patch (of 2):
In Facebook production, we've seen cases where cgroups have been put into
allocator throttling even when they appear to have a lot of slack file
caches which should be trivially reclaimable.
Looking more closely, the problem is that we only try a single cgroup
reclaim walk for each return to usermode before calculating whether or not
we should throttle. This single attempt doesn't produce enough pressure
to shrink for cgroups with a rapidly growing amount of file caches prior
to entering allocator throttling.
As an example, we see that threads in an affected cgroup are stuck in
allocator throttling:
# for i in $(cat cgroup.threads); do
> grep over_high "/proc/$i/stack"
> done
[<0>] mem_cgroup_handle_over_high+0x10b/0x150
[<0>] mem_cgroup_handle_over_high+0x10b/0x150
[<0>] mem_cgroup_handle_over_high+0x10b/0x150
...however, there is no I/O pressure reported by PSI, despite a lot of
slack file pages:
# cat memory.pressure
some avg10=78.50 avg60=84.99 avg300=84.53 total=5702440903
full avg10=78.50 avg60=84.99 avg300=84.53 total=5702116959
# cat io.pressure
some avg10=0.00 avg60=0.00 avg300=0.00 total=78051391
full avg10=0.00 avg60=0.00 avg300=0.00 total=78049640
# grep _file memory.stat
inactive_file 1370939392
active_file 661635072
This patch changes the behaviour to retry reclaim either until the current
task goes below the 10ms grace period, or we are making no reclaim
progress at all. In the latter case, we enter reclaim throttling as
before.
To a user, there's no intuitive reason for the reclaim behaviour to differ
from hitting memory.high as part of a new allocation, as opposed to
hitting memory.high because someone lowered its value. As such this also
brings an added benefit: it unifies the reclaim behaviour between the two.
There's precedent for this behaviour: we already do reclaim retries when
writing to memory.{high,max}, in max reclaim, and in the page allocator
itself.
Signed-off-by: Chris Down <chris@chrisdown.name>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Roman Gushchin <guro@fb.com>
Link: http://lkml.kernel.org/r/cover.1594640214.git.chris@chrisdown.name
Link: http://lkml.kernel.org/r/a4e23b59e9ef499b575ae73a8120ee089b7d3373.1594640214.git.chris@chrisdown.name
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-06 23:21:54 -07:00
|
|
|
|
2022-07-14 06:49:18 +00:00
|
|
|
|
2024-01-03 08:48:37 -08:00
|
|
|
|
|
|
|
|
|
2020-08-06 23:22:15 -07:00
|
|
|
|
2020-04-06 20:03:30 -07:00
|
|
|
|
|
|
|
|
|
mm, memcg: reclaim more aggressively before high allocator throttling
Patch series "mm, memcg: reclaim harder before high throttling", v2.
This patch (of 2):
In Facebook production, we've seen cases where cgroups have been put into
allocator throttling even when they appear to have a lot of slack file
caches which should be trivially reclaimable.
Looking more closely, the problem is that we only try a single cgroup
reclaim walk for each return to usermode before calculating whether or not
we should throttle. This single attempt doesn't produce enough pressure
to shrink for cgroups with a rapidly growing amount of file caches prior
to entering allocator throttling.
As an example, we see that threads in an affected cgroup are stuck in
allocator throttling:
# for i in $(cat cgroup.threads); do
> grep over_high "/proc/$i/stack"
> done
[<0>] mem_cgroup_handle_over_high+0x10b/0x150
[<0>] mem_cgroup_handle_over_high+0x10b/0x150
[<0>] mem_cgroup_handle_over_high+0x10b/0x150
...however, there is no I/O pressure reported by PSI, despite a lot of
slack file pages:
# cat memory.pressure
some avg10=78.50 avg60=84.99 avg300=84.53 total=5702440903
full avg10=78.50 avg60=84.99 avg300=84.53 total=5702116959
# cat io.pressure
some avg10=0.00 avg60=0.00 avg300=0.00 total=78051391
full avg10=0.00 avg60=0.00 avg300=0.00 total=78049640
# grep _file memory.stat
inactive_file 1370939392
active_file 661635072
This patch changes the behaviour to retry reclaim either until the current
task goes below the 10ms grace period, or we are making no reclaim
progress at all. In the latter case, we enter reclaim throttling as
before.
To a user, there's no intuitive reason for the reclaim behaviour to differ
from hitting memory.high as part of a new allocation, as opposed to
hitting memory.high because someone lowered its value. As such this also
brings an added benefit: it unifies the reclaim behaviour between the two.
There's precedent for this behaviour: we already do reclaim retries when
writing to memory.{high,max}, in max reclaim, and in the page allocator
itself.
Signed-off-by: Chris Down <chris@chrisdown.name>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Roman Gushchin <guro@fb.com>
Link: http://lkml.kernel.org/r/cover.1594640214.git.chris@chrisdown.name
Link: http://lkml.kernel.org/r/a4e23b59e9ef499b575ae73a8120ee089b7d3373.1594640214.git.chris@chrisdown.name
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-06 23:21:54 -07:00
|
|
|
|
|
|
|
|
|
2016-01-14 15:21:29 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-01-31 16:16:45 -08:00
|
|
|
|
2016-01-14 15:21:29 -08:00
|
|
|
|
|
|
|
|
|
mm, memcg: throttle allocators when failing reclaim over memory.high
We're trying to use memory.high to limit workloads, but have found that
containment can frequently fail completely and cause OOM situations
outside of the cgroup. This happens especially with swap space -- either
when none is configured, or swap is full. These failures often also don't
have enough warning to allow one to react, whether for a human or for a
daemon monitoring PSI.
Here is output from a simple program showing how long it takes in usec
(column 2) to allocate a megabyte of anonymous memory (column 1) when a
cgroup is already beyond its memory high setting, and no swap is
available:
[root@ktst ~]# systemd-run -p MemoryHigh=100M -p MemorySwapMax=1 \
> --wait -t timeout 300 /root/mdf
[...]
95 1035
96 1038
97 1000
98 1036
99 1048
100 1590
101 1968
102 1776
103 1863
104 1757
105 1921
106 1893
107 1760
108 1748
109 1843
110 1716
111 1924
112 1776
113 1831
114 1766
115 1836
116 1588
117 1912
118 1802
119 1857
120 1731
[...]
[System OOM in 2-3 seconds]
The delay does go up extremely marginally past the 100MB memory.high
threshold, as now we spend time scanning before returning to usermode, but
it's nowhere near enough to contain growth. It also doesn't get worse the
more pages you have, since it only considers nr_pages.
The current situation goes against both the expectations of users of
memory.high, and our intentions as cgroup v2 developers. In
cgroup-v2.txt, we claim that we will throttle and only under "extreme
conditions" will memory.high protection be breached. Likewise, cgroup v2
users generally also expect that memory.high should throttle workloads as
they exceed their high threshold. However, as seen above, this isn't
always how it works in practice -- even on banal setups like those with no
swap, or where swap has become exhausted, we can end up with memory.high
being breached and us having no weapons left in our arsenal to combat
runaway growth with, since reclaim is futile.
It's also hard for system monitoring software or users to tell how bad the
situation is, as "high" events for the memcg may in some cases be benign,
and in others be catastrophic. The current status quo is that we fail
containment in a way that doesn't provide any advance warning that things
are about to go horribly wrong (for example, we are about to invoke the
kernel OOM killer).
This patch introduces explicit throttling when reclaim is failing to keep
memcg size contained at the memory.high setting. It does so by applying
an exponential delay curve derived from the memcg's overage compared to
memory.high. In the normal case where the memcg is either below or only
marginally over its memory.high setting, no throttling will be performed.
This composes well with system health monitoring and remediation, as these
allocator delays are factored into PSI's memory pressure calculations.
This both creates a mechanism system administrators or applications
consuming the PSI interface to trivially see that the memcg in question is
struggling and use that to make more reasonable decisions, and permits
them enough time to act. Either of these can act with significantly more
nuance than that we can provide using the system OOM killer.
This is a similar idea to memory.oom_control in cgroup v1 which would put
the cgroup to sleep if the threshold was violated, but it's also
significantly improved as it results in visible memory pressure, and also
doesn't schedule indefinitely, which previously made tracing and other
introspection difficult (ie. it's clamped at 2*HZ per allocation through
MEMCG_MAX_HIGH_DELAY_JIFFIES).
Contrast the previous results with a kernel with this patch:
[root@ktst ~]# systemd-run -p MemoryHigh=100M -p MemorySwapMax=1 \
> --wait -t timeout 300 /root/mdf
[...]
95 1002
96 1000
97 1002
98 1003
99 1000
100 1043
101 84724
102 330628
103 610511
104 1016265
105 1503969
106 2391692
107 2872061
108 3248003
109 4791904
110 5759832
111 6912509
112 8127818
113 9472203
114 12287622
115 12480079
116 14144008
117 15808029
118 16384500
119 16383242
120 16384979
[...]
As you can see, in the normal case, memory allocation takes around 1000
usec. However, as we exceed our memory.high, things start to increase
exponentially, but fairly leniently at first. Our first megabyte over
memory.high takes us 0.16 seconds, then the next is 0.46 seconds, then the
next is almost an entire second. This gets worse until we reach our
eventual 2*HZ clamp per batch, resulting in 16 seconds per megabyte.
However, this is still making forward progress, so permits tracing or
further analysis with programs like GDB.
We use an exponential curve for our delay penalty for a few reasons:
1. We run mem_cgroup_handle_over_high to potentially do reclaim after
we've already performed allocations, which means that temporarily
going over memory.high by a small amount may be perfectly legitimate,
even for compliant workloads. We don't want to unduly penalise such
cases.
2. An exponential curve (as opposed to a static or linear delay) allows
ramping up memory pressure stats more gradually, which can be useful
to work out that you have set memory.high too low, without destroying
application performance entirely.
This patch expands on earlier work by Johannes Weiner. Thanks!
[akpm@linux-foundation.org: fix max() warning]
[akpm@linux-foundation.org: fix __udivdi3 ref on 32-bit]
[akpm@linux-foundation.org: fix it even more]
[chris@chrisdown.name: fix 64-bit divide even more]
Link: http://lkml.kernel.org/r/20190723180700.GA29459@chrisdown.name
Signed-off-by: Chris Down <chris@chrisdown.name>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Nathan Chancellor <natechancellor@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-09-23 15:34:55 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-08-11 18:33:02 -07:00
|
|
|
|
mm, memcg: throttle allocators when failing reclaim over memory.high
We're trying to use memory.high to limit workloads, but have found that
containment can frequently fail completely and cause OOM situations
outside of the cgroup. This happens especially with swap space -- either
when none is configured, or swap is full. These failures often also don't
have enough warning to allow one to react, whether for a human or for a
daemon monitoring PSI.
Here is output from a simple program showing how long it takes in usec
(column 2) to allocate a megabyte of anonymous memory (column 1) when a
cgroup is already beyond its memory high setting, and no swap is
available:
[root@ktst ~]# systemd-run -p MemoryHigh=100M -p MemorySwapMax=1 \
> --wait -t timeout 300 /root/mdf
[...]
95 1035
96 1038
97 1000
98 1036
99 1048
100 1590
101 1968
102 1776
103 1863
104 1757
105 1921
106 1893
107 1760
108 1748
109 1843
110 1716
111 1924
112 1776
113 1831
114 1766
115 1836
116 1588
117 1912
118 1802
119 1857
120 1731
[...]
[System OOM in 2-3 seconds]
The delay does go up extremely marginally past the 100MB memory.high
threshold, as now we spend time scanning before returning to usermode, but
it's nowhere near enough to contain growth. It also doesn't get worse the
more pages you have, since it only considers nr_pages.
The current situation goes against both the expectations of users of
memory.high, and our intentions as cgroup v2 developers. In
cgroup-v2.txt, we claim that we will throttle and only under "extreme
conditions" will memory.high protection be breached. Likewise, cgroup v2
users generally also expect that memory.high should throttle workloads as
they exceed their high threshold. However, as seen above, this isn't
always how it works in practice -- even on banal setups like those with no
swap, or where swap has become exhausted, we can end up with memory.high
being breached and us having no weapons left in our arsenal to combat
runaway growth with, since reclaim is futile.
It's also hard for system monitoring software or users to tell how bad the
situation is, as "high" events for the memcg may in some cases be benign,
and in others be catastrophic. The current status quo is that we fail
containment in a way that doesn't provide any advance warning that things
are about to go horribly wrong (for example, we are about to invoke the
kernel OOM killer).
This patch introduces explicit throttling when reclaim is failing to keep
memcg size contained at the memory.high setting. It does so by applying
an exponential delay curve derived from the memcg's overage compared to
memory.high. In the normal case where the memcg is either below or only
marginally over its memory.high setting, no throttling will be performed.
This composes well with system health monitoring and remediation, as these
allocator delays are factored into PSI's memory pressure calculations.
This both creates a mechanism system administrators or applications
consuming the PSI interface to trivially see that the memcg in question is
struggling and use that to make more reasonable decisions, and permits
them enough time to act. Either of these can act with significantly more
nuance than that we can provide using the system OOM killer.
This is a similar idea to memory.oom_control in cgroup v1 which would put
the cgroup to sleep if the threshold was violated, but it's also
significantly improved as it results in visible memory pressure, and also
doesn't schedule indefinitely, which previously made tracing and other
introspection difficult (ie. it's clamped at 2*HZ per allocation through
MEMCG_MAX_HIGH_DELAY_JIFFIES).
Contrast the previous results with a kernel with this patch:
[root@ktst ~]# systemd-run -p MemoryHigh=100M -p MemorySwapMax=1 \
> --wait -t timeout 300 /root/mdf
[...]
95 1002
96 1000
97 1002
98 1003
99 1000
100 1043
101 84724
102 330628
103 610511
104 1016265
105 1503969
106 2391692
107 2872061
108 3248003
109 4791904
110 5759832
111 6912509
112 8127818
113 9472203
114 12287622
115 12480079
116 14144008
117 15808029
118 16384500
119 16383242
120 16384979
[...]
As you can see, in the normal case, memory allocation takes around 1000
usec. However, as we exceed our memory.high, things start to increase
exponentially, but fairly leniently at first. Our first megabyte over
memory.high takes us 0.16 seconds, then the next is 0.46 seconds, then the
next is almost an entire second. This gets worse until we reach our
eventual 2*HZ clamp per batch, resulting in 16 seconds per megabyte.
However, this is still making forward progress, so permits tracing or
further analysis with programs like GDB.
We use an exponential curve for our delay penalty for a few reasons:
1. We run mem_cgroup_handle_over_high to potentially do reclaim after
we've already performed allocations, which means that temporarily
going over memory.high by a small amount may be perfectly legitimate,
even for compliant workloads. We don't want to unduly penalise such
cases.
2. An exponential curve (as opposed to a static or linear delay) allows
ramping up memory pressure stats more gradually, which can be useful
to work out that you have set memory.high too low, without destroying
application performance entirely.
This patch expands on earlier work by Johannes Weiner. Thanks!
[akpm@linux-foundation.org: fix max() warning]
[akpm@linux-foundation.org: fix __udivdi3 ref on 32-bit]
[akpm@linux-foundation.org: fix it even more]
[chris@chrisdown.name: fix 64-bit divide even more]
Link: http://lkml.kernel.org/r/20190723180700.GA29459@chrisdown.name
Signed-off-by: Chris Down <chris@chrisdown.name>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Nathan Chancellor <natechancellor@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-09-23 15:34:55 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-06-01 21:49:42 -07:00
|
|
|
|
memcg: punt high overage reclaim to return-to-userland path
Currently, try_charge() tries to reclaim memory synchronously when the
high limit is breached; however, if the allocation doesn't have
__GFP_WAIT, synchronous reclaim is skipped. If a process performs only
speculative allocations, it can blow way past the high limit. This is
actually easily reproducible by simply doing "find /". slab/slub
allocator tries speculative allocations first, so as long as there's
memory which can be consumed without blocking, it can keep allocating
memory regardless of the high limit.
This patch makes try_charge() always punt the over-high reclaim to the
return-to-userland path. If try_charge() detects that high limit is
breached, it adds the overage to current->memcg_nr_pages_over_high and
schedules execution of mem_cgroup_handle_over_high() which performs
synchronous reclaim from the return-to-userland path.
As long as kernel doesn't have a run-away allocation spree, this should
provide enough protection while making kmemcg behave more consistently.
It also has the following benefits.
- All over-high reclaims can use GFP_KERNEL regardless of the specific
gfp mask in use, e.g. GFP_NOFS, when the limit was breached.
- It copes with prio inversion. Previously, a low-prio task with
small memory.high might perform over-high reclaim with a bunch of
locks held. If a higher prio task needed any of these locks, it
would have to wait until the low prio task finished reclaim and
released the locks. By handing over-high reclaim to the task exit
path this issue can be avoided.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Michal Hocko <mhocko@kernel.org>
Reviewed-by: Vladimir Davydov <vdavydov@parallels.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-05 18:46:11 -08:00
|
|
|
|
2020-06-01 21:49:42 -07:00
|
|
|
|
memcg: punt high overage reclaim to return-to-userland path
Currently, try_charge() tries to reclaim memory synchronously when the
high limit is breached; however, if the allocation doesn't have
__GFP_WAIT, synchronous reclaim is skipped. If a process performs only
speculative allocations, it can blow way past the high limit. This is
actually easily reproducible by simply doing "find /". slab/slub
allocator tries speculative allocations first, so as long as there's
memory which can be consumed without blocking, it can keep allocating
memory regardless of the high limit.
This patch makes try_charge() always punt the over-high reclaim to the
return-to-userland path. If try_charge() detects that high limit is
breached, it adds the overage to current->memcg_nr_pages_over_high and
schedules execution of mem_cgroup_handle_over_high() which performs
synchronous reclaim from the return-to-userland path.
As long as kernel doesn't have a run-away allocation spree, this should
provide enough protection while making kmemcg behave more consistently.
It also has the following benefits.
- All over-high reclaims can use GFP_KERNEL regardless of the specific
gfp mask in use, e.g. GFP_NOFS, when the limit was breached.
- It copes with prio inversion. Previously, a low-prio task with
small memory.high might perform over-high reclaim with a bunch of
locks held. If a higher prio task needed any of these locks, it
would have to wait until the low prio task finished reclaim and
released the locks. By handing over-high reclaim to the task exit
path this issue can be avoided.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Michal Hocko <mhocko@kernel.org>
Reviewed-by: Vladimir Davydov <vdavydov@parallels.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-05 18:46:11 -08:00
|
|
|
|
2020-06-01 21:49:42 -07:00
|
|
|
|
|
|
|
|
|
2020-03-21 18:22:23 -07:00
|
|
|
|
2020-06-01 21:49:42 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, memcg: do not high throttle allocators based on wraparound
If a cgroup violates its memory.high constraints, we may end up unduly
penalising it. For example, for the following hierarchy:
A: max high, 20 usage
A/B: 9 high, 10 usage
A/C: max high, 10 usage
We would end up doing the following calculation below when calculating
high delay for A/B:
A/B: 10 - 9 = 1...
A: 20 - PAGE_COUNTER_MAX = 21, so set max_overage to 21.
This gets worse with higher disparities in usage in the parent.
I have no idea how this disappeared from the final version of the patch,
but it is certainly Not Good(tm). This wasn't obvious in testing because,
for a simple cgroup hierarchy with only one child, the result is usually
roughly the same. It's only in more complex hierarchies that things go
really awry (although still, the effects are limited to a maximum of 2
seconds in schedule_timeout_killable at a maximum).
[chris@chrisdown.name: changelog]
Fixes: e26733e0d0ec ("mm, memcg: throttle allocators based on ancestral memory.high")
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Chris Down <chris@chrisdown.name>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: <stable@vger.kernel.org> [5.4.x]
Link: http://lkml.kernel.org/r/20200331152424.GA1019937@chrisdown.name
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-10 14:32:19 -07:00
|
|
|
|
2020-06-01 21:49:42 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-03-21 18:22:23 -07:00
|
|
|
|
2020-06-01 21:49:42 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-03-21 18:22:23 -07:00
|
|
|
|
2020-06-01 21:49:42 -07:00
|
|
|
|
|
|
|
|
|
2020-06-01 21:49:49 -07:00
|
|
|
|
2020-06-01 21:49:42 -07:00
|
|
|
|
2020-03-21 18:22:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-06-01 21:49:42 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-06-01 21:49:52 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-06-01 21:49:42 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-03-21 18:22:23 -07:00
|
|
|
|
|
|
|
|
|
mm, memcg: throttle allocators when failing reclaim over memory.high
We're trying to use memory.high to limit workloads, but have found that
containment can frequently fail completely and cause OOM situations
outside of the cgroup. This happens especially with swap space -- either
when none is configured, or swap is full. These failures often also don't
have enough warning to allow one to react, whether for a human or for a
daemon monitoring PSI.
Here is output from a simple program showing how long it takes in usec
(column 2) to allocate a megabyte of anonymous memory (column 1) when a
cgroup is already beyond its memory high setting, and no swap is
available:
[root@ktst ~]# systemd-run -p MemoryHigh=100M -p MemorySwapMax=1 \
> --wait -t timeout 300 /root/mdf
[...]
95 1035
96 1038
97 1000
98 1036
99 1048
100 1590
101 1968
102 1776
103 1863
104 1757
105 1921
106 1893
107 1760
108 1748
109 1843
110 1716
111 1924
112 1776
113 1831
114 1766
115 1836
116 1588
117 1912
118 1802
119 1857
120 1731
[...]
[System OOM in 2-3 seconds]
The delay does go up extremely marginally past the 100MB memory.high
threshold, as now we spend time scanning before returning to usermode, but
it's nowhere near enough to contain growth. It also doesn't get worse the
more pages you have, since it only considers nr_pages.
The current situation goes against both the expectations of users of
memory.high, and our intentions as cgroup v2 developers. In
cgroup-v2.txt, we claim that we will throttle and only under "extreme
conditions" will memory.high protection be breached. Likewise, cgroup v2
users generally also expect that memory.high should throttle workloads as
they exceed their high threshold. However, as seen above, this isn't
always how it works in practice -- even on banal setups like those with no
swap, or where swap has become exhausted, we can end up with memory.high
being breached and us having no weapons left in our arsenal to combat
runaway growth with, since reclaim is futile.
It's also hard for system monitoring software or users to tell how bad the
situation is, as "high" events for the memcg may in some cases be benign,
and in others be catastrophic. The current status quo is that we fail
containment in a way that doesn't provide any advance warning that things
are about to go horribly wrong (for example, we are about to invoke the
kernel OOM killer).
This patch introduces explicit throttling when reclaim is failing to keep
memcg size contained at the memory.high setting. It does so by applying
an exponential delay curve derived from the memcg's overage compared to
memory.high. In the normal case where the memcg is either below or only
marginally over its memory.high setting, no throttling will be performed.
This composes well with system health monitoring and remediation, as these
allocator delays are factored into PSI's memory pressure calculations.
This both creates a mechanism system administrators or applications
consuming the PSI interface to trivially see that the memcg in question is
struggling and use that to make more reasonable decisions, and permits
them enough time to act. Either of these can act with significantly more
nuance than that we can provide using the system OOM killer.
This is a similar idea to memory.oom_control in cgroup v1 which would put
the cgroup to sleep if the threshold was violated, but it's also
significantly improved as it results in visible memory pressure, and also
doesn't schedule indefinitely, which previously made tracing and other
introspection difficult (ie. it's clamped at 2*HZ per allocation through
MEMCG_MAX_HIGH_DELAY_JIFFIES).
Contrast the previous results with a kernel with this patch:
[root@ktst ~]# systemd-run -p MemoryHigh=100M -p MemorySwapMax=1 \
> --wait -t timeout 300 /root/mdf
[...]
95 1002
96 1000
97 1002
98 1003
99 1000
100 1043
101 84724
102 330628
103 610511
104 1016265
105 1503969
106 2391692
107 2872061
108 3248003
109 4791904
110 5759832
111 6912509
112 8127818
113 9472203
114 12287622
115 12480079
116 14144008
117 15808029
118 16384500
119 16383242
120 16384979
[...]
As you can see, in the normal case, memory allocation takes around 1000
usec. However, as we exceed our memory.high, things start to increase
exponentially, but fairly leniently at first. Our first megabyte over
memory.high takes us 0.16 seconds, then the next is 0.46 seconds, then the
next is almost an entire second. This gets worse until we reach our
eventual 2*HZ clamp per batch, resulting in 16 seconds per megabyte.
However, this is still making forward progress, so permits tracing or
further analysis with programs like GDB.
We use an exponential curve for our delay penalty for a few reasons:
1. We run mem_cgroup_handle_over_high to potentially do reclaim after
we've already performed allocations, which means that temporarily
going over memory.high by a small amount may be perfectly legitimate,
even for compliant workloads. We don't want to unduly penalise such
cases.
2. An exponential curve (as opposed to a static or linear delay) allows
ramping up memory pressure stats more gradually, which can be useful
to work out that you have set memory.high too low, without destroying
application performance entirely.
This patch expands on earlier work by Johannes Weiner. Thanks!
[akpm@linux-foundation.org: fix max() warning]
[akpm@linux-foundation.org: fix __udivdi3 ref on 32-bit]
[akpm@linux-foundation.org: fix it even more]
[chris@chrisdown.name: fix 64-bit divide even more]
Link: http://lkml.kernel.org/r/20190723180700.GA29459@chrisdown.name
Signed-off-by: Chris Down <chris@chrisdown.name>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Nathan Chancellor <natechancellor@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-09-23 15:34:55 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-03-21 18:22:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, memcg: throttle allocators when failing reclaim over memory.high
We're trying to use memory.high to limit workloads, but have found that
containment can frequently fail completely and cause OOM situations
outside of the cgroup. This happens especially with swap space -- either
when none is configured, or swap is full. These failures often also don't
have enough warning to allow one to react, whether for a human or for a
daemon monitoring PSI.
Here is output from a simple program showing how long it takes in usec
(column 2) to allocate a megabyte of anonymous memory (column 1) when a
cgroup is already beyond its memory high setting, and no swap is
available:
[root@ktst ~]# systemd-run -p MemoryHigh=100M -p MemorySwapMax=1 \
> --wait -t timeout 300 /root/mdf
[...]
95 1035
96 1038
97 1000
98 1036
99 1048
100 1590
101 1968
102 1776
103 1863
104 1757
105 1921
106 1893
107 1760
108 1748
109 1843
110 1716
111 1924
112 1776
113 1831
114 1766
115 1836
116 1588
117 1912
118 1802
119 1857
120 1731
[...]
[System OOM in 2-3 seconds]
The delay does go up extremely marginally past the 100MB memory.high
threshold, as now we spend time scanning before returning to usermode, but
it's nowhere near enough to contain growth. It also doesn't get worse the
more pages you have, since it only considers nr_pages.
The current situation goes against both the expectations of users of
memory.high, and our intentions as cgroup v2 developers. In
cgroup-v2.txt, we claim that we will throttle and only under "extreme
conditions" will memory.high protection be breached. Likewise, cgroup v2
users generally also expect that memory.high should throttle workloads as
they exceed their high threshold. However, as seen above, this isn't
always how it works in practice -- even on banal setups like those with no
swap, or where swap has become exhausted, we can end up with memory.high
being breached and us having no weapons left in our arsenal to combat
runaway growth with, since reclaim is futile.
It's also hard for system monitoring software or users to tell how bad the
situation is, as "high" events for the memcg may in some cases be benign,
and in others be catastrophic. The current status quo is that we fail
containment in a way that doesn't provide any advance warning that things
are about to go horribly wrong (for example, we are about to invoke the
kernel OOM killer).
This patch introduces explicit throttling when reclaim is failing to keep
memcg size contained at the memory.high setting. It does so by applying
an exponential delay curve derived from the memcg's overage compared to
memory.high. In the normal case where the memcg is either below or only
marginally over its memory.high setting, no throttling will be performed.
This composes well with system health monitoring and remediation, as these
allocator delays are factored into PSI's memory pressure calculations.
This both creates a mechanism system administrators or applications
consuming the PSI interface to trivially see that the memcg in question is
struggling and use that to make more reasonable decisions, and permits
them enough time to act. Either of these can act with significantly more
nuance than that we can provide using the system OOM killer.
This is a similar idea to memory.oom_control in cgroup v1 which would put
the cgroup to sleep if the threshold was violated, but it's also
significantly improved as it results in visible memory pressure, and also
doesn't schedule indefinitely, which previously made tracing and other
introspection difficult (ie. it's clamped at 2*HZ per allocation through
MEMCG_MAX_HIGH_DELAY_JIFFIES).
Contrast the previous results with a kernel with this patch:
[root@ktst ~]# systemd-run -p MemoryHigh=100M -p MemorySwapMax=1 \
> --wait -t timeout 300 /root/mdf
[...]
95 1002
96 1000
97 1002
98 1003
99 1000
100 1043
101 84724
102 330628
103 610511
104 1016265
105 1503969
106 2391692
107 2872061
108 3248003
109 4791904
110 5759832
111 6912509
112 8127818
113 9472203
114 12287622
115 12480079
116 14144008
117 15808029
118 16384500
119 16383242
120 16384979
[...]
As you can see, in the normal case, memory allocation takes around 1000
usec. However, as we exceed our memory.high, things start to increase
exponentially, but fairly leniently at first. Our first megabyte over
memory.high takes us 0.16 seconds, then the next is 0.46 seconds, then the
next is almost an entire second. This gets worse until we reach our
eventual 2*HZ clamp per batch, resulting in 16 seconds per megabyte.
However, this is still making forward progress, so permits tracing or
further analysis with programs like GDB.
We use an exponential curve for our delay penalty for a few reasons:
1. We run mem_cgroup_handle_over_high to potentially do reclaim after
we've already performed allocations, which means that temporarily
going over memory.high by a small amount may be perfectly legitimate,
even for compliant workloads. We don't want to unduly penalise such
cases.
2. An exponential curve (as opposed to a static or linear delay) allows
ramping up memory pressure stats more gradually, which can be useful
to work out that you have set memory.high too low, without destroying
application performance entirely.
This patch expands on earlier work by Johannes Weiner. Thanks!
[akpm@linux-foundation.org: fix max() warning]
[akpm@linux-foundation.org: fix __udivdi3 ref on 32-bit]
[akpm@linux-foundation.org: fix it even more]
[chris@chrisdown.name: fix 64-bit divide even more]
Link: http://lkml.kernel.org/r/20190723180700.GA29459@chrisdown.name
Signed-off-by: Chris Down <chris@chrisdown.name>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Nathan Chancellor <natechancellor@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-09-23 15:34:55 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-06-01 21:49:45 -07:00
|
|
|
|
2020-03-21 18:22:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcontrol: don't throttle dying tasks on memory.high
While investigating hosts with high cgroup memory pressures, Tejun
found culprit zombie tasks that had were holding on to a lot of
memory, had SIGKILL pending, but were stuck in memory.high reclaim.
In the past, we used to always force-charge allocations from tasks
that were exiting in order to accelerate them dying and freeing up
their rss. This changed for memory.max in a4ebf1b6ca1e ("memcg:
prohibit unconditional exceeding the limit of dying tasks"); it noted
that this can cause (userspace inducable) containment failures, so it
added a mandatory reclaim and OOM kill cycle before forcing charges.
At the time, memory.high enforcement was handled in the userspace
return path, which isn't reached by dying tasks, and so memory.high
was still never enforced by dying tasks.
When c9afe31ec443 ("memcg: synchronously enforce memory.high for large
overcharges") added synchronous reclaim for memory.high, it added
unconditional memory.high enforcement for dying tasks as well. The
callstack shows that this path is where the zombie is stuck in.
We need to accelerate dying tasks getting past memory.high, but we
cannot do it quite the same way as we do for memory.max: memory.max is
enforced strictly, and tasks aren't allowed to move past it without
FIRST reclaiming and OOM killing if necessary. This ensures very small
levels of excess. With memory.high, though, enforcement happens lazily
after the charge, and OOM killing is never triggered. A lot of
concurrent threads could have pushed, or could actively be pushing,
the cgroup into excess. The dying task will enter reclaim on every
allocation attempt, with little hope of restoring balance.
To fix this, skip synchronous memory.high enforcement on dying tasks
altogether again. Update memory.high path documentation while at it.
[hannes@cmpxchg.org: also handle tasks are being killed during the reclaim]
Link: https://lkml.kernel.org/r/20240111192807.GA424308@cmpxchg.org
Link: https://lkml.kernel.org/r/20240111132902.389862-1-hannes@cmpxchg.org
Fixes: c9afe31ec443 ("memcg: synchronously enforce memory.high for large overcharges")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reported-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Yosry Ahmed <yosryahmed@google.com>
Acked-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Dan Schatzberg <schatzberg.dan@gmail.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-01-11 08:29:02 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-03-21 18:22:23 -07:00
|
|
|
|
2025-08-13 16:57:55 +02:00
|
|
|
|
2020-03-21 18:22:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, memcg: reclaim more aggressively before high allocator throttling
Patch series "mm, memcg: reclaim harder before high throttling", v2.
This patch (of 2):
In Facebook production, we've seen cases where cgroups have been put into
allocator throttling even when they appear to have a lot of slack file
caches which should be trivially reclaimable.
Looking more closely, the problem is that we only try a single cgroup
reclaim walk for each return to usermode before calculating whether or not
we should throttle. This single attempt doesn't produce enough pressure
to shrink for cgroups with a rapidly growing amount of file caches prior
to entering allocator throttling.
As an example, we see that threads in an affected cgroup are stuck in
allocator throttling:
# for i in $(cat cgroup.threads); do
> grep over_high "/proc/$i/stack"
> done
[<0>] mem_cgroup_handle_over_high+0x10b/0x150
[<0>] mem_cgroup_handle_over_high+0x10b/0x150
[<0>] mem_cgroup_handle_over_high+0x10b/0x150
...however, there is no I/O pressure reported by PSI, despite a lot of
slack file pages:
# cat memory.pressure
some avg10=78.50 avg60=84.99 avg300=84.53 total=5702440903
full avg10=78.50 avg60=84.99 avg300=84.53 total=5702116959
# cat io.pressure
some avg10=0.00 avg60=0.00 avg300=0.00 total=78051391
full avg10=0.00 avg60=0.00 avg300=0.00 total=78049640
# grep _file memory.stat
inactive_file 1370939392
active_file 661635072
This patch changes the behaviour to retry reclaim either until the current
task goes below the 10ms grace period, or we are making no reclaim
progress at all. In the latter case, we enter reclaim throttling as
before.
To a user, there's no intuitive reason for the reclaim behaviour to differ
from hitting memory.high as part of a new allocation, as opposed to
hitting memory.high because someone lowered its value. As such this also
brings an added benefit: it unifies the reclaim behaviour between the two.
There's precedent for this behaviour: we already do reclaim retries when
writing to memory.{high,max}, in max reclaim, and in the page allocator
itself.
Signed-off-by: Chris Down <chris@chrisdown.name>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Roman Gushchin <guro@fb.com>
Link: http://lkml.kernel.org/r/cover.1594640214.git.chris@chrisdown.name
Link: http://lkml.kernel.org/r/a4e23b59e9ef499b575ae73a8120ee089b7d3373.1594640214.git.chris@chrisdown.name
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-06 23:21:54 -07:00
|
|
|
|
2020-03-21 18:22:23 -07:00
|
|
|
|
2020-08-06 23:21:58 -07:00
|
|
|
|
2020-03-21 18:22:23 -07:00
|
|
|
|
mm, memcg: reclaim more aggressively before high allocator throttling
Patch series "mm, memcg: reclaim harder before high throttling", v2.
This patch (of 2):
In Facebook production, we've seen cases where cgroups have been put into
allocator throttling even when they appear to have a lot of slack file
caches which should be trivially reclaimable.
Looking more closely, the problem is that we only try a single cgroup
reclaim walk for each return to usermode before calculating whether or not
we should throttle. This single attempt doesn't produce enough pressure
to shrink for cgroups with a rapidly growing amount of file caches prior
to entering allocator throttling.
As an example, we see that threads in an affected cgroup are stuck in
allocator throttling:
# for i in $(cat cgroup.threads); do
> grep over_high "/proc/$i/stack"
> done
[<0>] mem_cgroup_handle_over_high+0x10b/0x150
[<0>] mem_cgroup_handle_over_high+0x10b/0x150
[<0>] mem_cgroup_handle_over_high+0x10b/0x150
...however, there is no I/O pressure reported by PSI, despite a lot of
slack file pages:
# cat memory.pressure
some avg10=78.50 avg60=84.99 avg300=84.53 total=5702440903
full avg10=78.50 avg60=84.99 avg300=84.53 total=5702116959
# cat io.pressure
some avg10=0.00 avg60=0.00 avg300=0.00 total=78051391
full avg10=0.00 avg60=0.00 avg300=0.00 total=78049640
# grep _file memory.stat
inactive_file 1370939392
active_file 661635072
This patch changes the behaviour to retry reclaim either until the current
task goes below the 10ms grace period, or we are making no reclaim
progress at all. In the latter case, we enter reclaim throttling as
before.
To a user, there's no intuitive reason for the reclaim behaviour to differ
from hitting memory.high as part of a new allocation, as opposed to
hitting memory.high because someone lowered its value. As such this also
brings an added benefit: it unifies the reclaim behaviour between the two.
There's precedent for this behaviour: we already do reclaim retries when
writing to memory.{high,max}, in max reclaim, and in the page allocator
itself.
Signed-off-by: Chris Down <chris@chrisdown.name>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Roman Gushchin <guro@fb.com>
Link: http://lkml.kernel.org/r/cover.1594640214.git.chris@chrisdown.name
Link: http://lkml.kernel.org/r/a4e23b59e9ef499b575ae73a8120ee089b7d3373.1594640214.git.chris@chrisdown.name
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-06 23:21:54 -07:00
|
|
|
|
2020-03-21 18:22:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, memcg: reclaim more aggressively before high allocator throttling
Patch series "mm, memcg: reclaim harder before high throttling", v2.
This patch (of 2):
In Facebook production, we've seen cases where cgroups have been put into
allocator throttling even when they appear to have a lot of slack file
caches which should be trivially reclaimable.
Looking more closely, the problem is that we only try a single cgroup
reclaim walk for each return to usermode before calculating whether or not
we should throttle. This single attempt doesn't produce enough pressure
to shrink for cgroups with a rapidly growing amount of file caches prior
to entering allocator throttling.
As an example, we see that threads in an affected cgroup are stuck in
allocator throttling:
# for i in $(cat cgroup.threads); do
> grep over_high "/proc/$i/stack"
> done
[<0>] mem_cgroup_handle_over_high+0x10b/0x150
[<0>] mem_cgroup_handle_over_high+0x10b/0x150
[<0>] mem_cgroup_handle_over_high+0x10b/0x150
...however, there is no I/O pressure reported by PSI, despite a lot of
slack file pages:
# cat memory.pressure
some avg10=78.50 avg60=84.99 avg300=84.53 total=5702440903
full avg10=78.50 avg60=84.99 avg300=84.53 total=5702116959
# cat io.pressure
some avg10=0.00 avg60=0.00 avg300=0.00 total=78051391
full avg10=0.00 avg60=0.00 avg300=0.00 total=78049640
# grep _file memory.stat
inactive_file 1370939392
active_file 661635072
This patch changes the behaviour to retry reclaim either until the current
task goes below the 10ms grace period, or we are making no reclaim
progress at all. In the latter case, we enter reclaim throttling as
before.
To a user, there's no intuitive reason for the reclaim behaviour to differ
from hitting memory.high as part of a new allocation, as opposed to
hitting memory.high because someone lowered its value. As such this also
brings an added benefit: it unifies the reclaim behaviour between the two.
There's precedent for this behaviour: we already do reclaim retries when
writing to memory.{high,max}, in max reclaim, and in the page allocator
itself.
Signed-off-by: Chris Down <chris@chrisdown.name>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Roman Gushchin <guro@fb.com>
Link: http://lkml.kernel.org/r/cover.1594640214.git.chris@chrisdown.name
Link: http://lkml.kernel.org/r/a4e23b59e9ef499b575ae73a8120ee089b7d3373.1594640214.git.chris@chrisdown.name
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-06 23:21:54 -07:00
|
|
|
|
mm: memcontrol: don't throttle dying tasks on memory.high
While investigating hosts with high cgroup memory pressures, Tejun
found culprit zombie tasks that had were holding on to a lot of
memory, had SIGKILL pending, but were stuck in memory.high reclaim.
In the past, we used to always force-charge allocations from tasks
that were exiting in order to accelerate them dying and freeing up
their rss. This changed for memory.max in a4ebf1b6ca1e ("memcg:
prohibit unconditional exceeding the limit of dying tasks"); it noted
that this can cause (userspace inducable) containment failures, so it
added a mandatory reclaim and OOM kill cycle before forcing charges.
At the time, memory.high enforcement was handled in the userspace
return path, which isn't reached by dying tasks, and so memory.high
was still never enforced by dying tasks.
When c9afe31ec443 ("memcg: synchronously enforce memory.high for large
overcharges") added synchronous reclaim for memory.high, it added
unconditional memory.high enforcement for dying tasks as well. The
callstack shows that this path is where the zombie is stuck in.
We need to accelerate dying tasks getting past memory.high, but we
cannot do it quite the same way as we do for memory.max: memory.max is
enforced strictly, and tasks aren't allowed to move past it without
FIRST reclaiming and OOM killing if necessary. This ensures very small
levels of excess. With memory.high, though, enforcement happens lazily
after the charge, and OOM killing is never triggered. A lot of
concurrent threads could have pushed, or could actively be pushing,
the cgroup into excess. The dying task will enter reclaim on every
allocation attempt, with little hope of restoring balance.
To fix this, skip synchronous memory.high enforcement on dying tasks
altogether again. Update memory.high path documentation while at it.
[hannes@cmpxchg.org: also handle tasks are being killed during the reclaim]
Link: https://lkml.kernel.org/r/20240111192807.GA424308@cmpxchg.org
Link: https://lkml.kernel.org/r/20240111132902.389862-1-hannes@cmpxchg.org
Fixes: c9afe31ec443 ("memcg: synchronously enforce memory.high for large overcharges")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reported-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Yosry Ahmed <yosryahmed@google.com>
Acked-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Dan Schatzberg <schatzberg.dan@gmail.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-01-11 08:29:02 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, memcg: reclaim more aggressively before high allocator throttling
Patch series "mm, memcg: reclaim harder before high throttling", v2.
This patch (of 2):
In Facebook production, we've seen cases where cgroups have been put into
allocator throttling even when they appear to have a lot of slack file
caches which should be trivially reclaimable.
Looking more closely, the problem is that we only try a single cgroup
reclaim walk for each return to usermode before calculating whether or not
we should throttle. This single attempt doesn't produce enough pressure
to shrink for cgroups with a rapidly growing amount of file caches prior
to entering allocator throttling.
As an example, we see that threads in an affected cgroup are stuck in
allocator throttling:
# for i in $(cat cgroup.threads); do
> grep over_high "/proc/$i/stack"
> done
[<0>] mem_cgroup_handle_over_high+0x10b/0x150
[<0>] mem_cgroup_handle_over_high+0x10b/0x150
[<0>] mem_cgroup_handle_over_high+0x10b/0x150
...however, there is no I/O pressure reported by PSI, despite a lot of
slack file pages:
# cat memory.pressure
some avg10=78.50 avg60=84.99 avg300=84.53 total=5702440903
full avg10=78.50 avg60=84.99 avg300=84.53 total=5702116959
# cat io.pressure
some avg10=0.00 avg60=0.00 avg300=0.00 total=78051391
full avg10=0.00 avg60=0.00 avg300=0.00 total=78049640
# grep _file memory.stat
inactive_file 1370939392
active_file 661635072
This patch changes the behaviour to retry reclaim either until the current
task goes below the 10ms grace period, or we are making no reclaim
progress at all. In the latter case, we enter reclaim throttling as
before.
To a user, there's no intuitive reason for the reclaim behaviour to differ
from hitting memory.high as part of a new allocation, as opposed to
hitting memory.high because someone lowered its value. As such this also
brings an added benefit: it unifies the reclaim behaviour between the two.
There's precedent for this behaviour: we already do reclaim retries when
writing to memory.{high,max}, in max reclaim, and in the page allocator
itself.
Signed-off-by: Chris Down <chris@chrisdown.name>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Roman Gushchin <guro@fb.com>
Link: http://lkml.kernel.org/r/cover.1594640214.git.chris@chrisdown.name
Link: http://lkml.kernel.org/r/a4e23b59e9ef499b575ae73a8120ee089b7d3373.1594640214.git.chris@chrisdown.name
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-06 23:21:54 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-09-14 11:21:39 -04:00
|
|
|
|
mm, memcg: reclaim more aggressively before high allocator throttling
Patch series "mm, memcg: reclaim harder before high throttling", v2.
This patch (of 2):
In Facebook production, we've seen cases where cgroups have been put into
allocator throttling even when they appear to have a lot of slack file
caches which should be trivially reclaimable.
Looking more closely, the problem is that we only try a single cgroup
reclaim walk for each return to usermode before calculating whether or not
we should throttle. This single attempt doesn't produce enough pressure
to shrink for cgroups with a rapidly growing amount of file caches prior
to entering allocator throttling.
As an example, we see that threads in an affected cgroup are stuck in
allocator throttling:
# for i in $(cat cgroup.threads); do
> grep over_high "/proc/$i/stack"
> done
[<0>] mem_cgroup_handle_over_high+0x10b/0x150
[<0>] mem_cgroup_handle_over_high+0x10b/0x150
[<0>] mem_cgroup_handle_over_high+0x10b/0x150
...however, there is no I/O pressure reported by PSI, despite a lot of
slack file pages:
# cat memory.pressure
some avg10=78.50 avg60=84.99 avg300=84.53 total=5702440903
full avg10=78.50 avg60=84.99 avg300=84.53 total=5702116959
# cat io.pressure
some avg10=0.00 avg60=0.00 avg300=0.00 total=78051391
full avg10=0.00 avg60=0.00 avg300=0.00 total=78049640
# grep _file memory.stat
inactive_file 1370939392
active_file 661635072
This patch changes the behaviour to retry reclaim either until the current
task goes below the 10ms grace period, or we are making no reclaim
progress at all. In the latter case, we enter reclaim throttling as
before.
To a user, there's no intuitive reason for the reclaim behaviour to differ
from hitting memory.high as part of a new allocation, as opposed to
hitting memory.high because someone lowered its value. As such this also
brings an added benefit: it unifies the reclaim behaviour between the two.
There's precedent for this behaviour: we already do reclaim retries when
writing to memory.{high,max}, in max reclaim, and in the page allocator
itself.
Signed-off-by: Chris Down <chris@chrisdown.name>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Roman Gushchin <guro@fb.com>
Link: http://lkml.kernel.org/r/cover.1594640214.git.chris@chrisdown.name
Link: http://lkml.kernel.org/r/a4e23b59e9ef499b575ae73a8120ee089b7d3373.1594640214.git.chris@chrisdown.name
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-06 23:21:54 -07:00
|
|
|
|
2020-03-21 18:22:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-06-01 21:49:42 -07:00
|
|
|
|
|
|
|
|
|
mm, memcg: throttle allocators when failing reclaim over memory.high
We're trying to use memory.high to limit workloads, but have found that
containment can frequently fail completely and cause OOM situations
outside of the cgroup. This happens especially with swap space -- either
when none is configured, or swap is full. These failures often also don't
have enough warning to allow one to react, whether for a human or for a
daemon monitoring PSI.
Here is output from a simple program showing how long it takes in usec
(column 2) to allocate a megabyte of anonymous memory (column 1) when a
cgroup is already beyond its memory high setting, and no swap is
available:
[root@ktst ~]# systemd-run -p MemoryHigh=100M -p MemorySwapMax=1 \
> --wait -t timeout 300 /root/mdf
[...]
95 1035
96 1038
97 1000
98 1036
99 1048
100 1590
101 1968
102 1776
103 1863
104 1757
105 1921
106 1893
107 1760
108 1748
109 1843
110 1716
111 1924
112 1776
113 1831
114 1766
115 1836
116 1588
117 1912
118 1802
119 1857
120 1731
[...]
[System OOM in 2-3 seconds]
The delay does go up extremely marginally past the 100MB memory.high
threshold, as now we spend time scanning before returning to usermode, but
it's nowhere near enough to contain growth. It also doesn't get worse the
more pages you have, since it only considers nr_pages.
The current situation goes against both the expectations of users of
memory.high, and our intentions as cgroup v2 developers. In
cgroup-v2.txt, we claim that we will throttle and only under "extreme
conditions" will memory.high protection be breached. Likewise, cgroup v2
users generally also expect that memory.high should throttle workloads as
they exceed their high threshold. However, as seen above, this isn't
always how it works in practice -- even on banal setups like those with no
swap, or where swap has become exhausted, we can end up with memory.high
being breached and us having no weapons left in our arsenal to combat
runaway growth with, since reclaim is futile.
It's also hard for system monitoring software or users to tell how bad the
situation is, as "high" events for the memcg may in some cases be benign,
and in others be catastrophic. The current status quo is that we fail
containment in a way that doesn't provide any advance warning that things
are about to go horribly wrong (for example, we are about to invoke the
kernel OOM killer).
This patch introduces explicit throttling when reclaim is failing to keep
memcg size contained at the memory.high setting. It does so by applying
an exponential delay curve derived from the memcg's overage compared to
memory.high. In the normal case where the memcg is either below or only
marginally over its memory.high setting, no throttling will be performed.
This composes well with system health monitoring and remediation, as these
allocator delays are factored into PSI's memory pressure calculations.
This both creates a mechanism system administrators or applications
consuming the PSI interface to trivially see that the memcg in question is
struggling and use that to make more reasonable decisions, and permits
them enough time to act. Either of these can act with significantly more
nuance than that we can provide using the system OOM killer.
This is a similar idea to memory.oom_control in cgroup v1 which would put
the cgroup to sleep if the threshold was violated, but it's also
significantly improved as it results in visible memory pressure, and also
doesn't schedule indefinitely, which previously made tracing and other
introspection difficult (ie. it's clamped at 2*HZ per allocation through
MEMCG_MAX_HIGH_DELAY_JIFFIES).
Contrast the previous results with a kernel with this patch:
[root@ktst ~]# systemd-run -p MemoryHigh=100M -p MemorySwapMax=1 \
> --wait -t timeout 300 /root/mdf
[...]
95 1002
96 1000
97 1002
98 1003
99 1000
100 1043
101 84724
102 330628
103 610511
104 1016265
105 1503969
106 2391692
107 2872061
108 3248003
109 4791904
110 5759832
111 6912509
112 8127818
113 9472203
114 12287622
115 12480079
116 14144008
117 15808029
118 16384500
119 16383242
120 16384979
[...]
As you can see, in the normal case, memory allocation takes around 1000
usec. However, as we exceed our memory.high, things start to increase
exponentially, but fairly leniently at first. Our first megabyte over
memory.high takes us 0.16 seconds, then the next is 0.46 seconds, then the
next is almost an entire second. This gets worse until we reach our
eventual 2*HZ clamp per batch, resulting in 16 seconds per megabyte.
However, this is still making forward progress, so permits tracing or
further analysis with programs like GDB.
We use an exponential curve for our delay penalty for a few reasons:
1. We run mem_cgroup_handle_over_high to potentially do reclaim after
we've already performed allocations, which means that temporarily
going over memory.high by a small amount may be perfectly legitimate,
even for compliant workloads. We don't want to unduly penalise such
cases.
2. An exponential curve (as opposed to a static or linear delay) allows
ramping up memory pressure stats more gradually, which can be useful
to work out that you have set memory.high too low, without destroying
application performance entirely.
This patch expands on earlier work by Johannes Weiner. Thanks!
[akpm@linux-foundation.org: fix max() warning]
[akpm@linux-foundation.org: fix __udivdi3 ref on 32-bit]
[akpm@linux-foundation.org: fix it even more]
[chris@chrisdown.name: fix 64-bit divide even more]
Link: http://lkml.kernel.org/r/20190723180700.GA29459@chrisdown.name
Signed-off-by: Chris Down <chris@chrisdown.name>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Nathan Chancellor <natechancellor@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-09-23 15:34:55 -07:00
|
|
|
|
2020-06-01 21:49:52 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-06-01 21:49:45 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, memcg: throttle allocators when failing reclaim over memory.high
We're trying to use memory.high to limit workloads, but have found that
containment can frequently fail completely and cause OOM situations
outside of the cgroup. This happens especially with swap space -- either
when none is configured, or swap is full. These failures often also don't
have enough warning to allow one to react, whether for a human or for a
daemon monitoring PSI.
Here is output from a simple program showing how long it takes in usec
(column 2) to allocate a megabyte of anonymous memory (column 1) when a
cgroup is already beyond its memory high setting, and no swap is
available:
[root@ktst ~]# systemd-run -p MemoryHigh=100M -p MemorySwapMax=1 \
> --wait -t timeout 300 /root/mdf
[...]
95 1035
96 1038
97 1000
98 1036
99 1048
100 1590
101 1968
102 1776
103 1863
104 1757
105 1921
106 1893
107 1760
108 1748
109 1843
110 1716
111 1924
112 1776
113 1831
114 1766
115 1836
116 1588
117 1912
118 1802
119 1857
120 1731
[...]
[System OOM in 2-3 seconds]
The delay does go up extremely marginally past the 100MB memory.high
threshold, as now we spend time scanning before returning to usermode, but
it's nowhere near enough to contain growth. It also doesn't get worse the
more pages you have, since it only considers nr_pages.
The current situation goes against both the expectations of users of
memory.high, and our intentions as cgroup v2 developers. In
cgroup-v2.txt, we claim that we will throttle and only under "extreme
conditions" will memory.high protection be breached. Likewise, cgroup v2
users generally also expect that memory.high should throttle workloads as
they exceed their high threshold. However, as seen above, this isn't
always how it works in practice -- even on banal setups like those with no
swap, or where swap has become exhausted, we can end up with memory.high
being breached and us having no weapons left in our arsenal to combat
runaway growth with, since reclaim is futile.
It's also hard for system monitoring software or users to tell how bad the
situation is, as "high" events for the memcg may in some cases be benign,
and in others be catastrophic. The current status quo is that we fail
containment in a way that doesn't provide any advance warning that things
are about to go horribly wrong (for example, we are about to invoke the
kernel OOM killer).
This patch introduces explicit throttling when reclaim is failing to keep
memcg size contained at the memory.high setting. It does so by applying
an exponential delay curve derived from the memcg's overage compared to
memory.high. In the normal case where the memcg is either below or only
marginally over its memory.high setting, no throttling will be performed.
This composes well with system health monitoring and remediation, as these
allocator delays are factored into PSI's memory pressure calculations.
This both creates a mechanism system administrators or applications
consuming the PSI interface to trivially see that the memcg in question is
struggling and use that to make more reasonable decisions, and permits
them enough time to act. Either of these can act with significantly more
nuance than that we can provide using the system OOM killer.
This is a similar idea to memory.oom_control in cgroup v1 which would put
the cgroup to sleep if the threshold was violated, but it's also
significantly improved as it results in visible memory pressure, and also
doesn't schedule indefinitely, which previously made tracing and other
introspection difficult (ie. it's clamped at 2*HZ per allocation through
MEMCG_MAX_HIGH_DELAY_JIFFIES).
Contrast the previous results with a kernel with this patch:
[root@ktst ~]# systemd-run -p MemoryHigh=100M -p MemorySwapMax=1 \
> --wait -t timeout 300 /root/mdf
[...]
95 1002
96 1000
97 1002
98 1003
99 1000
100 1043
101 84724
102 330628
103 610511
104 1016265
105 1503969
106 2391692
107 2872061
108 3248003
109 4791904
110 5759832
111 6912509
112 8127818
113 9472203
114 12287622
115 12480079
116 14144008
117 15808029
118 16384500
119 16383242
120 16384979
[...]
As you can see, in the normal case, memory allocation takes around 1000
usec. However, as we exceed our memory.high, things start to increase
exponentially, but fairly leniently at first. Our first megabyte over
memory.high takes us 0.16 seconds, then the next is 0.46 seconds, then the
next is almost an entire second. This gets worse until we reach our
eventual 2*HZ clamp per batch, resulting in 16 seconds per megabyte.
However, this is still making forward progress, so permits tracing or
further analysis with programs like GDB.
We use an exponential curve for our delay penalty for a few reasons:
1. We run mem_cgroup_handle_over_high to potentially do reclaim after
we've already performed allocations, which means that temporarily
going over memory.high by a small amount may be perfectly legitimate,
even for compliant workloads. We don't want to unduly penalise such
cases.
2. An exponential curve (as opposed to a static or linear delay) allows
ramping up memory pressure stats more gradually, which can be useful
to work out that you have set memory.high too low, without destroying
application performance entirely.
This patch expands on earlier work by Johannes Weiner. Thanks!
[akpm@linux-foundation.org: fix max() warning]
[akpm@linux-foundation.org: fix __udivdi3 ref on 32-bit]
[akpm@linux-foundation.org: fix it even more]
[chris@chrisdown.name: fix 64-bit divide even more]
Link: http://lkml.kernel.org/r/20190723180700.GA29459@chrisdown.name
Signed-off-by: Chris Down <chris@chrisdown.name>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Nathan Chancellor <natechancellor@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-09-23 15:34:55 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, memcg: reclaim more aggressively before high allocator throttling
Patch series "mm, memcg: reclaim harder before high throttling", v2.
This patch (of 2):
In Facebook production, we've seen cases where cgroups have been put into
allocator throttling even when they appear to have a lot of slack file
caches which should be trivially reclaimable.
Looking more closely, the problem is that we only try a single cgroup
reclaim walk for each return to usermode before calculating whether or not
we should throttle. This single attempt doesn't produce enough pressure
to shrink for cgroups with a rapidly growing amount of file caches prior
to entering allocator throttling.
As an example, we see that threads in an affected cgroup are stuck in
allocator throttling:
# for i in $(cat cgroup.threads); do
> grep over_high "/proc/$i/stack"
> done
[<0>] mem_cgroup_handle_over_high+0x10b/0x150
[<0>] mem_cgroup_handle_over_high+0x10b/0x150
[<0>] mem_cgroup_handle_over_high+0x10b/0x150
...however, there is no I/O pressure reported by PSI, despite a lot of
slack file pages:
# cat memory.pressure
some avg10=78.50 avg60=84.99 avg300=84.53 total=5702440903
full avg10=78.50 avg60=84.99 avg300=84.53 total=5702116959
# cat io.pressure
some avg10=0.00 avg60=0.00 avg300=0.00 total=78051391
full avg10=0.00 avg60=0.00 avg300=0.00 total=78049640
# grep _file memory.stat
inactive_file 1370939392
active_file 661635072
This patch changes the behaviour to retry reclaim either until the current
task goes below the 10ms grace period, or we are making no reclaim
progress at all. In the latter case, we enter reclaim throttling as
before.
To a user, there's no intuitive reason for the reclaim behaviour to differ
from hitting memory.high as part of a new allocation, as opposed to
hitting memory.high because someone lowered its value. As such this also
brings an added benefit: it unifies the reclaim behaviour between the two.
There's precedent for this behaviour: we already do reclaim retries when
writing to memory.{high,max}, in max reclaim, and in the page allocator
itself.
Signed-off-by: Chris Down <chris@chrisdown.name>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Roman Gushchin <guro@fb.com>
Link: http://lkml.kernel.org/r/cover.1594640214.git.chris@chrisdown.name
Link: http://lkml.kernel.org/r/a4e23b59e9ef499b575ae73a8120ee089b7d3373.1594640214.git.chris@chrisdown.name
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-06 23:21:54 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, memcg: throttle allocators when failing reclaim over memory.high
We're trying to use memory.high to limit workloads, but have found that
containment can frequently fail completely and cause OOM situations
outside of the cgroup. This happens especially with swap space -- either
when none is configured, or swap is full. These failures often also don't
have enough warning to allow one to react, whether for a human or for a
daemon monitoring PSI.
Here is output from a simple program showing how long it takes in usec
(column 2) to allocate a megabyte of anonymous memory (column 1) when a
cgroup is already beyond its memory high setting, and no swap is
available:
[root@ktst ~]# systemd-run -p MemoryHigh=100M -p MemorySwapMax=1 \
> --wait -t timeout 300 /root/mdf
[...]
95 1035
96 1038
97 1000
98 1036
99 1048
100 1590
101 1968
102 1776
103 1863
104 1757
105 1921
106 1893
107 1760
108 1748
109 1843
110 1716
111 1924
112 1776
113 1831
114 1766
115 1836
116 1588
117 1912
118 1802
119 1857
120 1731
[...]
[System OOM in 2-3 seconds]
The delay does go up extremely marginally past the 100MB memory.high
threshold, as now we spend time scanning before returning to usermode, but
it's nowhere near enough to contain growth. It also doesn't get worse the
more pages you have, since it only considers nr_pages.
The current situation goes against both the expectations of users of
memory.high, and our intentions as cgroup v2 developers. In
cgroup-v2.txt, we claim that we will throttle and only under "extreme
conditions" will memory.high protection be breached. Likewise, cgroup v2
users generally also expect that memory.high should throttle workloads as
they exceed their high threshold. However, as seen above, this isn't
always how it works in practice -- even on banal setups like those with no
swap, or where swap has become exhausted, we can end up with memory.high
being breached and us having no weapons left in our arsenal to combat
runaway growth with, since reclaim is futile.
It's also hard for system monitoring software or users to tell how bad the
situation is, as "high" events for the memcg may in some cases be benign,
and in others be catastrophic. The current status quo is that we fail
containment in a way that doesn't provide any advance warning that things
are about to go horribly wrong (for example, we are about to invoke the
kernel OOM killer).
This patch introduces explicit throttling when reclaim is failing to keep
memcg size contained at the memory.high setting. It does so by applying
an exponential delay curve derived from the memcg's overage compared to
memory.high. In the normal case where the memcg is either below or only
marginally over its memory.high setting, no throttling will be performed.
This composes well with system health monitoring and remediation, as these
allocator delays are factored into PSI's memory pressure calculations.
This both creates a mechanism system administrators or applications
consuming the PSI interface to trivially see that the memcg in question is
struggling and use that to make more reasonable decisions, and permits
them enough time to act. Either of these can act with significantly more
nuance than that we can provide using the system OOM killer.
This is a similar idea to memory.oom_control in cgroup v1 which would put
the cgroup to sleep if the threshold was violated, but it's also
significantly improved as it results in visible memory pressure, and also
doesn't schedule indefinitely, which previously made tracing and other
introspection difficult (ie. it's clamped at 2*HZ per allocation through
MEMCG_MAX_HIGH_DELAY_JIFFIES).
Contrast the previous results with a kernel with this patch:
[root@ktst ~]# systemd-run -p MemoryHigh=100M -p MemorySwapMax=1 \
> --wait -t timeout 300 /root/mdf
[...]
95 1002
96 1000
97 1002
98 1003
99 1000
100 1043
101 84724
102 330628
103 610511
104 1016265
105 1503969
106 2391692
107 2872061
108 3248003
109 4791904
110 5759832
111 6912509
112 8127818
113 9472203
114 12287622
115 12480079
116 14144008
117 15808029
118 16384500
119 16383242
120 16384979
[...]
As you can see, in the normal case, memory allocation takes around 1000
usec. However, as we exceed our memory.high, things start to increase
exponentially, but fairly leniently at first. Our first megabyte over
memory.high takes us 0.16 seconds, then the next is 0.46 seconds, then the
next is almost an entire second. This gets worse until we reach our
eventual 2*HZ clamp per batch, resulting in 16 seconds per megabyte.
However, this is still making forward progress, so permits tracing or
further analysis with programs like GDB.
We use an exponential curve for our delay penalty for a few reasons:
1. We run mem_cgroup_handle_over_high to potentially do reclaim after
we've already performed allocations, which means that temporarily
going over memory.high by a small amount may be perfectly legitimate,
even for compliant workloads. We don't want to unduly penalise such
cases.
2. An exponential curve (as opposed to a static or linear delay) allows
ramping up memory pressure stats more gradually, which can be useful
to work out that you have set memory.high too low, without destroying
application performance entirely.
This patch expands on earlier work by Johannes Weiner. Thanks!
[akpm@linux-foundation.org: fix max() warning]
[akpm@linux-foundation.org: fix __udivdi3 ref on 32-bit]
[akpm@linux-foundation.org: fix it even more]
[chris@chrisdown.name: fix 64-bit divide even more]
Link: http://lkml.kernel.org/r/20190723180700.GA29459@chrisdown.name
Signed-off-by: Chris Down <chris@chrisdown.name>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Nathan Chancellor <natechancellor@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-09-23 15:34:55 -07:00
|
|
|
|
mm: memcontrol: don't throttle dying tasks on memory.high
While investigating hosts with high cgroup memory pressures, Tejun
found culprit zombie tasks that had were holding on to a lot of
memory, had SIGKILL pending, but were stuck in memory.high reclaim.
In the past, we used to always force-charge allocations from tasks
that were exiting in order to accelerate them dying and freeing up
their rss. This changed for memory.max in a4ebf1b6ca1e ("memcg:
prohibit unconditional exceeding the limit of dying tasks"); it noted
that this can cause (userspace inducable) containment failures, so it
added a mandatory reclaim and OOM kill cycle before forcing charges.
At the time, memory.high enforcement was handled in the userspace
return path, which isn't reached by dying tasks, and so memory.high
was still never enforced by dying tasks.
When c9afe31ec443 ("memcg: synchronously enforce memory.high for large
overcharges") added synchronous reclaim for memory.high, it added
unconditional memory.high enforcement for dying tasks as well. The
callstack shows that this path is where the zombie is stuck in.
We need to accelerate dying tasks getting past memory.high, but we
cannot do it quite the same way as we do for memory.max: memory.max is
enforced strictly, and tasks aren't allowed to move past it without
FIRST reclaiming and OOM killing if necessary. This ensures very small
levels of excess. With memory.high, though, enforcement happens lazily
after the charge, and OOM killing is never triggered. A lot of
concurrent threads could have pushed, or could actively be pushing,
the cgroup into excess. The dying task will enter reclaim on every
allocation attempt, with little hope of restoring balance.
To fix this, skip synchronous memory.high enforcement on dying tasks
altogether again. Update memory.high path documentation while at it.
[hannes@cmpxchg.org: also handle tasks are being killed during the reclaim]
Link: https://lkml.kernel.org/r/20240111192807.GA424308@cmpxchg.org
Link: https://lkml.kernel.org/r/20240111132902.389862-1-hannes@cmpxchg.org
Fixes: c9afe31ec443 ("memcg: synchronously enforce memory.high for large overcharges")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reported-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Yosry Ahmed <yosryahmed@google.com>
Acked-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Dan Schatzberg <schatzberg.dan@gmail.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-01-11 08:29:02 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, memcg: throttle allocators when failing reclaim over memory.high
We're trying to use memory.high to limit workloads, but have found that
containment can frequently fail completely and cause OOM situations
outside of the cgroup. This happens especially with swap space -- either
when none is configured, or swap is full. These failures often also don't
have enough warning to allow one to react, whether for a human or for a
daemon monitoring PSI.
Here is output from a simple program showing how long it takes in usec
(column 2) to allocate a megabyte of anonymous memory (column 1) when a
cgroup is already beyond its memory high setting, and no swap is
available:
[root@ktst ~]# systemd-run -p MemoryHigh=100M -p MemorySwapMax=1 \
> --wait -t timeout 300 /root/mdf
[...]
95 1035
96 1038
97 1000
98 1036
99 1048
100 1590
101 1968
102 1776
103 1863
104 1757
105 1921
106 1893
107 1760
108 1748
109 1843
110 1716
111 1924
112 1776
113 1831
114 1766
115 1836
116 1588
117 1912
118 1802
119 1857
120 1731
[...]
[System OOM in 2-3 seconds]
The delay does go up extremely marginally past the 100MB memory.high
threshold, as now we spend time scanning before returning to usermode, but
it's nowhere near enough to contain growth. It also doesn't get worse the
more pages you have, since it only considers nr_pages.
The current situation goes against both the expectations of users of
memory.high, and our intentions as cgroup v2 developers. In
cgroup-v2.txt, we claim that we will throttle and only under "extreme
conditions" will memory.high protection be breached. Likewise, cgroup v2
users generally also expect that memory.high should throttle workloads as
they exceed their high threshold. However, as seen above, this isn't
always how it works in practice -- even on banal setups like those with no
swap, or where swap has become exhausted, we can end up with memory.high
being breached and us having no weapons left in our arsenal to combat
runaway growth with, since reclaim is futile.
It's also hard for system monitoring software or users to tell how bad the
situation is, as "high" events for the memcg may in some cases be benign,
and in others be catastrophic. The current status quo is that we fail
containment in a way that doesn't provide any advance warning that things
are about to go horribly wrong (for example, we are about to invoke the
kernel OOM killer).
This patch introduces explicit throttling when reclaim is failing to keep
memcg size contained at the memory.high setting. It does so by applying
an exponential delay curve derived from the memcg's overage compared to
memory.high. In the normal case where the memcg is either below or only
marginally over its memory.high setting, no throttling will be performed.
This composes well with system health monitoring and remediation, as these
allocator delays are factored into PSI's memory pressure calculations.
This both creates a mechanism system administrators or applications
consuming the PSI interface to trivially see that the memcg in question is
struggling and use that to make more reasonable decisions, and permits
them enough time to act. Either of these can act with significantly more
nuance than that we can provide using the system OOM killer.
This is a similar idea to memory.oom_control in cgroup v1 which would put
the cgroup to sleep if the threshold was violated, but it's also
significantly improved as it results in visible memory pressure, and also
doesn't schedule indefinitely, which previously made tracing and other
introspection difficult (ie. it's clamped at 2*HZ per allocation through
MEMCG_MAX_HIGH_DELAY_JIFFIES).
Contrast the previous results with a kernel with this patch:
[root@ktst ~]# systemd-run -p MemoryHigh=100M -p MemorySwapMax=1 \
> --wait -t timeout 300 /root/mdf
[...]
95 1002
96 1000
97 1002
98 1003
99 1000
100 1043
101 84724
102 330628
103 610511
104 1016265
105 1503969
106 2391692
107 2872061
108 3248003
109 4791904
110 5759832
111 6912509
112 8127818
113 9472203
114 12287622
115 12480079
116 14144008
117 15808029
118 16384500
119 16383242
120 16384979
[...]
As you can see, in the normal case, memory allocation takes around 1000
usec. However, as we exceed our memory.high, things start to increase
exponentially, but fairly leniently at first. Our first megabyte over
memory.high takes us 0.16 seconds, then the next is 0.46 seconds, then the
next is almost an entire second. This gets worse until we reach our
eventual 2*HZ clamp per batch, resulting in 16 seconds per megabyte.
However, this is still making forward progress, so permits tracing or
further analysis with programs like GDB.
We use an exponential curve for our delay penalty for a few reasons:
1. We run mem_cgroup_handle_over_high to potentially do reclaim after
we've already performed allocations, which means that temporarily
going over memory.high by a small amount may be perfectly legitimate,
even for compliant workloads. We don't want to unduly penalise such
cases.
2. An exponential curve (as opposed to a static or linear delay) allows
ramping up memory pressure stats more gradually, which can be useful
to work out that you have set memory.high too low, without destroying
application performance entirely.
This patch expands on earlier work by Johannes Weiner. Thanks!
[akpm@linux-foundation.org: fix max() warning]
[akpm@linux-foundation.org: fix __udivdi3 ref on 32-bit]
[akpm@linux-foundation.org: fix it even more]
[chris@chrisdown.name: fix 64-bit divide even more]
Link: http://lkml.kernel.org/r/20190723180700.GA29459@chrisdown.name
Signed-off-by: Chris Down <chris@chrisdown.name>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Nathan Chancellor <natechancellor@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-09-23 15:34:55 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
memcg: punt high overage reclaim to return-to-userland path
Currently, try_charge() tries to reclaim memory synchronously when the
high limit is breached; however, if the allocation doesn't have
__GFP_WAIT, synchronous reclaim is skipped. If a process performs only
speculative allocations, it can blow way past the high limit. This is
actually easily reproducible by simply doing "find /". slab/slub
allocator tries speculative allocations first, so as long as there's
memory which can be consumed without blocking, it can keep allocating
memory regardless of the high limit.
This patch makes try_charge() always punt the over-high reclaim to the
return-to-userland path. If try_charge() detects that high limit is
breached, it adds the overage to current->memcg_nr_pages_over_high and
schedules execution of mem_cgroup_handle_over_high() which performs
synchronous reclaim from the return-to-userland path.
As long as kernel doesn't have a run-away allocation spree, this should
provide enough protection while making kmemcg behave more consistently.
It also has the following benefits.
- All over-high reclaims can use GFP_KERNEL regardless of the specific
gfp mask in use, e.g. GFP_NOFS, when the limit was breached.
- It copes with prio inversion. Previously, a low-prio task with
small memory.high might perform over-high reclaim with a bunch of
locks held. If a higher prio task needed any of these locks, it
would have to wait until the low prio task finished reclaim and
released the locks. By handing over-high reclaim to the task exit
path this issue can be avoided.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Michal Hocko <mhocko@kernel.org>
Reviewed-by: Vladimir Davydov <vdavydov@parallels.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-05 18:46:11 -08:00
|
|
|
|
|
|
|
|
|
2025-01-23 23:38:58 -05:00
|
|
|
|
|
|
|
|
|
2008-02-07 00:13:53 -08:00
|
|
|
|
2018-01-31 16:16:45 -08:00
|
|
|
|
2020-08-06 23:21:58 -07:00
|
|
|
|
2014-08-06 16:05:42 -07:00
|
|
|
|
mm: memcontrol: lockless page counters
Memory is internally accounted in bytes, using spinlock-protected 64-bit
counters, even though the smallest accounting delta is a page. The
counter interface is also convoluted and does too many things.
Introduce a new lockless word-sized page counter API, then change all
memory accounting over to it. The translation from and to bytes then only
happens when interfacing with userspace.
The removed locking overhead is noticable when scaling beyond the per-cpu
charge caches - on a 4-socket machine with 144-threads, the following test
shows the performance differences of 288 memcgs concurrently running a
page fault benchmark:
vanilla:
18631648.500498 task-clock (msec) # 140.643 CPUs utilized ( +- 0.33% )
1,380,638 context-switches # 0.074 K/sec ( +- 0.75% )
24,390 cpu-migrations # 0.001 K/sec ( +- 8.44% )
1,843,305,768 page-faults # 0.099 M/sec ( +- 0.00% )
50,134,994,088,218 cycles # 2.691 GHz ( +- 0.33% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
8,049,712,224,651 instructions # 0.16 insns per cycle ( +- 0.04% )
1,586,970,584,979 branches # 85.176 M/sec ( +- 0.05% )
1,724,989,949 branch-misses # 0.11% of all branches ( +- 0.48% )
132.474343877 seconds time elapsed ( +- 0.21% )
lockless:
12195979.037525 task-clock (msec) # 133.480 CPUs utilized ( +- 0.18% )
832,850 context-switches # 0.068 K/sec ( +- 0.54% )
15,624 cpu-migrations # 0.001 K/sec ( +- 10.17% )
1,843,304,774 page-faults # 0.151 M/sec ( +- 0.00% )
32,811,216,801,141 cycles # 2.690 GHz ( +- 0.18% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
9,999,265,091,727 instructions # 0.30 insns per cycle ( +- 0.10% )
2,076,759,325,203 branches # 170.282 M/sec ( +- 0.12% )
1,656,917,214 branch-misses # 0.08% of all branches ( +- 0.55% )
91.369330729 seconds time elapsed ( +- 0.45% )
On top of improved scalability, this also gets rid of the icky long long
types in the very heart of memcg, which is great for 32 bit and also makes
the code a lot more readable.
Notable differences between the old and new API:
- res_counter_charge() and res_counter_charge_nofail() become
page_counter_try_charge() and page_counter_charge() resp. to match
the more common kernel naming scheme of try_do()/do()
- res_counter_uncharge_until() is only ever used to cancel a local
counter and never to uncharge bigger segments of a hierarchy, so
it's replaced by the simpler page_counter_cancel()
- res_counter_set_limit() is replaced by page_counter_limit(), which
expects its callers to serialize against themselves
- res_counter_memparse_write_strategy() is replaced by
page_counter_limit(), which rounds down to the nearest page size -
rather than up. This is more reasonable for explicitely requested
hard upper limits.
- to keep charging light-weight, page_counter_try_charge() charges
speculatively, only to roll back if the result exceeds the limit.
Because of this, a failing bigger charge can temporarily lock out
smaller charges that would otherwise succeed. The error is bounded
to the difference between the smallest and the biggest possible
charge size, so for memcg, this means that a failing THP charge can
send base page charges into reclaim upto 2MB (4MB) before the limit
would have been reached. This should be acceptable.
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE and memparse]
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE, memparse, strncmp, and PAGE_SIZE]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Vladimir Davydov <vdavydov@parallels.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-12-10 15:42:31 -08:00
|
|
|
|
2014-08-06 16:05:42 -07:00
|
|
|
|
2021-11-05 13:38:09 -07:00
|
|
|
|
2022-07-14 06:49:18 +00:00
|
|
|
|
2014-10-09 15:28:56 -07:00
|
|
|
|
2022-07-01 20:35:21 -07:00
|
|
|
|
2020-08-06 23:22:15 -07:00
|
|
|
|
2025-09-22 15:02:03 -07:00
|
|
|
|
2009-01-07 18:08:08 -08:00
|
|
|
|
2014-08-06 16:05:42 -07:00
|
|
|
|
2025-05-06 15:55:30 -07:00
|
|
|
|
memcg: ratify and consolidate over-charge handling
try_charge() is the main charging logic of memcg. When it hits the limit
but either can't fail the allocation due to __GFP_NOFAIL or the task is
likely to free memory very soon, being OOM killed, has SIGKILL pending or
exiting, it "bypasses" the charge to the root memcg and returns -EINTR.
While this is one approach which can be taken for these situations, it has
several issues.
* It unnecessarily lies about the reality. The number itself doesn't
go over the limit but the actual usage does. memcg is either forced
to or actively chooses to go over the limit because that is the
right behavior under the circumstances, which is completely fine,
but, if at all avoidable, it shouldn't be misrepresenting what's
happening by sneaking the charges into the root memcg.
* Despite trying, we already do over-charge. kmemcg can't deal with
switching over to the root memcg by the point try_charge() returns
-EINTR, so it open-codes over-charing.
* It complicates the callers. Each try_charge() user has to handle
the weird -EINTR exception. memcg_charge_kmem() does the manual
over-charging. mem_cgroup_do_precharge() performs unnecessary
uncharging of root memcg, which BTW is inconsistent with what
memcg_charge_kmem() does but not broken as [un]charging are noops on
root memcg. mem_cgroup_try_charge() needs to switch the returned
cgroup to the root one.
The reality is that in memcg there are cases where we are forced and/or
willing to go over the limit. Each such case needs to be scrutinized and
justified but there definitely are situations where that is the right
thing to do. We alredy do this but with a superficial and inconsistent
disguise which leads to unnecessary complications.
This patch updates try_charge() so that it over-charges and returns 0 when
deemed necessary. -EINTR return is removed along with all special case
handling in the callers.
While at it, remove the local variable @ret, which was initialized to zero
and never changed, along with done: label which just returned the always
zero @ret.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Vladimir Davydov <vdavydov@parallels.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-05 18:46:17 -08:00
|
|
|
|
2008-02-07 00:13:53 -08:00
|
|
|
|
2025-09-22 15:02:03 -07:00
|
|
|
|
2025-02-21 18:44:25 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-01-14 15:21:23 -08:00
|
|
|
|
2015-11-05 18:50:26 -08:00
|
|
|
|
|
|
|
|
|
2014-08-06 16:05:42 -07:00
|
|
|
|
2016-01-14 15:21:23 -08:00
|
|
|
|
mm: memcontrol: lockless page counters
Memory is internally accounted in bytes, using spinlock-protected 64-bit
counters, even though the smallest accounting delta is a page. The
counter interface is also convoluted and does too many things.
Introduce a new lockless word-sized page counter API, then change all
memory accounting over to it. The translation from and to bytes then only
happens when interfacing with userspace.
The removed locking overhead is noticable when scaling beyond the per-cpu
charge caches - on a 4-socket machine with 144-threads, the following test
shows the performance differences of 288 memcgs concurrently running a
page fault benchmark:
vanilla:
18631648.500498 task-clock (msec) # 140.643 CPUs utilized ( +- 0.33% )
1,380,638 context-switches # 0.074 K/sec ( +- 0.75% )
24,390 cpu-migrations # 0.001 K/sec ( +- 8.44% )
1,843,305,768 page-faults # 0.099 M/sec ( +- 0.00% )
50,134,994,088,218 cycles # 2.691 GHz ( +- 0.33% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
8,049,712,224,651 instructions # 0.16 insns per cycle ( +- 0.04% )
1,586,970,584,979 branches # 85.176 M/sec ( +- 0.05% )
1,724,989,949 branch-misses # 0.11% of all branches ( +- 0.48% )
132.474343877 seconds time elapsed ( +- 0.21% )
lockless:
12195979.037525 task-clock (msec) # 133.480 CPUs utilized ( +- 0.18% )
832,850 context-switches # 0.068 K/sec ( +- 0.54% )
15,624 cpu-migrations # 0.001 K/sec ( +- 10.17% )
1,843,304,774 page-faults # 0.151 M/sec ( +- 0.00% )
32,811,216,801,141 cycles # 2.690 GHz ( +- 0.18% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
9,999,265,091,727 instructions # 0.30 insns per cycle ( +- 0.10% )
2,076,759,325,203 branches # 170.282 M/sec ( +- 0.12% )
1,656,917,214 branch-misses # 0.08% of all branches ( +- 0.55% )
91.369330729 seconds time elapsed ( +- 0.45% )
On top of improved scalability, this also gets rid of the icky long long
types in the very heart of memcg, which is great for 32 bit and also makes
the code a lot more readable.
Notable differences between the old and new API:
- res_counter_charge() and res_counter_charge_nofail() become
page_counter_try_charge() and page_counter_charge() resp. to match
the more common kernel naming scheme of try_do()/do()
- res_counter_uncharge_until() is only ever used to cancel a local
counter and never to uncharge bigger segments of a hierarchy, so
it's replaced by the simpler page_counter_cancel()
- res_counter_set_limit() is replaced by page_counter_limit(), which
expects its callers to serialize against themselves
- res_counter_memparse_write_strategy() is replaced by
page_counter_limit(), which rounds down to the nearest page size -
rather than up. This is more reasonable for explicitely requested
hard upper limits.
- to keep charging light-weight, page_counter_try_charge() charges
speculatively, only to roll back if the result exceeds the limit.
Because of this, a failing bigger charge can temporarily lock out
smaller charges that would otherwise succeed. The error is bounded
to the difference between the smallest and the biggest possible
charge size, so for memcg, this means that a failing THP charge can
send base page charges into reclaim upto 2MB (4MB) before the limit
would have been reached. This should be acceptable.
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE and memparse]
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE, memparse, strncmp, and PAGE_SIZE]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Vladimir Davydov <vdavydov@parallels.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-12-10 15:42:31 -08:00
|
|
|
|
|
|
|
|
|
2014-10-09 15:28:54 -07:00
|
|
|
|
mm: memcontrol: lockless page counters
Memory is internally accounted in bytes, using spinlock-protected 64-bit
counters, even though the smallest accounting delta is a page. The
counter interface is also convoluted and does too many things.
Introduce a new lockless word-sized page counter API, then change all
memory accounting over to it. The translation from and to bytes then only
happens when interfacing with userspace.
The removed locking overhead is noticable when scaling beyond the per-cpu
charge caches - on a 4-socket machine with 144-threads, the following test
shows the performance differences of 288 memcgs concurrently running a
page fault benchmark:
vanilla:
18631648.500498 task-clock (msec) # 140.643 CPUs utilized ( +- 0.33% )
1,380,638 context-switches # 0.074 K/sec ( +- 0.75% )
24,390 cpu-migrations # 0.001 K/sec ( +- 8.44% )
1,843,305,768 page-faults # 0.099 M/sec ( +- 0.00% )
50,134,994,088,218 cycles # 2.691 GHz ( +- 0.33% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
8,049,712,224,651 instructions # 0.16 insns per cycle ( +- 0.04% )
1,586,970,584,979 branches # 85.176 M/sec ( +- 0.05% )
1,724,989,949 branch-misses # 0.11% of all branches ( +- 0.48% )
132.474343877 seconds time elapsed ( +- 0.21% )
lockless:
12195979.037525 task-clock (msec) # 133.480 CPUs utilized ( +- 0.18% )
832,850 context-switches # 0.068 K/sec ( +- 0.54% )
15,624 cpu-migrations # 0.001 K/sec ( +- 10.17% )
1,843,304,774 page-faults # 0.151 M/sec ( +- 0.00% )
32,811,216,801,141 cycles # 2.690 GHz ( +- 0.18% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
9,999,265,091,727 instructions # 0.30 insns per cycle ( +- 0.10% )
2,076,759,325,203 branches # 170.282 M/sec ( +- 0.12% )
1,656,917,214 branch-misses # 0.08% of all branches ( +- 0.55% )
91.369330729 seconds time elapsed ( +- 0.45% )
On top of improved scalability, this also gets rid of the icky long long
types in the very heart of memcg, which is great for 32 bit and also makes
the code a lot more readable.
Notable differences between the old and new API:
- res_counter_charge() and res_counter_charge_nofail() become
page_counter_try_charge() and page_counter_charge() resp. to match
the more common kernel naming scheme of try_do()/do()
- res_counter_uncharge_until() is only ever used to cancel a local
counter and never to uncharge bigger segments of a hierarchy, so
it's replaced by the simpler page_counter_cancel()
- res_counter_set_limit() is replaced by page_counter_limit(), which
expects its callers to serialize against themselves
- res_counter_memparse_write_strategy() is replaced by
page_counter_limit(), which rounds down to the nearest page size -
rather than up. This is more reasonable for explicitely requested
hard upper limits.
- to keep charging light-weight, page_counter_try_charge() charges
speculatively, only to roll back if the result exceeds the limit.
Because of this, a failing bigger charge can temporarily lock out
smaller charges that would otherwise succeed. The error is bounded
to the difference between the smallest and the biggest possible
charge size, so for memcg, this means that a failing THP charge can
send base page charges into reclaim upto 2MB (4MB) before the limit
would have been reached. This should be acceptable.
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE and memparse]
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE, memparse, strncmp, and PAGE_SIZE]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Vladimir Davydov <vdavydov@parallels.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-12-10 15:42:31 -08:00
|
|
|
|
2022-07-14 06:49:18 +00:00
|
|
|
|
2014-10-09 15:28:54 -07:00
|
|
|
|
2009-01-07 18:07:48 -08:00
|
|
|
|
2014-08-06 16:05:42 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2009-01-07 18:08:06 -08:00
|
|
|
|
2016-10-27 17:46:56 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-08-06 16:05:44 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-11-06 16:28:21 -08:00
|
|
|
|
2014-08-06 16:05:42 -07:00
|
|
|
|
2010-08-10 18:02:57 -07:00
|
|
|
|
2025-09-22 15:02:03 -07:00
|
|
|
|
2022-07-01 20:35:21 -07:00
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
2020-08-06 23:22:15 -07:00
|
|
|
|
2014-10-09 15:28:56 -07:00
|
|
|
|
2024-01-03 08:48:37 -08:00
|
|
|
|
2020-08-06 23:22:15 -07:00
|
|
|
|
2014-08-06 16:05:42 -07:00
|
|
|
|
2014-08-06 16:08:16 -07:00
|
|
|
|
2014-08-06 16:05:42 -07:00
|
|
|
|
2014-08-06 16:05:47 -07:00
|
|
|
|
2014-10-09 15:28:56 -07:00
|
|
|
|
2014-12-10 15:42:50 -08:00
|
|
|
|
2014-10-09 15:28:56 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-08-06 16:05:47 -07:00
|
|
|
|
|
|
|
|
|
2014-08-06 16:05:42 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-08-06 16:08:16 -07:00
|
|
|
|
2014-08-06 16:05:42 -07:00
|
|
|
|
|
|
|
|
|
2014-08-06 16:05:51 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-07-11 20:55:48 -07:00
|
|
|
|
2018-08-17 15:47:11 -07:00
|
|
|
|
|
|
|
|
|
2021-11-05 13:38:09 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-08-06 16:05:42 -07:00
|
|
|
|
2018-08-17 15:47:11 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-03-22 14:40:19 -07:00
|
|
|
|
|
|
|
|
|
2021-11-05 13:38:09 -07:00
|
|
|
|
2020-08-06 23:21:58 -07:00
|
|
|
|
2018-08-17 15:47:11 -07:00
|
|
|
|
|
|
|
|
|
2009-01-07 18:07:48 -08:00
|
|
|
|
2022-03-22 14:40:22 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-10-31 16:34:13 -07:00
|
|
|
|
memcg: ratify and consolidate over-charge handling
try_charge() is the main charging logic of memcg. When it hits the limit
but either can't fail the allocation due to __GFP_NOFAIL or the task is
likely to free memory very soon, being OOM killed, has SIGKILL pending or
exiting, it "bypasses" the charge to the root memcg and returns -EINTR.
While this is one approach which can be taken for these situations, it has
several issues.
* It unnecessarily lies about the reality. The number itself doesn't
go over the limit but the actual usage does. memcg is either forced
to or actively chooses to go over the limit because that is the
right behavior under the circumstances, which is completely fine,
but, if at all avoidable, it shouldn't be misrepresenting what's
happening by sneaking the charges into the root memcg.
* Despite trying, we already do over-charge. kmemcg can't deal with
switching over to the root memcg by the point try_charge() returns
-EINTR, so it open-codes over-charing.
* It complicates the callers. Each try_charge() user has to handle
the weird -EINTR exception. memcg_charge_kmem() does the manual
over-charging. mem_cgroup_do_precharge() performs unnecessary
uncharging of root memcg, which BTW is inconsistent with what
memcg_charge_kmem() does but not broken as [un]charging are noops on
root memcg. mem_cgroup_try_charge() needs to switch the returned
cgroup to the root one.
The reality is that in memcg there are cases where we are forced and/or
willing to go over the limit. Each such case needs to be scrutinized and
justified but there definitely are situations where that is the right
thing to do. We alredy do this but with a superficial and inconsistent
disguise which leads to unnecessary complications.
This patch updates try_charge() so that it over-charges and returns 0 when
deemed necessary. -EINTR return is removed along with all special case
handling in the callers.
While at it, remove the local variable @ret, which was initialized to zero
and never changed, along with done: label which just returned the always
zero @ret.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Vladimir Davydov <vdavydov@parallels.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-05 18:46:17 -08:00
|
|
|
|
2022-07-01 20:35:21 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-09-22 15:02:03 -07:00
|
|
|
|
2022-07-01 20:35:21 -07:00
|
|
|
|
memcg: ratify and consolidate over-charge handling
try_charge() is the main charging logic of memcg. When it hits the limit
but either can't fail the allocation due to __GFP_NOFAIL or the task is
likely to free memory very soon, being OOM killed, has SIGKILL pending or
exiting, it "bypasses" the charge to the root memcg and returns -EINTR.
While this is one approach which can be taken for these situations, it has
several issues.
* It unnecessarily lies about the reality. The number itself doesn't
go over the limit but the actual usage does. memcg is either forced
to or actively chooses to go over the limit because that is the
right behavior under the circumstances, which is completely fine,
but, if at all avoidable, it shouldn't be misrepresenting what's
happening by sneaking the charges into the root memcg.
* Despite trying, we already do over-charge. kmemcg can't deal with
switching over to the root memcg by the point try_charge() returns
-EINTR, so it open-codes over-charing.
* It complicates the callers. Each try_charge() user has to handle
the weird -EINTR exception. memcg_charge_kmem() does the manual
over-charging. mem_cgroup_do_precharge() performs unnecessary
uncharging of root memcg, which BTW is inconsistent with what
memcg_charge_kmem() does but not broken as [un]charging are noops on
root memcg. mem_cgroup_try_charge() needs to switch the returned
cgroup to the root one.
The reality is that in memcg there are cases where we are forced and/or
willing to go over the limit. Each such case needs to be scrutinized and
justified but there definitely are situations where that is the right
thing to do. We alredy do this but with a superficial and inconsistent
disguise which leads to unnecessary complications.
This patch updates try_charge() so that it over-charges and returns 0 when
deemed necessary. -EINTR return is removed along with all special case
handling in the callers.
While at it, remove the local variable @ret, which was initialized to zero
and never changed, along with done: label which just returned the always
zero @ret.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Vladimir Davydov <vdavydov@parallels.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-05 18:46:17 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-01-14 15:21:23 -08:00
|
|
|
|
memcg: ratify and consolidate over-charge handling
try_charge() is the main charging logic of memcg. When it hits the limit
but either can't fail the allocation due to __GFP_NOFAIL or the task is
likely to free memory very soon, being OOM killed, has SIGKILL pending or
exiting, it "bypasses" the charge to the root memcg and returns -EINTR.
While this is one approach which can be taken for these situations, it has
several issues.
* It unnecessarily lies about the reality. The number itself doesn't
go over the limit but the actual usage does. memcg is either forced
to or actively chooses to go over the limit because that is the
right behavior under the circumstances, which is completely fine,
but, if at all avoidable, it shouldn't be misrepresenting what's
happening by sneaking the charges into the root memcg.
* Despite trying, we already do over-charge. kmemcg can't deal with
switching over to the root memcg by the point try_charge() returns
-EINTR, so it open-codes over-charing.
* It complicates the callers. Each try_charge() user has to handle
the weird -EINTR exception. memcg_charge_kmem() does the manual
over-charging. mem_cgroup_do_precharge() performs unnecessary
uncharging of root memcg, which BTW is inconsistent with what
memcg_charge_kmem() does but not broken as [un]charging are noops on
root memcg. mem_cgroup_try_charge() needs to switch the returned
cgroup to the root one.
The reality is that in memcg there are cases where we are forced and/or
willing to go over the limit. Each such case needs to be scrutinized and
justified but there definitely are situations where that is the right
thing to do. We alredy do this but with a superficial and inconsistent
disguise which leads to unnecessary complications.
This patch updates try_charge() so that it over-charges and returns 0 when
deemed necessary. -EINTR return is removed along with all special case
handling in the callers.
While at it, remove the local variable @ret, which was initialized to zero
and never changed, along with done: label which just returned the always
zero @ret.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Vladimir Davydov <vdavydov@parallels.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-05 18:46:17 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-08-06 16:05:42 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
memcg: punt high overage reclaim to return-to-userland path
Currently, try_charge() tries to reclaim memory synchronously when the
high limit is breached; however, if the allocation doesn't have
__GFP_WAIT, synchronous reclaim is skipped. If a process performs only
speculative allocations, it can blow way past the high limit. This is
actually easily reproducible by simply doing "find /". slab/slub
allocator tries speculative allocations first, so as long as there's
memory which can be consumed without blocking, it can keep allocating
memory regardless of the high limit.
This patch makes try_charge() always punt the over-high reclaim to the
return-to-userland path. If try_charge() detects that high limit is
breached, it adds the overage to current->memcg_nr_pages_over_high and
schedules execution of mem_cgroup_handle_over_high() which performs
synchronous reclaim from the return-to-userland path.
As long as kernel doesn't have a run-away allocation spree, this should
provide enough protection while making kmemcg behave more consistently.
It also has the following benefits.
- All over-high reclaims can use GFP_KERNEL regardless of the specific
gfp mask in use, e.g. GFP_NOFS, when the limit was breached.
- It copes with prio inversion. Previously, a low-prio task with
small memory.high might perform over-high reclaim with a bunch of
locks held. If a higher prio task needed any of these locks, it
would have to wait until the low prio task finished reclaim and
released the locks. By handing over-high reclaim to the task exit
path this issue can be avoided.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Michal Hocko <mhocko@kernel.org>
Reviewed-by: Vladimir Davydov <vdavydov@parallels.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-05 18:46:11 -08:00
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
memcg: punt high overage reclaim to return-to-userland path
Currently, try_charge() tries to reclaim memory synchronously when the
high limit is breached; however, if the allocation doesn't have
__GFP_WAIT, synchronous reclaim is skipped. If a process performs only
speculative allocations, it can blow way past the high limit. This is
actually easily reproducible by simply doing "find /". slab/slub
allocator tries speculative allocations first, so as long as there's
memory which can be consumed without blocking, it can keep allocating
memory regardless of the high limit.
This patch makes try_charge() always punt the over-high reclaim to the
return-to-userland path. If try_charge() detects that high limit is
breached, it adds the overage to current->memcg_nr_pages_over_high and
schedules execution of mem_cgroup_handle_over_high() which performs
synchronous reclaim from the return-to-userland path.
As long as kernel doesn't have a run-away allocation spree, this should
provide enough protection while making kmemcg behave more consistently.
It also has the following benefits.
- All over-high reclaims can use GFP_KERNEL regardless of the specific
gfp mask in use, e.g. GFP_NOFS, when the limit was breached.
- It copes with prio inversion. Previously, a low-prio task with
small memory.high might perform over-high reclaim with a bunch of
locks held. If a higher prio task needed any of these locks, it
would have to wait until the low prio task finished reclaim and
released the locks. By handing over-high reclaim to the task exit
path this issue can be avoided.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Michal Hocko <mhocko@kernel.org>
Reviewed-by: Vladimir Davydov <vdavydov@parallels.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-05 18:46:11 -08:00
|
|
|
|
|
|
|
|
|
2015-11-06 16:28:28 -08:00
|
|
|
|
memcg: punt high overage reclaim to return-to-userland path
Currently, try_charge() tries to reclaim memory synchronously when the
high limit is breached; however, if the allocation doesn't have
__GFP_WAIT, synchronous reclaim is skipped. If a process performs only
speculative allocations, it can blow way past the high limit. This is
actually easily reproducible by simply doing "find /". slab/slub
allocator tries speculative allocations first, so as long as there's
memory which can be consumed without blocking, it can keep allocating
memory regardless of the high limit.
This patch makes try_charge() always punt the over-high reclaim to the
return-to-userland path. If try_charge() detects that high limit is
breached, it adds the overage to current->memcg_nr_pages_over_high and
schedules execution of mem_cgroup_handle_over_high() which performs
synchronous reclaim from the return-to-userland path.
As long as kernel doesn't have a run-away allocation spree, this should
provide enough protection while making kmemcg behave more consistently.
It also has the following benefits.
- All over-high reclaims can use GFP_KERNEL regardless of the specific
gfp mask in use, e.g. GFP_NOFS, when the limit was breached.
- It copes with prio inversion. Previously, a low-prio task with
small memory.high might perform over-high reclaim with a bunch of
locks held. If a higher prio task needed any of these locks, it
would have to wait until the low prio task finished reclaim and
released the locks. By handing over-high reclaim to the task exit
path this issue can be avoided.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Michal Hocko <mhocko@kernel.org>
Reviewed-by: Vladimir Davydov <vdavydov@parallels.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-05 18:46:11 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
|
|
|
|
|
2020-06-01 21:49:52 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-03-22 14:40:07 -07:00
|
|
|
|
2020-06-01 21:49:52 -07:00
|
|
|
|
2016-01-14 15:21:29 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-06-01 21:49:52 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-12-11 13:40:24 -08:00
|
|
|
|
memcg: punt high overage reclaim to return-to-userland path
Currently, try_charge() tries to reclaim memory synchronously when the
high limit is breached; however, if the allocation doesn't have
__GFP_WAIT, synchronous reclaim is skipped. If a process performs only
speculative allocations, it can blow way past the high limit. This is
actually easily reproducible by simply doing "find /". slab/slub
allocator tries speculative allocations first, so as long as there's
memory which can be consumed without blocking, it can keep allocating
memory regardless of the high limit.
This patch makes try_charge() always punt the over-high reclaim to the
return-to-userland path. If try_charge() detects that high limit is
breached, it adds the overage to current->memcg_nr_pages_over_high and
schedules execution of mem_cgroup_handle_over_high() which performs
synchronous reclaim from the return-to-userland path.
As long as kernel doesn't have a run-away allocation spree, this should
provide enough protection while making kmemcg behave more consistently.
It also has the following benefits.
- All over-high reclaims can use GFP_KERNEL regardless of the specific
gfp mask in use, e.g. GFP_NOFS, when the limit was breached.
- It copes with prio inversion. Previously, a low-prio task with
small memory.high might perform over-high reclaim with a bunch of
locks held. If a higher prio task needed any of these locks, it
would have to wait until the low prio task finished reclaim and
released the locks. By handing over-high reclaim to the task exit
path this issue can be avoided.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Michal Hocko <mhocko@kernel.org>
Reviewed-by: Vladimir Davydov <vdavydov@parallels.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-05 18:46:11 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
memcg: ratify and consolidate over-charge handling
try_charge() is the main charging logic of memcg. When it hits the limit
but either can't fail the allocation due to __GFP_NOFAIL or the task is
likely to free memory very soon, being OOM killed, has SIGKILL pending or
exiting, it "bypasses" the charge to the root memcg and returns -EINTR.
While this is one approach which can be taken for these situations, it has
several issues.
* It unnecessarily lies about the reality. The number itself doesn't
go over the limit but the actual usage does. memcg is either forced
to or actively chooses to go over the limit because that is the
right behavior under the circumstances, which is completely fine,
but, if at all avoidable, it shouldn't be misrepresenting what's
happening by sneaking the charges into the root memcg.
* Despite trying, we already do over-charge. kmemcg can't deal with
switching over to the root memcg by the point try_charge() returns
-EINTR, so it open-codes over-charing.
* It complicates the callers. Each try_charge() user has to handle
the weird -EINTR exception. memcg_charge_kmem() does the manual
over-charging. mem_cgroup_do_precharge() performs unnecessary
uncharging of root memcg, which BTW is inconsistent with what
memcg_charge_kmem() does but not broken as [un]charging are noops on
root memcg. mem_cgroup_try_charge() needs to switch the returned
cgroup to the root one.
The reality is that in memcg there are cases where we are forced and/or
willing to go over the limit. Each such case needs to be scrutinized and
justified but there definitely are situations where that is the right
thing to do. We alredy do this but with a superficial and inconsistent
disguise which leads to unnecessary complications.
This patch updates try_charge() so that it over-charges and returns 0 when
deemed necessary. -EINTR return is removed along with all special case
handling in the callers.
While at it, remove the local variable @ret, which was initialized to zero
and never changed, along with done: label which just returned the always
zero @ret.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Vladimir Davydov <vdavydov@parallels.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-05 18:46:17 -08:00
|
|
|
|
mm: memcontrol: don't throttle dying tasks on memory.high
While investigating hosts with high cgroup memory pressures, Tejun
found culprit zombie tasks that had were holding on to a lot of
memory, had SIGKILL pending, but were stuck in memory.high reclaim.
In the past, we used to always force-charge allocations from tasks
that were exiting in order to accelerate them dying and freeing up
their rss. This changed for memory.max in a4ebf1b6ca1e ("memcg:
prohibit unconditional exceeding the limit of dying tasks"); it noted
that this can cause (userspace inducable) containment failures, so it
added a mandatory reclaim and OOM kill cycle before forcing charges.
At the time, memory.high enforcement was handled in the userspace
return path, which isn't reached by dying tasks, and so memory.high
was still never enforced by dying tasks.
When c9afe31ec443 ("memcg: synchronously enforce memory.high for large
overcharges") added synchronous reclaim for memory.high, it added
unconditional memory.high enforcement for dying tasks as well. The
callstack shows that this path is where the zombie is stuck in.
We need to accelerate dying tasks getting past memory.high, but we
cannot do it quite the same way as we do for memory.max: memory.max is
enforced strictly, and tasks aren't allowed to move past it without
FIRST reclaiming and OOM killing if necessary. This ensures very small
levels of excess. With memory.high, though, enforcement happens lazily
after the charge, and OOM killing is never triggered. A lot of
concurrent threads could have pushed, or could actively be pushing,
the cgroup into excess. The dying task will enter reclaim on every
allocation attempt, with little hope of restoring balance.
To fix this, skip synchronous memory.high enforcement on dying tasks
altogether again. Update memory.high path documentation while at it.
[hannes@cmpxchg.org: also handle tasks are being killed during the reclaim]
Link: https://lkml.kernel.org/r/20240111192807.GA424308@cmpxchg.org
Link: https://lkml.kernel.org/r/20240111132902.389862-1-hannes@cmpxchg.org
Fixes: c9afe31ec443 ("memcg: synchronously enforce memory.high for large overcharges")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reported-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Yosry Ahmed <yosryahmed@google.com>
Acked-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Dan Schatzberg <schatzberg.dan@gmail.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-01-11 08:29:02 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-03-22 14:40:28 -07:00
|
|
|
|
|
|
|
|
|
mm: memcontrol: don't throttle dying tasks on memory.high
While investigating hosts with high cgroup memory pressures, Tejun
found culprit zombie tasks that had were holding on to a lot of
memory, had SIGKILL pending, but were stuck in memory.high reclaim.
In the past, we used to always force-charge allocations from tasks
that were exiting in order to accelerate them dying and freeing up
their rss. This changed for memory.max in a4ebf1b6ca1e ("memcg:
prohibit unconditional exceeding the limit of dying tasks"); it noted
that this can cause (userspace inducable) containment failures, so it
added a mandatory reclaim and OOM kill cycle before forcing charges.
At the time, memory.high enforcement was handled in the userspace
return path, which isn't reached by dying tasks, and so memory.high
was still never enforced by dying tasks.
When c9afe31ec443 ("memcg: synchronously enforce memory.high for large
overcharges") added synchronous reclaim for memory.high, it added
unconditional memory.high enforcement for dying tasks as well. The
callstack shows that this path is where the zombie is stuck in.
We need to accelerate dying tasks getting past memory.high, but we
cannot do it quite the same way as we do for memory.max: memory.max is
enforced strictly, and tasks aren't allowed to move past it without
FIRST reclaiming and OOM killing if necessary. This ensures very small
levels of excess. With memory.high, though, enforcement happens lazily
after the charge, and OOM killing is never triggered. A lot of
concurrent threads could have pushed, or could actively be pushing,
the cgroup into excess. The dying task will enter reclaim on every
allocation attempt, with little hope of restoring balance.
To fix this, skip synchronous memory.high enforcement on dying tasks
altogether again. Update memory.high path documentation while at it.
[hannes@cmpxchg.org: also handle tasks are being killed during the reclaim]
Link: https://lkml.kernel.org/r/20240111192807.GA424308@cmpxchg.org
Link: https://lkml.kernel.org/r/20240111132902.389862-1-hannes@cmpxchg.org
Fixes: c9afe31ec443 ("memcg: synchronously enforce memory.high for large overcharges")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reported-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Yosry Ahmed <yosryahmed@google.com>
Acked-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Dan Schatzberg <schatzberg.dan@gmail.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-01-11 08:29:02 -05:00
|
|
|
|
2025-08-13 16:57:55 +02:00
|
|
|
|
memcg: ratify and consolidate over-charge handling
try_charge() is the main charging logic of memcg. When it hits the limit
but either can't fail the allocation due to __GFP_NOFAIL or the task is
likely to free memory very soon, being OOM killed, has SIGKILL pending or
exiting, it "bypasses" the charge to the root memcg and returns -EINTR.
While this is one approach which can be taken for these situations, it has
several issues.
* It unnecessarily lies about the reality. The number itself doesn't
go over the limit but the actual usage does. memcg is either forced
to or actively chooses to go over the limit because that is the
right behavior under the circumstances, which is completely fine,
but, if at all avoidable, it shouldn't be misrepresenting what's
happening by sneaking the charges into the root memcg.
* Despite trying, we already do over-charge. kmemcg can't deal with
switching over to the root memcg by the point try_charge() returns
-EINTR, so it open-codes over-charing.
* It complicates the callers. Each try_charge() user has to handle
the weird -EINTR exception. memcg_charge_kmem() does the manual
over-charging. mem_cgroup_do_precharge() performs unnecessary
uncharging of root memcg, which BTW is inconsistent with what
memcg_charge_kmem() does but not broken as [un]charging are noops on
root memcg. mem_cgroup_try_charge() needs to switch the returned
cgroup to the root one.
The reality is that in memcg there are cases where we are forced and/or
willing to go over the limit. Each such case needs to be scrutinized and
justified but there definitely are situations where that is the right
thing to do. We alredy do this but with a superficial and inconsistent
disguise which leads to unnecessary complications.
This patch updates try_charge() so that it over-charges and returns 0 when
deemed necessary. -EINTR return is removed along with all special case
handling in the callers.
While at it, remove the local variable @ret, which was initialized to zero
and never changed, along with done: label which just returned the always
zero @ret.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Vladimir Davydov <vdavydov@parallels.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-05 18:46:17 -08:00
|
|
|
|
2009-01-07 18:07:48 -08:00
|
|
|
|
2008-02-07 00:13:53 -08:00
|
|
|
|
2025-01-23 23:38:58 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-04-29 14:07:12 -04:00
|
|
|
|
mm: memcontrol: rewrite uncharge API
The memcg uncharging code that is involved towards the end of a page's
lifetime - truncation, reclaim, swapout, migration - is impressively
complicated and fragile.
Because anonymous and file pages were always charged before they had their
page->mapping established, uncharges had to happen when the page type
could still be known from the context; as in unmap for anonymous, page
cache removal for file and shmem pages, and swap cache truncation for swap
pages. However, these operations happen well before the page is actually
freed, and so a lot of synchronization is necessary:
- Charging, uncharging, page migration, and charge migration all need
to take a per-page bit spinlock as they could race with uncharging.
- Swap cache truncation happens during both swap-in and swap-out, and
possibly repeatedly before the page is actually freed. This means
that the memcg swapout code is called from many contexts that make
no sense and it has to figure out the direction from page state to
make sure memory and memory+swap are always correctly charged.
- On page migration, the old page might be unmapped but then reused,
so memcg code has to prevent untimely uncharging in that case.
Because this code - which should be a simple charge transfer - is so
special-cased, it is not reusable for replace_page_cache().
But now that charged pages always have a page->mapping, introduce
mem_cgroup_uncharge(), which is called after the final put_page(), when we
know for sure that nobody is looking at the page anymore.
For page migration, introduce mem_cgroup_migrate(), which is called after
the migration is successful and the new page is fully rmapped. Because
the old page is no longer uncharged after migration, prevent double
charges by decoupling the page's memcg association (PCG_USED and
pc->mem_cgroup) from the page holding an actual charge. The new bits
PCG_MEM and PCG_MEMSW represent the respective charges and are transferred
to the new page during migration.
mem_cgroup_migrate() is suitable for replace_page_cache() as well,
which gets rid of mem_cgroup_replace_page_cache(). However, care
needs to be taken because both the source and the target page can
already be charged and on the LRU when fuse is splicing: grab the page
lock on the charge moving side to prevent changing pc->mem_cgroup of a
page under migration. Also, the lruvecs of both pages change as we
uncharge the old and charge the new during migration, and putback may
race with us, so grab the lru lock and isolate the pages iff on LRU to
prevent races and ensure the pages are on the right lruvec afterward.
Swap accounting is massively simplified: because the page is no longer
uncharged as early as swap cache deletion, a new mem_cgroup_swapout() can
transfer the page's memory+swap charge (PCG_MEMSW) to the swap entry
before the final put_page() in page reclaim.
Finally, page_cgroup changes are now protected by whatever protection the
page itself offers: anonymous pages are charged under the page table lock,
whereas page cache insertions, swapin, and migration hold the page lock.
Uncharging happens under full exclusion with no outstanding references.
Charging and uncharging also ensure that the page is off-LRU, which
serializes against charge migration. Remove the very costly page_cgroup
lock and set pc->flags non-atomically.
[mhocko@suse.cz: mem_cgroup_charge_statistics needs preempt_disable]
[vdavydov@parallels.com: fix flags definition]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Tested-by: Jet Chen <jet.chen@intel.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Tested-by: Felipe Balbi <balbi@ti.com>
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-08-08 14:19:22 -07:00
|
|
|
|
2024-08-14 17:34:15 +08:00
|
|
|
|
mm: memcontrol: rewrite uncharge API
The memcg uncharging code that is involved towards the end of a page's
lifetime - truncation, reclaim, swapout, migration - is impressively
complicated and fragile.
Because anonymous and file pages were always charged before they had their
page->mapping established, uncharges had to happen when the page type
could still be known from the context; as in unmap for anonymous, page
cache removal for file and shmem pages, and swap cache truncation for swap
pages. However, these operations happen well before the page is actually
freed, and so a lot of synchronization is necessary:
- Charging, uncharging, page migration, and charge migration all need
to take a per-page bit spinlock as they could race with uncharging.
- Swap cache truncation happens during both swap-in and swap-out, and
possibly repeatedly before the page is actually freed. This means
that the memcg swapout code is called from many contexts that make
no sense and it has to figure out the direction from page state to
make sure memory and memory+swap are always correctly charged.
- On page migration, the old page might be unmapped but then reused,
so memcg code has to prevent untimely uncharging in that case.
Because this code - which should be a simple charge transfer - is so
special-cased, it is not reusable for replace_page_cache().
But now that charged pages always have a page->mapping, introduce
mem_cgroup_uncharge(), which is called after the final put_page(), when we
know for sure that nobody is looking at the page anymore.
For page migration, introduce mem_cgroup_migrate(), which is called after
the migration is successful and the new page is fully rmapped. Because
the old page is no longer uncharged after migration, prevent double
charges by decoupling the page's memcg association (PCG_USED and
pc->mem_cgroup) from the page holding an actual charge. The new bits
PCG_MEM and PCG_MEMSW represent the respective charges and are transferred
to the new page during migration.
mem_cgroup_migrate() is suitable for replace_page_cache() as well,
which gets rid of mem_cgroup_replace_page_cache(). However, care
needs to be taken because both the source and the target page can
already be charged and on the LRU when fuse is splicing: grab the page
lock on the charge moving side to prevent changing pc->mem_cgroup of a
page under migration. Also, the lruvecs of both pages change as we
uncharge the old and charge the new during migration, and putback may
race with us, so grab the lru lock and isolate the pages iff on LRU to
prevent races and ensure the pages are on the right lruvec afterward.
Swap accounting is massively simplified: because the page is no longer
uncharged as early as swap cache deletion, a new mem_cgroup_swapout() can
transfer the page's memory+swap charge (PCG_MEMSW) to the swap entry
before the final put_page() in page reclaim.
Finally, page_cgroup changes are now protected by whatever protection the
page itself offers: anonymous pages are charged under the page table lock,
whereas page cache insertions, swapin, and migration hold the page lock.
Uncharging happens under full exclusion with no outstanding references.
Charging and uncharging also ensure that the page is off-LRU, which
serializes against charge migration. Remove the very costly page_cgroup
lock and set pc->flags non-atomically.
[mhocko@suse.cz: mem_cgroup_charge_statistics needs preempt_disable]
[vdavydov@parallels.com: fix flags definition]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Tested-by: Jet Chen <jet.chen@intel.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Tested-by: Felipe Balbi <balbi@ti.com>
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-08-08 14:19:22 -07:00
|
|
|
|
2020-12-14 19:06:42 -08:00
|
|
|
|
mm: memcontrol: rewrite uncharge API
The memcg uncharging code that is involved towards the end of a page's
lifetime - truncation, reclaim, swapout, migration - is impressively
complicated and fragile.
Because anonymous and file pages were always charged before they had their
page->mapping established, uncharges had to happen when the page type
could still be known from the context; as in unmap for anonymous, page
cache removal for file and shmem pages, and swap cache truncation for swap
pages. However, these operations happen well before the page is actually
freed, and so a lot of synchronization is necessary:
- Charging, uncharging, page migration, and charge migration all need
to take a per-page bit spinlock as they could race with uncharging.
- Swap cache truncation happens during both swap-in and swap-out, and
possibly repeatedly before the page is actually freed. This means
that the memcg swapout code is called from many contexts that make
no sense and it has to figure out the direction from page state to
make sure memory and memory+swap are always correctly charged.
- On page migration, the old page might be unmapped but then reused,
so memcg code has to prevent untimely uncharging in that case.
Because this code - which should be a simple charge transfer - is so
special-cased, it is not reusable for replace_page_cache().
But now that charged pages always have a page->mapping, introduce
mem_cgroup_uncharge(), which is called after the final put_page(), when we
know for sure that nobody is looking at the page anymore.
For page migration, introduce mem_cgroup_migrate(), which is called after
the migration is successful and the new page is fully rmapped. Because
the old page is no longer uncharged after migration, prevent double
charges by decoupling the page's memcg association (PCG_USED and
pc->mem_cgroup) from the page holding an actual charge. The new bits
PCG_MEM and PCG_MEMSW represent the respective charges and are transferred
to the new page during migration.
mem_cgroup_migrate() is suitable for replace_page_cache() as well,
which gets rid of mem_cgroup_replace_page_cache(). However, care
needs to be taken because both the source and the target page can
already be charged and on the LRU when fuse is splicing: grab the page
lock on the charge moving side to prevent changing pc->mem_cgroup of a
page under migration. Also, the lruvecs of both pages change as we
uncharge the old and charge the new during migration, and putback may
race with us, so grab the lru lock and isolate the pages iff on LRU to
prevent races and ensure the pages are on the right lruvec afterward.
Swap accounting is massively simplified: because the page is no longer
uncharged as early as swap cache deletion, a new mem_cgroup_swapout() can
transfer the page's memory+swap charge (PCG_MEMSW) to the swap entry
before the final put_page() in page reclaim.
Finally, page_cgroup changes are now protected by whatever protection the
page itself offers: anonymous pages are charged under the page table lock,
whereas page cache insertions, swapin, and migration hold the page lock.
Uncharging happens under full exclusion with no outstanding references.
Charging and uncharging also ensure that the page is off-LRU, which
serializes against charge migration. Remove the very costly page_cgroup
lock and set pc->flags non-atomically.
[mhocko@suse.cz: mem_cgroup_charge_statistics needs preempt_disable]
[vdavydov@parallels.com: fix flags definition]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Tested-by: Jet Chen <jet.chen@intel.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Tested-by: Felipe Balbi <balbi@ti.com>
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-08-08 14:19:22 -07:00
|
|
|
|
2020-06-03 16:02:27 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcontrol: rewrite uncharge API
The memcg uncharging code that is involved towards the end of a page's
lifetime - truncation, reclaim, swapout, migration - is impressively
complicated and fragile.
Because anonymous and file pages were always charged before they had their
page->mapping established, uncharges had to happen when the page type
could still be known from the context; as in unmap for anonymous, page
cache removal for file and shmem pages, and swap cache truncation for swap
pages. However, these operations happen well before the page is actually
freed, and so a lot of synchronization is necessary:
- Charging, uncharging, page migration, and charge migration all need
to take a per-page bit spinlock as they could race with uncharging.
- Swap cache truncation happens during both swap-in and swap-out, and
possibly repeatedly before the page is actually freed. This means
that the memcg swapout code is called from many contexts that make
no sense and it has to figure out the direction from page state to
make sure memory and memory+swap are always correctly charged.
- On page migration, the old page might be unmapped but then reused,
so memcg code has to prevent untimely uncharging in that case.
Because this code - which should be a simple charge transfer - is so
special-cased, it is not reusable for replace_page_cache().
But now that charged pages always have a page->mapping, introduce
mem_cgroup_uncharge(), which is called after the final put_page(), when we
know for sure that nobody is looking at the page anymore.
For page migration, introduce mem_cgroup_migrate(), which is called after
the migration is successful and the new page is fully rmapped. Because
the old page is no longer uncharged after migration, prevent double
charges by decoupling the page's memcg association (PCG_USED and
pc->mem_cgroup) from the page holding an actual charge. The new bits
PCG_MEM and PCG_MEMSW represent the respective charges and are transferred
to the new page during migration.
mem_cgroup_migrate() is suitable for replace_page_cache() as well,
which gets rid of mem_cgroup_replace_page_cache(). However, care
needs to be taken because both the source and the target page can
already be charged and on the LRU when fuse is splicing: grab the page
lock on the charge moving side to prevent changing pc->mem_cgroup of a
page under migration. Also, the lruvecs of both pages change as we
uncharge the old and charge the new during migration, and putback may
race with us, so grab the lru lock and isolate the pages iff on LRU to
prevent races and ensure the pages are on the right lruvec afterward.
Swap accounting is massively simplified: because the page is no longer
uncharged as early as swap cache deletion, a new mem_cgroup_swapout() can
transfer the page's memory+swap charge (PCG_MEMSW) to the swap entry
before the final put_page() in page reclaim.
Finally, page_cgroup changes are now protected by whatever protection the
page itself offers: anonymous pages are charged under the page table lock,
whereas page cache insertions, swapin, and migration hold the page lock.
Uncharging happens under full exclusion with no outstanding references.
Charging and uncharging also ensure that the page is off-LRU, which
serializes against charge migration. Remove the very costly page_cgroup
lock and set pc->flags non-atomically.
[mhocko@suse.cz: mem_cgroup_charge_statistics needs preempt_disable]
[vdavydov@parallels.com: fix flags definition]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Tested-by: Jet Chen <jet.chen@intel.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Tested-by: Felipe Balbi <balbi@ti.com>
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-08-08 14:19:22 -07:00
|
|
|
|
2021-04-29 14:07:12 -04:00
|
|
|
|
2009-01-07 18:07:48 -08:00
|
|
|
|
2008-02-07 00:13:56 -08:00
|
|
|
|
2025-05-18 23:31:41 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-06-17 12:57:25 -07:00
|
|
|
|
|
|
|
|
|
2025-05-18 23:31:41 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-05-14 11:41:56 -07:00
|
|
|
|
2024-04-20 16:25:05 -07:00
|
|
|
|
|
|
|
|
|
2021-12-10 14:47:05 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-05-18 23:31:41 -07:00
|
|
|
|
2021-12-10 14:47:05 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-06-10 11:03:10 -07:00
|
|
|
|
|
|
|
|
|
2020-03-28 19:17:25 -07:00
|
|
|
|
|
|
|
|
|
2020-08-06 23:21:10 -07:00
|
|
|
|
|
|
|
|
|
2024-03-21 09:36:28 -07:00
|
|
|
|
2020-03-28 19:17:25 -07:00
|
|
|
|
2021-11-02 22:42:04 +01:00
|
|
|
|
2024-03-21 09:36:28 -07:00
|
|
|
|
2021-11-02 22:42:04 +01:00
|
|
|
|
2020-08-06 23:21:10 -07:00
|
|
|
|
|
|
|
|
|
2021-11-02 22:42:04 +01:00
|
|
|
|
2024-03-21 09:36:28 -07:00
|
|
|
|
|
|
|
|
|
2021-11-02 22:42:04 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-21 09:36:28 -07:00
|
|
|
|
|
|
|
|
|
2020-08-06 23:21:27 -07:00
|
|
|
|
|
|
|
|
|
2020-08-06 23:21:10 -07:00
|
|
|
|
2020-03-28 19:17:25 -07:00
|
|
|
|
mm: memcontrol: Use helpers to read page's memcg data
Patch series "mm: allow mapping accounted kernel pages to userspace", v6.
Currently a non-slab kernel page which has been charged to a memory cgroup
can't be mapped to userspace. The underlying reason is simple: PageKmemcg
flag is defined as a page type (like buddy, offline, etc), so it takes a
bit from a page->mapped counter. Pages with a type set can't be mapped to
userspace.
But in general the kmemcg flag has nothing to do with mapping to
userspace. It only means that the page has been accounted by the page
allocator, so it has to be properly uncharged on release.
Some bpf maps are mapping the vmalloc-based memory to userspace, and their
memory can't be accounted because of this implementation detail.
This patchset removes this limitation by moving the PageKmemcg flag into
one of the free bits of the page->mem_cgroup pointer. Also it formalizes
accesses to the page->mem_cgroup and page->obj_cgroups using new helpers,
adds several checks and removes a couple of obsolete functions. As the
result the code became more robust with fewer open-coded bit tricks.
This patch (of 4):
Currently there are many open-coded reads of the page->mem_cgroup pointer,
as well as a couple of read helpers, which are barely used.
It creates an obstacle on a way to reuse some bits of the pointer for
storing additional bits of information. In fact, we already do this for
slab pages, where the last bit indicates that a pointer has an attached
vector of objcg pointers instead of a regular memcg pointer.
This commits uses 2 existing helpers and introduces a new helper to
converts all read sides to calls of these helpers:
struct mem_cgroup *page_memcg(struct page *page);
struct mem_cgroup *page_memcg_rcu(struct page *page);
struct mem_cgroup *page_memcg_check(struct page *page);
page_memcg_check() is intended to be used in cases when the page can be a
slab page and have a memcg pointer pointing at objcg vector. It does
check the lowest bit, and if set, returns NULL. page_memcg() contains a
VM_BUG_ON_PAGE() check for the page not being a slab page.
To make sure nobody uses a direct access, struct page's
mem_cgroup/obj_cgroups is converted to unsigned long memcg_data.
Signed-off-by: Roman Gushchin <guro@fb.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Link: https://lkml.kernel.org/r/20201027001657.3398190-1-guro@fb.com
Link: https://lkml.kernel.org/r/20201027001657.3398190-2-guro@fb.com
Link: https://lore.kernel.org/bpf/20201201215900.3569844-2-guro@fb.com
2020-12-01 13:58:27 -08:00
|
|
|
|
2022-12-30 15:08:42 +08:00
|
|
|
|
2021-11-02 22:42:04 +01:00
|
|
|
|
2024-03-21 09:36:28 -07:00
|
|
|
|
2022-12-30 15:08:42 +08:00
|
|
|
|
mm: memcontrol: Use helpers to read page's memcg data
Patch series "mm: allow mapping accounted kernel pages to userspace", v6.
Currently a non-slab kernel page which has been charged to a memory cgroup
can't be mapped to userspace. The underlying reason is simple: PageKmemcg
flag is defined as a page type (like buddy, offline, etc), so it takes a
bit from a page->mapped counter. Pages with a type set can't be mapped to
userspace.
But in general the kmemcg flag has nothing to do with mapping to
userspace. It only means that the page has been accounted by the page
allocator, so it has to be properly uncharged on release.
Some bpf maps are mapping the vmalloc-based memory to userspace, and their
memory can't be accounted because of this implementation detail.
This patchset removes this limitation by moving the PageKmemcg flag into
one of the free bits of the page->mem_cgroup pointer. Also it formalizes
accesses to the page->mem_cgroup and page->obj_cgroups using new helpers,
adds several checks and removes a couple of obsolete functions. As the
result the code became more robust with fewer open-coded bit tricks.
This patch (of 4):
Currently there are many open-coded reads of the page->mem_cgroup pointer,
as well as a couple of read helpers, which are barely used.
It creates an obstacle on a way to reuse some bits of the pointer for
storing additional bits of information. In fact, we already do this for
slab pages, where the last bit indicates that a pointer has an attached
vector of objcg pointers instead of a regular memcg pointer.
This commits uses 2 existing helpers and introduces a new helper to
converts all read sides to calls of these helpers:
struct mem_cgroup *page_memcg(struct page *page);
struct mem_cgroup *page_memcg_rcu(struct page *page);
struct mem_cgroup *page_memcg_check(struct page *page);
page_memcg_check() is intended to be used in cases when the page can be a
slab page and have a memcg pointer pointing at objcg vector. It does
check the lowest bit, and if set, returns NULL. page_memcg() contains a
VM_BUG_ON_PAGE() check for the page not being a slab page.
To make sure nobody uses a direct access, struct page's
mem_cgroup/obj_cgroups is converted to unsigned long memcg_data.
Signed-off-by: Roman Gushchin <guro@fb.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Link: https://lkml.kernel.org/r/20201027001657.3398190-1-guro@fb.com
Link: https://lkml.kernel.org/r/20201027001657.3398190-2-guro@fb.com
Link: https://lore.kernel.org/bpf/20201201215900.3569844-2-guro@fb.com
2020-12-01 13:58:27 -08:00
|
|
|
|
|
|
|
|
|
2022-12-30 15:08:42 +08:00
|
|
|
|
2020-03-28 19:17:25 -07:00
|
|
|
|
|
|
|
|
|
2022-06-10 11:03:10 -07:00
|
|
|
|
|
|
|
|
|
2024-07-18 17:18:21 +08:00
|
|
|
|
2022-06-10 11:03:10 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-09-30 19:14:33 +05:30
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
2023-10-19 15:53:41 -07:00
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-19 15:53:42 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-16 01:58:03 +00:00
|
|
|
|
2023-10-19 15:53:42 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-19 15:53:44 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-05-18 23:31:38 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-19 15:53:44 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-11-15 18:51:09 -08:00
|
|
|
|
2023-10-19 15:53:44 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-07-15 05:23:41 +01:00
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-02-13 11:29:22 -08:00
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
2023-07-15 05:23:41 +01:00
|
|
|
|
|
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
2023-07-15 05:23:41 +01:00
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-05-18 23:31:40 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-06-17 12:57:25 -07:00
|
|
|
|
|
|
|
|
|
2025-05-18 23:31:40 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-04-29 22:56:55 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-04-29 22:56:42 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-05-18 23:31:40 -07:00
|
|
|
|
2024-06-28 21:03:09 +00:00
|
|
|
|
2025-03-12 22:48:12 -07:00
|
|
|
|
|
|
|
|
|
2021-04-29 22:56:42 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-04-29 22:56:55 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-07-26 15:24:21 -07:00
|
|
|
|
2020-04-01 21:06:49 -07:00
|
|
|
|
2016-07-26 15:24:21 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-04-29 22:56:55 -07:00
|
|
|
|
|
|
|
|
|
2012-12-18 14:21:56 -08:00
|
|
|
|
2021-04-29 22:56:55 -07:00
|
|
|
|
2012-12-18 14:21:56 -08:00
|
|
|
|
|
|
|
|
|
2021-04-29 22:56:55 -07:00
|
|
|
|
|
|
|
|
|
2021-06-28 19:37:44 -07:00
|
|
|
|
2016-01-20 15:02:35 -08:00
|
|
|
|
2021-04-29 22:56:55 -07:00
|
|
|
|
2016-01-20 15:02:35 -08:00
|
|
|
|
2025-05-18 23:31:40 -07:00
|
|
|
|
2024-06-28 21:03:09 +00:00
|
|
|
|
2021-04-29 22:56:55 -07:00
|
|
|
|
|
|
|
|
|
2020-04-01 21:06:56 -07:00
|
|
|
|
2021-04-29 22:56:55 -07:00
|
|
|
|
2020-04-01 21:06:56 -07:00
|
|
|
|
|
|
|
|
|
2025-03-14 13:36:13 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-07-26 15:24:21 -07:00
|
|
|
|
2020-04-01 21:06:46 -07:00
|
|
|
|
2016-07-26 15:24:21 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-04-01 21:06:46 -07:00
|
|
|
|
2012-12-18 14:21:56 -08:00
|
|
|
|
mm: memcontrol: use obj_cgroup APIs to charge kmem pages
Since Roman's series "The new cgroup slab memory controller" applied.
All slab objects are charged via the new APIs of obj_cgroup. The new
APIs introduce a struct obj_cgroup to charge slab objects. It prevents
long-living objects from pinning the original memory cgroup in the
memory. But there are still some corner objects (e.g. allocations
larger than order-1 page on SLUB) which are not charged via the new
APIs. Those objects (include the pages which are allocated from buddy
allocator directly) are charged as kmem pages which still hold a
reference to the memory cgroup.
We want to reuse the obj_cgroup APIs to charge the kmem pages. If we do
that, we should store an object cgroup pointer to page->memcg_data for
the kmem pages.
Finally, page->memcg_data will have 3 different meanings.
1) For the slab pages, page->memcg_data points to an object cgroups
vector.
2) For the kmem pages (exclude the slab pages), page->memcg_data
points to an object cgroup.
3) For the user pages (e.g. the LRU pages), page->memcg_data points
to a memory cgroup.
We do not change the behavior of page_memcg() and page_memcg_rcu(). They
are also suitable for LRU pages and kmem pages. Why?
Because memory allocations pinning memcgs for a long time - it exists at a
larger scale and is causing recurring problems in the real world: page
cache doesn't get reclaimed for a long time, or is used by the second,
third, fourth, ... instance of the same job that was restarted into a new
cgroup every time. Unreclaimable dying cgroups pile up, waste memory, and
make page reclaim very inefficient.
We can convert LRU pages and most other raw memcg pins to the objcg
direction to fix this problem, and then the page->memcg will always point
to an object cgroup pointer. At that time, LRU pages and kmem pages will
be treated the same. The implementation of page_memcg() will remove the
kmem page check.
This patch aims to charge the kmem pages by using the new APIs of
obj_cgroup. Finally, the page->memcg_data of the kmem page points to an
object cgroup. We can use the __page_objcg() to get the object cgroup
associated with a kmem page. Or we can use page_memcg() to get the memory
cgroup associated with a kmem page, but caller must ensure that the
returned memcg won't be released (e.g. acquire the rcu_read_lock or
css_set_lock).
Link: https://lkml.kernel.org/r/20210401030141.37061-1-songmuchun@bytedance.com
Link: https://lkml.kernel.org/r/20210319163821.20704-6-songmuchun@bytedance.com
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Roman Gushchin <guro@fb.com>
Reviewed-by: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Xiongchun Duan <duanxiongchun@bytedance.com>
Cc: Christian Borntraeger <borntraeger@de.ibm.com>
[songmuchun@bytedance.com: fix forget to obtain the ref to objcg in split_page_memcg]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:52 -07:00
|
|
|
|
2016-03-17 14:17:29 -07:00
|
|
|
|
2012-12-18 14:21:56 -08:00
|
|
|
|
2023-10-19 15:53:44 -07:00
|
|
|
|
mm: memcontrol: use obj_cgroup APIs to charge kmem pages
Since Roman's series "The new cgroup slab memory controller" applied.
All slab objects are charged via the new APIs of obj_cgroup. The new
APIs introduce a struct obj_cgroup to charge slab objects. It prevents
long-living objects from pinning the original memory cgroup in the
memory. But there are still some corner objects (e.g. allocations
larger than order-1 page on SLUB) which are not charged via the new
APIs. Those objects (include the pages which are allocated from buddy
allocator directly) are charged as kmem pages which still hold a
reference to the memory cgroup.
We want to reuse the obj_cgroup APIs to charge the kmem pages. If we do
that, we should store an object cgroup pointer to page->memcg_data for
the kmem pages.
Finally, page->memcg_data will have 3 different meanings.
1) For the slab pages, page->memcg_data points to an object cgroups
vector.
2) For the kmem pages (exclude the slab pages), page->memcg_data
points to an object cgroup.
3) For the user pages (e.g. the LRU pages), page->memcg_data points
to a memory cgroup.
We do not change the behavior of page_memcg() and page_memcg_rcu(). They
are also suitable for LRU pages and kmem pages. Why?
Because memory allocations pinning memcgs for a long time - it exists at a
larger scale and is causing recurring problems in the real world: page
cache doesn't get reclaimed for a long time, or is used by the second,
third, fourth, ... instance of the same job that was restarted into a new
cgroup every time. Unreclaimable dying cgroups pile up, waste memory, and
make page reclaim very inefficient.
We can convert LRU pages and most other raw memcg pins to the objcg
direction to fix this problem, and then the page->memcg will always point
to an object cgroup pointer. At that time, LRU pages and kmem pages will
be treated the same. The implementation of page_memcg() will remove the
kmem page check.
This patch aims to charge the kmem pages by using the new APIs of
obj_cgroup. Finally, the page->memcg_data of the kmem page points to an
object cgroup. We can use the __page_objcg() to get the object cgroup
associated with a kmem page. Or we can use page_memcg() to get the memory
cgroup associated with a kmem page, but caller must ensure that the
returned memcg won't be released (e.g. acquire the rcu_read_lock or
css_set_lock).
Link: https://lkml.kernel.org/r/20210401030141.37061-1-songmuchun@bytedance.com
Link: https://lkml.kernel.org/r/20210319163821.20704-6-songmuchun@bytedance.com
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Roman Gushchin <guro@fb.com>
Reviewed-by: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Xiongchun Duan <duanxiongchun@bytedance.com>
Cc: Christian Borntraeger <borntraeger@de.ibm.com>
[songmuchun@bytedance.com: fix forget to obtain the ref to objcg in split_page_memcg]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:52 -07:00
|
|
|
|
|
|
|
|
|
2019-07-11 20:56:31 -07:00
|
|
|
|
2023-10-19 15:53:44 -07:00
|
|
|
|
2025-03-14 13:36:13 +00:00
|
|
|
|
2020-08-06 23:20:45 -07:00
|
|
|
|
2019-07-11 20:56:31 -07:00
|
|
|
|
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
|
|
|
|
2015-11-05 18:48:59 -08:00
|
|
|
|
2012-12-18 14:21:56 -08:00
|
|
|
|
2019-07-11 20:56:13 -07:00
|
|
|
|
2016-07-26 15:24:21 -07:00
|
|
|
|
2020-04-01 21:06:46 -07:00
|
|
|
|
2016-07-26 15:24:21 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-04-01 21:06:46 -07:00
|
|
|
|
2012-12-18 14:21:56 -08:00
|
|
|
|
2025-03-14 13:36:15 +00:00
|
|
|
|
memcg: unify slab and other kmem pages charging
We have memcg_kmem_charge and memcg_kmem_uncharge methods for charging and
uncharging kmem pages to memcg, but currently they are not used for
charging slab pages (i.e. they are only used for charging pages allocated
with alloc_kmem_pages). The only reason why the slab subsystem uses
special helpers, memcg_charge_slab and memcg_uncharge_slab, is that it
needs to charge to the memcg of kmem cache while memcg_charge_kmem charges
to the memcg that the current task belongs to.
To remove this diversity, this patch adds an extra argument to
__memcg_kmem_charge that can be a pointer to a memcg or NULL. If it is
not NULL, the function tries to charge to the memcg it points to,
otherwise it charge to the current context. Next, it makes the slab
subsystem use this function to charge slab pages.
Since memcg_charge_kmem and memcg_uncharge_kmem helpers are now used only
in __memcg_kmem_charge and __memcg_kmem_uncharge, they are inlined. Since
__memcg_kmem_charge stores a pointer to the memcg in the page struct, we
don't need memcg_uncharge_slab anymore and can use free_kmem_pages.
Besides, one can now detect which memcg a slab page belongs to by reading
/proc/kpagecgroup.
Note, this patch switches slab to charge-after-alloc design. Since this
design is already used for all other memcg charges, it should not make any
difference.
[hannes@cmpxchg.org: better to have an outer function than a magic parameter for the memcg lookup]
Signed-off-by: Vladimir Davydov <vdavydov@virtuozzo.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
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>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-05 18:49:01 -08:00
|
|
|
|
2012-12-18 14:21:56 -08:00
|
|
|
|
2025-03-14 13:36:15 +00:00
|
|
|
|
2012-12-18 14:21:56 -08:00
|
|
|
|
|
|
|
|
|
mm: memcontrol: use obj_cgroup APIs to charge kmem pages
Since Roman's series "The new cgroup slab memory controller" applied.
All slab objects are charged via the new APIs of obj_cgroup. The new
APIs introduce a struct obj_cgroup to charge slab objects. It prevents
long-living objects from pinning the original memory cgroup in the
memory. But there are still some corner objects (e.g. allocations
larger than order-1 page on SLUB) which are not charged via the new
APIs. Those objects (include the pages which are allocated from buddy
allocator directly) are charged as kmem pages which still hold a
reference to the memory cgroup.
We want to reuse the obj_cgroup APIs to charge the kmem pages. If we do
that, we should store an object cgroup pointer to page->memcg_data for
the kmem pages.
Finally, page->memcg_data will have 3 different meanings.
1) For the slab pages, page->memcg_data points to an object cgroups
vector.
2) For the kmem pages (exclude the slab pages), page->memcg_data
points to an object cgroup.
3) For the user pages (e.g. the LRU pages), page->memcg_data points
to a memory cgroup.
We do not change the behavior of page_memcg() and page_memcg_rcu(). They
are also suitable for LRU pages and kmem pages. Why?
Because memory allocations pinning memcgs for a long time - it exists at a
larger scale and is causing recurring problems in the real world: page
cache doesn't get reclaimed for a long time, or is used by the second,
third, fourth, ... instance of the same job that was restarted into a new
cgroup every time. Unreclaimable dying cgroups pile up, waste memory, and
make page reclaim very inefficient.
We can convert LRU pages and most other raw memcg pins to the objcg
direction to fix this problem, and then the page->memcg will always point
to an object cgroup pointer. At that time, LRU pages and kmem pages will
be treated the same. The implementation of page_memcg() will remove the
kmem page check.
This patch aims to charge the kmem pages by using the new APIs of
obj_cgroup. Finally, the page->memcg_data of the kmem page points to an
object cgroup. We can use the __page_objcg() to get the object cgroup
associated with a kmem page. Or we can use page_memcg() to get the memory
cgroup associated with a kmem page, but caller must ensure that the
returned memcg won't be released (e.g. acquire the rcu_read_lock or
css_set_lock).
Link: https://lkml.kernel.org/r/20210401030141.37061-1-songmuchun@bytedance.com
Link: https://lkml.kernel.org/r/20210319163821.20704-6-songmuchun@bytedance.com
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Roman Gushchin <guro@fb.com>
Reviewed-by: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Xiongchun Duan <duanxiongchun@bytedance.com>
Cc: Christian Borntraeger <borntraeger@de.ibm.com>
[songmuchun@bytedance.com: fix forget to obtain the ref to objcg in split_page_memcg]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:52 -07:00
|
|
|
|
2025-03-14 13:36:15 +00:00
|
|
|
|
mm: memcontrol: use obj_cgroup APIs to charge kmem pages
Since Roman's series "The new cgroup slab memory controller" applied.
All slab objects are charged via the new APIs of obj_cgroup. The new
APIs introduce a struct obj_cgroup to charge slab objects. It prevents
long-living objects from pinning the original memory cgroup in the
memory. But there are still some corner objects (e.g. allocations
larger than order-1 page on SLUB) which are not charged via the new
APIs. Those objects (include the pages which are allocated from buddy
allocator directly) are charged as kmem pages which still hold a
reference to the memory cgroup.
We want to reuse the obj_cgroup APIs to charge the kmem pages. If we do
that, we should store an object cgroup pointer to page->memcg_data for
the kmem pages.
Finally, page->memcg_data will have 3 different meanings.
1) For the slab pages, page->memcg_data points to an object cgroups
vector.
2) For the kmem pages (exclude the slab pages), page->memcg_data
points to an object cgroup.
3) For the user pages (e.g. the LRU pages), page->memcg_data points
to a memory cgroup.
We do not change the behavior of page_memcg() and page_memcg_rcu(). They
are also suitable for LRU pages and kmem pages. Why?
Because memory allocations pinning memcgs for a long time - it exists at a
larger scale and is causing recurring problems in the real world: page
cache doesn't get reclaimed for a long time, or is used by the second,
third, fourth, ... instance of the same job that was restarted into a new
cgroup every time. Unreclaimable dying cgroups pile up, waste memory, and
make page reclaim very inefficient.
We can convert LRU pages and most other raw memcg pins to the objcg
direction to fix this problem, and then the page->memcg will always point
to an object cgroup pointer. At that time, LRU pages and kmem pages will
be treated the same. The implementation of page_memcg() will remove the
kmem page check.
This patch aims to charge the kmem pages by using the new APIs of
obj_cgroup. Finally, the page->memcg_data of the kmem page points to an
object cgroup. We can use the __page_objcg() to get the object cgroup
associated with a kmem page. Or we can use page_memcg() to get the memory
cgroup associated with a kmem page, but caller must ensure that the
returned memcg won't be released (e.g. acquire the rcu_read_lock or
css_set_lock).
Link: https://lkml.kernel.org/r/20210401030141.37061-1-songmuchun@bytedance.com
Link: https://lkml.kernel.org/r/20210319163821.20704-6-songmuchun@bytedance.com
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Roman Gushchin <guro@fb.com>
Reviewed-by: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Xiongchun Duan <duanxiongchun@bytedance.com>
Cc: Christian Borntraeger <borntraeger@de.ibm.com>
[songmuchun@bytedance.com: fix forget to obtain the ref to objcg in split_page_memcg]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:52 -07:00
|
|
|
|
list_lru: introduce per-memcg lists
There are several FS shrinkers, including super_block::s_shrink, that
keep reclaimable objects in the list_lru structure. Hence to turn them
to memcg-aware shrinkers, it is enough to make list_lru per-memcg.
This patch does the trick. It adds an array of lru lists to the
list_lru_node structure (per-node part of the list_lru), one for each
kmem-active memcg, and dispatches every item addition or removal to the
list corresponding to the memcg which the item is accounted to. So now
the list_lru structure is not just per node, but per node and per memcg.
Not all list_lrus need this feature, so this patch also adds a new
method, list_lru_init_memcg, which initializes a list_lru as memcg
aware. Otherwise (i.e. if initialized with old list_lru_init), the
list_lru won't have per memcg lists.
Just like per memcg caches arrays, the arrays of per-memcg lists are
indexed by memcg_cache_id, so we must grow them whenever
memcg_nr_cache_ids is increased. So we introduce a callback,
memcg_update_all_list_lrus, invoked by memcg_alloc_cache_id if the id
space is full.
The locking is implemented in a manner similar to lruvecs, i.e. we have
one lock per node that protects all lists (both global and per cgroup) on
the node.
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Greg Thelen <gthelen@google.com>
Cc: Glauber Costa <glommer@gmail.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
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: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-12 14:59:10 -08:00
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
2025-04-03 18:39:12 -07:00
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
2025-04-03 18:39:12 -07:00
|
|
|
|
2021-06-28 19:37:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-04-03 18:39:12 -07:00
|
|
|
|
2021-06-28 19:37:23 -07:00
|
|
|
|
2025-04-03 18:39:12 -07:00
|
|
|
|
2021-06-28 19:37:23 -07:00
|
|
|
|
2021-08-13 16:54:41 -07:00
|
|
|
|
|
|
|
|
|
2021-06-28 19:37:23 -07:00
|
|
|
|
2025-05-14 11:41:56 -07:00
|
|
|
|
2021-06-28 19:37:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-05-14 11:41:56 -07:00
|
|
|
|
2021-06-28 19:37:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-05-14 11:41:56 -07:00
|
|
|
|
2021-06-28 19:37:23 -07:00
|
|
|
|
|
|
|
|
|
2025-04-03 18:39:12 -07:00
|
|
|
|
|
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
|
|
|
|
|
2025-05-14 11:41:58 -07:00
|
|
|
|
|
|
|
|
|
2022-03-22 14:40:35 -07:00
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
2023-05-02 09:08:38 -07:00
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
|
|
|
|
|
2025-04-03 18:39:12 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
|
|
|
|
|
2025-05-14 11:41:58 -07:00
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
2023-05-02 09:08:38 -07:00
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
|
|
|
|
|
2025-04-03 18:39:10 -07:00
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-03-22 14:40:44 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-06-28 21:03:09 +00:00
|
|
|
|
|
|
|
|
|
2025-04-03 18:39:07 -07:00
|
|
|
|
|
|
|
|
|
2022-03-22 14:40:44 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-28 19:37:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-05-14 11:41:56 -07:00
|
|
|
|
2021-06-28 19:37:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-05-14 11:41:56 -07:00
|
|
|
|
2021-06-28 19:37:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-05-02 09:08:38 -07:00
|
|
|
|
2025-04-03 18:39:10 -07:00
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
|
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
|
|
|
|
|
2023-05-02 09:08:38 -07:00
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
2023-05-02 09:08:38 -07:00
|
|
|
|
|
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
|
|
|
|
|
mm/memcg: improve refill_obj_stock() performance
There are two issues with the current refill_obj_stock() code. First of
all, when nr_bytes reaches over PAGE_SIZE, it calls drain_obj_stock() to
atomically flush out remaining bytes to obj_cgroup, clear cached_objcg and
do a obj_cgroup_put(). It is likely that the same obj_cgroup will be used
again which leads to another call to drain_obj_stock() and
obj_cgroup_get() as well as atomically retrieve the available byte from
obj_cgroup. That is costly. Instead, we should just uncharge the excess
pages, reduce the stock bytes and be done with it. The drain_obj_stock()
function should only be called when obj_cgroup changes.
Secondly, when charging an object of size not less than a page in
obj_cgroup_charge(), it is possible that the remaining bytes to be
refilled to the stock will overflow a page and cause refill_obj_stock() to
uncharge 1 page. To avoid the additional uncharge in this case, a new
allow_uncharge flag is added to refill_obj_stock() which will be set to
false when called from obj_cgroup_charge() so that an uncharge_pages()
call won't be issued right after a charge_pages() call unless the objcg
changes.
A multithreaded kmalloc+kfree microbenchmark on a 2-socket 48-core
96-thread x86-64 system with 96 testing threads were run. Before this
patch, the total number of kilo kmalloc+kfree operations done for a 4k
large object by all the testing threads per second were 4,304 kops/s
(cgroup v1) and 8,478 kops/s (cgroup v2). After applying this patch, the
number were 4,731 (cgroup v1) and 418,142 (cgroup v2) respectively. This
represents a performance improvement of 1.10X (cgroup v1) and 49.3X
(cgroup v2).
Link: https://lkml.kernel.org/r/20210506150007.16288-4-longman@redhat.com
Signed-off-by: Waiman Long <longman@redhat.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Alex Shi <alex.shi@linux.alibaba.com>
Cc: Chris Down <chris@chrisdown.name>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Masayoshi Mizuma <msys.mizuma@gmail.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <songmuchun@bytedance.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Roman Gushchin <guro@fb.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Wei Yang <richard.weiyang@gmail.com>
Cc: Xing Zhengjun <zhengjun.xing@linux.intel.com>
Cc: Yafang Shao <laoar.shao@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-28 19:37:27 -07:00
|
|
|
|
2025-04-03 18:39:12 -07:00
|
|
|
|
|
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
mm/memcg: improve refill_obj_stock() performance
There are two issues with the current refill_obj_stock() code. First of
all, when nr_bytes reaches over PAGE_SIZE, it calls drain_obj_stock() to
atomically flush out remaining bytes to obj_cgroup, clear cached_objcg and
do a obj_cgroup_put(). It is likely that the same obj_cgroup will be used
again which leads to another call to drain_obj_stock() and
obj_cgroup_get() as well as atomically retrieve the available byte from
obj_cgroup. That is costly. Instead, we should just uncharge the excess
pages, reduce the stock bytes and be done with it. The drain_obj_stock()
function should only be called when obj_cgroup changes.
Secondly, when charging an object of size not less than a page in
obj_cgroup_charge(), it is possible that the remaining bytes to be
refilled to the stock will overflow a page and cause refill_obj_stock() to
uncharge 1 page. To avoid the additional uncharge in this case, a new
allow_uncharge flag is added to refill_obj_stock() which will be set to
false when called from obj_cgroup_charge() so that an uncharge_pages()
call won't be issued right after a charge_pages() call unless the objcg
changes.
A multithreaded kmalloc+kfree microbenchmark on a 2-socket 48-core
96-thread x86-64 system with 96 testing threads were run. Before this
patch, the total number of kilo kmalloc+kfree operations done for a 4k
large object by all the testing threads per second were 4,304 kops/s
(cgroup v1) and 8,478 kops/s (cgroup v2). After applying this patch, the
number were 4,731 (cgroup v1) and 418,142 (cgroup v2) respectively. This
represents a performance improvement of 1.10X (cgroup v1) and 49.3X
(cgroup v2).
Link: https://lkml.kernel.org/r/20210506150007.16288-4-longman@redhat.com
Signed-off-by: Waiman Long <longman@redhat.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Alex Shi <alex.shi@linux.alibaba.com>
Cc: Chris Down <chris@chrisdown.name>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Masayoshi Mizuma <msys.mizuma@gmail.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <songmuchun@bytedance.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Roman Gushchin <guro@fb.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Wei Yang <richard.weiyang@gmail.com>
Cc: Xing Zhengjun <zhengjun.xing@linux.intel.com>
Cc: Yafang Shao <laoar.shao@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-28 19:37:27 -07:00
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
2025-05-14 11:41:58 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-03-22 14:40:35 -07:00
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
2023-05-02 09:08:38 -07:00
|
|
|
|
2025-04-03 18:39:13 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/memcg: improve refill_obj_stock() performance
There are two issues with the current refill_obj_stock() code. First of
all, when nr_bytes reaches over PAGE_SIZE, it calls drain_obj_stock() to
atomically flush out remaining bytes to obj_cgroup, clear cached_objcg and
do a obj_cgroup_put(). It is likely that the same obj_cgroup will be used
again which leads to another call to drain_obj_stock() and
obj_cgroup_get() as well as atomically retrieve the available byte from
obj_cgroup. That is costly. Instead, we should just uncharge the excess
pages, reduce the stock bytes and be done with it. The drain_obj_stock()
function should only be called when obj_cgroup changes.
Secondly, when charging an object of size not less than a page in
obj_cgroup_charge(), it is possible that the remaining bytes to be
refilled to the stock will overflow a page and cause refill_obj_stock() to
uncharge 1 page. To avoid the additional uncharge in this case, a new
allow_uncharge flag is added to refill_obj_stock() which will be set to
false when called from obj_cgroup_charge() so that an uncharge_pages()
call won't be issued right after a charge_pages() call unless the objcg
changes.
A multithreaded kmalloc+kfree microbenchmark on a 2-socket 48-core
96-thread x86-64 system with 96 testing threads were run. Before this
patch, the total number of kilo kmalloc+kfree operations done for a 4k
large object by all the testing threads per second were 4,304 kops/s
(cgroup v1) and 8,478 kops/s (cgroup v2). After applying this patch, the
number were 4,731 (cgroup v1) and 418,142 (cgroup v2) respectively. This
represents a performance improvement of 1.10X (cgroup v1) and 49.3X
(cgroup v2).
Link: https://lkml.kernel.org/r/20210506150007.16288-4-longman@redhat.com
Signed-off-by: Waiman Long <longman@redhat.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Alex Shi <alex.shi@linux.alibaba.com>
Cc: Chris Down <chris@chrisdown.name>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Masayoshi Mizuma <msys.mizuma@gmail.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <songmuchun@bytedance.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Roman Gushchin <guro@fb.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Wei Yang <richard.weiyang@gmail.com>
Cc: Xing Zhengjun <zhengjun.xing@linux.intel.com>
Cc: Yafang Shao <laoar.shao@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-28 19:37:27 -07:00
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-04-03 18:39:12 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/memcg: improve refill_obj_stock() performance
There are two issues with the current refill_obj_stock() code. First of
all, when nr_bytes reaches over PAGE_SIZE, it calls drain_obj_stock() to
atomically flush out remaining bytes to obj_cgroup, clear cached_objcg and
do a obj_cgroup_put(). It is likely that the same obj_cgroup will be used
again which leads to another call to drain_obj_stock() and
obj_cgroup_get() as well as atomically retrieve the available byte from
obj_cgroup. That is costly. Instead, we should just uncharge the excess
pages, reduce the stock bytes and be done with it. The drain_obj_stock()
function should only be called when obj_cgroup changes.
Secondly, when charging an object of size not less than a page in
obj_cgroup_charge(), it is possible that the remaining bytes to be
refilled to the stock will overflow a page and cause refill_obj_stock() to
uncharge 1 page. To avoid the additional uncharge in this case, a new
allow_uncharge flag is added to refill_obj_stock() which will be set to
false when called from obj_cgroup_charge() so that an uncharge_pages()
call won't be issued right after a charge_pages() call unless the objcg
changes.
A multithreaded kmalloc+kfree microbenchmark on a 2-socket 48-core
96-thread x86-64 system with 96 testing threads were run. Before this
patch, the total number of kilo kmalloc+kfree operations done for a 4k
large object by all the testing threads per second were 4,304 kops/s
(cgroup v1) and 8,478 kops/s (cgroup v2). After applying this patch, the
number were 4,731 (cgroup v1) and 418,142 (cgroup v2) respectively. This
represents a performance improvement of 1.10X (cgroup v1) and 49.3X
(cgroup v2).
Link: https://lkml.kernel.org/r/20210506150007.16288-4-longman@redhat.com
Signed-off-by: Waiman Long <longman@redhat.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Alex Shi <alex.shi@linux.alibaba.com>
Cc: Chris Down <chris@chrisdown.name>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Masayoshi Mizuma <msys.mizuma@gmail.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <songmuchun@bytedance.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Roman Gushchin <guro@fb.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Wei Yang <richard.weiyang@gmail.com>
Cc: Xing Zhengjun <zhengjun.xing@linux.intel.com>
Cc: Yafang Shao <laoar.shao@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-28 19:37:27 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
2025-05-14 11:41:58 -07:00
|
|
|
|
|
|
|
|
|
mm/memcg: improve refill_obj_stock() performance
There are two issues with the current refill_obj_stock() code. First of
all, when nr_bytes reaches over PAGE_SIZE, it calls drain_obj_stock() to
atomically flush out remaining bytes to obj_cgroup, clear cached_objcg and
do a obj_cgroup_put(). It is likely that the same obj_cgroup will be used
again which leads to another call to drain_obj_stock() and
obj_cgroup_get() as well as atomically retrieve the available byte from
obj_cgroup. That is costly. Instead, we should just uncharge the excess
pages, reduce the stock bytes and be done with it. The drain_obj_stock()
function should only be called when obj_cgroup changes.
Secondly, when charging an object of size not less than a page in
obj_cgroup_charge(), it is possible that the remaining bytes to be
refilled to the stock will overflow a page and cause refill_obj_stock() to
uncharge 1 page. To avoid the additional uncharge in this case, a new
allow_uncharge flag is added to refill_obj_stock() which will be set to
false when called from obj_cgroup_charge() so that an uncharge_pages()
call won't be issued right after a charge_pages() call unless the objcg
changes.
A multithreaded kmalloc+kfree microbenchmark on a 2-socket 48-core
96-thread x86-64 system with 96 testing threads were run. Before this
patch, the total number of kilo kmalloc+kfree operations done for a 4k
large object by all the testing threads per second were 4,304 kops/s
(cgroup v1) and 8,478 kops/s (cgroup v2). After applying this patch, the
number were 4,731 (cgroup v1) and 418,142 (cgroup v2) respectively. This
represents a performance improvement of 1.10X (cgroup v1) and 49.3X
(cgroup v2).
Link: https://lkml.kernel.org/r/20210506150007.16288-4-longman@redhat.com
Signed-off-by: Waiman Long <longman@redhat.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Alex Shi <alex.shi@linux.alibaba.com>
Cc: Chris Down <chris@chrisdown.name>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Masayoshi Mizuma <msys.mizuma@gmail.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <songmuchun@bytedance.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Roman Gushchin <guro@fb.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Wei Yang <richard.weiyang@gmail.com>
Cc: Xing Zhengjun <zhengjun.xing@linux.intel.com>
Cc: Yafang Shao <laoar.shao@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-28 19:37:27 -07:00
|
|
|
|
|
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
|
|
|
|
|
2025-04-03 18:39:12 -07:00
|
|
|
|
|
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-04-03 18:39:12 -07:00
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/memcg: improve refill_obj_stock() performance
There are two issues with the current refill_obj_stock() code. First of
all, when nr_bytes reaches over PAGE_SIZE, it calls drain_obj_stock() to
atomically flush out remaining bytes to obj_cgroup, clear cached_objcg and
do a obj_cgroup_put(). It is likely that the same obj_cgroup will be used
again which leads to another call to drain_obj_stock() and
obj_cgroup_get() as well as atomically retrieve the available byte from
obj_cgroup. That is costly. Instead, we should just uncharge the excess
pages, reduce the stock bytes and be done with it. The drain_obj_stock()
function should only be called when obj_cgroup changes.
Secondly, when charging an object of size not less than a page in
obj_cgroup_charge(), it is possible that the remaining bytes to be
refilled to the stock will overflow a page and cause refill_obj_stock() to
uncharge 1 page. To avoid the additional uncharge in this case, a new
allow_uncharge flag is added to refill_obj_stock() which will be set to
false when called from obj_cgroup_charge() so that an uncharge_pages()
call won't be issued right after a charge_pages() call unless the objcg
changes.
A multithreaded kmalloc+kfree microbenchmark on a 2-socket 48-core
96-thread x86-64 system with 96 testing threads were run. Before this
patch, the total number of kilo kmalloc+kfree operations done for a 4k
large object by all the testing threads per second were 4,304 kops/s
(cgroup v1) and 8,478 kops/s (cgroup v2). After applying this patch, the
number were 4,731 (cgroup v1) and 418,142 (cgroup v2) respectively. This
represents a performance improvement of 1.10X (cgroup v1) and 49.3X
(cgroup v2).
Link: https://lkml.kernel.org/r/20210506150007.16288-4-longman@redhat.com
Signed-off-by: Waiman Long <longman@redhat.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Alex Shi <alex.shi@linux.alibaba.com>
Cc: Chris Down <chris@chrisdown.name>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Masayoshi Mizuma <msys.mizuma@gmail.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <songmuchun@bytedance.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Roman Gushchin <guro@fb.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Wei Yang <richard.weiyang@gmail.com>
Cc: Xing Zhengjun <zhengjun.xing@linux.intel.com>
Cc: Yafang Shao <laoar.shao@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-28 19:37:27 -07:00
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
mm/memcg: improve refill_obj_stock() performance
There are two issues with the current refill_obj_stock() code. First of
all, when nr_bytes reaches over PAGE_SIZE, it calls drain_obj_stock() to
atomically flush out remaining bytes to obj_cgroup, clear cached_objcg and
do a obj_cgroup_put(). It is likely that the same obj_cgroup will be used
again which leads to another call to drain_obj_stock() and
obj_cgroup_get() as well as atomically retrieve the available byte from
obj_cgroup. That is costly. Instead, we should just uncharge the excess
pages, reduce the stock bytes and be done with it. The drain_obj_stock()
function should only be called when obj_cgroup changes.
Secondly, when charging an object of size not less than a page in
obj_cgroup_charge(), it is possible that the remaining bytes to be
refilled to the stock will overflow a page and cause refill_obj_stock() to
uncharge 1 page. To avoid the additional uncharge in this case, a new
allow_uncharge flag is added to refill_obj_stock() which will be set to
false when called from obj_cgroup_charge() so that an uncharge_pages()
call won't be issued right after a charge_pages() call unless the objcg
changes.
A multithreaded kmalloc+kfree microbenchmark on a 2-socket 48-core
96-thread x86-64 system with 96 testing threads were run. Before this
patch, the total number of kilo kmalloc+kfree operations done for a 4k
large object by all the testing threads per second were 4,304 kops/s
(cgroup v1) and 8,478 kops/s (cgroup v2). After applying this patch, the
number were 4,731 (cgroup v1) and 418,142 (cgroup v2) respectively. This
represents a performance improvement of 1.10X (cgroup v1) and 49.3X
(cgroup v2).
Link: https://lkml.kernel.org/r/20210506150007.16288-4-longman@redhat.com
Signed-off-by: Waiman Long <longman@redhat.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Alex Shi <alex.shi@linux.alibaba.com>
Cc: Chris Down <chris@chrisdown.name>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Masayoshi Mizuma <msys.mizuma@gmail.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <songmuchun@bytedance.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Roman Gushchin <guro@fb.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Wei Yang <richard.weiyang@gmail.com>
Cc: Xing Zhengjun <zhengjun.xing@linux.intel.com>
Cc: Yafang Shao <laoar.shao@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-28 19:37:27 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-04-29 22:56:42 -07:00
|
|
|
|
2025-04-03 18:39:12 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-04-03 18:39:12 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
|
|
|
|
|
2025-04-03 18:39:12 -07:00
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
|
|
|
|
|
2024-03-26 11:37:39 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-04-03 18:39:12 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-26 11:37:39 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-04-03 18:39:12 -07:00
|
|
|
|
|
|
|
|
|
2024-03-26 11:37:39 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-04-03 18:39:12 -07:00
|
|
|
|
|
|
|
|
|
2024-03-26 11:37:39 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-12-18 14:21:56 -08:00
|
|
|
|
2011-01-20 14:44:24 -08:00
|
|
|
|
2025-03-14 13:36:12 +00:00
|
|
|
|
|
|
|
|
|
2011-01-20 14:44:24 -08:00
|
|
|
|
2025-03-14 13:36:13 +00:00
|
|
|
|
2011-01-20 14:44:24 -08:00
|
|
|
|
2025-03-14 13:36:13 +00:00
|
|
|
|
2025-03-14 13:36:12 +00:00
|
|
|
|
2011-01-20 14:44:24 -08:00
|
|
|
|
2025-03-14 13:36:13 +00:00
|
|
|
|
2011-01-25 15:07:28 -08:00
|
|
|
|
2013-05-07 16:18:09 -07:00
|
|
|
|
2025-03-14 13:36:12 +00:00
|
|
|
|
2025-03-14 13:36:13 +00:00
|
|
|
|
mm: memcontrol: use obj_cgroup APIs to charge kmem pages
Since Roman's series "The new cgroup slab memory controller" applied.
All slab objects are charged via the new APIs of obj_cgroup. The new
APIs introduce a struct obj_cgroup to charge slab objects. It prevents
long-living objects from pinning the original memory cgroup in the
memory. But there are still some corner objects (e.g. allocations
larger than order-1 page on SLUB) which are not charged via the new
APIs. Those objects (include the pages which are allocated from buddy
allocator directly) are charged as kmem pages which still hold a
reference to the memory cgroup.
We want to reuse the obj_cgroup APIs to charge the kmem pages. If we do
that, we should store an object cgroup pointer to page->memcg_data for
the kmem pages.
Finally, page->memcg_data will have 3 different meanings.
1) For the slab pages, page->memcg_data points to an object cgroups
vector.
2) For the kmem pages (exclude the slab pages), page->memcg_data
points to an object cgroup.
3) For the user pages (e.g. the LRU pages), page->memcg_data points
to a memory cgroup.
We do not change the behavior of page_memcg() and page_memcg_rcu(). They
are also suitable for LRU pages and kmem pages. Why?
Because memory allocations pinning memcgs for a long time - it exists at a
larger scale and is causing recurring problems in the real world: page
cache doesn't get reclaimed for a long time, or is used by the second,
third, fourth, ... instance of the same job that was restarted into a new
cgroup every time. Unreclaimable dying cgroups pile up, waste memory, and
make page reclaim very inefficient.
We can convert LRU pages and most other raw memcg pins to the objcg
direction to fix this problem, and then the page->memcg will always point
to an object cgroup pointer. At that time, LRU pages and kmem pages will
be treated the same. The implementation of page_memcg() will remove the
kmem page check.
This patch aims to charge the kmem pages by using the new APIs of
obj_cgroup. Finally, the page->memcg_data of the kmem page points to an
object cgroup. We can use the __page_objcg() to get the object cgroup
associated with a kmem page. Or we can use page_memcg() to get the memory
cgroup associated with a kmem page, but caller must ensure that the
returned memcg won't be released (e.g. acquire the rcu_read_lock or
css_set_lock).
Link: https://lkml.kernel.org/r/20210401030141.37061-1-songmuchun@bytedance.com
Link: https://lkml.kernel.org/r/20210319163821.20704-6-songmuchun@bytedance.com
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Roman Gushchin <guro@fb.com>
Reviewed-by: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Xiongchun Duan <duanxiongchun@bytedance.com>
Cc: Christian Borntraeger <borntraeger@de.ibm.com>
[songmuchun@bytedance.com: fix forget to obtain the ref to objcg in split_page_memcg]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:52 -07:00
|
|
|
|
2025-03-14 13:36:13 +00:00
|
|
|
|
2025-03-14 13:36:11 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-01-20 14:44:24 -08:00
|
|
|
|
|
|
|
|
|
2024-06-24 17:58:58 -07:00
|
|
|
|
2014-09-05 08:43:57 -04:00
|
|
|
|
mm: memcontrol: fix recursive statistics correctness & scalabilty
Right now, when somebody needs to know the recursive memory statistics
and events of a cgroup subtree, they need to walk the entire subtree and
sum up the counters manually.
There are two issues with this:
1. When a cgroup gets deleted, its stats are lost. The state counters
should all be 0 at that point, of course, but the events are not.
When this happens, the event counters, which are supposed to be
monotonic, can go backwards in the parent cgroups.
2. During regular operation, we always have a certain number of lazily
freed cgroups sitting around that have been deleted, have no tasks,
but have a few cache pages remaining. These groups' statistics do not
change until we eventually hit memory pressure, but somebody
watching, say, memory.stat on an ancestor has to iterate those every
time.
This patch addresses both issues by introducing recursive counters at
each level that are propagated from the write side when stats change.
Upward propagation happens when the per-cpu caches spill over into the
local atomic counter. This is the same thing we do during charge and
uncharge, except that the latter uses atomic RMWs, which are more
expensive; stat changes happen at around the same rate. In a sparse
file test (page faults and reclaim at maximum CPU speed) with 5 cgroup
nesting levels, perf shows __mod_memcg_page state at ~1%.
Link: http://lkml.kernel.org/r/20190412151507.2769-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 15:47:12 -07:00
|
|
|
|
2014-09-05 08:43:57 -04:00
|
|
|
|
mm: memcontrol: lockless page counters
Memory is internally accounted in bytes, using spinlock-protected 64-bit
counters, even though the smallest accounting delta is a page. The
counter interface is also convoluted and does too many things.
Introduce a new lockless word-sized page counter API, then change all
memory accounting over to it. The translation from and to bytes then only
happens when interfacing with userspace.
The removed locking overhead is noticable when scaling beyond the per-cpu
charge caches - on a 4-socket machine with 144-threads, the following test
shows the performance differences of 288 memcgs concurrently running a
page fault benchmark:
vanilla:
18631648.500498 task-clock (msec) # 140.643 CPUs utilized ( +- 0.33% )
1,380,638 context-switches # 0.074 K/sec ( +- 0.75% )
24,390 cpu-migrations # 0.001 K/sec ( +- 8.44% )
1,843,305,768 page-faults # 0.099 M/sec ( +- 0.00% )
50,134,994,088,218 cycles # 2.691 GHz ( +- 0.33% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
8,049,712,224,651 instructions # 0.16 insns per cycle ( +- 0.04% )
1,586,970,584,979 branches # 85.176 M/sec ( +- 0.05% )
1,724,989,949 branch-misses # 0.11% of all branches ( +- 0.48% )
132.474343877 seconds time elapsed ( +- 0.21% )
lockless:
12195979.037525 task-clock (msec) # 133.480 CPUs utilized ( +- 0.18% )
832,850 context-switches # 0.068 K/sec ( +- 0.54% )
15,624 cpu-migrations # 0.001 K/sec ( +- 10.17% )
1,843,304,774 page-faults # 0.151 M/sec ( +- 0.00% )
32,811,216,801,141 cycles # 2.690 GHz ( +- 0.18% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
9,999,265,091,727 instructions # 0.30 insns per cycle ( +- 0.10% )
2,076,759,325,203 branches # 170.282 M/sec ( +- 0.12% )
1,656,917,214 branch-misses # 0.08% of all branches ( +- 0.55% )
91.369330729 seconds time elapsed ( +- 0.45% )
On top of improved scalability, this also gets rid of the icky long long
types in the very heart of memcg, which is great for 32 bit and also makes
the code a lot more readable.
Notable differences between the old and new API:
- res_counter_charge() and res_counter_charge_nofail() become
page_counter_try_charge() and page_counter_charge() resp. to match
the more common kernel naming scheme of try_do()/do()
- res_counter_uncharge_until() is only ever used to cancel a local
counter and never to uncharge bigger segments of a hierarchy, so
it's replaced by the simpler page_counter_cancel()
- res_counter_set_limit() is replaced by page_counter_limit(), which
expects its callers to serialize against themselves
- res_counter_memparse_write_strategy() is replaced by
page_counter_limit(), which rounds down to the nearest page size -
rather than up. This is more reasonable for explicitely requested
hard upper limits.
- to keep charging light-weight, page_counter_try_charge() charges
speculatively, only to roll back if the result exceeds the limit.
Because of this, a failing bigger charge can temporarily lock out
smaller charges that would otherwise succeed. The error is bounded
to the difference between the smallest and the biggest possible
charge size, so for memcg, this means that a failing THP charge can
send base page charges into reclaim upto 2MB (4MB) before the limit
would have been reached. This should be acceptable.
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE and memparse]
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE, memparse, strncmp, and PAGE_SIZE]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Vladimir Davydov <vdavydov@parallels.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-12-10 15:42:31 -08:00
|
|
|
|
2023-03-30 19:17:56 +00:00
|
|
|
|
2023-04-21 17:40:18 +00:00
|
|
|
|
|
|
|
|
|
2023-03-30 19:17:56 +00:00
|
|
|
|
2023-04-21 17:40:18 +00:00
|
|
|
|
|
|
|
|
|
mm: memcontrol: fix recursive statistics correctness & scalabilty
Right now, when somebody needs to know the recursive memory statistics
and events of a cgroup subtree, they need to walk the entire subtree and
sum up the counters manually.
There are two issues with this:
1. When a cgroup gets deleted, its stats are lost. The state counters
should all be 0 at that point, of course, but the events are not.
When this happens, the event counters, which are supposed to be
monotonic, can go backwards in the parent cgroups.
2. During regular operation, we always have a certain number of lazily
freed cgroups sitting around that have been deleted, have no tasks,
but have a few cache pages remaining. These groups' statistics do not
change until we eventually hit memory pressure, but somebody
watching, say, memory.stat on an ancestor has to iterate those every
time.
This patch addresses both issues by introducing recursive counters at
each level that are propagated from the write side when stats change.
Upward propagation happens when the per-cpu caches spill over into the
local atomic counter. This is the same thing we do during charge and
uncharge, except that the latter uses atomic RMWs, which are more
expensive; stat changes happen at around the same rate. In a sparse
file test (page faults and reclaim at maximum CPU speed) with 5 cgroup
nesting levels, perf shows __mod_memcg_page state at ~1%.
Link: http://lkml.kernel.org/r/20190412151507.2769-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 15:47:12 -07:00
|
|
|
|
2023-04-21 17:40:18 +00:00
|
|
|
|
mm: memcontrol: lockless page counters
Memory is internally accounted in bytes, using spinlock-protected 64-bit
counters, even though the smallest accounting delta is a page. The
counter interface is also convoluted and does too many things.
Introduce a new lockless word-sized page counter API, then change all
memory accounting over to it. The translation from and to bytes then only
happens when interfacing with userspace.
The removed locking overhead is noticable when scaling beyond the per-cpu
charge caches - on a 4-socket machine with 144-threads, the following test
shows the performance differences of 288 memcgs concurrently running a
page fault benchmark:
vanilla:
18631648.500498 task-clock (msec) # 140.643 CPUs utilized ( +- 0.33% )
1,380,638 context-switches # 0.074 K/sec ( +- 0.75% )
24,390 cpu-migrations # 0.001 K/sec ( +- 8.44% )
1,843,305,768 page-faults # 0.099 M/sec ( +- 0.00% )
50,134,994,088,218 cycles # 2.691 GHz ( +- 0.33% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
8,049,712,224,651 instructions # 0.16 insns per cycle ( +- 0.04% )
1,586,970,584,979 branches # 85.176 M/sec ( +- 0.05% )
1,724,989,949 branch-misses # 0.11% of all branches ( +- 0.48% )
132.474343877 seconds time elapsed ( +- 0.21% )
lockless:
12195979.037525 task-clock (msec) # 133.480 CPUs utilized ( +- 0.18% )
832,850 context-switches # 0.068 K/sec ( +- 0.54% )
15,624 cpu-migrations # 0.001 K/sec ( +- 10.17% )
1,843,304,774 page-faults # 0.151 M/sec ( +- 0.00% )
32,811,216,801,141 cycles # 2.690 GHz ( +- 0.18% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
9,999,265,091,727 instructions # 0.30 insns per cycle ( +- 0.10% )
2,076,759,325,203 branches # 170.282 M/sec ( +- 0.12% )
1,656,917,214 branch-misses # 0.08% of all branches ( +- 0.55% )
91.369330729 seconds time elapsed ( +- 0.45% )
On top of improved scalability, this also gets rid of the icky long long
types in the very heart of memcg, which is great for 32 bit and also makes
the code a lot more readable.
Notable differences between the old and new API:
- res_counter_charge() and res_counter_charge_nofail() become
page_counter_try_charge() and page_counter_charge() resp. to match
the more common kernel naming scheme of try_do()/do()
- res_counter_uncharge_until() is only ever used to cancel a local
counter and never to uncharge bigger segments of a hierarchy, so
it's replaced by the simpler page_counter_cancel()
- res_counter_set_limit() is replaced by page_counter_limit(), which
expects its callers to serialize against themselves
- res_counter_memparse_write_strategy() is replaced by
page_counter_limit(), which rounds down to the nearest page size -
rather than up. This is more reasonable for explicitely requested
hard upper limits.
- to keep charging light-weight, page_counter_try_charge() charges
speculatively, only to roll back if the result exceeds the limit.
Because of this, a failing bigger charge can temporarily lock out
smaller charges that would otherwise succeed. The error is bounded
to the difference between the smallest and the biggest possible
charge size, so for memcg, this means that a failing THP charge can
send base page charges into reclaim upto 2MB (4MB) before the limit
would have been reached. This should be acceptable.
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE and memparse]
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE, memparse, strncmp, and PAGE_SIZE]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Vladimir Davydov <vdavydov@parallels.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-12-10 15:42:31 -08:00
|
|
|
|
2014-09-05 08:43:57 -04:00
|
|
|
|
mm: memcontrol: lockless page counters
Memory is internally accounted in bytes, using spinlock-protected 64-bit
counters, even though the smallest accounting delta is a page. The
counter interface is also convoluted and does too many things.
Introduce a new lockless word-sized page counter API, then change all
memory accounting over to it. The translation from and to bytes then only
happens when interfacing with userspace.
The removed locking overhead is noticable when scaling beyond the per-cpu
charge caches - on a 4-socket machine with 144-threads, the following test
shows the performance differences of 288 memcgs concurrently running a
page fault benchmark:
vanilla:
18631648.500498 task-clock (msec) # 140.643 CPUs utilized ( +- 0.33% )
1,380,638 context-switches # 0.074 K/sec ( +- 0.75% )
24,390 cpu-migrations # 0.001 K/sec ( +- 8.44% )
1,843,305,768 page-faults # 0.099 M/sec ( +- 0.00% )
50,134,994,088,218 cycles # 2.691 GHz ( +- 0.33% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
8,049,712,224,651 instructions # 0.16 insns per cycle ( +- 0.04% )
1,586,970,584,979 branches # 85.176 M/sec ( +- 0.05% )
1,724,989,949 branch-misses # 0.11% of all branches ( +- 0.48% )
132.474343877 seconds time elapsed ( +- 0.21% )
lockless:
12195979.037525 task-clock (msec) # 133.480 CPUs utilized ( +- 0.18% )
832,850 context-switches # 0.068 K/sec ( +- 0.54% )
15,624 cpu-migrations # 0.001 K/sec ( +- 10.17% )
1,843,304,774 page-faults # 0.151 M/sec ( +- 0.00% )
32,811,216,801,141 cycles # 2.690 GHz ( +- 0.18% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
9,999,265,091,727 instructions # 0.30 insns per cycle ( +- 0.10% )
2,076,759,325,203 branches # 170.282 M/sec ( +- 0.12% )
1,656,917,214 branch-misses # 0.08% of all branches ( +- 0.55% )
91.369330729 seconds time elapsed ( +- 0.45% )
On top of improved scalability, this also gets rid of the icky long long
types in the very heart of memcg, which is great for 32 bit and also makes
the code a lot more readable.
Notable differences between the old and new API:
- res_counter_charge() and res_counter_charge_nofail() become
page_counter_try_charge() and page_counter_charge() resp. to match
the more common kernel naming scheme of try_do()/do()
- res_counter_uncharge_until() is only ever used to cancel a local
counter and never to uncharge bigger segments of a hierarchy, so
it's replaced by the simpler page_counter_cancel()
- res_counter_set_limit() is replaced by page_counter_limit(), which
expects its callers to serialize against themselves
- res_counter_memparse_write_strategy() is replaced by
page_counter_limit(), which rounds down to the nearest page size -
rather than up. This is more reasonable for explicitely requested
hard upper limits.
- to keep charging light-weight, page_counter_try_charge() charges
speculatively, only to roll back if the result exceeds the limit.
Because of this, a failing bigger charge can temporarily lock out
smaller charges that would otherwise succeed. The error is bounded
to the difference between the smallest and the biggest possible
charge size, so for memcg, this means that a failing THP charge can
send base page charges into reclaim upto 2MB (4MB) before the limit
would have been reached. This should be acceptable.
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE and memparse]
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE, memparse, strncmp, and PAGE_SIZE]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Vladimir Davydov <vdavydov@parallels.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-12-10 15:42:31 -08:00
|
|
|
|
2014-09-05 08:43:57 -04:00
|
|
|
|
mm: memcontrol: lockless page counters
Memory is internally accounted in bytes, using spinlock-protected 64-bit
counters, even though the smallest accounting delta is a page. The
counter interface is also convoluted and does too many things.
Introduce a new lockless word-sized page counter API, then change all
memory accounting over to it. The translation from and to bytes then only
happens when interfacing with userspace.
The removed locking overhead is noticable when scaling beyond the per-cpu
charge caches - on a 4-socket machine with 144-threads, the following test
shows the performance differences of 288 memcgs concurrently running a
page fault benchmark:
vanilla:
18631648.500498 task-clock (msec) # 140.643 CPUs utilized ( +- 0.33% )
1,380,638 context-switches # 0.074 K/sec ( +- 0.75% )
24,390 cpu-migrations # 0.001 K/sec ( +- 8.44% )
1,843,305,768 page-faults # 0.099 M/sec ( +- 0.00% )
50,134,994,088,218 cycles # 2.691 GHz ( +- 0.33% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
8,049,712,224,651 instructions # 0.16 insns per cycle ( +- 0.04% )
1,586,970,584,979 branches # 85.176 M/sec ( +- 0.05% )
1,724,989,949 branch-misses # 0.11% of all branches ( +- 0.48% )
132.474343877 seconds time elapsed ( +- 0.21% )
lockless:
12195979.037525 task-clock (msec) # 133.480 CPUs utilized ( +- 0.18% )
832,850 context-switches # 0.068 K/sec ( +- 0.54% )
15,624 cpu-migrations # 0.001 K/sec ( +- 10.17% )
1,843,304,774 page-faults # 0.151 M/sec ( +- 0.00% )
32,811,216,801,141 cycles # 2.690 GHz ( +- 0.18% )
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
9,999,265,091,727 instructions # 0.30 insns per cycle ( +- 0.10% )
2,076,759,325,203 branches # 170.282 M/sec ( +- 0.12% )
1,656,917,214 branch-misses # 0.08% of all branches ( +- 0.55% )
91.369330729 seconds time elapsed ( +- 0.45% )
On top of improved scalability, this also gets rid of the icky long long
types in the very heart of memcg, which is great for 32 bit and also makes
the code a lot more readable.
Notable differences between the old and new API:
- res_counter_charge() and res_counter_charge_nofail() become
page_counter_try_charge() and page_counter_charge() resp. to match
the more common kernel naming scheme of try_do()/do()
- res_counter_uncharge_until() is only ever used to cancel a local
counter and never to uncharge bigger segments of a hierarchy, so
it's replaced by the simpler page_counter_cancel()
- res_counter_set_limit() is replaced by page_counter_limit(), which
expects its callers to serialize against themselves
- res_counter_memparse_write_strategy() is replaced by
page_counter_limit(), which rounds down to the nearest page size -
rather than up. This is more reasonable for explicitely requested
hard upper limits.
- to keep charging light-weight, page_counter_try_charge() charges
speculatively, only to roll back if the result exceeds the limit.
Because of this, a failing bigger charge can temporarily lock out
smaller charges that would otherwise succeed. The error is bounded
to the difference between the smallest and the biggest possible
charge size, so for memcg, this means that a failing THP charge can
send base page charges into reclaim upto 2MB (4MB) before the limit
would have been reached. This should be acceptable.
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE and memparse]
[akpm@linux-foundation.org: add includes for WARN_ON_ONCE, memparse, strncmp, and PAGE_SIZE]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Vladimir Davydov <vdavydov@parallels.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-12-10 15:42:31 -08:00
|
|
|
|
2014-09-05 08:43:57 -04:00
|
|
|
|
2015-11-05 18:50:29 -08:00
|
|
|
|
2014-09-05 08:43:57 -04:00
|
|
|
|
|
|
|
|
|
2016-01-20 15:02:24 -08:00
|
|
|
|
2014-01-23 15:53:09 -08:00
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
2014-01-23 15:53:09 -08:00
|
|
|
|
2022-06-25 14:18:44 +08:00
|
|
|
|
mm: memcontrol: enable kmem accounting for all cgroups in the legacy hierarchy
Workingset code was recently made memcg aware, but shadow node shrinker
is still global. As a result, one small cgroup can consume all memory
available for shadow nodes, possibly hurting other cgroups by reclaiming
their shadow nodes, even though reclaim distances stored in its shadow
nodes have no effect. To avoid this, we need to make shadow node
shrinker memcg aware.
The actual work is done in patch 6 of the series. Patches 1 and 2
prepare memcg/shrinker infrastructure for the change. Patch 3 is just a
collateral cleanup. Patch 4 makes radix_tree_node accounted, which is
necessary for making shadow node shrinker memcg aware. Patch 5 reduces
shadow nodes overhead in case workload mostly uses anonymous pages.
This patch:
Currently, in the legacy hierarchy kmem accounting is off for all
cgroups by default and must be enabled explicitly by writing something
to memory.kmem.limit_in_bytes. Since we don't support reclaim on
hitting kmem limit, nor do we have any plans to implement it, this is
likely to be -1, just to enable kmem accounting and limit kernel memory
consumption by the memory.limit_in_bytes along with user memory.
This user API was introduced when the implementation of kmem accounting
lacked slab shrinker support and hence was useless in practice. Things
have changed since then - slab shrinkers were made memcg aware, the
accounting overhead seems to be negligible, and a failure to charge a
kmem allocation should not have critical consequences, because we only
account those kernel objects that should be safe to fail. That's why
kmem accounting is enabled by default for all cgroups in the default
hierarchy, which will eventually replace the legacy one.
The ability to enable kmem accounting for some cgroups while keeping it
disabled for others is getting difficult to maintain. E.g. to make
shadow node shrinker memcg aware (see mm/workingset.c), we need to know
the relationship between the number of shadow nodes allocated for a
cgroup and the size of its lru list. If kmem accounting is enabled for
all cgroups there is no problem, but what should we do if kmem
accounting is enabled only for half of cgroups? We've no other choice
but use global lru stats while scanning root cgroup's shadow nodes, but
that would be wrong if kmem accounting was enabled for all cgroups
(which is the case if the unified hierarchy is used), in which case we
should use lru stats of the root cgroup's lruvec.
That being said, let's enable kmem accounting for all memory cgroups by
default. If one finds it unstable or too costly, it can always be
disabled system-wide by passing cgroup.memory=nokmem to the kernel at
boot time.
Signed-off-by: Vladimir Davydov <vdavydov@virtuozzo.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.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>
2016-03-17 14:18:27 -07:00
|
|
|
|
|
|
|
|
|
2022-03-22 14:41:15 -07:00
|
|
|
|
|
|
|
|
|
2014-01-23 15:53:09 -08:00
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
2022-03-22 14:41:28 -07:00
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
2022-03-22 14:41:28 -07:00
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
|
|
|
|
|
2023-10-19 15:53:43 -07:00
|
|
|
|
|
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
2023-02-13 11:29:22 -08:00
|
|
|
|
2020-08-06 23:20:28 -07:00
|
|
|
|
2022-03-22 14:41:28 -07:00
|
|
|
|
2016-01-20 15:02:53 -08:00
|
|
|
|
|
|
|
|
|
2014-01-23 15:53:09 -08:00
|
|
|
|
|
|
|
|
|
2016-01-20 15:02:26 -08:00
|
|
|
|
|
|
|
|
|
2021-11-05 13:37:53 -07:00
|
|
|
|
2016-01-20 15:02:26 -08:00
|
|
|
|
2022-06-25 14:18:44 +08:00
|
|
|
|
2022-03-22 14:41:15 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-01-20 15:02:26 -08:00
|
|
|
|
2020-08-06 23:21:10 -07:00
|
|
|
|
2016-01-20 15:02:26 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-03-22 14:41:22 -07:00
|
|
|
|
mm/list_lru: split the lock to per-cgroup scope
Currently, every list_lru has a per-node lock that protects adding,
deletion, isolation, and reparenting of all list_lru_one instances
belonging to this list_lru on this node. This lock contention is heavy
when multiple cgroups modify the same list_lru.
This lock can be split into per-cgroup scope to reduce contention.
To achieve this, we need a stable list_lru_one for every cgroup. This
commit adds a lock to each list_lru_one and introduced a helper function
lock_list_lru_of_memcg, making it possible to pin the list_lru of a memcg.
Then reworked the reparenting process.
Reparenting will switch the list_lru_one instances one by one. By locking
each instance and marking it dead using the nr_items counter, reparenting
ensures that all items in the corresponding cgroup (on-list or not,
because items have a stable cgroup, see below) will see the list_lru_one
switch synchronously.
Objcg reparent is also moved after list_lru reparent so items will have a
stable mem cgroup until all list_lru_one instances are drained.
The only caller that doesn't work the *_obj interfaces are direct calls to
list_lru_{add,del}. But it's only used by zswap and that's also based on
objcg, so it's fine.
This also changes the bahaviour of the isolation function when LRU_RETRY
or LRU_REMOVED_RETRY is returned, because now releasing the lock could
unblock reparenting and free the list_lru_one, isolation function will
have to return withoug re-lock the lru.
prepare() {
mkdir /tmp/test-fs
modprobe brd rd_nr=1 rd_size=33554432
mkfs.xfs -f /dev/ram0
mount -t xfs /dev/ram0 /tmp/test-fs
for i in $(seq 1 512); do
mkdir "/tmp/test-fs/$i"
for j in $(seq 1 10240); do
echo TEST-CONTENT > "/tmp/test-fs/$i/$j"
done &
done; wait
}
do_test() {
read_worker() {
sleep 1
tar -cv "$1" &>/dev/null
}
read_in_all() {
cd "/tmp/test-fs" && ls
for i in $(seq 1 512); do
(exec sh -c 'echo "$PPID"') > "/sys/fs/cgroup/benchmark/$i/cgroup.procs"
read_worker "$i" &
done; wait
}
for i in $(seq 1 512); do
mkdir -p "/sys/fs/cgroup/benchmark/$i"
done
echo +memory > /sys/fs/cgroup/benchmark/cgroup.subtree_control
echo 512M > /sys/fs/cgroup/benchmark/memory.max
echo 3 > /proc/sys/vm/drop_caches
time read_in_all
}
Above script simulates compression of small files in multiple cgroups
with memory pressure. Run prepare() then do_test for 6 times:
Before:
real 0m7.762s user 0m11.340s sys 3m11.224s
real 0m8.123s user 0m11.548s sys 3m2.549s
real 0m7.736s user 0m11.515s sys 3m11.171s
real 0m8.539s user 0m11.508s sys 3m7.618s
real 0m7.928s user 0m11.349s sys 3m13.063s
real 0m8.105s user 0m11.128s sys 3m14.313s
After this commit (about ~15% faster):
real 0m6.953s user 0m11.327s sys 2m42.912s
real 0m7.453s user 0m11.343s sys 2m51.942s
real 0m6.916s user 0m11.269s sys 2m43.957s
real 0m6.894s user 0m11.528s sys 2m45.346s
real 0m6.911s user 0m11.095s sys 2m43.168s
real 0m6.773s user 0m11.518s sys 2m40.774s
Link: https://lkml.kernel.org/r/20241104175257.60853-6-ryncsn@gmail.com
Signed-off-by: Kairui Song <kasong@tencent.com>
Cc: Chengming Zhou <zhouchengming@bytedance.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Waiman Long <longman@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-11-05 01:52:56 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-01-20 15:02:26 -08:00
|
|
|
|
2016-01-20 15:02:32 -08:00
|
|
|
|
2015-05-22 17:13:37 -04:00
|
|
|
|
|
|
|
|
|
2019-08-29 15:47:19 -07:00
|
|
|
|
|
|
|
|
|
2015-05-22 18:23:33 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-05-22 18:23:34 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-05-22 18:23:33 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-05-22 18:23:35 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
writeback: fix incorrect calculation of available memory for memcg domains
For memcg domains, the amount of available memory was calculated as
min(the amount currently in use + headroom according to memcg,
total clean memory)
This isn't quite correct as what should be capped by the amount of
clean memory is the headroom, not the sum of memory in use and
headroom. For example, if a memcg domain has a significant amount of
dirty memory, the above can lead to a value which is lower than the
current amount in use which doesn't make much sense. In most
circumstances, the above leads to a number which is somewhat but not
drastically lower.
As the amount of memory which can be readily allocated to the memcg
domain is capped by the amount of system-wide clean memory which is
not already assigned to the memcg itself, the number we want is
the amount currently in use +
min(headroom according to memcg, clean memory elsewhere in the system)
This patch updates mem_cgroup_wb_stats() to return the number of
filepages and headroom instead of the calculated available pages.
mdtc_cap_avail() is renamed to mdtc_calc_avail() and performs the
above calculation from file, headroom, dirty and globally clean pages.
v2: Dummy mem_cgroup_wb_stats() implementation wasn't updated leading
to build failure when !CGROUP_WRITEBACK. Fixed.
Signed-off-by: Tejun Heo <tj@kernel.org>
Fixes: c2aa723a6093 ("writeback: implement memcg writeback domain based throttling")
Signed-off-by: Jens Axboe <axboe@fb.com>
2015-09-29 13:04:26 -04:00
|
|
|
|
|
|
|
|
|
2015-05-22 18:23:35 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
writeback: fix incorrect calculation of available memory for memcg domains
For memcg domains, the amount of available memory was calculated as
min(the amount currently in use + headroom according to memcg,
total clean memory)
This isn't quite correct as what should be capped by the amount of
clean memory is the headroom, not the sum of memory in use and
headroom. For example, if a memcg domain has a significant amount of
dirty memory, the above can lead to a value which is lower than the
current amount in use which doesn't make much sense. In most
circumstances, the above leads to a number which is somewhat but not
drastically lower.
As the amount of memory which can be readily allocated to the memcg
domain is capped by the amount of system-wide clean memory which is
not already assigned to the memcg itself, the number we want is
the amount currently in use +
min(headroom according to memcg, clean memory elsewhere in the system)
This patch updates mem_cgroup_wb_stats() to return the number of
filepages and headroom instead of the calculated available pages.
mdtc_cap_avail() is renamed to mdtc_calc_avail() and performs the
above calculation from file, headroom, dirty and globally clean pages.
v2: Dummy mem_cgroup_wb_stats() implementation wasn't updated leading
to build failure when !CGROUP_WRITEBACK. Fixed.
Signed-off-by: Tejun Heo <tj@kernel.org>
Fixes: c2aa723a6093 ("writeback: implement memcg writeback domain based throttling")
Signed-off-by: Jens Axboe <axboe@fb.com>
2015-09-29 13:04:26 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-05-22 18:23:35 -04:00
|
|
|
|
writeback: fix incorrect calculation of available memory for memcg domains
For memcg domains, the amount of available memory was calculated as
min(the amount currently in use + headroom according to memcg,
total clean memory)
This isn't quite correct as what should be capped by the amount of
clean memory is the headroom, not the sum of memory in use and
headroom. For example, if a memcg domain has a significant amount of
dirty memory, the above can lead to a value which is lower than the
current amount in use which doesn't make much sense. In most
circumstances, the above leads to a number which is somewhat but not
drastically lower.
As the amount of memory which can be readily allocated to the memcg
domain is capped by the amount of system-wide clean memory which is
not already assigned to the memcg itself, the number we want is
the amount currently in use +
min(headroom according to memcg, clean memory elsewhere in the system)
This patch updates mem_cgroup_wb_stats() to return the number of
filepages and headroom instead of the calculated available pages.
mdtc_cap_avail() is renamed to mdtc_calc_avail() and performs the
above calculation from file, headroom, dirty and globally clean pages.
v2: Dummy mem_cgroup_wb_stats() implementation wasn't updated leading
to build failure when !CGROUP_WRITEBACK. Fixed.
Signed-off-by: Tejun Heo <tj@kernel.org>
Fixes: c2aa723a6093 ("writeback: implement memcg writeback domain based throttling")
Signed-off-by: Jens Axboe <axboe@fb.com>
2015-09-29 13:04:26 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-05-22 18:23:35 -04:00
|
|
|
|
writeback: fix incorrect calculation of available memory for memcg domains
For memcg domains, the amount of available memory was calculated as
min(the amount currently in use + headroom according to memcg,
total clean memory)
This isn't quite correct as what should be capped by the amount of
clean memory is the headroom, not the sum of memory in use and
headroom. For example, if a memcg domain has a significant amount of
dirty memory, the above can lead to a value which is lower than the
current amount in use which doesn't make much sense. In most
circumstances, the above leads to a number which is somewhat but not
drastically lower.
As the amount of memory which can be readily allocated to the memcg
domain is capped by the amount of system-wide clean memory which is
not already assigned to the memcg itself, the number we want is
the amount currently in use +
min(headroom according to memcg, clean memory elsewhere in the system)
This patch updates mem_cgroup_wb_stats() to return the number of
filepages and headroom instead of the calculated available pages.
mdtc_cap_avail() is renamed to mdtc_calc_avail() and performs the
above calculation from file, headroom, dirty and globally clean pages.
v2: Dummy mem_cgroup_wb_stats() implementation wasn't updated leading
to build failure when !CGROUP_WRITEBACK. Fixed.
Signed-off-by: Tejun Heo <tj@kernel.org>
Fixes: c2aa723a6093 ("writeback: implement memcg writeback domain based throttling")
Signed-off-by: Jens Axboe <axboe@fb.com>
2015-09-29 13:04:26 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-05-22 18:23:35 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-01-18 18:42:35 +00:00
|
|
|
|
2015-05-22 18:23:35 -04:00
|
|
|
|
mm: memcontrol: switch to rstat
Replace the memory controller's custom hierarchical stats code with the
generic rstat infrastructure provided by the cgroup core.
The current implementation does batched upward propagation from the
write side (i.e. as stats change). The per-cpu batches introduce an
error, which is multiplied by the number of subgroups in a tree. In
systems with many CPUs and sizable cgroup trees, the error can be large
enough to confuse users (e.g. 32 batch pages * 32 CPUs * 32 subgroups
results in an error of up to 128M per stat item). This can entirely
swallow allocation bursts inside a workload that the user is expecting
to see reflected in the statistics.
In the past, we've done read-side aggregation, where a memory.stat read
would have to walk the entire subtree and add up per-cpu counts. This
became problematic with lazily-freed cgroups: we could have large
subtrees where most cgroups were entirely idle. Hence the switch to
change-driven upward propagation. Unfortunately, it needed to trade
accuracy for speed due to the write side being so hot.
Rstat combines the best of both worlds: from the write side, it cheaply
maintains a queue of cgroups that have pending changes, so that the read
side can do selective tree aggregation. This way the reported stats
will always be precise and recent as can be, while the aggregation can
skip over potentially large numbers of idle cgroups.
The way rstat works is that it implements a tree for tracking cgroups
with pending local changes, as well as a flush function that walks the
tree upwards. The controller then drives this by 1) telling rstat when
a local cgroup stat changes (e.g. mod_memcg_state) and 2) when a flush
is required to get uptodate hierarchy stats for a given subtree (e.g.
when memory.stat is read). The controller also provides a flush
callback that is called during the rstat flush walk for each cgroup and
aggregates its local per-cpu counters and propagates them upwards.
This adds a second vmstats to struct mem_cgroup (MEMCG_NR_STAT +
NR_VM_EVENT_ITEMS) to track pending subtree deltas during upward
aggregation. It removes 3 words from the per-cpu data. It eliminates
memcg_exact_page_state(), since memcg_page_state() is now exact.
[akpm@linux-foundation.org: merge fix]
[hannes@cmpxchg.org: fix a sleep in atomic section problem]
Link: https://lkml.kernel.org/r/20210315234100.64307-1-hannes@cmpxchg.org
Link: https://lkml.kernel.org/r/20210209163304.77088-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Roman Gushchin <guro@fb.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Balbir Singh <bsingharora@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:26 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-05-22 18:23:35 -04:00
|
|
|
|
mm: memcontrol: switch to rstat
Replace the memory controller's custom hierarchical stats code with the
generic rstat infrastructure provided by the cgroup core.
The current implementation does batched upward propagation from the
write side (i.e. as stats change). The per-cpu batches introduce an
error, which is multiplied by the number of subgroups in a tree. In
systems with many CPUs and sizable cgroup trees, the error can be large
enough to confuse users (e.g. 32 batch pages * 32 CPUs * 32 subgroups
results in an error of up to 128M per stat item). This can entirely
swallow allocation bursts inside a workload that the user is expecting
to see reflected in the statistics.
In the past, we've done read-side aggregation, where a memory.stat read
would have to walk the entire subtree and add up per-cpu counts. This
became problematic with lazily-freed cgroups: we could have large
subtrees where most cgroups were entirely idle. Hence the switch to
change-driven upward propagation. Unfortunately, it needed to trade
accuracy for speed due to the write side being so hot.
Rstat combines the best of both worlds: from the write side, it cheaply
maintains a queue of cgroups that have pending changes, so that the read
side can do selective tree aggregation. This way the reported stats
will always be precise and recent as can be, while the aggregation can
skip over potentially large numbers of idle cgroups.
The way rstat works is that it implements a tree for tracking cgroups
with pending local changes, as well as a flush function that walks the
tree upwards. The controller then drives this by 1) telling rstat when
a local cgroup stat changes (e.g. mod_memcg_state) and 2) when a flush
is required to get uptodate hierarchy stats for a given subtree (e.g.
when memory.stat is read). The controller also provides a flush
callback that is called during the rstat flush walk for each cgroup and
aggregates its local per-cpu counters and propagates them upwards.
This adds a second vmstats to struct mem_cgroup (MEMCG_NR_STAT +
NR_VM_EVENT_ITEMS) to track pending subtree deltas during upward
aggregation. It removes 3 words from the per-cpu data. It eliminates
memcg_exact_page_state(), since memcg_page_state() is now exact.
[akpm@linux-foundation.org: merge fix]
[hannes@cmpxchg.org: fix a sleep in atomic section problem]
Link: https://lkml.kernel.org/r/20210315234100.64307-1-hannes@cmpxchg.org
Link: https://lkml.kernel.org/r/20210209163304.77088-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Roman Gushchin <guro@fb.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Balbir Singh <bsingharora@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:26 -07:00
|
|
|
|
2015-05-22 18:23:35 -04:00
|
|
|
|
2020-04-01 21:07:20 -07:00
|
|
|
|
2020-06-01 21:49:49 -07:00
|
|
|
|
2015-05-22 18:23:35 -04:00
|
|
|
|
|
|
|
|
|
writeback: fix incorrect calculation of available memory for memcg domains
For memcg domains, the amount of available memory was calculated as
min(the amount currently in use + headroom according to memcg,
total clean memory)
This isn't quite correct as what should be capped by the amount of
clean memory is the headroom, not the sum of memory in use and
headroom. For example, if a memcg domain has a significant amount of
dirty memory, the above can lead to a value which is lower than the
current amount in use which doesn't make much sense. In most
circumstances, the above leads to a number which is somewhat but not
drastically lower.
As the amount of memory which can be readily allocated to the memcg
domain is capped by the amount of system-wide clean memory which is
not already assigned to the memcg itself, the number we want is
the amount currently in use +
min(headroom according to memcg, clean memory elsewhere in the system)
This patch updates mem_cgroup_wb_stats() to return the number of
filepages and headroom instead of the calculated available pages.
mdtc_cap_avail() is renamed to mdtc_calc_avail() and performs the
above calculation from file, headroom, dirty and globally clean pages.
v2: Dummy mem_cgroup_wb_stats() implementation wasn't updated leading
to build failure when !CGROUP_WRITEBACK. Fixed.
Signed-off-by: Tejun Heo <tj@kernel.org>
Fixes: c2aa723a6093 ("writeback: implement memcg writeback domain based throttling")
Signed-off-by: Jens Axboe <axboe@fb.com>
2015-09-29 13:04:26 -04:00
|
|
|
|
2015-05-22 18:23:35 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
writeback, memcg: Implement foreign dirty flushing
There's an inherent mismatch between memcg and writeback. The former
trackes ownership per-page while the latter per-inode. This was a
deliberate design decision because honoring per-page ownership in the
writeback path is complicated, may lead to higher CPU and IO overheads
and deemed unnecessary given that write-sharing an inode across
different cgroups isn't a common use-case.
Combined with inode majority-writer ownership switching, this works
well enough in most cases but there are some pathological cases. For
example, let's say there are two cgroups A and B which keep writing to
different but confined parts of the same inode. B owns the inode and
A's memory is limited far below B's. A's dirty ratio can rise enough
to trigger balance_dirty_pages() sleeps but B's can be low enough to
avoid triggering background writeback. A will be slowed down without
a way to make writeback of the dirty pages happen.
This patch implements foreign dirty recording and foreign mechanism so
that when a memcg encounters a condition as above it can trigger
flushes on bdi_writebacks which can clean its pages. Please see the
comment on top of mem_cgroup_track_foreign_dirty_slowpath() for
details.
A reproducer follows.
write-range.c::
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <fcntl.h>
#include <sys/types.h>
static const char *usage = "write-range FILE START SIZE\n";
int main(int argc, char **argv)
{
int fd;
unsigned long start, size, end, pos;
char *endp;
char buf[4096];
if (argc < 4) {
fprintf(stderr, usage);
return 1;
}
fd = open(argv[1], O_WRONLY);
if (fd < 0) {
perror("open");
return 1;
}
start = strtoul(argv[2], &endp, 0);
if (*endp != '\0') {
fprintf(stderr, usage);
return 1;
}
size = strtoul(argv[3], &endp, 0);
if (*endp != '\0') {
fprintf(stderr, usage);
return 1;
}
end = start + size;
while (1) {
for (pos = start; pos < end; ) {
long bread, bwritten = 0;
if (lseek(fd, pos, SEEK_SET) < 0) {
perror("lseek");
return 1;
}
bread = read(0, buf, sizeof(buf) < end - pos ?
sizeof(buf) : end - pos);
if (bread < 0) {
perror("read");
return 1;
}
if (bread == 0)
return 0;
while (bwritten < bread) {
long this;
this = write(fd, buf + bwritten,
bread - bwritten);
if (this < 0) {
perror("write");
return 1;
}
bwritten += this;
pos += bwritten;
}
}
}
}
repro.sh::
#!/bin/bash
set -e
set -x
sysctl -w vm.dirty_expire_centisecs=300000
sysctl -w vm.dirty_writeback_centisecs=300000
sysctl -w vm.dirtytime_expire_seconds=300000
echo 3 > /proc/sys/vm/drop_caches
TEST=/sys/fs/cgroup/test
A=$TEST/A
B=$TEST/B
mkdir -p $A $B
echo "+memory +io" > $TEST/cgroup.subtree_control
echo $((1<<30)) > $A/memory.high
echo $((32<<30)) > $B/memory.high
rm -f testfile
touch testfile
fallocate -l 4G testfile
echo "Starting B"
(echo $BASHPID > $B/cgroup.procs
pv -q --rate-limit 70M < /dev/urandom | ./write-range testfile $((2<<30)) $((2<<30))) &
echo "Waiting 10s to ensure B claims the testfile inode"
sleep 5
sync
sleep 5
sync
echo "Starting A"
(echo $BASHPID > $A/cgroup.procs
pv < /dev/urandom | ./write-range testfile 0 $((2<<30)))
v2: Added comments explaining why the specific intervals are being used.
v3: Use 0 @nr when calling cgroup_writeback_by_id() to use best-effort
flushing while avoding possible livelocks.
v4: Use get_jiffies_64() and time_before/after64() instead of raw
jiffies_64 and arthimetic comparisons as suggested by Jan.
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-08-26 09:06:56 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-05-06 18:06:47 -07:00
|
|
|
|
writeback, memcg: Implement foreign dirty flushing
There's an inherent mismatch between memcg and writeback. The former
trackes ownership per-page while the latter per-inode. This was a
deliberate design decision because honoring per-page ownership in the
writeback path is complicated, may lead to higher CPU and IO overheads
and deemed unnecessary given that write-sharing an inode across
different cgroups isn't a common use-case.
Combined with inode majority-writer ownership switching, this works
well enough in most cases but there are some pathological cases. For
example, let's say there are two cgroups A and B which keep writing to
different but confined parts of the same inode. B owns the inode and
A's memory is limited far below B's. A's dirty ratio can rise enough
to trigger balance_dirty_pages() sleeps but B's can be low enough to
avoid triggering background writeback. A will be slowed down without
a way to make writeback of the dirty pages happen.
This patch implements foreign dirty recording and foreign mechanism so
that when a memcg encounters a condition as above it can trigger
flushes on bdi_writebacks which can clean its pages. Please see the
comment on top of mem_cgroup_track_foreign_dirty_slowpath() for
details.
A reproducer follows.
write-range.c::
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <fcntl.h>
#include <sys/types.h>
static const char *usage = "write-range FILE START SIZE\n";
int main(int argc, char **argv)
{
int fd;
unsigned long start, size, end, pos;
char *endp;
char buf[4096];
if (argc < 4) {
fprintf(stderr, usage);
return 1;
}
fd = open(argv[1], O_WRONLY);
if (fd < 0) {
perror("open");
return 1;
}
start = strtoul(argv[2], &endp, 0);
if (*endp != '\0') {
fprintf(stderr, usage);
return 1;
}
size = strtoul(argv[3], &endp, 0);
if (*endp != '\0') {
fprintf(stderr, usage);
return 1;
}
end = start + size;
while (1) {
for (pos = start; pos < end; ) {
long bread, bwritten = 0;
if (lseek(fd, pos, SEEK_SET) < 0) {
perror("lseek");
return 1;
}
bread = read(0, buf, sizeof(buf) < end - pos ?
sizeof(buf) : end - pos);
if (bread < 0) {
perror("read");
return 1;
}
if (bread == 0)
return 0;
while (bwritten < bread) {
long this;
this = write(fd, buf + bwritten,
bread - bwritten);
if (this < 0) {
perror("write");
return 1;
}
bwritten += this;
pos += bwritten;
}
}
}
}
repro.sh::
#!/bin/bash
set -e
set -x
sysctl -w vm.dirty_expire_centisecs=300000
sysctl -w vm.dirty_writeback_centisecs=300000
sysctl -w vm.dirtytime_expire_seconds=300000
echo 3 > /proc/sys/vm/drop_caches
TEST=/sys/fs/cgroup/test
A=$TEST/A
B=$TEST/B
mkdir -p $A $B
echo "+memory +io" > $TEST/cgroup.subtree_control
echo $((1<<30)) > $A/memory.high
echo $((32<<30)) > $B/memory.high
rm -f testfile
touch testfile
fallocate -l 4G testfile
echo "Starting B"
(echo $BASHPID > $B/cgroup.procs
pv -q --rate-limit 70M < /dev/urandom | ./write-range testfile $((2<<30)) $((2<<30))) &
echo "Waiting 10s to ensure B claims the testfile inode"
sleep 5
sync
sleep 5
sync
echo "Starting A"
(echo $BASHPID > $A/cgroup.procs
pv < /dev/urandom | ./write-range testfile 0 $((2<<30)))
v2: Added comments explaining why the specific intervals are being used.
v3: Use 0 @nr when calling cgroup_writeback_by_id() to use best-effort
flushing while avoding possible livelocks.
v4: Use get_jiffies_64() and time_before/after64() instead of raw
jiffies_64 and arthimetic comparisons as suggested by Jan.
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-08-26 09:06:56 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-05-06 18:06:47 -07:00
|
|
|
|
writeback, memcg: Implement foreign dirty flushing
There's an inherent mismatch between memcg and writeback. The former
trackes ownership per-page while the latter per-inode. This was a
deliberate design decision because honoring per-page ownership in the
writeback path is complicated, may lead to higher CPU and IO overheads
and deemed unnecessary given that write-sharing an inode across
different cgroups isn't a common use-case.
Combined with inode majority-writer ownership switching, this works
well enough in most cases but there are some pathological cases. For
example, let's say there are two cgroups A and B which keep writing to
different but confined parts of the same inode. B owns the inode and
A's memory is limited far below B's. A's dirty ratio can rise enough
to trigger balance_dirty_pages() sleeps but B's can be low enough to
avoid triggering background writeback. A will be slowed down without
a way to make writeback of the dirty pages happen.
This patch implements foreign dirty recording and foreign mechanism so
that when a memcg encounters a condition as above it can trigger
flushes on bdi_writebacks which can clean its pages. Please see the
comment on top of mem_cgroup_track_foreign_dirty_slowpath() for
details.
A reproducer follows.
write-range.c::
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <fcntl.h>
#include <sys/types.h>
static const char *usage = "write-range FILE START SIZE\n";
int main(int argc, char **argv)
{
int fd;
unsigned long start, size, end, pos;
char *endp;
char buf[4096];
if (argc < 4) {
fprintf(stderr, usage);
return 1;
}
fd = open(argv[1], O_WRONLY);
if (fd < 0) {
perror("open");
return 1;
}
start = strtoul(argv[2], &endp, 0);
if (*endp != '\0') {
fprintf(stderr, usage);
return 1;
}
size = strtoul(argv[3], &endp, 0);
if (*endp != '\0') {
fprintf(stderr, usage);
return 1;
}
end = start + size;
while (1) {
for (pos = start; pos < end; ) {
long bread, bwritten = 0;
if (lseek(fd, pos, SEEK_SET) < 0) {
perror("lseek");
return 1;
}
bread = read(0, buf, sizeof(buf) < end - pos ?
sizeof(buf) : end - pos);
if (bread < 0) {
perror("read");
return 1;
}
if (bread == 0)
return 0;
while (bwritten < bread) {
long this;
this = write(fd, buf + bwritten,
bread - bwritten);
if (this < 0) {
perror("write");
return 1;
}
bwritten += this;
pos += bwritten;
}
}
}
}
repro.sh::
#!/bin/bash
set -e
set -x
sysctl -w vm.dirty_expire_centisecs=300000
sysctl -w vm.dirty_writeback_centisecs=300000
sysctl -w vm.dirtytime_expire_seconds=300000
echo 3 > /proc/sys/vm/drop_caches
TEST=/sys/fs/cgroup/test
A=$TEST/A
B=$TEST/B
mkdir -p $A $B
echo "+memory +io" > $TEST/cgroup.subtree_control
echo $((1<<30)) > $A/memory.high
echo $((32<<30)) > $B/memory.high
rm -f testfile
touch testfile
fallocate -l 4G testfile
echo "Starting B"
(echo $BASHPID > $B/cgroup.procs
pv -q --rate-limit 70M < /dev/urandom | ./write-range testfile $((2<<30)) $((2<<30))) &
echo "Waiting 10s to ensure B claims the testfile inode"
sleep 5
sync
sleep 5
sync
echo "Starting A"
(echo $BASHPID > $A/cgroup.procs
pv < /dev/urandom | ./write-range testfile 0 $((2<<30)))
v2: Added comments explaining why the specific intervals are being used.
v3: Use 0 @nr when calling cgroup_writeback_by_id() to use best-effort
flushing while avoding possible livelocks.
v4: Use get_jiffies_64() and time_before/after64() instead of raw
jiffies_64 and arthimetic comparisons as suggested by Jan.
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-08-26 09:06:56 -07:00
|
|
|
|
2021-05-06 18:06:47 -07:00
|
|
|
|
writeback, memcg: Implement foreign dirty flushing
There's an inherent mismatch between memcg and writeback. The former
trackes ownership per-page while the latter per-inode. This was a
deliberate design decision because honoring per-page ownership in the
writeback path is complicated, may lead to higher CPU and IO overheads
and deemed unnecessary given that write-sharing an inode across
different cgroups isn't a common use-case.
Combined with inode majority-writer ownership switching, this works
well enough in most cases but there are some pathological cases. For
example, let's say there are two cgroups A and B which keep writing to
different but confined parts of the same inode. B owns the inode and
A's memory is limited far below B's. A's dirty ratio can rise enough
to trigger balance_dirty_pages() sleeps but B's can be low enough to
avoid triggering background writeback. A will be slowed down without
a way to make writeback of the dirty pages happen.
This patch implements foreign dirty recording and foreign mechanism so
that when a memcg encounters a condition as above it can trigger
flushes on bdi_writebacks which can clean its pages. Please see the
comment on top of mem_cgroup_track_foreign_dirty_slowpath() for
details.
A reproducer follows.
write-range.c::
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <fcntl.h>
#include <sys/types.h>
static const char *usage = "write-range FILE START SIZE\n";
int main(int argc, char **argv)
{
int fd;
unsigned long start, size, end, pos;
char *endp;
char buf[4096];
if (argc < 4) {
fprintf(stderr, usage);
return 1;
}
fd = open(argv[1], O_WRONLY);
if (fd < 0) {
perror("open");
return 1;
}
start = strtoul(argv[2], &endp, 0);
if (*endp != '\0') {
fprintf(stderr, usage);
return 1;
}
size = strtoul(argv[3], &endp, 0);
if (*endp != '\0') {
fprintf(stderr, usage);
return 1;
}
end = start + size;
while (1) {
for (pos = start; pos < end; ) {
long bread, bwritten = 0;
if (lseek(fd, pos, SEEK_SET) < 0) {
perror("lseek");
return 1;
}
bread = read(0, buf, sizeof(buf) < end - pos ?
sizeof(buf) : end - pos);
if (bread < 0) {
perror("read");
return 1;
}
if (bread == 0)
return 0;
while (bwritten < bread) {
long this;
this = write(fd, buf + bwritten,
bread - bwritten);
if (this < 0) {
perror("write");
return 1;
}
bwritten += this;
pos += bwritten;
}
}
}
}
repro.sh::
#!/bin/bash
set -e
set -x
sysctl -w vm.dirty_expire_centisecs=300000
sysctl -w vm.dirty_writeback_centisecs=300000
sysctl -w vm.dirtytime_expire_seconds=300000
echo 3 > /proc/sys/vm/drop_caches
TEST=/sys/fs/cgroup/test
A=$TEST/A
B=$TEST/B
mkdir -p $A $B
echo "+memory +io" > $TEST/cgroup.subtree_control
echo $((1<<30)) > $A/memory.high
echo $((32<<30)) > $B/memory.high
rm -f testfile
touch testfile
fallocate -l 4G testfile
echo "Starting B"
(echo $BASHPID > $B/cgroup.procs
pv -q --rate-limit 70M < /dev/urandom | ./write-range testfile $((2<<30)) $((2<<30))) &
echo "Waiting 10s to ensure B claims the testfile inode"
sleep 5
sync
sleep 5
sync
echo "Starting A"
(echo $BASHPID > $A/cgroup.procs
pv < /dev/urandom | ./write-range testfile 0 $((2<<30)))
v2: Added comments explaining why the specific intervals are being used.
v3: Use 0 @nr when calling cgroup_writeback_by_id() to use best-effort
flushing while avoding possible livelocks.
v4: Use get_jiffies_64() and time_before/after64() instead of raw
jiffies_64 and arthimetic comparisons as suggested by Jan.
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-08-26 09:06:56 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-05-04 11:43:01 -04:00
|
|
|
|
writeback, memcg: Implement foreign dirty flushing
There's an inherent mismatch between memcg and writeback. The former
trackes ownership per-page while the latter per-inode. This was a
deliberate design decision because honoring per-page ownership in the
writeback path is complicated, may lead to higher CPU and IO overheads
and deemed unnecessary given that write-sharing an inode across
different cgroups isn't a common use-case.
Combined with inode majority-writer ownership switching, this works
well enough in most cases but there are some pathological cases. For
example, let's say there are two cgroups A and B which keep writing to
different but confined parts of the same inode. B owns the inode and
A's memory is limited far below B's. A's dirty ratio can rise enough
to trigger balance_dirty_pages() sleeps but B's can be low enough to
avoid triggering background writeback. A will be slowed down without
a way to make writeback of the dirty pages happen.
This patch implements foreign dirty recording and foreign mechanism so
that when a memcg encounters a condition as above it can trigger
flushes on bdi_writebacks which can clean its pages. Please see the
comment on top of mem_cgroup_track_foreign_dirty_slowpath() for
details.
A reproducer follows.
write-range.c::
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <fcntl.h>
#include <sys/types.h>
static const char *usage = "write-range FILE START SIZE\n";
int main(int argc, char **argv)
{
int fd;
unsigned long start, size, end, pos;
char *endp;
char buf[4096];
if (argc < 4) {
fprintf(stderr, usage);
return 1;
}
fd = open(argv[1], O_WRONLY);
if (fd < 0) {
perror("open");
return 1;
}
start = strtoul(argv[2], &endp, 0);
if (*endp != '\0') {
fprintf(stderr, usage);
return 1;
}
size = strtoul(argv[3], &endp, 0);
if (*endp != '\0') {
fprintf(stderr, usage);
return 1;
}
end = start + size;
while (1) {
for (pos = start; pos < end; ) {
long bread, bwritten = 0;
if (lseek(fd, pos, SEEK_SET) < 0) {
perror("lseek");
return 1;
}
bread = read(0, buf, sizeof(buf) < end - pos ?
sizeof(buf) : end - pos);
if (bread < 0) {
perror("read");
return 1;
}
if (bread == 0)
return 0;
while (bwritten < bread) {
long this;
this = write(fd, buf + bwritten,
bread - bwritten);
if (this < 0) {
perror("write");
return 1;
}
bwritten += this;
pos += bwritten;
}
}
}
}
repro.sh::
#!/bin/bash
set -e
set -x
sysctl -w vm.dirty_expire_centisecs=300000
sysctl -w vm.dirty_writeback_centisecs=300000
sysctl -w vm.dirtytime_expire_seconds=300000
echo 3 > /proc/sys/vm/drop_caches
TEST=/sys/fs/cgroup/test
A=$TEST/A
B=$TEST/B
mkdir -p $A $B
echo "+memory +io" > $TEST/cgroup.subtree_control
echo $((1<<30)) > $A/memory.high
echo $((32<<30)) > $B/memory.high
rm -f testfile
touch testfile
fallocate -l 4G testfile
echo "Starting B"
(echo $BASHPID > $B/cgroup.procs
pv -q --rate-limit 70M < /dev/urandom | ./write-range testfile $((2<<30)) $((2<<30))) &
echo "Waiting 10s to ensure B claims the testfile inode"
sleep 5
sync
sleep 5
sync
echo "Starting A"
(echo $BASHPID > $A/cgroup.procs
pv < /dev/urandom | ./write-range testfile 0 $((2<<30)))
v2: Added comments explaining why the specific intervals are being used.
v3: Use 0 @nr when calling cgroup_writeback_by_id() to use best-effort
flushing while avoding possible livelocks.
v4: Use get_jiffies_64() and time_before/after64() instead of raw
jiffies_64 and arthimetic comparisons as suggested by Jan.
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-08-26 09:06:56 -07:00
|
|
|
|
|
|
|
|
|
2021-05-04 11:43:01 -04:00
|
|
|
|
writeback, memcg: Implement foreign dirty flushing
There's an inherent mismatch between memcg and writeback. The former
trackes ownership per-page while the latter per-inode. This was a
deliberate design decision because honoring per-page ownership in the
writeback path is complicated, may lead to higher CPU and IO overheads
and deemed unnecessary given that write-sharing an inode across
different cgroups isn't a common use-case.
Combined with inode majority-writer ownership switching, this works
well enough in most cases but there are some pathological cases. For
example, let's say there are two cgroups A and B which keep writing to
different but confined parts of the same inode. B owns the inode and
A's memory is limited far below B's. A's dirty ratio can rise enough
to trigger balance_dirty_pages() sleeps but B's can be low enough to
avoid triggering background writeback. A will be slowed down without
a way to make writeback of the dirty pages happen.
This patch implements foreign dirty recording and foreign mechanism so
that when a memcg encounters a condition as above it can trigger
flushes on bdi_writebacks which can clean its pages. Please see the
comment on top of mem_cgroup_track_foreign_dirty_slowpath() for
details.
A reproducer follows.
write-range.c::
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <fcntl.h>
#include <sys/types.h>
static const char *usage = "write-range FILE START SIZE\n";
int main(int argc, char **argv)
{
int fd;
unsigned long start, size, end, pos;
char *endp;
char buf[4096];
if (argc < 4) {
fprintf(stderr, usage);
return 1;
}
fd = open(argv[1], O_WRONLY);
if (fd < 0) {
perror("open");
return 1;
}
start = strtoul(argv[2], &endp, 0);
if (*endp != '\0') {
fprintf(stderr, usage);
return 1;
}
size = strtoul(argv[3], &endp, 0);
if (*endp != '\0') {
fprintf(stderr, usage);
return 1;
}
end = start + size;
while (1) {
for (pos = start; pos < end; ) {
long bread, bwritten = 0;
if (lseek(fd, pos, SEEK_SET) < 0) {
perror("lseek");
return 1;
}
bread = read(0, buf, sizeof(buf) < end - pos ?
sizeof(buf) : end - pos);
if (bread < 0) {
perror("read");
return 1;
}
if (bread == 0)
return 0;
while (bwritten < bread) {
long this;
this = write(fd, buf + bwritten,
bread - bwritten);
if (this < 0) {
perror("write");
return 1;
}
bwritten += this;
pos += bwritten;
}
}
}
}
repro.sh::
#!/bin/bash
set -e
set -x
sysctl -w vm.dirty_expire_centisecs=300000
sysctl -w vm.dirty_writeback_centisecs=300000
sysctl -w vm.dirtytime_expire_seconds=300000
echo 3 > /proc/sys/vm/drop_caches
TEST=/sys/fs/cgroup/test
A=$TEST/A
B=$TEST/B
mkdir -p $A $B
echo "+memory +io" > $TEST/cgroup.subtree_control
echo $((1<<30)) > $A/memory.high
echo $((32<<30)) > $B/memory.high
rm -f testfile
touch testfile
fallocate -l 4G testfile
echo "Starting B"
(echo $BASHPID > $B/cgroup.procs
pv -q --rate-limit 70M < /dev/urandom | ./write-range testfile $((2<<30)) $((2<<30))) &
echo "Waiting 10s to ensure B claims the testfile inode"
sleep 5
sync
sleep 5
sync
echo "Starting A"
(echo $BASHPID > $A/cgroup.procs
pv < /dev/urandom | ./write-range testfile 0 $((2<<30)))
v2: Added comments explaining why the specific intervals are being used.
v3: Use 0 @nr when calling cgroup_writeback_by_id() to use best-effort
flushing while avoding possible livelocks.
v4: Use get_jiffies_64() and time_before/after64() instead of raw
jiffies_64 and arthimetic comparisons as suggested by Jan.
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-08-26 09:06:56 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-05-04 11:43:01 -04:00
|
|
|
|
2019-08-29 15:47:19 -07:00
|
|
|
|
writeback, memcg: Implement foreign dirty flushing
There's an inherent mismatch between memcg and writeback. The former
trackes ownership per-page while the latter per-inode. This was a
deliberate design decision because honoring per-page ownership in the
writeback path is complicated, may lead to higher CPU and IO overheads
and deemed unnecessary given that write-sharing an inode across
different cgroups isn't a common use-case.
Combined with inode majority-writer ownership switching, this works
well enough in most cases but there are some pathological cases. For
example, let's say there are two cgroups A and B which keep writing to
different but confined parts of the same inode. B owns the inode and
A's memory is limited far below B's. A's dirty ratio can rise enough
to trigger balance_dirty_pages() sleeps but B's can be low enough to
avoid triggering background writeback. A will be slowed down without
a way to make writeback of the dirty pages happen.
This patch implements foreign dirty recording and foreign mechanism so
that when a memcg encounters a condition as above it can trigger
flushes on bdi_writebacks which can clean its pages. Please see the
comment on top of mem_cgroup_track_foreign_dirty_slowpath() for
details.
A reproducer follows.
write-range.c::
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <fcntl.h>
#include <sys/types.h>
static const char *usage = "write-range FILE START SIZE\n";
int main(int argc, char **argv)
{
int fd;
unsigned long start, size, end, pos;
char *endp;
char buf[4096];
if (argc < 4) {
fprintf(stderr, usage);
return 1;
}
fd = open(argv[1], O_WRONLY);
if (fd < 0) {
perror("open");
return 1;
}
start = strtoul(argv[2], &endp, 0);
if (*endp != '\0') {
fprintf(stderr, usage);
return 1;
}
size = strtoul(argv[3], &endp, 0);
if (*endp != '\0') {
fprintf(stderr, usage);
return 1;
}
end = start + size;
while (1) {
for (pos = start; pos < end; ) {
long bread, bwritten = 0;
if (lseek(fd, pos, SEEK_SET) < 0) {
perror("lseek");
return 1;
}
bread = read(0, buf, sizeof(buf) < end - pos ?
sizeof(buf) : end - pos);
if (bread < 0) {
perror("read");
return 1;
}
if (bread == 0)
return 0;
while (bwritten < bread) {
long this;
this = write(fd, buf + bwritten,
bread - bwritten);
if (this < 0) {
perror("write");
return 1;
}
bwritten += this;
pos += bwritten;
}
}
}
}
repro.sh::
#!/bin/bash
set -e
set -x
sysctl -w vm.dirty_expire_centisecs=300000
sysctl -w vm.dirty_writeback_centisecs=300000
sysctl -w vm.dirtytime_expire_seconds=300000
echo 3 > /proc/sys/vm/drop_caches
TEST=/sys/fs/cgroup/test
A=$TEST/A
B=$TEST/B
mkdir -p $A $B
echo "+memory +io" > $TEST/cgroup.subtree_control
echo $((1<<30)) > $A/memory.high
echo $((32<<30)) > $B/memory.high
rm -f testfile
touch testfile
fallocate -l 4G testfile
echo "Starting B"
(echo $BASHPID > $B/cgroup.procs
pv -q --rate-limit 70M < /dev/urandom | ./write-range testfile $((2<<30)) $((2<<30))) &
echo "Waiting 10s to ensure B claims the testfile inode"
sleep 5
sync
sleep 5
sync
echo "Starting A"
(echo $BASHPID > $A/cgroup.procs
pv < /dev/urandom | ./write-range testfile 0 $((2<<30)))
v2: Added comments explaining why the specific intervals are being used.
v3: Use 0 @nr when calling cgroup_writeback_by_id() to use best-effort
flushing while avoding possible livelocks.
v4: Use get_jiffies_64() and time_before/after64() instead of raw
jiffies_64 and arthimetic comparisons as suggested by Jan.
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-08-26 09:06:56 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-08-29 15:47:19 -07:00
|
|
|
|
2021-09-02 14:53:27 -07:00
|
|
|
|
writeback, memcg: Implement foreign dirty flushing
There's an inherent mismatch between memcg and writeback. The former
trackes ownership per-page while the latter per-inode. This was a
deliberate design decision because honoring per-page ownership in the
writeback path is complicated, may lead to higher CPU and IO overheads
and deemed unnecessary given that write-sharing an inode across
different cgroups isn't a common use-case.
Combined with inode majority-writer ownership switching, this works
well enough in most cases but there are some pathological cases. For
example, let's say there are two cgroups A and B which keep writing to
different but confined parts of the same inode. B owns the inode and
A's memory is limited far below B's. A's dirty ratio can rise enough
to trigger balance_dirty_pages() sleeps but B's can be low enough to
avoid triggering background writeback. A will be slowed down without
a way to make writeback of the dirty pages happen.
This patch implements foreign dirty recording and foreign mechanism so
that when a memcg encounters a condition as above it can trigger
flushes on bdi_writebacks which can clean its pages. Please see the
comment on top of mem_cgroup_track_foreign_dirty_slowpath() for
details.
A reproducer follows.
write-range.c::
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <fcntl.h>
#include <sys/types.h>
static const char *usage = "write-range FILE START SIZE\n";
int main(int argc, char **argv)
{
int fd;
unsigned long start, size, end, pos;
char *endp;
char buf[4096];
if (argc < 4) {
fprintf(stderr, usage);
return 1;
}
fd = open(argv[1], O_WRONLY);
if (fd < 0) {
perror("open");
return 1;
}
start = strtoul(argv[2], &endp, 0);
if (*endp != '\0') {
fprintf(stderr, usage);
return 1;
}
size = strtoul(argv[3], &endp, 0);
if (*endp != '\0') {
fprintf(stderr, usage);
return 1;
}
end = start + size;
while (1) {
for (pos = start; pos < end; ) {
long bread, bwritten = 0;
if (lseek(fd, pos, SEEK_SET) < 0) {
perror("lseek");
return 1;
}
bread = read(0, buf, sizeof(buf) < end - pos ?
sizeof(buf) : end - pos);
if (bread < 0) {
perror("read");
return 1;
}
if (bread == 0)
return 0;
while (bwritten < bread) {
long this;
this = write(fd, buf + bwritten,
bread - bwritten);
if (this < 0) {
perror("write");
return 1;
}
bwritten += this;
pos += bwritten;
}
}
}
}
repro.sh::
#!/bin/bash
set -e
set -x
sysctl -w vm.dirty_expire_centisecs=300000
sysctl -w vm.dirty_writeback_centisecs=300000
sysctl -w vm.dirtytime_expire_seconds=300000
echo 3 > /proc/sys/vm/drop_caches
TEST=/sys/fs/cgroup/test
A=$TEST/A
B=$TEST/B
mkdir -p $A $B
echo "+memory +io" > $TEST/cgroup.subtree_control
echo $((1<<30)) > $A/memory.high
echo $((32<<30)) > $B/memory.high
rm -f testfile
touch testfile
fallocate -l 4G testfile
echo "Starting B"
(echo $BASHPID > $B/cgroup.procs
pv -q --rate-limit 70M < /dev/urandom | ./write-range testfile $((2<<30)) $((2<<30))) &
echo "Waiting 10s to ensure B claims the testfile inode"
sleep 5
sync
sleep 5
sync
echo "Starting A"
(echo $BASHPID > $A/cgroup.procs
pv < /dev/urandom | ./write-range testfile 0 $((2<<30)))
v2: Added comments explaining why the specific intervals are being used.
v3: Use 0 @nr when calling cgroup_writeback_by_id() to use best-effort
flushing while avoding possible livelocks.
v4: Use get_jiffies_64() and time_before/after64() instead of raw
jiffies_64 and arthimetic comparisons as suggested by Jan.
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-08-26 09:06:56 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-05-22 18:23:33 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-05-22 18:23:34 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-05-22 17:13:37 -04:00
|
|
|
|
|
|
|
|
|
2016-07-20 15:44:57 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-06-04 16:49:28 -07:00
|
|
|
|
2016-07-20 15:44:57 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-07-08 10:33:04 +08:00
|
|
|
|
2024-08-09 10:26:18 -07:00
|
|
|
|
2016-07-20 15:44:57 -07:00
|
|
|
|
2018-08-02 15:36:01 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-08-09 10:26:18 -07:00
|
|
|
|
2018-08-02 15:36:01 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-06-24 17:58:56 -07:00
|
|
|
|
|
|
|
|
|
2016-07-20 15:44:57 -07:00
|
|
|
|
2018-10-26 15:09:28 -07:00
|
|
|
|
2016-07-20 15:44:57 -07:00
|
|
|
|
|
|
|
|
|
2025-01-23 23:38:58 -05:00
|
|
|
|
2016-07-20 15:44:57 -07:00
|
|
|
|
2018-10-26 15:09:28 -07:00
|
|
|
|
2018-08-02 15:36:01 -07:00
|
|
|
|
2016-07-20 15:44:57 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-08-11 15:33:03 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-01-24 00:41:32 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-07-20 15:44:57 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-08-09 10:26:18 -07:00
|
|
|
|
2016-07-20 15:44:57 -07:00
|
|
|
|
|
|
|
|
|
mm: memcontrol: introduce mem_cgroup_ino() and mem_cgroup_get_from_ino()
Patch series "mm: introduce shrinker debugfs interface", v5.
The only existing debugging mechanism is a couple of tracepoints in
do_shrink_slab(): mm_shrink_slab_start and mm_shrink_slab_end. They
aren't covering everything though: shrinkers which report 0 objects will
never show up, there is no support for memcg-aware shrinkers. Shrinkers
are identified by their scan function, which is not always enough (e.g.
hard to guess which super block's shrinker it is having only
"super_cache_scan").
To provide a better visibility and debug options for memory shrinkers this
patchset introduces a /sys/kernel/debug/shrinker interface, to some extent
similar to /sys/kernel/slab.
For each shrinker registered in the system a directory is created. As
now, the directory will contain only a "scan" file, which allows to get
the number of managed objects for each memory cgroup (for memcg-aware
shrinkers) and each numa node (for numa-aware shrinkers on a numa
machine). Other interfaces might be added in the future.
To make debugging more pleasant, the patchset also names all shrinkers, so
that debugfs entries can have meaningful names.
This patch (of 5):
Shrinker debugfs requires a way to represent memory cgroups without using
full paths, both for displaying information and getting input from a user.
Cgroup inode number is a perfect way, already used by bpf.
This commit adds a couple of helper functions which will be used to handle
memcg-aware shrinkers.
Link: https://lkml.kernel.org/r/20220601032227.4076670-1-roman.gushchin@linux.dev
Link: https://lkml.kernel.org/r/20220601032227.4076670-2-roman.gushchin@linux.dev
Signed-off-by: Roman Gushchin <roman.gushchin@linux.dev>
Acked-by: Muchun Song <songmuchun@bytedance.com>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: Kent Overstreet <kent.overstreet@gmail.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Christophe JAILLET <christophe.jaillet@wanadoo.fr>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-31 20:22:22 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-08-26 18:52:37 +02:00
|
|
|
|
2022-08-28 17:54:15 -10:00
|
|
|
|
mm: memcontrol: introduce mem_cgroup_ino() and mem_cgroup_get_from_ino()
Patch series "mm: introduce shrinker debugfs interface", v5.
The only existing debugging mechanism is a couple of tracepoints in
do_shrink_slab(): mm_shrink_slab_start and mm_shrink_slab_end. They
aren't covering everything though: shrinkers which report 0 objects will
never show up, there is no support for memcg-aware shrinkers. Shrinkers
are identified by their scan function, which is not always enough (e.g.
hard to guess which super block's shrinker it is having only
"super_cache_scan").
To provide a better visibility and debug options for memory shrinkers this
patchset introduces a /sys/kernel/debug/shrinker interface, to some extent
similar to /sys/kernel/slab.
For each shrinker registered in the system a directory is created. As
now, the directory will contain only a "scan" file, which allows to get
the number of managed objects for each memory cgroup (for memcg-aware
shrinkers) and each numa node (for numa-aware shrinkers on a numa
machine). Other interfaces might be added in the future.
To make debugging more pleasant, the patchset also names all shrinkers, so
that debugfs entries can have meaningful names.
This patch (of 5):
Shrinker debugfs requires a way to represent memory cgroups without using
full paths, both for displaying information and getting input from a user.
Cgroup inode number is a perfect way, already used by bpf.
This commit adds a couple of helper functions which will be used to handle
memcg-aware shrinkers.
Link: https://lkml.kernel.org/r/20220601032227.4076670-1-roman.gushchin@linux.dev
Link: https://lkml.kernel.org/r/20220601032227.4076670-2-roman.gushchin@linux.dev
Signed-off-by: Roman Gushchin <roman.gushchin@linux.dev>
Acked-by: Muchun Song <songmuchun@bytedance.com>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: Kent Overstreet <kent.overstreet@gmail.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Christophe JAILLET <christophe.jaillet@wanadoo.fr>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-31 20:22:22 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-01-24 07:35:12 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-05-07 13:23:24 +00:00
|
|
|
|
2008-02-07 00:14:31 -08:00
|
|
|
|
|
|
|
|
|
2022-03-22 14:47:06 -07:00
|
|
|
|
2025-04-25 11:19:25 +08:00
|
|
|
|
|
|
|
|
|
2008-02-07 00:14:31 -08:00
|
|
|
|
2024-05-07 13:23:24 +00:00
|
|
|
|
2008-02-07 00:14:38 -08:00
|
|
|
|
2024-05-01 10:26:12 -07:00
|
|
|
|
|
|
|
|
|
2024-05-01 10:26:11 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-09-02 14:55:00 -07:00
|
|
|
|
|
|
|
|
|
2024-05-01 10:26:11 -07:00
|
|
|
|
|
|
|
|
|
2017-07-06 15:40:52 -07:00
|
|
|
|
2016-07-28 15:46:05 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-07-08 15:59:49 -07:00
|
|
|
|
2024-05-07 13:23:24 +00:00
|
|
|
|
2024-05-01 10:26:11 -07:00
|
|
|
|
2025-01-24 07:35:12 +00:00
|
|
|
|
2024-05-07 13:23:24 +00:00
|
|
|
|
2008-02-07 00:14:31 -08:00
|
|
|
|
|
|
|
|
|
2017-03-09 16:17:26 -08:00
|
|
|
|
memcg: free mem_cgroup by RCU to fix oops
After fixing the GPF in mem_cgroup_lru_del_list(), three times one
machine running a similar load (moving and removing memcgs while
swapping) has oopsed in mem_cgroup_zone_nr_lru_pages(), when retrieving
memcg zone numbers for get_scan_count() for shrink_mem_cgroup_zone():
this is where a struct mem_cgroup is first accessed after being chosen
by mem_cgroup_iter().
Just what protects a struct mem_cgroup from being freed, in between
mem_cgroup_iter()'s css_get_next() and its css_tryget()? css_tryget()
fails once css->refcnt is zero with CSS_REMOVED set in flags, yes: but
what if that memory is freed and reused for something else, which sets
"refcnt" non-zero? Hmm, and scope for an indefinite freeze if refcnt is
left at zero but flags are cleared.
It's tempting to move the css_tryget() into css_get_next(), to make it
really "get" the css, but I don't think that actually solves anything:
the same difficulty in moving from css_id found to stable css remains.
But we already have rcu_read_lock() around the two, so it's easily fixed
if __mem_cgroup_free() just uses kfree_rcu() to free mem_cgroup.
However, a big struct mem_cgroup is allocated with vzalloc() instead of
kzalloc(), and we're not allowed to vfree() at interrupt time: there
doesn't appear to be a general vfree_rcu() to help with this, so roll
our own using schedule_work(). The compiler decently removes
vfree_work() and vfree_rcu() when the config doesn't need them.
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Konstantin Khlebnikov <khlebnikov@openvz.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Ying Han <yinghan@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-03-15 15:17:07 -07:00
|
|
|
|
memcg: execute the whole memcg freeing in free_worker()
A lot of the initialization we do in mem_cgroup_create() is done with
softirqs enabled. This include grabbing a css id, which holds
&ss->id_lock->rlock, and the per-zone trees, which holds
rtpz->lock->rlock. All of those signal to the lockdep mechanism that
those locks can be used in SOFTIRQ-ON-W context.
This means that the freeing of memcg structure must happen in a
compatible context, otherwise we'll get a deadlock, like the one below,
caught by lockdep:
free_accounted_pages+0x47/0x4c
free_task+0x31/0x5c
__put_task_struct+0xc2/0xdb
put_task_struct+0x1e/0x22
delayed_put_task_struct+0x7a/0x98
__rcu_process_callbacks+0x269/0x3df
rcu_process_callbacks+0x31/0x5b
__do_softirq+0x122/0x277
This usage pattern could not be triggered before kmem came into play.
With the introduction of kmem stack handling, it is possible that we call
the last mem_cgroup_put() from the task destructor, which is run in an rcu
callback. Such callbacks are run with softirqs disabled, leading to the
offensive usage pattern.
In general, we have little, if any, means to guarantee in which context
the last memcg_put will happen. The best we can do is test it and try to
make sure no invalid context releases are happening. But as we add more
code to memcg, the possible interactions grow in number and expose more
ways to get context conflicts. One thing to keep in mind, is that part of
the freeing process is already deferred to a worker, such as vfree(), that
can only be called from process context.
For the moment, the only two functions we really need moved away are:
* free_css_id(), and
* mem_cgroup_remove_from_trees().
But because the later accesses per-zone info,
free_mem_cgroup_per_zone_info() needs to be moved as well. With that, we
are left with the per_cpu stats only. Better move it all.
Signed-off-by: Glauber Costa <glommer@parallels.com>
Tested-by: Greg Thelen <gthelen@google.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Frederic Weisbecker <fweisbec@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: JoonSoo Kim <js1304@gmail.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: Rik van Riel <riel@redhat.com>
Cc: Suleiman Souhlal <suleiman@google.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-12-18 14:22:13 -08:00
|
|
|
|
memcg: free mem_cgroup by RCU to fix oops
After fixing the GPF in mem_cgroup_lru_del_list(), three times one
machine running a similar load (moving and removing memcgs while
swapping) has oopsed in mem_cgroup_zone_nr_lru_pages(), when retrieving
memcg zone numbers for get_scan_count() for shrink_mem_cgroup_zone():
this is where a struct mem_cgroup is first accessed after being chosen
by mem_cgroup_iter().
Just what protects a struct mem_cgroup from being freed, in between
mem_cgroup_iter()'s css_get_next() and its css_tryget()? css_tryget()
fails once css->refcnt is zero with CSS_REMOVED set in flags, yes: but
what if that memory is freed and reused for something else, which sets
"refcnt" non-zero? Hmm, and scope for an indefinite freeze if refcnt is
left at zero but flags are cleared.
It's tempting to move the css_tryget() into css_get_next(), to make it
really "get" the css, but I don't think that actually solves anything:
the same difficulty in moving from css_id found to stable css remains.
But we already have rcu_read_lock() around the two, so it's easily fixed
if __mem_cgroup_free() just uses kfree_rcu() to free mem_cgroup.
However, a big struct mem_cgroup is allocated with vzalloc() instead of
kzalloc(), and we're not allowed to vfree() at interrupt time: there
doesn't appear to be a general vfree_rcu() to help with this, so roll
our own using schedule_work(). The compiler decently removes
vfree_work() and vfree_rcu() when the config doesn't need them.
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Konstantin Khlebnikov <khlebnikov@openvz.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Ying Han <yinghan@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-03-15 15:17:07 -07:00
|
|
|
|
2024-03-16 01:58:03 +00:00
|
|
|
|
2023-10-19 15:53:43 -07:00
|
|
|
|
memcg: execute the whole memcg freeing in free_worker()
A lot of the initialization we do in mem_cgroup_create() is done with
softirqs enabled. This include grabbing a css id, which holds
&ss->id_lock->rlock, and the per-zone trees, which holds
rtpz->lock->rlock. All of those signal to the lockdep mechanism that
those locks can be used in SOFTIRQ-ON-W context.
This means that the freeing of memcg structure must happen in a
compatible context, otherwise we'll get a deadlock, like the one below,
caught by lockdep:
free_accounted_pages+0x47/0x4c
free_task+0x31/0x5c
__put_task_struct+0xc2/0xdb
put_task_struct+0x1e/0x22
delayed_put_task_struct+0x7a/0x98
__rcu_process_callbacks+0x269/0x3df
rcu_process_callbacks+0x31/0x5b
__do_softirq+0x122/0x277
This usage pattern could not be triggered before kmem came into play.
With the introduction of kmem stack handling, it is possible that we call
the last mem_cgroup_put() from the task destructor, which is run in an rcu
callback. Such callbacks are run with softirqs disabled, leading to the
offensive usage pattern.
In general, we have little, if any, means to guarantee in which context
the last memcg_put will happen. The best we can do is test it and try to
make sure no invalid context releases are happening. But as we add more
code to memcg, the possible interactions grow in number and expose more
ways to get context conflicts. One thing to keep in mind, is that part of
the freeing process is already deferred to a worker, such as vfree(), that
can only be called from process context.
For the moment, the only two functions we really need moved away are:
* free_css_id(), and
* mem_cgroup_remove_from_trees().
But because the later accesses per-zone info,
free_mem_cgroup_per_zone_info() needs to be moved as well. With that, we
are left with the per_cpu stats only. Better move it all.
Signed-off-by: Glauber Costa <glommer@parallels.com>
Tested-by: Greg Thelen <gthelen@google.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Frederic Weisbecker <fweisbec@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: JoonSoo Kim <js1304@gmail.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: Rik van Riel <riel@redhat.com>
Cc: Suleiman Souhlal <suleiman@google.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-12-18 14:22:13 -08:00
|
|
|
|
2025-01-24 07:35:12 +00:00
|
|
|
|
2024-08-14 22:04:47 -07:00
|
|
|
|
2022-09-07 04:35:35 +00:00
|
|
|
|
2019-05-14 15:46:57 -07:00
|
|
|
|
2014-01-23 15:52:52 -08:00
|
|
|
|
memcg: free mem_cgroup by RCU to fix oops
After fixing the GPF in mem_cgroup_lru_del_list(), three times one
machine running a similar load (moving and removing memcgs while
swapping) has oopsed in mem_cgroup_zone_nr_lru_pages(), when retrieving
memcg zone numbers for get_scan_count() for shrink_mem_cgroup_zone():
this is where a struct mem_cgroup is first accessed after being chosen
by mem_cgroup_iter().
Just what protects a struct mem_cgroup from being freed, in between
mem_cgroup_iter()'s css_get_next() and its css_tryget()? css_tryget()
fails once css->refcnt is zero with CSS_REMOVED set in flags, yes: but
what if that memory is freed and reused for something else, which sets
"refcnt" non-zero? Hmm, and scope for an indefinite freeze if refcnt is
left at zero but flags are cleared.
It's tempting to move the css_tryget() into css_get_next(), to make it
really "get" the css, but I don't think that actually solves anything:
the same difficulty in moving from css_id found to stable css remains.
But we already have rcu_read_lock() around the two, so it's easily fixed
if __mem_cgroup_free() just uses kfree_rcu() to free mem_cgroup.
However, a big struct mem_cgroup is allocated with vzalloc() instead of
kzalloc(), and we're not allowed to vfree() at interrupt time: there
doesn't appear to be a general vfree_rcu() to help with this, so roll
our own using schedule_work(). The compiler decently removes
vfree_work() and vfree_rcu() when the config doesn't need them.
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Konstantin Khlebnikov <khlebnikov@openvz.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Ying Han <yinghan@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-03-15 15:17:07 -07:00
|
|
|
|
2012-05-29 15:07:10 -07:00
|
|
|
|
2017-03-09 16:17:26 -08:00
|
|
|
|
|
|
|
|
|
mm: multi-gen LRU: groundwork
Evictable pages are divided into multiple generations for each lruvec.
The youngest generation number is stored in lrugen->max_seq for both
anon and file types as they are aged on an equal footing. The oldest
generation numbers are stored in lrugen->min_seq[] separately for anon
and file types as clean file pages can be evicted regardless of swap
constraints. These three variables are monotonically increasing.
Generation numbers are truncated into order_base_2(MAX_NR_GENS+1) bits
in order to fit into the gen counter in folio->flags. Each truncated
generation number is an index to lrugen->lists[]. The sliding window
technique is used to track at least MIN_NR_GENS and at most
MAX_NR_GENS generations. The gen counter stores a value within [1,
MAX_NR_GENS] while a page is on one of lrugen->lists[]. Otherwise it
stores 0.
There are two conceptually independent procedures: "the aging", which
produces young generations, and "the eviction", which consumes old
generations. They form a closed-loop system, i.e., "the page reclaim".
Both procedures can be invoked from userspace for the purposes of working
set estimation and proactive reclaim. These techniques are commonly used
to optimize job scheduling (bin packing) in data centers [1][2].
To avoid confusion, the terms "hot" and "cold" will be applied to the
multi-gen LRU, as a new convention; the terms "active" and "inactive" will
be applied to the active/inactive LRU, as usual.
The protection of hot pages and the selection of cold pages are based
on page access channels and patterns. There are two access channels:
one through page tables and the other through file descriptors. The
protection of the former channel is by design stronger because:
1. The uncertainty in determining the access patterns of the former
channel is higher due to the approximation of the accessed bit.
2. The cost of evicting the former channel is higher due to the TLB
flushes required and the likelihood of encountering the dirty bit.
3. The penalty of underprotecting the former channel is higher because
applications usually do not prepare themselves for major page
faults like they do for blocked I/O. E.g., GUI applications
commonly use dedicated I/O threads to avoid blocking rendering
threads.
There are also two access patterns: one with temporal locality and the
other without. For the reasons listed above, the former channel is
assumed to follow the former pattern unless VM_SEQ_READ or VM_RAND_READ is
present; the latter channel is assumed to follow the latter pattern unless
outlying refaults have been observed [3][4].
The next patch will address the "outlying refaults". Three macros, i.e.,
LRU_REFS_WIDTH, LRU_REFS_PGOFF and LRU_REFS_MASK, used later are added in
this patch to make the entire patchset less diffy.
A page is added to the youngest generation on faulting. The aging needs
to check the accessed bit at least twice before handing this page over to
the eviction. The first check takes care of the accessed bit set on the
initial fault; the second check makes sure this page has not been used
since then. This protocol, AKA second chance, requires a minimum of two
generations, hence MIN_NR_GENS.
[1] https://dl.acm.org/doi/10.1145/3297858.3304053
[2] https://dl.acm.org/doi/10.1145/3503222.3507731
[3] https://lwn.net/Articles/495543/
[4] https://lwn.net/Articles/815342/
Link: https://lkml.kernel.org/r/20220918080010.2920238-6-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 02:00:02 -06:00
|
|
|
|
2017-03-09 16:17:26 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcg: optimize parent iteration in memcg_rstat_updated()
In memcg_rstat_updated(), we iterate the memcg being updated and its
parents to update memcg->vmstats_percpu->stats_updates in the fast path
(i.e. no atomic updates). According to my math, this is 3 memory loads
(and potentially 3 cache misses) per memcg:
- Load the address of memcg->vmstats_percpu.
- Load vmstats_percpu->stats_updates (based on some percpu calculation).
- Load the address of the parent memcg.
Avoid most of the cache misses by caching a pointer from each struct
memcg_vmstats_percpu to its parent on the corresponding CPU. In this
case, for the first memcg we have 2 memory loads (same as above):
- Load the address of memcg->vmstats_percpu.
- Load vmstats_percpu->stats_updates (based on some percpu calculation).
Then for each additional memcg, we need a single load to get the
parent's stats_updates directly. This reduces the number of loads from
O(3N) to O(2+N) -- where N is the number of memcgs we need to iterate.
Additionally, stash a pointer to memcg->vmstats in each struct
memcg_vmstats_percpu such that we can access the atomic counter that all
CPUs fold into, memcg->vmstats->stats_updates.
memcg_should_flush_stats() is changed to memcg_vmstats_needs_flush() to
accept a struct memcg_vmstats pointer accordingly.
In struct memcg_vmstats_percpu, make sure both pointers together with
stats_updates live on the same cacheline. Finally, update
mem_cgroup_alloc() to take in a parent pointer and initialize the new
cache pointers on each CPU. The percpu loop in mem_cgroup_alloc() may
look concerning, but there are multiple similar loops in the cgroup
creation path (e.g. cgroup_rstat_init()), most of which are hidden
within alloc_percpu().
According to Oliver's testing [1], this fixes multiple 30-38%
regressions in vm-scalability, will-it-scale-tlb_flush2, and
will-it-scale-fallocate1. This comes at a cost of 2 more pointers per
CPU (<2KB on a machine with 128 CPUs).
[1] https://lore.kernel.org/lkml/ZbDJsfsZt2ITyo61@xsang-OptiPlex-9020/
[yosryahmed@google.com: fix struct memcg_vmstats_percpu size and alignment]
Link: https://lkml.kernel.org/r/20240203044612.1234216-1-yosryahmed@google.com
Link: https://lkml.kernel.org/r/20240124100023.660032-1-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Fixes: 8d59d2214c23 ("mm: memcg: make stats flushing threshold per-memcg")
Tested-by: kernel test robot <oliver.sang@intel.com>
Reported-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202401221624.cb53a8ca-oliver.sang@intel.com
Acked-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-01-24 10:00:22 +00:00
|
|
|
|
2008-02-07 00:13:50 -08:00
|
|
|
|
2025-05-14 11:41:52 -07:00
|
|
|
|
|
|
|
|
|
2013-02-22 16:34:52 -08:00
|
|
|
|
mm: memcg: optimize parent iteration in memcg_rstat_updated()
In memcg_rstat_updated(), we iterate the memcg being updated and its
parents to update memcg->vmstats_percpu->stats_updates in the fast path
(i.e. no atomic updates). According to my math, this is 3 memory loads
(and potentially 3 cache misses) per memcg:
- Load the address of memcg->vmstats_percpu.
- Load vmstats_percpu->stats_updates (based on some percpu calculation).
- Load the address of the parent memcg.
Avoid most of the cache misses by caching a pointer from each struct
memcg_vmstats_percpu to its parent on the corresponding CPU. In this
case, for the first memcg we have 2 memory loads (same as above):
- Load the address of memcg->vmstats_percpu.
- Load vmstats_percpu->stats_updates (based on some percpu calculation).
Then for each additional memcg, we need a single load to get the
parent's stats_updates directly. This reduces the number of loads from
O(3N) to O(2+N) -- where N is the number of memcgs we need to iterate.
Additionally, stash a pointer to memcg->vmstats in each struct
memcg_vmstats_percpu such that we can access the atomic counter that all
CPUs fold into, memcg->vmstats->stats_updates.
memcg_should_flush_stats() is changed to memcg_vmstats_needs_flush() to
accept a struct memcg_vmstats pointer accordingly.
In struct memcg_vmstats_percpu, make sure both pointers together with
stats_updates live on the same cacheline. Finally, update
mem_cgroup_alloc() to take in a parent pointer and initialize the new
cache pointers on each CPU. The percpu loop in mem_cgroup_alloc() may
look concerning, but there are multiple similar loops in the cgroup
creation path (e.g. cgroup_rstat_init()), most of which are hidden
within alloc_percpu().
According to Oliver's testing [1], this fixes multiple 30-38%
regressions in vm-scalability, will-it-scale-tlb_flush2, and
will-it-scale-fallocate1. This comes at a cost of 2 more pointers per
CPU (<2KB on a machine with 128 CPUs).
[1] https://lore.kernel.org/lkml/ZbDJsfsZt2ITyo61@xsang-OptiPlex-9020/
[yosryahmed@google.com: fix struct memcg_vmstats_percpu size and alignment]
Link: https://lkml.kernel.org/r/20240203044612.1234216-1-yosryahmed@google.com
Link: https://lkml.kernel.org/r/20240124100023.660032-1-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Fixes: 8d59d2214c23 ("mm: memcg: make stats flushing threshold per-memcg")
Tested-by: kernel test robot <oliver.sang@intel.com>
Reported-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202401221624.cb53a8ca-oliver.sang@intel.com
Acked-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-01-24 10:00:22 +00:00
|
|
|
|
writeback, memcg: Implement foreign dirty flushing
There's an inherent mismatch between memcg and writeback. The former
trackes ownership per-page while the latter per-inode. This was a
deliberate design decision because honoring per-page ownership in the
writeback path is complicated, may lead to higher CPU and IO overheads
and deemed unnecessary given that write-sharing an inode across
different cgroups isn't a common use-case.
Combined with inode majority-writer ownership switching, this works
well enough in most cases but there are some pathological cases. For
example, let's say there are two cgroups A and B which keep writing to
different but confined parts of the same inode. B owns the inode and
A's memory is limited far below B's. A's dirty ratio can rise enough
to trigger balance_dirty_pages() sleeps but B's can be low enough to
avoid triggering background writeback. A will be slowed down without
a way to make writeback of the dirty pages happen.
This patch implements foreign dirty recording and foreign mechanism so
that when a memcg encounters a condition as above it can trigger
flushes on bdi_writebacks which can clean its pages. Please see the
comment on top of mem_cgroup_track_foreign_dirty_slowpath() for
details.
A reproducer follows.
write-range.c::
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <fcntl.h>
#include <sys/types.h>
static const char *usage = "write-range FILE START SIZE\n";
int main(int argc, char **argv)
{
int fd;
unsigned long start, size, end, pos;
char *endp;
char buf[4096];
if (argc < 4) {
fprintf(stderr, usage);
return 1;
}
fd = open(argv[1], O_WRONLY);
if (fd < 0) {
perror("open");
return 1;
}
start = strtoul(argv[2], &endp, 0);
if (*endp != '\0') {
fprintf(stderr, usage);
return 1;
}
size = strtoul(argv[3], &endp, 0);
if (*endp != '\0') {
fprintf(stderr, usage);
return 1;
}
end = start + size;
while (1) {
for (pos = start; pos < end; ) {
long bread, bwritten = 0;
if (lseek(fd, pos, SEEK_SET) < 0) {
perror("lseek");
return 1;
}
bread = read(0, buf, sizeof(buf) < end - pos ?
sizeof(buf) : end - pos);
if (bread < 0) {
perror("read");
return 1;
}
if (bread == 0)
return 0;
while (bwritten < bread) {
long this;
this = write(fd, buf + bwritten,
bread - bwritten);
if (this < 0) {
perror("write");
return 1;
}
bwritten += this;
pos += bwritten;
}
}
}
}
repro.sh::
#!/bin/bash
set -e
set -x
sysctl -w vm.dirty_expire_centisecs=300000
sysctl -w vm.dirty_writeback_centisecs=300000
sysctl -w vm.dirtytime_expire_seconds=300000
echo 3 > /proc/sys/vm/drop_caches
TEST=/sys/fs/cgroup/test
A=$TEST/A
B=$TEST/B
mkdir -p $A $B
echo "+memory +io" > $TEST/cgroup.subtree_control
echo $((1<<30)) > $A/memory.high
echo $((32<<30)) > $B/memory.high
rm -f testfile
touch testfile
fallocate -l 4G testfile
echo "Starting B"
(echo $BASHPID > $B/cgroup.procs
pv -q --rate-limit 70M < /dev/urandom | ./write-range testfile $((2<<30)) $((2<<30))) &
echo "Waiting 10s to ensure B claims the testfile inode"
sleep 5
sync
sleep 5
sync
echo "Starting A"
(echo $BASHPID > $A/cgroup.procs
pv < /dev/urandom | ./write-range testfile 0 $((2<<30)))
v2: Added comments explaining why the specific intervals are being used.
v3: Use 0 @nr when calling cgroup_writeback_by_id() to use best-effort
flushing while avoding possible livelocks.
v4: Use get_jiffies_64() and time_before/after64() instead of raw
jiffies_64 and arthimetic comparisons as suggested by Jan.
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-08-26 09:06:56 -07:00
|
|
|
|
2024-08-09 10:26:18 -07:00
|
|
|
|
2008-02-07 00:13:50 -08:00
|
|
|
|
2025-04-25 11:19:24 +08:00
|
|
|
|
2011-11-02 13:38:15 -07:00
|
|
|
|
2024-08-09 10:26:18 -07:00
|
|
|
|
2016-01-20 15:02:53 -08:00
|
|
|
|
2024-08-09 10:26:18 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-07-20 15:44:57 -07:00
|
|
|
|
2024-08-09 10:26:18 -07:00
|
|
|
|
2016-07-20 15:44:57 -07:00
|
|
|
|
2024-05-01 10:26:12 -07:00
|
|
|
|
|
|
|
|
|
2022-09-07 04:35:35 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-08-11 18:30:25 -07:00
|
|
|
|
|
|
|
|
|
2019-05-14 15:46:57 -07:00
|
|
|
|
2016-01-20 15:02:53 -08:00
|
|
|
|
2008-02-07 00:13:51 -08:00
|
|
|
|
2024-08-14 22:04:47 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcg: optimize parent iteration in memcg_rstat_updated()
In memcg_rstat_updated(), we iterate the memcg being updated and its
parents to update memcg->vmstats_percpu->stats_updates in the fast path
(i.e. no atomic updates). According to my math, this is 3 memory loads
(and potentially 3 cache misses) per memcg:
- Load the address of memcg->vmstats_percpu.
- Load vmstats_percpu->stats_updates (based on some percpu calculation).
- Load the address of the parent memcg.
Avoid most of the cache misses by caching a pointer from each struct
memcg_vmstats_percpu to its parent on the corresponding CPU. In this
case, for the first memcg we have 2 memory loads (same as above):
- Load the address of memcg->vmstats_percpu.
- Load vmstats_percpu->stats_updates (based on some percpu calculation).
Then for each additional memcg, we need a single load to get the
parent's stats_updates directly. This reduces the number of loads from
O(3N) to O(2+N) -- where N is the number of memcgs we need to iterate.
Additionally, stash a pointer to memcg->vmstats in each struct
memcg_vmstats_percpu such that we can access the atomic counter that all
CPUs fold into, memcg->vmstats->stats_updates.
memcg_should_flush_stats() is changed to memcg_vmstats_needs_flush() to
accept a struct memcg_vmstats pointer accordingly.
In struct memcg_vmstats_percpu, make sure both pointers together with
stats_updates live on the same cacheline. Finally, update
mem_cgroup_alloc() to take in a parent pointer and initialize the new
cache pointers on each CPU. The percpu loop in mem_cgroup_alloc() may
look concerning, but there are multiple similar loops in the cgroup
creation path (e.g. cgroup_rstat_init()), most of which are hidden
within alloc_percpu().
According to Oliver's testing [1], this fixes multiple 30-38%
regressions in vm-scalability, will-it-scale-tlb_flush2, and
will-it-scale-fallocate1. This comes at a cost of 2 more pointers per
CPU (<2KB on a machine with 128 CPUs).
[1] https://lore.kernel.org/lkml/ZbDJsfsZt2ITyo61@xsang-OptiPlex-9020/
[yosryahmed@google.com: fix struct memcg_vmstats_percpu size and alignment]
Link: https://lkml.kernel.org/r/20240203044612.1234216-1-yosryahmed@google.com
Link: https://lkml.kernel.org/r/20240124100023.660032-1-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Fixes: 8d59d2214c23 ("mm: memcg: make stats flushing threshold per-memcg")
Tested-by: kernel test robot <oliver.sang@intel.com>
Reported-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202401221624.cb53a8ca-oliver.sang@intel.com
Acked-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-01-24 10:00:22 +00:00
|
|
|
|
|
|
|
|
|
2025-05-14 11:41:52 -07:00
|
|
|
|
mm: memcg: optimize parent iteration in memcg_rstat_updated()
In memcg_rstat_updated(), we iterate the memcg being updated and its
parents to update memcg->vmstats_percpu->stats_updates in the fast path
(i.e. no atomic updates). According to my math, this is 3 memory loads
(and potentially 3 cache misses) per memcg:
- Load the address of memcg->vmstats_percpu.
- Load vmstats_percpu->stats_updates (based on some percpu calculation).
- Load the address of the parent memcg.
Avoid most of the cache misses by caching a pointer from each struct
memcg_vmstats_percpu to its parent on the corresponding CPU. In this
case, for the first memcg we have 2 memory loads (same as above):
- Load the address of memcg->vmstats_percpu.
- Load vmstats_percpu->stats_updates (based on some percpu calculation).
Then for each additional memcg, we need a single load to get the
parent's stats_updates directly. This reduces the number of loads from
O(3N) to O(2+N) -- where N is the number of memcgs we need to iterate.
Additionally, stash a pointer to memcg->vmstats in each struct
memcg_vmstats_percpu such that we can access the atomic counter that all
CPUs fold into, memcg->vmstats->stats_updates.
memcg_should_flush_stats() is changed to memcg_vmstats_needs_flush() to
accept a struct memcg_vmstats pointer accordingly.
In struct memcg_vmstats_percpu, make sure both pointers together with
stats_updates live on the same cacheline. Finally, update
mem_cgroup_alloc() to take in a parent pointer and initialize the new
cache pointers on each CPU. The percpu loop in mem_cgroup_alloc() may
look concerning, but there are multiple similar loops in the cgroup
creation path (e.g. cgroup_rstat_init()), most of which are hidden
within alloc_percpu().
According to Oliver's testing [1], this fixes multiple 30-38%
regressions in vm-scalability, will-it-scale-tlb_flush2, and
will-it-scale-fallocate1. This comes at a cost of 2 more pointers per
CPU (<2KB on a machine with 128 CPUs).
[1] https://lore.kernel.org/lkml/ZbDJsfsZt2ITyo61@xsang-OptiPlex-9020/
[yosryahmed@google.com: fix struct memcg_vmstats_percpu size and alignment]
Link: https://lkml.kernel.org/r/20240203044612.1234216-1-yosryahmed@google.com
Link: https://lkml.kernel.org/r/20240124100023.660032-1-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Fixes: 8d59d2214c23 ("mm: memcg: make stats flushing threshold per-memcg")
Tested-by: kernel test robot <oliver.sang@intel.com>
Reported-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202401221624.cb53a8ca-oliver.sang@intel.com
Acked-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-01-24 10:00:22 +00:00
|
|
|
|
2025-05-14 11:41:52 -07:00
|
|
|
|
mm: memcg: optimize parent iteration in memcg_rstat_updated()
In memcg_rstat_updated(), we iterate the memcg being updated and its
parents to update memcg->vmstats_percpu->stats_updates in the fast path
(i.e. no atomic updates). According to my math, this is 3 memory loads
(and potentially 3 cache misses) per memcg:
- Load the address of memcg->vmstats_percpu.
- Load vmstats_percpu->stats_updates (based on some percpu calculation).
- Load the address of the parent memcg.
Avoid most of the cache misses by caching a pointer from each struct
memcg_vmstats_percpu to its parent on the corresponding CPU. In this
case, for the first memcg we have 2 memory loads (same as above):
- Load the address of memcg->vmstats_percpu.
- Load vmstats_percpu->stats_updates (based on some percpu calculation).
Then for each additional memcg, we need a single load to get the
parent's stats_updates directly. This reduces the number of loads from
O(3N) to O(2+N) -- where N is the number of memcgs we need to iterate.
Additionally, stash a pointer to memcg->vmstats in each struct
memcg_vmstats_percpu such that we can access the atomic counter that all
CPUs fold into, memcg->vmstats->stats_updates.
memcg_should_flush_stats() is changed to memcg_vmstats_needs_flush() to
accept a struct memcg_vmstats pointer accordingly.
In struct memcg_vmstats_percpu, make sure both pointers together with
stats_updates live on the same cacheline. Finally, update
mem_cgroup_alloc() to take in a parent pointer and initialize the new
cache pointers on each CPU. The percpu loop in mem_cgroup_alloc() may
look concerning, but there are multiple similar loops in the cgroup
creation path (e.g. cgroup_rstat_init()), most of which are hidden
within alloc_percpu().
According to Oliver's testing [1], this fixes multiple 30-38%
regressions in vm-scalability, will-it-scale-tlb_flush2, and
will-it-scale-fallocate1. This comes at a cost of 2 more pointers per
CPU (<2KB on a machine with 128 CPUs).
[1] https://lore.kernel.org/lkml/ZbDJsfsZt2ITyo61@xsang-OptiPlex-9020/
[yosryahmed@google.com: fix struct memcg_vmstats_percpu size and alignment]
Link: https://lkml.kernel.org/r/20240203044612.1234216-1-yosryahmed@google.com
Link: https://lkml.kernel.org/r/20240124100023.660032-1-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Fixes: 8d59d2214c23 ("mm: memcg: make stats flushing threshold per-memcg")
Tested-by: kernel test robot <oliver.sang@intel.com>
Reported-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202401221624.cb53a8ca-oliver.sang@intel.com
Acked-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-01-24 10:00:22 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-01-12 17:19:04 -08:00
|
|
|
|
2024-05-07 13:23:24 +00:00
|
|
|
|
2016-01-20 15:02:53 -08:00
|
|
|
|
2009-09-23 15:56:37 -07:00
|
|
|
|
2016-01-20 15:02:53 -08:00
|
|
|
|
|
|
|
|
|
2009-01-07 18:08:05 -08:00
|
|
|
|
2016-01-14 15:21:29 -08:00
|
|
|
|
memcg: add memory.pressure_level events
With this patch userland applications that want to maintain the
interactivity/memory allocation cost can use the pressure level
notifications. The levels are defined like this:
The "low" level means that the system is reclaiming memory for new
allocations. Monitoring this reclaiming activity might be useful for
maintaining cache level. Upon notification, the program (typically
"Activity Manager") might analyze vmstat and act in advance (i.e.
prematurely shutdown unimportant services).
The "medium" level means that the system is experiencing medium memory
pressure, the system might be making swap, paging out active file
caches, etc. Upon this event applications may decide to further analyze
vmstat/zoneinfo/memcg or internal memory usage statistics and free any
resources that can be easily reconstructed or re-read from a disk.
The "critical" level means that the system is actively thrashing, it is
about to out of memory (OOM) or even the in-kernel OOM killer is on its
way to trigger. Applications should do whatever they can to help the
system. It might be too late to consult with vmstat or any other
statistics, so it's advisable to take an immediate action.
The events are propagated upward until the event is handled, i.e. the
events are not pass-through. Here is what this means: for example you
have three cgroups: A->B->C. Now you set up an event listener on
cgroups A, B and C, and suppose group C experiences some pressure. In
this situation, only group C will receive the notification, i.e. groups
A and B will not receive it. This is done to avoid excessive
"broadcasting" of messages, which disturbs the system and which is
especially bad if we are low on memory or thrashing. So, organize the
cgroups wisely, or propagate the events manually (or, ask us to
implement the pass-through events, explaining why would you need them.)
Performance wise, the memory pressure notifications feature itself is
lightweight and does not require much of bookkeeping, in contrast to the
rest of memcg features. Unfortunately, as of current memcg
implementation, pages accounting is an inseparable part and cannot be
turned off. The good news is that there are some efforts[1] to improve
the situation; plus, implementing the same, fully API-compatible[2]
interface for CONFIG_MEMCG=n case (e.g. embedded) is also a viable
option, so it will not require any changes on the userland side.
[1] http://permalink.gmane.org/gmane.linux.kernel.cgroups/6291
[2] http://lkml.org/lkml/2013/2/21/454
[akpm@linux-foundation.org: coding-style fixes]
[akpm@linux-foundation.org: fix CONFIG_CGROPUPS=n warnings]
Signed-off-by: Anton Vorontsov <anton.vorontsov@linaro.org>
Acked-by: Kirill A. Shutemov <kirill@shutemov.name>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Glauber Costa <glommer@parallels.com>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Luiz Capitulino <lcapitulino@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Leonid Moiseichuk <leonid.moiseichuk@nokia.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com>
Cc: John Stultz <john.stultz@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-04-29 15:08:31 -07:00
|
|
|
|
mm, memcg: cg2 memory{.swap,}.peak write handlers
Patch series "mm, memcg: cg2 memory{.swap,}.peak write handlers", v7.
This patch (of 2):
Other mechanisms for querying the peak memory usage of either a process or
v1 memory cgroup allow for resetting the high watermark. Restore parity
with those mechanisms, but with a less racy API.
For example:
- Any write to memory.max_usage_in_bytes in a cgroup v1 mount resets
the high watermark.
- writing "5" to the clear_refs pseudo-file in a processes's proc
directory resets the peak RSS.
This change is an evolution of a previous patch, which mostly copied the
cgroup v1 behavior, however, there were concerns about races/ownership
issues with a global reset, so instead this change makes the reset
filedescriptor-local.
Writing any non-empty string to the memory.peak and memory.swap.peak
pseudo-files reset the high watermark to the current usage for subsequent
reads through that same FD.
Notably, following Johannes's suggestion, this implementation moves the
O(FDs that have written) behavior onto the FD write(2) path. Instead, on
the page-allocation path, we simply add one additional watermark to
conditionally bump per-hierarchy level in the page-counter.
Additionally, this takes Longman's suggestion of nesting the
page-charging-path checks for the two watermarks to reduce the number of
common-case comparisons.
This behavior is particularly useful for work scheduling systems that need
to track memory usage of worker processes/cgroups per-work-item. Since
memory can't be squeezed like CPU can (the OOM-killer has opinions), these
systems need to track the peak memory usage to compute system/container
fullness when binpacking workitems.
Most notably, Vimeo's use-case involves a system that's doing global
binpacking across many Kubernetes pods/containers, and while we can use
PSI for some local decisions about overload, we strive to avoid packing
workloads too tightly in the first place. To facilitate this, we track
the peak memory usage. However, since we run with long-lived workers (to
amortize startup costs) we need a way to track the high watermark while a
work-item is executing. Polling runs the risk of missing short spikes
that last for timescales below the polling interval, and peak memory
tracking at the cgroup level is otherwise perfect for this use-case.
As this data is used to ensure that binpacked work ends up with sufficient
headroom, this use-case mostly avoids the inaccuracies surrounding
reclaimable memory.
Link: https://lkml.kernel.org/r/20240730231304.761942-1-davidf@vimeo.com
Link: https://lkml.kernel.org/r/20240729143743.34236-1-davidf@vimeo.com
Link: https://lkml.kernel.org/r/20240729143743.34236-2-davidf@vimeo.com
Signed-off-by: David Finkel <davidf@vimeo.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Waiman Long <longman@redhat.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-07-29 10:37:42 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-07-17 19:46:43 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-06-28 21:03:12 +00:00
|
|
|
|
2014-12-12 16:55:10 -08:00
|
|
|
|
2020-08-06 23:20:49 -07:00
|
|
|
|
2015-05-22 17:13:37 -04:00
|
|
|
|
|
|
|
|
|
writeback, memcg: Implement foreign dirty flushing
There's an inherent mismatch between memcg and writeback. The former
trackes ownership per-page while the latter per-inode. This was a
deliberate design decision because honoring per-page ownership in the
writeback path is complicated, may lead to higher CPU and IO overheads
and deemed unnecessary given that write-sharing an inode across
different cgroups isn't a common use-case.
Combined with inode majority-writer ownership switching, this works
well enough in most cases but there are some pathological cases. For
example, let's say there are two cgroups A and B which keep writing to
different but confined parts of the same inode. B owns the inode and
A's memory is limited far below B's. A's dirty ratio can rise enough
to trigger balance_dirty_pages() sleeps but B's can be low enough to
avoid triggering background writeback. A will be slowed down without
a way to make writeback of the dirty pages happen.
This patch implements foreign dirty recording and foreign mechanism so
that when a memcg encounters a condition as above it can trigger
flushes on bdi_writebacks which can clean its pages. Please see the
comment on top of mem_cgroup_track_foreign_dirty_slowpath() for
details.
A reproducer follows.
write-range.c::
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <fcntl.h>
#include <sys/types.h>
static const char *usage = "write-range FILE START SIZE\n";
int main(int argc, char **argv)
{
int fd;
unsigned long start, size, end, pos;
char *endp;
char buf[4096];
if (argc < 4) {
fprintf(stderr, usage);
return 1;
}
fd = open(argv[1], O_WRONLY);
if (fd < 0) {
perror("open");
return 1;
}
start = strtoul(argv[2], &endp, 0);
if (*endp != '\0') {
fprintf(stderr, usage);
return 1;
}
size = strtoul(argv[3], &endp, 0);
if (*endp != '\0') {
fprintf(stderr, usage);
return 1;
}
end = start + size;
while (1) {
for (pos = start; pos < end; ) {
long bread, bwritten = 0;
if (lseek(fd, pos, SEEK_SET) < 0) {
perror("lseek");
return 1;
}
bread = read(0, buf, sizeof(buf) < end - pos ?
sizeof(buf) : end - pos);
if (bread < 0) {
perror("read");
return 1;
}
if (bread == 0)
return 0;
while (bwritten < bread) {
long this;
this = write(fd, buf + bwritten,
bread - bwritten);
if (this < 0) {
perror("write");
return 1;
}
bwritten += this;
pos += bwritten;
}
}
}
}
repro.sh::
#!/bin/bash
set -e
set -x
sysctl -w vm.dirty_expire_centisecs=300000
sysctl -w vm.dirty_writeback_centisecs=300000
sysctl -w vm.dirtytime_expire_seconds=300000
echo 3 > /proc/sys/vm/drop_caches
TEST=/sys/fs/cgroup/test
A=$TEST/A
B=$TEST/B
mkdir -p $A $B
echo "+memory +io" > $TEST/cgroup.subtree_control
echo $((1<<30)) > $A/memory.high
echo $((32<<30)) > $B/memory.high
rm -f testfile
touch testfile
fallocate -l 4G testfile
echo "Starting B"
(echo $BASHPID > $B/cgroup.procs
pv -q --rate-limit 70M < /dev/urandom | ./write-range testfile $((2<<30)) $((2<<30))) &
echo "Waiting 10s to ensure B claims the testfile inode"
sleep 5
sync
sleep 5
sync
echo "Starting A"
(echo $BASHPID > $A/cgroup.procs
pv < /dev/urandom | ./write-range testfile 0 $((2<<30)))
v2: Added comments explaining why the specific intervals are being used.
v3: Use 0 @nr when calling cgroup_writeback_by_id() to use best-effort
flushing while avoding possible livelocks.
v4: Use get_jiffies_64() and time_before/after64() instead of raw
jiffies_64 and arthimetic comparisons as suggested by Jan.
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-08-26 09:06:56 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-09-23 15:38:15 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-05-22 17:13:37 -04:00
|
|
|
|
mm: multi-gen LRU: groundwork
Evictable pages are divided into multiple generations for each lruvec.
The youngest generation number is stored in lrugen->max_seq for both
anon and file types as they are aged on an equal footing. The oldest
generation numbers are stored in lrugen->min_seq[] separately for anon
and file types as clean file pages can be evicted regardless of swap
constraints. These three variables are monotonically increasing.
Generation numbers are truncated into order_base_2(MAX_NR_GENS+1) bits
in order to fit into the gen counter in folio->flags. Each truncated
generation number is an index to lrugen->lists[]. The sliding window
technique is used to track at least MIN_NR_GENS and at most
MAX_NR_GENS generations. The gen counter stores a value within [1,
MAX_NR_GENS] while a page is on one of lrugen->lists[]. Otherwise it
stores 0.
There are two conceptually independent procedures: "the aging", which
produces young generations, and "the eviction", which consumes old
generations. They form a closed-loop system, i.e., "the page reclaim".
Both procedures can be invoked from userspace for the purposes of working
set estimation and proactive reclaim. These techniques are commonly used
to optimize job scheduling (bin packing) in data centers [1][2].
To avoid confusion, the terms "hot" and "cold" will be applied to the
multi-gen LRU, as a new convention; the terms "active" and "inactive" will
be applied to the active/inactive LRU, as usual.
The protection of hot pages and the selection of cold pages are based
on page access channels and patterns. There are two access channels:
one through page tables and the other through file descriptors. The
protection of the former channel is by design stronger because:
1. The uncertainty in determining the access patterns of the former
channel is higher due to the approximation of the accessed bit.
2. The cost of evicting the former channel is higher due to the TLB
flushes required and the likelihood of encountering the dirty bit.
3. The penalty of underprotecting the former channel is higher because
applications usually do not prepare themselves for major page
faults like they do for blocked I/O. E.g., GUI applications
commonly use dedicated I/O threads to avoid blocking rendering
threads.
There are also two access patterns: one with temporal locality and the
other without. For the reasons listed above, the former channel is
assumed to follow the former pattern unless VM_SEQ_READ or VM_RAND_READ is
present; the latter channel is assumed to follow the latter pattern unless
outlying refaults have been observed [3][4].
The next patch will address the "outlying refaults". Three macros, i.e.,
LRU_REFS_WIDTH, LRU_REFS_PGOFF and LRU_REFS_MASK, used later are added in
this patch to make the entire patchset less diffy.
A page is added to the youngest generation on faulting. The aging needs
to check the accessed bit at least twice before handing this page over to
the eviction. The first check takes care of the accessed bit set on the
initial fault; the second check makes sure this page has not been used
since then. This protocol, AKA second chance, requires a minimum of two
generations, hence MIN_NR_GENS.
[1] https://dl.acm.org/doi/10.1145/3297858.3304053
[2] https://dl.acm.org/doi/10.1145/3503222.3507731
[3] https://lwn.net/Articles/495543/
[4] https://lwn.net/Articles/815342/
Link: https://lkml.kernel.org/r/20220918080010.2920238-6-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 02:00:02 -06:00
|
|
|
|
2016-01-20 15:02:53 -08:00
|
|
|
|
|
|
|
|
|
2018-08-02 15:36:01 -07:00
|
|
|
|
2017-03-09 16:17:26 -08:00
|
|
|
|
2020-05-07 18:35:43 -07:00
|
|
|
|
2013-02-22 16:34:52 -08:00
|
|
|
|
|
|
|
|
|
2016-01-20 15:02:53 -08:00
|
|
|
|
|
|
|
|
|
2013-02-22 16:34:52 -08:00
|
|
|
|
2016-01-20 15:02:53 -08:00
|
|
|
|
2020-10-17 16:13:40 -07:00
|
|
|
|
2025-02-27 23:58:06 -08:00
|
|
|
|
2013-02-22 16:34:52 -08:00
|
|
|
|
2020-10-17 16:13:40 -07:00
|
|
|
|
mm: memcg: optimize parent iteration in memcg_rstat_updated()
In memcg_rstat_updated(), we iterate the memcg being updated and its
parents to update memcg->vmstats_percpu->stats_updates in the fast path
(i.e. no atomic updates). According to my math, this is 3 memory loads
(and potentially 3 cache misses) per memcg:
- Load the address of memcg->vmstats_percpu.
- Load vmstats_percpu->stats_updates (based on some percpu calculation).
- Load the address of the parent memcg.
Avoid most of the cache misses by caching a pointer from each struct
memcg_vmstats_percpu to its parent on the corresponding CPU. In this
case, for the first memcg we have 2 memory loads (same as above):
- Load the address of memcg->vmstats_percpu.
- Load vmstats_percpu->stats_updates (based on some percpu calculation).
Then for each additional memcg, we need a single load to get the
parent's stats_updates directly. This reduces the number of loads from
O(3N) to O(2+N) -- where N is the number of memcgs we need to iterate.
Additionally, stash a pointer to memcg->vmstats in each struct
memcg_vmstats_percpu such that we can access the atomic counter that all
CPUs fold into, memcg->vmstats->stats_updates.
memcg_should_flush_stats() is changed to memcg_vmstats_needs_flush() to
accept a struct memcg_vmstats pointer accordingly.
In struct memcg_vmstats_percpu, make sure both pointers together with
stats_updates live on the same cacheline. Finally, update
mem_cgroup_alloc() to take in a parent pointer and initialize the new
cache pointers on each CPU. The percpu loop in mem_cgroup_alloc() may
look concerning, but there are multiple similar loops in the cgroup
creation path (e.g. cgroup_rstat_init()), most of which are hidden
within alloc_percpu().
According to Oliver's testing [1], this fixes multiple 30-38%
regressions in vm-scalability, will-it-scale-tlb_flush2, and
will-it-scale-fallocate1. This comes at a cost of 2 more pointers per
CPU (<2KB on a machine with 128 CPUs).
[1] https://lore.kernel.org/lkml/ZbDJsfsZt2ITyo61@xsang-OptiPlex-9020/
[yosryahmed@google.com: fix struct memcg_vmstats_percpu size and alignment]
Link: https://lkml.kernel.org/r/20240203044612.1234216-1-yosryahmed@google.com
Link: https://lkml.kernel.org/r/20240124100023.660032-1-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Fixes: 8d59d2214c23 ("mm: memcg: make stats flushing threshold per-memcg")
Tested-by: kernel test robot <oliver.sang@intel.com>
Reported-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202401221624.cb53a8ca-oliver.sang@intel.com
Acked-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-01-24 10:00:22 +00:00
|
|
|
|
2020-10-17 16:13:40 -07:00
|
|
|
|
2020-05-07 18:35:43 -07:00
|
|
|
|
|
|
|
|
|
2013-02-22 16:34:52 -08:00
|
|
|
|
2020-06-01 21:49:49 -07:00
|
|
|
|
2024-06-24 17:58:54 -07:00
|
|
|
|
2024-07-01 11:31:15 -04:00
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
2024-08-23 16:27:06 +00:00
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
2020-06-01 21:49:52 -07:00
|
|
|
|
2016-01-20 15:02:53 -08:00
|
|
|
|
2023-03-06 23:41:36 +08:00
|
|
|
|
2020-12-14 19:06:49 -08:00
|
|
|
|
2025-02-27 23:58:06 -08:00
|
|
|
|
2024-07-26 20:31:08 +00:00
|
|
|
|
2024-06-28 21:03:13 +00:00
|
|
|
|
2025-02-27 23:58:07 -08:00
|
|
|
|
2024-06-28 21:03:13 +00:00
|
|
|
|
2024-07-26 20:31:08 +00:00
|
|
|
|
|
|
|
|
|
2024-06-28 21:03:13 +00:00
|
|
|
|
2009-01-07 18:08:07 -08:00
|
|
|
|
2024-05-01 10:26:13 -07:00
|
|
|
|
2022-09-07 04:35:37 +00:00
|
|
|
|
2024-07-26 20:31:08 +00:00
|
|
|
|
|
|
|
|
|
2024-06-28 21:03:13 +00:00
|
|
|
|
2024-07-26 20:31:08 +00:00
|
|
|
|
|
|
|
|
|
2024-06-28 21:03:13 +00:00
|
|
|
|
2016-01-20 15:02:53 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-02-27 23:58:06 -08:00
|
|
|
|
2016-01-14 15:21:34 -08:00
|
|
|
|
2016-01-14 15:21:29 -08:00
|
|
|
|
2023-02-10 15:47:31 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-01-20 15:02:53 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-07-20 15:44:57 -07:00
|
|
|
|
2016-01-20 15:02:53 -08:00
|
|
|
|
2016-10-07 16:57:29 -07:00
|
|
|
|
|
|
|
|
|
2022-03-22 14:41:15 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, memcg: assign memcg-aware shrinkers bitmap to memcg
Imagine a big node with many cpus, memory cgroups and containers. Let
we have 200 containers, every container has 10 mounts, and 10 cgroups.
All container tasks don't touch foreign containers mounts. If there is
intensive pages write, and global reclaim happens, a writing task has to
iterate over all memcgs to shrink slab, before it's able to go to
shrink_page_list().
Iteration over all the memcg slabs is very expensive: the task has to
visit 200 * 10 = 2000 shrinkers for every memcg, and since there are
2000 memcgs, the total calls are 2000 * 2000 = 4000000.
So, the shrinker makes 4 million do_shrink_slab() calls just to try to
isolate SWAP_CLUSTER_MAX pages in one of the actively writing memcg via
shrink_page_list(). I've observed a node spending almost 100% in
kernel, making useless iteration over already shrinked slab.
This patch adds bitmap of memcg-aware shrinkers to memcg. The size of
the bitmap depends on bitmap_nr_ids, and during memcg life it's
maintained to be enough to fit bitmap_nr_ids shrinkers. Every bit in
the map is related to corresponding shrinker id.
Next patches will maintain set bit only for really charged memcg. This
will allow shrink_slab() to increase its performance in significant way.
See the last patch for the numbers.
[ktkhai@virtuozzo.com: v9]
Link: http://lkml.kernel.org/r/153112549031.4097.3576147070498769979.stgit@localhost.localdomain
[ktkhai@virtuozzo.com: add comment to mem_cgroup_css_online()]
Link: http://lkml.kernel.org/r/521f9e5f-c436-b388-fe83-4dc870bfb489@virtuozzo.com
Link: http://lkml.kernel.org/r/153063056619.1818.12550500883688681076.stgit@localhost.localdomain
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Vladimir Davydov <vdavydov.dev@gmail.com>
Tested-by: Shakeel Butt <shakeelb@google.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Guenter Roeck <linux@roeck-us.net>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Josef Bacik <jbacik@fb.com>
Cc: Li RongQing <lirongqing@baidu.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Matthias Kaehlcke <mka@chromium.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Philippe Ombredanne <pombredanne@nexb.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Sahitya Tummala <stummala@codeaurora.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Waiman Long <longman@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-17 15:47:37 -07:00
|
|
|
|
2021-05-04 18:36:23 -07:00
|
|
|
|
mm, memcg: assign memcg-aware shrinkers bitmap to memcg
Imagine a big node with many cpus, memory cgroups and containers. Let
we have 200 containers, every container has 10 mounts, and 10 cgroups.
All container tasks don't touch foreign containers mounts. If there is
intensive pages write, and global reclaim happens, a writing task has to
iterate over all memcgs to shrink slab, before it's able to go to
shrink_page_list().
Iteration over all the memcg slabs is very expensive: the task has to
visit 200 * 10 = 2000 shrinkers for every memcg, and since there are
2000 memcgs, the total calls are 2000 * 2000 = 4000000.
So, the shrinker makes 4 million do_shrink_slab() calls just to try to
isolate SWAP_CLUSTER_MAX pages in one of the actively writing memcg via
shrink_page_list(). I've observed a node spending almost 100% in
kernel, making useless iteration over already shrinked slab.
This patch adds bitmap of memcg-aware shrinkers to memcg. The size of
the bitmap depends on bitmap_nr_ids, and during memcg life it's
maintained to be enough to fit bitmap_nr_ids shrinkers. Every bit in
the map is related to corresponding shrinker id.
Next patches will maintain set bit only for really charged memcg. This
will allow shrink_slab() to increase its performance in significant way.
See the last patch for the numbers.
[ktkhai@virtuozzo.com: v9]
Link: http://lkml.kernel.org/r/153112549031.4097.3576147070498769979.stgit@localhost.localdomain
[ktkhai@virtuozzo.com: add comment to mem_cgroup_css_online()]
Link: http://lkml.kernel.org/r/521f9e5f-c436-b388-fe83-4dc870bfb489@virtuozzo.com
Link: http://lkml.kernel.org/r/153063056619.1818.12550500883688681076.stgit@localhost.localdomain
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Vladimir Davydov <vdavydov.dev@gmail.com>
Tested-by: Shakeel Butt <shakeelb@google.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Guenter Roeck <linux@roeck-us.net>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Josef Bacik <jbacik@fb.com>
Cc: Li RongQing <lirongqing@baidu.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Matthias Kaehlcke <mka@chromium.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Philippe Ombredanne <pombredanne@nexb.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Sahitya Tummala <stummala@codeaurora.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Waiman Long <longman@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-17 15:47:37 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-03-22 14:41:15 -07:00
|
|
|
|
|
|
|
|
|
mm, memcg: assign memcg-aware shrinkers bitmap to memcg
Imagine a big node with many cpus, memory cgroups and containers. Let
we have 200 containers, every container has 10 mounts, and 10 cgroups.
All container tasks don't touch foreign containers mounts. If there is
intensive pages write, and global reclaim happens, a writing task has to
iterate over all memcgs to shrink slab, before it's able to go to
shrink_page_list().
Iteration over all the memcg slabs is very expensive: the task has to
visit 200 * 10 = 2000 shrinkers for every memcg, and since there are
2000 memcgs, the total calls are 2000 * 2000 = 4000000.
So, the shrinker makes 4 million do_shrink_slab() calls just to try to
isolate SWAP_CLUSTER_MAX pages in one of the actively writing memcg via
shrink_page_list(). I've observed a node spending almost 100% in
kernel, making useless iteration over already shrinked slab.
This patch adds bitmap of memcg-aware shrinkers to memcg. The size of
the bitmap depends on bitmap_nr_ids, and during memcg life it's
maintained to be enough to fit bitmap_nr_ids shrinkers. Every bit in
the map is related to corresponding shrinker id.
Next patches will maintain set bit only for really charged memcg. This
will allow shrink_slab() to increase its performance in significant way.
See the last patch for the numbers.
[ktkhai@virtuozzo.com: v9]
Link: http://lkml.kernel.org/r/153112549031.4097.3576147070498769979.stgit@localhost.localdomain
[ktkhai@virtuozzo.com: add comment to mem_cgroup_css_online()]
Link: http://lkml.kernel.org/r/521f9e5f-c436-b388-fe83-4dc870bfb489@virtuozzo.com
Link: http://lkml.kernel.org/r/153063056619.1818.12550500883688681076.stgit@localhost.localdomain
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Vladimir Davydov <vdavydov.dev@gmail.com>
Tested-by: Shakeel Butt <shakeelb@google.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Guenter Roeck <linux@roeck-us.net>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Josef Bacik <jbacik@fb.com>
Cc: Li RongQing <lirongqing@baidu.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Matthias Kaehlcke <mka@chromium.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Philippe Ombredanne <pombredanne@nexb.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Sahitya Tummala <stummala@codeaurora.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Waiman Long <longman@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-17 15:47:37 -07:00
|
|
|
|
2024-01-26 21:19:25 +00:00
|
|
|
|
2021-09-02 14:55:04 -07:00
|
|
|
|
2023-06-03 15:21:16 +08:00
|
|
|
|
mm: multi-gen LRU: per-node lru_gen_folio lists
For each node, memcgs are divided into two generations: the old and
the young. For each generation, memcgs are randomly sharded into
multiple bins to improve scalability. For each bin, an RCU hlist_nulls
is virtually divided into three segments: the head, the tail and the
default.
An onlining memcg is added to the tail of a random bin in the old
generation. The eviction starts at the head of a random bin in the old
generation. The per-node memcg generation counter, whose reminder (mod
2) indexes the old generation, is incremented when all its bins become
empty.
There are four operations:
1. MEMCG_LRU_HEAD, which moves an memcg to the head of a random bin in
its current generation (old or young) and updates its "seg" to
"head";
2. MEMCG_LRU_TAIL, which moves an memcg to the tail of a random bin in
its current generation (old or young) and updates its "seg" to
"tail";
3. MEMCG_LRU_OLD, which moves an memcg to the head of a random bin in
the old generation, updates its "gen" to "old" and resets its "seg"
to "default";
4. MEMCG_LRU_YOUNG, which moves an memcg to the tail of a random bin
in the young generation, updates its "gen" to "young" and resets
its "seg" to "default".
The events that trigger the above operations are:
1. Exceeding the soft limit, which triggers MEMCG_LRU_HEAD;
2. The first attempt to reclaim an memcg below low, which triggers
MEMCG_LRU_TAIL;
3. The first attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_TAIL;
4. The second attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_YOUNG;
5. Attempting to reclaim an memcg below min, which triggers
MEMCG_LRU_YOUNG;
6. Finishing the aging on the eviction path, which triggers
MEMCG_LRU_YOUNG;
7. Offlining an memcg, which triggers MEMCG_LRU_OLD.
Note that memcg LRU only applies to global reclaim, and the
round-robin incrementing of their max_seq counters ensures the
eventual fairness to all eligible memcgs. For memcg reclaim, it still
relies on mem_cgroup_iter().
Link: https://lkml.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-21 21:19:04 -07:00
|
|
|
|
2023-08-23 15:54:30 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-08-09 10:26:18 -07:00
|
|
|
|
2023-08-23 15:54:30 -07:00
|
|
|
|
2014-10-02 16:16:57 -07:00
|
|
|
|
2022-03-22 14:41:15 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-02-07 00:13:50 -08:00
|
|
|
|
|
|
|
|
|
2013-08-08 20:11:23 -04:00
|
|
|
|
2008-02-07 00:14:28 -08:00
|
|
|
|
2013-08-08 20:11:23 -04:00
|
|
|
|
2013-11-22 18:20:42 -05:00
|
|
|
|
2024-06-24 17:58:58 -07:00
|
|
|
|
2009-04-02 16:57:26 -07:00
|
|
|
|
2018-06-07 17:07:46 -07:00
|
|
|
|
mm: memory.low hierarchical behavior
This patch aims to address an issue in current memory.low semantics,
which makes it hard to use it in a hierarchy, where some leaf memory
cgroups are more valuable than others.
For example, there are memcgs A, A/B, A/C, A/D and A/E:
A A/memory.low = 2G, A/memory.current = 6G
//\\
BC DE B/memory.low = 3G B/memory.current = 2G
C/memory.low = 1G C/memory.current = 2G
D/memory.low = 0 D/memory.current = 2G
E/memory.low = 10G E/memory.current = 0
If we apply memory pressure, B, C and D are reclaimed at the same pace
while A's usage exceeds 2G. This is obviously wrong, as B's usage is
fully below B's memory.low, and C has 1G of protection as well. Also, A
is pushed to the size, which is less than A's 2G memory.low, which is
also wrong.
A simple bash script (provided below) can be used to reproduce
the problem. Current results are:
A: 1430097920
A/B: 711929856
A/C: 717426688
A/D: 741376
A/E: 0
To address the issue a concept of effective memory.low is introduced.
Effective memory.low is always equal or less than original memory.low.
In a case, when there is no memory.low overcommittment (and also for
top-level cgroups), these two values are equal.
Otherwise it's a part of parent's effective memory.low, calculated as a
cgroup's memory.low usage divided by sum of sibling's memory.low usages
(under memory.low usage I mean the size of actually protected memory:
memory.current if memory.current < memory.low, 0 otherwise). It's
necessary to track the actual usage, because otherwise an empty cgroup
with memory.low set (A/E in my example) will affect actual memory
distribution, which makes no sense. To avoid traversing the cgroup tree
twice, page_counters code is reused.
Calculating effective memory.low can be done in the reclaim path, as we
conveniently traversing the cgroup tree from top to bottom and check
memory.low on each level. So, it's a perfect place to calculate
effective memory low and save it to use it for children cgroups.
This also eliminates a need to traverse the cgroup tree from bottom to
top each time to check if parent's guarantee is not exceeded.
Setting/resetting effective memory.low is intentionally racy, but it's
fine and shouldn't lead to any significant differences in actual memory
distribution.
With this patch applied results are matching the expectations:
A: 2147930112
A/B: 1428721664
A/C: 718393344
A/D: 815104
A/E: 0
Test script:
#!/bin/bash
CGPATH="/sys/fs/cgroup"
truncate /file1 --size 2G
truncate /file2 --size 2G
truncate /file3 --size 2G
truncate /file4 --size 50G
mkdir "${CGPATH}/A"
echo "+memory" > "${CGPATH}/A/cgroup.subtree_control"
mkdir "${CGPATH}/A/B" "${CGPATH}/A/C" "${CGPATH}/A/D" "${CGPATH}/A/E"
echo 2G > "${CGPATH}/A/memory.low"
echo 3G > "${CGPATH}/A/B/memory.low"
echo 1G > "${CGPATH}/A/C/memory.low"
echo 0 > "${CGPATH}/A/D/memory.low"
echo 10G > "${CGPATH}/A/E/memory.low"
echo $$ > "${CGPATH}/A/B/cgroup.procs" && vmtouch -qt /file1
echo $$ > "${CGPATH}/A/C/cgroup.procs" && vmtouch -qt /file2
echo $$ > "${CGPATH}/A/D/cgroup.procs" && vmtouch -qt /file3
echo $$ > "${CGPATH}/cgroup.procs" && vmtouch -qt /file4
echo "A: " `cat "${CGPATH}/A/memory.current"`
echo "A/B: " `cat "${CGPATH}/A/B/memory.current"`
echo "A/C: " `cat "${CGPATH}/A/C/memory.current"`
echo "A/D: " `cat "${CGPATH}/A/D/memory.current"`
echo "A/E: " `cat "${CGPATH}/A/E/memory.current"`
rmdir "${CGPATH}/A/B" "${CGPATH}/A/C" "${CGPATH}/A/D" "${CGPATH}/A/E"
rmdir "${CGPATH}/A"
rm /file1 /file2 /file3 /file4
Link: http://lkml.kernel.org/r/20180405185921.4942-2-guro@fb.com
Signed-off-by: Roman Gushchin <guro@fb.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-06-07 17:06:22 -07:00
|
|
|
|
2017-09-06 16:21:47 -07:00
|
|
|
|
2023-11-30 11:40:20 -08:00
|
|
|
|
|
|
|
|
|
2016-01-20 15:02:24 -08:00
|
|
|
|
2021-05-04 18:36:42 -07:00
|
|
|
|
2015-05-22 17:13:37 -04:00
|
|
|
|
mm: multi-gen LRU: per-node lru_gen_folio lists
For each node, memcgs are divided into two generations: the old and
the young. For each generation, memcgs are randomly sharded into
multiple bins to improve scalability. For each bin, an RCU hlist_nulls
is virtually divided into three segments: the head, the tail and the
default.
An onlining memcg is added to the tail of a random bin in the old
generation. The eviction starts at the head of a random bin in the old
generation. The per-node memcg generation counter, whose reminder (mod
2) indexes the old generation, is incremented when all its bins become
empty.
There are four operations:
1. MEMCG_LRU_HEAD, which moves an memcg to the head of a random bin in
its current generation (old or young) and updates its "seg" to
"head";
2. MEMCG_LRU_TAIL, which moves an memcg to the tail of a random bin in
its current generation (old or young) and updates its "seg" to
"tail";
3. MEMCG_LRU_OLD, which moves an memcg to the head of a random bin in
the old generation, updates its "gen" to "old" and resets its "seg"
to "default";
4. MEMCG_LRU_YOUNG, which moves an memcg to the tail of a random bin
in the young generation, updates its "gen" to "young" and resets
its "seg" to "default".
The events that trigger the above operations are:
1. Exceeding the soft limit, which triggers MEMCG_LRU_HEAD;
2. The first attempt to reclaim an memcg below low, which triggers
MEMCG_LRU_TAIL;
3. The first attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_TAIL;
4. The second attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_YOUNG;
5. Attempting to reclaim an memcg below min, which triggers
MEMCG_LRU_YOUNG;
6. Finishing the aging on the eviction path, which triggers
MEMCG_LRU_YOUNG;
7. Offlining an memcg, which triggers MEMCG_LRU_OLD.
Note that memcg LRU only applies to global reclaim, and the
round-robin incrementing of their max_seq counters ensures the
eventual fairness to all eligible memcgs. For memcg reclaim, it still
relies on mem_cgroup_iter().
Link: https://lkml.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-21 21:19:04 -07:00
|
|
|
|
2016-07-20 15:44:57 -07:00
|
|
|
|
2018-10-26 15:03:23 -07:00
|
|
|
|
|
|
|
|
|
2016-07-20 15:44:57 -07:00
|
|
|
|
2008-02-07 00:14:28 -08:00
|
|
|
|
|
|
|
|
|
2015-12-29 14:54:10 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: multi-gen LRU: per-node lru_gen_folio lists
For each node, memcgs are divided into two generations: the old and
the young. For each generation, memcgs are randomly sharded into
multiple bins to improve scalability. For each bin, an RCU hlist_nulls
is virtually divided into three segments: the head, the tail and the
default.
An onlining memcg is added to the tail of a random bin in the old
generation. The eviction starts at the head of a random bin in the old
generation. The per-node memcg generation counter, whose reminder (mod
2) indexes the old generation, is incremented when all its bins become
empty.
There are four operations:
1. MEMCG_LRU_HEAD, which moves an memcg to the head of a random bin in
its current generation (old or young) and updates its "seg" to
"head";
2. MEMCG_LRU_TAIL, which moves an memcg to the tail of a random bin in
its current generation (old or young) and updates its "seg" to
"tail";
3. MEMCG_LRU_OLD, which moves an memcg to the head of a random bin in
the old generation, updates its "gen" to "old" and resets its "seg"
to "default";
4. MEMCG_LRU_YOUNG, which moves an memcg to the tail of a random bin
in the young generation, updates its "gen" to "young" and resets
its "seg" to "default".
The events that trigger the above operations are:
1. Exceeding the soft limit, which triggers MEMCG_LRU_HEAD;
2. The first attempt to reclaim an memcg below low, which triggers
MEMCG_LRU_TAIL;
3. The first attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_TAIL;
4. The second attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_YOUNG;
5. Attempting to reclaim an memcg below min, which triggers
MEMCG_LRU_YOUNG;
6. Finishing the aging on the eviction path, which triggers
MEMCG_LRU_YOUNG;
7. Offlining an memcg, which triggers MEMCG_LRU_OLD.
Note that memcg LRU only applies to global reclaim, and the
round-robin incrementing of their max_seq counters ensures the
eventual fairness to all eligible memcgs. For memcg reclaim, it still
relies on mem_cgroup_iter().
Link: https://lkml.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-21 21:19:04 -07:00
|
|
|
|
2015-12-29 14:54:10 -08:00
|
|
|
|
|
|
|
|
|
2013-08-08 20:11:23 -04:00
|
|
|
|
2008-02-07 00:13:50 -08:00
|
|
|
|
2013-08-08 20:11:23 -04:00
|
|
|
|
writeback, memcg: Implement foreign dirty flushing
There's an inherent mismatch between memcg and writeback. The former
trackes ownership per-page while the latter per-inode. This was a
deliberate design decision because honoring per-page ownership in the
writeback path is complicated, may lead to higher CPU and IO overheads
and deemed unnecessary given that write-sharing an inode across
different cgroups isn't a common use-case.
Combined with inode majority-writer ownership switching, this works
well enough in most cases but there are some pathological cases. For
example, let's say there are two cgroups A and B which keep writing to
different but confined parts of the same inode. B owns the inode and
A's memory is limited far below B's. A's dirty ratio can rise enough
to trigger balance_dirty_pages() sleeps but B's can be low enough to
avoid triggering background writeback. A will be slowed down without
a way to make writeback of the dirty pages happen.
This patch implements foreign dirty recording and foreign mechanism so
that when a memcg encounters a condition as above it can trigger
flushes on bdi_writebacks which can clean its pages. Please see the
comment on top of mem_cgroup_track_foreign_dirty_slowpath() for
details.
A reproducer follows.
write-range.c::
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <fcntl.h>
#include <sys/types.h>
static const char *usage = "write-range FILE START SIZE\n";
int main(int argc, char **argv)
{
int fd;
unsigned long start, size, end, pos;
char *endp;
char buf[4096];
if (argc < 4) {
fprintf(stderr, usage);
return 1;
}
fd = open(argv[1], O_WRONLY);
if (fd < 0) {
perror("open");
return 1;
}
start = strtoul(argv[2], &endp, 0);
if (*endp != '\0') {
fprintf(stderr, usage);
return 1;
}
size = strtoul(argv[3], &endp, 0);
if (*endp != '\0') {
fprintf(stderr, usage);
return 1;
}
end = start + size;
while (1) {
for (pos = start; pos < end; ) {
long bread, bwritten = 0;
if (lseek(fd, pos, SEEK_SET) < 0) {
perror("lseek");
return 1;
}
bread = read(0, buf, sizeof(buf) < end - pos ?
sizeof(buf) : end - pos);
if (bread < 0) {
perror("read");
return 1;
}
if (bread == 0)
return 0;
while (bwritten < bread) {
long this;
this = write(fd, buf + bwritten,
bread - bwritten);
if (this < 0) {
perror("write");
return 1;
}
bwritten += this;
pos += bwritten;
}
}
}
}
repro.sh::
#!/bin/bash
set -e
set -x
sysctl -w vm.dirty_expire_centisecs=300000
sysctl -w vm.dirty_writeback_centisecs=300000
sysctl -w vm.dirtytime_expire_seconds=300000
echo 3 > /proc/sys/vm/drop_caches
TEST=/sys/fs/cgroup/test
A=$TEST/A
B=$TEST/B
mkdir -p $A $B
echo "+memory +io" > $TEST/cgroup.subtree_control
echo $((1<<30)) > $A/memory.high
echo $((32<<30)) > $B/memory.high
rm -f testfile
touch testfile
fallocate -l 4G testfile
echo "Starting B"
(echo $BASHPID > $B/cgroup.procs
pv -q --rate-limit 70M < /dev/urandom | ./write-range testfile $((2<<30)) $((2<<30))) &
echo "Waiting 10s to ensure B claims the testfile inode"
sleep 5
sync
sleep 5
sync
echo "Starting A"
(echo $BASHPID > $A/cgroup.procs
pv < /dev/urandom | ./write-range testfile 0 $((2<<30)))
v2: Added comments explaining why the specific intervals are being used.
v3: Use 0 @nr when calling cgroup_writeback_by_id() to use best-effort
flushing while avoding possible livelocks.
v4: Use get_jiffies_64() and time_before/after64() instead of raw
jiffies_64 and arthimetic comparisons as suggested by Jan.
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-08-26 09:06:56 -07:00
|
|
|
|
2009-01-15 13:51:13 -08:00
|
|
|
|
writeback, memcg: Implement foreign dirty flushing
There's an inherent mismatch between memcg and writeback. The former
trackes ownership per-page while the latter per-inode. This was a
deliberate design decision because honoring per-page ownership in the
writeback path is complicated, may lead to higher CPU and IO overheads
and deemed unnecessary given that write-sharing an inode across
different cgroups isn't a common use-case.
Combined with inode majority-writer ownership switching, this works
well enough in most cases but there are some pathological cases. For
example, let's say there are two cgroups A and B which keep writing to
different but confined parts of the same inode. B owns the inode and
A's memory is limited far below B's. A's dirty ratio can rise enough
to trigger balance_dirty_pages() sleeps but B's can be low enough to
avoid triggering background writeback. A will be slowed down without
a way to make writeback of the dirty pages happen.
This patch implements foreign dirty recording and foreign mechanism so
that when a memcg encounters a condition as above it can trigger
flushes on bdi_writebacks which can clean its pages. Please see the
comment on top of mem_cgroup_track_foreign_dirty_slowpath() for
details.
A reproducer follows.
write-range.c::
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <fcntl.h>
#include <sys/types.h>
static const char *usage = "write-range FILE START SIZE\n";
int main(int argc, char **argv)
{
int fd;
unsigned long start, size, end, pos;
char *endp;
char buf[4096];
if (argc < 4) {
fprintf(stderr, usage);
return 1;
}
fd = open(argv[1], O_WRONLY);
if (fd < 0) {
perror("open");
return 1;
}
start = strtoul(argv[2], &endp, 0);
if (*endp != '\0') {
fprintf(stderr, usage);
return 1;
}
size = strtoul(argv[3], &endp, 0);
if (*endp != '\0') {
fprintf(stderr, usage);
return 1;
}
end = start + size;
while (1) {
for (pos = start; pos < end; ) {
long bread, bwritten = 0;
if (lseek(fd, pos, SEEK_SET) < 0) {
perror("lseek");
return 1;
}
bread = read(0, buf, sizeof(buf) < end - pos ?
sizeof(buf) : end - pos);
if (bread < 0) {
perror("read");
return 1;
}
if (bread == 0)
return 0;
while (bwritten < bread) {
long this;
this = write(fd, buf + bwritten,
bread - bwritten);
if (this < 0) {
perror("write");
return 1;
}
bwritten += this;
pos += bwritten;
}
}
}
}
repro.sh::
#!/bin/bash
set -e
set -x
sysctl -w vm.dirty_expire_centisecs=300000
sysctl -w vm.dirty_writeback_centisecs=300000
sysctl -w vm.dirtytime_expire_seconds=300000
echo 3 > /proc/sys/vm/drop_caches
TEST=/sys/fs/cgroup/test
A=$TEST/A
B=$TEST/B
mkdir -p $A $B
echo "+memory +io" > $TEST/cgroup.subtree_control
echo $((1<<30)) > $A/memory.high
echo $((32<<30)) > $B/memory.high
rm -f testfile
touch testfile
fallocate -l 4G testfile
echo "Starting B"
(echo $BASHPID > $B/cgroup.procs
pv -q --rate-limit 70M < /dev/urandom | ./write-range testfile $((2<<30)) $((2<<30))) &
echo "Waiting 10s to ensure B claims the testfile inode"
sleep 5
sync
sleep 5
sync
echo "Starting A"
(echo $BASHPID > $A/cgroup.procs
pv < /dev/urandom | ./write-range testfile 0 $((2<<30)))
v2: Added comments explaining why the specific intervals are being used.
v3: Use 0 @nr when calling cgroup_writeback_by_id() to use best-effort
flushing while avoding possible livelocks.
v4: Use get_jiffies_64() and time_before/after64() instead of raw
jiffies_64 and arthimetic comparisons as suggested by Jan.
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-08-26 09:06:56 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-01-14 15:21:29 -08:00
|
|
|
|
2016-01-14 15:21:34 -08:00
|
|
|
|
2016-01-20 15:02:32 -08:00
|
|
|
|
2024-06-28 21:03:10 +00:00
|
|
|
|
2016-01-20 15:02:44 -08:00
|
|
|
|
2016-01-20 15:02:29 -08:00
|
|
|
|
2023-02-10 15:47:31 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-01-20 15:02:53 -08:00
|
|
|
|
|
|
|
|
|
2024-06-24 17:58:55 -07:00
|
|
|
|
2021-05-04 18:36:23 -07:00
|
|
|
|
2016-01-20 15:02:53 -08:00
|
|
|
|
2008-02-07 00:13:50 -08:00
|
|
|
|
|
|
|
|
|
2014-07-08 18:02:57 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-06-07 17:06:18 -07:00
|
|
|
|
|
|
|
|
|
2024-06-28 21:03:13 +00:00
|
|
|
|
2018-06-07 17:06:18 -07:00
|
|
|
|
|
|
|
|
|
2024-06-28 21:03:13 +00:00
|
|
|
|
2018-06-07 17:07:46 -07:00
|
|
|
|
mm: memory.low hierarchical behavior
This patch aims to address an issue in current memory.low semantics,
which makes it hard to use it in a hierarchy, where some leaf memory
cgroups are more valuable than others.
For example, there are memcgs A, A/B, A/C, A/D and A/E:
A A/memory.low = 2G, A/memory.current = 6G
//\\
BC DE B/memory.low = 3G B/memory.current = 2G
C/memory.low = 1G C/memory.current = 2G
D/memory.low = 0 D/memory.current = 2G
E/memory.low = 10G E/memory.current = 0
If we apply memory pressure, B, C and D are reclaimed at the same pace
while A's usage exceeds 2G. This is obviously wrong, as B's usage is
fully below B's memory.low, and C has 1G of protection as well. Also, A
is pushed to the size, which is less than A's 2G memory.low, which is
also wrong.
A simple bash script (provided below) can be used to reproduce
the problem. Current results are:
A: 1430097920
A/B: 711929856
A/C: 717426688
A/D: 741376
A/E: 0
To address the issue a concept of effective memory.low is introduced.
Effective memory.low is always equal or less than original memory.low.
In a case, when there is no memory.low overcommittment (and also for
top-level cgroups), these two values are equal.
Otherwise it's a part of parent's effective memory.low, calculated as a
cgroup's memory.low usage divided by sum of sibling's memory.low usages
(under memory.low usage I mean the size of actually protected memory:
memory.current if memory.current < memory.low, 0 otherwise). It's
necessary to track the actual usage, because otherwise an empty cgroup
with memory.low set (A/E in my example) will affect actual memory
distribution, which makes no sense. To avoid traversing the cgroup tree
twice, page_counters code is reused.
Calculating effective memory.low can be done in the reclaim path, as we
conveniently traversing the cgroup tree from top to bottom and check
memory.low on each level. So, it's a perfect place to calculate
effective memory low and save it to use it for children cgroups.
This also eliminates a need to traverse the cgroup tree from bottom to
top each time to check if parent's guarantee is not exceeded.
Setting/resetting effective memory.low is intentionally racy, but it's
fine and shouldn't lead to any significant differences in actual memory
distribution.
With this patch applied results are matching the expectations:
A: 2147930112
A/B: 1428721664
A/C: 718393344
A/D: 815104
A/E: 0
Test script:
#!/bin/bash
CGPATH="/sys/fs/cgroup"
truncate /file1 --size 2G
truncate /file2 --size 2G
truncate /file3 --size 2G
truncate /file4 --size 50G
mkdir "${CGPATH}/A"
echo "+memory" > "${CGPATH}/A/cgroup.subtree_control"
mkdir "${CGPATH}/A/B" "${CGPATH}/A/C" "${CGPATH}/A/D" "${CGPATH}/A/E"
echo 2G > "${CGPATH}/A/memory.low"
echo 3G > "${CGPATH}/A/B/memory.low"
echo 1G > "${CGPATH}/A/C/memory.low"
echo 0 > "${CGPATH}/A/D/memory.low"
echo 10G > "${CGPATH}/A/E/memory.low"
echo $$ > "${CGPATH}/A/B/cgroup.procs" && vmtouch -qt /file1
echo $$ > "${CGPATH}/A/C/cgroup.procs" && vmtouch -qt /file2
echo $$ > "${CGPATH}/A/D/cgroup.procs" && vmtouch -qt /file3
echo $$ > "${CGPATH}/cgroup.procs" && vmtouch -qt /file4
echo "A: " `cat "${CGPATH}/A/memory.current"`
echo "A/B: " `cat "${CGPATH}/A/B/memory.current"`
echo "A/C: " `cat "${CGPATH}/A/C/memory.current"`
echo "A/D: " `cat "${CGPATH}/A/D/memory.current"`
echo "A/E: " `cat "${CGPATH}/A/E/memory.current"`
rmdir "${CGPATH}/A/B" "${CGPATH}/A/C" "${CGPATH}/A/D" "${CGPATH}/A/E"
rmdir "${CGPATH}/A"
rm /file1 /file2 /file3 /file4
Link: http://lkml.kernel.org/r/20180405185921.4942-2-guro@fb.com
Signed-off-by: Roman Gushchin <guro@fb.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-06-07 17:06:22 -07:00
|
|
|
|
2020-06-01 21:49:49 -07:00
|
|
|
|
2024-06-24 17:58:54 -07:00
|
|
|
|
2020-06-01 21:49:52 -07:00
|
|
|
|
2015-05-22 18:23:34 -04:00
|
|
|
|
2014-07-08 18:02:57 -04:00
|
|
|
|
|
|
|
|
|
2024-10-26 09:34:07 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcontrol: switch to rstat
Replace the memory controller's custom hierarchical stats code with the
generic rstat infrastructure provided by the cgroup core.
The current implementation does batched upward propagation from the
write side (i.e. as stats change). The per-cpu batches introduce an
error, which is multiplied by the number of subgroups in a tree. In
systems with many CPUs and sizable cgroup trees, the error can be large
enough to confuse users (e.g. 32 batch pages * 32 CPUs * 32 subgroups
results in an error of up to 128M per stat item). This can entirely
swallow allocation bursts inside a workload that the user is expecting
to see reflected in the statistics.
In the past, we've done read-side aggregation, where a memory.stat read
would have to walk the entire subtree and add up per-cpu counts. This
became problematic with lazily-freed cgroups: we could have large
subtrees where most cgroups were entirely idle. Hence the switch to
change-driven upward propagation. Unfortunately, it needed to trade
accuracy for speed due to the write side being so hot.
Rstat combines the best of both worlds: from the write side, it cheaply
maintains a queue of cgroups that have pending changes, so that the read
side can do selective tree aggregation. This way the reported stats
will always be precise and recent as can be, while the aggregation can
skip over potentially large numbers of idle cgroups.
The way rstat works is that it implements a tree for tracking cgroups
with pending local changes, as well as a flush function that walks the
tree upwards. The controller then drives this by 1) telling rstat when
a local cgroup stat changes (e.g. mod_memcg_state) and 2) when a flush
is required to get uptodate hierarchy stats for a given subtree (e.g.
when memory.stat is read). The controller also provides a flush
callback that is called during the rstat flush walk for each cgroup and
aggregates its local per-cpu counters and propagates them upwards.
This adds a second vmstats to struct mem_cgroup (MEMCG_NR_STAT +
NR_VM_EVENT_ITEMS) to track pending subtree deltas during upward
aggregation. It removes 3 words from the per-cpu data. It eliminates
memcg_exact_page_state(), since memcg_page_state() is now exact.
[akpm@linux-foundation.org: merge fix]
[hannes@cmpxchg.org: fix a sleep in atomic section problem]
Link: https://lkml.kernel.org/r/20210315234100.64307-1-hannes@cmpxchg.org
Link: https://lkml.kernel.org/r/20210209163304.77088-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Roman Gushchin <guro@fb.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Balbir Singh <bsingharora@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:26 -07:00
|
|
|
|
2024-10-26 09:34:07 +00:00
|
|
|
|
mm: memcg: use rstat for non-hierarchical stats
Currently, memcg uses rstat to maintain aggregated hierarchical stats.
Counters are maintained for hierarchical stats at each memcg. Rstat
tracks which cgroups have updates on which cpus to keep those counters
fresh on the read-side.
Non-hierarchical stats are currently not covered by rstat. Their per-cpu
counters are summed up on every read, which is expensive. The original
implementation did the same. At some point before rstat, non-hierarchical
aggregated counters were introduced by commit a983b5ebee57 ("mm:
memcontrol: fix excessive complexity in memory.stat reporting"). However,
those counters were updated on the performance critical write-side, which
caused regressions, so they were later removed by commit 815744d75152
("mm: memcontrol: don't batch updates of local VM stats and events"). See
[1] for more detailed history.
Kernel versions in between a983b5ebee57 & 815744d75152 (a year and a half)
enjoyed cheap reads of non-hierarchical stats, specifically on cgroup v1.
When moving to more recent kernels, a performance regression for reading
non-hierarchical stats is observed.
Now that we have rstat, we know exactly which percpu counters have updates
for each stat. We can maintain non-hierarchical counters again, making
reads much more efficient, without affecting the performance critical
write-side. Hence, add non-hierarchical (i.e local) counters for the
stats, and extend rstat flushing to keep those up-to-date.
A caveat is that we now need a stats flush before reading
local/non-hierarchical stats through {memcg/lruvec}_page_state_local() or
memcg_events_local(), where we previously only needed a flush to read
hierarchical stats. Most contexts reading non-hierarchical stats are
already doing a flush, add a flush to the only missing context in
count_shadow_nodes().
With this patch, reading memory.stat from 1000 memcgs is 3x faster on a
machine with 256 cpus on cgroup v1:
# for i in $(seq 1000); do mkdir /sys/fs/cgroup/memory/cg$i; done
# time cat /sys/fs/cgroup/memory/cg*/memory.stat > /dev/null
real 0m0.125s
user 0m0.005s
sys 0m0.120s
After:
real 0m0.032s
user 0m0.005s
sys 0m0.027s
To make sure there are no regressions on cgroup v2, I ran an artificial
reclaim/refault stress test [2] that creates (NR_CPUS * 2) cgroups,
assigns them limits, runs a worker process in each cgroup that allocates
tmpfs memory equal to quadruple the limit (to invoke reclaim
continuously), and then reads back the entire file (to invoke refaults).
All workers are run in parallel, and zram is used as a swapping backend.
Both reclaim and refault have conditional stats flushing. I ran this on a
machine with 112 cpus, once on mm-unstable, and once on mm-unstable with
this patch reverted.
(1) A few runs without this patch:
# time ./stress_reclaim_refault.sh
real 0m9.949s
user 0m0.496s
sys 14m44.974s
# time ./stress_reclaim_refault.sh
real 0m10.049s
user 0m0.486s
sys 14m55.791s
# time ./stress_reclaim_refault.sh
real 0m9.984s
user 0m0.481s
sys 14m53.841s
(2) A few runs with this patch:
# time ./stress_reclaim_refault.sh
real 0m9.885s
user 0m0.486s
sys 14m48.753s
# time ./stress_reclaim_refault.sh
real 0m9.903s
user 0m0.495s
sys 14m48.339s
# time ./stress_reclaim_refault.sh
real 0m9.861s
user 0m0.507s
sys 14m49.317s
No regressions are observed with this patch. There is actually a very
slight improvement. If I have to guess, maybe it's because we avoid
the percpu loop in count_shadow_nodes() when calling
lruvec_page_state_local(), but I could not prove this using perf, it's
probably in the noise.
[1] https://lore.kernel.org/lkml/20230725201811.GA1231514@cmpxchg.org/
[2] https://lore.kernel.org/lkml/CAJD7tkb17x=qwoO37uxyYXLEUVp15BQKR+Xfh7Sg9Hx-wTQ_=w@mail.gmail.com/
Link: https://lkml.kernel.org/r/20230803185046.1385770-1-yosryahmed@google.com
Link: https://lkml.kernel.org/r/20230726153223.821757-2-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Roman Gushchin <roman.gushchin@linux.dev>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Shakeel Butt <shakeelb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-07-26 15:32:23 +00:00
|
|
|
|
mm: memcontrol: switch to rstat
Replace the memory controller's custom hierarchical stats code with the
generic rstat infrastructure provided by the cgroup core.
The current implementation does batched upward propagation from the
write side (i.e. as stats change). The per-cpu batches introduce an
error, which is multiplied by the number of subgroups in a tree. In
systems with many CPUs and sizable cgroup trees, the error can be large
enough to confuse users (e.g. 32 batch pages * 32 CPUs * 32 subgroups
results in an error of up to 128M per stat item). This can entirely
swallow allocation bursts inside a workload that the user is expecting
to see reflected in the statistics.
In the past, we've done read-side aggregation, where a memory.stat read
would have to walk the entire subtree and add up per-cpu counts. This
became problematic with lazily-freed cgroups: we could have large
subtrees where most cgroups were entirely idle. Hence the switch to
change-driven upward propagation. Unfortunately, it needed to trade
accuracy for speed due to the write side being so hot.
Rstat combines the best of both worlds: from the write side, it cheaply
maintains a queue of cgroups that have pending changes, so that the read
side can do selective tree aggregation. This way the reported stats
will always be precise and recent as can be, while the aggregation can
skip over potentially large numbers of idle cgroups.
The way rstat works is that it implements a tree for tracking cgroups
with pending local changes, as well as a flush function that walks the
tree upwards. The controller then drives this by 1) telling rstat when
a local cgroup stat changes (e.g. mod_memcg_state) and 2) when a flush
is required to get uptodate hierarchy stats for a given subtree (e.g.
when memory.stat is read). The controller also provides a flush
callback that is called during the rstat flush walk for each cgroup and
aggregates its local per-cpu counters and propagates them upwards.
This adds a second vmstats to struct mem_cgroup (MEMCG_NR_STAT +
NR_VM_EVENT_ITEMS) to track pending subtree deltas during upward
aggregation. It removes 3 words from the per-cpu data. It eliminates
memcg_exact_page_state(), since memcg_page_state() is now exact.
[akpm@linux-foundation.org: merge fix]
[hannes@cmpxchg.org: fix a sleep in atomic section problem]
Link: https://lkml.kernel.org/r/20210315234100.64307-1-hannes@cmpxchg.org
Link: https://lkml.kernel.org/r/20210209163304.77088-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Roman Gushchin <guro@fb.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Balbir Singh <bsingharora@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:26 -07:00
|
|
|
|
2024-10-26 09:34:07 +00:00
|
|
|
|
mm: memcontrol: switch to rstat
Replace the memory controller's custom hierarchical stats code with the
generic rstat infrastructure provided by the cgroup core.
The current implementation does batched upward propagation from the
write side (i.e. as stats change). The per-cpu batches introduce an
error, which is multiplied by the number of subgroups in a tree. In
systems with many CPUs and sizable cgroup trees, the error can be large
enough to confuse users (e.g. 32 batch pages * 32 CPUs * 32 subgroups
results in an error of up to 128M per stat item). This can entirely
swallow allocation bursts inside a workload that the user is expecting
to see reflected in the statistics.
In the past, we've done read-side aggregation, where a memory.stat read
would have to walk the entire subtree and add up per-cpu counts. This
became problematic with lazily-freed cgroups: we could have large
subtrees where most cgroups were entirely idle. Hence the switch to
change-driven upward propagation. Unfortunately, it needed to trade
accuracy for speed due to the write side being so hot.
Rstat combines the best of both worlds: from the write side, it cheaply
maintains a queue of cgroups that have pending changes, so that the read
side can do selective tree aggregation. This way the reported stats
will always be precise and recent as can be, while the aggregation can
skip over potentially large numbers of idle cgroups.
The way rstat works is that it implements a tree for tracking cgroups
with pending local changes, as well as a flush function that walks the
tree upwards. The controller then drives this by 1) telling rstat when
a local cgroup stat changes (e.g. mod_memcg_state) and 2) when a flush
is required to get uptodate hierarchy stats for a given subtree (e.g.
when memory.stat is read). The controller also provides a flush
callback that is called during the rstat flush walk for each cgroup and
aggregates its local per-cpu counters and propagates them upwards.
This adds a second vmstats to struct mem_cgroup (MEMCG_NR_STAT +
NR_VM_EVENT_ITEMS) to track pending subtree deltas during upward
aggregation. It removes 3 words from the per-cpu data. It eliminates
memcg_exact_page_state(), since memcg_page_state() is now exact.
[akpm@linux-foundation.org: merge fix]
[hannes@cmpxchg.org: fix a sleep in atomic section problem]
Link: https://lkml.kernel.org/r/20210315234100.64307-1-hannes@cmpxchg.org
Link: https://lkml.kernel.org/r/20210209163304.77088-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Roman Gushchin <guro@fb.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Balbir Singh <bsingharora@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:26 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-10-26 09:34:07 +00:00
|
|
|
|
mm: memcontrol: switch to rstat
Replace the memory controller's custom hierarchical stats code with the
generic rstat infrastructure provided by the cgroup core.
The current implementation does batched upward propagation from the
write side (i.e. as stats change). The per-cpu batches introduce an
error, which is multiplied by the number of subgroups in a tree. In
systems with many CPUs and sizable cgroup trees, the error can be large
enough to confuse users (e.g. 32 batch pages * 32 CPUs * 32 subgroups
results in an error of up to 128M per stat item). This can entirely
swallow allocation bursts inside a workload that the user is expecting
to see reflected in the statistics.
In the past, we've done read-side aggregation, where a memory.stat read
would have to walk the entire subtree and add up per-cpu counts. This
became problematic with lazily-freed cgroups: we could have large
subtrees where most cgroups were entirely idle. Hence the switch to
change-driven upward propagation. Unfortunately, it needed to trade
accuracy for speed due to the write side being so hot.
Rstat combines the best of both worlds: from the write side, it cheaply
maintains a queue of cgroups that have pending changes, so that the read
side can do selective tree aggregation. This way the reported stats
will always be precise and recent as can be, while the aggregation can
skip over potentially large numbers of idle cgroups.
The way rstat works is that it implements a tree for tracking cgroups
with pending local changes, as well as a flush function that walks the
tree upwards. The controller then drives this by 1) telling rstat when
a local cgroup stat changes (e.g. mod_memcg_state) and 2) when a flush
is required to get uptodate hierarchy stats for a given subtree (e.g.
when memory.stat is read). The controller also provides a flush
callback that is called during the rstat flush walk for each cgroup and
aggregates its local per-cpu counters and propagates them upwards.
This adds a second vmstats to struct mem_cgroup (MEMCG_NR_STAT +
NR_VM_EVENT_ITEMS) to track pending subtree deltas during upward
aggregation. It removes 3 words from the per-cpu data. It eliminates
memcg_exact_page_state(), since memcg_page_state() is now exact.
[akpm@linux-foundation.org: merge fix]
[hannes@cmpxchg.org: fix a sleep in atomic section problem]
Link: https://lkml.kernel.org/r/20210315234100.64307-1-hannes@cmpxchg.org
Link: https://lkml.kernel.org/r/20210209163304.77088-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Roman Gushchin <guro@fb.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Balbir Singh <bsingharora@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:26 -07:00
|
|
|
|
2024-10-26 09:34:07 +00:00
|
|
|
|
mm: memcontrol: switch to rstat
Replace the memory controller's custom hierarchical stats code with the
generic rstat infrastructure provided by the cgroup core.
The current implementation does batched upward propagation from the
write side (i.e. as stats change). The per-cpu batches introduce an
error, which is multiplied by the number of subgroups in a tree. In
systems with many CPUs and sizable cgroup trees, the error can be large
enough to confuse users (e.g. 32 batch pages * 32 CPUs * 32 subgroups
results in an error of up to 128M per stat item). This can entirely
swallow allocation bursts inside a workload that the user is expecting
to see reflected in the statistics.
In the past, we've done read-side aggregation, where a memory.stat read
would have to walk the entire subtree and add up per-cpu counts. This
became problematic with lazily-freed cgroups: we could have large
subtrees where most cgroups were entirely idle. Hence the switch to
change-driven upward propagation. Unfortunately, it needed to trade
accuracy for speed due to the write side being so hot.
Rstat combines the best of both worlds: from the write side, it cheaply
maintains a queue of cgroups that have pending changes, so that the read
side can do selective tree aggregation. This way the reported stats
will always be precise and recent as can be, while the aggregation can
skip over potentially large numbers of idle cgroups.
The way rstat works is that it implements a tree for tracking cgroups
with pending local changes, as well as a flush function that walks the
tree upwards. The controller then drives this by 1) telling rstat when
a local cgroup stat changes (e.g. mod_memcg_state) and 2) when a flush
is required to get uptodate hierarchy stats for a given subtree (e.g.
when memory.stat is read). The controller also provides a flush
callback that is called during the rstat flush walk for each cgroup and
aggregates its local per-cpu counters and propagates them upwards.
This adds a second vmstats to struct mem_cgroup (MEMCG_NR_STAT +
NR_VM_EVENT_ITEMS) to track pending subtree deltas during upward
aggregation. It removes 3 words from the per-cpu data. It eliminates
memcg_exact_page_state(), since memcg_page_state() is now exact.
[akpm@linux-foundation.org: merge fix]
[hannes@cmpxchg.org: fix a sleep in atomic section problem]
Link: https://lkml.kernel.org/r/20210315234100.64307-1-hannes@cmpxchg.org
Link: https://lkml.kernel.org/r/20210209163304.77088-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Roman Gushchin <guro@fb.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Balbir Singh <bsingharora@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:26 -07:00
|
|
|
|
|
|
|
|
|
mm: memcg: use rstat for non-hierarchical stats
Currently, memcg uses rstat to maintain aggregated hierarchical stats.
Counters are maintained for hierarchical stats at each memcg. Rstat
tracks which cgroups have updates on which cpus to keep those counters
fresh on the read-side.
Non-hierarchical stats are currently not covered by rstat. Their per-cpu
counters are summed up on every read, which is expensive. The original
implementation did the same. At some point before rstat, non-hierarchical
aggregated counters were introduced by commit a983b5ebee57 ("mm:
memcontrol: fix excessive complexity in memory.stat reporting"). However,
those counters were updated on the performance critical write-side, which
caused regressions, so they were later removed by commit 815744d75152
("mm: memcontrol: don't batch updates of local VM stats and events"). See
[1] for more detailed history.
Kernel versions in between a983b5ebee57 & 815744d75152 (a year and a half)
enjoyed cheap reads of non-hierarchical stats, specifically on cgroup v1.
When moving to more recent kernels, a performance regression for reading
non-hierarchical stats is observed.
Now that we have rstat, we know exactly which percpu counters have updates
for each stat. We can maintain non-hierarchical counters again, making
reads much more efficient, without affecting the performance critical
write-side. Hence, add non-hierarchical (i.e local) counters for the
stats, and extend rstat flushing to keep those up-to-date.
A caveat is that we now need a stats flush before reading
local/non-hierarchical stats through {memcg/lruvec}_page_state_local() or
memcg_events_local(), where we previously only needed a flush to read
hierarchical stats. Most contexts reading non-hierarchical stats are
already doing a flush, add a flush to the only missing context in
count_shadow_nodes().
With this patch, reading memory.stat from 1000 memcgs is 3x faster on a
machine with 256 cpus on cgroup v1:
# for i in $(seq 1000); do mkdir /sys/fs/cgroup/memory/cg$i; done
# time cat /sys/fs/cgroup/memory/cg*/memory.stat > /dev/null
real 0m0.125s
user 0m0.005s
sys 0m0.120s
After:
real 0m0.032s
user 0m0.005s
sys 0m0.027s
To make sure there are no regressions on cgroup v2, I ran an artificial
reclaim/refault stress test [2] that creates (NR_CPUS * 2) cgroups,
assigns them limits, runs a worker process in each cgroup that allocates
tmpfs memory equal to quadruple the limit (to invoke reclaim
continuously), and then reads back the entire file (to invoke refaults).
All workers are run in parallel, and zram is used as a swapping backend.
Both reclaim and refault have conditional stats flushing. I ran this on a
machine with 112 cpus, once on mm-unstable, and once on mm-unstable with
this patch reverted.
(1) A few runs without this patch:
# time ./stress_reclaim_refault.sh
real 0m9.949s
user 0m0.496s
sys 14m44.974s
# time ./stress_reclaim_refault.sh
real 0m10.049s
user 0m0.486s
sys 14m55.791s
# time ./stress_reclaim_refault.sh
real 0m9.984s
user 0m0.481s
sys 14m53.841s
(2) A few runs with this patch:
# time ./stress_reclaim_refault.sh
real 0m9.885s
user 0m0.486s
sys 14m48.753s
# time ./stress_reclaim_refault.sh
real 0m9.903s
user 0m0.495s
sys 14m48.339s
# time ./stress_reclaim_refault.sh
real 0m9.861s
user 0m0.507s
sys 14m49.317s
No regressions are observed with this patch. There is actually a very
slight improvement. If I have to guess, maybe it's because we avoid
the percpu loop in count_shadow_nodes() when calling
lruvec_page_state_local(), but I could not prove this using perf, it's
probably in the noise.
[1] https://lore.kernel.org/lkml/20230725201811.GA1231514@cmpxchg.org/
[2] https://lore.kernel.org/lkml/CAJD7tkb17x=qwoO37uxyYXLEUVp15BQKR+Xfh7Sg9Hx-wTQ_=w@mail.gmail.com/
Link: https://lkml.kernel.org/r/20230803185046.1385770-1-yosryahmed@google.com
Link: https://lkml.kernel.org/r/20230726153223.821757-2-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Roman Gushchin <roman.gushchin@linux.dev>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Shakeel Butt <shakeelb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-07-26 15:32:23 +00:00
|
|
|
|
2024-10-26 09:34:07 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcg: use rstat for non-hierarchical stats
Currently, memcg uses rstat to maintain aggregated hierarchical stats.
Counters are maintained for hierarchical stats at each memcg. Rstat
tracks which cgroups have updates on which cpus to keep those counters
fresh on the read-side.
Non-hierarchical stats are currently not covered by rstat. Their per-cpu
counters are summed up on every read, which is expensive. The original
implementation did the same. At some point before rstat, non-hierarchical
aggregated counters were introduced by commit a983b5ebee57 ("mm:
memcontrol: fix excessive complexity in memory.stat reporting"). However,
those counters were updated on the performance critical write-side, which
caused regressions, so they were later removed by commit 815744d75152
("mm: memcontrol: don't batch updates of local VM stats and events"). See
[1] for more detailed history.
Kernel versions in between a983b5ebee57 & 815744d75152 (a year and a half)
enjoyed cheap reads of non-hierarchical stats, specifically on cgroup v1.
When moving to more recent kernels, a performance regression for reading
non-hierarchical stats is observed.
Now that we have rstat, we know exactly which percpu counters have updates
for each stat. We can maintain non-hierarchical counters again, making
reads much more efficient, without affecting the performance critical
write-side. Hence, add non-hierarchical (i.e local) counters for the
stats, and extend rstat flushing to keep those up-to-date.
A caveat is that we now need a stats flush before reading
local/non-hierarchical stats through {memcg/lruvec}_page_state_local() or
memcg_events_local(), where we previously only needed a flush to read
hierarchical stats. Most contexts reading non-hierarchical stats are
already doing a flush, add a flush to the only missing context in
count_shadow_nodes().
With this patch, reading memory.stat from 1000 memcgs is 3x faster on a
machine with 256 cpus on cgroup v1:
# for i in $(seq 1000); do mkdir /sys/fs/cgroup/memory/cg$i; done
# time cat /sys/fs/cgroup/memory/cg*/memory.stat > /dev/null
real 0m0.125s
user 0m0.005s
sys 0m0.120s
After:
real 0m0.032s
user 0m0.005s
sys 0m0.027s
To make sure there are no regressions on cgroup v2, I ran an artificial
reclaim/refault stress test [2] that creates (NR_CPUS * 2) cgroups,
assigns them limits, runs a worker process in each cgroup that allocates
tmpfs memory equal to quadruple the limit (to invoke reclaim
continuously), and then reads back the entire file (to invoke refaults).
All workers are run in parallel, and zram is used as a swapping backend.
Both reclaim and refault have conditional stats flushing. I ran this on a
machine with 112 cpus, once on mm-unstable, and once on mm-unstable with
this patch reverted.
(1) A few runs without this patch:
# time ./stress_reclaim_refault.sh
real 0m9.949s
user 0m0.496s
sys 14m44.974s
# time ./stress_reclaim_refault.sh
real 0m10.049s
user 0m0.486s
sys 14m55.791s
# time ./stress_reclaim_refault.sh
real 0m9.984s
user 0m0.481s
sys 14m53.841s
(2) A few runs with this patch:
# time ./stress_reclaim_refault.sh
real 0m9.885s
user 0m0.486s
sys 14m48.753s
# time ./stress_reclaim_refault.sh
real 0m9.903s
user 0m0.495s
sys 14m48.339s
# time ./stress_reclaim_refault.sh
real 0m9.861s
user 0m0.507s
sys 14m49.317s
No regressions are observed with this patch. There is actually a very
slight improvement. If I have to guess, maybe it's because we avoid
the percpu loop in count_shadow_nodes() when calling
lruvec_page_state_local(), but I could not prove this using perf, it's
probably in the noise.
[1] https://lore.kernel.org/lkml/20230725201811.GA1231514@cmpxchg.org/
[2] https://lore.kernel.org/lkml/CAJD7tkb17x=qwoO37uxyYXLEUVp15BQKR+Xfh7Sg9Hx-wTQ_=w@mail.gmail.com/
Link: https://lkml.kernel.org/r/20230803185046.1385770-1-yosryahmed@google.com
Link: https://lkml.kernel.org/r/20230726153223.821757-2-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Roman Gushchin <roman.gushchin@linux.dev>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Shakeel Butt <shakeelb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-07-26 15:32:23 +00:00
|
|
|
|
2024-10-26 09:34:07 +00:00
|
|
|
|
mm: memcontrol: switch to rstat
Replace the memory controller's custom hierarchical stats code with the
generic rstat infrastructure provided by the cgroup core.
The current implementation does batched upward propagation from the
write side (i.e. as stats change). The per-cpu batches introduce an
error, which is multiplied by the number of subgroups in a tree. In
systems with many CPUs and sizable cgroup trees, the error can be large
enough to confuse users (e.g. 32 batch pages * 32 CPUs * 32 subgroups
results in an error of up to 128M per stat item). This can entirely
swallow allocation bursts inside a workload that the user is expecting
to see reflected in the statistics.
In the past, we've done read-side aggregation, where a memory.stat read
would have to walk the entire subtree and add up per-cpu counts. This
became problematic with lazily-freed cgroups: we could have large
subtrees where most cgroups were entirely idle. Hence the switch to
change-driven upward propagation. Unfortunately, it needed to trade
accuracy for speed due to the write side being so hot.
Rstat combines the best of both worlds: from the write side, it cheaply
maintains a queue of cgroups that have pending changes, so that the read
side can do selective tree aggregation. This way the reported stats
will always be precise and recent as can be, while the aggregation can
skip over potentially large numbers of idle cgroups.
The way rstat works is that it implements a tree for tracking cgroups
with pending local changes, as well as a flush function that walks the
tree upwards. The controller then drives this by 1) telling rstat when
a local cgroup stat changes (e.g. mod_memcg_state) and 2) when a flush
is required to get uptodate hierarchy stats for a given subtree (e.g.
when memory.stat is read). The controller also provides a flush
callback that is called during the rstat flush walk for each cgroup and
aggregates its local per-cpu counters and propagates them upwards.
This adds a second vmstats to struct mem_cgroup (MEMCG_NR_STAT +
NR_VM_EVENT_ITEMS) to track pending subtree deltas during upward
aggregation. It removes 3 words from the per-cpu data. It eliminates
memcg_exact_page_state(), since memcg_page_state() is now exact.
[akpm@linux-foundation.org: merge fix]
[hannes@cmpxchg.org: fix a sleep in atomic section problem]
Link: https://lkml.kernel.org/r/20210315234100.64307-1-hannes@cmpxchg.org
Link: https://lkml.kernel.org/r/20210209163304.77088-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Roman Gushchin <guro@fb.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Balbir Singh <bsingharora@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:26 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcg: use rstat for non-hierarchical stats
Currently, memcg uses rstat to maintain aggregated hierarchical stats.
Counters are maintained for hierarchical stats at each memcg. Rstat
tracks which cgroups have updates on which cpus to keep those counters
fresh on the read-side.
Non-hierarchical stats are currently not covered by rstat. Their per-cpu
counters are summed up on every read, which is expensive. The original
implementation did the same. At some point before rstat, non-hierarchical
aggregated counters were introduced by commit a983b5ebee57 ("mm:
memcontrol: fix excessive complexity in memory.stat reporting"). However,
those counters were updated on the performance critical write-side, which
caused regressions, so they were later removed by commit 815744d75152
("mm: memcontrol: don't batch updates of local VM stats and events"). See
[1] for more detailed history.
Kernel versions in between a983b5ebee57 & 815744d75152 (a year and a half)
enjoyed cheap reads of non-hierarchical stats, specifically on cgroup v1.
When moving to more recent kernels, a performance regression for reading
non-hierarchical stats is observed.
Now that we have rstat, we know exactly which percpu counters have updates
for each stat. We can maintain non-hierarchical counters again, making
reads much more efficient, without affecting the performance critical
write-side. Hence, add non-hierarchical (i.e local) counters for the
stats, and extend rstat flushing to keep those up-to-date.
A caveat is that we now need a stats flush before reading
local/non-hierarchical stats through {memcg/lruvec}_page_state_local() or
memcg_events_local(), where we previously only needed a flush to read
hierarchical stats. Most contexts reading non-hierarchical stats are
already doing a flush, add a flush to the only missing context in
count_shadow_nodes().
With this patch, reading memory.stat from 1000 memcgs is 3x faster on a
machine with 256 cpus on cgroup v1:
# for i in $(seq 1000); do mkdir /sys/fs/cgroup/memory/cg$i; done
# time cat /sys/fs/cgroup/memory/cg*/memory.stat > /dev/null
real 0m0.125s
user 0m0.005s
sys 0m0.120s
After:
real 0m0.032s
user 0m0.005s
sys 0m0.027s
To make sure there are no regressions on cgroup v2, I ran an artificial
reclaim/refault stress test [2] that creates (NR_CPUS * 2) cgroups,
assigns them limits, runs a worker process in each cgroup that allocates
tmpfs memory equal to quadruple the limit (to invoke reclaim
continuously), and then reads back the entire file (to invoke refaults).
All workers are run in parallel, and zram is used as a swapping backend.
Both reclaim and refault have conditional stats flushing. I ran this on a
machine with 112 cpus, once on mm-unstable, and once on mm-unstable with
this patch reverted.
(1) A few runs without this patch:
# time ./stress_reclaim_refault.sh
real 0m9.949s
user 0m0.496s
sys 14m44.974s
# time ./stress_reclaim_refault.sh
real 0m10.049s
user 0m0.486s
sys 14m55.791s
# time ./stress_reclaim_refault.sh
real 0m9.984s
user 0m0.481s
sys 14m53.841s
(2) A few runs with this patch:
# time ./stress_reclaim_refault.sh
real 0m9.885s
user 0m0.486s
sys 14m48.753s
# time ./stress_reclaim_refault.sh
real 0m9.903s
user 0m0.495s
sys 14m48.339s
# time ./stress_reclaim_refault.sh
real 0m9.861s
user 0m0.507s
sys 14m49.317s
No regressions are observed with this patch. There is actually a very
slight improvement. If I have to guess, maybe it's because we avoid
the percpu loop in count_shadow_nodes() when calling
lruvec_page_state_local(), but I could not prove this using perf, it's
probably in the noise.
[1] https://lore.kernel.org/lkml/20230725201811.GA1231514@cmpxchg.org/
[2] https://lore.kernel.org/lkml/CAJD7tkb17x=qwoO37uxyYXLEUVp15BQKR+Xfh7Sg9Hx-wTQ_=w@mail.gmail.com/
Link: https://lkml.kernel.org/r/20230803185046.1385770-1-yosryahmed@google.com
Link: https://lkml.kernel.org/r/20230726153223.821757-2-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Roman Gushchin <roman.gushchin@linux.dev>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Shakeel Butt <shakeelb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-07-26 15:32:23 +00:00
|
|
|
|
2024-10-26 09:34:07 +00:00
|
|
|
|
mm: memcg: use rstat for non-hierarchical stats
Currently, memcg uses rstat to maintain aggregated hierarchical stats.
Counters are maintained for hierarchical stats at each memcg. Rstat
tracks which cgroups have updates on which cpus to keep those counters
fresh on the read-side.
Non-hierarchical stats are currently not covered by rstat. Their per-cpu
counters are summed up on every read, which is expensive. The original
implementation did the same. At some point before rstat, non-hierarchical
aggregated counters were introduced by commit a983b5ebee57 ("mm:
memcontrol: fix excessive complexity in memory.stat reporting"). However,
those counters were updated on the performance critical write-side, which
caused regressions, so they were later removed by commit 815744d75152
("mm: memcontrol: don't batch updates of local VM stats and events"). See
[1] for more detailed history.
Kernel versions in between a983b5ebee57 & 815744d75152 (a year and a half)
enjoyed cheap reads of non-hierarchical stats, specifically on cgroup v1.
When moving to more recent kernels, a performance regression for reading
non-hierarchical stats is observed.
Now that we have rstat, we know exactly which percpu counters have updates
for each stat. We can maintain non-hierarchical counters again, making
reads much more efficient, without affecting the performance critical
write-side. Hence, add non-hierarchical (i.e local) counters for the
stats, and extend rstat flushing to keep those up-to-date.
A caveat is that we now need a stats flush before reading
local/non-hierarchical stats through {memcg/lruvec}_page_state_local() or
memcg_events_local(), where we previously only needed a flush to read
hierarchical stats. Most contexts reading non-hierarchical stats are
already doing a flush, add a flush to the only missing context in
count_shadow_nodes().
With this patch, reading memory.stat from 1000 memcgs is 3x faster on a
machine with 256 cpus on cgroup v1:
# for i in $(seq 1000); do mkdir /sys/fs/cgroup/memory/cg$i; done
# time cat /sys/fs/cgroup/memory/cg*/memory.stat > /dev/null
real 0m0.125s
user 0m0.005s
sys 0m0.120s
After:
real 0m0.032s
user 0m0.005s
sys 0m0.027s
To make sure there are no regressions on cgroup v2, I ran an artificial
reclaim/refault stress test [2] that creates (NR_CPUS * 2) cgroups,
assigns them limits, runs a worker process in each cgroup that allocates
tmpfs memory equal to quadruple the limit (to invoke reclaim
continuously), and then reads back the entire file (to invoke refaults).
All workers are run in parallel, and zram is used as a swapping backend.
Both reclaim and refault have conditional stats flushing. I ran this on a
machine with 112 cpus, once on mm-unstable, and once on mm-unstable with
this patch reverted.
(1) A few runs without this patch:
# time ./stress_reclaim_refault.sh
real 0m9.949s
user 0m0.496s
sys 14m44.974s
# time ./stress_reclaim_refault.sh
real 0m10.049s
user 0m0.486s
sys 14m55.791s
# time ./stress_reclaim_refault.sh
real 0m9.984s
user 0m0.481s
sys 14m53.841s
(2) A few runs with this patch:
# time ./stress_reclaim_refault.sh
real 0m9.885s
user 0m0.486s
sys 14m48.753s
# time ./stress_reclaim_refault.sh
real 0m9.903s
user 0m0.495s
sys 14m48.339s
# time ./stress_reclaim_refault.sh
real 0m9.861s
user 0m0.507s
sys 14m49.317s
No regressions are observed with this patch. There is actually a very
slight improvement. If I have to guess, maybe it's because we avoid
the percpu loop in count_shadow_nodes() when calling
lruvec_page_state_local(), but I could not prove this using perf, it's
probably in the noise.
[1] https://lore.kernel.org/lkml/20230725201811.GA1231514@cmpxchg.org/
[2] https://lore.kernel.org/lkml/CAJD7tkb17x=qwoO37uxyYXLEUVp15BQKR+Xfh7Sg9Hx-wTQ_=w@mail.gmail.com/
Link: https://lkml.kernel.org/r/20230803185046.1385770-1-yosryahmed@google.com
Link: https://lkml.kernel.org/r/20230726153223.821757-2-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Roman Gushchin <roman.gushchin@linux.dev>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Shakeel Butt <shakeelb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-07-26 15:32:23 +00:00
|
|
|
|
|
|
|
|
|
2024-10-26 09:34:07 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcg: use rstat for non-hierarchical stats
Currently, memcg uses rstat to maintain aggregated hierarchical stats.
Counters are maintained for hierarchical stats at each memcg. Rstat
tracks which cgroups have updates on which cpus to keep those counters
fresh on the read-side.
Non-hierarchical stats are currently not covered by rstat. Their per-cpu
counters are summed up on every read, which is expensive. The original
implementation did the same. At some point before rstat, non-hierarchical
aggregated counters were introduced by commit a983b5ebee57 ("mm:
memcontrol: fix excessive complexity in memory.stat reporting"). However,
those counters were updated on the performance critical write-side, which
caused regressions, so they were later removed by commit 815744d75152
("mm: memcontrol: don't batch updates of local VM stats and events"). See
[1] for more detailed history.
Kernel versions in between a983b5ebee57 & 815744d75152 (a year and a half)
enjoyed cheap reads of non-hierarchical stats, specifically on cgroup v1.
When moving to more recent kernels, a performance regression for reading
non-hierarchical stats is observed.
Now that we have rstat, we know exactly which percpu counters have updates
for each stat. We can maintain non-hierarchical counters again, making
reads much more efficient, without affecting the performance critical
write-side. Hence, add non-hierarchical (i.e local) counters for the
stats, and extend rstat flushing to keep those up-to-date.
A caveat is that we now need a stats flush before reading
local/non-hierarchical stats through {memcg/lruvec}_page_state_local() or
memcg_events_local(), where we previously only needed a flush to read
hierarchical stats. Most contexts reading non-hierarchical stats are
already doing a flush, add a flush to the only missing context in
count_shadow_nodes().
With this patch, reading memory.stat from 1000 memcgs is 3x faster on a
machine with 256 cpus on cgroup v1:
# for i in $(seq 1000); do mkdir /sys/fs/cgroup/memory/cg$i; done
# time cat /sys/fs/cgroup/memory/cg*/memory.stat > /dev/null
real 0m0.125s
user 0m0.005s
sys 0m0.120s
After:
real 0m0.032s
user 0m0.005s
sys 0m0.027s
To make sure there are no regressions on cgroup v2, I ran an artificial
reclaim/refault stress test [2] that creates (NR_CPUS * 2) cgroups,
assigns them limits, runs a worker process in each cgroup that allocates
tmpfs memory equal to quadruple the limit (to invoke reclaim
continuously), and then reads back the entire file (to invoke refaults).
All workers are run in parallel, and zram is used as a swapping backend.
Both reclaim and refault have conditional stats flushing. I ran this on a
machine with 112 cpus, once on mm-unstable, and once on mm-unstable with
this patch reverted.
(1) A few runs without this patch:
# time ./stress_reclaim_refault.sh
real 0m9.949s
user 0m0.496s
sys 14m44.974s
# time ./stress_reclaim_refault.sh
real 0m10.049s
user 0m0.486s
sys 14m55.791s
# time ./stress_reclaim_refault.sh
real 0m9.984s
user 0m0.481s
sys 14m53.841s
(2) A few runs with this patch:
# time ./stress_reclaim_refault.sh
real 0m9.885s
user 0m0.486s
sys 14m48.753s
# time ./stress_reclaim_refault.sh
real 0m9.903s
user 0m0.495s
sys 14m48.339s
# time ./stress_reclaim_refault.sh
real 0m9.861s
user 0m0.507s
sys 14m49.317s
No regressions are observed with this patch. There is actually a very
slight improvement. If I have to guess, maybe it's because we avoid
the percpu loop in count_shadow_nodes() when calling
lruvec_page_state_local(), but I could not prove this using perf, it's
probably in the noise.
[1] https://lore.kernel.org/lkml/20230725201811.GA1231514@cmpxchg.org/
[2] https://lore.kernel.org/lkml/CAJD7tkb17x=qwoO37uxyYXLEUVp15BQKR+Xfh7Sg9Hx-wTQ_=w@mail.gmail.com/
Link: https://lkml.kernel.org/r/20230803185046.1385770-1-yosryahmed@google.com
Link: https://lkml.kernel.org/r/20230726153223.821757-2-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Roman Gushchin <roman.gushchin@linux.dev>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Shakeel Butt <shakeelb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-07-26 15:32:23 +00:00
|
|
|
|
mm: memcontrol: switch to rstat
Replace the memory controller's custom hierarchical stats code with the
generic rstat infrastructure provided by the cgroup core.
The current implementation does batched upward propagation from the
write side (i.e. as stats change). The per-cpu batches introduce an
error, which is multiplied by the number of subgroups in a tree. In
systems with many CPUs and sizable cgroup trees, the error can be large
enough to confuse users (e.g. 32 batch pages * 32 CPUs * 32 subgroups
results in an error of up to 128M per stat item). This can entirely
swallow allocation bursts inside a workload that the user is expecting
to see reflected in the statistics.
In the past, we've done read-side aggregation, where a memory.stat read
would have to walk the entire subtree and add up per-cpu counts. This
became problematic with lazily-freed cgroups: we could have large
subtrees where most cgroups were entirely idle. Hence the switch to
change-driven upward propagation. Unfortunately, it needed to trade
accuracy for speed due to the write side being so hot.
Rstat combines the best of both worlds: from the write side, it cheaply
maintains a queue of cgroups that have pending changes, so that the read
side can do selective tree aggregation. This way the reported stats
will always be precise and recent as can be, while the aggregation can
skip over potentially large numbers of idle cgroups.
The way rstat works is that it implements a tree for tracking cgroups
with pending local changes, as well as a flush function that walks the
tree upwards. The controller then drives this by 1) telling rstat when
a local cgroup stat changes (e.g. mod_memcg_state) and 2) when a flush
is required to get uptodate hierarchy stats for a given subtree (e.g.
when memory.stat is read). The controller also provides a flush
callback that is called during the rstat flush walk for each cgroup and
aggregates its local per-cpu counters and propagates them upwards.
This adds a second vmstats to struct mem_cgroup (MEMCG_NR_STAT +
NR_VM_EVENT_ITEMS) to track pending subtree deltas during upward
aggregation. It removes 3 words from the per-cpu data. It eliminates
memcg_exact_page_state(), since memcg_page_state() is now exact.
[akpm@linux-foundation.org: merge fix]
[hannes@cmpxchg.org: fix a sleep in atomic section problem]
Link: https://lkml.kernel.org/r/20210315234100.64307-1-hannes@cmpxchg.org
Link: https://lkml.kernel.org/r/20210209163304.77088-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Roman Gushchin <guro@fb.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Balbir Singh <bsingharora@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:26 -07:00
|
|
|
|
2024-10-26 09:34:07 +00:00
|
|
|
|
mm: memcontrol: switch to rstat
Replace the memory controller's custom hierarchical stats code with the
generic rstat infrastructure provided by the cgroup core.
The current implementation does batched upward propagation from the
write side (i.e. as stats change). The per-cpu batches introduce an
error, which is multiplied by the number of subgroups in a tree. In
systems with many CPUs and sizable cgroup trees, the error can be large
enough to confuse users (e.g. 32 batch pages * 32 CPUs * 32 subgroups
results in an error of up to 128M per stat item). This can entirely
swallow allocation bursts inside a workload that the user is expecting
to see reflected in the statistics.
In the past, we've done read-side aggregation, where a memory.stat read
would have to walk the entire subtree and add up per-cpu counts. This
became problematic with lazily-freed cgroups: we could have large
subtrees where most cgroups were entirely idle. Hence the switch to
change-driven upward propagation. Unfortunately, it needed to trade
accuracy for speed due to the write side being so hot.
Rstat combines the best of both worlds: from the write side, it cheaply
maintains a queue of cgroups that have pending changes, so that the read
side can do selective tree aggregation. This way the reported stats
will always be precise and recent as can be, while the aggregation can
skip over potentially large numbers of idle cgroups.
The way rstat works is that it implements a tree for tracking cgroups
with pending local changes, as well as a flush function that walks the
tree upwards. The controller then drives this by 1) telling rstat when
a local cgroup stat changes (e.g. mod_memcg_state) and 2) when a flush
is required to get uptodate hierarchy stats for a given subtree (e.g.
when memory.stat is read). The controller also provides a flush
callback that is called during the rstat flush walk for each cgroup and
aggregates its local per-cpu counters and propagates them upwards.
This adds a second vmstats to struct mem_cgroup (MEMCG_NR_STAT +
NR_VM_EVENT_ITEMS) to track pending subtree deltas during upward
aggregation. It removes 3 words from the per-cpu data. It eliminates
memcg_exact_page_state(), since memcg_page_state() is now exact.
[akpm@linux-foundation.org: merge fix]
[hannes@cmpxchg.org: fix a sleep in atomic section problem]
Link: https://lkml.kernel.org/r/20210315234100.64307-1-hannes@cmpxchg.org
Link: https://lkml.kernel.org/r/20210209163304.77088-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Roman Gushchin <guro@fb.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Balbir Singh <bsingharora@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:26 -07:00
|
|
|
|
2025-05-18 23:31:39 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-10-26 09:34:07 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcontrol: switch to rstat
Replace the memory controller's custom hierarchical stats code with the
generic rstat infrastructure provided by the cgroup core.
The current implementation does batched upward propagation from the
write side (i.e. as stats change). The per-cpu batches introduce an
error, which is multiplied by the number of subgroups in a tree. In
systems with many CPUs and sizable cgroup trees, the error can be large
enough to confuse users (e.g. 32 batch pages * 32 CPUs * 32 subgroups
results in an error of up to 128M per stat item). This can entirely
swallow allocation bursts inside a workload that the user is expecting
to see reflected in the statistics.
In the past, we've done read-side aggregation, where a memory.stat read
would have to walk the entire subtree and add up per-cpu counts. This
became problematic with lazily-freed cgroups: we could have large
subtrees where most cgroups were entirely idle. Hence the switch to
change-driven upward propagation. Unfortunately, it needed to trade
accuracy for speed due to the write side being so hot.
Rstat combines the best of both worlds: from the write side, it cheaply
maintains a queue of cgroups that have pending changes, so that the read
side can do selective tree aggregation. This way the reported stats
will always be precise and recent as can be, while the aggregation can
skip over potentially large numbers of idle cgroups.
The way rstat works is that it implements a tree for tracking cgroups
with pending local changes, as well as a flush function that walks the
tree upwards. The controller then drives this by 1) telling rstat when
a local cgroup stat changes (e.g. mod_memcg_state) and 2) when a flush
is required to get uptodate hierarchy stats for a given subtree (e.g.
when memory.stat is read). The controller also provides a flush
callback that is called during the rstat flush walk for each cgroup and
aggregates its local per-cpu counters and propagates them upwards.
This adds a second vmstats to struct mem_cgroup (MEMCG_NR_STAT +
NR_VM_EVENT_ITEMS) to track pending subtree deltas during upward
aggregation. It removes 3 words from the per-cpu data. It eliminates
memcg_exact_page_state(), since memcg_page_state() is now exact.
[akpm@linux-foundation.org: merge fix]
[hannes@cmpxchg.org: fix a sleep in atomic section problem]
Link: https://lkml.kernel.org/r/20210315234100.64307-1-hannes@cmpxchg.org
Link: https://lkml.kernel.org/r/20210209163304.77088-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Roman Gushchin <guro@fb.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Balbir Singh <bsingharora@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:26 -07:00
|
|
|
|
2025-05-18 23:31:39 -07:00
|
|
|
|
|
|
|
|
|
2024-10-26 09:34:07 +00:00
|
|
|
|
mm: memcontrol: switch to rstat
Replace the memory controller's custom hierarchical stats code with the
generic rstat infrastructure provided by the cgroup core.
The current implementation does batched upward propagation from the
write side (i.e. as stats change). The per-cpu batches introduce an
error, which is multiplied by the number of subgroups in a tree. In
systems with many CPUs and sizable cgroup trees, the error can be large
enough to confuse users (e.g. 32 batch pages * 32 CPUs * 32 subgroups
results in an error of up to 128M per stat item). This can entirely
swallow allocation bursts inside a workload that the user is expecting
to see reflected in the statistics.
In the past, we've done read-side aggregation, where a memory.stat read
would have to walk the entire subtree and add up per-cpu counts. This
became problematic with lazily-freed cgroups: we could have large
subtrees where most cgroups were entirely idle. Hence the switch to
change-driven upward propagation. Unfortunately, it needed to trade
accuracy for speed due to the write side being so hot.
Rstat combines the best of both worlds: from the write side, it cheaply
maintains a queue of cgroups that have pending changes, so that the read
side can do selective tree aggregation. This way the reported stats
will always be precise and recent as can be, while the aggregation can
skip over potentially large numbers of idle cgroups.
The way rstat works is that it implements a tree for tracking cgroups
with pending local changes, as well as a flush function that walks the
tree upwards. The controller then drives this by 1) telling rstat when
a local cgroup stat changes (e.g. mod_memcg_state) and 2) when a flush
is required to get uptodate hierarchy stats for a given subtree (e.g.
when memory.stat is read). The controller also provides a flush
callback that is called during the rstat flush walk for each cgroup and
aggregates its local per-cpu counters and propagates them upwards.
This adds a second vmstats to struct mem_cgroup (MEMCG_NR_STAT +
NR_VM_EVENT_ITEMS) to track pending subtree deltas during upward
aggregation. It removes 3 words from the per-cpu data. It eliminates
memcg_exact_page_state(), since memcg_page_state() is now exact.
[akpm@linux-foundation.org: merge fix]
[hannes@cmpxchg.org: fix a sleep in atomic section problem]
Link: https://lkml.kernel.org/r/20210315234100.64307-1-hannes@cmpxchg.org
Link: https://lkml.kernel.org/r/20210209163304.77088-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Roman Gushchin <guro@fb.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Balbir Singh <bsingharora@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:26 -07:00
|
|
|
|
2024-10-26 09:34:07 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-09-02 14:55:00 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-05-01 10:26:11 -07:00
|
|
|
|
|
|
|
|
|
2021-09-02 14:55:00 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-05-01 10:26:11 -07:00
|
|
|
|
2021-09-02 14:55:00 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-10-26 09:34:07 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-09-02 14:55:00 -07:00
|
|
|
|
|
|
|
|
|
2024-04-24 05:59:39 -07:00
|
|
|
|
mm: memcg: make stats flushing threshold per-memcg
A global counter for the magnitude of memcg stats update is maintained on
the memcg side to avoid invoking rstat flushes when the pending updates
are not significant. This avoids unnecessary flushes, which are not very
cheap even if there isn't a lot of stats to flush. It also avoids
unnecessary lock contention on the underlying global rstat lock.
Make this threshold per-memcg. The scheme is followed where percpu (now
also per-memcg) counters are incremented in the update path, and only
propagated to per-memcg atomics when they exceed a certain threshold.
This provides two benefits: (a) On large machines with a lot of memcgs,
the global threshold can be reached relatively fast, so guarding the
underlying lock becomes less effective. Making the threshold per-memcg
avoids this.
(b) Having a global threshold makes it hard to do subtree flushes, as we
cannot reset the global counter except for a full flush. Per-memcg
counters removes this as a blocker from doing subtree flushes, which helps
avoid unnecessary work when the stats of a small subtree are needed.
Nothing is free, of course. This comes at a cost: (a) A new per-cpu
counter per memcg, consuming NR_CPUS * NR_MEMCGS * 4 bytes. The extra
memory usage is insigificant.
(b) More work on the update side, although in the common case it will only
be percpu counter updates. The amount of work scales with the number of
ancestors (i.e. tree depth). This is not a new concept, adding a cgroup
to the rstat tree involves a parent loop, so is charging. Testing results
below show no significant regressions.
(c) The error margin in the stats for the system as a whole increases from
NR_CPUS * MEMCG_CHARGE_BATCH to NR_CPUS * MEMCG_CHARGE_BATCH * NR_MEMCGS.
This is probably fine because we have a similar per-memcg error in charges
coming from percpu stocks, and we have a periodic flusher that makes sure
we always flush all the stats every 2s anyway.
This patch was tested to make sure no significant regressions are
introduced on the update path as follows. The following benchmarks were
ran in a cgroup that is 2 levels deep (/sys/fs/cgroup/a/b/):
(1) Running 22 instances of netperf on a 44 cpu machine with
hyperthreading disabled. All instances are run in a level 2 cgroup, as
well as netserver:
# netserver -6
# netperf -6 -H ::1 -l 60 -t TCP_SENDFILE -- -m 10K
Averaging 20 runs, the numbers are as follows:
Base: 40198.0 mbps
Patched: 38629.7 mbps (-3.9%)
The regression is minimal, especially for 22 instances in the same
cgroup sharing all ancestors (so updating the same atomics).
(2) will-it-scale page_fault tests. These tests (specifically
per_process_ops in page_fault3 test) detected a 25.9% regression before
for a change in the stats update path [1]. These are the
numbers from 10 runs (+ is good) on a machine with 256 cpus:
LABEL | MEAN | MEDIAN | STDDEV |
------------------------------+-------------+-------------+-------------
page_fault1_per_process_ops | | | |
(A) base | 270249.164 | 265437.000 | 13451.836 |
(B) patched | 261368.709 | 255725.000 | 13394.767 |
| -3.29% | -3.66% | |
page_fault1_per_thread_ops | | | |
(A) base | 242111.345 | 239737.000 | 10026.031 |
(B) patched | 237057.109 | 235305.000 | 9769.687 |
| -2.09% | -1.85% | |
page_fault1_scalability | | |
(A) base | 0.034387 | 0.035168 | 0.0018283 |
(B) patched | 0.033988 | 0.034573 | 0.0018056 |
| -1.16% | -1.69% | |
page_fault2_per_process_ops | | |
(A) base | 203561.836 | 203301.000 | 2550.764 |
(B) patched | 197195.945 | 197746.000 | 2264.263 |
| -3.13% | -2.73% | |
page_fault2_per_thread_ops | | |
(A) base | 171046.473 | 170776.000 | 1509.679 |
(B) patched | 166626.327 | 166406.000 | 768.753 |
| -2.58% | -2.56% | |
page_fault2_scalability | | |
(A) base | 0.054026 | 0.053821 | 0.00062121 |
(B) patched | 0.053329 | 0.05306 | 0.00048394 |
| -1.29% | -1.41% | |
page_fault3_per_process_ops | | |
(A) base | 1295807.782 | 1297550.000 | 5907.585 |
(B) patched | 1275579.873 | 1273359.000 | 8759.160 |
| -1.56% | -1.86% | |
page_fault3_per_thread_ops | | |
(A) base | 391234.164 | 390860.000 | 1760.720 |
(B) patched | 377231.273 | 376369.000 | 1874.971 |
| -3.58% | -3.71% | |
page_fault3_scalability | | |
(A) base | 0.60369 | 0.60072 | 0.0083029 |
(B) patched | 0.61733 | 0.61544 | 0.009855 |
| +2.26% | +2.45% | |
All regressions seem to be minimal, and within the normal variance for the
benchmark. The fix for [1] assumes that 3% is noise -- and there were no
further practical complaints), so hopefully this means that such
variations in these microbenchmarks do not reflect on practical workloads.
(3) I also ran stress-ng in a nested cgroup and did not observe any
obvious regressions.
[1]https://lore.kernel.org/all/20190520063534.GB19312@shao2-debian/
Link: https://lkml.kernel.org/r/20231129032154.3710765-4-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Tested-by: Domenico Cerasuolo <cerasuolodomenico@gmail.com>
Acked-by: Shakeel Butt <shakeelb@google.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: Greg Thelen <gthelen@google.com>
Cc: Ivan Babrou <ivan@cloudflare.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Michal Koutny <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Tejun Heo <tj@kernel.org>
Cc: Waiman Long <longman@redhat.com>
Cc: Wei Xu <weixugc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-11-29 03:21:51 +00:00
|
|
|
|
2025-05-18 23:31:42 -07:00
|
|
|
|
|
|
|
|
|
mm: memcontrol: switch to rstat
Replace the memory controller's custom hierarchical stats code with the
generic rstat infrastructure provided by the cgroup core.
The current implementation does batched upward propagation from the
write side (i.e. as stats change). The per-cpu batches introduce an
error, which is multiplied by the number of subgroups in a tree. In
systems with many CPUs and sizable cgroup trees, the error can be large
enough to confuse users (e.g. 32 batch pages * 32 CPUs * 32 subgroups
results in an error of up to 128M per stat item). This can entirely
swallow allocation bursts inside a workload that the user is expecting
to see reflected in the statistics.
In the past, we've done read-side aggregation, where a memory.stat read
would have to walk the entire subtree and add up per-cpu counts. This
became problematic with lazily-freed cgroups: we could have large
subtrees where most cgroups were entirely idle. Hence the switch to
change-driven upward propagation. Unfortunately, it needed to trade
accuracy for speed due to the write side being so hot.
Rstat combines the best of both worlds: from the write side, it cheaply
maintains a queue of cgroups that have pending changes, so that the read
side can do selective tree aggregation. This way the reported stats
will always be precise and recent as can be, while the aggregation can
skip over potentially large numbers of idle cgroups.
The way rstat works is that it implements a tree for tracking cgroups
with pending local changes, as well as a flush function that walks the
tree upwards. The controller then drives this by 1) telling rstat when
a local cgroup stat changes (e.g. mod_memcg_state) and 2) when a flush
is required to get uptodate hierarchy stats for a given subtree (e.g.
when memory.stat is read). The controller also provides a flush
callback that is called during the rstat flush walk for each cgroup and
aggregates its local per-cpu counters and propagates them upwards.
This adds a second vmstats to struct mem_cgroup (MEMCG_NR_STAT +
NR_VM_EVENT_ITEMS) to track pending subtree deltas during upward
aggregation. It removes 3 words from the per-cpu data. It eliminates
memcg_exact_page_state(), since memcg_page_state() is now exact.
[akpm@linux-foundation.org: merge fix]
[hannes@cmpxchg.org: fix a sleep in atomic section problem]
Link: https://lkml.kernel.org/r/20210315234100.64307-1-hannes@cmpxchg.org
Link: https://lkml.kernel.org/r/20210209163304.77088-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Roman Gushchin <guro@fb.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Balbir Singh <bsingharora@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:26 -07:00
|
|
|
|
|
|
|
|
|
2023-10-19 15:53:42 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-16 01:58:03 +00:00
|
|
|
|
2023-10-19 15:53:42 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 02:00:05 -06:00
|
|
|
|
2023-10-19 15:53:42 -07:00
|
|
|
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 02:00:05 -06:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-19 15:53:42 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 02:00:05 -06:00
|
|
|
|
|
|
|
|
|
2023-10-19 15:53:42 -07:00
|
|
|
|
|
|
|
|
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 02:00:05 -06:00
|
|
|
|
|
|
|
|
|
2019-03-05 15:45:55 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-11-05 18:50:23 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
|
|
|
|
|
mm, memcg: cg2 memory{.swap,}.peak write handlers
Patch series "mm, memcg: cg2 memory{.swap,}.peak write handlers", v7.
This patch (of 2):
Other mechanisms for querying the peak memory usage of either a process or
v1 memory cgroup allow for resetting the high watermark. Restore parity
with those mechanisms, but with a less racy API.
For example:
- Any write to memory.max_usage_in_bytes in a cgroup v1 mount resets
the high watermark.
- writing "5" to the clear_refs pseudo-file in a processes's proc
directory resets the peak RSS.
This change is an evolution of a previous patch, which mostly copied the
cgroup v1 behavior, however, there were concerns about races/ownership
issues with a global reset, so instead this change makes the reset
filedescriptor-local.
Writing any non-empty string to the memory.peak and memory.swap.peak
pseudo-files reset the high watermark to the current usage for subsequent
reads through that same FD.
Notably, following Johannes's suggestion, this implementation moves the
O(FDs that have written) behavior onto the FD write(2) path. Instead, on
the page-allocation path, we simply add one additional watermark to
conditionally bump per-hierarchy level in the page-counter.
Additionally, this takes Longman's suggestion of nesting the
page-charging-path checks for the two watermarks to reduce the number of
common-case comparisons.
This behavior is particularly useful for work scheduling systems that need
to track memory usage of worker processes/cgroups per-work-item. Since
memory can't be squeezed like CPU can (the OOM-killer has opinions), these
systems need to track the peak memory usage to compute system/container
fullness when binpacking workitems.
Most notably, Vimeo's use-case involves a system that's doing global
binpacking across many Kubernetes pods/containers, and while we can use
PSI for some local decisions about overload, we strive to avoid packing
workloads too tightly in the first place. To facilitate this, we track
the peak memory usage. However, since we run with long-lived workers (to
amortize startup costs) we need a way to track the high watermark while a
work-item is executing. Polling runs the risk of missing short spikes
that last for timescales below the polling interval, and peak memory
tracking at the cgroup level is otherwise perfect for this use-case.
As this data is used to ensure that binpacked work ends up with sufficient
headroom, this use-case mostly avoids the inaccuracies surrounding
reclaimable memory.
Link: https://lkml.kernel.org/r/20240730231304.761942-1-davidf@vimeo.com
Link: https://lkml.kernel.org/r/20240729143743.34236-1-davidf@vimeo.com
Link: https://lkml.kernel.org/r/20240729143743.34236-2-davidf@vimeo.com
Signed-off-by: David Finkel <davidf@vimeo.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Waiman Long <longman@redhat.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-07-29 10:37:42 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-05-13 16:48:57 -07:00
|
|
|
|
mm, memcg: cg2 memory{.swap,}.peak write handlers
Patch series "mm, memcg: cg2 memory{.swap,}.peak write handlers", v7.
This patch (of 2):
Other mechanisms for querying the peak memory usage of either a process or
v1 memory cgroup allow for resetting the high watermark. Restore parity
with those mechanisms, but with a less racy API.
For example:
- Any write to memory.max_usage_in_bytes in a cgroup v1 mount resets
the high watermark.
- writing "5" to the clear_refs pseudo-file in a processes's proc
directory resets the peak RSS.
This change is an evolution of a previous patch, which mostly copied the
cgroup v1 behavior, however, there were concerns about races/ownership
issues with a global reset, so instead this change makes the reset
filedescriptor-local.
Writing any non-empty string to the memory.peak and memory.swap.peak
pseudo-files reset the high watermark to the current usage for subsequent
reads through that same FD.
Notably, following Johannes's suggestion, this implementation moves the
O(FDs that have written) behavior onto the FD write(2) path. Instead, on
the page-allocation path, we simply add one additional watermark to
conditionally bump per-hierarchy level in the page-counter.
Additionally, this takes Longman's suggestion of nesting the
page-charging-path checks for the two watermarks to reduce the number of
common-case comparisons.
This behavior is particularly useful for work scheduling systems that need
to track memory usage of worker processes/cgroups per-work-item. Since
memory can't be squeezed like CPU can (the OOM-killer has opinions), these
systems need to track the peak memory usage to compute system/container
fullness when binpacking workitems.
Most notably, Vimeo's use-case involves a system that's doing global
binpacking across many Kubernetes pods/containers, and while we can use
PSI for some local decisions about overload, we strive to avoid packing
workloads too tightly in the first place. To facilitate this, we track
the peak memory usage. However, since we run with long-lived workers (to
amortize startup costs) we need a way to track the high watermark while a
work-item is executing. Polling runs the risk of missing short spikes
that last for timescales below the polling interval, and peak memory
tracking at the cgroup level is otherwise perfect for this use-case.
As this data is used to ensure that binpacked work ends up with sufficient
headroom, this use-case mostly avoids the inaccuracies surrounding
reclaimable memory.
Link: https://lkml.kernel.org/r/20240730231304.761942-1-davidf@vimeo.com
Link: https://lkml.kernel.org/r/20240729143743.34236-1-davidf@vimeo.com
Link: https://lkml.kernel.org/r/20240729143743.34236-2-davidf@vimeo.com
Signed-off-by: David Finkel <davidf@vimeo.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Waiman Long <longman@redhat.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-07-29 10:37:42 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-05-13 16:48:57 -07:00
|
|
|
|
mm, memcg: cg2 memory{.swap,}.peak write handlers
Patch series "mm, memcg: cg2 memory{.swap,}.peak write handlers", v7.
This patch (of 2):
Other mechanisms for querying the peak memory usage of either a process or
v1 memory cgroup allow for resetting the high watermark. Restore parity
with those mechanisms, but with a less racy API.
For example:
- Any write to memory.max_usage_in_bytes in a cgroup v1 mount resets
the high watermark.
- writing "5" to the clear_refs pseudo-file in a processes's proc
directory resets the peak RSS.
This change is an evolution of a previous patch, which mostly copied the
cgroup v1 behavior, however, there were concerns about races/ownership
issues with a global reset, so instead this change makes the reset
filedescriptor-local.
Writing any non-empty string to the memory.peak and memory.swap.peak
pseudo-files reset the high watermark to the current usage for subsequent
reads through that same FD.
Notably, following Johannes's suggestion, this implementation moves the
O(FDs that have written) behavior onto the FD write(2) path. Instead, on
the page-allocation path, we simply add one additional watermark to
conditionally bump per-hierarchy level in the page-counter.
Additionally, this takes Longman's suggestion of nesting the
page-charging-path checks for the two watermarks to reduce the number of
common-case comparisons.
This behavior is particularly useful for work scheduling systems that need
to track memory usage of worker processes/cgroups per-work-item. Since
memory can't be squeezed like CPU can (the OOM-killer has opinions), these
systems need to track the peak memory usage to compute system/container
fullness when binpacking workitems.
Most notably, Vimeo's use-case involves a system that's doing global
binpacking across many Kubernetes pods/containers, and while we can use
PSI for some local decisions about overload, we strive to avoid packing
workloads too tightly in the first place. To facilitate this, we track
the peak memory usage. However, since we run with long-lived workers (to
amortize startup costs) we need a way to track the high watermark while a
work-item is executing. Polling runs the risk of missing short spikes
that last for timescales below the polling interval, and peak memory
tracking at the cgroup level is otherwise perfect for this use-case.
As this data is used to ensure that binpacked work ends up with sufficient
headroom, this use-case mostly avoids the inaccuracies surrounding
reclaimable memory.
Link: https://lkml.kernel.org/r/20240730231304.761942-1-davidf@vimeo.com
Link: https://lkml.kernel.org/r/20240729143743.34236-1-davidf@vimeo.com
Link: https://lkml.kernel.org/r/20240729143743.34236-2-davidf@vimeo.com
Signed-off-by: David Finkel <davidf@vimeo.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Waiman Long <longman@redhat.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-07-29 10:37:42 -04:00
|
|
|
|
2022-05-13 16:48:57 -07:00
|
|
|
|
|
|
|
|
|
mm, memcg: cg2 memory{.swap,}.peak write handlers
Patch series "mm, memcg: cg2 memory{.swap,}.peak write handlers", v7.
This patch (of 2):
Other mechanisms for querying the peak memory usage of either a process or
v1 memory cgroup allow for resetting the high watermark. Restore parity
with those mechanisms, but with a less racy API.
For example:
- Any write to memory.max_usage_in_bytes in a cgroup v1 mount resets
the high watermark.
- writing "5" to the clear_refs pseudo-file in a processes's proc
directory resets the peak RSS.
This change is an evolution of a previous patch, which mostly copied the
cgroup v1 behavior, however, there were concerns about races/ownership
issues with a global reset, so instead this change makes the reset
filedescriptor-local.
Writing any non-empty string to the memory.peak and memory.swap.peak
pseudo-files reset the high watermark to the current usage for subsequent
reads through that same FD.
Notably, following Johannes's suggestion, this implementation moves the
O(FDs that have written) behavior onto the FD write(2) path. Instead, on
the page-allocation path, we simply add one additional watermark to
conditionally bump per-hierarchy level in the page-counter.
Additionally, this takes Longman's suggestion of nesting the
page-charging-path checks for the two watermarks to reduce the number of
common-case comparisons.
This behavior is particularly useful for work scheduling systems that need
to track memory usage of worker processes/cgroups per-work-item. Since
memory can't be squeezed like CPU can (the OOM-killer has opinions), these
systems need to track the peak memory usage to compute system/container
fullness when binpacking workitems.
Most notably, Vimeo's use-case involves a system that's doing global
binpacking across many Kubernetes pods/containers, and while we can use
PSI for some local decisions about overload, we strive to avoid packing
workloads too tightly in the first place. To facilitate this, we track
the peak memory usage. However, since we run with long-lived workers (to
amortize startup costs) we need a way to track the high watermark while a
work-item is executing. Polling runs the risk of missing short spikes
that last for timescales below the polling interval, and peak memory
tracking at the cgroup level is otherwise perfect for this use-case.
As this data is used to ensure that binpacked work ends up with sufficient
headroom, this use-case mostly avoids the inaccuracies surrounding
reclaimable memory.
Link: https://lkml.kernel.org/r/20240730231304.761942-1-davidf@vimeo.com
Link: https://lkml.kernel.org/r/20240729143743.34236-1-davidf@vimeo.com
Link: https://lkml.kernel.org/r/20240729143743.34236-2-davidf@vimeo.com
Signed-off-by: David Finkel <davidf@vimeo.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Waiman Long <longman@redhat.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-07-29 10:37:42 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-01-24 07:35:11 +00:00
|
|
|
|
mm, memcg: cg2 memory{.swap,}.peak write handlers
Patch series "mm, memcg: cg2 memory{.swap,}.peak write handlers", v7.
This patch (of 2):
Other mechanisms for querying the peak memory usage of either a process or
v1 memory cgroup allow for resetting the high watermark. Restore parity
with those mechanisms, but with a less racy API.
For example:
- Any write to memory.max_usage_in_bytes in a cgroup v1 mount resets
the high watermark.
- writing "5" to the clear_refs pseudo-file in a processes's proc
directory resets the peak RSS.
This change is an evolution of a previous patch, which mostly copied the
cgroup v1 behavior, however, there were concerns about races/ownership
issues with a global reset, so instead this change makes the reset
filedescriptor-local.
Writing any non-empty string to the memory.peak and memory.swap.peak
pseudo-files reset the high watermark to the current usage for subsequent
reads through that same FD.
Notably, following Johannes's suggestion, this implementation moves the
O(FDs that have written) behavior onto the FD write(2) path. Instead, on
the page-allocation path, we simply add one additional watermark to
conditionally bump per-hierarchy level in the page-counter.
Additionally, this takes Longman's suggestion of nesting the
page-charging-path checks for the two watermarks to reduce the number of
common-case comparisons.
This behavior is particularly useful for work scheduling systems that need
to track memory usage of worker processes/cgroups per-work-item. Since
memory can't be squeezed like CPU can (the OOM-killer has opinions), these
systems need to track the peak memory usage to compute system/container
fullness when binpacking workitems.
Most notably, Vimeo's use-case involves a system that's doing global
binpacking across many Kubernetes pods/containers, and while we can use
PSI for some local decisions about overload, we strive to avoid packing
workloads too tightly in the first place. To facilitate this, we track
the peak memory usage. However, since we run with long-lived workers (to
amortize startup costs) we need a way to track the high watermark while a
work-item is executing. Polling runs the risk of missing short spikes
that last for timescales below the polling interval, and peak memory
tracking at the cgroup level is otherwise perfect for this use-case.
As this data is used to ensure that binpacked work ends up with sufficient
headroom, this use-case mostly avoids the inaccuracies surrounding
reclaimable memory.
Link: https://lkml.kernel.org/r/20240730231304.761942-1-davidf@vimeo.com
Link: https://lkml.kernel.org/r/20240729143743.34236-1-davidf@vimeo.com
Link: https://lkml.kernel.org/r/20240729143743.34236-2-davidf@vimeo.com
Signed-off-by: David Finkel <davidf@vimeo.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Waiman Long <longman@redhat.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-07-29 10:37:42 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-05-13 16:48:57 -07:00
|
|
|
|
mm, memcg: cg2 memory{.swap,}.peak write handlers
Patch series "mm, memcg: cg2 memory{.swap,}.peak write handlers", v7.
This patch (of 2):
Other mechanisms for querying the peak memory usage of either a process or
v1 memory cgroup allow for resetting the high watermark. Restore parity
with those mechanisms, but with a less racy API.
For example:
- Any write to memory.max_usage_in_bytes in a cgroup v1 mount resets
the high watermark.
- writing "5" to the clear_refs pseudo-file in a processes's proc
directory resets the peak RSS.
This change is an evolution of a previous patch, which mostly copied the
cgroup v1 behavior, however, there were concerns about races/ownership
issues with a global reset, so instead this change makes the reset
filedescriptor-local.
Writing any non-empty string to the memory.peak and memory.swap.peak
pseudo-files reset the high watermark to the current usage for subsequent
reads through that same FD.
Notably, following Johannes's suggestion, this implementation moves the
O(FDs that have written) behavior onto the FD write(2) path. Instead, on
the page-allocation path, we simply add one additional watermark to
conditionally bump per-hierarchy level in the page-counter.
Additionally, this takes Longman's suggestion of nesting the
page-charging-path checks for the two watermarks to reduce the number of
common-case comparisons.
This behavior is particularly useful for work scheduling systems that need
to track memory usage of worker processes/cgroups per-work-item. Since
memory can't be squeezed like CPU can (the OOM-killer has opinions), these
systems need to track the peak memory usage to compute system/container
fullness when binpacking workitems.
Most notably, Vimeo's use-case involves a system that's doing global
binpacking across many Kubernetes pods/containers, and while we can use
PSI for some local decisions about overload, we strive to avoid packing
workloads too tightly in the first place. To facilitate this, we track
the peak memory usage. However, since we run with long-lived workers (to
amortize startup costs) we need a way to track the high watermark while a
work-item is executing. Polling runs the risk of missing short spikes
that last for timescales below the polling interval, and peak memory
tracking at the cgroup level is otherwise perfect for this use-case.
As this data is used to ensure that binpacked work ends up with sufficient
headroom, this use-case mostly avoids the inaccuracies surrounding
reclaimable memory.
Link: https://lkml.kernel.org/r/20240730231304.761942-1-davidf@vimeo.com
Link: https://lkml.kernel.org/r/20240729143743.34236-1-davidf@vimeo.com
Link: https://lkml.kernel.org/r/20240729143743.34236-2-davidf@vimeo.com
Signed-off-by: David Finkel <davidf@vimeo.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Waiman Long <longman@redhat.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-07-29 10:37:42 -04:00
|
|
|
|
|
|
|
|
|
2022-05-13 16:48:57 -07:00
|
|
|
|
|
|
|
|
|
mm, memcg: cg2 memory{.swap,}.peak write handlers
Patch series "mm, memcg: cg2 memory{.swap,}.peak write handlers", v7.
This patch (of 2):
Other mechanisms for querying the peak memory usage of either a process or
v1 memory cgroup allow for resetting the high watermark. Restore parity
with those mechanisms, but with a less racy API.
For example:
- Any write to memory.max_usage_in_bytes in a cgroup v1 mount resets
the high watermark.
- writing "5" to the clear_refs pseudo-file in a processes's proc
directory resets the peak RSS.
This change is an evolution of a previous patch, which mostly copied the
cgroup v1 behavior, however, there were concerns about races/ownership
issues with a global reset, so instead this change makes the reset
filedescriptor-local.
Writing any non-empty string to the memory.peak and memory.swap.peak
pseudo-files reset the high watermark to the current usage for subsequent
reads through that same FD.
Notably, following Johannes's suggestion, this implementation moves the
O(FDs that have written) behavior onto the FD write(2) path. Instead, on
the page-allocation path, we simply add one additional watermark to
conditionally bump per-hierarchy level in the page-counter.
Additionally, this takes Longman's suggestion of nesting the
page-charging-path checks for the two watermarks to reduce the number of
common-case comparisons.
This behavior is particularly useful for work scheduling systems that need
to track memory usage of worker processes/cgroups per-work-item. Since
memory can't be squeezed like CPU can (the OOM-killer has opinions), these
systems need to track the peak memory usage to compute system/container
fullness when binpacking workitems.
Most notably, Vimeo's use-case involves a system that's doing global
binpacking across many Kubernetes pods/containers, and while we can use
PSI for some local decisions about overload, we strive to avoid packing
workloads too tightly in the first place. To facilitate this, we track
the peak memory usage. However, since we run with long-lived workers (to
amortize startup costs) we need a way to track the high watermark while a
work-item is executing. Polling runs the risk of missing short spikes
that last for timescales below the polling interval, and peak memory
tracking at the cgroup level is otherwise perfect for this use-case.
As this data is used to ensure that binpacked work ends up with sufficient
headroom, this use-case mostly avoids the inaccuracies surrounding
reclaimable memory.
Link: https://lkml.kernel.org/r/20240730231304.761942-1-davidf@vimeo.com
Link: https://lkml.kernel.org/r/20240729143743.34236-1-davidf@vimeo.com
Link: https://lkml.kernel.org/r/20240729143743.34236-2-davidf@vimeo.com
Signed-off-by: David Finkel <davidf@vimeo.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Waiman Long <longman@redhat.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-07-29 10:37:42 -04:00
|
|
|
|
|
|
|
|
|
2018-06-07 17:07:46 -07:00
|
|
|
|
|
|
|
|
|
2019-03-05 15:45:55 -08:00
|
|
|
|
|
|
|
|
|
2018-06-07 17:07:46 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
|
|
|
|
|
2019-03-05 15:45:55 -08:00
|
|
|
|
|
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-02-27 15:52:04 -08:00
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memory.low hierarchical behavior
This patch aims to address an issue in current memory.low semantics,
which makes it hard to use it in a hierarchy, where some leaf memory
cgroups are more valuable than others.
For example, there are memcgs A, A/B, A/C, A/D and A/E:
A A/memory.low = 2G, A/memory.current = 6G
//\\
BC DE B/memory.low = 3G B/memory.current = 2G
C/memory.low = 1G C/memory.current = 2G
D/memory.low = 0 D/memory.current = 2G
E/memory.low = 10G E/memory.current = 0
If we apply memory pressure, B, C and D are reclaimed at the same pace
while A's usage exceeds 2G. This is obviously wrong, as B's usage is
fully below B's memory.low, and C has 1G of protection as well. Also, A
is pushed to the size, which is less than A's 2G memory.low, which is
also wrong.
A simple bash script (provided below) can be used to reproduce
the problem. Current results are:
A: 1430097920
A/B: 711929856
A/C: 717426688
A/D: 741376
A/E: 0
To address the issue a concept of effective memory.low is introduced.
Effective memory.low is always equal or less than original memory.low.
In a case, when there is no memory.low overcommittment (and also for
top-level cgroups), these two values are equal.
Otherwise it's a part of parent's effective memory.low, calculated as a
cgroup's memory.low usage divided by sum of sibling's memory.low usages
(under memory.low usage I mean the size of actually protected memory:
memory.current if memory.current < memory.low, 0 otherwise). It's
necessary to track the actual usage, because otherwise an empty cgroup
with memory.low set (A/E in my example) will affect actual memory
distribution, which makes no sense. To avoid traversing the cgroup tree
twice, page_counters code is reused.
Calculating effective memory.low can be done in the reclaim path, as we
conveniently traversing the cgroup tree from top to bottom and check
memory.low on each level. So, it's a perfect place to calculate
effective memory low and save it to use it for children cgroups.
This also eliminates a need to traverse the cgroup tree from bottom to
top each time to check if parent's guarantee is not exceeded.
Setting/resetting effective memory.low is intentionally racy, but it's
fine and shouldn't lead to any significant differences in actual memory
distribution.
With this patch applied results are matching the expectations:
A: 2147930112
A/B: 1428721664
A/C: 718393344
A/D: 815104
A/E: 0
Test script:
#!/bin/bash
CGPATH="/sys/fs/cgroup"
truncate /file1 --size 2G
truncate /file2 --size 2G
truncate /file3 --size 2G
truncate /file4 --size 50G
mkdir "${CGPATH}/A"
echo "+memory" > "${CGPATH}/A/cgroup.subtree_control"
mkdir "${CGPATH}/A/B" "${CGPATH}/A/C" "${CGPATH}/A/D" "${CGPATH}/A/E"
echo 2G > "${CGPATH}/A/memory.low"
echo 3G > "${CGPATH}/A/B/memory.low"
echo 1G > "${CGPATH}/A/C/memory.low"
echo 0 > "${CGPATH}/A/D/memory.low"
echo 10G > "${CGPATH}/A/E/memory.low"
echo $$ > "${CGPATH}/A/B/cgroup.procs" && vmtouch -qt /file1
echo $$ > "${CGPATH}/A/C/cgroup.procs" && vmtouch -qt /file2
echo $$ > "${CGPATH}/A/D/cgroup.procs" && vmtouch -qt /file3
echo $$ > "${CGPATH}/cgroup.procs" && vmtouch -qt /file4
echo "A: " `cat "${CGPATH}/A/memory.current"`
echo "A/B: " `cat "${CGPATH}/A/B/memory.current"`
echo "A/C: " `cat "${CGPATH}/A/C/memory.current"`
echo "A/D: " `cat "${CGPATH}/A/D/memory.current"`
echo "A/E: " `cat "${CGPATH}/A/E/memory.current"`
rmdir "${CGPATH}/A/B" "${CGPATH}/A/C" "${CGPATH}/A/D" "${CGPATH}/A/E"
rmdir "${CGPATH}/A"
rm /file1 /file2 /file3 /file4
Link: http://lkml.kernel.org/r/20180405185921.4942-2-guro@fb.com
Signed-off-by: Roman Gushchin <guro@fb.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-06-07 17:06:22 -07:00
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-06-01 21:49:49 -07:00
|
|
|
|
|
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-08-06 23:21:58 -07:00
|
|
|
|
2019-11-30 17:50:09 -08:00
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-02-27 15:52:04 -08:00
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Revert "mm: memcontrol: avoid workload stalls when lowering memory.high"
This reverts commit 536d3bf261a2fc3b05b3e91e7eef7383443015cf, as it can
cause writers to memory.high to get stuck in the kernel forever,
performing page reclaim and consuming excessive amounts of CPU cycles.
Before the patch, a write to memory.high would first put the new limit
in place for the workload, and then reclaim the requested delta. After
the patch, the kernel tries to reclaim the delta before putting the new
limit into place, in order to not overwhelm the workload with a sudden,
large excess over the limit. However, if reclaim is actively racing
with new allocations from the uncurbed workload, it can keep the write()
working inside the kernel indefinitely.
This is causing problems in Facebook production. A privileged
system-level daemon that adjusts memory.high for various workloads
running on a host can get unexpectedly stuck in the kernel and
essentially turn into a sort of involuntary kswapd for one of the
workloads. We've observed that daemon busy-spin in a write() for
minutes at a time, neglecting its other duties on the system, and
expending privileged system resources on behalf of a workload.
To remedy this, we have first considered changing the reclaim logic to
break out after a couple of loops - whether the workload has converged
to the new limit or not - and bound the write() call this way. However,
the root cause that inspired the sequence change in the first place has
been fixed through other means, and so a revert back to the proven
limit-setting sequence, also used by memory.max, is preferable.
The sequence was changed to avoid extreme latencies in the workload when
the limit was lowered: the sudden, large excess created by the limit
lowering would erroneously trigger the penalty sleeping code that is
meant to throttle excessive growth from below. Allocating threads could
end up sleeping long after the write() had already reclaimed the delta
for which they were being punished.
However, erroneous throttling also caused problems in other scenarios at
around the same time. This resulted in commit b3ff92916af3 ("mm, memcg:
reclaim more aggressively before high allocator throttling"), included
in the same release as the offending commit. When allocating threads
now encounter large excess caused by a racing write() to memory.high,
instead of entering punitive sleeps, they will simply be tasked with
helping reclaim down the excess, and will be held no longer than it
takes to accomplish that. This is in line with regular limit
enforcement - i.e. if the workload allocates up against or over an
otherwise unchanged limit from below.
With the patch breaking userspace, and the root cause addressed by other
means already, revert it again.
Link: https://lkml.kernel.org/r/20210122184341.292461-1-hannes@cmpxchg.org
Fixes: 536d3bf261a2 ("mm: memcontrol: avoid workload stalls when lowering memory.high")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reported-by: Tejun Heo <tj@kernel.org>
Acked-by: Chris Down <chris@chrisdown.name>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Michal Koutný <mkoutny@suse.com>
Cc: <stable@vger.kernel.org> [5.8+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-09 13:42:28 -08:00
|
|
|
|
|
|
|
|
|
memcg: introduce non-blocking limit setting option
Setting the max and high limits can trigger synchronous reclaim and/or
oom-kill if the usage is higher than the given limit. This behavior is
fine for newly created cgroups but it can cause issues for the node
controller while setting limits for existing cgroups.
In our production multi-tenant and overcommitted environment, we are
seeing priority inversion when the node controller dynamically adjusts the
limits of running jobs of different priorities. Based on the system
situation, the node controller may reduce the limits of lower priority
jobs and increase the limits of higher priority jobs. However we are
seeing node controller getting stuck for long period of time while
reclaiming from lower priority jobs while setting their limits and also
spends a lot of its own CPU.
One of the workaround we are trying is to fork a new process which sets
the limit of the lower priority job along with setting an alarm to get
itself killed if it get stuck in the reclaim for lower priority job.
However we are finding it very unreliable and costly. Either we need a
good enough time buffer for the alarm to be delivered after setting limit
and potentialy spend a lot of CPU in the reclaim or be unreliable in
setting the limit for much shorter but cheaper (less reclaim) alarms.
Let's introduce new limit setting option which does not trigger reclaim
and/or oom-kill and let the processes in the target cgroup to trigger
reclaim and/or throttling and/or oom-kill in their next charge request.
This will make the node controller on multi-tenant overcommitted
environment much more reliable.
Explanation from Johannes on side-effects of O_NONBLOCK limit change:
It's usually the allocating tasks inside the group bearing the cost of
limit enforcement and reclaim. This allows a (privileged) updater from
outside the group to keep that cost in there - instead of having to
help, from a context that doesn't necessarily make sense.
I suppose the tradeoff with that - and the reason why this was doing
sync reclaim in the first place - is that, if the group is idle and
not trying to allocate more, it can take indefinitely for the new
limit to actually be met.
It should be okay in most scenarios in practice. As the capacity is
reallocated from group A to B, B will exert pressure on A once it
tries to claim it and thereby shrink it down. If A is idle, that
shouldn't be hard. If A is running, it's likely to fault/allocate
soon-ish and then join the effort.
It does leave a (malicious) corner case where A is just busy-hitting
its memory to interfere with the clawback. This is comparable to
reclaiming memory.low overage from the outside, though, which is an
acceptable risk. Users of O_NONBLOCK just need to be aware.
Link: https://lkml.kernel.org/r/20250419183545.1982187-1-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Acked-by: Roman Gushchin <roman.gushchin@linux.dev>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Michal Koutný <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Tejun Heo <tj@kernel.org>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Yosry Ahmed <yosry.ahmed@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-19 11:35:45 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-11-30 17:50:09 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-01-03 08:48:37 -08:00
|
|
|
|
2019-11-30 17:50:09 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
memcg: introduce non-blocking limit setting option
Setting the max and high limits can trigger synchronous reclaim and/or
oom-kill if the usage is higher than the given limit. This behavior is
fine for newly created cgroups but it can cause issues for the node
controller while setting limits for existing cgroups.
In our production multi-tenant and overcommitted environment, we are
seeing priority inversion when the node controller dynamically adjusts the
limits of running jobs of different priorities. Based on the system
situation, the node controller may reduce the limits of lower priority
jobs and increase the limits of higher priority jobs. However we are
seeing node controller getting stuck for long period of time while
reclaiming from lower priority jobs while setting their limits and also
spends a lot of its own CPU.
One of the workaround we are trying is to fork a new process which sets
the limit of the lower priority job along with setting an alarm to get
itself killed if it get stuck in the reclaim for lower priority job.
However we are finding it very unreliable and costly. Either we need a
good enough time buffer for the alarm to be delivered after setting limit
and potentialy spend a lot of CPU in the reclaim or be unreliable in
setting the limit for much shorter but cheaper (less reclaim) alarms.
Let's introduce new limit setting option which does not trigger reclaim
and/or oom-kill and let the processes in the target cgroup to trigger
reclaim and/or throttling and/or oom-kill in their next charge request.
This will make the node controller on multi-tenant overcommitted
environment much more reliable.
Explanation from Johannes on side-effects of O_NONBLOCK limit change:
It's usually the allocating tasks inside the group bearing the cost of
limit enforcement and reclaim. This allows a (privileged) updater from
outside the group to keep that cost in there - instead of having to
help, from a context that doesn't necessarily make sense.
I suppose the tradeoff with that - and the reason why this was doing
sync reclaim in the first place - is that, if the group is idle and
not trying to allocate more, it can take indefinitely for the new
limit to actually be met.
It should be okay in most scenarios in practice. As the capacity is
reallocated from group A to B, B will exert pressure on A once it
tries to claim it and thereby shrink it down. If A is idle, that
shouldn't be hard. If A is running, it's likely to fault/allocate
soon-ish and then join the effort.
It does leave a (malicious) corner case where A is just busy-hitting
its memory to interfere with the clawback. This is comparable to
reclaiming memory.low overage from the outside, though, which is an
acceptable risk. Users of O_NONBLOCK just need to be aware.
Link: https://lkml.kernel.org/r/20250419183545.1982187-1-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Acked-by: Roman Gushchin <roman.gushchin@linux.dev>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Michal Koutný <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Tejun Heo <tj@kernel.org>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Yosry Ahmed <yosry.ahmed@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-19 11:35:45 -07:00
|
|
|
|
2020-08-06 23:22:12 -07:00
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-03-05 15:45:55 -08:00
|
|
|
|
|
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-08-06 23:21:58 -07:00
|
|
|
|
2016-03-17 14:20:28 -07:00
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-02-27 15:52:04 -08:00
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-06-07 17:06:18 -07:00
|
|
|
|
2016-03-17 14:20:28 -07:00
|
|
|
|
memcg: introduce non-blocking limit setting option
Setting the max and high limits can trigger synchronous reclaim and/or
oom-kill if the usage is higher than the given limit. This behavior is
fine for newly created cgroups but it can cause issues for the node
controller while setting limits for existing cgroups.
In our production multi-tenant and overcommitted environment, we are
seeing priority inversion when the node controller dynamically adjusts the
limits of running jobs of different priorities. Based on the system
situation, the node controller may reduce the limits of lower priority
jobs and increase the limits of higher priority jobs. However we are
seeing node controller getting stuck for long period of time while
reclaiming from lower priority jobs while setting their limits and also
spends a lot of its own CPU.
One of the workaround we are trying is to fork a new process which sets
the limit of the lower priority job along with setting an alarm to get
itself killed if it get stuck in the reclaim for lower priority job.
However we are finding it very unreliable and costly. Either we need a
good enough time buffer for the alarm to be delivered after setting limit
and potentialy spend a lot of CPU in the reclaim or be unreliable in
setting the limit for much shorter but cheaper (less reclaim) alarms.
Let's introduce new limit setting option which does not trigger reclaim
and/or oom-kill and let the processes in the target cgroup to trigger
reclaim and/or throttling and/or oom-kill in their next charge request.
This will make the node controller on multi-tenant overcommitted
environment much more reliable.
Explanation from Johannes on side-effects of O_NONBLOCK limit change:
It's usually the allocating tasks inside the group bearing the cost of
limit enforcement and reclaim. This allows a (privileged) updater from
outside the group to keep that cost in there - instead of having to
help, from a context that doesn't necessarily make sense.
I suppose the tradeoff with that - and the reason why this was doing
sync reclaim in the first place - is that, if the group is idle and
not trying to allocate more, it can take indefinitely for the new
limit to actually be met.
It should be okay in most scenarios in practice. As the capacity is
reallocated from group A to B, B will exert pressure on A once it
tries to claim it and thereby shrink it down. If A is idle, that
shouldn't be hard. If A is running, it's likely to fault/allocate
soon-ish and then join the effort.
It does leave a (malicious) corner case where A is just busy-hitting
its memory to interfere with the clawback. This is comparable to
reclaiming memory.low overage from the outside, though, which is an
acceptable risk. Users of O_NONBLOCK just need to be aware.
Link: https://lkml.kernel.org/r/20250419183545.1982187-1-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Acked-by: Roman Gushchin <roman.gushchin@linux.dev>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Michal Koutný <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Tejun Heo <tj@kernel.org>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Yosry Ahmed <yosry.ahmed@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-19 11:35:45 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-03-17 14:20:28 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-11-30 17:50:06 -08:00
|
|
|
|
2016-03-17 14:20:28 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-01-03 08:48:37 -08:00
|
|
|
|
2016-03-17 14:20:28 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-04-10 16:29:45 -07:00
|
|
|
|
2016-03-17 14:20:28 -07:00
|
|
|
|
|
|
|
|
|
2025-02-11 08:18:19 +00:00
|
|
|
|
2016-03-17 14:20:28 -07:00
|
|
|
|
memcg: introduce non-blocking limit setting option
Setting the max and high limits can trigger synchronous reclaim and/or
oom-kill if the usage is higher than the given limit. This behavior is
fine for newly created cgroups but it can cause issues for the node
controller while setting limits for existing cgroups.
In our production multi-tenant and overcommitted environment, we are
seeing priority inversion when the node controller dynamically adjusts the
limits of running jobs of different priorities. Based on the system
situation, the node controller may reduce the limits of lower priority
jobs and increase the limits of higher priority jobs. However we are
seeing node controller getting stuck for long period of time while
reclaiming from lower priority jobs while setting their limits and also
spends a lot of its own CPU.
One of the workaround we are trying is to fork a new process which sets
the limit of the lower priority job along with setting an alarm to get
itself killed if it get stuck in the reclaim for lower priority job.
However we are finding it very unreliable and costly. Either we need a
good enough time buffer for the alarm to be delivered after setting limit
and potentialy spend a lot of CPU in the reclaim or be unreliable in
setting the limit for much shorter but cheaper (less reclaim) alarms.
Let's introduce new limit setting option which does not trigger reclaim
and/or oom-kill and let the processes in the target cgroup to trigger
reclaim and/or throttling and/or oom-kill in their next charge request.
This will make the node controller on multi-tenant overcommitted
environment much more reliable.
Explanation from Johannes on side-effects of O_NONBLOCK limit change:
It's usually the allocating tasks inside the group bearing the cost of
limit enforcement and reclaim. This allows a (privileged) updater from
outside the group to keep that cost in there - instead of having to
help, from a context that doesn't necessarily make sense.
I suppose the tradeoff with that - and the reason why this was doing
sync reclaim in the first place - is that, if the group is idle and
not trying to allocate more, it can take indefinitely for the new
limit to actually be met.
It should be okay in most scenarios in practice. As the capacity is
reallocated from group A to B, B will exert pressure on A once it
tries to claim it and thereby shrink it down. If A is idle, that
shouldn't be hard. If A is running, it's likely to fault/allocate
soon-ish and then join the effort.
It does leave a (malicious) corner case where A is just busy-hitting
its memory to interfere with the clawback. This is comparable to
reclaiming memory.low overage from the outside, though, which is an
acceptable risk. Users of O_NONBLOCK just need to be aware.
Link: https://lkml.kernel.org/r/20250419183545.1982187-1-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Acked-by: Roman Gushchin <roman.gushchin@linux.dev>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Michal Koutný <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Tejun Heo <tj@kernel.org>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Yosry Ahmed <yosry.ahmed@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-19 11:35:45 -07:00
|
|
|
|
2015-05-22 18:23:34 -04:00
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-11-23 10:19:45 +03:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-07-11 20:55:55 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-01-14 14:05:35 -08:00
|
|
|
|
|
|
|
|
|
2019-07-11 20:55:55 -07:00
|
|
|
|
|
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
|
|
|
|
|
2019-03-05 15:45:52 -08:00
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
2019-07-11 20:55:55 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
2019-07-11 20:55:55 -07:00
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-06-24 17:59:02 -07:00
|
|
|
|
2016-01-20 15:03:19 -08:00
|
|
|
|
2019-03-05 15:45:52 -08:00
|
|
|
|
2024-10-21 14:00:26 +01:00
|
|
|
|
2023-04-28 13:24:05 +00:00
|
|
|
|
2019-03-05 15:48:09 -08:00
|
|
|
|
mm: memcontrol: dump memory.stat during cgroup OOM
The current cgroup OOM memory info dump doesn't include all the memory
we are tracking, nor does it give insight into what the VM tried to do
leading up to the OOM. All that useful info is in memory.stat.
Furthermore, the recursive printing for every child cgroup can
generate absurd amounts of data on the console for larger cgroup
trees, and it's not like we provide a per-cgroup breakdown during
global OOM kills.
When an OOM kill is triggered, print one set of recursive memory.stat
items at the level whose limit triggered the OOM condition.
Example output:
stress invoked oom-killer: gfp_mask=0x100cca(GFP_HIGHUSER_MOVABLE), order=0, oom_score_adj=0
CPU: 2 PID: 210 Comm: stress Not tainted 5.2.0-rc2-mm1-00247-g47d49835983c #135
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.12.0-20181126_142135-anatol 04/01/2014
Call Trace:
dump_stack+0x46/0x60
dump_header+0x4c/0x2d0
oom_kill_process.cold.10+0xb/0x10
out_of_memory+0x200/0x270
? try_to_free_mem_cgroup_pages+0xdf/0x130
mem_cgroup_out_of_memory+0xb7/0xc0
try_charge+0x680/0x6f0
mem_cgroup_try_charge+0xb5/0x160
__add_to_page_cache_locked+0xc6/0x300
? list_lru_destroy+0x80/0x80
add_to_page_cache_lru+0x45/0xc0
pagecache_get_page+0x11b/0x290
filemap_fault+0x458/0x6d0
ext4_filemap_fault+0x27/0x36
__do_fault+0x2f/0xb0
__handle_mm_fault+0x9c5/0x1140
? apic_timer_interrupt+0xa/0x20
handle_mm_fault+0xc5/0x180
__do_page_fault+0x1ab/0x440
? page_fault+0x8/0x30
page_fault+0x1e/0x30
RIP: 0033:0x55c32167fc10
Code: Bad RIP value.
RSP: 002b:00007fff1d031c50 EFLAGS: 00010206
RAX: 000000000dc00000 RBX: 00007fd2db000010 RCX: 00007fd2db000010
RDX: 0000000000000000 RSI: 0000000010001000 RDI: 0000000000000000
RBP: 000055c321680a54 R08: 00000000ffffffff R09: 0000000000000000
R10: 0000000000000022 R11: 0000000000000246 R12: ffffffffffffffff
R13: 0000000000000002 R14: 0000000000001000 R15: 0000000010000000
memory: usage 1024kB, limit 1024kB, failcnt 75131
swap: usage 0kB, limit 9007199254740988kB, failcnt 0
Memory cgroup stats for /foo:
anon 0
file 0
kernel_stack 36864
slab 274432
sock 0
shmem 0
file_mapped 0
file_dirty 0
file_writeback 0
anon_thp 0
inactive_anon 126976
active_anon 0
inactive_file 0
active_file 0
unevictable 0
slab_reclaimable 0
slab_unreclaimable 274432
pgfault 59466
pgmajfault 1617
workingset_refault 2145
workingset_activate 0
workingset_nodereclaim 0
pgrefill 98952
pgscan 200060
pgsteal 59340
pgactivate 40095
pgdeactivate 96787
pglazyfree 0
pglazyfreed 0
thp_fault_alloc 0
thp_collapse_alloc 0
Tasks state (memory values in pages):
[ pid ] uid tgid total_vm rss pgtables_bytes swapents oom_score_adj name
[ 200] 0 200 1121 884 53248 29 0 bash
[ 209] 0 209 905 246 45056 19 0 stress
[ 210] 0 210 66442 56 499712 56349 0 stress
oom-kill:constraint=CONSTRAINT_NONE,nodemask=(null),oom_memcg=/foo,task_memcg=/foo,task=stress,pid=210,uid=0
Memory cgroup out of memory: Killed process 210 (stress) total-vm:265768kB, anon-rss:0kB, file-rss:224kB, shmem-rss:0kB
oom_reaper: reaped process 210 (stress), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
[hannes@cmpxchg.org: s/kvmalloc/kmalloc/ per Michal]
Link: http://lkml.kernel.org/r/20190605161133.GA12453@cmpxchg.org
Link: http://lkml.kernel.org/r/20190604210509.9744-1-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-11 20:55:59 -07:00
|
|
|
|
|
|
|
|
|
2024-10-21 14:00:26 +01:00
|
|
|
|
2023-04-28 13:24:05 +00:00
|
|
|
|
mm: memcontrol: dump memory.stat during cgroup OOM
The current cgroup OOM memory info dump doesn't include all the memory
we are tracking, nor does it give insight into what the VM tried to do
leading up to the OOM. All that useful info is in memory.stat.
Furthermore, the recursive printing for every child cgroup can
generate absurd amounts of data on the console for larger cgroup
trees, and it's not like we provide a per-cgroup breakdown during
global OOM kills.
When an OOM kill is triggered, print one set of recursive memory.stat
items at the level whose limit triggered the OOM condition.
Example output:
stress invoked oom-killer: gfp_mask=0x100cca(GFP_HIGHUSER_MOVABLE), order=0, oom_score_adj=0
CPU: 2 PID: 210 Comm: stress Not tainted 5.2.0-rc2-mm1-00247-g47d49835983c #135
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.12.0-20181126_142135-anatol 04/01/2014
Call Trace:
dump_stack+0x46/0x60
dump_header+0x4c/0x2d0
oom_kill_process.cold.10+0xb/0x10
out_of_memory+0x200/0x270
? try_to_free_mem_cgroup_pages+0xdf/0x130
mem_cgroup_out_of_memory+0xb7/0xc0
try_charge+0x680/0x6f0
mem_cgroup_try_charge+0xb5/0x160
__add_to_page_cache_locked+0xc6/0x300
? list_lru_destroy+0x80/0x80
add_to_page_cache_lru+0x45/0xc0
pagecache_get_page+0x11b/0x290
filemap_fault+0x458/0x6d0
ext4_filemap_fault+0x27/0x36
__do_fault+0x2f/0xb0
__handle_mm_fault+0x9c5/0x1140
? apic_timer_interrupt+0xa/0x20
handle_mm_fault+0xc5/0x180
__do_page_fault+0x1ab/0x440
? page_fault+0x8/0x30
page_fault+0x1e/0x30
RIP: 0033:0x55c32167fc10
Code: Bad RIP value.
RSP: 002b:00007fff1d031c50 EFLAGS: 00010206
RAX: 000000000dc00000 RBX: 00007fd2db000010 RCX: 00007fd2db000010
RDX: 0000000000000000 RSI: 0000000010001000 RDI: 0000000000000000
RBP: 000055c321680a54 R08: 00000000ffffffff R09: 0000000000000000
R10: 0000000000000022 R11: 0000000000000246 R12: ffffffffffffffff
R13: 0000000000000002 R14: 0000000000001000 R15: 0000000010000000
memory: usage 1024kB, limit 1024kB, failcnt 75131
swap: usage 0kB, limit 9007199254740988kB, failcnt 0
Memory cgroup stats for /foo:
anon 0
file 0
kernel_stack 36864
slab 274432
sock 0
shmem 0
file_mapped 0
file_dirty 0
file_writeback 0
anon_thp 0
inactive_anon 126976
active_anon 0
inactive_file 0
active_file 0
unevictable 0
slab_reclaimable 0
slab_unreclaimable 274432
pgfault 59466
pgmajfault 1617
workingset_refault 2145
workingset_activate 0
workingset_nodereclaim 0
pgrefill 98952
pgscan 200060
pgsteal 59340
pgactivate 40095
pgdeactivate 96787
pglazyfree 0
pglazyfreed 0
thp_fault_alloc 0
thp_collapse_alloc 0
Tasks state (memory values in pages):
[ pid ] uid tgid total_vm rss pgtables_bytes swapents oom_score_adj name
[ 200] 0 200 1121 884 53248 29 0 bash
[ 209] 0 209 905 246 45056 19 0 stress
[ 210] 0 210 66442 56 499712 56349 0 stress
oom-kill:constraint=CONSTRAINT_NONE,nodemask=(null),oom_memcg=/foo,task_memcg=/foo,task=stress,pid=210,uid=0
Memory cgroup out of memory: Killed process 210 (stress) total-vm:265768kB, anon-rss:0kB, file-rss:224kB, shmem-rss:0kB
oom_reaper: reaped process 210 (stress), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
[hannes@cmpxchg.org: s/kvmalloc/kmalloc/ per Michal]
Link: http://lkml.kernel.org/r/20190605161133.GA12453@cmpxchg.org
Link: http://lkml.kernel.org/r/20190604210509.9744-1-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-11 20:55:59 -07:00
|
|
|
|
|
|
|
|
|
2016-01-20 15:03:19 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-10-13 16:52:59 -07:00
|
|
|
|
2021-02-24 12:03:43 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-09-22 17:57:39 +00:00
|
|
|
|
|
|
|
|
|
2021-02-24 12:03:43 -08:00
|
|
|
|
|
|
|
|
|
2020-10-13 16:52:59 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcg: restore subtree stats flushing
Stats flushing for memcg currently follows the following rules:
- Always flush the entire memcg hierarchy (i.e. flush the root).
- Only one flusher is allowed at a time. If someone else tries to flush
concurrently, they skip and return immediately.
- A periodic flusher flushes all the stats every 2 seconds.
The reason this approach is followed is because all flushes are serialized
by a global rstat spinlock. On the memcg side, flushing is invoked from
userspace reads as well as in-kernel flushers (e.g. reclaim, refault,
etc). This approach aims to avoid serializing all flushers on the global
lock, which can cause a significant performance hit under high
concurrency.
This approach has the following problems:
- Occasionally a userspace read of the stats of a non-root cgroup will
be too expensive as it has to flush the entire hierarchy [1].
- Sometimes the stats accuracy are compromised if there is an ongoing
flush, and we skip and return before the subtree of interest is
actually flushed, yielding stale stats (by up to 2s due to periodic
flushing). This is more visible when reading stats from userspace,
but can also affect in-kernel flushers.
The latter problem is particulary a concern when userspace reads stats
after an event occurs, but gets stats from before the event. Examples:
- When memory usage / pressure spikes, a userspace OOM handler may look
at the stats of different memcgs to select a victim based on various
heuristics (e.g. how much private memory will be freed by killing
this). Reading stale stats from before the usage spike in this case
may cause a wrongful OOM kill.
- A proactive reclaimer may read the stats after writing to
memory.reclaim to measure the success of the reclaim operation. Stale
stats from before reclaim may give a false negative.
- Reading the stats of a parent and a child memcg may be inconsistent
(child larger than parent), if the flush doesn't happen when the
parent is read, but happens when the child is read.
As for in-kernel flushers, they will occasionally get stale stats. No
regressions are currently known from this, but if there are regressions,
they would be very difficult to debug and link to the source of the
problem.
This patch aims to fix these problems by restoring subtree flushing, and
removing the unified/coalesced flushing logic that skips flushing if there
is an ongoing flush. This change would introduce a significant regression
with global stats flushing thresholds. With per-memcg stats flushing
thresholds, this seems to perform really well. The thresholds protect the
underlying lock from unnecessary contention.
This patch was tested in two ways to ensure the latency of flushing is
up to par, on a machine with 384 cpus:
- A synthetic test with 5000 concurrent workers in 500 cgroups doing
allocations and reclaim, as well as 1000 readers for memory.stat
(variation of [2]). No regressions were noticed in the total runtime.
Note that significant regressions in this test are observed with
global stats thresholds, but not with per-memcg thresholds.
- A synthetic stress test for concurrently reading memcg stats while
memory allocation/freeing workers are running in the background,
provided by Wei Xu [3]. With 250k threads reading the stats every
100ms in 50k cgroups, 99.9% of reads take <= 50us. Less than 0.01%
of reads take more than 1ms, and no reads take more than 100ms.
[1] https://lore.kernel.org/lkml/CABWYdi0c6__rh-K7dcM_pkf9BJdTRtAU08M43KO9ME4-dsgfoQ@mail.gmail.com/
[2] https://lore.kernel.org/lkml/CAJD7tka13M-zVZTyQJYL1iUAYvuQ1fcHbCjcOBZcz6POYTV-4g@mail.gmail.com/
[3] https://lore.kernel.org/lkml/CAAPL-u9D2b=iF5Lf_cRnKxUfkiEe0AMDTu6yhrUAzX0b6a6rDg@mail.gmail.com/
[akpm@linux-foundation.org: fix mm/zswap.c]
[yosryahmed@google.com: remove stats flushing mutex]
Link: https://lkml.kernel.org/r/CAJD7tkZgP3m-VVPn+fF_YuvXeQYK=tZZjJHj=dzD=CcSSpp2qg@mail.gmail.com
Link: https://lkml.kernel.org/r/20231129032154.3710765-6-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Tested-by: Domenico Cerasuolo <cerasuolodomenico@gmail.com>
Acked-by: Shakeel Butt <shakeelb@google.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: Greg Thelen <gthelen@google.com>
Cc: Ivan Babrou <ivan@cloudflare.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Michal Koutny <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Tejun Heo <tj@kernel.org>
Cc: Waiman Long <longman@redhat.com>
Cc: Wei Xu <weixugc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-11-29 03:21:53 +00:00
|
|
|
|
2021-09-02 14:55:00 -07:00
|
|
|
|
2020-10-13 16:52:59 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-02-24 12:03:43 -08:00
|
|
|
|
|
|
|
|
|
2020-10-13 16:52:59 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-08-21 21:53:54 -07:00
|
|
|
|
|
|
|
|
|
2019-03-05 15:45:52 -08:00
|
|
|
|
2018-08-21 21:53:54 -07:00
|
|
|
|
2023-03-06 23:41:35 +08:00
|
|
|
|
2018-08-21 21:53:54 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-03-06 23:41:35 +08:00
|
|
|
|
2018-08-21 21:53:54 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
memcg: introduce per-memcg reclaim interface
This patch series adds a memory.reclaim proactive reclaim interface.
The rationale behind the interface and how it works are in the first
patch.
This patch (of 4):
Introduce a memcg interface to trigger memory reclaim on a memory cgroup.
Use case: Proactive Reclaim
---------------------------
A userspace proactive reclaimer can continuously probe the memcg to
reclaim a small amount of memory. This gives more accurate and up-to-date
workingset estimation as the LRUs are continuously sorted and can
potentially provide more deterministic memory overcommit behavior. The
memory overcommit controller can provide more proactive response to the
changing behavior of the running applications instead of being reactive.
A userspace reclaimer's purpose in this case is not a complete replacement
for kswapd or direct reclaim, it is to proactively identify memory savings
opportunities and reclaim some amount of cold pages set by the policy to
free up the memory for more demanding jobs or scheduling new jobs.
A user space proactive reclaimer is used in Google data centers.
Additionally, Meta's TMO paper recently referenced a very similar
interface used for user space proactive reclaim:
https://dl.acm.org/doi/pdf/10.1145/3503222.3507731
Benefits of a user space reclaimer:
-----------------------------------
1) More flexible on who should be charged for the cpu of the memory
reclaim. For proactive reclaim, it makes more sense to be centralized.
2) More flexible on dedicating the resources (like cpu). The memory
overcommit controller can balance the cost between the cpu usage and
the memory reclaimed.
3) Provides a way to the applications to keep their LRUs sorted, so,
under memory pressure better reclaim candidates are selected. This
also gives more accurate and uptodate notion of working set for an
application.
Why memory.high is not enough?
------------------------------
- memory.high can be used to trigger reclaim in a memcg and can
potentially be used for proactive reclaim. However there is a big
downside in using memory.high. It can potentially introduce high
reclaim stalls in the target application as the allocations from the
processes or the threads of the application can hit the temporary
memory.high limit.
- Userspace proactive reclaimers usually use feedback loops to decide
how much memory to proactively reclaim from a workload. The metrics
used for this are usually either refaults or PSI, and these metrics will
become messy if the application gets throttled by hitting the high
limit.
- memory.high is a stateful interface, if the userspace proactive
reclaimer crashes for any reason while triggering reclaim it can leave
the application in a bad state.
- If a workload is rapidly expanding, setting memory.high to proactively
reclaim memory can result in actually reclaiming more memory than
intended.
The benefits of such interface and shortcomings of existing interface were
further discussed in this RFC thread:
https://lore.kernel.org/linux-mm/5df21376-7dd1-bf81-8414-32a73cea45dd@google.com/
Interface:
----------
Introducing a very simple memcg interface 'echo 10M > memory.reclaim' to
trigger reclaim in the target memory cgroup.
The interface is introduced as a nested-keyed file to allow for future
optional arguments to be easily added to configure the behavior of
reclaim.
Possible Extensions:
--------------------
- This interface can be extended with an additional parameter or flags
to allow specifying one or more types of memory to reclaim from (e.g.
file, anon, ..).
- The interface can also be extended with a node mask to reclaim from
specific nodes. This has use cases for reclaim-based demotion in memory
tiering systens.
- A similar per-node interface can also be added to support proactive
reclaim and reclaim-based demotion in systems without memcg.
- Add a timeout parameter to make it easier for user space to call the
interface without worrying about being blocked for an undefined amount
of time.
For now, let's keep things simple by adding the basic functionality.
[yosryahmed@google.com: worked on versions v2 onwards, refreshed to
current master, updated commit message based on recent
discussions and use cases]
Link: https://lkml.kernel.org/r/20220425190040.2475377-1-yosryahmed@google.com
Link: https://lkml.kernel.org/r/20220425190040.2475377-2-yosryahmed@google.com
Signed-off-by: Shakeel Butt <shakeelb@google.com>
Co-developed-by: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Wei Xu <weixugc@google.com>
Acked-by: Roman Gushchin <roman.gushchin@linux.dev>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Zefan Li <lizefan.x@bytedance.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Yu Zhao <yuzhao@google.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Chen Wandun <chenwandun@huawei.com>
Cc: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: "Michal Koutn" <mkoutny@suse.com>
Cc: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-04-29 14:36:59 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-06-23 11:58:49 -07:00
|
|
|
|
memcg: introduce per-memcg reclaim interface
This patch series adds a memory.reclaim proactive reclaim interface.
The rationale behind the interface and how it works are in the first
patch.
This patch (of 4):
Introduce a memcg interface to trigger memory reclaim on a memory cgroup.
Use case: Proactive Reclaim
---------------------------
A userspace proactive reclaimer can continuously probe the memcg to
reclaim a small amount of memory. This gives more accurate and up-to-date
workingset estimation as the LRUs are continuously sorted and can
potentially provide more deterministic memory overcommit behavior. The
memory overcommit controller can provide more proactive response to the
changing behavior of the running applications instead of being reactive.
A userspace reclaimer's purpose in this case is not a complete replacement
for kswapd or direct reclaim, it is to proactively identify memory savings
opportunities and reclaim some amount of cold pages set by the policy to
free up the memory for more demanding jobs or scheduling new jobs.
A user space proactive reclaimer is used in Google data centers.
Additionally, Meta's TMO paper recently referenced a very similar
interface used for user space proactive reclaim:
https://dl.acm.org/doi/pdf/10.1145/3503222.3507731
Benefits of a user space reclaimer:
-----------------------------------
1) More flexible on who should be charged for the cpu of the memory
reclaim. For proactive reclaim, it makes more sense to be centralized.
2) More flexible on dedicating the resources (like cpu). The memory
overcommit controller can balance the cost between the cpu usage and
the memory reclaimed.
3) Provides a way to the applications to keep their LRUs sorted, so,
under memory pressure better reclaim candidates are selected. This
also gives more accurate and uptodate notion of working set for an
application.
Why memory.high is not enough?
------------------------------
- memory.high can be used to trigger reclaim in a memcg and can
potentially be used for proactive reclaim. However there is a big
downside in using memory.high. It can potentially introduce high
reclaim stalls in the target application as the allocations from the
processes or the threads of the application can hit the temporary
memory.high limit.
- Userspace proactive reclaimers usually use feedback loops to decide
how much memory to proactively reclaim from a workload. The metrics
used for this are usually either refaults or PSI, and these metrics will
become messy if the application gets throttled by hitting the high
limit.
- memory.high is a stateful interface, if the userspace proactive
reclaimer crashes for any reason while triggering reclaim it can leave
the application in a bad state.
- If a workload is rapidly expanding, setting memory.high to proactively
reclaim memory can result in actually reclaiming more memory than
intended.
The benefits of such interface and shortcomings of existing interface were
further discussed in this RFC thread:
https://lore.kernel.org/linux-mm/5df21376-7dd1-bf81-8414-32a73cea45dd@google.com/
Interface:
----------
Introducing a very simple memcg interface 'echo 10M > memory.reclaim' to
trigger reclaim in the target memory cgroup.
The interface is introduced as a nested-keyed file to allow for future
optional arguments to be easily added to configure the behavior of
reclaim.
Possible Extensions:
--------------------
- This interface can be extended with an additional parameter or flags
to allow specifying one or more types of memory to reclaim from (e.g.
file, anon, ..).
- The interface can also be extended with a node mask to reclaim from
specific nodes. This has use cases for reclaim-based demotion in memory
tiering systens.
- A similar per-node interface can also be added to support proactive
reclaim and reclaim-based demotion in systems without memcg.
- Add a timeout parameter to make it easier for user space to call the
interface without worrying about being blocked for an undefined amount
of time.
For now, let's keep things simple by adding the basic functionality.
[yosryahmed@google.com: worked on versions v2 onwards, refreshed to
current master, updated commit message based on recent
discussions and use cases]
Link: https://lkml.kernel.org/r/20220425190040.2475377-1-yosryahmed@google.com
Link: https://lkml.kernel.org/r/20220425190040.2475377-2-yosryahmed@google.com
Signed-off-by: Shakeel Butt <shakeelb@google.com>
Co-developed-by: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Wei Xu <weixugc@google.com>
Acked-by: Roman Gushchin <roman.gushchin@linux.dev>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Zefan Li <lizefan.x@bytedance.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Yu Zhao <yuzhao@google.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Chen Wandun <chenwandun@huawei.com>
Cc: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: "Michal Koutn" <mkoutny@suse.com>
Cc: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-04-29 14:36:59 -07:00
|
|
|
|
2025-06-23 11:58:49 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
memcg: introduce per-memcg reclaim interface
This patch series adds a memory.reclaim proactive reclaim interface.
The rationale behind the interface and how it works are in the first
patch.
This patch (of 4):
Introduce a memcg interface to trigger memory reclaim on a memory cgroup.
Use case: Proactive Reclaim
---------------------------
A userspace proactive reclaimer can continuously probe the memcg to
reclaim a small amount of memory. This gives more accurate and up-to-date
workingset estimation as the LRUs are continuously sorted and can
potentially provide more deterministic memory overcommit behavior. The
memory overcommit controller can provide more proactive response to the
changing behavior of the running applications instead of being reactive.
A userspace reclaimer's purpose in this case is not a complete replacement
for kswapd or direct reclaim, it is to proactively identify memory savings
opportunities and reclaim some amount of cold pages set by the policy to
free up the memory for more demanding jobs or scheduling new jobs.
A user space proactive reclaimer is used in Google data centers.
Additionally, Meta's TMO paper recently referenced a very similar
interface used for user space proactive reclaim:
https://dl.acm.org/doi/pdf/10.1145/3503222.3507731
Benefits of a user space reclaimer:
-----------------------------------
1) More flexible on who should be charged for the cpu of the memory
reclaim. For proactive reclaim, it makes more sense to be centralized.
2) More flexible on dedicating the resources (like cpu). The memory
overcommit controller can balance the cost between the cpu usage and
the memory reclaimed.
3) Provides a way to the applications to keep their LRUs sorted, so,
under memory pressure better reclaim candidates are selected. This
also gives more accurate and uptodate notion of working set for an
application.
Why memory.high is not enough?
------------------------------
- memory.high can be used to trigger reclaim in a memcg and can
potentially be used for proactive reclaim. However there is a big
downside in using memory.high. It can potentially introduce high
reclaim stalls in the target application as the allocations from the
processes or the threads of the application can hit the temporary
memory.high limit.
- Userspace proactive reclaimers usually use feedback loops to decide
how much memory to proactively reclaim from a workload. The metrics
used for this are usually either refaults or PSI, and these metrics will
become messy if the application gets throttled by hitting the high
limit.
- memory.high is a stateful interface, if the userspace proactive
reclaimer crashes for any reason while triggering reclaim it can leave
the application in a bad state.
- If a workload is rapidly expanding, setting memory.high to proactively
reclaim memory can result in actually reclaiming more memory than
intended.
The benefits of such interface and shortcomings of existing interface were
further discussed in this RFC thread:
https://lore.kernel.org/linux-mm/5df21376-7dd1-bf81-8414-32a73cea45dd@google.com/
Interface:
----------
Introducing a very simple memcg interface 'echo 10M > memory.reclaim' to
trigger reclaim in the target memory cgroup.
The interface is introduced as a nested-keyed file to allow for future
optional arguments to be easily added to configure the behavior of
reclaim.
Possible Extensions:
--------------------
- This interface can be extended with an additional parameter or flags
to allow specifying one or more types of memory to reclaim from (e.g.
file, anon, ..).
- The interface can also be extended with a node mask to reclaim from
specific nodes. This has use cases for reclaim-based demotion in memory
tiering systens.
- A similar per-node interface can also be added to support proactive
reclaim and reclaim-based demotion in systems without memcg.
- Add a timeout parameter to make it easier for user space to call the
interface without worrying about being blocked for an undefined amount
of time.
For now, let's keep things simple by adding the basic functionality.
[yosryahmed@google.com: worked on versions v2 onwards, refreshed to
current master, updated commit message based on recent
discussions and use cases]
Link: https://lkml.kernel.org/r/20220425190040.2475377-1-yosryahmed@google.com
Link: https://lkml.kernel.org/r/20220425190040.2475377-2-yosryahmed@google.com
Signed-off-by: Shakeel Butt <shakeelb@google.com>
Co-developed-by: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Wei Xu <weixugc@google.com>
Acked-by: Roman Gushchin <roman.gushchin@linux.dev>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Zefan Li <lizefan.x@bytedance.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Yu Zhao <yuzhao@google.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Chen Wandun <chenwandun@huawei.com>
Cc: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: "Michal Koutn" <mkoutny@suse.com>
Cc: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-04-29 14:36:59 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-11-05 18:50:23 -08:00
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
|
|
|
|
|
2022-05-13 16:48:57 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, memcg: cg2 memory{.swap,}.peak write handlers
Patch series "mm, memcg: cg2 memory{.swap,}.peak write handlers", v7.
This patch (of 2):
Other mechanisms for querying the peak memory usage of either a process or
v1 memory cgroup allow for resetting the high watermark. Restore parity
with those mechanisms, but with a less racy API.
For example:
- Any write to memory.max_usage_in_bytes in a cgroup v1 mount resets
the high watermark.
- writing "5" to the clear_refs pseudo-file in a processes's proc
directory resets the peak RSS.
This change is an evolution of a previous patch, which mostly copied the
cgroup v1 behavior, however, there were concerns about races/ownership
issues with a global reset, so instead this change makes the reset
filedescriptor-local.
Writing any non-empty string to the memory.peak and memory.swap.peak
pseudo-files reset the high watermark to the current usage for subsequent
reads through that same FD.
Notably, following Johannes's suggestion, this implementation moves the
O(FDs that have written) behavior onto the FD write(2) path. Instead, on
the page-allocation path, we simply add one additional watermark to
conditionally bump per-hierarchy level in the page-counter.
Additionally, this takes Longman's suggestion of nesting the
page-charging-path checks for the two watermarks to reduce the number of
common-case comparisons.
This behavior is particularly useful for work scheduling systems that need
to track memory usage of worker processes/cgroups per-work-item. Since
memory can't be squeezed like CPU can (the OOM-killer has opinions), these
systems need to track the peak memory usage to compute system/container
fullness when binpacking workitems.
Most notably, Vimeo's use-case involves a system that's doing global
binpacking across many Kubernetes pods/containers, and while we can use
PSI for some local decisions about overload, we strive to avoid packing
workloads too tightly in the first place. To facilitate this, we track
the peak memory usage. However, since we run with long-lived workers (to
amortize startup costs) we need a way to track the high watermark while a
work-item is executing. Polling runs the risk of missing short spikes
that last for timescales below the polling interval, and peak memory
tracking at the cgroup level is otherwise perfect for this use-case.
As this data is used to ensure that binpacked work ends up with sufficient
headroom, this use-case mostly avoids the inaccuracies surrounding
reclaimable memory.
Link: https://lkml.kernel.org/r/20240730231304.761942-1-davidf@vimeo.com
Link: https://lkml.kernel.org/r/20240729143743.34236-1-davidf@vimeo.com
Link: https://lkml.kernel.org/r/20240729143743.34236-2-davidf@vimeo.com
Signed-off-by: David Finkel <davidf@vimeo.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Waiman Long <longman@redhat.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-07-29 10:37:42 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-05-13 16:48:57 -07:00
|
|
|
|
2018-06-07 17:07:46 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-09-18 18:01:59 -04:00
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
|
|
|
|
|
2019-07-11 20:55:55 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-01-20 15:03:19 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-10-13 16:52:59 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-08-21 21:53:54 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
memcg: introduce per-memcg reclaim interface
This patch series adds a memory.reclaim proactive reclaim interface.
The rationale behind the interface and how it works are in the first
patch.
This patch (of 4):
Introduce a memcg interface to trigger memory reclaim on a memory cgroup.
Use case: Proactive Reclaim
---------------------------
A userspace proactive reclaimer can continuously probe the memcg to
reclaim a small amount of memory. This gives more accurate and up-to-date
workingset estimation as the LRUs are continuously sorted and can
potentially provide more deterministic memory overcommit behavior. The
memory overcommit controller can provide more proactive response to the
changing behavior of the running applications instead of being reactive.
A userspace reclaimer's purpose in this case is not a complete replacement
for kswapd or direct reclaim, it is to proactively identify memory savings
opportunities and reclaim some amount of cold pages set by the policy to
free up the memory for more demanding jobs or scheduling new jobs.
A user space proactive reclaimer is used in Google data centers.
Additionally, Meta's TMO paper recently referenced a very similar
interface used for user space proactive reclaim:
https://dl.acm.org/doi/pdf/10.1145/3503222.3507731
Benefits of a user space reclaimer:
-----------------------------------
1) More flexible on who should be charged for the cpu of the memory
reclaim. For proactive reclaim, it makes more sense to be centralized.
2) More flexible on dedicating the resources (like cpu). The memory
overcommit controller can balance the cost between the cpu usage and
the memory reclaimed.
3) Provides a way to the applications to keep their LRUs sorted, so,
under memory pressure better reclaim candidates are selected. This
also gives more accurate and uptodate notion of working set for an
application.
Why memory.high is not enough?
------------------------------
- memory.high can be used to trigger reclaim in a memcg and can
potentially be used for proactive reclaim. However there is a big
downside in using memory.high. It can potentially introduce high
reclaim stalls in the target application as the allocations from the
processes or the threads of the application can hit the temporary
memory.high limit.
- Userspace proactive reclaimers usually use feedback loops to decide
how much memory to proactively reclaim from a workload. The metrics
used for this are usually either refaults or PSI, and these metrics will
become messy if the application gets throttled by hitting the high
limit.
- memory.high is a stateful interface, if the userspace proactive
reclaimer crashes for any reason while triggering reclaim it can leave
the application in a bad state.
- If a workload is rapidly expanding, setting memory.high to proactively
reclaim memory can result in actually reclaiming more memory than
intended.
The benefits of such interface and shortcomings of existing interface were
further discussed in this RFC thread:
https://lore.kernel.org/linux-mm/5df21376-7dd1-bf81-8414-32a73cea45dd@google.com/
Interface:
----------
Introducing a very simple memcg interface 'echo 10M > memory.reclaim' to
trigger reclaim in the target memory cgroup.
The interface is introduced as a nested-keyed file to allow for future
optional arguments to be easily added to configure the behavior of
reclaim.
Possible Extensions:
--------------------
- This interface can be extended with an additional parameter or flags
to allow specifying one or more types of memory to reclaim from (e.g.
file, anon, ..).
- The interface can also be extended with a node mask to reclaim from
specific nodes. This has use cases for reclaim-based demotion in memory
tiering systens.
- A similar per-node interface can also be added to support proactive
reclaim and reclaim-based demotion in systems without memcg.
- Add a timeout parameter to make it easier for user space to call the
interface without worrying about being blocked for an undefined amount
of time.
For now, let's keep things simple by adding the basic functionality.
[yosryahmed@google.com: worked on versions v2 onwards, refreshed to
current master, updated commit message based on recent
discussions and use cases]
Link: https://lkml.kernel.org/r/20220425190040.2475377-1-yosryahmed@google.com
Link: https://lkml.kernel.org/r/20220425190040.2475377-2-yosryahmed@google.com
Signed-off-by: Shakeel Butt <shakeelb@google.com>
Co-developed-by: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Wei Xu <weixugc@google.com>
Acked-by: Roman Gushchin <roman.gushchin@linux.dev>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Zefan Li <lizefan.x@bytedance.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Yu Zhao <yuzhao@google.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Chen Wandun <chenwandun@huawei.com>
Cc: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: "Michal Koutn" <mkoutny@suse.com>
Cc: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-04-29 14:36:59 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-02-08 10:36:58 -05:00
|
|
|
|
2012-11-19 08:13:38 -08:00
|
|
|
|
2013-02-22 16:34:52 -08:00
|
|
|
|
2012-11-19 08:13:38 -08:00
|
|
|
|
2015-12-29 14:54:10 -08:00
|
|
|
|
2012-11-19 08:13:38 -08:00
|
|
|
|
2014-07-08 18:02:57 -04:00
|
|
|
|
mm: memcontrol: switch to rstat
Replace the memory controller's custom hierarchical stats code with the
generic rstat infrastructure provided by the cgroup core.
The current implementation does batched upward propagation from the
write side (i.e. as stats change). The per-cpu batches introduce an
error, which is multiplied by the number of subgroups in a tree. In
systems with many CPUs and sizable cgroup trees, the error can be large
enough to confuse users (e.g. 32 batch pages * 32 CPUs * 32 subgroups
results in an error of up to 128M per stat item). This can entirely
swallow allocation bursts inside a workload that the user is expecting
to see reflected in the statistics.
In the past, we've done read-side aggregation, where a memory.stat read
would have to walk the entire subtree and add up per-cpu counts. This
became problematic with lazily-freed cgroups: we could have large
subtrees where most cgroups were entirely idle. Hence the switch to
change-driven upward propagation. Unfortunately, it needed to trade
accuracy for speed due to the write side being so hot.
Rstat combines the best of both worlds: from the write side, it cheaply
maintains a queue of cgroups that have pending changes, so that the read
side can do selective tree aggregation. This way the reported stats
will always be precise and recent as can be, while the aggregation can
skip over potentially large numbers of idle cgroups.
The way rstat works is that it implements a tree for tracking cgroups
with pending local changes, as well as a flush function that walks the
tree upwards. The controller then drives this by 1) telling rstat when
a local cgroup stat changes (e.g. mod_memcg_state) and 2) when a flush
is required to get uptodate hierarchy stats for a given subtree (e.g.
when memory.stat is read). The controller also provides a flush
callback that is called during the rstat flush walk for each cgroup and
aggregates its local per-cpu counters and propagates them upwards.
This adds a second vmstats to struct mem_cgroup (MEMCG_NR_STAT +
NR_VM_EVENT_ITEMS) to track pending subtree deltas during upward
aggregation. It removes 3 words from the per-cpu data. It eliminates
memcg_exact_page_state(), since memcg_page_state() is now exact.
[akpm@linux-foundation.org: merge fix]
[hannes@cmpxchg.org: fix a sleep in atomic section problem]
Link: https://lkml.kernel.org/r/20210315234100.64307-1-hannes@cmpxchg.org
Link: https://lkml.kernel.org/r/20210209163304.77088-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Roman Gushchin <guro@fb.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Balbir Singh <bsingharora@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:26 -07:00
|
|
|
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 02:00:05 -06:00
|
|
|
|
2023-10-19 15:53:42 -07:00
|
|
|
|
|
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
2024-06-24 17:59:05 -07:00
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
2024-06-24 17:59:05 -07:00
|
|
|
|
2008-02-07 00:14:31 -08:00
|
|
|
|
2008-02-07 00:13:50 -08:00
|
|
|
|
2009-01-07 18:07:57 -08:00
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
2021-06-30 18:53:32 -07:00
|
|
|
|
mm/memcontrol: exclude @root from checks in mem_cgroup_low
Make @root exclusive in mem_cgroup_low; it is never considered low when
looked at directly and is not checked when traversing the tree. In
effect, @root is handled identically to how root_mem_cgroup was
previously handled by mem_cgroup_low.
If @root is not excluded from the checks, a cgroup underneath @root will
never be considered low during targeted reclaim of @root, e.g. due to
memory.current > memory.high, unless @root is misconfigured to have
memory.low > memory.high.
Excluding @root enables using memory.low to prioritize memory usage
between cgroups within a subtree of the hierarchy that is limited by
memory.high or memory.max, e.g. when ROOT owns @root's controls but
delegates the @root directory to a USER so that USER can create and
administer children of @root.
For example, given cgroup A with children B and C:
A
/ \
B C
and
1. A/memory.current > A/memory.high
2. A/B/memory.current < A/B/memory.low
3. A/C/memory.current >= A/C/memory.low
As 'A' is high, i.e. triggers reclaim from 'A', and 'B' is low, we
should reclaim from 'C' until 'A' is no longer high or until we can no
longer reclaim from 'C'. If 'A', i.e. @root, isn't excluded by
mem_cgroup_low when reclaming from 'A', then 'B' won't be considered low
and we will reclaim indiscriminately from both 'B' and 'C'.
Here is the test I used to confirm the bug and the patch.
20:00:55@sjchrist-vm ? ~ $ cat ~/.bin/memcg_low_test
#!/bin/bash
x62mb=$((62<<20))
x66mb=$((66<<20))
x94mb=$((94<<20))
x98mb=$((98<<20))
setup() {
set -e
if [[ -n $DEBUG ]]; then
set -x
fi
trap teardown EXIT HUP INT TERM
if [[ ! -e /mnt/1gb.swap ]]; then
sudo fallocate -l 1G /mnt/1gb.swap > /dev/null
sudo mkswap /mnt/1gb.swap > /dev/null
fi
if ! swapon --show=NAME | grep -q "/mnt/1gb.swap"; then
sudo swapon /mnt/1gb.swap
fi
if [[ ! -e /cgroup/cgroup.controllers ]]; then
sudo mount -t cgroup2 none /cgroup
fi
grep -q memory /cgroup/cgroup.controllers
sudo sh -c "echo '+memory' > /cgroup/cgroup.subtree_control"
sudo mkdir /cgroup/A && sudo chown $USER:$USER /cgroup/A
sudo sh -c "echo '+memory' > /cgroup/A/cgroup.subtree_control"
sudo sh -c "echo '96m' > /cgroup/A/memory.high"
mkdir /cgroup/A/0
mkdir /cgroup/A/1
echo 64m > /cgroup/A/0/memory.low
}
teardown() {
set +e
trap - EXIT HUP INT TERM
if [[ -z $1 ]]; then
printf "\n"
printf "%0.s*" {1..35}
printf "\nFAILED!\n\n"
tail /cgroup/A/**/memory.current
printf "%0.s*" {1..35}
printf "\n\n"
fi
ps | grep stress | tr -s ' ' | cut -f 2 -d ' ' | xargs -I % kill %
sleep 2
if [[ -e /cgroup/A/0 ]]; then
rmdir /cgroup/A/0
fi
if [[ -e /cgroup/A/1 ]]; then
rmdir /cgroup/A/1
fi
if [[ -e /cgroup/A ]]; then
sudo rmdir /cgroup/A
fi
}
stress_test() {
sudo sh -c "echo $$ > /cgroup/A/$1/cgroup.procs"
stress --vm 1 --vm-bytes 64M --vm-keep > /dev/null &
sudo sh -c "echo $$ > /cgroup/A/$2/cgroup.procs"
stress --vm 1 --vm-bytes 64M --vm-keep > /dev/null &
sudo sh -c "echo $$ > /cgroup/cgroup.procs"
sleep 1
# A/0 should be consuming more memory than A/1
[[ $(cat /cgroup/A/0/memory.current) -ge $(cat /cgroup/A/1/memory.current) ]]
# A/0 should be consuming ~64mb
[[ $(cat /cgroup/A/0/memory.current) -ge $x62mb ]] && [[ $(cat /cgroup/A/0/memory.current) -le $x66mb ]]
# A should cumulatively be consuming ~96mb
[[ $(cat /cgroup/A/memory.current) -ge $x94mb ]] && [[ $(cat /cgroup/A/memory.current) -le $x98mb ]]
# Stop the stressors
ps | grep stress | tr -s ' ' | cut -f 2 -d ' ' | xargs -I % kill %
}
teardown 1
setup
for ((i=1;i<=$1;i++)); do
printf "ITERATION $i of $1 - stress_test 0 1"
stress_test 0 1
printf "\x1b[2K\r"
printf "ITERATION $i of $1 - stress_test 1 0"
stress_test 1 0
printf "\x1b[2K\r"
printf "ITERATION $i of $1 - PASSED\n"
done
teardown 1
echo PASSED!
20:11:26@sjchrist-vm ? ~ $ memcg_low_test 10
Link: http://lkml.kernel.org/r/1496434412-21005-1-git-send-email-sean.j.christopherson@intel.com
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Acked-by: Vladimir Davydov <vdavydov.dev@gmail.com>
Acked-by: Balbir Singh <bsingharora@gmail.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-10 15:48:05 -07:00
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
|
|
|
|
|
mm: memory.low hierarchical behavior
This patch aims to address an issue in current memory.low semantics,
which makes it hard to use it in a hierarchy, where some leaf memory
cgroups are more valuable than others.
For example, there are memcgs A, A/B, A/C, A/D and A/E:
A A/memory.low = 2G, A/memory.current = 6G
//\\
BC DE B/memory.low = 3G B/memory.current = 2G
C/memory.low = 1G C/memory.current = 2G
D/memory.low = 0 D/memory.current = 2G
E/memory.low = 10G E/memory.current = 0
If we apply memory pressure, B, C and D are reclaimed at the same pace
while A's usage exceeds 2G. This is obviously wrong, as B's usage is
fully below B's memory.low, and C has 1G of protection as well. Also, A
is pushed to the size, which is less than A's 2G memory.low, which is
also wrong.
A simple bash script (provided below) can be used to reproduce
the problem. Current results are:
A: 1430097920
A/B: 711929856
A/C: 717426688
A/D: 741376
A/E: 0
To address the issue a concept of effective memory.low is introduced.
Effective memory.low is always equal or less than original memory.low.
In a case, when there is no memory.low overcommittment (and also for
top-level cgroups), these two values are equal.
Otherwise it's a part of parent's effective memory.low, calculated as a
cgroup's memory.low usage divided by sum of sibling's memory.low usages
(under memory.low usage I mean the size of actually protected memory:
memory.current if memory.current < memory.low, 0 otherwise). It's
necessary to track the actual usage, because otherwise an empty cgroup
with memory.low set (A/E in my example) will affect actual memory
distribution, which makes no sense. To avoid traversing the cgroup tree
twice, page_counters code is reused.
Calculating effective memory.low can be done in the reclaim path, as we
conveniently traversing the cgroup tree from top to bottom and check
memory.low on each level. So, it's a perfect place to calculate
effective memory low and save it to use it for children cgroups.
This also eliminates a need to traverse the cgroup tree from bottom to
top each time to check if parent's guarantee is not exceeded.
Setting/resetting effective memory.low is intentionally racy, but it's
fine and shouldn't lead to any significant differences in actual memory
distribution.
With this patch applied results are matching the expectations:
A: 2147930112
A/B: 1428721664
A/C: 718393344
A/D: 815104
A/E: 0
Test script:
#!/bin/bash
CGPATH="/sys/fs/cgroup"
truncate /file1 --size 2G
truncate /file2 --size 2G
truncate /file3 --size 2G
truncate /file4 --size 50G
mkdir "${CGPATH}/A"
echo "+memory" > "${CGPATH}/A/cgroup.subtree_control"
mkdir "${CGPATH}/A/B" "${CGPATH}/A/C" "${CGPATH}/A/D" "${CGPATH}/A/E"
echo 2G > "${CGPATH}/A/memory.low"
echo 3G > "${CGPATH}/A/B/memory.low"
echo 1G > "${CGPATH}/A/C/memory.low"
echo 0 > "${CGPATH}/A/D/memory.low"
echo 10G > "${CGPATH}/A/E/memory.low"
echo $$ > "${CGPATH}/A/B/cgroup.procs" && vmtouch -qt /file1
echo $$ > "${CGPATH}/A/C/cgroup.procs" && vmtouch -qt /file2
echo $$ > "${CGPATH}/A/D/cgroup.procs" && vmtouch -qt /file3
echo $$ > "${CGPATH}/cgroup.procs" && vmtouch -qt /file4
echo "A: " `cat "${CGPATH}/A/memory.current"`
echo "A/B: " `cat "${CGPATH}/A/B/memory.current"`
echo "A/C: " `cat "${CGPATH}/A/C/memory.current"`
echo "A/D: " `cat "${CGPATH}/A/D/memory.current"`
echo "A/E: " `cat "${CGPATH}/A/E/memory.current"`
rmdir "${CGPATH}/A/B" "${CGPATH}/A/C" "${CGPATH}/A/D" "${CGPATH}/A/E"
rmdir "${CGPATH}/A"
rm /file1 /file2 /file3 /file4
Link: http://lkml.kernel.org/r/20180405185921.4942-2-guro@fb.com
Signed-off-by: Roman Gushchin <guro@fb.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-06-07 17:06:22 -07:00
|
|
|
|
|
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
2020-08-06 23:22:05 -07:00
|
|
|
|
|
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
2024-07-03 13:25:10 +02:00
|
|
|
|
|
|
|
|
|
mm: memory.low hierarchical behavior
This patch aims to address an issue in current memory.low semantics,
which makes it hard to use it in a hierarchy, where some leaf memory
cgroups are more valuable than others.
For example, there are memcgs A, A/B, A/C, A/D and A/E:
A A/memory.low = 2G, A/memory.current = 6G
//\\
BC DE B/memory.low = 3G B/memory.current = 2G
C/memory.low = 1G C/memory.current = 2G
D/memory.low = 0 D/memory.current = 2G
E/memory.low = 10G E/memory.current = 0
If we apply memory pressure, B, C and D are reclaimed at the same pace
while A's usage exceeds 2G. This is obviously wrong, as B's usage is
fully below B's memory.low, and C has 1G of protection as well. Also, A
is pushed to the size, which is less than A's 2G memory.low, which is
also wrong.
A simple bash script (provided below) can be used to reproduce
the problem. Current results are:
A: 1430097920
A/B: 711929856
A/C: 717426688
A/D: 741376
A/E: 0
To address the issue a concept of effective memory.low is introduced.
Effective memory.low is always equal or less than original memory.low.
In a case, when there is no memory.low overcommittment (and also for
top-level cgroups), these two values are equal.
Otherwise it's a part of parent's effective memory.low, calculated as a
cgroup's memory.low usage divided by sum of sibling's memory.low usages
(under memory.low usage I mean the size of actually protected memory:
memory.current if memory.current < memory.low, 0 otherwise). It's
necessary to track the actual usage, because otherwise an empty cgroup
with memory.low set (A/E in my example) will affect actual memory
distribution, which makes no sense. To avoid traversing the cgroup tree
twice, page_counters code is reused.
Calculating effective memory.low can be done in the reclaim path, as we
conveniently traversing the cgroup tree from top to bottom and check
memory.low on each level. So, it's a perfect place to calculate
effective memory low and save it to use it for children cgroups.
This also eliminates a need to traverse the cgroup tree from bottom to
top each time to check if parent's guarantee is not exceeded.
Setting/resetting effective memory.low is intentionally racy, but it's
fine and shouldn't lead to any significant differences in actual memory
distribution.
With this patch applied results are matching the expectations:
A: 2147930112
A/B: 1428721664
A/C: 718393344
A/D: 815104
A/E: 0
Test script:
#!/bin/bash
CGPATH="/sys/fs/cgroup"
truncate /file1 --size 2G
truncate /file2 --size 2G
truncate /file3 --size 2G
truncate /file4 --size 50G
mkdir "${CGPATH}/A"
echo "+memory" > "${CGPATH}/A/cgroup.subtree_control"
mkdir "${CGPATH}/A/B" "${CGPATH}/A/C" "${CGPATH}/A/D" "${CGPATH}/A/E"
echo 2G > "${CGPATH}/A/memory.low"
echo 3G > "${CGPATH}/A/B/memory.low"
echo 1G > "${CGPATH}/A/C/memory.low"
echo 0 > "${CGPATH}/A/D/memory.low"
echo 10G > "${CGPATH}/A/E/memory.low"
echo $$ > "${CGPATH}/A/B/cgroup.procs" && vmtouch -qt /file1
echo $$ > "${CGPATH}/A/C/cgroup.procs" && vmtouch -qt /file2
echo $$ > "${CGPATH}/A/D/cgroup.procs" && vmtouch -qt /file3
echo $$ > "${CGPATH}/cgroup.procs" && vmtouch -qt /file4
echo "A: " `cat "${CGPATH}/A/memory.current"`
echo "A/B: " `cat "${CGPATH}/A/B/memory.current"`
echo "A/C: " `cat "${CGPATH}/A/C/memory.current"`
echo "A/D: " `cat "${CGPATH}/A/D/memory.current"`
echo "A/E: " `cat "${CGPATH}/A/E/memory.current"`
rmdir "${CGPATH}/A/B" "${CGPATH}/A/C" "${CGPATH}/A/D" "${CGPATH}/A/E"
rmdir "${CGPATH}/A"
rm /file1 /file2 /file3 /file4
Link: http://lkml.kernel.org/r/20180405185921.4942-2-guro@fb.com
Signed-off-by: Roman Gushchin <guro@fb.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-06-07 17:06:22 -07:00
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
2020-08-06 23:22:05 -07:00
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
mm/memcontrol: exclude @root from checks in mem_cgroup_low
Make @root exclusive in mem_cgroup_low; it is never considered low when
looked at directly and is not checked when traversing the tree. In
effect, @root is handled identically to how root_mem_cgroup was
previously handled by mem_cgroup_low.
If @root is not excluded from the checks, a cgroup underneath @root will
never be considered low during targeted reclaim of @root, e.g. due to
memory.current > memory.high, unless @root is misconfigured to have
memory.low > memory.high.
Excluding @root enables using memory.low to prioritize memory usage
between cgroups within a subtree of the hierarchy that is limited by
memory.high or memory.max, e.g. when ROOT owns @root's controls but
delegates the @root directory to a USER so that USER can create and
administer children of @root.
For example, given cgroup A with children B and C:
A
/ \
B C
and
1. A/memory.current > A/memory.high
2. A/B/memory.current < A/B/memory.low
3. A/C/memory.current >= A/C/memory.low
As 'A' is high, i.e. triggers reclaim from 'A', and 'B' is low, we
should reclaim from 'C' until 'A' is no longer high or until we can no
longer reclaim from 'C'. If 'A', i.e. @root, isn't excluded by
mem_cgroup_low when reclaming from 'A', then 'B' won't be considered low
and we will reclaim indiscriminately from both 'B' and 'C'.
Here is the test I used to confirm the bug and the patch.
20:00:55@sjchrist-vm ? ~ $ cat ~/.bin/memcg_low_test
#!/bin/bash
x62mb=$((62<<20))
x66mb=$((66<<20))
x94mb=$((94<<20))
x98mb=$((98<<20))
setup() {
set -e
if [[ -n $DEBUG ]]; then
set -x
fi
trap teardown EXIT HUP INT TERM
if [[ ! -e /mnt/1gb.swap ]]; then
sudo fallocate -l 1G /mnt/1gb.swap > /dev/null
sudo mkswap /mnt/1gb.swap > /dev/null
fi
if ! swapon --show=NAME | grep -q "/mnt/1gb.swap"; then
sudo swapon /mnt/1gb.swap
fi
if [[ ! -e /cgroup/cgroup.controllers ]]; then
sudo mount -t cgroup2 none /cgroup
fi
grep -q memory /cgroup/cgroup.controllers
sudo sh -c "echo '+memory' > /cgroup/cgroup.subtree_control"
sudo mkdir /cgroup/A && sudo chown $USER:$USER /cgroup/A
sudo sh -c "echo '+memory' > /cgroup/A/cgroup.subtree_control"
sudo sh -c "echo '96m' > /cgroup/A/memory.high"
mkdir /cgroup/A/0
mkdir /cgroup/A/1
echo 64m > /cgroup/A/0/memory.low
}
teardown() {
set +e
trap - EXIT HUP INT TERM
if [[ -z $1 ]]; then
printf "\n"
printf "%0.s*" {1..35}
printf "\nFAILED!\n\n"
tail /cgroup/A/**/memory.current
printf "%0.s*" {1..35}
printf "\n\n"
fi
ps | grep stress | tr -s ' ' | cut -f 2 -d ' ' | xargs -I % kill %
sleep 2
if [[ -e /cgroup/A/0 ]]; then
rmdir /cgroup/A/0
fi
if [[ -e /cgroup/A/1 ]]; then
rmdir /cgroup/A/1
fi
if [[ -e /cgroup/A ]]; then
sudo rmdir /cgroup/A
fi
}
stress_test() {
sudo sh -c "echo $$ > /cgroup/A/$1/cgroup.procs"
stress --vm 1 --vm-bytes 64M --vm-keep > /dev/null &
sudo sh -c "echo $$ > /cgroup/A/$2/cgroup.procs"
stress --vm 1 --vm-bytes 64M --vm-keep > /dev/null &
sudo sh -c "echo $$ > /cgroup/cgroup.procs"
sleep 1
# A/0 should be consuming more memory than A/1
[[ $(cat /cgroup/A/0/memory.current) -ge $(cat /cgroup/A/1/memory.current) ]]
# A/0 should be consuming ~64mb
[[ $(cat /cgroup/A/0/memory.current) -ge $x62mb ]] && [[ $(cat /cgroup/A/0/memory.current) -le $x66mb ]]
# A should cumulatively be consuming ~96mb
[[ $(cat /cgroup/A/memory.current) -ge $x94mb ]] && [[ $(cat /cgroup/A/memory.current) -le $x98mb ]]
# Stop the stressors
ps | grep stress | tr -s ' ' | cut -f 2 -d ' ' | xargs -I % kill %
}
teardown 1
setup
for ((i=1;i<=$1;i++)); do
printf "ITERATION $i of $1 - stress_test 0 1"
stress_test 0 1
printf "\x1b[2K\r"
printf "ITERATION $i of $1 - stress_test 1 0"
stress_test 1 0
printf "\x1b[2K\r"
printf "ITERATION $i of $1 - PASSED\n"
done
teardown 1
echo PASSED!
20:11:26@sjchrist-vm ? ~ $ memcg_low_test 10
Link: http://lkml.kernel.org/r/1496434412-21005-1-git-send-email-sean.j.christopherson@intel.com
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Acked-by: Vladimir Davydov <vdavydov.dev@gmail.com>
Acked-by: Balbir Singh <bsingharora@gmail.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-10 15:48:05 -07:00
|
|
|
|
|
|
|
|
|
mm, memcg: avoid stale protection values when cgroup is above protection
Patch series "mm, memcg: memory.{low,min} reclaim fix & cleanup", v4.
This series contains a fix for a edge case in my earlier protection
calculation patches, and a patch to make the area overall a little more
robust to hopefully help avoid this in future.
This patch (of 2):
A cgroup can have both memory protection and a memory limit to isolate it
from its siblings in both directions - for example, to prevent it from
being shrunk below 2G under high pressure from outside, but also from
growing beyond 4G under low pressure.
Commit 9783aa9917f8 ("mm, memcg: proportional memory.{low,min} reclaim")
implemented proportional scan pressure so that multiple siblings in excess
of their protection settings don't get reclaimed equally but instead in
accordance to their unprotected portion.
During limit reclaim, this proportionality shouldn't apply of course:
there is no competition, all pressure is from within the cgroup and should
be applied as such. Reclaim should operate at full efficiency.
However, mem_cgroup_protected() never expected anybody to look at the
effective protection values when it indicated that the cgroup is above its
protection. As a result, a query during limit reclaim may return stale
protection values that were calculated by a previous reclaim cycle in
which the cgroup did have siblings.
When this happens, reclaim is unnecessarily hesitant and potentially slow
to meet the desired limit. In theory this could lead to premature OOM
kills, although it's not obvious this has occurred in practice.
Workaround the problem by special casing reclaim roots in
mem_cgroup_protection. These memcgs are never participating in the
reclaim protection because the reclaim is internal.
We have to ignore effective protection values for reclaim roots because
mem_cgroup_protected might be called from racing reclaim contexts with
different roots. Calculation is relying on root -> leaf tree traversal
therefore top-down reclaim protection invariants should hold. The only
exception is the reclaim root which should have effective protection set
to 0 but that would be problematic for the following setup:
Let's have global and A's reclaim in parallel:
|
A (low=2G, usage = 3G, max = 3G, children_low_usage = 1.5G)
|\
| C (low = 1G, usage = 2.5G)
B (low = 1G, usage = 0.5G)
for A reclaim we have
B.elow = B.low
C.elow = C.low
For the global reclaim
A.elow = A.low
B.elow = min(B.usage, B.low) because children_low_usage <= A.elow
C.elow = min(C.usage, C.low)
With the effective values resetting we have A reclaim
A.elow = 0
B.elow = B.low
C.elow = C.low
and global reclaim could see the above and then
B.elow = C.elow = 0 because children_low_usage > A.elow
Which means that protected memcgs would get reclaimed.
In future we would like to make mem_cgroup_protected more robust against
racing reclaim contexts but that is likely more complex solution than this
simple workaround.
[hannes@cmpxchg.org - large part of the changelog]
[mhocko@suse.com - workaround explanation]
[chris@chrisdown.name - retitle]
Fixes: 9783aa9917f8 ("mm, memcg: proportional memory.{low,min} reclaim")
Signed-off-by: Yafang Shao <laoar.shao@gmail.com>
Signed-off-by: Chris Down <chris@chrisdown.name>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Chris Down <chris@chrisdown.name>
Acked-by: Roman Gushchin <guro@fb.com>
Link: http://lkml.kernel.org/r/cover.1594638158.git.chris@chrisdown.name
Link: http://lkml.kernel.org/r/044fb8ecffd001c7905d27c0c2ad998069fdc396.1594638158.git.chris@chrisdown.name
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-06 23:22:01 -07:00
|
|
|
|
2024-07-03 13:25:10 +02:00
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 15:26:06 -08:00
|
|
|
|
|
|
|
|
|
2021-06-25 09:27:04 -04:00
|
|
|
|
|
|
|
|
|
2021-04-29 22:56:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
memcontrol: add helpers for hugetlb memcg accounting
Patch series "hugetlb memcg accounting", v4.
Currently, hugetlb memory usage is not acounted for in the memory
controller, which could lead to memory overprotection for cgroups with
hugetlb-backed memory. This has been observed in our production system.
For instance, here is one of our usecases: suppose there are two 32G
containers. The machine is booted with hugetlb_cma=6G, and each container
may or may not use up to 3 gigantic page, depending on the workload within
it. The rest is anon, cache, slab, etc. We can set the hugetlb cgroup
limit of each cgroup to 3G to enforce hugetlb fairness. But it is very
difficult to configure memory.max to keep overall consumption, including
anon, cache, slab etcetera fair.
What we have had to resort to is to constantly poll hugetlb usage and
readjust memory.max. Similar procedure is done to other memory limits
(memory.low for e.g). However, this is rather cumbersome and buggy.
Furthermore, when there is a delay in memory limits correction, (for e.g
when hugetlb usage changes within consecutive runs of the userspace
agent), the system could be in an over/underprotected state.
This patch series rectifies this issue by charging the memcg when the
hugetlb folio is allocated, and uncharging when the folio is freed. In
addition, a new selftest is added to demonstrate and verify this new
behavior.
This patch (of 4):
This patch exposes charge committing and cancelling as parts of the memory
controller interface. These functionalities are useful when the
try_charge() and commit_charge() stages have to be separated by other
actions in between (which can fail). One such example is the new hugetlb
accounting behavior in the following patch.
The patch also adds a helper function to obtain a reference to the
current task's memcg.
Link: https://lkml.kernel.org/r/20231006184629.155543-1-nphamcs@gmail.com
Link: https://lkml.kernel.org/r/20231006184629.155543-2-nphamcs@gmail.com
Signed-off-by: Nhat Pham <nphamcs@gmail.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Frank van der Linden <fvdl@google.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Rik van Riel <riel@surriel.com>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Tejun heo <tj@kernel.org>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-06 11:46:26 -07:00
|
|
|
|
2021-04-29 22:56:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-12-11 12:39:51 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-04-29 22:56:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-25 09:27:04 -04:00
|
|
|
|
mm: memcontrol: rewrite charge API
These patches rework memcg charge lifetime to integrate more naturally
with the lifetime of user pages. This drastically simplifies the code and
reduces charging and uncharging overhead. The most expensive part of
charging and uncharging is the page_cgroup bit spinlock, which is removed
entirely after this series.
Here are the top-10 profile entries of a stress test that reads a 128G
sparse file on a freshly booted box, without even a dedicated cgroup (i.e.
executing in the root memcg). Before:
15.36% cat [kernel.kallsyms] [k] copy_user_generic_string
13.31% cat [kernel.kallsyms] [k] memset
11.48% cat [kernel.kallsyms] [k] do_mpage_readpage
4.23% cat [kernel.kallsyms] [k] get_page_from_freelist
2.38% cat [kernel.kallsyms] [k] put_page
2.32% cat [kernel.kallsyms] [k] __mem_cgroup_commit_charge
2.18% kswapd0 [kernel.kallsyms] [k] __mem_cgroup_uncharge_common
1.92% kswapd0 [kernel.kallsyms] [k] shrink_page_list
1.86% cat [kernel.kallsyms] [k] __radix_tree_lookup
1.62% cat [kernel.kallsyms] [k] __pagevec_lru_add_fn
After:
15.67% cat [kernel.kallsyms] [k] copy_user_generic_string
13.48% cat [kernel.kallsyms] [k] memset
11.42% cat [kernel.kallsyms] [k] do_mpage_readpage
3.98% cat [kernel.kallsyms] [k] get_page_from_freelist
2.46% cat [kernel.kallsyms] [k] put_page
2.13% kswapd0 [kernel.kallsyms] [k] shrink_page_list
1.88% cat [kernel.kallsyms] [k] __radix_tree_lookup
1.67% cat [kernel.kallsyms] [k] __pagevec_lru_add_fn
1.39% kswapd0 [kernel.kallsyms] [k] free_pcppages_bulk
1.30% cat [kernel.kallsyms] [k] kfree
As you can see, the memcg footprint has shrunk quite a bit.
text data bss dec hex filename
37970 9892 400 48262 bc86 mm/memcontrol.o.old
35239 9892 400 45531 b1db mm/memcontrol.o
This patch (of 4):
The memcg charge API charges pages before they are rmapped - i.e. have an
actual "type" - and so every callsite needs its own set of charge and
uncharge functions to know what type is being operated on. Worse,
uncharge has to happen from a context that is still type-specific, rather
than at the end of the page's lifetime with exclusive access, and so
requires a lot of synchronization.
Rewrite the charge API to provide a generic set of try_charge(),
commit_charge() and cancel_charge() transaction operations, much like
what's currently done for swap-in:
mem_cgroup_try_charge() attempts to reserve a charge, reclaiming
pages from the memcg if necessary.
mem_cgroup_commit_charge() commits the page to the charge once it
has a valid page->mapping and PageAnon() reliably tells the type.
mem_cgroup_cancel_charge() aborts the transaction.
This reduces the charge API and enables subsequent patches to
drastically simplify uncharging.
As pages need to be committed after rmap is established but before they
are added to the LRU, page_add_new_anon_rmap() must stop doing LRU
additions again. Revive lru_cache_add_active_or_unevictable().
[hughd@google.com: fix shmem_unuse]
[hughd@google.com: Add comments on the private use of -EAGAIN]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-08-08 14:19:20 -07:00
|
|
|
|
2021-04-29 22:56:36 -07:00
|
|
|
|
|
|
|
|
|
mm: memcontrol: rewrite charge API
These patches rework memcg charge lifetime to integrate more naturally
with the lifetime of user pages. This drastically simplifies the code and
reduces charging and uncharging overhead. The most expensive part of
charging and uncharging is the page_cgroup bit spinlock, which is removed
entirely after this series.
Here are the top-10 profile entries of a stress test that reads a 128G
sparse file on a freshly booted box, without even a dedicated cgroup (i.e.
executing in the root memcg). Before:
15.36% cat [kernel.kallsyms] [k] copy_user_generic_string
13.31% cat [kernel.kallsyms] [k] memset
11.48% cat [kernel.kallsyms] [k] do_mpage_readpage
4.23% cat [kernel.kallsyms] [k] get_page_from_freelist
2.38% cat [kernel.kallsyms] [k] put_page
2.32% cat [kernel.kallsyms] [k] __mem_cgroup_commit_charge
2.18% kswapd0 [kernel.kallsyms] [k] __mem_cgroup_uncharge_common
1.92% kswapd0 [kernel.kallsyms] [k] shrink_page_list
1.86% cat [kernel.kallsyms] [k] __radix_tree_lookup
1.62% cat [kernel.kallsyms] [k] __pagevec_lru_add_fn
After:
15.67% cat [kernel.kallsyms] [k] copy_user_generic_string
13.48% cat [kernel.kallsyms] [k] memset
11.42% cat [kernel.kallsyms] [k] do_mpage_readpage
3.98% cat [kernel.kallsyms] [k] get_page_from_freelist
2.46% cat [kernel.kallsyms] [k] put_page
2.13% kswapd0 [kernel.kallsyms] [k] shrink_page_list
1.88% cat [kernel.kallsyms] [k] __radix_tree_lookup
1.67% cat [kernel.kallsyms] [k] __pagevec_lru_add_fn
1.39% kswapd0 [kernel.kallsyms] [k] free_pcppages_bulk
1.30% cat [kernel.kallsyms] [k] kfree
As you can see, the memcg footprint has shrunk quite a bit.
text data bss dec hex filename
37970 9892 400 48262 bc86 mm/memcontrol.o.old
35239 9892 400 45531 b1db mm/memcontrol.o
This patch (of 4):
The memcg charge API charges pages before they are rmapped - i.e. have an
actual "type" - and so every callsite needs its own set of charge and
uncharge functions to know what type is being operated on. Worse,
uncharge has to happen from a context that is still type-specific, rather
than at the end of the page's lifetime with exclusive access, and so
requires a lot of synchronization.
Rewrite the charge API to provide a generic set of try_charge(),
commit_charge() and cancel_charge() transaction operations, much like
what's currently done for swap-in:
mem_cgroup_try_charge() attempts to reserve a charge, reclaiming
pages from the memcg if necessary.
mem_cgroup_commit_charge() commits the page to the charge once it
has a valid page->mapping and PageAnon() reliably tells the type.
mem_cgroup_cancel_charge() aborts the transaction.
This reduces the charge API and enables subsequent patches to
drastically simplify uncharging.
As pages need to be committed after rmap is established but before they
are added to the LRU, page_add_new_anon_rmap() must stop doing LRU
additions again. Revive lru_cache_add_active_or_unevictable().
[hughd@google.com: fix shmem_unuse]
[hughd@google.com: Add comments on the private use of -EAGAIN]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-08-08 14:19:20 -07:00
|
|
|
|
2021-04-29 22:56:36 -07:00
|
|
|
|
2021-06-25 09:27:04 -04:00
|
|
|
|
2021-04-29 22:56:36 -07:00
|
|
|
|
2020-06-03 16:02:14 -07:00
|
|
|
|
2021-04-29 22:56:36 -07:00
|
|
|
|
|
|
|
|
|
2015-09-09 15:35:35 -07:00
|
|
|
|
2024-12-11 12:39:50 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-04-29 22:56:36 -07:00
|
|
|
|
2022-09-02 20:46:12 +01:00
|
|
|
|
|
|
|
|
|
2021-04-29 22:56:36 -07:00
|
|
|
|
|
|
|
|
|
2022-09-02 20:46:12 +01:00
|
|
|
|
2021-04-29 22:56:36 -07:00
|
|
|
|
2022-09-02 20:46:12 +01:00
|
|
|
|
|
|
|
|
|
2021-04-29 22:56:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-09-02 20:46:12 +01:00
|
|
|
|
2021-04-29 22:56:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcontrol: rewrite charge API
These patches rework memcg charge lifetime to integrate more naturally
with the lifetime of user pages. This drastically simplifies the code and
reduces charging and uncharging overhead. The most expensive part of
charging and uncharging is the page_cgroup bit spinlock, which is removed
entirely after this series.
Here are the top-10 profile entries of a stress test that reads a 128G
sparse file on a freshly booted box, without even a dedicated cgroup (i.e.
executing in the root memcg). Before:
15.36% cat [kernel.kallsyms] [k] copy_user_generic_string
13.31% cat [kernel.kallsyms] [k] memset
11.48% cat [kernel.kallsyms] [k] do_mpage_readpage
4.23% cat [kernel.kallsyms] [k] get_page_from_freelist
2.38% cat [kernel.kallsyms] [k] put_page
2.32% cat [kernel.kallsyms] [k] __mem_cgroup_commit_charge
2.18% kswapd0 [kernel.kallsyms] [k] __mem_cgroup_uncharge_common
1.92% kswapd0 [kernel.kallsyms] [k] shrink_page_list
1.86% cat [kernel.kallsyms] [k] __radix_tree_lookup
1.62% cat [kernel.kallsyms] [k] __pagevec_lru_add_fn
After:
15.67% cat [kernel.kallsyms] [k] copy_user_generic_string
13.48% cat [kernel.kallsyms] [k] memset
11.42% cat [kernel.kallsyms] [k] do_mpage_readpage
3.98% cat [kernel.kallsyms] [k] get_page_from_freelist
2.46% cat [kernel.kallsyms] [k] put_page
2.13% kswapd0 [kernel.kallsyms] [k] shrink_page_list
1.88% cat [kernel.kallsyms] [k] __radix_tree_lookup
1.67% cat [kernel.kallsyms] [k] __pagevec_lru_add_fn
1.39% kswapd0 [kernel.kallsyms] [k] free_pcppages_bulk
1.30% cat [kernel.kallsyms] [k] kfree
As you can see, the memcg footprint has shrunk quite a bit.
text data bss dec hex filename
37970 9892 400 48262 bc86 mm/memcontrol.o.old
35239 9892 400 45531 b1db mm/memcontrol.o
This patch (of 4):
The memcg charge API charges pages before they are rmapped - i.e. have an
actual "type" - and so every callsite needs its own set of charge and
uncharge functions to know what type is being operated on. Worse,
uncharge has to happen from a context that is still type-specific, rather
than at the end of the page's lifetime with exclusive access, and so
requires a lot of synchronization.
Rewrite the charge API to provide a generic set of try_charge(),
commit_charge() and cancel_charge() transaction operations, much like
what's currently done for swap-in:
mem_cgroup_try_charge() attempts to reserve a charge, reclaiming
pages from the memcg if necessary.
mem_cgroup_commit_charge() commits the page to the charge once it
has a valid page->mapping and PageAnon() reliably tells the type.
mem_cgroup_cancel_charge() aborts the transaction.
This reduces the charge API and enables subsequent patches to
drastically simplify uncharging.
As pages need to be committed after rmap is established but before they
are added to the LRU, page_add_new_anon_rmap() must stop doing LRU
additions again. Revive lru_cache_add_active_or_unevictable().
[hughd@google.com: fix shmem_unuse]
[hughd@google.com: Add comments on the private use of -EAGAIN]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-08-08 14:19:20 -07:00
|
|
|
|
2021-04-29 22:56:36 -07:00
|
|
|
|
|
|
|
|
|
mm: memcontrol: rewrite charge API
These patches rework memcg charge lifetime to integrate more naturally
with the lifetime of user pages. This drastically simplifies the code and
reduces charging and uncharging overhead. The most expensive part of
charging and uncharging is the page_cgroup bit spinlock, which is removed
entirely after this series.
Here are the top-10 profile entries of a stress test that reads a 128G
sparse file on a freshly booted box, without even a dedicated cgroup (i.e.
executing in the root memcg). Before:
15.36% cat [kernel.kallsyms] [k] copy_user_generic_string
13.31% cat [kernel.kallsyms] [k] memset
11.48% cat [kernel.kallsyms] [k] do_mpage_readpage
4.23% cat [kernel.kallsyms] [k] get_page_from_freelist
2.38% cat [kernel.kallsyms] [k] put_page
2.32% cat [kernel.kallsyms] [k] __mem_cgroup_commit_charge
2.18% kswapd0 [kernel.kallsyms] [k] __mem_cgroup_uncharge_common
1.92% kswapd0 [kernel.kallsyms] [k] shrink_page_list
1.86% cat [kernel.kallsyms] [k] __radix_tree_lookup
1.62% cat [kernel.kallsyms] [k] __pagevec_lru_add_fn
After:
15.67% cat [kernel.kallsyms] [k] copy_user_generic_string
13.48% cat [kernel.kallsyms] [k] memset
11.42% cat [kernel.kallsyms] [k] do_mpage_readpage
3.98% cat [kernel.kallsyms] [k] get_page_from_freelist
2.46% cat [kernel.kallsyms] [k] put_page
2.13% kswapd0 [kernel.kallsyms] [k] shrink_page_list
1.88% cat [kernel.kallsyms] [k] __radix_tree_lookup
1.67% cat [kernel.kallsyms] [k] __pagevec_lru_add_fn
1.39% kswapd0 [kernel.kallsyms] [k] free_pcppages_bulk
1.30% cat [kernel.kallsyms] [k] kfree
As you can see, the memcg footprint has shrunk quite a bit.
text data bss dec hex filename
37970 9892 400 48262 bc86 mm/memcontrol.o.old
35239 9892 400 45531 b1db mm/memcontrol.o
This patch (of 4):
The memcg charge API charges pages before they are rmapped - i.e. have an
actual "type" - and so every callsite needs its own set of charge and
uncharge functions to know what type is being operated on. Worse,
uncharge has to happen from a context that is still type-specific, rather
than at the end of the page's lifetime with exclusive access, and so
requires a lot of synchronization.
Rewrite the charge API to provide a generic set of try_charge(),
commit_charge() and cancel_charge() transaction operations, much like
what's currently done for swap-in:
mem_cgroup_try_charge() attempts to reserve a charge, reclaiming
pages from the memcg if necessary.
mem_cgroup_commit_charge() commits the page to the charge once it
has a valid page->mapping and PageAnon() reliably tells the type.
mem_cgroup_cancel_charge() aborts the transaction.
This reduces the charge API and enables subsequent patches to
drastically simplify uncharging.
As pages need to be committed after rmap is established but before they
are added to the LRU, page_add_new_anon_rmap() must stop doing LRU
additions again. Revive lru_cache_add_active_or_unevictable().
[hughd@google.com: fix shmem_unuse]
[hughd@google.com: Add comments on the private use of -EAGAIN]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-08-08 14:19:20 -07:00
|
|
|
|
2021-04-29 22:56:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcontrol: rewrite charge API
These patches rework memcg charge lifetime to integrate more naturally
with the lifetime of user pages. This drastically simplifies the code and
reduces charging and uncharging overhead. The most expensive part of
charging and uncharging is the page_cgroup bit spinlock, which is removed
entirely after this series.
Here are the top-10 profile entries of a stress test that reads a 128G
sparse file on a freshly booted box, without even a dedicated cgroup (i.e.
executing in the root memcg). Before:
15.36% cat [kernel.kallsyms] [k] copy_user_generic_string
13.31% cat [kernel.kallsyms] [k] memset
11.48% cat [kernel.kallsyms] [k] do_mpage_readpage
4.23% cat [kernel.kallsyms] [k] get_page_from_freelist
2.38% cat [kernel.kallsyms] [k] put_page
2.32% cat [kernel.kallsyms] [k] __mem_cgroup_commit_charge
2.18% kswapd0 [kernel.kallsyms] [k] __mem_cgroup_uncharge_common
1.92% kswapd0 [kernel.kallsyms] [k] shrink_page_list
1.86% cat [kernel.kallsyms] [k] __radix_tree_lookup
1.62% cat [kernel.kallsyms] [k] __pagevec_lru_add_fn
After:
15.67% cat [kernel.kallsyms] [k] copy_user_generic_string
13.48% cat [kernel.kallsyms] [k] memset
11.42% cat [kernel.kallsyms] [k] do_mpage_readpage
3.98% cat [kernel.kallsyms] [k] get_page_from_freelist
2.46% cat [kernel.kallsyms] [k] put_page
2.13% kswapd0 [kernel.kallsyms] [k] shrink_page_list
1.88% cat [kernel.kallsyms] [k] __radix_tree_lookup
1.67% cat [kernel.kallsyms] [k] __pagevec_lru_add_fn
1.39% kswapd0 [kernel.kallsyms] [k] free_pcppages_bulk
1.30% cat [kernel.kallsyms] [k] kfree
As you can see, the memcg footprint has shrunk quite a bit.
text data bss dec hex filename
37970 9892 400 48262 bc86 mm/memcontrol.o.old
35239 9892 400 45531 b1db mm/memcontrol.o
This patch (of 4):
The memcg charge API charges pages before they are rmapped - i.e. have an
actual "type" - and so every callsite needs its own set of charge and
uncharge functions to know what type is being operated on. Worse,
uncharge has to happen from a context that is still type-specific, rather
than at the end of the page's lifetime with exclusive access, and so
requires a lot of synchronization.
Rewrite the charge API to provide a generic set of try_charge(),
commit_charge() and cancel_charge() transaction operations, much like
what's currently done for swap-in:
mem_cgroup_try_charge() attempts to reserve a charge, reclaiming
pages from the memcg if necessary.
mem_cgroup_commit_charge() commits the page to the charge once it
has a valid page->mapping and PageAnon() reliably tells the type.
mem_cgroup_cancel_charge() aborts the transaction.
This reduces the charge API and enables subsequent patches to
drastically simplify uncharging.
As pages need to be committed after rmap is established but before they
are added to the LRU, page_add_new_anon_rmap() must stop doing LRU
additions again. Revive lru_cache_add_active_or_unevictable().
[hughd@google.com: fix shmem_unuse]
[hughd@google.com: Add comments on the private use of -EAGAIN]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-08-08 14:19:20 -07:00
|
|
|
|
2021-06-25 09:27:04 -04:00
|
|
|
|
2014-08-08 14:19:33 -07:00
|
|
|
|
2021-04-29 22:56:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcontrol: rewrite charge API
These patches rework memcg charge lifetime to integrate more naturally
with the lifetime of user pages. This drastically simplifies the code and
reduces charging and uncharging overhead. The most expensive part of
charging and uncharging is the page_cgroup bit spinlock, which is removed
entirely after this series.
Here are the top-10 profile entries of a stress test that reads a 128G
sparse file on a freshly booted box, without even a dedicated cgroup (i.e.
executing in the root memcg). Before:
15.36% cat [kernel.kallsyms] [k] copy_user_generic_string
13.31% cat [kernel.kallsyms] [k] memset
11.48% cat [kernel.kallsyms] [k] do_mpage_readpage
4.23% cat [kernel.kallsyms] [k] get_page_from_freelist
2.38% cat [kernel.kallsyms] [k] put_page
2.32% cat [kernel.kallsyms] [k] __mem_cgroup_commit_charge
2.18% kswapd0 [kernel.kallsyms] [k] __mem_cgroup_uncharge_common
1.92% kswapd0 [kernel.kallsyms] [k] shrink_page_list
1.86% cat [kernel.kallsyms] [k] __radix_tree_lookup
1.62% cat [kernel.kallsyms] [k] __pagevec_lru_add_fn
After:
15.67% cat [kernel.kallsyms] [k] copy_user_generic_string
13.48% cat [kernel.kallsyms] [k] memset
11.42% cat [kernel.kallsyms] [k] do_mpage_readpage
3.98% cat [kernel.kallsyms] [k] get_page_from_freelist
2.46% cat [kernel.kallsyms] [k] put_page
2.13% kswapd0 [kernel.kallsyms] [k] shrink_page_list
1.88% cat [kernel.kallsyms] [k] __radix_tree_lookup
1.67% cat [kernel.kallsyms] [k] __pagevec_lru_add_fn
1.39% kswapd0 [kernel.kallsyms] [k] free_pcppages_bulk
1.30% cat [kernel.kallsyms] [k] kfree
As you can see, the memcg footprint has shrunk quite a bit.
text data bss dec hex filename
37970 9892 400 48262 bc86 mm/memcontrol.o.old
35239 9892 400 45531 b1db mm/memcontrol.o
This patch (of 4):
The memcg charge API charges pages before they are rmapped - i.e. have an
actual "type" - and so every callsite needs its own set of charge and
uncharge functions to know what type is being operated on. Worse,
uncharge has to happen from a context that is still type-specific, rather
than at the end of the page's lifetime with exclusive access, and so
requires a lot of synchronization.
Rewrite the charge API to provide a generic set of try_charge(),
commit_charge() and cancel_charge() transaction operations, much like
what's currently done for swap-in:
mem_cgroup_try_charge() attempts to reserve a charge, reclaiming
pages from the memcg if necessary.
mem_cgroup_commit_charge() commits the page to the charge once it
has a valid page->mapping and PageAnon() reliably tells the type.
mem_cgroup_cancel_charge() aborts the transaction.
This reduces the charge API and enables subsequent patches to
drastically simplify uncharging.
As pages need to be committed after rmap is established but before they
are added to the LRU, page_add_new_anon_rmap() must stop doing LRU
additions again. Revive lru_cache_add_active_or_unevictable().
[hughd@google.com: fix shmem_unuse]
[hughd@google.com: Add comments on the private use of -EAGAIN]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-08-08 14:19:20 -07:00
|
|
|
|
2017-09-08 16:11:50 -07:00
|
|
|
|
|
|
|
|
|
mm: memcontrol: use obj_cgroup APIs to charge kmem pages
Since Roman's series "The new cgroup slab memory controller" applied.
All slab objects are charged via the new APIs of obj_cgroup. The new
APIs introduce a struct obj_cgroup to charge slab objects. It prevents
long-living objects from pinning the original memory cgroup in the
memory. But there are still some corner objects (e.g. allocations
larger than order-1 page on SLUB) which are not charged via the new
APIs. Those objects (include the pages which are allocated from buddy
allocator directly) are charged as kmem pages which still hold a
reference to the memory cgroup.
We want to reuse the obj_cgroup APIs to charge the kmem pages. If we do
that, we should store an object cgroup pointer to page->memcg_data for
the kmem pages.
Finally, page->memcg_data will have 3 different meanings.
1) For the slab pages, page->memcg_data points to an object cgroups
vector.
2) For the kmem pages (exclude the slab pages), page->memcg_data
points to an object cgroup.
3) For the user pages (e.g. the LRU pages), page->memcg_data points
to a memory cgroup.
We do not change the behavior of page_memcg() and page_memcg_rcu(). They
are also suitable for LRU pages and kmem pages. Why?
Because memory allocations pinning memcgs for a long time - it exists at a
larger scale and is causing recurring problems in the real world: page
cache doesn't get reclaimed for a long time, or is used by the second,
third, fourth, ... instance of the same job that was restarted into a new
cgroup every time. Unreclaimable dying cgroups pile up, waste memory, and
make page reclaim very inefficient.
We can convert LRU pages and most other raw memcg pins to the objcg
direction to fix this problem, and then the page->memcg will always point
to an object cgroup pointer. At that time, LRU pages and kmem pages will
be treated the same. The implementation of page_memcg() will remove the
kmem page check.
This patch aims to charge the kmem pages by using the new APIs of
obj_cgroup. Finally, the page->memcg_data of the kmem page points to an
object cgroup. We can use the __page_objcg() to get the object cgroup
associated with a kmem page. Or we can use page_memcg() to get the memory
cgroup associated with a kmem page, but caller must ensure that the
returned memcg won't be released (e.g. acquire the rcu_read_lock or
css_set_lock).
Link: https://lkml.kernel.org/r/20210401030141.37061-1-songmuchun@bytedance.com
Link: https://lkml.kernel.org/r/20210319163821.20704-6-songmuchun@bytedance.com
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Roman Gushchin <guro@fb.com>
Reviewed-by: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Xiongchun Duan <duanxiongchun@bytedance.com>
Cc: Christian Borntraeger <borntraeger@de.ibm.com>
[songmuchun@bytedance.com: fix forget to obtain the ref to objcg in split_page_memcg]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:52 -07:00
|
|
|
|
2017-09-08 16:11:50 -07:00
|
|
|
|
|
|
|
|
|
2021-06-25 09:05:47 -04:00
|
|
|
|
2017-09-08 16:11:50 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-08-08 14:19:24 -07:00
|
|
|
|
2017-09-08 16:11:50 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcontrol: use obj_cgroup APIs to charge kmem pages
Since Roman's series "The new cgroup slab memory controller" applied.
All slab objects are charged via the new APIs of obj_cgroup. The new
APIs introduce a struct obj_cgroup to charge slab objects. It prevents
long-living objects from pinning the original memory cgroup in the
memory. But there are still some corner objects (e.g. allocations
larger than order-1 page on SLUB) which are not charged via the new
APIs. Those objects (include the pages which are allocated from buddy
allocator directly) are charged as kmem pages which still hold a
reference to the memory cgroup.
We want to reuse the obj_cgroup APIs to charge the kmem pages. If we do
that, we should store an object cgroup pointer to page->memcg_data for
the kmem pages.
Finally, page->memcg_data will have 3 different meanings.
1) For the slab pages, page->memcg_data points to an object cgroups
vector.
2) For the kmem pages (exclude the slab pages), page->memcg_data
points to an object cgroup.
3) For the user pages (e.g. the LRU pages), page->memcg_data points
to a memory cgroup.
We do not change the behavior of page_memcg() and page_memcg_rcu(). They
are also suitable for LRU pages and kmem pages. Why?
Because memory allocations pinning memcgs for a long time - it exists at a
larger scale and is causing recurring problems in the real world: page
cache doesn't get reclaimed for a long time, or is used by the second,
third, fourth, ... instance of the same job that was restarted into a new
cgroup every time. Unreclaimable dying cgroups pile up, waste memory, and
make page reclaim very inefficient.
We can convert LRU pages and most other raw memcg pins to the objcg
direction to fix this problem, and then the page->memcg will always point
to an object cgroup pointer. At that time, LRU pages and kmem pages will
be treated the same. The implementation of page_memcg() will remove the
kmem page check.
This patch aims to charge the kmem pages by using the new APIs of
obj_cgroup. Finally, the page->memcg_data of the kmem page points to an
object cgroup. We can use the __page_objcg() to get the object cgroup
associated with a kmem page. Or we can use page_memcg() to get the memory
cgroup associated with a kmem page, but caller must ensure that the
returned memcg won't be released (e.g. acquire the rcu_read_lock or
css_set_lock).
Link: https://lkml.kernel.org/r/20210401030141.37061-1-songmuchun@bytedance.com
Link: https://lkml.kernel.org/r/20210319163821.20704-6-songmuchun@bytedance.com
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Roman Gushchin <guro@fb.com>
Reviewed-by: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Xiongchun Duan <duanxiongchun@bytedance.com>
Cc: Christian Borntraeger <borntraeger@de.ibm.com>
[songmuchun@bytedance.com: fix forget to obtain the ref to objcg in split_page_memcg]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:52 -07:00
|
|
|
|
2025-04-03 18:39:07 -07:00
|
|
|
|
2024-06-28 21:03:09 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-06-24 17:59:01 -07:00
|
|
|
|
2014-09-05 08:43:57 -04:00
|
|
|
|
2014-08-08 14:19:24 -07:00
|
|
|
|
2024-08-14 22:04:50 -07:00
|
|
|
|
2020-09-04 16:35:24 -07:00
|
|
|
|
2021-06-29 21:47:12 -04:00
|
|
|
|
2020-09-04 16:35:24 -07:00
|
|
|
|
2017-09-08 16:11:50 -07:00
|
|
|
|
|
|
|
|
|
2021-06-29 21:47:12 -04:00
|
|
|
|
2017-09-08 16:11:50 -07:00
|
|
|
|
2021-06-29 21:47:12 -04:00
|
|
|
|
mm: memcontrol: use obj_cgroup APIs to charge kmem pages
Since Roman's series "The new cgroup slab memory controller" applied.
All slab objects are charged via the new APIs of obj_cgroup. The new
APIs introduce a struct obj_cgroup to charge slab objects. It prevents
long-living objects from pinning the original memory cgroup in the
memory. But there are still some corner objects (e.g. allocations
larger than order-1 page on SLUB) which are not charged via the new
APIs. Those objects (include the pages which are allocated from buddy
allocator directly) are charged as kmem pages which still hold a
reference to the memory cgroup.
We want to reuse the obj_cgroup APIs to charge the kmem pages. If we do
that, we should store an object cgroup pointer to page->memcg_data for
the kmem pages.
Finally, page->memcg_data will have 3 different meanings.
1) For the slab pages, page->memcg_data points to an object cgroups
vector.
2) For the kmem pages (exclude the slab pages), page->memcg_data
points to an object cgroup.
3) For the user pages (e.g. the LRU pages), page->memcg_data points
to a memory cgroup.
We do not change the behavior of page_memcg() and page_memcg_rcu(). They
are also suitable for LRU pages and kmem pages. Why?
Because memory allocations pinning memcgs for a long time - it exists at a
larger scale and is causing recurring problems in the real world: page
cache doesn't get reclaimed for a long time, or is used by the second,
third, fourth, ... instance of the same job that was restarted into a new
cgroup every time. Unreclaimable dying cgroups pile up, waste memory, and
make page reclaim very inefficient.
We can convert LRU pages and most other raw memcg pins to the objcg
direction to fix this problem, and then the page->memcg will always point
to an object cgroup pointer. At that time, LRU pages and kmem pages will
be treated the same. The implementation of page_memcg() will remove the
kmem page check.
This patch aims to charge the kmem pages by using the new APIs of
obj_cgroup. Finally, the page->memcg_data of the kmem page points to an
object cgroup. We can use the __page_objcg() to get the object cgroup
associated with a kmem page. Or we can use page_memcg() to get the memory
cgroup associated with a kmem page, but caller must ensure that the
returned memcg won't be released (e.g. acquire the rcu_read_lock or
css_set_lock).
Link: https://lkml.kernel.org/r/20210401030141.37061-1-songmuchun@bytedance.com
Link: https://lkml.kernel.org/r/20210319163821.20704-6-songmuchun@bytedance.com
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Roman Gushchin <guro@fb.com>
Reviewed-by: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Xiongchun Duan <duanxiongchun@bytedance.com>
Cc: Christian Borntraeger <borntraeger@de.ibm.com>
[songmuchun@bytedance.com: fix forget to obtain the ref to objcg in split_page_memcg]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:52 -07:00
|
|
|
|
|
|
|
|
|
2020-06-03 16:01:44 -07:00
|
|
|
|
2021-06-29 21:47:12 -04:00
|
|
|
|
2017-09-08 16:11:50 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-29 21:47:12 -04:00
|
|
|
|
|
|
|
|
|
2017-09-08 16:11:50 -07:00
|
|
|
|
2022-03-22 14:40:35 -07:00
|
|
|
|
2021-06-28 14:59:26 -04:00
|
|
|
|
mm: memcontrol: use obj_cgroup APIs to charge kmem pages
Since Roman's series "The new cgroup slab memory controller" applied.
All slab objects are charged via the new APIs of obj_cgroup. The new
APIs introduce a struct obj_cgroup to charge slab objects. It prevents
long-living objects from pinning the original memory cgroup in the
memory. But there are still some corner objects (e.g. allocations
larger than order-1 page on SLUB) which are not charged via the new
APIs. Those objects (include the pages which are allocated from buddy
allocator directly) are charged as kmem pages which still hold a
reference to the memory cgroup.
We want to reuse the obj_cgroup APIs to charge the kmem pages. If we do
that, we should store an object cgroup pointer to page->memcg_data for
the kmem pages.
Finally, page->memcg_data will have 3 different meanings.
1) For the slab pages, page->memcg_data points to an object cgroups
vector.
2) For the kmem pages (exclude the slab pages), page->memcg_data
points to an object cgroup.
3) For the user pages (e.g. the LRU pages), page->memcg_data points
to a memory cgroup.
We do not change the behavior of page_memcg() and page_memcg_rcu(). They
are also suitable for LRU pages and kmem pages. Why?
Because memory allocations pinning memcgs for a long time - it exists at a
larger scale and is causing recurring problems in the real world: page
cache doesn't get reclaimed for a long time, or is used by the second,
third, fourth, ... instance of the same job that was restarted into a new
cgroup every time. Unreclaimable dying cgroups pile up, waste memory, and
make page reclaim very inefficient.
We can convert LRU pages and most other raw memcg pins to the objcg
direction to fix this problem, and then the page->memcg will always point
to an object cgroup pointer. At that time, LRU pages and kmem pages will
be treated the same. The implementation of page_memcg() will remove the
kmem page check.
This patch aims to charge the kmem pages by using the new APIs of
obj_cgroup. Finally, the page->memcg_data of the kmem page points to an
object cgroup. We can use the __page_objcg() to get the object cgroup
associated with a kmem page. Or we can use page_memcg() to get the memory
cgroup associated with a kmem page, but caller must ensure that the
returned memcg won't be released (e.g. acquire the rcu_read_lock or
css_set_lock).
Link: https://lkml.kernel.org/r/20210401030141.37061-1-songmuchun@bytedance.com
Link: https://lkml.kernel.org/r/20210319163821.20704-6-songmuchun@bytedance.com
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Roman Gushchin <guro@fb.com>
Reviewed-by: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Xiongchun Duan <duanxiongchun@bytedance.com>
Cc: Christian Borntraeger <borntraeger@de.ibm.com>
[songmuchun@bytedance.com: fix forget to obtain the ref to objcg in split_page_memcg]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:52 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-28 14:59:26 -04:00
|
|
|
|
mm: memcontrol: use obj_cgroup APIs to charge kmem pages
Since Roman's series "The new cgroup slab memory controller" applied.
All slab objects are charged via the new APIs of obj_cgroup. The new
APIs introduce a struct obj_cgroup to charge slab objects. It prevents
long-living objects from pinning the original memory cgroup in the
memory. But there are still some corner objects (e.g. allocations
larger than order-1 page on SLUB) which are not charged via the new
APIs. Those objects (include the pages which are allocated from buddy
allocator directly) are charged as kmem pages which still hold a
reference to the memory cgroup.
We want to reuse the obj_cgroup APIs to charge the kmem pages. If we do
that, we should store an object cgroup pointer to page->memcg_data for
the kmem pages.
Finally, page->memcg_data will have 3 different meanings.
1) For the slab pages, page->memcg_data points to an object cgroups
vector.
2) For the kmem pages (exclude the slab pages), page->memcg_data
points to an object cgroup.
3) For the user pages (e.g. the LRU pages), page->memcg_data points
to a memory cgroup.
We do not change the behavior of page_memcg() and page_memcg_rcu(). They
are also suitable for LRU pages and kmem pages. Why?
Because memory allocations pinning memcgs for a long time - it exists at a
larger scale and is causing recurring problems in the real world: page
cache doesn't get reclaimed for a long time, or is used by the second,
third, fourth, ... instance of the same job that was restarted into a new
cgroup every time. Unreclaimable dying cgroups pile up, waste memory, and
make page reclaim very inefficient.
We can convert LRU pages and most other raw memcg pins to the objcg
direction to fix this problem, and then the page->memcg will always point
to an object cgroup pointer. At that time, LRU pages and kmem pages will
be treated the same. The implementation of page_memcg() will remove the
kmem page check.
This patch aims to charge the kmem pages by using the new APIs of
obj_cgroup. Finally, the page->memcg_data of the kmem page points to an
object cgroup. We can use the __page_objcg() to get the object cgroup
associated with a kmem page. Or we can use page_memcg() to get the memory
cgroup associated with a kmem page, but caller must ensure that the
returned memcg won't be released (e.g. acquire the rcu_read_lock or
css_set_lock).
Link: https://lkml.kernel.org/r/20210401030141.37061-1-songmuchun@bytedance.com
Link: https://lkml.kernel.org/r/20210319163821.20704-6-songmuchun@bytedance.com
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Roman Gushchin <guro@fb.com>
Reviewed-by: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Xiongchun Duan <duanxiongchun@bytedance.com>
Cc: Christian Borntraeger <borntraeger@de.ibm.com>
[songmuchun@bytedance.com: fix forget to obtain the ref to objcg in split_page_memcg]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:52 -07:00
|
|
|
|
2017-09-08 16:11:50 -07:00
|
|
|
|
mm: memcontrol: use obj_cgroup APIs to charge kmem pages
Since Roman's series "The new cgroup slab memory controller" applied.
All slab objects are charged via the new APIs of obj_cgroup. The new
APIs introduce a struct obj_cgroup to charge slab objects. It prevents
long-living objects from pinning the original memory cgroup in the
memory. But there are still some corner objects (e.g. allocations
larger than order-1 page on SLUB) which are not charged via the new
APIs. Those objects (include the pages which are allocated from buddy
allocator directly) are charged as kmem pages which still hold a
reference to the memory cgroup.
We want to reuse the obj_cgroup APIs to charge the kmem pages. If we do
that, we should store an object cgroup pointer to page->memcg_data for
the kmem pages.
Finally, page->memcg_data will have 3 different meanings.
1) For the slab pages, page->memcg_data points to an object cgroups
vector.
2) For the kmem pages (exclude the slab pages), page->memcg_data
points to an object cgroup.
3) For the user pages (e.g. the LRU pages), page->memcg_data points
to a memory cgroup.
We do not change the behavior of page_memcg() and page_memcg_rcu(). They
are also suitable for LRU pages and kmem pages. Why?
Because memory allocations pinning memcgs for a long time - it exists at a
larger scale and is causing recurring problems in the real world: page
cache doesn't get reclaimed for a long time, or is used by the second,
third, fourth, ... instance of the same job that was restarted into a new
cgroup every time. Unreclaimable dying cgroups pile up, waste memory, and
make page reclaim very inefficient.
We can convert LRU pages and most other raw memcg pins to the objcg
direction to fix this problem, and then the page->memcg will always point
to an object cgroup pointer. At that time, LRU pages and kmem pages will
be treated the same. The implementation of page_memcg() will remove the
kmem page check.
This patch aims to charge the kmem pages by using the new APIs of
obj_cgroup. Finally, the page->memcg_data of the kmem page points to an
object cgroup. We can use the __page_objcg() to get the object cgroup
associated with a kmem page. Or we can use page_memcg() to get the memory
cgroup associated with a kmem page, but caller must ensure that the
returned memcg won't be released (e.g. acquire the rcu_read_lock or
css_set_lock).
Link: https://lkml.kernel.org/r/20210401030141.37061-1-songmuchun@bytedance.com
Link: https://lkml.kernel.org/r/20210319163821.20704-6-songmuchun@bytedance.com
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Roman Gushchin <guro@fb.com>
Reviewed-by: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Xiongchun Duan <duanxiongchun@bytedance.com>
Cc: Christian Borntraeger <borntraeger@de.ibm.com>
[songmuchun@bytedance.com: fix forget to obtain the ref to objcg in split_page_memcg]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:52 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-09-08 16:11:50 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcontrol: use obj_cgroup APIs to charge kmem pages
Since Roman's series "The new cgroup slab memory controller" applied.
All slab objects are charged via the new APIs of obj_cgroup. The new
APIs introduce a struct obj_cgroup to charge slab objects. It prevents
long-living objects from pinning the original memory cgroup in the
memory. But there are still some corner objects (e.g. allocations
larger than order-1 page on SLUB) which are not charged via the new
APIs. Those objects (include the pages which are allocated from buddy
allocator directly) are charged as kmem pages which still hold a
reference to the memory cgroup.
We want to reuse the obj_cgroup APIs to charge the kmem pages. If we do
that, we should store an object cgroup pointer to page->memcg_data for
the kmem pages.
Finally, page->memcg_data will have 3 different meanings.
1) For the slab pages, page->memcg_data points to an object cgroups
vector.
2) For the kmem pages (exclude the slab pages), page->memcg_data
points to an object cgroup.
3) For the user pages (e.g. the LRU pages), page->memcg_data points
to a memory cgroup.
We do not change the behavior of page_memcg() and page_memcg_rcu(). They
are also suitable for LRU pages and kmem pages. Why?
Because memory allocations pinning memcgs for a long time - it exists at a
larger scale and is causing recurring problems in the real world: page
cache doesn't get reclaimed for a long time, or is used by the second,
third, fourth, ... instance of the same job that was restarted into a new
cgroup every time. Unreclaimable dying cgroups pile up, waste memory, and
make page reclaim very inefficient.
We can convert LRU pages and most other raw memcg pins to the objcg
direction to fix this problem, and then the page->memcg will always point
to an object cgroup pointer. At that time, LRU pages and kmem pages will
be treated the same. The implementation of page_memcg() will remove the
kmem page check.
This patch aims to charge the kmem pages by using the new APIs of
obj_cgroup. Finally, the page->memcg_data of the kmem page points to an
object cgroup. We can use the __page_objcg() to get the object cgroup
associated with a kmem page. Or we can use page_memcg() to get the memory
cgroup associated with a kmem page, but caller must ensure that the
returned memcg won't be released (e.g. acquire the rcu_read_lock or
css_set_lock).
Link: https://lkml.kernel.org/r/20210401030141.37061-1-songmuchun@bytedance.com
Link: https://lkml.kernel.org/r/20210319163821.20704-6-songmuchun@bytedance.com
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Roman Gushchin <guro@fb.com>
Reviewed-by: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Xiongchun Duan <duanxiongchun@bytedance.com>
Cc: Christian Borntraeger <borntraeger@de.ibm.com>
[songmuchun@bytedance.com: fix forget to obtain the ref to objcg in split_page_memcg]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:52 -07:00
|
|
|
|
2021-06-29 21:47:12 -04:00
|
|
|
|
2020-09-04 16:35:24 -07:00
|
|
|
|
|
|
|
|
|
mm: memcontrol: use obj_cgroup APIs to charge kmem pages
Since Roman's series "The new cgroup slab memory controller" applied.
All slab objects are charged via the new APIs of obj_cgroup. The new
APIs introduce a struct obj_cgroup to charge slab objects. It prevents
long-living objects from pinning the original memory cgroup in the
memory. But there are still some corner objects (e.g. allocations
larger than order-1 page on SLUB) which are not charged via the new
APIs. Those objects (include the pages which are allocated from buddy
allocator directly) are charged as kmem pages which still hold a
reference to the memory cgroup.
We want to reuse the obj_cgroup APIs to charge the kmem pages. If we do
that, we should store an object cgroup pointer to page->memcg_data for
the kmem pages.
Finally, page->memcg_data will have 3 different meanings.
1) For the slab pages, page->memcg_data points to an object cgroups
vector.
2) For the kmem pages (exclude the slab pages), page->memcg_data
points to an object cgroup.
3) For the user pages (e.g. the LRU pages), page->memcg_data points
to a memory cgroup.
We do not change the behavior of page_memcg() and page_memcg_rcu(). They
are also suitable for LRU pages and kmem pages. Why?
Because memory allocations pinning memcgs for a long time - it exists at a
larger scale and is causing recurring problems in the real world: page
cache doesn't get reclaimed for a long time, or is used by the second,
third, fourth, ... instance of the same job that was restarted into a new
cgroup every time. Unreclaimable dying cgroups pile up, waste memory, and
make page reclaim very inefficient.
We can convert LRU pages and most other raw memcg pins to the objcg
direction to fix this problem, and then the page->memcg will always point
to an object cgroup pointer. At that time, LRU pages and kmem pages will
be treated the same. The implementation of page_memcg() will remove the
kmem page check.
This patch aims to charge the kmem pages by using the new APIs of
obj_cgroup. Finally, the page->memcg_data of the kmem page points to an
object cgroup. We can use the __page_objcg() to get the object cgroup
associated with a kmem page. Or we can use page_memcg() to get the memory
cgroup associated with a kmem page, but caller must ensure that the
returned memcg won't be released (e.g. acquire the rcu_read_lock or
css_set_lock).
Link: https://lkml.kernel.org/r/20210401030141.37061-1-songmuchun@bytedance.com
Link: https://lkml.kernel.org/r/20210319163821.20704-6-songmuchun@bytedance.com
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Roman Gushchin <guro@fb.com>
Reviewed-by: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Xiongchun Duan <duanxiongchun@bytedance.com>
Cc: Christian Borntraeger <borntraeger@de.ibm.com>
[songmuchun@bytedance.com: fix forget to obtain the ref to objcg in split_page_memcg]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:52 -07:00
|
|
|
|
2017-09-08 16:11:50 -07:00
|
|
|
|
|
|
|
|
|
2021-06-29 21:47:12 -04:00
|
|
|
|
2017-09-08 16:11:50 -07:00
|
|
|
|
2022-03-22 14:40:35 -07:00
|
|
|
|
mm: memcontrol: use obj_cgroup APIs to charge kmem pages
Since Roman's series "The new cgroup slab memory controller" applied.
All slab objects are charged via the new APIs of obj_cgroup. The new
APIs introduce a struct obj_cgroup to charge slab objects. It prevents
long-living objects from pinning the original memory cgroup in the
memory. But there are still some corner objects (e.g. allocations
larger than order-1 page on SLUB) which are not charged via the new
APIs. Those objects (include the pages which are allocated from buddy
allocator directly) are charged as kmem pages which still hold a
reference to the memory cgroup.
We want to reuse the obj_cgroup APIs to charge the kmem pages. If we do
that, we should store an object cgroup pointer to page->memcg_data for
the kmem pages.
Finally, page->memcg_data will have 3 different meanings.
1) For the slab pages, page->memcg_data points to an object cgroups
vector.
2) For the kmem pages (exclude the slab pages), page->memcg_data
points to an object cgroup.
3) For the user pages (e.g. the LRU pages), page->memcg_data points
to a memory cgroup.
We do not change the behavior of page_memcg() and page_memcg_rcu(). They
are also suitable for LRU pages and kmem pages. Why?
Because memory allocations pinning memcgs for a long time - it exists at a
larger scale and is causing recurring problems in the real world: page
cache doesn't get reclaimed for a long time, or is used by the second,
third, fourth, ... instance of the same job that was restarted into a new
cgroup every time. Unreclaimable dying cgroups pile up, waste memory, and
make page reclaim very inefficient.
We can convert LRU pages and most other raw memcg pins to the objcg
direction to fix this problem, and then the page->memcg will always point
to an object cgroup pointer. At that time, LRU pages and kmem pages will
be treated the same. The implementation of page_memcg() will remove the
kmem page check.
This patch aims to charge the kmem pages by using the new APIs of
obj_cgroup. Finally, the page->memcg_data of the kmem page points to an
object cgroup. We can use the __page_objcg() to get the object cgroup
associated with a kmem page. Or we can use page_memcg() to get the memory
cgroup associated with a kmem page, but caller must ensure that the
returned memcg won't be released (e.g. acquire the rcu_read_lock or
css_set_lock).
Link: https://lkml.kernel.org/r/20210401030141.37061-1-songmuchun@bytedance.com
Link: https://lkml.kernel.org/r/20210319163821.20704-6-songmuchun@bytedance.com
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Roman Gushchin <guro@fb.com>
Reviewed-by: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Xiongchun Duan <duanxiongchun@bytedance.com>
Cc: Christian Borntraeger <borntraeger@de.ibm.com>
[songmuchun@bytedance.com: fix forget to obtain the ref to objcg in split_page_memcg]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:52 -07:00
|
|
|
|
2020-06-03 16:01:44 -07:00
|
|
|
|
mm: memcontrol: use obj_cgroup APIs to charge kmem pages
Since Roman's series "The new cgroup slab memory controller" applied.
All slab objects are charged via the new APIs of obj_cgroup. The new
APIs introduce a struct obj_cgroup to charge slab objects. It prevents
long-living objects from pinning the original memory cgroup in the
memory. But there are still some corner objects (e.g. allocations
larger than order-1 page on SLUB) which are not charged via the new
APIs. Those objects (include the pages which are allocated from buddy
allocator directly) are charged as kmem pages which still hold a
reference to the memory cgroup.
We want to reuse the obj_cgroup APIs to charge the kmem pages. If we do
that, we should store an object cgroup pointer to page->memcg_data for
the kmem pages.
Finally, page->memcg_data will have 3 different meanings.
1) For the slab pages, page->memcg_data points to an object cgroups
vector.
2) For the kmem pages (exclude the slab pages), page->memcg_data
points to an object cgroup.
3) For the user pages (e.g. the LRU pages), page->memcg_data points
to a memory cgroup.
We do not change the behavior of page_memcg() and page_memcg_rcu(). They
are also suitable for LRU pages and kmem pages. Why?
Because memory allocations pinning memcgs for a long time - it exists at a
larger scale and is causing recurring problems in the real world: page
cache doesn't get reclaimed for a long time, or is used by the second,
third, fourth, ... instance of the same job that was restarted into a new
cgroup every time. Unreclaimable dying cgroups pile up, waste memory, and
make page reclaim very inefficient.
We can convert LRU pages and most other raw memcg pins to the objcg
direction to fix this problem, and then the page->memcg will always point
to an object cgroup pointer. At that time, LRU pages and kmem pages will
be treated the same. The implementation of page_memcg() will remove the
kmem page check.
This patch aims to charge the kmem pages by using the new APIs of
obj_cgroup. Finally, the page->memcg_data of the kmem page points to an
object cgroup. We can use the __page_objcg() to get the object cgroup
associated with a kmem page. Or we can use page_memcg() to get the memory
cgroup associated with a kmem page, but caller must ensure that the
returned memcg won't be released (e.g. acquire the rcu_read_lock or
css_set_lock).
Link: https://lkml.kernel.org/r/20210401030141.37061-1-songmuchun@bytedance.com
Link: https://lkml.kernel.org/r/20210319163821.20704-6-songmuchun@bytedance.com
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Roman Gushchin <guro@fb.com>
Reviewed-by: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Xiongchun Duan <duanxiongchun@bytedance.com>
Cc: Christian Borntraeger <borntraeger@de.ibm.com>
[songmuchun@bytedance.com: fix forget to obtain the ref to objcg in split_page_memcg]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:52 -07:00
|
|
|
|
2021-06-29 21:47:12 -04:00
|
|
|
|
mm: memcontrol: use obj_cgroup APIs to charge kmem pages
Since Roman's series "The new cgroup slab memory controller" applied.
All slab objects are charged via the new APIs of obj_cgroup. The new
APIs introduce a struct obj_cgroup to charge slab objects. It prevents
long-living objects from pinning the original memory cgroup in the
memory. But there are still some corner objects (e.g. allocations
larger than order-1 page on SLUB) which are not charged via the new
APIs. Those objects (include the pages which are allocated from buddy
allocator directly) are charged as kmem pages which still hold a
reference to the memory cgroup.
We want to reuse the obj_cgroup APIs to charge the kmem pages. If we do
that, we should store an object cgroup pointer to page->memcg_data for
the kmem pages.
Finally, page->memcg_data will have 3 different meanings.
1) For the slab pages, page->memcg_data points to an object cgroups
vector.
2) For the kmem pages (exclude the slab pages), page->memcg_data
points to an object cgroup.
3) For the user pages (e.g. the LRU pages), page->memcg_data points
to a memory cgroup.
We do not change the behavior of page_memcg() and page_memcg_rcu(). They
are also suitable for LRU pages and kmem pages. Why?
Because memory allocations pinning memcgs for a long time - it exists at a
larger scale and is causing recurring problems in the real world: page
cache doesn't get reclaimed for a long time, or is used by the second,
third, fourth, ... instance of the same job that was restarted into a new
cgroup every time. Unreclaimable dying cgroups pile up, waste memory, and
make page reclaim very inefficient.
We can convert LRU pages and most other raw memcg pins to the objcg
direction to fix this problem, and then the page->memcg will always point
to an object cgroup pointer. At that time, LRU pages and kmem pages will
be treated the same. The implementation of page_memcg() will remove the
kmem page check.
This patch aims to charge the kmem pages by using the new APIs of
obj_cgroup. Finally, the page->memcg_data of the kmem page points to an
object cgroup. We can use the __page_objcg() to get the object cgroup
associated with a kmem page. Or we can use page_memcg() to get the memory
cgroup associated with a kmem page, but caller must ensure that the
returned memcg won't be released (e.g. acquire the rcu_read_lock or
css_set_lock).
Link: https://lkml.kernel.org/r/20210401030141.37061-1-songmuchun@bytedance.com
Link: https://lkml.kernel.org/r/20210319163821.20704-6-songmuchun@bytedance.com
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Roman Gushchin <guro@fb.com>
Reviewed-by: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Xiongchun Duan <duanxiongchun@bytedance.com>
Cc: Christian Borntraeger <borntraeger@de.ibm.com>
[songmuchun@bytedance.com: fix forget to obtain the ref to objcg in split_page_memcg]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:52 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-12-01 13:58:30 -08:00
|
|
|
|
2017-09-08 16:11:50 -07:00
|
|
|
|
mm/thp: fix deferred split unqueue naming and locking
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.
Before fixing locking: rename misleading folio_undo_large_rmappable(),
which does not undo large_rmappable, to folio_unqueue_deferred_split(),
which is what it does. But that and its out-of-line __callee are mm
internals of very limited usability: add comment and WARN_ON_ONCEs to
check usage; and return a bool to say if a deferred split was unqueued,
which can then be used in WARN_ON_ONCEs around safety checks (sparing
callers the arcane conditionals in __folio_unqueue_deferred_split()).
Just omit the folio_unqueue_deferred_split() from free_unref_folios(), all
of whose callers now call it beforehand (and if any forget then bad_page()
will tell) - except for its caller put_pages_list(), which itself no
longer has any callers (and will be deleted separately).
Swapout: mem_cgroup_swapout() has been resetting folio->memcg_data 0
without checking and unqueueing a THP folio from deferred split list;
which is unfortunate, since the split_queue_lock depends on the memcg
(when memcg is enabled); so swapout has been unqueueing such THPs later,
when freeing the folio, using the pgdat's lock instead: potentially
corrupting the memcg's list. __remove_mapping() has frozen refcount to 0
here, so no problem with calling folio_unqueue_deferred_split() before
resetting memcg_data.
That goes back to 5.4 commit 87eaceb3faa5 ("mm: thp: make deferred split
shrinker memcg aware"): which included a check on swapcache before adding
to deferred queue, but no check on deferred queue before adding THP to
swapcache. That worked fine with the usual sequence of events in reclaim
(though there were a couple of rare ways in which a THP on deferred queue
could have been swapped out), but 6.12 commit dafff3f4c850 ("mm: split
underused THPs") avoids splitting underused THPs in reclaim, which makes
swapcache THPs on deferred queue commonplace.
Keep the check on swapcache before adding to deferred queue? Yes: it is
no longer essential, but preserves the existing behaviour, and is likely
to be a worthwhile optimization (vmstat showed much more traffic on the
queue under swapping load if the check was removed); update its comment.
Memcg-v1 move (deprecated): mem_cgroup_move_account() has been changing
folio->memcg_data without checking and unqueueing a THP folio from the
deferred list, sometimes corrupting "from" memcg's list, like swapout.
Refcount is non-zero here, so folio_unqueue_deferred_split() can only be
used in a WARN_ON_ONCE to validate the fix, which must be done earlier:
mem_cgroup_move_charge_pte_range() first try to split the THP (splitting
of course unqueues), or skip it if that fails. Not ideal, but moving
charge has been requested, and khugepaged should repair the THP later:
nobody wants new custom unqueueing code just for this deprecated case.
The 87eaceb3faa5 commit did have the code to move from one deferred list
to another (but was not conscious of its unsafety while refcount non-0);
but that was removed by 5.6 commit fac0516b5534 ("mm: thp: don't need care
deferred split queue in memcg charge move path"), which argued that the
existence of a PMD mapping guarantees that the THP cannot be on a deferred
list. As above, false in rare cases, and now commonly false.
Backport to 6.11 should be straightforward. Earlier backports must take
care that other _deferred_list fixes and dependencies are included. There
is not a strong case for backports, but they can fix cornercases.
Link: https://lkml.kernel.org/r/8dc111ae-f6db-2da7-b25c-7a20b1effe3b@google.com
Fixes: 87eaceb3faa5 ("mm: thp: make deferred split shrinker memcg aware")
Fixes: dafff3f4c850 ("mm: split underused THPs")
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: David Hildenbrand <david@redhat.com>
Reviewed-by: Yang Shi <shy828301@gmail.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.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: Usama Arif <usamaarif642@gmail.com>
Cc: Wei Yang <richard.weiyang@gmail.com>
Cc: Zi Yan <ziy@nvidia.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-10-27 13:02:13 -07:00
|
|
|
|
2021-06-29 21:47:12 -04:00
|
|
|
|
mm: memcontrol: use obj_cgroup APIs to charge kmem pages
Since Roman's series "The new cgroup slab memory controller" applied.
All slab objects are charged via the new APIs of obj_cgroup. The new
APIs introduce a struct obj_cgroup to charge slab objects. It prevents
long-living objects from pinning the original memory cgroup in the
memory. But there are still some corner objects (e.g. allocations
larger than order-1 page on SLUB) which are not charged via the new
APIs. Those objects (include the pages which are allocated from buddy
allocator directly) are charged as kmem pages which still hold a
reference to the memory cgroup.
We want to reuse the obj_cgroup APIs to charge the kmem pages. If we do
that, we should store an object cgroup pointer to page->memcg_data for
the kmem pages.
Finally, page->memcg_data will have 3 different meanings.
1) For the slab pages, page->memcg_data points to an object cgroups
vector.
2) For the kmem pages (exclude the slab pages), page->memcg_data
points to an object cgroup.
3) For the user pages (e.g. the LRU pages), page->memcg_data points
to a memory cgroup.
We do not change the behavior of page_memcg() and page_memcg_rcu(). They
are also suitable for LRU pages and kmem pages. Why?
Because memory allocations pinning memcgs for a long time - it exists at a
larger scale and is causing recurring problems in the real world: page
cache doesn't get reclaimed for a long time, or is used by the second,
third, fourth, ... instance of the same job that was restarted into a new
cgroup every time. Unreclaimable dying cgroups pile up, waste memory, and
make page reclaim very inefficient.
We can convert LRU pages and most other raw memcg pins to the objcg
direction to fix this problem, and then the page->memcg will always point
to an object cgroup pointer. At that time, LRU pages and kmem pages will
be treated the same. The implementation of page_memcg() will remove the
kmem page check.
This patch aims to charge the kmem pages by using the new APIs of
obj_cgroup. Finally, the page->memcg_data of the kmem page points to an
object cgroup. We can use the __page_objcg() to get the object cgroup
associated with a kmem page. Or we can use page_memcg() to get the memory
cgroup associated with a kmem page, but caller must ensure that the
returned memcg won't be released (e.g. acquire the rcu_read_lock or
css_set_lock).
Link: https://lkml.kernel.org/r/20210401030141.37061-1-songmuchun@bytedance.com
Link: https://lkml.kernel.org/r/20210319163821.20704-6-songmuchun@bytedance.com
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Roman Gushchin <guro@fb.com>
Reviewed-by: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Xiongchun Duan <duanxiongchun@bytedance.com>
Cc: Christian Borntraeger <borntraeger@de.ibm.com>
[songmuchun@bytedance.com: fix forget to obtain the ref to objcg in split_page_memcg]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-29 22:56:52 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-08-08 14:19:24 -07:00
|
|
|
|
|
|
|
|
|
2021-05-01 20:42:23 -04:00
|
|
|
|
mm: memcontrol: rewrite uncharge API
The memcg uncharging code that is involved towards the end of a page's
lifetime - truncation, reclaim, swapout, migration - is impressively
complicated and fragile.
Because anonymous and file pages were always charged before they had their
page->mapping established, uncharges had to happen when the page type
could still be known from the context; as in unmap for anonymous, page
cache removal for file and shmem pages, and swap cache truncation for swap
pages. However, these operations happen well before the page is actually
freed, and so a lot of synchronization is necessary:
- Charging, uncharging, page migration, and charge migration all need
to take a per-page bit spinlock as they could race with uncharging.
- Swap cache truncation happens during both swap-in and swap-out, and
possibly repeatedly before the page is actually freed. This means
that the memcg swapout code is called from many contexts that make
no sense and it has to figure out the direction from page state to
make sure memory and memory+swap are always correctly charged.
- On page migration, the old page might be unmapped but then reused,
so memcg code has to prevent untimely uncharging in that case.
Because this code - which should be a simple charge transfer - is so
special-cased, it is not reusable for replace_page_cache().
But now that charged pages always have a page->mapping, introduce
mem_cgroup_uncharge(), which is called after the final put_page(), when we
know for sure that nobody is looking at the page anymore.
For page migration, introduce mem_cgroup_migrate(), which is called after
the migration is successful and the new page is fully rmapped. Because
the old page is no longer uncharged after migration, prevent double
charges by decoupling the page's memcg association (PCG_USED and
pc->mem_cgroup) from the page holding an actual charge. The new bits
PCG_MEM and PCG_MEMSW represent the respective charges and are transferred
to the new page during migration.
mem_cgroup_migrate() is suitable for replace_page_cache() as well,
which gets rid of mem_cgroup_replace_page_cache(). However, care
needs to be taken because both the source and the target page can
already be charged and on the LRU when fuse is splicing: grab the page
lock on the charge moving side to prevent changing pc->mem_cgroup of a
page under migration. Also, the lruvecs of both pages change as we
uncharge the old and charge the new during migration, and putback may
race with us, so grab the lru lock and isolate the pages iff on LRU to
prevent races and ensure the pages are on the right lruvec afterward.
Swap accounting is massively simplified: because the page is no longer
uncharged as early as swap cache deletion, a new mem_cgroup_swapout() can
transfer the page's memory+swap charge (PCG_MEMSW) to the swap entry
before the final put_page() in page reclaim.
Finally, page_cgroup changes are now protected by whatever protection the
page itself offers: anonymous pages are charged under the page table lock,
whereas page cache insertions, swapin, and migration hold the page lock.
Uncharging happens under full exclusion with no outstanding references.
Charging and uncharging also ensure that the page is off-LRU, which
serializes against charge migration. Remove the very costly page_cgroup
lock and set pc->flags non-atomically.
[mhocko@suse.cz: mem_cgroup_charge_statistics needs preempt_disable]
[vdavydov@parallels.com: fix flags definition]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Tested-by: Jet Chen <jet.chen@intel.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Tested-by: Felipe Balbi <balbi@ti.com>
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-08-08 14:19:22 -07:00
|
|
|
|
2017-09-08 16:11:50 -07:00
|
|
|
|
|
|
|
|
|
2021-05-01 20:42:23 -04:00
|
|
|
|
2024-08-14 17:34:15 +08:00
|
|
|
|
mm: memcontrol: rewrite uncharge API
The memcg uncharging code that is involved towards the end of a page's
lifetime - truncation, reclaim, swapout, migration - is impressively
complicated and fragile.
Because anonymous and file pages were always charged before they had their
page->mapping established, uncharges had to happen when the page type
could still be known from the context; as in unmap for anonymous, page
cache removal for file and shmem pages, and swap cache truncation for swap
pages. However, these operations happen well before the page is actually
freed, and so a lot of synchronization is necessary:
- Charging, uncharging, page migration, and charge migration all need
to take a per-page bit spinlock as they could race with uncharging.
- Swap cache truncation happens during both swap-in and swap-out, and
possibly repeatedly before the page is actually freed. This means
that the memcg swapout code is called from many contexts that make
no sense and it has to figure out the direction from page state to
make sure memory and memory+swap are always correctly charged.
- On page migration, the old page might be unmapped but then reused,
so memcg code has to prevent untimely uncharging in that case.
Because this code - which should be a simple charge transfer - is so
special-cased, it is not reusable for replace_page_cache().
But now that charged pages always have a page->mapping, introduce
mem_cgroup_uncharge(), which is called after the final put_page(), when we
know for sure that nobody is looking at the page anymore.
For page migration, introduce mem_cgroup_migrate(), which is called after
the migration is successful and the new page is fully rmapped. Because
the old page is no longer uncharged after migration, prevent double
charges by decoupling the page's memcg association (PCG_USED and
pc->mem_cgroup) from the page holding an actual charge. The new bits
PCG_MEM and PCG_MEMSW represent the respective charges and are transferred
to the new page during migration.
mem_cgroup_migrate() is suitable for replace_page_cache() as well,
which gets rid of mem_cgroup_replace_page_cache(). However, care
needs to be taken because both the source and the target page can
already be charged and on the LRU when fuse is splicing: grab the page
lock on the charge moving side to prevent changing pc->mem_cgroup of a
page under migration. Also, the lruvecs of both pages change as we
uncharge the old and charge the new during migration, and putback may
race with us, so grab the lru lock and isolate the pages iff on LRU to
prevent races and ensure the pages are on the right lruvec afterward.
Swap accounting is massively simplified: because the page is no longer
uncharged as early as swap cache deletion, a new mem_cgroup_swapout() can
transfer the page's memory+swap charge (PCG_MEMSW) to the swap entry
before the final put_page() in page reclaim.
Finally, page_cgroup changes are now protected by whatever protection the
page itself offers: anonymous pages are charged under the page table lock,
whereas page cache insertions, swapin, and migration hold the page lock.
Uncharging happens under full exclusion with no outstanding references.
Charging and uncharging also ensure that the page is off-LRU, which
serializes against charge migration. Remove the very costly page_cgroup
lock and set pc->flags non-atomically.
[mhocko@suse.cz: mem_cgroup_charge_statistics needs preempt_disable]
[vdavydov@parallels.com: fix flags definition]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Tested-by: Jet Chen <jet.chen@intel.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Tested-by: Felipe Balbi <balbi@ti.com>
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-08-08 14:19:22 -07:00
|
|
|
|
|
|
|
|
|
2017-09-08 16:11:50 -07:00
|
|
|
|
2021-05-01 20:42:23 -04:00
|
|
|
|
2017-09-08 16:11:50 -07:00
|
|
|
|
2014-08-08 14:19:24 -07:00
|
|
|
|
mm: memcontrol: rewrite uncharge API
The memcg uncharging code that is involved towards the end of a page's
lifetime - truncation, reclaim, swapout, migration - is impressively
complicated and fragile.
Because anonymous and file pages were always charged before they had their
page->mapping established, uncharges had to happen when the page type
could still be known from the context; as in unmap for anonymous, page
cache removal for file and shmem pages, and swap cache truncation for swap
pages. However, these operations happen well before the page is actually
freed, and so a lot of synchronization is necessary:
- Charging, uncharging, page migration, and charge migration all need
to take a per-page bit spinlock as they could race with uncharging.
- Swap cache truncation happens during both swap-in and swap-out, and
possibly repeatedly before the page is actually freed. This means
that the memcg swapout code is called from many contexts that make
no sense and it has to figure out the direction from page state to
make sure memory and memory+swap are always correctly charged.
- On page migration, the old page might be unmapped but then reused,
so memcg code has to prevent untimely uncharging in that case.
Because this code - which should be a simple charge transfer - is so
special-cased, it is not reusable for replace_page_cache().
But now that charged pages always have a page->mapping, introduce
mem_cgroup_uncharge(), which is called after the final put_page(), when we
know for sure that nobody is looking at the page anymore.
For page migration, introduce mem_cgroup_migrate(), which is called after
the migration is successful and the new page is fully rmapped. Because
the old page is no longer uncharged after migration, prevent double
charges by decoupling the page's memcg association (PCG_USED and
pc->mem_cgroup) from the page holding an actual charge. The new bits
PCG_MEM and PCG_MEMSW represent the respective charges and are transferred
to the new page during migration.
mem_cgroup_migrate() is suitable for replace_page_cache() as well,
which gets rid of mem_cgroup_replace_page_cache(). However, care
needs to be taken because both the source and the target page can
already be charged and on the LRU when fuse is splicing: grab the page
lock on the charge moving side to prevent changing pc->mem_cgroup of a
page under migration. Also, the lruvecs of both pages change as we
uncharge the old and charge the new during migration, and putback may
race with us, so grab the lru lock and isolate the pages iff on LRU to
prevent races and ensure the pages are on the right lruvec afterward.
Swap accounting is massively simplified: because the page is no longer
uncharged as early as swap cache deletion, a new mem_cgroup_swapout() can
transfer the page's memory+swap charge (PCG_MEMSW) to the swap entry
before the final put_page() in page reclaim.
Finally, page_cgroup changes are now protected by whatever protection the
page itself offers: anonymous pages are charged under the page table lock,
whereas page cache insertions, swapin, and migration hold the page lock.
Uncharging happens under full exclusion with no outstanding references.
Charging and uncharging also ensure that the page is off-LRU, which
serializes against charge migration. Remove the very costly page_cgroup
lock and set pc->flags non-atomically.
[mhocko@suse.cz: mem_cgroup_charge_statistics needs preempt_disable]
[vdavydov@parallels.com: fix flags definition]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Tested-by: Jet Chen <jet.chen@intel.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Tested-by: Felipe Balbi <balbi@ti.com>
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-08-08 14:19:22 -07:00
|
|
|
|
2024-02-27 17:42:39 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcontrol: rewrite uncharge API
The memcg uncharging code that is involved towards the end of a page's
lifetime - truncation, reclaim, swapout, migration - is impressively
complicated and fragile.
Because anonymous and file pages were always charged before they had their
page->mapping established, uncharges had to happen when the page type
could still be known from the context; as in unmap for anonymous, page
cache removal for file and shmem pages, and swap cache truncation for swap
pages. However, these operations happen well before the page is actually
freed, and so a lot of synchronization is necessary:
- Charging, uncharging, page migration, and charge migration all need
to take a per-page bit spinlock as they could race with uncharging.
- Swap cache truncation happens during both swap-in and swap-out, and
possibly repeatedly before the page is actually freed. This means
that the memcg swapout code is called from many contexts that make
no sense and it has to figure out the direction from page state to
make sure memory and memory+swap are always correctly charged.
- On page migration, the old page might be unmapped but then reused,
so memcg code has to prevent untimely uncharging in that case.
Because this code - which should be a simple charge transfer - is so
special-cased, it is not reusable for replace_page_cache().
But now that charged pages always have a page->mapping, introduce
mem_cgroup_uncharge(), which is called after the final put_page(), when we
know for sure that nobody is looking at the page anymore.
For page migration, introduce mem_cgroup_migrate(), which is called after
the migration is successful and the new page is fully rmapped. Because
the old page is no longer uncharged after migration, prevent double
charges by decoupling the page's memcg association (PCG_USED and
pc->mem_cgroup) from the page holding an actual charge. The new bits
PCG_MEM and PCG_MEMSW represent the respective charges and are transferred
to the new page during migration.
mem_cgroup_migrate() is suitable for replace_page_cache() as well,
which gets rid of mem_cgroup_replace_page_cache(). However, care
needs to be taken because both the source and the target page can
already be charged and on the LRU when fuse is splicing: grab the page
lock on the charge moving side to prevent changing pc->mem_cgroup of a
page under migration. Also, the lruvecs of both pages change as we
uncharge the old and charge the new during migration, and putback may
race with us, so grab the lru lock and isolate the pages iff on LRU to
prevent races and ensure the pages are on the right lruvec afterward.
Swap accounting is massively simplified: because the page is no longer
uncharged as early as swap cache deletion, a new mem_cgroup_swapout() can
transfer the page's memory+swap charge (PCG_MEMSW) to the swap entry
before the final put_page() in page reclaim.
Finally, page_cgroup changes are now protected by whatever protection the
page itself offers: anonymous pages are charged under the page table lock,
whereas page cache insertions, swapin, and migration hold the page lock.
Uncharging happens under full exclusion with no outstanding references.
Charging and uncharging also ensure that the page is off-LRU, which
serializes against charge migration. Remove the very costly page_cgroup
lock and set pc->flags non-atomically.
[mhocko@suse.cz: mem_cgroup_charge_statistics needs preempt_disable]
[vdavydov@parallels.com: fix flags definition]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Tested-by: Jet Chen <jet.chen@intel.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Tested-by: Felipe Balbi <balbi@ti.com>
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-08-08 14:19:22 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-06 11:46:27 -07:00
|
|
|
|
2021-05-06 18:14:59 -04:00
|
|
|
|
|
|
|
|
|
mm: memcontrol: rewrite uncharge API
The memcg uncharging code that is involved towards the end of a page's
lifetime - truncation, reclaim, swapout, migration - is impressively
complicated and fragile.
Because anonymous and file pages were always charged before they had their
page->mapping established, uncharges had to happen when the page type
could still be known from the context; as in unmap for anonymous, page
cache removal for file and shmem pages, and swap cache truncation for swap
pages. However, these operations happen well before the page is actually
freed, and so a lot of synchronization is necessary:
- Charging, uncharging, page migration, and charge migration all need
to take a per-page bit spinlock as they could race with uncharging.
- Swap cache truncation happens during both swap-in and swap-out, and
possibly repeatedly before the page is actually freed. This means
that the memcg swapout code is called from many contexts that make
no sense and it has to figure out the direction from page state to
make sure memory and memory+swap are always correctly charged.
- On page migration, the old page might be unmapped but then reused,
so memcg code has to prevent untimely uncharging in that case.
Because this code - which should be a simple charge transfer - is so
special-cased, it is not reusable for replace_page_cache().
But now that charged pages always have a page->mapping, introduce
mem_cgroup_uncharge(), which is called after the final put_page(), when we
know for sure that nobody is looking at the page anymore.
For page migration, introduce mem_cgroup_migrate(), which is called after
the migration is successful and the new page is fully rmapped. Because
the old page is no longer uncharged after migration, prevent double
charges by decoupling the page's memcg association (PCG_USED and
pc->mem_cgroup) from the page holding an actual charge. The new bits
PCG_MEM and PCG_MEMSW represent the respective charges and are transferred
to the new page during migration.
mem_cgroup_migrate() is suitable for replace_page_cache() as well,
which gets rid of mem_cgroup_replace_page_cache(). However, care
needs to be taken because both the source and the target page can
already be charged and on the LRU when fuse is splicing: grab the page
lock on the charge moving side to prevent changing pc->mem_cgroup of a
page under migration. Also, the lruvecs of both pages change as we
uncharge the old and charge the new during migration, and putback may
race with us, so grab the lru lock and isolate the pages iff on LRU to
prevent races and ensure the pages are on the right lruvec afterward.
Swap accounting is massively simplified: because the page is no longer
uncharged as early as swap cache deletion, a new mem_cgroup_swapout() can
transfer the page's memory+swap charge (PCG_MEMSW) to the swap entry
before the final put_page() in page reclaim.
Finally, page_cgroup changes are now protected by whatever protection the
page itself offers: anonymous pages are charged under the page table lock,
whereas page cache insertions, swapin, and migration hold the page lock.
Uncharging happens under full exclusion with no outstanding references.
Charging and uncharging also ensure that the page is off-LRU, which
serializes against charge migration. Remove the very costly page_cgroup
lock and set pc->flags non-atomically.
[mhocko@suse.cz: mem_cgroup_charge_statistics needs preempt_disable]
[vdavydov@parallels.com: fix flags definition]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Tested-by: Jet Chen <jet.chen@intel.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Tested-by: Felipe Balbi <balbi@ti.com>
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-08-08 14:19:22 -07:00
|
|
|
|
2021-05-06 18:14:59 -04:00
|
|
|
|
2024-06-13 16:21:19 +08:00
|
|
|
|
mm: memcontrol: rewrite uncharge API
The memcg uncharging code that is involved towards the end of a page's
lifetime - truncation, reclaim, swapout, migration - is impressively
complicated and fragile.
Because anonymous and file pages were always charged before they had their
page->mapping established, uncharges had to happen when the page type
could still be known from the context; as in unmap for anonymous, page
cache removal for file and shmem pages, and swap cache truncation for swap
pages. However, these operations happen well before the page is actually
freed, and so a lot of synchronization is necessary:
- Charging, uncharging, page migration, and charge migration all need
to take a per-page bit spinlock as they could race with uncharging.
- Swap cache truncation happens during both swap-in and swap-out, and
possibly repeatedly before the page is actually freed. This means
that the memcg swapout code is called from many contexts that make
no sense and it has to figure out the direction from page state to
make sure memory and memory+swap are always correctly charged.
- On page migration, the old page might be unmapped but then reused,
so memcg code has to prevent untimely uncharging in that case.
Because this code - which should be a simple charge transfer - is so
special-cased, it is not reusable for replace_page_cache().
But now that charged pages always have a page->mapping, introduce
mem_cgroup_uncharge(), which is called after the final put_page(), when we
know for sure that nobody is looking at the page anymore.
For page migration, introduce mem_cgroup_migrate(), which is called after
the migration is successful and the new page is fully rmapped. Because
the old page is no longer uncharged after migration, prevent double
charges by decoupling the page's memcg association (PCG_USED and
pc->mem_cgroup) from the page holding an actual charge. The new bits
PCG_MEM and PCG_MEMSW represent the respective charges and are transferred
to the new page during migration.
mem_cgroup_migrate() is suitable for replace_page_cache() as well,
which gets rid of mem_cgroup_replace_page_cache(). However, care
needs to be taken because both the source and the target page can
already be charged and on the LRU when fuse is splicing: grab the page
lock on the charge moving side to prevent changing pc->mem_cgroup of a
page under migration. Also, the lruvecs of both pages change as we
uncharge the old and charge the new during migration, and putback may
race with us, so grab the lru lock and isolate the pages iff on LRU to
prevent races and ensure the pages are on the right lruvec afterward.
Swap accounting is massively simplified: because the page is no longer
uncharged as early as swap cache deletion, a new mem_cgroup_swapout() can
transfer the page's memory+swap charge (PCG_MEMSW) to the swap entry
before the final put_page() in page reclaim.
Finally, page_cgroup changes are now protected by whatever protection the
page itself offers: anonymous pages are charged under the page table lock,
whereas page cache insertions, swapin, and migration hold the page lock.
Uncharging happens under full exclusion with no outstanding references.
Charging and uncharging also ensure that the page is off-LRU, which
serializes against charge migration. Remove the very costly page_cgroup
lock and set pc->flags non-atomically.
[mhocko@suse.cz: mem_cgroup_charge_statistics needs preempt_disable]
[vdavydov@parallels.com: fix flags definition]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Tested-by: Jet Chen <jet.chen@intel.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Tested-by: Felipe Balbi <balbi@ti.com>
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-08-08 14:19:22 -07:00
|
|
|
|
2021-05-06 18:14:59 -04:00
|
|
|
|
mm: memcontrol: rewrite uncharge API
The memcg uncharging code that is involved towards the end of a page's
lifetime - truncation, reclaim, swapout, migration - is impressively
complicated and fragile.
Because anonymous and file pages were always charged before they had their
page->mapping established, uncharges had to happen when the page type
could still be known from the context; as in unmap for anonymous, page
cache removal for file and shmem pages, and swap cache truncation for swap
pages. However, these operations happen well before the page is actually
freed, and so a lot of synchronization is necessary:
- Charging, uncharging, page migration, and charge migration all need
to take a per-page bit spinlock as they could race with uncharging.
- Swap cache truncation happens during both swap-in and swap-out, and
possibly repeatedly before the page is actually freed. This means
that the memcg swapout code is called from many contexts that make
no sense and it has to figure out the direction from page state to
make sure memory and memory+swap are always correctly charged.
- On page migration, the old page might be unmapped but then reused,
so memcg code has to prevent untimely uncharging in that case.
Because this code - which should be a simple charge transfer - is so
special-cased, it is not reusable for replace_page_cache().
But now that charged pages always have a page->mapping, introduce
mem_cgroup_uncharge(), which is called after the final put_page(), when we
know for sure that nobody is looking at the page anymore.
For page migration, introduce mem_cgroup_migrate(), which is called after
the migration is successful and the new page is fully rmapped. Because
the old page is no longer uncharged after migration, prevent double
charges by decoupling the page's memcg association (PCG_USED and
pc->mem_cgroup) from the page holding an actual charge. The new bits
PCG_MEM and PCG_MEMSW represent the respective charges and are transferred
to the new page during migration.
mem_cgroup_migrate() is suitable for replace_page_cache() as well,
which gets rid of mem_cgroup_replace_page_cache(). However, care
needs to be taken because both the source and the target page can
already be charged and on the LRU when fuse is splicing: grab the page
lock on the charge moving side to prevent changing pc->mem_cgroup of a
page under migration. Also, the lruvecs of both pages change as we
uncharge the old and charge the new during migration, and putback may
race with us, so grab the lru lock and isolate the pages iff on LRU to
prevent races and ensure the pages are on the right lruvec afterward.
Swap accounting is massively simplified: because the page is no longer
uncharged as early as swap cache deletion, a new mem_cgroup_swapout() can
transfer the page's memory+swap charge (PCG_MEMSW) to the swap entry
before the final put_page() in page reclaim.
Finally, page_cgroup changes are now protected by whatever protection the
page itself offers: anonymous pages are charged under the page table lock,
whereas page cache insertions, swapin, and migration hold the page lock.
Uncharging happens under full exclusion with no outstanding references.
Charging and uncharging also ensure that the page is off-LRU, which
serializes against charge migration. Remove the very costly page_cgroup
lock and set pc->flags non-atomically.
[mhocko@suse.cz: mem_cgroup_charge_statistics needs preempt_disable]
[vdavydov@parallels.com: fix flags definition]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Tested-by: Jet Chen <jet.chen@intel.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Tested-by: Felipe Balbi <balbi@ti.com>
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-08-08 14:19:22 -07:00
|
|
|
|
2023-10-06 11:46:27 -07:00
|
|
|
|
mm: memcontrol: rewrite uncharge API
The memcg uncharging code that is involved towards the end of a page's
lifetime - truncation, reclaim, swapout, migration - is impressively
complicated and fragile.
Because anonymous and file pages were always charged before they had their
page->mapping established, uncharges had to happen when the page type
could still be known from the context; as in unmap for anonymous, page
cache removal for file and shmem pages, and swap cache truncation for swap
pages. However, these operations happen well before the page is actually
freed, and so a lot of synchronization is necessary:
- Charging, uncharging, page migration, and charge migration all need
to take a per-page bit spinlock as they could race with uncharging.
- Swap cache truncation happens during both swap-in and swap-out, and
possibly repeatedly before the page is actually freed. This means
that the memcg swapout code is called from many contexts that make
no sense and it has to figure out the direction from page state to
make sure memory and memory+swap are always correctly charged.
- On page migration, the old page might be unmapped but then reused,
so memcg code has to prevent untimely uncharging in that case.
Because this code - which should be a simple charge transfer - is so
special-cased, it is not reusable for replace_page_cache().
But now that charged pages always have a page->mapping, introduce
mem_cgroup_uncharge(), which is called after the final put_page(), when we
know for sure that nobody is looking at the page anymore.
For page migration, introduce mem_cgroup_migrate(), which is called after
the migration is successful and the new page is fully rmapped. Because
the old page is no longer uncharged after migration, prevent double
charges by decoupling the page's memcg association (PCG_USED and
pc->mem_cgroup) from the page holding an actual charge. The new bits
PCG_MEM and PCG_MEMSW represent the respective charges and are transferred
to the new page during migration.
mem_cgroup_migrate() is suitable for replace_page_cache() as well,
which gets rid of mem_cgroup_replace_page_cache(). However, care
needs to be taken because both the source and the target page can
already be charged and on the LRU when fuse is splicing: grab the page
lock on the charge moving side to prevent changing pc->mem_cgroup of a
page under migration. Also, the lruvecs of both pages change as we
uncharge the old and charge the new during migration, and putback may
race with us, so grab the lru lock and isolate the pages iff on LRU to
prevent races and ensure the pages are on the right lruvec afterward.
Swap accounting is massively simplified: because the page is no longer
uncharged as early as swap cache deletion, a new mem_cgroup_swapout() can
transfer the page's memory+swap charge (PCG_MEMSW) to the swap entry
before the final put_page() in page reclaim.
Finally, page_cgroup changes are now protected by whatever protection the
page itself offers: anonymous pages are charged under the page table lock,
whereas page cache insertions, swapin, and migration hold the page lock.
Uncharging happens under full exclusion with no outstanding references.
Charging and uncharging also ensure that the page is off-LRU, which
serializes against charge migration. Remove the very costly page_cgroup
lock and set pc->flags non-atomically.
[mhocko@suse.cz: mem_cgroup_charge_statistics needs preempt_disable]
[vdavydov@parallels.com: fix flags definition]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Tested-by: Jet Chen <jet.chen@intel.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Tested-by: Felipe Balbi <balbi@ti.com>
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-08-08 14:19:22 -07:00
|
|
|
|
2014-12-10 15:44:02 -08:00
|
|
|
|
2021-05-06 18:14:59 -04:00
|
|
|
|
mm: memcontrol: rewrite uncharge API
The memcg uncharging code that is involved towards the end of a page's
lifetime - truncation, reclaim, swapout, migration - is impressively
complicated and fragile.
Because anonymous and file pages were always charged before they had their
page->mapping established, uncharges had to happen when the page type
could still be known from the context; as in unmap for anonymous, page
cache removal for file and shmem pages, and swap cache truncation for swap
pages. However, these operations happen well before the page is actually
freed, and so a lot of synchronization is necessary:
- Charging, uncharging, page migration, and charge migration all need
to take a per-page bit spinlock as they could race with uncharging.
- Swap cache truncation happens during both swap-in and swap-out, and
possibly repeatedly before the page is actually freed. This means
that the memcg swapout code is called from many contexts that make
no sense and it has to figure out the direction from page state to
make sure memory and memory+swap are always correctly charged.
- On page migration, the old page might be unmapped but then reused,
so memcg code has to prevent untimely uncharging in that case.
Because this code - which should be a simple charge transfer - is so
special-cased, it is not reusable for replace_page_cache().
But now that charged pages always have a page->mapping, introduce
mem_cgroup_uncharge(), which is called after the final put_page(), when we
know for sure that nobody is looking at the page anymore.
For page migration, introduce mem_cgroup_migrate(), which is called after
the migration is successful and the new page is fully rmapped. Because
the old page is no longer uncharged after migration, prevent double
charges by decoupling the page's memcg association (PCG_USED and
pc->mem_cgroup) from the page holding an actual charge. The new bits
PCG_MEM and PCG_MEMSW represent the respective charges and are transferred
to the new page during migration.
mem_cgroup_migrate() is suitable for replace_page_cache() as well,
which gets rid of mem_cgroup_replace_page_cache(). However, care
needs to be taken because both the source and the target page can
already be charged and on the LRU when fuse is splicing: grab the page
lock on the charge moving side to prevent changing pc->mem_cgroup of a
page under migration. Also, the lruvecs of both pages change as we
uncharge the old and charge the new during migration, and putback may
race with us, so grab the lru lock and isolate the pages iff on LRU to
prevent races and ensure the pages are on the right lruvec afterward.
Swap accounting is massively simplified: because the page is no longer
uncharged as early as swap cache deletion, a new mem_cgroup_swapout() can
transfer the page's memory+swap charge (PCG_MEMSW) to the swap entry
before the final put_page() in page reclaim.
Finally, page_cgroup changes are now protected by whatever protection the
page itself offers: anonymous pages are charged under the page table lock,
whereas page cache insertions, swapin, and migration hold the page lock.
Uncharging happens under full exclusion with no outstanding references.
Charging and uncharging also ensure that the page is off-LRU, which
serializes against charge migration. Remove the very costly page_cgroup
lock and set pc->flags non-atomically.
[mhocko@suse.cz: mem_cgroup_charge_statistics needs preempt_disable]
[vdavydov@parallels.com: fix flags definition]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Tested-by: Jet Chen <jet.chen@intel.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Tested-by: Felipe Balbi <balbi@ti.com>
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-08-08 14:19:22 -07:00
|
|
|
|
2021-05-06 18:14:59 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcontrol: rewrite uncharge API
The memcg uncharging code that is involved towards the end of a page's
lifetime - truncation, reclaim, swapout, migration - is impressively
complicated and fragile.
Because anonymous and file pages were always charged before they had their
page->mapping established, uncharges had to happen when the page type
could still be known from the context; as in unmap for anonymous, page
cache removal for file and shmem pages, and swap cache truncation for swap
pages. However, these operations happen well before the page is actually
freed, and so a lot of synchronization is necessary:
- Charging, uncharging, page migration, and charge migration all need
to take a per-page bit spinlock as they could race with uncharging.
- Swap cache truncation happens during both swap-in and swap-out, and
possibly repeatedly before the page is actually freed. This means
that the memcg swapout code is called from many contexts that make
no sense and it has to figure out the direction from page state to
make sure memory and memory+swap are always correctly charged.
- On page migration, the old page might be unmapped but then reused,
so memcg code has to prevent untimely uncharging in that case.
Because this code - which should be a simple charge transfer - is so
special-cased, it is not reusable for replace_page_cache().
But now that charged pages always have a page->mapping, introduce
mem_cgroup_uncharge(), which is called after the final put_page(), when we
know for sure that nobody is looking at the page anymore.
For page migration, introduce mem_cgroup_migrate(), which is called after
the migration is successful and the new page is fully rmapped. Because
the old page is no longer uncharged after migration, prevent double
charges by decoupling the page's memcg association (PCG_USED and
pc->mem_cgroup) from the page holding an actual charge. The new bits
PCG_MEM and PCG_MEMSW represent the respective charges and are transferred
to the new page during migration.
mem_cgroup_migrate() is suitable for replace_page_cache() as well,
which gets rid of mem_cgroup_replace_page_cache(). However, care
needs to be taken because both the source and the target page can
already be charged and on the LRU when fuse is splicing: grab the page
lock on the charge moving side to prevent changing pc->mem_cgroup of a
page under migration. Also, the lruvecs of both pages change as we
uncharge the old and charge the new during migration, and putback may
race with us, so grab the lru lock and isolate the pages iff on LRU to
prevent races and ensure the pages are on the right lruvec afterward.
Swap accounting is massively simplified: because the page is no longer
uncharged as early as swap cache deletion, a new mem_cgroup_swapout() can
transfer the page's memory+swap charge (PCG_MEMSW) to the swap entry
before the final put_page() in page reclaim.
Finally, page_cgroup changes are now protected by whatever protection the
page itself offers: anonymous pages are charged under the page table lock,
whereas page cache insertions, swapin, and migration hold the page lock.
Uncharging happens under full exclusion with no outstanding references.
Charging and uncharging also ensure that the page is off-LRU, which
serializes against charge migration. Remove the very costly page_cgroup
lock and set pc->flags non-atomically.
[mhocko@suse.cz: mem_cgroup_charge_statistics needs preempt_disable]
[vdavydov@parallels.com: fix flags definition]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Tested-by: Jet Chen <jet.chen@intel.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Tested-by: Felipe Balbi <balbi@ti.com>
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-08-08 14:19:22 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-05-06 18:14:59 -04:00
|
|
|
|
2024-08-14 17:34:15 +08:00
|
|
|
|
mm: memcontrol: rewrite uncharge API
The memcg uncharging code that is involved towards the end of a page's
lifetime - truncation, reclaim, swapout, migration - is impressively
complicated and fragile.
Because anonymous and file pages were always charged before they had their
page->mapping established, uncharges had to happen when the page type
could still be known from the context; as in unmap for anonymous, page
cache removal for file and shmem pages, and swap cache truncation for swap
pages. However, these operations happen well before the page is actually
freed, and so a lot of synchronization is necessary:
- Charging, uncharging, page migration, and charge migration all need
to take a per-page bit spinlock as they could race with uncharging.
- Swap cache truncation happens during both swap-in and swap-out, and
possibly repeatedly before the page is actually freed. This means
that the memcg swapout code is called from many contexts that make
no sense and it has to figure out the direction from page state to
make sure memory and memory+swap are always correctly charged.
- On page migration, the old page might be unmapped but then reused,
so memcg code has to prevent untimely uncharging in that case.
Because this code - which should be a simple charge transfer - is so
special-cased, it is not reusable for replace_page_cache().
But now that charged pages always have a page->mapping, introduce
mem_cgroup_uncharge(), which is called after the final put_page(), when we
know for sure that nobody is looking at the page anymore.
For page migration, introduce mem_cgroup_migrate(), which is called after
the migration is successful and the new page is fully rmapped. Because
the old page is no longer uncharged after migration, prevent double
charges by decoupling the page's memcg association (PCG_USED and
pc->mem_cgroup) from the page holding an actual charge. The new bits
PCG_MEM and PCG_MEMSW represent the respective charges and are transferred
to the new page during migration.
mem_cgroup_migrate() is suitable for replace_page_cache() as well,
which gets rid of mem_cgroup_replace_page_cache(). However, care
needs to be taken because both the source and the target page can
already be charged and on the LRU when fuse is splicing: grab the page
lock on the charge moving side to prevent changing pc->mem_cgroup of a
page under migration. Also, the lruvecs of both pages change as we
uncharge the old and charge the new during migration, and putback may
race with us, so grab the lru lock and isolate the pages iff on LRU to
prevent races and ensure the pages are on the right lruvec afterward.
Swap accounting is massively simplified: because the page is no longer
uncharged as early as swap cache deletion, a new mem_cgroup_swapout() can
transfer the page's memory+swap charge (PCG_MEMSW) to the swap entry
before the final put_page() in page reclaim.
Finally, page_cgroup changes are now protected by whatever protection the
page itself offers: anonymous pages are charged under the page table lock,
whereas page cache insertions, swapin, and migration hold the page lock.
Uncharging happens under full exclusion with no outstanding references.
Charging and uncharging also ensure that the page is off-LRU, which
serializes against charge migration. Remove the very costly page_cgroup
lock and set pc->flags non-atomically.
[mhocko@suse.cz: mem_cgroup_charge_statistics needs preempt_disable]
[vdavydov@parallels.com: fix flags definition]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Tested-by: Jet Chen <jet.chen@intel.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Tested-by: Felipe Balbi <balbi@ti.com>
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-08-08 14:19:22 -07:00
|
|
|
|
|
|
|
|
|
2021-05-06 18:14:59 -04:00
|
|
|
|
|
|
|
|
|
2014-12-10 15:44:02 -08:00
|
|
|
|
mm: memcontrol: rewrite uncharge API
The memcg uncharging code that is involved towards the end of a page's
lifetime - truncation, reclaim, swapout, migration - is impressively
complicated and fragile.
Because anonymous and file pages were always charged before they had their
page->mapping established, uncharges had to happen when the page type
could still be known from the context; as in unmap for anonymous, page
cache removal for file and shmem pages, and swap cache truncation for swap
pages. However, these operations happen well before the page is actually
freed, and so a lot of synchronization is necessary:
- Charging, uncharging, page migration, and charge migration all need
to take a per-page bit spinlock as they could race with uncharging.
- Swap cache truncation happens during both swap-in and swap-out, and
possibly repeatedly before the page is actually freed. This means
that the memcg swapout code is called from many contexts that make
no sense and it has to figure out the direction from page state to
make sure memory and memory+swap are always correctly charged.
- On page migration, the old page might be unmapped but then reused,
so memcg code has to prevent untimely uncharging in that case.
Because this code - which should be a simple charge transfer - is so
special-cased, it is not reusable for replace_page_cache().
But now that charged pages always have a page->mapping, introduce
mem_cgroup_uncharge(), which is called after the final put_page(), when we
know for sure that nobody is looking at the page anymore.
For page migration, introduce mem_cgroup_migrate(), which is called after
the migration is successful and the new page is fully rmapped. Because
the old page is no longer uncharged after migration, prevent double
charges by decoupling the page's memcg association (PCG_USED and
pc->mem_cgroup) from the page holding an actual charge. The new bits
PCG_MEM and PCG_MEMSW represent the respective charges and are transferred
to the new page during migration.
mem_cgroup_migrate() is suitable for replace_page_cache() as well,
which gets rid of mem_cgroup_replace_page_cache(). However, care
needs to be taken because both the source and the target page can
already be charged and on the LRU when fuse is splicing: grab the page
lock on the charge moving side to prevent changing pc->mem_cgroup of a
page under migration. Also, the lruvecs of both pages change as we
uncharge the old and charge the new during migration, and putback may
race with us, so grab the lru lock and isolate the pages iff on LRU to
prevent races and ensure the pages are on the right lruvec afterward.
Swap accounting is massively simplified: because the page is no longer
uncharged as early as swap cache deletion, a new mem_cgroup_swapout() can
transfer the page's memory+swap charge (PCG_MEMSW) to the swap entry
before the final put_page() in page reclaim.
Finally, page_cgroup changes are now protected by whatever protection the
page itself offers: anonymous pages are charged under the page table lock,
whereas page cache insertions, swapin, and migration hold the page lock.
Uncharging happens under full exclusion with no outstanding references.
Charging and uncharging also ensure that the page is off-LRU, which
serializes against charge migration. Remove the very costly page_cgroup
lock and set pc->flags non-atomically.
[mhocko@suse.cz: mem_cgroup_charge_statistics needs preempt_disable]
[vdavydov@parallels.com: fix flags definition]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Tested-by: Jet Chen <jet.chen@intel.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Tested-by: Felipe Balbi <balbi@ti.com>
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-08-08 14:19:22 -07:00
|
|
|
|
|
|
|
|
|
2016-01-20 15:03:16 -08:00
|
|
|
|
2021-06-28 19:37:47 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcontrol: rewrite uncharge API
The memcg uncharging code that is involved towards the end of a page's
lifetime - truncation, reclaim, swapout, migration - is impressively
complicated and fragile.
Because anonymous and file pages were always charged before they had their
page->mapping established, uncharges had to happen when the page type
could still be known from the context; as in unmap for anonymous, page
cache removal for file and shmem pages, and swap cache truncation for swap
pages. However, these operations happen well before the page is actually
freed, and so a lot of synchronization is necessary:
- Charging, uncharging, page migration, and charge migration all need
to take a per-page bit spinlock as they could race with uncharging.
- Swap cache truncation happens during both swap-in and swap-out, and
possibly repeatedly before the page is actually freed. This means
that the memcg swapout code is called from many contexts that make
no sense and it has to figure out the direction from page state to
make sure memory and memory+swap are always correctly charged.
- On page migration, the old page might be unmapped but then reused,
so memcg code has to prevent untimely uncharging in that case.
Because this code - which should be a simple charge transfer - is so
special-cased, it is not reusable for replace_page_cache().
But now that charged pages always have a page->mapping, introduce
mem_cgroup_uncharge(), which is called after the final put_page(), when we
know for sure that nobody is looking at the page anymore.
For page migration, introduce mem_cgroup_migrate(), which is called after
the migration is successful and the new page is fully rmapped. Because
the old page is no longer uncharged after migration, prevent double
charges by decoupling the page's memcg association (PCG_USED and
pc->mem_cgroup) from the page holding an actual charge. The new bits
PCG_MEM and PCG_MEMSW represent the respective charges and are transferred
to the new page during migration.
mem_cgroup_migrate() is suitable for replace_page_cache() as well,
which gets rid of mem_cgroup_replace_page_cache(). However, care
needs to be taken because both the source and the target page can
already be charged and on the LRU when fuse is splicing: grab the page
lock on the charge moving side to prevent changing pc->mem_cgroup of a
page under migration. Also, the lruvecs of both pages change as we
uncharge the old and charge the new during migration, and putback may
race with us, so grab the lru lock and isolate the pages iff on LRU to
prevent races and ensure the pages are on the right lruvec afterward.
Swap accounting is massively simplified: because the page is no longer
uncharged as early as swap cache deletion, a new mem_cgroup_swapout() can
transfer the page's memory+swap charge (PCG_MEMSW) to the swap entry
before the final put_page() in page reclaim.
Finally, page_cgroup changes are now protected by whatever protection the
page itself offers: anonymous pages are charged under the page table lock,
whereas page cache insertions, swapin, and migration hold the page lock.
Uncharging happens under full exclusion with no outstanding references.
Charging and uncharging also ensure that the page is off-LRU, which
serializes against charge migration. Remove the very costly page_cgroup
lock and set pc->flags non-atomically.
[mhocko@suse.cz: mem_cgroup_charge_statistics needs preempt_disable]
[vdavydov@parallels.com: fix flags definition]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Tested-by: Jet Chen <jet.chen@intel.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Tested-by: Felipe Balbi <balbi@ti.com>
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-08-08 14:19:22 -07:00
|
|
|
|
2020-08-06 23:20:45 -07:00
|
|
|
|
2021-05-06 18:14:59 -04:00
|
|
|
|
2024-08-14 22:04:50 -07:00
|
|
|
|
mm: memcontrol: rewrite uncharge API
The memcg uncharging code that is involved towards the end of a page's
lifetime - truncation, reclaim, swapout, migration - is impressively
complicated and fragile.
Because anonymous and file pages were always charged before they had their
page->mapping established, uncharges had to happen when the page type
could still be known from the context; as in unmap for anonymous, page
cache removal for file and shmem pages, and swap cache truncation for swap
pages. However, these operations happen well before the page is actually
freed, and so a lot of synchronization is necessary:
- Charging, uncharging, page migration, and charge migration all need
to take a per-page bit spinlock as they could race with uncharging.
- Swap cache truncation happens during both swap-in and swap-out, and
possibly repeatedly before the page is actually freed. This means
that the memcg swapout code is called from many contexts that make
no sense and it has to figure out the direction from page state to
make sure memory and memory+swap are always correctly charged.
- On page migration, the old page might be unmapped but then reused,
so memcg code has to prevent untimely uncharging in that case.
Because this code - which should be a simple charge transfer - is so
special-cased, it is not reusable for replace_page_cache().
But now that charged pages always have a page->mapping, introduce
mem_cgroup_uncharge(), which is called after the final put_page(), when we
know for sure that nobody is looking at the page anymore.
For page migration, introduce mem_cgroup_migrate(), which is called after
the migration is successful and the new page is fully rmapped. Because
the old page is no longer uncharged after migration, prevent double
charges by decoupling the page's memcg association (PCG_USED and
pc->mem_cgroup) from the page holding an actual charge. The new bits
PCG_MEM and PCG_MEMSW represent the respective charges and are transferred
to the new page during migration.
mem_cgroup_migrate() is suitable for replace_page_cache() as well,
which gets rid of mem_cgroup_replace_page_cache(). However, care
needs to be taken because both the source and the target page can
already be charged and on the LRU when fuse is splicing: grab the page
lock on the charge moving side to prevent changing pc->mem_cgroup of a
page under migration. Also, the lruvecs of both pages change as we
uncharge the old and charge the new during migration, and putback may
race with us, so grab the lru lock and isolate the pages iff on LRU to
prevent races and ensure the pages are on the right lruvec afterward.
Swap accounting is massively simplified: because the page is no longer
uncharged as early as swap cache deletion, a new mem_cgroup_swapout() can
transfer the page's memory+swap charge (PCG_MEMSW) to the swap entry
before the final put_page() in page reclaim.
Finally, page_cgroup changes are now protected by whatever protection the
page itself offers: anonymous pages are charged under the page table lock,
whereas page cache insertions, swapin, and migration hold the page lock.
Uncharging happens under full exclusion with no outstanding references.
Charging and uncharging also ensure that the page is off-LRU, which
serializes against charge migration. Remove the very costly page_cgroup
lock and set pc->flags non-atomically.
[mhocko@suse.cz: mem_cgroup_charge_statistics needs preempt_disable]
[vdavydov@parallels.com: fix flags definition]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Tested-by: Jet Chen <jet.chen@intel.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Tested-by: Felipe Balbi <balbi@ti.com>
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-08-08 14:19:22 -07:00
|
|
|
|
|
|
|
|
|
2023-10-06 11:46:27 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-06-14 09:07:42 +08:00
|
|
|
|
2023-10-06 11:46:27 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
hugetlb: memcg: account hugetlb-backed memory in memory controller
Currently, hugetlb memory usage is not acounted for in the memory
controller, which could lead to memory overprotection for cgroups with
hugetlb-backed memory. This has been observed in our production system.
For instance, here is one of our usecases: suppose there are two 32G
containers. The machine is booted with hugetlb_cma=6G, and each container
may or may not use up to 3 gigantic page, depending on the workload within
it. The rest is anon, cache, slab, etc. We can set the hugetlb cgroup
limit of each cgroup to 3G to enforce hugetlb fairness. But it is very
difficult to configure memory.max to keep overall consumption, including
anon, cache, slab etc. fair.
What we have had to resort to is to constantly poll hugetlb usage and
readjust memory.max. Similar procedure is done to other memory limits
(memory.low for e.g). However, this is rather cumbersome and buggy.
Furthermore, when there is a delay in memory limits correction, (for e.g
when hugetlb usage changes within consecutive runs of the userspace
agent), the system could be in an over/underprotected state.
This patch rectifies this issue by charging the memcg when the hugetlb
folio is utilized, and uncharging when the folio is freed (analogous to
the hugetlb controller). Note that we do not charge when the folio is
allocated to the hugetlb pool, because at this point it is not owned by
any memcg.
Some caveats to consider:
* This feature is only available on cgroup v2.
* There is no hugetlb pool management involved in the memory
controller. As stated above, hugetlb folios are only charged towards
the memory controller when it is used. Host overcommit management
has to consider it when configuring hard limits.
* Failure to charge towards the memcg results in SIGBUS. This could
happen even if the hugetlb pool still has pages (but the cgroup
limit is hit and reclaim attempt fails).
* When this feature is enabled, hugetlb pages contribute to memory
reclaim protection. low, min limits tuning must take into account
hugetlb memory.
* Hugetlb pages utilized while this option is not selected will not
be tracked by the memory controller (even if cgroup v2 is remounted
later on).
Link: https://lkml.kernel.org/r/20231006184629.155543-4-nphamcs@gmail.com
Signed-off-by: Nhat Pham <nphamcs@gmail.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Frank van der Linden <fvdl@google.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Rik van Riel <riel@surriel.com>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Tejun heo <tj@kernel.org>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-06 11:46:28 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-06 11:46:27 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/thp: fix deferred split unqueue naming and locking
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.
Before fixing locking: rename misleading folio_undo_large_rmappable(),
which does not undo large_rmappable, to folio_unqueue_deferred_split(),
which is what it does. But that and its out-of-line __callee are mm
internals of very limited usability: add comment and WARN_ON_ONCEs to
check usage; and return a bool to say if a deferred split was unqueued,
which can then be used in WARN_ON_ONCEs around safety checks (sparing
callers the arcane conditionals in __folio_unqueue_deferred_split()).
Just omit the folio_unqueue_deferred_split() from free_unref_folios(), all
of whose callers now call it beforehand (and if any forget then bad_page()
will tell) - except for its caller put_pages_list(), which itself no
longer has any callers (and will be deleted separately).
Swapout: mem_cgroup_swapout() has been resetting folio->memcg_data 0
without checking and unqueueing a THP folio from deferred split list;
which is unfortunate, since the split_queue_lock depends on the memcg
(when memcg is enabled); so swapout has been unqueueing such THPs later,
when freeing the folio, using the pgdat's lock instead: potentially
corrupting the memcg's list. __remove_mapping() has frozen refcount to 0
here, so no problem with calling folio_unqueue_deferred_split() before
resetting memcg_data.
That goes back to 5.4 commit 87eaceb3faa5 ("mm: thp: make deferred split
shrinker memcg aware"): which included a check on swapcache before adding
to deferred queue, but no check on deferred queue before adding THP to
swapcache. That worked fine with the usual sequence of events in reclaim
(though there were a couple of rare ways in which a THP on deferred queue
could have been swapped out), but 6.12 commit dafff3f4c850 ("mm: split
underused THPs") avoids splitting underused THPs in reclaim, which makes
swapcache THPs on deferred queue commonplace.
Keep the check on swapcache before adding to deferred queue? Yes: it is
no longer essential, but preserves the existing behaviour, and is likely
to be a worthwhile optimization (vmstat showed much more traffic on the
queue under swapping load if the check was removed); update its comment.
Memcg-v1 move (deprecated): mem_cgroup_move_account() has been changing
folio->memcg_data without checking and unqueueing a THP folio from the
deferred list, sometimes corrupting "from" memcg's list, like swapout.
Refcount is non-zero here, so folio_unqueue_deferred_split() can only be
used in a WARN_ON_ONCE to validate the fix, which must be done earlier:
mem_cgroup_move_charge_pte_range() first try to split the THP (splitting
of course unqueues), or skip it if that fails. Not ideal, but moving
charge has been requested, and khugepaged should repair the THP later:
nobody wants new custom unqueueing code just for this deprecated case.
The 87eaceb3faa5 commit did have the code to move from one deferred list
to another (but was not conscious of its unsafety while refcount non-0);
but that was removed by 5.6 commit fac0516b5534 ("mm: thp: don't need care
deferred split queue in memcg charge move path"), which argued that the
existence of a PMD mapping guarantees that the THP cannot be on a deferred
list. As above, false in rare cases, and now commonly false.
Backport to 6.11 should be straightforward. Earlier backports must take
care that other _deferred_list fixes and dependencies are included. There
is not a strong case for backports, but they can fix cornercases.
Link: https://lkml.kernel.org/r/8dc111ae-f6db-2da7-b25c-7a20b1effe3b@google.com
Fixes: 87eaceb3faa5 ("mm: thp: make deferred split shrinker memcg aware")
Fixes: dafff3f4c850 ("mm: split underused THPs")
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: David Hildenbrand <david@redhat.com>
Reviewed-by: Yang Shi <shy828301@gmail.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.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: Usama Arif <usamaarif642@gmail.com>
Cc: Wei Yang <richard.weiyang@gmail.com>
Cc: Zi Yan <ziy@nvidia.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-10-27 13:02:13 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-06 11:46:27 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-01-14 15:21:34 -08:00
|
|
|
|
2016-01-14 15:21:26 -08:00
|
|
|
|
|
|
|
|
|
2016-10-07 17:00:58 -07:00
|
|
|
|
2016-01-14 15:21:26 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-10-07 17:00:58 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-03-09 22:16:05 -07:00
|
|
|
|
2022-03-22 14:40:07 -07:00
|
|
|
|
2020-03-09 22:16:05 -07:00
|
|
|
|
|
|
|
|
|
2016-01-14 15:21:26 -08:00
|
|
|
|
|
|
|
|
|
2022-09-30 19:14:33 +05:30
|
|
|
|
2016-01-14 15:21:29 -08:00
|
|
|
|
2024-06-28 21:03:10 +00:00
|
|
|
|
2016-01-14 15:21:29 -08:00
|
|
|
|
2020-04-01 21:07:10 -07:00
|
|
|
|
2016-01-14 15:21:26 -08:00
|
|
|
|
2016-01-14 15:21:29 -08:00
|
|
|
|
2016-01-14 15:21:26 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-10-07 17:00:58 -07:00
|
|
|
|
2016-01-14 15:21:26 -08:00
|
|
|
|
2025-08-15 20:16:14 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-01-14 15:21:26 -08:00
|
|
|
|
|
|
|
|
|
2025-08-15 20:16:09 +00:00
|
|
|
|
|
|
|
|
|
2025-08-15 20:16:14 +00:00
|
|
|
|
|
|
|
|
|
2025-08-15 20:16:09 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-08-15 20:16:14 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-08-15 20:16:09 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-01-14 15:21:26 -08:00
|
|
|
|
2025-08-15 20:16:16 +00:00
|
|
|
|
|
|
|
|
|
2016-01-14 15:21:26 -08:00
|
|
|
|
2021-08-17 12:40:03 -07:00
|
|
|
|
2016-01-14 15:21:26 -08:00
|
|
|
|
|
|
|
|
|
2021-08-17 12:40:03 -07:00
|
|
|
|
2016-01-14 15:21:26 -08:00
|
|
|
|
2025-08-15 20:16:16 +00:00
|
|
|
|
|
|
|
|
|
2016-01-14 15:21:26 -08:00
|
|
|
|
2025-08-15 20:16:16 +00:00
|
|
|
|
|
|
|
|
|
2024-06-28 21:03:10 +00:00
|
|
|
|
|
|
|
|
|
2016-01-20 15:02:47 -08:00
|
|
|
|
2025-02-27 18:23:54 -08:00
|
|
|
|
2021-08-17 12:40:03 -07:00
|
|
|
|
2016-01-14 15:21:29 -08:00
|
|
|
|
2021-08-17 12:40:03 -07:00
|
|
|
|
2016-01-14 15:21:29 -08:00
|
|
|
|
2016-01-14 15:21:26 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-08-15 20:16:16 +00:00
|
|
|
|
|
|
|
|
|
2018-02-06 15:42:13 -08:00
|
|
|
|
2016-01-14 15:21:26 -08:00
|
|
|
|
2025-08-15 20:16:16 +00:00
|
|
|
|
2016-01-14 15:21:26 -08:00
|
|
|
|
2025-08-15 20:16:16 +00:00
|
|
|
|
|
|
|
|
|
2016-01-14 15:21:29 -08:00
|
|
|
|
2024-06-28 21:03:10 +00:00
|
|
|
|
2016-01-14 15:21:29 -08:00
|
|
|
|
|
|
|
|
|
2016-01-20 15:02:47 -08:00
|
|
|
|
2018-01-31 16:16:37 -08:00
|
|
|
|
2016-01-20 15:03:22 -08:00
|
|
|
|
2017-09-08 16:13:09 -07:00
|
|
|
|
2016-01-14 15:21:26 -08:00
|
|
|
|
|
|
|
|
|
2016-01-14 15:21:29 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-01-20 15:02:38 -08:00
|
|
|
|
|
|
|
|
|
2023-02-10 15:47:31 +00:00
|
|
|
|
|
|
|
|
|
2016-01-14 15:21:29 -08:00
|
|
|
|
2022-03-22 14:40:31 -07:00
|
|
|
|
2016-01-14 15:21:29 -08:00
|
|
|
|
|
|
|
|
|
2016-01-14 15:21:26 -08:00
|
|
|
|
2013-02-22 16:34:43 -08:00
|
|
|
|
mm/memcg: move mem_cgroup_init() ahead of cgroup_init()
Patch series "Use kmem_cache for memcg alloc", v3.
(willy tldr: "you've gone from allocating 8 objects per 32KiB to
allocating 13 objects per 32KiB, a 62% improvement in memory consumption"
[1])
The mem_cgroup_alloc function creates mem_cgroup struct and it's
associated structures including mem_cgroup_per_node. Through detailed
analysis on our test machine (Arm64, 16GB RAM, 6.6 kernel, 1 NUMA node,
memcgv2 with nokmem,nosocket,cgroup_disable=pressure), we can observe the
memory allocation for these structures using the following shell commands:
# Enable tracing
echo 1 > /sys/kernel/tracing/events/kmem/kmalloc/enable
echo 1 > /sys/kernel/tracing/tracing_on
cat /sys/kernel/tracing/trace_pipe | grep kmalloc | grep mem_cgroup
# Trigger allocation if cgroup subtree do not enable memcg
echo +memory > /sys/fs/cgroup/cgroup.subtree_control
Ftrace Output:
# mem_cgroup struct allocation
sh-6312 [000] ..... 58015.698365: kmalloc:
call_site=mem_cgroup_css_alloc+0xd8/0x5b4
ptr=000000003e4c3799 bytes_req=2312 bytes_alloc=4096
gfp_flags=GFP_KERNEL|__GFP_ZERO node=-1 accounted=false
# mem_cgroup_per_node allocation
sh-6312 [000] ..... 58015.698389: kmalloc:
call_site=mem_cgroup_css_alloc+0x1d8/0x5b4
ptr=00000000d798700c bytes_req=2896 bytes_alloc=4096
gfp_flags=GFP_KERNEL|__GFP_ZERO node=0 accounted=false
Key Observations:
1. Both structures use kmalloc with requested sizes between 2KB-4KB
2. Allocation alignment forces 4KB slab usage due to pre-defined sizes
(64B, 128B,..., 2KB, 4KB, 8KB)
3. Memory waste per memcg instance:
Base struct: 4096 - 2312 = 1784 bytes
Per-node struct: 4096 - 2896 = 1200 bytes
Total waste: 2984 bytes (1-node system)
NUMA scaling: (1200 + 8) * nr_node_ids bytes
So, it's a little waste.
This patchset introduces dedicated kmem_cache:
Patch2 - mem_cgroup kmem_cache - memcg_cachep
Patch3 - mem_cgroup_per_node kmem_cache - memcg_pn_cachep
The benefits of this change can be observed with the following tracing
commands:
# Enable tracing
echo 1 > /sys/kernel/tracing/events/kmem/kmem_cache_alloc/enable
echo 1 > /sys/kernel/tracing/tracing_on
cat /sys/kernel/tracing/trace_pipe | grep kmem_cache_alloc | grep mem_cgroup
# In another terminal:
echo +memory > /sys/fs/cgroup/cgroup.subtree_control
The output might now look like this:
# mem_cgroup struct allocation
sh-9827 [000] ..... 289.513598: kmem_cache_alloc:
call_site=mem_cgroup_css_alloc+0xbc/0x5d4 ptr=00000000695c1806
bytes_req=2312 bytes_alloc=2368 gfp_flags=GFP_KERNEL|__GFP_ZERO node=-1
accounted=false
# mem_cgroup_per_node allocation
sh-9827 [000] ..... 289.513602: kmem_cache_alloc:
call_site=mem_cgroup_css_alloc+0x1b8/0x5d4 ptr=000000002989e63a
bytes_req=2896 bytes_alloc=2944 gfp_flags=GFP_KERNEL|__GFP_ZERO node=0
accounted=false
This indicates that the `mem_cgroup` struct now requests 2312 bytes and is
allocated 2368 bytes, while `mem_cgroup_per_node` requests 2896 bytes and
is allocated 2944 bytes. The slight increase in allocated size is due to
`SLAB_HWCACHE_ALIGN` in the `kmem_cache`.
Without `SLAB_HWCACHE_ALIGN`, the allocation might appear as:
# mem_cgroup struct allocation
sh-9269 [003] ..... 80.396366: kmem_cache_alloc:
call_site=mem_cgroup_css_alloc+0xbc/0x5d4 ptr=000000005b12b475
bytes_req=2312 bytes_alloc=2312 gfp_flags=GFP_KERNEL|__GFP_ZERO node=-1
accounted=false
# mem_cgroup_per_node allocation
sh-9269 [003] ..... 80.396411: kmem_cache_alloc:
call_site=mem_cgroup_css_alloc+0x1b8/0x5d4 ptr=00000000f347adc6
bytes_req=2896 bytes_alloc=2896 gfp_flags=GFP_KERNEL|__GFP_ZERO node=0
accounted=false
While the `bytes_alloc` now matches the `bytes_req`, this patchset
defaults to using `SLAB_HWCACHE_ALIGN` as it is generally considered more
beneficial for performance. Please let me know if there are any issues or
if I've misunderstood anything.
This patchset also move mem_cgroup_init ahead of cgroup_init() due to
cgroup_init() will allocate root_mem_cgroup, but each initcall invoke
after cgroup_init, so if each kmem_cache do not prepare, we need testing
NULL before use it.
This patch (of 3):
When cgroup_init() creates root_mem_cgroup through css_alloc callback,
some critical resources might not be fully initialized, forcing later
operations to perform conditional checks for resource availability.
This patch move mem_cgroup_init() to address the init order, it invoke
before cgroup_init, so, compare to subsys_initcall, it can use to prepare
some key resources before root_mem_cgroup alloc.
Link: https://lkml.kernel.org/r/aAsRCj-niMMTtmK8@casper.infradead.org [1]
Link: https://lkml.kernel.org/r/20250425031935.76411-1-link@vivo.com
Link: https://lkml.kernel.org/r/20250425031935.76411-2-link@vivo.com
Signed-off-by: Huan Yang <link@vivo.com>
Suggested-by: Shakeel Butt <shakeel.butt@linux.dev>
Acked-by: Shakeel Butt <shakeel.butt@linux.dev>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Francesco Valla <francesco@valla.it>
Cc: guoweikang <guoweikang.kernel@gmail.com>
Cc: Huang Shijie <shijie@os.amperecomputing.com>
Cc: KP Singh <kpsingh@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: "Paul E . McKenney" <paulmck@kernel.org>
Cc: Petr Mladek <pmladek@suse.com>
Cc: Rasmus Villemoes <linux@rasmusvillemoes.dk>
Cc: Raul E Rangel <rrangel@chromium.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: "Uladzislau Rezki (Sony)" <urezki@gmail.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Matthew Wilcox <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-25 11:19:23 +08:00
|
|
|
|
2013-02-22 16:35:41 -08:00
|
|
|
|
2016-11-03 15:49:59 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-02-22 16:34:43 -08:00
|
|
|
|
mm/memcg: move mem_cgroup_init() ahead of cgroup_init()
Patch series "Use kmem_cache for memcg alloc", v3.
(willy tldr: "you've gone from allocating 8 objects per 32KiB to
allocating 13 objects per 32KiB, a 62% improvement in memory consumption"
[1])
The mem_cgroup_alloc function creates mem_cgroup struct and it's
associated structures including mem_cgroup_per_node. Through detailed
analysis on our test machine (Arm64, 16GB RAM, 6.6 kernel, 1 NUMA node,
memcgv2 with nokmem,nosocket,cgroup_disable=pressure), we can observe the
memory allocation for these structures using the following shell commands:
# Enable tracing
echo 1 > /sys/kernel/tracing/events/kmem/kmalloc/enable
echo 1 > /sys/kernel/tracing/tracing_on
cat /sys/kernel/tracing/trace_pipe | grep kmalloc | grep mem_cgroup
# Trigger allocation if cgroup subtree do not enable memcg
echo +memory > /sys/fs/cgroup/cgroup.subtree_control
Ftrace Output:
# mem_cgroup struct allocation
sh-6312 [000] ..... 58015.698365: kmalloc:
call_site=mem_cgroup_css_alloc+0xd8/0x5b4
ptr=000000003e4c3799 bytes_req=2312 bytes_alloc=4096
gfp_flags=GFP_KERNEL|__GFP_ZERO node=-1 accounted=false
# mem_cgroup_per_node allocation
sh-6312 [000] ..... 58015.698389: kmalloc:
call_site=mem_cgroup_css_alloc+0x1d8/0x5b4
ptr=00000000d798700c bytes_req=2896 bytes_alloc=4096
gfp_flags=GFP_KERNEL|__GFP_ZERO node=0 accounted=false
Key Observations:
1. Both structures use kmalloc with requested sizes between 2KB-4KB
2. Allocation alignment forces 4KB slab usage due to pre-defined sizes
(64B, 128B,..., 2KB, 4KB, 8KB)
3. Memory waste per memcg instance:
Base struct: 4096 - 2312 = 1784 bytes
Per-node struct: 4096 - 2896 = 1200 bytes
Total waste: 2984 bytes (1-node system)
NUMA scaling: (1200 + 8) * nr_node_ids bytes
So, it's a little waste.
This patchset introduces dedicated kmem_cache:
Patch2 - mem_cgroup kmem_cache - memcg_cachep
Patch3 - mem_cgroup_per_node kmem_cache - memcg_pn_cachep
The benefits of this change can be observed with the following tracing
commands:
# Enable tracing
echo 1 > /sys/kernel/tracing/events/kmem/kmem_cache_alloc/enable
echo 1 > /sys/kernel/tracing/tracing_on
cat /sys/kernel/tracing/trace_pipe | grep kmem_cache_alloc | grep mem_cgroup
# In another terminal:
echo +memory > /sys/fs/cgroup/cgroup.subtree_control
The output might now look like this:
# mem_cgroup struct allocation
sh-9827 [000] ..... 289.513598: kmem_cache_alloc:
call_site=mem_cgroup_css_alloc+0xbc/0x5d4 ptr=00000000695c1806
bytes_req=2312 bytes_alloc=2368 gfp_flags=GFP_KERNEL|__GFP_ZERO node=-1
accounted=false
# mem_cgroup_per_node allocation
sh-9827 [000] ..... 289.513602: kmem_cache_alloc:
call_site=mem_cgroup_css_alloc+0x1b8/0x5d4 ptr=000000002989e63a
bytes_req=2896 bytes_alloc=2944 gfp_flags=GFP_KERNEL|__GFP_ZERO node=0
accounted=false
This indicates that the `mem_cgroup` struct now requests 2312 bytes and is
allocated 2368 bytes, while `mem_cgroup_per_node` requests 2896 bytes and
is allocated 2944 bytes. The slight increase in allocated size is due to
`SLAB_HWCACHE_ALIGN` in the `kmem_cache`.
Without `SLAB_HWCACHE_ALIGN`, the allocation might appear as:
# mem_cgroup struct allocation
sh-9269 [003] ..... 80.396366: kmem_cache_alloc:
call_site=mem_cgroup_css_alloc+0xbc/0x5d4 ptr=000000005b12b475
bytes_req=2312 bytes_alloc=2312 gfp_flags=GFP_KERNEL|__GFP_ZERO node=-1
accounted=false
# mem_cgroup_per_node allocation
sh-9269 [003] ..... 80.396411: kmem_cache_alloc:
call_site=mem_cgroup_css_alloc+0x1b8/0x5d4 ptr=00000000f347adc6
bytes_req=2896 bytes_alloc=2896 gfp_flags=GFP_KERNEL|__GFP_ZERO node=0
accounted=false
While the `bytes_alloc` now matches the `bytes_req`, this patchset
defaults to using `SLAB_HWCACHE_ALIGN` as it is generally considered more
beneficial for performance. Please let me know if there are any issues or
if I've misunderstood anything.
This patchset also move mem_cgroup_init ahead of cgroup_init() due to
cgroup_init() will allocate root_mem_cgroup, but each initcall invoke
after cgroup_init, so if each kmem_cache do not prepare, we need testing
NULL before use it.
This patch (of 3):
When cgroup_init() creates root_mem_cgroup through css_alloc callback,
some critical resources might not be fully initialized, forcing later
operations to perform conditional checks for resource availability.
This patch move mem_cgroup_init() to address the init order, it invoke
before cgroup_init, so, compare to subsys_initcall, it can use to prepare
some key resources before root_mem_cgroup alloc.
Link: https://lkml.kernel.org/r/aAsRCj-niMMTtmK8@casper.infradead.org [1]
Link: https://lkml.kernel.org/r/20250425031935.76411-1-link@vivo.com
Link: https://lkml.kernel.org/r/20250425031935.76411-2-link@vivo.com
Signed-off-by: Huan Yang <link@vivo.com>
Suggested-by: Shakeel Butt <shakeel.butt@linux.dev>
Acked-by: Shakeel Butt <shakeel.butt@linux.dev>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Francesco Valla <francesco@valla.it>
Cc: guoweikang <guoweikang.kernel@gmail.com>
Cc: Huang Shijie <shijie@os.amperecomputing.com>
Cc: KP Singh <kpsingh@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: "Paul E . McKenney" <paulmck@kernel.org>
Cc: Petr Mladek <pmladek@suse.com>
Cc: Rasmus Villemoes <linux@rasmusvillemoes.dk>
Cc: Raul E Rangel <rrangel@chromium.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: "Uladzislau Rezki (Sony)" <urezki@gmail.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Matthew Wilcox <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-25 11:19:23 +08:00
|
|
|
|
2013-02-22 16:34:43 -08:00
|
|
|
|
2025-04-25 11:19:24 +08:00
|
|
|
|
2024-06-24 17:58:54 -07:00
|
|
|
|
2015-02-11 15:26:33 -08:00
|
|
|
|
2021-02-24 12:03:15 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-11-03 15:49:59 +01:00
|
|
|
|
|
|
|
|
|
2015-02-11 15:26:33 -08:00
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
2015-02-11 15:26:33 -08:00
|
|
|
|
2025-05-06 15:55:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-02-11 15:26:33 -08:00
|
|
|
|
2025-04-25 11:19:24 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-04-25 11:19:25 +08:00
|
|
|
|
|
|
|
|
|
2015-02-11 15:26:33 -08:00
|
|
|
|
2013-02-22 16:34:43 -08:00
|
|
|
|
|
|
|
|
|
2015-02-11 15:26:36 -08:00
|
|
|
|
2022-09-26 09:57:04 -04:00
|
|
|
|
mm, THP, swap: delay splitting THP during swap out
Patch series "THP swap: Delay splitting THP during swapping out", v11.
This patchset is to optimize the performance of Transparent Huge Page
(THP) swap.
Recently, the performance of the storage devices improved so fast that
we cannot saturate the disk bandwidth with single logical CPU when do
page swap out even on a high-end server machine. Because the
performance of the storage device improved faster than that of single
logical CPU. And it seems that the trend will not change in the near
future. On the other hand, the THP becomes more and more popular
because of increased memory size. So it becomes necessary to optimize
THP swap performance.
The advantages of the THP swap support include:
- Batch the swap operations for the THP to reduce lock
acquiring/releasing, including allocating/freeing the swap space,
adding/deleting to/from the swap cache, and writing/reading the swap
space, etc. This will help improve the performance of the THP swap.
- The THP swap space read/write will be 2M sequential IO. It is
particularly helpful for the swap read, which are usually 4k random
IO. This will improve the performance of the THP swap too.
- It will help the memory fragmentation, especially when the THP is
heavily used by the applications. The 2M continuous pages will be
free up after THP swapping out.
- It will improve the THP utilization on the system with the swap
turned on. Because the speed for khugepaged to collapse the normal
pages into the THP is quite slow. After the THP is split during the
swapping out, it will take quite long time for the normal pages to
collapse back into the THP after being swapped in. The high THP
utilization helps the efficiency of the page based memory management
too.
There are some concerns regarding THP swap in, mainly because possible
enlarged read/write IO size (for swap in/out) may put more overhead on
the storage device. To deal with that, the THP swap in should be turned
on only when necessary. For example, it can be selected via
"always/never/madvise" logic, to be turned on globally, turned off
globally, or turned on only for VMA with MADV_HUGEPAGE, etc.
This patchset is the first step for the THP swap support. The plan is
to delay splitting THP step by step, finally avoid splitting THP during
the THP swapping out and swap out/in the THP as a whole.
As the first step, in this patchset, the splitting huge page is delayed
from almost the first step of swapping out to after allocating the swap
space for the THP and adding the THP into the swap cache. This will
reduce lock acquiring/releasing for the locks used for the swap cache
management.
With the patchset, the swap out throughput improves 15.5% (from about
3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case
with 8 processes. The test is done on a Xeon E5 v3 system. The swap
device used is a RAM simulated PMEM (persistent memory) device. To test
the sequential swapping out, the test case creates 8 processes, which
sequentially allocate and write to the anonymous pages until the RAM and
part of the swap device is used up.
This patch (of 5):
In this patch, splitting huge page is delayed from almost the first step
of swapping out to after allocating the swap space for the THP
(Transparent Huge Page) and adding the THP into the swap cache. This
will batch the corresponding operation, thus improve THP swap out
throughput.
This is the first step for the THP swap optimization. The plan is to
delay splitting the THP step by step and avoid splitting the THP
finally.
In this patch, one swap cluster is used to hold the contents of each THP
swapped out. So, the size of the swap cluster is changed to that of the
THP (Transparent Huge Page) on x86_64 architecture (512). For other
architectures which want such THP swap optimization,
ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for
the architecture. In effect, this will enlarge swap cluster size by 2
times on x86_64. Which may make it harder to find a free cluster when
the swap space becomes fragmented. So that, this may reduce the
continuous swap space allocation and sequential write in theory. The
performance test in 0day shows no regressions caused by this.
In the future of THP swap optimization, some information of the swapped
out THP (such as compound map count) will be recorded in the
swap_cluster_info data structure.
The mem cgroup swap accounting functions are enhanced to support charge
or uncharge a swap cluster backing a THP as a whole.
The swap cluster allocate/free functions are added to allocate/free a
swap cluster for a THP. A fair simple algorithm is used for swap
cluster allocation, that is, only the first swap device in priority list
will be tried to allocate the swap cluster. The function will fail if
the trying is not successful, and the caller will fallback to allocate a
single swap slot instead. This works good enough for normal cases. If
the difference of the number of the free swap clusters among multiple
swap devices is significant, it is possible that some THPs are split
earlier than necessary. For example, this could be caused by big size
difference among multiple swap devices.
The swap cache functions is enhanced to support add/delete THP to/from
the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be
enhanced in the future with multi-order radix tree. But because we will
split the THP soon during swapping out, that optimization doesn't make
much sense for this first step.
The THP splitting functions are enhanced to support to split THP in swap
cache during swapping out. The page lock will be held during allocating
the swap cluster, adding the THP into the swap cache and splitting the
THP. So in the code path other than swapping out, if the THP need to be
split, the PageSwapCache(THP) will be always false.
The swap cluster is only available for SSD, so the THP swap optimization
in this patchset has no effect for HDD.
[ying.huang@intel.com: fix two issues in THP optimize patch]
Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com
[hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size]
Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option]
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h]
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
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>
2017-07-06 15:37:18 -07:00
|
|
|
|
2022-05-12 20:23:02 -07:00
|
|
|
|
|
|
|
|
|
mm: memcontrol: charge swap to cgroup2
This patchset introduces swap accounting to cgroup2.
This patch (of 7):
In the legacy hierarchy we charge memsw, which is dubious, because:
- memsw.limit must be >= memory.limit, so it is impossible to limit
swap usage less than memory usage. Taking into account the fact that
the primary limiting mechanism in the unified hierarchy is
memory.high while memory.limit is either left unset or set to a very
large value, moving memsw.limit knob to the unified hierarchy would
effectively make it impossible to limit swap usage according to the
user preference.
- memsw.usage != memory.usage + swap.usage, because a page occupying
both swap entry and a swap cache page is charged only once to memsw
counter. As a result, it is possible to effectively eat up to
memory.limit of memory pages *and* memsw.limit of swap entries, which
looks unexpected.
That said, we should provide a different swap limiting mechanism for
cgroup2.
This patch adds mem_cgroup->swap counter, which charges the actual number
of swap entries used by a cgroup. It is only charged in the unified
hierarchy, while the legacy hierarchy memsw logic is left intact.
The swap usage can be monitored using new memory.swap.current file and
limited using memory.swap.max.
Note, to charge swap resource properly in the unified hierarchy, we have
to make swap_entry_free uncharge swap only when ->usage reaches zero, not
just ->count, i.e. when all references to a swap entry, including the one
taken by swap cache, are gone. This is necessary, because otherwise
swap-in could result in uncharging swap even if the page is still in swap
cache and hence still occupies a swap entry. At the same time, this
shouldn't break memsw counter logic, where a page is never charged twice
for using both memory and swap, because in case of legacy hierarchy we
uncharge swap on commit (see mem_cgroup_commit_charge).
Signed-off-by: Vladimir Davydov <vdavydov@virtuozzo.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-20 15:02:56 -08:00
|
|
|
|
|
|
|
|
|
2022-05-12 20:23:02 -07:00
|
|
|
|
mm: memcontrol: charge swap to cgroup2
This patchset introduces swap accounting to cgroup2.
This patch (of 7):
In the legacy hierarchy we charge memsw, which is dubious, because:
- memsw.limit must be >= memory.limit, so it is impossible to limit
swap usage less than memory usage. Taking into account the fact that
the primary limiting mechanism in the unified hierarchy is
memory.high while memory.limit is either left unset or set to a very
large value, moving memsw.limit knob to the unified hierarchy would
effectively make it impossible to limit swap usage according to the
user preference.
- memsw.usage != memory.usage + swap.usage, because a page occupying
both swap entry and a swap cache page is charged only once to memsw
counter. As a result, it is possible to effectively eat up to
memory.limit of memory pages *and* memsw.limit of swap entries, which
looks unexpected.
That said, we should provide a different swap limiting mechanism for
cgroup2.
This patch adds mem_cgroup->swap counter, which charges the actual number
of swap entries used by a cgroup. It is only charged in the unified
hierarchy, while the legacy hierarchy memsw logic is left intact.
The swap usage can be monitored using new memory.swap.current file and
limited using memory.swap.max.
Note, to charge swap resource properly in the unified hierarchy, we have
to make swap_entry_free uncharge swap only when ->usage reaches zero, not
just ->count, i.e. when all references to a swap entry, including the one
taken by swap cache, are gone. This is necessary, because otherwise
swap-in could result in uncharging swap even if the page is still in swap
cache and hence still occupies a swap entry. At the same time, this
shouldn't break memsw counter logic, where a page is never charged twice
for using both memory and swap, because in case of legacy hierarchy we
uncharge swap on commit (see mem_cgroup_commit_charge).
Signed-off-by: Vladimir Davydov <vdavydov@virtuozzo.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-20 15:02:56 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-05-12 20:23:02 -07:00
|
|
|
|
mm: memcontrol: charge swap to cgroup2
This patchset introduces swap accounting to cgroup2.
This patch (of 7):
In the legacy hierarchy we charge memsw, which is dubious, because:
- memsw.limit must be >= memory.limit, so it is impossible to limit
swap usage less than memory usage. Taking into account the fact that
the primary limiting mechanism in the unified hierarchy is
memory.high while memory.limit is either left unset or set to a very
large value, moving memsw.limit knob to the unified hierarchy would
effectively make it impossible to limit swap usage according to the
user preference.
- memsw.usage != memory.usage + swap.usage, because a page occupying
both swap entry and a swap cache page is charged only once to memsw
counter. As a result, it is possible to effectively eat up to
memory.limit of memory pages *and* memsw.limit of swap entries, which
looks unexpected.
That said, we should provide a different swap limiting mechanism for
cgroup2.
This patch adds mem_cgroup->swap counter, which charges the actual number
of swap entries used by a cgroup. It is only charged in the unified
hierarchy, while the legacy hierarchy memsw logic is left intact.
The swap usage can be monitored using new memory.swap.current file and
limited using memory.swap.max.
Note, to charge swap resource properly in the unified hierarchy, we have
to make swap_entry_free uncharge swap only when ->usage reaches zero, not
just ->count, i.e. when all references to a swap entry, including the one
taken by swap cache, are gone. This is necessary, because otherwise
swap-in could result in uncharging swap even if the page is still in swap
cache and hence still occupies a swap entry. At the same time, this
shouldn't break memsw counter logic, where a page is never charged twice
for using both memory and swap, because in case of legacy hierarchy we
uncharge swap on commit (see mem_cgroup_commit_charge).
Signed-off-by: Vladimir Davydov <vdavydov@virtuozzo.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-20 15:02:56 -08:00
|
|
|
|
2022-05-12 20:23:02 -07:00
|
|
|
|
mm: memcontrol: charge swap to cgroup2
This patchset introduces swap accounting to cgroup2.
This patch (of 7):
In the legacy hierarchy we charge memsw, which is dubious, because:
- memsw.limit must be >= memory.limit, so it is impossible to limit
swap usage less than memory usage. Taking into account the fact that
the primary limiting mechanism in the unified hierarchy is
memory.high while memory.limit is either left unset or set to a very
large value, moving memsw.limit knob to the unified hierarchy would
effectively make it impossible to limit swap usage according to the
user preference.
- memsw.usage != memory.usage + swap.usage, because a page occupying
both swap entry and a swap cache page is charged only once to memsw
counter. As a result, it is possible to effectively eat up to
memory.limit of memory pages *and* memsw.limit of swap entries, which
looks unexpected.
That said, we should provide a different swap limiting mechanism for
cgroup2.
This patch adds mem_cgroup->swap counter, which charges the actual number
of swap entries used by a cgroup. It is only charged in the unified
hierarchy, while the legacy hierarchy memsw logic is left intact.
The swap usage can be monitored using new memory.swap.current file and
limited using memory.swap.max.
Note, to charge swap resource properly in the unified hierarchy, we have
to make swap_entry_free uncharge swap only when ->usage reaches zero, not
just ->count, i.e. when all references to a swap entry, including the one
taken by swap cache, are gone. This is necessary, because otherwise
swap-in could result in uncharging swap even if the page is still in swap
cache and hence still occupies a swap entry. At the same time, this
shouldn't break memsw counter logic, where a page is never charged twice
for using both memory and swap, because in case of legacy hierarchy we
uncharge swap on commit (see mem_cgroup_commit_charge).
Signed-off-by: Vladimir Davydov <vdavydov@virtuozzo.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-20 15:02:56 -08:00
|
|
|
|
mm, THP, swap: delay splitting THP during swap out
Patch series "THP swap: Delay splitting THP during swapping out", v11.
This patchset is to optimize the performance of Transparent Huge Page
(THP) swap.
Recently, the performance of the storage devices improved so fast that
we cannot saturate the disk bandwidth with single logical CPU when do
page swap out even on a high-end server machine. Because the
performance of the storage device improved faster than that of single
logical CPU. And it seems that the trend will not change in the near
future. On the other hand, the THP becomes more and more popular
because of increased memory size. So it becomes necessary to optimize
THP swap performance.
The advantages of the THP swap support include:
- Batch the swap operations for the THP to reduce lock
acquiring/releasing, including allocating/freeing the swap space,
adding/deleting to/from the swap cache, and writing/reading the swap
space, etc. This will help improve the performance of the THP swap.
- The THP swap space read/write will be 2M sequential IO. It is
particularly helpful for the swap read, which are usually 4k random
IO. This will improve the performance of the THP swap too.
- It will help the memory fragmentation, especially when the THP is
heavily used by the applications. The 2M continuous pages will be
free up after THP swapping out.
- It will improve the THP utilization on the system with the swap
turned on. Because the speed for khugepaged to collapse the normal
pages into the THP is quite slow. After the THP is split during the
swapping out, it will take quite long time for the normal pages to
collapse back into the THP after being swapped in. The high THP
utilization helps the efficiency of the page based memory management
too.
There are some concerns regarding THP swap in, mainly because possible
enlarged read/write IO size (for swap in/out) may put more overhead on
the storage device. To deal with that, the THP swap in should be turned
on only when necessary. For example, it can be selected via
"always/never/madvise" logic, to be turned on globally, turned off
globally, or turned on only for VMA with MADV_HUGEPAGE, etc.
This patchset is the first step for the THP swap support. The plan is
to delay splitting THP step by step, finally avoid splitting THP during
the THP swapping out and swap out/in the THP as a whole.
As the first step, in this patchset, the splitting huge page is delayed
from almost the first step of swapping out to after allocating the swap
space for the THP and adding the THP into the swap cache. This will
reduce lock acquiring/releasing for the locks used for the swap cache
management.
With the patchset, the swap out throughput improves 15.5% (from about
3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case
with 8 processes. The test is done on a Xeon E5 v3 system. The swap
device used is a RAM simulated PMEM (persistent memory) device. To test
the sequential swapping out, the test case creates 8 processes, which
sequentially allocate and write to the anonymous pages until the RAM and
part of the swap device is used up.
This patch (of 5):
In this patch, splitting huge page is delayed from almost the first step
of swapping out to after allocating the swap space for the THP
(Transparent Huge Page) and adding the THP into the swap cache. This
will batch the corresponding operation, thus improve THP swap out
throughput.
This is the first step for the THP swap optimization. The plan is to
delay splitting the THP step by step and avoid splitting the THP
finally.
In this patch, one swap cluster is used to hold the contents of each THP
swapped out. So, the size of the swap cluster is changed to that of the
THP (Transparent Huge Page) on x86_64 architecture (512). For other
architectures which want such THP swap optimization,
ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for
the architecture. In effect, this will enlarge swap cluster size by 2
times on x86_64. Which may make it harder to find a free cluster when
the swap space becomes fragmented. So that, this may reduce the
continuous swap space allocation and sequential write in theory. The
performance test in 0day shows no regressions caused by this.
In the future of THP swap optimization, some information of the swapped
out THP (such as compound map count) will be recorded in the
swap_cluster_info data structure.
The mem cgroup swap accounting functions are enhanced to support charge
or uncharge a swap cluster backing a THP as a whole.
The swap cluster allocate/free functions are added to allocate/free a
swap cluster for a THP. A fair simple algorithm is used for swap
cluster allocation, that is, only the first swap device in priority list
will be tried to allocate the swap cluster. The function will fail if
the trying is not successful, and the caller will fallback to allocate a
single swap slot instead. This works good enough for normal cases. If
the difference of the number of the free swap clusters among multiple
swap devices is significant, it is possible that some THPs are split
earlier than necessary. For example, this could be caused by big size
difference among multiple swap devices.
The swap cache functions is enhanced to support add/delete THP to/from
the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be
enhanced in the future with multi-order radix tree. But because we will
split the THP soon during swapping out, that optimization doesn't make
much sense for this first step.
The THP splitting functions are enhanced to support to split THP in swap
cache during swapping out. The page lock will be held during allocating
the swap cluster, adding the THP into the swap cache and splitting the
THP. So in the code path other than swapping out, if the THP need to be
split, the PageSwapCache(THP) will be always false.
The swap cluster is only available for SSD, so the THP swap optimization
in this patchset has no effect for HDD.
[ying.huang@intel.com: fix two issues in THP optimize patch]
Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com
[hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size]
Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option]
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h]
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
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>
2017-07-06 15:37:18 -07:00
|
|
|
|
mm: memcontrol: charge swap to cgroup2
This patchset introduces swap accounting to cgroup2.
This patch (of 7):
In the legacy hierarchy we charge memsw, which is dubious, because:
- memsw.limit must be >= memory.limit, so it is impossible to limit
swap usage less than memory usage. Taking into account the fact that
the primary limiting mechanism in the unified hierarchy is
memory.high while memory.limit is either left unset or set to a very
large value, moving memsw.limit knob to the unified hierarchy would
effectively make it impossible to limit swap usage according to the
user preference.
- memsw.usage != memory.usage + swap.usage, because a page occupying
both swap entry and a swap cache page is charged only once to memsw
counter. As a result, it is possible to effectively eat up to
memory.limit of memory pages *and* memsw.limit of swap entries, which
looks unexpected.
That said, we should provide a different swap limiting mechanism for
cgroup2.
This patch adds mem_cgroup->swap counter, which charges the actual number
of swap entries used by a cgroup. It is only charged in the unified
hierarchy, while the legacy hierarchy memsw logic is left intact.
The swap usage can be monitored using new memory.swap.current file and
limited using memory.swap.max.
Note, to charge swap resource properly in the unified hierarchy, we have
to make swap_entry_free uncharge swap only when ->usage reaches zero, not
just ->count, i.e. when all references to a swap entry, including the one
taken by swap cache, are gone. This is necessary, because otherwise
swap-in could result in uncharging swap even if the page is still in swap
cache and hence still occupies a swap entry. At the same time, this
shouldn't break memsw counter logic, where a page is never charged twice
for using both memory and swap, because in case of legacy hierarchy we
uncharge swap on commit (see mem_cgroup_commit_charge).
Signed-off-by: Vladimir Davydov <vdavydov@virtuozzo.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-20 15:02:56 -08:00
|
|
|
|
2022-09-26 09:57:03 -04:00
|
|
|
|
mm: memcontrol: charge swap to cgroup2
This patchset introduces swap accounting to cgroup2.
This patch (of 7):
In the legacy hierarchy we charge memsw, which is dubious, because:
- memsw.limit must be >= memory.limit, so it is impossible to limit
swap usage less than memory usage. Taking into account the fact that
the primary limiting mechanism in the unified hierarchy is
memory.high while memory.limit is either left unset or set to a very
large value, moving memsw.limit knob to the unified hierarchy would
effectively make it impossible to limit swap usage according to the
user preference.
- memsw.usage != memory.usage + swap.usage, because a page occupying
both swap entry and a swap cache page is charged only once to memsw
counter. As a result, it is possible to effectively eat up to
memory.limit of memory pages *and* memsw.limit of swap entries, which
looks unexpected.
That said, we should provide a different swap limiting mechanism for
cgroup2.
This patch adds mem_cgroup->swap counter, which charges the actual number
of swap entries used by a cgroup. It is only charged in the unified
hierarchy, while the legacy hierarchy memsw logic is left intact.
The swap usage can be monitored using new memory.swap.current file and
limited using memory.swap.max.
Note, to charge swap resource properly in the unified hierarchy, we have
to make swap_entry_free uncharge swap only when ->usage reaches zero, not
just ->count, i.e. when all references to a swap entry, including the one
taken by swap cache, are gone. This is necessary, because otherwise
swap-in could result in uncharging swap even if the page is still in swap
cache and hence still occupies a swap entry. At the same time, this
shouldn't break memsw counter logic, where a page is never charged twice
for using both memory and swap, because in case of legacy hierarchy we
uncharge swap on commit (see mem_cgroup_commit_charge).
Signed-off-by: Vladimir Davydov <vdavydov@virtuozzo.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-20 15:02:56 -08:00
|
|
|
|
|
|
|
|
|
2022-05-12 20:23:02 -07:00
|
|
|
|
mm: memcontrol: charge swap to cgroup2
This patchset introduces swap accounting to cgroup2.
This patch (of 7):
In the legacy hierarchy we charge memsw, which is dubious, because:
- memsw.limit must be >= memory.limit, so it is impossible to limit
swap usage less than memory usage. Taking into account the fact that
the primary limiting mechanism in the unified hierarchy is
memory.high while memory.limit is either left unset or set to a very
large value, moving memsw.limit knob to the unified hierarchy would
effectively make it impossible to limit swap usage according to the
user preference.
- memsw.usage != memory.usage + swap.usage, because a page occupying
both swap entry and a swap cache page is charged only once to memsw
counter. As a result, it is possible to effectively eat up to
memory.limit of memory pages *and* memsw.limit of swap entries, which
looks unexpected.
That said, we should provide a different swap limiting mechanism for
cgroup2.
This patch adds mem_cgroup->swap counter, which charges the actual number
of swap entries used by a cgroup. It is only charged in the unified
hierarchy, while the legacy hierarchy memsw logic is left intact.
The swap usage can be monitored using new memory.swap.current file and
limited using memory.swap.max.
Note, to charge swap resource properly in the unified hierarchy, we have
to make swap_entry_free uncharge swap only when ->usage reaches zero, not
just ->count, i.e. when all references to a swap entry, including the one
taken by swap cache, are gone. This is necessary, because otherwise
swap-in could result in uncharging swap even if the page is still in swap
cache and hence still occupies a swap entry. At the same time, this
shouldn't break memsw counter logic, where a page is never charged twice
for using both memory and swap, because in case of legacy hierarchy we
uncharge swap on commit (see mem_cgroup_commit_charge).
Signed-off-by: Vladimir Davydov <vdavydov@virtuozzo.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-20 15:02:56 -08:00
|
|
|
|
2022-05-12 20:23:02 -07:00
|
|
|
|
mm: memcontrol: charge swap to cgroup2
This patchset introduces swap accounting to cgroup2.
This patch (of 7):
In the legacy hierarchy we charge memsw, which is dubious, because:
- memsw.limit must be >= memory.limit, so it is impossible to limit
swap usage less than memory usage. Taking into account the fact that
the primary limiting mechanism in the unified hierarchy is
memory.high while memory.limit is either left unset or set to a very
large value, moving memsw.limit knob to the unified hierarchy would
effectively make it impossible to limit swap usage according to the
user preference.
- memsw.usage != memory.usage + swap.usage, because a page occupying
both swap entry and a swap cache page is charged only once to memsw
counter. As a result, it is possible to effectively eat up to
memory.limit of memory pages *and* memsw.limit of swap entries, which
looks unexpected.
That said, we should provide a different swap limiting mechanism for
cgroup2.
This patch adds mem_cgroup->swap counter, which charges the actual number
of swap entries used by a cgroup. It is only charged in the unified
hierarchy, while the legacy hierarchy memsw logic is left intact.
The swap usage can be monitored using new memory.swap.current file and
limited using memory.swap.max.
Note, to charge swap resource properly in the unified hierarchy, we have
to make swap_entry_free uncharge swap only when ->usage reaches zero, not
just ->count, i.e. when all references to a swap entry, including the one
taken by swap cache, are gone. This is necessary, because otherwise
swap-in could result in uncharging swap even if the page is still in swap
cache and hence still occupies a swap entry. At the same time, this
shouldn't break memsw counter logic, where a page is never charged twice
for using both memory and swap, because in case of legacy hierarchy we
uncharge swap on commit (see mem_cgroup_commit_charge).
Signed-off-by: Vladimir Davydov <vdavydov@virtuozzo.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-20 15:02:56 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-06-07 17:05:35 -07:00
|
|
|
|
|
|
|
|
|
2018-06-07 17:05:31 -07:00
|
|
|
|
2018-06-07 17:05:35 -07:00
|
|
|
|
2018-06-07 17:05:31 -07:00
|
|
|
|
2016-08-11 15:33:00 -07:00
|
|
|
|
|
|
|
|
|
mm: memcontrol: deprecate swapaccounting=0 mode
The swapaccounting= commandline option already does very little today. To
close a trivial containment failure case, the swap ownership tracking part
of the swap controller has recently become mandatory (see commit
2d1c498072de ("mm: memcontrol: make swap tracking an integral part of
memory control") for details), which makes up the majority of the work
during swapout, swapin, and the swap slot map.
The only thing left under this flag is the page_counter operations and the
visibility of the swap control files in the first place, which are rather
meager savings. There also aren't many scenarios, if any, where
controlling the memory of a cgroup while allowing it unlimited access to a
global swap space is a workable resource isolation strategy.
On the other hand, there have been several bugs and confusion around the
many possible swap controller states (cgroup1 vs cgroup2 behavior, memory
accounting without swap accounting, memcg runtime disabled).
This puts the maintenance overhead of retaining the toggle above its
practical benefits. Deprecate it.
Link: https://lkml.kernel.org/r/20220926135704.400818-3-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-26 09:57:02 -04:00
|
|
|
|
mm, THP, swap: delay splitting THP during swap out
Patch series "THP swap: Delay splitting THP during swapping out", v11.
This patchset is to optimize the performance of Transparent Huge Page
(THP) swap.
Recently, the performance of the storage devices improved so fast that
we cannot saturate the disk bandwidth with single logical CPU when do
page swap out even on a high-end server machine. Because the
performance of the storage device improved faster than that of single
logical CPU. And it seems that the trend will not change in the near
future. On the other hand, the THP becomes more and more popular
because of increased memory size. So it becomes necessary to optimize
THP swap performance.
The advantages of the THP swap support include:
- Batch the swap operations for the THP to reduce lock
acquiring/releasing, including allocating/freeing the swap space,
adding/deleting to/from the swap cache, and writing/reading the swap
space, etc. This will help improve the performance of the THP swap.
- The THP swap space read/write will be 2M sequential IO. It is
particularly helpful for the swap read, which are usually 4k random
IO. This will improve the performance of the THP swap too.
- It will help the memory fragmentation, especially when the THP is
heavily used by the applications. The 2M continuous pages will be
free up after THP swapping out.
- It will improve the THP utilization on the system with the swap
turned on. Because the speed for khugepaged to collapse the normal
pages into the THP is quite slow. After the THP is split during the
swapping out, it will take quite long time for the normal pages to
collapse back into the THP after being swapped in. The high THP
utilization helps the efficiency of the page based memory management
too.
There are some concerns regarding THP swap in, mainly because possible
enlarged read/write IO size (for swap in/out) may put more overhead on
the storage device. To deal with that, the THP swap in should be turned
on only when necessary. For example, it can be selected via
"always/never/madvise" logic, to be turned on globally, turned off
globally, or turned on only for VMA with MADV_HUGEPAGE, etc.
This patchset is the first step for the THP swap support. The plan is
to delay splitting THP step by step, finally avoid splitting THP during
the THP swapping out and swap out/in the THP as a whole.
As the first step, in this patchset, the splitting huge page is delayed
from almost the first step of swapping out to after allocating the swap
space for the THP and adding the THP into the swap cache. This will
reduce lock acquiring/releasing for the locks used for the swap cache
management.
With the patchset, the swap out throughput improves 15.5% (from about
3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case
with 8 processes. The test is done on a Xeon E5 v3 system. The swap
device used is a RAM simulated PMEM (persistent memory) device. To test
the sequential swapping out, the test case creates 8 processes, which
sequentially allocate and write to the anonymous pages until the RAM and
part of the swap device is used up.
This patch (of 5):
In this patch, splitting huge page is delayed from almost the first step
of swapping out to after allocating the swap space for the THP
(Transparent Huge Page) and adding the THP into the swap cache. This
will batch the corresponding operation, thus improve THP swap out
throughput.
This is the first step for the THP swap optimization. The plan is to
delay splitting the THP step by step and avoid splitting the THP
finally.
In this patch, one swap cluster is used to hold the contents of each THP
swapped out. So, the size of the swap cluster is changed to that of the
THP (Transparent Huge Page) on x86_64 architecture (512). For other
architectures which want such THP swap optimization,
ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for
the architecture. In effect, this will enlarge swap cluster size by 2
times on x86_64. Which may make it harder to find a free cluster when
the swap space becomes fragmented. So that, this may reduce the
continuous swap space allocation and sequential write in theory. The
performance test in 0day shows no regressions caused by this.
In the future of THP swap optimization, some information of the swapped
out THP (such as compound map count) will be recorded in the
swap_cluster_info data structure.
The mem cgroup swap accounting functions are enhanced to support charge
or uncharge a swap cluster backing a THP as a whole.
The swap cluster allocate/free functions are added to allocate/free a
swap cluster for a THP. A fair simple algorithm is used for swap
cluster allocation, that is, only the first swap device in priority list
will be tried to allocate the swap cluster. The function will fail if
the trying is not successful, and the caller will fallback to allocate a
single swap slot instead. This works good enough for normal cases. If
the difference of the number of the free swap clusters among multiple
swap devices is significant, it is possible that some THPs are split
earlier than necessary. For example, this could be caused by big size
difference among multiple swap devices.
The swap cache functions is enhanced to support add/delete THP to/from
the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be
enhanced in the future with multi-order radix tree. But because we will
split the THP soon during swapping out, that optimization doesn't make
much sense for this first step.
The THP splitting functions are enhanced to support to split THP in swap
cache during swapping out. The page lock will be held during allocating
the swap cluster, adding the THP into the swap cache and splitting the
THP. So in the code path other than swapping out, if the THP need to be
split, the PageSwapCache(THP) will be always false.
The swap cluster is only available for SSD, so the THP swap optimization
in this patchset has no effect for HDD.
[ying.huang@intel.com: fix two issues in THP optimize patch]
Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com
[hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size]
Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option]
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h]
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
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>
2017-07-06 15:37:18 -07:00
|
|
|
|
2018-06-07 17:05:35 -07:00
|
|
|
|
|
|
|
|
|
2016-08-11 15:33:00 -07:00
|
|
|
|
mm: memcontrol: charge swap to cgroup2
This patchset introduces swap accounting to cgroup2.
This patch (of 7):
In the legacy hierarchy we charge memsw, which is dubious, because:
- memsw.limit must be >= memory.limit, so it is impossible to limit
swap usage less than memory usage. Taking into account the fact that
the primary limiting mechanism in the unified hierarchy is
memory.high while memory.limit is either left unset or set to a very
large value, moving memsw.limit knob to the unified hierarchy would
effectively make it impossible to limit swap usage according to the
user preference.
- memsw.usage != memory.usage + swap.usage, because a page occupying
both swap entry and a swap cache page is charged only once to memsw
counter. As a result, it is possible to effectively eat up to
memory.limit of memory pages *and* memsw.limit of swap entries, which
looks unexpected.
That said, we should provide a different swap limiting mechanism for
cgroup2.
This patch adds mem_cgroup->swap counter, which charges the actual number
of swap entries used by a cgroup. It is only charged in the unified
hierarchy, while the legacy hierarchy memsw logic is left intact.
The swap usage can be monitored using new memory.swap.current file and
limited using memory.swap.max.
Note, to charge swap resource properly in the unified hierarchy, we have
to make swap_entry_free uncharge swap only when ->usage reaches zero, not
just ->count, i.e. when all references to a swap entry, including the one
taken by swap cache, are gone. This is necessary, because otherwise
swap-in could result in uncharging swap even if the page is still in swap
cache and hence still occupies a swap entry. At the same time, this
shouldn't break memsw counter logic, where a page is never charged twice
for using both memory and swap, because in case of legacy hierarchy we
uncharge swap on commit (see mem_cgroup_commit_charge).
Signed-off-by: Vladimir Davydov <vdavydov@virtuozzo.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-20 15:02:56 -08:00
|
|
|
|
2016-08-11 15:33:00 -07:00
|
|
|
|
mm: memcontrol: charge swap to cgroup2
This patchset introduces swap accounting to cgroup2.
This patch (of 7):
In the legacy hierarchy we charge memsw, which is dubious, because:
- memsw.limit must be >= memory.limit, so it is impossible to limit
swap usage less than memory usage. Taking into account the fact that
the primary limiting mechanism in the unified hierarchy is
memory.high while memory.limit is either left unset or set to a very
large value, moving memsw.limit knob to the unified hierarchy would
effectively make it impossible to limit swap usage according to the
user preference.
- memsw.usage != memory.usage + swap.usage, because a page occupying
both swap entry and a swap cache page is charged only once to memsw
counter. As a result, it is possible to effectively eat up to
memory.limit of memory pages *and* memsw.limit of swap entries, which
looks unexpected.
That said, we should provide a different swap limiting mechanism for
cgroup2.
This patch adds mem_cgroup->swap counter, which charges the actual number
of swap entries used by a cgroup. It is only charged in the unified
hierarchy, while the legacy hierarchy memsw logic is left intact.
The swap usage can be monitored using new memory.swap.current file and
limited using memory.swap.max.
Note, to charge swap resource properly in the unified hierarchy, we have
to make swap_entry_free uncharge swap only when ->usage reaches zero, not
just ->count, i.e. when all references to a swap entry, including the one
taken by swap cache, are gone. This is necessary, because otherwise
swap-in could result in uncharging swap even if the page is still in swap
cache and hence still occupies a swap entry. At the same time, this
shouldn't break memsw counter logic, where a page is never charged twice
for using both memory and swap, because in case of legacy hierarchy we
uncharge swap on commit (see mem_cgroup_commit_charge).
Signed-off-by: Vladimir Davydov <vdavydov@virtuozzo.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-20 15:02:56 -08:00
|
|
|
|
mm, THP, swap: delay splitting THP during swap out
Patch series "THP swap: Delay splitting THP during swapping out", v11.
This patchset is to optimize the performance of Transparent Huge Page
(THP) swap.
Recently, the performance of the storage devices improved so fast that
we cannot saturate the disk bandwidth with single logical CPU when do
page swap out even on a high-end server machine. Because the
performance of the storage device improved faster than that of single
logical CPU. And it seems that the trend will not change in the near
future. On the other hand, the THP becomes more and more popular
because of increased memory size. So it becomes necessary to optimize
THP swap performance.
The advantages of the THP swap support include:
- Batch the swap operations for the THP to reduce lock
acquiring/releasing, including allocating/freeing the swap space,
adding/deleting to/from the swap cache, and writing/reading the swap
space, etc. This will help improve the performance of the THP swap.
- The THP swap space read/write will be 2M sequential IO. It is
particularly helpful for the swap read, which are usually 4k random
IO. This will improve the performance of the THP swap too.
- It will help the memory fragmentation, especially when the THP is
heavily used by the applications. The 2M continuous pages will be
free up after THP swapping out.
- It will improve the THP utilization on the system with the swap
turned on. Because the speed for khugepaged to collapse the normal
pages into the THP is quite slow. After the THP is split during the
swapping out, it will take quite long time for the normal pages to
collapse back into the THP after being swapped in. The high THP
utilization helps the efficiency of the page based memory management
too.
There are some concerns regarding THP swap in, mainly because possible
enlarged read/write IO size (for swap in/out) may put more overhead on
the storage device. To deal with that, the THP swap in should be turned
on only when necessary. For example, it can be selected via
"always/never/madvise" logic, to be turned on globally, turned off
globally, or turned on only for VMA with MADV_HUGEPAGE, etc.
This patchset is the first step for the THP swap support. The plan is
to delay splitting THP step by step, finally avoid splitting THP during
the THP swapping out and swap out/in the THP as a whole.
As the first step, in this patchset, the splitting huge page is delayed
from almost the first step of swapping out to after allocating the swap
space for the THP and adding the THP into the swap cache. This will
reduce lock acquiring/releasing for the locks used for the swap cache
management.
With the patchset, the swap out throughput improves 15.5% (from about
3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case
with 8 processes. The test is done on a Xeon E5 v3 system. The swap
device used is a RAM simulated PMEM (persistent memory) device. To test
the sequential swapping out, the test case creates 8 processes, which
sequentially allocate and write to the anonymous pages until the RAM and
part of the swap device is used up.
This patch (of 5):
In this patch, splitting huge page is delayed from almost the first step
of swapping out to after allocating the swap space for the THP
(Transparent Huge Page) and adding the THP into the swap cache. This
will batch the corresponding operation, thus improve THP swap out
throughput.
This is the first step for the THP swap optimization. The plan is to
delay splitting the THP step by step and avoid splitting the THP
finally.
In this patch, one swap cluster is used to hold the contents of each THP
swapped out. So, the size of the swap cluster is changed to that of the
THP (Transparent Huge Page) on x86_64 architecture (512). For other
architectures which want such THP swap optimization,
ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for
the architecture. In effect, this will enlarge swap cluster size by 2
times on x86_64. Which may make it harder to find a free cluster when
the swap space becomes fragmented. So that, this may reduce the
continuous swap space allocation and sequential write in theory. The
performance test in 0day shows no regressions caused by this.
In the future of THP swap optimization, some information of the swapped
out THP (such as compound map count) will be recorded in the
swap_cluster_info data structure.
The mem cgroup swap accounting functions are enhanced to support charge
or uncharge a swap cluster backing a THP as a whole.
The swap cluster allocate/free functions are added to allocate/free a
swap cluster for a THP. A fair simple algorithm is used for swap
cluster allocation, that is, only the first swap device in priority list
will be tried to allocate the swap cluster. The function will fail if
the trying is not successful, and the caller will fallback to allocate a
single swap slot instead. This works good enough for normal cases. If
the difference of the number of the free swap clusters among multiple
swap devices is significant, it is possible that some THPs are split
earlier than necessary. For example, this could be caused by big size
difference among multiple swap devices.
The swap cache functions is enhanced to support add/delete THP to/from
the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be
enhanced in the future with multi-order radix tree. But because we will
split the THP soon during swapping out, that optimization doesn't make
much sense for this first step.
The THP splitting functions are enhanced to support to split THP in swap
cache during swapping out. The page lock will be held during allocating
the swap cluster, adding the THP into the swap cache and splitting the
THP. So in the code path other than swapping out, if the THP need to be
split, the PageSwapCache(THP) will be always false.
The swap cluster is only available for SSD, so the THP swap optimization
in this patchset has no effect for HDD.
[ying.huang@intel.com: fix two issues in THP optimize patch]
Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com
[hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size]
Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option]
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h]
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
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>
2017-07-06 15:37:18 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-01-31 16:16:37 -08:00
|
|
|
|
mm: memcontrol: charge swap to cgroup2
This patchset introduces swap accounting to cgroup2.
This patch (of 7):
In the legacy hierarchy we charge memsw, which is dubious, because:
- memsw.limit must be >= memory.limit, so it is impossible to limit
swap usage less than memory usage. Taking into account the fact that
the primary limiting mechanism in the unified hierarchy is
memory.high while memory.limit is either left unset or set to a very
large value, moving memsw.limit knob to the unified hierarchy would
effectively make it impossible to limit swap usage according to the
user preference.
- memsw.usage != memory.usage + swap.usage, because a page occupying
both swap entry and a swap cache page is charged only once to memsw
counter. As a result, it is possible to effectively eat up to
memory.limit of memory pages *and* memsw.limit of swap entries, which
looks unexpected.
That said, we should provide a different swap limiting mechanism for
cgroup2.
This patch adds mem_cgroup->swap counter, which charges the actual number
of swap entries used by a cgroup. It is only charged in the unified
hierarchy, while the legacy hierarchy memsw logic is left intact.
The swap usage can be monitored using new memory.swap.current file and
limited using memory.swap.max.
Note, to charge swap resource properly in the unified hierarchy, we have
to make swap_entry_free uncharge swap only when ->usage reaches zero, not
just ->count, i.e. when all references to a swap entry, including the one
taken by swap cache, are gone. This is necessary, because otherwise
swap-in could result in uncharging swap even if the page is still in swap
cache and hence still occupies a swap entry. At the same time, this
shouldn't break memsw counter logic, where a page is never charged twice
for using both memory and swap, because in case of legacy hierarchy we
uncharge swap on commit (see mem_cgroup_commit_charge).
Signed-off-by: Vladimir Davydov <vdavydov@virtuozzo.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-20 15:02:56 -08:00
|
|
|
|
2025-03-06 10:31:33 +08:00
|
|
|
|
2024-12-18 19:46:33 +08:00
|
|
|
|
mm: memcontrol: charge swap to cgroup2
This patchset introduces swap accounting to cgroup2.
This patch (of 7):
In the legacy hierarchy we charge memsw, which is dubious, because:
- memsw.limit must be >= memory.limit, so it is impossible to limit
swap usage less than memory usage. Taking into account the fact that
the primary limiting mechanism in the unified hierarchy is
memory.high while memory.limit is either left unset or set to a very
large value, moving memsw.limit knob to the unified hierarchy would
effectively make it impossible to limit swap usage according to the
user preference.
- memsw.usage != memory.usage + swap.usage, because a page occupying
both swap entry and a swap cache page is charged only once to memsw
counter. As a result, it is possible to effectively eat up to
memory.limit of memory pages *and* memsw.limit of swap entries, which
looks unexpected.
That said, we should provide a different swap limiting mechanism for
cgroup2.
This patch adds mem_cgroup->swap counter, which charges the actual number
of swap entries used by a cgroup. It is only charged in the unified
hierarchy, while the legacy hierarchy memsw logic is left intact.
The swap usage can be monitored using new memory.swap.current file and
limited using memory.swap.max.
Note, to charge swap resource properly in the unified hierarchy, we have
to make swap_entry_free uncharge swap only when ->usage reaches zero, not
just ->count, i.e. when all references to a swap entry, including the one
taken by swap cache, are gone. This is necessary, because otherwise
swap-in could result in uncharging swap even if the page is still in swap
cache and hence still occupies a swap entry. At the same time, this
shouldn't break memsw counter logic, where a page is never charged twice
for using both memory and swap, because in case of legacy hierarchy we
uncharge swap on commit (see mem_cgroup_commit_charge).
Signed-off-by: Vladimir Davydov <vdavydov@virtuozzo.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-20 15:02:56 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-02-11 15:26:36 -08:00
|
|
|
|
2021-09-02 14:54:54 -07:00
|
|
|
|
2015-02-11 15:26:36 -08:00
|
|
|
|
mm, THP, swap: delay splitting THP during swap out
Patch series "THP swap: Delay splitting THP during swapping out", v11.
This patchset is to optimize the performance of Transparent Huge Page
(THP) swap.
Recently, the performance of the storage devices improved so fast that
we cannot saturate the disk bandwidth with single logical CPU when do
page swap out even on a high-end server machine. Because the
performance of the storage device improved faster than that of single
logical CPU. And it seems that the trend will not change in the near
future. On the other hand, the THP becomes more and more popular
because of increased memory size. So it becomes necessary to optimize
THP swap performance.
The advantages of the THP swap support include:
- Batch the swap operations for the THP to reduce lock
acquiring/releasing, including allocating/freeing the swap space,
adding/deleting to/from the swap cache, and writing/reading the swap
space, etc. This will help improve the performance of the THP swap.
- The THP swap space read/write will be 2M sequential IO. It is
particularly helpful for the swap read, which are usually 4k random
IO. This will improve the performance of the THP swap too.
- It will help the memory fragmentation, especially when the THP is
heavily used by the applications. The 2M continuous pages will be
free up after THP swapping out.
- It will improve the THP utilization on the system with the swap
turned on. Because the speed for khugepaged to collapse the normal
pages into the THP is quite slow. After the THP is split during the
swapping out, it will take quite long time for the normal pages to
collapse back into the THP after being swapped in. The high THP
utilization helps the efficiency of the page based memory management
too.
There are some concerns regarding THP swap in, mainly because possible
enlarged read/write IO size (for swap in/out) may put more overhead on
the storage device. To deal with that, the THP swap in should be turned
on only when necessary. For example, it can be selected via
"always/never/madvise" logic, to be turned on globally, turned off
globally, or turned on only for VMA with MADV_HUGEPAGE, etc.
This patchset is the first step for the THP swap support. The plan is
to delay splitting THP step by step, finally avoid splitting THP during
the THP swapping out and swap out/in the THP as a whole.
As the first step, in this patchset, the splitting huge page is delayed
from almost the first step of swapping out to after allocating the swap
space for the THP and adding the THP into the swap cache. This will
reduce lock acquiring/releasing for the locks used for the swap cache
management.
With the patchset, the swap out throughput improves 15.5% (from about
3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case
with 8 processes. The test is done on a Xeon E5 v3 system. The swap
device used is a RAM simulated PMEM (persistent memory) device. To test
the sequential swapping out, the test case creates 8 processes, which
sequentially allocate and write to the anonymous pages until the RAM and
part of the swap device is used up.
This patch (of 5):
In this patch, splitting huge page is delayed from almost the first step
of swapping out to after allocating the swap space for the THP
(Transparent Huge Page) and adding the THP into the swap cache. This
will batch the corresponding operation, thus improve THP swap out
throughput.
This is the first step for the THP swap optimization. The plan is to
delay splitting the THP step by step and avoid splitting the THP
finally.
In this patch, one swap cluster is used to hold the contents of each THP
swapped out. So, the size of the swap cluster is changed to that of the
THP (Transparent Huge Page) on x86_64 architecture (512). For other
architectures which want such THP swap optimization,
ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for
the architecture. In effect, this will enlarge swap cluster size by 2
times on x86_64. Which may make it harder to find a free cluster when
the swap space becomes fragmented. So that, this may reduce the
continuous swap space allocation and sequential write in theory. The
performance test in 0day shows no regressions caused by this.
In the future of THP swap optimization, some information of the swapped
out THP (such as compound map count) will be recorded in the
swap_cluster_info data structure.
The mem cgroup swap accounting functions are enhanced to support charge
or uncharge a swap cluster backing a THP as a whole.
The swap cluster allocate/free functions are added to allocate/free a
swap cluster for a THP. A fair simple algorithm is used for swap
cluster allocation, that is, only the first swap device in priority list
will be tried to allocate the swap cluster. The function will fail if
the trying is not successful, and the caller will fallback to allocate a
single swap slot instead. This works good enough for normal cases. If
the difference of the number of the free swap clusters among multiple
swap devices is significant, it is possible that some THPs are split
earlier than necessary. For example, this could be caused by big size
difference among multiple swap devices.
The swap cache functions is enhanced to support add/delete THP to/from
the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be
enhanced in the future with multi-order radix tree. But because we will
split the THP soon during swapping out, that optimization doesn't make
much sense for this first step.
The THP splitting functions are enhanced to support to split THP in swap
cache during swapping out. The page lock will be held during allocating
the swap cluster, adding the THP into the swap cache and splitting the
THP. So in the code path other than swapping out, if the THP need to be
split, the PageSwapCache(THP) will be always false.
The swap cluster is only available for SSD, so the THP swap optimization
in this patchset has no effect for HDD.
[ying.huang@intel.com: fix two issues in THP optimize patch]
Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com
[hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size]
Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option]
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h]
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
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>
2017-07-06 15:37:18 -07:00
|
|
|
|
2015-02-11 15:26:36 -08:00
|
|
|
|
2021-09-02 14:54:54 -07:00
|
|
|
|
2015-02-11 15:26:36 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-12-18 19:46:33 +08:00
|
|
|
|
2015-02-11 15:26:36 -08:00
|
|
|
|
2015-04-15 16:13:00 -07:00
|
|
|
|
2015-02-11 15:26:36 -08:00
|
|
|
|
mm: memcontrol: deprecate swapaccounting=0 mode
The swapaccounting= commandline option already does very little today. To
close a trivial containment failure case, the swap ownership tracking part
of the swap controller has recently become mandatory (see commit
2d1c498072de ("mm: memcontrol: make swap tracking an integral part of
memory control") for details), which makes up the majority of the work
during swapout, swapin, and the swap slot map.
The only thing left under this flag is the page_counter operations and the
visibility of the swap control files in the first place, which are rather
meager savings. There also aren't many scenarios, if any, where
controlling the memory of a cgroup while allowing it unlimited access to a
global swap space is a workable resource isolation strategy.
On the other hand, there have been several bugs and confusion around the
many possible swap controller states (cgroup1 vs cgroup2 behavior, memory
accounting without swap accounting, memcg runtime disabled).
This puts the maintenance overhead of retaining the toggle above its
practical benefits. Deprecate it.
Link: https://lkml.kernel.org/r/20220926135704.400818-3-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-26 09:57:02 -04:00
|
|
|
|
2022-09-26 09:57:03 -04:00
|
|
|
|
mm, THP, swap: delay splitting THP during swap out
Patch series "THP swap: Delay splitting THP during swapping out", v11.
This patchset is to optimize the performance of Transparent Huge Page
(THP) swap.
Recently, the performance of the storage devices improved so fast that
we cannot saturate the disk bandwidth with single logical CPU when do
page swap out even on a high-end server machine. Because the
performance of the storage device improved faster than that of single
logical CPU. And it seems that the trend will not change in the near
future. On the other hand, the THP becomes more and more popular
because of increased memory size. So it becomes necessary to optimize
THP swap performance.
The advantages of the THP swap support include:
- Batch the swap operations for the THP to reduce lock
acquiring/releasing, including allocating/freeing the swap space,
adding/deleting to/from the swap cache, and writing/reading the swap
space, etc. This will help improve the performance of the THP swap.
- The THP swap space read/write will be 2M sequential IO. It is
particularly helpful for the swap read, which are usually 4k random
IO. This will improve the performance of the THP swap too.
- It will help the memory fragmentation, especially when the THP is
heavily used by the applications. The 2M continuous pages will be
free up after THP swapping out.
- It will improve the THP utilization on the system with the swap
turned on. Because the speed for khugepaged to collapse the normal
pages into the THP is quite slow. After the THP is split during the
swapping out, it will take quite long time for the normal pages to
collapse back into the THP after being swapped in. The high THP
utilization helps the efficiency of the page based memory management
too.
There are some concerns regarding THP swap in, mainly because possible
enlarged read/write IO size (for swap in/out) may put more overhead on
the storage device. To deal with that, the THP swap in should be turned
on only when necessary. For example, it can be selected via
"always/never/madvise" logic, to be turned on globally, turned off
globally, or turned on only for VMA with MADV_HUGEPAGE, etc.
This patchset is the first step for the THP swap support. The plan is
to delay splitting THP step by step, finally avoid splitting THP during
the THP swapping out and swap out/in the THP as a whole.
As the first step, in this patchset, the splitting huge page is delayed
from almost the first step of swapping out to after allocating the swap
space for the THP and adding the THP into the swap cache. This will
reduce lock acquiring/releasing for the locks used for the swap cache
management.
With the patchset, the swap out throughput improves 15.5% (from about
3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case
with 8 processes. The test is done on a Xeon E5 v3 system. The swap
device used is a RAM simulated PMEM (persistent memory) device. To test
the sequential swapping out, the test case creates 8 processes, which
sequentially allocate and write to the anonymous pages until the RAM and
part of the swap device is used up.
This patch (of 5):
In this patch, splitting huge page is delayed from almost the first step
of swapping out to after allocating the swap space for the THP
(Transparent Huge Page) and adding the THP into the swap cache. This
will batch the corresponding operation, thus improve THP swap out
throughput.
This is the first step for the THP swap optimization. The plan is to
delay splitting the THP step by step and avoid splitting the THP
finally.
In this patch, one swap cluster is used to hold the contents of each THP
swapped out. So, the size of the swap cluster is changed to that of the
THP (Transparent Huge Page) on x86_64 architecture (512). For other
architectures which want such THP swap optimization,
ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for
the architecture. In effect, this will enlarge swap cluster size by 2
times on x86_64. Which may make it harder to find a free cluster when
the swap space becomes fragmented. So that, this may reduce the
continuous swap space allocation and sequential write in theory. The
performance test in 0day shows no regressions caused by this.
In the future of THP swap optimization, some information of the swapped
out THP (such as compound map count) will be recorded in the
swap_cluster_info data structure.
The mem cgroup swap accounting functions are enhanced to support charge
or uncharge a swap cluster backing a THP as a whole.
The swap cluster allocate/free functions are added to allocate/free a
swap cluster for a THP. A fair simple algorithm is used for swap
cluster allocation, that is, only the first swap device in priority list
will be tried to allocate the swap cluster. The function will fail if
the trying is not successful, and the caller will fallback to allocate a
single swap slot instead. This works good enough for normal cases. If
the difference of the number of the free swap clusters among multiple
swap devices is significant, it is possible that some THPs are split
earlier than necessary. For example, this could be caused by big size
difference among multiple swap devices.
The swap cache functions is enhanced to support add/delete THP to/from
the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be
enhanced in the future with multi-order radix tree. But because we will
split the THP soon during swapping out, that optimization doesn't make
much sense for this first step.
The THP splitting functions are enhanced to support to split THP in swap
cache during swapping out. The page lock will be held during allocating
the swap cluster, adding the THP into the swap cache and splitting the
THP. So in the code path other than swapping out, if the THP need to be
split, the PageSwapCache(THP) will be always false.
The swap cluster is only available for SSD, so the THP swap optimization
in this patchset has no effect for HDD.
[ying.huang@intel.com: fix two issues in THP optimize patch]
Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com
[hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size]
Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option]
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h]
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
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>
2017-07-06 15:37:18 -07:00
|
|
|
|
2022-09-26 09:57:03 -04:00
|
|
|
|
|
|
|
|
|
mm: memcontrol: charge swap to cgroup2
This patchset introduces swap accounting to cgroup2.
This patch (of 7):
In the legacy hierarchy we charge memsw, which is dubious, because:
- memsw.limit must be >= memory.limit, so it is impossible to limit
swap usage less than memory usage. Taking into account the fact that
the primary limiting mechanism in the unified hierarchy is
memory.high while memory.limit is either left unset or set to a very
large value, moving memsw.limit knob to the unified hierarchy would
effectively make it impossible to limit swap usage according to the
user preference.
- memsw.usage != memory.usage + swap.usage, because a page occupying
both swap entry and a swap cache page is charged only once to memsw
counter. As a result, it is possible to effectively eat up to
memory.limit of memory pages *and* memsw.limit of swap entries, which
looks unexpected.
That said, we should provide a different swap limiting mechanism for
cgroup2.
This patch adds mem_cgroup->swap counter, which charges the actual number
of swap entries used by a cgroup. It is only charged in the unified
hierarchy, while the legacy hierarchy memsw logic is left intact.
The swap usage can be monitored using new memory.swap.current file and
limited using memory.swap.max.
Note, to charge swap resource properly in the unified hierarchy, we have
to make swap_entry_free uncharge swap only when ->usage reaches zero, not
just ->count, i.e. when all references to a swap entry, including the one
taken by swap cache, are gone. This is necessary, because otherwise
swap-in could result in uncharging swap even if the page is still in swap
cache and hence still occupies a swap entry. At the same time, this
shouldn't break memsw counter logic, where a page is never charged twice
for using both memory and swap, because in case of legacy hierarchy we
uncharge swap on commit (see mem_cgroup_commit_charge).
Signed-off-by: Vladimir Davydov <vdavydov@virtuozzo.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-20 15:02:56 -08:00
|
|
|
|
2018-01-31 16:16:37 -08:00
|
|
|
|
mm, THP, swap: delay splitting THP during swap out
Patch series "THP swap: Delay splitting THP during swapping out", v11.
This patchset is to optimize the performance of Transparent Huge Page
(THP) swap.
Recently, the performance of the storage devices improved so fast that
we cannot saturate the disk bandwidth with single logical CPU when do
page swap out even on a high-end server machine. Because the
performance of the storage device improved faster than that of single
logical CPU. And it seems that the trend will not change in the near
future. On the other hand, the THP becomes more and more popular
because of increased memory size. So it becomes necessary to optimize
THP swap performance.
The advantages of the THP swap support include:
- Batch the swap operations for the THP to reduce lock
acquiring/releasing, including allocating/freeing the swap space,
adding/deleting to/from the swap cache, and writing/reading the swap
space, etc. This will help improve the performance of the THP swap.
- The THP swap space read/write will be 2M sequential IO. It is
particularly helpful for the swap read, which are usually 4k random
IO. This will improve the performance of the THP swap too.
- It will help the memory fragmentation, especially when the THP is
heavily used by the applications. The 2M continuous pages will be
free up after THP swapping out.
- It will improve the THP utilization on the system with the swap
turned on. Because the speed for khugepaged to collapse the normal
pages into the THP is quite slow. After the THP is split during the
swapping out, it will take quite long time for the normal pages to
collapse back into the THP after being swapped in. The high THP
utilization helps the efficiency of the page based memory management
too.
There are some concerns regarding THP swap in, mainly because possible
enlarged read/write IO size (for swap in/out) may put more overhead on
the storage device. To deal with that, the THP swap in should be turned
on only when necessary. For example, it can be selected via
"always/never/madvise" logic, to be turned on globally, turned off
globally, or turned on only for VMA with MADV_HUGEPAGE, etc.
This patchset is the first step for the THP swap support. The plan is
to delay splitting THP step by step, finally avoid splitting THP during
the THP swapping out and swap out/in the THP as a whole.
As the first step, in this patchset, the splitting huge page is delayed
from almost the first step of swapping out to after allocating the swap
space for the THP and adding the THP into the swap cache. This will
reduce lock acquiring/releasing for the locks used for the swap cache
management.
With the patchset, the swap out throughput improves 15.5% (from about
3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case
with 8 processes. The test is done on a Xeon E5 v3 system. The swap
device used is a RAM simulated PMEM (persistent memory) device. To test
the sequential swapping out, the test case creates 8 processes, which
sequentially allocate and write to the anonymous pages until the RAM and
part of the swap device is used up.
This patch (of 5):
In this patch, splitting huge page is delayed from almost the first step
of swapping out to after allocating the swap space for the THP
(Transparent Huge Page) and adding the THP into the swap cache. This
will batch the corresponding operation, thus improve THP swap out
throughput.
This is the first step for the THP swap optimization. The plan is to
delay splitting the THP step by step and avoid splitting the THP
finally.
In this patch, one swap cluster is used to hold the contents of each THP
swapped out. So, the size of the swap cluster is changed to that of the
THP (Transparent Huge Page) on x86_64 architecture (512). For other
architectures which want such THP swap optimization,
ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for
the architecture. In effect, this will enlarge swap cluster size by 2
times on x86_64. Which may make it harder to find a free cluster when
the swap space becomes fragmented. So that, this may reduce the
continuous swap space allocation and sequential write in theory. The
performance test in 0day shows no regressions caused by this.
In the future of THP swap optimization, some information of the swapped
out THP (such as compound map count) will be recorded in the
swap_cluster_info data structure.
The mem cgroup swap accounting functions are enhanced to support charge
or uncharge a swap cluster backing a THP as a whole.
The swap cluster allocate/free functions are added to allocate/free a
swap cluster for a THP. A fair simple algorithm is used for swap
cluster allocation, that is, only the first swap device in priority list
will be tried to allocate the swap cluster. The function will fail if
the trying is not successful, and the caller will fallback to allocate a
single swap slot instead. This works good enough for normal cases. If
the difference of the number of the free swap clusters among multiple
swap devices is significant, it is possible that some THPs are split
earlier than necessary. For example, this could be caused by big size
difference among multiple swap devices.
The swap cache functions is enhanced to support add/delete THP to/from
the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be
enhanced in the future with multi-order radix tree. But because we will
split the THP soon during swapping out, that optimization doesn't make
much sense for this first step.
The THP splitting functions are enhanced to support to split THP in swap
cache during swapping out. The page lock will be held during allocating
the swap cluster, adding the THP into the swap cache and splitting the
THP. So in the code path other than swapping out, if the THP need to be
split, the PageSwapCache(THP) will be always false.
The swap cluster is only available for SSD, so the THP swap optimization
in this patchset has no effect for HDD.
[ying.huang@intel.com: fix two issues in THP optimize patch]
Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com
[hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size]
Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option]
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h]
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
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>
2017-07-06 15:37:18 -07:00
|
|
|
|
2015-02-11 15:26:36 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-01-20 15:03:07 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcontrol: deprecate swapaccounting=0 mode
The swapaccounting= commandline option already does very little today. To
close a trivial containment failure case, the swap ownership tracking part
of the swap controller has recently become mandatory (see commit
2d1c498072de ("mm: memcontrol: make swap tracking an integral part of
memory control") for details), which makes up the majority of the work
during swapout, swapin, and the swap slot map.
The only thing left under this flag is the page_counter operations and the
visibility of the swap control files in the first place, which are rather
meager savings. There also aren't many scenarios, if any, where
controlling the memory of a cgroup while allowing it unlimited access to a
global swap space is a workable resource isolation strategy.
On the other hand, there have been several bugs and confusion around the
many possible swap controller states (cgroup1 vs cgroup2 behavior, memory
accounting without swap accounting, memcg runtime disabled).
This puts the maintenance overhead of retaining the toggle above its
practical benefits. Deprecate it.
Link: https://lkml.kernel.org/r/20220926135704.400818-3-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-26 09:57:02 -04:00
|
|
|
|
2016-01-20 15:03:07 -08:00
|
|
|
|
2022-09-30 19:14:33 +05:30
|
|
|
|
2016-01-20 15:03:07 -08:00
|
|
|
|
2018-06-07 17:06:18 -07:00
|
|
|
|
2016-01-20 15:03:07 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-09-02 20:46:43 +01:00
|
|
|
|
2016-01-20 15:03:10 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-09-02 20:46:43 +01:00
|
|
|
|
2016-01-20 15:03:10 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcontrol: deprecate swapaccounting=0 mode
The swapaccounting= commandline option already does very little today. To
close a trivial containment failure case, the swap ownership tracking part
of the swap controller has recently become mandatory (see commit
2d1c498072de ("mm: memcontrol: make swap tracking an integral part of
memory control") for details), which makes up the majority of the work
during swapout, swapin, and the swap slot map.
The only thing left under this flag is the page_counter operations and the
visibility of the swap control files in the first place, which are rather
meager savings. There also aren't many scenarios, if any, where
controlling the memory of a cgroup while allowing it unlimited access to a
global swap space is a workable resource isolation strategy.
On the other hand, there have been several bugs and confusion around the
many possible swap controller states (cgroup1 vs cgroup2 behavior, memory
accounting without swap accounting, memcg runtime disabled).
This puts the maintenance overhead of retaining the toggle above its
practical benefits. Deprecate it.
Link: https://lkml.kernel.org/r/20220926135704.400818-3-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-26 09:57:02 -04:00
|
|
|
|
2016-01-20 15:03:10 -08:00
|
|
|
|
|
|
|
|
|
2022-09-02 20:46:43 +01:00
|
|
|
|
2016-01-20 15:03:10 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-09-30 19:14:33 +05:30
|
|
|
|
2020-06-01 21:49:52 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-01-20 15:03:10 -08:00
|
|
|
|
2020-06-01 21:49:52 -07:00
|
|
|
|
2016-01-20 15:03:10 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-06-03 16:02:11 -07:00
|
|
|
|
2015-02-11 15:26:36 -08:00
|
|
|
|
2024-02-13 03:16:34 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-02-11 15:26:36 -08:00
|
|
|
|
|
|
|
|
|
2020-06-03 16:02:11 -07:00
|
|
|
|
2015-02-11 15:26:36 -08:00
|
|
|
|
mm: memcontrol: charge swap to cgroup2
This patchset introduces swap accounting to cgroup2.
This patch (of 7):
In the legacy hierarchy we charge memsw, which is dubious, because:
- memsw.limit must be >= memory.limit, so it is impossible to limit
swap usage less than memory usage. Taking into account the fact that
the primary limiting mechanism in the unified hierarchy is
memory.high while memory.limit is either left unset or set to a very
large value, moving memsw.limit knob to the unified hierarchy would
effectively make it impossible to limit swap usage according to the
user preference.
- memsw.usage != memory.usage + swap.usage, because a page occupying
both swap entry and a swap cache page is charged only once to memsw
counter. As a result, it is possible to effectively eat up to
memory.limit of memory pages *and* memsw.limit of swap entries, which
looks unexpected.
That said, we should provide a different swap limiting mechanism for
cgroup2.
This patch adds mem_cgroup->swap counter, which charges the actual number
of swap entries used by a cgroup. It is only charged in the unified
hierarchy, while the legacy hierarchy memsw logic is left intact.
The swap usage can be monitored using new memory.swap.current file and
limited using memory.swap.max.
Note, to charge swap resource properly in the unified hierarchy, we have
to make swap_entry_free uncharge swap only when ->usage reaches zero, not
just ->count, i.e. when all references to a swap entry, including the one
taken by swap cache, are gone. This is necessary, because otherwise
swap-in could result in uncharging swap even if the page is still in swap
cache and hence still occupies a swap entry. At the same time, this
shouldn't break memsw counter logic, where a page is never charged twice
for using both memory and swap, because in case of legacy hierarchy we
uncharge swap on commit (see mem_cgroup_commit_charge).
Signed-off-by: Vladimir Davydov <vdavydov@virtuozzo.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-20 15:02:56 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, memcg: cg2 memory{.swap,}.peak write handlers
Patch series "mm, memcg: cg2 memory{.swap,}.peak write handlers", v7.
This patch (of 2):
Other mechanisms for querying the peak memory usage of either a process or
v1 memory cgroup allow for resetting the high watermark. Restore parity
with those mechanisms, but with a less racy API.
For example:
- Any write to memory.max_usage_in_bytes in a cgroup v1 mount resets
the high watermark.
- writing "5" to the clear_refs pseudo-file in a processes's proc
directory resets the peak RSS.
This change is an evolution of a previous patch, which mostly copied the
cgroup v1 behavior, however, there were concerns about races/ownership
issues with a global reset, so instead this change makes the reset
filedescriptor-local.
Writing any non-empty string to the memory.peak and memory.swap.peak
pseudo-files reset the high watermark to the current usage for subsequent
reads through that same FD.
Notably, following Johannes's suggestion, this implementation moves the
O(FDs that have written) behavior onto the FD write(2) path. Instead, on
the page-allocation path, we simply add one additional watermark to
conditionally bump per-hierarchy level in the page-counter.
Additionally, this takes Longman's suggestion of nesting the
page-charging-path checks for the two watermarks to reduce the number of
common-case comparisons.
This behavior is particularly useful for work scheduling systems that need
to track memory usage of worker processes/cgroups per-work-item. Since
memory can't be squeezed like CPU can (the OOM-killer has opinions), these
systems need to track the peak memory usage to compute system/container
fullness when binpacking workitems.
Most notably, Vimeo's use-case involves a system that's doing global
binpacking across many Kubernetes pods/containers, and while we can use
PSI for some local decisions about overload, we strive to avoid packing
workloads too tightly in the first place. To facilitate this, we track
the peak memory usage. However, since we run with long-lived workers (to
amortize startup costs) we need a way to track the high watermark while a
work-item is executing. Polling runs the risk of missing short spikes
that last for timescales below the polling interval, and peak memory
tracking at the cgroup level is otherwise perfect for this use-case.
As this data is used to ensure that binpacked work ends up with sufficient
headroom, this use-case mostly avoids the inaccuracies surrounding
reclaimable memory.
Link: https://lkml.kernel.org/r/20240730231304.761942-1-davidf@vimeo.com
Link: https://lkml.kernel.org/r/20240729143743.34236-1-davidf@vimeo.com
Link: https://lkml.kernel.org/r/20240729143743.34236-2-davidf@vimeo.com
Signed-off-by: David Finkel <davidf@vimeo.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Waiman Long <longman@redhat.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-07-29 10:37:42 -04:00
|
|
|
|
2023-05-24 11:17:33 -07:00
|
|
|
|
mm, memcg: cg2 memory{.swap,}.peak write handlers
Patch series "mm, memcg: cg2 memory{.swap,}.peak write handlers", v7.
This patch (of 2):
Other mechanisms for querying the peak memory usage of either a process or
v1 memory cgroup allow for resetting the high watermark. Restore parity
with those mechanisms, but with a less racy API.
For example:
- Any write to memory.max_usage_in_bytes in a cgroup v1 mount resets
the high watermark.
- writing "5" to the clear_refs pseudo-file in a processes's proc
directory resets the peak RSS.
This change is an evolution of a previous patch, which mostly copied the
cgroup v1 behavior, however, there were concerns about races/ownership
issues with a global reset, so instead this change makes the reset
filedescriptor-local.
Writing any non-empty string to the memory.peak and memory.swap.peak
pseudo-files reset the high watermark to the current usage for subsequent
reads through that same FD.
Notably, following Johannes's suggestion, this implementation moves the
O(FDs that have written) behavior onto the FD write(2) path. Instead, on
the page-allocation path, we simply add one additional watermark to
conditionally bump per-hierarchy level in the page-counter.
Additionally, this takes Longman's suggestion of nesting the
page-charging-path checks for the two watermarks to reduce the number of
common-case comparisons.
This behavior is particularly useful for work scheduling systems that need
to track memory usage of worker processes/cgroups per-work-item. Since
memory can't be squeezed like CPU can (the OOM-killer has opinions), these
systems need to track the peak memory usage to compute system/container
fullness when binpacking workitems.
Most notably, Vimeo's use-case involves a system that's doing global
binpacking across many Kubernetes pods/containers, and while we can use
PSI for some local decisions about overload, we strive to avoid packing
workloads too tightly in the first place. To facilitate this, we track
the peak memory usage. However, since we run with long-lived workers (to
amortize startup costs) we need a way to track the high watermark while a
work-item is executing. Polling runs the risk of missing short spikes
that last for timescales below the polling interval, and peak memory
tracking at the cgroup level is otherwise perfect for this use-case.
As this data is used to ensure that binpacked work ends up with sufficient
headroom, this use-case mostly avoids the inaccuracies surrounding
reclaimable memory.
Link: https://lkml.kernel.org/r/20240730231304.761942-1-davidf@vimeo.com
Link: https://lkml.kernel.org/r/20240729143743.34236-1-davidf@vimeo.com
Link: https://lkml.kernel.org/r/20240729143743.34236-2-davidf@vimeo.com
Signed-off-by: David Finkel <davidf@vimeo.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Waiman Long <longman@redhat.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-07-29 10:37:42 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-05-24 11:17:33 -07:00
|
|
|
|
mm, memcg: cg2 memory{.swap,}.peak write handlers
Patch series "mm, memcg: cg2 memory{.swap,}.peak write handlers", v7.
This patch (of 2):
Other mechanisms for querying the peak memory usage of either a process or
v1 memory cgroup allow for resetting the high watermark. Restore parity
with those mechanisms, but with a less racy API.
For example:
- Any write to memory.max_usage_in_bytes in a cgroup v1 mount resets
the high watermark.
- writing "5" to the clear_refs pseudo-file in a processes's proc
directory resets the peak RSS.
This change is an evolution of a previous patch, which mostly copied the
cgroup v1 behavior, however, there were concerns about races/ownership
issues with a global reset, so instead this change makes the reset
filedescriptor-local.
Writing any non-empty string to the memory.peak and memory.swap.peak
pseudo-files reset the high watermark to the current usage for subsequent
reads through that same FD.
Notably, following Johannes's suggestion, this implementation moves the
O(FDs that have written) behavior onto the FD write(2) path. Instead, on
the page-allocation path, we simply add one additional watermark to
conditionally bump per-hierarchy level in the page-counter.
Additionally, this takes Longman's suggestion of nesting the
page-charging-path checks for the two watermarks to reduce the number of
common-case comparisons.
This behavior is particularly useful for work scheduling systems that need
to track memory usage of worker processes/cgroups per-work-item. Since
memory can't be squeezed like CPU can (the OOM-killer has opinions), these
systems need to track the peak memory usage to compute system/container
fullness when binpacking workitems.
Most notably, Vimeo's use-case involves a system that's doing global
binpacking across many Kubernetes pods/containers, and while we can use
PSI for some local decisions about overload, we strive to avoid packing
workloads too tightly in the first place. To facilitate this, we track
the peak memory usage. However, since we run with long-lived workers (to
amortize startup costs) we need a way to track the high watermark while a
work-item is executing. Polling runs the risk of missing short spikes
that last for timescales below the polling interval, and peak memory
tracking at the cgroup level is otherwise perfect for this use-case.
As this data is used to ensure that binpacked work ends up with sufficient
headroom, this use-case mostly avoids the inaccuracies surrounding
reclaimable memory.
Link: https://lkml.kernel.org/r/20240730231304.761942-1-davidf@vimeo.com
Link: https://lkml.kernel.org/r/20240729143743.34236-1-davidf@vimeo.com
Link: https://lkml.kernel.org/r/20240729143743.34236-2-davidf@vimeo.com
Signed-off-by: David Finkel <davidf@vimeo.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Waiman Long <longman@redhat.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-07-29 10:37:42 -04:00
|
|
|
|
|
|
|
|
|
2023-05-24 11:17:33 -07:00
|
|
|
|
|
|
|
|
|
2020-06-01 21:49:52 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcontrol: charge swap to cgroup2
This patchset introduces swap accounting to cgroup2.
This patch (of 7):
In the legacy hierarchy we charge memsw, which is dubious, because:
- memsw.limit must be >= memory.limit, so it is impossible to limit
swap usage less than memory usage. Taking into account the fact that
the primary limiting mechanism in the unified hierarchy is
memory.high while memory.limit is either left unset or set to a very
large value, moving memsw.limit knob to the unified hierarchy would
effectively make it impossible to limit swap usage according to the
user preference.
- memsw.usage != memory.usage + swap.usage, because a page occupying
both swap entry and a swap cache page is charged only once to memsw
counter. As a result, it is possible to effectively eat up to
memory.limit of memory pages *and* memsw.limit of swap entries, which
looks unexpected.
That said, we should provide a different swap limiting mechanism for
cgroup2.
This patch adds mem_cgroup->swap counter, which charges the actual number
of swap entries used by a cgroup. It is only charged in the unified
hierarchy, while the legacy hierarchy memsw logic is left intact.
The swap usage can be monitored using new memory.swap.current file and
limited using memory.swap.max.
Note, to charge swap resource properly in the unified hierarchy, we have
to make swap_entry_free uncharge swap only when ->usage reaches zero, not
just ->count, i.e. when all references to a swap entry, including the one
taken by swap cache, are gone. This is necessary, because otherwise
swap-in could result in uncharging swap even if the page is still in swap
cache and hence still occupies a swap entry. At the same time, this
shouldn't break memsw counter logic, where a page is never charged twice
for using both memory and swap, because in case of legacy hierarchy we
uncharge swap on commit (see mem_cgroup_commit_charge).
Signed-off-by: Vladimir Davydov <vdavydov@virtuozzo.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-20 15:02:56 -08:00
|
|
|
|
|
|
|
|
|
2019-03-05 15:45:55 -08:00
|
|
|
|
|
|
|
|
|
mm: memcontrol: charge swap to cgroup2
This patchset introduces swap accounting to cgroup2.
This patch (of 7):
In the legacy hierarchy we charge memsw, which is dubious, because:
- memsw.limit must be >= memory.limit, so it is impossible to limit
swap usage less than memory usage. Taking into account the fact that
the primary limiting mechanism in the unified hierarchy is
memory.high while memory.limit is either left unset or set to a very
large value, moving memsw.limit knob to the unified hierarchy would
effectively make it impossible to limit swap usage according to the
user preference.
- memsw.usage != memory.usage + swap.usage, because a page occupying
both swap entry and a swap cache page is charged only once to memsw
counter. As a result, it is possible to effectively eat up to
memory.limit of memory pages *and* memsw.limit of swap entries, which
looks unexpected.
That said, we should provide a different swap limiting mechanism for
cgroup2.
This patch adds mem_cgroup->swap counter, which charges the actual number
of swap entries used by a cgroup. It is only charged in the unified
hierarchy, while the legacy hierarchy memsw logic is left intact.
The swap usage can be monitored using new memory.swap.current file and
limited using memory.swap.max.
Note, to charge swap resource properly in the unified hierarchy, we have
to make swap_entry_free uncharge swap only when ->usage reaches zero, not
just ->count, i.e. when all references to a swap entry, including the one
taken by swap cache, are gone. This is necessary, because otherwise
swap-in could result in uncharging swap even if the page is still in swap
cache and hence still occupies a swap entry. At the same time, this
shouldn't break memsw counter logic, where a page is never charged twice
for using both memory and swap, because in case of legacy hierarchy we
uncharge swap on commit (see mem_cgroup_commit_charge).
Signed-off-by: Vladimir Davydov <vdavydov@virtuozzo.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-20 15:02:56 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-06-07 17:09:21 -07:00
|
|
|
|
mm: memcontrol: charge swap to cgroup2
This patchset introduces swap accounting to cgroup2.
This patch (of 7):
In the legacy hierarchy we charge memsw, which is dubious, because:
- memsw.limit must be >= memory.limit, so it is impossible to limit
swap usage less than memory usage. Taking into account the fact that
the primary limiting mechanism in the unified hierarchy is
memory.high while memory.limit is either left unset or set to a very
large value, moving memsw.limit knob to the unified hierarchy would
effectively make it impossible to limit swap usage according to the
user preference.
- memsw.usage != memory.usage + swap.usage, because a page occupying
both swap entry and a swap cache page is charged only once to memsw
counter. As a result, it is possible to effectively eat up to
memory.limit of memory pages *and* memsw.limit of swap entries, which
looks unexpected.
That said, we should provide a different swap limiting mechanism for
cgroup2.
This patch adds mem_cgroup->swap counter, which charges the actual number
of swap entries used by a cgroup. It is only charged in the unified
hierarchy, while the legacy hierarchy memsw logic is left intact.
The swap usage can be monitored using new memory.swap.current file and
limited using memory.swap.max.
Note, to charge swap resource properly in the unified hierarchy, we have
to make swap_entry_free uncharge swap only when ->usage reaches zero, not
just ->count, i.e. when all references to a swap entry, including the one
taken by swap cache, are gone. This is necessary, because otherwise
swap-in could result in uncharging swap even if the page is still in swap
cache and hence still occupies a swap entry. At the same time, this
shouldn't break memsw counter logic, where a page is never charged twice
for using both memory and swap, because in case of legacy hierarchy we
uncharge swap on commit (see mem_cgroup_commit_charge).
Signed-off-by: Vladimir Davydov <vdavydov@virtuozzo.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-20 15:02:56 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-06-07 17:05:35 -07:00
|
|
|
|
|
|
|
|
|
2019-03-05 15:45:52 -08:00
|
|
|
|
2018-06-07 17:05:35 -07:00
|
|
|
|
2020-06-01 21:49:52 -07:00
|
|
|
|
|
|
|
|
|
2018-06-07 17:05:35 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcontrol: charge swap to cgroup2
This patchset introduces swap accounting to cgroup2.
This patch (of 7):
In the legacy hierarchy we charge memsw, which is dubious, because:
- memsw.limit must be >= memory.limit, so it is impossible to limit
swap usage less than memory usage. Taking into account the fact that
the primary limiting mechanism in the unified hierarchy is
memory.high while memory.limit is either left unset or set to a very
large value, moving memsw.limit knob to the unified hierarchy would
effectively make it impossible to limit swap usage according to the
user preference.
- memsw.usage != memory.usage + swap.usage, because a page occupying
both swap entry and a swap cache page is charged only once to memsw
counter. As a result, it is possible to effectively eat up to
memory.limit of memory pages *and* memsw.limit of swap entries, which
looks unexpected.
That said, we should provide a different swap limiting mechanism for
cgroup2.
This patch adds mem_cgroup->swap counter, which charges the actual number
of swap entries used by a cgroup. It is only charged in the unified
hierarchy, while the legacy hierarchy memsw logic is left intact.
The swap usage can be monitored using new memory.swap.current file and
limited using memory.swap.max.
Note, to charge swap resource properly in the unified hierarchy, we have
to make swap_entry_free uncharge swap only when ->usage reaches zero, not
just ->count, i.e. when all references to a swap entry, including the one
taken by swap cache, are gone. This is necessary, because otherwise
swap-in could result in uncharging swap even if the page is still in swap
cache and hence still occupies a swap entry. At the same time, this
shouldn't break memsw counter logic, where a page is never charged twice
for using both memory and swap, because in case of legacy hierarchy we
uncharge swap on commit (see mem_cgroup_commit_charge).
Signed-off-by: Vladimir Davydov <vdavydov@virtuozzo.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-20 15:02:56 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-06-01 21:49:52 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcontrol: charge swap to cgroup2
This patchset introduces swap accounting to cgroup2.
This patch (of 7):
In the legacy hierarchy we charge memsw, which is dubious, because:
- memsw.limit must be >= memory.limit, so it is impossible to limit
swap usage less than memory usage. Taking into account the fact that
the primary limiting mechanism in the unified hierarchy is
memory.high while memory.limit is either left unset or set to a very
large value, moving memsw.limit knob to the unified hierarchy would
effectively make it impossible to limit swap usage according to the
user preference.
- memsw.usage != memory.usage + swap.usage, because a page occupying
both swap entry and a swap cache page is charged only once to memsw
counter. As a result, it is possible to effectively eat up to
memory.limit of memory pages *and* memsw.limit of swap entries, which
looks unexpected.
That said, we should provide a different swap limiting mechanism for
cgroup2.
This patch adds mem_cgroup->swap counter, which charges the actual number
of swap entries used by a cgroup. It is only charged in the unified
hierarchy, while the legacy hierarchy memsw logic is left intact.
The swap usage can be monitored using new memory.swap.current file and
limited using memory.swap.max.
Note, to charge swap resource properly in the unified hierarchy, we have
to make swap_entry_free uncharge swap only when ->usage reaches zero, not
just ->count, i.e. when all references to a swap entry, including the one
taken by swap cache, are gone. This is necessary, because otherwise
swap-in could result in uncharging swap even if the page is still in swap
cache and hence still occupies a swap entry. At the same time, this
shouldn't break memsw counter logic, where a page is never charged twice
for using both memory and swap, because in case of legacy hierarchy we
uncharge swap on commit (see mem_cgroup_commit_charge).
Signed-off-by: Vladimir Davydov <vdavydov@virtuozzo.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-20 15:02:56 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-05-24 11:17:33 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, memcg: cg2 memory{.swap,}.peak write handlers
Patch series "mm, memcg: cg2 memory{.swap,}.peak write handlers", v7.
This patch (of 2):
Other mechanisms for querying the peak memory usage of either a process or
v1 memory cgroup allow for resetting the high watermark. Restore parity
with those mechanisms, but with a less racy API.
For example:
- Any write to memory.max_usage_in_bytes in a cgroup v1 mount resets
the high watermark.
- writing "5" to the clear_refs pseudo-file in a processes's proc
directory resets the peak RSS.
This change is an evolution of a previous patch, which mostly copied the
cgroup v1 behavior, however, there were concerns about races/ownership
issues with a global reset, so instead this change makes the reset
filedescriptor-local.
Writing any non-empty string to the memory.peak and memory.swap.peak
pseudo-files reset the high watermark to the current usage for subsequent
reads through that same FD.
Notably, following Johannes's suggestion, this implementation moves the
O(FDs that have written) behavior onto the FD write(2) path. Instead, on
the page-allocation path, we simply add one additional watermark to
conditionally bump per-hierarchy level in the page-counter.
Additionally, this takes Longman's suggestion of nesting the
page-charging-path checks for the two watermarks to reduce the number of
common-case comparisons.
This behavior is particularly useful for work scheduling systems that need
to track memory usage of worker processes/cgroups per-work-item. Since
memory can't be squeezed like CPU can (the OOM-killer has opinions), these
systems need to track the peak memory usage to compute system/container
fullness when binpacking workitems.
Most notably, Vimeo's use-case involves a system that's doing global
binpacking across many Kubernetes pods/containers, and while we can use
PSI for some local decisions about overload, we strive to avoid packing
workloads too tightly in the first place. To facilitate this, we track
the peak memory usage. However, since we run with long-lived workers (to
amortize startup costs) we need a way to track the high watermark while a
work-item is executing. Polling runs the risk of missing short spikes
that last for timescales below the polling interval, and peak memory
tracking at the cgroup level is otherwise perfect for this use-case.
As this data is used to ensure that binpacked work ends up with sufficient
headroom, this use-case mostly avoids the inaccuracies surrounding
reclaimable memory.
Link: https://lkml.kernel.org/r/20240730231304.761942-1-davidf@vimeo.com
Link: https://lkml.kernel.org/r/20240729143743.34236-1-davidf@vimeo.com
Link: https://lkml.kernel.org/r/20240729143743.34236-2-davidf@vimeo.com
Signed-off-by: David Finkel <davidf@vimeo.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Waiman Long <longman@redhat.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-07-29 10:37:42 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-05-24 11:17:33 -07:00
|
|
|
|
2018-06-07 17:05:35 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcontrol: charge swap to cgroup2
This patchset introduces swap accounting to cgroup2.
This patch (of 7):
In the legacy hierarchy we charge memsw, which is dubious, because:
- memsw.limit must be >= memory.limit, so it is impossible to limit
swap usage less than memory usage. Taking into account the fact that
the primary limiting mechanism in the unified hierarchy is
memory.high while memory.limit is either left unset or set to a very
large value, moving memsw.limit knob to the unified hierarchy would
effectively make it impossible to limit swap usage according to the
user preference.
- memsw.usage != memory.usage + swap.usage, because a page occupying
both swap entry and a swap cache page is charged only once to memsw
counter. As a result, it is possible to effectively eat up to
memory.limit of memory pages *and* memsw.limit of swap entries, which
looks unexpected.
That said, we should provide a different swap limiting mechanism for
cgroup2.
This patch adds mem_cgroup->swap counter, which charges the actual number
of swap entries used by a cgroup. It is only charged in the unified
hierarchy, while the legacy hierarchy memsw logic is left intact.
The swap usage can be monitored using new memory.swap.current file and
limited using memory.swap.max.
Note, to charge swap resource properly in the unified hierarchy, we have
to make swap_entry_free uncharge swap only when ->usage reaches zero, not
just ->count, i.e. when all references to a swap entry, including the one
taken by swap cache, are gone. This is necessary, because otherwise
swap-in could result in uncharging swap even if the page is still in swap
cache and hence still occupies a swap entry. At the same time, this
shouldn't break memsw counter logic, where a page is never charged twice
for using both memory and swap, because in case of legacy hierarchy we
uncharge swap on commit (see mem_cgroup_commit_charge).
Signed-off-by: Vladimir Davydov <vdavydov@virtuozzo.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-20 15:02:56 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-07-01 11:31:15 -04:00
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-23 17:44:05 +05:00
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-09-30 19:14:33 +05:30
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-10-28 19:11:05 -07:00
|
|
|
|
|
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-08-03 20:00:21 +08:00
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
zswap: memcontrol: implement zswap writeback disabling
During our experiment with zswap, we sometimes observe swap IOs due to
occasional zswap store failures and writebacks-to-swap. These swapping
IOs prevent many users who cannot tolerate swapping from adopting zswap to
save memory and improve performance where possible.
This patch adds the option to disable this behavior entirely: do not
writeback to backing swapping device when a zswap store attempt fail, and
do not write pages in the zswap pool back to the backing swap device (both
when the pool is full, and when the new zswap shrinker is called).
This new behavior can be opted-in/out on a per-cgroup basis via a new
cgroup file. By default, writebacks to swap device is enabled, which is
the previous behavior. Initially, writeback is enabled for the root
cgroup, and a newly created cgroup will inherit the current setting of its
parent.
Note that this is subtly different from setting memory.swap.max to 0, as
it still allows for pages to be stored in the zswap pool (which itself
consumes swap space in its current form).
This patch should be applied on top of the zswap shrinker series:
https://lore.kernel.org/linux-mm/20231130194023.4102148-1-nphamcs@gmail.com/
as it also disables the zswap shrinker, a major source of zswap
writebacks.
For the most part, this feature is motivated by internal parties who
have already established their opinions regarding swapping - the
workloads that are highly sensitive to IO, and especially those who are
using servers with really slow disk performance (for instance, massive
but slow HDDs). For these folks, it's impossible to convince them to
even entertain zswap if swapping also comes as a packaged deal.
Writeback disabling is quite a useful feature in these situations - on
a mixed workloads deployment, they can disable writeback for the more
IO-sensitive workloads, and enable writeback for other background
workloads.
For instance, on a server with HDD, I allocate memories and populate
them with random values (so that zswap store will always fail), and
specify memory.high low enough to trigger reclaim. The time it takes
to allocate the memories and just read through it a couple of times
(doing silly things like computing the values' average etc.):
zswap.writeback disabled:
real 0m30.537s
user 0m23.687s
sys 0m6.637s
0 pages swapped in
0 pages swapped out
zswap.writeback enabled:
real 0m45.061s
user 0m24.310s
sys 0m8.892s
712686 pages swapped in
461093 pages swapped out
(the last two lines are from vmstat -s).
[nphamcs@gmail.com: add a comment about recurring zswap store failures leading to reclaim inefficiency]
Link: https://lkml.kernel.org/r/20231221005725.3446672-1-nphamcs@gmail.com
Link: https://lkml.kernel.org/r/20231207192406.3809579-1-nphamcs@gmail.com
Signed-off-by: Nhat Pham <nphamcs@gmail.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Yosry Ahmed <yosryahmed@google.com>
Acked-by: Chris Li <chrisl@kernel.org>
Cc: Dan Streetman <ddstreet@ieee.org>
Cc: David Heidelberg <david@ixit.cz>
Cc: Domenico Cerasuolo <cerasuolodomenico@gmail.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport (IBM) <rppt@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Seth Jennings <sjenning@redhat.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vitaly Wool <vitaly.wool@konsulko.com>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-12-07 11:24:06 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-08-23 16:27:06 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
zswap: memcontrol: implement zswap writeback disabling
During our experiment with zswap, we sometimes observe swap IOs due to
occasional zswap store failures and writebacks-to-swap. These swapping
IOs prevent many users who cannot tolerate swapping from adopting zswap to
save memory and improve performance where possible.
This patch adds the option to disable this behavior entirely: do not
writeback to backing swapping device when a zswap store attempt fail, and
do not write pages in the zswap pool back to the backing swap device (both
when the pool is full, and when the new zswap shrinker is called).
This new behavior can be opted-in/out on a per-cgroup basis via a new
cgroup file. By default, writebacks to swap device is enabled, which is
the previous behavior. Initially, writeback is enabled for the root
cgroup, and a newly created cgroup will inherit the current setting of its
parent.
Note that this is subtly different from setting memory.swap.max to 0, as
it still allows for pages to be stored in the zswap pool (which itself
consumes swap space in its current form).
This patch should be applied on top of the zswap shrinker series:
https://lore.kernel.org/linux-mm/20231130194023.4102148-1-nphamcs@gmail.com/
as it also disables the zswap shrinker, a major source of zswap
writebacks.
For the most part, this feature is motivated by internal parties who
have already established their opinions regarding swapping - the
workloads that are highly sensitive to IO, and especially those who are
using servers with really slow disk performance (for instance, massive
but slow HDDs). For these folks, it's impossible to convince them to
even entertain zswap if swapping also comes as a packaged deal.
Writeback disabling is quite a useful feature in these situations - on
a mixed workloads deployment, they can disable writeback for the more
IO-sensitive workloads, and enable writeback for other background
workloads.
For instance, on a server with HDD, I allocate memories and populate
them with random values (so that zswap store will always fail), and
specify memory.high low enough to trigger reclaim. The time it takes
to allocate the memories and just read through it a couple of times
(doing silly things like computing the values' average etc.):
zswap.writeback disabled:
real 0m30.537s
user 0m23.687s
sys 0m6.637s
0 pages swapped in
0 pages swapped out
zswap.writeback enabled:
real 0m45.061s
user 0m24.310s
sys 0m8.892s
712686 pages swapped in
461093 pages swapped out
(the last two lines are from vmstat -s).
[nphamcs@gmail.com: add a comment about recurring zswap store failures leading to reclaim inefficiency]
Link: https://lkml.kernel.org/r/20231221005725.3446672-1-nphamcs@gmail.com
Link: https://lkml.kernel.org/r/20231207192406.3809579-1-nphamcs@gmail.com
Signed-off-by: Nhat Pham <nphamcs@gmail.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Yosry Ahmed <yosryahmed@google.com>
Acked-by: Chris Li <chrisl@kernel.org>
Cc: Dan Streetman <ddstreet@ieee.org>
Cc: David Heidelberg <david@ixit.cz>
Cc: Domenico Cerasuolo <cerasuolodomenico@gmail.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport (IBM) <rppt@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Seth Jennings <sjenning@redhat.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vitaly Wool <vitaly.wool@konsulko.com>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-12-07 11:24:06 -08:00
|
|
|
|
|
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: memcg: restore subtree stats flushing
Stats flushing for memcg currently follows the following rules:
- Always flush the entire memcg hierarchy (i.e. flush the root).
- Only one flusher is allowed at a time. If someone else tries to flush
concurrently, they skip and return immediately.
- A periodic flusher flushes all the stats every 2 seconds.
The reason this approach is followed is because all flushes are serialized
by a global rstat spinlock. On the memcg side, flushing is invoked from
userspace reads as well as in-kernel flushers (e.g. reclaim, refault,
etc). This approach aims to avoid serializing all flushers on the global
lock, which can cause a significant performance hit under high
concurrency.
This approach has the following problems:
- Occasionally a userspace read of the stats of a non-root cgroup will
be too expensive as it has to flush the entire hierarchy [1].
- Sometimes the stats accuracy are compromised if there is an ongoing
flush, and we skip and return before the subtree of interest is
actually flushed, yielding stale stats (by up to 2s due to periodic
flushing). This is more visible when reading stats from userspace,
but can also affect in-kernel flushers.
The latter problem is particulary a concern when userspace reads stats
after an event occurs, but gets stats from before the event. Examples:
- When memory usage / pressure spikes, a userspace OOM handler may look
at the stats of different memcgs to select a victim based on various
heuristics (e.g. how much private memory will be freed by killing
this). Reading stale stats from before the usage spike in this case
may cause a wrongful OOM kill.
- A proactive reclaimer may read the stats after writing to
memory.reclaim to measure the success of the reclaim operation. Stale
stats from before reclaim may give a false negative.
- Reading the stats of a parent and a child memcg may be inconsistent
(child larger than parent), if the flush doesn't happen when the
parent is read, but happens when the child is read.
As for in-kernel flushers, they will occasionally get stale stats. No
regressions are currently known from this, but if there are regressions,
they would be very difficult to debug and link to the source of the
problem.
This patch aims to fix these problems by restoring subtree flushing, and
removing the unified/coalesced flushing logic that skips flushing if there
is an ongoing flush. This change would introduce a significant regression
with global stats flushing thresholds. With per-memcg stats flushing
thresholds, this seems to perform really well. The thresholds protect the
underlying lock from unnecessary contention.
This patch was tested in two ways to ensure the latency of flushing is
up to par, on a machine with 384 cpus:
- A synthetic test with 5000 concurrent workers in 500 cgroups doing
allocations and reclaim, as well as 1000 readers for memory.stat
(variation of [2]). No regressions were noticed in the total runtime.
Note that significant regressions in this test are observed with
global stats thresholds, but not with per-memcg thresholds.
- A synthetic stress test for concurrently reading memcg stats while
memory allocation/freeing workers are running in the background,
provided by Wei Xu [3]. With 250k threads reading the stats every
100ms in 50k cgroups, 99.9% of reads take <= 50us. Less than 0.01%
of reads take more than 1ms, and no reads take more than 100ms.
[1] https://lore.kernel.org/lkml/CABWYdi0c6__rh-K7dcM_pkf9BJdTRtAU08M43KO9ME4-dsgfoQ@mail.gmail.com/
[2] https://lore.kernel.org/lkml/CAJD7tka13M-zVZTyQJYL1iUAYvuQ1fcHbCjcOBZcz6POYTV-4g@mail.gmail.com/
[3] https://lore.kernel.org/lkml/CAAPL-u9D2b=iF5Lf_cRnKxUfkiEe0AMDTu6yhrUAzX0b6a6rDg@mail.gmail.com/
[akpm@linux-foundation.org: fix mm/zswap.c]
[yosryahmed@google.com: remove stats flushing mutex]
Link: https://lkml.kernel.org/r/CAJD7tkZgP3m-VVPn+fF_YuvXeQYK=tZZjJHj=dzD=CcSSpp2qg@mail.gmail.com
Link: https://lkml.kernel.org/r/20231129032154.3710765-6-yosryahmed@google.com
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Tested-by: Domenico Cerasuolo <cerasuolodomenico@gmail.com>
Acked-by: Shakeel Butt <shakeelb@google.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: Greg Thelen <gthelen@google.com>
Cc: Ivan Babrou <ivan@cloudflare.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Michal Koutny <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Tejun Heo <tj@kernel.org>
Cc: Waiman Long <longman@redhat.com>
Cc: Wei Xu <weixugc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-11-29 03:21:53 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
zswap: memcontrol: implement zswap writeback disabling
During our experiment with zswap, we sometimes observe swap IOs due to
occasional zswap store failures and writebacks-to-swap. These swapping
IOs prevent many users who cannot tolerate swapping from adopting zswap to
save memory and improve performance where possible.
This patch adds the option to disable this behavior entirely: do not
writeback to backing swapping device when a zswap store attempt fail, and
do not write pages in the zswap pool back to the backing swap device (both
when the pool is full, and when the new zswap shrinker is called).
This new behavior can be opted-in/out on a per-cgroup basis via a new
cgroup file. By default, writebacks to swap device is enabled, which is
the previous behavior. Initially, writeback is enabled for the root
cgroup, and a newly created cgroup will inherit the current setting of its
parent.
Note that this is subtly different from setting memory.swap.max to 0, as
it still allows for pages to be stored in the zswap pool (which itself
consumes swap space in its current form).
This patch should be applied on top of the zswap shrinker series:
https://lore.kernel.org/linux-mm/20231130194023.4102148-1-nphamcs@gmail.com/
as it also disables the zswap shrinker, a major source of zswap
writebacks.
For the most part, this feature is motivated by internal parties who
have already established their opinions regarding swapping - the
workloads that are highly sensitive to IO, and especially those who are
using servers with really slow disk performance (for instance, massive
but slow HDDs). For these folks, it's impossible to convince them to
even entertain zswap if swapping also comes as a packaged deal.
Writeback disabling is quite a useful feature in these situations - on
a mixed workloads deployment, they can disable writeback for the more
IO-sensitive workloads, and enable writeback for other background
workloads.
For instance, on a server with HDD, I allocate memories and populate
them with random values (so that zswap store will always fail), and
specify memory.high low enough to trigger reclaim. The time it takes
to allocate the memories and just read through it a couple of times
(doing silly things like computing the values' average etc.):
zswap.writeback disabled:
real 0m30.537s
user 0m23.687s
sys 0m6.637s
0 pages swapped in
0 pages swapped out
zswap.writeback enabled:
real 0m45.061s
user 0m24.310s
sys 0m8.892s
712686 pages swapped in
461093 pages swapped out
(the last two lines are from vmstat -s).
[nphamcs@gmail.com: add a comment about recurring zswap store failures leading to reclaim inefficiency]
Link: https://lkml.kernel.org/r/20231221005725.3446672-1-nphamcs@gmail.com
Link: https://lkml.kernel.org/r/20231207192406.3809579-1-nphamcs@gmail.com
Signed-off-by: Nhat Pham <nphamcs@gmail.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Yosry Ahmed <yosryahmed@google.com>
Acked-by: Chris Li <chrisl@kernel.org>
Cc: Dan Streetman <ddstreet@ieee.org>
Cc: David Heidelberg <david@ixit.cz>
Cc: Domenico Cerasuolo <cerasuolodomenico@gmail.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport (IBM) <rppt@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Seth Jennings <sjenning@redhat.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vitaly Wool <vitaly.wool@konsulko.com>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-12-07 11:24:06 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
zswap: memcontrol: implement zswap writeback disabling
During our experiment with zswap, we sometimes observe swap IOs due to
occasional zswap store failures and writebacks-to-swap. These swapping
IOs prevent many users who cannot tolerate swapping from adopting zswap to
save memory and improve performance where possible.
This patch adds the option to disable this behavior entirely: do not
writeback to backing swapping device when a zswap store attempt fail, and
do not write pages in the zswap pool back to the backing swap device (both
when the pool is full, and when the new zswap shrinker is called).
This new behavior can be opted-in/out on a per-cgroup basis via a new
cgroup file. By default, writebacks to swap device is enabled, which is
the previous behavior. Initially, writeback is enabled for the root
cgroup, and a newly created cgroup will inherit the current setting of its
parent.
Note that this is subtly different from setting memory.swap.max to 0, as
it still allows for pages to be stored in the zswap pool (which itself
consumes swap space in its current form).
This patch should be applied on top of the zswap shrinker series:
https://lore.kernel.org/linux-mm/20231130194023.4102148-1-nphamcs@gmail.com/
as it also disables the zswap shrinker, a major source of zswap
writebacks.
For the most part, this feature is motivated by internal parties who
have already established their opinions regarding swapping - the
workloads that are highly sensitive to IO, and especially those who are
using servers with really slow disk performance (for instance, massive
but slow HDDs). For these folks, it's impossible to convince them to
even entertain zswap if swapping also comes as a packaged deal.
Writeback disabling is quite a useful feature in these situations - on
a mixed workloads deployment, they can disable writeback for the more
IO-sensitive workloads, and enable writeback for other background
workloads.
For instance, on a server with HDD, I allocate memories and populate
them with random values (so that zswap store will always fail), and
specify memory.high low enough to trigger reclaim. The time it takes
to allocate the memories and just read through it a couple of times
(doing silly things like computing the values' average etc.):
zswap.writeback disabled:
real 0m30.537s
user 0m23.687s
sys 0m6.637s
0 pages swapped in
0 pages swapped out
zswap.writeback enabled:
real 0m45.061s
user 0m24.310s
sys 0m8.892s
712686 pages swapped in
461093 pages swapped out
(the last two lines are from vmstat -s).
[nphamcs@gmail.com: add a comment about recurring zswap store failures leading to reclaim inefficiency]
Link: https://lkml.kernel.org/r/20231221005725.3446672-1-nphamcs@gmail.com
Link: https://lkml.kernel.org/r/20231207192406.3809579-1-nphamcs@gmail.com
Signed-off-by: Nhat Pham <nphamcs@gmail.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Yosry Ahmed <yosryahmed@google.com>
Acked-by: Chris Li <chrisl@kernel.org>
Cc: Dan Streetman <ddstreet@ieee.org>
Cc: David Heidelberg <david@ixit.cz>
Cc: Domenico Cerasuolo <cerasuolodomenico@gmail.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport (IBM) <rppt@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Seth Jennings <sjenning@redhat.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vitaly Wool <vitaly.wool@konsulko.com>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-12-07 11:24:06 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
2024-07-01 11:31:15 -04:00
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
2015-02-11 15:26:36 -08:00
|
|
|
|
|
|
|
|
|
2020-06-03 16:02:14 -07:00
|
|
|
|
2020-06-03 16:02:11 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-06-24 17:59:05 -07:00
|
|
|
|
2020-06-03 16:02:11 -07:00
|
|
|
|
2024-06-24 17:59:05 -07:00
|
|
|
|
2024-07-01 11:31:15 -04:00
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
2015-02-11 15:26:36 -08:00
|
|
|
|
|
|
|
|
|
mm: memcontrol: deprecate swapaccounting=0 mode
The swapaccounting= commandline option already does very little today. To
close a trivial containment failure case, the swap ownership tracking part
of the swap controller has recently become mandatory (see commit
2d1c498072de ("mm: memcontrol: make swap tracking an integral part of
memory control") for details), which makes up the majority of the work
during swapout, swapin, and the swap slot map.
The only thing left under this flag is the page_counter operations and the
visibility of the swap control files in the first place, which are rather
meager savings. There also aren't many scenarios, if any, where
controlling the memory of a cgroup while allowing it unlimited access to a
global swap space is a workable resource isolation strategy.
On the other hand, there have been several bugs and confusion around the
many possible swap controller states (cgroup1 vs cgroup2 behavior, memory
accounting without swap accounting, memcg runtime disabled).
This puts the maintenance overhead of retaining the toggle above its
practical benefits. Deprecate it.
Link: https://lkml.kernel.org/r/20220926135704.400818-3-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-26 09:57:02 -04:00
|
|
|
|
2015-02-11 15:26:36 -08:00
|
|
|
|
2022-09-26 09:57:04 -04:00
|
|
|
|
2025-04-24 16:28:06 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
