2019-05-28 10:10:27 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-10-29 18:16:59 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-09-27 00:05:16 +05:30
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-10-29 18:16:59 -07:00
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2005-10-29 18:16:59 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-03 02:20:14 -07:00
|
|
|
|
2013-02-22 16:35:36 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-10-29 18:16:59 -07:00
|
|
|
|
mm/mempolicy: add MPOL_PREFERRED_MANY for multiple preferred nodes
Patch series "Introduce multi-preference mempolicy", v7.
This patch series introduces the concept of the MPOL_PREFERRED_MANY
mempolicy. This mempolicy mode can be used with either the
set_mempolicy(2) or mbind(2) interfaces. Like the MPOL_PREFERRED
interface, it allows an application to set a preference for nodes which
will fulfil memory allocation requests. Unlike the MPOL_PREFERRED mode,
it takes a set of nodes. Like the MPOL_BIND interface, it works over a
set of nodes. Unlike MPOL_BIND, it will not cause a SIGSEGV or invoke the
OOM killer if those preferred nodes are not available.
Along with these patches are patches for libnuma, numactl, numademo, and
memhog. They still need some polish, but can be found here:
https://gitlab.com/bwidawsk/numactl/-/tree/prefer-many It allows new
usage: `numactl -P 0,3,4`
The goal of the new mode is to enable some use-cases when using tiered memory
usage models which I've lovingly named.
1a. The Hare - The interconnect is fast enough to meet bandwidth and
latency requirements allowing preference to be given to all nodes with
"fast" memory.
1b. The Indiscriminate Hare - An application knows it wants fast
memory (or perhaps slow memory), but doesn't care which node it runs
on. The application can prefer a set of nodes and then xpu bind to
the local node (cpu, accelerator, etc). This reverses the nodes are
chosen today where the kernel attempts to use local memory to the CPU
whenever possible. This will attempt to use the local accelerator to
the memory.
2. The Tortoise - The administrator (or the application itself) is
aware it only needs slow memory, and so can prefer that.
Much of this is almost achievable with the bind interface, but the bind
interface suffers from an inability to fallback to another set of nodes if
binding fails to all nodes in the nodemask.
Like MPOL_BIND a nodemask is given. Inherently this removes ordering from the
preference.
> /* Set first two nodes as preferred in an 8 node system. */
> const unsigned long nodes = 0x3
> set_mempolicy(MPOL_PREFER_MANY, &nodes, 8);
> /* Mimic interleave policy, but have fallback *.
> const unsigned long nodes = 0xaa
> set_mempolicy(MPOL_PREFER_MANY, &nodes, 8);
Some internal discussion took place around the interface. There are two
alternatives which we have discussed, plus one I stuck in:
1. Ordered list of nodes. Currently it's believed that the added
complexity is nod needed for expected usecases.
2. A flag for bind to allow falling back to other nodes. This
confuses the notion of binding and is less flexible than the current
solution.
3. Create flags or new modes that helps with some ordering. This
offers both a friendlier API as well as a solution for more customized
usage. It's unknown if it's worth the complexity to support this.
Here is sample code for how this might work:
> // Prefer specific nodes for some something wacky
> set_mempolicy(MPOL_PREFER_MANY, 0x17c, 1024);
>
> // Default
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_SOCKET, NULL, 0);
> // which is the same as
> set_mempolicy(MPOL_DEFAULT, NULL, 0);
>
> // The Hare
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE, NULL, 0);
>
> // The Tortoise
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE_REV, NULL, 0);
>
> // Prefer the fast memory of the first two sockets
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE, -1, 2);
>
This patch (of 5):
The NUMA APIs currently allow passing in a "preferred node" as a single
bit set in a nodemask. If more than one bit it set, bits after the first
are ignored.
This single node is generally OK for location-based NUMA where memory
being allocated will eventually be operated on by a single CPU. However,
in systems with multiple memory types, folks want to target a *type* of
memory instead of a location. For instance, someone might want some
high-bandwidth memory but do not care about the CPU next to which it is
allocated. Or, they want a cheap, high capacity allocation and want to
target all NUMA nodes which have persistent memory in volatile mode. In
both of these cases, the application wants to target a *set* of nodes, but
does not want strict MPOL_BIND behavior as that could lead to OOM killer
or SIGSEGV.
So add MPOL_PREFERRED_MANY policy to support the multiple preferred nodes
requirement. This is not a pie-in-the-sky dream for an API. This was a
response to a specific ask of more than one group at Intel. Specifically:
1. There are existing libraries that target memory types such as
https://github.com/memkind/memkind. These are known to suffer from
SIGSEGV's when memory is low on targeted memory "kinds" that span more
than one node. The MCDRAM on a Xeon Phi in "Cluster on Die" mode is an
example of this.
2. Volatile-use persistent memory users want to have a memory policy
which is targeted at either "cheap and slow" (PMEM) or "expensive and
fast" (DRAM). However, they do not want to experience allocation
failures when the targeted type is unavailable.
3. Allocate-then-run. Generally, we let the process scheduler decide
on which physical CPU to run a task. That location provides a default
allocation policy, and memory availability is not generally considered
when placing tasks. For situations where memory is valuable and
constrained, some users want to allocate memory first, *then* allocate
close compute resources to the allocation. This is the reverse of the
normal (CPU) model. Accelerators such as GPUs that operate on
core-mm-managed memory are interested in this model.
A check is added in sanitize_mpol_flags() to not permit 'prefer_many'
policy to be used for now, and will be removed in later patch after all
implementations for 'prefer_many' are ready, as suggested by Michal Hocko.
[mhocko@kernel.org: suggest to refine policy_node/policy_nodemask handling]
Link: https://lkml.kernel.org/r/1627970362-61305-1-git-send-email-feng.tang@intel.com
Link: https://lore.kernel.org/r/20200630212517.308045-4-ben.widawsky@intel.com
Link: https://lkml.kernel.org/r/1627970362-61305-2-git-send-email-feng.tang@intel.com
Co-developed-by: Ben Widawsky <ben.widawsky@intel.com>
Signed-off-by: Ben Widawsky <ben.widawsky@intel.com>
Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com>
Signed-off-by: Feng Tang <feng.tang@intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Huang Ying <ying.huang@intel.com>b
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-09-02 15:00:06 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-03 02:20:14 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-06-06 14:38:30 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2019-08-28 16:19:53 +02:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-02-08 18:51:29 +01:00
|
|
|
|
2017-02-08 18:51:31 +01:00
|
|
|
|
2025-10-23 19:37:36 +08:00
|
|
|
|
2017-02-06 10:57:33 +01:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-10-16 02:01:52 -04:00
|
|
|
|
2007-10-18 23:40:14 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-11-15 17:38:14 -08:00
|
|
|
|
2006-01-08 01:00:50 -08:00
|
|
|
|
2006-01-08 01:01:02 -08:00
|
|
|
|
|
|
|
|
|
2025-10-23 19:37:36 +08:00
|
|
|
|
2006-03-22 00:09:12 -08:00
|
|
|
|
ksm: memory hotremove migration only
The previous patch enables page migration of ksm pages, but that soon gets
into trouble: not surprising, since we're using the ksm page lock to lock
operations on its stable_node, but page migration switches the page whose
lock is to be used for that. Another layer of locking would fix it, but
do we need that yet?
Do we actually need page migration of ksm pages? Yes, memory hotremove
needs to offline sections of memory: and since we stopped allocating ksm
pages with GFP_HIGHUSER, they will tend to be GFP_HIGHUSER_MOVABLE
candidates for migration.
But KSM is currently unconscious of NUMA issues, happily merging pages
from different NUMA nodes: at present the rule must be, not to use
MADV_MERGEABLE where you care about NUMA. So no, NUMA page migration of
ksm pages does not make sense yet.
So, to complete support for ksm swapping we need to make hotremove safe.
ksm_memory_callback() take ksm_thread_mutex when MEM_GOING_OFFLINE and
release it when MEM_OFFLINE or MEM_CANCEL_OFFLINE. But if mapped pages
are freed before migration reaches them, stable_nodes may be left still
pointing to struct pages which have been removed from the system: the
stable_node needs to identify a page by pfn rather than page pointer, then
it can safely prune them when MEM_OFFLINE.
And make NUMA migration skip PageKsm pages where it skips PageReserved.
But it's only when we reach unmap_and_move() that the page lock is taken
and we can be sure that raised pagecount has prevented a PageAnon from
being upgraded: so add offlining arg to migrate_pages(), to migrate ksm
page when offlining (has sufficient locking) but reject it otherwise.
Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Cc: Izik Eidus <ieidus@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Chris Wright <chrisw@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-14 17:59:33 -08:00
|
|
|
|
2006-06-23 02:03:53 -07:00
|
|
|
|
2006-06-23 02:04:02 -07:00
|
|
|
|
2007-10-16 01:26:26 -07:00
|
|
|
|
mempolicy: rework shmem mpol parsing and display
mm/shmem.c currently contains functions to parse and display memory policy
strings for the tmpfs 'mpol' mount option. Move this to mm/mempolicy.c with
the rest of the mempolicy support. With subsequent patches, we'll be able to
remove knowledge of the details [mode, flags, policy, ...] completely from
shmem.c
1) replace shmem_parse_mpol() in mm/shmem.c with mpol_parse_str() in
mm/mempolicy.c. Rework to use the policy_types[] array [used by
mpol_to_str()] to look up mode by name.
2) use mpol_to_str() to format policy for shmem_show_mpol(). mpol_to_str()
expects a pointer to a struct mempolicy, so temporarily construct one.
This will be replaced with a reference to a struct mempolicy in the tmpfs
superblock in a subsequent patch.
NOTE 1: I changed mpol_to_str() to use a colon ':' rather than an equal
sign '=' as the nodemask delimiter to match mpol_parse_str() and the
tmpfs/shmem mpol mount option formatting that now uses mpol_to_str(). This
is a user visible change to numa_maps, but then the addition of the mode
flags already changed the display. It makes sense to me to have the mounts
and numa_maps display the policy in the same format. However, if anyone
objects strongly, I can pass the desired nodemask delimeter as an arg to
mpol_to_str().
Note 2: Like show_numa_map(), I don't check the return code from
mpol_to_str(). I do use a longer buffer than the one provided by
show_numa_map(), which seems to have sufficed so far.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:23 -07:00
|
|
|
|
2009-12-14 17:58:11 -08:00
|
|
|
|
2012-10-25 14:16:32 +02:00
|
|
|
|
2014-06-06 14:38:30 -07:00
|
|
|
|
2025-11-10 22:21:28 +00:00
|
|
|
|
mm/mempolicy: Weighted Interleave Auto-tuning
On machines with multiple memory nodes, interleaving page allocations
across nodes allows for better utilization of each node's bandwidth.
Previous work by Gregory Price [1] introduced weighted interleave, which
allowed for pages to be allocated across nodes according to user-set
ratios.
Ideally, these weights should be proportional to their bandwidth, so that
under bandwidth pressure, each node uses its maximal efficient bandwidth
and prevents latency from increasing exponentially.
Previously, weighted interleave's default weights were just 1s -- which
would be equivalent to the (unweighted) interleave mempolicy, which goes
through the nodes in a round-robin fashion, ignoring bandwidth
information.
This patch has two main goals: First, it makes weighted interleave easier
to use for users who wish to relieve bandwidth pressure when using nodes
with varying bandwidth (CXL). By providing a set of "real" default
weights that just work out of the box, users who might not have the
capability (or wish to) perform experimentation to find the most optimal
weights for their system can still take advantage of bandwidth-informed
weighted interleave.
Second, it allows for weighted interleave to dynamically adjust to
hotplugged memory with new bandwidth information. Instead of manually
updating node weights every time new bandwidth information is reported or
taken off, weighted interleave adjusts and provides a new set of default
weights for weighted interleave to use when there is a change in bandwidth
information.
To meet these goals, this patch introduces an auto-configuration mode for
the interleave weights that provides a reasonable set of default weights,
calculated using bandwidth data reported by the system. In auto mode,
weights are dynamically adjusted based on whatever the current bandwidth
information reports (and responds to hotplug events).
This patch still supports users manually writing weights into the nodeN
sysfs interface by entering into manual mode. When a user enters manual
mode, the system stops dynamically updating any of the node weights, even
during hotplug events that shift the optimal weight distribution.
A new sysfs interface "auto" is introduced, which allows users to switch
between the auto (writing 1 or Y) and manual (writing 0 or N) modes. The
system also automatically enters manual mode when a nodeN interface is
manually written to.
There is one functional change that this patch makes to the existing
weighted_interleave ABI: previously, writing 0 directly to a nodeN
interface was said to reset the weight to the system default. Before this
patch, the default for all weights were 1, which meant that writing 0 and
1 were functionally equivalent. With this patch, writing 0 is invalid.
Link: https://lkml.kernel.org/r/20250520141236.2987309-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: wordsmithing changes, simplification, fixes]
Link: https://lkml.kernel.org/r/20250511025840.2410154-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: remove auto_kobj_attr field from struct sysfs_wi_group]
Link: https://lkml.kernel.org/r/20250512142511.3959833-1-joshua.hahnjy@gmail.com
https://lore.kernel.org/linux-mm/20240202170238.90004-1-gregory.price@memverge.com/ [1]
Link: https://lkml.kernel.org/r/20250505182328.4148265-1-joshua.hahnjy@gmail.com
Co-developed-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Yunjeong Mun <yunjeong.mun@sk.com>
Suggested-by: Oscar Salvador <osalvador@suse.de>
Suggested-by: Ying Huang <ying.huang@linux.alibaba.com>
Suggested-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Reviewed-by: Honggyu Kim <honggyu.kim@sk.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-05 11:23:28 -07:00
|
|
|
|
2006-01-08 01:00:50 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
mm/mprotect: use mmu_gather
Patch series "mm/mprotect: avoid unnecessary TLB flushes", v6.
This patchset is intended to remove unnecessary TLB flushes during
mprotect() syscalls. Once this patch-set make it through, similar and
further optimizations for MADV_COLD and userfaultfd would be possible.
Basically, there are 3 optimizations in this patch-set:
1. Use TLB batching infrastructure to batch flushes across VMAs and do
better/fewer flushes. This would also be handy for later userfaultfd
enhancements.
2. Avoid unnecessary TLB flushes. This optimization is the one that
provides most of the performance benefits. Unlike previous versions,
we now only avoid flushes that would not result in spurious
page-faults.
3. Avoiding TLB flushes on change_huge_pmd() that are only needed to
prevent the A/D bits from changing.
Andrew asked for some benchmark numbers. I do not have an easy
determinate macrobenchmark in which it is easy to show benefit. I
therefore ran a microbenchmark: a loop that does the following on
anonymous memory, just as a sanity check to see that time is saved by
avoiding TLB flushes. The loop goes:
mprotect(p, PAGE_SIZE, PROT_READ)
mprotect(p, PAGE_SIZE, PROT_READ|PROT_WRITE)
*p = 0; // make the page writable
The test was run in KVM guest with 1 or 2 threads (the second thread was
busy-looping). I measured the time (cycles) of each operation:
1 thread 2 threads
mmots +patch mmots +patch
PROT_READ 3494 2725 (-22%) 8630 7788 (-10%)
PROT_READ|WRITE 3952 2724 (-31%) 9075 2865 (-68%)
[ mmots = v5.17-rc6-mmots-2022-03-06-20-38 ]
The exact numbers are really meaningless, but the benefit is clear. There
are 2 interesting results though.
(1) PROT_READ is cheaper, while one can expect it not to be affected.
This is presumably due to TLB miss that is saved
(2) Without memory access (*p = 0), the speedup of the patch is even
greater. In that scenario mprotect(PROT_READ) also avoids the TLB flush.
As a result both operations on the patched kernel take roughly ~1500
cycles (with either 1 or 2 threads), whereas on mmotm their cost is as
high as presented in the table.
This patch (of 3):
change_pXX_range() currently does not use mmu_gather, but instead
implements its own deferred TLB flushes scheme. This both complicates the
code, as developers need to be aware of different invalidation schemes,
and prevents opportunities to avoid TLB flushes or perform them in finer
granularity.
The use of mmu_gather for modified PTEs has benefits in various scenarios
even if pages are not released. For instance, if only a single page needs
to be flushed out of a range of many pages, only that page would be
flushed. If a THP page is flushed, on x86 a single TLB invlpg instruction
can be used instead of 512 instructions (or a full TLB flush, which would
Linux would actually use by default). mprotect() over multiple VMAs
requires a single flush.
Use mmu_gather in change_pXX_range(). As the pages are not released, only
record the flushed range using tlb_flush_pXX_range().
Handle THP similarly and get rid of flush_cache_range() which becomes
redundant since tlb_start_vma() calls it when needed.
Link: https://lkml.kernel.org/r/20220401180821.1986781-1-namit@vmware.com
Link: https://lkml.kernel.org/r/20220401180821.1986781-2-namit@vmware.com
Signed-off-by: Nadav Amit <namit@vmware.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andrew Cooper <andrew.cooper3@citrix.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Will Deacon <will@kernel.org>
Cc: Yu Zhao <yuzhao@google.com>
Cc: Nick Piggin <npiggin@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-09 18:20:50 -07:00
|
|
|
|
2016-12-24 11:46:01 -08:00
|
|
|
|
mm/mempolicy: support memory hotplug in weighted interleave
The weighted interleave policy distributes page allocations across
multiple NUMA nodes based on their performance weight, thereby improving
memory bandwidth utilization. The weight values for each node are
configured through sysfs.
Previously, sysfs entries for configuring weighted interleave were created
for all possible nodes (N_POSSIBLE) at initialization, including nodes
that might not have memory. However, not all nodes in N_POSSIBLE are
usable at runtime, as some may remain memoryless or offline. This led to
sysfs entries being created for unusable nodes, causing potential
misconfiguration issues.
To address this issue, this patch modifies the sysfs creation logic to:
1) Limit sysfs entries to nodes that are online and have memory, avoiding
the creation of sysfs entries for nodes that cannot be used.
2) Support memory hotplug by dynamically adding and removing sysfs entries
based on whether a node transitions into or out of the N_MEMORY state.
Additionally, the patch ensures that sysfs attributes are properly managed
when nodes go offline, preventing stale or redundant entries from
persisting in the system.
By making these changes, the weighted interleave policy now manages its
sysfs entries more efficiently, ensuring that only relevant nodes are
considered for interleaving, and dynamically adapting to memory hotplug
events.
[dan.carpenter@linaro.org: fix error code in sysfs_wi_node_add()]
Link: https://lkml.kernel.org/r/aBjL7Bwc0QBzgajK@stanley.mountain
Link: https://lkml.kernel.org/r/20250417072839.711-4-rakie.kim@sk.com
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Dan Carpenter <dan.carpenter@linaro.org>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:37 +09:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
vmscan: move isolate_lru_page() to vmscan.c
On large memory systems, the VM can spend way too much time scanning
through pages that it cannot (or should not) evict from memory. Not only
does it use up CPU time, but it also provokes lock contention and can
leave large systems under memory presure in a catatonic state.
This patch series improves VM scalability by:
1) putting filesystem backed, swap backed and unevictable pages
onto their own LRUs, so the system only scans the pages that it
can/should evict from memory
2) switching to two handed clock replacement for the anonymous LRUs,
so the number of pages that need to be scanned when the system
starts swapping is bound to a reasonable number
3) keeping unevictable pages off the LRU completely, so the
VM does not waste CPU time scanning them. ramfs, ramdisk,
SHM_LOCKED shared memory segments and mlock()ed VMA pages
are keept on the unevictable list.
This patch:
isolate_lru_page logically belongs to be in vmscan.c than migrate.c.
It is tough, because we don't need that function without memory migration
so there is a valid argument to have it in migrate.c. However a
subsequent patch needs to make use of it in the core mm, so we can happily
move it to vmscan.c.
Also, make the function a little more generic by not requiring that it
adds an isolated page to a given list. Callers can do that.
Note that we now have '__isolate_lru_page()', that does
something quite different, visible outside of vmscan.c
for use with memory controller. Methinks we need to
rationalize these names/purposes. --lts
[akpm@linux-foundation.org: fix mm/memory_hotplug.c build]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-18 20:26:09 -07:00
|
|
|
|
|
|
|
|
|
2006-01-08 01:01:01 -08:00
|
|
|
|
2006-01-08 01:00:50 -08:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
|
|
|
|
|
2006-01-08 01:00:50 -08:00
|
|
|
|
2006-03-22 00:08:13 -08:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-02-10 01:43:07 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
mempolicy: use MPOL_PREFERRED for system-wide default policy
Currently, when one specifies MPOL_DEFAULT via a NUMA memory policy API
[set_mempolicy(), mbind() and internal versions], the kernel simply installs a
NULL struct mempolicy pointer in the appropriate context: task policy, vma
policy, or shared policy. This causes any use of that policy to "fall back"
to the next most specific policy scope.
The only use of MPOL_DEFAULT to mean "local allocation" is in the system
default policy. This requires extra checks/cases for MPOL_DEFAULT in many
mempolicy.c functions.
There is another, "preferred" way to specify local allocation via the APIs.
That is using the MPOL_PREFERRED policy mode with an empty nodemask.
Internally, the empty nodemask gets converted to a preferred_node id of '-1'.
All internal usage of MPOL_PREFERRED will convert the '-1' to the id of the
node local to the cpu where the allocation occurs.
System default policy, except during boot, is hard-coded to "local
allocation". By using the MPOL_PREFERRED mode with a negative value of
preferred node for system default policy, MPOL_DEFAULT will never occur in the
'policy' member of a struct mempolicy. Thus, we can remove all checks for
MPOL_DEFAULT when converting policy to a node id/zonelist in the allocation
paths.
In slab_node() return local node id when policy pointer is NULL. No need to
set a pol value to take the switch default. Replace switch default with
BUG()--i.e., shouldn't happen.
With this patch MPOL_DEFAULT is only used in the APIs, including internal
calls to do_set_mempolicy() and in the display of policy in
/proc/<pid>/numa_maps. It always means "fall back" to the the next most
specific policy scope. This simplifies the description of memory policies
quite a bit, with no visible change in behavior.
get_mempolicy() continues to return MPOL_DEFAULT and an empty nodemask when
the requested policy [task or vma/shared] is NULL. These are the values one
would supply via set_mempolicy() or mbind() to achieve that condition--default
behavior.
This patch updates Documentation to reflect this change.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:18 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-10-31 17:09:23 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
mm/mempolicy: don't handle MPOL_LOCAL like a fake MPOL_PREFERRED policy
MPOL_LOCAL policy has been setup as a real policy, but it is still handled
like a faked POL_PREFERRED policy with one internal MPOL_F_LOCAL flag bit
set, and there are many places having to judge the real 'prefer' or the
'local' policy, which are quite confusing.
In current code, there are 4 cases that MPOL_LOCAL are used:
1. user specifies 'local' policy
2. user specifies 'prefer' policy, but with empty nodemask
3. system 'default' policy is used
4. 'prefer' policy + valid 'preferred' node with MPOL_F_STATIC_NODES
flag set, and when it is 'rebind' to a nodemask which doesn't contains
the 'preferred' node, it will perform as 'local' policy
So make 'local' a real policy instead of a fake 'prefer' one, and kill
MPOL_F_LOCAL bit, which can greatly reduce the confusion for code reading.
For case 4, the logic of mpol_rebind_preferred() is confusing, as Michal
Hocko pointed out:
: I do believe that rebinding preferred policy is just bogus and it should
: be dropped altogether on the ground that a preference is a mere hint from
: userspace where to start the allocation. Unless I am missing something
: cpusets will be always authoritative for the final placement. The
: preferred node just acts as a starting point and it should be really
: preserved when cpusets changes. Otherwise we have a very subtle behavior
: corner cases.
So dump all the tricky transformation between 'prefer' and 'local', and
just record the new nodemask of rebinding.
[feng.tang@intel.com: fix a problem in mpol_set_nodemask(), per Michal Hocko]
Link: https://lkml.kernel.org/r/1622560492-1294-3-git-send-email-feng.tang@intel.com
[feng.tang@intel.com: refine code and comments of mpol_set_nodemask(), per Michal]
Link: https://lkml.kernel.org/r/20210603081807.GE56979@shbuild999.sh.intel.com
Link: https://lkml.kernel.org/r/1622469956-82897-3-git-send-email-feng.tang@intel.com
Signed-off-by: Feng Tang <feng.tang@intel.com>
Suggested-by: Michal Hocko <mhocko@suse.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ben Widawsky <ben.widawsky@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-30 18:51:00 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2012-11-02 18:19:13 +00:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
mm/mempolicy: Weighted Interleave Auto-tuning
On machines with multiple memory nodes, interleaving page allocations
across nodes allows for better utilization of each node's bandwidth.
Previous work by Gregory Price [1] introduced weighted interleave, which
allowed for pages to be allocated across nodes according to user-set
ratios.
Ideally, these weights should be proportional to their bandwidth, so that
under bandwidth pressure, each node uses its maximal efficient bandwidth
and prevents latency from increasing exponentially.
Previously, weighted interleave's default weights were just 1s -- which
would be equivalent to the (unweighted) interleave mempolicy, which goes
through the nodes in a round-robin fashion, ignoring bandwidth
information.
This patch has two main goals: First, it makes weighted interleave easier
to use for users who wish to relieve bandwidth pressure when using nodes
with varying bandwidth (CXL). By providing a set of "real" default
weights that just work out of the box, users who might not have the
capability (or wish to) perform experimentation to find the most optimal
weights for their system can still take advantage of bandwidth-informed
weighted interleave.
Second, it allows for weighted interleave to dynamically adjust to
hotplugged memory with new bandwidth information. Instead of manually
updating node weights every time new bandwidth information is reported or
taken off, weighted interleave adjusts and provides a new set of default
weights for weighted interleave to use when there is a change in bandwidth
information.
To meet these goals, this patch introduces an auto-configuration mode for
the interleave weights that provides a reasonable set of default weights,
calculated using bandwidth data reported by the system. In auto mode,
weights are dynamically adjusted based on whatever the current bandwidth
information reports (and responds to hotplug events).
This patch still supports users manually writing weights into the nodeN
sysfs interface by entering into manual mode. When a user enters manual
mode, the system stops dynamically updating any of the node weights, even
during hotplug events that shift the optimal weight distribution.
A new sysfs interface "auto" is introduced, which allows users to switch
between the auto (writing 1 or Y) and manual (writing 0 or N) modes. The
system also automatically enters manual mode when a nodeN interface is
manually written to.
There is one functional change that this patch makes to the existing
weighted_interleave ABI: previously, writing 0 directly to a nodeN
interface was said to reset the weight to the system default. Before this
patch, the default for all weights were 1, which meant that writing 0 and
1 were functionally equivalent. With this patch, writing 0 is invalid.
Link: https://lkml.kernel.org/r/20250520141236.2987309-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: wordsmithing changes, simplification, fixes]
Link: https://lkml.kernel.org/r/20250511025840.2410154-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: remove auto_kobj_attr field from struct sysfs_wi_group]
Link: https://lkml.kernel.org/r/20250512142511.3959833-1-joshua.hahnjy@gmail.com
https://lore.kernel.org/linux-mm/20240202170238.90004-1-gregory.price@memverge.com/ [1]
Link: https://lkml.kernel.org/r/20250505182328.4148265-1-joshua.hahnjy@gmail.com
Co-developed-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Yunjeong Mun <yunjeong.mun@sk.com>
Suggested-by: Oscar Salvador <osalvador@suse.de>
Suggested-by: Ying Huang <ying.huang@linux.alibaba.com>
Suggested-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Reviewed-by: Honggyu Kim <honggyu.kim@sk.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-05 11:23:28 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
mm/mempolicy: Weighted Interleave Auto-tuning
On machines with multiple memory nodes, interleaving page allocations
across nodes allows for better utilization of each node's bandwidth.
Previous work by Gregory Price [1] introduced weighted interleave, which
allowed for pages to be allocated across nodes according to user-set
ratios.
Ideally, these weights should be proportional to their bandwidth, so that
under bandwidth pressure, each node uses its maximal efficient bandwidth
and prevents latency from increasing exponentially.
Previously, weighted interleave's default weights were just 1s -- which
would be equivalent to the (unweighted) interleave mempolicy, which goes
through the nodes in a round-robin fashion, ignoring bandwidth
information.
This patch has two main goals: First, it makes weighted interleave easier
to use for users who wish to relieve bandwidth pressure when using nodes
with varying bandwidth (CXL). By providing a set of "real" default
weights that just work out of the box, users who might not have the
capability (or wish to) perform experimentation to find the most optimal
weights for their system can still take advantage of bandwidth-informed
weighted interleave.
Second, it allows for weighted interleave to dynamically adjust to
hotplugged memory with new bandwidth information. Instead of manually
updating node weights every time new bandwidth information is reported or
taken off, weighted interleave adjusts and provides a new set of default
weights for weighted interleave to use when there is a change in bandwidth
information.
To meet these goals, this patch introduces an auto-configuration mode for
the interleave weights that provides a reasonable set of default weights,
calculated using bandwidth data reported by the system. In auto mode,
weights are dynamically adjusted based on whatever the current bandwidth
information reports (and responds to hotplug events).
This patch still supports users manually writing weights into the nodeN
sysfs interface by entering into manual mode. When a user enters manual
mode, the system stops dynamically updating any of the node weights, even
during hotplug events that shift the optimal weight distribution.
A new sysfs interface "auto" is introduced, which allows users to switch
between the auto (writing 1 or Y) and manual (writing 0 or N) modes. The
system also automatically enters manual mode when a nodeN interface is
manually written to.
There is one functional change that this patch makes to the existing
weighted_interleave ABI: previously, writing 0 directly to a nodeN
interface was said to reset the weight to the system default. Before this
patch, the default for all weights were 1, which meant that writing 0 and
1 were functionally equivalent. With this patch, writing 0 is invalid.
Link: https://lkml.kernel.org/r/20250520141236.2987309-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: wordsmithing changes, simplification, fixes]
Link: https://lkml.kernel.org/r/20250511025840.2410154-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: remove auto_kobj_attr field from struct sysfs_wi_group]
Link: https://lkml.kernel.org/r/20250512142511.3959833-1-joshua.hahnjy@gmail.com
https://lore.kernel.org/linux-mm/20240202170238.90004-1-gregory.price@memverge.com/ [1]
Link: https://lkml.kernel.org/r/20250505182328.4148265-1-joshua.hahnjy@gmail.com
Co-developed-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Yunjeong Mun <yunjeong.mun@sk.com>
Suggested-by: Oscar Salvador <osalvador@suse.de>
Suggested-by: Ying Huang <ying.huang@linux.alibaba.com>
Suggested-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Reviewed-by: Honggyu Kim <honggyu.kim@sk.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-05 11:23:28 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: Weighted Interleave Auto-tuning
On machines with multiple memory nodes, interleaving page allocations
across nodes allows for better utilization of each node's bandwidth.
Previous work by Gregory Price [1] introduced weighted interleave, which
allowed for pages to be allocated across nodes according to user-set
ratios.
Ideally, these weights should be proportional to their bandwidth, so that
under bandwidth pressure, each node uses its maximal efficient bandwidth
and prevents latency from increasing exponentially.
Previously, weighted interleave's default weights were just 1s -- which
would be equivalent to the (unweighted) interleave mempolicy, which goes
through the nodes in a round-robin fashion, ignoring bandwidth
information.
This patch has two main goals: First, it makes weighted interleave easier
to use for users who wish to relieve bandwidth pressure when using nodes
with varying bandwidth (CXL). By providing a set of "real" default
weights that just work out of the box, users who might not have the
capability (or wish to) perform experimentation to find the most optimal
weights for their system can still take advantage of bandwidth-informed
weighted interleave.
Second, it allows for weighted interleave to dynamically adjust to
hotplugged memory with new bandwidth information. Instead of manually
updating node weights every time new bandwidth information is reported or
taken off, weighted interleave adjusts and provides a new set of default
weights for weighted interleave to use when there is a change in bandwidth
information.
To meet these goals, this patch introduces an auto-configuration mode for
the interleave weights that provides a reasonable set of default weights,
calculated using bandwidth data reported by the system. In auto mode,
weights are dynamically adjusted based on whatever the current bandwidth
information reports (and responds to hotplug events).
This patch still supports users manually writing weights into the nodeN
sysfs interface by entering into manual mode. When a user enters manual
mode, the system stops dynamically updating any of the node weights, even
during hotplug events that shift the optimal weight distribution.
A new sysfs interface "auto" is introduced, which allows users to switch
between the auto (writing 1 or Y) and manual (writing 0 or N) modes. The
system also automatically enters manual mode when a nodeN interface is
manually written to.
There is one functional change that this patch makes to the existing
weighted_interleave ABI: previously, writing 0 directly to a nodeN
interface was said to reset the weight to the system default. Before this
patch, the default for all weights were 1, which meant that writing 0 and
1 were functionally equivalent. With this patch, writing 0 is invalid.
Link: https://lkml.kernel.org/r/20250520141236.2987309-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: wordsmithing changes, simplification, fixes]
Link: https://lkml.kernel.org/r/20250511025840.2410154-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: remove auto_kobj_attr field from struct sysfs_wi_group]
Link: https://lkml.kernel.org/r/20250512142511.3959833-1-joshua.hahnjy@gmail.com
https://lore.kernel.org/linux-mm/20240202170238.90004-1-gregory.price@memverge.com/ [1]
Link: https://lkml.kernel.org/r/20250505182328.4148265-1-joshua.hahnjy@gmail.com
Co-developed-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Yunjeong Mun <yunjeong.mun@sk.com>
Suggested-by: Oscar Salvador <osalvador@suse.de>
Suggested-by: Ying Huang <ying.huang@linux.alibaba.com>
Suggested-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Reviewed-by: Honggyu Kim <honggyu.kim@sk.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-05 11:23:28 -07:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: Weighted Interleave Auto-tuning
On machines with multiple memory nodes, interleaving page allocations
across nodes allows for better utilization of each node's bandwidth.
Previous work by Gregory Price [1] introduced weighted interleave, which
allowed for pages to be allocated across nodes according to user-set
ratios.
Ideally, these weights should be proportional to their bandwidth, so that
under bandwidth pressure, each node uses its maximal efficient bandwidth
and prevents latency from increasing exponentially.
Previously, weighted interleave's default weights were just 1s -- which
would be equivalent to the (unweighted) interleave mempolicy, which goes
through the nodes in a round-robin fashion, ignoring bandwidth
information.
This patch has two main goals: First, it makes weighted interleave easier
to use for users who wish to relieve bandwidth pressure when using nodes
with varying bandwidth (CXL). By providing a set of "real" default
weights that just work out of the box, users who might not have the
capability (or wish to) perform experimentation to find the most optimal
weights for their system can still take advantage of bandwidth-informed
weighted interleave.
Second, it allows for weighted interleave to dynamically adjust to
hotplugged memory with new bandwidth information. Instead of manually
updating node weights every time new bandwidth information is reported or
taken off, weighted interleave adjusts and provides a new set of default
weights for weighted interleave to use when there is a change in bandwidth
information.
To meet these goals, this patch introduces an auto-configuration mode for
the interleave weights that provides a reasonable set of default weights,
calculated using bandwidth data reported by the system. In auto mode,
weights are dynamically adjusted based on whatever the current bandwidth
information reports (and responds to hotplug events).
This patch still supports users manually writing weights into the nodeN
sysfs interface by entering into manual mode. When a user enters manual
mode, the system stops dynamically updating any of the node weights, even
during hotplug events that shift the optimal weight distribution.
A new sysfs interface "auto" is introduced, which allows users to switch
between the auto (writing 1 or Y) and manual (writing 0 or N) modes. The
system also automatically enters manual mode when a nodeN interface is
manually written to.
There is one functional change that this patch makes to the existing
weighted_interleave ABI: previously, writing 0 directly to a nodeN
interface was said to reset the weight to the system default. Before this
patch, the default for all weights were 1, which meant that writing 0 and
1 were functionally equivalent. With this patch, writing 0 is invalid.
Link: https://lkml.kernel.org/r/20250520141236.2987309-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: wordsmithing changes, simplification, fixes]
Link: https://lkml.kernel.org/r/20250511025840.2410154-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: remove auto_kobj_attr field from struct sysfs_wi_group]
Link: https://lkml.kernel.org/r/20250512142511.3959833-1-joshua.hahnjy@gmail.com
https://lore.kernel.org/linux-mm/20240202170238.90004-1-gregory.price@memverge.com/ [1]
Link: https://lkml.kernel.org/r/20250505182328.4148265-1-joshua.hahnjy@gmail.com
Co-developed-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Yunjeong Mun <yunjeong.mun@sk.com>
Suggested-by: Oscar Salvador <osalvador@suse.de>
Suggested-by: Ying Huang <ying.huang@linux.alibaba.com>
Suggested-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Reviewed-by: Honggyu Kim <honggyu.kim@sk.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-05 11:23:28 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: Weighted Interleave Auto-tuning
On machines with multiple memory nodes, interleaving page allocations
across nodes allows for better utilization of each node's bandwidth.
Previous work by Gregory Price [1] introduced weighted interleave, which
allowed for pages to be allocated across nodes according to user-set
ratios.
Ideally, these weights should be proportional to their bandwidth, so that
under bandwidth pressure, each node uses its maximal efficient bandwidth
and prevents latency from increasing exponentially.
Previously, weighted interleave's default weights were just 1s -- which
would be equivalent to the (unweighted) interleave mempolicy, which goes
through the nodes in a round-robin fashion, ignoring bandwidth
information.
This patch has two main goals: First, it makes weighted interleave easier
to use for users who wish to relieve bandwidth pressure when using nodes
with varying bandwidth (CXL). By providing a set of "real" default
weights that just work out of the box, users who might not have the
capability (or wish to) perform experimentation to find the most optimal
weights for their system can still take advantage of bandwidth-informed
weighted interleave.
Second, it allows for weighted interleave to dynamically adjust to
hotplugged memory with new bandwidth information. Instead of manually
updating node weights every time new bandwidth information is reported or
taken off, weighted interleave adjusts and provides a new set of default
weights for weighted interleave to use when there is a change in bandwidth
information.
To meet these goals, this patch introduces an auto-configuration mode for
the interleave weights that provides a reasonable set of default weights,
calculated using bandwidth data reported by the system. In auto mode,
weights are dynamically adjusted based on whatever the current bandwidth
information reports (and responds to hotplug events).
This patch still supports users manually writing weights into the nodeN
sysfs interface by entering into manual mode. When a user enters manual
mode, the system stops dynamically updating any of the node weights, even
during hotplug events that shift the optimal weight distribution.
A new sysfs interface "auto" is introduced, which allows users to switch
between the auto (writing 1 or Y) and manual (writing 0 or N) modes. The
system also automatically enters manual mode when a nodeN interface is
manually written to.
There is one functional change that this patch makes to the existing
weighted_interleave ABI: previously, writing 0 directly to a nodeN
interface was said to reset the weight to the system default. Before this
patch, the default for all weights were 1, which meant that writing 0 and
1 were functionally equivalent. With this patch, writing 0 is invalid.
Link: https://lkml.kernel.org/r/20250520141236.2987309-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: wordsmithing changes, simplification, fixes]
Link: https://lkml.kernel.org/r/20250511025840.2410154-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: remove auto_kobj_attr field from struct sysfs_wi_group]
Link: https://lkml.kernel.org/r/20250512142511.3959833-1-joshua.hahnjy@gmail.com
https://lore.kernel.org/linux-mm/20240202170238.90004-1-gregory.price@memverge.com/ [1]
Link: https://lkml.kernel.org/r/20250505182328.4148265-1-joshua.hahnjy@gmail.com
Co-developed-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Yunjeong Mun <yunjeong.mun@sk.com>
Suggested-by: Oscar Salvador <osalvador@suse.de>
Suggested-by: Ying Huang <ying.huang@linux.alibaba.com>
Suggested-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Reviewed-by: Honggyu Kim <honggyu.kim@sk.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-05 11:23:28 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-02-21 20:03:00 -08:00
|
|
|
|
mm/mempolicy: Weighted Interleave Auto-tuning
On machines with multiple memory nodes, interleaving page allocations
across nodes allows for better utilization of each node's bandwidth.
Previous work by Gregory Price [1] introduced weighted interleave, which
allowed for pages to be allocated across nodes according to user-set
ratios.
Ideally, these weights should be proportional to their bandwidth, so that
under bandwidth pressure, each node uses its maximal efficient bandwidth
and prevents latency from increasing exponentially.
Previously, weighted interleave's default weights were just 1s -- which
would be equivalent to the (unweighted) interleave mempolicy, which goes
through the nodes in a round-robin fashion, ignoring bandwidth
information.
This patch has two main goals: First, it makes weighted interleave easier
to use for users who wish to relieve bandwidth pressure when using nodes
with varying bandwidth (CXL). By providing a set of "real" default
weights that just work out of the box, users who might not have the
capability (or wish to) perform experimentation to find the most optimal
weights for their system can still take advantage of bandwidth-informed
weighted interleave.
Second, it allows for weighted interleave to dynamically adjust to
hotplugged memory with new bandwidth information. Instead of manually
updating node weights every time new bandwidth information is reported or
taken off, weighted interleave adjusts and provides a new set of default
weights for weighted interleave to use when there is a change in bandwidth
information.
To meet these goals, this patch introduces an auto-configuration mode for
the interleave weights that provides a reasonable set of default weights,
calculated using bandwidth data reported by the system. In auto mode,
weights are dynamically adjusted based on whatever the current bandwidth
information reports (and responds to hotplug events).
This patch still supports users manually writing weights into the nodeN
sysfs interface by entering into manual mode. When a user enters manual
mode, the system stops dynamically updating any of the node weights, even
during hotplug events that shift the optimal weight distribution.
A new sysfs interface "auto" is introduced, which allows users to switch
between the auto (writing 1 or Y) and manual (writing 0 or N) modes. The
system also automatically enters manual mode when a nodeN interface is
manually written to.
There is one functional change that this patch makes to the existing
weighted_interleave ABI: previously, writing 0 directly to a nodeN
interface was said to reset the weight to the system default. Before this
patch, the default for all weights were 1, which meant that writing 0 and
1 were functionally equivalent. With this patch, writing 0 is invalid.
Link: https://lkml.kernel.org/r/20250520141236.2987309-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: wordsmithing changes, simplification, fixes]
Link: https://lkml.kernel.org/r/20250511025840.2410154-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: remove auto_kobj_attr field from struct sysfs_wi_group]
Link: https://lkml.kernel.org/r/20250512142511.3959833-1-joshua.hahnjy@gmail.com
https://lore.kernel.org/linux-mm/20240202170238.90004-1-gregory.price@memverge.com/ [1]
Link: https://lkml.kernel.org/r/20250505182328.4148265-1-joshua.hahnjy@gmail.com
Co-developed-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Yunjeong Mun <yunjeong.mun@sk.com>
Suggested-by: Oscar Salvador <osalvador@suse.de>
Suggested-by: Ying Huang <ying.huang@linux.alibaba.com>
Suggested-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Reviewed-by: Honggyu Kim <honggyu.kim@sk.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-05 11:23:28 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-02-16 12:00:48 -08:00
|
|
|
|
2023-08-19 07:12:33 -07:00
|
|
|
|
2020-08-11 18:31:13 -07:00
|
|
|
|
2023-08-19 07:12:33 -07:00
|
|
|
|
2020-02-16 12:00:48 -08:00
|
|
|
|
2023-08-19 07:12:33 -07:00
|
|
|
|
2022-01-14 14:08:24 -08:00
|
|
|
|
2023-08-19 07:12:33 -07:00
|
|
|
|
2020-02-16 12:00:48 -08:00
|
|
|
|
2023-08-19 07:12:33 -07:00
|
|
|
|
2020-02-16 12:00:48 -08:00
|
|
|
|
2020-02-16 12:00:53 -08:00
|
|
|
|
2020-02-16 12:00:48 -08:00
|
|
|
|
2023-08-19 07:12:33 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-02-16 12:00:53 -08:00
|
|
|
|
2020-02-16 12:00:48 -08:00
|
|
|
|
|
|
|
|
|
2023-08-19 07:12:33 -07:00
|
|
|
|
2020-02-16 12:00:53 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-02-16 12:00:48 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-08-19 07:12:33 -07:00
|
|
|
|
2020-02-16 12:00:48 -08:00
|
|
|
|
2025-02-14 20:40:02 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-10-09 15:27:50 -07:00
|
|
|
|
2012-11-02 18:19:13 +00:00
|
|
|
|
|
|
|
|
|
2014-10-09 15:27:43 -07:00
|
|
|
|
2012-11-02 18:19:13 +00:00
|
|
|
|
2014-10-09 15:27:43 -07:00
|
|
|
|
|
|
|
|
|
2012-11-02 18:19:13 +00:00
|
|
|
|
2014-10-09 15:27:43 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-11-02 18:19:13 +00:00
|
|
|
|
|
|
|
|
|
2014-10-09 15:27:43 -07:00
|
|
|
|
2012-11-02 18:19:13 +00:00
|
|
|
|
2025-08-27 17:52:45 +00:00
|
|
|
|
2012-11-02 18:19:13 +00:00
|
|
|
|
2008-04-28 02:12:33 -07:00
|
|
|
|
|
|
|
|
|
2017-07-06 15:40:06 -07:00
|
|
|
|
2008-04-28 02:12:33 -07:00
|
|
|
|
|
|
|
|
|
mempolicy: add MPOL_F_STATIC_NODES flag
Add an optional mempolicy mode flag, MPOL_F_STATIC_NODES, that suppresses the
node remap when the policy is rebound.
Adds another member to struct mempolicy, nodemask_t user_nodemask, as part of
a union with cpuset_mems_allowed:
struct mempolicy {
...
union {
nodemask_t cpuset_mems_allowed;
nodemask_t user_nodemask;
} w;
}
that stores the the nodemask that the user passed when he or she created the
mempolicy via set_mempolicy() or mbind(). When using MPOL_F_STATIC_NODES,
which is passed with any mempolicy mode, the user's passed nodemask
intersected with the VMA or task's allowed nodes is always used when
determining the preferred node, setting the MPOL_BIND zonelist, or creating
the interleave nodemask. This happens whenever the policy is rebound,
including when a task's cpuset assignment changes or the cpuset's mems are
changed.
This creates an interesting side-effect in that it allows the mempolicy
"intent" to lie dormant and uneffected until it has access to the node(s) that
it desires. For example, if you currently ask for an interleaved policy over
a set of nodes that you do not have access to, the mempolicy is not created
and the task continues to use the previous policy. With this change, however,
it is possible to create the same mempolicy; it is only effected when access
to nodes in the nodemask is acquired.
It is also possible to mount tmpfs with the static nodemask behavior when
specifying a node or nodemask. To do this, simply add "=static" immediately
following the mempolicy mode at mount time:
mount -o remount mpol=interleave=static:1-3
Also removes mpol_check_policy() and folds its logic into mpol_new() since it
is now obsoleted. The unused vma_mpol_equal() is also removed.
Cc: Paul Jackson <pj@sgi.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:12:27 -07:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: fix mpol_rebind_nodemask() for MPOL_F_NUMA_BALANCING
commit bda420b98505 ("numa balancing: migrate on fault among multiple
bound nodes") adds new flag MPOL_F_NUMA_BALANCING to enable NUMA balancing
for MPOL_BIND memory policy.
When the cpuset of tasks changes, the mempolicy of the task is rebound by
mpol_rebind_nodemask(). When MPOL_F_STATIC_NODES and
MPOL_F_RELATIVE_NODES are both not set, the behaviour of rebinding should
be same whenever MPOL_F_NUMA_BALANCING is set or not. So, when an
application calls set_mempolicy() with MPOL_F_NUMA_BALANCING set but both
MPOL_F_STATIC_NODES and MPOL_F_RELATIVE_NODES cleared,
mempolicy.w.cpuset_mems_allowed should be set to
cpuset_current_mems_allowed nodemask. However, in current implementation,
mpol_store_user_nodemask() wrongly returns true, causing
mempolicy->w.user_nodemask to be incorrectly set to the user-specified
nodemask. Later, when the cpuset of the application changes,
mpol_rebind_nodemask() ends up rebinding based on the user-specified
nodemask rather than the cpuset_mems_allowed nodemask as intended.
I can reproduce with the following steps in qemu with 4 NUMA nodes:
1. echo '+cpuset' > /sys/fs/cgroup/cgroup.subtree_control
2. mkdir /sys/fs/cgroup/test
3. ./reproducer &
4. cat /proc/$pid/numa_maps, the task is bound to NUMA 1
5. echo $pid > /sys/fs/cgroup/test/cgroup.procs
6. cat /proc/$pid/numa_maps, the task is bound to NUMA 0 now.
The reproducer code:
int main()
{
struct bitmask *bmp;
int ret;
bmp = numa_parse_nodestring("1");
ret = set_mempolicy(MPOL_BIND | MPOL_F_NUMA_BALANCING,
bmp->maskp, bmp->size + 1);
if (ret < 0) {
perror("Failed to call set_mempolicy");
exit(-1);
}
while (1);
return 0;
}
If I call set_mempolicy() without MPOL_F_NUMA_BALANCING in the reproducer
code. After step 5, the task is still bound to NUMA 1.
To fix this, only set mempolicy->w.user_nodemask to the user-specified
nodemask if MPOL_F_STATIC_NODES or MPOL_F_RELATIVE_NODES is present.
Link: https://lkml.kernel.org/r/20260120011018.1256654-1-tujinjiang@huawei.com
Link: https://lkml.kernel.org/r/20251223110523.1161421-1-tujinjiang@huawei.com
Fixes: bda420b98505 ("numa balancing: migrate on fault among multiple bound nodes")
Signed-off-by: Jinjiang Tu <tujinjiang@huawei.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Acked-by: David Hildenbrand (Red Hat) <david@kernel.org>
Cc: Alistair Popple <apopple@nvidia.com>
Cc: Byungchul Park <byungchul@sk.com>
Cc: Joshua Hahn <joshua.hahnjy@gmail.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mathew Brost <matthew.brost@intel.com>
Cc: Mel Gorman <mgorman <mgorman@suse.de>
Cc: Rakie Kim <rakie.kim@sk.com>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-12-23 19:05:23 +08:00
|
|
|
|
mempolicy: add MPOL_F_RELATIVE_NODES flag
Adds another optional mode flag, MPOL_F_RELATIVE_NODES, that specifies
nodemasks passed via set_mempolicy() or mbind() should be considered relative
to the current task's mems_allowed.
When the mempolicy is created, the passed nodemask is folded and mapped onto
the current task's mems_allowed. For example, consider a task using
set_mempolicy() to pass MPOL_INTERLEAVE | MPOL_F_RELATIVE_NODES with a
nodemask of 1-3. If current's mems_allowed is 4-7, the effected nodemask is
5-7 (the second, third, and fourth node of mems_allowed).
If the same task is attached to a cpuset, the mempolicy nodemask is rebound
each time the mems are changed. Some possible rebinds and results are:
mems result
1-3 1-3
1-7 2-4
1,5-6 1,5-6
1,5-7 5-7
Likewise, the zonelist built for MPOL_BIND acts on the set of zones assigned
to the resultant nodemask from the relative remap.
In the MPOL_PREFERRED case, the preferred node is remapped from the currently
effected nodemask to the relative nodemask.
This mempolicy mode flag was conceived of by Paul Jackson <pj@sgi.com>.
Cc: Paul Jackson <pj@sgi.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:12:30 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: add MPOL_F_STATIC_NODES flag
Add an optional mempolicy mode flag, MPOL_F_STATIC_NODES, that suppresses the
node remap when the policy is rebound.
Adds another member to struct mempolicy, nodemask_t user_nodemask, as part of
a union with cpuset_mems_allowed:
struct mempolicy {
...
union {
nodemask_t cpuset_mems_allowed;
nodemask_t user_nodemask;
} w;
}
that stores the the nodemask that the user passed when he or she created the
mempolicy via set_mempolicy() or mbind(). When using MPOL_F_STATIC_NODES,
which is passed with any mempolicy mode, the user's passed nodemask
intersected with the VMA or task's allowed nodes is always used when
determining the preferred node, setting the MPOL_BIND zonelist, or creating
the interleave nodemask. This happens whenever the policy is rebound,
including when a task's cpuset assignment changes or the cpuset's mems are
changed.
This creates an interesting side-effect in that it allows the mempolicy
"intent" to lie dormant and uneffected until it has access to the node(s) that
it desires. For example, if you currently ask for an interleaved policy over
a set of nodes that you do not have access to, the mempolicy is not created
and the task continues to use the previous policy. With this change, however,
it is possible to create the same mempolicy; it is only effected when access
to nodes in the nodemask is acquired.
It is also possible to mount tmpfs with the static nodemask behavior when
specifying a node or nodemask. To do this, simply add "=static" immediately
following the mempolicy mode at mount time:
mount -o remount mpol=interleave=static:1-3
Also removes mpol_check_policy() and folds its logic into mpol_new() since it
is now obsoleted. The unused vma_mpol_equal() is also removed.
Cc: Paul Jackson <pj@sgi.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:12:27 -07:00
|
|
|
|
|
|
|
|
|
2021-09-02 15:00:19 -07:00
|
|
|
|
2008-04-28 02:12:33 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-30 18:51:10 -07:00
|
|
|
|
2008-04-28 02:12:33 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: don't handle MPOL_LOCAL like a fake MPOL_PREFERRED policy
MPOL_LOCAL policy has been setup as a real policy, but it is still handled
like a faked POL_PREFERRED policy with one internal MPOL_F_LOCAL flag bit
set, and there are many places having to judge the real 'prefer' or the
'local' policy, which are quite confusing.
In current code, there are 4 cases that MPOL_LOCAL are used:
1. user specifies 'local' policy
2. user specifies 'prefer' policy, but with empty nodemask
3. system 'default' policy is used
4. 'prefer' policy + valid 'preferred' node with MPOL_F_STATIC_NODES
flag set, and when it is 'rebind' to a nodemask which doesn't contains
the 'preferred' node, it will perform as 'local' policy
So make 'local' a real policy instead of a fake 'prefer' one, and kill
MPOL_F_LOCAL bit, which can greatly reduce the confusion for code reading.
For case 4, the logic of mpol_rebind_preferred() is confusing, as Michal
Hocko pointed out:
: I do believe that rebinding preferred policy is just bogus and it should
: be dropped altogether on the ground that a preference is a mere hint from
: userspace where to start the allocation. Unless I am missing something
: cpusets will be always authoritative for the final placement. The
: preferred node just acts as a starting point and it should be really
: preserved when cpusets changes. Otherwise we have a very subtle behavior
: corner cases.
So dump all the tricky transformation between 'prefer' and 'local', and
just record the new nodemask of rebinding.
[feng.tang@intel.com: fix a problem in mpol_set_nodemask(), per Michal Hocko]
Link: https://lkml.kernel.org/r/1622560492-1294-3-git-send-email-feng.tang@intel.com
[feng.tang@intel.com: refine code and comments of mpol_set_nodemask(), per Michal]
Link: https://lkml.kernel.org/r/20210603081807.GE56979@shbuild999.sh.intel.com
Link: https://lkml.kernel.org/r/1622469956-82897-3-git-send-email-feng.tang@intel.com
Signed-off-by: Feng Tang <feng.tang@intel.com>
Suggested-by: Michal Hocko <mhocko@suse.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ben Widawsky <ben.widawsky@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-30 18:51:00 -07:00
|
|
|
|
|
|
|
|
|
2021-06-30 18:51:10 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-04-28 02:12:33 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
cpuset,mm: update tasks' mems_allowed in time
Fix allocating page cache/slab object on the unallowed node when memory
spread is set by updating tasks' mems_allowed after its cpuset's mems is
changed.
In order to update tasks' mems_allowed in time, we must modify the code of
memory policy. Because the memory policy is applied in the process's
context originally. After applying this patch, one task directly
manipulates anothers mems_allowed, and we use alloc_lock in the
task_struct to protect mems_allowed and memory policy of the task.
But in the fast path, we didn't use lock to protect them, because adding a
lock may lead to performance regression. But if we don't add a lock,the
task might see no nodes when changing cpuset's mems_allowed to some
non-overlapping set. In order to avoid it, we set all new allowed nodes,
then clear newly disallowed ones.
[lee.schermerhorn@hp.com:
The rework of mpol_new() to extract the adjusting of the node mask to
apply cpuset and mpol flags "context" breaks set_mempolicy() and mbind()
with MPOL_PREFERRED and a NULL nodemask--i.e., explicit local
allocation. Fix this by adding the check for MPOL_PREFERRED and empty
node mask to mpol_new_mpolicy().
Remove the now unneeded 'nodes = NULL' from mpol_new().
Note that mpol_new_mempolicy() is always called with a non-NULL
'nodes' parameter now that it has been removed from mpol_new().
Therefore, we don't need to test nodes for NULL before testing it for
'empty'. However, just to be extra paranoid, add a VM_BUG_ON() to
verify this assumption.]
[lee.schermerhorn@hp.com:
I don't think the function name 'mpol_new_mempolicy' is descriptive
enough to differentiate it from mpol_new().
This function applies cpuset set context, usually constraining nodes
to those allowed by the cpuset. However, when the 'RELATIVE_NODES flag
is set, it also translates the nodes. So I settled on
'mpol_set_nodemask()', because the comment block for mpol_new() mentions
that we need to call this function to "set nodes".
Some additional minor line length, whitespace and typo cleanup.]
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 15:31:49 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: don't handle MPOL_LOCAL like a fake MPOL_PREFERRED policy
MPOL_LOCAL policy has been setup as a real policy, but it is still handled
like a faked POL_PREFERRED policy with one internal MPOL_F_LOCAL flag bit
set, and there are many places having to judge the real 'prefer' or the
'local' policy, which are quite confusing.
In current code, there are 4 cases that MPOL_LOCAL are used:
1. user specifies 'local' policy
2. user specifies 'prefer' policy, but with empty nodemask
3. system 'default' policy is used
4. 'prefer' policy + valid 'preferred' node with MPOL_F_STATIC_NODES
flag set, and when it is 'rebind' to a nodemask which doesn't contains
the 'preferred' node, it will perform as 'local' policy
So make 'local' a real policy instead of a fake 'prefer' one, and kill
MPOL_F_LOCAL bit, which can greatly reduce the confusion for code reading.
For case 4, the logic of mpol_rebind_preferred() is confusing, as Michal
Hocko pointed out:
: I do believe that rebinding preferred policy is just bogus and it should
: be dropped altogether on the ground that a preference is a mere hint from
: userspace where to start the allocation. Unless I am missing something
: cpusets will be always authoritative for the final placement. The
: preferred node just acts as a starting point and it should be really
: preserved when cpusets changes. Otherwise we have a very subtle behavior
: corner cases.
So dump all the tricky transformation between 'prefer' and 'local', and
just record the new nodemask of rebinding.
[feng.tang@intel.com: fix a problem in mpol_set_nodemask(), per Michal Hocko]
Link: https://lkml.kernel.org/r/1622560492-1294-3-git-send-email-feng.tang@intel.com
[feng.tang@intel.com: refine code and comments of mpol_set_nodemask(), per Michal]
Link: https://lkml.kernel.org/r/20210603081807.GE56979@shbuild999.sh.intel.com
Link: https://lkml.kernel.org/r/1622469956-82897-3-git-send-email-feng.tang@intel.com
Signed-off-by: Feng Tang <feng.tang@intel.com>
Suggested-by: Michal Hocko <mhocko@suse.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ben Widawsky <ben.widawsky@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-30 18:51:00 -07:00
|
|
|
|
cpuset,mm: update tasks' mems_allowed in time
Fix allocating page cache/slab object on the unallowed node when memory
spread is set by updating tasks' mems_allowed after its cpuset's mems is
changed.
In order to update tasks' mems_allowed in time, we must modify the code of
memory policy. Because the memory policy is applied in the process's
context originally. After applying this patch, one task directly
manipulates anothers mems_allowed, and we use alloc_lock in the
task_struct to protect mems_allowed and memory policy of the task.
But in the fast path, we didn't use lock to protect them, because adding a
lock may lead to performance regression. But if we don't add a lock,the
task might see no nodes when changing cpuset's mems_allowed to some
non-overlapping set. In order to avoid it, we set all new allowed nodes,
then clear newly disallowed ones.
[lee.schermerhorn@hp.com:
The rework of mpol_new() to extract the adjusting of the node mask to
apply cpuset and mpol flags "context" breaks set_mempolicy() and mbind()
with MPOL_PREFERRED and a NULL nodemask--i.e., explicit local
allocation. Fix this by adding the check for MPOL_PREFERRED and empty
node mask to mpol_new_mpolicy().
Remove the now unneeded 'nodes = NULL' from mpol_new().
Note that mpol_new_mempolicy() is always called with a non-NULL
'nodes' parameter now that it has been removed from mpol_new().
Therefore, we don't need to test nodes for NULL before testing it for
'empty'. However, just to be extra paranoid, add a VM_BUG_ON() to
verify this assumption.]
[lee.schermerhorn@hp.com:
I don't think the function name 'mpol_new_mempolicy' is descriptive
enough to differentiate it from mpol_new().
This function applies cpuset set context, usually constraining nodes
to those allowed by the cpuset. However, when the 'RELATIVE_NODES flag
is set, it also translates the nodes. So I settled on
'mpol_set_nodemask()', because the comment block for mpol_new() mentions
that we need to call this function to "set nodes".
Some additional minor line length, whitespace and typo cleanup.]
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 15:31:49 -07:00
|
|
|
|
|
|
|
|
|
2020-06-08 21:33:54 -07:00
|
|
|
|
cpuset,mm: update tasks' mems_allowed in time
Fix allocating page cache/slab object on the unallowed node when memory
spread is set by updating tasks' mems_allowed after its cpuset's mems is
changed.
In order to update tasks' mems_allowed in time, we must modify the code of
memory policy. Because the memory policy is applied in the process's
context originally. After applying this patch, one task directly
manipulates anothers mems_allowed, and we use alloc_lock in the
task_struct to protect mems_allowed and memory policy of the task.
But in the fast path, we didn't use lock to protect them, because adding a
lock may lead to performance regression. But if we don't add a lock,the
task might see no nodes when changing cpuset's mems_allowed to some
non-overlapping set. In order to avoid it, we set all new allowed nodes,
then clear newly disallowed ones.
[lee.schermerhorn@hp.com:
The rework of mpol_new() to extract the adjusting of the node mask to
apply cpuset and mpol flags "context" breaks set_mempolicy() and mbind()
with MPOL_PREFERRED and a NULL nodemask--i.e., explicit local
allocation. Fix this by adding the check for MPOL_PREFERRED and empty
node mask to mpol_new_mpolicy().
Remove the now unneeded 'nodes = NULL' from mpol_new().
Note that mpol_new_mempolicy() is always called with a non-NULL
'nodes' parameter now that it has been removed from mpol_new().
Therefore, we don't need to test nodes for NULL before testing it for
'empty'. However, just to be extra paranoid, add a VM_BUG_ON() to
verify this assumption.]
[lee.schermerhorn@hp.com:
I don't think the function name 'mpol_new_mempolicy' is descriptive
enough to differentiate it from mpol_new().
This function applies cpuset set context, usually constraining nodes
to those allowed by the cpuset. However, when the 'RELATIVE_NODES flag
is set, it also translates the nodes. So I settled on
'mpol_set_nodemask()', because the comment block for mpol_new() mentions
that we need to call this function to "set nodes".
Some additional minor line length, whitespace and typo cleanup.]
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 15:31:49 -07:00
|
|
|
|
mm: make set_mempolicy(MPOL_INTERLEAV) N_HIGH_MEMORY aware
At first, init_task's mems_allowed is initialized as this.
init_task->mems_allowed == node_state[N_POSSIBLE]
And cpuset's top_cpuset mask is initialized as this
top_cpuset->mems_allowed = node_state[N_HIGH_MEMORY]
Before 2.6.29:
policy's mems_allowed is initialized as this.
1. update tasks->mems_allowed by its cpuset->mems_allowed.
2. policy->mems_allowed = nodes_and(tasks->mems_allowed, user's mask)
Updating task's mems_allowed in reference to top_cpuset's one.
cpuset's mems_allowed is aware of N_HIGH_MEMORY, always.
In 2.6.30: After commit 58568d2a8215cb6f55caf2332017d7bdff954e1c
("cpuset,mm: update tasks' mems_allowed in time"), policy's mems_allowed
is initialized as this.
1. policy->mems_allowd = nodes_and(task->mems_allowed, user's mask)
Here, if task is in top_cpuset, task->mems_allowed is not updated from
init's one. Assume user excutes command as #numactrl --interleave=all
,....
policy->mems_allowd = nodes_and(N_POSSIBLE, ALL_SET_MASK)
Then, policy's mems_allowd can includes a possible node, which has no pgdat.
MPOL's INTERLEAVE just scans nodemask of task->mems_allowd and access this
directly.
NODE_DATA(nid)->zonelist even if NODE_DATA(nid)==NULL
Then, what's we need is making policy->mems_allowed be aware of
N_HIGH_MEMORY. This patch does that. But to do so, extra nodemask will
be on statck. Because I know cpumask has a new interface of
CPUMASK_ALLOC(), I added it to node.
This patch stands on old behavior. But I feel this fix itself is just a
Band-Aid. But to do fundametal fix, we have to take care of memory
hotplug and it takes time. (task->mems_allowd should be N_HIGH_MEMORY, I
think.)
mpol_set_nodemask() should be aware of N_HIGH_MEMORY and policy's nodemask
should be includes only online nodes.
In old behavior, this is guaranteed by frequent reference to cpuset's
code. Now, most of them are removed and mempolicy has to check it by
itself.
To do check, a few nodemask_t will be used for calculating nodemask. But,
size of nodemask_t can be big and it's not good to allocate them on stack.
Now, cpumask_t has CPUMASK_ALLOC/FREE an easy code for get scratch area.
NODEMASK_ALLOC/FREE shoudl be there.
[akpm@linux-foundation.org: cleanups & tweaks]
Tested-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-08-06 15:07:33 -07:00
|
|
|
|
|
|
|
|
|
cpuset,mm: update tasks' mems_allowed in time
Fix allocating page cache/slab object on the unallowed node when memory
spread is set by updating tasks' mems_allowed after its cpuset's mems is
changed.
In order to update tasks' mems_allowed in time, we must modify the code of
memory policy. Because the memory policy is applied in the process's
context originally. After applying this patch, one task directly
manipulates anothers mems_allowed, and we use alloc_lock in the
task_struct to protect mems_allowed and memory policy of the task.
But in the fast path, we didn't use lock to protect them, because adding a
lock may lead to performance regression. But if we don't add a lock,the
task might see no nodes when changing cpuset's mems_allowed to some
non-overlapping set. In order to avoid it, we set all new allowed nodes,
then clear newly disallowed ones.
[lee.schermerhorn@hp.com:
The rework of mpol_new() to extract the adjusting of the node mask to
apply cpuset and mpol flags "context" breaks set_mempolicy() and mbind()
with MPOL_PREFERRED and a NULL nodemask--i.e., explicit local
allocation. Fix this by adding the check for MPOL_PREFERRED and empty
node mask to mpol_new_mpolicy().
Remove the now unneeded 'nodes = NULL' from mpol_new().
Note that mpol_new_mempolicy() is always called with a non-NULL
'nodes' parameter now that it has been removed from mpol_new().
Therefore, we don't need to test nodes for NULL before testing it for
'empty'. However, just to be extra paranoid, add a VM_BUG_ON() to
verify this assumption.]
[lee.schermerhorn@hp.com:
I don't think the function name 'mpol_new_mempolicy' is descriptive
enough to differentiate it from mpol_new().
This function applies cpuset set context, usually constraining nodes
to those allowed by the cpuset. However, when the 'RELATIVE_NODES flag
is set, it also translates the nodes. So I settled on
'mpol_set_nodemask()', because the comment block for mpol_new() mentions
that we need to call this function to "set nodes".
Some additional minor line length, whitespace and typo cleanup.]
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 15:31:49 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: don't handle MPOL_LOCAL like a fake MPOL_PREFERRED policy
MPOL_LOCAL policy has been setup as a real policy, but it is still handled
like a faked POL_PREFERRED policy with one internal MPOL_F_LOCAL flag bit
set, and there are many places having to judge the real 'prefer' or the
'local' policy, which are quite confusing.
In current code, there are 4 cases that MPOL_LOCAL are used:
1. user specifies 'local' policy
2. user specifies 'prefer' policy, but with empty nodemask
3. system 'default' policy is used
4. 'prefer' policy + valid 'preferred' node with MPOL_F_STATIC_NODES
flag set, and when it is 'rebind' to a nodemask which doesn't contains
the 'preferred' node, it will perform as 'local' policy
So make 'local' a real policy instead of a fake 'prefer' one, and kill
MPOL_F_LOCAL bit, which can greatly reduce the confusion for code reading.
For case 4, the logic of mpol_rebind_preferred() is confusing, as Michal
Hocko pointed out:
: I do believe that rebinding preferred policy is just bogus and it should
: be dropped altogether on the ground that a preference is a mere hint from
: userspace where to start the allocation. Unless I am missing something
: cpusets will be always authoritative for the final placement. The
: preferred node just acts as a starting point and it should be really
: preserved when cpusets changes. Otherwise we have a very subtle behavior
: corner cases.
So dump all the tricky transformation between 'prefer' and 'local', and
just record the new nodemask of rebinding.
[feng.tang@intel.com: fix a problem in mpol_set_nodemask(), per Michal Hocko]
Link: https://lkml.kernel.org/r/1622560492-1294-3-git-send-email-feng.tang@intel.com
[feng.tang@intel.com: refine code and comments of mpol_set_nodemask(), per Michal]
Link: https://lkml.kernel.org/r/20210603081807.GE56979@shbuild999.sh.intel.com
Link: https://lkml.kernel.org/r/1622469956-82897-3-git-send-email-feng.tang@intel.com
Signed-off-by: Feng Tang <feng.tang@intel.com>
Suggested-by: Michal Hocko <mhocko@suse.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ben Widawsky <ben.widawsky@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-30 18:51:00 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
cpuset,mm: update tasks' mems_allowed in time
Fix allocating page cache/slab object on the unallowed node when memory
spread is set by updating tasks' mems_allowed after its cpuset's mems is
changed.
In order to update tasks' mems_allowed in time, we must modify the code of
memory policy. Because the memory policy is applied in the process's
context originally. After applying this patch, one task directly
manipulates anothers mems_allowed, and we use alloc_lock in the
task_struct to protect mems_allowed and memory policy of the task.
But in the fast path, we didn't use lock to protect them, because adding a
lock may lead to performance regression. But if we don't add a lock,the
task might see no nodes when changing cpuset's mems_allowed to some
non-overlapping set. In order to avoid it, we set all new allowed nodes,
then clear newly disallowed ones.
[lee.schermerhorn@hp.com:
The rework of mpol_new() to extract the adjusting of the node mask to
apply cpuset and mpol flags "context" breaks set_mempolicy() and mbind()
with MPOL_PREFERRED and a NULL nodemask--i.e., explicit local
allocation. Fix this by adding the check for MPOL_PREFERRED and empty
node mask to mpol_new_mpolicy().
Remove the now unneeded 'nodes = NULL' from mpol_new().
Note that mpol_new_mempolicy() is always called with a non-NULL
'nodes' parameter now that it has been removed from mpol_new().
Therefore, we don't need to test nodes for NULL before testing it for
'empty'. However, just to be extra paranoid, add a VM_BUG_ON() to
verify this assumption.]
[lee.schermerhorn@hp.com:
I don't think the function name 'mpol_new_mempolicy' is descriptive
enough to differentiate it from mpol_new().
This function applies cpuset set context, usually constraining nodes
to those allowed by the cpuset. However, when the 'RELATIVE_NODES flag
is set, it also translates the nodes. So I settled on
'mpol_set_nodemask()', because the comment block for mpol_new() mentions
that we need to call this function to "set nodes".
Some additional minor line length, whitespace and typo cleanup.]
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 15:31:49 -07:00
|
|
|
|
mm/mempolicy: don't handle MPOL_LOCAL like a fake MPOL_PREFERRED policy
MPOL_LOCAL policy has been setup as a real policy, but it is still handled
like a faked POL_PREFERRED policy with one internal MPOL_F_LOCAL flag bit
set, and there are many places having to judge the real 'prefer' or the
'local' policy, which are quite confusing.
In current code, there are 4 cases that MPOL_LOCAL are used:
1. user specifies 'local' policy
2. user specifies 'prefer' policy, but with empty nodemask
3. system 'default' policy is used
4. 'prefer' policy + valid 'preferred' node with MPOL_F_STATIC_NODES
flag set, and when it is 'rebind' to a nodemask which doesn't contains
the 'preferred' node, it will perform as 'local' policy
So make 'local' a real policy instead of a fake 'prefer' one, and kill
MPOL_F_LOCAL bit, which can greatly reduce the confusion for code reading.
For case 4, the logic of mpol_rebind_preferred() is confusing, as Michal
Hocko pointed out:
: I do believe that rebinding preferred policy is just bogus and it should
: be dropped altogether on the ground that a preference is a mere hint from
: userspace where to start the allocation. Unless I am missing something
: cpusets will be always authoritative for the final placement. The
: preferred node just acts as a starting point and it should be really
: preserved when cpusets changes. Otherwise we have a very subtle behavior
: corner cases.
So dump all the tricky transformation between 'prefer' and 'local', and
just record the new nodemask of rebinding.
[feng.tang@intel.com: fix a problem in mpol_set_nodemask(), per Michal Hocko]
Link: https://lkml.kernel.org/r/1622560492-1294-3-git-send-email-feng.tang@intel.com
[feng.tang@intel.com: refine code and comments of mpol_set_nodemask(), per Michal]
Link: https://lkml.kernel.org/r/20210603081807.GE56979@shbuild999.sh.intel.com
Link: https://lkml.kernel.org/r/1622469956-82897-3-git-send-email-feng.tang@intel.com
Signed-off-by: Feng Tang <feng.tang@intel.com>
Suggested-by: Michal Hocko <mhocko@suse.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ben Widawsky <ben.widawsky@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-30 18:51:00 -07:00
|
|
|
|
2012-12-12 13:51:33 -08:00
|
|
|
|
mm: make set_mempolicy(MPOL_INTERLEAV) N_HIGH_MEMORY aware
At first, init_task's mems_allowed is initialized as this.
init_task->mems_allowed == node_state[N_POSSIBLE]
And cpuset's top_cpuset mask is initialized as this
top_cpuset->mems_allowed = node_state[N_HIGH_MEMORY]
Before 2.6.29:
policy's mems_allowed is initialized as this.
1. update tasks->mems_allowed by its cpuset->mems_allowed.
2. policy->mems_allowed = nodes_and(tasks->mems_allowed, user's mask)
Updating task's mems_allowed in reference to top_cpuset's one.
cpuset's mems_allowed is aware of N_HIGH_MEMORY, always.
In 2.6.30: After commit 58568d2a8215cb6f55caf2332017d7bdff954e1c
("cpuset,mm: update tasks' mems_allowed in time"), policy's mems_allowed
is initialized as this.
1. policy->mems_allowd = nodes_and(task->mems_allowed, user's mask)
Here, if task is in top_cpuset, task->mems_allowed is not updated from
init's one. Assume user excutes command as #numactrl --interleave=all
,....
policy->mems_allowd = nodes_and(N_POSSIBLE, ALL_SET_MASK)
Then, policy's mems_allowd can includes a possible node, which has no pgdat.
MPOL's INTERLEAVE just scans nodemask of task->mems_allowd and access this
directly.
NODE_DATA(nid)->zonelist even if NODE_DATA(nid)==NULL
Then, what's we need is making policy->mems_allowed be aware of
N_HIGH_MEMORY. This patch does that. But to do so, extra nodemask will
be on statck. Because I know cpumask has a new interface of
CPUMASK_ALLOC(), I added it to node.
This patch stands on old behavior. But I feel this fix itself is just a
Band-Aid. But to do fundametal fix, we have to take care of memory
hotplug and it takes time. (task->mems_allowd should be N_HIGH_MEMORY, I
think.)
mpol_set_nodemask() should be aware of N_HIGH_MEMORY and policy's nodemask
should be includes only online nodes.
In old behavior, this is guaranteed by frequent reference to cpuset's
code. Now, most of them are removed and mempolicy has to check it by
itself.
To do check, a few nodemask_t will be used for calculating nodemask. But,
size of nodemask_t can be big and it's not good to allocate them on stack.
Now, cpumask_t has CPUMASK_ALLOC/FREE an easy code for get scratch area.
NODEMASK_ALLOC/FREE shoudl be there.
[akpm@linux-foundation.org: cleanups & tweaks]
Tested-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-08-06 15:07:33 -07:00
|
|
|
|
2012-12-12 13:51:33 -08:00
|
|
|
|
cpuset,mm: update tasks' mems_allowed in time
Fix allocating page cache/slab object on the unallowed node when memory
spread is set by updating tasks' mems_allowed after its cpuset's mems is
changed.
In order to update tasks' mems_allowed in time, we must modify the code of
memory policy. Because the memory policy is applied in the process's
context originally. After applying this patch, one task directly
manipulates anothers mems_allowed, and we use alloc_lock in the
task_struct to protect mems_allowed and memory policy of the task.
But in the fast path, we didn't use lock to protect them, because adding a
lock may lead to performance regression. But if we don't add a lock,the
task might see no nodes when changing cpuset's mems_allowed to some
non-overlapping set. In order to avoid it, we set all new allowed nodes,
then clear newly disallowed ones.
[lee.schermerhorn@hp.com:
The rework of mpol_new() to extract the adjusting of the node mask to
apply cpuset and mpol flags "context" breaks set_mempolicy() and mbind()
with MPOL_PREFERRED and a NULL nodemask--i.e., explicit local
allocation. Fix this by adding the check for MPOL_PREFERRED and empty
node mask to mpol_new_mpolicy().
Remove the now unneeded 'nodes = NULL' from mpol_new().
Note that mpol_new_mempolicy() is always called with a non-NULL
'nodes' parameter now that it has been removed from mpol_new().
Therefore, we don't need to test nodes for NULL before testing it for
'empty'. However, just to be extra paranoid, add a VM_BUG_ON() to
verify this assumption.]
[lee.schermerhorn@hp.com:
I don't think the function name 'mpol_new_mempolicy' is descriptive
enough to differentiate it from mpol_new().
This function applies cpuset set context, usually constraining nodes
to those allowed by the cpuset. However, when the 'RELATIVE_NODES flag
is set, it also translates the nodes. So I settled on
'mpol_set_nodemask()', because the comment block for mpol_new() mentions
that we need to call this function to "set nodes".
Some additional minor line length, whitespace and typo cleanup.]
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 15:31:49 -07:00
|
|
|
|
|
|
|
|
|
mm: make set_mempolicy(MPOL_INTERLEAV) N_HIGH_MEMORY aware
At first, init_task's mems_allowed is initialized as this.
init_task->mems_allowed == node_state[N_POSSIBLE]
And cpuset's top_cpuset mask is initialized as this
top_cpuset->mems_allowed = node_state[N_HIGH_MEMORY]
Before 2.6.29:
policy's mems_allowed is initialized as this.
1. update tasks->mems_allowed by its cpuset->mems_allowed.
2. policy->mems_allowed = nodes_and(tasks->mems_allowed, user's mask)
Updating task's mems_allowed in reference to top_cpuset's one.
cpuset's mems_allowed is aware of N_HIGH_MEMORY, always.
In 2.6.30: After commit 58568d2a8215cb6f55caf2332017d7bdff954e1c
("cpuset,mm: update tasks' mems_allowed in time"), policy's mems_allowed
is initialized as this.
1. policy->mems_allowd = nodes_and(task->mems_allowed, user's mask)
Here, if task is in top_cpuset, task->mems_allowed is not updated from
init's one. Assume user excutes command as #numactrl --interleave=all
,....
policy->mems_allowd = nodes_and(N_POSSIBLE, ALL_SET_MASK)
Then, policy's mems_allowd can includes a possible node, which has no pgdat.
MPOL's INTERLEAVE just scans nodemask of task->mems_allowd and access this
directly.
NODE_DATA(nid)->zonelist even if NODE_DATA(nid)==NULL
Then, what's we need is making policy->mems_allowed be aware of
N_HIGH_MEMORY. This patch does that. But to do so, extra nodemask will
be on statck. Because I know cpumask has a new interface of
CPUMASK_ALLOC(), I added it to node.
This patch stands on old behavior. But I feel this fix itself is just a
Band-Aid. But to do fundametal fix, we have to take care of memory
hotplug and it takes time. (task->mems_allowd should be N_HIGH_MEMORY, I
think.)
mpol_set_nodemask() should be aware of N_HIGH_MEMORY and policy's nodemask
should be includes only online nodes.
In old behavior, this is guaranteed by frequent reference to cpuset's
code. Now, most of them are removed and mempolicy has to check it by
itself.
To do check, a few nodemask_t will be used for calculating nodemask. But,
size of nodemask_t can be big and it's not good to allocate them on stack.
Now, cpumask_t has CPUMASK_ALLOC/FREE an easy code for get scratch area.
NODEMASK_ALLOC/FREE shoudl be there.
[akpm@linux-foundation.org: cleanups & tweaks]
Tested-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-08-06 15:07:33 -07:00
|
|
|
|
mm/mempolicy: don't handle MPOL_LOCAL like a fake MPOL_PREFERRED policy
MPOL_LOCAL policy has been setup as a real policy, but it is still handled
like a faked POL_PREFERRED policy with one internal MPOL_F_LOCAL flag bit
set, and there are many places having to judge the real 'prefer' or the
'local' policy, which are quite confusing.
In current code, there are 4 cases that MPOL_LOCAL are used:
1. user specifies 'local' policy
2. user specifies 'prefer' policy, but with empty nodemask
3. system 'default' policy is used
4. 'prefer' policy + valid 'preferred' node with MPOL_F_STATIC_NODES
flag set, and when it is 'rebind' to a nodemask which doesn't contains
the 'preferred' node, it will perform as 'local' policy
So make 'local' a real policy instead of a fake 'prefer' one, and kill
MPOL_F_LOCAL bit, which can greatly reduce the confusion for code reading.
For case 4, the logic of mpol_rebind_preferred() is confusing, as Michal
Hocko pointed out:
: I do believe that rebinding preferred policy is just bogus and it should
: be dropped altogether on the ground that a preference is a mere hint from
: userspace where to start the allocation. Unless I am missing something
: cpusets will be always authoritative for the final placement. The
: preferred node just acts as a starting point and it should be really
: preserved when cpusets changes. Otherwise we have a very subtle behavior
: corner cases.
So dump all the tricky transformation between 'prefer' and 'local', and
just record the new nodemask of rebinding.
[feng.tang@intel.com: fix a problem in mpol_set_nodemask(), per Michal Hocko]
Link: https://lkml.kernel.org/r/1622560492-1294-3-git-send-email-feng.tang@intel.com
[feng.tang@intel.com: refine code and comments of mpol_set_nodemask(), per Michal]
Link: https://lkml.kernel.org/r/20210603081807.GE56979@shbuild999.sh.intel.com
Link: https://lkml.kernel.org/r/1622469956-82897-3-git-send-email-feng.tang@intel.com
Signed-off-by: Feng Tang <feng.tang@intel.com>
Suggested-by: Michal Hocko <mhocko@suse.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ben Widawsky <ben.widawsky@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-30 18:51:00 -07:00
|
|
|
|
|
|
|
|
|
mm: make set_mempolicy(MPOL_INTERLEAV) N_HIGH_MEMORY aware
At first, init_task's mems_allowed is initialized as this.
init_task->mems_allowed == node_state[N_POSSIBLE]
And cpuset's top_cpuset mask is initialized as this
top_cpuset->mems_allowed = node_state[N_HIGH_MEMORY]
Before 2.6.29:
policy's mems_allowed is initialized as this.
1. update tasks->mems_allowed by its cpuset->mems_allowed.
2. policy->mems_allowed = nodes_and(tasks->mems_allowed, user's mask)
Updating task's mems_allowed in reference to top_cpuset's one.
cpuset's mems_allowed is aware of N_HIGH_MEMORY, always.
In 2.6.30: After commit 58568d2a8215cb6f55caf2332017d7bdff954e1c
("cpuset,mm: update tasks' mems_allowed in time"), policy's mems_allowed
is initialized as this.
1. policy->mems_allowd = nodes_and(task->mems_allowed, user's mask)
Here, if task is in top_cpuset, task->mems_allowed is not updated from
init's one. Assume user excutes command as #numactrl --interleave=all
,....
policy->mems_allowd = nodes_and(N_POSSIBLE, ALL_SET_MASK)
Then, policy's mems_allowd can includes a possible node, which has no pgdat.
MPOL's INTERLEAVE just scans nodemask of task->mems_allowd and access this
directly.
NODE_DATA(nid)->zonelist even if NODE_DATA(nid)==NULL
Then, what's we need is making policy->mems_allowed be aware of
N_HIGH_MEMORY. This patch does that. But to do so, extra nodemask will
be on statck. Because I know cpumask has a new interface of
CPUMASK_ALLOC(), I added it to node.
This patch stands on old behavior. But I feel this fix itself is just a
Band-Aid. But to do fundametal fix, we have to take care of memory
hotplug and it takes time. (task->mems_allowd should be N_HIGH_MEMORY, I
think.)
mpol_set_nodemask() should be aware of N_HIGH_MEMORY and policy's nodemask
should be includes only online nodes.
In old behavior, this is guaranteed by frequent reference to cpuset's
code. Now, most of them are removed and mempolicy has to check it by
itself.
To do check, a few nodemask_t will be used for calculating nodemask. But,
size of nodemask_t can be big and it's not good to allocate them on stack.
Now, cpumask_t has CPUMASK_ALLOC/FREE an easy code for get scratch area.
NODEMASK_ALLOC/FREE shoudl be there.
[akpm@linux-foundation.org: cleanups & tweaks]
Tested-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-08-06 15:07:33 -07:00
|
|
|
|
mm/mempolicy: don't handle MPOL_LOCAL like a fake MPOL_PREFERRED policy
MPOL_LOCAL policy has been setup as a real policy, but it is still handled
like a faked POL_PREFERRED policy with one internal MPOL_F_LOCAL flag bit
set, and there are many places having to judge the real 'prefer' or the
'local' policy, which are quite confusing.
In current code, there are 4 cases that MPOL_LOCAL are used:
1. user specifies 'local' policy
2. user specifies 'prefer' policy, but with empty nodemask
3. system 'default' policy is used
4. 'prefer' policy + valid 'preferred' node with MPOL_F_STATIC_NODES
flag set, and when it is 'rebind' to a nodemask which doesn't contains
the 'preferred' node, it will perform as 'local' policy
So make 'local' a real policy instead of a fake 'prefer' one, and kill
MPOL_F_LOCAL bit, which can greatly reduce the confusion for code reading.
For case 4, the logic of mpol_rebind_preferred() is confusing, as Michal
Hocko pointed out:
: I do believe that rebinding preferred policy is just bogus and it should
: be dropped altogether on the ground that a preference is a mere hint from
: userspace where to start the allocation. Unless I am missing something
: cpusets will be always authoritative for the final placement. The
: preferred node just acts as a starting point and it should be really
: preserved when cpusets changes. Otherwise we have a very subtle behavior
: corner cases.
So dump all the tricky transformation between 'prefer' and 'local', and
just record the new nodemask of rebinding.
[feng.tang@intel.com: fix a problem in mpol_set_nodemask(), per Michal Hocko]
Link: https://lkml.kernel.org/r/1622560492-1294-3-git-send-email-feng.tang@intel.com
[feng.tang@intel.com: refine code and comments of mpol_set_nodemask(), per Michal]
Link: https://lkml.kernel.org/r/20210603081807.GE56979@shbuild999.sh.intel.com
Link: https://lkml.kernel.org/r/1622469956-82897-3-git-send-email-feng.tang@intel.com
Signed-off-by: Feng Tang <feng.tang@intel.com>
Suggested-by: Michal Hocko <mhocko@suse.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ben Widawsky <ben.widawsky@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-30 18:51:00 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
cpuset,mm: update tasks' mems_allowed in time
Fix allocating page cache/slab object on the unallowed node when memory
spread is set by updating tasks' mems_allowed after its cpuset's mems is
changed.
In order to update tasks' mems_allowed in time, we must modify the code of
memory policy. Because the memory policy is applied in the process's
context originally. After applying this patch, one task directly
manipulates anothers mems_allowed, and we use alloc_lock in the
task_struct to protect mems_allowed and memory policy of the task.
But in the fast path, we didn't use lock to protect them, because adding a
lock may lead to performance regression. But if we don't add a lock,the
task might see no nodes when changing cpuset's mems_allowed to some
non-overlapping set. In order to avoid it, we set all new allowed nodes,
then clear newly disallowed ones.
[lee.schermerhorn@hp.com:
The rework of mpol_new() to extract the adjusting of the node mask to
apply cpuset and mpol flags "context" breaks set_mempolicy() and mbind()
with MPOL_PREFERRED and a NULL nodemask--i.e., explicit local
allocation. Fix this by adding the check for MPOL_PREFERRED and empty
node mask to mpol_new_mpolicy().
Remove the now unneeded 'nodes = NULL' from mpol_new().
Note that mpol_new_mempolicy() is always called with a non-NULL
'nodes' parameter now that it has been removed from mpol_new().
Therefore, we don't need to test nodes for NULL before testing it for
'empty'. However, just to be extra paranoid, add a VM_BUG_ON() to
verify this assumption.]
[lee.schermerhorn@hp.com:
I don't think the function name 'mpol_new_mempolicy' is descriptive
enough to differentiate it from mpol_new().
This function applies cpuset set context, usually constraining nodes
to those allowed by the cpuset. However, when the 'RELATIVE_NODES flag
is set, it also translates the nodes. So I settled on
'mpol_set_nodemask()', because the comment block for mpol_new() mentions
that we need to call this function to "set nodes".
Some additional minor line length, whitespace and typo cleanup.]
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 15:31:49 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-04-28 02:12:25 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-04-28 02:12:34 -07:00
|
|
|
|
|
|
|
|
|
2008-04-28 02:12:33 -07:00
|
|
|
|
2012-10-25 14:16:29 +02:00
|
|
|
|
2008-04-28 02:12:33 -07:00
|
|
|
|
2008-04-28 02:12:34 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: don't handle MPOL_LOCAL like a fake MPOL_PREFERRED policy
MPOL_LOCAL policy has been setup as a real policy, but it is still handled
like a faked POL_PREFERRED policy with one internal MPOL_F_LOCAL flag bit
set, and there are many places having to judge the real 'prefer' or the
'local' policy, which are quite confusing.
In current code, there are 4 cases that MPOL_LOCAL are used:
1. user specifies 'local' policy
2. user specifies 'prefer' policy, but with empty nodemask
3. system 'default' policy is used
4. 'prefer' policy + valid 'preferred' node with MPOL_F_STATIC_NODES
flag set, and when it is 'rebind' to a nodemask which doesn't contains
the 'preferred' node, it will perform as 'local' policy
So make 'local' a real policy instead of a fake 'prefer' one, and kill
MPOL_F_LOCAL bit, which can greatly reduce the confusion for code reading.
For case 4, the logic of mpol_rebind_preferred() is confusing, as Michal
Hocko pointed out:
: I do believe that rebinding preferred policy is just bogus and it should
: be dropped altogether on the ground that a preference is a mere hint from
: userspace where to start the allocation. Unless I am missing something
: cpusets will be always authoritative for the final placement. The
: preferred node just acts as a starting point and it should be really
: preserved when cpusets changes. Otherwise we have a very subtle behavior
: corner cases.
So dump all the tricky transformation between 'prefer' and 'local', and
just record the new nodemask of rebinding.
[feng.tang@intel.com: fix a problem in mpol_set_nodemask(), per Michal Hocko]
Link: https://lkml.kernel.org/r/1622560492-1294-3-git-send-email-feng.tang@intel.com
[feng.tang@intel.com: refine code and comments of mpol_set_nodemask(), per Michal]
Link: https://lkml.kernel.org/r/20210603081807.GE56979@shbuild999.sh.intel.com
Link: https://lkml.kernel.org/r/1622469956-82897-3-git-send-email-feng.tang@intel.com
Signed-off-by: Feng Tang <feng.tang@intel.com>
Suggested-by: Michal Hocko <mhocko@suse.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ben Widawsky <ben.widawsky@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-30 18:51:00 -07:00
|
|
|
|
|
|
|
|
|
2008-04-28 02:12:34 -07:00
|
|
|
|
2012-10-25 14:16:28 +02:00
|
|
|
|
2016-12-12 16:42:49 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-10-25 14:16:28 +02:00
|
|
|
|
2008-04-28 02:12:34 -07:00
|
|
|
|
|
|
|
|
|
2023-10-03 02:20:14 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-04-28 02:13:12 -07:00
|
|
|
|
2008-04-28 02:12:34 -07:00
|
|
|
|
mm/mempolicy: add set_mempolicy_home_node syscall
This syscall can be used to set a home node for the MPOL_BIND and
MPOL_PREFERRED_MANY memory policy. Users should use this syscall after
setting up a memory policy for the specified range as shown below.
mbind(p, nr_pages * page_size, MPOL_BIND, new_nodes->maskp,
new_nodes->size + 1, 0);
sys_set_mempolicy_home_node((unsigned long)p, nr_pages * page_size,
home_node, 0);
The syscall allows specifying a home node/preferred node from which
kernel will fulfill memory allocation requests first.
For address range with MPOL_BIND memory policy, if nodemask specifies
more than one node, page allocations will come from the node in the
nodemask with sufficient free memory that is closest to the home
node/preferred node.
For MPOL_PREFERRED_MANY if the nodemask specifies more than one node,
page allocation will come from the node in the nodemask with sufficient
free memory that is closest to the home node/preferred node. If there
is not enough memory in all the nodes specified in the nodemask, the
allocation will be attempted from the closest numa node to the home node
in the system.
This helps applications to hint at a memory allocation preference node
and fallback to _only_ a set of nodes if the memory is not available on
the preferred node. Fallback allocation is attempted from the node
which is nearest to the preferred node.
This helps applications to have control on memory allocation numa nodes
and avoids default fallback to slow memory NUMA nodes. For example a
system with NUMA nodes 1,2 and 3 with DRAM memory and 10, 11 and 12 of
slow memory
new_nodes = numa_bitmask_alloc(nr_nodes);
numa_bitmask_setbit(new_nodes, 1);
numa_bitmask_setbit(new_nodes, 2);
numa_bitmask_setbit(new_nodes, 3);
p = mmap(NULL, nr_pages * page_size, protflag, mapflag, -1, 0);
mbind(p, nr_pages * page_size, MPOL_BIND, new_nodes->maskp, new_nodes->size + 1, 0);
sys_set_mempolicy_home_node(p, nr_pages * page_size, 2, 0);
This will allocate from nodes closer to node 2 and will make sure the
kernel will only allocate from nodes 1, 2, and 3. Memory will not be
allocated from slow memory nodes 10, 11, and 12. This differs from
default MPOL_BIND behavior in that with default MPOL_BIND the allocation
will be attempted from node closer to the local node. One of the
reasons to specify a home node is to allow allocations from cpu less
NUMA node and its nearby NUMA nodes.
With MPOL_PREFERRED_MANY on the other hand will first try to allocate
from the closest node to node 2 from the node list 1, 2 and 3. If those
nodes don't have enough memory, kernel will allocate from slow memory
node 10, 11 and 12 which ever is closer to node 2.
Link: https://lkml.kernel.org/r/20211202123810.267175-3-aneesh.kumar@linux.ibm.com
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Ben Widawsky <ben.widawsky@intel.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Feng Tang <feng.tang@intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: <linux-api@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-14 14:08:17 -08:00
|
|
|
|
mempolicy: add MPOL_F_STATIC_NODES flag
Add an optional mempolicy mode flag, MPOL_F_STATIC_NODES, that suppresses the
node remap when the policy is rebound.
Adds another member to struct mempolicy, nodemask_t user_nodemask, as part of
a union with cpuset_mems_allowed:
struct mempolicy {
...
union {
nodemask_t cpuset_mems_allowed;
nodemask_t user_nodemask;
} w;
}
that stores the the nodemask that the user passed when he or she created the
mempolicy via set_mempolicy() or mbind(). When using MPOL_F_STATIC_NODES,
which is passed with any mempolicy mode, the user's passed nodemask
intersected with the VMA or task's allowed nodes is always used when
determining the preferred node, setting the MPOL_BIND zonelist, or creating
the interleave nodemask. This happens whenever the policy is rebound,
including when a task's cpuset assignment changes or the cpuset's mems are
changed.
This creates an interesting side-effect in that it allows the mempolicy
"intent" to lie dormant and uneffected until it has access to the node(s) that
it desires. For example, if you currently ask for an interleaved policy over
a set of nodes that you do not have access to, the mempolicy is not created
and the task continues to use the previous policy. With this change, however,
it is possible to create the same mempolicy; it is only effected when access
to nodes in the nodemask is acquired.
It is also possible to mount tmpfs with the static nodemask behavior when
specifying a node or nodemask. To do this, simply add "=static" immediately
following the mempolicy mode at mount time:
mount -o remount mpol=interleave=static:1-3
Also removes mpol_check_policy() and folds its logic into mpol_new() since it
is now obsoleted. The unused vma_mpol_equal() is also removed.
Cc: Paul Jackson <pj@sgi.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:12:27 -07:00
|
|
|
|
2008-04-28 02:12:33 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: rework mempolicy Reference Counting [yet again]
After further discussion with Christoph Lameter, it has become clear that my
earlier attempts to clean up the mempolicy reference counting were a bit of
overkill in some areas, resulting in superflous ref/unref in what are usually
fast paths. In other areas, further inspection reveals that I botched the
unref for interleave policies.
A separate patch, suitable for upstream/stable trees, fixes up the known
errors in the previous attempt to fix reference counting.
This patch reworks the memory policy referencing counting and, one hopes,
simplifies the code. Maybe I'll get it right this time.
See the update to the numa_memory_policy.txt document for a discussion of
memory policy reference counting that motivates this patch.
Summary:
Lookup of mempolicy, based on (vma, address) need only add a reference for
shared policy, and we need only unref the policy when finished for shared
policies. So, this patch backs out all of the unneeded extra reference
counting added by my previous attempt. It then unrefs only shared policies
when we're finished with them, using the mpol_cond_put() [conditional put]
helper function introduced by this patch.
Note that shmem_swapin() calls read_swap_cache_async() with a dummy vma
containing just the policy. read_swap_cache_async() can call alloc_page_vma()
multiple times, so we can't let alloc_page_vma() unref the shared policy in
this case. To avoid this, we make a copy of any non-null shared policy and
remove the MPOL_F_SHARED flag from the copy. This copy occurs before reading
a page [or multiple pages] from swap, so the overhead should not be an issue
here.
I introduced a new static inline function "mpol_cond_copy()" to copy the
shared policy to an on-stack policy and remove the flags that would require a
conditional free. The current implementation of mpol_cond_copy() assumes that
the struct mempolicy contains no pointers to dynamically allocated structures
that must be duplicated or reference counted during copy.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:16 -07:00
|
|
|
|
2023-10-03 02:20:14 -07:00
|
|
|
|
mempolicy: rework mempolicy Reference Counting [yet again]
After further discussion with Christoph Lameter, it has become clear that my
earlier attempts to clean up the mempolicy reference counting were a bit of
overkill in some areas, resulting in superflous ref/unref in what are usually
fast paths. In other areas, further inspection reveals that I botched the
unref for interleave policies.
A separate patch, suitable for upstream/stable trees, fixes up the known
errors in the previous attempt to fix reference counting.
This patch reworks the memory policy referencing counting and, one hopes,
simplifies the code. Maybe I'll get it right this time.
See the update to the numa_memory_policy.txt document for a discussion of
memory policy reference counting that motivates this patch.
Summary:
Lookup of mempolicy, based on (vma, address) need only add a reference for
shared policy, and we need only unref the policy when finished for shared
policies. So, this patch backs out all of the unneeded extra reference
counting added by my previous attempt. It then unrefs only shared policies
when we're finished with them, using the mpol_cond_put() [conditional put]
helper function introduced by this patch.
Note that shmem_swapin() calls read_swap_cache_async() with a dummy vma
containing just the policy. read_swap_cache_async() can call alloc_page_vma()
multiple times, so we can't let alloc_page_vma() unref the shared policy in
this case. To avoid this, we make a copy of any non-null shared policy and
remove the MPOL_F_SHARED flag from the copy. This copy occurs before reading
a page [or multiple pages] from swap, so the overhead should not be an issue
here.
I introduced a new static inline function "mpol_cond_copy()" to copy the
shared policy to an on-stack policy and remove the flags that would require a
conditional free. The current implementation of mpol_cond_copy() assumes that
the struct mempolicy contains no pointers to dynamically allocated structures
that must be duplicated or reference counted during copy.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:16 -07:00
|
|
|
|
2023-10-03 02:20:14 -07:00
|
|
|
|
mempolicy: rework mempolicy Reference Counting [yet again]
After further discussion with Christoph Lameter, it has become clear that my
earlier attempts to clean up the mempolicy reference counting were a bit of
overkill in some areas, resulting in superflous ref/unref in what are usually
fast paths. In other areas, further inspection reveals that I botched the
unref for interleave policies.
A separate patch, suitable for upstream/stable trees, fixes up the known
errors in the previous attempt to fix reference counting.
This patch reworks the memory policy referencing counting and, one hopes,
simplifies the code. Maybe I'll get it right this time.
See the update to the numa_memory_policy.txt document for a discussion of
memory policy reference counting that motivates this patch.
Summary:
Lookup of mempolicy, based on (vma, address) need only add a reference for
shared policy, and we need only unref the policy when finished for shared
policies. So, this patch backs out all of the unneeded extra reference
counting added by my previous attempt. It then unrefs only shared policies
when we're finished with them, using the mpol_cond_put() [conditional put]
helper function introduced by this patch.
Note that shmem_swapin() calls read_swap_cache_async() with a dummy vma
containing just the policy. read_swap_cache_async() can call alloc_page_vma()
multiple times, so we can't let alloc_page_vma() unref the shared policy in
this case. To avoid this, we make a copy of any non-null shared policy and
remove the MPOL_F_SHARED flag from the copy. This copy occurs before reading
a page [or multiple pages] from swap, so the overhead should not be an issue
here.
I introduced a new static inline function "mpol_cond_copy()" to copy the
shared policy to an on-stack policy and remove the flags that would require a
conditional free. The current implementation of mpol_cond_copy() assumes that
the struct mempolicy contains no pointers to dynamically allocated structures
that must be duplicated or reference counted during copy.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:16 -07:00
|
|
|
|
2026-03-13 20:47:56 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: rework mempolicy Reference Counting [yet again]
After further discussion with Christoph Lameter, it has become clear that my
earlier attempts to clean up the mempolicy reference counting were a bit of
overkill in some areas, resulting in superflous ref/unref in what are usually
fast paths. In other areas, further inspection reveals that I botched the
unref for interleave policies.
A separate patch, suitable for upstream/stable trees, fixes up the known
errors in the previous attempt to fix reference counting.
This patch reworks the memory policy referencing counting and, one hopes,
simplifies the code. Maybe I'll get it right this time.
See the update to the numa_memory_policy.txt document for a discussion of
memory policy reference counting that motivates this patch.
Summary:
Lookup of mempolicy, based on (vma, address) need only add a reference for
shared policy, and we need only unref the policy when finished for shared
policies. So, this patch backs out all of the unneeded extra reference
counting added by my previous attempt. It then unrefs only shared policies
when we're finished with them, using the mpol_cond_put() [conditional put]
helper function introduced by this patch.
Note that shmem_swapin() calls read_swap_cache_async() with a dummy vma
containing just the policy. read_swap_cache_async() can call alloc_page_vma()
multiple times, so we can't let alloc_page_vma() unref the shared policy in
this case. To avoid this, we make a copy of any non-null shared policy and
remove the MPOL_F_SHARED flag from the copy. This copy occurs before reading
a page [or multiple pages] from swap, so the overhead should not be an issue
here.
I introduced a new static inline function "mpol_cond_copy()" to copy the
shared policy to an on-stack policy and remove the flags that would require a
conditional free. The current implementation of mpol_cond_copy() assumes that
the struct mempolicy contains no pointers to dynamically allocated structures
that must be duplicated or reference counted during copy.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:16 -07:00
|
|
|
|
2025-08-27 17:52:45 +00:00
|
|
|
|
mempolicy: rework mempolicy Reference Counting [yet again]
After further discussion with Christoph Lameter, it has become clear that my
earlier attempts to clean up the mempolicy reference counting were a bit of
overkill in some areas, resulting in superflous ref/unref in what are usually
fast paths. In other areas, further inspection reveals that I botched the
unref for interleave policies.
A separate patch, suitable for upstream/stable trees, fixes up the known
errors in the previous attempt to fix reference counting.
This patch reworks the memory policy referencing counting and, one hopes,
simplifies the code. Maybe I'll get it right this time.
See the update to the numa_memory_policy.txt document for a discussion of
memory policy reference counting that motivates this patch.
Summary:
Lookup of mempolicy, based on (vma, address) need only add a reference for
shared policy, and we need only unref the policy when finished for shared
policies. So, this patch backs out all of the unneeded extra reference
counting added by my previous attempt. It then unrefs only shared policies
when we're finished with them, using the mpol_cond_put() [conditional put]
helper function introduced by this patch.
Note that shmem_swapin() calls read_swap_cache_async() with a dummy vma
containing just the policy. read_swap_cache_async() can call alloc_page_vma()
multiple times, so we can't let alloc_page_vma() unref the shared policy in
this case. To avoid this, we make a copy of any non-null shared policy and
remove the MPOL_F_SHARED flag from the copy. This copy occurs before reading
a page [or multiple pages] from swap, so the overhead should not be an issue
here.
I introduced a new static inline function "mpol_cond_copy()" to copy the
shared policy to an on-stack policy and remove the flags that would require a
conditional free. The current implementation of mpol_cond_copy() assumes that
the struct mempolicy contains no pointers to dynamically allocated structures
that must be duplicated or reference counted during copy.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:16 -07:00
|
|
|
|
2017-07-06 15:40:06 -07:00
|
|
|
|
2008-04-28 02:12:33 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-07-06 15:40:06 -07:00
|
|
|
|
2008-04-28 02:12:33 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-30 18:51:10 -07:00
|
|
|
|
2017-07-06 15:40:06 -07:00
|
|
|
|
2019-06-28 12:06:43 -07:00
|
|
|
|
2008-04-28 02:12:33 -07:00
|
|
|
|
|
|
|
|
|
mempolicy: restructure rebinding-mempolicy functions
Nick Piggin reported that the allocator may see an empty nodemask when
changing cpuset's mems[1]. It happens only on the kernel that do not do
atomic nodemask_t stores. (MAX_NUMNODES > BITS_PER_LONG)
But I found that there is also a problem on the kernel that can do atomic
nodemask_t stores. The problem is that the allocator can't find a node to
alloc page when changing cpuset's mems though there is a lot of free
memory. The reason is like this:
(mpol: mempolicy)
task1 task1's mpol task2
alloc page 1
alloc on node0? NO 1
1 change mems from 1 to 0
1 rebind task1's mpol
0-1 set new bits
0 clear disallowed bits
alloc on node1? NO 0
...
can't alloc page
goto oom
I can use the attached program reproduce it by the following step:
# mkdir /dev/cpuset
# mount -t cpuset cpuset /dev/cpuset
# mkdir /dev/cpuset/1
# echo `cat /dev/cpuset/cpus` > /dev/cpuset/1/cpus
# echo `cat /dev/cpuset/mems` > /dev/cpuset/1/mems
# echo $$ > /dev/cpuset/1/tasks
# numactl --membind=`cat /dev/cpuset/mems` ./cpuset_mem_hog <nr_tasks> &
<nr_tasks> = max(nr_cpus - 1, 1)
# killall -s SIGUSR1 cpuset_mem_hog
# ./change_mems.sh
several hours later, oom will happen though there is a lot of free memory.
This patchset fixes this problem by expanding the nodes range first(set
newly allowed bits) and shrink it lazily(clear newly disallowed bits). So
we use a variable to tell the write-side task that read-side task is
reading nodemask, and the write-side task clears newly disallowed nodes
after read-side task ends the current memory allocation.
This patch:
In order to fix no node to alloc memory, when we want to update mempolicy
and mems_allowed, we expand the set of nodes first (set all the newly
nodes) and shrink the set of nodes lazily(clean disallowed nodes), But the
mempolicy's rebind functions may breaks the expanding.
So we restructure the mempolicy's rebind functions and split the rebind
work to two steps, just like the update of cpuset's mems: The 1st step:
expand the set of the mempolicy's nodes. The 2nd step: shrink the set of
the mempolicy's nodes. It is used when there is no real lock to protect
the mempolicy in the read-side. Otherwise we can do rebind work at once.
In order to implement it, we define
enum mpol_rebind_step {
MPOL_REBIND_ONCE,
MPOL_REBIND_STEP1,
MPOL_REBIND_STEP2,
MPOL_REBIND_NSTEP,
};
If the mempolicy needn't be updated by two steps, we can pass
MPOL_REBIND_ONCE to the rebind functions. Or we can pass
MPOL_REBIND_STEP1 to do the first step of the rebind work and pass
MPOL_REBIND_STEP2 to do the second step work.
Besides that, it maybe long time between these two step and we have to
release the lock that protects mempolicy and mems_allowed. If we hold the
lock once again, we must check whether the current mempolicy is under the
rebinding (the first step has been done) or not, because the task may
alloc a new mempolicy when we don't hold the lock. So we defined the
following flag to identify it:
#define MPOL_F_REBINDING (1 << 2)
The new functions will be used in the next patch.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Nick Piggin <npiggin@suse.de>
Cc: Paul Menage <menage@google.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Cc: Ravikiran Thirumalai <kiran@scalex86.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Andi Kleen <andi@firstfloor.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-24 14:32:07 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-30 18:51:10 -07:00
|
|
|
|
2008-04-28 02:12:33 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-07-06 15:40:06 -07:00
|
|
|
|
2008-04-28 02:12:33 -07:00
|
|
|
|
mm/mempolicy: don't handle MPOL_LOCAL like a fake MPOL_PREFERRED policy
MPOL_LOCAL policy has been setup as a real policy, but it is still handled
like a faked POL_PREFERRED policy with one internal MPOL_F_LOCAL flag bit
set, and there are many places having to judge the real 'prefer' or the
'local' policy, which are quite confusing.
In current code, there are 4 cases that MPOL_LOCAL are used:
1. user specifies 'local' policy
2. user specifies 'prefer' policy, but with empty nodemask
3. system 'default' policy is used
4. 'prefer' policy + valid 'preferred' node with MPOL_F_STATIC_NODES
flag set, and when it is 'rebind' to a nodemask which doesn't contains
the 'preferred' node, it will perform as 'local' policy
So make 'local' a real policy instead of a fake 'prefer' one, and kill
MPOL_F_LOCAL bit, which can greatly reduce the confusion for code reading.
For case 4, the logic of mpol_rebind_preferred() is confusing, as Michal
Hocko pointed out:
: I do believe that rebinding preferred policy is just bogus and it should
: be dropped altogether on the ground that a preference is a mere hint from
: userspace where to start the allocation. Unless I am missing something
: cpusets will be always authoritative for the final placement. The
: preferred node just acts as a starting point and it should be really
: preserved when cpusets changes. Otherwise we have a very subtle behavior
: corner cases.
So dump all the tricky transformation between 'prefer' and 'local', and
just record the new nodemask of rebinding.
[feng.tang@intel.com: fix a problem in mpol_set_nodemask(), per Michal Hocko]
Link: https://lkml.kernel.org/r/1622560492-1294-3-git-send-email-feng.tang@intel.com
[feng.tang@intel.com: refine code and comments of mpol_set_nodemask(), per Michal]
Link: https://lkml.kernel.org/r/20210603081807.GE56979@shbuild999.sh.intel.com
Link: https://lkml.kernel.org/r/1622469956-82897-3-git-send-email-feng.tang@intel.com
Signed-off-by: Feng Tang <feng.tang@intel.com>
Suggested-by: Michal Hocko <mhocko@suse.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ben Widawsky <ben.widawsky@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-30 18:51:00 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
mempolicy: restructure rebinding-mempolicy functions
Nick Piggin reported that the allocator may see an empty nodemask when
changing cpuset's mems[1]. It happens only on the kernel that do not do
atomic nodemask_t stores. (MAX_NUMNODES > BITS_PER_LONG)
But I found that there is also a problem on the kernel that can do atomic
nodemask_t stores. The problem is that the allocator can't find a node to
alloc page when changing cpuset's mems though there is a lot of free
memory. The reason is like this:
(mpol: mempolicy)
task1 task1's mpol task2
alloc page 1
alloc on node0? NO 1
1 change mems from 1 to 0
1 rebind task1's mpol
0-1 set new bits
0 clear disallowed bits
alloc on node1? NO 0
...
can't alloc page
goto oom
I can use the attached program reproduce it by the following step:
# mkdir /dev/cpuset
# mount -t cpuset cpuset /dev/cpuset
# mkdir /dev/cpuset/1
# echo `cat /dev/cpuset/cpus` > /dev/cpuset/1/cpus
# echo `cat /dev/cpuset/mems` > /dev/cpuset/1/mems
# echo $$ > /dev/cpuset/1/tasks
# numactl --membind=`cat /dev/cpuset/mems` ./cpuset_mem_hog <nr_tasks> &
<nr_tasks> = max(nr_cpus - 1, 1)
# killall -s SIGUSR1 cpuset_mem_hog
# ./change_mems.sh
several hours later, oom will happen though there is a lot of free memory.
This patchset fixes this problem by expanding the nodes range first(set
newly allowed bits) and shrink it lazily(clear newly disallowed bits). So
we use a variable to tell the write-side task that read-side task is
reading nodemask, and the write-side task clears newly disallowed nodes
after read-side task ends the current memory allocation.
This patch:
In order to fix no node to alloc memory, when we want to update mempolicy
and mems_allowed, we expand the set of nodes first (set all the newly
nodes) and shrink the set of nodes lazily(clean disallowed nodes), But the
mempolicy's rebind functions may breaks the expanding.
So we restructure the mempolicy's rebind functions and split the rebind
work to two steps, just like the update of cpuset's mems: The 1st step:
expand the set of the mempolicy's nodes. The 2nd step: shrink the set of
the mempolicy's nodes. It is used when there is no real lock to protect
the mempolicy in the read-side. Otherwise we can do rebind work at once.
In order to implement it, we define
enum mpol_rebind_step {
MPOL_REBIND_ONCE,
MPOL_REBIND_STEP1,
MPOL_REBIND_STEP2,
MPOL_REBIND_NSTEP,
};
If the mempolicy needn't be updated by two steps, we can pass
MPOL_REBIND_ONCE to the rebind functions. Or we can pass
MPOL_REBIND_STEP1 to do the first step of the rebind work and pass
MPOL_REBIND_STEP2 to do the second step work.
Besides that, it maybe long time between these two step and we have to
release the lock that protects mempolicy and mems_allowed. If we hold the
lock once again, we must check whether the current mempolicy is under the
rebinding (the first step has been done) or not, because the task may
alloc a new mempolicy when we don't hold the lock. So we defined the
following flag to identify it:
#define MPOL_F_REBINDING (1 << 2)
The new functions will be used in the next patch.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Nick Piggin <npiggin@suse.de>
Cc: Paul Menage <menage@google.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Cc: Ravikiran Thirumalai <kiran@scalex86.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Andi Kleen <andi@firstfloor.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-24 14:32:07 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-06-08 21:33:54 -07:00
|
|
|
|
2017-07-06 15:40:06 -07:00
|
|
|
|
|
|
|
|
|
mempolicy: restructure rebinding-mempolicy functions
Nick Piggin reported that the allocator may see an empty nodemask when
changing cpuset's mems[1]. It happens only on the kernel that do not do
atomic nodemask_t stores. (MAX_NUMNODES > BITS_PER_LONG)
But I found that there is also a problem on the kernel that can do atomic
nodemask_t stores. The problem is that the allocator can't find a node to
alloc page when changing cpuset's mems though there is a lot of free
memory. The reason is like this:
(mpol: mempolicy)
task1 task1's mpol task2
alloc page 1
alloc on node0? NO 1
1 change mems from 1 to 0
1 rebind task1's mpol
0-1 set new bits
0 clear disallowed bits
alloc on node1? NO 0
...
can't alloc page
goto oom
I can use the attached program reproduce it by the following step:
# mkdir /dev/cpuset
# mount -t cpuset cpuset /dev/cpuset
# mkdir /dev/cpuset/1
# echo `cat /dev/cpuset/cpus` > /dev/cpuset/1/cpus
# echo `cat /dev/cpuset/mems` > /dev/cpuset/1/mems
# echo $$ > /dev/cpuset/1/tasks
# numactl --membind=`cat /dev/cpuset/mems` ./cpuset_mem_hog <nr_tasks> &
<nr_tasks> = max(nr_cpus - 1, 1)
# killall -s SIGUSR1 cpuset_mem_hog
# ./change_mems.sh
several hours later, oom will happen though there is a lot of free memory.
This patchset fixes this problem by expanding the nodes range first(set
newly allowed bits) and shrink it lazily(clear newly disallowed bits). So
we use a variable to tell the write-side task that read-side task is
reading nodemask, and the write-side task clears newly disallowed nodes
after read-side task ends the current memory allocation.
This patch:
In order to fix no node to alloc memory, when we want to update mempolicy
and mems_allowed, we expand the set of nodes first (set all the newly
nodes) and shrink the set of nodes lazily(clean disallowed nodes), But the
mempolicy's rebind functions may breaks the expanding.
So we restructure the mempolicy's rebind functions and split the rebind
work to two steps, just like the update of cpuset's mems: The 1st step:
expand the set of the mempolicy's nodes. The 2nd step: shrink the set of
the mempolicy's nodes. It is used when there is no real lock to protect
the mempolicy in the read-side. Otherwise we can do rebind work at once.
In order to implement it, we define
enum mpol_rebind_step {
MPOL_REBIND_ONCE,
MPOL_REBIND_STEP1,
MPOL_REBIND_STEP2,
MPOL_REBIND_NSTEP,
};
If the mempolicy needn't be updated by two steps, we can pass
MPOL_REBIND_ONCE to the rebind functions. Or we can pass
MPOL_REBIND_STEP1 to do the first step of the rebind work and pass
MPOL_REBIND_STEP2 to do the second step work.
Besides that, it maybe long time between these two step and we have to
release the lock that protects mempolicy and mems_allowed. If we hold the
lock once again, we must check whether the current mempolicy is under the
rebinding (the first step has been done) or not, because the task may
alloc a new mempolicy when we don't hold the lock. So we defined the
following flag to identify it:
#define MPOL_F_REBINDING (1 << 2)
The new functions will be used in the next patch.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Nick Piggin <npiggin@suse.de>
Cc: Paul Menage <menage@google.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Cc: Ravikiran Thirumalai <kiran@scalex86.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Andi Kleen <andi@firstfloor.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-24 14:32:07 -07:00
|
|
|
|
2017-07-06 15:40:06 -07:00
|
|
|
|
2008-04-28 02:12:32 -07:00
|
|
|
|
2022-05-19 14:08:54 -07:00
|
|
|
|
2008-04-28 02:12:32 -07:00
|
|
|
|
mm/mempolicy: don't handle MPOL_LOCAL like a fake MPOL_PREFERRED policy
MPOL_LOCAL policy has been setup as a real policy, but it is still handled
like a faked POL_PREFERRED policy with one internal MPOL_F_LOCAL flag bit
set, and there are many places having to judge the real 'prefer' or the
'local' policy, which are quite confusing.
In current code, there are 4 cases that MPOL_LOCAL are used:
1. user specifies 'local' policy
2. user specifies 'prefer' policy, but with empty nodemask
3. system 'default' policy is used
4. 'prefer' policy + valid 'preferred' node with MPOL_F_STATIC_NODES
flag set, and when it is 'rebind' to a nodemask which doesn't contains
the 'preferred' node, it will perform as 'local' policy
So make 'local' a real policy instead of a fake 'prefer' one, and kill
MPOL_F_LOCAL bit, which can greatly reduce the confusion for code reading.
For case 4, the logic of mpol_rebind_preferred() is confusing, as Michal
Hocko pointed out:
: I do believe that rebinding preferred policy is just bogus and it should
: be dropped altogether on the ground that a preference is a mere hint from
: userspace where to start the allocation. Unless I am missing something
: cpusets will be always authoritative for the final placement. The
: preferred node just acts as a starting point and it should be really
: preserved when cpusets changes. Otherwise we have a very subtle behavior
: corner cases.
So dump all the tricky transformation between 'prefer' and 'local', and
just record the new nodemask of rebinding.
[feng.tang@intel.com: fix a problem in mpol_set_nodemask(), per Michal Hocko]
Link: https://lkml.kernel.org/r/1622560492-1294-3-git-send-email-feng.tang@intel.com
[feng.tang@intel.com: refine code and comments of mpol_set_nodemask(), per Michal]
Link: https://lkml.kernel.org/r/20210603081807.GE56979@shbuild999.sh.intel.com
Link: https://lkml.kernel.org/r/1622469956-82897-3-git-send-email-feng.tang@intel.com
Signed-off-by: Feng Tang <feng.tang@intel.com>
Suggested-by: Michal Hocko <mhocko@suse.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ben Widawsky <ben.widawsky@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-30 18:51:00 -07:00
|
|
|
|
2008-04-28 02:12:32 -07:00
|
|
|
|
|
|
|
|
|
mempolicy: restructure rebinding-mempolicy functions
Nick Piggin reported that the allocator may see an empty nodemask when
changing cpuset's mems[1]. It happens only on the kernel that do not do
atomic nodemask_t stores. (MAX_NUMNODES > BITS_PER_LONG)
But I found that there is also a problem on the kernel that can do atomic
nodemask_t stores. The problem is that the allocator can't find a node to
alloc page when changing cpuset's mems though there is a lot of free
memory. The reason is like this:
(mpol: mempolicy)
task1 task1's mpol task2
alloc page 1
alloc on node0? NO 1
1 change mems from 1 to 0
1 rebind task1's mpol
0-1 set new bits
0 clear disallowed bits
alloc on node1? NO 0
...
can't alloc page
goto oom
I can use the attached program reproduce it by the following step:
# mkdir /dev/cpuset
# mount -t cpuset cpuset /dev/cpuset
# mkdir /dev/cpuset/1
# echo `cat /dev/cpuset/cpus` > /dev/cpuset/1/cpus
# echo `cat /dev/cpuset/mems` > /dev/cpuset/1/mems
# echo $$ > /dev/cpuset/1/tasks
# numactl --membind=`cat /dev/cpuset/mems` ./cpuset_mem_hog <nr_tasks> &
<nr_tasks> = max(nr_cpus - 1, 1)
# killall -s SIGUSR1 cpuset_mem_hog
# ./change_mems.sh
several hours later, oom will happen though there is a lot of free memory.
This patchset fixes this problem by expanding the nodes range first(set
newly allowed bits) and shrink it lazily(clear newly disallowed bits). So
we use a variable to tell the write-side task that read-side task is
reading nodemask, and the write-side task clears newly disallowed nodes
after read-side task ends the current memory allocation.
This patch:
In order to fix no node to alloc memory, when we want to update mempolicy
and mems_allowed, we expand the set of nodes first (set all the newly
nodes) and shrink the set of nodes lazily(clean disallowed nodes), But the
mempolicy's rebind functions may breaks the expanding.
So we restructure the mempolicy's rebind functions and split the rebind
work to two steps, just like the update of cpuset's mems: The 1st step:
expand the set of the mempolicy's nodes. The 2nd step: shrink the set of
the mempolicy's nodes. It is used when there is no real lock to protect
the mempolicy in the read-side. Otherwise we can do rebind work at once.
In order to implement it, we define
enum mpol_rebind_step {
MPOL_REBIND_ONCE,
MPOL_REBIND_STEP1,
MPOL_REBIND_STEP2,
MPOL_REBIND_NSTEP,
};
If the mempolicy needn't be updated by two steps, we can pass
MPOL_REBIND_ONCE to the rebind functions. Or we can pass
MPOL_REBIND_STEP1 to do the first step of the rebind work and pass
MPOL_REBIND_STEP2 to do the second step work.
Besides that, it maybe long time between these two step and we have to
release the lock that protects mempolicy and mems_allowed. If we hold the
lock once again, we must check whether the current mempolicy is under the
rebinding (the first step has been done) or not, because the task may
alloc a new mempolicy when we don't hold the lock. So we defined the
following flag to identify it:
#define MPOL_F_REBINDING (1 << 2)
The new functions will be used in the next patch.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Nick Piggin <npiggin@suse.de>
Cc: Paul Menage <menage@google.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Cc: Ravikiran Thirumalai <kiran@scalex86.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Andi Kleen <andi@firstfloor.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-24 14:32:07 -07:00
|
|
|
|
2017-07-06 15:40:06 -07:00
|
|
|
|
2008-04-28 02:12:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
cpuset,mm: update tasks' mems_allowed in time
Fix allocating page cache/slab object on the unallowed node when memory
spread is set by updating tasks' mems_allowed after its cpuset's mems is
changed.
In order to update tasks' mems_allowed in time, we must modify the code of
memory policy. Because the memory policy is applied in the process's
context originally. After applying this patch, one task directly
manipulates anothers mems_allowed, and we use alloc_lock in the
task_struct to protect mems_allowed and memory policy of the task.
But in the fast path, we didn't use lock to protect them, because adding a
lock may lead to performance regression. But if we don't add a lock,the
task might see no nodes when changing cpuset's mems_allowed to some
non-overlapping set. In order to avoid it, we set all new allowed nodes,
then clear newly disallowed ones.
[lee.schermerhorn@hp.com:
The rework of mpol_new() to extract the adjusting of the node mask to
apply cpuset and mpol flags "context" breaks set_mempolicy() and mbind()
with MPOL_PREFERRED and a NULL nodemask--i.e., explicit local
allocation. Fix this by adding the check for MPOL_PREFERRED and empty
node mask to mpol_new_mpolicy().
Remove the now unneeded 'nodes = NULL' from mpol_new().
Note that mpol_new_mempolicy() is always called with a non-NULL
'nodes' parameter now that it has been removed from mpol_new().
Therefore, we don't need to test nodes for NULL before testing it for
'empty'. However, just to be extra paranoid, add a VM_BUG_ON() to
verify this assumption.]
[lee.schermerhorn@hp.com:
I don't think the function name 'mpol_new_mempolicy' is descriptive
enough to differentiate it from mpol_new().
This function applies cpuset set context, usually constraining nodes
to those allowed by the cpuset. However, when the 'RELATIVE_NODES flag
is set, it also translates the nodes. So I settled on
'mpol_set_nodemask()', because the comment block for mpol_new() mentions
that we need to call this function to "set nodes".
Some additional minor line length, whitespace and typo cleanup.]
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 15:31:49 -07:00
|
|
|
|
|
|
|
|
|
2008-04-28 02:12:32 -07:00
|
|
|
|
2017-07-06 15:40:06 -07:00
|
|
|
|
2008-04-28 02:12:32 -07:00
|
|
|
|
2017-07-06 15:40:06 -07:00
|
|
|
|
2008-04-28 02:12:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-06-08 21:33:54 -07:00
|
|
|
|
2008-04-28 02:12:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-09-06 19:49:02 +00:00
|
|
|
|
2008-04-28 02:12:32 -07:00
|
|
|
|
2020-06-08 21:33:25 -07:00
|
|
|
|
2023-07-28 06:13:21 +02:00
|
|
|
|
|
|
|
|
|
2017-07-06 15:40:06 -07:00
|
|
|
|
2023-07-28 06:13:21 +02:00
|
|
|
|
2020-06-08 21:33:25 -07:00
|
|
|
|
2008-04-28 02:12:32 -07:00
|
|
|
|
|
|
|
|
|
2008-04-28 02:12:33 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-09-02 15:00:19 -07:00
|
|
|
|
2008-04-28 02:12:33 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-09-02 15:00:19 -07:00
|
|
|
|
2008-04-28 02:12:33 -07:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: don't handle MPOL_LOCAL like a fake MPOL_PREFERRED policy
MPOL_LOCAL policy has been setup as a real policy, but it is still handled
like a faked POL_PREFERRED policy with one internal MPOL_F_LOCAL flag bit
set, and there are many places having to judge the real 'prefer' or the
'local' policy, which are quite confusing.
In current code, there are 4 cases that MPOL_LOCAL are used:
1. user specifies 'local' policy
2. user specifies 'prefer' policy, but with empty nodemask
3. system 'default' policy is used
4. 'prefer' policy + valid 'preferred' node with MPOL_F_STATIC_NODES
flag set, and when it is 'rebind' to a nodemask which doesn't contains
the 'preferred' node, it will perform as 'local' policy
So make 'local' a real policy instead of a fake 'prefer' one, and kill
MPOL_F_LOCAL bit, which can greatly reduce the confusion for code reading.
For case 4, the logic of mpol_rebind_preferred() is confusing, as Michal
Hocko pointed out:
: I do believe that rebinding preferred policy is just bogus and it should
: be dropped altogether on the ground that a preference is a mere hint from
: userspace where to start the allocation. Unless I am missing something
: cpusets will be always authoritative for the final placement. The
: preferred node just acts as a starting point and it should be really
: preserved when cpusets changes. Otherwise we have a very subtle behavior
: corner cases.
So dump all the tricky transformation between 'prefer' and 'local', and
just record the new nodemask of rebinding.
[feng.tang@intel.com: fix a problem in mpol_set_nodemask(), per Michal Hocko]
Link: https://lkml.kernel.org/r/1622560492-1294-3-git-send-email-feng.tang@intel.com
[feng.tang@intel.com: refine code and comments of mpol_set_nodemask(), per Michal]
Link: https://lkml.kernel.org/r/20210603081807.GE56979@shbuild999.sh.intel.com
Link: https://lkml.kernel.org/r/1622469956-82897-3-git-send-email-feng.tang@intel.com
Signed-off-by: Feng Tang <feng.tang@intel.com>
Suggested-by: Michal Hocko <mhocko@suse.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ben Widawsky <ben.widawsky@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-30 18:51:00 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: add MPOL_PREFERRED_MANY for multiple preferred nodes
Patch series "Introduce multi-preference mempolicy", v7.
This patch series introduces the concept of the MPOL_PREFERRED_MANY
mempolicy. This mempolicy mode can be used with either the
set_mempolicy(2) or mbind(2) interfaces. Like the MPOL_PREFERRED
interface, it allows an application to set a preference for nodes which
will fulfil memory allocation requests. Unlike the MPOL_PREFERRED mode,
it takes a set of nodes. Like the MPOL_BIND interface, it works over a
set of nodes. Unlike MPOL_BIND, it will not cause a SIGSEGV or invoke the
OOM killer if those preferred nodes are not available.
Along with these patches are patches for libnuma, numactl, numademo, and
memhog. They still need some polish, but can be found here:
https://gitlab.com/bwidawsk/numactl/-/tree/prefer-many It allows new
usage: `numactl -P 0,3,4`
The goal of the new mode is to enable some use-cases when using tiered memory
usage models which I've lovingly named.
1a. The Hare - The interconnect is fast enough to meet bandwidth and
latency requirements allowing preference to be given to all nodes with
"fast" memory.
1b. The Indiscriminate Hare - An application knows it wants fast
memory (or perhaps slow memory), but doesn't care which node it runs
on. The application can prefer a set of nodes and then xpu bind to
the local node (cpu, accelerator, etc). This reverses the nodes are
chosen today where the kernel attempts to use local memory to the CPU
whenever possible. This will attempt to use the local accelerator to
the memory.
2. The Tortoise - The administrator (or the application itself) is
aware it only needs slow memory, and so can prefer that.
Much of this is almost achievable with the bind interface, but the bind
interface suffers from an inability to fallback to another set of nodes if
binding fails to all nodes in the nodemask.
Like MPOL_BIND a nodemask is given. Inherently this removes ordering from the
preference.
> /* Set first two nodes as preferred in an 8 node system. */
> const unsigned long nodes = 0x3
> set_mempolicy(MPOL_PREFER_MANY, &nodes, 8);
> /* Mimic interleave policy, but have fallback *.
> const unsigned long nodes = 0xaa
> set_mempolicy(MPOL_PREFER_MANY, &nodes, 8);
Some internal discussion took place around the interface. There are two
alternatives which we have discussed, plus one I stuck in:
1. Ordered list of nodes. Currently it's believed that the added
complexity is nod needed for expected usecases.
2. A flag for bind to allow falling back to other nodes. This
confuses the notion of binding and is less flexible than the current
solution.
3. Create flags or new modes that helps with some ordering. This
offers both a friendlier API as well as a solution for more customized
usage. It's unknown if it's worth the complexity to support this.
Here is sample code for how this might work:
> // Prefer specific nodes for some something wacky
> set_mempolicy(MPOL_PREFER_MANY, 0x17c, 1024);
>
> // Default
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_SOCKET, NULL, 0);
> // which is the same as
> set_mempolicy(MPOL_DEFAULT, NULL, 0);
>
> // The Hare
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE, NULL, 0);
>
> // The Tortoise
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE_REV, NULL, 0);
>
> // Prefer the fast memory of the first two sockets
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE, -1, 2);
>
This patch (of 5):
The NUMA APIs currently allow passing in a "preferred node" as a single
bit set in a nodemask. If more than one bit it set, bits after the first
are ignored.
This single node is generally OK for location-based NUMA where memory
being allocated will eventually be operated on by a single CPU. However,
in systems with multiple memory types, folks want to target a *type* of
memory instead of a location. For instance, someone might want some
high-bandwidth memory but do not care about the CPU next to which it is
allocated. Or, they want a cheap, high capacity allocation and want to
target all NUMA nodes which have persistent memory in volatile mode. In
both of these cases, the application wants to target a *set* of nodes, but
does not want strict MPOL_BIND behavior as that could lead to OOM killer
or SIGSEGV.
So add MPOL_PREFERRED_MANY policy to support the multiple preferred nodes
requirement. This is not a pie-in-the-sky dream for an API. This was a
response to a specific ask of more than one group at Intel. Specifically:
1. There are existing libraries that target memory types such as
https://github.com/memkind/memkind. These are known to suffer from
SIGSEGV's when memory is low on targeted memory "kinds" that span more
than one node. The MCDRAM on a Xeon Phi in "Cluster on Die" mode is an
example of this.
2. Volatile-use persistent memory users want to have a memory policy
which is targeted at either "cheap and slow" (PMEM) or "expensive and
fast" (DRAM). However, they do not want to experience allocation
failures when the targeted type is unavailable.
3. Allocate-then-run. Generally, we let the process scheduler decide
on which physical CPU to run a task. That location provides a default
allocation policy, and memory availability is not generally considered
when placing tasks. For situations where memory is valuable and
constrained, some users want to allocate memory first, *then* allocate
close compute resources to the allocation. This is the reverse of the
normal (CPU) model. Accelerators such as GPUs that operate on
core-mm-managed memory are interested in this model.
A check is added in sanitize_mpol_flags() to not permit 'prefer_many'
policy to be used for now, and will be removed in later patch after all
implementations for 'prefer_many' are ready, as suggested by Michal Hocko.
[mhocko@kernel.org: suggest to refine policy_node/policy_nodemask handling]
Link: https://lkml.kernel.org/r/1627970362-61305-1-git-send-email-feng.tang@intel.com
Link: https://lore.kernel.org/r/20200630212517.308045-4-ben.widawsky@intel.com
Link: https://lkml.kernel.org/r/1627970362-61305-2-git-send-email-feng.tang@intel.com
Co-developed-by: Ben Widawsky <ben.widawsky@intel.com>
Signed-off-by: Ben Widawsky <ben.widawsky@intel.com>
Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com>
Signed-off-by: Feng Tang <feng.tang@intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Huang Ying <ying.huang@intel.com>b
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-09-02 15:00:06 -07:00
|
|
|
|
2021-09-02 15:00:19 -07:00
|
|
|
|
mm/mempolicy: add MPOL_PREFERRED_MANY for multiple preferred nodes
Patch series "Introduce multi-preference mempolicy", v7.
This patch series introduces the concept of the MPOL_PREFERRED_MANY
mempolicy. This mempolicy mode can be used with either the
set_mempolicy(2) or mbind(2) interfaces. Like the MPOL_PREFERRED
interface, it allows an application to set a preference for nodes which
will fulfil memory allocation requests. Unlike the MPOL_PREFERRED mode,
it takes a set of nodes. Like the MPOL_BIND interface, it works over a
set of nodes. Unlike MPOL_BIND, it will not cause a SIGSEGV or invoke the
OOM killer if those preferred nodes are not available.
Along with these patches are patches for libnuma, numactl, numademo, and
memhog. They still need some polish, but can be found here:
https://gitlab.com/bwidawsk/numactl/-/tree/prefer-many It allows new
usage: `numactl -P 0,3,4`
The goal of the new mode is to enable some use-cases when using tiered memory
usage models which I've lovingly named.
1a. The Hare - The interconnect is fast enough to meet bandwidth and
latency requirements allowing preference to be given to all nodes with
"fast" memory.
1b. The Indiscriminate Hare - An application knows it wants fast
memory (or perhaps slow memory), but doesn't care which node it runs
on. The application can prefer a set of nodes and then xpu bind to
the local node (cpu, accelerator, etc). This reverses the nodes are
chosen today where the kernel attempts to use local memory to the CPU
whenever possible. This will attempt to use the local accelerator to
the memory.
2. The Tortoise - The administrator (or the application itself) is
aware it only needs slow memory, and so can prefer that.
Much of this is almost achievable with the bind interface, but the bind
interface suffers from an inability to fallback to another set of nodes if
binding fails to all nodes in the nodemask.
Like MPOL_BIND a nodemask is given. Inherently this removes ordering from the
preference.
> /* Set first two nodes as preferred in an 8 node system. */
> const unsigned long nodes = 0x3
> set_mempolicy(MPOL_PREFER_MANY, &nodes, 8);
> /* Mimic interleave policy, but have fallback *.
> const unsigned long nodes = 0xaa
> set_mempolicy(MPOL_PREFER_MANY, &nodes, 8);
Some internal discussion took place around the interface. There are two
alternatives which we have discussed, plus one I stuck in:
1. Ordered list of nodes. Currently it's believed that the added
complexity is nod needed for expected usecases.
2. A flag for bind to allow falling back to other nodes. This
confuses the notion of binding and is less flexible than the current
solution.
3. Create flags or new modes that helps with some ordering. This
offers both a friendlier API as well as a solution for more customized
usage. It's unknown if it's worth the complexity to support this.
Here is sample code for how this might work:
> // Prefer specific nodes for some something wacky
> set_mempolicy(MPOL_PREFER_MANY, 0x17c, 1024);
>
> // Default
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_SOCKET, NULL, 0);
> // which is the same as
> set_mempolicy(MPOL_DEFAULT, NULL, 0);
>
> // The Hare
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE, NULL, 0);
>
> // The Tortoise
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE_REV, NULL, 0);
>
> // Prefer the fast memory of the first two sockets
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE, -1, 2);
>
This patch (of 5):
The NUMA APIs currently allow passing in a "preferred node" as a single
bit set in a nodemask. If more than one bit it set, bits after the first
are ignored.
This single node is generally OK for location-based NUMA where memory
being allocated will eventually be operated on by a single CPU. However,
in systems with multiple memory types, folks want to target a *type* of
memory instead of a location. For instance, someone might want some
high-bandwidth memory but do not care about the CPU next to which it is
allocated. Or, they want a cheap, high capacity allocation and want to
target all NUMA nodes which have persistent memory in volatile mode. In
both of these cases, the application wants to target a *set* of nodes, but
does not want strict MPOL_BIND behavior as that could lead to OOM killer
or SIGSEGV.
So add MPOL_PREFERRED_MANY policy to support the multiple preferred nodes
requirement. This is not a pie-in-the-sky dream for an API. This was a
response to a specific ask of more than one group at Intel. Specifically:
1. There are existing libraries that target memory types such as
https://github.com/memkind/memkind. These are known to suffer from
SIGSEGV's when memory is low on targeted memory "kinds" that span more
than one node. The MCDRAM on a Xeon Phi in "Cluster on Die" mode is an
example of this.
2. Volatile-use persistent memory users want to have a memory policy
which is targeted at either "cheap and slow" (PMEM) or "expensive and
fast" (DRAM). However, they do not want to experience allocation
failures when the targeted type is unavailable.
3. Allocate-then-run. Generally, we let the process scheduler decide
on which physical CPU to run a task. That location provides a default
allocation policy, and memory availability is not generally considered
when placing tasks. For situations where memory is valuable and
constrained, some users want to allocate memory first, *then* allocate
close compute resources to the allocation. This is the reverse of the
normal (CPU) model. Accelerators such as GPUs that operate on
core-mm-managed memory are interested in this model.
A check is added in sanitize_mpol_flags() to not permit 'prefer_many'
policy to be used for now, and will be removed in later patch after all
implementations for 'prefer_many' are ready, as suggested by Michal Hocko.
[mhocko@kernel.org: suggest to refine policy_node/policy_nodemask handling]
Link: https://lkml.kernel.org/r/1627970362-61305-1-git-send-email-feng.tang@intel.com
Link: https://lore.kernel.org/r/20200630212517.308045-4-ben.widawsky@intel.com
Link: https://lkml.kernel.org/r/1627970362-61305-2-git-send-email-feng.tang@intel.com
Co-developed-by: Ben Widawsky <ben.widawsky@intel.com>
Signed-off-by: Ben Widawsky <ben.widawsky@intel.com>
Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com>
Signed-off-by: Feng Tang <feng.tang@intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Huang Ying <ying.huang@intel.com>b
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-09-02 15:00:06 -07:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-04-28 02:12:33 -07:00
|
|
|
|
|
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
2006-01-18 17:42:29 -08:00
|
|
|
|
mempolicy: mmap_lock is not needed while migrating folios
mbind(2) holds down_write of current task's mmap_lock throughout
(exclusive because it needs to set the new mempolicy on the vmas);
migrate_pages(2) holds down_read of pid's mmap_lock throughout.
They both hold mmap_lock across the internal migrate_pages(), under which
all new page allocations (huge or small) are made. I'm nervous about it;
and migrate_pages() certainly does not need mmap_lock itself. It's done
this way for mbind(2), because its page allocator is vma_alloc_folio() or
alloc_hugetlb_folio_vma(), both of which depend on vma and address.
Now that we have alloc_pages_mpol(), depending on (refcounted) memory
policy and interleave index, mbind(2) can be modified to use that or
alloc_hugetlb_folio_nodemask(), and then not need mmap_lock across the
internal migrate_pages() at all: add alloc_migration_target_by_mpol() to
replace mbind's new_page().
(After that change, alloc_hugetlb_folio_vma() is used by nothing but a
userfaultfd function: move it out of hugetlb.h and into the #ifdef.)
migrate_pages(2) has chosen its target node before migrating, so can
continue to use the standard alloc_migration_target(); but let it take and
drop mmap_lock just around migrate_to_node()'s queue_pages_range():
neither the node-to-node calculations nor the page migrations need it.
It seems unlikely, but it is conceivable that some userspace depends on
the kernel's mmap_lock exclusion here, instead of doing its own locking:
more likely in a testsuite than in real life. It is also possible, of
course, that some pages on the list will be munmapped by another thread
before they are migrated, or a newer memory policy applied to the range by
that time: but such races could happen before, as soon as mmap_lock was
dropped, so it does not appear to be a concern.
Link: https://lkml.kernel.org/r/21e564e8-269f-6a89-7ee2-fd612831c289@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:27:47 -07:00
|
|
|
|
|
|
|
|
|
2006-01-08 01:01:02 -08:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-03 02:29:00 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-01-08 01:01:02 -08:00
|
|
|
|
2015-02-11 15:28:03 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy.c: fix checking unmapped holes for mbind
mbind() is required to report EFAULT if range, specified by addr and
len, contains unmapped holes. In current implementation, below rules
are applied for this checking:
1: Unmapped holes at any part of the specified range should be reported
as EFAULT if mbind() for none MPOL_DEFAULT cases;
2: Unmapped holes at any part of the specified range should be ignored
(do not reprot EFAULT) if mbind() for MPOL_DEFAULT case;
3: The whole range in an unmapped hole should be reported as EFAULT;
Note that rule 2 does not fullfill the mbind() API definition, but since
that behavior has existed for long days (the internal flag
MPOL_MF_DISCONTIG_OK is for this purpose), this patch does not plan to
change it.
In current code, application observed inconsistent behavior on rule 1
and rule 2 respectively. That inconsistency is fixed as below details.
Cases of rule 1:
- Hole at head side of range. Current code reprot EFAULT, no change by
this patch.
[ vma ][ hole ][ vma ]
[ range ]
- Hole at middle of range. Current code report EFAULT, no change by
this patch.
[ vma ][ hole ][ vma ]
[ range ]
- Hole at tail side of range. Current code do not report EFAULT, this
patch fixes it.
[ vma ][ hole ][ vma ]
[ range ]
Cases of rule 2:
- Hole at head side of range. Current code reports EFAULT, this patch
fixes it.
[ vma ][ hole ][ vma ]
[ range ]
- Hole at middle of range. Current code does not report EFAULT, no
change by this patch.
[ vma ][ hole ][ vma]
[ range ]
- Hole at tail side of range. Current code does not report EFAULT, no
change by this patch.
[ vma ][ hole ][ vma]
[ range ]
This patch has no changes to rule 3.
The unmapped hole checking can also be handled by using .pte_hole(),
instead of .test_walk(). But .pte_hole() is called for holes inside and
outside vma, which causes more cost, so this patch keeps the original
design with .test_walk().
Link: http://lkml.kernel.org/r/1573218104-11021-3-git-send-email-lixinhai.lxh@gmail.com
Fixes: 6f4576e3687b ("mempolicy: apply page table walker on queue_pages_range()")
Signed-off-by: Li Xinhai <lixinhai.lxh@gmail.com>
Reviewed-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Hugh Dickins <hughd@google.com>
Cc: linux-man <linux-man@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-11-30 17:56:18 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
|
|
|
|
|
2015-02-11 15:28:03 -08:00
|
|
|
|
|
|
|
|
|
mm: mempolicy: add queue_pages_required()
Patch series "mm: page migration enhancement for thp", v9.
Motivations:
1. THP migration becomes important in the upcoming heterogeneous memory
systems. As David Nellans from NVIDIA pointed out from other threads
(http://www.mail-archive.com/linux-kernel@vger.kernel.org/msg1349227.html),
future GPUs or other accelerators will have their memory managed by
operating systems. Moving data into and out of these memory nodes
efficiently is critical to applications that use GPUs or other
accelerators. Existing page migration only supports base pages, which
has a very low memory bandwidth utilization. My experiments (see
below) show THP migration can migrate pages more efficiently.
2. Base page migration vs THP migration throughput.
Here are cross-socket page migration results from calling
move_pages() syscall:
In x86_64, a Intel two-socket E5-2640v3 box,
- single 4KB base page migration takes 62.47 us, using 0.06 GB/s BW,
- single 2MB THP migration takes 658.54 us, using 2.97 GB/s BW,
- 512 4KB base page migration takes 1987.38 us, using 0.98 GB/s BW.
In ppc64, a two-socket Power8 box,
- single 64KB base page migration takes 49.3 us, using 1.24 GB/s BW,
- single 16MB THP migration takes 2202.17 us, using 7.10 GB/s BW,
- 256 64KB base page migration takes 2543.65 us, using 6.14 GB/s BW.
THP migration can give us 3x and 1.15x throughput over base page
migration in x86_64 and ppc64 respectivley.
You can test it out by using the code here:
https://github.com/x-y-z/thp-migration-bench
3. Existing page migration splits THP before migration and cannot
guarantee the migrated pages are still contiguous. Contiguity is
always what GPUs and accelerators look for. Without THP migration,
khugepaged needs to do extra work to reassemble the migrated pages
back to THPs.
This patch (of 10):
Introduce a separate check routine related to MPOL_MF_INVERT flag. This
patch just does cleanup, no behavioral change.
Link: http://lkml.kernel.org/r/20170717193955.20207-2-zi.yan@sent.com
Signed-off-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Signed-off-by: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Anshuman Khandual <khandual@linux.vnet.ibm.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Nellans <dnellans@nvidia.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Andrea Arcangeli <aarcange@redhat.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>
2017-09-08 16:10:42 -07:00
|
|
|
|
2023-01-30 12:18:32 -08:00
|
|
|
|
mm: mempolicy: add queue_pages_required()
Patch series "mm: page migration enhancement for thp", v9.
Motivations:
1. THP migration becomes important in the upcoming heterogeneous memory
systems. As David Nellans from NVIDIA pointed out from other threads
(http://www.mail-archive.com/linux-kernel@vger.kernel.org/msg1349227.html),
future GPUs or other accelerators will have their memory managed by
operating systems. Moving data into and out of these memory nodes
efficiently is critical to applications that use GPUs or other
accelerators. Existing page migration only supports base pages, which
has a very low memory bandwidth utilization. My experiments (see
below) show THP migration can migrate pages more efficiently.
2. Base page migration vs THP migration throughput.
Here are cross-socket page migration results from calling
move_pages() syscall:
In x86_64, a Intel two-socket E5-2640v3 box,
- single 4KB base page migration takes 62.47 us, using 0.06 GB/s BW,
- single 2MB THP migration takes 658.54 us, using 2.97 GB/s BW,
- 512 4KB base page migration takes 1987.38 us, using 0.98 GB/s BW.
In ppc64, a two-socket Power8 box,
- single 64KB base page migration takes 49.3 us, using 1.24 GB/s BW,
- single 16MB THP migration takes 2202.17 us, using 7.10 GB/s BW,
- 256 64KB base page migration takes 2543.65 us, using 6.14 GB/s BW.
THP migration can give us 3x and 1.15x throughput over base page
migration in x86_64 and ppc64 respectivley.
You can test it out by using the code here:
https://github.com/x-y-z/thp-migration-bench
3. Existing page migration splits THP before migration and cannot
guarantee the migrated pages are still contiguous. Contiguity is
always what GPUs and accelerators look for. Without THP migration,
khugepaged needs to do extra work to reassemble the migrated pages
back to THPs.
This patch (of 10):
Introduce a separate check routine related to MPOL_MF_INVERT flag. This
patch just does cleanup, no behavioral change.
Link: http://lkml.kernel.org/r/20170717193955.20207-2-zi.yan@sent.com
Signed-off-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Signed-off-by: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Anshuman Khandual <khandual@linux.vnet.ibm.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Nellans <dnellans@nvidia.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Andrea Arcangeli <aarcange@redhat.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>
2017-09-08 16:10:42 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-01-30 12:18:32 -08:00
|
|
|
|
mm: mempolicy: add queue_pages_required()
Patch series "mm: page migration enhancement for thp", v9.
Motivations:
1. THP migration becomes important in the upcoming heterogeneous memory
systems. As David Nellans from NVIDIA pointed out from other threads
(http://www.mail-archive.com/linux-kernel@vger.kernel.org/msg1349227.html),
future GPUs or other accelerators will have their memory managed by
operating systems. Moving data into and out of these memory nodes
efficiently is critical to applications that use GPUs or other
accelerators. Existing page migration only supports base pages, which
has a very low memory bandwidth utilization. My experiments (see
below) show THP migration can migrate pages more efficiently.
2. Base page migration vs THP migration throughput.
Here are cross-socket page migration results from calling
move_pages() syscall:
In x86_64, a Intel two-socket E5-2640v3 box,
- single 4KB base page migration takes 62.47 us, using 0.06 GB/s BW,
- single 2MB THP migration takes 658.54 us, using 2.97 GB/s BW,
- 512 4KB base page migration takes 1987.38 us, using 0.98 GB/s BW.
In ppc64, a two-socket Power8 box,
- single 64KB base page migration takes 49.3 us, using 1.24 GB/s BW,
- single 16MB THP migration takes 2202.17 us, using 7.10 GB/s BW,
- 256 64KB base page migration takes 2543.65 us, using 6.14 GB/s BW.
THP migration can give us 3x and 1.15x throughput over base page
migration in x86_64 and ppc64 respectivley.
You can test it out by using the code here:
https://github.com/x-y-z/thp-migration-bench
3. Existing page migration splits THP before migration and cannot
guarantee the migrated pages are still contiguous. Contiguity is
always what GPUs and accelerators look for. Without THP migration,
khugepaged needs to do extra work to reassemble the migrated pages
back to THPs.
This patch (of 10):
Introduce a separate check routine related to MPOL_MF_INVERT flag. This
patch just does cleanup, no behavioral change.
Link: http://lkml.kernel.org/r/20170717193955.20207-2-zi.yan@sent.com
Signed-off-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Signed-off-by: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Anshuman Khandual <khandual@linux.vnet.ibm.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Nellans <dnellans@nvidia.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Andrea Arcangeli <aarcange@redhat.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>
2017-09-08 16:10:42 -07:00
|
|
|
|
|
|
|
|
|
2023-01-30 12:18:32 -08:00
|
|
|
|
mm: mempolicy: add queue_pages_required()
Patch series "mm: page migration enhancement for thp", v9.
Motivations:
1. THP migration becomes important in the upcoming heterogeneous memory
systems. As David Nellans from NVIDIA pointed out from other threads
(http://www.mail-archive.com/linux-kernel@vger.kernel.org/msg1349227.html),
future GPUs or other accelerators will have their memory managed by
operating systems. Moving data into and out of these memory nodes
efficiently is critical to applications that use GPUs or other
accelerators. Existing page migration only supports base pages, which
has a very low memory bandwidth utilization. My experiments (see
below) show THP migration can migrate pages more efficiently.
2. Base page migration vs THP migration throughput.
Here are cross-socket page migration results from calling
move_pages() syscall:
In x86_64, a Intel two-socket E5-2640v3 box,
- single 4KB base page migration takes 62.47 us, using 0.06 GB/s BW,
- single 2MB THP migration takes 658.54 us, using 2.97 GB/s BW,
- 512 4KB base page migration takes 1987.38 us, using 0.98 GB/s BW.
In ppc64, a two-socket Power8 box,
- single 64KB base page migration takes 49.3 us, using 1.24 GB/s BW,
- single 16MB THP migration takes 2202.17 us, using 7.10 GB/s BW,
- 256 64KB base page migration takes 2543.65 us, using 6.14 GB/s BW.
THP migration can give us 3x and 1.15x throughput over base page
migration in x86_64 and ppc64 respectivley.
You can test it out by using the code here:
https://github.com/x-y-z/thp-migration-bench
3. Existing page migration splits THP before migration and cannot
guarantee the migrated pages are still contiguous. Contiguity is
always what GPUs and accelerators look for. Without THP migration,
khugepaged needs to do extra work to reassemble the migrated pages
back to THPs.
This patch (of 10):
Introduce a separate check routine related to MPOL_MF_INVERT flag. This
patch just does cleanup, no behavioral change.
Link: http://lkml.kernel.org/r/20170717193955.20207-2-zi.yan@sent.com
Signed-off-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Signed-off-by: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Anshuman Khandual <khandual@linux.vnet.ibm.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Nellans <dnellans@nvidia.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Andrea Arcangeli <aarcange@redhat.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>
2017-09-08 16:10:42 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
2017-09-08 16:11:08 -07:00
|
|
|
|
2023-01-30 12:18:29 -08:00
|
|
|
|
2017-09-08 16:11:08 -07:00
|
|
|
|
|
|
|
|
|
2025-11-10 22:21:28 +00:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
|
|
|
|
|
2017-09-08 16:11:08 -07:00
|
|
|
|
2024-03-26 20:28:23 +00:00
|
|
|
|
2024-03-26 20:28:22 +00:00
|
|
|
|
2021-06-30 18:51:07 -07:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
2017-09-08 16:11:08 -07:00
|
|
|
|
2023-01-30 12:18:32 -08:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-09-08 16:11:08 -07:00
|
|
|
|
|
|
|
|
|
2013-09-11 14:22:14 -07:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
|
|
|
|
|
mm: mempolicy: make the behavior consistent when MPOL_MF_MOVE* and MPOL_MF_STRICT were specified
When both MPOL_MF_MOVE* and MPOL_MF_STRICT was specified, mbind() should
try best to migrate misplaced pages, if some of the pages could not be
migrated, then return -EIO.
There are three different sub-cases:
1. vma is not migratable
2. vma is migratable, but there are unmovable pages
3. vma is migratable, pages are movable, but migrate_pages() fails
If #1 happens, kernel would just abort immediately, then return -EIO,
after a7f40cfe3b7a ("mm: mempolicy: make mbind() return -EIO when
MPOL_MF_STRICT is specified").
If #3 happens, kernel would set policy and migrate pages with
best-effort, but won't rollback the migrated pages and reset the policy
back.
Before that commit, they behaves in the same way. It'd better to keep
their behavior consistent. But, rolling back the migrated pages and
resetting the policy back sounds not feasible, so just make #1 behave as
same as #3.
Userspace will know that not everything was successfully migrated (via
-EIO), and can take whatever steps it deems necessary - attempt
rollback, determine which exact page(s) are violating the policy, etc.
Make queue_pages_range() return 1 to indicate there are unmovable pages
or vma is not migratable.
The #2 is not handled correctly in the current kernel, the following
patch will fix it.
[yang.shi@linux.alibaba.com: fix review comments from Vlastimil]
Link: http://lkml.kernel.org/r/1563556862-54056-2-git-send-email-yang.shi@linux.alibaba.com
Link: http://lkml.kernel.org/r/1561162809-59140-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
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 15:37:15 -07:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-09-11 14:22:14 -07:00
|
|
|
|
2023-01-30 12:18:30 -08:00
|
|
|
|
2015-02-11 15:28:03 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2015-02-11 15:28:03 -08:00
|
|
|
|
2023-01-30 12:18:30 -08:00
|
|
|
|
2015-02-11 15:28:03 -08:00
|
|
|
|
|
|
|
|
|
2020-11-01 17:07:40 -08:00
|
|
|
|
mm: ptep_get() conversion
Convert all instances of direct pte_t* dereferencing to instead use
ptep_get() helper. This means that by default, the accesses change from a
C dereference to a READ_ONCE(). This is technically the correct thing to
do since where pgtables are modified by HW (for access/dirty) they are
volatile and therefore we should always ensure READ_ONCE() semantics.
But more importantly, by always using the helper, it can be overridden by
the architecture to fully encapsulate the contents of the pte. Arch code
is deliberately not converted, as the arch code knows best. It is
intended that arch code (arm64) will override the default with its own
implementation that can (e.g.) hide certain bits from the core code, or
determine young/dirty status by mixing in state from another source.
Conversion was done using Coccinelle:
----
// $ make coccicheck \
// COCCI=ptepget.cocci \
// SPFLAGS="--include-headers" \
// MODE=patch
virtual patch
@ depends on patch @
pte_t *v;
@@
- *v
+ ptep_get(v)
----
Then reviewed and hand-edited to avoid multiple unnecessary calls to
ptep_get(), instead opting to store the result of a single call in a
variable, where it is correct to do so. This aims to negate any cost of
READ_ONCE() and will benefit arch-overrides that may be more complex.
Included is a fix for an issue in an earlier version of this patch that
was pointed out by kernel test robot. The issue arose because config
MMU=n elides definition of the ptep helper functions, including
ptep_get(). HUGETLB_PAGE=n configs still define a simple
huge_ptep_clear_flush() for linking purposes, which dereferences the ptep.
So when both configs are disabled, this caused a build error because
ptep_get() is not defined. Fix by continuing to do a direct dereference
when MMU=n. This is safe because for this config the arch code cannot be
trying to virtualize the ptes because none of the ptep helpers are
defined.
Link: https://lkml.kernel.org/r/20230612151545.3317766-4-ryan.roberts@arm.com
Reported-by: kernel test robot <lkp@intel.com>
Link: https://lore.kernel.org/oe-kbuild-all/202305120142.yXsNEo6H-lkp@intel.com/
Signed-off-by: Ryan Roberts <ryan.roberts@arm.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Alex Williamson <alex.williamson@redhat.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Andrey Konovalov <andreyknvl@gmail.com>
Cc: Andrey Ryabinin <ryabinin.a.a@gmail.com>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Daniel Vetter <daniel@ffwll.ch>
Cc: Dave Airlie <airlied@gmail.com>
Cc: Dimitri Sivanich <dimitri.sivanich@hpe.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Ian Rogers <irogers@google.com>
Cc: Jason Gunthorpe <jgg@ziepe.ca>
Cc: Jérôme Glisse <jglisse@redhat.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Lorenzo Stoakes <lstoakes@gmail.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mike Rapoport (IBM) <rppt@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Naoya Horiguchi <naoya.horiguchi@nec.com>
Cc: Oleksandr Tyshchenko <oleksandr_tyshchenko@epam.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: SeongJae Park <sj@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Uladzislau Rezki (Sony) <urezki@gmail.com>
Cc: Vincenzo Frascino <vincenzo.frascino@arm.com>
Cc: Yu Zhao <yuzhao@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-06-12 16:15:45 +01:00
|
|
|
|
2005-10-29 18:16:27 -07:00
|
|
|
|
2025-04-16 11:00:48 +05:30
|
|
|
|
2005-06-21 17:15:06 -07:00
|
|
|
|
2017-09-08 16:11:08 -07:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-06-21 17:15:07 -07:00
|
|
|
|
2020-11-01 17:07:40 -08:00
|
|
|
|
2023-06-08 18:17:26 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-04-16 11:00:48 +05:30
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: ptep_get() conversion
Convert all instances of direct pte_t* dereferencing to instead use
ptep_get() helper. This means that by default, the accesses change from a
C dereference to a READ_ONCE(). This is technically the correct thing to
do since where pgtables are modified by HW (for access/dirty) they are
volatile and therefore we should always ensure READ_ONCE() semantics.
But more importantly, by always using the helper, it can be overridden by
the architecture to fully encapsulate the contents of the pte. Arch code
is deliberately not converted, as the arch code knows best. It is
intended that arch code (arm64) will override the default with its own
implementation that can (e.g.) hide certain bits from the core code, or
determine young/dirty status by mixing in state from another source.
Conversion was done using Coccinelle:
----
// $ make coccicheck \
// COCCI=ptepget.cocci \
// SPFLAGS="--include-headers" \
// MODE=patch
virtual patch
@ depends on patch @
pte_t *v;
@@
- *v
+ ptep_get(v)
----
Then reviewed and hand-edited to avoid multiple unnecessary calls to
ptep_get(), instead opting to store the result of a single call in a
variable, where it is correct to do so. This aims to negate any cost of
READ_ONCE() and will benefit arch-overrides that may be more complex.
Included is a fix for an issue in an earlier version of this patch that
was pointed out by kernel test robot. The issue arose because config
MMU=n elides definition of the ptep helper functions, including
ptep_get(). HUGETLB_PAGE=n configs still define a simple
huge_ptep_clear_flush() for linking purposes, which dereferences the ptep.
So when both configs are disabled, this caused a build error because
ptep_get() is not defined. Fix by continuing to do a direct dereference
when MMU=n. This is safe because for this config the arch code cannot be
trying to virtualize the ptes because none of the ptep helpers are
defined.
Link: https://lkml.kernel.org/r/20230612151545.3317766-4-ryan.roberts@arm.com
Reported-by: kernel test robot <lkp@intel.com>
Link: https://lore.kernel.org/oe-kbuild-all/202305120142.yXsNEo6H-lkp@intel.com/
Signed-off-by: Ryan Roberts <ryan.roberts@arm.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Alex Williamson <alex.williamson@redhat.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Andrey Konovalov <andreyknvl@gmail.com>
Cc: Andrey Ryabinin <ryabinin.a.a@gmail.com>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Daniel Vetter <daniel@ffwll.ch>
Cc: Dave Airlie <airlied@gmail.com>
Cc: Dimitri Sivanich <dimitri.sivanich@hpe.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Ian Rogers <irogers@google.com>
Cc: Jason Gunthorpe <jgg@ziepe.ca>
Cc: Jérôme Glisse <jglisse@redhat.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Lorenzo Stoakes <lstoakes@gmail.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mike Rapoport (IBM) <rppt@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Naoya Horiguchi <naoya.horiguchi@nec.com>
Cc: Oleksandr Tyshchenko <oleksandr_tyshchenko@epam.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: SeongJae Park <sj@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Uladzislau Rezki (Sony) <urezki@gmail.com>
Cc: Vincenzo Frascino <vincenzo.frascino@arm.com>
Cc: Yu Zhao <yuzhao@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-06-12 16:15:45 +01:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
2025-11-10 22:21:33 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: ptep_get() conversion
Convert all instances of direct pte_t* dereferencing to instead use
ptep_get() helper. This means that by default, the accesses change from a
C dereference to a READ_ONCE(). This is technically the correct thing to
do since where pgtables are modified by HW (for access/dirty) they are
volatile and therefore we should always ensure READ_ONCE() semantics.
But more importantly, by always using the helper, it can be overridden by
the architecture to fully encapsulate the contents of the pte. Arch code
is deliberately not converted, as the arch code knows best. It is
intended that arch code (arm64) will override the default with its own
implementation that can (e.g.) hide certain bits from the core code, or
determine young/dirty status by mixing in state from another source.
Conversion was done using Coccinelle:
----
// $ make coccicheck \
// COCCI=ptepget.cocci \
// SPFLAGS="--include-headers" \
// MODE=patch
virtual patch
@ depends on patch @
pte_t *v;
@@
- *v
+ ptep_get(v)
----
Then reviewed and hand-edited to avoid multiple unnecessary calls to
ptep_get(), instead opting to store the result of a single call in a
variable, where it is correct to do so. This aims to negate any cost of
READ_ONCE() and will benefit arch-overrides that may be more complex.
Included is a fix for an issue in an earlier version of this patch that
was pointed out by kernel test robot. The issue arose because config
MMU=n elides definition of the ptep helper functions, including
ptep_get(). HUGETLB_PAGE=n configs still define a simple
huge_ptep_clear_flush() for linking purposes, which dereferences the ptep.
So when both configs are disabled, this caused a build error because
ptep_get() is not defined. Fix by continuing to do a direct dereference
when MMU=n. This is safe because for this config the arch code cannot be
trying to virtualize the ptes because none of the ptep helpers are
defined.
Link: https://lkml.kernel.org/r/20230612151545.3317766-4-ryan.roberts@arm.com
Reported-by: kernel test robot <lkp@intel.com>
Link: https://lore.kernel.org/oe-kbuild-all/202305120142.yXsNEo6H-lkp@intel.com/
Signed-off-by: Ryan Roberts <ryan.roberts@arm.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Alex Williamson <alex.williamson@redhat.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Andrey Konovalov <andreyknvl@gmail.com>
Cc: Andrey Ryabinin <ryabinin.a.a@gmail.com>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Daniel Vetter <daniel@ffwll.ch>
Cc: Dave Airlie <airlied@gmail.com>
Cc: Dimitri Sivanich <dimitri.sivanich@hpe.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Ian Rogers <irogers@google.com>
Cc: Jason Gunthorpe <jgg@ziepe.ca>
Cc: Jérôme Glisse <jglisse@redhat.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Lorenzo Stoakes <lstoakes@gmail.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mike Rapoport (IBM) <rppt@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Naoya Horiguchi <naoya.horiguchi@nec.com>
Cc: Oleksandr Tyshchenko <oleksandr_tyshchenko@epam.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: SeongJae Park <sj@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Uladzislau Rezki (Sony) <urezki@gmail.com>
Cc: Vincenzo Frascino <vincenzo.frascino@arm.com>
Cc: Yu Zhao <yuzhao@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-06-12 16:15:45 +01:00
|
|
|
|
2023-01-30 12:18:30 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2025-04-16 11:00:48 +05:30
|
|
|
|
mm: split folio_pte_batch() into folio_pte_batch() and folio_pte_batch_flags()
Many users (including upcoming ones) don't really need the flags etc, and
can live with the possible overhead of a function call.
So let's provide a basic, non-inlined folio_pte_batch(), to avoid code
bloat while still providing a variant that optimizes out all flag checks
at runtime. folio_pte_batch_flags() will get inlined into
folio_pte_batch(), optimizing out any conditionals that depend on input
flags.
folio_pte_batch() will behave like folio_pte_batch_flags() when no flags
are specified. It's okay to add new users of folio_pte_batch_flags(), but
using folio_pte_batch() if applicable is preferred.
So, before this change, folio_pte_batch() was inlined into the C file
optimized by propagating constants within the resulting object file.
With this change, we now also have a folio_pte_batch() that is optimized
by propagating all constants. But instead of having one instance per
object file, we have a single shared one.
In zap_present_ptes(), where we care about performance, the compiler
already seem to generate a call to a common inlined folio_pte_batch()
variant, shared with fork() code. So calling the new non-inlined variant
should not make a difference.
While at it, drop the "addr" parameter that is unused.
Link: https://lkml.kernel.org/r/20250702104926.212243-4-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Suggested-by: Andrew Morton <akpm@linux-foundation.org>
Link: https://lore.kernel.org/linux-mm/20250503182858.5a02729fcffd6d4723afcfc2@linux-foundation.org/
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Dev Jain <dev.jain@arm.com>
Cc: Alistair Popple <apopple@nvidia.com>
Cc: Byungchul Park <byungchul@sk.com>
Cc: Gregory Price <gourry@gourry.net>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Jann Horn <jannh@google.com>
Cc: Joshua Hahn <joshua.hahnjy@gmail.com>
Cc: Lance Yang <lance.yang@linux.dev>
Cc: Liam Howlett <liam.howlett@oracle.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Mathew Brost <matthew.brost@intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Rakie Kim <rakie.kim@sk.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-07-02 12:49:25 +02:00
|
|
|
|
2006-01-18 17:42:27 -08:00
|
|
|
|
2023-01-30 12:18:30 -08:00
|
|
|
|
|
|
|
|
|
2006-01-18 17:42:27 -08:00
|
|
|
|
2023-01-30 12:18:30 -08:00
|
|
|
|
2006-01-12 01:05:20 -08:00
|
|
|
|
2023-01-30 12:18:32 -08:00
|
|
|
|
2006-01-08 01:01:01 -08:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
2023-09-20 15:32:42 -07:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-09-20 15:32:42 -07:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-04-16 11:00:48 +05:30
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-02-11 15:28:03 -08:00
|
|
|
|
2020-11-01 17:07:40 -08:00
|
|
|
|
2015-02-11 15:28:03 -08:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-06-21 17:15:07 -07:00
|
|
|
|
|
|
|
|
|
2023-01-30 12:18:31 -08:00
|
|
|
|
2015-02-11 15:28:03 -08:00
|
|
|
|
|
|
|
|
|
2013-09-11 14:22:03 -07:00
|
|
|
|
|
|
|
|
|
2015-02-11 15:28:03 -08:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
2023-01-30 12:18:31 -08:00
|
|
|
|
2013-11-14 14:31:02 -08:00
|
|
|
|
2025-11-10 22:21:32 +00:00
|
|
|
|
2013-09-11 14:22:03 -07:00
|
|
|
|
2015-02-11 15:28:03 -08:00
|
|
|
|
2025-11-10 22:21:32 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-06-06 10:00:01 -04:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
2025-11-10 22:21:32 +00:00
|
|
|
|
2023-01-30 12:18:32 -08:00
|
|
|
|
2013-09-11 14:22:03 -07:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: support MPOL_MF_STRICT for huge page mapping
MPOL_MF_STRICT is used in mbind() for purposes:
(1) MPOL_MF_STRICT is set alone without MPOL_MF_MOVE or
MPOL_MF_MOVE_ALL, to check if there is misplaced page and return -EIO;
(2) MPOL_MF_STRICT is set with MPOL_MF_MOVE or MPOL_MF_MOVE_ALL, to
check if there is misplaced page which is failed to isolate, or page
is success on isolate but failed to move, and return -EIO.
For non hugepage mapping, (1) and (2) are implemented as expectation. For
hugepage mapping, (1) is not implemented. And in (2), the part about
failed to isolate and report -EIO is not implemented.
This patch implements the missed parts for hugepage mapping. Benefits
with it applied:
- User space can apply same code logic to handle mbind() on hugepage and
non hugepage mapping;
- Reliably using MPOL_MF_STRICT alone to check whether there is
misplaced page or not when bind policy on address range, especially for
address range which contains both hugepage and non hugepage mapping.
Analysis of potential impact to existing users:
- If MPOL_MF_STRICT alone was previously used, hugetlb pages not
following the memory policy would not cause an EIO error. After this
change, hugetlb pages are treated like all other pages. If
MPOL_MF_STRICT alone is used and hugetlb pages do not follow memory
policy an EIO error will be returned.
- For users who using MPOL_MF_STRICT with MPOL_MF_MOVE or
MPOL_MF_MOVE_ALL, the semantic about some pages could not be moved will
not be changed by this patch, because failed to isolate and failed to
move have same effects to users, so their existing code will not be
impacted.
In mbind man page, the note about 'MPOL_MF_STRICT is ignored on huge page
mappings' can be removed after this patch is applied.
Mike:
: The current behavior with MPOL_MF_STRICT and hugetlb pages is inconsistent
: and does not match documentation (as described above). The special
: behavior for hugetlb pages ideally should have been removed when hugetlb
: page migration was introduced. It is unlikely that anyone relies on
: today's inconsistent behavior, and removing one more case of special
: handling for hugetlb pages is a good thing.
Signed-off-by: Li Xinhai <lixinhai.lxh@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Reviewed-by: Naoya Horiguchi <naoya.horiguchi@nec.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: linux-man <linux-man@vger.kernel.org>
Link: http://lkml.kernel.org/r/1581559627-6206-1-git-send-email-lixinhai.lxh@gmail.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-01 21:10:48 -07:00
|
|
|
|
|
|
|
|
|
2023-01-30 12:18:31 -08:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
|
|
|
|
|
2023-01-30 12:18:31 -08:00
|
|
|
|
mm: convert folio_likely_mapped_shared() to folio_maybe_mapped_shared()
Let's reuse our new MM ownership tracking infrastructure for large folios
to make folio_likely_mapped_shared() never return false negatives -- never
indicating "not mapped shared" although the folio *is* mapped shared.
With that, we can rename it to folio_maybe_mapped_shared() and get rid of
the dependency on the mapcount of the first folio page.
The semantics are now arguably clearer: no mixture of "false negatives"
and "false positives", only the remaining possibility for "false
positives".
Thoroughly document the new semantics. We might now detect that a large
folio is "maybe mapped shared" although it *no longer* is -- but once was.
Now, if more than two MMs mapped a folio at the same time, and the MM
mapping the folio exclusively at the end is not one tracked in the two
folio MM slots, we will detect the folio as "maybe mapped shared".
For anonymous folios, usually (except weird corner cases) all PTEs that
target a "maybe mapped shared" folio are R/O. As soon as a child process
would write to them (iow, actively use them), we would CoW and effectively
replace these PTEs. Most cases (below) are not expected to really matter
with large anonymous folios for this reason.
Most importantly, there will be no change at all for:
* small folios
* hugetlb folios
* PMD-mapped PMD-sized THPs (single mapping)
This change has the potential to affect existing callers of
folio_likely_mapped_shared() -> folio_maybe_mapped_shared():
(1) fs/proc/task_mmu.c: no change (hugetlb)
(2) khugepaged counts PTEs that target shared folios towards
max_ptes_shared (default: HPAGE_PMD_NR / 2), meaning we could skip a
collapse where we would have previously collapsed. This only applies
to anonymous folios and is not expected to matter in practice.
Worth noting that this change sorts out case (A) documented in
commit 1bafe96e89f0 ("mm/khugepaged: replace page_mapcount() check by
folio_likely_mapped_shared()") by removing the possibility for "false
negatives".
(3) MADV_COLD / MADV_PAGEOUT / MADV_FREE will not try splitting
PTE-mapped THPs that are considered shared but not fully covered by
the requested range, consequently not processing them.
PMD-mapped PMD-sized THP are not affected, or when all PTEs are
covered. These functions are usually only called on anon/file folios
that are exclusively mapped most of the time (no other file mappings
or no fork()), so the "false negatives" are not expected to matter in
practice.
(4) mbind() / migrate_pages() / move_pages() will refuse to migrate
shared folios unless MPOL_MF_MOVE_ALL is effective (requires
CAP_SYS_NICE). We will now reject some folios that could be migrated.
Similar to (3), especially with MPOL_MF_MOVE_ALL, so this is not
expected to matter in practice.
Note that cpuset_migrate_mm_workfn() calls do_migrate_pages() with
MPOL_MF_MOVE_ALL.
(5) NUMA hinting
mm/migrate.c:migrate_misplaced_folio_prepare() will skip file
folios that are probably shared libraries (-> "mapped shared" and
executable). This check would have detected it as a shared library at
some point (at least 3 MMs mapping it), so detecting it afterwards
does not sound wrong (still a shared library). Not expected to
matter.
mm/memory.c:numa_migrate_check() will indicate TNF_SHARED in
MAP_SHARED file mappings when encountering a shared folio. Similar
reasoning, not expected to matter.
mm/mprotect.c:change_pte_range() will skip folios detected as
shared in CoW mappings. Similarly, this is not expected to matter in
practice, but if it would ever be a problem we could relax that check
a bit (e.g., basing it on the average page-mapcount in a folio),
because it was only an optimization when many (e.g., 288) processes
were mapping the same folios -- see commit 859d4adc3415 ("mm: numa: do
not trap faults on shared data section pages.")
(6) mm/rmap.c:folio_referenced_one() will skip exclusive swapbacked
folios in dying processes. Applies to anonymous folios only. Without
"false negatives", we'll now skip all actually shared ones. Skipping
ones that are actually exclusive won't really matter, it's a pure
optimization, and is not expected to matter in practice.
In theory, one can detect the problematic scenario: folio_mapcount() > 0
and no folio MM slot is occupied ("state unknown"). One could reset the
MM slots while doing an rmap walk, which migration / folio split already
do when setting everything up. Further, when batching PTEs we might
naturally learn about a owner (e.g., folio_mapcount() == nr_ptes) and
could update the owner. However, we'll defer that until the scenarios
where it would really matter are clear.
Link: https://lkml.kernel.org/r/20250303163014.1128035-15-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Cc: Andy Lutomirks^H^Hski <luto@kernel.org>
Cc: Borislav Betkov <bp@alien8.de>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jann Horn <jannh@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: Liam Howlett <liam.howlett@oracle.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Michal Koutn <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: tejun heo <tj@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-03-03 17:30:07 +01:00
|
|
|
|
2024-02-27 21:15:48 +01:00
|
|
|
|
2023-01-30 12:18:31 -08:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
mm: convert folio_likely_mapped_shared() to folio_maybe_mapped_shared()
Let's reuse our new MM ownership tracking infrastructure for large folios
to make folio_likely_mapped_shared() never return false negatives -- never
indicating "not mapped shared" although the folio *is* mapped shared.
With that, we can rename it to folio_maybe_mapped_shared() and get rid of
the dependency on the mapcount of the first folio page.
The semantics are now arguably clearer: no mixture of "false negatives"
and "false positives", only the remaining possibility for "false
positives".
Thoroughly document the new semantics. We might now detect that a large
folio is "maybe mapped shared" although it *no longer* is -- but once was.
Now, if more than two MMs mapped a folio at the same time, and the MM
mapping the folio exclusively at the end is not one tracked in the two
folio MM slots, we will detect the folio as "maybe mapped shared".
For anonymous folios, usually (except weird corner cases) all PTEs that
target a "maybe mapped shared" folio are R/O. As soon as a child process
would write to them (iow, actively use them), we would CoW and effectively
replace these PTEs. Most cases (below) are not expected to really matter
with large anonymous folios for this reason.
Most importantly, there will be no change at all for:
* small folios
* hugetlb folios
* PMD-mapped PMD-sized THPs (single mapping)
This change has the potential to affect existing callers of
folio_likely_mapped_shared() -> folio_maybe_mapped_shared():
(1) fs/proc/task_mmu.c: no change (hugetlb)
(2) khugepaged counts PTEs that target shared folios towards
max_ptes_shared (default: HPAGE_PMD_NR / 2), meaning we could skip a
collapse where we would have previously collapsed. This only applies
to anonymous folios and is not expected to matter in practice.
Worth noting that this change sorts out case (A) documented in
commit 1bafe96e89f0 ("mm/khugepaged: replace page_mapcount() check by
folio_likely_mapped_shared()") by removing the possibility for "false
negatives".
(3) MADV_COLD / MADV_PAGEOUT / MADV_FREE will not try splitting
PTE-mapped THPs that are considered shared but not fully covered by
the requested range, consequently not processing them.
PMD-mapped PMD-sized THP are not affected, or when all PTEs are
covered. These functions are usually only called on anon/file folios
that are exclusively mapped most of the time (no other file mappings
or no fork()), so the "false negatives" are not expected to matter in
practice.
(4) mbind() / migrate_pages() / move_pages() will refuse to migrate
shared folios unless MPOL_MF_MOVE_ALL is effective (requires
CAP_SYS_NICE). We will now reject some folios that could be migrated.
Similar to (3), especially with MPOL_MF_MOVE_ALL, so this is not
expected to matter in practice.
Note that cpuset_migrate_mm_workfn() calls do_migrate_pages() with
MPOL_MF_MOVE_ALL.
(5) NUMA hinting
mm/migrate.c:migrate_misplaced_folio_prepare() will skip file
folios that are probably shared libraries (-> "mapped shared" and
executable). This check would have detected it as a shared library at
some point (at least 3 MMs mapping it), so detecting it afterwards
does not sound wrong (still a shared library). Not expected to
matter.
mm/memory.c:numa_migrate_check() will indicate TNF_SHARED in
MAP_SHARED file mappings when encountering a shared folio. Similar
reasoning, not expected to matter.
mm/mprotect.c:change_pte_range() will skip folios detected as
shared in CoW mappings. Similarly, this is not expected to matter in
practice, but if it would ever be a problem we could relax that check
a bit (e.g., basing it on the average page-mapcount in a folio),
because it was only an optimization when many (e.g., 288) processes
were mapping the same folios -- see commit 859d4adc3415 ("mm: numa: do
not trap faults on shared data section pages.")
(6) mm/rmap.c:folio_referenced_one() will skip exclusive swapbacked
folios in dying processes. Applies to anonymous folios only. Without
"false negatives", we'll now skip all actually shared ones. Skipping
ones that are actually exclusive won't really matter, it's a pure
optimization, and is not expected to matter in practice.
In theory, one can detect the problematic scenario: folio_mapcount() > 0
and no folio MM slot is occupied ("state unknown"). One could reset the
MM slots while doing an rmap walk, which migration / folio split already
do when setting everything up. Further, when batching PTEs we might
naturally learn about a owner (e.g., folio_mapcount() == nr_ptes) and
could update the owner. However, we'll defer that until the scenarios
where it would really matter are clear.
Link: https://lkml.kernel.org/r/20250303163014.1128035-15-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Cc: Andy Lutomirks^H^Hski <luto@kernel.org>
Cc: Borislav Betkov <bp@alien8.de>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jann Horn <jannh@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: Liam Howlett <liam.howlett@oracle.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Michal Koutn <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: tejun heo <tj@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-03-03 17:30:07 +01:00
|
|
|
|
2025-01-13 14:16:07 +01:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
2013-09-11 14:22:03 -07:00
|
|
|
|
2013-11-14 14:31:02 -08:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
|
|
|
|
|
2013-09-11 14:22:03 -07:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2013-12-06 00:08:22 +05:30
|
|
|
|
2025-10-23 19:37:36 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-10-25 14:16:32 +02:00
|
|
|
|
2012-10-25 14:16:32 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-10-25 14:16:32 +02:00
|
|
|
|
2012-10-25 14:16:32 +02:00
|
|
|
|
|
|
|
|
|
2012-10-25 14:16:32 +02:00
|
|
|
|
mm/mprotect: use mmu_gather
Patch series "mm/mprotect: avoid unnecessary TLB flushes", v6.
This patchset is intended to remove unnecessary TLB flushes during
mprotect() syscalls. Once this patch-set make it through, similar and
further optimizations for MADV_COLD and userfaultfd would be possible.
Basically, there are 3 optimizations in this patch-set:
1. Use TLB batching infrastructure to batch flushes across VMAs and do
better/fewer flushes. This would also be handy for later userfaultfd
enhancements.
2. Avoid unnecessary TLB flushes. This optimization is the one that
provides most of the performance benefits. Unlike previous versions,
we now only avoid flushes that would not result in spurious
page-faults.
3. Avoiding TLB flushes on change_huge_pmd() that are only needed to
prevent the A/D bits from changing.
Andrew asked for some benchmark numbers. I do not have an easy
determinate macrobenchmark in which it is easy to show benefit. I
therefore ran a microbenchmark: a loop that does the following on
anonymous memory, just as a sanity check to see that time is saved by
avoiding TLB flushes. The loop goes:
mprotect(p, PAGE_SIZE, PROT_READ)
mprotect(p, PAGE_SIZE, PROT_READ|PROT_WRITE)
*p = 0; // make the page writable
The test was run in KVM guest with 1 or 2 threads (the second thread was
busy-looping). I measured the time (cycles) of each operation:
1 thread 2 threads
mmots +patch mmots +patch
PROT_READ 3494 2725 (-22%) 8630 7788 (-10%)
PROT_READ|WRITE 3952 2724 (-31%) 9075 2865 (-68%)
[ mmots = v5.17-rc6-mmots-2022-03-06-20-38 ]
The exact numbers are really meaningless, but the benefit is clear. There
are 2 interesting results though.
(1) PROT_READ is cheaper, while one can expect it not to be affected.
This is presumably due to TLB miss that is saved
(2) Without memory access (*p = 0), the speedup of the patch is even
greater. In that scenario mprotect(PROT_READ) also avoids the TLB flush.
As a result both operations on the patched kernel take roughly ~1500
cycles (with either 1 or 2 threads), whereas on mmotm their cost is as
high as presented in the table.
This patch (of 3):
change_pXX_range() currently does not use mmu_gather, but instead
implements its own deferred TLB flushes scheme. This both complicates the
code, as developers need to be aware of different invalidation schemes,
and prevents opportunities to avoid TLB flushes or perform them in finer
granularity.
The use of mmu_gather for modified PTEs has benefits in various scenarios
even if pages are not released. For instance, if only a single page needs
to be flushed out of a range of many pages, only that page would be
flushed. If a THP page is flushed, on x86 a single TLB invlpg instruction
can be used instead of 512 instructions (or a full TLB flush, which would
Linux would actually use by default). mprotect() over multiple VMAs
requires a single flush.
Use mmu_gather in change_pXX_range(). As the pages are not released, only
record the flushed range using tlb_flush_pXX_range().
Handle THP similarly and get rid of flush_cache_range() which becomes
redundant since tlb_start_vma() calls it when needed.
Link: https://lkml.kernel.org/r/20220401180821.1986781-1-namit@vmware.com
Link: https://lkml.kernel.org/r/20220401180821.1986781-2-namit@vmware.com
Signed-off-by: Nadav Amit <namit@vmware.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andrew Cooper <andrew.cooper3@citrix.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Will Deacon <will@kernel.org>
Cc: Yu Zhao <yuzhao@google.com>
Cc: Nick Piggin <npiggin@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-09 18:20:50 -07:00
|
|
|
|
2023-01-04 17:52:06 -05:00
|
|
|
|
2012-10-25 14:16:32 +02:00
|
|
|
|
mm/mprotect: use mmu_gather
Patch series "mm/mprotect: avoid unnecessary TLB flushes", v6.
This patchset is intended to remove unnecessary TLB flushes during
mprotect() syscalls. Once this patch-set make it through, similar and
further optimizations for MADV_COLD and userfaultfd would be possible.
Basically, there are 3 optimizations in this patch-set:
1. Use TLB batching infrastructure to batch flushes across VMAs and do
better/fewer flushes. This would also be handy for later userfaultfd
enhancements.
2. Avoid unnecessary TLB flushes. This optimization is the one that
provides most of the performance benefits. Unlike previous versions,
we now only avoid flushes that would not result in spurious
page-faults.
3. Avoiding TLB flushes on change_huge_pmd() that are only needed to
prevent the A/D bits from changing.
Andrew asked for some benchmark numbers. I do not have an easy
determinate macrobenchmark in which it is easy to show benefit. I
therefore ran a microbenchmark: a loop that does the following on
anonymous memory, just as a sanity check to see that time is saved by
avoiding TLB flushes. The loop goes:
mprotect(p, PAGE_SIZE, PROT_READ)
mprotect(p, PAGE_SIZE, PROT_READ|PROT_WRITE)
*p = 0; // make the page writable
The test was run in KVM guest with 1 or 2 threads (the second thread was
busy-looping). I measured the time (cycles) of each operation:
1 thread 2 threads
mmots +patch mmots +patch
PROT_READ 3494 2725 (-22%) 8630 7788 (-10%)
PROT_READ|WRITE 3952 2724 (-31%) 9075 2865 (-68%)
[ mmots = v5.17-rc6-mmots-2022-03-06-20-38 ]
The exact numbers are really meaningless, but the benefit is clear. There
are 2 interesting results though.
(1) PROT_READ is cheaper, while one can expect it not to be affected.
This is presumably due to TLB miss that is saved
(2) Without memory access (*p = 0), the speedup of the patch is even
greater. In that scenario mprotect(PROT_READ) also avoids the TLB flush.
As a result both operations on the patched kernel take roughly ~1500
cycles (with either 1 or 2 threads), whereas on mmotm their cost is as
high as presented in the table.
This patch (of 3):
change_pXX_range() currently does not use mmu_gather, but instead
implements its own deferred TLB flushes scheme. This both complicates the
code, as developers need to be aware of different invalidation schemes,
and prevents opportunities to avoid TLB flushes or perform them in finer
granularity.
The use of mmu_gather for modified PTEs has benefits in various scenarios
even if pages are not released. For instance, if only a single page needs
to be flushed out of a range of many pages, only that page would be
flushed. If a THP page is flushed, on x86 a single TLB invlpg instruction
can be used instead of 512 instructions (or a full TLB flush, which would
Linux would actually use by default). mprotect() over multiple VMAs
requires a single flush.
Use mmu_gather in change_pXX_range(). As the pages are not released, only
record the flushed range using tlb_flush_pXX_range().
Handle THP similarly and get rid of flush_cache_range() which becomes
redundant since tlb_start_vma() calls it when needed.
Link: https://lkml.kernel.org/r/20220401180821.1986781-1-namit@vmware.com
Link: https://lkml.kernel.org/r/20220401180821.1986781-2-namit@vmware.com
Signed-off-by: Nadav Amit <namit@vmware.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andrew Cooper <andrew.cooper3@citrix.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Will Deacon <will@kernel.org>
Cc: Yu Zhao <yuzhao@google.com>
Cc: Nick Piggin <npiggin@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-09 18:20:50 -07:00
|
|
|
|
|
|
|
|
|
2022-12-23 16:56:16 +01:00
|
|
|
|
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
|
|
|
|
2012-11-02 14:52:48 +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
|
|
|
|
|
|
|
|
|
2012-10-25 14:16:32 +02:00
|
|
|
|
mm/mprotect: use mmu_gather
Patch series "mm/mprotect: avoid unnecessary TLB flushes", v6.
This patchset is intended to remove unnecessary TLB flushes during
mprotect() syscalls. Once this patch-set make it through, similar and
further optimizations for MADV_COLD and userfaultfd would be possible.
Basically, there are 3 optimizations in this patch-set:
1. Use TLB batching infrastructure to batch flushes across VMAs and do
better/fewer flushes. This would also be handy for later userfaultfd
enhancements.
2. Avoid unnecessary TLB flushes. This optimization is the one that
provides most of the performance benefits. Unlike previous versions,
we now only avoid flushes that would not result in spurious
page-faults.
3. Avoiding TLB flushes on change_huge_pmd() that are only needed to
prevent the A/D bits from changing.
Andrew asked for some benchmark numbers. I do not have an easy
determinate macrobenchmark in which it is easy to show benefit. I
therefore ran a microbenchmark: a loop that does the following on
anonymous memory, just as a sanity check to see that time is saved by
avoiding TLB flushes. The loop goes:
mprotect(p, PAGE_SIZE, PROT_READ)
mprotect(p, PAGE_SIZE, PROT_READ|PROT_WRITE)
*p = 0; // make the page writable
The test was run in KVM guest with 1 or 2 threads (the second thread was
busy-looping). I measured the time (cycles) of each operation:
1 thread 2 threads
mmots +patch mmots +patch
PROT_READ 3494 2725 (-22%) 8630 7788 (-10%)
PROT_READ|WRITE 3952 2724 (-31%) 9075 2865 (-68%)
[ mmots = v5.17-rc6-mmots-2022-03-06-20-38 ]
The exact numbers are really meaningless, but the benefit is clear. There
are 2 interesting results though.
(1) PROT_READ is cheaper, while one can expect it not to be affected.
This is presumably due to TLB miss that is saved
(2) Without memory access (*p = 0), the speedup of the patch is even
greater. In that scenario mprotect(PROT_READ) also avoids the TLB flush.
As a result both operations on the patched kernel take roughly ~1500
cycles (with either 1 or 2 threads), whereas on mmotm their cost is as
high as presented in the table.
This patch (of 3):
change_pXX_range() currently does not use mmu_gather, but instead
implements its own deferred TLB flushes scheme. This both complicates the
code, as developers need to be aware of different invalidation schemes,
and prevents opportunities to avoid TLB flushes or perform them in finer
granularity.
The use of mmu_gather for modified PTEs has benefits in various scenarios
even if pages are not released. For instance, if only a single page needs
to be flushed out of a range of many pages, only that page would be
flushed. If a THP page is flushed, on x86 a single TLB invlpg instruction
can be used instead of 512 instructions (or a full TLB flush, which would
Linux would actually use by default). mprotect() over multiple VMAs
requires a single flush.
Use mmu_gather in change_pXX_range(). As the pages are not released, only
record the flushed range using tlb_flush_pXX_range().
Handle THP similarly and get rid of flush_cache_range() which becomes
redundant since tlb_start_vma() calls it when needed.
Link: https://lkml.kernel.org/r/20220401180821.1986781-1-namit@vmware.com
Link: https://lkml.kernel.org/r/20220401180821.1986781-2-namit@vmware.com
Signed-off-by: Nadav Amit <namit@vmware.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andrew Cooper <andrew.cooper3@citrix.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Will Deacon <will@kernel.org>
Cc: Yu Zhao <yuzhao@google.com>
Cc: Nick Piggin <npiggin@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-09 18:20:50 -07:00
|
|
|
|
|
|
|
|
|
2012-10-25 14:16:32 +02:00
|
|
|
|
2012-10-25 14:16:32 +02:00
|
|
|
|
2013-12-06 00:08:22 +05:30
|
|
|
|
2012-10-25 14:16:32 +02:00
|
|
|
|
2015-02-11 15:28:03 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-09-06 19:49:02 +00:00
|
|
|
|
2015-02-11 15:28:03 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy.c: check range first in queue_pages_test_walk
Patch series "mm: Fix checking unmapped holes for mbind", v4.
This patchset fix checking unmapped holes for mbind().
First patch makes sure the vma been correctly tracked in .test_walk(),
so each time when .test_walk() is called, the neighborhood of two vma
is correct.
Current problem is that the !vma_migratable() check could cause return
immediately without update tracking to vma.
Second patch fix the inconsistent report of EFAULT when mbind() is
called for MPOL_DEFAULT and non MPOL_DEFAULT cases, so application do
not need to have workaround code to handle this special behavior.
Currently there are two problems, one is that the .test_walk() can not
know there is hole at tail side of range, because .test_walk() only
call for vma not for hole. The other one is that mbind_range() checks
for hole at head side of range but do not consider the
MPOL_MF_DISCONTIG_OK flag as done in .test_walk().
This patch (of 2):
Checking unmapped hole and updating the previous vma must be handled
first, otherwise the unmapped hole could be calculated from a wrong
previous vma.
Several commits were relevant to this error:
- commit 6f4576e3687b ("mempolicy: apply page table walker on
queue_pages_range()")
This commit was correct, the VM_PFNMAP check was after updating
previous vma
- commit 48684a65b4e3 ("mm: pagewalk: fix misbehavior of
walk_page_range for vma(VM_PFNMAP)")
This commit added VM_PFNMAP check before updating previous vma. Then,
there were two VM_PFNMAP check did same thing twice.
- commit acda0c334028 ("mm/mempolicy.c: get rid of duplicated check for
vma(VM_PFNMAP) in queue_page s_range()")
This commit tried to fix the duplicated VM_PFNMAP check, but it
wrongly removed the one which was after updating vma.
Link: http://lkml.kernel.org/r/1573218104-11021-2-git-send-email-lixinhai.lxh@gmail.com
Fixes: acda0c334028 (mm/mempolicy.c: get rid of duplicated check for vma(VM_PFNMAP) in queue_pages_range())
Signed-off-by: Li Xinhai <lixinhai.lxh@gmail.com>
Reviewed-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Hugh Dickins <hughd@google.com>
Cc: linux-man <linux-man@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-11-30 17:56:15 -08:00
|
|
|
|
2021-02-24 12:09:47 -08:00
|
|
|
|
mm/mempolicy.c: fix checking unmapped holes for mbind
mbind() is required to report EFAULT if range, specified by addr and
len, contains unmapped holes. In current implementation, below rules
are applied for this checking:
1: Unmapped holes at any part of the specified range should be reported
as EFAULT if mbind() for none MPOL_DEFAULT cases;
2: Unmapped holes at any part of the specified range should be ignored
(do not reprot EFAULT) if mbind() for MPOL_DEFAULT case;
3: The whole range in an unmapped hole should be reported as EFAULT;
Note that rule 2 does not fullfill the mbind() API definition, but since
that behavior has existed for long days (the internal flag
MPOL_MF_DISCONTIG_OK is for this purpose), this patch does not plan to
change it.
In current code, application observed inconsistent behavior on rule 1
and rule 2 respectively. That inconsistency is fixed as below details.
Cases of rule 1:
- Hole at head side of range. Current code reprot EFAULT, no change by
this patch.
[ vma ][ hole ][ vma ]
[ range ]
- Hole at middle of range. Current code report EFAULT, no change by
this patch.
[ vma ][ hole ][ vma ]
[ range ]
- Hole at tail side of range. Current code do not report EFAULT, this
patch fixes it.
[ vma ][ hole ][ vma ]
[ range ]
Cases of rule 2:
- Hole at head side of range. Current code reports EFAULT, this patch
fixes it.
[ vma ][ hole ][ vma ]
[ range ]
- Hole at middle of range. Current code does not report EFAULT, no
change by this patch.
[ vma ][ hole ][ vma]
[ range ]
- Hole at tail side of range. Current code does not report EFAULT, no
change by this patch.
[ vma ][ hole ][ vma]
[ range ]
This patch has no changes to rule 3.
The unmapped hole checking can also be handled by using .pte_hole(),
instead of .test_walk(). But .pte_hole() is called for holes inside and
outside vma, which causes more cost, so this patch keeps the original
design with .test_walk().
Link: http://lkml.kernel.org/r/1573218104-11021-3-git-send-email-lixinhai.lxh@gmail.com
Fixes: 6f4576e3687b ("mempolicy: apply page table walker on queue_pages_range()")
Signed-off-by: Li Xinhai <lixinhai.lxh@gmail.com>
Reviewed-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Hugh Dickins <hughd@google.com>
Cc: linux-man <linux-man@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-11-30 17:56:18 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy.c: check range first in queue_pages_test_walk
Patch series "mm: Fix checking unmapped holes for mbind", v4.
This patchset fix checking unmapped holes for mbind().
First patch makes sure the vma been correctly tracked in .test_walk(),
so each time when .test_walk() is called, the neighborhood of two vma
is correct.
Current problem is that the !vma_migratable() check could cause return
immediately without update tracking to vma.
Second patch fix the inconsistent report of EFAULT when mbind() is
called for MPOL_DEFAULT and non MPOL_DEFAULT cases, so application do
not need to have workaround code to handle this special behavior.
Currently there are two problems, one is that the .test_walk() can not
know there is hole at tail side of range, because .test_walk() only
call for vma not for hole. The other one is that mbind_range() checks
for hole at head side of range but do not consider the
MPOL_MF_DISCONTIG_OK flag as done in .test_walk().
This patch (of 2):
Checking unmapped hole and updating the previous vma must be handled
first, otherwise the unmapped hole could be calculated from a wrong
previous vma.
Several commits were relevant to this error:
- commit 6f4576e3687b ("mempolicy: apply page table walker on
queue_pages_range()")
This commit was correct, the VM_PFNMAP check was after updating
previous vma
- commit 48684a65b4e3 ("mm: pagewalk: fix misbehavior of
walk_page_range for vma(VM_PFNMAP)")
This commit added VM_PFNMAP check before updating previous vma. Then,
there were two VM_PFNMAP check did same thing twice.
- commit acda0c334028 ("mm/mempolicy.c: get rid of duplicated check for
vma(VM_PFNMAP) in queue_page s_range()")
This commit tried to fix the duplicated VM_PFNMAP check, but it
wrongly removed the one which was after updating vma.
Link: http://lkml.kernel.org/r/1573218104-11021-2-git-send-email-lixinhai.lxh@gmail.com
Fixes: acda0c334028 (mm/mempolicy.c: get rid of duplicated check for vma(VM_PFNMAP) in queue_pages_range())
Signed-off-by: Li Xinhai <lixinhai.lxh@gmail.com>
Reviewed-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Hugh Dickins <hughd@google.com>
Cc: linux-man <linux-man@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-11-30 17:56:15 -08:00
|
|
|
|
|
|
|
|
|
2022-09-06 19:49:02 +00:00
|
|
|
|
mm/mempolicy.c: fix checking unmapped holes for mbind
mbind() is required to report EFAULT if range, specified by addr and
len, contains unmapped holes. In current implementation, below rules
are applied for this checking:
1: Unmapped holes at any part of the specified range should be reported
as EFAULT if mbind() for none MPOL_DEFAULT cases;
2: Unmapped holes at any part of the specified range should be ignored
(do not reprot EFAULT) if mbind() for MPOL_DEFAULT case;
3: The whole range in an unmapped hole should be reported as EFAULT;
Note that rule 2 does not fullfill the mbind() API definition, but since
that behavior has existed for long days (the internal flag
MPOL_MF_DISCONTIG_OK is for this purpose), this patch does not plan to
change it.
In current code, application observed inconsistent behavior on rule 1
and rule 2 respectively. That inconsistency is fixed as below details.
Cases of rule 1:
- Hole at head side of range. Current code reprot EFAULT, no change by
this patch.
[ vma ][ hole ][ vma ]
[ range ]
- Hole at middle of range. Current code report EFAULT, no change by
this patch.
[ vma ][ hole ][ vma ]
[ range ]
- Hole at tail side of range. Current code do not report EFAULT, this
patch fixes it.
[ vma ][ hole ][ vma ]
[ range ]
Cases of rule 2:
- Hole at head side of range. Current code reports EFAULT, this patch
fixes it.
[ vma ][ hole ][ vma ]
[ range ]
- Hole at middle of range. Current code does not report EFAULT, no
change by this patch.
[ vma ][ hole ][ vma]
[ range ]
- Hole at tail side of range. Current code does not report EFAULT, no
change by this patch.
[ vma ][ hole ][ vma]
[ range ]
This patch has no changes to rule 3.
The unmapped hole checking can also be handled by using .pte_hole(),
instead of .test_walk(). But .pte_hole() is called for holes inside and
outside vma, which causes more cost, so this patch keeps the original
design with .test_walk().
Link: http://lkml.kernel.org/r/1573218104-11021-3-git-send-email-lixinhai.lxh@gmail.com
Fixes: 6f4576e3687b ("mempolicy: apply page table walker on queue_pages_range()")
Signed-off-by: Li Xinhai <lixinhai.lxh@gmail.com>
Reviewed-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Hugh Dickins <hughd@google.com>
Cc: linux-man <linux-man@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-11-30 17:56:18 -08:00
|
|
|
|
|
|
|
|
|
2022-09-06 19:49:02 +00:00
|
|
|
|
mm/mempolicy.c: fix checking unmapped holes for mbind
mbind() is required to report EFAULT if range, specified by addr and
len, contains unmapped holes. In current implementation, below rules
are applied for this checking:
1: Unmapped holes at any part of the specified range should be reported
as EFAULT if mbind() for none MPOL_DEFAULT cases;
2: Unmapped holes at any part of the specified range should be ignored
(do not reprot EFAULT) if mbind() for MPOL_DEFAULT case;
3: The whole range in an unmapped hole should be reported as EFAULT;
Note that rule 2 does not fullfill the mbind() API definition, but since
that behavior has existed for long days (the internal flag
MPOL_MF_DISCONTIG_OK is for this purpose), this patch does not plan to
change it.
In current code, application observed inconsistent behavior on rule 1
and rule 2 respectively. That inconsistency is fixed as below details.
Cases of rule 1:
- Hole at head side of range. Current code reprot EFAULT, no change by
this patch.
[ vma ][ hole ][ vma ]
[ range ]
- Hole at middle of range. Current code report EFAULT, no change by
this patch.
[ vma ][ hole ][ vma ]
[ range ]
- Hole at tail side of range. Current code do not report EFAULT, this
patch fixes it.
[ vma ][ hole ][ vma ]
[ range ]
Cases of rule 2:
- Hole at head side of range. Current code reports EFAULT, this patch
fixes it.
[ vma ][ hole ][ vma ]
[ range ]
- Hole at middle of range. Current code does not report EFAULT, no
change by this patch.
[ vma ][ hole ][ vma]
[ range ]
- Hole at tail side of range. Current code does not report EFAULT, no
change by this patch.
[ vma ][ hole ][ vma]
[ range ]
This patch has no changes to rule 3.
The unmapped hole checking can also be handled by using .pte_hole(),
instead of .test_walk(). But .pte_hole() is called for holes inside and
outside vma, which causes more cost, so this patch keeps the original
design with .test_walk().
Link: http://lkml.kernel.org/r/1573218104-11021-3-git-send-email-lixinhai.lxh@gmail.com
Fixes: 6f4576e3687b ("mempolicy: apply page table walker on queue_pages_range()")
Signed-off-by: Li Xinhai <lixinhai.lxh@gmail.com>
Reviewed-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Hugh Dickins <hughd@google.com>
Cc: linux-man <linux-man@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-11-30 17:56:18 -08:00
|
|
|
|
|
|
|
|
|
mm/mempolicy.c: check range first in queue_pages_test_walk
Patch series "mm: Fix checking unmapped holes for mbind", v4.
This patchset fix checking unmapped holes for mbind().
First patch makes sure the vma been correctly tracked in .test_walk(),
so each time when .test_walk() is called, the neighborhood of two vma
is correct.
Current problem is that the !vma_migratable() check could cause return
immediately without update tracking to vma.
Second patch fix the inconsistent report of EFAULT when mbind() is
called for MPOL_DEFAULT and non MPOL_DEFAULT cases, so application do
not need to have workaround code to handle this special behavior.
Currently there are two problems, one is that the .test_walk() can not
know there is hole at tail side of range, because .test_walk() only
call for vma not for hole. The other one is that mbind_range() checks
for hole at head side of range but do not consider the
MPOL_MF_DISCONTIG_OK flag as done in .test_walk().
This patch (of 2):
Checking unmapped hole and updating the previous vma must be handled
first, otherwise the unmapped hole could be calculated from a wrong
previous vma.
Several commits were relevant to this error:
- commit 6f4576e3687b ("mempolicy: apply page table walker on
queue_pages_range()")
This commit was correct, the VM_PFNMAP check was after updating
previous vma
- commit 48684a65b4e3 ("mm: pagewalk: fix misbehavior of
walk_page_range for vma(VM_PFNMAP)")
This commit added VM_PFNMAP check before updating previous vma. Then,
there were two VM_PFNMAP check did same thing twice.
- commit acda0c334028 ("mm/mempolicy.c: get rid of duplicated check for
vma(VM_PFNMAP) in queue_page s_range()")
This commit tried to fix the duplicated VM_PFNMAP check, but it
wrongly removed the one which was after updating vma.
Link: http://lkml.kernel.org/r/1573218104-11021-2-git-send-email-lixinhai.lxh@gmail.com
Fixes: acda0c334028 (mm/mempolicy.c: get rid of duplicated check for vma(VM_PFNMAP) in queue_pages_range())
Signed-off-by: Li Xinhai <lixinhai.lxh@gmail.com>
Reviewed-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Hugh Dickins <hughd@google.com>
Cc: linux-man <linux-man@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-11-30 17:56:15 -08:00
|
|
|
|
2019-03-28 20:43:55 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-02-11 15:28:06 -08:00
|
|
|
|
|
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-02-11 15:28:03 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-08-28 16:19:54 +02:00
|
|
|
|
2023-01-30 12:18:31 -08:00
|
|
|
|
2023-01-30 12:18:30 -08:00
|
|
|
|
2019-08-28 16:19:54 +02:00
|
|
|
|
2023-08-04 08:27:19 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-08-28 16:19:54 +02:00
|
|
|
|
|
|
|
|
|
2006-01-08 01:00:50 -08:00
|
|
|
|
2013-09-11 14:22:14 -07:00
|
|
|
|
|
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
|
|
|
|
|
mm: mempolicy: make the behavior consistent when MPOL_MF_MOVE* and MPOL_MF_STRICT were specified
When both MPOL_MF_MOVE* and MPOL_MF_STRICT was specified, mbind() should
try best to migrate misplaced pages, if some of the pages could not be
migrated, then return -EIO.
There are three different sub-cases:
1. vma is not migratable
2. vma is migratable, but there are unmovable pages
3. vma is migratable, pages are movable, but migrate_pages() fails
If #1 happens, kernel would just abort immediately, then return -EIO,
after a7f40cfe3b7a ("mm: mempolicy: make mbind() return -EIO when
MPOL_MF_STRICT is specified").
If #3 happens, kernel would set policy and migrate pages with
best-effort, but won't rollback the migrated pages and reset the policy
back.
Before that commit, they behaves in the same way. It'd better to keep
their behavior consistent. But, rolling back the migrated pages and
resetting the policy back sounds not feasible, so just make #1 behave as
same as #3.
Userspace will know that not everything was successfully migrated (via
-EIO), and can take whatever steps it deems necessary - attempt
rollback, determine which exact page(s) are violating the policy, etc.
Make queue_pages_range() return 1 to indicate there are unmovable pages
or vma is not migratable.
The #2 is not handled correctly in the current kernel, the following
patch will fix it.
[yang.shi@linux.alibaba.com: fix review comments from Vlastimil]
Link: http://lkml.kernel.org/r/1563556862-54056-2-git-send-email-yang.shi@linux.alibaba.com
Link: http://lkml.kernel.org/r/1561162809-59140-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
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 15:37:15 -07:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-01-08 01:00:50 -08:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
2013-09-11 14:22:14 -07:00
|
|
|
|
2015-02-11 15:28:03 -08:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
mm/mempolicy.c: fix checking unmapped holes for mbind
mbind() is required to report EFAULT if range, specified by addr and
len, contains unmapped holes. In current implementation, below rules
are applied for this checking:
1: Unmapped holes at any part of the specified range should be reported
as EFAULT if mbind() for none MPOL_DEFAULT cases;
2: Unmapped holes at any part of the specified range should be ignored
(do not reprot EFAULT) if mbind() for MPOL_DEFAULT case;
3: The whole range in an unmapped hole should be reported as EFAULT;
Note that rule 2 does not fullfill the mbind() API definition, but since
that behavior has existed for long days (the internal flag
MPOL_MF_DISCONTIG_OK is for this purpose), this patch does not plan to
change it.
In current code, application observed inconsistent behavior on rule 1
and rule 2 respectively. That inconsistency is fixed as below details.
Cases of rule 1:
- Hole at head side of range. Current code reprot EFAULT, no change by
this patch.
[ vma ][ hole ][ vma ]
[ range ]
- Hole at middle of range. Current code report EFAULT, no change by
this patch.
[ vma ][ hole ][ vma ]
[ range ]
- Hole at tail side of range. Current code do not report EFAULT, this
patch fixes it.
[ vma ][ hole ][ vma ]
[ range ]
Cases of rule 2:
- Hole at head side of range. Current code reports EFAULT, this patch
fixes it.
[ vma ][ hole ][ vma ]
[ range ]
- Hole at middle of range. Current code does not report EFAULT, no
change by this patch.
[ vma ][ hole ][ vma]
[ range ]
- Hole at tail side of range. Current code does not report EFAULT, no
change by this patch.
[ vma ][ hole ][ vma]
[ range ]
This patch has no changes to rule 3.
The unmapped hole checking can also be handled by using .pte_hole(),
instead of .test_walk(). But .pte_hole() is called for holes inside and
outside vma, which causes more cost, so this patch keeps the original
design with .test_walk().
Link: http://lkml.kernel.org/r/1573218104-11021-3-git-send-email-lixinhai.lxh@gmail.com
Fixes: 6f4576e3687b ("mempolicy: apply page table walker on queue_pages_range()")
Signed-off-by: Li Xinhai <lixinhai.lxh@gmail.com>
Reviewed-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Hugh Dickins <hughd@google.com>
Cc: linux-man <linux-man@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-11-30 17:56:18 -08:00
|
|
|
|
2015-02-11 15:28:03 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy.c: fix checking unmapped holes for mbind
mbind() is required to report EFAULT if range, specified by addr and
len, contains unmapped holes. In current implementation, below rules
are applied for this checking:
1: Unmapped holes at any part of the specified range should be reported
as EFAULT if mbind() for none MPOL_DEFAULT cases;
2: Unmapped holes at any part of the specified range should be ignored
(do not reprot EFAULT) if mbind() for MPOL_DEFAULT case;
3: The whole range in an unmapped hole should be reported as EFAULT;
Note that rule 2 does not fullfill the mbind() API definition, but since
that behavior has existed for long days (the internal flag
MPOL_MF_DISCONTIG_OK is for this purpose), this patch does not plan to
change it.
In current code, application observed inconsistent behavior on rule 1
and rule 2 respectively. That inconsistency is fixed as below details.
Cases of rule 1:
- Hole at head side of range. Current code reprot EFAULT, no change by
this patch.
[ vma ][ hole ][ vma ]
[ range ]
- Hole at middle of range. Current code report EFAULT, no change by
this patch.
[ vma ][ hole ][ vma ]
[ range ]
- Hole at tail side of range. Current code do not report EFAULT, this
patch fixes it.
[ vma ][ hole ][ vma ]
[ range ]
Cases of rule 2:
- Hole at head side of range. Current code reports EFAULT, this patch
fixes it.
[ vma ][ hole ][ vma ]
[ range ]
- Hole at middle of range. Current code does not report EFAULT, no
change by this patch.
[ vma ][ hole ][ vma]
[ range ]
- Hole at tail side of range. Current code does not report EFAULT, no
change by this patch.
[ vma ][ hole ][ vma]
[ range ]
This patch has no changes to rule 3.
The unmapped hole checking can also be handled by using .pte_hole(),
instead of .test_walk(). But .pte_hole() is called for holes inside and
outside vma, which causes more cost, so this patch keeps the original
design with .test_walk().
Link: http://lkml.kernel.org/r/1573218104-11021-3-git-send-email-lixinhai.lxh@gmail.com
Fixes: 6f4576e3687b ("mempolicy: apply page table walker on queue_pages_range()")
Signed-off-by: Li Xinhai <lixinhai.lxh@gmail.com>
Reviewed-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Hugh Dickins <hughd@google.com>
Cc: linux-man <linux-man@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-11-30 17:56:18 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-02-11 15:28:03 -08:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
2023-08-04 08:27:19 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2023-08-04 08:27:19 -07:00
|
|
|
|
mm/mempolicy.c: fix checking unmapped holes for mbind
mbind() is required to report EFAULT if range, specified by addr and
len, contains unmapped holes. In current implementation, below rules
are applied for this checking:
1: Unmapped holes at any part of the specified range should be reported
as EFAULT if mbind() for none MPOL_DEFAULT cases;
2: Unmapped holes at any part of the specified range should be ignored
(do not reprot EFAULT) if mbind() for MPOL_DEFAULT case;
3: The whole range in an unmapped hole should be reported as EFAULT;
Note that rule 2 does not fullfill the mbind() API definition, but since
that behavior has existed for long days (the internal flag
MPOL_MF_DISCONTIG_OK is for this purpose), this patch does not plan to
change it.
In current code, application observed inconsistent behavior on rule 1
and rule 2 respectively. That inconsistency is fixed as below details.
Cases of rule 1:
- Hole at head side of range. Current code reprot EFAULT, no change by
this patch.
[ vma ][ hole ][ vma ]
[ range ]
- Hole at middle of range. Current code report EFAULT, no change by
this patch.
[ vma ][ hole ][ vma ]
[ range ]
- Hole at tail side of range. Current code do not report EFAULT, this
patch fixes it.
[ vma ][ hole ][ vma ]
[ range ]
Cases of rule 2:
- Hole at head side of range. Current code reports EFAULT, this patch
fixes it.
[ vma ][ hole ][ vma ]
[ range ]
- Hole at middle of range. Current code does not report EFAULT, no
change by this patch.
[ vma ][ hole ][ vma]
[ range ]
- Hole at tail side of range. Current code does not report EFAULT, no
change by this patch.
[ vma ][ hole ][ vma]
[ range ]
This patch has no changes to rule 3.
The unmapped hole checking can also be handled by using .pte_hole(),
instead of .test_walk(). But .pte_hole() is called for holes inside and
outside vma, which causes more cost, so this patch keeps the original
design with .test_walk().
Link: http://lkml.kernel.org/r/1573218104-11021-3-git-send-email-lixinhai.lxh@gmail.com
Fixes: 6f4576e3687b ("mempolicy: apply page table walker on queue_pages_range()")
Signed-off-by: Li Xinhai <lixinhai.lxh@gmail.com>
Reviewed-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Hugh Dickins <hughd@google.com>
Cc: linux-man <linux-man@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-11-30 17:56:18 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
mempolicy: remove mempolicy sharing
Dave Jones' system call fuzz testing tool "trinity" triggered the
following bug error with slab debugging enabled
=============================================================================
BUG numa_policy (Not tainted): Poison overwritten
-----------------------------------------------------------------------------
INFO: 0xffff880146498250-0xffff880146498250. First byte 0x6a instead of 0x6b
INFO: Allocated in mpol_new+0xa3/0x140 age=46310 cpu=6 pid=32154
__slab_alloc+0x3d3/0x445
kmem_cache_alloc+0x29d/0x2b0
mpol_new+0xa3/0x140
sys_mbind+0x142/0x620
system_call_fastpath+0x16/0x1b
INFO: Freed in __mpol_put+0x27/0x30 age=46268 cpu=6 pid=32154
__slab_free+0x2e/0x1de
kmem_cache_free+0x25a/0x260
__mpol_put+0x27/0x30
remove_vma+0x68/0x90
exit_mmap+0x118/0x140
mmput+0x73/0x110
exit_mm+0x108/0x130
do_exit+0x162/0xb90
do_group_exit+0x4f/0xc0
sys_exit_group+0x17/0x20
system_call_fastpath+0x16/0x1b
INFO: Slab 0xffffea0005192600 objects=27 used=27 fp=0x (null) flags=0x20000000004080
INFO: Object 0xffff880146498250 @offset=592 fp=0xffff88014649b9d0
The problem is that the structure is being prematurely freed due to a
reference count imbalance. In the following case mbind(addr, len) should
replace the memory policies of both vma1 and vma2 and thus they will
become to share the same mempolicy and the new mempolicy will have the
MPOL_F_SHARED flag.
+-------------------+-------------------+
| vma1 | vma2(shmem) |
+-------------------+-------------------+
| |
addr addr+len
alloc_pages_vma() uses get_vma_policy() and mpol_cond_put() pair for
maintaining the mempolicy reference count. The current rule is that
get_vma_policy() only increments refcount for shmem VMA and
mpol_conf_put() only decrements refcount if the policy has
MPOL_F_SHARED.
In above case, vma1 is not shmem vma and vma->policy has MPOL_F_SHARED!
The reference count will be decreased even though was not increased
whenever alloc_page_vma() is called. This has been broken since commit
[52cd3b07: mempolicy: rework mempolicy Reference Counting] in 2008.
There is another serious bug with the sharing of memory policies.
Currently, mempolicy rebind logic (it is called from cpuset rebinding)
ignores a refcount of mempolicy and override it forcibly. Thus, any
mempolicy sharing may cause mempolicy corruption. The bug was
introduced by commit [68860ec1: cpusets: automatic numa mempolicy
rebinding].
Ideally, the shared policy handling would be rewritten to either
properly handle COW of the policy structures or at least reference count
MPOL_F_SHARED based exclusively on information within the policy.
However, this patch takes the easier approach of disabling any policy
sharing between VMAs. Each new range allocated with sp_alloc will
allocate a new policy, set the reference count to 1 and drop the
reference count of the old policy. This increases the memory footprint
but is not expected to be a major problem as mbind() is unlikely to be
used for fine-grained ranges. It is also inefficient because it means
we allocate a new policy even in cases where mbind_range() could use the
new_policy passed to it. However, it is more straight-forward and the
change should be invisible to the user.
[mgorman@suse.de: Edited changelog]
Reported-by: Dave Jones <davej@redhat.com>,
Cc: Christoph Lameter <cl@linux.com>,
Reviewed-by: Christoph Lameter <cl@linux.com>
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Josh Boyer <jwboyer@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-10-08 16:29:16 -07:00
|
|
|
|
|
|
|
|
|
2020-06-08 21:33:54 -07:00
|
|
|
|
mempolicy: remove mempolicy sharing
Dave Jones' system call fuzz testing tool "trinity" triggered the
following bug error with slab debugging enabled
=============================================================================
BUG numa_policy (Not tainted): Poison overwritten
-----------------------------------------------------------------------------
INFO: 0xffff880146498250-0xffff880146498250. First byte 0x6a instead of 0x6b
INFO: Allocated in mpol_new+0xa3/0x140 age=46310 cpu=6 pid=32154
__slab_alloc+0x3d3/0x445
kmem_cache_alloc+0x29d/0x2b0
mpol_new+0xa3/0x140
sys_mbind+0x142/0x620
system_call_fastpath+0x16/0x1b
INFO: Freed in __mpol_put+0x27/0x30 age=46268 cpu=6 pid=32154
__slab_free+0x2e/0x1de
kmem_cache_free+0x25a/0x260
__mpol_put+0x27/0x30
remove_vma+0x68/0x90
exit_mmap+0x118/0x140
mmput+0x73/0x110
exit_mm+0x108/0x130
do_exit+0x162/0xb90
do_group_exit+0x4f/0xc0
sys_exit_group+0x17/0x20
system_call_fastpath+0x16/0x1b
INFO: Slab 0xffffea0005192600 objects=27 used=27 fp=0x (null) flags=0x20000000004080
INFO: Object 0xffff880146498250 @offset=592 fp=0xffff88014649b9d0
The problem is that the structure is being prematurely freed due to a
reference count imbalance. In the following case mbind(addr, len) should
replace the memory policies of both vma1 and vma2 and thus they will
become to share the same mempolicy and the new mempolicy will have the
MPOL_F_SHARED flag.
+-------------------+-------------------+
| vma1 | vma2(shmem) |
+-------------------+-------------------+
| |
addr addr+len
alloc_pages_vma() uses get_vma_policy() and mpol_cond_put() pair for
maintaining the mempolicy reference count. The current rule is that
get_vma_policy() only increments refcount for shmem VMA and
mpol_conf_put() only decrements refcount if the policy has
MPOL_F_SHARED.
In above case, vma1 is not shmem vma and vma->policy has MPOL_F_SHARED!
The reference count will be decreased even though was not increased
whenever alloc_page_vma() is called. This has been broken since commit
[52cd3b07: mempolicy: rework mempolicy Reference Counting] in 2008.
There is another serious bug with the sharing of memory policies.
Currently, mempolicy rebind logic (it is called from cpuset rebinding)
ignores a refcount of mempolicy and override it forcibly. Thus, any
mempolicy sharing may cause mempolicy corruption. The bug was
introduced by commit [68860ec1: cpusets: automatic numa mempolicy
rebinding].
Ideally, the shared policy handling would be rewritten to either
properly handle COW of the policy structures or at least reference count
MPOL_F_SHARED based exclusively on information within the policy.
However, this patch takes the easier approach of disabling any policy
sharing between VMAs. Each new range allocated with sp_alloc will
allocate a new policy, set the reference count to 1 and drop the
reference count of the old policy. This increases the memory footprint
but is not expected to be a major problem as mbind() is unlikely to be
used for fine-grained ranges. It is also inefficient because it means
we allocate a new policy even in cases where mbind_range() could use the
new_policy passed to it. However, it is more straight-forward and the
change should be invisible to the user.
[mgorman@suse.de: Edited changelog]
Reported-by: Dave Jones <davej@redhat.com>,
Cc: Christoph Lameter <cl@linux.com>,
Reviewed-by: Christoph Lameter <cl@linux.com>
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Josh Boyer <jwboyer@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-10-08 16:29:16 -07:00
|
|
|
|
|
|
|
|
|
2023-10-03 02:20:14 -07:00
|
|
|
|
2012-10-08 16:29:14 -07:00
|
|
|
|
mempolicy: remove mempolicy sharing
Dave Jones' system call fuzz testing tool "trinity" triggered the
following bug error with slab debugging enabled
=============================================================================
BUG numa_policy (Not tainted): Poison overwritten
-----------------------------------------------------------------------------
INFO: 0xffff880146498250-0xffff880146498250. First byte 0x6a instead of 0x6b
INFO: Allocated in mpol_new+0xa3/0x140 age=46310 cpu=6 pid=32154
__slab_alloc+0x3d3/0x445
kmem_cache_alloc+0x29d/0x2b0
mpol_new+0xa3/0x140
sys_mbind+0x142/0x620
system_call_fastpath+0x16/0x1b
INFO: Freed in __mpol_put+0x27/0x30 age=46268 cpu=6 pid=32154
__slab_free+0x2e/0x1de
kmem_cache_free+0x25a/0x260
__mpol_put+0x27/0x30
remove_vma+0x68/0x90
exit_mmap+0x118/0x140
mmput+0x73/0x110
exit_mm+0x108/0x130
do_exit+0x162/0xb90
do_group_exit+0x4f/0xc0
sys_exit_group+0x17/0x20
system_call_fastpath+0x16/0x1b
INFO: Slab 0xffffea0005192600 objects=27 used=27 fp=0x (null) flags=0x20000000004080
INFO: Object 0xffff880146498250 @offset=592 fp=0xffff88014649b9d0
The problem is that the structure is being prematurely freed due to a
reference count imbalance. In the following case mbind(addr, len) should
replace the memory policies of both vma1 and vma2 and thus they will
become to share the same mempolicy and the new mempolicy will have the
MPOL_F_SHARED flag.
+-------------------+-------------------+
| vma1 | vma2(shmem) |
+-------------------+-------------------+
| |
addr addr+len
alloc_pages_vma() uses get_vma_policy() and mpol_cond_put() pair for
maintaining the mempolicy reference count. The current rule is that
get_vma_policy() only increments refcount for shmem VMA and
mpol_conf_put() only decrements refcount if the policy has
MPOL_F_SHARED.
In above case, vma1 is not shmem vma and vma->policy has MPOL_F_SHARED!
The reference count will be decreased even though was not increased
whenever alloc_page_vma() is called. This has been broken since commit
[52cd3b07: mempolicy: rework mempolicy Reference Counting] in 2008.
There is another serious bug with the sharing of memory policies.
Currently, mempolicy rebind logic (it is called from cpuset rebinding)
ignores a refcount of mempolicy and override it forcibly. Thus, any
mempolicy sharing may cause mempolicy corruption. The bug was
introduced by commit [68860ec1: cpusets: automatic numa mempolicy
rebinding].
Ideally, the shared policy handling would be rewritten to either
properly handle COW of the policy structures or at least reference count
MPOL_F_SHARED based exclusively on information within the policy.
However, this patch takes the easier approach of disabling any policy
sharing between VMAs. Each new range allocated with sp_alloc will
allocate a new policy, set the reference count to 1 and drop the
reference count of the old policy. This increases the memory footprint
but is not expected to be a major problem as mbind() is unlikely to be
used for fine-grained ranges. It is also inefficient because it means
we allocate a new policy even in cases where mbind_range() could use the
new_policy passed to it. However, it is more straight-forward and the
change should be invisible to the user.
[mgorman@suse.de: Edited changelog]
Reported-by: Dave Jones <davej@redhat.com>,
Cc: Christoph Lameter <cl@linux.com>,
Reviewed-by: Christoph Lameter <cl@linux.com>
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Josh Boyer <jwboyer@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-10-08 16:29:16 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-10-08 16:29:14 -07:00
|
|
|
|
2023-07-28 06:13:21 +02:00
|
|
|
|
|
|
|
|
|
mempolicy: remove mempolicy sharing
Dave Jones' system call fuzz testing tool "trinity" triggered the
following bug error with slab debugging enabled
=============================================================================
BUG numa_policy (Not tainted): Poison overwritten
-----------------------------------------------------------------------------
INFO: 0xffff880146498250-0xffff880146498250. First byte 0x6a instead of 0x6b
INFO: Allocated in mpol_new+0xa3/0x140 age=46310 cpu=6 pid=32154
__slab_alloc+0x3d3/0x445
kmem_cache_alloc+0x29d/0x2b0
mpol_new+0xa3/0x140
sys_mbind+0x142/0x620
system_call_fastpath+0x16/0x1b
INFO: Freed in __mpol_put+0x27/0x30 age=46268 cpu=6 pid=32154
__slab_free+0x2e/0x1de
kmem_cache_free+0x25a/0x260
__mpol_put+0x27/0x30
remove_vma+0x68/0x90
exit_mmap+0x118/0x140
mmput+0x73/0x110
exit_mm+0x108/0x130
do_exit+0x162/0xb90
do_group_exit+0x4f/0xc0
sys_exit_group+0x17/0x20
system_call_fastpath+0x16/0x1b
INFO: Slab 0xffffea0005192600 objects=27 used=27 fp=0x (null) flags=0x20000000004080
INFO: Object 0xffff880146498250 @offset=592 fp=0xffff88014649b9d0
The problem is that the structure is being prematurely freed due to a
reference count imbalance. In the following case mbind(addr, len) should
replace the memory policies of both vma1 and vma2 and thus they will
become to share the same mempolicy and the new mempolicy will have the
MPOL_F_SHARED flag.
+-------------------+-------------------+
| vma1 | vma2(shmem) |
+-------------------+-------------------+
| |
addr addr+len
alloc_pages_vma() uses get_vma_policy() and mpol_cond_put() pair for
maintaining the mempolicy reference count. The current rule is that
get_vma_policy() only increments refcount for shmem VMA and
mpol_conf_put() only decrements refcount if the policy has
MPOL_F_SHARED.
In above case, vma1 is not shmem vma and vma->policy has MPOL_F_SHARED!
The reference count will be decreased even though was not increased
whenever alloc_page_vma() is called. This has been broken since commit
[52cd3b07: mempolicy: rework mempolicy Reference Counting] in 2008.
There is another serious bug with the sharing of memory policies.
Currently, mempolicy rebind logic (it is called from cpuset rebinding)
ignores a refcount of mempolicy and override it forcibly. Thus, any
mempolicy sharing may cause mempolicy corruption. The bug was
introduced by commit [68860ec1: cpusets: automatic numa mempolicy
rebinding].
Ideally, the shared policy handling would be rewritten to either
properly handle COW of the policy structures or at least reference count
MPOL_F_SHARED based exclusively on information within the policy.
However, this patch takes the easier approach of disabling any policy
sharing between VMAs. Each new range allocated with sp_alloc will
allocate a new policy, set the reference count to 1 and drop the
reference count of the old policy. This increases the memory footprint
but is not expected to be a major problem as mbind() is unlikely to be
used for fine-grained ranges. It is also inefficient because it means
we allocate a new policy even in cases where mbind_range() could use the
new_policy passed to it. However, it is more straight-forward and the
change should be invisible to the user.
[mgorman@suse.de: Edited changelog]
Reported-by: Dave Jones <davej@redhat.com>,
Cc: Christoph Lameter <cl@linux.com>,
Reviewed-by: Christoph Lameter <cl@linux.com>
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Josh Boyer <jwboyer@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-10-08 16:29:16 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-10-08 16:29:14 -07:00
|
|
|
|
mempolicy: remove mempolicy sharing
Dave Jones' system call fuzz testing tool "trinity" triggered the
following bug error with slab debugging enabled
=============================================================================
BUG numa_policy (Not tainted): Poison overwritten
-----------------------------------------------------------------------------
INFO: 0xffff880146498250-0xffff880146498250. First byte 0x6a instead of 0x6b
INFO: Allocated in mpol_new+0xa3/0x140 age=46310 cpu=6 pid=32154
__slab_alloc+0x3d3/0x445
kmem_cache_alloc+0x29d/0x2b0
mpol_new+0xa3/0x140
sys_mbind+0x142/0x620
system_call_fastpath+0x16/0x1b
INFO: Freed in __mpol_put+0x27/0x30 age=46268 cpu=6 pid=32154
__slab_free+0x2e/0x1de
kmem_cache_free+0x25a/0x260
__mpol_put+0x27/0x30
remove_vma+0x68/0x90
exit_mmap+0x118/0x140
mmput+0x73/0x110
exit_mm+0x108/0x130
do_exit+0x162/0xb90
do_group_exit+0x4f/0xc0
sys_exit_group+0x17/0x20
system_call_fastpath+0x16/0x1b
INFO: Slab 0xffffea0005192600 objects=27 used=27 fp=0x (null) flags=0x20000000004080
INFO: Object 0xffff880146498250 @offset=592 fp=0xffff88014649b9d0
The problem is that the structure is being prematurely freed due to a
reference count imbalance. In the following case mbind(addr, len) should
replace the memory policies of both vma1 and vma2 and thus they will
become to share the same mempolicy and the new mempolicy will have the
MPOL_F_SHARED flag.
+-------------------+-------------------+
| vma1 | vma2(shmem) |
+-------------------+-------------------+
| |
addr addr+len
alloc_pages_vma() uses get_vma_policy() and mpol_cond_put() pair for
maintaining the mempolicy reference count. The current rule is that
get_vma_policy() only increments refcount for shmem VMA and
mpol_conf_put() only decrements refcount if the policy has
MPOL_F_SHARED.
In above case, vma1 is not shmem vma and vma->policy has MPOL_F_SHARED!
The reference count will be decreased even though was not increased
whenever alloc_page_vma() is called. This has been broken since commit
[52cd3b07: mempolicy: rework mempolicy Reference Counting] in 2008.
There is another serious bug with the sharing of memory policies.
Currently, mempolicy rebind logic (it is called from cpuset rebinding)
ignores a refcount of mempolicy and override it forcibly. Thus, any
mempolicy sharing may cause mempolicy corruption. The bug was
introduced by commit [68860ec1: cpusets: automatic numa mempolicy
rebinding].
Ideally, the shared policy handling would be rewritten to either
properly handle COW of the policy structures or at least reference count
MPOL_F_SHARED based exclusively on information within the policy.
However, this patch takes the easier approach of disabling any policy
sharing between VMAs. Each new range allocated with sp_alloc will
allocate a new policy, set the reference count to 1 and drop the
reference count of the old policy. This increases the memory footprint
but is not expected to be a major problem as mbind() is unlikely to be
used for fine-grained ranges. It is also inefficient because it means
we allocate a new policy even in cases where mbind_range() could use the
new_policy passed to it. However, it is more straight-forward and the
change should be invisible to the user.
[mgorman@suse.de: Edited changelog]
Reported-by: Dave Jones <davej@redhat.com>,
Cc: Christoph Lameter <cl@linux.com>,
Reviewed-by: Christoph Lameter <cl@linux.com>
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Josh Boyer <jwboyer@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-10-08 16:29:16 -07:00
|
|
|
|
|
|
|
|
|
2012-10-08 16:29:14 -07:00
|
|
|
|
mempolicy: remove mempolicy sharing
Dave Jones' system call fuzz testing tool "trinity" triggered the
following bug error with slab debugging enabled
=============================================================================
BUG numa_policy (Not tainted): Poison overwritten
-----------------------------------------------------------------------------
INFO: 0xffff880146498250-0xffff880146498250. First byte 0x6a instead of 0x6b
INFO: Allocated in mpol_new+0xa3/0x140 age=46310 cpu=6 pid=32154
__slab_alloc+0x3d3/0x445
kmem_cache_alloc+0x29d/0x2b0
mpol_new+0xa3/0x140
sys_mbind+0x142/0x620
system_call_fastpath+0x16/0x1b
INFO: Freed in __mpol_put+0x27/0x30 age=46268 cpu=6 pid=32154
__slab_free+0x2e/0x1de
kmem_cache_free+0x25a/0x260
__mpol_put+0x27/0x30
remove_vma+0x68/0x90
exit_mmap+0x118/0x140
mmput+0x73/0x110
exit_mm+0x108/0x130
do_exit+0x162/0xb90
do_group_exit+0x4f/0xc0
sys_exit_group+0x17/0x20
system_call_fastpath+0x16/0x1b
INFO: Slab 0xffffea0005192600 objects=27 used=27 fp=0x (null) flags=0x20000000004080
INFO: Object 0xffff880146498250 @offset=592 fp=0xffff88014649b9d0
The problem is that the structure is being prematurely freed due to a
reference count imbalance. In the following case mbind(addr, len) should
replace the memory policies of both vma1 and vma2 and thus they will
become to share the same mempolicy and the new mempolicy will have the
MPOL_F_SHARED flag.
+-------------------+-------------------+
| vma1 | vma2(shmem) |
+-------------------+-------------------+
| |
addr addr+len
alloc_pages_vma() uses get_vma_policy() and mpol_cond_put() pair for
maintaining the mempolicy reference count. The current rule is that
get_vma_policy() only increments refcount for shmem VMA and
mpol_conf_put() only decrements refcount if the policy has
MPOL_F_SHARED.
In above case, vma1 is not shmem vma and vma->policy has MPOL_F_SHARED!
The reference count will be decreased even though was not increased
whenever alloc_page_vma() is called. This has been broken since commit
[52cd3b07: mempolicy: rework mempolicy Reference Counting] in 2008.
There is another serious bug with the sharing of memory policies.
Currently, mempolicy rebind logic (it is called from cpuset rebinding)
ignores a refcount of mempolicy and override it forcibly. Thus, any
mempolicy sharing may cause mempolicy corruption. The bug was
introduced by commit [68860ec1: cpusets: automatic numa mempolicy
rebinding].
Ideally, the shared policy handling would be rewritten to either
properly handle COW of the policy structures or at least reference count
MPOL_F_SHARED based exclusively on information within the policy.
However, this patch takes the easier approach of disabling any policy
sharing between VMAs. Each new range allocated with sp_alloc will
allocate a new policy, set the reference count to 1 and drop the
reference count of the old policy. This increases the memory footprint
but is not expected to be a major problem as mbind() is unlikely to be
used for fine-grained ranges. It is also inefficient because it means
we allocate a new policy even in cases where mbind_range() could use the
new_policy passed to it. However, it is more straight-forward and the
change should be invisible to the user.
[mgorman@suse.de: Edited changelog]
Reported-by: Dave Jones <davej@redhat.com>,
Cc: Christoph Lameter <cl@linux.com>,
Reviewed-by: Christoph Lameter <cl@linux.com>
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Josh Boyer <jwboyer@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-10-08 16:29:16 -07:00
|
|
|
|
|
|
|
|
|
2026-03-13 20:47:56 +08:00
|
|
|
|
mempolicy: remove mempolicy sharing
Dave Jones' system call fuzz testing tool "trinity" triggered the
following bug error with slab debugging enabled
=============================================================================
BUG numa_policy (Not tainted): Poison overwritten
-----------------------------------------------------------------------------
INFO: 0xffff880146498250-0xffff880146498250. First byte 0x6a instead of 0x6b
INFO: Allocated in mpol_new+0xa3/0x140 age=46310 cpu=6 pid=32154
__slab_alloc+0x3d3/0x445
kmem_cache_alloc+0x29d/0x2b0
mpol_new+0xa3/0x140
sys_mbind+0x142/0x620
system_call_fastpath+0x16/0x1b
INFO: Freed in __mpol_put+0x27/0x30 age=46268 cpu=6 pid=32154
__slab_free+0x2e/0x1de
kmem_cache_free+0x25a/0x260
__mpol_put+0x27/0x30
remove_vma+0x68/0x90
exit_mmap+0x118/0x140
mmput+0x73/0x110
exit_mm+0x108/0x130
do_exit+0x162/0xb90
do_group_exit+0x4f/0xc0
sys_exit_group+0x17/0x20
system_call_fastpath+0x16/0x1b
INFO: Slab 0xffffea0005192600 objects=27 used=27 fp=0x (null) flags=0x20000000004080
INFO: Object 0xffff880146498250 @offset=592 fp=0xffff88014649b9d0
The problem is that the structure is being prematurely freed due to a
reference count imbalance. In the following case mbind(addr, len) should
replace the memory policies of both vma1 and vma2 and thus they will
become to share the same mempolicy and the new mempolicy will have the
MPOL_F_SHARED flag.
+-------------------+-------------------+
| vma1 | vma2(shmem) |
+-------------------+-------------------+
| |
addr addr+len
alloc_pages_vma() uses get_vma_policy() and mpol_cond_put() pair for
maintaining the mempolicy reference count. The current rule is that
get_vma_policy() only increments refcount for shmem VMA and
mpol_conf_put() only decrements refcount if the policy has
MPOL_F_SHARED.
In above case, vma1 is not shmem vma and vma->policy has MPOL_F_SHARED!
The reference count will be decreased even though was not increased
whenever alloc_page_vma() is called. This has been broken since commit
[52cd3b07: mempolicy: rework mempolicy Reference Counting] in 2008.
There is another serious bug with the sharing of memory policies.
Currently, mempolicy rebind logic (it is called from cpuset rebinding)
ignores a refcount of mempolicy and override it forcibly. Thus, any
mempolicy sharing may cause mempolicy corruption. The bug was
introduced by commit [68860ec1: cpusets: automatic numa mempolicy
rebinding].
Ideally, the shared policy handling would be rewritten to either
properly handle COW of the policy structures or at least reference count
MPOL_F_SHARED based exclusively on information within the policy.
However, this patch takes the easier approach of disabling any policy
sharing between VMAs. Each new range allocated with sp_alloc will
allocate a new policy, set the reference count to 1 and drop the
reference count of the old policy. This increases the memory footprint
but is not expected to be a major problem as mbind() is unlikely to be
used for fine-grained ranges. It is also inefficient because it means
we allocate a new policy even in cases where mbind_range() could use the
new_policy passed to it. However, it is more straight-forward and the
change should be invisible to the user.
[mgorman@suse.de: Edited changelog]
Reported-by: Dave Jones <davej@redhat.com>,
Cc: Christoph Lameter <cl@linux.com>,
Reviewed-by: Christoph Lameter <cl@linux.com>
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Josh Boyer <jwboyer@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-10-08 16:29:16 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-10-08 16:29:14 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-04-10 11:22:05 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2023-04-10 11:22:05 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2023-04-10 11:22:05 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-03 02:20:14 -07:00
|
|
|
|
mm/mempolicy: correctly update prev when policy is equal on mbind
The refactoring in commit f4e9e0e69468 ("mm/mempolicy: fix use-after-free
of VMA iterator") introduces a subtle bug which arises when attempting to
apply a new NUMA policy across a range of VMAs in mbind_range().
The refactoring passes a **prev pointer to keep track of the previous VMA
in order to reduce duplication, and in all but one case it keeps this
correctly updated.
The bug arises when a VMA within the specified range has an equivalent
policy as determined by mpol_equal() - which unlike other cases, does not
update prev.
This can result in a situation where, later in the iteration, a VMA is
found whose policy does need to change. At this point, vma_merge() is
invoked with prev pointing to a VMA which is before the previous VMA.
Since vma_merge() discovers the curr VMA by looking for the one
immediately after prev, it will now be in a situation where this VMA is
incorrect and the merge will not proceed correctly.
This is checked in the VM_WARN_ON() invariant case with end >
curr->vm_end, which, if a merge is possible, results in a warning (if
CONFIG_DEBUG_VM is specified).
I note that vma_merge() performs these invariant checks only after
merge_prev/merge_next are checked, which is debatable as it hides this
issue if no merge is possible even though a buggy situation has arisen.
The solution is simply to update the prev pointer even when policies are
equal.
This caused a bug to arise in the 6.2.y stable tree, and this patch
resolves this bug.
Link: https://lkml.kernel.org/r/83f1d612acb519d777bebf7f3359317c4e7f4265.1682866629.git.lstoakes@gmail.com
Fixes: f4e9e0e69468 ("mm/mempolicy: fix use-after-free of VMA iterator")
Signed-off-by: Lorenzo Stoakes <lstoakes@gmail.com>
Reported-by: kernel test robot <oliver.sang@intel.com>
Link: https://lore.kernel.org/oe-lkp/202304292203.44ddeff6-oliver.sang@intel.com
Cc: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-04-30 16:07:07 +01:00
|
|
|
|
2022-10-15 02:12:33 +00:00
|
|
|
|
mm/mempolicy: correctly update prev when policy is equal on mbind
The refactoring in commit f4e9e0e69468 ("mm/mempolicy: fix use-after-free
of VMA iterator") introduces a subtle bug which arises when attempting to
apply a new NUMA policy across a range of VMAs in mbind_range().
The refactoring passes a **prev pointer to keep track of the previous VMA
in order to reduce duplication, and in all but one case it keeps this
correctly updated.
The bug arises when a VMA within the specified range has an equivalent
policy as determined by mpol_equal() - which unlike other cases, does not
update prev.
This can result in a situation where, later in the iteration, a VMA is
found whose policy does need to change. At this point, vma_merge() is
invoked with prev pointing to a VMA which is before the previous VMA.
Since vma_merge() discovers the curr VMA by looking for the one
immediately after prev, it will now be in a situation where this VMA is
incorrect and the merge will not proceed correctly.
This is checked in the VM_WARN_ON() invariant case with end >
curr->vm_end, which, if a merge is possible, results in a warning (if
CONFIG_DEBUG_VM is specified).
I note that vma_merge() performs these invariant checks only after
merge_prev/merge_next are checked, which is debatable as it hides this
issue if no merge is possible even though a buggy situation has arisen.
The solution is simply to update the prev pointer even when policies are
equal.
This caused a bug to arise in the 6.2.y stable tree, and this patch
resolves this bug.
Link: https://lkml.kernel.org/r/83f1d612acb519d777bebf7f3359317c4e7f4265.1682866629.git.lstoakes@gmail.com
Fixes: f4e9e0e69468 ("mm/mempolicy: fix use-after-free of VMA iterator")
Signed-off-by: Lorenzo Stoakes <lstoakes@gmail.com>
Reported-by: kernel test robot <oliver.sang@intel.com>
Link: https://lore.kernel.org/oe-lkp/202304292203.44ddeff6-oliver.sang@intel.com
Cc: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-04-30 16:07:07 +01:00
|
|
|
|
2022-10-15 02:12:33 +00:00
|
|
|
|
mm: abstract the vma_merge()/split_vma() pattern for mprotect() et al.
mprotect() and other functions which change VMA parameters over a range
each employ a pattern of:-
1. Attempt to merge the range with adjacent VMAs.
2. If this fails, and the range spans a subset of the VMA, split it
accordingly.
This is open-coded and duplicated in each case. Also in each case most of
the parameters passed to vma_merge() remain the same.
Create a new function, vma_modify(), which abstracts this operation,
accepting only those parameters which can be changed.
To avoid the mess of invoking each function call with unnecessary
parameters, create inline wrapper functions for each of the modify
operations, parameterised only by what is required to perform the action.
We can also significantly simplify the logic - by returning the VMA if we
split (or merged VMA if we do not) we no longer need specific handling for
merge/split cases in any of the call sites.
Note that the userfaultfd_release() case works even though it does not
split VMAs - since start is set to vma->vm_start and end is set to
vma->vm_end, the split logic does not trigger.
In addition, since we calculate pgoff to be equal to vma->vm_pgoff + (start
- vma->vm_start) >> PAGE_SHIFT, and start - vma->vm_start will be 0 in this
instance, this invocation will remain unchanged.
We eliminate a VM_WARN_ON() in mprotect_fixup() as this simply asserts that
vma_merge() correctly ensures that flags remain the same, something that is
already checked in is_mergeable_vma() and elsewhere, and in any case is not
specific to mprotect().
Link: https://lkml.kernel.org/r/0dfa9368f37199a423674bf0ee312e8ea0619044.1697043508.git.lstoakes@gmail.com
Signed-off-by: Lorenzo Stoakes <lstoakes@gmail.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Liam R. Howlett <Liam.Howlett@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-11 18:04:28 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-04-10 11:22:05 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-04-28 02:12:25 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
cpuset,mm: update tasks' mems_allowed in time
Fix allocating page cache/slab object on the unallowed node when memory
spread is set by updating tasks' mems_allowed after its cpuset's mems is
changed.
In order to update tasks' mems_allowed in time, we must modify the code of
memory policy. Because the memory policy is applied in the process's
context originally. After applying this patch, one task directly
manipulates anothers mems_allowed, and we use alloc_lock in the
task_struct to protect mems_allowed and memory policy of the task.
But in the fast path, we didn't use lock to protect them, because adding a
lock may lead to performance regression. But if we don't add a lock,the
task might see no nodes when changing cpuset's mems_allowed to some
non-overlapping set. In order to avoid it, we set all new allowed nodes,
then clear newly disallowed ones.
[lee.schermerhorn@hp.com:
The rework of mpol_new() to extract the adjusting of the node mask to
apply cpuset and mpol flags "context" breaks set_mempolicy() and mbind()
with MPOL_PREFERRED and a NULL nodemask--i.e., explicit local
allocation. Fix this by adding the check for MPOL_PREFERRED and empty
node mask to mpol_new_mpolicy().
Remove the now unneeded 'nodes = NULL' from mpol_new().
Note that mpol_new_mempolicy() is always called with a non-NULL
'nodes' parameter now that it has been removed from mpol_new().
Therefore, we don't need to test nodes for NULL before testing it for
'empty'. However, just to be extra paranoid, add a VM_BUG_ON() to
verify this assumption.]
[lee.schermerhorn@hp.com:
I don't think the function name 'mpol_new_mempolicy' is descriptive
enough to differentiate it from mpol_new().
This function applies cpuset set context, usually constraining nodes
to those allowed by the cpuset. However, when the 'RELATIVE_NODES flag
is set, it also translates the nodes. So I settled on
'mpol_set_nodemask()', because the comment block for mpol_new() mentions
that we need to call this function to "set nodes".
Some additional minor line length, whitespace and typo cleanup.]
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 15:31:49 -07:00
|
|
|
|
mm: make set_mempolicy(MPOL_INTERLEAV) N_HIGH_MEMORY aware
At first, init_task's mems_allowed is initialized as this.
init_task->mems_allowed == node_state[N_POSSIBLE]
And cpuset's top_cpuset mask is initialized as this
top_cpuset->mems_allowed = node_state[N_HIGH_MEMORY]
Before 2.6.29:
policy's mems_allowed is initialized as this.
1. update tasks->mems_allowed by its cpuset->mems_allowed.
2. policy->mems_allowed = nodes_and(tasks->mems_allowed, user's mask)
Updating task's mems_allowed in reference to top_cpuset's one.
cpuset's mems_allowed is aware of N_HIGH_MEMORY, always.
In 2.6.30: After commit 58568d2a8215cb6f55caf2332017d7bdff954e1c
("cpuset,mm: update tasks' mems_allowed in time"), policy's mems_allowed
is initialized as this.
1. policy->mems_allowd = nodes_and(task->mems_allowed, user's mask)
Here, if task is in top_cpuset, task->mems_allowed is not updated from
init's one. Assume user excutes command as #numactrl --interleave=all
,....
policy->mems_allowd = nodes_and(N_POSSIBLE, ALL_SET_MASK)
Then, policy's mems_allowd can includes a possible node, which has no pgdat.
MPOL's INTERLEAVE just scans nodemask of task->mems_allowd and access this
directly.
NODE_DATA(nid)->zonelist even if NODE_DATA(nid)==NULL
Then, what's we need is making policy->mems_allowed be aware of
N_HIGH_MEMORY. This patch does that. But to do so, extra nodemask will
be on statck. Because I know cpumask has a new interface of
CPUMASK_ALLOC(), I added it to node.
This patch stands on old behavior. But I feel this fix itself is just a
Band-Aid. But to do fundametal fix, we have to take care of memory
hotplug and it takes time. (task->mems_allowd should be N_HIGH_MEMORY, I
think.)
mpol_set_nodemask() should be aware of N_HIGH_MEMORY and policy's nodemask
should be includes only online nodes.
In old behavior, this is guaranteed by frequent reference to cpuset's
code. Now, most of them are removed and mempolicy has to check it by
itself.
To do check, a few nodemask_t will be used for calculating nodemask. But,
size of nodemask_t can be big and it's not good to allocate them on stack.
Now, cpumask_t has CPUMASK_ALLOC/FREE an easy code for get scratch area.
NODEMASK_ALLOC/FREE shoudl be there.
[akpm@linux-foundation.org: cleanups & tweaks]
Tested-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-08-06 15:07:33 -07:00
|
|
|
|
cpuset,mm: update tasks' mems_allowed in time
Fix allocating page cache/slab object on the unallowed node when memory
spread is set by updating tasks' mems_allowed after its cpuset's mems is
changed.
In order to update tasks' mems_allowed in time, we must modify the code of
memory policy. Because the memory policy is applied in the process's
context originally. After applying this patch, one task directly
manipulates anothers mems_allowed, and we use alloc_lock in the
task_struct to protect mems_allowed and memory policy of the task.
But in the fast path, we didn't use lock to protect them, because adding a
lock may lead to performance regression. But if we don't add a lock,the
task might see no nodes when changing cpuset's mems_allowed to some
non-overlapping set. In order to avoid it, we set all new allowed nodes,
then clear newly disallowed ones.
[lee.schermerhorn@hp.com:
The rework of mpol_new() to extract the adjusting of the node mask to
apply cpuset and mpol flags "context" breaks set_mempolicy() and mbind()
with MPOL_PREFERRED and a NULL nodemask--i.e., explicit local
allocation. Fix this by adding the check for MPOL_PREFERRED and empty
node mask to mpol_new_mpolicy().
Remove the now unneeded 'nodes = NULL' from mpol_new().
Note that mpol_new_mempolicy() is always called with a non-NULL
'nodes' parameter now that it has been removed from mpol_new().
Therefore, we don't need to test nodes for NULL before testing it for
'empty'. However, just to be extra paranoid, add a VM_BUG_ON() to
verify this assumption.]
[lee.schermerhorn@hp.com:
I don't think the function name 'mpol_new_mempolicy' is descriptive
enough to differentiate it from mpol_new().
This function applies cpuset set context, usually constraining nodes
to those allowed by the cpuset. However, when the 'RELATIVE_NODES flag
is set, it also translates the nodes. So I settled on
'mpol_set_nodemask()', because the comment block for mpol_new() mentions
that we need to call this function to "set nodes".
Some additional minor line length, whitespace and typo cleanup.]
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 15:31:49 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
mm: make set_mempolicy(MPOL_INTERLEAV) N_HIGH_MEMORY aware
At first, init_task's mems_allowed is initialized as this.
init_task->mems_allowed == node_state[N_POSSIBLE]
And cpuset's top_cpuset mask is initialized as this
top_cpuset->mems_allowed = node_state[N_HIGH_MEMORY]
Before 2.6.29:
policy's mems_allowed is initialized as this.
1. update tasks->mems_allowed by its cpuset->mems_allowed.
2. policy->mems_allowed = nodes_and(tasks->mems_allowed, user's mask)
Updating task's mems_allowed in reference to top_cpuset's one.
cpuset's mems_allowed is aware of N_HIGH_MEMORY, always.
In 2.6.30: After commit 58568d2a8215cb6f55caf2332017d7bdff954e1c
("cpuset,mm: update tasks' mems_allowed in time"), policy's mems_allowed
is initialized as this.
1. policy->mems_allowd = nodes_and(task->mems_allowed, user's mask)
Here, if task is in top_cpuset, task->mems_allowed is not updated from
init's one. Assume user excutes command as #numactrl --interleave=all
,....
policy->mems_allowd = nodes_and(N_POSSIBLE, ALL_SET_MASK)
Then, policy's mems_allowd can includes a possible node, which has no pgdat.
MPOL's INTERLEAVE just scans nodemask of task->mems_allowd and access this
directly.
NODE_DATA(nid)->zonelist even if NODE_DATA(nid)==NULL
Then, what's we need is making policy->mems_allowed be aware of
N_HIGH_MEMORY. This patch does that. But to do so, extra nodemask will
be on statck. Because I know cpumask has a new interface of
CPUMASK_ALLOC(), I added it to node.
This patch stands on old behavior. But I feel this fix itself is just a
Band-Aid. But to do fundametal fix, we have to take care of memory
hotplug and it takes time. (task->mems_allowd should be N_HIGH_MEMORY, I
think.)
mpol_set_nodemask() should be aware of N_HIGH_MEMORY and policy's nodemask
should be includes only online nodes.
In old behavior, this is guaranteed by frequent reference to cpuset's
code. Now, most of them are removed and mempolicy has to check it by
itself.
To do check, a few nodemask_t will be used for calculating nodemask. But,
size of nodemask_t can be big and it's not good to allocate them on stack.
Now, cpumask_t has CPUMASK_ALLOC/FREE an easy code for get scratch area.
NODEMASK_ALLOC/FREE shoudl be there.
[akpm@linux-foundation.org: cleanups & tweaks]
Tested-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-08-06 15:07:33 -07:00
|
|
|
|
|
|
|
|
|
2008-04-28 02:13:10 -07:00
|
|
|
|
mm: make set_mempolicy(MPOL_INTERLEAV) N_HIGH_MEMORY aware
At first, init_task's mems_allowed is initialized as this.
init_task->mems_allowed == node_state[N_POSSIBLE]
And cpuset's top_cpuset mask is initialized as this
top_cpuset->mems_allowed = node_state[N_HIGH_MEMORY]
Before 2.6.29:
policy's mems_allowed is initialized as this.
1. update tasks->mems_allowed by its cpuset->mems_allowed.
2. policy->mems_allowed = nodes_and(tasks->mems_allowed, user's mask)
Updating task's mems_allowed in reference to top_cpuset's one.
cpuset's mems_allowed is aware of N_HIGH_MEMORY, always.
In 2.6.30: After commit 58568d2a8215cb6f55caf2332017d7bdff954e1c
("cpuset,mm: update tasks' mems_allowed in time"), policy's mems_allowed
is initialized as this.
1. policy->mems_allowd = nodes_and(task->mems_allowed, user's mask)
Here, if task is in top_cpuset, task->mems_allowed is not updated from
init's one. Assume user excutes command as #numactrl --interleave=all
,....
policy->mems_allowd = nodes_and(N_POSSIBLE, ALL_SET_MASK)
Then, policy's mems_allowd can includes a possible node, which has no pgdat.
MPOL's INTERLEAVE just scans nodemask of task->mems_allowd and access this
directly.
NODE_DATA(nid)->zonelist even if NODE_DATA(nid)==NULL
Then, what's we need is making policy->mems_allowed be aware of
N_HIGH_MEMORY. This patch does that. But to do so, extra nodemask will
be on statck. Because I know cpumask has a new interface of
CPUMASK_ALLOC(), I added it to node.
This patch stands on old behavior. But I feel this fix itself is just a
Band-Aid. But to do fundametal fix, we have to take care of memory
hotplug and it takes time. (task->mems_allowd should be N_HIGH_MEMORY, I
think.)
mpol_set_nodemask() should be aware of N_HIGH_MEMORY and policy's nodemask
should be includes only online nodes.
In old behavior, this is guaranteed by frequent reference to cpuset's
code. Now, most of them are removed and mempolicy has to check it by
itself.
To do check, a few nodemask_t will be used for calculating nodemask. But,
size of nodemask_t can be big and it's not good to allocate them on stack.
Now, cpumask_t has CPUMASK_ALLOC/FREE an easy code for get scratch area.
NODEMASK_ALLOC/FREE shoudl be there.
[akpm@linux-foundation.org: cleanups & tweaks]
Tested-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-08-06 15:07:33 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-10-09 15:27:55 -07:00
|
|
|
|
mm/mempolicy: fix lock contention on mems_allowed
The mems_allowed field can be modified by other tasks, so it isn't safe to
access it with alloc_lock unlocked even in the current process context.
Say there are two tasks: A from cpusetA is performing set_mempolicy(2),
and B is changing cpusetA's cpuset.mems:
A (set_mempolicy) B (echo xx > cpuset.mems)
-------------------------------------------------------
pol = mpol_new();
update_tasks_nodemask(cpusetA) {
foreach t in cpusetA {
cpuset_change_task_nodemask(t) {
mpol_set_nodemask(pol) {
task_lock(t); // t could be A
new = f(A->mems_allowed);
update t->mems_allowed;
pol.create(pol, new);
task_unlock(t);
}
}
}
}
task_lock(A);
A->mempolicy = pol;
task_unlock(A);
In this case A's pol->nodes is computed by old mems_allowed, and could
be inconsistent with A's new mems_allowed.
While it is different when replacing vmas' policy: the pol->nodes is
gone wild only when current_cpuset_is_being_rebound():
A (mbind) B (echo xx > cpuset.mems)
-------------------------------------------------------
pol = mpol_new();
mmap_write_lock(A->mm);
cpuset_being_rebound = cpusetA;
update_tasks_nodemask(cpusetA) {
foreach t in cpusetA {
cpuset_change_task_nodemask(t) {
mpol_set_nodemask(pol) {
task_lock(t); // t could be A
mask = f(A->mems_allowed);
update t->mems_allowed;
pol.create(pol, mask);
task_unlock(t);
}
}
foreach v in A->mm {
if (cpuset_being_rebound == cpusetA)
pol.rebind(pol, cpuset.mems);
v->vma_policy = pol;
}
mmap_write_unlock(A->mm);
mmap_write_lock(t->mm);
mpol_rebind_mm(t->mm);
mmap_write_unlock(t->mm);
}
}
cpuset_being_rebound = NULL;
In this case, the cpuset.mems, which has already done updating, is finally
used for calculating pol->nodes, rather than A->mems_allowed. So it is OK
to call mpol_set_nodemask() with alloc_lock unlocked when doing mbind(2).
Link: https://lkml.kernel.org/r/20220811124157.74888-1-wuyun.abel@bytedance.com
Fixes: 78b132e9bae9 ("mm/mempolicy: remove or narrow the lock on current")
Signed-off-by: Abel Wu <wuyun.abel@bytedance.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Wei Yang <richard.weiyang@gmail.com>
Reviewed-by: Muchun Song <songmuchun@bytedance.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-08-11 20:41:57 +08:00
|
|
|
|
mm: make set_mempolicy(MPOL_INTERLEAV) N_HIGH_MEMORY aware
At first, init_task's mems_allowed is initialized as this.
init_task->mems_allowed == node_state[N_POSSIBLE]
And cpuset's top_cpuset mask is initialized as this
top_cpuset->mems_allowed = node_state[N_HIGH_MEMORY]
Before 2.6.29:
policy's mems_allowed is initialized as this.
1. update tasks->mems_allowed by its cpuset->mems_allowed.
2. policy->mems_allowed = nodes_and(tasks->mems_allowed, user's mask)
Updating task's mems_allowed in reference to top_cpuset's one.
cpuset's mems_allowed is aware of N_HIGH_MEMORY, always.
In 2.6.30: After commit 58568d2a8215cb6f55caf2332017d7bdff954e1c
("cpuset,mm: update tasks' mems_allowed in time"), policy's mems_allowed
is initialized as this.
1. policy->mems_allowd = nodes_and(task->mems_allowed, user's mask)
Here, if task is in top_cpuset, task->mems_allowed is not updated from
init's one. Assume user excutes command as #numactrl --interleave=all
,....
policy->mems_allowd = nodes_and(N_POSSIBLE, ALL_SET_MASK)
Then, policy's mems_allowd can includes a possible node, which has no pgdat.
MPOL's INTERLEAVE just scans nodemask of task->mems_allowd and access this
directly.
NODE_DATA(nid)->zonelist even if NODE_DATA(nid)==NULL
Then, what's we need is making policy->mems_allowed be aware of
N_HIGH_MEMORY. This patch does that. But to do so, extra nodemask will
be on statck. Because I know cpumask has a new interface of
CPUMASK_ALLOC(), I added it to node.
This patch stands on old behavior. But I feel this fix itself is just a
Band-Aid. But to do fundametal fix, we have to take care of memory
hotplug and it takes time. (task->mems_allowd should be N_HIGH_MEMORY, I
think.)
mpol_set_nodemask() should be aware of N_HIGH_MEMORY and policy's nodemask
should be includes only online nodes.
In old behavior, this is guaranteed by frequent reference to cpuset's
code. Now, most of them are removed and mempolicy has to check it by
itself.
To do check, a few nodemask_t will be used for calculating nodemask. But,
size of nodemask_t can be big and it's not good to allocate them on stack.
Now, cpumask_t has CPUMASK_ALLOC/FREE an easy code for get scratch area.
NODEMASK_ALLOC/FREE shoudl be there.
[akpm@linux-foundation.org: cleanups & tweaks]
Tested-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-08-06 15:07:33 -07:00
|
|
|
|
cpuset,mm: update tasks' mems_allowed in time
Fix allocating page cache/slab object on the unallowed node when memory
spread is set by updating tasks' mems_allowed after its cpuset's mems is
changed.
In order to update tasks' mems_allowed in time, we must modify the code of
memory policy. Because the memory policy is applied in the process's
context originally. After applying this patch, one task directly
manipulates anothers mems_allowed, and we use alloc_lock in the
task_struct to protect mems_allowed and memory policy of the task.
But in the fast path, we didn't use lock to protect them, because adding a
lock may lead to performance regression. But if we don't add a lock,the
task might see no nodes when changing cpuset's mems_allowed to some
non-overlapping set. In order to avoid it, we set all new allowed nodes,
then clear newly disallowed ones.
[lee.schermerhorn@hp.com:
The rework of mpol_new() to extract the adjusting of the node mask to
apply cpuset and mpol flags "context" breaks set_mempolicy() and mbind()
with MPOL_PREFERRED and a NULL nodemask--i.e., explicit local
allocation. Fix this by adding the check for MPOL_PREFERRED and empty
node mask to mpol_new_mpolicy().
Remove the now unneeded 'nodes = NULL' from mpol_new().
Note that mpol_new_mempolicy() is always called with a non-NULL
'nodes' parameter now that it has been removed from mpol_new().
Therefore, we don't need to test nodes for NULL before testing it for
'empty'. However, just to be extra paranoid, add a VM_BUG_ON() to
verify this assumption.]
[lee.schermerhorn@hp.com:
I don't think the function name 'mpol_new_mempolicy' is descriptive
enough to differentiate it from mpol_new().
This function applies cpuset set context, usually constraining nodes
to those allowed by the cpuset. However, when the 'RELATIVE_NODES flag
is set, it also translates the nodes. So I settled on
'mpol_set_nodemask()', because the comment block for mpol_new() mentions
that we need to call this function to "set nodes".
Some additional minor line length, whitespace and typo cleanup.]
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 15:31:49 -07:00
|
|
|
|
mm/mempolicy: fix lock contention on mems_allowed
The mems_allowed field can be modified by other tasks, so it isn't safe to
access it with alloc_lock unlocked even in the current process context.
Say there are two tasks: A from cpusetA is performing set_mempolicy(2),
and B is changing cpusetA's cpuset.mems:
A (set_mempolicy) B (echo xx > cpuset.mems)
-------------------------------------------------------
pol = mpol_new();
update_tasks_nodemask(cpusetA) {
foreach t in cpusetA {
cpuset_change_task_nodemask(t) {
mpol_set_nodemask(pol) {
task_lock(t); // t could be A
new = f(A->mems_allowed);
update t->mems_allowed;
pol.create(pol, new);
task_unlock(t);
}
}
}
}
task_lock(A);
A->mempolicy = pol;
task_unlock(A);
In this case A's pol->nodes is computed by old mems_allowed, and could
be inconsistent with A's new mems_allowed.
While it is different when replacing vmas' policy: the pol->nodes is
gone wild only when current_cpuset_is_being_rebound():
A (mbind) B (echo xx > cpuset.mems)
-------------------------------------------------------
pol = mpol_new();
mmap_write_lock(A->mm);
cpuset_being_rebound = cpusetA;
update_tasks_nodemask(cpusetA) {
foreach t in cpusetA {
cpuset_change_task_nodemask(t) {
mpol_set_nodemask(pol) {
task_lock(t); // t could be A
mask = f(A->mems_allowed);
update t->mems_allowed;
pol.create(pol, mask);
task_unlock(t);
}
}
foreach v in A->mm {
if (cpuset_being_rebound == cpusetA)
pol.rebind(pol, cpuset.mems);
v->vma_policy = pol;
}
mmap_write_unlock(A->mm);
mmap_write_lock(t->mm);
mpol_rebind_mm(t->mm);
mmap_write_unlock(t->mm);
}
}
cpuset_being_rebound = NULL;
In this case, the cpuset.mems, which has already done updating, is finally
used for calculating pol->nodes, rather than A->mems_allowed. So it is OK
to call mpol_set_nodemask() with alloc_lock unlocked when doing mbind(2).
Link: https://lkml.kernel.org/r/20220811124157.74888-1-wuyun.abel@bytedance.com
Fixes: 78b132e9bae9 ("mm/mempolicy: remove or narrow the lock on current")
Signed-off-by: Abel Wu <wuyun.abel@bytedance.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Wei Yang <richard.weiyang@gmail.com>
Reviewed-by: Muchun Song <songmuchun@bytedance.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-08-11 20:41:57 +08:00
|
|
|
|
cpuset,mm: update tasks' mems_allowed in time
Fix allocating page cache/slab object on the unallowed node when memory
spread is set by updating tasks' mems_allowed after its cpuset's mems is
changed.
In order to update tasks' mems_allowed in time, we must modify the code of
memory policy. Because the memory policy is applied in the process's
context originally. After applying this patch, one task directly
manipulates anothers mems_allowed, and we use alloc_lock in the
task_struct to protect mems_allowed and memory policy of the task.
But in the fast path, we didn't use lock to protect them, because adding a
lock may lead to performance regression. But if we don't add a lock,the
task might see no nodes when changing cpuset's mems_allowed to some
non-overlapping set. In order to avoid it, we set all new allowed nodes,
then clear newly disallowed ones.
[lee.schermerhorn@hp.com:
The rework of mpol_new() to extract the adjusting of the node mask to
apply cpuset and mpol flags "context" breaks set_mempolicy() and mbind()
with MPOL_PREFERRED and a NULL nodemask--i.e., explicit local
allocation. Fix this by adding the check for MPOL_PREFERRED and empty
node mask to mpol_new_mpolicy().
Remove the now unneeded 'nodes = NULL' from mpol_new().
Note that mpol_new_mempolicy() is always called with a non-NULL
'nodes' parameter now that it has been removed from mpol_new().
Therefore, we don't need to test nodes for NULL before testing it for
'empty'. However, just to be extra paranoid, add a VM_BUG_ON() to
verify this assumption.]
[lee.schermerhorn@hp.com:
I don't think the function name 'mpol_new_mempolicy' is descriptive
enough to differentiate it from mpol_new().
This function applies cpuset set context, usually constraining nodes
to those allowed by the cpuset. However, when the 'RELATIVE_NODES flag
is set, it also translates the nodes. So I settled on
'mpol_set_nodemask()', because the comment block for mpol_new() mentions
that we need to call this function to "set nodes".
Some additional minor line length, whitespace and typo cleanup.]
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 15:31:49 -07:00
|
|
|
|
mm: make set_mempolicy(MPOL_INTERLEAV) N_HIGH_MEMORY aware
At first, init_task's mems_allowed is initialized as this.
init_task->mems_allowed == node_state[N_POSSIBLE]
And cpuset's top_cpuset mask is initialized as this
top_cpuset->mems_allowed = node_state[N_HIGH_MEMORY]
Before 2.6.29:
policy's mems_allowed is initialized as this.
1. update tasks->mems_allowed by its cpuset->mems_allowed.
2. policy->mems_allowed = nodes_and(tasks->mems_allowed, user's mask)
Updating task's mems_allowed in reference to top_cpuset's one.
cpuset's mems_allowed is aware of N_HIGH_MEMORY, always.
In 2.6.30: After commit 58568d2a8215cb6f55caf2332017d7bdff954e1c
("cpuset,mm: update tasks' mems_allowed in time"), policy's mems_allowed
is initialized as this.
1. policy->mems_allowd = nodes_and(task->mems_allowed, user's mask)
Here, if task is in top_cpuset, task->mems_allowed is not updated from
init's one. Assume user excutes command as #numactrl --interleave=all
,....
policy->mems_allowd = nodes_and(N_POSSIBLE, ALL_SET_MASK)
Then, policy's mems_allowd can includes a possible node, which has no pgdat.
MPOL's INTERLEAVE just scans nodemask of task->mems_allowd and access this
directly.
NODE_DATA(nid)->zonelist even if NODE_DATA(nid)==NULL
Then, what's we need is making policy->mems_allowed be aware of
N_HIGH_MEMORY. This patch does that. But to do so, extra nodemask will
be on statck. Because I know cpumask has a new interface of
CPUMASK_ALLOC(), I added it to node.
This patch stands on old behavior. But I feel this fix itself is just a
Band-Aid. But to do fundametal fix, we have to take care of memory
hotplug and it takes time. (task->mems_allowd should be N_HIGH_MEMORY, I
think.)
mpol_set_nodemask() should be aware of N_HIGH_MEMORY and policy's nodemask
should be includes only online nodes.
In old behavior, this is guaranteed by frequent reference to cpuset's
code. Now, most of them are removed and mempolicy has to check it by
itself.
To do check, a few nodemask_t will be used for calculating nodemask. But,
size of nodemask_t can be big and it's not good to allocate them on stack.
Now, cpumask_t has CPUMASK_ALLOC/FREE an easy code for get scratch area.
NODEMASK_ALLOC/FREE shoudl be there.
[akpm@linux-foundation.org: cleanups & tweaks]
Tested-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-08-06 15:07:33 -07:00
|
|
|
|
cpuset,mm: update tasks' mems_allowed in time
Fix allocating page cache/slab object on the unallowed node when memory
spread is set by updating tasks' mems_allowed after its cpuset's mems is
changed.
In order to update tasks' mems_allowed in time, we must modify the code of
memory policy. Because the memory policy is applied in the process's
context originally. After applying this patch, one task directly
manipulates anothers mems_allowed, and we use alloc_lock in the
task_struct to protect mems_allowed and memory policy of the task.
But in the fast path, we didn't use lock to protect them, because adding a
lock may lead to performance regression. But if we don't add a lock,the
task might see no nodes when changing cpuset's mems_allowed to some
non-overlapping set. In order to avoid it, we set all new allowed nodes,
then clear newly disallowed ones.
[lee.schermerhorn@hp.com:
The rework of mpol_new() to extract the adjusting of the node mask to
apply cpuset and mpol flags "context" breaks set_mempolicy() and mbind()
with MPOL_PREFERRED and a NULL nodemask--i.e., explicit local
allocation. Fix this by adding the check for MPOL_PREFERRED and empty
node mask to mpol_new_mpolicy().
Remove the now unneeded 'nodes = NULL' from mpol_new().
Note that mpol_new_mempolicy() is always called with a non-NULL
'nodes' parameter now that it has been removed from mpol_new().
Therefore, we don't need to test nodes for NULL before testing it for
'empty'. However, just to be extra paranoid, add a VM_BUG_ON() to
verify this assumption.]
[lee.schermerhorn@hp.com:
I don't think the function name 'mpol_new_mempolicy' is descriptive
enough to differentiate it from mpol_new().
This function applies cpuset set context, usually constraining nodes
to those allowed by the cpuset. However, when the 'RELATIVE_NODES flag
is set, it also translates the nodes. So I settled on
'mpol_set_nodemask()', because the comment block for mpol_new() mentions
that we need to call this function to "set nodes".
Some additional minor line length, whitespace and typo cleanup.]
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 15:31:49 -07:00
|
|
|
|
mm/mempolicy: fix lock contention on mems_allowed
The mems_allowed field can be modified by other tasks, so it isn't safe to
access it with alloc_lock unlocked even in the current process context.
Say there are two tasks: A from cpusetA is performing set_mempolicy(2),
and B is changing cpusetA's cpuset.mems:
A (set_mempolicy) B (echo xx > cpuset.mems)
-------------------------------------------------------
pol = mpol_new();
update_tasks_nodemask(cpusetA) {
foreach t in cpusetA {
cpuset_change_task_nodemask(t) {
mpol_set_nodemask(pol) {
task_lock(t); // t could be A
new = f(A->mems_allowed);
update t->mems_allowed;
pol.create(pol, new);
task_unlock(t);
}
}
}
}
task_lock(A);
A->mempolicy = pol;
task_unlock(A);
In this case A's pol->nodes is computed by old mems_allowed, and could
be inconsistent with A's new mems_allowed.
While it is different when replacing vmas' policy: the pol->nodes is
gone wild only when current_cpuset_is_being_rebound():
A (mbind) B (echo xx > cpuset.mems)
-------------------------------------------------------
pol = mpol_new();
mmap_write_lock(A->mm);
cpuset_being_rebound = cpusetA;
update_tasks_nodemask(cpusetA) {
foreach t in cpusetA {
cpuset_change_task_nodemask(t) {
mpol_set_nodemask(pol) {
task_lock(t); // t could be A
mask = f(A->mems_allowed);
update t->mems_allowed;
pol.create(pol, mask);
task_unlock(t);
}
}
foreach v in A->mm {
if (cpuset_being_rebound == cpusetA)
pol.rebind(pol, cpuset.mems);
v->vma_policy = pol;
}
mmap_write_unlock(A->mm);
mmap_write_lock(t->mm);
mpol_rebind_mm(t->mm);
mmap_write_unlock(t->mm);
}
}
cpuset_being_rebound = NULL;
In this case, the cpuset.mems, which has already done updating, is finally
used for calculating pol->nodes, rather than A->mems_allowed. So it is OK
to call mpol_set_nodemask() with alloc_lock unlocked when doing mbind(2).
Link: https://lkml.kernel.org/r/20220811124157.74888-1-wuyun.abel@bytedance.com
Fixes: 78b132e9bae9 ("mm/mempolicy: remove or narrow the lock on current")
Signed-off-by: Abel Wu <wuyun.abel@bytedance.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Wei Yang <richard.weiyang@gmail.com>
Reviewed-by: Muchun Song <songmuchun@bytedance.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-08-11 20:41:57 +08:00
|
|
|
|
cpuset,mm: update tasks' mems_allowed in time
Fix allocating page cache/slab object on the unallowed node when memory
spread is set by updating tasks' mems_allowed after its cpuset's mems is
changed.
In order to update tasks' mems_allowed in time, we must modify the code of
memory policy. Because the memory policy is applied in the process's
context originally. After applying this patch, one task directly
manipulates anothers mems_allowed, and we use alloc_lock in the
task_struct to protect mems_allowed and memory policy of the task.
But in the fast path, we didn't use lock to protect them, because adding a
lock may lead to performance regression. But if we don't add a lock,the
task might see no nodes when changing cpuset's mems_allowed to some
non-overlapping set. In order to avoid it, we set all new allowed nodes,
then clear newly disallowed ones.
[lee.schermerhorn@hp.com:
The rework of mpol_new() to extract the adjusting of the node mask to
apply cpuset and mpol flags "context" breaks set_mempolicy() and mbind()
with MPOL_PREFERRED and a NULL nodemask--i.e., explicit local
allocation. Fix this by adding the check for MPOL_PREFERRED and empty
node mask to mpol_new_mpolicy().
Remove the now unneeded 'nodes = NULL' from mpol_new().
Note that mpol_new_mempolicy() is always called with a non-NULL
'nodes' parameter now that it has been removed from mpol_new().
Therefore, we don't need to test nodes for NULL before testing it for
'empty'. However, just to be extra paranoid, add a VM_BUG_ON() to
verify this assumption.]
[lee.schermerhorn@hp.com:
I don't think the function name 'mpol_new_mempolicy' is descriptive
enough to differentiate it from mpol_new().
This function applies cpuset set context, usually constraining nodes
to those allowed by the cpuset. However, when the 'RELATIVE_NODES flag
is set, it also translates the nodes. So I settled on
'mpol_set_nodemask()', because the comment block for mpol_new() mentions
that we need to call this function to "set nodes".
Some additional minor line length, whitespace and typo cleanup.]
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 15:31:49 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
|
|
|
|
|
2017-07-06 15:39:59 -07:00
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
|
|
|
|
|
cpuset,mm: update tasks' mems_allowed in time
Fix allocating page cache/slab object on the unallowed node when memory
spread is set by updating tasks' mems_allowed after its cpuset's mems is
changed.
In order to update tasks' mems_allowed in time, we must modify the code of
memory policy. Because the memory policy is applied in the process's
context originally. After applying this patch, one task directly
manipulates anothers mems_allowed, and we use alloc_lock in the
task_struct to protect mems_allowed and memory policy of the task.
But in the fast path, we didn't use lock to protect them, because adding a
lock may lead to performance regression. But if we don't add a lock,the
task might see no nodes when changing cpuset's mems_allowed to some
non-overlapping set. In order to avoid it, we set all new allowed nodes,
then clear newly disallowed ones.
[lee.schermerhorn@hp.com:
The rework of mpol_new() to extract the adjusting of the node mask to
apply cpuset and mpol flags "context" breaks set_mempolicy() and mbind()
with MPOL_PREFERRED and a NULL nodemask--i.e., explicit local
allocation. Fix this by adding the check for MPOL_PREFERRED and empty
node mask to mpol_new_mpolicy().
Remove the now unneeded 'nodes = NULL' from mpol_new().
Note that mpol_new_mempolicy() is always called with a non-NULL
'nodes' parameter now that it has been removed from mpol_new().
Therefore, we don't need to test nodes for NULL before testing it for
'empty'. However, just to be extra paranoid, add a VM_BUG_ON() to
verify this assumption.]
[lee.schermerhorn@hp.com:
I don't think the function name 'mpol_new_mempolicy' is descriptive
enough to differentiate it from mpol_new().
This function applies cpuset set context, usually constraining nodes
to those allowed by the cpuset. However, when the 'RELATIVE_NODES flag
is set, it also translates the nodes. So I settled on
'mpol_set_nodemask()', because the comment block for mpol_new() mentions
that we need to call this function to "set nodes".
Some additional minor line length, whitespace and typo cleanup.]
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 15:31:49 -07:00
|
|
|
|
|
|
|
|
|
mm: make set_mempolicy(MPOL_INTERLEAV) N_HIGH_MEMORY aware
At first, init_task's mems_allowed is initialized as this.
init_task->mems_allowed == node_state[N_POSSIBLE]
And cpuset's top_cpuset mask is initialized as this
top_cpuset->mems_allowed = node_state[N_HIGH_MEMORY]
Before 2.6.29:
policy's mems_allowed is initialized as this.
1. update tasks->mems_allowed by its cpuset->mems_allowed.
2. policy->mems_allowed = nodes_and(tasks->mems_allowed, user's mask)
Updating task's mems_allowed in reference to top_cpuset's one.
cpuset's mems_allowed is aware of N_HIGH_MEMORY, always.
In 2.6.30: After commit 58568d2a8215cb6f55caf2332017d7bdff954e1c
("cpuset,mm: update tasks' mems_allowed in time"), policy's mems_allowed
is initialized as this.
1. policy->mems_allowd = nodes_and(task->mems_allowed, user's mask)
Here, if task is in top_cpuset, task->mems_allowed is not updated from
init's one. Assume user excutes command as #numactrl --interleave=all
,....
policy->mems_allowd = nodes_and(N_POSSIBLE, ALL_SET_MASK)
Then, policy's mems_allowd can includes a possible node, which has no pgdat.
MPOL's INTERLEAVE just scans nodemask of task->mems_allowd and access this
directly.
NODE_DATA(nid)->zonelist even if NODE_DATA(nid)==NULL
Then, what's we need is making policy->mems_allowed be aware of
N_HIGH_MEMORY. This patch does that. But to do so, extra nodemask will
be on statck. Because I know cpumask has a new interface of
CPUMASK_ALLOC(), I added it to node.
This patch stands on old behavior. But I feel this fix itself is just a
Band-Aid. But to do fundametal fix, we have to take care of memory
hotplug and it takes time. (task->mems_allowd should be N_HIGH_MEMORY, I
think.)
mpol_set_nodemask() should be aware of N_HIGH_MEMORY and policy's nodemask
should be includes only online nodes.
In old behavior, this is guaranteed by frequent reference to cpuset's
code. Now, most of them are removed and mempolicy has to check it by
itself.
To do check, a few nodemask_t will be used for calculating nodemask. But,
size of nodemask_t can be big and it's not good to allocate them on stack.
Now, cpumask_t has CPUMASK_ALLOC/FREE an easy code for get scratch area.
NODEMASK_ALLOC/FREE shoudl be there.
[akpm@linux-foundation.org: cleanups & tweaks]
Tested-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-08-06 15:07:33 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
mempolicy: use MPOL_PREFERRED for system-wide default policy
Currently, when one specifies MPOL_DEFAULT via a NUMA memory policy API
[set_mempolicy(), mbind() and internal versions], the kernel simply installs a
NULL struct mempolicy pointer in the appropriate context: task policy, vma
policy, or shared policy. This causes any use of that policy to "fall back"
to the next most specific policy scope.
The only use of MPOL_DEFAULT to mean "local allocation" is in the system
default policy. This requires extra checks/cases for MPOL_DEFAULT in many
mempolicy.c functions.
There is another, "preferred" way to specify local allocation via the APIs.
That is using the MPOL_PREFERRED policy mode with an empty nodemask.
Internally, the empty nodemask gets converted to a preferred_node id of '-1'.
All internal usage of MPOL_PREFERRED will convert the '-1' to the id of the
node local to the cpu where the allocation occurs.
System default policy, except during boot, is hard-coded to "local
allocation". By using the MPOL_PREFERRED mode with a negative value of
preferred node for system default policy, MPOL_DEFAULT will never occur in the
'policy' member of a struct mempolicy. Thus, we can remove all checks for
MPOL_DEFAULT when converting policy to a node id/zonelist in the allocation
paths.
In slab_node() return local node id when policy pointer is NULL. No need to
set a pol value to take the switch default. Replace switch default with
BUG()--i.e., shouldn't happen.
With this patch MPOL_DEFAULT is only used in the APIs, including internal
calls to do_set_mempolicy() and in the display of policy in
/proc/<pid>/numa_maps. It always means "fall back" to the the next most
specific policy scope. This simplifies the description of memory policies
quite a bit, with no visible change in behavior.
get_mempolicy() continues to return MPOL_DEFAULT and an empty nodemask when
the requested policy [task or vma/shared] is NULL. These are the values one
would supply via set_mempolicy() or mbind() to achieve that condition--default
behavior.
This patch updates Documentation to reflect this change.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:18 -07:00
|
|
|
|
|
|
|
|
|
cpuset,mm: update tasks' mems_allowed in time
Fix allocating page cache/slab object on the unallowed node when memory
spread is set by updating tasks' mems_allowed after its cpuset's mems is
changed.
In order to update tasks' mems_allowed in time, we must modify the code of
memory policy. Because the memory policy is applied in the process's
context originally. After applying this patch, one task directly
manipulates anothers mems_allowed, and we use alloc_lock in the
task_struct to protect mems_allowed and memory policy of the task.
But in the fast path, we didn't use lock to protect them, because adding a
lock may lead to performance regression. But if we don't add a lock,the
task might see no nodes when changing cpuset's mems_allowed to some
non-overlapping set. In order to avoid it, we set all new allowed nodes,
then clear newly disallowed ones.
[lee.schermerhorn@hp.com:
The rework of mpol_new() to extract the adjusting of the node mask to
apply cpuset and mpol flags "context" breaks set_mempolicy() and mbind()
with MPOL_PREFERRED and a NULL nodemask--i.e., explicit local
allocation. Fix this by adding the check for MPOL_PREFERRED and empty
node mask to mpol_new_mpolicy().
Remove the now unneeded 'nodes = NULL' from mpol_new().
Note that mpol_new_mempolicy() is always called with a non-NULL
'nodes' parameter now that it has been removed from mpol_new().
Therefore, we don't need to test nodes for NULL before testing it for
'empty'. However, just to be extra paranoid, add a VM_BUG_ON() to
verify this assumption.]
[lee.schermerhorn@hp.com:
I don't think the function name 'mpol_new_mempolicy' is descriptive
enough to differentiate it from mpol_new().
This function applies cpuset set context, usually constraining nodes
to those allowed by the cpuset. However, when the 'RELATIVE_NODES flag
is set, it also translates the nodes. So I settled on
'mpol_set_nodemask()', because the comment block for mpol_new() mentions
that we need to call this function to "set nodes".
Some additional minor line length, whitespace and typo cleanup.]
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 15:31:49 -07:00
|
|
|
|
|
|
|
|
|
mempolicy: use MPOL_PREFERRED for system-wide default policy
Currently, when one specifies MPOL_DEFAULT via a NUMA memory policy API
[set_mempolicy(), mbind() and internal versions], the kernel simply installs a
NULL struct mempolicy pointer in the appropriate context: task policy, vma
policy, or shared policy. This causes any use of that policy to "fall back"
to the next most specific policy scope.
The only use of MPOL_DEFAULT to mean "local allocation" is in the system
default policy. This requires extra checks/cases for MPOL_DEFAULT in many
mempolicy.c functions.
There is another, "preferred" way to specify local allocation via the APIs.
That is using the MPOL_PREFERRED policy mode with an empty nodemask.
Internally, the empty nodemask gets converted to a preferred_node id of '-1'.
All internal usage of MPOL_PREFERRED will convert the '-1' to the id of the
node local to the cpu where the allocation occurs.
System default policy, except during boot, is hard-coded to "local
allocation". By using the MPOL_PREFERRED mode with a negative value of
preferred node for system default policy, MPOL_DEFAULT will never occur in the
'policy' member of a struct mempolicy. Thus, we can remove all checks for
MPOL_DEFAULT when converting policy to a node id/zonelist in the allocation
paths.
In slab_node() return local node id when policy pointer is NULL. No need to
set a pol value to take the switch default. Replace switch default with
BUG()--i.e., shouldn't happen.
With this patch MPOL_DEFAULT is only used in the APIs, including internal
calls to do_set_mempolicy() and in the display of policy in
/proc/<pid>/numa_maps. It always means "fall back" to the the next most
specific policy scope. This simplifies the description of memory policies
quite a bit, with no visible change in behavior.
get_mempolicy() continues to return MPOL_DEFAULT and an empty nodemask when
the requested policy [task or vma/shared] is NULL. These are the values one
would supply via set_mempolicy() or mbind() to achieve that condition--default
behavior.
This patch updates Documentation to reflect this change.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:18 -07:00
|
|
|
|
2023-10-03 02:20:14 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2005-10-29 18:15:48 -07:00
|
|
|
|
2023-10-03 02:20:14 -07:00
|
|
|
|
mempolicy: use MPOL_PREFERRED for system-wide default policy
Currently, when one specifies MPOL_DEFAULT via a NUMA memory policy API
[set_mempolicy(), mbind() and internal versions], the kernel simply installs a
NULL struct mempolicy pointer in the appropriate context: task policy, vma
policy, or shared policy. This causes any use of that policy to "fall back"
to the next most specific policy scope.
The only use of MPOL_DEFAULT to mean "local allocation" is in the system
default policy. This requires extra checks/cases for MPOL_DEFAULT in many
mempolicy.c functions.
There is another, "preferred" way to specify local allocation via the APIs.
That is using the MPOL_PREFERRED policy mode with an empty nodemask.
Internally, the empty nodemask gets converted to a preferred_node id of '-1'.
All internal usage of MPOL_PREFERRED will convert the '-1' to the id of the
node local to the cpu where the allocation occurs.
System default policy, except during boot, is hard-coded to "local
allocation". By using the MPOL_PREFERRED mode with a negative value of
preferred node for system default policy, MPOL_DEFAULT will never occur in the
'policy' member of a struct mempolicy. Thus, we can remove all checks for
MPOL_DEFAULT when converting policy to a node id/zonelist in the allocation
paths.
In slab_node() return local node id when policy pointer is NULL. No need to
set a pol value to take the switch default. Replace switch default with
BUG()--i.e., shouldn't happen.
With this patch MPOL_DEFAULT is only used in the APIs, including internal
calls to do_set_mempolicy() and in the display of policy in
/proc/<pid>/numa_maps. It always means "fall back" to the the next most
specific policy scope. This simplifies the description of memory policies
quite a bit, with no visible change in behavior.
get_mempolicy() continues to return MPOL_DEFAULT and an empty nodemask when
the requested policy [task or vma/shared] is NULL. These are the values one
would supply via set_mempolicy() or mbind() to achieve that condition--default
behavior.
This patch updates Documentation to reflect this change.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:18 -07:00
|
|
|
|
|
|
|
|
|
2023-10-03 02:20:14 -07:00
|
|
|
|
2008-04-28 02:12:18 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2021-06-30 18:51:10 -07:00
|
|
|
|
mm/mempolicy: add MPOL_PREFERRED_MANY for multiple preferred nodes
Patch series "Introduce multi-preference mempolicy", v7.
This patch series introduces the concept of the MPOL_PREFERRED_MANY
mempolicy. This mempolicy mode can be used with either the
set_mempolicy(2) or mbind(2) interfaces. Like the MPOL_PREFERRED
interface, it allows an application to set a preference for nodes which
will fulfil memory allocation requests. Unlike the MPOL_PREFERRED mode,
it takes a set of nodes. Like the MPOL_BIND interface, it works over a
set of nodes. Unlike MPOL_BIND, it will not cause a SIGSEGV or invoke the
OOM killer if those preferred nodes are not available.
Along with these patches are patches for libnuma, numactl, numademo, and
memhog. They still need some polish, but can be found here:
https://gitlab.com/bwidawsk/numactl/-/tree/prefer-many It allows new
usage: `numactl -P 0,3,4`
The goal of the new mode is to enable some use-cases when using tiered memory
usage models which I've lovingly named.
1a. The Hare - The interconnect is fast enough to meet bandwidth and
latency requirements allowing preference to be given to all nodes with
"fast" memory.
1b. The Indiscriminate Hare - An application knows it wants fast
memory (or perhaps slow memory), but doesn't care which node it runs
on. The application can prefer a set of nodes and then xpu bind to
the local node (cpu, accelerator, etc). This reverses the nodes are
chosen today where the kernel attempts to use local memory to the CPU
whenever possible. This will attempt to use the local accelerator to
the memory.
2. The Tortoise - The administrator (or the application itself) is
aware it only needs slow memory, and so can prefer that.
Much of this is almost achievable with the bind interface, but the bind
interface suffers from an inability to fallback to another set of nodes if
binding fails to all nodes in the nodemask.
Like MPOL_BIND a nodemask is given. Inherently this removes ordering from the
preference.
> /* Set first two nodes as preferred in an 8 node system. */
> const unsigned long nodes = 0x3
> set_mempolicy(MPOL_PREFER_MANY, &nodes, 8);
> /* Mimic interleave policy, but have fallback *.
> const unsigned long nodes = 0xaa
> set_mempolicy(MPOL_PREFER_MANY, &nodes, 8);
Some internal discussion took place around the interface. There are two
alternatives which we have discussed, plus one I stuck in:
1. Ordered list of nodes. Currently it's believed that the added
complexity is nod needed for expected usecases.
2. A flag for bind to allow falling back to other nodes. This
confuses the notion of binding and is less flexible than the current
solution.
3. Create flags or new modes that helps with some ordering. This
offers both a friendlier API as well as a solution for more customized
usage. It's unknown if it's worth the complexity to support this.
Here is sample code for how this might work:
> // Prefer specific nodes for some something wacky
> set_mempolicy(MPOL_PREFER_MANY, 0x17c, 1024);
>
> // Default
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_SOCKET, NULL, 0);
> // which is the same as
> set_mempolicy(MPOL_DEFAULT, NULL, 0);
>
> // The Hare
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE, NULL, 0);
>
> // The Tortoise
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE_REV, NULL, 0);
>
> // Prefer the fast memory of the first two sockets
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE, -1, 2);
>
This patch (of 5):
The NUMA APIs currently allow passing in a "preferred node" as a single
bit set in a nodemask. If more than one bit it set, bits after the first
are ignored.
This single node is generally OK for location-based NUMA where memory
being allocated will eventually be operated on by a single CPU. However,
in systems with multiple memory types, folks want to target a *type* of
memory instead of a location. For instance, someone might want some
high-bandwidth memory but do not care about the CPU next to which it is
allocated. Or, they want a cheap, high capacity allocation and want to
target all NUMA nodes which have persistent memory in volatile mode. In
both of these cases, the application wants to target a *set* of nodes, but
does not want strict MPOL_BIND behavior as that could lead to OOM killer
or SIGSEGV.
So add MPOL_PREFERRED_MANY policy to support the multiple preferred nodes
requirement. This is not a pie-in-the-sky dream for an API. This was a
response to a specific ask of more than one group at Intel. Specifically:
1. There are existing libraries that target memory types such as
https://github.com/memkind/memkind. These are known to suffer from
SIGSEGV's when memory is low on targeted memory "kinds" that span more
than one node. The MCDRAM on a Xeon Phi in "Cluster on Die" mode is an
example of this.
2. Volatile-use persistent memory users want to have a memory policy
which is targeted at either "cheap and slow" (PMEM) or "expensive and
fast" (DRAM). However, they do not want to experience allocation
failures when the targeted type is unavailable.
3. Allocate-then-run. Generally, we let the process scheduler decide
on which physical CPU to run a task. That location provides a default
allocation policy, and memory availability is not generally considered
when placing tasks. For situations where memory is valuable and
constrained, some users want to allocate memory first, *then* allocate
close compute resources to the allocation. This is the reverse of the
normal (CPU) model. Accelerators such as GPUs that operate on
core-mm-managed memory are interested in this model.
A check is added in sanitize_mpol_flags() to not permit 'prefer_many'
policy to be used for now, and will be removed in later patch after all
implementations for 'prefer_many' are ready, as suggested by Michal Hocko.
[mhocko@kernel.org: suggest to refine policy_node/policy_nodemask handling]
Link: https://lkml.kernel.org/r/1627970362-61305-1-git-send-email-feng.tang@intel.com
Link: https://lore.kernel.org/r/20200630212517.308045-4-ben.widawsky@intel.com
Link: https://lkml.kernel.org/r/1627970362-61305-2-git-send-email-feng.tang@intel.com
Co-developed-by: Ben Widawsky <ben.widawsky@intel.com>
Signed-off-by: Ben Widawsky <ben.widawsky@intel.com>
Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com>
Signed-off-by: Feng Tang <feng.tang@intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Huang Ying <ying.huang@intel.com>b
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-09-02 15:00:06 -07:00
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
2023-10-03 02:20:14 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
mm/mempolicy: don't handle MPOL_LOCAL like a fake MPOL_PREFERRED policy
MPOL_LOCAL policy has been setup as a real policy, but it is still handled
like a faked POL_PREFERRED policy with one internal MPOL_F_LOCAL flag bit
set, and there are many places having to judge the real 'prefer' or the
'local' policy, which are quite confusing.
In current code, there are 4 cases that MPOL_LOCAL are used:
1. user specifies 'local' policy
2. user specifies 'prefer' policy, but with empty nodemask
3. system 'default' policy is used
4. 'prefer' policy + valid 'preferred' node with MPOL_F_STATIC_NODES
flag set, and when it is 'rebind' to a nodemask which doesn't contains
the 'preferred' node, it will perform as 'local' policy
So make 'local' a real policy instead of a fake 'prefer' one, and kill
MPOL_F_LOCAL bit, which can greatly reduce the confusion for code reading.
For case 4, the logic of mpol_rebind_preferred() is confusing, as Michal
Hocko pointed out:
: I do believe that rebinding preferred policy is just bogus and it should
: be dropped altogether on the ground that a preference is a mere hint from
: userspace where to start the allocation. Unless I am missing something
: cpusets will be always authoritative for the final placement. The
: preferred node just acts as a starting point and it should be really
: preserved when cpusets changes. Otherwise we have a very subtle behavior
: corner cases.
So dump all the tricky transformation between 'prefer' and 'local', and
just record the new nodemask of rebinding.
[feng.tang@intel.com: fix a problem in mpol_set_nodemask(), per Michal Hocko]
Link: https://lkml.kernel.org/r/1622560492-1294-3-git-send-email-feng.tang@intel.com
[feng.tang@intel.com: refine code and comments of mpol_set_nodemask(), per Michal]
Link: https://lkml.kernel.org/r/20210603081807.GE56979@shbuild999.sh.intel.com
Link: https://lkml.kernel.org/r/1622469956-82897-3-git-send-email-feng.tang@intel.com
Signed-off-by: Feng Tang <feng.tang@intel.com>
Suggested-by: Michal Hocko <mhocko@suse.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ben Widawsky <ben.widawsky@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-30 18:51:00 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
userfaultfd: allow get_mempolicy(MPOL_F_NODE|MPOL_F_ADDR) to trigger userfaults
get_mempolicy(MPOL_F_NODE|MPOL_F_ADDR) called a get_user_pages that would
not be waiting for userfaults before failing and it would hit on a SIGBUS
instead. Using get_user_pages_locked/unlocked instead will allow
get_mempolicy to allow userfaults to resolve the fault and fill the hole,
before grabbing the node id of the page.
If the user calls get_mempolicy() with MPOL_F_ADDR | MPOL_F_NODE for an
address inside an area managed by uffd and there is no page at that
address, the page allocation from within get_mempolicy() will fail
because get_user_pages() does not allow for page fault retry required
for uffd; the user will get SIGBUS.
With this patch, the page fault will be resolved by the uffd and the
get_mempolicy() will continue normally.
Background:
Via code review, previously the syscall would have returned -EFAULT
(vm_fault_to_errno), now it will block and wait for an userfault (if
it's waken before the fault is resolved it'll still -EFAULT).
This way get_mempolicy will give a chance to an "unaware" app to be
compliant with userfaults.
The reason this visible change is that becoming "userfault compliant"
cannot regress anything: all other syscalls including read(2)/write(2)
had to become "userfault compliant" long time ago (that's one of the
things userfaultfd can do that PROT_NONE and trapping segfaults can't).
So this is just one more syscall that become "userfault compliant" like
all other major ones already were.
This has been happening on virtio-bridge dpdk process which just called
get_mempolicy on the guest space post live migration, but before the
memory had a chance to be migrated to destination.
I didn't run an strace to be able to show the -EFAULT going away, but
I've the confirmation of the below debug aid information (only visible
with CONFIG_DEBUG_VM=y) going away with the patch:
[20116.371461] FAULT_FLAG_ALLOW_RETRY missing 0
[20116.371464] CPU: 1 PID: 13381 Comm: vhost-events Not tainted 4.17.12-200.fc28.x86_64 #1
[20116.371465] Hardware name: LENOVO 20FAS2BN0A/20FAS2BN0A, BIOS N1CET54W (1.22 ) 02/10/2017
[20116.371466] Call Trace:
[20116.371473] dump_stack+0x5c/0x80
[20116.371476] handle_userfault.cold.37+0x1b/0x22
[20116.371479] ? remove_wait_queue+0x20/0x60
[20116.371481] ? poll_freewait+0x45/0xa0
[20116.371483] ? do_sys_poll+0x31c/0x520
[20116.371485] ? radix_tree_lookup_slot+0x1e/0x50
[20116.371488] shmem_getpage_gfp+0xce7/0xe50
[20116.371491] ? page_add_file_rmap+0x1a/0x2c0
[20116.371493] shmem_fault+0x78/0x1e0
[20116.371495] ? filemap_map_pages+0x3a1/0x450
[20116.371498] __do_fault+0x1f/0xc0
[20116.371500] __handle_mm_fault+0xe2e/0x12f0
[20116.371502] handle_mm_fault+0xda/0x200
[20116.371504] __get_user_pages+0x238/0x790
[20116.371506] get_user_pages+0x3e/0x50
[20116.371510] kernel_get_mempolicy+0x40b/0x700
[20116.371512] ? vfs_write+0x170/0x1a0
[20116.371515] __x64_sys_get_mempolicy+0x21/0x30
[20116.371517] do_syscall_64+0x5b/0x160
[20116.371520] entry_SYSCALL_64_after_hwframe+0x44/0xa9
The above harmless debug message (not a kernel crash, just a
dump_stack()) is shown with CONFIG_DEBUG_VM=y to more quickly identify
and improve kernel spots that may have to become "userfaultfd
compliant" like this one (without having to run an strace and search
for syscall misbehavior). Spots like the above are more closer to a
kernel bug for the non-cooperative usages that Mike focuses on, than
for for dpdk qemu-cooperative usages that reproduced it, but it's still
nicer to get this fixed for dpdk too.
The part of the patch that caused me to think is only the
implementation issue of mpol_get, but it looks like it should work safe
no matter the kind of mempolicy structure that is (the default static
policy also starts at 1 so it'll go to 2 and back to 1 without crashing
everything at 0).
[rppt@linux.vnet.ibm.com: changelog addition]
http://lkml.kernel.org/r/20180904073718.GA26916@rapoport-lnx
Link: http://lkml.kernel.org/r/20180831214848.23676-1-aarcange@redhat.com
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Tested-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Reviewed-by: Mike Rapoport <rppt@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-26 15:05:16 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2020-04-07 21:40:09 -04:00
|
|
|
|
2022-03-22 14:39:46 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2022-03-22 14:39:46 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2022-03-22 14:39:46 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-10-16 01:26:26 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2005-10-29 18:16:59 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
userfaultfd: allow get_mempolicy(MPOL_F_NODE|MPOL_F_ADDR) to trigger userfaults
get_mempolicy(MPOL_F_NODE|MPOL_F_ADDR) called a get_user_pages that would
not be waiting for userfaults before failing and it would hit on a SIGBUS
instead. Using get_user_pages_locked/unlocked instead will allow
get_mempolicy to allow userfaults to resolve the fault and fill the hole,
before grabbing the node id of the page.
If the user calls get_mempolicy() with MPOL_F_ADDR | MPOL_F_NODE for an
address inside an area managed by uffd and there is no page at that
address, the page allocation from within get_mempolicy() will fail
because get_user_pages() does not allow for page fault retry required
for uffd; the user will get SIGBUS.
With this patch, the page fault will be resolved by the uffd and the
get_mempolicy() will continue normally.
Background:
Via code review, previously the syscall would have returned -EFAULT
(vm_fault_to_errno), now it will block and wait for an userfault (if
it's waken before the fault is resolved it'll still -EFAULT).
This way get_mempolicy will give a chance to an "unaware" app to be
compliant with userfaults.
The reason this visible change is that becoming "userfault compliant"
cannot regress anything: all other syscalls including read(2)/write(2)
had to become "userfault compliant" long time ago (that's one of the
things userfaultfd can do that PROT_NONE and trapping segfaults can't).
So this is just one more syscall that become "userfault compliant" like
all other major ones already were.
This has been happening on virtio-bridge dpdk process which just called
get_mempolicy on the guest space post live migration, but before the
memory had a chance to be migrated to destination.
I didn't run an strace to be able to show the -EFAULT going away, but
I've the confirmation of the below debug aid information (only visible
with CONFIG_DEBUG_VM=y) going away with the patch:
[20116.371461] FAULT_FLAG_ALLOW_RETRY missing 0
[20116.371464] CPU: 1 PID: 13381 Comm: vhost-events Not tainted 4.17.12-200.fc28.x86_64 #1
[20116.371465] Hardware name: LENOVO 20FAS2BN0A/20FAS2BN0A, BIOS N1CET54W (1.22 ) 02/10/2017
[20116.371466] Call Trace:
[20116.371473] dump_stack+0x5c/0x80
[20116.371476] handle_userfault.cold.37+0x1b/0x22
[20116.371479] ? remove_wait_queue+0x20/0x60
[20116.371481] ? poll_freewait+0x45/0xa0
[20116.371483] ? do_sys_poll+0x31c/0x520
[20116.371485] ? radix_tree_lookup_slot+0x1e/0x50
[20116.371488] shmem_getpage_gfp+0xce7/0xe50
[20116.371491] ? page_add_file_rmap+0x1a/0x2c0
[20116.371493] shmem_fault+0x78/0x1e0
[20116.371495] ? filemap_map_pages+0x3a1/0x450
[20116.371498] __do_fault+0x1f/0xc0
[20116.371500] __handle_mm_fault+0xe2e/0x12f0
[20116.371502] handle_mm_fault+0xda/0x200
[20116.371504] __get_user_pages+0x238/0x790
[20116.371506] get_user_pages+0x3e/0x50
[20116.371510] kernel_get_mempolicy+0x40b/0x700
[20116.371512] ? vfs_write+0x170/0x1a0
[20116.371515] __x64_sys_get_mempolicy+0x21/0x30
[20116.371517] do_syscall_64+0x5b/0x160
[20116.371520] entry_SYSCALL_64_after_hwframe+0x44/0xa9
The above harmless debug message (not a kernel crash, just a
dump_stack()) is shown with CONFIG_DEBUG_VM=y to more quickly identify
and improve kernel spots that may have to become "userfaultfd
compliant" like this one (without having to run an strace and search
for syscall misbehavior). Spots like the above are more closer to a
kernel bug for the non-cooperative usages that Mike focuses on, than
for for dpdk qemu-cooperative usages that reproduced it, but it's still
nicer to get this fixed for dpdk too.
The part of the patch that caused me to think is only the
implementation issue of mpol_get, but it looks like it should work safe
no matter the kind of mempolicy structure that is (the default static
policy also starts at 1 so it'll go to 2 and back to 1 without crashing
everything at 0).
[rppt@linux.vnet.ibm.com: changelog addition]
http://lkml.kernel.org/r/20180904073718.GA26916@rapoport-lnx
Link: http://lkml.kernel.org/r/20180831214848.23676-1-aarcange@redhat.com
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Tested-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Reviewed-by: Mike Rapoport <rppt@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-26 15:05:16 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2007-10-16 01:24:51 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2007-10-16 01:24:51 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
cpuset,mm: update tasks' mems_allowed in time
Fix allocating page cache/slab object on the unallowed node when memory
spread is set by updating tasks' mems_allowed after its cpuset's mems is
changed.
In order to update tasks' mems_allowed in time, we must modify the code of
memory policy. Because the memory policy is applied in the process's
context originally. After applying this patch, one task directly
manipulates anothers mems_allowed, and we use alloc_lock in the
task_struct to protect mems_allowed and memory policy of the task.
But in the fast path, we didn't use lock to protect them, because adding a
lock may lead to performance regression. But if we don't add a lock,the
task might see no nodes when changing cpuset's mems_allowed to some
non-overlapping set. In order to avoid it, we set all new allowed nodes,
then clear newly disallowed ones.
[lee.schermerhorn@hp.com:
The rework of mpol_new() to extract the adjusting of the node mask to
apply cpuset and mpol flags "context" breaks set_mempolicy() and mbind()
with MPOL_PREFERRED and a NULL nodemask--i.e., explicit local
allocation. Fix this by adding the check for MPOL_PREFERRED and empty
node mask to mpol_new_mpolicy().
Remove the now unneeded 'nodes = NULL' from mpol_new().
Note that mpol_new_mempolicy() is always called with a non-NULL
'nodes' parameter now that it has been removed from mpol_new().
Therefore, we don't need to test nodes for NULL before testing it for
'empty'. However, just to be extra paranoid, add a VM_BUG_ON() to
verify this assumption.]
[lee.schermerhorn@hp.com:
I don't think the function name 'mpol_new_mempolicy' is descriptive
enough to differentiate it from mpol_new().
This function applies cpuset set context, usually constraining nodes
to those allowed by the cpuset. However, when the 'RELATIVE_NODES flag
is set, it also translates the nodes. So I settled on
'mpol_set_nodemask()', because the comment block for mpol_new() mentions
that we need to call this function to "set nodes".
Some additional minor line length, whitespace and typo cleanup.]
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 15:31:49 -07:00
|
|
|
|
2007-10-16 01:24:51 -07:00
|
|
|
|
cpuset,mm: update tasks' mems_allowed in time
Fix allocating page cache/slab object on the unallowed node when memory
spread is set by updating tasks' mems_allowed after its cpuset's mems is
changed.
In order to update tasks' mems_allowed in time, we must modify the code of
memory policy. Because the memory policy is applied in the process's
context originally. After applying this patch, one task directly
manipulates anothers mems_allowed, and we use alloc_lock in the
task_struct to protect mems_allowed and memory policy of the task.
But in the fast path, we didn't use lock to protect them, because adding a
lock may lead to performance regression. But if we don't add a lock,the
task might see no nodes when changing cpuset's mems_allowed to some
non-overlapping set. In order to avoid it, we set all new allowed nodes,
then clear newly disallowed ones.
[lee.schermerhorn@hp.com:
The rework of mpol_new() to extract the adjusting of the node mask to
apply cpuset and mpol flags "context" breaks set_mempolicy() and mbind()
with MPOL_PREFERRED and a NULL nodemask--i.e., explicit local
allocation. Fix this by adding the check for MPOL_PREFERRED and empty
node mask to mpol_new_mpolicy().
Remove the now unneeded 'nodes = NULL' from mpol_new().
Note that mpol_new_mempolicy() is always called with a non-NULL
'nodes' parameter now that it has been removed from mpol_new().
Therefore, we don't need to test nodes for NULL before testing it for
'empty'. However, just to be extra paranoid, add a VM_BUG_ON() to
verify this assumption.]
[lee.schermerhorn@hp.com:
I don't think the function name 'mpol_new_mempolicy' is descriptive
enough to differentiate it from mpol_new().
This function applies cpuset set context, usually constraining nodes
to those allowed by the cpuset. However, when the 'RELATIVE_NODES flag
is set, it also translates the nodes. So I settled on
'mpol_set_nodemask()', because the comment block for mpol_new() mentions
that we need to call this function to "set nodes".
Some additional minor line length, whitespace and typo cleanup.]
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 15:31:49 -07:00
|
|
|
|
2007-10-16 01:24:51 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
mempolicy: use MPOL_PREFERRED for system-wide default policy
Currently, when one specifies MPOL_DEFAULT via a NUMA memory policy API
[set_mempolicy(), mbind() and internal versions], the kernel simply installs a
NULL struct mempolicy pointer in the appropriate context: task policy, vma
policy, or shared policy. This causes any use of that policy to "fall back"
to the next most specific policy scope.
The only use of MPOL_DEFAULT to mean "local allocation" is in the system
default policy. This requires extra checks/cases for MPOL_DEFAULT in many
mempolicy.c functions.
There is another, "preferred" way to specify local allocation via the APIs.
That is using the MPOL_PREFERRED policy mode with an empty nodemask.
Internally, the empty nodemask gets converted to a preferred_node id of '-1'.
All internal usage of MPOL_PREFERRED will convert the '-1' to the id of the
node local to the cpu where the allocation occurs.
System default policy, except during boot, is hard-coded to "local
allocation". By using the MPOL_PREFERRED mode with a negative value of
preferred node for system default policy, MPOL_DEFAULT will never occur in the
'policy' member of a struct mempolicy. Thus, we can remove all checks for
MPOL_DEFAULT when converting policy to a node id/zonelist in the allocation
paths.
In slab_node() return local node id when policy pointer is NULL. No need to
set a pol value to take the switch default. Replace switch default with
BUG()--i.e., shouldn't happen.
With this patch MPOL_DEFAULT is only used in the APIs, including internal
calls to do_set_mempolicy() and in the display of policy in
/proc/<pid>/numa_maps. It always means "fall back" to the the next most
specific policy scope. This simplifies the description of memory policies
quite a bit, with no visible change in behavior.
get_mempolicy() continues to return MPOL_DEFAULT and an empty nodemask when
the requested policy [task or vma/shared] is NULL. These are the values one
would supply via set_mempolicy() or mbind() to achieve that condition--default
behavior.
This patch updates Documentation to reflect this change.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:18 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-06-08 21:33:25 -07:00
|
|
|
|
2021-06-28 19:39:53 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2020-06-08 21:33:25 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: use MPOL_PREFERRED for system-wide default policy
Currently, when one specifies MPOL_DEFAULT via a NUMA memory policy API
[set_mempolicy(), mbind() and internal versions], the kernel simply installs a
NULL struct mempolicy pointer in the appropriate context: task policy, vma
policy, or shared policy. This causes any use of that policy to "fall back"
to the next most specific policy scope.
The only use of MPOL_DEFAULT to mean "local allocation" is in the system
default policy. This requires extra checks/cases for MPOL_DEFAULT in many
mempolicy.c functions.
There is another, "preferred" way to specify local allocation via the APIs.
That is using the MPOL_PREFERRED policy mode with an empty nodemask.
Internally, the empty nodemask gets converted to a preferred_node id of '-1'.
All internal usage of MPOL_PREFERRED will convert the '-1' to the id of the
node local to the cpu where the allocation occurs.
System default policy, except during boot, is hard-coded to "local
allocation". By using the MPOL_PREFERRED mode with a negative value of
preferred node for system default policy, MPOL_DEFAULT will never occur in the
'policy' member of a struct mempolicy. Thus, we can remove all checks for
MPOL_DEFAULT when converting policy to a node id/zonelist in the allocation
paths.
In slab_node() return local node id when policy pointer is NULL. No need to
set a pol value to take the switch default. Replace switch default with
BUG()--i.e., shouldn't happen.
With this patch MPOL_DEFAULT is only used in the APIs, including internal
calls to do_set_mempolicy() and in the display of policy in
/proc/<pid>/numa_maps. It always means "fall back" to the the next most
specific policy scope. This simplifies the description of memory policies
quite a bit, with no visible change in behavior.
get_mempolicy() continues to return MPOL_DEFAULT and an empty nodemask when
the requested policy [task or vma/shared] is NULL. These are the values one
would supply via set_mempolicy() or mbind() to achieve that condition--default
behavior.
This patch updates Documentation to reflect this change.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:18 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
userfaultfd: allow get_mempolicy(MPOL_F_NODE|MPOL_F_ADDR) to trigger userfaults
get_mempolicy(MPOL_F_NODE|MPOL_F_ADDR) called a get_user_pages that would
not be waiting for userfaults before failing and it would hit on a SIGBUS
instead. Using get_user_pages_locked/unlocked instead will allow
get_mempolicy to allow userfaults to resolve the fault and fill the hole,
before grabbing the node id of the page.
If the user calls get_mempolicy() with MPOL_F_ADDR | MPOL_F_NODE for an
address inside an area managed by uffd and there is no page at that
address, the page allocation from within get_mempolicy() will fail
because get_user_pages() does not allow for page fault retry required
for uffd; the user will get SIGBUS.
With this patch, the page fault will be resolved by the uffd and the
get_mempolicy() will continue normally.
Background:
Via code review, previously the syscall would have returned -EFAULT
(vm_fault_to_errno), now it will block and wait for an userfault (if
it's waken before the fault is resolved it'll still -EFAULT).
This way get_mempolicy will give a chance to an "unaware" app to be
compliant with userfaults.
The reason this visible change is that becoming "userfault compliant"
cannot regress anything: all other syscalls including read(2)/write(2)
had to become "userfault compliant" long time ago (that's one of the
things userfaultfd can do that PROT_NONE and trapping segfaults can't).
So this is just one more syscall that become "userfault compliant" like
all other major ones already were.
This has been happening on virtio-bridge dpdk process which just called
get_mempolicy on the guest space post live migration, but before the
memory had a chance to be migrated to destination.
I didn't run an strace to be able to show the -EFAULT going away, but
I've the confirmation of the below debug aid information (only visible
with CONFIG_DEBUG_VM=y) going away with the patch:
[20116.371461] FAULT_FLAG_ALLOW_RETRY missing 0
[20116.371464] CPU: 1 PID: 13381 Comm: vhost-events Not tainted 4.17.12-200.fc28.x86_64 #1
[20116.371465] Hardware name: LENOVO 20FAS2BN0A/20FAS2BN0A, BIOS N1CET54W (1.22 ) 02/10/2017
[20116.371466] Call Trace:
[20116.371473] dump_stack+0x5c/0x80
[20116.371476] handle_userfault.cold.37+0x1b/0x22
[20116.371479] ? remove_wait_queue+0x20/0x60
[20116.371481] ? poll_freewait+0x45/0xa0
[20116.371483] ? do_sys_poll+0x31c/0x520
[20116.371485] ? radix_tree_lookup_slot+0x1e/0x50
[20116.371488] shmem_getpage_gfp+0xce7/0xe50
[20116.371491] ? page_add_file_rmap+0x1a/0x2c0
[20116.371493] shmem_fault+0x78/0x1e0
[20116.371495] ? filemap_map_pages+0x3a1/0x450
[20116.371498] __do_fault+0x1f/0xc0
[20116.371500] __handle_mm_fault+0xe2e/0x12f0
[20116.371502] handle_mm_fault+0xda/0x200
[20116.371504] __get_user_pages+0x238/0x790
[20116.371506] get_user_pages+0x3e/0x50
[20116.371510] kernel_get_mempolicy+0x40b/0x700
[20116.371512] ? vfs_write+0x170/0x1a0
[20116.371515] __x64_sys_get_mempolicy+0x21/0x30
[20116.371517] do_syscall_64+0x5b/0x160
[20116.371520] entry_SYSCALL_64_after_hwframe+0x44/0xa9
The above harmless debug message (not a kernel crash, just a
dump_stack()) is shown with CONFIG_DEBUG_VM=y to more quickly identify
and improve kernel spots that may have to become "userfaultfd
compliant" like this one (without having to run an strace and search
for syscall misbehavior). Spots like the above are more closer to a
kernel bug for the non-cooperative usages that Mike focuses on, than
for for dpdk qemu-cooperative usages that reproduced it, but it's still
nicer to get this fixed for dpdk too.
The part of the patch that caused me to think is only the
implementation issue of mpol_get, but it looks like it should work safe
no matter the kind of mempolicy structure that is (the default static
policy also starts at 1 so it'll go to 2 and back to 1 without crashing
everything at 0).
[rppt@linux.vnet.ibm.com: changelog addition]
http://lkml.kernel.org/r/20180904073718.GA26916@rapoport-lnx
Link: http://lkml.kernel.org/r/20180831214848.23676-1-aarcange@redhat.com
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Tested-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Reviewed-by: Mike Rapoport <rppt@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-26 15:05:16 -07:00
|
|
|
|
2022-03-22 14:39:46 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
userfaultfd: allow get_mempolicy(MPOL_F_NODE|MPOL_F_ADDR) to trigger userfaults
get_mempolicy(MPOL_F_NODE|MPOL_F_ADDR) called a get_user_pages that would
not be waiting for userfaults before failing and it would hit on a SIGBUS
instead. Using get_user_pages_locked/unlocked instead will allow
get_mempolicy to allow userfaults to resolve the fault and fill the hole,
before grabbing the node id of the page.
If the user calls get_mempolicy() with MPOL_F_ADDR | MPOL_F_NODE for an
address inside an area managed by uffd and there is no page at that
address, the page allocation from within get_mempolicy() will fail
because get_user_pages() does not allow for page fault retry required
for uffd; the user will get SIGBUS.
With this patch, the page fault will be resolved by the uffd and the
get_mempolicy() will continue normally.
Background:
Via code review, previously the syscall would have returned -EFAULT
(vm_fault_to_errno), now it will block and wait for an userfault (if
it's waken before the fault is resolved it'll still -EFAULT).
This way get_mempolicy will give a chance to an "unaware" app to be
compliant with userfaults.
The reason this visible change is that becoming "userfault compliant"
cannot regress anything: all other syscalls including read(2)/write(2)
had to become "userfault compliant" long time ago (that's one of the
things userfaultfd can do that PROT_NONE and trapping segfaults can't).
So this is just one more syscall that become "userfault compliant" like
all other major ones already were.
This has been happening on virtio-bridge dpdk process which just called
get_mempolicy on the guest space post live migration, but before the
memory had a chance to be migrated to destination.
I didn't run an strace to be able to show the -EFAULT going away, but
I've the confirmation of the below debug aid information (only visible
with CONFIG_DEBUG_VM=y) going away with the patch:
[20116.371461] FAULT_FLAG_ALLOW_RETRY missing 0
[20116.371464] CPU: 1 PID: 13381 Comm: vhost-events Not tainted 4.17.12-200.fc28.x86_64 #1
[20116.371465] Hardware name: LENOVO 20FAS2BN0A/20FAS2BN0A, BIOS N1CET54W (1.22 ) 02/10/2017
[20116.371466] Call Trace:
[20116.371473] dump_stack+0x5c/0x80
[20116.371476] handle_userfault.cold.37+0x1b/0x22
[20116.371479] ? remove_wait_queue+0x20/0x60
[20116.371481] ? poll_freewait+0x45/0xa0
[20116.371483] ? do_sys_poll+0x31c/0x520
[20116.371485] ? radix_tree_lookup_slot+0x1e/0x50
[20116.371488] shmem_getpage_gfp+0xce7/0xe50
[20116.371491] ? page_add_file_rmap+0x1a/0x2c0
[20116.371493] shmem_fault+0x78/0x1e0
[20116.371495] ? filemap_map_pages+0x3a1/0x450
[20116.371498] __do_fault+0x1f/0xc0
[20116.371500] __handle_mm_fault+0xe2e/0x12f0
[20116.371502] handle_mm_fault+0xda/0x200
[20116.371504] __get_user_pages+0x238/0x790
[20116.371506] get_user_pages+0x3e/0x50
[20116.371510] kernel_get_mempolicy+0x40b/0x700
[20116.371512] ? vfs_write+0x170/0x1a0
[20116.371515] __x64_sys_get_mempolicy+0x21/0x30
[20116.371517] do_syscall_64+0x5b/0x160
[20116.371520] entry_SYSCALL_64_after_hwframe+0x44/0xa9
The above harmless debug message (not a kernel crash, just a
dump_stack()) is shown with CONFIG_DEBUG_VM=y to more quickly identify
and improve kernel spots that may have to become "userfaultfd
compliant" like this one (without having to run an strace and search
for syscall misbehavior). Spots like the above are more closer to a
kernel bug for the non-cooperative usages that Mike focuses on, than
for for dpdk qemu-cooperative usages that reproduced it, but it's still
nicer to get this fixed for dpdk too.
The part of the patch that caused me to think is only the
implementation issue of mpol_get, but it looks like it should work safe
no matter the kind of mempolicy structure that is (the default static
policy also starts at 1 so it'll go to 2 and back to 1 without crashing
everything at 0).
[rppt@linux.vnet.ibm.com: changelog addition]
http://lkml.kernel.org/r/20180904073718.GA26916@rapoport-lnx
Link: http://lkml.kernel.org/r/20180831214848.23676-1-aarcange@redhat.com
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Tested-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Reviewed-by: Mike Rapoport <rppt@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-26 15:05:16 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-03-22 14:39:46 -07:00
|
|
|
|
userfaultfd: allow get_mempolicy(MPOL_F_NODE|MPOL_F_ADDR) to trigger userfaults
get_mempolicy(MPOL_F_NODE|MPOL_F_ADDR) called a get_user_pages that would
not be waiting for userfaults before failing and it would hit on a SIGBUS
instead. Using get_user_pages_locked/unlocked instead will allow
get_mempolicy to allow userfaults to resolve the fault and fill the hole,
before grabbing the node id of the page.
If the user calls get_mempolicy() with MPOL_F_ADDR | MPOL_F_NODE for an
address inside an area managed by uffd and there is no page at that
address, the page allocation from within get_mempolicy() will fail
because get_user_pages() does not allow for page fault retry required
for uffd; the user will get SIGBUS.
With this patch, the page fault will be resolved by the uffd and the
get_mempolicy() will continue normally.
Background:
Via code review, previously the syscall would have returned -EFAULT
(vm_fault_to_errno), now it will block and wait for an userfault (if
it's waken before the fault is resolved it'll still -EFAULT).
This way get_mempolicy will give a chance to an "unaware" app to be
compliant with userfaults.
The reason this visible change is that becoming "userfault compliant"
cannot regress anything: all other syscalls including read(2)/write(2)
had to become "userfault compliant" long time ago (that's one of the
things userfaultfd can do that PROT_NONE and trapping segfaults can't).
So this is just one more syscall that become "userfault compliant" like
all other major ones already were.
This has been happening on virtio-bridge dpdk process which just called
get_mempolicy on the guest space post live migration, but before the
memory had a chance to be migrated to destination.
I didn't run an strace to be able to show the -EFAULT going away, but
I've the confirmation of the below debug aid information (only visible
with CONFIG_DEBUG_VM=y) going away with the patch:
[20116.371461] FAULT_FLAG_ALLOW_RETRY missing 0
[20116.371464] CPU: 1 PID: 13381 Comm: vhost-events Not tainted 4.17.12-200.fc28.x86_64 #1
[20116.371465] Hardware name: LENOVO 20FAS2BN0A/20FAS2BN0A, BIOS N1CET54W (1.22 ) 02/10/2017
[20116.371466] Call Trace:
[20116.371473] dump_stack+0x5c/0x80
[20116.371476] handle_userfault.cold.37+0x1b/0x22
[20116.371479] ? remove_wait_queue+0x20/0x60
[20116.371481] ? poll_freewait+0x45/0xa0
[20116.371483] ? do_sys_poll+0x31c/0x520
[20116.371485] ? radix_tree_lookup_slot+0x1e/0x50
[20116.371488] shmem_getpage_gfp+0xce7/0xe50
[20116.371491] ? page_add_file_rmap+0x1a/0x2c0
[20116.371493] shmem_fault+0x78/0x1e0
[20116.371495] ? filemap_map_pages+0x3a1/0x450
[20116.371498] __do_fault+0x1f/0xc0
[20116.371500] __handle_mm_fault+0xe2e/0x12f0
[20116.371502] handle_mm_fault+0xda/0x200
[20116.371504] __get_user_pages+0x238/0x790
[20116.371506] get_user_pages+0x3e/0x50
[20116.371510] kernel_get_mempolicy+0x40b/0x700
[20116.371512] ? vfs_write+0x170/0x1a0
[20116.371515] __x64_sys_get_mempolicy+0x21/0x30
[20116.371517] do_syscall_64+0x5b/0x160
[20116.371520] entry_SYSCALL_64_after_hwframe+0x44/0xa9
The above harmless debug message (not a kernel crash, just a
dump_stack()) is shown with CONFIG_DEBUG_VM=y to more quickly identify
and improve kernel spots that may have to become "userfaultfd
compliant" like this one (without having to run an strace and search
for syscall misbehavior). Spots like the above are more closer to a
kernel bug for the non-cooperative usages that Mike focuses on, than
for for dpdk qemu-cooperative usages that reproduced it, but it's still
nicer to get this fixed for dpdk too.
The part of the patch that caused me to think is only the
implementation issue of mpol_get, but it looks like it should work safe
no matter the kind of mempolicy structure that is (the default static
policy also starts at 1 so it'll go to 2 and back to 1 without crashing
everything at 0).
[rppt@linux.vnet.ibm.com: changelog addition]
http://lkml.kernel.org/r/20180904073718.GA26916@rapoport-lnx
Link: http://lkml.kernel.org/r/20180831214848.23676-1-aarcange@redhat.com
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Tested-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Reviewed-by: Mike Rapoport <rppt@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-26 15:05:16 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2005-10-29 18:16:59 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2008-04-28 02:13:12 -07:00
|
|
|
|
2021-06-30 18:51:10 -07:00
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: use MPOL_PREFERRED for system-wide default policy
Currently, when one specifies MPOL_DEFAULT via a NUMA memory policy API
[set_mempolicy(), mbind() and internal versions], the kernel simply installs a
NULL struct mempolicy pointer in the appropriate context: task policy, vma
policy, or shared policy. This causes any use of that policy to "fall back"
to the next most specific policy scope.
The only use of MPOL_DEFAULT to mean "local allocation" is in the system
default policy. This requires extra checks/cases for MPOL_DEFAULT in many
mempolicy.c functions.
There is another, "preferred" way to specify local allocation via the APIs.
That is using the MPOL_PREFERRED policy mode with an empty nodemask.
Internally, the empty nodemask gets converted to a preferred_node id of '-1'.
All internal usage of MPOL_PREFERRED will convert the '-1' to the id of the
node local to the cpu where the allocation occurs.
System default policy, except during boot, is hard-coded to "local
allocation". By using the MPOL_PREFERRED mode with a negative value of
preferred node for system default policy, MPOL_DEFAULT will never occur in the
'policy' member of a struct mempolicy. Thus, we can remove all checks for
MPOL_DEFAULT when converting policy to a node id/zonelist in the allocation
paths.
In slab_node() return local node id when policy pointer is NULL. No need to
set a pol value to take the switch default. Replace switch default with
BUG()--i.e., shouldn't happen.
With this patch MPOL_DEFAULT is only used in the APIs, including internal
calls to do_set_mempolicy() and in the display of policy in
/proc/<pid>/numa_maps. It always means "fall back" to the the next most
specific policy scope. This simplifies the description of memory policies
quite a bit, with no visible change in behavior.
get_mempolicy() continues to return MPOL_DEFAULT and an empty nodemask when
the requested policy [task or vma/shared] is NULL. These are the values one
would supply via set_mempolicy() or mbind() to achieve that condition--default
behavior.
This patch updates Documentation to reflect this change.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:18 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-07-04 12:24:13 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: use MPOL_PREFERRED for system-wide default policy
Currently, when one specifies MPOL_DEFAULT via a NUMA memory policy API
[set_mempolicy(), mbind() and internal versions], the kernel simply installs a
NULL struct mempolicy pointer in the appropriate context: task policy, vma
policy, or shared policy. This causes any use of that policy to "fall back"
to the next most specific policy scope.
The only use of MPOL_DEFAULT to mean "local allocation" is in the system
default policy. This requires extra checks/cases for MPOL_DEFAULT in many
mempolicy.c functions.
There is another, "preferred" way to specify local allocation via the APIs.
That is using the MPOL_PREFERRED policy mode with an empty nodemask.
Internally, the empty nodemask gets converted to a preferred_node id of '-1'.
All internal usage of MPOL_PREFERRED will convert the '-1' to the id of the
node local to the cpu where the allocation occurs.
System default policy, except during boot, is hard-coded to "local
allocation". By using the MPOL_PREFERRED mode with a negative value of
preferred node for system default policy, MPOL_DEFAULT will never occur in the
'policy' member of a struct mempolicy. Thus, we can remove all checks for
MPOL_DEFAULT when converting policy to a node id/zonelist in the allocation
paths.
In slab_node() return local node id when policy pointer is NULL. No need to
set a pol value to take the switch default. Replace switch default with
BUG()--i.e., shouldn't happen.
With this patch MPOL_DEFAULT is only used in the APIs, including internal
calls to do_set_mempolicy() and in the display of policy in
/proc/<pid>/numa_maps. It always means "fall back" to the the next most
specific policy scope. This simplifies the description of memory policies
quite a bit, with no visible change in behavior.
get_mempolicy() continues to return MPOL_DEFAULT and an empty nodemask when
the requested policy [task or vma/shared] is NULL. These are the values one
would supply via set_mempolicy() or mbind() to achieve that condition--default
behavior.
This patch updates Documentation to reflect this change.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:18 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
cpuset,mm: update tasks' mems_allowed in time
Fix allocating page cache/slab object on the unallowed node when memory
spread is set by updating tasks' mems_allowed after its cpuset's mems is
changed.
In order to update tasks' mems_allowed in time, we must modify the code of
memory policy. Because the memory policy is applied in the process's
context originally. After applying this patch, one task directly
manipulates anothers mems_allowed, and we use alloc_lock in the
task_struct to protect mems_allowed and memory policy of the task.
But in the fast path, we didn't use lock to protect them, because adding a
lock may lead to performance regression. But if we don't add a lock,the
task might see no nodes when changing cpuset's mems_allowed to some
non-overlapping set. In order to avoid it, we set all new allowed nodes,
then clear newly disallowed ones.
[lee.schermerhorn@hp.com:
The rework of mpol_new() to extract the adjusting of the node mask to
apply cpuset and mpol flags "context" breaks set_mempolicy() and mbind()
with MPOL_PREFERRED and a NULL nodemask--i.e., explicit local
allocation. Fix this by adding the check for MPOL_PREFERRED and empty
node mask to mpol_new_mpolicy().
Remove the now unneeded 'nodes = NULL' from mpol_new().
Note that mpol_new_mempolicy() is always called with a non-NULL
'nodes' parameter now that it has been removed from mpol_new().
Therefore, we don't need to test nodes for NULL before testing it for
'empty'. However, just to be extra paranoid, add a VM_BUG_ON() to
verify this assumption.]
[lee.schermerhorn@hp.com:
I don't think the function name 'mpol_new_mempolicy' is descriptive
enough to differentiate it from mpol_new().
This function applies cpuset set context, usually constraining nodes
to those allowed by the cpuset. However, when the 'RELATIVE_NODES flag
is set, it also translates the nodes. So I settled on
'mpol_set_nodemask()', because the comment block for mpol_new() mentions
that we need to call this function to "set nodes".
Some additional minor line length, whitespace and typo cleanup.]
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 15:31:49 -07:00
|
|
|
|
2010-03-23 13:35:41 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
cpuset,mm: update tasks' mems_allowed in time
Fix allocating page cache/slab object on the unallowed node when memory
spread is set by updating tasks' mems_allowed after its cpuset's mems is
changed.
In order to update tasks' mems_allowed in time, we must modify the code of
memory policy. Because the memory policy is applied in the process's
context originally. After applying this patch, one task directly
manipulates anothers mems_allowed, and we use alloc_lock in the
task_struct to protect mems_allowed and memory policy of the task.
But in the fast path, we didn't use lock to protect them, because adding a
lock may lead to performance regression. But if we don't add a lock,the
task might see no nodes when changing cpuset's mems_allowed to some
non-overlapping set. In order to avoid it, we set all new allowed nodes,
then clear newly disallowed ones.
[lee.schermerhorn@hp.com:
The rework of mpol_new() to extract the adjusting of the node mask to
apply cpuset and mpol flags "context" breaks set_mempolicy() and mbind()
with MPOL_PREFERRED and a NULL nodemask--i.e., explicit local
allocation. Fix this by adding the check for MPOL_PREFERRED and empty
node mask to mpol_new_mpolicy().
Remove the now unneeded 'nodes = NULL' from mpol_new().
Note that mpol_new_mempolicy() is always called with a non-NULL
'nodes' parameter now that it has been removed from mpol_new().
Therefore, we don't need to test nodes for NULL before testing it for
'empty'. However, just to be extra paranoid, add a VM_BUG_ON() to
verify this assumption.]
[lee.schermerhorn@hp.com:
I don't think the function name 'mpol_new_mempolicy' is descriptive
enough to differentiate it from mpol_new().
This function applies cpuset set context, usually constraining nodes
to those allowed by the cpuset. However, when the 'RELATIVE_NODES flag
is set, it also translates the nodes. So I settled on
'mpol_set_nodemask()', because the comment block for mpol_new() mentions
that we need to call this function to "set nodes".
Some additional minor line length, whitespace and typo cleanup.]
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 15:31:49 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
mempolicy: rework mempolicy Reference Counting [yet again]
After further discussion with Christoph Lameter, it has become clear that my
earlier attempts to clean up the mempolicy reference counting were a bit of
overkill in some areas, resulting in superflous ref/unref in what are usually
fast paths. In other areas, further inspection reveals that I botched the
unref for interleave policies.
A separate patch, suitable for upstream/stable trees, fixes up the known
errors in the previous attempt to fix reference counting.
This patch reworks the memory policy referencing counting and, one hopes,
simplifies the code. Maybe I'll get it right this time.
See the update to the numa_memory_policy.txt document for a discussion of
memory policy reference counting that motivates this patch.
Summary:
Lookup of mempolicy, based on (vma, address) need only add a reference for
shared policy, and we need only unref the policy when finished for shared
policies. So, this patch backs out all of the unneeded extra reference
counting added by my previous attempt. It then unrefs only shared policies
when we're finished with them, using the mpol_cond_put() [conditional put]
helper function introduced by this patch.
Note that shmem_swapin() calls read_swap_cache_async() with a dummy vma
containing just the policy. read_swap_cache_async() can call alloc_page_vma()
multiple times, so we can't let alloc_page_vma() unref the shared policy in
this case. To avoid this, we make a copy of any non-null shared policy and
remove the MPOL_F_SHARED flag from the copy. This copy occurs before reading
a page [or multiple pages] from swap, so the overhead should not be an issue
here.
I introduced a new static inline function "mpol_cond_copy()" to copy the
shared policy to an on-stack policy and remove the flags that would require a
conditional free. The current implementation of mpol_cond_copy() assumes that
the struct mempolicy contains no pointers to dynamically allocated structures
that must be duplicated or reference counted during copy.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:16 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2020-06-08 21:33:25 -07:00
|
|
|
|
userfaultfd: allow get_mempolicy(MPOL_F_NODE|MPOL_F_ADDR) to trigger userfaults
get_mempolicy(MPOL_F_NODE|MPOL_F_ADDR) called a get_user_pages that would
not be waiting for userfaults before failing and it would hit on a SIGBUS
instead. Using get_user_pages_locked/unlocked instead will allow
get_mempolicy to allow userfaults to resolve the fault and fill the hole,
before grabbing the node id of the page.
If the user calls get_mempolicy() with MPOL_F_ADDR | MPOL_F_NODE for an
address inside an area managed by uffd and there is no page at that
address, the page allocation from within get_mempolicy() will fail
because get_user_pages() does not allow for page fault retry required
for uffd; the user will get SIGBUS.
With this patch, the page fault will be resolved by the uffd and the
get_mempolicy() will continue normally.
Background:
Via code review, previously the syscall would have returned -EFAULT
(vm_fault_to_errno), now it will block and wait for an userfault (if
it's waken before the fault is resolved it'll still -EFAULT).
This way get_mempolicy will give a chance to an "unaware" app to be
compliant with userfaults.
The reason this visible change is that becoming "userfault compliant"
cannot regress anything: all other syscalls including read(2)/write(2)
had to become "userfault compliant" long time ago (that's one of the
things userfaultfd can do that PROT_NONE and trapping segfaults can't).
So this is just one more syscall that become "userfault compliant" like
all other major ones already were.
This has been happening on virtio-bridge dpdk process which just called
get_mempolicy on the guest space post live migration, but before the
memory had a chance to be migrated to destination.
I didn't run an strace to be able to show the -EFAULT going away, but
I've the confirmation of the below debug aid information (only visible
with CONFIG_DEBUG_VM=y) going away with the patch:
[20116.371461] FAULT_FLAG_ALLOW_RETRY missing 0
[20116.371464] CPU: 1 PID: 13381 Comm: vhost-events Not tainted 4.17.12-200.fc28.x86_64 #1
[20116.371465] Hardware name: LENOVO 20FAS2BN0A/20FAS2BN0A, BIOS N1CET54W (1.22 ) 02/10/2017
[20116.371466] Call Trace:
[20116.371473] dump_stack+0x5c/0x80
[20116.371476] handle_userfault.cold.37+0x1b/0x22
[20116.371479] ? remove_wait_queue+0x20/0x60
[20116.371481] ? poll_freewait+0x45/0xa0
[20116.371483] ? do_sys_poll+0x31c/0x520
[20116.371485] ? radix_tree_lookup_slot+0x1e/0x50
[20116.371488] shmem_getpage_gfp+0xce7/0xe50
[20116.371491] ? page_add_file_rmap+0x1a/0x2c0
[20116.371493] shmem_fault+0x78/0x1e0
[20116.371495] ? filemap_map_pages+0x3a1/0x450
[20116.371498] __do_fault+0x1f/0xc0
[20116.371500] __handle_mm_fault+0xe2e/0x12f0
[20116.371502] handle_mm_fault+0xda/0x200
[20116.371504] __get_user_pages+0x238/0x790
[20116.371506] get_user_pages+0x3e/0x50
[20116.371510] kernel_get_mempolicy+0x40b/0x700
[20116.371512] ? vfs_write+0x170/0x1a0
[20116.371515] __x64_sys_get_mempolicy+0x21/0x30
[20116.371517] do_syscall_64+0x5b/0x160
[20116.371520] entry_SYSCALL_64_after_hwframe+0x44/0xa9
The above harmless debug message (not a kernel crash, just a
dump_stack()) is shown with CONFIG_DEBUG_VM=y to more quickly identify
and improve kernel spots that may have to become "userfaultfd
compliant" like this one (without having to run an strace and search
for syscall misbehavior). Spots like the above are more closer to a
kernel bug for the non-cooperative usages that Mike focuses on, than
for for dpdk qemu-cooperative usages that reproduced it, but it's still
nicer to get this fixed for dpdk too.
The part of the patch that caused me to think is only the
implementation issue of mpol_get, but it looks like it should work safe
no matter the kind of mempolicy structure that is (the default static
policy also starts at 1 so it'll go to 2 and back to 1 without crashing
everything at 0).
[rppt@linux.vnet.ibm.com: changelog addition]
http://lkml.kernel.org/r/20180904073718.GA26916@rapoport-lnx
Link: http://lkml.kernel.org/r/20180831214848.23676-1-aarcange@redhat.com
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Tested-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Reviewed-by: Mike Rapoport <rppt@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-26 15:05:16 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: introduce CONFIG_NUMA_MIGRATION and simplify CONFIG_MIGRATION
CONFIG_MEMORY_HOTREMOVE, CONFIG_COMPACTION and CONFIG_CMA all select
CONFIG_MIGRATION, because they require it to work (users).
Only CONFIG_NUMA_BALANCING and CONFIG_BALLOON_MIGRATION depend on
CONFIG_MIGRATION. CONFIG_BALLOON_MIGRATION is not an actual user, but an
implementation of migration support, so the dependency is correct
(CONFIG_BALLOON_MIGRATION does not make any sense without
CONFIG_MIGRATION).
However, kconfig-language.rst clearly states "In general use select only
for non-visible symbols". So far CONFIG_MIGRATION is user-visible ...
and the dependencies rather confusing.
The whole reason why CONFIG_MIGRATION is user-visible is because of
CONFIG_NUMA: some users might want CONFIG_NUMA but not page migration
support.
Let's clean all that up by introducing a dedicated CONFIG_NUMA_MIGRATION
config option for that purpose only. Make CONFIG_NUMA_BALANCING that so
far depended on CONFIG_NUMA && CONFIG_MIGRATION to depend on
CONFIG_MIGRATION instead. CONFIG_NUMA_MIGRATION will depend on
CONFIG_NUMA && CONFIG_MMU.
CONFIG_NUMA_MIGRATION is user-visible and will default to "y". We use
that default so new configs will automatically enable it, just like it was
the case with CONFIG_MIGRATION. The downside is that some configs that
used to have CONFIG_MIGRATION=n might get it re-enabled by
CONFIG_NUMA_MIGRATION=y, which shouldn't be a problem.
CONFIG_MIGRATION is now a non-visible config option. Any code that select
CONFIG_MIGRATION (as before) must depend directly or indirectly on
CONFIG_MMU.
CONFIG_NUMA_MIGRATION is responsible for any NUMA migration code, which is
mempolicy migration code, memory-tiering code, and move_pages() code in
migrate.c. CONFIG_NUMA_BALANCING uses its functionality.
Note that this implies that with CONFIG_NUMA_MIGRATION=n, move_pages()
will not be available even though CONFIG_MIGRATION=y, which is an expected
change.
In migrate.c, we can remove the CONFIG_NUMA check as both
CONFIG_NUMA_MIGRATION and CONFIG_NUMA_BALANCING depend on it.
With this change, CONFIG_MIGRATION is an internal config, all users of
migration selects CONFIG_MIGRATION, and only CONFIG_BALLOON_MIGRATION
depends on it.
Link: https://lkml.kernel.org/r/20260319-config_migration-v1-2-42270124966f@kernel.org
Signed-off-by: David Hildenbrand (Arm) <david@kernel.org>
Reviewed-by: Lorenzo Stoakes (Oracle) <ljs@kernel.org>
Acked-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Jonathan Cameron <jonathan.cameron@huawei.com>
Cc: Albert Ou <aou@eecs.berkeley.edu>
Cc: Alexander Gordeev <agordeev@linux.ibm.com>
Cc: Alexandre Ghiti <alex@ghiti.fr>
Cc: Alistair Popple <apopple@nvidia.com>
Cc: "Borislav Petkov (AMD)" <bp@alien8.de>
Cc: Byungchul Park <byungchul@sk.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Christian Borntraeger <borntraeger@linux.ibm.com>
Cc: Gregory Price <gourry@gourry.net>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Huacai Chen <chenhuacai@kernel.org>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Joshua Hahn <joshua.hahnjy@gmail.com>
Cc: Liam Howlett <liam.howlett@oracle.com>
Cc: Madhavan Srinivasan <maddy@linux.ibm.com>
Cc: Matthew Brost <matthew.brost@intel.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Nicholas Piggin <npiggin@gmail.com>
Cc: Palmer Dabbelt <palmer@dabbelt.com>
Cc: Rakie Kim <rakie.kim@sk.com>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Sven Schnelle <svens@linux.ibm.com>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Cc: WANG Xuerui <kernel@xen0n.name>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2026-03-19 09:19:41 +01:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
2006-01-18 17:42:29 -08:00
|
|
|
|
2006-01-08 01:01:04 -08:00
|
|
|
|
|
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
|
|
|
|
|
2023-01-30 12:18:33 -08:00
|
|
|
|
mm: convert folio_likely_mapped_shared() to folio_maybe_mapped_shared()
Let's reuse our new MM ownership tracking infrastructure for large folios
to make folio_likely_mapped_shared() never return false negatives -- never
indicating "not mapped shared" although the folio *is* mapped shared.
With that, we can rename it to folio_maybe_mapped_shared() and get rid of
the dependency on the mapcount of the first folio page.
The semantics are now arguably clearer: no mixture of "false negatives"
and "false positives", only the remaining possibility for "false
positives".
Thoroughly document the new semantics. We might now detect that a large
folio is "maybe mapped shared" although it *no longer* is -- but once was.
Now, if more than two MMs mapped a folio at the same time, and the MM
mapping the folio exclusively at the end is not one tracked in the two
folio MM slots, we will detect the folio as "maybe mapped shared".
For anonymous folios, usually (except weird corner cases) all PTEs that
target a "maybe mapped shared" folio are R/O. As soon as a child process
would write to them (iow, actively use them), we would CoW and effectively
replace these PTEs. Most cases (below) are not expected to really matter
with large anonymous folios for this reason.
Most importantly, there will be no change at all for:
* small folios
* hugetlb folios
* PMD-mapped PMD-sized THPs (single mapping)
This change has the potential to affect existing callers of
folio_likely_mapped_shared() -> folio_maybe_mapped_shared():
(1) fs/proc/task_mmu.c: no change (hugetlb)
(2) khugepaged counts PTEs that target shared folios towards
max_ptes_shared (default: HPAGE_PMD_NR / 2), meaning we could skip a
collapse where we would have previously collapsed. This only applies
to anonymous folios and is not expected to matter in practice.
Worth noting that this change sorts out case (A) documented in
commit 1bafe96e89f0 ("mm/khugepaged: replace page_mapcount() check by
folio_likely_mapped_shared()") by removing the possibility for "false
negatives".
(3) MADV_COLD / MADV_PAGEOUT / MADV_FREE will not try splitting
PTE-mapped THPs that are considered shared but not fully covered by
the requested range, consequently not processing them.
PMD-mapped PMD-sized THP are not affected, or when all PTEs are
covered. These functions are usually only called on anon/file folios
that are exclusively mapped most of the time (no other file mappings
or no fork()), so the "false negatives" are not expected to matter in
practice.
(4) mbind() / migrate_pages() / move_pages() will refuse to migrate
shared folios unless MPOL_MF_MOVE_ALL is effective (requires
CAP_SYS_NICE). We will now reject some folios that could be migrated.
Similar to (3), especially with MPOL_MF_MOVE_ALL, so this is not
expected to matter in practice.
Note that cpuset_migrate_mm_workfn() calls do_migrate_pages() with
MPOL_MF_MOVE_ALL.
(5) NUMA hinting
mm/migrate.c:migrate_misplaced_folio_prepare() will skip file
folios that are probably shared libraries (-> "mapped shared" and
executable). This check would have detected it as a shared library at
some point (at least 3 MMs mapping it), so detecting it afterwards
does not sound wrong (still a shared library). Not expected to
matter.
mm/memory.c:numa_migrate_check() will indicate TNF_SHARED in
MAP_SHARED file mappings when encountering a shared folio. Similar
reasoning, not expected to matter.
mm/mprotect.c:change_pte_range() will skip folios detected as
shared in CoW mappings. Similarly, this is not expected to matter in
practice, but if it would ever be a problem we could relax that check
a bit (e.g., basing it on the average page-mapcount in a folio),
because it was only an optimization when many (e.g., 288) processes
were mapping the same folios -- see commit 859d4adc3415 ("mm: numa: do
not trap faults on shared data section pages.")
(6) mm/rmap.c:folio_referenced_one() will skip exclusive swapbacked
folios in dying processes. Applies to anonymous folios only. Without
"false negatives", we'll now skip all actually shared ones. Skipping
ones that are actually exclusive won't really matter, it's a pure
optimization, and is not expected to matter in practice.
In theory, one can detect the problematic scenario: folio_mapcount() > 0
and no folio MM slot is occupied ("state unknown"). One could reset the
MM slots while doing an rmap walk, which migration / folio split already
do when setting everything up. Further, when batching PTEs we might
naturally learn about a owner (e.g., folio_mapcount() == nr_ptes) and
could update the owner. However, we'll defer that until the scenarios
where it would really matter are clear.
Link: https://lkml.kernel.org/r/20250303163014.1128035-15-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Cc: Andy Lutomirks^H^Hski <luto@kernel.org>
Cc: Borislav Betkov <bp@alien8.de>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jann Horn <jannh@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: Liam Howlett <liam.howlett@oracle.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Michal Koutn <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: tejun heo <tj@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-03-03 17:30:07 +01:00
|
|
|
|
2024-02-27 21:15:48 +01:00
|
|
|
|
2006-01-08 01:01:04 -08:00
|
|
|
|
mm: convert folio_likely_mapped_shared() to folio_maybe_mapped_shared()
Let's reuse our new MM ownership tracking infrastructure for large folios
to make folio_likely_mapped_shared() never return false negatives -- never
indicating "not mapped shared" although the folio *is* mapped shared.
With that, we can rename it to folio_maybe_mapped_shared() and get rid of
the dependency on the mapcount of the first folio page.
The semantics are now arguably clearer: no mixture of "false negatives"
and "false positives", only the remaining possibility for "false
positives".
Thoroughly document the new semantics. We might now detect that a large
folio is "maybe mapped shared" although it *no longer* is -- but once was.
Now, if more than two MMs mapped a folio at the same time, and the MM
mapping the folio exclusively at the end is not one tracked in the two
folio MM slots, we will detect the folio as "maybe mapped shared".
For anonymous folios, usually (except weird corner cases) all PTEs that
target a "maybe mapped shared" folio are R/O. As soon as a child process
would write to them (iow, actively use them), we would CoW and effectively
replace these PTEs. Most cases (below) are not expected to really matter
with large anonymous folios for this reason.
Most importantly, there will be no change at all for:
* small folios
* hugetlb folios
* PMD-mapped PMD-sized THPs (single mapping)
This change has the potential to affect existing callers of
folio_likely_mapped_shared() -> folio_maybe_mapped_shared():
(1) fs/proc/task_mmu.c: no change (hugetlb)
(2) khugepaged counts PTEs that target shared folios towards
max_ptes_shared (default: HPAGE_PMD_NR / 2), meaning we could skip a
collapse where we would have previously collapsed. This only applies
to anonymous folios and is not expected to matter in practice.
Worth noting that this change sorts out case (A) documented in
commit 1bafe96e89f0 ("mm/khugepaged: replace page_mapcount() check by
folio_likely_mapped_shared()") by removing the possibility for "false
negatives".
(3) MADV_COLD / MADV_PAGEOUT / MADV_FREE will not try splitting
PTE-mapped THPs that are considered shared but not fully covered by
the requested range, consequently not processing them.
PMD-mapped PMD-sized THP are not affected, or when all PTEs are
covered. These functions are usually only called on anon/file folios
that are exclusively mapped most of the time (no other file mappings
or no fork()), so the "false negatives" are not expected to matter in
practice.
(4) mbind() / migrate_pages() / move_pages() will refuse to migrate
shared folios unless MPOL_MF_MOVE_ALL is effective (requires
CAP_SYS_NICE). We will now reject some folios that could be migrated.
Similar to (3), especially with MPOL_MF_MOVE_ALL, so this is not
expected to matter in practice.
Note that cpuset_migrate_mm_workfn() calls do_migrate_pages() with
MPOL_MF_MOVE_ALL.
(5) NUMA hinting
mm/migrate.c:migrate_misplaced_folio_prepare() will skip file
folios that are probably shared libraries (-> "mapped shared" and
executable). This check would have detected it as a shared library at
some point (at least 3 MMs mapping it), so detecting it afterwards
does not sound wrong (still a shared library). Not expected to
matter.
mm/memory.c:numa_migrate_check() will indicate TNF_SHARED in
MAP_SHARED file mappings when encountering a shared folio. Similar
reasoning, not expected to matter.
mm/mprotect.c:change_pte_range() will skip folios detected as
shared in CoW mappings. Similarly, this is not expected to matter in
practice, but if it would ever be a problem we could relax that check
a bit (e.g., basing it on the average page-mapcount in a folio),
because it was only an optimization when many (e.g., 288) processes
were mapping the same folios -- see commit 859d4adc3415 ("mm: numa: do
not trap faults on shared data section pages.")
(6) mm/rmap.c:folio_referenced_one() will skip exclusive swapbacked
folios in dying processes. Applies to anonymous folios only. Without
"false negatives", we'll now skip all actually shared ones. Skipping
ones that are actually exclusive won't really matter, it's a pure
optimization, and is not expected to matter in practice.
In theory, one can detect the problematic scenario: folio_mapcount() > 0
and no folio MM slot is occupied ("state unknown"). One could reset the
MM slots while doing an rmap walk, which migration / folio split already
do when setting everything up. Further, when batching PTEs we might
naturally learn about a owner (e.g., folio_mapcount() == nr_ptes) and
could update the owner. However, we'll defer that until the scenarios
where it would really matter are clear.
Link: https://lkml.kernel.org/r/20250303163014.1128035-15-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Cc: Andy Lutomirks^H^Hski <luto@kernel.org>
Cc: Borislav Betkov <bp@alien8.de>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jann Horn <jannh@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: Liam Howlett <liam.howlett@oracle.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Michal Koutn <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: tejun heo <tj@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-03-03 17:30:07 +01:00
|
|
|
|
mm: change to return bool for folio_isolate_lru()
Patch series "Change the return value for page isolation functions", v3.
Now the page isolation functions did not return a boolean to indicate
success or not, instead it will return a negative error when failed
to isolate a page. So below code used in most places seem a boolean
success/failure thing, which can confuse people whether the isolation
is successful.
if (folio_isolate_lru(folio))
continue;
Moreover the page isolation functions only return 0 or -EBUSY, and
most users did not care about the negative error except for few users,
thus we can convert all page isolation functions to return a boolean
value, which can remove the confusion to make code more clear.
No functional changes intended in this patch series.
This patch (of 4):
Now the folio_isolate_lru() did not return a boolean value to indicate
isolation success or not, however below code checking the return value can
make people think that it was a boolean success/failure thing, which makes
people easy to make mistakes (see the fix patch[1]).
if (folio_isolate_lru(folio))
continue;
Thus it's better to check the negative error value expilictly returned by
folio_isolate_lru(), which makes code more clear per Linus's
suggestion[2]. Moreover Matthew suggested we can convert the isolation
functions to return a boolean[3], since most users did not care about the
negative error value, and can also remove the confusing of checking return
value.
So this patch converts the folio_isolate_lru() to return a boolean value,
which means return 'true' to indicate the folio isolation is successful,
and 'false' means a failure to isolation. Meanwhile changing all users'
logic of checking the isolation state.
No functional changes intended.
[1] https://lore.kernel.org/all/20230131063206.28820-1-Kuan-Ying.Lee@mediatek.com/T/#u
[2] https://lore.kernel.org/all/CAHk-=wiBrY+O-4=2mrbVyxR+hOqfdJ=Do6xoucfJ9_5az01L4Q@mail.gmail.com/
[3] https://lore.kernel.org/all/Y+sTFqwMNAjDvxw3@casper.infradead.org/
Link: https://lkml.kernel.org/r/cover.1676424378.git.baolin.wang@linux.alibaba.com
Link: https://lkml.kernel.org/r/8a4e3679ed4196168efadf7ea36c038f2f7d5aa9.1676424378.git.baolin.wang@linux.alibaba.com
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Reviewed-by: SeongJae Park <sj@kernel.org>
Acked-by: David Hildenbrand <david@redhat.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Naoya Horiguchi <naoya.horiguchi@nec.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Shakeel Butt <shakeelb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-02-15 18:39:34 +08:00
|
|
|
|
2023-01-30 12:18:33 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
mm: mempolicy: handle vma with unmovable pages mapped correctly in mbind
When running syzkaller internally, we ran into the below bug on 4.9.x
kernel:
kernel BUG at mm/huge_memory.c:2124!
invalid opcode: 0000 [#1] SMP KASAN
CPU: 0 PID: 1518 Comm: syz-executor107 Not tainted 4.9.168+ #2
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 0.5.1 01/01/2011
task: ffff880067b34900 task.stack: ffff880068998000
RIP: split_huge_page_to_list+0x8fb/0x1030 mm/huge_memory.c:2124
Call Trace:
split_huge_page include/linux/huge_mm.h:100 [inline]
queue_pages_pte_range+0x7e1/0x1480 mm/mempolicy.c:538
walk_pmd_range mm/pagewalk.c:50 [inline]
walk_pud_range mm/pagewalk.c:90 [inline]
walk_pgd_range mm/pagewalk.c:116 [inline]
__walk_page_range+0x44a/0xdb0 mm/pagewalk.c:208
walk_page_range+0x154/0x370 mm/pagewalk.c:285
queue_pages_range+0x115/0x150 mm/mempolicy.c:694
do_mbind mm/mempolicy.c:1241 [inline]
SYSC_mbind+0x3c3/0x1030 mm/mempolicy.c:1370
SyS_mbind+0x46/0x60 mm/mempolicy.c:1352
do_syscall_64+0x1d2/0x600 arch/x86/entry/common.c:282
entry_SYSCALL_64_after_swapgs+0x5d/0xdb
Code: c7 80 1c 02 00 e8 26 0a 76 01 <0f> 0b 48 c7 c7 40 46 45 84 e8 4c
RIP [<ffffffff81895d6b>] split_huge_page_to_list+0x8fb/0x1030 mm/huge_memory.c:2124
RSP <ffff88006899f980>
with the below test:
uint64_t r[1] = {0xffffffffffffffff};
int main(void)
{
syscall(__NR_mmap, 0x20000000, 0x1000000, 3, 0x32, -1, 0);
intptr_t res = 0;
res = syscall(__NR_socket, 0x11, 3, 0x300);
if (res != -1)
r[0] = res;
*(uint32_t*)0x20000040 = 0x10000;
*(uint32_t*)0x20000044 = 1;
*(uint32_t*)0x20000048 = 0xc520;
*(uint32_t*)0x2000004c = 1;
syscall(__NR_setsockopt, r[0], 0x107, 0xd, 0x20000040, 0x10);
syscall(__NR_mmap, 0x20fed000, 0x10000, 0, 0x8811, r[0], 0);
*(uint64_t*)0x20000340 = 2;
syscall(__NR_mbind, 0x20ff9000, 0x4000, 0x4002, 0x20000340, 0x45d4, 3);
return 0;
}
Actually the test does:
mmap(0x20000000, 16777216, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_FIXED|MAP_ANONYMOUS, -1, 0) = 0x20000000
socket(AF_PACKET, SOCK_RAW, 768) = 3
setsockopt(3, SOL_PACKET, PACKET_TX_RING, {block_size=65536, block_nr=1, frame_size=50464, frame_nr=1}, 16) = 0
mmap(0x20fed000, 65536, PROT_NONE, MAP_SHARED|MAP_FIXED|MAP_POPULATE|MAP_DENYWRITE, 3, 0) = 0x20fed000
mbind(..., MPOL_MF_STRICT|MPOL_MF_MOVE) = 0
The setsockopt() would allocate compound pages (16 pages in this test)
for packet tx ring, then the mmap() would call packet_mmap() to map the
pages into the user address space specified by the mmap() call.
When calling mbind(), it would scan the vma to queue the pages for
migration to the new node. It would split any huge page since 4.9
doesn't support THP migration, however, the packet tx ring compound
pages are not THP and even not movable. So, the above bug is triggered.
However, the later kernel is not hit by this issue due to commit
d44d363f6578 ("mm: don't assume anonymous pages have SwapBacked flag"),
which just removes the PageSwapBacked check for a different reason.
But, there is a deeper issue. According to the semantic of mbind(), it
should return -EIO if MPOL_MF_MOVE or MPOL_MF_MOVE_ALL was specified and
MPOL_MF_STRICT was also specified, but the kernel was unable to move all
existing pages in the range. The tx ring of the packet socket is
definitely not movable, however, mbind() returns success for this case.
Although the most socket file associates with non-movable pages, but XDP
may have movable pages from gup. So, it sounds not fine to just check
the underlying file type of vma in vma_migratable().
Change migrate_page_add() to check if the page is movable or not, if it
is unmovable, just return -EIO. But do not abort pte walk immediately,
since there may be pages off LRU temporarily. We should migrate other
pages if MPOL_MF_MOVE* is specified. Set has_unmovable flag if some
paged could not be not moved, then return -EIO for mbind() eventually.
With this change the above test would return -EIO as expected.
[yang.shi@linux.alibaba.com: fix review comments from Vlastimil]
Link: http://lkml.kernel.org/r/1563556862-54056-3-git-send-email-yang.shi@linux.alibaba.com
Link: http://lkml.kernel.org/r/1561162809-59140-3-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
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 15:37:18 -07:00
|
|
|
|
2023-01-30 12:18:33 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
mm: mempolicy: handle vma with unmovable pages mapped correctly in mbind
When running syzkaller internally, we ran into the below bug on 4.9.x
kernel:
kernel BUG at mm/huge_memory.c:2124!
invalid opcode: 0000 [#1] SMP KASAN
CPU: 0 PID: 1518 Comm: syz-executor107 Not tainted 4.9.168+ #2
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 0.5.1 01/01/2011
task: ffff880067b34900 task.stack: ffff880068998000
RIP: split_huge_page_to_list+0x8fb/0x1030 mm/huge_memory.c:2124
Call Trace:
split_huge_page include/linux/huge_mm.h:100 [inline]
queue_pages_pte_range+0x7e1/0x1480 mm/mempolicy.c:538
walk_pmd_range mm/pagewalk.c:50 [inline]
walk_pud_range mm/pagewalk.c:90 [inline]
walk_pgd_range mm/pagewalk.c:116 [inline]
__walk_page_range+0x44a/0xdb0 mm/pagewalk.c:208
walk_page_range+0x154/0x370 mm/pagewalk.c:285
queue_pages_range+0x115/0x150 mm/mempolicy.c:694
do_mbind mm/mempolicy.c:1241 [inline]
SYSC_mbind+0x3c3/0x1030 mm/mempolicy.c:1370
SyS_mbind+0x46/0x60 mm/mempolicy.c:1352
do_syscall_64+0x1d2/0x600 arch/x86/entry/common.c:282
entry_SYSCALL_64_after_swapgs+0x5d/0xdb
Code: c7 80 1c 02 00 e8 26 0a 76 01 <0f> 0b 48 c7 c7 40 46 45 84 e8 4c
RIP [<ffffffff81895d6b>] split_huge_page_to_list+0x8fb/0x1030 mm/huge_memory.c:2124
RSP <ffff88006899f980>
with the below test:
uint64_t r[1] = {0xffffffffffffffff};
int main(void)
{
syscall(__NR_mmap, 0x20000000, 0x1000000, 3, 0x32, -1, 0);
intptr_t res = 0;
res = syscall(__NR_socket, 0x11, 3, 0x300);
if (res != -1)
r[0] = res;
*(uint32_t*)0x20000040 = 0x10000;
*(uint32_t*)0x20000044 = 1;
*(uint32_t*)0x20000048 = 0xc520;
*(uint32_t*)0x2000004c = 1;
syscall(__NR_setsockopt, r[0], 0x107, 0xd, 0x20000040, 0x10);
syscall(__NR_mmap, 0x20fed000, 0x10000, 0, 0x8811, r[0], 0);
*(uint64_t*)0x20000340 = 2;
syscall(__NR_mbind, 0x20ff9000, 0x4000, 0x4002, 0x20000340, 0x45d4, 3);
return 0;
}
Actually the test does:
mmap(0x20000000, 16777216, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_FIXED|MAP_ANONYMOUS, -1, 0) = 0x20000000
socket(AF_PACKET, SOCK_RAW, 768) = 3
setsockopt(3, SOL_PACKET, PACKET_TX_RING, {block_size=65536, block_nr=1, frame_size=50464, frame_nr=1}, 16) = 0
mmap(0x20fed000, 65536, PROT_NONE, MAP_SHARED|MAP_FIXED|MAP_POPULATE|MAP_DENYWRITE, 3, 0) = 0x20fed000
mbind(..., MPOL_MF_STRICT|MPOL_MF_MOVE) = 0
The setsockopt() would allocate compound pages (16 pages in this test)
for packet tx ring, then the mmap() would call packet_mmap() to map the
pages into the user address space specified by the mmap() call.
When calling mbind(), it would scan the vma to queue the pages for
migration to the new node. It would split any huge page since 4.9
doesn't support THP migration, however, the packet tx ring compound
pages are not THP and even not movable. So, the above bug is triggered.
However, the later kernel is not hit by this issue due to commit
d44d363f6578 ("mm: don't assume anonymous pages have SwapBacked flag"),
which just removes the PageSwapBacked check for a different reason.
But, there is a deeper issue. According to the semantic of mbind(), it
should return -EIO if MPOL_MF_MOVE or MPOL_MF_MOVE_ALL was specified and
MPOL_MF_STRICT was also specified, but the kernel was unable to move all
existing pages in the range. The tx ring of the packet socket is
definitely not movable, however, mbind() returns success for this case.
Although the most socket file associates with non-movable pages, but XDP
may have movable pages from gup. So, it sounds not fine to just check
the underlying file type of vma in vma_migratable().
Change migrate_page_add() to check if the page is movable or not, if it
is unmovable, just return -EIO. But do not abort pte walk immediately,
since there may be pages off LRU temporarily. We should migrate other
pages if MPOL_MF_MOVE* is specified. Set has_unmovable flag if some
paged could not be not moved, then return -EIO for mbind() eventually.
With this change the above test would return -EIO as expected.
[yang.shi@linux.alibaba.com: fix review comments from Vlastimil]
Link: http://lkml.kernel.org/r/1563556862-54056-3-git-send-email-yang.shi@linux.alibaba.com
Link: http://lkml.kernel.org/r/1561162809-59140-3-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
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 15:37:18 -07:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
vmscan: move isolate_lru_page() to vmscan.c
On large memory systems, the VM can spend way too much time scanning
through pages that it cannot (or should not) evict from memory. Not only
does it use up CPU time, but it also provokes lock contention and can
leave large systems under memory presure in a catatonic state.
This patch series improves VM scalability by:
1) putting filesystem backed, swap backed and unevictable pages
onto their own LRUs, so the system only scans the pages that it
can/should evict from memory
2) switching to two handed clock replacement for the anonymous LRUs,
so the number of pages that need to be scanned when the system
starts swapping is bound to a reasonable number
3) keeping unevictable pages off the LRU completely, so the
VM does not waste CPU time scanning them. ramfs, ramdisk,
SHM_LOCKED shared memory segments and mlock()ed VMA pages
are keept on the unevictable list.
This patch:
isolate_lru_page logically belongs to be in vmscan.c than migrate.c.
It is tough, because we don't need that function without memory migration
so there is a valid argument to have it in migrate.c. However a
subsequent patch needs to make use of it in the core mm, so we can happily
move it to vmscan.c.
Also, make the function a little more generic by not requiring that it
adds an isolated page to a given list. Callers can do that.
Note that we now have '__isolate_lru_page()', that does
something quite different, visible outside of vmscan.c
for use with memory controller. Methinks we need to
rationalize these names/purposes. --lts
[akpm@linux-foundation.org: fix mm/memory_hotplug.c build]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-18 20:26:09 -07:00
|
|
|
|
|
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
2006-01-08 01:01:04 -08:00
|
|
|
|
|
|
|
|
|
2006-01-08 01:00:51 -08:00
|
|
|
|
2006-02-01 03:05:40 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
|
|
|
|
|
2006-02-01 03:05:40 -08:00
|
|
|
|
|
|
|
|
|
2022-09-06 19:49:02 +00:00
|
|
|
|
2006-02-01 03:05:40 -08:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
|
|
|
|
|
2020-08-11 18:37:28 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-06 18:13:26 +08:00
|
|
|
|
2020-08-11 18:37:28 -07:00
|
|
|
|
2006-02-01 03:05:40 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-10-08 16:33:38 -07:00
|
|
|
|
mempolicy: mmap_lock is not needed while migrating folios
mbind(2) holds down_write of current task's mmap_lock throughout
(exclusive because it needs to set the new mempolicy on the vmas);
migrate_pages(2) holds down_read of pid's mmap_lock throughout.
They both hold mmap_lock across the internal migrate_pages(), under which
all new page allocations (huge or small) are made. I'm nervous about it;
and migrate_pages() certainly does not need mmap_lock itself. It's done
this way for mbind(2), because its page allocator is vma_alloc_folio() or
alloc_hugetlb_folio_vma(), both of which depend on vma and address.
Now that we have alloc_pages_mpol(), depending on (refcounted) memory
policy and interleave index, mbind(2) can be modified to use that or
alloc_hugetlb_folio_nodemask(), and then not need mmap_lock across the
internal migrate_pages() at all: add alloc_migration_target_by_mpol() to
replace mbind's new_page().
(After that change, alloc_hugetlb_folio_vma() is used by nothing but a
userfaultfd function: move it out of hugetlb.h and into the #ifdef.)
migrate_pages(2) has chosen its target node before migrating, so can
continue to use the standard alloc_migration_target(); but let it take and
drop mmap_lock just around migrate_to_node()'s queue_pages_range():
neither the node-to-node calculations nor the page migrations need it.
It seems unlikely, but it is conceivable that some userspace depends on
the kernel's mmap_lock exclusion here, instead of doing its own locking:
more likely in a testsuite than in real life. It is also possible, of
course, that some pages on the list will be munmapped by another thread
before they are migrated, or a newer memory policy applied to the range by
that time: but such races could happen before, as soon as mmap_lock was
dropped, so it does not appear to be a concern.
Link: https://lkml.kernel.org/r/21e564e8-269f-6a89-7ee2-fd612831c289@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:27:47 -07:00
|
|
|
|
|
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
2024-11-20 21:11:51 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: mmap_lock is not needed while migrating folios
mbind(2) holds down_write of current task's mmap_lock throughout
(exclusive because it needs to set the new mempolicy on the vmas);
migrate_pages(2) holds down_read of pid's mmap_lock throughout.
They both hold mmap_lock across the internal migrate_pages(), under which
all new page allocations (huge or small) are made. I'm nervous about it;
and migrate_pages() certainly does not need mmap_lock itself. It's done
this way for mbind(2), because its page allocator is vma_alloc_folio() or
alloc_hugetlb_folio_vma(), both of which depend on vma and address.
Now that we have alloc_pages_mpol(), depending on (refcounted) memory
policy and interleave index, mbind(2) can be modified to use that or
alloc_hugetlb_folio_nodemask(), and then not need mmap_lock across the
internal migrate_pages() at all: add alloc_migration_target_by_mpol() to
replace mbind's new_page().
(After that change, alloc_hugetlb_folio_vma() is used by nothing but a
userfaultfd function: move it out of hugetlb.h and into the #ifdef.)
migrate_pages(2) has chosen its target node before migrating, so can
continue to use the standard alloc_migration_target(); but let it take and
drop mmap_lock just around migrate_to_node()'s queue_pages_range():
neither the node-to-node calculations nor the page migrations need it.
It seems unlikely, but it is conceivable that some userspace depends on
the kernel's mmap_lock exclusion here, instead of doing its own locking:
more likely in a testsuite than in real life. It is also possible, of
course, that some pages on the list will be munmapped by another thread
before they are migrated, or a newer memory policy applied to the range by
that time: but such races could happen before, as soon as mmap_lock was
dropped, so it does not appear to be a concern.
Link: https://lkml.kernel.org/r/21e564e8-269f-6a89-7ee2-fd612831c289@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:27:47 -07:00
|
|
|
|
2006-02-01 03:05:40 -08:00
|
|
|
|
2010-10-26 14:21:29 -07:00
|
|
|
|
2020-08-11 18:37:28 -07:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
2010-10-26 14:21:29 -07:00
|
|
|
|
2013-09-11 14:22:03 -07:00
|
|
|
|
2010-10-26 14:21:29 -07:00
|
|
|
|
2006-06-23 02:03:53 -07:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
|
|
|
|
|
2006-02-01 03:05:40 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-01-08 01:00:51 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-05-29 15:06:24 -07:00
|
|
|
|
|
|
|
|
|
2006-01-08 01:00:51 -08:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
|
|
|
|
|
2006-02-01 03:05:40 -08:00
|
|
|
|
2006-01-08 01:00:51 -08:00
|
|
|
|
2021-05-04 18:36:57 -07:00
|
|
|
|
2008-11-06 12:53:30 -08:00
|
|
|
|
2010-03-05 13:41:59 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-03-01 15:06:02 +01:00
|
|
|
|
2010-03-05 13:41:59 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-01-08 01:00:55 -08:00
|
|
|
|
2012-05-29 15:06:24 -07:00
|
|
|
|
2006-02-01 03:05:40 -08:00
|
|
|
|
2021-05-04 18:40:12 -07:00
|
|
|
|
2013-11-12 15:07:39 -08:00
|
|
|
|
2006-02-01 03:05:40 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: do_migrate_pages() calls migrate_to_node() even if task is already on a correct node
While running an application that moves tasks from one cpuset to another
I noticed that it takes much longer and moves many more pages than
expected.
The reason for this is do_migrate_pages() does its best to preserve the
relative node differential from the first node of the cpuset because the
application may have been written with that in mind. If memory was
interleaved on the nodes of the source cpuset by an application
do_migrate_pages() will try its best to maintain that interleaving on
the nodes of the destination cpuset. This means copying the memory from
all source nodes to the destination nodes even if the source and
destination nodes overlap.
This is a problem for userspace NUMA placement tools. The amount of
time spent doing extra memory moves cancels out some of the NUMA
performance improvements. Furthermore, if the number of source and
destination nodes are to maintain the previous interleaving layout
anyway.
This patch changes do_migrate_pages() to only preserve the relative
layout inside the program if the number of NUMA nodes in the source and
destination mask are the same. If the number is different, we do a much
more efficient migration by not touching memory that is in an allowed
node.
This preserves the old behaviour for programs that want it, while
allowing a userspace NUMA placement tool to use the new, faster
migration. This improves performance in our tests by up to a factor of
7.
Without this change migrating tasks from a cpuset containing nodes 0-7
to a cpuset containing nodes 3-4, we migrate from ALL the nodes even if
they are in the both the source and destination nodesets:
Migrating 7 to 4
Migrating 6 to 3
Migrating 5 to 4
Migrating 4 to 3
Migrating 1 to 4
Migrating 3 to 4
Migrating 0 to 3
Migrating 2 to 3
With this change we only migrate from nodes that are not in the
destination nodesets:
Migrating 7 to 4
Migrating 6 to 3
Migrating 5 to 4
Migrating 2 to 3
Migrating 1 to 4
Migrating 0 to 3
Yet if we move from a cpuset containing nodes 2,3,4 to a cpuset
containing 3,4,5 we still do move everything so that we preserve the
desired NUMA offsets:
Migrating 4 to 5
Migrating 3 to 4
Migrating 2 to 3
As far as performance is concerned this simple patch improves the time
it takes to move 14, 20 and 26 large tasks from a cpuset containing
nodes 0-7 to a cpuset containing nodes 1 & 3 by up to a factor of 7.
Here are the timings with and without the patch:
BEFORE PATCH -- Move times: 59, 140, 651 seconds
============
Moving 14 tasks from nodes (0-7) to nodes (1,3)
numad(8780) do_migrate_pages (mm=0xffff88081d414400
from_nodes=0xffff880818c81d28 to_nodes=0xffff880818c81ce8 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d414400 source=0x7 dest=0x3 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d414400 source=0x6 dest=0x1 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d414400 source=0x5 dest=0x3 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d414400 source=0x4 dest=0x1 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d414400 source=0x2 dest=0x1 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d414400 source=0x1 dest=0x3 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d414400 source=0x0 dest=0x1 flags=0x4)
(Above moves repeated for each of the 14 tasks...)
PID 8890 moved to node(s) 1,3 in 59.2 seconds
Moving 20 tasks from nodes (0-7) to nodes (1,4-5)
numad(8780) do_migrate_pages (mm=0xffff88081d88c700
from_nodes=0xffff880818c81d28 to_nodes=0xffff880818c81ce8 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d88c700 source=0x7 dest=0x4 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d88c700 source=0x6 dest=0x1 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d88c700 source=0x3 dest=0x1 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d88c700 source=0x2 dest=0x5 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d88c700 source=0x1 dest=0x4 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d88c700 source=0x0 dest=0x1 flags=0x4)
(Above moves repeated for each of the 20 tasks...)
PID 8962 moved to node(s) 1,4-5 in 139.88 seconds
Moving 26 tasks from nodes (0-7) to nodes (1-3,5)
numad(8780) do_migrate_pages (mm=0xffff88081d5bc740
from_nodes=0xffff880818c81d28 to_nodes=0xffff880818c81ce8 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d5bc740 source=0x7 dest=0x5 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d5bc740 source=0x6 dest=0x3 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d5bc740 source=0x5 dest=0x2 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d5bc740 source=0x3 dest=0x5 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d5bc740 source=0x2 dest=0x3 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d5bc740 source=0x1 dest=0x2 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d5bc740 source=0x0 dest=0x1 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d5bc740 source=0x4 dest=0x1 flags=0x4)
(Above moves repeated for each of the 26 tasks...)
PID 9058 moved to node(s) 1-3,5 in 651.45 seconds
AFTER PATCH -- Move times: 42, 56, 93 seconds
===========
Moving 14 tasks from nodes (0-7) to nodes (5,7)
numad(33209) do_migrate_pages (mm=0xffff88101d5ff140
from_nodes=0xffff88101e7b5d28 to_nodes=0xffff88101e7b5ce8 flags=0x4)
numad(33209) migrate_to_node (mm=0xffff88101d5ff140 source=0x6 dest=0x5 flags=0x4)
numad(33209) migrate_to_node (mm=0xffff88101d5ff140 source=0x4 dest=0x5 flags=0x4)
numad(33209) migrate_to_node (mm=0xffff88101d5ff140 source=0x3 dest=0x7 flags=0x4)
numad(33209) migrate_to_node (mm=0xffff88101d5ff140 source=0x2 dest=0x5 flags=0x4)
numad(33209) migrate_to_node (mm=0xffff88101d5ff140 source=0x1 dest=0x7 flags=0x4)
numad(33209) migrate_to_node (mm=0xffff88101d5ff140 source=0x0 dest=0x5 flags=0x4)
(Above moves repeated for each of the 14 tasks...)
PID 33221 moved to node(s) 5,7 in 41.67 seconds
Moving 20 tasks from nodes (0-7) to nodes (1,3,5)
numad(33209) do_migrate_pages (mm=0xffff88101d6c37c0
from_nodes=0xffff88101e7b5d28 to_nodes=0xffff88101e7b5ce8 flags=0x4)
numad(33209) migrate_to_node (mm=0xffff88101d6c37c0 source=0x7 dest=0x3 flags=0x4)
numad(33209) migrate_to_node (mm=0xffff88101d6c37c0 source=0x6 dest=0x1 flags=0x4)
numad(33209) migrate_to_node (mm=0xffff88101d6c37c0 source=0x4 dest=0x3 flags=0x4)
numad(33209) migrate_to_node (mm=0xffff88101d6c37c0 source=0x2 dest=0x5 flags=0x4)
numad(33209) migrate_to_node (mm=0xffff88101d6c37c0 source=0x0 dest=0x1 flags=0x4)
(Above moves repeated for each of the 20 tasks...)
PID 33289 moved to node(s) 1,3,5 in 56.3 seconds
Moving 26 tasks from nodes (0-7) to nodes (1,3,5,7)
numad(33209) do_migrate_pages (mm=0xffff88101d924400
from_nodes=0xffff88101e7b5d28 to_nodes=0xffff88101e7b5ce8 flags=0x4)
numad(33209) migrate_to_node (mm=0xffff88101d924400 source=0x6 dest=0x5 flags=0x4)
numad(33209) migrate_to_node (mm=0xffff88101d924400 source=0x4 dest=0x1 flags=0x4)
numad(33209) migrate_to_node (mm=0xffff88101d924400 source=0x2 dest=0x5 flags=0x4)
numad(33209) migrate_to_node (mm=0xffff88101d924400 source=0x0 dest=0x1 flags=0x4)
(Above moves repeated for each of the 26 tasks...)
PID 33372 moved to node(s) 1,3,5,7 in 92.67 seconds
[akpm@linux-foundation.org: clean up comment layout]
Signed-off-by: Larry Woodman <lwoodman@redhat.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Acked-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-05-29 15:06:24 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-05-29 15:06:24 -07:00
|
|
|
|
|
|
|
|
|
mm: do_migrate_pages() calls migrate_to_node() even if task is already on a correct node
While running an application that moves tasks from one cpuset to another
I noticed that it takes much longer and moves many more pages than
expected.
The reason for this is do_migrate_pages() does its best to preserve the
relative node differential from the first node of the cpuset because the
application may have been written with that in mind. If memory was
interleaved on the nodes of the source cpuset by an application
do_migrate_pages() will try its best to maintain that interleaving on
the nodes of the destination cpuset. This means copying the memory from
all source nodes to the destination nodes even if the source and
destination nodes overlap.
This is a problem for userspace NUMA placement tools. The amount of
time spent doing extra memory moves cancels out some of the NUMA
performance improvements. Furthermore, if the number of source and
destination nodes are to maintain the previous interleaving layout
anyway.
This patch changes do_migrate_pages() to only preserve the relative
layout inside the program if the number of NUMA nodes in the source and
destination mask are the same. If the number is different, we do a much
more efficient migration by not touching memory that is in an allowed
node.
This preserves the old behaviour for programs that want it, while
allowing a userspace NUMA placement tool to use the new, faster
migration. This improves performance in our tests by up to a factor of
7.
Without this change migrating tasks from a cpuset containing nodes 0-7
to a cpuset containing nodes 3-4, we migrate from ALL the nodes even if
they are in the both the source and destination nodesets:
Migrating 7 to 4
Migrating 6 to 3
Migrating 5 to 4
Migrating 4 to 3
Migrating 1 to 4
Migrating 3 to 4
Migrating 0 to 3
Migrating 2 to 3
With this change we only migrate from nodes that are not in the
destination nodesets:
Migrating 7 to 4
Migrating 6 to 3
Migrating 5 to 4
Migrating 2 to 3
Migrating 1 to 4
Migrating 0 to 3
Yet if we move from a cpuset containing nodes 2,3,4 to a cpuset
containing 3,4,5 we still do move everything so that we preserve the
desired NUMA offsets:
Migrating 4 to 5
Migrating 3 to 4
Migrating 2 to 3
As far as performance is concerned this simple patch improves the time
it takes to move 14, 20 and 26 large tasks from a cpuset containing
nodes 0-7 to a cpuset containing nodes 1 & 3 by up to a factor of 7.
Here are the timings with and without the patch:
BEFORE PATCH -- Move times: 59, 140, 651 seconds
============
Moving 14 tasks from nodes (0-7) to nodes (1,3)
numad(8780) do_migrate_pages (mm=0xffff88081d414400
from_nodes=0xffff880818c81d28 to_nodes=0xffff880818c81ce8 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d414400 source=0x7 dest=0x3 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d414400 source=0x6 dest=0x1 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d414400 source=0x5 dest=0x3 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d414400 source=0x4 dest=0x1 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d414400 source=0x2 dest=0x1 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d414400 source=0x1 dest=0x3 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d414400 source=0x0 dest=0x1 flags=0x4)
(Above moves repeated for each of the 14 tasks...)
PID 8890 moved to node(s) 1,3 in 59.2 seconds
Moving 20 tasks from nodes (0-7) to nodes (1,4-5)
numad(8780) do_migrate_pages (mm=0xffff88081d88c700
from_nodes=0xffff880818c81d28 to_nodes=0xffff880818c81ce8 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d88c700 source=0x7 dest=0x4 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d88c700 source=0x6 dest=0x1 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d88c700 source=0x3 dest=0x1 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d88c700 source=0x2 dest=0x5 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d88c700 source=0x1 dest=0x4 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d88c700 source=0x0 dest=0x1 flags=0x4)
(Above moves repeated for each of the 20 tasks...)
PID 8962 moved to node(s) 1,4-5 in 139.88 seconds
Moving 26 tasks from nodes (0-7) to nodes (1-3,5)
numad(8780) do_migrate_pages (mm=0xffff88081d5bc740
from_nodes=0xffff880818c81d28 to_nodes=0xffff880818c81ce8 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d5bc740 source=0x7 dest=0x5 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d5bc740 source=0x6 dest=0x3 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d5bc740 source=0x5 dest=0x2 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d5bc740 source=0x3 dest=0x5 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d5bc740 source=0x2 dest=0x3 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d5bc740 source=0x1 dest=0x2 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d5bc740 source=0x0 dest=0x1 flags=0x4)
numad(8780) migrate_to_node (mm=0xffff88081d5bc740 source=0x4 dest=0x1 flags=0x4)
(Above moves repeated for each of the 26 tasks...)
PID 9058 moved to node(s) 1-3,5 in 651.45 seconds
AFTER PATCH -- Move times: 42, 56, 93 seconds
===========
Moving 14 tasks from nodes (0-7) to nodes (5,7)
numad(33209) do_migrate_pages (mm=0xffff88101d5ff140
from_nodes=0xffff88101e7b5d28 to_nodes=0xffff88101e7b5ce8 flags=0x4)
numad(33209) migrate_to_node (mm=0xffff88101d5ff140 source=0x6 dest=0x5 flags=0x4)
numad(33209) migrate_to_node (mm=0xffff88101d5ff140 source=0x4 dest=0x5 flags=0x4)
numad(33209) migrate_to_node (mm=0xffff88101d5ff140 source=0x3 dest=0x7 flags=0x4)
numad(33209) migrate_to_node (mm=0xffff88101d5ff140 source=0x2 dest=0x5 flags=0x4)
numad(33209) migrate_to_node (mm=0xffff88101d5ff140 source=0x1 dest=0x7 flags=0x4)
numad(33209) migrate_to_node (mm=0xffff88101d5ff140 source=0x0 dest=0x5 flags=0x4)
(Above moves repeated for each of the 14 tasks...)
PID 33221 moved to node(s) 5,7 in 41.67 seconds
Moving 20 tasks from nodes (0-7) to nodes (1,3,5)
numad(33209) do_migrate_pages (mm=0xffff88101d6c37c0
from_nodes=0xffff88101e7b5d28 to_nodes=0xffff88101e7b5ce8 flags=0x4)
numad(33209) migrate_to_node (mm=0xffff88101d6c37c0 source=0x7 dest=0x3 flags=0x4)
numad(33209) migrate_to_node (mm=0xffff88101d6c37c0 source=0x6 dest=0x1 flags=0x4)
numad(33209) migrate_to_node (mm=0xffff88101d6c37c0 source=0x4 dest=0x3 flags=0x4)
numad(33209) migrate_to_node (mm=0xffff88101d6c37c0 source=0x2 dest=0x5 flags=0x4)
numad(33209) migrate_to_node (mm=0xffff88101d6c37c0 source=0x0 dest=0x1 flags=0x4)
(Above moves repeated for each of the 20 tasks...)
PID 33289 moved to node(s) 1,3,5 in 56.3 seconds
Moving 26 tasks from nodes (0-7) to nodes (1,3,5,7)
numad(33209) do_migrate_pages (mm=0xffff88101d924400
from_nodes=0xffff88101e7b5d28 to_nodes=0xffff88101e7b5ce8 flags=0x4)
numad(33209) migrate_to_node (mm=0xffff88101d924400 source=0x6 dest=0x5 flags=0x4)
numad(33209) migrate_to_node (mm=0xffff88101d924400 source=0x4 dest=0x1 flags=0x4)
numad(33209) migrate_to_node (mm=0xffff88101d924400 source=0x2 dest=0x5 flags=0x4)
numad(33209) migrate_to_node (mm=0xffff88101d924400 source=0x0 dest=0x1 flags=0x4)
(Above moves repeated for each of the 26 tasks...)
PID 33372 moved to node(s) 1,3,5,7 in 92.67 seconds
[akpm@linux-foundation.org: clean up comment layout]
Signed-off-by: Larry Woodman <lwoodman@redhat.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Acked-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-05-29 15:06:24 -07:00
|
|
|
|
|
|
|
|
|
2012-05-29 15:06:24 -07:00
|
|
|
|
2006-02-01 03:05:40 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-11-12 15:07:39 -08:00
|
|
|
|
2006-02-01 03:05:40 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
2006-02-01 03:05:40 -08:00
|
|
|
|
|
|
|
|
|
2006-01-08 01:00:51 -08:00
|
|
|
|
mm: disable LRU pagevec during the migration temporarily
LRU pagevec holds refcount of pages until the pagevec are drained. It
could prevent migration since the refcount of the page is greater than
the expection in migration logic. To mitigate the issue, callers of
migrate_pages drains LRU pagevec via migrate_prep or lru_add_drain_all
before migrate_pages call.
However, it's not enough because pages coming into pagevec after the
draining call still could stay at the pagevec so it could keep
preventing page migration. Since some callers of migrate_pages have
retrial logic with LRU draining, the page would migrate at next trail
but it is still fragile in that it doesn't close the fundamental race
between upcoming LRU pages into pagvec and migration so the migration
failure could cause contiguous memory allocation failure in the end.
To close the race, this patch disables lru caches(i.e, pagevec) during
ongoing migration until migrate is done.
Since it's really hard to reproduce, I measured how many times
migrate_pages retried with force mode(it is about a fallback to a sync
migration) with below debug code.
int migrate_pages(struct list_head *from, new_page_t get_new_page,
..
..
if (rc && reason == MR_CONTIG_RANGE && pass > 2) {
printk(KERN_ERR, "pfn 0x%lx reason %d", page_to_pfn(page), rc);
dump_page(page, "fail to migrate");
}
The test was repeating android apps launching with cma allocation in
background every five seconds. Total cma allocation count was about 500
during the testing. With this patch, the dump_page count was reduced
from 400 to 30.
The new interface is also useful for memory hotplug which currently
drains lru pcp caches after each migration failure. This is rather
suboptimal as it has to disrupt others running during the operation.
With the new interface the operation happens only once. This is also in
line with pcp allocator cache which are disabled for the offlining as
well.
Link: https://lkml.kernel.org/r/20210319175127.886124-1-minchan@kernel.org
Signed-off-by: Minchan Kim <minchan@kernel.org>
Reviewed-by: Chris Goldsworthy <cgoldswo@codeaurora.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: John Dias <joaodias@google.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: David Hildenbrand <david@redhat.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Oliver Sang <oliver.sang@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-05-04 18:36:54 -07:00
|
|
|
|
2021-05-04 18:36:57 -07:00
|
|
|
|
2006-02-01 03:05:40 -08:00
|
|
|
|
|
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
2006-01-08 01:00:51 -08:00
|
|
|
|
|
|
|
|
|
Migration: find correct vma in new_vma_page()
We hit the BUG_ON() in mm/rmap.c:vma_address() when trying to migrate via
mbind(MPOL_MF_MOVE) a non-anon region that spans multiple vmas. For
anon-regions, we just fail to migrate any pages beyond the 1st vma in the
range.
This occurs because do_mbind() collects a list of pages to migrate by
calling check_range(). check_range() walks the task's mm, spanning vmas as
necessary, to collect the migratable pages into a list. Then, do_mbind()
calls migrate_pages() passing the list of pages, a function to allocate new
pages based on vma policy [new_vma_page()], and a pointer to the first vma
of the range.
For each page in the list, new_vma_page() calls page_address_in_vma()
passing the page and the vma [first in range] to obtain the address to get
for alloc_page_vma(). The page address is needed to get interleaving
policy correct. If the pages in the list come from multiple vmas,
eventually, new_page_address() will pass that page to page_address_in_vma()
with the incorrect vma. For !PageAnon pages, this will result in a bug
check in rmap.c:vma_address(). For anon pages, vma_address() will just
return EFAULT and fail the migration.
This patch modifies new_vma_page() to check the return value from
page_address_in_vma(). If the return value is EFAULT, new_vma_page()
searchs forward via vm_next for the vma that maps the page--i.e., that does
not return EFAULT. This assumes that the pages in the list handed to
migrate_pages() is in address order. This is currently case. The patch
documents this assumption in a new comment block for new_vma_page().
If new_vma_page() cannot locate the vma mapping the page in a forward
search in the mm, it will pass a NULL vma to alloc_page_vma(). This will
result in the allocation using the task policy, if any, else system default
policy. This situation is unlikely, but the patch documents this behavior
with a comment.
Note, this patch results in restarting from the first vma in a multi-vma
range each time new_vma_page() is called. If this is not acceptable, we
can make the vma argument a pointer, both in new_vma_page() and it's caller
unmap_and_move() so that the value held by the loop in migrate_pages()
always passes down the last vma in which a page was found. This will
require changes to all new_page_t functions passed to migrate_pages(). Is
this necessary?
For this patch to work, we can't bug check in vma_address() for pages
outside the argument vma. This patch removes the BUG_ON(). All other
callers [besides new_vma_page()] already check the return status.
Tested on x86_64, 4 node NUMA platform.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Acked-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-11-14 16:59:10 -08:00
|
|
|
|
mempolicy: mmap_lock is not needed while migrating folios
mbind(2) holds down_write of current task's mmap_lock throughout
(exclusive because it needs to set the new mempolicy on the vmas);
migrate_pages(2) holds down_read of pid's mmap_lock throughout.
They both hold mmap_lock across the internal migrate_pages(), under which
all new page allocations (huge or small) are made. I'm nervous about it;
and migrate_pages() certainly does not need mmap_lock itself. It's done
this way for mbind(2), because its page allocator is vma_alloc_folio() or
alloc_hugetlb_folio_vma(), both of which depend on vma and address.
Now that we have alloc_pages_mpol(), depending on (refcounted) memory
policy and interleave index, mbind(2) can be modified to use that or
alloc_hugetlb_folio_nodemask(), and then not need mmap_lock across the
internal migrate_pages() at all: add alloc_migration_target_by_mpol() to
replace mbind's new_page().
(After that change, alloc_hugetlb_folio_vma() is used by nothing but a
userfaultfd function: move it out of hugetlb.h and into the #ifdef.)
migrate_pages(2) has chosen its target node before migrating, so can
continue to use the standard alloc_migration_target(); but let it take and
drop mmap_lock just around migrate_to_node()'s queue_pages_range():
neither the node-to-node calculations nor the page migrations need it.
It seems unlikely, but it is conceivable that some userspace depends on
the kernel's mmap_lock exclusion here, instead of doing its own locking:
more likely in a testsuite than in real life. It is also possible, of
course, that some pages on the list will be munmapped by another thread
before they are migrated, or a newer memory policy applied to the range by
that time: but such races could happen before, as soon as mmap_lock was
dropped, so it does not appear to be a concern.
Link: https://lkml.kernel.org/r/21e564e8-269f-6a89-7ee2-fd612831c289@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:27:47 -07:00
|
|
|
|
Migration: find correct vma in new_vma_page()
We hit the BUG_ON() in mm/rmap.c:vma_address() when trying to migrate via
mbind(MPOL_MF_MOVE) a non-anon region that spans multiple vmas. For
anon-regions, we just fail to migrate any pages beyond the 1st vma in the
range.
This occurs because do_mbind() collects a list of pages to migrate by
calling check_range(). check_range() walks the task's mm, spanning vmas as
necessary, to collect the migratable pages into a list. Then, do_mbind()
calls migrate_pages() passing the list of pages, a function to allocate new
pages based on vma policy [new_vma_page()], and a pointer to the first vma
of the range.
For each page in the list, new_vma_page() calls page_address_in_vma()
passing the page and the vma [first in range] to obtain the address to get
for alloc_page_vma(). The page address is needed to get interleaving
policy correct. If the pages in the list come from multiple vmas,
eventually, new_page_address() will pass that page to page_address_in_vma()
with the incorrect vma. For !PageAnon pages, this will result in a bug
check in rmap.c:vma_address(). For anon pages, vma_address() will just
return EFAULT and fail the migration.
This patch modifies new_vma_page() to check the return value from
page_address_in_vma(). If the return value is EFAULT, new_vma_page()
searchs forward via vm_next for the vma that maps the page--i.e., that does
not return EFAULT. This assumes that the pages in the list handed to
migrate_pages() is in address order. This is currently case. The patch
documents this assumption in a new comment block for new_vma_page().
If new_vma_page() cannot locate the vma mapping the page in a forward
search in the mm, it will pass a NULL vma to alloc_page_vma(). This will
result in the allocation using the task policy, if any, else system default
policy. This situation is unlikely, but the patch documents this behavior
with a comment.
Note, this patch results in restarting from the first vma in a multi-vma
range each time new_vma_page() is called. If this is not acceptable, we
can make the vma argument a pointer, both in new_vma_page() and it's caller
unmap_and_move() so that the value held by the loop in migrate_pages()
always passes down the last vma in which a page was found. This will
require changes to all new_page_t functions passed to migrate_pages(). Is
this necessary?
For this patch to work, we can't bug check in vma_address() for pages
outside the argument vma. This patch removes the BUG_ON(). All other
callers [besides new_vma_page()] already check the return status.
Tested on x86_64, 4 node NUMA platform.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Acked-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-11-14 16:59:10 -08:00
|
|
|
|
mempolicy: mmap_lock is not needed while migrating folios
mbind(2) holds down_write of current task's mmap_lock throughout
(exclusive because it needs to set the new mempolicy on the vmas);
migrate_pages(2) holds down_read of pid's mmap_lock throughout.
They both hold mmap_lock across the internal migrate_pages(), under which
all new page allocations (huge or small) are made. I'm nervous about it;
and migrate_pages() certainly does not need mmap_lock itself. It's done
this way for mbind(2), because its page allocator is vma_alloc_folio() or
alloc_hugetlb_folio_vma(), both of which depend on vma and address.
Now that we have alloc_pages_mpol(), depending on (refcounted) memory
policy and interleave index, mbind(2) can be modified to use that or
alloc_hugetlb_folio_nodemask(), and then not need mmap_lock across the
internal migrate_pages() at all: add alloc_migration_target_by_mpol() to
replace mbind's new_page().
(After that change, alloc_hugetlb_folio_vma() is used by nothing but a
userfaultfd function: move it out of hugetlb.h and into the #ifdef.)
migrate_pages(2) has chosen its target node before migrating, so can
continue to use the standard alloc_migration_target(); but let it take and
drop mmap_lock just around migrate_to_node()'s queue_pages_range():
neither the node-to-node calculations nor the page migrations need it.
It seems unlikely, but it is conceivable that some userspace depends on
the kernel's mmap_lock exclusion here, instead of doing its own locking:
more likely in a testsuite than in real life. It is also possible, of
course, that some pages on the list will be munmapped by another thread
before they are migrated, or a newer memory policy applied to the range by
that time: but such races could happen before, as soon as mmap_lock was
dropped, so it does not appear to be a concern.
Link: https://lkml.kernel.org/r/21e564e8-269f-6a89-7ee2-fd612831c289@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:27:47 -07:00
|
|
|
|
|
|
|
|
|
2006-06-23 02:03:53 -07:00
|
|
|
|
2023-10-03 02:29:00 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: mmap_lock is not needed while migrating folios
mbind(2) holds down_write of current task's mmap_lock throughout
(exclusive because it needs to set the new mempolicy on the vmas);
migrate_pages(2) holds down_read of pid's mmap_lock throughout.
They both hold mmap_lock across the internal migrate_pages(), under which
all new page allocations (huge or small) are made. I'm nervous about it;
and migrate_pages() certainly does not need mmap_lock itself. It's done
this way for mbind(2), because its page allocator is vma_alloc_folio() or
alloc_hugetlb_folio_vma(), both of which depend on vma and address.
Now that we have alloc_pages_mpol(), depending on (refcounted) memory
policy and interleave index, mbind(2) can be modified to use that or
alloc_hugetlb_folio_nodemask(), and then not need mmap_lock across the
internal migrate_pages() at all: add alloc_migration_target_by_mpol() to
replace mbind's new_page().
(After that change, alloc_hugetlb_folio_vma() is used by nothing but a
userfaultfd function: move it out of hugetlb.h and into the #ifdef.)
migrate_pages(2) has chosen its target node before migrating, so can
continue to use the standard alloc_migration_target(); but let it take and
drop mmap_lock just around migrate_to_node()'s queue_pages_range():
neither the node-to-node calculations nor the page migrations need it.
It seems unlikely, but it is conceivable that some userspace depends on
the kernel's mmap_lock exclusion here, instead of doing its own locking:
more likely in a testsuite than in real life. It is also possible, of
course, that some pages on the list will be munmapped by another thread
before they are migrated, or a newer memory policy applied to the range by
that time: but such races could happen before, as soon as mmap_lock was
dropped, so it does not appear to be a concern.
Link: https://lkml.kernel.org/r/21e564e8-269f-6a89-7ee2-fd612831c289@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:27:47 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-06-23 02:03:53 -07:00
|
|
|
|
mempolicy: mmap_lock is not needed while migrating folios
mbind(2) holds down_write of current task's mmap_lock throughout
(exclusive because it needs to set the new mempolicy on the vmas);
migrate_pages(2) holds down_read of pid's mmap_lock throughout.
They both hold mmap_lock across the internal migrate_pages(), under which
all new page allocations (huge or small) are made. I'm nervous about it;
and migrate_pages() certainly does not need mmap_lock itself. It's done
this way for mbind(2), because its page allocator is vma_alloc_folio() or
alloc_hugetlb_folio_vma(), both of which depend on vma and address.
Now that we have alloc_pages_mpol(), depending on (refcounted) memory
policy and interleave index, mbind(2) can be modified to use that or
alloc_hugetlb_folio_nodemask(), and then not need mmap_lock across the
internal migrate_pages() at all: add alloc_migration_target_by_mpol() to
replace mbind's new_page().
(After that change, alloc_hugetlb_folio_vma() is used by nothing but a
userfaultfd function: move it out of hugetlb.h and into the #ifdef.)
migrate_pages(2) has chosen its target node before migrating, so can
continue to use the standard alloc_migration_target(); but let it take and
drop mmap_lock just around migrate_to_node()'s queue_pages_range():
neither the node-to-node calculations nor the page migrations need it.
It seems unlikely, but it is conceivable that some userspace depends on
the kernel's mmap_lock exclusion here, instead of doing its own locking:
more likely in a testsuite than in real life. It is also possible, of
course, that some pages on the list will be munmapped by another thread
before they are migrated, or a newer memory policy applied to the range by
that time: but such races could happen before, as soon as mmap_lock was
dropped, so it does not appear to be a concern.
Link: https://lkml.kernel.org/r/21e564e8-269f-6a89-7ee2-fd612831c289@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:27:47 -07:00
|
|
|
|
|
|
|
|
|
Migration: find correct vma in new_vma_page()
We hit the BUG_ON() in mm/rmap.c:vma_address() when trying to migrate via
mbind(MPOL_MF_MOVE) a non-anon region that spans multiple vmas. For
anon-regions, we just fail to migrate any pages beyond the 1st vma in the
range.
This occurs because do_mbind() collects a list of pages to migrate by
calling check_range(). check_range() walks the task's mm, spanning vmas as
necessary, to collect the migratable pages into a list. Then, do_mbind()
calls migrate_pages() passing the list of pages, a function to allocate new
pages based on vma policy [new_vma_page()], and a pointer to the first vma
of the range.
For each page in the list, new_vma_page() calls page_address_in_vma()
passing the page and the vma [first in range] to obtain the address to get
for alloc_page_vma(). The page address is needed to get interleaving
policy correct. If the pages in the list come from multiple vmas,
eventually, new_page_address() will pass that page to page_address_in_vma()
with the incorrect vma. For !PageAnon pages, this will result in a bug
check in rmap.c:vma_address(). For anon pages, vma_address() will just
return EFAULT and fail the migration.
This patch modifies new_vma_page() to check the return value from
page_address_in_vma(). If the return value is EFAULT, new_vma_page()
searchs forward via vm_next for the vma that maps the page--i.e., that does
not return EFAULT. This assumes that the pages in the list handed to
migrate_pages() is in address order. This is currently case. The patch
documents this assumption in a new comment block for new_vma_page().
If new_vma_page() cannot locate the vma mapping the page in a forward
search in the mm, it will pass a NULL vma to alloc_page_vma(). This will
result in the allocation using the task policy, if any, else system default
policy. This situation is unlikely, but the patch documents this behavior
with a comment.
Note, this patch results in restarting from the first vma in a multi-vma
range each time new_vma_page() is called. If this is not acceptable, we
can make the vma argument a pointer, both in new_vma_page() and it's caller
unmap_and_move() so that the value held by the loop in migrate_pages()
always passes down the last vma in which a page was found. This will
require changes to all new_page_t functions passed to migrate_pages(). Is
this necessary?
For this patch to work, we can't bug check in vma_address() for pages
outside the argument vma. This patch removes the BUG_ON(). All other
callers [besides new_vma_page()] already check the return status.
Tested on x86_64, 4 node NUMA platform.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Acked-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-11-14 16:59:10 -08:00
|
|
|
|
2023-01-25 09:05:33 -08:00
|
|
|
|
mempolicy: mmap_lock is not needed while migrating folios
mbind(2) holds down_write of current task's mmap_lock throughout
(exclusive because it needs to set the new mempolicy on the vmas);
migrate_pages(2) holds down_read of pid's mmap_lock throughout.
They both hold mmap_lock across the internal migrate_pages(), under which
all new page allocations (huge or small) are made. I'm nervous about it;
and migrate_pages() certainly does not need mmap_lock itself. It's done
this way for mbind(2), because its page allocator is vma_alloc_folio() or
alloc_hugetlb_folio_vma(), both of which depend on vma and address.
Now that we have alloc_pages_mpol(), depending on (refcounted) memory
policy and interleave index, mbind(2) can be modified to use that or
alloc_hugetlb_folio_nodemask(), and then not need mmap_lock across the
internal migrate_pages() at all: add alloc_migration_target_by_mpol() to
replace mbind's new_page().
(After that change, alloc_hugetlb_folio_vma() is used by nothing but a
userfaultfd function: move it out of hugetlb.h and into the #ifdef.)
migrate_pages(2) has chosen its target node before migrating, so can
continue to use the standard alloc_migration_target(); but let it take and
drop mmap_lock just around migrate_to_node()'s queue_pages_range():
neither the node-to-node calculations nor the page migrations need it.
It seems unlikely, but it is conceivable that some userspace depends on
the kernel's mmap_lock exclusion here, instead of doing its own locking:
more likely in a testsuite than in real life. It is also possible, of
course, that some pages on the list will be munmapped by another thread
before they are migrated, or a newer memory policy applied to the range by
that time: but such races could happen before, as soon as mmap_lock was
dropped, so it does not appear to be a concern.
Link: https://lkml.kernel.org/r/21e564e8-269f-6a89-7ee2-fd612831c289@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:27:47 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: hugetlb: make the hugetlb migration strategy consistent
As discussed in previous thread [1], there is an inconsistency when
handing hugetlb migration. When handling the migration of freed hugetlb,
it prevents fallback to other NUMA nodes in
alloc_and_dissolve_hugetlb_folio(). However, when dealing with in-use
hugetlb, it allows fallback to other NUMA nodes in
alloc_hugetlb_folio_nodemask(), which can break the per-node hugetlb pool
and might result in unexpected failures when node bound workloads doesn't
get what is asssumed available.
To make hugetlb migration strategy more clear, we should list all the scenarios
of hugetlb migration and analyze whether allocation fallback is permitted:
1) Memory offline: will call dissolve_free_huge_pages() to free the
freed hugetlb, and call do_migrate_range() to migrate the in-use
hugetlb. Both can break the per-node hugetlb pool, but as this is an
explicit offlining operation, no better choice. So should allow the
hugetlb allocation fallback.
2) Memory failure: same as memory offline. Should allow fallback to a
different node might be the only option to handle it, otherwise the
impact of poisoned memory can be amplified.
3) Longterm pinning: will call migrate_longterm_unpinnable_pages() to
migrate in-use and not-longterm-pinnable hugetlb, which can break the
per-node pool. But we should fail to longterm pinning if can not
allocate on current node to avoid breaking the per-node pool.
4) Syscalls (mbind, migrate_pages, move_pages): these are explicit
users operation to move pages to other nodes, so fallback to other
nodes should not be prohibited.
5) alloc_contig_range: used by CMA allocation and virtio-mem
fake-offline to allocate given range of pages. Now the freed hugetlb
migration is not allowed to fallback, to keep consistency, the in-use
hugetlb migration should be also not allowed to fallback.
6) alloc_contig_pages: used by kfence, pgtable_debug etc. The strategy
should be consistent with that of alloc_contig_range().
Based on the analysis of the various scenarios above, introducing a new
helper to determine whether fallback is permitted according to the
migration reason..
[1] https://lore.kernel.org/all/6f26ce22d2fcd523418a085f2c588fe0776d46e7.1706794035.git.baolin.wang@linux.alibaba.com/
Link: https://lkml.kernel.org/r/3519fcd41522817307a05b40fb551e2e17e68101.1709719720.git.baolin.wang@linux.alibaba.com
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Naoya Horiguchi <nao.horiguchi@gmail.com>
Cc: Oscar Salvador <osalvador@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-06 18:13:27 +08:00
|
|
|
|
|
|
|
|
|
2023-01-25 09:05:33 -08:00
|
|
|
|
2017-09-08 16:11:08 -07:00
|
|
|
|
2022-04-04 15:23:39 -04:00
|
|
|
|
|
|
|
|
|
mempolicy: mmap_lock is not needed while migrating folios
mbind(2) holds down_write of current task's mmap_lock throughout
(exclusive because it needs to set the new mempolicy on the vmas);
migrate_pages(2) holds down_read of pid's mmap_lock throughout.
They both hold mmap_lock across the internal migrate_pages(), under which
all new page allocations (huge or small) are made. I'm nervous about it;
and migrate_pages() certainly does not need mmap_lock itself. It's done
this way for mbind(2), because its page allocator is vma_alloc_folio() or
alloc_hugetlb_folio_vma(), both of which depend on vma and address.
Now that we have alloc_pages_mpol(), depending on (refcounted) memory
policy and interleave index, mbind(2) can be modified to use that or
alloc_hugetlb_folio_nodemask(), and then not need mmap_lock across the
internal migrate_pages() at all: add alloc_migration_target_by_mpol() to
replace mbind's new_page().
(After that change, alloc_hugetlb_folio_vma() is used by nothing but a
userfaultfd function: move it out of hugetlb.h and into the #ifdef.)
migrate_pages(2) has chosen its target node before migrating, so can
continue to use the standard alloc_migration_target(); but let it take and
drop mmap_lock just around migrate_to_node()'s queue_pages_range():
neither the node-to-node calculations nor the page migrations need it.
It seems unlikely, but it is conceivable that some userspace depends on
the kernel's mmap_lock exclusion here, instead of doing its own locking:
more likely in a testsuite than in real life. It is also possible, of
course, that some pages on the list will be munmapped by another thread
before they are migrated, or a newer memory policy applied to the range by
that time: but such races could happen before, as soon as mmap_lock was
dropped, so it does not appear to be a concern.
Link: https://lkml.kernel.org/r/21e564e8-269f-6a89-7ee2-fd612831c289@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:27:47 -07:00
|
|
|
|
|
|
|
|
|
2022-04-04 15:23:39 -04:00
|
|
|
|
2024-05-15 15:07:08 +08:00
|
|
|
|
2006-06-23 02:03:53 -07:00
|
|
|
|
2006-03-22 00:09:12 -08:00
|
|
|
|
|
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
2006-03-22 00:09:12 -08:00
|
|
|
|
|
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
2006-03-22 00:09:12 -08:00
|
|
|
|
|
|
|
|
|
2012-05-29 15:06:24 -07:00
|
|
|
|
|
|
|
|
|
2006-03-22 00:09:12 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-06-23 02:03:53 -07:00
|
|
|
|
mempolicy: mmap_lock is not needed while migrating folios
mbind(2) holds down_write of current task's mmap_lock throughout
(exclusive because it needs to set the new mempolicy on the vmas);
migrate_pages(2) holds down_read of pid's mmap_lock throughout.
They both hold mmap_lock across the internal migrate_pages(), under which
all new page allocations (huge or small) are made. I'm nervous about it;
and migrate_pages() certainly does not need mmap_lock itself. It's done
this way for mbind(2), because its page allocator is vma_alloc_folio() or
alloc_hugetlb_folio_vma(), both of which depend on vma and address.
Now that we have alloc_pages_mpol(), depending on (refcounted) memory
policy and interleave index, mbind(2) can be modified to use that or
alloc_hugetlb_folio_nodemask(), and then not need mmap_lock across the
internal migrate_pages() at all: add alloc_migration_target_by_mpol() to
replace mbind's new_page().
(After that change, alloc_hugetlb_folio_vma() is used by nothing but a
userfaultfd function: move it out of hugetlb.h and into the #ifdef.)
migrate_pages(2) has chosen its target node before migrating, so can
continue to use the standard alloc_migration_target(); but let it take and
drop mmap_lock just around migrate_to_node()'s queue_pages_range():
neither the node-to-node calculations nor the page migrations need it.
It seems unlikely, but it is conceivable that some userspace depends on
the kernel's mmap_lock exclusion here, instead of doing its own locking:
more likely in a testsuite than in real life. It is also possible, of
course, that some pages on the list will be munmapped by another thread
before they are migrated, or a newer memory policy applied to the range by
that time: but such races could happen before, as soon as mmap_lock was
dropped, so it does not appear to be a concern.
Link: https://lkml.kernel.org/r/21e564e8-269f-6a89-7ee2-fd612831c289@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:27:47 -07:00
|
|
|
|
|
|
|
|
|
2006-06-23 02:03:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-03-22 00:09:12 -08:00
|
|
|
|
|
|
|
|
|
2007-10-16 01:26:26 -07:00
|
|
|
|
2008-04-28 02:12:25 -07:00
|
|
|
|
|
|
|
|
|
2006-01-08 01:01:04 -08:00
|
|
|
|
|
|
|
|
|
2023-04-10 11:22:05 -04:00
|
|
|
|
|
|
|
|
|
2023-10-03 02:29:00 -07:00
|
|
|
|
2006-01-08 01:01:04 -08:00
|
|
|
|
|
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
|
|
|
|
|
2006-01-08 01:01:04 -08:00
|
|
|
|
|
|
|
|
|
2012-10-25 14:16:32 +02:00
|
|
|
|
2006-01-08 01:01:04 -08:00
|
|
|
|
2006-03-14 19:50:21 -08:00
|
|
|
|
2006-01-08 01:01:04 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-09-13 01:55:05 +00:00
|
|
|
|
2006-01-08 01:01:04 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-04-28 02:12:25 -07:00
|
|
|
|
2006-01-08 01:01:04 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
2021-05-04 18:36:57 -07:00
|
|
|
|
mm: make set_mempolicy(MPOL_INTERLEAV) N_HIGH_MEMORY aware
At first, init_task's mems_allowed is initialized as this.
init_task->mems_allowed == node_state[N_POSSIBLE]
And cpuset's top_cpuset mask is initialized as this
top_cpuset->mems_allowed = node_state[N_HIGH_MEMORY]
Before 2.6.29:
policy's mems_allowed is initialized as this.
1. update tasks->mems_allowed by its cpuset->mems_allowed.
2. policy->mems_allowed = nodes_and(tasks->mems_allowed, user's mask)
Updating task's mems_allowed in reference to top_cpuset's one.
cpuset's mems_allowed is aware of N_HIGH_MEMORY, always.
In 2.6.30: After commit 58568d2a8215cb6f55caf2332017d7bdff954e1c
("cpuset,mm: update tasks' mems_allowed in time"), policy's mems_allowed
is initialized as this.
1. policy->mems_allowd = nodes_and(task->mems_allowed, user's mask)
Here, if task is in top_cpuset, task->mems_allowed is not updated from
init's one. Assume user excutes command as #numactrl --interleave=all
,....
policy->mems_allowd = nodes_and(N_POSSIBLE, ALL_SET_MASK)
Then, policy's mems_allowd can includes a possible node, which has no pgdat.
MPOL's INTERLEAVE just scans nodemask of task->mems_allowd and access this
directly.
NODE_DATA(nid)->zonelist even if NODE_DATA(nid)==NULL
Then, what's we need is making policy->mems_allowed be aware of
N_HIGH_MEMORY. This patch does that. But to do so, extra nodemask will
be on statck. Because I know cpumask has a new interface of
CPUMASK_ALLOC(), I added it to node.
This patch stands on old behavior. But I feel this fix itself is just a
Band-Aid. But to do fundametal fix, we have to take care of memory
hotplug and it takes time. (task->mems_allowd should be N_HIGH_MEMORY, I
think.)
mpol_set_nodemask() should be aware of N_HIGH_MEMORY and policy's nodemask
should be includes only online nodes.
In old behavior, this is guaranteed by frequent reference to cpuset's
code. Now, most of them are removed and mempolicy has to check it by
itself.
To do check, a few nodemask_t will be used for calculating nodemask. But,
size of nodemask_t can be big and it's not good to allocate them on stack.
Now, cpumask_t has CPUMASK_ALLOC/FREE an easy code for get scratch area.
NODEMASK_ALLOC/FREE shoudl be there.
[akpm@linux-foundation.org: cleanups & tweaks]
Tested-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-08-06 15:07:33 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-06-08 21:33:25 -07:00
|
|
|
|
mm: make set_mempolicy(MPOL_INTERLEAV) N_HIGH_MEMORY aware
At first, init_task's mems_allowed is initialized as this.
init_task->mems_allowed == node_state[N_POSSIBLE]
And cpuset's top_cpuset mask is initialized as this
top_cpuset->mems_allowed = node_state[N_HIGH_MEMORY]
Before 2.6.29:
policy's mems_allowed is initialized as this.
1. update tasks->mems_allowed by its cpuset->mems_allowed.
2. policy->mems_allowed = nodes_and(tasks->mems_allowed, user's mask)
Updating task's mems_allowed in reference to top_cpuset's one.
cpuset's mems_allowed is aware of N_HIGH_MEMORY, always.
In 2.6.30: After commit 58568d2a8215cb6f55caf2332017d7bdff954e1c
("cpuset,mm: update tasks' mems_allowed in time"), policy's mems_allowed
is initialized as this.
1. policy->mems_allowd = nodes_and(task->mems_allowed, user's mask)
Here, if task is in top_cpuset, task->mems_allowed is not updated from
init's one. Assume user excutes command as #numactrl --interleave=all
,....
policy->mems_allowd = nodes_and(N_POSSIBLE, ALL_SET_MASK)
Then, policy's mems_allowd can includes a possible node, which has no pgdat.
MPOL's INTERLEAVE just scans nodemask of task->mems_allowd and access this
directly.
NODE_DATA(nid)->zonelist even if NODE_DATA(nid)==NULL
Then, what's we need is making policy->mems_allowed be aware of
N_HIGH_MEMORY. This patch does that. But to do so, extra nodemask will
be on statck. Because I know cpumask has a new interface of
CPUMASK_ALLOC(), I added it to node.
This patch stands on old behavior. But I feel this fix itself is just a
Band-Aid. But to do fundametal fix, we have to take care of memory
hotplug and it takes time. (task->mems_allowd should be N_HIGH_MEMORY, I
think.)
mpol_set_nodemask() should be aware of N_HIGH_MEMORY and policy's nodemask
should be includes only online nodes.
In old behavior, this is guaranteed by frequent reference to cpuset's
code. Now, most of them are removed and mempolicy has to check it by
itself.
To do check, a few nodemask_t will be used for calculating nodemask. But,
size of nodemask_t can be big and it's not good to allocate them on stack.
Now, cpumask_t has CPUMASK_ALLOC/FREE an easy code for get scratch area.
NODEMASK_ALLOC/FREE shoudl be there.
[akpm@linux-foundation.org: cleanups & tweaks]
Tested-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-08-06 15:07:33 -07:00
|
|
|
|
|
|
|
|
|
2020-06-08 21:33:25 -07:00
|
|
|
|
mm: make set_mempolicy(MPOL_INTERLEAV) N_HIGH_MEMORY aware
At first, init_task's mems_allowed is initialized as this.
init_task->mems_allowed == node_state[N_POSSIBLE]
And cpuset's top_cpuset mask is initialized as this
top_cpuset->mems_allowed = node_state[N_HIGH_MEMORY]
Before 2.6.29:
policy's mems_allowed is initialized as this.
1. update tasks->mems_allowed by its cpuset->mems_allowed.
2. policy->mems_allowed = nodes_and(tasks->mems_allowed, user's mask)
Updating task's mems_allowed in reference to top_cpuset's one.
cpuset's mems_allowed is aware of N_HIGH_MEMORY, always.
In 2.6.30: After commit 58568d2a8215cb6f55caf2332017d7bdff954e1c
("cpuset,mm: update tasks' mems_allowed in time"), policy's mems_allowed
is initialized as this.
1. policy->mems_allowd = nodes_and(task->mems_allowed, user's mask)
Here, if task is in top_cpuset, task->mems_allowed is not updated from
init's one. Assume user excutes command as #numactrl --interleave=all
,....
policy->mems_allowd = nodes_and(N_POSSIBLE, ALL_SET_MASK)
Then, policy's mems_allowd can includes a possible node, which has no pgdat.
MPOL's INTERLEAVE just scans nodemask of task->mems_allowd and access this
directly.
NODE_DATA(nid)->zonelist even if NODE_DATA(nid)==NULL
Then, what's we need is making policy->mems_allowed be aware of
N_HIGH_MEMORY. This patch does that. But to do so, extra nodemask will
be on statck. Because I know cpumask has a new interface of
CPUMASK_ALLOC(), I added it to node.
This patch stands on old behavior. But I feel this fix itself is just a
Band-Aid. But to do fundametal fix, we have to take care of memory
hotplug and it takes time. (task->mems_allowd should be N_HIGH_MEMORY, I
think.)
mpol_set_nodemask() should be aware of N_HIGH_MEMORY and policy's nodemask
should be includes only online nodes.
In old behavior, this is guaranteed by frequent reference to cpuset's
code. Now, most of them are removed and mempolicy has to check it by
itself.
To do check, a few nodemask_t will be used for calculating nodemask. But,
size of nodemask_t can be big and it's not good to allocate them on stack.
Now, cpumask_t has CPUMASK_ALLOC/FREE an easy code for get scratch area.
NODEMASK_ALLOC/FREE shoudl be there.
[akpm@linux-foundation.org: cleanups & tweaks]
Tested-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-08-06 15:07:33 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2009-10-26 16:49:59 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-07-28 06:13:21 +02:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
|
|
|
|
|
2023-07-28 06:13:21 +02:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
|
|
|
|
|
mm: mempolicy: make the behavior consistent when MPOL_MF_MOVE* and MPOL_MF_STRICT were specified
When both MPOL_MF_MOVE* and MPOL_MF_STRICT was specified, mbind() should
try best to migrate misplaced pages, if some of the pages could not be
migrated, then return -EIO.
There are three different sub-cases:
1. vma is not migratable
2. vma is migratable, but there are unmovable pages
3. vma is migratable, pages are movable, but migrate_pages() fails
If #1 happens, kernel would just abort immediately, then return -EIO,
after a7f40cfe3b7a ("mm: mempolicy: make mbind() return -EIO when
MPOL_MF_STRICT is specified").
If #3 happens, kernel would set policy and migrate pages with
best-effort, but won't rollback the migrated pages and reset the policy
back.
Before that commit, they behaves in the same way. It'd better to keep
their behavior consistent. But, rolling back the migrated pages and
resetting the policy back sounds not feasible, so just make #1 behave as
same as #3.
Userspace will know that not everything was successfully migrated (via
-EIO), and can take whatever steps it deems necessary - attempt
rollback, determine which exact page(s) are violating the policy, etc.
Make queue_pages_range() return 1 to indicate there are unmovable pages
or vma is not migratable.
The #2 is not handled correctly in the current kernel, the following
patch will fix it.
[yang.shi@linux.alibaba.com: fix review comments from Vlastimil]
Link: http://lkml.kernel.org/r/1563556862-54056-2-git-send-email-yang.shi@linux.alibaba.com
Link: http://lkml.kernel.org/r/1561162809-59140-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
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 15:37:15 -07:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
|
|
|
|
|
mempolicy: mmap_lock is not needed while migrating folios
mbind(2) holds down_write of current task's mmap_lock throughout
(exclusive because it needs to set the new mempolicy on the vmas);
migrate_pages(2) holds down_read of pid's mmap_lock throughout.
They both hold mmap_lock across the internal migrate_pages(), under which
all new page allocations (huge or small) are made. I'm nervous about it;
and migrate_pages() certainly does not need mmap_lock itself. It's done
this way for mbind(2), because its page allocator is vma_alloc_folio() or
alloc_hugetlb_folio_vma(), both of which depend on vma and address.
Now that we have alloc_pages_mpol(), depending on (refcounted) memory
policy and interleave index, mbind(2) can be modified to use that or
alloc_hugetlb_folio_nodemask(), and then not need mmap_lock across the
internal migrate_pages() at all: add alloc_migration_target_by_mpol() to
replace mbind's new_page().
(After that change, alloc_hugetlb_folio_vma() is used by nothing but a
userfaultfd function: move it out of hugetlb.h and into the #ifdef.)
migrate_pages(2) has chosen its target node before migrating, so can
continue to use the standard alloc_migration_target(); but let it take and
drop mmap_lock just around migrate_to_node()'s queue_pages_range():
neither the node-to-node calculations nor the page migrations need it.
It seems unlikely, but it is conceivable that some userspace depends on
the kernel's mmap_lock exclusion here, instead of doing its own locking:
more likely in a testsuite than in real life. It is also possible, of
course, that some pages on the list will be munmapped by another thread
before they are migrated, or a newer memory policy applied to the range by
that time: but such races could happen before, as soon as mmap_lock was
dropped, so it does not appear to be a concern.
Link: https://lkml.kernel.org/r/21e564e8-269f-6a89-7ee2-fd612831c289@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:27:47 -07:00
|
|
|
|
2019-11-15 17:34:33 -08:00
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-04-10 11:22:05 -04:00
|
|
|
|
2006-02-01 03:05:40 -08:00
|
|
|
|
mempolicy: mmap_lock is not needed while migrating folios
mbind(2) holds down_write of current task's mmap_lock throughout
(exclusive because it needs to set the new mempolicy on the vmas);
migrate_pages(2) holds down_read of pid's mmap_lock throughout.
They both hold mmap_lock across the internal migrate_pages(), under which
all new page allocations (huge or small) are made. I'm nervous about it;
and migrate_pages() certainly does not need mmap_lock itself. It's done
this way for mbind(2), because its page allocator is vma_alloc_folio() or
alloc_hugetlb_folio_vma(), both of which depend on vma and address.
Now that we have alloc_pages_mpol(), depending on (refcounted) memory
policy and interleave index, mbind(2) can be modified to use that or
alloc_hugetlb_folio_nodemask(), and then not need mmap_lock across the
internal migrate_pages() at all: add alloc_migration_target_by_mpol() to
replace mbind's new_page().
(After that change, alloc_hugetlb_folio_vma() is used by nothing but a
userfaultfd function: move it out of hugetlb.h and into the #ifdef.)
migrate_pages(2) has chosen its target node before migrating, so can
continue to use the standard alloc_migration_target(); but let it take and
drop mmap_lock just around migrate_to_node()'s queue_pages_range():
neither the node-to-node calculations nor the page migrations need it.
It seems unlikely, but it is conceivable that some userspace depends on
the kernel's mmap_lock exclusion here, instead of doing its own locking:
more likely in a testsuite than in real life. It is also possible, of
course, that some pages on the list will be munmapped by another thread
before they are migrated, or a newer memory policy applied to the range by
that time: but such races could happen before, as soon as mmap_lock was
dropped, so it does not appear to be a concern.
Link: https://lkml.kernel.org/r/21e564e8-269f-6a89-7ee2-fd612831c289@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:27:47 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-10-26 14:21:29 -07:00
|
|
|
|
2023-10-03 02:29:00 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
|
|
|
|
|
2024-02-29 15:29:45 +00:00
|
|
|
|
2023-10-03 02:29:00 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-02-29 15:29:45 +00:00
|
|
|
|
|
|
|
|
|
2023-10-03 02:29:00 -07:00
|
|
|
|
|
|
|
|
|
2024-02-29 15:29:45 +00:00
|
|
|
|
2023-10-03 02:29:00 -07:00
|
|
|
|
|
|
|
|
|
2024-10-05 21:01:14 +01:00
|
|
|
|
2024-02-29 15:29:45 +00:00
|
|
|
|
2023-10-03 02:29:00 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-02-29 15:29:45 +00:00
|
|
|
|
2023-10-03 02:29:00 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-02-29 15:29:45 +00:00
|
|
|
|
2023-10-03 02:29:00 -07:00
|
|
|
|
|
|
|
|
|
2019-11-15 17:34:33 -08:00
|
|
|
|
|
|
|
|
|
2020-06-08 21:33:25 -07:00
|
|
|
|
2023-10-03 02:29:00 -07:00
|
|
|
|
|
|
|
|
|
mempolicy: mmap_lock is not needed while migrating folios
mbind(2) holds down_write of current task's mmap_lock throughout
(exclusive because it needs to set the new mempolicy on the vmas);
migrate_pages(2) holds down_read of pid's mmap_lock throughout.
They both hold mmap_lock across the internal migrate_pages(), under which
all new page allocations (huge or small) are made. I'm nervous about it;
and migrate_pages() certainly does not need mmap_lock itself. It's done
this way for mbind(2), because its page allocator is vma_alloc_folio() or
alloc_hugetlb_folio_vma(), both of which depend on vma and address.
Now that we have alloc_pages_mpol(), depending on (refcounted) memory
policy and interleave index, mbind(2) can be modified to use that or
alloc_hugetlb_folio_nodemask(), and then not need mmap_lock across the
internal migrate_pages() at all: add alloc_migration_target_by_mpol() to
replace mbind's new_page().
(After that change, alloc_hugetlb_folio_vma() is used by nothing but a
userfaultfd function: move it out of hugetlb.h and into the #ifdef.)
migrate_pages(2) has chosen its target node before migrating, so can
continue to use the standard alloc_migration_target(); but let it take and
drop mmap_lock just around migrate_to_node()'s queue_pages_range():
neither the node-to-node calculations nor the page migrations need it.
It seems unlikely, but it is conceivable that some userspace depends on
the kernel's mmap_lock exclusion here, instead of doing its own locking:
more likely in a testsuite than in real life. It is also possible, of
course, that some pages on the list will be munmapped by another thread
before they are migrated, or a newer memory policy applied to the range by
that time: but such races could happen before, as soon as mmap_lock was
dropped, so it does not appear to be a concern.
Link: https://lkml.kernel.org/r/21e564e8-269f-6a89-7ee2-fd612831c289@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:27:47 -07:00
|
|
|
|
|
|
|
|
|
2023-10-03 02:29:00 -07:00
|
|
|
|
mempolicy: mmap_lock is not needed while migrating folios
mbind(2) holds down_write of current task's mmap_lock throughout
(exclusive because it needs to set the new mempolicy on the vmas);
migrate_pages(2) holds down_read of pid's mmap_lock throughout.
They both hold mmap_lock across the internal migrate_pages(), under which
all new page allocations (huge or small) are made. I'm nervous about it;
and migrate_pages() certainly does not need mmap_lock itself. It's done
this way for mbind(2), because its page allocator is vma_alloc_folio() or
alloc_hugetlb_folio_vma(), both of which depend on vma and address.
Now that we have alloc_pages_mpol(), depending on (refcounted) memory
policy and interleave index, mbind(2) can be modified to use that or
alloc_hugetlb_folio_nodemask(), and then not need mmap_lock across the
internal migrate_pages() at all: add alloc_migration_target_by_mpol() to
replace mbind's new_page().
(After that change, alloc_hugetlb_folio_vma() is used by nothing but a
userfaultfd function: move it out of hugetlb.h and into the #ifdef.)
migrate_pages(2) has chosen its target node before migrating, so can
continue to use the standard alloc_migration_target(); but let it take and
drop mmap_lock just around migrate_to_node()'s queue_pages_range():
neither the node-to-node calculations nor the page migrations need it.
It seems unlikely, but it is conceivable that some userspace depends on
the kernel's mmap_lock exclusion here, instead of doing its own locking:
more likely in a testsuite than in real life. It is also possible, of
course, that some pages on the list will be munmapped by another thread
before they are migrated, or a newer memory policy applied to the range by
that time: but such races could happen before, as soon as mmap_lock was
dropped, so it does not appear to be a concern.
Link: https://lkml.kernel.org/r/21e564e8-269f-6a89-7ee2-fd612831c289@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:27:47 -07:00
|
|
|
|
2019-11-15 17:34:33 -08:00
|
|
|
|
|
|
|
|
|
mempolicy: mmap_lock is not needed while migrating folios
mbind(2) holds down_write of current task's mmap_lock throughout
(exclusive because it needs to set the new mempolicy on the vmas);
migrate_pages(2) holds down_read of pid's mmap_lock throughout.
They both hold mmap_lock across the internal migrate_pages(), under which
all new page allocations (huge or small) are made. I'm nervous about it;
and migrate_pages() certainly does not need mmap_lock itself. It's done
this way for mbind(2), because its page allocator is vma_alloc_folio() or
alloc_hugetlb_folio_vma(), both of which depend on vma and address.
Now that we have alloc_pages_mpol(), depending on (refcounted) memory
policy and interleave index, mbind(2) can be modified to use that or
alloc_hugetlb_folio_nodemask(), and then not need mmap_lock across the
internal migrate_pages() at all: add alloc_migration_target_by_mpol() to
replace mbind's new_page().
(After that change, alloc_hugetlb_folio_vma() is used by nothing but a
userfaultfd function: move it out of hugetlb.h and into the #ifdef.)
migrate_pages(2) has chosen its target node before migrating, so can
continue to use the standard alloc_migration_target(); but let it take and
drop mmap_lock just around migrate_to_node()'s queue_pages_range():
neither the node-to-node calculations nor the page migrations need it.
It seems unlikely, but it is conceivable that some userspace depends on
the kernel's mmap_lock exclusion here, instead of doing its own locking:
more likely in a testsuite than in real life. It is also possible, of
course, that some pages on the list will be munmapped by another thread
before they are migrated, or a newer memory policy applied to the range by
that time: but such races could happen before, as soon as mmap_lock was
dropped, so it does not appear to be a concern.
Link: https://lkml.kernel.org/r/21e564e8-269f-6a89-7ee2-fd612831c289@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:27:47 -07:00
|
|
|
|
|
|
|
|
|
mempolicy: fix migrate_pages(2) syscall return nr_failed
"man 2 migrate_pages" says "On success migrate_pages() returns the number
of pages that could not be moved". Although 5.3 and 5.4 commits fixed
mbind(MPOL_MF_STRICT|MPOL_MF_MOVE*) to fail with EIO when not all pages
could be moved (because some could not be isolated for migration),
migrate_pages(2) was left still reporting only those pages failing at the
migration stage, forgetting those failing at the earlier isolation stage.
Fix that by accumulating a long nr_failed count in struct queue_pages,
returned by queue_pages_range() when it's not returning an error, for
adding on to the nr_failed count from migrate_pages() in mm/migrate.c. A
count of pages? It's more a count of folios, but changing it to pages
would entail more work (also in mm/migrate.c): does not seem justified.
queue_pages_range() itself should only return -EIO in the "strictly
unmovable" case (STRICT without any MOVEs): in that case it's best to
break out as soon as nr_failed gets set; but otherwise it should continue
to isolate pages for MOVing even when nr_failed - as the mbind(2) manpage
promises.
There's a case when nr_failed should be incremented when it was missed:
queue_folios_pte_range() and queue_folios_hugetlb() count the transient
migration entries, like queue_folios_pmd() already did. And there's a
case when nr_failed should not be incremented when it would have been: in
meeting later PTEs of the same large folio, which can only be isolated
once: fixed by recording the current large folio in struct queue_pages.
Clean up the affected functions, fixing or updating many comments. Bool
migrate_folio_add(), without -EIO: true if adding, or if skipping shared
(but its arguable folio_estimated_sharers() heuristic left unchanged).
Use MPOL_MF_WRLOCK flag to queue_pages_range(), instead of bool lock_vma.
Use explicit STRICT|MOVE* flags where queue_pages_test_walk() checks for
skipping, instead of hiding them behind MPOL_MF_VALID.
Link: https://lkml.kernel.org/r/9a6b0b9-3bb-dbef-8adf-efab4397b8d@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-03 02:17:43 -07:00
|
|
|
|
|
|
|
|
|
mm: mempolicy: make the behavior consistent when MPOL_MF_MOVE* and MPOL_MF_STRICT were specified
When both MPOL_MF_MOVE* and MPOL_MF_STRICT was specified, mbind() should
try best to migrate misplaced pages, if some of the pages could not be
migrated, then return -EIO.
There are three different sub-cases:
1. vma is not migratable
2. vma is migratable, but there are unmovable pages
3. vma is migratable, pages are movable, but migrate_pages() fails
If #1 happens, kernel would just abort immediately, then return -EIO,
after a7f40cfe3b7a ("mm: mempolicy: make mbind() return -EIO when
MPOL_MF_STRICT is specified").
If #3 happens, kernel would set policy and migrate pages with
best-effort, but won't rollback the migrated pages and reset the policy
back.
Before that commit, they behaves in the same way. It'd better to keep
their behavior consistent. But, rolling back the migrated pages and
resetting the policy back sounds not feasible, so just make #1 behave as
same as #3.
Userspace will know that not everything was successfully migrated (via
-EIO), and can take whatever steps it deems necessary - attempt
rollback, determine which exact page(s) are violating the policy, etc.
Make queue_pages_range() return 1 to indicate there are unmovable pages
or vma is not migratable.
The #2 is not handled correctly in the current kernel, the following
patch will fix it.
[yang.shi@linux.alibaba.com: fix review comments from Vlastimil]
Link: http://lkml.kernel.org/r/1563556862-54056-2-git-send-email-yang.shi@linux.alibaba.com
Link: http://lkml.kernel.org/r/1561162809-59140-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
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 15:37:15 -07:00
|
|
|
|
2008-04-28 02:13:08 -07:00
|
|
|
|
mm: disable LRU pagevec during the migration temporarily
LRU pagevec holds refcount of pages until the pagevec are drained. It
could prevent migration since the refcount of the page is greater than
the expection in migration logic. To mitigate the issue, callers of
migrate_pages drains LRU pagevec via migrate_prep or lru_add_drain_all
before migrate_pages call.
However, it's not enough because pages coming into pagevec after the
draining call still could stay at the pagevec so it could keep
preventing page migration. Since some callers of migrate_pages have
retrial logic with LRU draining, the page would migrate at next trail
but it is still fragile in that it doesn't close the fundamental race
between upcoming LRU pages into pagvec and migration so the migration
failure could cause contiguous memory allocation failure in the end.
To close the race, this patch disables lru caches(i.e, pagevec) during
ongoing migration until migrate is done.
Since it's really hard to reproduce, I measured how many times
migrate_pages retried with force mode(it is about a fallback to a sync
migration) with below debug code.
int migrate_pages(struct list_head *from, new_page_t get_new_page,
..
..
if (rc && reason == MR_CONTIG_RANGE && pass > 2) {
printk(KERN_ERR, "pfn 0x%lx reason %d", page_to_pfn(page), rc);
dump_page(page, "fail to migrate");
}
The test was repeating android apps launching with cma allocation in
background every five seconds. Total cma allocation count was about 500
during the testing. With this patch, the dump_page count was reduced
from 400 to 30.
The new interface is also useful for memory hotplug which currently
drains lru pcp caches after each migration failure. This is rather
suboptimal as it has to disrupt others running during the operation.
With the new interface the operation happens only once. This is also in
line with pcp allocator cache which are disabled for the offlining as
well.
Link: https://lkml.kernel.org/r/20210319175127.886124-1-minchan@kernel.org
Signed-off-by: Minchan Kim <minchan@kernel.org>
Reviewed-by: Chris Goldsworthy <cgoldswo@codeaurora.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: John Dias <joaodias@google.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: David Hildenbrand <david@redhat.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Oliver Sang <oliver.sang@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-05-04 18:36:54 -07:00
|
|
|
|
2021-05-04 18:36:57 -07:00
|
|
|
|
2006-01-08 01:01:04 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-10-29 18:16:59 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-09-08 15:18:21 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-10-29 18:16:59 -07:00
|
|
|
|
|
|
|
|
|
2006-01-08 01:00:51 -08:00
|
|
|
|
2005-10-29 18:16:59 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-02-20 18:27:59 -08:00
|
|
|
|
2006-02-17 13:59:36 -08:00
|
|
|
|
2005-10-29 18:16:59 -07:00
|
|
|
|
mm/mempolicy: fix the check of nodemask from user
As Xiaojun reported the ltp of migrate_pages01 will fail on arm64 system
which has 4 nodes[0...3], all have memory and CONFIG_NODES_SHIFT=2:
migrate_pages01 0 TINFO : test_invalid_nodes
migrate_pages01 14 TFAIL : migrate_pages_common.c:45: unexpected failure - returned value = 0, expected: -1
migrate_pages01 15 TFAIL : migrate_pages_common.c:55: call succeeded unexpectedly
In this case the test_invalid_nodes of migrate_pages01 will call:
SYSC_migrate_pages as:
migrate_pages(0, , {0x0000000000000001}, 64, , {0x0000000000000010}, 64) = 0
The new nodes specifies one or more node IDs that are greater than the
maximum supported node ID, however, the errno is not set to EINVAL as
expected.
As man pages of set_mempolicy[1], mbind[2], and migrate_pages[3]
mentioned, when nodemask specifies one or more node IDs that are greater
than the maximum supported node ID, the errno should set to EINVAL.
However, get_nodes only check whether the part of bits
[BITS_PER_LONG*BITS_TO_LONGS(MAX_NUMNODES), maxnode) is zero or not, and
remain [MAX_NUMNODES, BITS_PER_LONG*BITS_TO_LONGS(MAX_NUMNODES)
unchecked.
This patch is to check the bits of [MAX_NUMNODES, maxnode) in get_nodes
to let migrate_pages set the errno to EINVAL when nodemask specifies one
or more node IDs that are greater than the maximum supported node ID,
which follows the manpage's guide.
[1] http://man7.org/linux/man-pages/man2/set_mempolicy.2.html
[2] http://man7.org/linux/man-pages/man2/mbind.2.html
[3] http://man7.org/linux/man-pages/man2/migrate_pages.2.html
Link: http://lkml.kernel.org/r/1510882624-44342-3-git-send-email-xieyisheng1@huawei.com
Signed-off-by: Yisheng Xie <xieyisheng1@huawei.com>
Reported-by: Tan Xiaojun <tanxiaojun@huawei.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Chris Salls <salls@cs.ucsb.edu>
Cc: Christopher Lameter <cl@linux.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Ingo Molnar <mingo@kernel.org>
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>
2018-01-31 16:16:11 -08:00
|
|
|
|
|
|
|
|
|
2021-09-08 15:18:21 -07:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: fix the check of nodemask from user
As Xiaojun reported the ltp of migrate_pages01 will fail on arm64 system
which has 4 nodes[0...3], all have memory and CONFIG_NODES_SHIFT=2:
migrate_pages01 0 TINFO : test_invalid_nodes
migrate_pages01 14 TFAIL : migrate_pages_common.c:45: unexpected failure - returned value = 0, expected: -1
migrate_pages01 15 TFAIL : migrate_pages_common.c:55: call succeeded unexpectedly
In this case the test_invalid_nodes of migrate_pages01 will call:
SYSC_migrate_pages as:
migrate_pages(0, , {0x0000000000000001}, 64, , {0x0000000000000010}, 64) = 0
The new nodes specifies one or more node IDs that are greater than the
maximum supported node ID, however, the errno is not set to EINVAL as
expected.
As man pages of set_mempolicy[1], mbind[2], and migrate_pages[3]
mentioned, when nodemask specifies one or more node IDs that are greater
than the maximum supported node ID, the errno should set to EINVAL.
However, get_nodes only check whether the part of bits
[BITS_PER_LONG*BITS_TO_LONGS(MAX_NUMNODES), maxnode) is zero or not, and
remain [MAX_NUMNODES, BITS_PER_LONG*BITS_TO_LONGS(MAX_NUMNODES)
unchecked.
This patch is to check the bits of [MAX_NUMNODES, maxnode) in get_nodes
to let migrate_pages set the errno to EINVAL when nodemask specifies one
or more node IDs that are greater than the maximum supported node ID,
which follows the manpage's guide.
[1] http://man7.org/linux/man-pages/man2/set_mempolicy.2.html
[2] http://man7.org/linux/man-pages/man2/mbind.2.html
[3] http://man7.org/linux/man-pages/man2/migrate_pages.2.html
Link: http://lkml.kernel.org/r/1510882624-44342-3-git-send-email-xieyisheng1@huawei.com
Signed-off-by: Yisheng Xie <xieyisheng1@huawei.com>
Reported-by: Tan Xiaojun <tanxiaojun@huawei.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Chris Salls <salls@cs.ucsb.edu>
Cc: Christopher Lameter <cl@linux.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Ingo Molnar <mingo@kernel.org>
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>
2018-01-31 16:16:11 -08:00
|
|
|
|
2021-09-08 15:18:21 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-10-29 18:16:59 -07:00
|
|
|
|
2022-06-01 17:32:11 +08:00
|
|
|
|
mm/mempolicy: fix the check of nodemask from user
As Xiaojun reported the ltp of migrate_pages01 will fail on arm64 system
which has 4 nodes[0...3], all have memory and CONFIG_NODES_SHIFT=2:
migrate_pages01 0 TINFO : test_invalid_nodes
migrate_pages01 14 TFAIL : migrate_pages_common.c:45: unexpected failure - returned value = 0, expected: -1
migrate_pages01 15 TFAIL : migrate_pages_common.c:55: call succeeded unexpectedly
In this case the test_invalid_nodes of migrate_pages01 will call:
SYSC_migrate_pages as:
migrate_pages(0, , {0x0000000000000001}, 64, , {0x0000000000000010}, 64) = 0
The new nodes specifies one or more node IDs that are greater than the
maximum supported node ID, however, the errno is not set to EINVAL as
expected.
As man pages of set_mempolicy[1], mbind[2], and migrate_pages[3]
mentioned, when nodemask specifies one or more node IDs that are greater
than the maximum supported node ID, the errno should set to EINVAL.
However, get_nodes only check whether the part of bits
[BITS_PER_LONG*BITS_TO_LONGS(MAX_NUMNODES), maxnode) is zero or not, and
remain [MAX_NUMNODES, BITS_PER_LONG*BITS_TO_LONGS(MAX_NUMNODES)
unchecked.
This patch is to check the bits of [MAX_NUMNODES, maxnode) in get_nodes
to let migrate_pages set the errno to EINVAL when nodemask specifies one
or more node IDs that are greater than the maximum supported node ID,
which follows the manpage's guide.
[1] http://man7.org/linux/man-pages/man2/set_mempolicy.2.html
[2] http://man7.org/linux/man-pages/man2/mbind.2.html
[3] http://man7.org/linux/man-pages/man2/migrate_pages.2.html
Link: http://lkml.kernel.org/r/1510882624-44342-3-git-send-email-xieyisheng1@huawei.com
Signed-off-by: Yisheng Xie <xieyisheng1@huawei.com>
Reported-by: Tan Xiaojun <tanxiaojun@huawei.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Chris Salls <salls@cs.ucsb.edu>
Cc: Christopher Lameter <cl@linux.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Ingo Molnar <mingo@kernel.org>
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>
2018-01-31 16:16:11 -08:00
|
|
|
|
2021-09-08 15:18:21 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: fix the check of nodemask from user
As Xiaojun reported the ltp of migrate_pages01 will fail on arm64 system
which has 4 nodes[0...3], all have memory and CONFIG_NODES_SHIFT=2:
migrate_pages01 0 TINFO : test_invalid_nodes
migrate_pages01 14 TFAIL : migrate_pages_common.c:45: unexpected failure - returned value = 0, expected: -1
migrate_pages01 15 TFAIL : migrate_pages_common.c:55: call succeeded unexpectedly
In this case the test_invalid_nodes of migrate_pages01 will call:
SYSC_migrate_pages as:
migrate_pages(0, , {0x0000000000000001}, 64, , {0x0000000000000010}, 64) = 0
The new nodes specifies one or more node IDs that are greater than the
maximum supported node ID, however, the errno is not set to EINVAL as
expected.
As man pages of set_mempolicy[1], mbind[2], and migrate_pages[3]
mentioned, when nodemask specifies one or more node IDs that are greater
than the maximum supported node ID, the errno should set to EINVAL.
However, get_nodes only check whether the part of bits
[BITS_PER_LONG*BITS_TO_LONGS(MAX_NUMNODES), maxnode) is zero or not, and
remain [MAX_NUMNODES, BITS_PER_LONG*BITS_TO_LONGS(MAX_NUMNODES)
unchecked.
This patch is to check the bits of [MAX_NUMNODES, maxnode) in get_nodes
to let migrate_pages set the errno to EINVAL when nodemask specifies one
or more node IDs that are greater than the maximum supported node ID,
which follows the manpage's guide.
[1] http://man7.org/linux/man-pages/man2/set_mempolicy.2.html
[2] http://man7.org/linux/man-pages/man2/mbind.2.html
[3] http://man7.org/linux/man-pages/man2/migrate_pages.2.html
Link: http://lkml.kernel.org/r/1510882624-44342-3-git-send-email-xieyisheng1@huawei.com
Signed-off-by: Yisheng Xie <xieyisheng1@huawei.com>
Reported-by: Tan Xiaojun <tanxiaojun@huawei.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Chris Salls <salls@cs.ucsb.edu>
Cc: Christopher Lameter <cl@linux.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Ingo Molnar <mingo@kernel.org>
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>
2018-01-31 16:16:11 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-09-08 15:18:21 -07:00
|
|
|
|
2005-10-29 18:16:59 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-02-20 22:18:58 -08:00
|
|
|
|
2021-09-08 15:18:21 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-10-29 18:16:59 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-09-08 15:18:21 -07:00
|
|
|
|
2005-10-29 18:16:59 -07:00
|
|
|
|
2021-09-08 15:18:21 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-10-29 18:16:59 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-30 18:51:03 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: add MPOL_PREFERRED_MANY for multiple preferred nodes
Patch series "Introduce multi-preference mempolicy", v7.
This patch series introduces the concept of the MPOL_PREFERRED_MANY
mempolicy. This mempolicy mode can be used with either the
set_mempolicy(2) or mbind(2) interfaces. Like the MPOL_PREFERRED
interface, it allows an application to set a preference for nodes which
will fulfil memory allocation requests. Unlike the MPOL_PREFERRED mode,
it takes a set of nodes. Like the MPOL_BIND interface, it works over a
set of nodes. Unlike MPOL_BIND, it will not cause a SIGSEGV or invoke the
OOM killer if those preferred nodes are not available.
Along with these patches are patches for libnuma, numactl, numademo, and
memhog. They still need some polish, but can be found here:
https://gitlab.com/bwidawsk/numactl/-/tree/prefer-many It allows new
usage: `numactl -P 0,3,4`
The goal of the new mode is to enable some use-cases when using tiered memory
usage models which I've lovingly named.
1a. The Hare - The interconnect is fast enough to meet bandwidth and
latency requirements allowing preference to be given to all nodes with
"fast" memory.
1b. The Indiscriminate Hare - An application knows it wants fast
memory (or perhaps slow memory), but doesn't care which node it runs
on. The application can prefer a set of nodes and then xpu bind to
the local node (cpu, accelerator, etc). This reverses the nodes are
chosen today where the kernel attempts to use local memory to the CPU
whenever possible. This will attempt to use the local accelerator to
the memory.
2. The Tortoise - The administrator (or the application itself) is
aware it only needs slow memory, and so can prefer that.
Much of this is almost achievable with the bind interface, but the bind
interface suffers from an inability to fallback to another set of nodes if
binding fails to all nodes in the nodemask.
Like MPOL_BIND a nodemask is given. Inherently this removes ordering from the
preference.
> /* Set first two nodes as preferred in an 8 node system. */
> const unsigned long nodes = 0x3
> set_mempolicy(MPOL_PREFER_MANY, &nodes, 8);
> /* Mimic interleave policy, but have fallback *.
> const unsigned long nodes = 0xaa
> set_mempolicy(MPOL_PREFER_MANY, &nodes, 8);
Some internal discussion took place around the interface. There are two
alternatives which we have discussed, plus one I stuck in:
1. Ordered list of nodes. Currently it's believed that the added
complexity is nod needed for expected usecases.
2. A flag for bind to allow falling back to other nodes. This
confuses the notion of binding and is less flexible than the current
solution.
3. Create flags or new modes that helps with some ordering. This
offers both a friendlier API as well as a solution for more customized
usage. It's unknown if it's worth the complexity to support this.
Here is sample code for how this might work:
> // Prefer specific nodes for some something wacky
> set_mempolicy(MPOL_PREFER_MANY, 0x17c, 1024);
>
> // Default
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_SOCKET, NULL, 0);
> // which is the same as
> set_mempolicy(MPOL_DEFAULT, NULL, 0);
>
> // The Hare
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE, NULL, 0);
>
> // The Tortoise
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE_REV, NULL, 0);
>
> // Prefer the fast memory of the first two sockets
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE, -1, 2);
>
This patch (of 5):
The NUMA APIs currently allow passing in a "preferred node" as a single
bit set in a nodemask. If more than one bit it set, bits after the first
are ignored.
This single node is generally OK for location-based NUMA where memory
being allocated will eventually be operated on by a single CPU. However,
in systems with multiple memory types, folks want to target a *type* of
memory instead of a location. For instance, someone might want some
high-bandwidth memory but do not care about the CPU next to which it is
allocated. Or, they want a cheap, high capacity allocation and want to
target all NUMA nodes which have persistent memory in volatile mode. In
both of these cases, the application wants to target a *set* of nodes, but
does not want strict MPOL_BIND behavior as that could lead to OOM killer
or SIGSEGV.
So add MPOL_PREFERRED_MANY policy to support the multiple preferred nodes
requirement. This is not a pie-in-the-sky dream for an API. This was a
response to a specific ask of more than one group at Intel. Specifically:
1. There are existing libraries that target memory types such as
https://github.com/memkind/memkind. These are known to suffer from
SIGSEGV's when memory is low on targeted memory "kinds" that span more
than one node. The MCDRAM on a Xeon Phi in "Cluster on Die" mode is an
example of this.
2. Volatile-use persistent memory users want to have a memory policy
which is targeted at either "cheap and slow" (PMEM) or "expensive and
fast" (DRAM). However, they do not want to experience allocation
failures when the targeted type is unavailable.
3. Allocate-then-run. Generally, we let the process scheduler decide
on which physical CPU to run a task. That location provides a default
allocation policy, and memory availability is not generally considered
when placing tasks. For situations where memory is valuable and
constrained, some users want to allocate memory first, *then* allocate
close compute resources to the allocation. This is the reverse of the
normal (CPU) model. Accelerators such as GPUs that operate on
core-mm-managed memory are interested in this model.
A check is added in sanitize_mpol_flags() to not permit 'prefer_many'
policy to be used for now, and will be removed in later patch after all
implementations for 'prefer_many' are ready, as suggested by Michal Hocko.
[mhocko@kernel.org: suggest to refine policy_node/policy_nodemask handling]
Link: https://lkml.kernel.org/r/1627970362-61305-1-git-send-email-feng.tang@intel.com
Link: https://lore.kernel.org/r/20200630212517.308045-4-ben.widawsky@intel.com
Link: https://lkml.kernel.org/r/1627970362-61305-2-git-send-email-feng.tang@intel.com
Co-developed-by: Ben Widawsky <ben.widawsky@intel.com>
Signed-off-by: Ben Widawsky <ben.widawsky@intel.com>
Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com>
Signed-off-by: Feng Tang <feng.tang@intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Huang Ying <ying.huang@intel.com>b
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-09-02 15:00:06 -07:00
|
|
|
|
2021-09-02 15:00:16 -07:00
|
|
|
|
2021-06-30 18:51:03 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-10-18 15:15:49 -07:00
|
|
|
|
mm/numa_balancing: allow migrate on protnone reference with MPOL_PREFERRED_MANY policy
commit bda420b98505 ("numa balancing: migrate on fault among multiple
bound nodes") added support for migrate on protnone reference with
MPOL_BIND memory policy. This allowed numa fault migration when the
executing node is part of the policy mask for MPOL_BIND. This patch
extends migration support to MPOL_PREFERRED_MANY policy.
Currently, we cannot specify MPOL_PREFERRED_MANY with the mempolicy flag
MPOL_F_NUMA_BALANCING. This causes issues when we want to use
NUMA_BALANCING_MEMORY_TIERING. To effectively use the slow memory tier,
the kernel should not allocate pages from the slower memory tier via
allocation control zonelist fallback. Instead, we should move cold pages
from the faster memory node via memory demotion. For a page allocation,
kswapd is only woken up after we try to allocate pages from all nodes in
the allocation zone list. This implies that, without using memory
policies, we will end up allocating hot pages in the slower memory tier.
MPOL_PREFERRED_MANY was added by commit b27abaccf8e8 ("mm/mempolicy: add
MPOL_PREFERRED_MANY for multiple preferred nodes") to allow better
allocation control when we have memory tiers in the system. With
MPOL_PREFERRED_MANY, the user can use a policy node mask consisting only
of faster memory nodes. When we fail to allocate pages from the faster
memory node, kswapd would be woken up, allowing demotion of cold pages to
slower memory nodes.
With the current kernel, such usage of memory policies implies we can't do
page promotion from a slower memory tier to a faster memory tier using
numa fault. This patch fixes this issue.
For MPOL_PREFERRED_MANY, if the executing node is in the policy node mask,
we allow numa migration to the executing nodes. If the executing node is
not in the policy node mask, we do not allow numa migration.
Example:
On a 2-sockets system, NUMA node N0, N1 and N2 are in socket 0,
N3 in socket 1. N0, N1 and N3 have fast memory and CPU, while
N2 has slow memory and no CPU. For a workload, we may use
MPOL_PREFERRED_MANY with nodemask N0 and N1 set because the workload
runs on CPUs of socket 0 at most times. Then, even if the workload
runs on CPUs of N3 occasionally, we will not try to migrate the workload
pages from N2 to N3 because users may want to avoid cross-socket access
as much as possible in the long term.
In below table, Process is the Process executing node and
Curr Loc Pgs is the numa node where page present(folio node)
===========================================================
Process Policy Curr Loc Pgs Observation
-----------------------------------------------------------
N0 N0 N1 N1 Pages Migrated from N1 to N0
N0 N0 N1 N2 Pages Migrated from N2 to N0
N0 N0 N1 N3 Pages Migrated from N3 to N0
N3 N0 N1 N0 Pages NOT Migrated to N3
N3 N0 N1 N1 Pages NOT Migrated to N3
N3 N0 N1 N2 Pages NOT Migrated to N3
------------------------------------------------------------
Link: https://lkml.kernel.org/r/158acc57319129aa46d50fd64c9330f3e7c7b4bf.1711373653.git.donettom@linux.ibm.com
Link: https://lkml.kernel.org/r/369d6a58758396335fd1176d97bbca4e7730d75a.1709909210.git.donettom@linux.ibm.com
Signed-off-by: Aneesh Kumar K.V (IBM) <aneesh.kumar@kernel.org>
Signed-off-by: Donet Tom <donettom@linux.ibm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Feng Tang <feng.tang@intel.com>
Cc: Huang, Ying <ying.huang@intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rik van Riel <riel@surriel.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-08 09:15:38 -06:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-10-18 15:15:49 -07:00
|
|
|
|
|
|
|
|
|
2021-06-30 18:51:03 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-03-17 16:12:22 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-10-29 18:16:59 -07:00
|
|
|
|
2008-04-28 02:12:25 -07:00
|
|
|
|
2021-06-30 18:51:03 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-10-29 18:16:59 -07:00
|
|
|
|
mm: untag user pointers passed to memory syscalls
This patch is a part of a series that extends kernel ABI to allow to pass
tagged user pointers (with the top byte set to something else other than
0x00) as syscall arguments.
This patch allows tagged pointers to be passed to the following memory
syscalls: get_mempolicy, madvise, mbind, mincore, mlock, mlock2, mprotect,
mremap, msync, munlock, move_pages.
The mmap and mremap syscalls do not currently accept tagged addresses.
Architectures may interpret the tag as a background colour for the
corresponding vma.
Link: http://lkml.kernel.org/r/aaf0c0969d46b2feb9017f3e1b3ef3970b633d91.1563904656.git.andreyknvl@google.com
Signed-off-by: Andrey Konovalov <andreyknvl@google.com>
Reviewed-by: Khalid Aziz <khalid.aziz@oracle.com>
Reviewed-by: Vincenzo Frascino <vincenzo.frascino@arm.com>
Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Reviewed-by: Kees Cook <keescook@chromium.org>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Eric Auger <eric.auger@redhat.com>
Cc: Felix Kuehling <Felix.Kuehling@amd.com>
Cc: Jens Wiklander <jens.wiklander@linaro.org>
Cc: Mauro Carvalho Chehab <mchehab+samsung@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-09-25 16:48:30 -07:00
|
|
|
|
2021-06-30 18:51:03 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-10-29 18:16:59 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-30 18:51:03 -07:00
|
|
|
|
|
|
|
|
|
2005-10-29 18:16:59 -07:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: add set_mempolicy_home_node syscall
This syscall can be used to set a home node for the MPOL_BIND and
MPOL_PREFERRED_MANY memory policy. Users should use this syscall after
setting up a memory policy for the specified range as shown below.
mbind(p, nr_pages * page_size, MPOL_BIND, new_nodes->maskp,
new_nodes->size + 1, 0);
sys_set_mempolicy_home_node((unsigned long)p, nr_pages * page_size,
home_node, 0);
The syscall allows specifying a home node/preferred node from which
kernel will fulfill memory allocation requests first.
For address range with MPOL_BIND memory policy, if nodemask specifies
more than one node, page allocations will come from the node in the
nodemask with sufficient free memory that is closest to the home
node/preferred node.
For MPOL_PREFERRED_MANY if the nodemask specifies more than one node,
page allocation will come from the node in the nodemask with sufficient
free memory that is closest to the home node/preferred node. If there
is not enough memory in all the nodes specified in the nodemask, the
allocation will be attempted from the closest numa node to the home node
in the system.
This helps applications to hint at a memory allocation preference node
and fallback to _only_ a set of nodes if the memory is not available on
the preferred node. Fallback allocation is attempted from the node
which is nearest to the preferred node.
This helps applications to have control on memory allocation numa nodes
and avoids default fallback to slow memory NUMA nodes. For example a
system with NUMA nodes 1,2 and 3 with DRAM memory and 10, 11 and 12 of
slow memory
new_nodes = numa_bitmask_alloc(nr_nodes);
numa_bitmask_setbit(new_nodes, 1);
numa_bitmask_setbit(new_nodes, 2);
numa_bitmask_setbit(new_nodes, 3);
p = mmap(NULL, nr_pages * page_size, protflag, mapflag, -1, 0);
mbind(p, nr_pages * page_size, MPOL_BIND, new_nodes->maskp, new_nodes->size + 1, 0);
sys_set_mempolicy_home_node(p, nr_pages * page_size, 2, 0);
This will allocate from nodes closer to node 2 and will make sure the
kernel will only allocate from nodes 1, 2, and 3. Memory will not be
allocated from slow memory nodes 10, 11, and 12. This differs from
default MPOL_BIND behavior in that with default MPOL_BIND the allocation
will be attempted from node closer to the local node. One of the
reasons to specify a home node is to allow allocations from cpu less
NUMA node and its nearby NUMA nodes.
With MPOL_PREFERRED_MANY on the other hand will first try to allocate
from the closest node to node 2 from the node list 1, 2 and 3. If those
nodes don't have enough memory, kernel will allocate from slow memory
node 10, 11 and 12 which ever is closer to node 2.
Link: https://lkml.kernel.org/r/20211202123810.267175-3-aneesh.kumar@linux.ibm.com
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Ben Widawsky <ben.widawsky@intel.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Feng Tang <feng.tang@intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: <linux-api@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-14 14:08:17 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-04-10 11:22:05 -04:00
|
|
|
|
2022-12-16 14:45:37 -05:00
|
|
|
|
mm/mempolicy: add set_mempolicy_home_node syscall
This syscall can be used to set a home node for the MPOL_BIND and
MPOL_PREFERRED_MANY memory policy. Users should use this syscall after
setting up a memory policy for the specified range as shown below.
mbind(p, nr_pages * page_size, MPOL_BIND, new_nodes->maskp,
new_nodes->size + 1, 0);
sys_set_mempolicy_home_node((unsigned long)p, nr_pages * page_size,
home_node, 0);
The syscall allows specifying a home node/preferred node from which
kernel will fulfill memory allocation requests first.
For address range with MPOL_BIND memory policy, if nodemask specifies
more than one node, page allocations will come from the node in the
nodemask with sufficient free memory that is closest to the home
node/preferred node.
For MPOL_PREFERRED_MANY if the nodemask specifies more than one node,
page allocation will come from the node in the nodemask with sufficient
free memory that is closest to the home node/preferred node. If there
is not enough memory in all the nodes specified in the nodemask, the
allocation will be attempted from the closest numa node to the home node
in the system.
This helps applications to hint at a memory allocation preference node
and fallback to _only_ a set of nodes if the memory is not available on
the preferred node. Fallback allocation is attempted from the node
which is nearest to the preferred node.
This helps applications to have control on memory allocation numa nodes
and avoids default fallback to slow memory NUMA nodes. For example a
system with NUMA nodes 1,2 and 3 with DRAM memory and 10, 11 and 12 of
slow memory
new_nodes = numa_bitmask_alloc(nr_nodes);
numa_bitmask_setbit(new_nodes, 1);
numa_bitmask_setbit(new_nodes, 2);
numa_bitmask_setbit(new_nodes, 3);
p = mmap(NULL, nr_pages * page_size, protflag, mapflag, -1, 0);
mbind(p, nr_pages * page_size, MPOL_BIND, new_nodes->maskp, new_nodes->size + 1, 0);
sys_set_mempolicy_home_node(p, nr_pages * page_size, 2, 0);
This will allocate from nodes closer to node 2 and will make sure the
kernel will only allocate from nodes 1, 2, and 3. Memory will not be
allocated from slow memory nodes 10, 11, and 12. This differs from
default MPOL_BIND behavior in that with default MPOL_BIND the allocation
will be attempted from node closer to the local node. One of the
reasons to specify a home node is to allow allocations from cpu less
NUMA node and its nearby NUMA nodes.
With MPOL_PREFERRED_MANY on the other hand will first try to allocate
from the closest node to node 2 from the node list 1, 2 and 3. If those
nodes don't have enough memory, kernel will allocate from slow memory
node 10, 11 and 12 which ever is closer to node 2.
Link: https://lkml.kernel.org/r/20211202123810.267175-3-aneesh.kumar@linux.ibm.com
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Ben Widawsky <ben.widawsky@intel.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Feng Tang <feng.tang@intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: <linux-api@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-14 14:08:17 -08:00
|
|
|
|
|
|
|
|
|
2022-09-06 19:49:02 +00:00
|
|
|
|
mm/mempolicy: add set_mempolicy_home_node syscall
This syscall can be used to set a home node for the MPOL_BIND and
MPOL_PREFERRED_MANY memory policy. Users should use this syscall after
setting up a memory policy for the specified range as shown below.
mbind(p, nr_pages * page_size, MPOL_BIND, new_nodes->maskp,
new_nodes->size + 1, 0);
sys_set_mempolicy_home_node((unsigned long)p, nr_pages * page_size,
home_node, 0);
The syscall allows specifying a home node/preferred node from which
kernel will fulfill memory allocation requests first.
For address range with MPOL_BIND memory policy, if nodemask specifies
more than one node, page allocations will come from the node in the
nodemask with sufficient free memory that is closest to the home
node/preferred node.
For MPOL_PREFERRED_MANY if the nodemask specifies more than one node,
page allocation will come from the node in the nodemask with sufficient
free memory that is closest to the home node/preferred node. If there
is not enough memory in all the nodes specified in the nodemask, the
allocation will be attempted from the closest numa node to the home node
in the system.
This helps applications to hint at a memory allocation preference node
and fallback to _only_ a set of nodes if the memory is not available on
the preferred node. Fallback allocation is attempted from the node
which is nearest to the preferred node.
This helps applications to have control on memory allocation numa nodes
and avoids default fallback to slow memory NUMA nodes. For example a
system with NUMA nodes 1,2 and 3 with DRAM memory and 10, 11 and 12 of
slow memory
new_nodes = numa_bitmask_alloc(nr_nodes);
numa_bitmask_setbit(new_nodes, 1);
numa_bitmask_setbit(new_nodes, 2);
numa_bitmask_setbit(new_nodes, 3);
p = mmap(NULL, nr_pages * page_size, protflag, mapflag, -1, 0);
mbind(p, nr_pages * page_size, MPOL_BIND, new_nodes->maskp, new_nodes->size + 1, 0);
sys_set_mempolicy_home_node(p, nr_pages * page_size, 2, 0);
This will allocate from nodes closer to node 2 and will make sure the
kernel will only allocate from nodes 1, 2, and 3. Memory will not be
allocated from slow memory nodes 10, 11, and 12. This differs from
default MPOL_BIND behavior in that with default MPOL_BIND the allocation
will be attempted from node closer to the local node. One of the
reasons to specify a home node is to allow allocations from cpu less
NUMA node and its nearby NUMA nodes.
With MPOL_PREFERRED_MANY on the other hand will first try to allocate
from the closest node to node 2 from the node list 1, 2 and 3. If those
nodes don't have enough memory, kernel will allocate from slow memory
node 10, 11 and 12 which ever is closer to node 2.
Link: https://lkml.kernel.org/r/20211202123810.267175-3-aneesh.kumar@linux.ibm.com
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Ben Widawsky <ben.widawsky@intel.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Feng Tang <feng.tang@intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: <linux-api@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-14 14:08:17 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-09-13 01:55:05 +00:00
|
|
|
|
mm/mempolicy: add set_mempolicy_home_node syscall
This syscall can be used to set a home node for the MPOL_BIND and
MPOL_PREFERRED_MANY memory policy. Users should use this syscall after
setting up a memory policy for the specified range as shown below.
mbind(p, nr_pages * page_size, MPOL_BIND, new_nodes->maskp,
new_nodes->size + 1, 0);
sys_set_mempolicy_home_node((unsigned long)p, nr_pages * page_size,
home_node, 0);
The syscall allows specifying a home node/preferred node from which
kernel will fulfill memory allocation requests first.
For address range with MPOL_BIND memory policy, if nodemask specifies
more than one node, page allocations will come from the node in the
nodemask with sufficient free memory that is closest to the home
node/preferred node.
For MPOL_PREFERRED_MANY if the nodemask specifies more than one node,
page allocation will come from the node in the nodemask with sufficient
free memory that is closest to the home node/preferred node. If there
is not enough memory in all the nodes specified in the nodemask, the
allocation will be attempted from the closest numa node to the home node
in the system.
This helps applications to hint at a memory allocation preference node
and fallback to _only_ a set of nodes if the memory is not available on
the preferred node. Fallback allocation is attempted from the node
which is nearest to the preferred node.
This helps applications to have control on memory allocation numa nodes
and avoids default fallback to slow memory NUMA nodes. For example a
system with NUMA nodes 1,2 and 3 with DRAM memory and 10, 11 and 12 of
slow memory
new_nodes = numa_bitmask_alloc(nr_nodes);
numa_bitmask_setbit(new_nodes, 1);
numa_bitmask_setbit(new_nodes, 2);
numa_bitmask_setbit(new_nodes, 3);
p = mmap(NULL, nr_pages * page_size, protflag, mapflag, -1, 0);
mbind(p, nr_pages * page_size, MPOL_BIND, new_nodes->maskp, new_nodes->size + 1, 0);
sys_set_mempolicy_home_node(p, nr_pages * page_size, 2, 0);
This will allocate from nodes closer to node 2 and will make sure the
kernel will only allocate from nodes 1, 2, and 3. Memory will not be
allocated from slow memory nodes 10, 11, and 12. This differs from
default MPOL_BIND behavior in that with default MPOL_BIND the allocation
will be attempted from node closer to the local node. One of the
reasons to specify a home node is to allow allocations from cpu less
NUMA node and its nearby NUMA nodes.
With MPOL_PREFERRED_MANY on the other hand will first try to allocate
from the closest node to node 2 from the node list 1, 2 and 3. If those
nodes don't have enough memory, kernel will allocate from slow memory
node 10, 11 and 12 which ever is closer to node 2.
Link: https://lkml.kernel.org/r/20211202123810.267175-3-aneesh.kumar@linux.ibm.com
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Ben Widawsky <ben.widawsky@intel.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Feng Tang <feng.tang@intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: <linux-api@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-14 14:08:17 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-04-10 11:22:05 -04:00
|
|
|
|
2022-09-06 19:49:02 +00:00
|
|
|
|
mm/mempolicy: add set_mempolicy_home_node syscall
This syscall can be used to set a home node for the MPOL_BIND and
MPOL_PREFERRED_MANY memory policy. Users should use this syscall after
setting up a memory policy for the specified range as shown below.
mbind(p, nr_pages * page_size, MPOL_BIND, new_nodes->maskp,
new_nodes->size + 1, 0);
sys_set_mempolicy_home_node((unsigned long)p, nr_pages * page_size,
home_node, 0);
The syscall allows specifying a home node/preferred node from which
kernel will fulfill memory allocation requests first.
For address range with MPOL_BIND memory policy, if nodemask specifies
more than one node, page allocations will come from the node in the
nodemask with sufficient free memory that is closest to the home
node/preferred node.
For MPOL_PREFERRED_MANY if the nodemask specifies more than one node,
page allocation will come from the node in the nodemask with sufficient
free memory that is closest to the home node/preferred node. If there
is not enough memory in all the nodes specified in the nodemask, the
allocation will be attempted from the closest numa node to the home node
in the system.
This helps applications to hint at a memory allocation preference node
and fallback to _only_ a set of nodes if the memory is not available on
the preferred node. Fallback allocation is attempted from the node
which is nearest to the preferred node.
This helps applications to have control on memory allocation numa nodes
and avoids default fallback to slow memory NUMA nodes. For example a
system with NUMA nodes 1,2 and 3 with DRAM memory and 10, 11 and 12 of
slow memory
new_nodes = numa_bitmask_alloc(nr_nodes);
numa_bitmask_setbit(new_nodes, 1);
numa_bitmask_setbit(new_nodes, 2);
numa_bitmask_setbit(new_nodes, 3);
p = mmap(NULL, nr_pages * page_size, protflag, mapflag, -1, 0);
mbind(p, nr_pages * page_size, MPOL_BIND, new_nodes->maskp, new_nodes->size + 1, 0);
sys_set_mempolicy_home_node(p, nr_pages * page_size, 2, 0);
This will allocate from nodes closer to node 2 and will make sure the
kernel will only allocate from nodes 1, 2, and 3. Memory will not be
allocated from slow memory nodes 10, 11, and 12. This differs from
default MPOL_BIND behavior in that with default MPOL_BIND the allocation
will be attempted from node closer to the local node. One of the
reasons to specify a home node is to allow allocations from cpu less
NUMA node and its nearby NUMA nodes.
With MPOL_PREFERRED_MANY on the other hand will first try to allocate
from the closest node to node 2 from the node list 1, 2 and 3. If those
nodes don't have enough memory, kernel will allocate from slow memory
node 10, 11 and 12 which ever is closer to node 2.
Link: https://lkml.kernel.org/r/20211202123810.267175-3-aneesh.kumar@linux.ibm.com
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Ben Widawsky <ben.widawsky@intel.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Feng Tang <feng.tang@intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: <linux-api@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-14 14:08:17 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-12-16 14:45:37 -05:00
|
|
|
|
2023-09-28 13:24:32 -04:00
|
|
|
|
|
|
|
|
|
2022-12-16 14:45:37 -05:00
|
|
|
|
2023-09-28 13:24:32 -04:00
|
|
|
|
2022-12-16 14:45:37 -05:00
|
|
|
|
mm/mempolicy: add set_mempolicy_home_node syscall
This syscall can be used to set a home node for the MPOL_BIND and
MPOL_PREFERRED_MANY memory policy. Users should use this syscall after
setting up a memory policy for the specified range as shown below.
mbind(p, nr_pages * page_size, MPOL_BIND, new_nodes->maskp,
new_nodes->size + 1, 0);
sys_set_mempolicy_home_node((unsigned long)p, nr_pages * page_size,
home_node, 0);
The syscall allows specifying a home node/preferred node from which
kernel will fulfill memory allocation requests first.
For address range with MPOL_BIND memory policy, if nodemask specifies
more than one node, page allocations will come from the node in the
nodemask with sufficient free memory that is closest to the home
node/preferred node.
For MPOL_PREFERRED_MANY if the nodemask specifies more than one node,
page allocation will come from the node in the nodemask with sufficient
free memory that is closest to the home node/preferred node. If there
is not enough memory in all the nodes specified in the nodemask, the
allocation will be attempted from the closest numa node to the home node
in the system.
This helps applications to hint at a memory allocation preference node
and fallback to _only_ a set of nodes if the memory is not available on
the preferred node. Fallback allocation is attempted from the node
which is nearest to the preferred node.
This helps applications to have control on memory allocation numa nodes
and avoids default fallback to slow memory NUMA nodes. For example a
system with NUMA nodes 1,2 and 3 with DRAM memory and 10, 11 and 12 of
slow memory
new_nodes = numa_bitmask_alloc(nr_nodes);
numa_bitmask_setbit(new_nodes, 1);
numa_bitmask_setbit(new_nodes, 2);
numa_bitmask_setbit(new_nodes, 3);
p = mmap(NULL, nr_pages * page_size, protflag, mapflag, -1, 0);
mbind(p, nr_pages * page_size, MPOL_BIND, new_nodes->maskp, new_nodes->size + 1, 0);
sys_set_mempolicy_home_node(p, nr_pages * page_size, 2, 0);
This will allocate from nodes closer to node 2 and will make sure the
kernel will only allocate from nodes 1, 2, and 3. Memory will not be
allocated from slow memory nodes 10, 11, and 12. This differs from
default MPOL_BIND behavior in that with default MPOL_BIND the allocation
will be attempted from node closer to the local node. One of the
reasons to specify a home node is to allow allocations from cpu less
NUMA node and its nearby NUMA nodes.
With MPOL_PREFERRED_MANY on the other hand will first try to allocate
from the closest node to node 2 from the node list 1, 2 and 3. If those
nodes don't have enough memory, kernel will allocate from slow memory
node 10, 11 and 12 which ever is closer to node 2.
Link: https://lkml.kernel.org/r/20211202123810.267175-3-aneesh.kumar@linux.ibm.com
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Ben Widawsky <ben.widawsky@intel.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Feng Tang <feng.tang@intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: <linux-api@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-14 14:08:17 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-12-16 14:45:37 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: add set_mempolicy_home_node syscall
This syscall can be used to set a home node for the MPOL_BIND and
MPOL_PREFERRED_MANY memory policy. Users should use this syscall after
setting up a memory policy for the specified range as shown below.
mbind(p, nr_pages * page_size, MPOL_BIND, new_nodes->maskp,
new_nodes->size + 1, 0);
sys_set_mempolicy_home_node((unsigned long)p, nr_pages * page_size,
home_node, 0);
The syscall allows specifying a home node/preferred node from which
kernel will fulfill memory allocation requests first.
For address range with MPOL_BIND memory policy, if nodemask specifies
more than one node, page allocations will come from the node in the
nodemask with sufficient free memory that is closest to the home
node/preferred node.
For MPOL_PREFERRED_MANY if the nodemask specifies more than one node,
page allocation will come from the node in the nodemask with sufficient
free memory that is closest to the home node/preferred node. If there
is not enough memory in all the nodes specified in the nodemask, the
allocation will be attempted from the closest numa node to the home node
in the system.
This helps applications to hint at a memory allocation preference node
and fallback to _only_ a set of nodes if the memory is not available on
the preferred node. Fallback allocation is attempted from the node
which is nearest to the preferred node.
This helps applications to have control on memory allocation numa nodes
and avoids default fallback to slow memory NUMA nodes. For example a
system with NUMA nodes 1,2 and 3 with DRAM memory and 10, 11 and 12 of
slow memory
new_nodes = numa_bitmask_alloc(nr_nodes);
numa_bitmask_setbit(new_nodes, 1);
numa_bitmask_setbit(new_nodes, 2);
numa_bitmask_setbit(new_nodes, 3);
p = mmap(NULL, nr_pages * page_size, protflag, mapflag, -1, 0);
mbind(p, nr_pages * page_size, MPOL_BIND, new_nodes->maskp, new_nodes->size + 1, 0);
sys_set_mempolicy_home_node(p, nr_pages * page_size, 2, 0);
This will allocate from nodes closer to node 2 and will make sure the
kernel will only allocate from nodes 1, 2, and 3. Memory will not be
allocated from slow memory nodes 10, 11, and 12. This differs from
default MPOL_BIND behavior in that with default MPOL_BIND the allocation
will be attempted from node closer to the local node. One of the
reasons to specify a home node is to allow allocations from cpu less
NUMA node and its nearby NUMA nodes.
With MPOL_PREFERRED_MANY on the other hand will first try to allocate
from the closest node to node 2 from the node list 1, 2 and 3. If those
nodes don't have enough memory, kernel will allocate from slow memory
node 10, 11 and 12 which ever is closer to node 2.
Link: https://lkml.kernel.org/r/20211202123810.267175-3-aneesh.kumar@linux.ibm.com
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Ben Widawsky <ben.widawsky@intel.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Feng Tang <feng.tang@intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: <linux-api@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-14 14:08:17 -08:00
|
|
|
|
2023-07-28 06:13:21 +02:00
|
|
|
|
mm/mempolicy: add set_mempolicy_home_node syscall
This syscall can be used to set a home node for the MPOL_BIND and
MPOL_PREFERRED_MANY memory policy. Users should use this syscall after
setting up a memory policy for the specified range as shown below.
mbind(p, nr_pages * page_size, MPOL_BIND, new_nodes->maskp,
new_nodes->size + 1, 0);
sys_set_mempolicy_home_node((unsigned long)p, nr_pages * page_size,
home_node, 0);
The syscall allows specifying a home node/preferred node from which
kernel will fulfill memory allocation requests first.
For address range with MPOL_BIND memory policy, if nodemask specifies
more than one node, page allocations will come from the node in the
nodemask with sufficient free memory that is closest to the home
node/preferred node.
For MPOL_PREFERRED_MANY if the nodemask specifies more than one node,
page allocation will come from the node in the nodemask with sufficient
free memory that is closest to the home node/preferred node. If there
is not enough memory in all the nodes specified in the nodemask, the
allocation will be attempted from the closest numa node to the home node
in the system.
This helps applications to hint at a memory allocation preference node
and fallback to _only_ a set of nodes if the memory is not available on
the preferred node. Fallback allocation is attempted from the node
which is nearest to the preferred node.
This helps applications to have control on memory allocation numa nodes
and avoids default fallback to slow memory NUMA nodes. For example a
system with NUMA nodes 1,2 and 3 with DRAM memory and 10, 11 and 12 of
slow memory
new_nodes = numa_bitmask_alloc(nr_nodes);
numa_bitmask_setbit(new_nodes, 1);
numa_bitmask_setbit(new_nodes, 2);
numa_bitmask_setbit(new_nodes, 3);
p = mmap(NULL, nr_pages * page_size, protflag, mapflag, -1, 0);
mbind(p, nr_pages * page_size, MPOL_BIND, new_nodes->maskp, new_nodes->size + 1, 0);
sys_set_mempolicy_home_node(p, nr_pages * page_size, 2, 0);
This will allocate from nodes closer to node 2 and will make sure the
kernel will only allocate from nodes 1, 2, and 3. Memory will not be
allocated from slow memory nodes 10, 11, and 12. This differs from
default MPOL_BIND behavior in that with default MPOL_BIND the allocation
will be attempted from node closer to the local node. One of the
reasons to specify a home node is to allow allocations from cpu less
NUMA node and its nearby NUMA nodes.
With MPOL_PREFERRED_MANY on the other hand will first try to allocate
from the closest node to node 2 from the node list 1, 2 and 3. If those
nodes don't have enough memory, kernel will allocate from slow memory
node 10, 11 and 12 which ever is closer to node 2.
Link: https://lkml.kernel.org/r/20211202123810.267175-3-aneesh.kumar@linux.ibm.com
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Ben Widawsky <ben.widawsky@intel.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Feng Tang <feng.tang@intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: <linux-api@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-14 14:08:17 -08:00
|
|
|
|
2023-04-10 11:22:05 -04:00
|
|
|
|
mm/mempolicy: add set_mempolicy_home_node syscall
This syscall can be used to set a home node for the MPOL_BIND and
MPOL_PREFERRED_MANY memory policy. Users should use this syscall after
setting up a memory policy for the specified range as shown below.
mbind(p, nr_pages * page_size, MPOL_BIND, new_nodes->maskp,
new_nodes->size + 1, 0);
sys_set_mempolicy_home_node((unsigned long)p, nr_pages * page_size,
home_node, 0);
The syscall allows specifying a home node/preferred node from which
kernel will fulfill memory allocation requests first.
For address range with MPOL_BIND memory policy, if nodemask specifies
more than one node, page allocations will come from the node in the
nodemask with sufficient free memory that is closest to the home
node/preferred node.
For MPOL_PREFERRED_MANY if the nodemask specifies more than one node,
page allocation will come from the node in the nodemask with sufficient
free memory that is closest to the home node/preferred node. If there
is not enough memory in all the nodes specified in the nodemask, the
allocation will be attempted from the closest numa node to the home node
in the system.
This helps applications to hint at a memory allocation preference node
and fallback to _only_ a set of nodes if the memory is not available on
the preferred node. Fallback allocation is attempted from the node
which is nearest to the preferred node.
This helps applications to have control on memory allocation numa nodes
and avoids default fallback to slow memory NUMA nodes. For example a
system with NUMA nodes 1,2 and 3 with DRAM memory and 10, 11 and 12 of
slow memory
new_nodes = numa_bitmask_alloc(nr_nodes);
numa_bitmask_setbit(new_nodes, 1);
numa_bitmask_setbit(new_nodes, 2);
numa_bitmask_setbit(new_nodes, 3);
p = mmap(NULL, nr_pages * page_size, protflag, mapflag, -1, 0);
mbind(p, nr_pages * page_size, MPOL_BIND, new_nodes->maskp, new_nodes->size + 1, 0);
sys_set_mempolicy_home_node(p, nr_pages * page_size, 2, 0);
This will allocate from nodes closer to node 2 and will make sure the
kernel will only allocate from nodes 1, 2, and 3. Memory will not be
allocated from slow memory nodes 10, 11, and 12. This differs from
default MPOL_BIND behavior in that with default MPOL_BIND the allocation
will be attempted from node closer to the local node. One of the
reasons to specify a home node is to allow allocations from cpu less
NUMA node and its nearby NUMA nodes.
With MPOL_PREFERRED_MANY on the other hand will first try to allocate
from the closest node to node 2 from the node list 1, 2 and 3. If those
nodes don't have enough memory, kernel will allocate from slow memory
node 10, 11 and 12 which ever is closer to node 2.
Link: https://lkml.kernel.org/r/20211202123810.267175-3-aneesh.kumar@linux.ibm.com
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Ben Widawsky <ben.widawsky@intel.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Feng Tang <feng.tang@intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: <linux-api@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-14 14:08:17 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-03-17 16:12:22 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-10-29 18:16:59 -07:00
|
|
|
|
2018-03-17 16:20:01 +01:00
|
|
|
|
|
|
|
|
|
2005-10-29 18:16:59 -07:00
|
|
|
|
2021-06-30 18:51:03 -07:00
|
|
|
|
2005-10-29 18:16:59 -07:00
|
|
|
|
2021-06-30 18:51:03 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-10-29 18:16:59 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-30 18:51:03 -07:00
|
|
|
|
|
|
|
|
|
2005-10-29 18:16:59 -07:00
|
|
|
|
|
|
|
|
|
2018-03-17 16:20:01 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-03-17 16:00:25 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-01-08 01:00:51 -08:00
|
|
|
|
2010-08-09 17:19:01 -07:00
|
|
|
|
2006-01-08 01:00:51 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-08-09 17:19:01 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-01-08 01:00:51 -08:00
|
|
|
|
2010-08-09 17:19:01 -07:00
|
|
|
|
|
|
|
|
|
2006-01-08 01:00:51 -08:00
|
|
|
|
2010-08-09 17:19:01 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-01-08 01:00:51 -08:00
|
|
|
|
2010-08-09 17:19:01 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-01-08 01:00:51 -08:00
|
|
|
|
|
|
|
|
|
2010-12-02 14:31:13 -08:00
|
|
|
|
2007-10-18 23:40:16 -07:00
|
|
|
|
2006-01-08 01:00:51 -08:00
|
|
|
|
2010-12-02 14:31:13 -08:00
|
|
|
|
2010-08-09 17:19:01 -07:00
|
|
|
|
|
|
|
|
|
2006-01-08 01:00:51 -08:00
|
|
|
|
2012-03-21 16:34:06 -07:00
|
|
|
|
2006-01-08 01:00:51 -08:00
|
|
|
|
2010-08-09 17:19:01 -07:00
|
|
|
|
2006-01-08 01:00:51 -08:00
|
|
|
|
|
|
|
|
|
2017-11-15 17:38:14 -08:00
|
|
|
|
|
|
|
|
|
2006-01-08 01:00:51 -08:00
|
|
|
|
2017-11-15 17:38:14 -08:00
|
|
|
|
2008-11-14 10:39:19 +11:00
|
|
|
|
2006-01-08 01:00:51 -08:00
|
|
|
|
2012-03-21 16:34:06 -07:00
|
|
|
|
2006-01-08 01:00:51 -08:00
|
|
|
|
2008-11-14 10:39:19 +11:00
|
|
|
|
2006-01-08 01:00:51 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-08-09 17:19:01 -07:00
|
|
|
|
2006-01-08 01:00:51 -08:00
|
|
|
|
2012-03-21 16:34:06 -07:00
|
|
|
|
2006-01-08 01:00:51 -08:00
|
|
|
|
|
|
|
|
|
2018-01-31 16:16:15 -08:00
|
|
|
|
2026-01-14 12:22:14 -05:00
|
|
|
|
2018-01-31 16:16:15 -08:00
|
|
|
|
|
|
|
|
|
2006-06-23 02:04:02 -07:00
|
|
|
|
|
|
|
|
|
2012-03-21 16:34:06 -07:00
|
|
|
|
2006-06-23 02:04:02 -07:00
|
|
|
|
2012-03-21 16:34:06 -07:00
|
|
|
|
|
|
|
|
|
2012-04-25 16:01:52 -07:00
|
|
|
|
|
|
|
|
|
2012-03-21 16:34:06 -07:00
|
|
|
|
2012-04-25 16:01:52 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-03-21 16:34:06 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-08-09 17:19:01 -07:00
|
|
|
|
|
|
|
|
|
2006-01-08 01:00:51 -08:00
|
|
|
|
2012-03-21 16:34:06 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-01-08 01:00:51 -08:00
|
|
|
|
|
|
|
|
|
2018-03-17 16:00:25 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-10-29 18:16:59 -07:00
|
|
|
|
2018-03-17 16:20:01 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-10-29 18:16:59 -07:00
|
|
|
|
2007-10-16 01:26:26 -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
|
|
|
|
2005-10-29 18:16:59 -07:00
|
|
|
|
|
|
|
|
|
2019-02-20 22:18:58 -08:00
|
|
|
|
2005-10-29 18:16:59 -07:00
|
|
|
|
|
|
|
|
|
2020-08-11 18:31:16 -07:00
|
|
|
|
|
|
|
|
|
2005-10-29 18:16:59 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-03-17 16:20:01 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-04-01 21:10:52 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-10-09 15:27:50 -07:00
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-10-09 15:27:50 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
2014-10-09 15:27:50 -07:00
|
|
|
|
|
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
|
|
|
|
|
2014-10-09 15:27:50 -07:00
|
|
|
|
|
|
|
|
|
2014-10-09 15:27:57 -07:00
|
|
|
|
2014-10-09 15:27:50 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
|
|
|
|
|
2014-10-09 15:27:50 -07:00
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
2014-10-09 15:27:50 -07:00
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
2014-10-09 15:27:45 -07:00
|
|
|
|
2014-10-09 15:27:57 -07:00
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
|
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-10-09 15:27:48 -07:00
|
|
|
|
2013-10-07 11:29:09 +01:00
|
|
|
|
2014-10-09 15:27:48 -07:00
|
|
|
|
2014-10-09 15:27:43 -07:00
|
|
|
|
2014-10-09 15:27:48 -07:00
|
|
|
|
|
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
2013-10-07 11:29:09 +01:00
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
2014-10-09 15:27:48 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-10-07 11:29:09 +01:00
|
|
|
|
2014-10-09 15:27:48 -07:00
|
|
|
|
2013-10-07 11:29:09 +01:00
|
|
|
|
|
|
|
|
|
2014-10-09 15:27:48 -07:00
|
|
|
|
2014-10-09 15:27:45 -07:00
|
|
|
|
2014-10-09 15:27:48 -07:00
|
|
|
|
2014-10-09 15:27:45 -07:00
|
|
|
|
2013-10-07 11:29:09 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-08-05 08:59:03 +08:00
|
|
|
|
2013-02-22 16:33:22 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-30 18:51:10 -07:00
|
|
|
|
2013-02-22 16:33:22 -08:00
|
|
|
|
|
|
|
|
|
2021-06-30 18:51:10 -07:00
|
|
|
|
2021-05-06 18:06:47 -07:00
|
|
|
|
2021-06-30 18:51:10 -07:00
|
|
|
|
2013-02-22 16:33:22 -08:00
|
|
|
|
2021-06-30 18:51:10 -07:00
|
|
|
|
2013-02-22 16:33:22 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
|
|
|
|
|
2024-02-02 12:02:38 -05:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
2024-02-02 12:02:38 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
|
|
|
|
|
2024-02-02 12:02:38 -05:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2023-10-03 02:20:14 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2023-10-03 02:20:14 -07:00
|
|
|
|
2024-02-02 12:02:38 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2023-10-03 02:20:14 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2006-01-18 17:42:36 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-04-07 15:37:29 -07:00
|
|
|
|
2006-01-18 17:42:36 -08:00
|
|
|
|
2012-06-09 02:40:03 -07:00
|
|
|
|
2014-04-07 15:37:29 -07:00
|
|
|
|
2012-06-09 02:40:03 -07:00
|
|
|
|
2021-09-02 15:00:23 -07:00
|
|
|
|
2014-04-07 15:37:29 -07:00
|
|
|
|
2012-06-09 02:40:03 -07:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: don't handle MPOL_LOCAL like a fake MPOL_PREFERRED policy
MPOL_LOCAL policy has been setup as a real policy, but it is still handled
like a faked POL_PREFERRED policy with one internal MPOL_F_LOCAL flag bit
set, and there are many places having to judge the real 'prefer' or the
'local' policy, which are quite confusing.
In current code, there are 4 cases that MPOL_LOCAL are used:
1. user specifies 'local' policy
2. user specifies 'prefer' policy, but with empty nodemask
3. system 'default' policy is used
4. 'prefer' policy + valid 'preferred' node with MPOL_F_STATIC_NODES
flag set, and when it is 'rebind' to a nodemask which doesn't contains
the 'preferred' node, it will perform as 'local' policy
So make 'local' a real policy instead of a fake 'prefer' one, and kill
MPOL_F_LOCAL bit, which can greatly reduce the confusion for code reading.
For case 4, the logic of mpol_rebind_preferred() is confusing, as Michal
Hocko pointed out:
: I do believe that rebinding preferred policy is just bogus and it should
: be dropped altogether on the ground that a preference is a mere hint from
: userspace where to start the allocation. Unless I am missing something
: cpusets will be always authoritative for the final placement. The
: preferred node just acts as a starting point and it should be really
: preserved when cpusets changes. Otherwise we have a very subtle behavior
: corner cases.
So dump all the tricky transformation between 'prefer' and 'local', and
just record the new nodemask of rebinding.
[feng.tang@intel.com: fix a problem in mpol_set_nodemask(), per Michal Hocko]
Link: https://lkml.kernel.org/r/1622560492-1294-3-git-send-email-feng.tang@intel.com
[feng.tang@intel.com: refine code and comments of mpol_set_nodemask(), per Michal]
Link: https://lkml.kernel.org/r/20210603081807.GE56979@shbuild999.sh.intel.com
Link: https://lkml.kernel.org/r/1622469956-82897-3-git-send-email-feng.tang@intel.com
Signed-off-by: Feng Tang <feng.tang@intel.com>
Suggested-by: Michal Hocko <mhocko@suse.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ben Widawsky <ben.widawsky@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-30 18:51:00 -07:00
|
|
|
|
2014-04-07 15:37:29 -07:00
|
|
|
|
mempolicy: use MPOL_PREFERRED for system-wide default policy
Currently, when one specifies MPOL_DEFAULT via a NUMA memory policy API
[set_mempolicy(), mbind() and internal versions], the kernel simply installs a
NULL struct mempolicy pointer in the appropriate context: task policy, vma
policy, or shared policy. This causes any use of that policy to "fall back"
to the next most specific policy scope.
The only use of MPOL_DEFAULT to mean "local allocation" is in the system
default policy. This requires extra checks/cases for MPOL_DEFAULT in many
mempolicy.c functions.
There is another, "preferred" way to specify local allocation via the APIs.
That is using the MPOL_PREFERRED policy mode with an empty nodemask.
Internally, the empty nodemask gets converted to a preferred_node id of '-1'.
All internal usage of MPOL_PREFERRED will convert the '-1' to the id of the
node local to the cpu where the allocation occurs.
System default policy, except during boot, is hard-coded to "local
allocation". By using the MPOL_PREFERRED mode with a negative value of
preferred node for system default policy, MPOL_DEFAULT will never occur in the
'policy' member of a struct mempolicy. Thus, we can remove all checks for
MPOL_DEFAULT when converting policy to a node id/zonelist in the allocation
paths.
In slab_node() return local node id when policy pointer is NULL. No need to
set a pol value to take the switch default. Replace switch default with
BUG()--i.e., shouldn't happen.
With this patch MPOL_DEFAULT is only used in the APIs, including internal
calls to do_set_mempolicy() and in the display of policy in
/proc/<pid>/numa_maps. It always means "fall back" to the the next most
specific policy scope. This simplifies the description of memory policies
quite a bit, with no visible change in behavior.
get_mempolicy() continues to return MPOL_DEFAULT and an empty nodemask when
the requested policy [task or vma/shared] is NULL. These are the values one
would supply via set_mempolicy() or mbind() to achieve that condition--default
behavior.
This patch updates Documentation to reflect this change.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:18 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-30 18:51:10 -07:00
|
|
|
|
2006-09-27 01:50:08 -07:00
|
|
|
|
2006-01-18 17:42:36 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: add MPOL_PREFERRED_MANY for multiple preferred nodes
Patch series "Introduce multi-preference mempolicy", v7.
This patch series introduces the concept of the MPOL_PREFERRED_MANY
mempolicy. This mempolicy mode can be used with either the
set_mempolicy(2) or mbind(2) interfaces. Like the MPOL_PREFERRED
interface, it allows an application to set a preference for nodes which
will fulfil memory allocation requests. Unlike the MPOL_PREFERRED mode,
it takes a set of nodes. Like the MPOL_BIND interface, it works over a
set of nodes. Unlike MPOL_BIND, it will not cause a SIGSEGV or invoke the
OOM killer if those preferred nodes are not available.
Along with these patches are patches for libnuma, numactl, numademo, and
memhog. They still need some polish, but can be found here:
https://gitlab.com/bwidawsk/numactl/-/tree/prefer-many It allows new
usage: `numactl -P 0,3,4`
The goal of the new mode is to enable some use-cases when using tiered memory
usage models which I've lovingly named.
1a. The Hare - The interconnect is fast enough to meet bandwidth and
latency requirements allowing preference to be given to all nodes with
"fast" memory.
1b. The Indiscriminate Hare - An application knows it wants fast
memory (or perhaps slow memory), but doesn't care which node it runs
on. The application can prefer a set of nodes and then xpu bind to
the local node (cpu, accelerator, etc). This reverses the nodes are
chosen today where the kernel attempts to use local memory to the CPU
whenever possible. This will attempt to use the local accelerator to
the memory.
2. The Tortoise - The administrator (or the application itself) is
aware it only needs slow memory, and so can prefer that.
Much of this is almost achievable with the bind interface, but the bind
interface suffers from an inability to fallback to another set of nodes if
binding fails to all nodes in the nodemask.
Like MPOL_BIND a nodemask is given. Inherently this removes ordering from the
preference.
> /* Set first two nodes as preferred in an 8 node system. */
> const unsigned long nodes = 0x3
> set_mempolicy(MPOL_PREFER_MANY, &nodes, 8);
> /* Mimic interleave policy, but have fallback *.
> const unsigned long nodes = 0xaa
> set_mempolicy(MPOL_PREFER_MANY, &nodes, 8);
Some internal discussion took place around the interface. There are two
alternatives which we have discussed, plus one I stuck in:
1. Ordered list of nodes. Currently it's believed that the added
complexity is nod needed for expected usecases.
2. A flag for bind to allow falling back to other nodes. This
confuses the notion of binding and is less flexible than the current
solution.
3. Create flags or new modes that helps with some ordering. This
offers both a friendlier API as well as a solution for more customized
usage. It's unknown if it's worth the complexity to support this.
Here is sample code for how this might work:
> // Prefer specific nodes for some something wacky
> set_mempolicy(MPOL_PREFER_MANY, 0x17c, 1024);
>
> // Default
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_SOCKET, NULL, 0);
> // which is the same as
> set_mempolicy(MPOL_DEFAULT, NULL, 0);
>
> // The Hare
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE, NULL, 0);
>
> // The Tortoise
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE_REV, NULL, 0);
>
> // Prefer the fast memory of the first two sockets
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE, -1, 2);
>
This patch (of 5):
The NUMA APIs currently allow passing in a "preferred node" as a single
bit set in a nodemask. If more than one bit it set, bits after the first
are ignored.
This single node is generally OK for location-based NUMA where memory
being allocated will eventually be operated on by a single CPU. However,
in systems with multiple memory types, folks want to target a *type* of
memory instead of a location. For instance, someone might want some
high-bandwidth memory but do not care about the CPU next to which it is
allocated. Or, they want a cheap, high capacity allocation and want to
target all NUMA nodes which have persistent memory in volatile mode. In
both of these cases, the application wants to target a *set* of nodes, but
does not want strict MPOL_BIND behavior as that could lead to OOM killer
or SIGSEGV.
So add MPOL_PREFERRED_MANY policy to support the multiple preferred nodes
requirement. This is not a pie-in-the-sky dream for an API. This was a
response to a specific ask of more than one group at Intel. Specifically:
1. There are existing libraries that target memory types such as
https://github.com/memkind/memkind. These are known to suffer from
SIGSEGV's when memory is low on targeted memory "kinds" that span more
than one node. The MCDRAM on a Xeon Phi in "Cluster on Die" mode is an
example of this.
2. Volatile-use persistent memory users want to have a memory policy
which is targeted at either "cheap and slow" (PMEM) or "expensive and
fast" (DRAM). However, they do not want to experience allocation
failures when the targeted type is unavailable.
3. Allocate-then-run. Generally, we let the process scheduler decide
on which physical CPU to run a task. That location provides a default
allocation policy, and memory availability is not generally considered
when placing tasks. For situations where memory is valuable and
constrained, some users want to allocate memory first, *then* allocate
close compute resources to the allocation. This is the reverse of the
normal (CPU) model. Accelerators such as GPUs that operate on
core-mm-managed memory are interested in this model.
A check is added in sanitize_mpol_flags() to not permit 'prefer_many'
policy to be used for now, and will be removed in later patch after all
implementations for 'prefer_many' are ready, as suggested by Michal Hocko.
[mhocko@kernel.org: suggest to refine policy_node/policy_nodemask handling]
Link: https://lkml.kernel.org/r/1627970362-61305-1-git-send-email-feng.tang@intel.com
Link: https://lore.kernel.org/r/20200630212517.308045-4-ben.widawsky@intel.com
Link: https://lkml.kernel.org/r/1627970362-61305-2-git-send-email-feng.tang@intel.com
Co-developed-by: Ben Widawsky <ben.widawsky@intel.com>
Signed-off-by: Ben Widawsky <ben.widawsky@intel.com>
Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com>
Signed-off-by: Feng Tang <feng.tang@intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Huang Ying <ying.huang@intel.com>b
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-09-02 15:00:06 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-05-19 17:14:10 -07:00
|
|
|
|
|
|
|
|
|
2006-01-18 17:42:36 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-04-28 02:12:18 -07:00
|
|
|
|
|
|
|
|
|
2016-10-07 16:59:12 -07:00
|
|
|
|
2016-05-19 17:14:10 -07:00
|
|
|
|
2021-06-30 18:51:10 -07:00
|
|
|
|
2024-07-29 14:47:17 +05:30
|
|
|
|
2008-04-28 02:12:17 -07:00
|
|
|
|
mm/mempolicy: don't handle MPOL_LOCAL like a fake MPOL_PREFERRED policy
MPOL_LOCAL policy has been setup as a real policy, but it is still handled
like a faked POL_PREFERRED policy with one internal MPOL_F_LOCAL flag bit
set, and there are many places having to judge the real 'prefer' or the
'local' policy, which are quite confusing.
In current code, there are 4 cases that MPOL_LOCAL are used:
1. user specifies 'local' policy
2. user specifies 'prefer' policy, but with empty nodemask
3. system 'default' policy is used
4. 'prefer' policy + valid 'preferred' node with MPOL_F_STATIC_NODES
flag set, and when it is 'rebind' to a nodemask which doesn't contains
the 'preferred' node, it will perform as 'local' policy
So make 'local' a real policy instead of a fake 'prefer' one, and kill
MPOL_F_LOCAL bit, which can greatly reduce the confusion for code reading.
For case 4, the logic of mpol_rebind_preferred() is confusing, as Michal
Hocko pointed out:
: I do believe that rebinding preferred policy is just bogus and it should
: be dropped altogether on the ground that a preference is a mere hint from
: userspace where to start the allocation. Unless I am missing something
: cpusets will be always authoritative for the final placement. The
: preferred node just acts as a starting point and it should be really
: preserved when cpusets changes. Otherwise we have a very subtle behavior
: corner cases.
So dump all the tricky transformation between 'prefer' and 'local', and
just record the new nodemask of rebinding.
[feng.tang@intel.com: fix a problem in mpol_set_nodemask(), per Michal Hocko]
Link: https://lkml.kernel.org/r/1622560492-1294-3-git-send-email-feng.tang@intel.com
[feng.tang@intel.com: refine code and comments of mpol_set_nodemask(), per Michal]
Link: https://lkml.kernel.org/r/20210603081807.GE56979@shbuild999.sh.intel.com
Link: https://lkml.kernel.org/r/1622469956-82897-3-git-send-email-feng.tang@intel.com
Signed-off-by: Feng Tang <feng.tang@intel.com>
Suggested-by: Michal Hocko <mhocko@suse.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ben Widawsky <ben.widawsky@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-30 18:51:00 -07:00
|
|
|
|
|
|
|
|
|
2006-01-18 17:42:36 -08:00
|
|
|
|
|
|
|
|
|
mempolicy: use MPOL_PREFERRED for system-wide default policy
Currently, when one specifies MPOL_DEFAULT via a NUMA memory policy API
[set_mempolicy(), mbind() and internal versions], the kernel simply installs a
NULL struct mempolicy pointer in the appropriate context: task policy, vma
policy, or shared policy. This causes any use of that policy to "fall back"
to the next most specific policy scope.
The only use of MPOL_DEFAULT to mean "local allocation" is in the system
default policy. This requires extra checks/cases for MPOL_DEFAULT in many
mempolicy.c functions.
There is another, "preferred" way to specify local allocation via the APIs.
That is using the MPOL_PREFERRED policy mode with an empty nodemask.
Internally, the empty nodemask gets converted to a preferred_node id of '-1'.
All internal usage of MPOL_PREFERRED will convert the '-1' to the id of the
node local to the cpu where the allocation occurs.
System default policy, except during boot, is hard-coded to "local
allocation". By using the MPOL_PREFERRED mode with a negative value of
preferred node for system default policy, MPOL_DEFAULT will never occur in the
'policy' member of a struct mempolicy. Thus, we can remove all checks for
MPOL_DEFAULT when converting policy to a node id/zonelist in the allocation
paths.
In slab_node() return local node id when policy pointer is NULL. No need to
set a pol value to take the switch default. Replace switch default with
BUG()--i.e., shouldn't happen.
With this patch MPOL_DEFAULT is only used in the APIs, including internal
calls to do_set_mempolicy() and in the display of policy in
/proc/<pid>/numa_maps. It always means "fall back" to the the next most
specific policy scope. This simplifies the description of memory policies
quite a bit, with no visible change in behavior.
get_mempolicy() continues to return MPOL_DEFAULT and an empty nodemask when
the requested policy [task or vma/shared] is NULL. These are the values one
would supply via set_mempolicy() or mbind() to achieve that condition--default
behavior.
This patch updates Documentation to reflect this change.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:18 -07:00
|
|
|
|
2006-01-18 17:42:36 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-02-02 12:02:36 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: Weighted Interleave Auto-tuning
On machines with multiple memory nodes, interleaving page allocations
across nodes allows for better utilization of each node's bandwidth.
Previous work by Gregory Price [1] introduced weighted interleave, which
allowed for pages to be allocated across nodes according to user-set
ratios.
Ideally, these weights should be proportional to their bandwidth, so that
under bandwidth pressure, each node uses its maximal efficient bandwidth
and prevents latency from increasing exponentially.
Previously, weighted interleave's default weights were just 1s -- which
would be equivalent to the (unweighted) interleave mempolicy, which goes
through the nodes in a round-robin fashion, ignoring bandwidth
information.
This patch has two main goals: First, it makes weighted interleave easier
to use for users who wish to relieve bandwidth pressure when using nodes
with varying bandwidth (CXL). By providing a set of "real" default
weights that just work out of the box, users who might not have the
capability (or wish to) perform experimentation to find the most optimal
weights for their system can still take advantage of bandwidth-informed
weighted interleave.
Second, it allows for weighted interleave to dynamically adjust to
hotplugged memory with new bandwidth information. Instead of manually
updating node weights every time new bandwidth information is reported or
taken off, weighted interleave adjusts and provides a new set of default
weights for weighted interleave to use when there is a change in bandwidth
information.
To meet these goals, this patch introduces an auto-configuration mode for
the interleave weights that provides a reasonable set of default weights,
calculated using bandwidth data reported by the system. In auto mode,
weights are dynamically adjusted based on whatever the current bandwidth
information reports (and responds to hotplug events).
This patch still supports users manually writing weights into the nodeN
sysfs interface by entering into manual mode. When a user enters manual
mode, the system stops dynamically updating any of the node weights, even
during hotplug events that shift the optimal weight distribution.
A new sysfs interface "auto" is introduced, which allows users to switch
between the auto (writing 1 or Y) and manual (writing 0 or N) modes. The
system also automatically enters manual mode when a nodeN interface is
manually written to.
There is one functional change that this patch makes to the existing
weighted_interleave ABI: previously, writing 0 directly to a nodeN
interface was said to reset the weight to the system default. Before this
patch, the default for all weights were 1, which meant that writing 0 and
1 were functionally equivalent. With this patch, writing 0 is invalid.
Link: https://lkml.kernel.org/r/20250520141236.2987309-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: wordsmithing changes, simplification, fixes]
Link: https://lkml.kernel.org/r/20250511025840.2410154-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: remove auto_kobj_attr field from struct sysfs_wi_group]
Link: https://lkml.kernel.org/r/20250512142511.3959833-1-joshua.hahnjy@gmail.com
https://lore.kernel.org/linux-mm/20240202170238.90004-1-gregory.price@memverge.com/ [1]
Link: https://lkml.kernel.org/r/20250505182328.4148265-1-joshua.hahnjy@gmail.com
Co-developed-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Yunjeong Mun <yunjeong.mun@sk.com>
Suggested-by: Oscar Salvador <osalvador@suse.de>
Suggested-by: Ying Huang <ying.huang@linux.alibaba.com>
Suggested-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Reviewed-by: Honggyu Kim <honggyu.kim@sk.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-05 11:23:28 -07:00
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: Weighted Interleave Auto-tuning
On machines with multiple memory nodes, interleaving page allocations
across nodes allows for better utilization of each node's bandwidth.
Previous work by Gregory Price [1] introduced weighted interleave, which
allowed for pages to be allocated across nodes according to user-set
ratios.
Ideally, these weights should be proportional to their bandwidth, so that
under bandwidth pressure, each node uses its maximal efficient bandwidth
and prevents latency from increasing exponentially.
Previously, weighted interleave's default weights were just 1s -- which
would be equivalent to the (unweighted) interleave mempolicy, which goes
through the nodes in a round-robin fashion, ignoring bandwidth
information.
This patch has two main goals: First, it makes weighted interleave easier
to use for users who wish to relieve bandwidth pressure when using nodes
with varying bandwidth (CXL). By providing a set of "real" default
weights that just work out of the box, users who might not have the
capability (or wish to) perform experimentation to find the most optimal
weights for their system can still take advantage of bandwidth-informed
weighted interleave.
Second, it allows for weighted interleave to dynamically adjust to
hotplugged memory with new bandwidth information. Instead of manually
updating node weights every time new bandwidth information is reported or
taken off, weighted interleave adjusts and provides a new set of default
weights for weighted interleave to use when there is a change in bandwidth
information.
To meet these goals, this patch introduces an auto-configuration mode for
the interleave weights that provides a reasonable set of default weights,
calculated using bandwidth data reported by the system. In auto mode,
weights are dynamically adjusted based on whatever the current bandwidth
information reports (and responds to hotplug events).
This patch still supports users manually writing weights into the nodeN
sysfs interface by entering into manual mode. When a user enters manual
mode, the system stops dynamically updating any of the node weights, even
during hotplug events that shift the optimal weight distribution.
A new sysfs interface "auto" is introduced, which allows users to switch
between the auto (writing 1 or Y) and manual (writing 0 or N) modes. The
system also automatically enters manual mode when a nodeN interface is
manually written to.
There is one functional change that this patch makes to the existing
weighted_interleave ABI: previously, writing 0 directly to a nodeN
interface was said to reset the weight to the system default. Before this
patch, the default for all weights were 1, which meant that writing 0 and
1 were functionally equivalent. With this patch, writing 0 is invalid.
Link: https://lkml.kernel.org/r/20250520141236.2987309-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: wordsmithing changes, simplification, fixes]
Link: https://lkml.kernel.org/r/20250511025840.2410154-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: remove auto_kobj_attr field from struct sysfs_wi_group]
Link: https://lkml.kernel.org/r/20250512142511.3959833-1-joshua.hahnjy@gmail.com
https://lore.kernel.org/linux-mm/20240202170238.90004-1-gregory.price@memverge.com/ [1]
Link: https://lkml.kernel.org/r/20250505182328.4148265-1-joshua.hahnjy@gmail.com
Co-developed-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Yunjeong Mun <yunjeong.mun@sk.com>
Suggested-by: Oscar Salvador <osalvador@suse.de>
Suggested-by: Ying Huang <ying.huang@linux.alibaba.com>
Suggested-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Reviewed-by: Honggyu Kim <honggyu.kim@sk.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-05 11:23:28 -07:00
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: Weighted Interleave Auto-tuning
On machines with multiple memory nodes, interleaving page allocations
across nodes allows for better utilization of each node's bandwidth.
Previous work by Gregory Price [1] introduced weighted interleave, which
allowed for pages to be allocated across nodes according to user-set
ratios.
Ideally, these weights should be proportional to their bandwidth, so that
under bandwidth pressure, each node uses its maximal efficient bandwidth
and prevents latency from increasing exponentially.
Previously, weighted interleave's default weights were just 1s -- which
would be equivalent to the (unweighted) interleave mempolicy, which goes
through the nodes in a round-robin fashion, ignoring bandwidth
information.
This patch has two main goals: First, it makes weighted interleave easier
to use for users who wish to relieve bandwidth pressure when using nodes
with varying bandwidth (CXL). By providing a set of "real" default
weights that just work out of the box, users who might not have the
capability (or wish to) perform experimentation to find the most optimal
weights for their system can still take advantage of bandwidth-informed
weighted interleave.
Second, it allows for weighted interleave to dynamically adjust to
hotplugged memory with new bandwidth information. Instead of manually
updating node weights every time new bandwidth information is reported or
taken off, weighted interleave adjusts and provides a new set of default
weights for weighted interleave to use when there is a change in bandwidth
information.
To meet these goals, this patch introduces an auto-configuration mode for
the interleave weights that provides a reasonable set of default weights,
calculated using bandwidth data reported by the system. In auto mode,
weights are dynamically adjusted based on whatever the current bandwidth
information reports (and responds to hotplug events).
This patch still supports users manually writing weights into the nodeN
sysfs interface by entering into manual mode. When a user enters manual
mode, the system stops dynamically updating any of the node weights, even
during hotplug events that shift the optimal weight distribution.
A new sysfs interface "auto" is introduced, which allows users to switch
between the auto (writing 1 or Y) and manual (writing 0 or N) modes. The
system also automatically enters manual mode when a nodeN interface is
manually written to.
There is one functional change that this patch makes to the existing
weighted_interleave ABI: previously, writing 0 directly to a nodeN
interface was said to reset the weight to the system default. Before this
patch, the default for all weights were 1, which meant that writing 0 and
1 were functionally equivalent. With this patch, writing 0 is invalid.
Link: https://lkml.kernel.org/r/20250520141236.2987309-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: wordsmithing changes, simplification, fixes]
Link: https://lkml.kernel.org/r/20250511025840.2410154-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: remove auto_kobj_attr field from struct sysfs_wi_group]
Link: https://lkml.kernel.org/r/20250512142511.3959833-1-joshua.hahnjy@gmail.com
https://lore.kernel.org/linux-mm/20240202170238.90004-1-gregory.price@memverge.com/ [1]
Link: https://lkml.kernel.org/r/20250505182328.4148265-1-joshua.hahnjy@gmail.com
Co-developed-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Yunjeong Mun <yunjeong.mun@sk.com>
Suggested-by: Oscar Salvador <osalvador@suse.de>
Suggested-by: Ying Huang <ying.huang@linux.alibaba.com>
Suggested-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Reviewed-by: Honggyu Kim <honggyu.kim@sk.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-05 11:23:28 -07:00
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: Weighted Interleave Auto-tuning
On machines with multiple memory nodes, interleaving page allocations
across nodes allows for better utilization of each node's bandwidth.
Previous work by Gregory Price [1] introduced weighted interleave, which
allowed for pages to be allocated across nodes according to user-set
ratios.
Ideally, these weights should be proportional to their bandwidth, so that
under bandwidth pressure, each node uses its maximal efficient bandwidth
and prevents latency from increasing exponentially.
Previously, weighted interleave's default weights were just 1s -- which
would be equivalent to the (unweighted) interleave mempolicy, which goes
through the nodes in a round-robin fashion, ignoring bandwidth
information.
This patch has two main goals: First, it makes weighted interleave easier
to use for users who wish to relieve bandwidth pressure when using nodes
with varying bandwidth (CXL). By providing a set of "real" default
weights that just work out of the box, users who might not have the
capability (or wish to) perform experimentation to find the most optimal
weights for their system can still take advantage of bandwidth-informed
weighted interleave.
Second, it allows for weighted interleave to dynamically adjust to
hotplugged memory with new bandwidth information. Instead of manually
updating node weights every time new bandwidth information is reported or
taken off, weighted interleave adjusts and provides a new set of default
weights for weighted interleave to use when there is a change in bandwidth
information.
To meet these goals, this patch introduces an auto-configuration mode for
the interleave weights that provides a reasonable set of default weights,
calculated using bandwidth data reported by the system. In auto mode,
weights are dynamically adjusted based on whatever the current bandwidth
information reports (and responds to hotplug events).
This patch still supports users manually writing weights into the nodeN
sysfs interface by entering into manual mode. When a user enters manual
mode, the system stops dynamically updating any of the node weights, even
during hotplug events that shift the optimal weight distribution.
A new sysfs interface "auto" is introduced, which allows users to switch
between the auto (writing 1 or Y) and manual (writing 0 or N) modes. The
system also automatically enters manual mode when a nodeN interface is
manually written to.
There is one functional change that this patch makes to the existing
weighted_interleave ABI: previously, writing 0 directly to a nodeN
interface was said to reset the weight to the system default. Before this
patch, the default for all weights were 1, which meant that writing 0 and
1 were functionally equivalent. With this patch, writing 0 is invalid.
Link: https://lkml.kernel.org/r/20250520141236.2987309-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: wordsmithing changes, simplification, fixes]
Link: https://lkml.kernel.org/r/20250511025840.2410154-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: remove auto_kobj_attr field from struct sysfs_wi_group]
Link: https://lkml.kernel.org/r/20250512142511.3959833-1-joshua.hahnjy@gmail.com
https://lore.kernel.org/linux-mm/20240202170238.90004-1-gregory.price@memverge.com/ [1]
Link: https://lkml.kernel.org/r/20250505182328.4148265-1-joshua.hahnjy@gmail.com
Co-developed-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Yunjeong Mun <yunjeong.mun@sk.com>
Suggested-by: Oscar Salvador <osalvador@suse.de>
Suggested-by: Ying Huang <ying.huang@linux.alibaba.com>
Suggested-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Reviewed-by: Honggyu Kim <honggyu.kim@sk.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-05 11:23:28 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
mm/mempolicy: Weighted Interleave Auto-tuning
On machines with multiple memory nodes, interleaving page allocations
across nodes allows for better utilization of each node's bandwidth.
Previous work by Gregory Price [1] introduced weighted interleave, which
allowed for pages to be allocated across nodes according to user-set
ratios.
Ideally, these weights should be proportional to their bandwidth, so that
under bandwidth pressure, each node uses its maximal efficient bandwidth
and prevents latency from increasing exponentially.
Previously, weighted interleave's default weights were just 1s -- which
would be equivalent to the (unweighted) interleave mempolicy, which goes
through the nodes in a round-robin fashion, ignoring bandwidth
information.
This patch has two main goals: First, it makes weighted interleave easier
to use for users who wish to relieve bandwidth pressure when using nodes
with varying bandwidth (CXL). By providing a set of "real" default
weights that just work out of the box, users who might not have the
capability (or wish to) perform experimentation to find the most optimal
weights for their system can still take advantage of bandwidth-informed
weighted interleave.
Second, it allows for weighted interleave to dynamically adjust to
hotplugged memory with new bandwidth information. Instead of manually
updating node weights every time new bandwidth information is reported or
taken off, weighted interleave adjusts and provides a new set of default
weights for weighted interleave to use when there is a change in bandwidth
information.
To meet these goals, this patch introduces an auto-configuration mode for
the interleave weights that provides a reasonable set of default weights,
calculated using bandwidth data reported by the system. In auto mode,
weights are dynamically adjusted based on whatever the current bandwidth
information reports (and responds to hotplug events).
This patch still supports users manually writing weights into the nodeN
sysfs interface by entering into manual mode. When a user enters manual
mode, the system stops dynamically updating any of the node weights, even
during hotplug events that shift the optimal weight distribution.
A new sysfs interface "auto" is introduced, which allows users to switch
between the auto (writing 1 or Y) and manual (writing 0 or N) modes. The
system also automatically enters manual mode when a nodeN interface is
manually written to.
There is one functional change that this patch makes to the existing
weighted_interleave ABI: previously, writing 0 directly to a nodeN
interface was said to reset the weight to the system default. Before this
patch, the default for all weights were 1, which meant that writing 0 and
1 were functionally equivalent. With this patch, writing 0 is invalid.
Link: https://lkml.kernel.org/r/20250520141236.2987309-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: wordsmithing changes, simplification, fixes]
Link: https://lkml.kernel.org/r/20250511025840.2410154-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: remove auto_kobj_attr field from struct sysfs_wi_group]
Link: https://lkml.kernel.org/r/20250512142511.3959833-1-joshua.hahnjy@gmail.com
https://lore.kernel.org/linux-mm/20240202170238.90004-1-gregory.price@memverge.com/ [1]
Link: https://lkml.kernel.org/r/20250505182328.4148265-1-joshua.hahnjy@gmail.com
Co-developed-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Yunjeong Mun <yunjeong.mun@sk.com>
Suggested-by: Oscar Salvador <osalvador@suse.de>
Suggested-by: Ying Huang <ying.huang@linux.alibaba.com>
Suggested-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Reviewed-by: Honggyu Kim <honggyu.kim@sk.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-05 11:23:28 -07:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-05-19 17:11:43 -07:00
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-05-19 17:11:43 -07:00
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2024-02-02 12:02:36 -05:00
|
|
|
|
2021-09-08 18:10:20 -07:00
|
|
|
|
2016-05-19 17:11:43 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2024-02-02 12:02:36 -05:00
|
|
|
|
mempolicy: add MPOL_F_STATIC_NODES flag
Add an optional mempolicy mode flag, MPOL_F_STATIC_NODES, that suppresses the
node remap when the policy is rebound.
Adds another member to struct mempolicy, nodemask_t user_nodemask, as part of
a union with cpuset_mems_allowed:
struct mempolicy {
...
union {
nodemask_t cpuset_mems_allowed;
nodemask_t user_nodemask;
} w;
}
that stores the the nodemask that the user passed when he or she created the
mempolicy via set_mempolicy() or mbind(). When using MPOL_F_STATIC_NODES,
which is passed with any mempolicy mode, the user's passed nodemask
intersected with the VMA or task's allowed nodes is always used when
determining the preferred node, setting the MPOL_BIND zonelist, or creating
the interleave nodemask. This happens whenever the policy is rebound,
including when a task's cpuset assignment changes or the cpuset's mems are
changed.
This creates an interesting side-effect in that it allows the mempolicy
"intent" to lie dormant and uneffected until it has access to the node(s) that
it desires. For example, if you currently ask for an interleaved policy over
a set of nodes that you do not have access to, the mempolicy is not created
and the task continues to use the previous policy. With this change, however,
it is possible to create the same mempolicy; it is only effected when access
to nodes in the nodemask is acquired.
It is also possible to mount tmpfs with the static nodemask behavior when
specifying a node or nodemask. To do this, simply add "=static" immediately
following the mempolicy mode at mount time:
mount -o remount mpol=interleave=static:1-3
Also removes mpol_check_policy() and folds its logic into mpol_new() since it
is now obsoleted. The unused vma_mpol_equal() is also removed.
Cc: Paul Jackson <pj@sgi.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:12:27 -07:00
|
|
|
|
|
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
2021-09-08 18:10:20 -07:00
|
|
|
|
2016-05-19 17:11:43 -07:00
|
|
|
|
2021-09-08 18:10:20 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-01-06 00:10:46 -08:00
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
2006-01-06 00:10:46 -08:00
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-08-31 21:27:53 -07:00
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-08-31 21:27:53 -07:00
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-01-06 00:10:46 -08:00
|
|
|
|
|
|
|
|
|
2006-02-03 21:51:14 +01:00
|
|
|
|
Fix NUMA Memory Policy Reference Counting
This patch proposes fixes to the reference counting of memory policy in the
page allocation paths and in show_numa_map(). Extracted from my "Memory
Policy Cleanups and Enhancements" series as stand-alone.
Shared policy lookup [shmem] has always added a reference to the policy,
but this was never unrefed after page allocation or after formatting the
numa map data.
Default system policy should not require additional ref counting, nor
should the current task's task policy. However, show_numa_map() calls
get_vma_policy() to examine what may be [likely is] another task's policy.
The latter case needs protection against freeing of the policy.
This patch adds a reference count to a mempolicy returned by
get_vma_policy() when the policy is a vma policy or another task's
mempolicy. Again, shared policy is already reference counted on lookup. A
matching "unref" [__mpol_free()] is performed in alloc_page_vma() for
shared and vma policies, and in show_numa_map() for shared and another
task's mempolicy. We can call __mpol_free() directly, saving an admittedly
inexpensive inline NULL test, because we know we have a non-NULL policy.
Handling policy ref counts for hugepages is a bit trickier.
huge_zonelist() returns a zone list that might come from a shared or vma
'BIND policy. In this case, we should hold the reference until after the
huge page allocation in dequeue_hugepage(). The patch modifies
huge_zonelist() to return a pointer to the mempolicy if it needs to be
unref'd after allocation.
Kernel Build [16cpu, 32GB, ia64] - average of 10 runs:
w/o patch w/ refcount patch
Avg Std Devn Avg Std Devn
Real: 100.59 0.38 100.63 0.43
User: 1209.60 0.37 1209.91 0.31
System: 81.52 0.42 81.64 0.34
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Acked-by: Andi Kleen <ak@suse.de>
Cc: Christoph Lameter <clameter@sgi.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-09-18 22:46:47 -07:00
|
|
|
|
2017-07-06 15:40:03 -07:00
|
|
|
|
2014-06-04 16:08:18 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: add MPOL_PREFERRED_MANY for multiple preferred nodes
Patch series "Introduce multi-preference mempolicy", v7.
This patch series introduces the concept of the MPOL_PREFERRED_MANY
mempolicy. This mempolicy mode can be used with either the
set_mempolicy(2) or mbind(2) interfaces. Like the MPOL_PREFERRED
interface, it allows an application to set a preference for nodes which
will fulfil memory allocation requests. Unlike the MPOL_PREFERRED mode,
it takes a set of nodes. Like the MPOL_BIND interface, it works over a
set of nodes. Unlike MPOL_BIND, it will not cause a SIGSEGV or invoke the
OOM killer if those preferred nodes are not available.
Along with these patches are patches for libnuma, numactl, numademo, and
memhog. They still need some polish, but can be found here:
https://gitlab.com/bwidawsk/numactl/-/tree/prefer-many It allows new
usage: `numactl -P 0,3,4`
The goal of the new mode is to enable some use-cases when using tiered memory
usage models which I've lovingly named.
1a. The Hare - The interconnect is fast enough to meet bandwidth and
latency requirements allowing preference to be given to all nodes with
"fast" memory.
1b. The Indiscriminate Hare - An application knows it wants fast
memory (or perhaps slow memory), but doesn't care which node it runs
on. The application can prefer a set of nodes and then xpu bind to
the local node (cpu, accelerator, etc). This reverses the nodes are
chosen today where the kernel attempts to use local memory to the CPU
whenever possible. This will attempt to use the local accelerator to
the memory.
2. The Tortoise - The administrator (or the application itself) is
aware it only needs slow memory, and so can prefer that.
Much of this is almost achievable with the bind interface, but the bind
interface suffers from an inability to fallback to another set of nodes if
binding fails to all nodes in the nodemask.
Like MPOL_BIND a nodemask is given. Inherently this removes ordering from the
preference.
> /* Set first two nodes as preferred in an 8 node system. */
> const unsigned long nodes = 0x3
> set_mempolicy(MPOL_PREFER_MANY, &nodes, 8);
> /* Mimic interleave policy, but have fallback *.
> const unsigned long nodes = 0xaa
> set_mempolicy(MPOL_PREFER_MANY, &nodes, 8);
Some internal discussion took place around the interface. There are two
alternatives which we have discussed, plus one I stuck in:
1. Ordered list of nodes. Currently it's believed that the added
complexity is nod needed for expected usecases.
2. A flag for bind to allow falling back to other nodes. This
confuses the notion of binding and is less flexible than the current
solution.
3. Create flags or new modes that helps with some ordering. This
offers both a friendlier API as well as a solution for more customized
usage. It's unknown if it's worth the complexity to support this.
Here is sample code for how this might work:
> // Prefer specific nodes for some something wacky
> set_mempolicy(MPOL_PREFER_MANY, 0x17c, 1024);
>
> // Default
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_SOCKET, NULL, 0);
> // which is the same as
> set_mempolicy(MPOL_DEFAULT, NULL, 0);
>
> // The Hare
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE, NULL, 0);
>
> // The Tortoise
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE_REV, NULL, 0);
>
> // Prefer the fast memory of the first two sockets
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE, -1, 2);
>
This patch (of 5):
The NUMA APIs currently allow passing in a "preferred node" as a single
bit set in a nodemask. If more than one bit it set, bits after the first
are ignored.
This single node is generally OK for location-based NUMA where memory
being allocated will eventually be operated on by a single CPU. However,
in systems with multiple memory types, folks want to target a *type* of
memory instead of a location. For instance, someone might want some
high-bandwidth memory but do not care about the CPU next to which it is
allocated. Or, they want a cheap, high capacity allocation and want to
target all NUMA nodes which have persistent memory in volatile mode. In
both of these cases, the application wants to target a *set* of nodes, but
does not want strict MPOL_BIND behavior as that could lead to OOM killer
or SIGSEGV.
So add MPOL_PREFERRED_MANY policy to support the multiple preferred nodes
requirement. This is not a pie-in-the-sky dream for an API. This was a
response to a specific ask of more than one group at Intel. Specifically:
1. There are existing libraries that target memory types such as
https://github.com/memkind/memkind. These are known to suffer from
SIGSEGV's when memory is low on targeted memory "kinds" that span more
than one node. The MCDRAM on a Xeon Phi in "Cluster on Die" mode is an
example of this.
2. Volatile-use persistent memory users want to have a memory policy
which is targeted at either "cheap and slow" (PMEM) or "expensive and
fast" (DRAM). However, they do not want to experience allocation
failures when the targeted type is unavailable.
3. Allocate-then-run. Generally, we let the process scheduler decide
on which physical CPU to run a task. That location provides a default
allocation policy, and memory availability is not generally considered
when placing tasks. For situations where memory is valuable and
constrained, some users want to allocate memory first, *then* allocate
close compute resources to the allocation. This is the reverse of the
normal (CPU) model. Accelerators such as GPUs that operate on
core-mm-managed memory are interested in this model.
A check is added in sanitize_mpol_flags() to not permit 'prefer_many'
policy to be used for now, and will be removed in later patch after all
implementations for 'prefer_many' are ready, as suggested by Michal Hocko.
[mhocko@kernel.org: suggest to refine policy_node/policy_nodemask handling]
Link: https://lkml.kernel.org/r/1627970362-61305-1-git-send-email-feng.tang@intel.com
Link: https://lore.kernel.org/r/20200630212517.308045-4-ben.widawsky@intel.com
Link: https://lkml.kernel.org/r/1627970362-61305-2-git-send-email-feng.tang@intel.com
Co-developed-by: Ben Widawsky <ben.widawsky@intel.com>
Signed-off-by: Ben Widawsky <ben.widawsky@intel.com>
Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com>
Signed-off-by: Feng Tang <feng.tang@intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Huang Ying <ying.huang@intel.com>b
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-09-02 15:00:06 -07:00
|
|
|
|
Fix NUMA Memory Policy Reference Counting
This patch proposes fixes to the reference counting of memory policy in the
page allocation paths and in show_numa_map(). Extracted from my "Memory
Policy Cleanups and Enhancements" series as stand-alone.
Shared policy lookup [shmem] has always added a reference to the policy,
but this was never unrefed after page allocation or after formatting the
numa map data.
Default system policy should not require additional ref counting, nor
should the current task's task policy. However, show_numa_map() calls
get_vma_policy() to examine what may be [likely is] another task's policy.
The latter case needs protection against freeing of the policy.
This patch adds a reference count to a mempolicy returned by
get_vma_policy() when the policy is a vma policy or another task's
mempolicy. Again, shared policy is already reference counted on lookup. A
matching "unref" [__mpol_free()] is performed in alloc_page_vma() for
shared and vma policies, and in show_numa_map() for shared and another
task's mempolicy. We can call __mpol_free() directly, saving an admittedly
inexpensive inline NULL test, because we know we have a non-NULL policy.
Handling policy ref counts for hugepages is a bit trickier.
huge_zonelist() returns a zone list that might come from a shared or vma
'BIND policy. In this case, we should hold the reference until after the
huge page allocation in dequeue_hugepage(). The patch modifies
huge_zonelist() to return a pointer to the mempolicy if it needs to be
unref'd after allocation.
Kernel Build [16cpu, 32GB, ia64] - average of 10 runs:
w/o patch w/ refcount patch
Avg Std Devn Avg Std Devn
Real: 100.59 0.38 100.63 0.43
User: 1209.60 0.37 1209.91 0.31
System: 81.52 0.42 81.64 0.34
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Acked-by: Andi Kleen <ak@suse.de>
Cc: Christoph Lameter <clameter@sgi.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-09-18 22:46:47 -07:00
|
|
|
|
2017-07-06 15:40:03 -07:00
|
|
|
|
mempolicy: rework mempolicy Reference Counting [yet again]
After further discussion with Christoph Lameter, it has become clear that my
earlier attempts to clean up the mempolicy reference counting were a bit of
overkill in some areas, resulting in superflous ref/unref in what are usually
fast paths. In other areas, further inspection reveals that I botched the
unref for interleave policies.
A separate patch, suitable for upstream/stable trees, fixes up the known
errors in the previous attempt to fix reference counting.
This patch reworks the memory policy referencing counting and, one hopes,
simplifies the code. Maybe I'll get it right this time.
See the update to the numa_memory_policy.txt document for a discussion of
memory policy reference counting that motivates this patch.
Summary:
Lookup of mempolicy, based on (vma, address) need only add a reference for
shared policy, and we need only unref the policy when finished for shared
policies. So, this patch backs out all of the unneeded extra reference
counting added by my previous attempt. It then unrefs only shared policies
when we're finished with them, using the mpol_cond_put() [conditional put]
helper function introduced by this patch.
Note that shmem_swapin() calls read_swap_cache_async() with a dummy vma
containing just the policy. read_swap_cache_async() can call alloc_page_vma()
multiple times, so we can't let alloc_page_vma() unref the shared policy in
this case. To avoid this, we make a copy of any non-null shared policy and
remove the MPOL_F_SHARED flag from the copy. This copy occurs before reading
a page [or multiple pages] from swap, so the overhead should not be an issue
here.
I introduced a new static inline function "mpol_cond_copy()" to copy the
shared policy to an on-stack policy and remove the flags that would require a
conditional free. The current implementation of mpol_cond_copy() assumes that
the struct mempolicy contains no pointers to dynamically allocated structures
that must be duplicated or reference counted during copy.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:16 -07:00
|
|
|
|
mm/mempolicy: add MPOL_PREFERRED_MANY for multiple preferred nodes
Patch series "Introduce multi-preference mempolicy", v7.
This patch series introduces the concept of the MPOL_PREFERRED_MANY
mempolicy. This mempolicy mode can be used with either the
set_mempolicy(2) or mbind(2) interfaces. Like the MPOL_PREFERRED
interface, it allows an application to set a preference for nodes which
will fulfil memory allocation requests. Unlike the MPOL_PREFERRED mode,
it takes a set of nodes. Like the MPOL_BIND interface, it works over a
set of nodes. Unlike MPOL_BIND, it will not cause a SIGSEGV or invoke the
OOM killer if those preferred nodes are not available.
Along with these patches are patches for libnuma, numactl, numademo, and
memhog. They still need some polish, but can be found here:
https://gitlab.com/bwidawsk/numactl/-/tree/prefer-many It allows new
usage: `numactl -P 0,3,4`
The goal of the new mode is to enable some use-cases when using tiered memory
usage models which I've lovingly named.
1a. The Hare - The interconnect is fast enough to meet bandwidth and
latency requirements allowing preference to be given to all nodes with
"fast" memory.
1b. The Indiscriminate Hare - An application knows it wants fast
memory (or perhaps slow memory), but doesn't care which node it runs
on. The application can prefer a set of nodes and then xpu bind to
the local node (cpu, accelerator, etc). This reverses the nodes are
chosen today where the kernel attempts to use local memory to the CPU
whenever possible. This will attempt to use the local accelerator to
the memory.
2. The Tortoise - The administrator (or the application itself) is
aware it only needs slow memory, and so can prefer that.
Much of this is almost achievable with the bind interface, but the bind
interface suffers from an inability to fallback to another set of nodes if
binding fails to all nodes in the nodemask.
Like MPOL_BIND a nodemask is given. Inherently this removes ordering from the
preference.
> /* Set first two nodes as preferred in an 8 node system. */
> const unsigned long nodes = 0x3
> set_mempolicy(MPOL_PREFER_MANY, &nodes, 8);
> /* Mimic interleave policy, but have fallback *.
> const unsigned long nodes = 0xaa
> set_mempolicy(MPOL_PREFER_MANY, &nodes, 8);
Some internal discussion took place around the interface. There are two
alternatives which we have discussed, plus one I stuck in:
1. Ordered list of nodes. Currently it's believed that the added
complexity is nod needed for expected usecases.
2. A flag for bind to allow falling back to other nodes. This
confuses the notion of binding and is less flexible than the current
solution.
3. Create flags or new modes that helps with some ordering. This
offers both a friendlier API as well as a solution for more customized
usage. It's unknown if it's worth the complexity to support this.
Here is sample code for how this might work:
> // Prefer specific nodes for some something wacky
> set_mempolicy(MPOL_PREFER_MANY, 0x17c, 1024);
>
> // Default
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_SOCKET, NULL, 0);
> // which is the same as
> set_mempolicy(MPOL_DEFAULT, NULL, 0);
>
> // The Hare
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE, NULL, 0);
>
> // The Tortoise
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE_REV, NULL, 0);
>
> // Prefer the fast memory of the first two sockets
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE, -1, 2);
>
This patch (of 5):
The NUMA APIs currently allow passing in a "preferred node" as a single
bit set in a nodemask. If more than one bit it set, bits after the first
are ignored.
This single node is generally OK for location-based NUMA where memory
being allocated will eventually be operated on by a single CPU. However,
in systems with multiple memory types, folks want to target a *type* of
memory instead of a location. For instance, someone might want some
high-bandwidth memory but do not care about the CPU next to which it is
allocated. Or, they want a cheap, high capacity allocation and want to
target all NUMA nodes which have persistent memory in volatile mode. In
both of these cases, the application wants to target a *set* of nodes, but
does not want strict MPOL_BIND behavior as that could lead to OOM killer
or SIGSEGV.
So add MPOL_PREFERRED_MANY policy to support the multiple preferred nodes
requirement. This is not a pie-in-the-sky dream for an API. This was a
response to a specific ask of more than one group at Intel. Specifically:
1. There are existing libraries that target memory types such as
https://github.com/memkind/memkind. These are known to suffer from
SIGSEGV's when memory is low on targeted memory "kinds" that span more
than one node. The MCDRAM on a Xeon Phi in "Cluster on Die" mode is an
example of this.
2. Volatile-use persistent memory users want to have a memory policy
which is targeted at either "cheap and slow" (PMEM) or "expensive and
fast" (DRAM). However, they do not want to experience allocation
failures when the targeted type is unavailable.
3. Allocate-then-run. Generally, we let the process scheduler decide
on which physical CPU to run a task. That location provides a default
allocation policy, and memory availability is not generally considered
when placing tasks. For situations where memory is valuable and
constrained, some users want to allocate memory first, *then* allocate
close compute resources to the allocation. This is the reverse of the
normal (CPU) model. Accelerators such as GPUs that operate on
core-mm-managed memory are interested in this model.
A check is added in sanitize_mpol_flags() to not permit 'prefer_many'
policy to be used for now, and will be removed in later patch after all
implementations for 'prefer_many' are ready, as suggested by Michal Hocko.
[mhocko@kernel.org: suggest to refine policy_node/policy_nodemask handling]
Link: https://lkml.kernel.org/r/1627970362-61305-1-git-send-email-feng.tang@intel.com
Link: https://lore.kernel.org/r/20200630212517.308045-4-ben.widawsky@intel.com
Link: https://lkml.kernel.org/r/1627970362-61305-2-git-send-email-feng.tang@intel.com
Co-developed-by: Ben Widawsky <ben.widawsky@intel.com>
Signed-off-by: Ben Widawsky <ben.widawsky@intel.com>
Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com>
Signed-off-by: Feng Tang <feng.tang@intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Huang Ying <ying.huang@intel.com>b
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-09-02 15:00:06 -07:00
|
|
|
|
|
|
|
|
|
Fix NUMA Memory Policy Reference Counting
This patch proposes fixes to the reference counting of memory policy in the
page allocation paths and in show_numa_map(). Extracted from my "Memory
Policy Cleanups and Enhancements" series as stand-alone.
Shared policy lookup [shmem] has always added a reference to the policy,
but this was never unrefed after page allocation or after formatting the
numa map data.
Default system policy should not require additional ref counting, nor
should the current task's task policy. However, show_numa_map() calls
get_vma_policy() to examine what may be [likely is] another task's policy.
The latter case needs protection against freeing of the policy.
This patch adds a reference count to a mempolicy returned by
get_vma_policy() when the policy is a vma policy or another task's
mempolicy. Again, shared policy is already reference counted on lookup. A
matching "unref" [__mpol_free()] is performed in alloc_page_vma() for
shared and vma policies, and in show_numa_map() for shared and another
task's mempolicy. We can call __mpol_free() directly, saving an admittedly
inexpensive inline NULL test, because we know we have a non-NULL policy.
Handling policy ref counts for hugepages is a bit trickier.
huge_zonelist() returns a zone list that might come from a shared or vma
'BIND policy. In this case, we should hold the reference until after the
huge page allocation in dequeue_hugepage(). The patch modifies
huge_zonelist() to return a pointer to the mempolicy if it needs to be
unref'd after allocation.
Kernel Build [16cpu, 32GB, ia64] - average of 10 runs:
w/o patch w/ refcount patch
Avg Std Devn Avg Std Devn
Real: 100.59 0.38 100.63 0.43
User: 1209.60 0.37 1209.91 0.31
System: 81.52 0.42 81.64 0.34
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Acked-by: Andi Kleen <ak@suse.de>
Cc: Christoph Lameter <clameter@sgi.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-09-18 22:46:47 -07:00
|
|
|
|
2017-07-06 15:40:03 -07:00
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
2006-01-06 00:10:46 -08:00
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
2017-07-06 15:40:03 -07:00
|
|
|
|
2006-01-06 00:10:46 -08:00
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-07-06 15:40:03 -07:00
|
|
|
|
2006-01-06 00:10:46 -08:00
|
|
|
|
hugetlb: derive huge pages nodes allowed from task mempolicy
This patch derives a "nodes_allowed" node mask from the numa mempolicy of
the task modifying the number of persistent huge pages to control the
allocation, freeing and adjusting of surplus huge pages when the pool page
count is modified via the new sysctl or sysfs attribute
"nr_hugepages_mempolicy". The nodes_allowed mask is derived as follows:
* For "default" [NULL] task mempolicy, a NULL nodemask_t pointer
is produced. This will cause the hugetlb subsystem to use
node_online_map as the "nodes_allowed". This preserves the
behavior before this patch.
* For "preferred" mempolicy, including explicit local allocation,
a nodemask with the single preferred node will be produced.
"local" policy will NOT track any internode migrations of the
task adjusting nr_hugepages.
* For "bind" and "interleave" policy, the mempolicy's nodemask
will be used.
* Other than to inform the construction of the nodes_allowed node
mask, the actual mempolicy mode is ignored. That is, all modes
behave like interleave over the resulting nodes_allowed mask
with no "fallback".
See the updated documentation [next patch] for more information
about the implications of this patch.
Examples:
Starting with:
Node 0 HugePages_Total: 0
Node 1 HugePages_Total: 0
Node 2 HugePages_Total: 0
Node 3 HugePages_Total: 0
Default behavior [with or without this patch] balances persistent
hugepage allocation across nodes [with sufficient contiguous memory]:
sysctl vm.nr_hugepages[_mempolicy]=32
yields:
Node 0 HugePages_Total: 8
Node 1 HugePages_Total: 8
Node 2 HugePages_Total: 8
Node 3 HugePages_Total: 8
Of course, we only have nr_hugepages_mempolicy with the patch,
but with default mempolicy, nr_hugepages_mempolicy behaves the
same as nr_hugepages.
Applying mempolicy--e.g., with numactl [using '-m' a.k.a.
'--membind' because it allows multiple nodes to be specified
and it's easy to type]--we can allocate huge pages on
individual nodes or sets of nodes. So, starting from the
condition above, with 8 huge pages per node, add 8 more to
node 2 using:
numactl -m 2 sysctl vm.nr_hugepages_mempolicy=40
This yields:
Node 0 HugePages_Total: 8
Node 1 HugePages_Total: 8
Node 2 HugePages_Total: 16
Node 3 HugePages_Total: 8
The incremental 8 huge pages were restricted to node 2 by the
specified mempolicy.
Similarly, we can use mempolicy to free persistent huge pages
from specified nodes:
numactl -m 0,1 sysctl vm.nr_hugepages_mempolicy=32
yields:
Node 0 HugePages_Total: 4
Node 1 HugePages_Total: 4
Node 2 HugePages_Total: 16
Node 3 HugePages_Total: 8
The 8 huge pages freed were balanced over nodes 0 and 1.
[rientjes@google.com: accomodate reworked NODEMASK_ALLOC]
Signed-off-by: David Rientjes <rientjes@google.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Andi Kleen <andi@firstfloor.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Randy Dunlap <randy.dunlap@oracle.com>
Cc: Nishanth Aravamudan <nacc@us.ibm.com>
Cc: Adam Litke <agl@us.ibm.com>
Cc: Andy Whitcroft <apw@canonical.com>
Cc: Eric Whitney <eric.whitney@hp.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-14 17:58:21 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-05-24 14:32:08 -07:00
|
|
|
|
hugetlb: derive huge pages nodes allowed from task mempolicy
This patch derives a "nodes_allowed" node mask from the numa mempolicy of
the task modifying the number of persistent huge pages to control the
allocation, freeing and adjusting of surplus huge pages when the pool page
count is modified via the new sysctl or sysfs attribute
"nr_hugepages_mempolicy". The nodes_allowed mask is derived as follows:
* For "default" [NULL] task mempolicy, a NULL nodemask_t pointer
is produced. This will cause the hugetlb subsystem to use
node_online_map as the "nodes_allowed". This preserves the
behavior before this patch.
* For "preferred" mempolicy, including explicit local allocation,
a nodemask with the single preferred node will be produced.
"local" policy will NOT track any internode migrations of the
task adjusting nr_hugepages.
* For "bind" and "interleave" policy, the mempolicy's nodemask
will be used.
* Other than to inform the construction of the nodes_allowed node
mask, the actual mempolicy mode is ignored. That is, all modes
behave like interleave over the resulting nodes_allowed mask
with no "fallback".
See the updated documentation [next patch] for more information
about the implications of this patch.
Examples:
Starting with:
Node 0 HugePages_Total: 0
Node 1 HugePages_Total: 0
Node 2 HugePages_Total: 0
Node 3 HugePages_Total: 0
Default behavior [with or without this patch] balances persistent
hugepage allocation across nodes [with sufficient contiguous memory]:
sysctl vm.nr_hugepages[_mempolicy]=32
yields:
Node 0 HugePages_Total: 8
Node 1 HugePages_Total: 8
Node 2 HugePages_Total: 8
Node 3 HugePages_Total: 8
Of course, we only have nr_hugepages_mempolicy with the patch,
but with default mempolicy, nr_hugepages_mempolicy behaves the
same as nr_hugepages.
Applying mempolicy--e.g., with numactl [using '-m' a.k.a.
'--membind' because it allows multiple nodes to be specified
and it's easy to type]--we can allocate huge pages on
individual nodes or sets of nodes. So, starting from the
condition above, with 8 huge pages per node, add 8 more to
node 2 using:
numactl -m 2 sysctl vm.nr_hugepages_mempolicy=40
This yields:
Node 0 HugePages_Total: 8
Node 1 HugePages_Total: 8
Node 2 HugePages_Total: 16
Node 3 HugePages_Total: 8
The incremental 8 huge pages were restricted to node 2 by the
specified mempolicy.
Similarly, we can use mempolicy to free persistent huge pages
from specified nodes:
numactl -m 0,1 sysctl vm.nr_hugepages_mempolicy=32
yields:
Node 0 HugePages_Total: 4
Node 1 HugePages_Total: 4
Node 2 HugePages_Total: 16
Node 3 HugePages_Total: 8
The 8 huge pages freed were balanced over nodes 0 and 1.
[rientjes@google.com: accomodate reworked NODEMASK_ALLOC]
Signed-off-by: David Rientjes <rientjes@google.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Andi Kleen <andi@firstfloor.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Randy Dunlap <randy.dunlap@oracle.com>
Cc: Nishanth Aravamudan <nacc@us.ibm.com>
Cc: Adam Litke <agl@us.ibm.com>
Cc: Andy Whitcroft <apw@canonical.com>
Cc: Eric Whitney <eric.whitney@hp.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-14 17:58:21 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: add MPOL_PREFERRED_MANY for multiple preferred nodes
Patch series "Introduce multi-preference mempolicy", v7.
This patch series introduces the concept of the MPOL_PREFERRED_MANY
mempolicy. This mempolicy mode can be used with either the
set_mempolicy(2) or mbind(2) interfaces. Like the MPOL_PREFERRED
interface, it allows an application to set a preference for nodes which
will fulfil memory allocation requests. Unlike the MPOL_PREFERRED mode,
it takes a set of nodes. Like the MPOL_BIND interface, it works over a
set of nodes. Unlike MPOL_BIND, it will not cause a SIGSEGV or invoke the
OOM killer if those preferred nodes are not available.
Along with these patches are patches for libnuma, numactl, numademo, and
memhog. They still need some polish, but can be found here:
https://gitlab.com/bwidawsk/numactl/-/tree/prefer-many It allows new
usage: `numactl -P 0,3,4`
The goal of the new mode is to enable some use-cases when using tiered memory
usage models which I've lovingly named.
1a. The Hare - The interconnect is fast enough to meet bandwidth and
latency requirements allowing preference to be given to all nodes with
"fast" memory.
1b. The Indiscriminate Hare - An application knows it wants fast
memory (or perhaps slow memory), but doesn't care which node it runs
on. The application can prefer a set of nodes and then xpu bind to
the local node (cpu, accelerator, etc). This reverses the nodes are
chosen today where the kernel attempts to use local memory to the CPU
whenever possible. This will attempt to use the local accelerator to
the memory.
2. The Tortoise - The administrator (or the application itself) is
aware it only needs slow memory, and so can prefer that.
Much of this is almost achievable with the bind interface, but the bind
interface suffers from an inability to fallback to another set of nodes if
binding fails to all nodes in the nodemask.
Like MPOL_BIND a nodemask is given. Inherently this removes ordering from the
preference.
> /* Set first two nodes as preferred in an 8 node system. */
> const unsigned long nodes = 0x3
> set_mempolicy(MPOL_PREFER_MANY, &nodes, 8);
> /* Mimic interleave policy, but have fallback *.
> const unsigned long nodes = 0xaa
> set_mempolicy(MPOL_PREFER_MANY, &nodes, 8);
Some internal discussion took place around the interface. There are two
alternatives which we have discussed, plus one I stuck in:
1. Ordered list of nodes. Currently it's believed that the added
complexity is nod needed for expected usecases.
2. A flag for bind to allow falling back to other nodes. This
confuses the notion of binding and is less flexible than the current
solution.
3. Create flags or new modes that helps with some ordering. This
offers both a friendlier API as well as a solution for more customized
usage. It's unknown if it's worth the complexity to support this.
Here is sample code for how this might work:
> // Prefer specific nodes for some something wacky
> set_mempolicy(MPOL_PREFER_MANY, 0x17c, 1024);
>
> // Default
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_SOCKET, NULL, 0);
> // which is the same as
> set_mempolicy(MPOL_DEFAULT, NULL, 0);
>
> // The Hare
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE, NULL, 0);
>
> // The Tortoise
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE_REV, NULL, 0);
>
> // Prefer the fast memory of the first two sockets
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE, -1, 2);
>
This patch (of 5):
The NUMA APIs currently allow passing in a "preferred node" as a single
bit set in a nodemask. If more than one bit it set, bits after the first
are ignored.
This single node is generally OK for location-based NUMA where memory
being allocated will eventually be operated on by a single CPU. However,
in systems with multiple memory types, folks want to target a *type* of
memory instead of a location. For instance, someone might want some
high-bandwidth memory but do not care about the CPU next to which it is
allocated. Or, they want a cheap, high capacity allocation and want to
target all NUMA nodes which have persistent memory in volatile mode. In
both of these cases, the application wants to target a *set* of nodes, but
does not want strict MPOL_BIND behavior as that could lead to OOM killer
or SIGSEGV.
So add MPOL_PREFERRED_MANY policy to support the multiple preferred nodes
requirement. This is not a pie-in-the-sky dream for an API. This was a
response to a specific ask of more than one group at Intel. Specifically:
1. There are existing libraries that target memory types such as
https://github.com/memkind/memkind. These are known to suffer from
SIGSEGV's when memory is low on targeted memory "kinds" that span more
than one node. The MCDRAM on a Xeon Phi in "Cluster on Die" mode is an
example of this.
2. Volatile-use persistent memory users want to have a memory policy
which is targeted at either "cheap and slow" (PMEM) or "expensive and
fast" (DRAM). However, they do not want to experience allocation
failures when the targeted type is unavailable.
3. Allocate-then-run. Generally, we let the process scheduler decide
on which physical CPU to run a task. That location provides a default
allocation policy, and memory availability is not generally considered
when placing tasks. For situations where memory is valuable and
constrained, some users want to allocate memory first, *then* allocate
close compute resources to the allocation. This is the reverse of the
normal (CPU) model. Accelerators such as GPUs that operate on
core-mm-managed memory are interested in this model.
A check is added in sanitize_mpol_flags() to not permit 'prefer_many'
policy to be used for now, and will be removed in later patch after all
implementations for 'prefer_many' are ready, as suggested by Michal Hocko.
[mhocko@kernel.org: suggest to refine policy_node/policy_nodemask handling]
Link: https://lkml.kernel.org/r/1627970362-61305-1-git-send-email-feng.tang@intel.com
Link: https://lore.kernel.org/r/20200630212517.308045-4-ben.widawsky@intel.com
Link: https://lkml.kernel.org/r/1627970362-61305-2-git-send-email-feng.tang@intel.com
Co-developed-by: Ben Widawsky <ben.widawsky@intel.com>
Signed-off-by: Ben Widawsky <ben.widawsky@intel.com>
Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com>
Signed-off-by: Feng Tang <feng.tang@intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Huang Ying <ying.huang@intel.com>b
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-09-02 15:00:06 -07:00
|
|
|
|
hugetlb: derive huge pages nodes allowed from task mempolicy
This patch derives a "nodes_allowed" node mask from the numa mempolicy of
the task modifying the number of persistent huge pages to control the
allocation, freeing and adjusting of surplus huge pages when the pool page
count is modified via the new sysctl or sysfs attribute
"nr_hugepages_mempolicy". The nodes_allowed mask is derived as follows:
* For "default" [NULL] task mempolicy, a NULL nodemask_t pointer
is produced. This will cause the hugetlb subsystem to use
node_online_map as the "nodes_allowed". This preserves the
behavior before this patch.
* For "preferred" mempolicy, including explicit local allocation,
a nodemask with the single preferred node will be produced.
"local" policy will NOT track any internode migrations of the
task adjusting nr_hugepages.
* For "bind" and "interleave" policy, the mempolicy's nodemask
will be used.
* Other than to inform the construction of the nodes_allowed node
mask, the actual mempolicy mode is ignored. That is, all modes
behave like interleave over the resulting nodes_allowed mask
with no "fallback".
See the updated documentation [next patch] for more information
about the implications of this patch.
Examples:
Starting with:
Node 0 HugePages_Total: 0
Node 1 HugePages_Total: 0
Node 2 HugePages_Total: 0
Node 3 HugePages_Total: 0
Default behavior [with or without this patch] balances persistent
hugepage allocation across nodes [with sufficient contiguous memory]:
sysctl vm.nr_hugepages[_mempolicy]=32
yields:
Node 0 HugePages_Total: 8
Node 1 HugePages_Total: 8
Node 2 HugePages_Total: 8
Node 3 HugePages_Total: 8
Of course, we only have nr_hugepages_mempolicy with the patch,
but with default mempolicy, nr_hugepages_mempolicy behaves the
same as nr_hugepages.
Applying mempolicy--e.g., with numactl [using '-m' a.k.a.
'--membind' because it allows multiple nodes to be specified
and it's easy to type]--we can allocate huge pages on
individual nodes or sets of nodes. So, starting from the
condition above, with 8 huge pages per node, add 8 more to
node 2 using:
numactl -m 2 sysctl vm.nr_hugepages_mempolicy=40
This yields:
Node 0 HugePages_Total: 8
Node 1 HugePages_Total: 8
Node 2 HugePages_Total: 16
Node 3 HugePages_Total: 8
The incremental 8 huge pages were restricted to node 2 by the
specified mempolicy.
Similarly, we can use mempolicy to free persistent huge pages
from specified nodes:
numactl -m 0,1 sysctl vm.nr_hugepages_mempolicy=32
yields:
Node 0 HugePages_Total: 4
Node 1 HugePages_Total: 4
Node 2 HugePages_Total: 16
Node 3 HugePages_Total: 8
The 8 huge pages freed were balanced over nodes 0 and 1.
[rientjes@google.com: accomodate reworked NODEMASK_ALLOC]
Signed-off-by: David Rientjes <rientjes@google.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Andi Kleen <andi@firstfloor.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Randy Dunlap <randy.dunlap@oracle.com>
Cc: Nishanth Aravamudan <nacc@us.ibm.com>
Cc: Adam Litke <agl@us.ibm.com>
Cc: Andy Whitcroft <apw@canonical.com>
Cc: Eric Whitney <eric.whitney@hp.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-14 17:58:21 -08:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
2021-06-30 18:51:10 -07:00
|
|
|
|
mm/mempolicy: don't handle MPOL_LOCAL like a fake MPOL_PREFERRED policy
MPOL_LOCAL policy has been setup as a real policy, but it is still handled
like a faked POL_PREFERRED policy with one internal MPOL_F_LOCAL flag bit
set, and there are many places having to judge the real 'prefer' or the
'local' policy, which are quite confusing.
In current code, there are 4 cases that MPOL_LOCAL are used:
1. user specifies 'local' policy
2. user specifies 'prefer' policy, but with empty nodemask
3. system 'default' policy is used
4. 'prefer' policy + valid 'preferred' node with MPOL_F_STATIC_NODES
flag set, and when it is 'rebind' to a nodemask which doesn't contains
the 'preferred' node, it will perform as 'local' policy
So make 'local' a real policy instead of a fake 'prefer' one, and kill
MPOL_F_LOCAL bit, which can greatly reduce the confusion for code reading.
For case 4, the logic of mpol_rebind_preferred() is confusing, as Michal
Hocko pointed out:
: I do believe that rebinding preferred policy is just bogus and it should
: be dropped altogether on the ground that a preference is a mere hint from
: userspace where to start the allocation. Unless I am missing something
: cpusets will be always authoritative for the final placement. The
: preferred node just acts as a starting point and it should be really
: preserved when cpusets changes. Otherwise we have a very subtle behavior
: corner cases.
So dump all the tricky transformation between 'prefer' and 'local', and
just record the new nodemask of rebinding.
[feng.tang@intel.com: fix a problem in mpol_set_nodemask(), per Michal Hocko]
Link: https://lkml.kernel.org/r/1622560492-1294-3-git-send-email-feng.tang@intel.com
[feng.tang@intel.com: refine code and comments of mpol_set_nodemask(), per Michal]
Link: https://lkml.kernel.org/r/20210603081807.GE56979@shbuild999.sh.intel.com
Link: https://lkml.kernel.org/r/1622469956-82897-3-git-send-email-feng.tang@intel.com
Signed-off-by: Feng Tang <feng.tang@intel.com>
Suggested-by: Michal Hocko <mhocko@suse.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ben Widawsky <ben.widawsky@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-30 18:51:00 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-30 18:51:10 -07:00
|
|
|
|
hugetlb: derive huge pages nodes allowed from task mempolicy
This patch derives a "nodes_allowed" node mask from the numa mempolicy of
the task modifying the number of persistent huge pages to control the
allocation, freeing and adjusting of surplus huge pages when the pool page
count is modified via the new sysctl or sysfs attribute
"nr_hugepages_mempolicy". The nodes_allowed mask is derived as follows:
* For "default" [NULL] task mempolicy, a NULL nodemask_t pointer
is produced. This will cause the hugetlb subsystem to use
node_online_map as the "nodes_allowed". This preserves the
behavior before this patch.
* For "preferred" mempolicy, including explicit local allocation,
a nodemask with the single preferred node will be produced.
"local" policy will NOT track any internode migrations of the
task adjusting nr_hugepages.
* For "bind" and "interleave" policy, the mempolicy's nodemask
will be used.
* Other than to inform the construction of the nodes_allowed node
mask, the actual mempolicy mode is ignored. That is, all modes
behave like interleave over the resulting nodes_allowed mask
with no "fallback".
See the updated documentation [next patch] for more information
about the implications of this patch.
Examples:
Starting with:
Node 0 HugePages_Total: 0
Node 1 HugePages_Total: 0
Node 2 HugePages_Total: 0
Node 3 HugePages_Total: 0
Default behavior [with or without this patch] balances persistent
hugepage allocation across nodes [with sufficient contiguous memory]:
sysctl vm.nr_hugepages[_mempolicy]=32
yields:
Node 0 HugePages_Total: 8
Node 1 HugePages_Total: 8
Node 2 HugePages_Total: 8
Node 3 HugePages_Total: 8
Of course, we only have nr_hugepages_mempolicy with the patch,
but with default mempolicy, nr_hugepages_mempolicy behaves the
same as nr_hugepages.
Applying mempolicy--e.g., with numactl [using '-m' a.k.a.
'--membind' because it allows multiple nodes to be specified
and it's easy to type]--we can allocate huge pages on
individual nodes or sets of nodes. So, starting from the
condition above, with 8 huge pages per node, add 8 more to
node 2 using:
numactl -m 2 sysctl vm.nr_hugepages_mempolicy=40
This yields:
Node 0 HugePages_Total: 8
Node 1 HugePages_Total: 8
Node 2 HugePages_Total: 16
Node 3 HugePages_Total: 8
The incremental 8 huge pages were restricted to node 2 by the
specified mempolicy.
Similarly, we can use mempolicy to free persistent huge pages
from specified nodes:
numactl -m 0,1 sysctl vm.nr_hugepages_mempolicy=32
yields:
Node 0 HugePages_Total: 4
Node 1 HugePages_Total: 4
Node 2 HugePages_Total: 16
Node 3 HugePages_Total: 8
The 8 huge pages freed were balanced over nodes 0 and 1.
[rientjes@google.com: accomodate reworked NODEMASK_ALLOC]
Signed-off-by: David Rientjes <rientjes@google.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Andi Kleen <andi@firstfloor.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Randy Dunlap <randy.dunlap@oracle.com>
Cc: Nishanth Aravamudan <nacc@us.ibm.com>
Cc: Adam Litke <agl@us.ibm.com>
Cc: Andy Whitcroft <apw@canonical.com>
Cc: Eric Whitney <eric.whitney@hp.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-14 17:58:21 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-05-24 14:32:08 -07:00
|
|
|
|
hugetlb: derive huge pages nodes allowed from task mempolicy
This patch derives a "nodes_allowed" node mask from the numa mempolicy of
the task modifying the number of persistent huge pages to control the
allocation, freeing and adjusting of surplus huge pages when the pool page
count is modified via the new sysctl or sysfs attribute
"nr_hugepages_mempolicy". The nodes_allowed mask is derived as follows:
* For "default" [NULL] task mempolicy, a NULL nodemask_t pointer
is produced. This will cause the hugetlb subsystem to use
node_online_map as the "nodes_allowed". This preserves the
behavior before this patch.
* For "preferred" mempolicy, including explicit local allocation,
a nodemask with the single preferred node will be produced.
"local" policy will NOT track any internode migrations of the
task adjusting nr_hugepages.
* For "bind" and "interleave" policy, the mempolicy's nodemask
will be used.
* Other than to inform the construction of the nodes_allowed node
mask, the actual mempolicy mode is ignored. That is, all modes
behave like interleave over the resulting nodes_allowed mask
with no "fallback".
See the updated documentation [next patch] for more information
about the implications of this patch.
Examples:
Starting with:
Node 0 HugePages_Total: 0
Node 1 HugePages_Total: 0
Node 2 HugePages_Total: 0
Node 3 HugePages_Total: 0
Default behavior [with or without this patch] balances persistent
hugepage allocation across nodes [with sufficient contiguous memory]:
sysctl vm.nr_hugepages[_mempolicy]=32
yields:
Node 0 HugePages_Total: 8
Node 1 HugePages_Total: 8
Node 2 HugePages_Total: 8
Node 3 HugePages_Total: 8
Of course, we only have nr_hugepages_mempolicy with the patch,
but with default mempolicy, nr_hugepages_mempolicy behaves the
same as nr_hugepages.
Applying mempolicy--e.g., with numactl [using '-m' a.k.a.
'--membind' because it allows multiple nodes to be specified
and it's easy to type]--we can allocate huge pages on
individual nodes or sets of nodes. So, starting from the
condition above, with 8 huge pages per node, add 8 more to
node 2 using:
numactl -m 2 sysctl vm.nr_hugepages_mempolicy=40
This yields:
Node 0 HugePages_Total: 8
Node 1 HugePages_Total: 8
Node 2 HugePages_Total: 16
Node 3 HugePages_Total: 8
The incremental 8 huge pages were restricted to node 2 by the
specified mempolicy.
Similarly, we can use mempolicy to free persistent huge pages
from specified nodes:
numactl -m 0,1 sysctl vm.nr_hugepages_mempolicy=32
yields:
Node 0 HugePages_Total: 4
Node 1 HugePages_Total: 4
Node 2 HugePages_Total: 16
Node 3 HugePages_Total: 8
The 8 huge pages freed were balanced over nodes 0 and 1.
[rientjes@google.com: accomodate reworked NODEMASK_ALLOC]
Signed-off-by: David Rientjes <rientjes@google.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Andi Kleen <andi@firstfloor.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Randy Dunlap <randy.dunlap@oracle.com>
Cc: Nishanth Aravamudan <nacc@us.ibm.com>
Cc: Adam Litke <agl@us.ibm.com>
Cc: Andy Whitcroft <apw@canonical.com>
Cc: Eric Whitney <eric.whitney@hp.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-14 17:58:21 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-02-03 21:51:14 +01:00
|
|
|
|
2006-01-06 00:10:46 -08:00
|
|
|
|
2010-08-09 17:18:52 -07:00
|
|
|
|
2021-06-30 18:50:56 -07:00
|
|
|
|
2010-08-09 17:18:52 -07:00
|
|
|
|
2021-06-30 18:50:56 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-08-09 17:18:52 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-30 18:50:56 -07:00
|
|
|
|
2010-08-09 17:18:52 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-30 18:50:56 -07:00
|
|
|
|
2010-08-09 17:18:52 -07:00
|
|
|
|
|
|
|
|
|
2021-06-30 18:50:56 -07:00
|
|
|
|
2021-06-30 18:51:10 -07:00
|
|
|
|
2010-08-09 17:18:52 -07:00
|
|
|
|
2021-06-30 18:50:56 -07:00
|
|
|
|
2010-08-09 17:18:52 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-09-02 15:00:10 -07:00
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
2021-09-02 15:00:10 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-11-25 21:01:46 +00:00
|
|
|
|
2021-09-02 15:00:10 -07:00
|
|
|
|
2024-11-25 21:01:46 +00:00
|
|
|
|
2021-09-02 15:00:10 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
2021-04-29 23:01:24 -07:00
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2024-11-25 21:01:34 +00:00
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
|
|
|
|
|
cpuset: mm: reduce large amounts of memory barrier related damage v3
Commit c0ff7453bb5c ("cpuset,mm: fix no node to alloc memory when
changing cpuset's mems") wins a super prize for the largest number of
memory barriers entered into fast paths for one commit.
[get|put]_mems_allowed is incredibly heavy with pairs of full memory
barriers inserted into a number of hot paths. This was detected while
investigating at large page allocator slowdown introduced some time
after 2.6.32. The largest portion of this overhead was shown by
oprofile to be at an mfence introduced by this commit into the page
allocator hot path.
For extra style points, the commit introduced the use of yield() in an
implementation of what looks like a spinning mutex.
This patch replaces the full memory barriers on both read and write
sides with a sequence counter with just read barriers on the fast path
side. This is much cheaper on some architectures, including x86. The
main bulk of the patch is the retry logic if the nodemask changes in a
manner that can cause a false failure.
While updating the nodemask, a check is made to see if a false failure
is a risk. If it is, the sequence number gets bumped and parallel
allocators will briefly stall while the nodemask update takes place.
In a page fault test microbenchmark, oprofile samples from
__alloc_pages_nodemask went from 4.53% of all samples to 1.15%. The
actual results were
3.3.0-rc3 3.3.0-rc3
rc3-vanilla nobarrier-v2r1
Clients 1 UserTime 0.07 ( 0.00%) 0.08 (-14.19%)
Clients 2 UserTime 0.07 ( 0.00%) 0.07 ( 2.72%)
Clients 4 UserTime 0.08 ( 0.00%) 0.07 ( 3.29%)
Clients 1 SysTime 0.70 ( 0.00%) 0.65 ( 6.65%)
Clients 2 SysTime 0.85 ( 0.00%) 0.82 ( 3.65%)
Clients 4 SysTime 1.41 ( 0.00%) 1.41 ( 0.32%)
Clients 1 WallTime 0.77 ( 0.00%) 0.74 ( 4.19%)
Clients 2 WallTime 0.47 ( 0.00%) 0.45 ( 3.73%)
Clients 4 WallTime 0.38 ( 0.00%) 0.37 ( 1.58%)
Clients 1 Flt/sec/cpu 497620.28 ( 0.00%) 520294.53 ( 4.56%)
Clients 2 Flt/sec/cpu 414639.05 ( 0.00%) 429882.01 ( 3.68%)
Clients 4 Flt/sec/cpu 257959.16 ( 0.00%) 258761.48 ( 0.31%)
Clients 1 Flt/sec 495161.39 ( 0.00%) 517292.87 ( 4.47%)
Clients 2 Flt/sec 820325.95 ( 0.00%) 850289.77 ( 3.65%)
Clients 4 Flt/sec 1020068.93 ( 0.00%) 1022674.06 ( 0.26%)
MMTests Statistics: duration
Sys Time Running Test (seconds) 135.68 132.17
User+Sys Time Running Test (seconds) 164.2 160.13
Total Elapsed Time (seconds) 123.46 120.87
The overall improvement is small but the System CPU time is much
improved and roughly in correlation to what oprofile reported (these
performance figures are without profiling so skew is expected). The
actual number of page faults is noticeably improved.
For benchmarks like kernel builds, the overall benefit is marginal but
the system CPU time is slightly reduced.
To test the actual bug the commit fixed I opened two terminals. The
first ran within a cpuset and continually ran a small program that
faulted 100M of anonymous data. In a second window, the nodemask of the
cpuset was continually randomised in a loop.
Without the commit, the program would fail every so often (usually
within 10 seconds) and obviously with the commit everything worked fine.
With this patch applied, it also worked fine so the fix should be
functionally equivalent.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Miao Xie <miaox@cn.fujitsu.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-03-21 16:34:11 -07:00
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
|
|
|
|
|
2019-09-04 12:54:20 -07:00
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
|
|
|
|
|
2026-03-25 05:40:19 -06:00
|
|
|
|
2019-09-04 12:54:20 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: add MPOL_PREFERRED_MANY for multiple preferred nodes
Patch series "Introduce multi-preference mempolicy", v7.
This patch series introduces the concept of the MPOL_PREFERRED_MANY
mempolicy. This mempolicy mode can be used with either the
set_mempolicy(2) or mbind(2) interfaces. Like the MPOL_PREFERRED
interface, it allows an application to set a preference for nodes which
will fulfil memory allocation requests. Unlike the MPOL_PREFERRED mode,
it takes a set of nodes. Like the MPOL_BIND interface, it works over a
set of nodes. Unlike MPOL_BIND, it will not cause a SIGSEGV or invoke the
OOM killer if those preferred nodes are not available.
Along with these patches are patches for libnuma, numactl, numademo, and
memhog. They still need some polish, but can be found here:
https://gitlab.com/bwidawsk/numactl/-/tree/prefer-many It allows new
usage: `numactl -P 0,3,4`
The goal of the new mode is to enable some use-cases when using tiered memory
usage models which I've lovingly named.
1a. The Hare - The interconnect is fast enough to meet bandwidth and
latency requirements allowing preference to be given to all nodes with
"fast" memory.
1b. The Indiscriminate Hare - An application knows it wants fast
memory (or perhaps slow memory), but doesn't care which node it runs
on. The application can prefer a set of nodes and then xpu bind to
the local node (cpu, accelerator, etc). This reverses the nodes are
chosen today where the kernel attempts to use local memory to the CPU
whenever possible. This will attempt to use the local accelerator to
the memory.
2. The Tortoise - The administrator (or the application itself) is
aware it only needs slow memory, and so can prefer that.
Much of this is almost achievable with the bind interface, but the bind
interface suffers from an inability to fallback to another set of nodes if
binding fails to all nodes in the nodemask.
Like MPOL_BIND a nodemask is given. Inherently this removes ordering from the
preference.
> /* Set first two nodes as preferred in an 8 node system. */
> const unsigned long nodes = 0x3
> set_mempolicy(MPOL_PREFER_MANY, &nodes, 8);
> /* Mimic interleave policy, but have fallback *.
> const unsigned long nodes = 0xaa
> set_mempolicy(MPOL_PREFER_MANY, &nodes, 8);
Some internal discussion took place around the interface. There are two
alternatives which we have discussed, plus one I stuck in:
1. Ordered list of nodes. Currently it's believed that the added
complexity is nod needed for expected usecases.
2. A flag for bind to allow falling back to other nodes. This
confuses the notion of binding and is less flexible than the current
solution.
3. Create flags or new modes that helps with some ordering. This
offers both a friendlier API as well as a solution for more customized
usage. It's unknown if it's worth the complexity to support this.
Here is sample code for how this might work:
> // Prefer specific nodes for some something wacky
> set_mempolicy(MPOL_PREFER_MANY, 0x17c, 1024);
>
> // Default
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_SOCKET, NULL, 0);
> // which is the same as
> set_mempolicy(MPOL_DEFAULT, NULL, 0);
>
> // The Hare
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE, NULL, 0);
>
> // The Tortoise
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE_REV, NULL, 0);
>
> // Prefer the fast memory of the first two sockets
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE, -1, 2);
>
This patch (of 5):
The NUMA APIs currently allow passing in a "preferred node" as a single
bit set in a nodemask. If more than one bit it set, bits after the first
are ignored.
This single node is generally OK for location-based NUMA where memory
being allocated will eventually be operated on by a single CPU. However,
in systems with multiple memory types, folks want to target a *type* of
memory instead of a location. For instance, someone might want some
high-bandwidth memory but do not care about the CPU next to which it is
allocated. Or, they want a cheap, high capacity allocation and want to
target all NUMA nodes which have persistent memory in volatile mode. In
both of these cases, the application wants to target a *set* of nodes, but
does not want strict MPOL_BIND behavior as that could lead to OOM killer
or SIGSEGV.
So add MPOL_PREFERRED_MANY policy to support the multiple preferred nodes
requirement. This is not a pie-in-the-sky dream for an API. This was a
response to a specific ask of more than one group at Intel. Specifically:
1. There are existing libraries that target memory types such as
https://github.com/memkind/memkind. These are known to suffer from
SIGSEGV's when memory is low on targeted memory "kinds" that span more
than one node. The MCDRAM on a Xeon Phi in "Cluster on Die" mode is an
example of this.
2. Volatile-use persistent memory users want to have a memory policy
which is targeted at either "cheap and slow" (PMEM) or "expensive and
fast" (DRAM). However, they do not want to experience allocation
failures when the targeted type is unavailable.
3. Allocate-then-run. Generally, we let the process scheduler decide
on which physical CPU to run a task. That location provides a default
allocation policy, and memory availability is not generally considered
when placing tasks. For situations where memory is valuable and
constrained, some users want to allocate memory first, *then* allocate
close compute resources to the allocation. This is the reverse of the
normal (CPU) model. Accelerators such as GPUs that operate on
core-mm-managed memory are interested in this model.
A check is added in sanitize_mpol_flags() to not permit 'prefer_many'
policy to be used for now, and will be removed in later patch after all
implementations for 'prefer_many' are ready, as suggested by Michal Hocko.
[mhocko@kernel.org: suggest to refine policy_node/policy_nodemask handling]
Link: https://lkml.kernel.org/r/1627970362-61305-1-git-send-email-feng.tang@intel.com
Link: https://lore.kernel.org/r/20200630212517.308045-4-ben.widawsky@intel.com
Link: https://lkml.kernel.org/r/1627970362-61305-2-git-send-email-feng.tang@intel.com
Co-developed-by: Ben Widawsky <ben.widawsky@intel.com>
Signed-off-by: Ben Widawsky <ben.widawsky@intel.com>
Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com>
Signed-off-by: Feng Tang <feng.tang@intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Huang Ying <ying.huang@intel.com>b
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-09-02 15:00:06 -07:00
|
|
|
|
2019-09-04 12:54:20 -07:00
|
|
|
|
|
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
mm, thp: tweak reclaim/compaction effort of local-only and all-node allocations
THP page faults now attempt a __GFP_THISNODE allocation first, which
should only compact existing free memory, followed by another attempt
that can allocate from any node using reclaim/compaction effort
specified by global defrag setting and madvise.
This patch makes the following changes to the scheme:
- Before the patch, the first allocation relies on a check for
pageblock order and __GFP_IO to prevent excessive reclaim. This
however affects also the second attempt, which is not limited to
single node.
Instead of that, reuse the existing check for costly order
__GFP_NORETRY allocations, and make sure the first THP attempt uses
__GFP_NORETRY. As a side-effect, all costly order __GFP_NORETRY
allocations will bail out if compaction needs reclaim, while
previously they only bailed out when compaction was deferred due to
previous failures.
This should be still acceptable within the __GFP_NORETRY semantics.
- Before the patch, the second allocation attempt (on all nodes) was
passing __GFP_NORETRY. This is redundant as the check for pageblock
order (discussed above) was stronger. It's also contrary to
madvise(MADV_HUGEPAGE) which means some effort to allocate THP is
requested.
After this patch, the second attempt doesn't pass __GFP_THISNODE nor
__GFP_NORETRY.
To sum up, THP page faults now try the following attempts:
1. local node only THP allocation with no reclaim, just compaction.
2. for madvised VMA's or when synchronous compaction is enabled always - THP
allocation from any node with effort determined by global defrag setting
and VMA madvise
3. fallback to base pages on any node
Link: http://lkml.kernel.org/r/08a3f4dd-c3ce-0009-86c5-9ee51aba8557@suse.cz
Fixes: b39d0ee2632d ("mm, page_alloc: avoid expensive reclaim when compaction may not succeed")
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-01-13 16:29:04 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-11-25 21:01:46 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
|
|
|
|
|
2019-09-04 12:54:25 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, thp: tweak reclaim/compaction effort of local-only and all-node allocations
THP page faults now attempt a __GFP_THISNODE allocation first, which
should only compact existing free memory, followed by another attempt
that can allocate from any node using reclaim/compaction effort
specified by global defrag setting and madvise.
This patch makes the following changes to the scheme:
- Before the patch, the first allocation relies on a check for
pageblock order and __GFP_IO to prevent excessive reclaim. This
however affects also the second attempt, which is not limited to
single node.
Instead of that, reuse the existing check for costly order
__GFP_NORETRY allocations, and make sure the first THP attempt uses
__GFP_NORETRY. As a side-effect, all costly order __GFP_NORETRY
allocations will bail out if compaction needs reclaim, while
previously they only bailed out when compaction was deferred due to
previous failures.
This should be still acceptable within the __GFP_NORETRY semantics.
- Before the patch, the second allocation attempt (on all nodes) was
passing __GFP_NORETRY. This is redundant as the check for pageblock
order (discussed above) was stronger. It's also contrary to
madvise(MADV_HUGEPAGE) which means some effort to allocate THP is
requested.
After this patch, the second attempt doesn't pass __GFP_THISNODE nor
__GFP_NORETRY.
To sum up, THP page faults now try the following attempts:
1. local node only THP allocation with no reclaim, just compaction.
2. for madvised VMA's or when synchronous compaction is enabled always - THP
allocation from any node with effort determined by global defrag setting
and VMA madvise
3. fallback to base pages on any node
Link: http://lkml.kernel.org/r/08a3f4dd-c3ce-0009-86c5-9ee51aba8557@suse.cz
Fixes: b39d0ee2632d ("mm, page_alloc: avoid expensive reclaim when compaction may not succeed")
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-01-13 16:29:04 -08:00
|
|
|
|
2019-09-04 12:54:25 -07:00
|
|
|
|
2019-09-04 12:54:20 -07:00
|
|
|
|
Revert "mm, thp: consolidate THP gfp handling into alloc_hugepage_direct_gfpmask"
This reverts commit 89c83fb539f95491be80cdd5158e6f0ce329e317.
This should have been done as part of 2f0799a0ffc0 ("mm, thp: restore
node-local hugepage allocations"). The movement of the thp allocation
policy from alloc_pages_vma() to alloc_hugepage_direct_gfpmask() was
intended to only set __GFP_THISNODE for mempolicies that are not
MPOL_BIND whereas the revert could set this regardless of mempolicy.
While the check for MPOL_BIND between alloc_hugepage_direct_gfpmask()
and alloc_pages_vma() was racy, that has since been removed since the
revert. What is left is the possibility to use __GFP_THISNODE in
policy_node() when it is unexpected because the special handling for
hugepages in alloc_pages_vma() was removed as part of the consolidation.
Secondly, prior to 89c83fb539f9, alloc_pages_vma() implemented a somewhat
different policy for hugepage allocations, which were allocated through
alloc_hugepage_vma(). For hugepage allocations, if the allocating
process's node is in the set of allowed nodes, allocate with
__GFP_THISNODE for that node (for MPOL_PREFERRED, use that node with
__GFP_THISNODE instead). This was changed for shmem_alloc_hugepage() to
allow fallback to other nodes in 89c83fb539f9 as it did for new_page() in
mm/mempolicy.c which is functionally different behavior and removes the
requirement to only allocate hugepages locally.
So this commit does a full revert of 89c83fb539f9 instead of the partial
revert that was done in 2f0799a0ffc0. The result is the same thp
allocation policy for 4.20 that was in 4.19.
Fixes: 89c83fb539f9 ("mm, thp: consolidate THP gfp handling into alloc_hugepage_direct_gfpmask")
Fixes: 2f0799a0ffc0 ("mm, thp: restore node-local hugepage allocations")
Signed-off-by: David Rientjes <rientjes@google.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-07 14:50:16 -08:00
|
|
|
|
|
|
|
|
|
2024-11-25 21:01:46 +00:00
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
2024-12-27 18:57:37 +09:00
|
|
|
|
|
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-05-15 15:07:06 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-11-25 21:01:46 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-05-15 15:07:06 +08:00
|
|
|
|
|
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-11-25 21:01:34 +00:00
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
2024-10-10 14:15:56 +08:00
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-05-15 15:07:07 +08:00
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
mm: add MAP_DROPPABLE for designating always lazily freeable mappings
The vDSO getrandom() implementation works with a buffer allocated with a
new system call that has certain requirements:
- It shouldn't be written to core dumps.
* Easy: VM_DONTDUMP.
- It should be zeroed on fork.
* Easy: VM_WIPEONFORK.
- It shouldn't be written to swap.
* Uh-oh: mlock is rlimited.
* Uh-oh: mlock isn't inherited by forks.
- It shouldn't reserve actual memory, but it also shouldn't crash when
page faulting in memory if none is available
* Uh-oh: VM_NORESERVE means segfaults.
It turns out that the vDSO getrandom() function has three really nice
characteristics that we can exploit to solve this problem:
1) Due to being wiped during fork(), the vDSO code is already robust to
having the contents of the pages it reads zeroed out midway through
the function's execution.
2) In the absolute worst case of whatever contingency we're coding for,
we have the option to fallback to the getrandom() syscall, and
everything is fine.
3) The buffers the function uses are only ever useful for a maximum of
60 seconds -- a sort of cache, rather than a long term allocation.
These characteristics mean that we can introduce VM_DROPPABLE, which
has the following semantics:
a) It never is written out to swap.
b) Under memory pressure, mm can just drop the pages (so that they're
zero when read back again).
c) It is inherited by fork.
d) It doesn't count against the mlock budget, since nothing is locked.
e) If there's not enough memory to service a page fault, it's not fatal,
and no signal is sent.
This way, allocations used by vDSO getrandom() can use:
VM_DROPPABLE | VM_DONTDUMP | VM_WIPEONFORK | VM_NORESERVE
And there will be no problem with OOMing, crashing on overcommitment,
using memory when not in use, not wiping on fork(), coredumps, or
writing out to swap.
In order to let vDSO getrandom() use this, expose these via mmap(2) as
MAP_DROPPABLE.
Note that this involves removing the MADV_FREE special case from
sort_folio(), which according to Yu Zhao is unnecessary and will simply
result in an extra call to shrink_folio_list() in the worst case. The
chunk removed reenables the swapbacked flag, which we don't want for
VM_DROPPABLE, and we can't conditionalize it here because there isn't a
vma reference available.
Finally, the provided self test ensures that this is working as desired.
Cc: linux-mm@kvack.org
Acked-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
2022-12-08 17:55:04 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
2024-05-15 15:07:07 +08:00
|
|
|
|
2017-01-24 15:18:18 -08:00
|
|
|
|
2024-05-15 15:07:07 +08:00
|
|
|
|
2022-04-04 15:11:04 -04:00
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
2022-04-04 15:11:04 -04:00
|
|
|
|
2024-11-25 21:01:46 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2021-04-29 23:01:21 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2021-04-29 23:01:21 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2021-04-29 23:01:21 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2024-11-25 21:01:46 +00:00
|
|
|
|
mempolicy: rework mempolicy Reference Counting [yet again]
After further discussion with Christoph Lameter, it has become clear that my
earlier attempts to clean up the mempolicy reference counting were a bit of
overkill in some areas, resulting in superflous ref/unref in what are usually
fast paths. In other areas, further inspection reveals that I botched the
unref for interleave policies.
A separate patch, suitable for upstream/stable trees, fixes up the known
errors in the previous attempt to fix reference counting.
This patch reworks the memory policy referencing counting and, one hopes,
simplifies the code. Maybe I'll get it right this time.
See the update to the numa_memory_policy.txt document for a discussion of
memory policy reference counting that motivates this patch.
Summary:
Lookup of mempolicy, based on (vma, address) need only add a reference for
shared policy, and we need only unref the policy when finished for shared
policies. So, this patch backs out all of the unneeded extra reference
counting added by my previous attempt. It then unrefs only shared policies
when we're finished with them, using the mpol_cond_put() [conditional put]
helper function introduced by this patch.
Note that shmem_swapin() calls read_swap_cache_async() with a dummy vma
containing just the policy. read_swap_cache_async() can call alloc_page_vma()
multiple times, so we can't let alloc_page_vma() unref the shared policy in
this case. To avoid this, we make a copy of any non-null shared policy and
remove the MPOL_F_SHARED flag from the copy. This copy occurs before reading
a page [or multiple pages] from swap, so the overhead should not be an issue
here.
I introduced a new static inline function "mpol_cond_copy()" to copy the
shared policy to an on-stack policy and remove the flags that would require a
conditional free. The current implementation of mpol_cond_copy() assumes that
the struct mempolicy contains no pointers to dynamically allocated structures
that must be duplicated or reference counted during copy.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:16 -07:00
|
|
|
|
2024-11-25 21:01:46 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
2020-12-15 22:55:54 -05:00
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
2020-12-15 22:55:54 -05:00
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
2020-12-15 22:55:54 -05:00
|
|
|
|
2024-12-23 17:00:38 -05:00
|
|
|
|
2021-11-05 13:39:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
2021-11-05 13:39:53 -07:00
|
|
|
|
2024-12-23 17:00:37 -05:00
|
|
|
|
2021-11-05 13:39:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
2021-11-05 13:39:53 -07:00
|
|
|
|
2024-12-23 17:00:37 -05:00
|
|
|
|
2021-11-05 13:39:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-12-23 17:00:38 -05:00
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: Weighted Interleave Auto-tuning
On machines with multiple memory nodes, interleaving page allocations
across nodes allows for better utilization of each node's bandwidth.
Previous work by Gregory Price [1] introduced weighted interleave, which
allowed for pages to be allocated across nodes according to user-set
ratios.
Ideally, these weights should be proportional to their bandwidth, so that
under bandwidth pressure, each node uses its maximal efficient bandwidth
and prevents latency from increasing exponentially.
Previously, weighted interleave's default weights were just 1s -- which
would be equivalent to the (unweighted) interleave mempolicy, which goes
through the nodes in a round-robin fashion, ignoring bandwidth
information.
This patch has two main goals: First, it makes weighted interleave easier
to use for users who wish to relieve bandwidth pressure when using nodes
with varying bandwidth (CXL). By providing a set of "real" default
weights that just work out of the box, users who might not have the
capability (or wish to) perform experimentation to find the most optimal
weights for their system can still take advantage of bandwidth-informed
weighted interleave.
Second, it allows for weighted interleave to dynamically adjust to
hotplugged memory with new bandwidth information. Instead of manually
updating node weights every time new bandwidth information is reported or
taken off, weighted interleave adjusts and provides a new set of default
weights for weighted interleave to use when there is a change in bandwidth
information.
To meet these goals, this patch introduces an auto-configuration mode for
the interleave weights that provides a reasonable set of default weights,
calculated using bandwidth data reported by the system. In auto mode,
weights are dynamically adjusted based on whatever the current bandwidth
information reports (and responds to hotplug events).
This patch still supports users manually writing weights into the nodeN
sysfs interface by entering into manual mode. When a user enters manual
mode, the system stops dynamically updating any of the node weights, even
during hotplug events that shift the optimal weight distribution.
A new sysfs interface "auto" is introduced, which allows users to switch
between the auto (writing 1 or Y) and manual (writing 0 or N) modes. The
system also automatically enters manual mode when a nodeN interface is
manually written to.
There is one functional change that this patch makes to the existing
weighted_interleave ABI: previously, writing 0 directly to a nodeN
interface was said to reset the weight to the system default. Before this
patch, the default for all weights were 1, which meant that writing 0 and
1 were functionally equivalent. With this patch, writing 0 is invalid.
Link: https://lkml.kernel.org/r/20250520141236.2987309-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: wordsmithing changes, simplification, fixes]
Link: https://lkml.kernel.org/r/20250511025840.2410154-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: remove auto_kobj_attr field from struct sysfs_wi_group]
Link: https://lkml.kernel.org/r/20250512142511.3959833-1-joshua.hahnjy@gmail.com
https://lore.kernel.org/linux-mm/20240202170238.90004-1-gregory.price@memverge.com/ [1]
Link: https://lkml.kernel.org/r/20250505182328.4148265-1-joshua.hahnjy@gmail.com
Co-developed-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Yunjeong Mun <yunjeong.mun@sk.com>
Suggested-by: Oscar Salvador <osalvador@suse.de>
Suggested-by: Ying Huang <ying.huang@linux.alibaba.com>
Suggested-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Reviewed-by: Honggyu Kim <honggyu.kim@sk.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-05 11:23:28 -07:00
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
2024-02-02 12:02:38 -05:00
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: Weighted Interleave Auto-tuning
On machines with multiple memory nodes, interleaving page allocations
across nodes allows for better utilization of each node's bandwidth.
Previous work by Gregory Price [1] introduced weighted interleave, which
allowed for pages to be allocated across nodes according to user-set
ratios.
Ideally, these weights should be proportional to their bandwidth, so that
under bandwidth pressure, each node uses its maximal efficient bandwidth
and prevents latency from increasing exponentially.
Previously, weighted interleave's default weights were just 1s -- which
would be equivalent to the (unweighted) interleave mempolicy, which goes
through the nodes in a round-robin fashion, ignoring bandwidth
information.
This patch has two main goals: First, it makes weighted interleave easier
to use for users who wish to relieve bandwidth pressure when using nodes
with varying bandwidth (CXL). By providing a set of "real" default
weights that just work out of the box, users who might not have the
capability (or wish to) perform experimentation to find the most optimal
weights for their system can still take advantage of bandwidth-informed
weighted interleave.
Second, it allows for weighted interleave to dynamically adjust to
hotplugged memory with new bandwidth information. Instead of manually
updating node weights every time new bandwidth information is reported or
taken off, weighted interleave adjusts and provides a new set of default
weights for weighted interleave to use when there is a change in bandwidth
information.
To meet these goals, this patch introduces an auto-configuration mode for
the interleave weights that provides a reasonable set of default weights,
calculated using bandwidth data reported by the system. In auto mode,
weights are dynamically adjusted based on whatever the current bandwidth
information reports (and responds to hotplug events).
This patch still supports users manually writing weights into the nodeN
sysfs interface by entering into manual mode. When a user enters manual
mode, the system stops dynamically updating any of the node weights, even
during hotplug events that shift the optimal weight distribution.
A new sysfs interface "auto" is introduced, which allows users to switch
between the auto (writing 1 or Y) and manual (writing 0 or N) modes. The
system also automatically enters manual mode when a nodeN interface is
manually written to.
There is one functional change that this patch makes to the existing
weighted_interleave ABI: previously, writing 0 directly to a nodeN
interface was said to reset the weight to the system default. Before this
patch, the default for all weights were 1, which meant that writing 0 and
1 were functionally equivalent. With this patch, writing 0 is invalid.
Link: https://lkml.kernel.org/r/20250520141236.2987309-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: wordsmithing changes, simplification, fixes]
Link: https://lkml.kernel.org/r/20250511025840.2410154-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: remove auto_kobj_attr field from struct sysfs_wi_group]
Link: https://lkml.kernel.org/r/20250512142511.3959833-1-joshua.hahnjy@gmail.com
https://lore.kernel.org/linux-mm/20240202170238.90004-1-gregory.price@memverge.com/ [1]
Link: https://lkml.kernel.org/r/20250505182328.4148265-1-joshua.hahnjy@gmail.com
Co-developed-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Yunjeong Mun <yunjeong.mun@sk.com>
Suggested-by: Oscar Salvador <osalvador@suse.de>
Suggested-by: Ying Huang <ying.huang@linux.alibaba.com>
Suggested-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Reviewed-by: Honggyu Kim <honggyu.kim@sk.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-05 11:23:28 -07:00
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-02-02 12:02:38 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-12-23 17:00:37 -05:00
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: Weighted Interleave Auto-tuning
On machines with multiple memory nodes, interleaving page allocations
across nodes allows for better utilization of each node's bandwidth.
Previous work by Gregory Price [1] introduced weighted interleave, which
allowed for pages to be allocated across nodes according to user-set
ratios.
Ideally, these weights should be proportional to their bandwidth, so that
under bandwidth pressure, each node uses its maximal efficient bandwidth
and prevents latency from increasing exponentially.
Previously, weighted interleave's default weights were just 1s -- which
would be equivalent to the (unweighted) interleave mempolicy, which goes
through the nodes in a round-robin fashion, ignoring bandwidth
information.
This patch has two main goals: First, it makes weighted interleave easier
to use for users who wish to relieve bandwidth pressure when using nodes
with varying bandwidth (CXL). By providing a set of "real" default
weights that just work out of the box, users who might not have the
capability (or wish to) perform experimentation to find the most optimal
weights for their system can still take advantage of bandwidth-informed
weighted interleave.
Second, it allows for weighted interleave to dynamically adjust to
hotplugged memory with new bandwidth information. Instead of manually
updating node weights every time new bandwidth information is reported or
taken off, weighted interleave adjusts and provides a new set of default
weights for weighted interleave to use when there is a change in bandwidth
information.
To meet these goals, this patch introduces an auto-configuration mode for
the interleave weights that provides a reasonable set of default weights,
calculated using bandwidth data reported by the system. In auto mode,
weights are dynamically adjusted based on whatever the current bandwidth
information reports (and responds to hotplug events).
This patch still supports users manually writing weights into the nodeN
sysfs interface by entering into manual mode. When a user enters manual
mode, the system stops dynamically updating any of the node weights, even
during hotplug events that shift the optimal weight distribution.
A new sysfs interface "auto" is introduced, which allows users to switch
between the auto (writing 1 or Y) and manual (writing 0 or N) modes. The
system also automatically enters manual mode when a nodeN interface is
manually written to.
There is one functional change that this patch makes to the existing
weighted_interleave ABI: previously, writing 0 directly to a nodeN
interface was said to reset the weight to the system default. Before this
patch, the default for all weights were 1, which meant that writing 0 and
1 were functionally equivalent. With this patch, writing 0 is invalid.
Link: https://lkml.kernel.org/r/20250520141236.2987309-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: wordsmithing changes, simplification, fixes]
Link: https://lkml.kernel.org/r/20250511025840.2410154-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: remove auto_kobj_attr field from struct sysfs_wi_group]
Link: https://lkml.kernel.org/r/20250512142511.3959833-1-joshua.hahnjy@gmail.com
https://lore.kernel.org/linux-mm/20240202170238.90004-1-gregory.price@memverge.com/ [1]
Link: https://lkml.kernel.org/r/20250505182328.4148265-1-joshua.hahnjy@gmail.com
Co-developed-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Yunjeong Mun <yunjeong.mun@sk.com>
Suggested-by: Oscar Salvador <osalvador@suse.de>
Suggested-by: Ying Huang <ying.huang@linux.alibaba.com>
Suggested-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Reviewed-by: Honggyu Kim <honggyu.kim@sk.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-05 11:23:28 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: Weighted Interleave Auto-tuning
On machines with multiple memory nodes, interleaving page allocations
across nodes allows for better utilization of each node's bandwidth.
Previous work by Gregory Price [1] introduced weighted interleave, which
allowed for pages to be allocated across nodes according to user-set
ratios.
Ideally, these weights should be proportional to their bandwidth, so that
under bandwidth pressure, each node uses its maximal efficient bandwidth
and prevents latency from increasing exponentially.
Previously, weighted interleave's default weights were just 1s -- which
would be equivalent to the (unweighted) interleave mempolicy, which goes
through the nodes in a round-robin fashion, ignoring bandwidth
information.
This patch has two main goals: First, it makes weighted interleave easier
to use for users who wish to relieve bandwidth pressure when using nodes
with varying bandwidth (CXL). By providing a set of "real" default
weights that just work out of the box, users who might not have the
capability (or wish to) perform experimentation to find the most optimal
weights for their system can still take advantage of bandwidth-informed
weighted interleave.
Second, it allows for weighted interleave to dynamically adjust to
hotplugged memory with new bandwidth information. Instead of manually
updating node weights every time new bandwidth information is reported or
taken off, weighted interleave adjusts and provides a new set of default
weights for weighted interleave to use when there is a change in bandwidth
information.
To meet these goals, this patch introduces an auto-configuration mode for
the interleave weights that provides a reasonable set of default weights,
calculated using bandwidth data reported by the system. In auto mode,
weights are dynamically adjusted based on whatever the current bandwidth
information reports (and responds to hotplug events).
This patch still supports users manually writing weights into the nodeN
sysfs interface by entering into manual mode. When a user enters manual
mode, the system stops dynamically updating any of the node weights, even
during hotplug events that shift the optimal weight distribution.
A new sysfs interface "auto" is introduced, which allows users to switch
between the auto (writing 1 or Y) and manual (writing 0 or N) modes. The
system also automatically enters manual mode when a nodeN interface is
manually written to.
There is one functional change that this patch makes to the existing
weighted_interleave ABI: previously, writing 0 directly to a nodeN
interface was said to reset the weight to the system default. Before this
patch, the default for all weights were 1, which meant that writing 0 and
1 were functionally equivalent. With this patch, writing 0 is invalid.
Link: https://lkml.kernel.org/r/20250520141236.2987309-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: wordsmithing changes, simplification, fixes]
Link: https://lkml.kernel.org/r/20250511025840.2410154-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: remove auto_kobj_attr field from struct sysfs_wi_group]
Link: https://lkml.kernel.org/r/20250512142511.3959833-1-joshua.hahnjy@gmail.com
https://lore.kernel.org/linux-mm/20240202170238.90004-1-gregory.price@memverge.com/ [1]
Link: https://lkml.kernel.org/r/20250505182328.4148265-1-joshua.hahnjy@gmail.com
Co-developed-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Yunjeong Mun <yunjeong.mun@sk.com>
Suggested-by: Oscar Salvador <osalvador@suse.de>
Suggested-by: Ying Huang <ying.huang@linux.alibaba.com>
Suggested-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Reviewed-by: Honggyu Kim <honggyu.kim@sk.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-05 11:23:28 -07:00
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-12-23 17:00:37 -05:00
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-12-23 17:00:38 -05:00
|
|
|
|
2021-11-05 13:39:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
2024-12-23 17:00:37 -05:00
|
|
|
|
2021-11-05 13:39:53 -07:00
|
|
|
|
|
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
2024-12-23 17:00:37 -05:00
|
|
|
|
2021-11-05 13:39:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-12-23 17:00:38 -05:00
|
|
|
|
2021-11-05 13:39:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
|
|
|
|
|
2021-11-05 13:39:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-12-23 17:00:38 -05:00
|
|
|
|
2021-11-05 13:39:53 -07:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
2024-12-23 17:00:38 -05:00
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
|
|
|
|
|
2021-11-05 13:39:53 -07:00
|
|
|
|
2024-12-23 17:00:38 -05:00
|
|
|
|
2021-11-05 13:39:53 -07:00
|
|
|
|
|
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
|
|
|
|
|
2024-03-21 09:36:40 -07:00
|
|
|
|
2024-12-23 17:00:37 -05:00
|
|
|
|
2021-11-05 13:39:53 -07:00
|
|
|
|
|
|
|
|
|
2013-09-11 14:20:14 -07:00
|
|
|
|
|
|
|
|
|
2023-10-03 02:20:14 -07:00
|
|
|
|
2013-09-11 14:20:14 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
[PATCH] cpuset: rebind vma mempolicies fix
Fix more of longstanding bug in cpuset/mempolicy interaction.
NUMA mempolicies (mm/mempolicy.c) are constrained by the current tasks cpuset
to just the Memory Nodes allowed by that cpuset. The kernel maintains
internal state for each mempolicy, tracking what nodes are used for the
MPOL_INTERLEAVE, MPOL_BIND or MPOL_PREFERRED policies.
When a tasks cpuset memory placement changes, whether because the cpuset
changed, or because the task was attached to a different cpuset, then the
tasks mempolicies have to be rebound to the new cpuset placement, so as to
preserve the cpuset-relative numbering of the nodes in that policy.
An earlier fix handled such mempolicy rebinding for mempolicies attached to a
task.
This fix rebinds mempolicies attached to vma's (address ranges in a tasks
address space.) Due to the need to hold the task->mm->mmap_sem semaphore while
updating vma's, the rebinding of vma mempolicies has to be done when the
cpuset memory placement is changed, at which time mmap_sem can be safely
acquired. The tasks mempolicy is rebound later, when the task next attempts
to allocate memory and notices that its task->cpuset_mems_generation is
out-of-date with its cpusets mems_generation.
Because walking the tasklist to find all tasks attached to a changing cpuset
requires holding tasklist_lock, a spinlock, one cannot update the vma's of the
affected tasks while doing the tasklist scan. In general, one cannot acquire
a semaphore (which can sleep) while already holding a spinlock (such as
tasklist_lock). So a list of mm references has to be built up during the
tasklist scan, then the tasklist lock dropped, then for each mm, its mmap_sem
acquired, and the vma's in that mm rebound.
Once the tasklist lock is dropped, affected tasks may fork new tasks, before
their mm's are rebound. A kernel global 'cpuset_being_rebound' is set to
point to the cpuset being rebound (there can only be one; cpuset modifications
are done under a global 'manage_sem' semaphore), and the mpol_copy code that
is used to copy a tasks mempolicies during fork catches such forking tasks,
and ensures their children are also rebound.
When a task is moved to a different cpuset, it is easier, as there is only one
task involved. It's mm->vma's are scanned, using the same
mpol_rebind_policy() as used above.
It may happen that both the mpol_copy hook and the update done via the
tasklist scan update the same mm twice. This is ok, as the mempolicies of
each vma in an mm keep track of what mems_allowed they are relative to, and
safely no-op a second request to rebind to the same nodes.
Signed-off-by: Paul Jackson <pj@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-01-08 01:01:59 -08:00
|
|
|
|
2008-04-28 02:13:09 -07:00
|
|
|
|
[PATCH] cpuset: rebind vma mempolicies fix
Fix more of longstanding bug in cpuset/mempolicy interaction.
NUMA mempolicies (mm/mempolicy.c) are constrained by the current tasks cpuset
to just the Memory Nodes allowed by that cpuset. The kernel maintains
internal state for each mempolicy, tracking what nodes are used for the
MPOL_INTERLEAVE, MPOL_BIND or MPOL_PREFERRED policies.
When a tasks cpuset memory placement changes, whether because the cpuset
changed, or because the task was attached to a different cpuset, then the
tasks mempolicies have to be rebound to the new cpuset placement, so as to
preserve the cpuset-relative numbering of the nodes in that policy.
An earlier fix handled such mempolicy rebinding for mempolicies attached to a
task.
This fix rebinds mempolicies attached to vma's (address ranges in a tasks
address space.) Due to the need to hold the task->mm->mmap_sem semaphore while
updating vma's, the rebinding of vma mempolicies has to be done when the
cpuset memory placement is changed, at which time mmap_sem can be safely
acquired. The tasks mempolicy is rebound later, when the task next attempts
to allocate memory and notices that its task->cpuset_mems_generation is
out-of-date with its cpusets mems_generation.
Because walking the tasklist to find all tasks attached to a changing cpuset
requires holding tasklist_lock, a spinlock, one cannot update the vma's of the
affected tasks while doing the tasklist scan. In general, one cannot acquire
a semaphore (which can sleep) while already holding a spinlock (such as
tasklist_lock). So a list of mm references has to be built up during the
tasklist scan, then the tasklist lock dropped, then for each mm, its mmap_sem
acquired, and the vma's in that mm rebound.
Once the tasklist lock is dropped, affected tasks may fork new tasks, before
their mm's are rebound. A kernel global 'cpuset_being_rebound' is set to
point to the cpuset being rebound (there can only be one; cpuset modifications
are done under a global 'manage_sem' semaphore), and the mpol_copy code that
is used to copy a tasks mempolicies during fork catches such forking tasks,
and ensures their children are also rebound.
When a task is moved to a different cpuset, it is easier, as there is only one
task involved. It's mm->vma's are scanned, using the same
mpol_rebind_policy() as used above.
It may happen that both the mpol_copy hook and the update done via the
tasklist scan update the same mm twice. This is ok, as the mempolicies of
each vma in an mm keep track of what mems_allowed they are relative to, and
safely no-op a second request to rebind to the same nodes.
Signed-off-by: Paul Jackson <pj@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-01-08 01:01:59 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: restructure rebinding-mempolicy functions
Nick Piggin reported that the allocator may see an empty nodemask when
changing cpuset's mems[1]. It happens only on the kernel that do not do
atomic nodemask_t stores. (MAX_NUMNODES > BITS_PER_LONG)
But I found that there is also a problem on the kernel that can do atomic
nodemask_t stores. The problem is that the allocator can't find a node to
alloc page when changing cpuset's mems though there is a lot of free
memory. The reason is like this:
(mpol: mempolicy)
task1 task1's mpol task2
alloc page 1
alloc on node0? NO 1
1 change mems from 1 to 0
1 rebind task1's mpol
0-1 set new bits
0 clear disallowed bits
alloc on node1? NO 0
...
can't alloc page
goto oom
I can use the attached program reproduce it by the following step:
# mkdir /dev/cpuset
# mount -t cpuset cpuset /dev/cpuset
# mkdir /dev/cpuset/1
# echo `cat /dev/cpuset/cpus` > /dev/cpuset/1/cpus
# echo `cat /dev/cpuset/mems` > /dev/cpuset/1/mems
# echo $$ > /dev/cpuset/1/tasks
# numactl --membind=`cat /dev/cpuset/mems` ./cpuset_mem_hog <nr_tasks> &
<nr_tasks> = max(nr_cpus - 1, 1)
# killall -s SIGUSR1 cpuset_mem_hog
# ./change_mems.sh
several hours later, oom will happen though there is a lot of free memory.
This patchset fixes this problem by expanding the nodes range first(set
newly allowed bits) and shrink it lazily(clear newly disallowed bits). So
we use a variable to tell the write-side task that read-side task is
reading nodemask, and the write-side task clears newly disallowed nodes
after read-side task ends the current memory allocation.
This patch:
In order to fix no node to alloc memory, when we want to update mempolicy
and mems_allowed, we expand the set of nodes first (set all the newly
nodes) and shrink the set of nodes lazily(clean disallowed nodes), But the
mempolicy's rebind functions may breaks the expanding.
So we restructure the mempolicy's rebind functions and split the rebind
work to two steps, just like the update of cpuset's mems: The 1st step:
expand the set of the mempolicy's nodes. The 2nd step: shrink the set of
the mempolicy's nodes. It is used when there is no real lock to protect
the mempolicy in the read-side. Otherwise we can do rebind work at once.
In order to implement it, we define
enum mpol_rebind_step {
MPOL_REBIND_ONCE,
MPOL_REBIND_STEP1,
MPOL_REBIND_STEP2,
MPOL_REBIND_NSTEP,
};
If the mempolicy needn't be updated by two steps, we can pass
MPOL_REBIND_ONCE to the rebind functions. Or we can pass
MPOL_REBIND_STEP1 to do the first step of the rebind work and pass
MPOL_REBIND_STEP2 to do the second step work.
Besides that, it maybe long time between these two step and we have to
release the lock that protects mempolicy and mems_allowed. If we hold the
lock once again, we must check whether the current mempolicy is under the
rebinding (the first step has been done) or not, because the task may
alloc a new mempolicy when we don't hold the lock. So we defined the
following flag to identify it:
#define MPOL_F_REBINDING (1 << 2)
The new functions will be used in the next patch.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Nick Piggin <npiggin@suse.de>
Cc: Paul Menage <menage@google.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Cc: Ravikiran Thirumalai <kiran@scalex86.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Andi Kleen <andi@firstfloor.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-24 14:32:07 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
[PATCH] cpuset: rebind vma mempolicies fix
Fix more of longstanding bug in cpuset/mempolicy interaction.
NUMA mempolicies (mm/mempolicy.c) are constrained by the current tasks cpuset
to just the Memory Nodes allowed by that cpuset. The kernel maintains
internal state for each mempolicy, tracking what nodes are used for the
MPOL_INTERLEAVE, MPOL_BIND or MPOL_PREFERRED policies.
When a tasks cpuset memory placement changes, whether because the cpuset
changed, or because the task was attached to a different cpuset, then the
tasks mempolicies have to be rebound to the new cpuset placement, so as to
preserve the cpuset-relative numbering of the nodes in that policy.
An earlier fix handled such mempolicy rebinding for mempolicies attached to a
task.
This fix rebinds mempolicies attached to vma's (address ranges in a tasks
address space.) Due to the need to hold the task->mm->mmap_sem semaphore while
updating vma's, the rebinding of vma mempolicies has to be done when the
cpuset memory placement is changed, at which time mmap_sem can be safely
acquired. The tasks mempolicy is rebound later, when the task next attempts
to allocate memory and notices that its task->cpuset_mems_generation is
out-of-date with its cpusets mems_generation.
Because walking the tasklist to find all tasks attached to a changing cpuset
requires holding tasklist_lock, a spinlock, one cannot update the vma's of the
affected tasks while doing the tasklist scan. In general, one cannot acquire
a semaphore (which can sleep) while already holding a spinlock (such as
tasklist_lock). So a list of mm references has to be built up during the
tasklist scan, then the tasklist lock dropped, then for each mm, its mmap_sem
acquired, and the vma's in that mm rebound.
Once the tasklist lock is dropped, affected tasks may fork new tasks, before
their mm's are rebound. A kernel global 'cpuset_being_rebound' is set to
point to the cpuset being rebound (there can only be one; cpuset modifications
are done under a global 'manage_sem' semaphore), and the mpol_copy code that
is used to copy a tasks mempolicies during fork catches such forking tasks,
and ensures their children are also rebound.
When a task is moved to a different cpuset, it is easier, as there is only one
task involved. It's mm->vma's are scanned, using the same
mpol_rebind_policy() as used above.
It may happen that both the mpol_copy hook and the update done via the
tasklist scan update the same mm twice. This is ok, as the mempolicies of
each vma in an mm keep track of what mems_allowed they are relative to, and
safely no-op a second request to rebind to the same nodes.
Signed-off-by: Paul Jackson <pj@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-01-08 01:01:59 -08:00
|
|
|
|
|
|
|
|
|
2008-04-28 02:13:09 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: restructure rebinding-mempolicy functions
Nick Piggin reported that the allocator may see an empty nodemask when
changing cpuset's mems[1]. It happens only on the kernel that do not do
atomic nodemask_t stores. (MAX_NUMNODES > BITS_PER_LONG)
But I found that there is also a problem on the kernel that can do atomic
nodemask_t stores. The problem is that the allocator can't find a node to
alloc page when changing cpuset's mems though there is a lot of free
memory. The reason is like this:
(mpol: mempolicy)
task1 task1's mpol task2
alloc page 1
alloc on node0? NO 1
1 change mems from 1 to 0
1 rebind task1's mpol
0-1 set new bits
0 clear disallowed bits
alloc on node1? NO 0
...
can't alloc page
goto oom
I can use the attached program reproduce it by the following step:
# mkdir /dev/cpuset
# mount -t cpuset cpuset /dev/cpuset
# mkdir /dev/cpuset/1
# echo `cat /dev/cpuset/cpus` > /dev/cpuset/1/cpus
# echo `cat /dev/cpuset/mems` > /dev/cpuset/1/mems
# echo $$ > /dev/cpuset/1/tasks
# numactl --membind=`cat /dev/cpuset/mems` ./cpuset_mem_hog <nr_tasks> &
<nr_tasks> = max(nr_cpus - 1, 1)
# killall -s SIGUSR1 cpuset_mem_hog
# ./change_mems.sh
several hours later, oom will happen though there is a lot of free memory.
This patchset fixes this problem by expanding the nodes range first(set
newly allowed bits) and shrink it lazily(clear newly disallowed bits). So
we use a variable to tell the write-side task that read-side task is
reading nodemask, and the write-side task clears newly disallowed nodes
after read-side task ends the current memory allocation.
This patch:
In order to fix no node to alloc memory, when we want to update mempolicy
and mems_allowed, we expand the set of nodes first (set all the newly
nodes) and shrink the set of nodes lazily(clean disallowed nodes), But the
mempolicy's rebind functions may breaks the expanding.
So we restructure the mempolicy's rebind functions and split the rebind
work to two steps, just like the update of cpuset's mems: The 1st step:
expand the set of the mempolicy's nodes. The 2nd step: shrink the set of
the mempolicy's nodes. It is used when there is no real lock to protect
the mempolicy in the read-side. Otherwise we can do rebind work at once.
In order to implement it, we define
enum mpol_rebind_step {
MPOL_REBIND_ONCE,
MPOL_REBIND_STEP1,
MPOL_REBIND_STEP2,
MPOL_REBIND_NSTEP,
};
If the mempolicy needn't be updated by two steps, we can pass
MPOL_REBIND_ONCE to the rebind functions. Or we can pass
MPOL_REBIND_STEP1 to do the first step of the rebind work and pass
MPOL_REBIND_STEP2 to do the second step work.
Besides that, it maybe long time between these two step and we have to
release the lock that protects mempolicy and mems_allowed. If we hold the
lock once again, we must check whether the current mempolicy is under the
rebinding (the first step has been done) or not, because the task may
alloc a new mempolicy when we don't hold the lock. So we defined the
following flag to identify it:
#define MPOL_F_REBINDING (1 << 2)
The new functions will be used in the next patch.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Nick Piggin <npiggin@suse.de>
Cc: Paul Menage <menage@google.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Cc: Ravikiran Thirumalai <kiran@scalex86.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Andi Kleen <andi@firstfloor.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-24 14:32:07 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
[PATCH] cpuset: rebind vma mempolicies fix
Fix more of longstanding bug in cpuset/mempolicy interaction.
NUMA mempolicies (mm/mempolicy.c) are constrained by the current tasks cpuset
to just the Memory Nodes allowed by that cpuset. The kernel maintains
internal state for each mempolicy, tracking what nodes are used for the
MPOL_INTERLEAVE, MPOL_BIND or MPOL_PREFERRED policies.
When a tasks cpuset memory placement changes, whether because the cpuset
changed, or because the task was attached to a different cpuset, then the
tasks mempolicies have to be rebound to the new cpuset placement, so as to
preserve the cpuset-relative numbering of the nodes in that policy.
An earlier fix handled such mempolicy rebinding for mempolicies attached to a
task.
This fix rebinds mempolicies attached to vma's (address ranges in a tasks
address space.) Due to the need to hold the task->mm->mmap_sem semaphore while
updating vma's, the rebinding of vma mempolicies has to be done when the
cpuset memory placement is changed, at which time mmap_sem can be safely
acquired. The tasks mempolicy is rebound later, when the task next attempts
to allocate memory and notices that its task->cpuset_mems_generation is
out-of-date with its cpusets mems_generation.
Because walking the tasklist to find all tasks attached to a changing cpuset
requires holding tasklist_lock, a spinlock, one cannot update the vma's of the
affected tasks while doing the tasklist scan. In general, one cannot acquire
a semaphore (which can sleep) while already holding a spinlock (such as
tasklist_lock). So a list of mm references has to be built up during the
tasklist scan, then the tasklist lock dropped, then for each mm, its mmap_sem
acquired, and the vma's in that mm rebound.
Once the tasklist lock is dropped, affected tasks may fork new tasks, before
their mm's are rebound. A kernel global 'cpuset_being_rebound' is set to
point to the cpuset being rebound (there can only be one; cpuset modifications
are done under a global 'manage_sem' semaphore), and the mpol_copy code that
is used to copy a tasks mempolicies during fork catches such forking tasks,
and ensures their children are also rebound.
When a task is moved to a different cpuset, it is easier, as there is only one
task involved. It's mm->vma's are scanned, using the same
mpol_rebind_policy() as used above.
It may happen that both the mpol_copy hook and the update done via the
tasklist scan update the same mm twice. This is ok, as the mempolicies of
each vma in an mm keep track of what mems_allowed they are relative to, and
safely no-op a second request to rebind to the same nodes.
Signed-off-by: Paul Jackson <pj@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-01-08 01:01:59 -08:00
|
|
|
|
|
|
|
|
|
2017-07-06 15:40:06 -07:00
|
|
|
|
[PATCH] cpuset: rebind vma mempolicies fix
Fix more of longstanding bug in cpuset/mempolicy interaction.
NUMA mempolicies (mm/mempolicy.c) are constrained by the current tasks cpuset
to just the Memory Nodes allowed by that cpuset. The kernel maintains
internal state for each mempolicy, tracking what nodes are used for the
MPOL_INTERLEAVE, MPOL_BIND or MPOL_PREFERRED policies.
When a tasks cpuset memory placement changes, whether because the cpuset
changed, or because the task was attached to a different cpuset, then the
tasks mempolicies have to be rebound to the new cpuset placement, so as to
preserve the cpuset-relative numbering of the nodes in that policy.
An earlier fix handled such mempolicy rebinding for mempolicies attached to a
task.
This fix rebinds mempolicies attached to vma's (address ranges in a tasks
address space.) Due to the need to hold the task->mm->mmap_sem semaphore while
updating vma's, the rebinding of vma mempolicies has to be done when the
cpuset memory placement is changed, at which time mmap_sem can be safely
acquired. The tasks mempolicy is rebound later, when the task next attempts
to allocate memory and notices that its task->cpuset_mems_generation is
out-of-date with its cpusets mems_generation.
Because walking the tasklist to find all tasks attached to a changing cpuset
requires holding tasklist_lock, a spinlock, one cannot update the vma's of the
affected tasks while doing the tasklist scan. In general, one cannot acquire
a semaphore (which can sleep) while already holding a spinlock (such as
tasklist_lock). So a list of mm references has to be built up during the
tasklist scan, then the tasklist lock dropped, then for each mm, its mmap_sem
acquired, and the vma's in that mm rebound.
Once the tasklist lock is dropped, affected tasks may fork new tasks, before
their mm's are rebound. A kernel global 'cpuset_being_rebound' is set to
point to the cpuset being rebound (there can only be one; cpuset modifications
are done under a global 'manage_sem' semaphore), and the mpol_copy code that
is used to copy a tasks mempolicies during fork catches such forking tasks,
and ensures their children are also rebound.
When a task is moved to a different cpuset, it is easier, as there is only one
task involved. It's mm->vma's are scanned, using the same
mpol_rebind_policy() as used above.
It may happen that both the mpol_copy hook and the update done via the
tasklist scan update the same mm twice. This is ok, as the mempolicies of
each vma in an mm keep track of what mems_allowed they are relative to, and
safely no-op a second request to rebind to the same nodes.
Signed-off-by: Paul Jackson <pj@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-01-08 01:01:59 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-01-10 15:08:21 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2012-01-10 15:08:21 -08:00
|
|
|
|
2008-04-28 02:13:12 -07:00
|
|
|
|
2012-01-10 15:08:21 -08:00
|
|
|
|
2010-05-24 14:32:01 -07:00
|
|
|
|
2012-01-10 15:08:21 -08:00
|
|
|
|
mm/mempolicy: add set_mempolicy_home_node syscall
This syscall can be used to set a home node for the MPOL_BIND and
MPOL_PREFERRED_MANY memory policy. Users should use this syscall after
setting up a memory policy for the specified range as shown below.
mbind(p, nr_pages * page_size, MPOL_BIND, new_nodes->maskp,
new_nodes->size + 1, 0);
sys_set_mempolicy_home_node((unsigned long)p, nr_pages * page_size,
home_node, 0);
The syscall allows specifying a home node/preferred node from which
kernel will fulfill memory allocation requests first.
For address range with MPOL_BIND memory policy, if nodemask specifies
more than one node, page allocations will come from the node in the
nodemask with sufficient free memory that is closest to the home
node/preferred node.
For MPOL_PREFERRED_MANY if the nodemask specifies more than one node,
page allocation will come from the node in the nodemask with sufficient
free memory that is closest to the home node/preferred node. If there
is not enough memory in all the nodes specified in the nodemask, the
allocation will be attempted from the closest numa node to the home node
in the system.
This helps applications to hint at a memory allocation preference node
and fallback to _only_ a set of nodes if the memory is not available on
the preferred node. Fallback allocation is attempted from the node
which is nearest to the preferred node.
This helps applications to have control on memory allocation numa nodes
and avoids default fallback to slow memory NUMA nodes. For example a
system with NUMA nodes 1,2 and 3 with DRAM memory and 10, 11 and 12 of
slow memory
new_nodes = numa_bitmask_alloc(nr_nodes);
numa_bitmask_setbit(new_nodes, 1);
numa_bitmask_setbit(new_nodes, 2);
numa_bitmask_setbit(new_nodes, 3);
p = mmap(NULL, nr_pages * page_size, protflag, mapflag, -1, 0);
mbind(p, nr_pages * page_size, MPOL_BIND, new_nodes->maskp, new_nodes->size + 1, 0);
sys_set_mempolicy_home_node(p, nr_pages * page_size, 2, 0);
This will allocate from nodes closer to node 2 and will make sure the
kernel will only allocate from nodes 1, 2, and 3. Memory will not be
allocated from slow memory nodes 10, 11, and 12. This differs from
default MPOL_BIND behavior in that with default MPOL_BIND the allocation
will be attempted from node closer to the local node. One of the
reasons to specify a home node is to allow allocations from cpu less
NUMA node and its nearby NUMA nodes.
With MPOL_PREFERRED_MANY on the other hand will first try to allocate
from the closest node to node 2 from the node list 1, 2 and 3. If those
nodes don't have enough memory, kernel will allocate from slow memory
node 10, 11 and 12 which ever is closer to node 2.
Link: https://lkml.kernel.org/r/20211202123810.267175-3-aneesh.kumar@linux.ibm.com
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Ben Widawsky <ben.widawsky@intel.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Feng Tang <feng.tang@intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: <linux-api@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-14 14:08:17 -08:00
|
|
|
|
|
|
|
|
|
2010-05-24 14:32:01 -07:00
|
|
|
|
|
|
|
|
|
2012-01-10 15:08:21 -08:00
|
|
|
|
2010-05-24 14:32:01 -07:00
|
|
|
|
2008-04-28 02:13:12 -07:00
|
|
|
|
2008-04-28 02:12:18 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: add MPOL_PREFERRED_MANY for multiple preferred nodes
Patch series "Introduce multi-preference mempolicy", v7.
This patch series introduces the concept of the MPOL_PREFERRED_MANY
mempolicy. This mempolicy mode can be used with either the
set_mempolicy(2) or mbind(2) interfaces. Like the MPOL_PREFERRED
interface, it allows an application to set a preference for nodes which
will fulfil memory allocation requests. Unlike the MPOL_PREFERRED mode,
it takes a set of nodes. Like the MPOL_BIND interface, it works over a
set of nodes. Unlike MPOL_BIND, it will not cause a SIGSEGV or invoke the
OOM killer if those preferred nodes are not available.
Along with these patches are patches for libnuma, numactl, numademo, and
memhog. They still need some polish, but can be found here:
https://gitlab.com/bwidawsk/numactl/-/tree/prefer-many It allows new
usage: `numactl -P 0,3,4`
The goal of the new mode is to enable some use-cases when using tiered memory
usage models which I've lovingly named.
1a. The Hare - The interconnect is fast enough to meet bandwidth and
latency requirements allowing preference to be given to all nodes with
"fast" memory.
1b. The Indiscriminate Hare - An application knows it wants fast
memory (or perhaps slow memory), but doesn't care which node it runs
on. The application can prefer a set of nodes and then xpu bind to
the local node (cpu, accelerator, etc). This reverses the nodes are
chosen today where the kernel attempts to use local memory to the CPU
whenever possible. This will attempt to use the local accelerator to
the memory.
2. The Tortoise - The administrator (or the application itself) is
aware it only needs slow memory, and so can prefer that.
Much of this is almost achievable with the bind interface, but the bind
interface suffers from an inability to fallback to another set of nodes if
binding fails to all nodes in the nodemask.
Like MPOL_BIND a nodemask is given. Inherently this removes ordering from the
preference.
> /* Set first two nodes as preferred in an 8 node system. */
> const unsigned long nodes = 0x3
> set_mempolicy(MPOL_PREFER_MANY, &nodes, 8);
> /* Mimic interleave policy, but have fallback *.
> const unsigned long nodes = 0xaa
> set_mempolicy(MPOL_PREFER_MANY, &nodes, 8);
Some internal discussion took place around the interface. There are two
alternatives which we have discussed, plus one I stuck in:
1. Ordered list of nodes. Currently it's believed that the added
complexity is nod needed for expected usecases.
2. A flag for bind to allow falling back to other nodes. This
confuses the notion of binding and is less flexible than the current
solution.
3. Create flags or new modes that helps with some ordering. This
offers both a friendlier API as well as a solution for more customized
usage. It's unknown if it's worth the complexity to support this.
Here is sample code for how this might work:
> // Prefer specific nodes for some something wacky
> set_mempolicy(MPOL_PREFER_MANY, 0x17c, 1024);
>
> // Default
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_SOCKET, NULL, 0);
> // which is the same as
> set_mempolicy(MPOL_DEFAULT, NULL, 0);
>
> // The Hare
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE, NULL, 0);
>
> // The Tortoise
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE_REV, NULL, 0);
>
> // Prefer the fast memory of the first two sockets
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE, -1, 2);
>
This patch (of 5):
The NUMA APIs currently allow passing in a "preferred node" as a single
bit set in a nodemask. If more than one bit it set, bits after the first
are ignored.
This single node is generally OK for location-based NUMA where memory
being allocated will eventually be operated on by a single CPU. However,
in systems with multiple memory types, folks want to target a *type* of
memory instead of a location. For instance, someone might want some
high-bandwidth memory but do not care about the CPU next to which it is
allocated. Or, they want a cheap, high capacity allocation and want to
target all NUMA nodes which have persistent memory in volatile mode. In
both of these cases, the application wants to target a *set* of nodes, but
does not want strict MPOL_BIND behavior as that could lead to OOM killer
or SIGSEGV.
So add MPOL_PREFERRED_MANY policy to support the multiple preferred nodes
requirement. This is not a pie-in-the-sky dream for an API. This was a
response to a specific ask of more than one group at Intel. Specifically:
1. There are existing libraries that target memory types such as
https://github.com/memkind/memkind. These are known to suffer from
SIGSEGV's when memory is low on targeted memory "kinds" that span more
than one node. The MCDRAM on a Xeon Phi in "Cluster on Die" mode is an
example of this.
2. Volatile-use persistent memory users want to have a memory policy
which is targeted at either "cheap and slow" (PMEM) or "expensive and
fast" (DRAM). However, they do not want to experience allocation
failures when the targeted type is unavailable.
3. Allocate-then-run. Generally, we let the process scheduler decide
on which physical CPU to run a task. That location provides a default
allocation policy, and memory availability is not generally considered
when placing tasks. For situations where memory is valuable and
constrained, some users want to allocate memory first, *then* allocate
close compute resources to the allocation. This is the reverse of the
normal (CPU) model. Accelerators such as GPUs that operate on
core-mm-managed memory are interested in this model.
A check is added in sanitize_mpol_flags() to not permit 'prefer_many'
policy to be used for now, and will be removed in later patch after all
implementations for 'prefer_many' are ready, as suggested by Michal Hocko.
[mhocko@kernel.org: suggest to refine policy_node/policy_nodemask handling]
Link: https://lkml.kernel.org/r/1627970362-61305-1-git-send-email-feng.tang@intel.com
Link: https://lore.kernel.org/r/20200630212517.308045-4-ben.widawsky@intel.com
Link: https://lkml.kernel.org/r/1627970362-61305-2-git-send-email-feng.tang@intel.com
Co-developed-by: Ben Widawsky <ben.widawsky@intel.com>
Signed-off-by: Ben Widawsky <ben.widawsky@intel.com>
Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com>
Signed-off-by: Feng Tang <feng.tang@intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Huang Ying <ying.huang@intel.com>b
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-09-02 15:00:06 -07:00
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
2021-06-30 18:51:10 -07:00
|
|
|
|
mm/mempolicy: don't handle MPOL_LOCAL like a fake MPOL_PREFERRED policy
MPOL_LOCAL policy has been setup as a real policy, but it is still handled
like a faked POL_PREFERRED policy with one internal MPOL_F_LOCAL flag bit
set, and there are many places having to judge the real 'prefer' or the
'local' policy, which are quite confusing.
In current code, there are 4 cases that MPOL_LOCAL are used:
1. user specifies 'local' policy
2. user specifies 'prefer' policy, but with empty nodemask
3. system 'default' policy is used
4. 'prefer' policy + valid 'preferred' node with MPOL_F_STATIC_NODES
flag set, and when it is 'rebind' to a nodemask which doesn't contains
the 'preferred' node, it will perform as 'local' policy
So make 'local' a real policy instead of a fake 'prefer' one, and kill
MPOL_F_LOCAL bit, which can greatly reduce the confusion for code reading.
For case 4, the logic of mpol_rebind_preferred() is confusing, as Michal
Hocko pointed out:
: I do believe that rebinding preferred policy is just bogus and it should
: be dropped altogether on the ground that a preference is a mere hint from
: userspace where to start the allocation. Unless I am missing something
: cpusets will be always authoritative for the final placement. The
: preferred node just acts as a starting point and it should be really
: preserved when cpusets changes. Otherwise we have a very subtle behavior
: corner cases.
So dump all the tricky transformation between 'prefer' and 'local', and
just record the new nodemask of rebinding.
[feng.tang@intel.com: fix a problem in mpol_set_nodemask(), per Michal Hocko]
Link: https://lkml.kernel.org/r/1622560492-1294-3-git-send-email-feng.tang@intel.com
[feng.tang@intel.com: refine code and comments of mpol_set_nodemask(), per Michal]
Link: https://lkml.kernel.org/r/20210603081807.GE56979@shbuild999.sh.intel.com
Link: https://lkml.kernel.org/r/1622469956-82897-3-git-send-email-feng.tang@intel.com
Signed-off-by: Feng Tang <feng.tang@intel.com>
Suggested-by: Michal Hocko <mhocko@suse.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ben Widawsky <ben.widawsky@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-30 18:51:00 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2012-01-10 15:08:21 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-01-14 15:18:36 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-01-14 15:18:36 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-03 02:21:34 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-01-14 15:18:36 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-03 02:21:34 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-01-14 15:18:36 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-01-14 15:18:36 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2025-08-27 17:52:45 +00:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2012-10-08 16:29:19 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-10-25 14:16:30 +02:00
|
|
|
|
2023-09-21 15:44:16 +08:00
|
|
|
|
2012-10-25 14:16:30 +02:00
|
|
|
|
2023-09-21 15:44:16 +08:00
|
|
|
|
mm/mempolicy: use numa_node_id() instead of cpu_to_node()
Patch series "Allow migrate on protnone reference with MPOL_PREFERRED_MANY
policy:, v4.
This patchset is to optimize the cross-socket memory access with
MPOL_PREFERRED_MANY policy.
To test this patch we ran the following test on a 3 node system.
Node 0 - 2GB - Tier 1
Node 1 - 11GB - Tier 1
Node 6 - 10GB - Tier 2
Below changes are made to memcached to set the memory policy,
It select Node0 and Node1 as preferred nodes.
#include <numaif.h>
#include <numa.h>
unsigned long nodemask;
int ret;
nodemask = 0x03;
ret = set_mempolicy(MPOL_PREFERRED_MANY | MPOL_F_NUMA_BALANCING,
&nodemask, 10);
/* If MPOL_F_NUMA_BALANCING isn't supported,
* fall back to MPOL_PREFERRED_MANY */
if (ret < 0 && errno == EINVAL){
printf("set mem policy normal\n");
ret = set_mempolicy(MPOL_PREFERRED_MANY, &nodemask, 10);
}
if (ret < 0) {
perror("Failed to call set_mempolicy");
exit(-1);
}
Test Procedure:
===============
1. Make sure memory tiring and demotion are enabled.
2. Start memcached.
# ./memcached -b 100000 -m 204800 -u root -c 1000000 -t 7
-d -s "/tmp/memcached.sock"
3. Run memtier_benchmark to store 3200000 keys.
#./memtier_benchmark -S "/tmp/memcached.sock" --protocol=memcache_binary
--threads=1 --pipeline=1 --ratio=1:0 --key-pattern=S:S --key-minimum=1
--key-maximum=3200000 -n allkeys -c 1 -R -x 1 -d 1024
4. Start a memory eater on node 0 and 1. This will demote all memcached
pages to node 6.
5. Make sure all the memcached pages got demoted to lower tier by reading
/proc/<memcaced PID>/numa_maps.
# cat /proc/2771/numa_maps
---
default anon=1009 dirty=1009 active=0 N6=1009 kernelpagesize_kB=64
default anon=1009 dirty=1009 active=0 N6=1009 kernelpagesize_kB=64
---
6. Kill memory eater.
7. Read the pgpromote_success counter.
8. Start reading the keys by running memtier_benchmark.
#./memtier_benchmark -S "/tmp/memcached.sock" --protocol=memcache_binary
--pipeline=1 --distinct-client-seed --ratio=0:3 --key-pattern=R:R
--key-minimum=1 --key-maximum=3200000 -n allkeys
--threads=64 -c 1 -R -x 6
9. Read the pgpromote_success counter.
Test Results:
=============
Without Patch
------------------
1. pgpromote_success before test
Node 0: pgpromote_success 11
Node 1: pgpromote_success 140974
pgpromote_success after test
Node 0: pgpromote_success 11
Node 1: pgpromote_success 140974
2. Memtier-benchmark result.
AGGREGATED AVERAGE RESULTS (6 runs)
==================================================================
Type Ops/sec Hits/sec Misses/sec Avg. Latency p50 Latency
------------------------------------------------------------------
Sets 0.00 --- --- --- ---
Gets 305792.03 305791.93 0.10 0.18949 0.16700
Waits 0.00 --- --- --- ---
Totals 305792.03 305791.93 0.10 0.18949 0.16700
======================================
p99 Latency p99.9 Latency KB/sec
-------------------------------------
--- --- 0.00
0.44700 1.71100 11542.69
--- --- ---
0.44700 1.71100 11542.69
With Patch
---------------
1. pgpromote_success before test
Node 0: pgpromote_success 5
Node 1: pgpromote_success 89386
pgpromote_success after test
Node 0: pgpromote_success 57895
Node 1: pgpromote_success 141463
2. Memtier-benchmark result.
AGGREGATED AVERAGE RESULTS (6 runs)
====================================================================
Type Ops/sec Hits/sec Misses/sec Avg. Latency p50 Latency
--------------------------------------------------------------------
Sets 0.00 --- --- --- ---
Gets 521942.24 521942.07 0.17 0.11459 0.10300
Waits 0.00 --- --- --- ---
Totals 521942.24 521942.07 0.17 0.11459 0.10300
=======================================
p99 Latency p99.9 Latency KB/sec
---------------------------------------
--- --- 0.00
0.23100 0.31900 19701.68
--- --- ---
0.23100 0.31900 19701.68
Test Result Analysis:
=====================
1. With patch we could observe pages are getting promoted.
2. Memtier-benchmark results shows that, with the patch,
performance has increased more than 50%.
Ops/sec without fix - 305792.03
Ops/sec with fix - 521942.24
This patch (of 2):
Instead of using 'cpu_to_node()', we use 'numa_node_id()', which is
quicker. smp_processor_id is guaranteed to be stable in the
'mpol_misplaced()' function because it is called with ptl held.
lockdep_assert_held was added to ensure that.
No functional change in this patch.
[donettom@linux.ibm.com: add "* @vmf: structure describing the fault" comment]
Link: https://lkml.kernel.org/r/d8b993ea9dccfac0bc3ed61d3a81f4ac5f376e46.1711002865.git.donettom@linux.ibm.com
Link: https://lkml.kernel.org/r/cover.1711373653.git.donettom@linux.ibm.com
Link: https://lkml.kernel.org/r/6059f034f436734b472d066db69676fb3a459864.1711373653.git.donettom@linux.ibm.com
Link: https://lkml.kernel.org/r/cover.1709909210.git.donettom@linux.ibm.com
Link: https://lkml.kernel.org/r/744646531af02cc687cde8ae788fb1779e99d02c.1709909210.git.donettom@linux.ibm.com
Signed-off-by: Aneesh Kumar K.V (IBM) <aneesh.kumar@kernel.org>
Signed-off-by: Donet Tom <donettom@linux.ibm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Feng Tang <feng.tang@intel.com>
Cc: Huang, Ying <ying.huang@intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rik van Riel <riel@surriel.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-08 09:15:37 -06:00
|
|
|
|
2023-09-21 15:44:16 +08:00
|
|
|
|
2012-10-25 14:16:30 +02:00
|
|
|
|
2023-09-21 15:44:16 +08:00
|
|
|
|
2021-04-29 23:01:27 -07:00
|
|
|
|
2012-10-25 14:16:30 +02:00
|
|
|
|
2021-04-29 23:01:27 -07:00
|
|
|
|
2021-09-02 15:00:03 -07:00
|
|
|
|
2023-09-21 15:44:16 +08:00
|
|
|
|
2012-10-25 14:16:30 +02:00
|
|
|
|
mm/mempolicy: use numa_node_id() instead of cpu_to_node()
Patch series "Allow migrate on protnone reference with MPOL_PREFERRED_MANY
policy:, v4.
This patchset is to optimize the cross-socket memory access with
MPOL_PREFERRED_MANY policy.
To test this patch we ran the following test on a 3 node system.
Node 0 - 2GB - Tier 1
Node 1 - 11GB - Tier 1
Node 6 - 10GB - Tier 2
Below changes are made to memcached to set the memory policy,
It select Node0 and Node1 as preferred nodes.
#include <numaif.h>
#include <numa.h>
unsigned long nodemask;
int ret;
nodemask = 0x03;
ret = set_mempolicy(MPOL_PREFERRED_MANY | MPOL_F_NUMA_BALANCING,
&nodemask, 10);
/* If MPOL_F_NUMA_BALANCING isn't supported,
* fall back to MPOL_PREFERRED_MANY */
if (ret < 0 && errno == EINVAL){
printf("set mem policy normal\n");
ret = set_mempolicy(MPOL_PREFERRED_MANY, &nodemask, 10);
}
if (ret < 0) {
perror("Failed to call set_mempolicy");
exit(-1);
}
Test Procedure:
===============
1. Make sure memory tiring and demotion are enabled.
2. Start memcached.
# ./memcached -b 100000 -m 204800 -u root -c 1000000 -t 7
-d -s "/tmp/memcached.sock"
3. Run memtier_benchmark to store 3200000 keys.
#./memtier_benchmark -S "/tmp/memcached.sock" --protocol=memcache_binary
--threads=1 --pipeline=1 --ratio=1:0 --key-pattern=S:S --key-minimum=1
--key-maximum=3200000 -n allkeys -c 1 -R -x 1 -d 1024
4. Start a memory eater on node 0 and 1. This will demote all memcached
pages to node 6.
5. Make sure all the memcached pages got demoted to lower tier by reading
/proc/<memcaced PID>/numa_maps.
# cat /proc/2771/numa_maps
---
default anon=1009 dirty=1009 active=0 N6=1009 kernelpagesize_kB=64
default anon=1009 dirty=1009 active=0 N6=1009 kernelpagesize_kB=64
---
6. Kill memory eater.
7. Read the pgpromote_success counter.
8. Start reading the keys by running memtier_benchmark.
#./memtier_benchmark -S "/tmp/memcached.sock" --protocol=memcache_binary
--pipeline=1 --distinct-client-seed --ratio=0:3 --key-pattern=R:R
--key-minimum=1 --key-maximum=3200000 -n allkeys
--threads=64 -c 1 -R -x 6
9. Read the pgpromote_success counter.
Test Results:
=============
Without Patch
------------------
1. pgpromote_success before test
Node 0: pgpromote_success 11
Node 1: pgpromote_success 140974
pgpromote_success after test
Node 0: pgpromote_success 11
Node 1: pgpromote_success 140974
2. Memtier-benchmark result.
AGGREGATED AVERAGE RESULTS (6 runs)
==================================================================
Type Ops/sec Hits/sec Misses/sec Avg. Latency p50 Latency
------------------------------------------------------------------
Sets 0.00 --- --- --- ---
Gets 305792.03 305791.93 0.10 0.18949 0.16700
Waits 0.00 --- --- --- ---
Totals 305792.03 305791.93 0.10 0.18949 0.16700
======================================
p99 Latency p99.9 Latency KB/sec
-------------------------------------
--- --- 0.00
0.44700 1.71100 11542.69
--- --- ---
0.44700 1.71100 11542.69
With Patch
---------------
1. pgpromote_success before test
Node 0: pgpromote_success 5
Node 1: pgpromote_success 89386
pgpromote_success after test
Node 0: pgpromote_success 57895
Node 1: pgpromote_success 141463
2. Memtier-benchmark result.
AGGREGATED AVERAGE RESULTS (6 runs)
====================================================================
Type Ops/sec Hits/sec Misses/sec Avg. Latency p50 Latency
--------------------------------------------------------------------
Sets 0.00 --- --- --- ---
Gets 521942.24 521942.07 0.17 0.11459 0.10300
Waits 0.00 --- --- --- ---
Totals 521942.24 521942.07 0.17 0.11459 0.10300
=======================================
p99 Latency p99.9 Latency KB/sec
---------------------------------------
--- --- 0.00
0.23100 0.31900 19701.68
--- --- ---
0.23100 0.31900 19701.68
Test Result Analysis:
=====================
1. With patch we could observe pages are getting promoted.
2. Memtier-benchmark results shows that, with the patch,
performance has increased more than 50%.
Ops/sec without fix - 305792.03
Ops/sec with fix - 521942.24
This patch (of 2):
Instead of using 'cpu_to_node()', we use 'numa_node_id()', which is
quicker. smp_processor_id is guaranteed to be stable in the
'mpol_misplaced()' function because it is called with ptl held.
lockdep_assert_held was added to ensure that.
No functional change in this patch.
[donettom@linux.ibm.com: add "* @vmf: structure describing the fault" comment]
Link: https://lkml.kernel.org/r/d8b993ea9dccfac0bc3ed61d3a81f4ac5f376e46.1711002865.git.donettom@linux.ibm.com
Link: https://lkml.kernel.org/r/cover.1711373653.git.donettom@linux.ibm.com
Link: https://lkml.kernel.org/r/6059f034f436734b472d066db69676fb3a459864.1711373653.git.donettom@linux.ibm.com
Link: https://lkml.kernel.org/r/cover.1709909210.git.donettom@linux.ibm.com
Link: https://lkml.kernel.org/r/744646531af02cc687cde8ae788fb1779e99d02c.1709909210.git.donettom@linux.ibm.com
Signed-off-by: Aneesh Kumar K.V (IBM) <aneesh.kumar@kernel.org>
Signed-off-by: Donet Tom <donettom@linux.ibm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Feng Tang <feng.tang@intel.com>
Cc: Huang, Ying <ying.huang@intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rik van Riel <riel@surriel.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-08 09:15:37 -06:00
|
|
|
|
2023-09-21 15:44:16 +08:00
|
|
|
|
2012-10-25 14:16:30 +02:00
|
|
|
|
|
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
2016-05-19 17:14:10 -07:00
|
|
|
|
2023-09-21 15:44:16 +08:00
|
|
|
|
mm/mempolicy: use numa_node_id() instead of cpu_to_node()
Patch series "Allow migrate on protnone reference with MPOL_PREFERRED_MANY
policy:, v4.
This patchset is to optimize the cross-socket memory access with
MPOL_PREFERRED_MANY policy.
To test this patch we ran the following test on a 3 node system.
Node 0 - 2GB - Tier 1
Node 1 - 11GB - Tier 1
Node 6 - 10GB - Tier 2
Below changes are made to memcached to set the memory policy,
It select Node0 and Node1 as preferred nodes.
#include <numaif.h>
#include <numa.h>
unsigned long nodemask;
int ret;
nodemask = 0x03;
ret = set_mempolicy(MPOL_PREFERRED_MANY | MPOL_F_NUMA_BALANCING,
&nodemask, 10);
/* If MPOL_F_NUMA_BALANCING isn't supported,
* fall back to MPOL_PREFERRED_MANY */
if (ret < 0 && errno == EINVAL){
printf("set mem policy normal\n");
ret = set_mempolicy(MPOL_PREFERRED_MANY, &nodemask, 10);
}
if (ret < 0) {
perror("Failed to call set_mempolicy");
exit(-1);
}
Test Procedure:
===============
1. Make sure memory tiring and demotion are enabled.
2. Start memcached.
# ./memcached -b 100000 -m 204800 -u root -c 1000000 -t 7
-d -s "/tmp/memcached.sock"
3. Run memtier_benchmark to store 3200000 keys.
#./memtier_benchmark -S "/tmp/memcached.sock" --protocol=memcache_binary
--threads=1 --pipeline=1 --ratio=1:0 --key-pattern=S:S --key-minimum=1
--key-maximum=3200000 -n allkeys -c 1 -R -x 1 -d 1024
4. Start a memory eater on node 0 and 1. This will demote all memcached
pages to node 6.
5. Make sure all the memcached pages got demoted to lower tier by reading
/proc/<memcaced PID>/numa_maps.
# cat /proc/2771/numa_maps
---
default anon=1009 dirty=1009 active=0 N6=1009 kernelpagesize_kB=64
default anon=1009 dirty=1009 active=0 N6=1009 kernelpagesize_kB=64
---
6. Kill memory eater.
7. Read the pgpromote_success counter.
8. Start reading the keys by running memtier_benchmark.
#./memtier_benchmark -S "/tmp/memcached.sock" --protocol=memcache_binary
--pipeline=1 --distinct-client-seed --ratio=0:3 --key-pattern=R:R
--key-minimum=1 --key-maximum=3200000 -n allkeys
--threads=64 -c 1 -R -x 6
9. Read the pgpromote_success counter.
Test Results:
=============
Without Patch
------------------
1. pgpromote_success before test
Node 0: pgpromote_success 11
Node 1: pgpromote_success 140974
pgpromote_success after test
Node 0: pgpromote_success 11
Node 1: pgpromote_success 140974
2. Memtier-benchmark result.
AGGREGATED AVERAGE RESULTS (6 runs)
==================================================================
Type Ops/sec Hits/sec Misses/sec Avg. Latency p50 Latency
------------------------------------------------------------------
Sets 0.00 --- --- --- ---
Gets 305792.03 305791.93 0.10 0.18949 0.16700
Waits 0.00 --- --- --- ---
Totals 305792.03 305791.93 0.10 0.18949 0.16700
======================================
p99 Latency p99.9 Latency KB/sec
-------------------------------------
--- --- 0.00
0.44700 1.71100 11542.69
--- --- ---
0.44700 1.71100 11542.69
With Patch
---------------
1. pgpromote_success before test
Node 0: pgpromote_success 5
Node 1: pgpromote_success 89386
pgpromote_success after test
Node 0: pgpromote_success 57895
Node 1: pgpromote_success 141463
2. Memtier-benchmark result.
AGGREGATED AVERAGE RESULTS (6 runs)
====================================================================
Type Ops/sec Hits/sec Misses/sec Avg. Latency p50 Latency
--------------------------------------------------------------------
Sets 0.00 --- --- --- ---
Gets 521942.24 521942.07 0.17 0.11459 0.10300
Waits 0.00 --- --- --- ---
Totals 521942.24 521942.07 0.17 0.11459 0.10300
=======================================
p99 Latency p99.9 Latency KB/sec
---------------------------------------
--- --- 0.00
0.23100 0.31900 19701.68
--- --- ---
0.23100 0.31900 19701.68
Test Result Analysis:
=====================
1. With patch we could observe pages are getting promoted.
2. Memtier-benchmark results shows that, with the patch,
performance has increased more than 50%.
Ops/sec without fix - 305792.03
Ops/sec with fix - 521942.24
This patch (of 2):
Instead of using 'cpu_to_node()', we use 'numa_node_id()', which is
quicker. smp_processor_id is guaranteed to be stable in the
'mpol_misplaced()' function because it is called with ptl held.
lockdep_assert_held was added to ensure that.
No functional change in this patch.
[donettom@linux.ibm.com: add "* @vmf: structure describing the fault" comment]
Link: https://lkml.kernel.org/r/d8b993ea9dccfac0bc3ed61d3a81f4ac5f376e46.1711002865.git.donettom@linux.ibm.com
Link: https://lkml.kernel.org/r/cover.1711373653.git.donettom@linux.ibm.com
Link: https://lkml.kernel.org/r/6059f034f436734b472d066db69676fb3a459864.1711373653.git.donettom@linux.ibm.com
Link: https://lkml.kernel.org/r/cover.1709909210.git.donettom@linux.ibm.com
Link: https://lkml.kernel.org/r/744646531af02cc687cde8ae788fb1779e99d02c.1709909210.git.donettom@linux.ibm.com
Signed-off-by: Aneesh Kumar K.V (IBM) <aneesh.kumar@kernel.org>
Signed-off-by: Donet Tom <donettom@linux.ibm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Feng Tang <feng.tang@intel.com>
Cc: Huang, Ying <ying.huang@intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rik van Riel <riel@surriel.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-08 09:15:37 -06:00
|
|
|
|
2013-10-07 11:29:20 +01:00
|
|
|
|
mm/mempolicy: use numa_node_id() instead of cpu_to_node()
Patch series "Allow migrate on protnone reference with MPOL_PREFERRED_MANY
policy:, v4.
This patchset is to optimize the cross-socket memory access with
MPOL_PREFERRED_MANY policy.
To test this patch we ran the following test on a 3 node system.
Node 0 - 2GB - Tier 1
Node 1 - 11GB - Tier 1
Node 6 - 10GB - Tier 2
Below changes are made to memcached to set the memory policy,
It select Node0 and Node1 as preferred nodes.
#include <numaif.h>
#include <numa.h>
unsigned long nodemask;
int ret;
nodemask = 0x03;
ret = set_mempolicy(MPOL_PREFERRED_MANY | MPOL_F_NUMA_BALANCING,
&nodemask, 10);
/* If MPOL_F_NUMA_BALANCING isn't supported,
* fall back to MPOL_PREFERRED_MANY */
if (ret < 0 && errno == EINVAL){
printf("set mem policy normal\n");
ret = set_mempolicy(MPOL_PREFERRED_MANY, &nodemask, 10);
}
if (ret < 0) {
perror("Failed to call set_mempolicy");
exit(-1);
}
Test Procedure:
===============
1. Make sure memory tiring and demotion are enabled.
2. Start memcached.
# ./memcached -b 100000 -m 204800 -u root -c 1000000 -t 7
-d -s "/tmp/memcached.sock"
3. Run memtier_benchmark to store 3200000 keys.
#./memtier_benchmark -S "/tmp/memcached.sock" --protocol=memcache_binary
--threads=1 --pipeline=1 --ratio=1:0 --key-pattern=S:S --key-minimum=1
--key-maximum=3200000 -n allkeys -c 1 -R -x 1 -d 1024
4. Start a memory eater on node 0 and 1. This will demote all memcached
pages to node 6.
5. Make sure all the memcached pages got demoted to lower tier by reading
/proc/<memcaced PID>/numa_maps.
# cat /proc/2771/numa_maps
---
default anon=1009 dirty=1009 active=0 N6=1009 kernelpagesize_kB=64
default anon=1009 dirty=1009 active=0 N6=1009 kernelpagesize_kB=64
---
6. Kill memory eater.
7. Read the pgpromote_success counter.
8. Start reading the keys by running memtier_benchmark.
#./memtier_benchmark -S "/tmp/memcached.sock" --protocol=memcache_binary
--pipeline=1 --distinct-client-seed --ratio=0:3 --key-pattern=R:R
--key-minimum=1 --key-maximum=3200000 -n allkeys
--threads=64 -c 1 -R -x 6
9. Read the pgpromote_success counter.
Test Results:
=============
Without Patch
------------------
1. pgpromote_success before test
Node 0: pgpromote_success 11
Node 1: pgpromote_success 140974
pgpromote_success after test
Node 0: pgpromote_success 11
Node 1: pgpromote_success 140974
2. Memtier-benchmark result.
AGGREGATED AVERAGE RESULTS (6 runs)
==================================================================
Type Ops/sec Hits/sec Misses/sec Avg. Latency p50 Latency
------------------------------------------------------------------
Sets 0.00 --- --- --- ---
Gets 305792.03 305791.93 0.10 0.18949 0.16700
Waits 0.00 --- --- --- ---
Totals 305792.03 305791.93 0.10 0.18949 0.16700
======================================
p99 Latency p99.9 Latency KB/sec
-------------------------------------
--- --- 0.00
0.44700 1.71100 11542.69
--- --- ---
0.44700 1.71100 11542.69
With Patch
---------------
1. pgpromote_success before test
Node 0: pgpromote_success 5
Node 1: pgpromote_success 89386
pgpromote_success after test
Node 0: pgpromote_success 57895
Node 1: pgpromote_success 141463
2. Memtier-benchmark result.
AGGREGATED AVERAGE RESULTS (6 runs)
====================================================================
Type Ops/sec Hits/sec Misses/sec Avg. Latency p50 Latency
--------------------------------------------------------------------
Sets 0.00 --- --- --- ---
Gets 521942.24 521942.07 0.17 0.11459 0.10300
Waits 0.00 --- --- --- ---
Totals 521942.24 521942.07 0.17 0.11459 0.10300
=======================================
p99 Latency p99.9 Latency KB/sec
---------------------------------------
--- --- 0.00
0.23100 0.31900 19701.68
--- --- ---
0.23100 0.31900 19701.68
Test Result Analysis:
=====================
1. With patch we could observe pages are getting promoted.
2. Memtier-benchmark results shows that, with the patch,
performance has increased more than 50%.
Ops/sec without fix - 305792.03
Ops/sec with fix - 521942.24
This patch (of 2):
Instead of using 'cpu_to_node()', we use 'numa_node_id()', which is
quicker. smp_processor_id is guaranteed to be stable in the
'mpol_misplaced()' function because it is called with ptl held.
lockdep_assert_held was added to ensure that.
No functional change in this patch.
[donettom@linux.ibm.com: add "* @vmf: structure describing the fault" comment]
Link: https://lkml.kernel.org/r/d8b993ea9dccfac0bc3ed61d3a81f4ac5f376e46.1711002865.git.donettom@linux.ibm.com
Link: https://lkml.kernel.org/r/cover.1711373653.git.donettom@linux.ibm.com
Link: https://lkml.kernel.org/r/6059f034f436734b472d066db69676fb3a459864.1711373653.git.donettom@linux.ibm.com
Link: https://lkml.kernel.org/r/cover.1709909210.git.donettom@linux.ibm.com
Link: https://lkml.kernel.org/r/744646531af02cc687cde8ae788fb1779e99d02c.1709909210.git.donettom@linux.ibm.com
Signed-off-by: Aneesh Kumar K.V (IBM) <aneesh.kumar@kernel.org>
Signed-off-by: Donet Tom <donettom@linux.ibm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Feng Tang <feng.tang@intel.com>
Cc: Huang, Ying <ying.huang@intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rik van Riel <riel@surriel.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-08 09:15:37 -06:00
|
|
|
|
2019-03-05 15:42:58 -08:00
|
|
|
|
2021-09-02 15:00:03 -07:00
|
|
|
|
2012-10-25 14:16:30 +02:00
|
|
|
|
mm/mempolicy: use numa_node_id() instead of cpu_to_node()
Patch series "Allow migrate on protnone reference with MPOL_PREFERRED_MANY
policy:, v4.
This patchset is to optimize the cross-socket memory access with
MPOL_PREFERRED_MANY policy.
To test this patch we ran the following test on a 3 node system.
Node 0 - 2GB - Tier 1
Node 1 - 11GB - Tier 1
Node 6 - 10GB - Tier 2
Below changes are made to memcached to set the memory policy,
It select Node0 and Node1 as preferred nodes.
#include <numaif.h>
#include <numa.h>
unsigned long nodemask;
int ret;
nodemask = 0x03;
ret = set_mempolicy(MPOL_PREFERRED_MANY | MPOL_F_NUMA_BALANCING,
&nodemask, 10);
/* If MPOL_F_NUMA_BALANCING isn't supported,
* fall back to MPOL_PREFERRED_MANY */
if (ret < 0 && errno == EINVAL){
printf("set mem policy normal\n");
ret = set_mempolicy(MPOL_PREFERRED_MANY, &nodemask, 10);
}
if (ret < 0) {
perror("Failed to call set_mempolicy");
exit(-1);
}
Test Procedure:
===============
1. Make sure memory tiring and demotion are enabled.
2. Start memcached.
# ./memcached -b 100000 -m 204800 -u root -c 1000000 -t 7
-d -s "/tmp/memcached.sock"
3. Run memtier_benchmark to store 3200000 keys.
#./memtier_benchmark -S "/tmp/memcached.sock" --protocol=memcache_binary
--threads=1 --pipeline=1 --ratio=1:0 --key-pattern=S:S --key-minimum=1
--key-maximum=3200000 -n allkeys -c 1 -R -x 1 -d 1024
4. Start a memory eater on node 0 and 1. This will demote all memcached
pages to node 6.
5. Make sure all the memcached pages got demoted to lower tier by reading
/proc/<memcaced PID>/numa_maps.
# cat /proc/2771/numa_maps
---
default anon=1009 dirty=1009 active=0 N6=1009 kernelpagesize_kB=64
default anon=1009 dirty=1009 active=0 N6=1009 kernelpagesize_kB=64
---
6. Kill memory eater.
7. Read the pgpromote_success counter.
8. Start reading the keys by running memtier_benchmark.
#./memtier_benchmark -S "/tmp/memcached.sock" --protocol=memcache_binary
--pipeline=1 --distinct-client-seed --ratio=0:3 --key-pattern=R:R
--key-minimum=1 --key-maximum=3200000 -n allkeys
--threads=64 -c 1 -R -x 6
9. Read the pgpromote_success counter.
Test Results:
=============
Without Patch
------------------
1. pgpromote_success before test
Node 0: pgpromote_success 11
Node 1: pgpromote_success 140974
pgpromote_success after test
Node 0: pgpromote_success 11
Node 1: pgpromote_success 140974
2. Memtier-benchmark result.
AGGREGATED AVERAGE RESULTS (6 runs)
==================================================================
Type Ops/sec Hits/sec Misses/sec Avg. Latency p50 Latency
------------------------------------------------------------------
Sets 0.00 --- --- --- ---
Gets 305792.03 305791.93 0.10 0.18949 0.16700
Waits 0.00 --- --- --- ---
Totals 305792.03 305791.93 0.10 0.18949 0.16700
======================================
p99 Latency p99.9 Latency KB/sec
-------------------------------------
--- --- 0.00
0.44700 1.71100 11542.69
--- --- ---
0.44700 1.71100 11542.69
With Patch
---------------
1. pgpromote_success before test
Node 0: pgpromote_success 5
Node 1: pgpromote_success 89386
pgpromote_success after test
Node 0: pgpromote_success 57895
Node 1: pgpromote_success 141463
2. Memtier-benchmark result.
AGGREGATED AVERAGE RESULTS (6 runs)
====================================================================
Type Ops/sec Hits/sec Misses/sec Avg. Latency p50 Latency
--------------------------------------------------------------------
Sets 0.00 --- --- --- ---
Gets 521942.24 521942.07 0.17 0.11459 0.10300
Waits 0.00 --- --- --- ---
Totals 521942.24 521942.07 0.17 0.11459 0.10300
=======================================
p99 Latency p99.9 Latency KB/sec
---------------------------------------
--- --- 0.00
0.23100 0.31900 19701.68
--- --- ---
0.23100 0.31900 19701.68
Test Result Analysis:
=====================
1. With patch we could observe pages are getting promoted.
2. Memtier-benchmark results shows that, with the patch,
performance has increased more than 50%.
Ops/sec without fix - 305792.03
Ops/sec with fix - 521942.24
This patch (of 2):
Instead of using 'cpu_to_node()', we use 'numa_node_id()', which is
quicker. smp_processor_id is guaranteed to be stable in the
'mpol_misplaced()' function because it is called with ptl held.
lockdep_assert_held was added to ensure that.
No functional change in this patch.
[donettom@linux.ibm.com: add "* @vmf: structure describing the fault" comment]
Link: https://lkml.kernel.org/r/d8b993ea9dccfac0bc3ed61d3a81f4ac5f376e46.1711002865.git.donettom@linux.ibm.com
Link: https://lkml.kernel.org/r/cover.1711373653.git.donettom@linux.ibm.com
Link: https://lkml.kernel.org/r/6059f034f436734b472d066db69676fb3a459864.1711373653.git.donettom@linux.ibm.com
Link: https://lkml.kernel.org/r/cover.1709909210.git.donettom@linux.ibm.com
Link: https://lkml.kernel.org/r/744646531af02cc687cde8ae788fb1779e99d02c.1709909210.git.donettom@linux.ibm.com
Signed-off-by: Aneesh Kumar K.V (IBM) <aneesh.kumar@kernel.org>
Signed-off-by: Donet Tom <donettom@linux.ibm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Feng Tang <feng.tang@intel.com>
Cc: Huang, Ying <ying.huang@intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rik van Riel <riel@surriel.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-08 09:15:37 -06:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
2012-10-25 14:16:30 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: alloc_pages_mpol() for NUMA policy without vma
Shrink shmem's stack usage by eliminating the pseudo-vma from its folio
allocation. alloc_pages_mpol(gfp, order, pol, ilx, nid) becomes the
principal actor for passing mempolicy choice down to __alloc_pages(),
rather than vma_alloc_folio(gfp, order, vma, addr, hugepage).
vma_alloc_folio() and alloc_pages() remain, but as wrappers around
alloc_pages_mpol(). alloc_pages_bulk_*() untouched, except to provide the
additional args to policy_nodemask(), which subsumes policy_node().
Cleanup throughout, cutting out some unhelpful "helpers".
It would all be much simpler without MPOL_INTERLEAVE, but that adds a
dynamic to the constant mpol: complicated by v3.6 commit 09c231cb8bfd
("tmpfs: distribute interleave better across nodes"), which added ino bias
to the interleave, hidden from mm/mempolicy.c until this commit.
Hence "ilx" throughout, the "interleave index". Originally I thought it
could be done just with nid, but that's wrong: the nodemask may come from
the shared policy layer below a shmem vma, or it may come from the task
layer above a shmem vma; and without the final nodemask then nodeid cannot
be decided. And how ilx is applied depends also on page order.
The interleave index is almost always irrelevant unless MPOL_INTERLEAVE:
with one exception in alloc_pages_mpol(), where the NO_INTERLEAVE_INDEX
passed down from vma-less alloc_pages() is also used as hint not to use
THP-style hugepage allocation - to avoid the overhead of a hugepage arg
(though I don't understand why we never just added a GFP bit for THP - if
it actually needs a different allocation strategy from other pages of the
same order). vma_alloc_folio() still carries its hugepage arg here, but
it is not used, and should be removed when agreed.
get_vma_policy() no longer allows a NULL vma: over time I believe we've
eradicated all the places which used to need it e.g. swapoff and madvise
used to pass NULL vma to read_swap_cache_async(), but now know the vma.
[hughd@google.com: handle NULL mpol being passed to __read_swap_cache_async()]
Link: https://lkml.kernel.org/r/ea419956-4751-0102-21f7-9c93cb957892@google.com
Link: https://lkml.kernel.org/r/74e34633-6060-f5e3-aee-7040d43f2e93@google.com
Link: https://lkml.kernel.org/r/1738368e-bac0-fd11-ed7f-b87142a939fe@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Tejun heo <tj@kernel.org>
Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Domenico Cerasuolo <mimmocerasuolo@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-19 13:39:08 -07:00
|
|
|
|
2012-10-25 14:16:30 +02:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-10-25 14:16:30 +02:00
|
|
|
|
mm/mempolicy: add MPOL_PREFERRED_MANY for multiple preferred nodes
Patch series "Introduce multi-preference mempolicy", v7.
This patch series introduces the concept of the MPOL_PREFERRED_MANY
mempolicy. This mempolicy mode can be used with either the
set_mempolicy(2) or mbind(2) interfaces. Like the MPOL_PREFERRED
interface, it allows an application to set a preference for nodes which
will fulfil memory allocation requests. Unlike the MPOL_PREFERRED mode,
it takes a set of nodes. Like the MPOL_BIND interface, it works over a
set of nodes. Unlike MPOL_BIND, it will not cause a SIGSEGV or invoke the
OOM killer if those preferred nodes are not available.
Along with these patches are patches for libnuma, numactl, numademo, and
memhog. They still need some polish, but can be found here:
https://gitlab.com/bwidawsk/numactl/-/tree/prefer-many It allows new
usage: `numactl -P 0,3,4`
The goal of the new mode is to enable some use-cases when using tiered memory
usage models which I've lovingly named.
1a. The Hare - The interconnect is fast enough to meet bandwidth and
latency requirements allowing preference to be given to all nodes with
"fast" memory.
1b. The Indiscriminate Hare - An application knows it wants fast
memory (or perhaps slow memory), but doesn't care which node it runs
on. The application can prefer a set of nodes and then xpu bind to
the local node (cpu, accelerator, etc). This reverses the nodes are
chosen today where the kernel attempts to use local memory to the CPU
whenever possible. This will attempt to use the local accelerator to
the memory.
2. The Tortoise - The administrator (or the application itself) is
aware it only needs slow memory, and so can prefer that.
Much of this is almost achievable with the bind interface, but the bind
interface suffers from an inability to fallback to another set of nodes if
binding fails to all nodes in the nodemask.
Like MPOL_BIND a nodemask is given. Inherently this removes ordering from the
preference.
> /* Set first two nodes as preferred in an 8 node system. */
> const unsigned long nodes = 0x3
> set_mempolicy(MPOL_PREFER_MANY, &nodes, 8);
> /* Mimic interleave policy, but have fallback *.
> const unsigned long nodes = 0xaa
> set_mempolicy(MPOL_PREFER_MANY, &nodes, 8);
Some internal discussion took place around the interface. There are two
alternatives which we have discussed, plus one I stuck in:
1. Ordered list of nodes. Currently it's believed that the added
complexity is nod needed for expected usecases.
2. A flag for bind to allow falling back to other nodes. This
confuses the notion of binding and is less flexible than the current
solution.
3. Create flags or new modes that helps with some ordering. This
offers both a friendlier API as well as a solution for more customized
usage. It's unknown if it's worth the complexity to support this.
Here is sample code for how this might work:
> // Prefer specific nodes for some something wacky
> set_mempolicy(MPOL_PREFER_MANY, 0x17c, 1024);
>
> // Default
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_SOCKET, NULL, 0);
> // which is the same as
> set_mempolicy(MPOL_DEFAULT, NULL, 0);
>
> // The Hare
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE, NULL, 0);
>
> // The Tortoise
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE_REV, NULL, 0);
>
> // Prefer the fast memory of the first two sockets
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE, -1, 2);
>
This patch (of 5):
The NUMA APIs currently allow passing in a "preferred node" as a single
bit set in a nodemask. If more than one bit it set, bits after the first
are ignored.
This single node is generally OK for location-based NUMA where memory
being allocated will eventually be operated on by a single CPU. However,
in systems with multiple memory types, folks want to target a *type* of
memory instead of a location. For instance, someone might want some
high-bandwidth memory but do not care about the CPU next to which it is
allocated. Or, they want a cheap, high capacity allocation and want to
target all NUMA nodes which have persistent memory in volatile mode. In
both of these cases, the application wants to target a *set* of nodes, but
does not want strict MPOL_BIND behavior as that could lead to OOM killer
or SIGSEGV.
So add MPOL_PREFERRED_MANY policy to support the multiple preferred nodes
requirement. This is not a pie-in-the-sky dream for an API. This was a
response to a specific ask of more than one group at Intel. Specifically:
1. There are existing libraries that target memory types such as
https://github.com/memkind/memkind. These are known to suffer from
SIGSEGV's when memory is low on targeted memory "kinds" that span more
than one node. The MCDRAM on a Xeon Phi in "Cluster on Die" mode is an
example of this.
2. Volatile-use persistent memory users want to have a memory policy
which is targeted at either "cheap and slow" (PMEM) or "expensive and
fast" (DRAM). However, they do not want to experience allocation
failures when the targeted type is unavailable.
3. Allocate-then-run. Generally, we let the process scheduler decide
on which physical CPU to run a task. That location provides a default
allocation policy, and memory availability is not generally considered
when placing tasks. For situations where memory is valuable and
constrained, some users want to allocate memory first, *then* allocate
close compute resources to the allocation. This is the reverse of the
normal (CPU) model. Accelerators such as GPUs that operate on
core-mm-managed memory are interested in this model.
A check is added in sanitize_mpol_flags() to not permit 'prefer_many'
policy to be used for now, and will be removed in later patch after all
implementations for 'prefer_many' are ready, as suggested by Michal Hocko.
[mhocko@kernel.org: suggest to refine policy_node/policy_nodemask handling]
Link: https://lkml.kernel.org/r/1627970362-61305-1-git-send-email-feng.tang@intel.com
Link: https://lore.kernel.org/r/20200630212517.308045-4-ben.widawsky@intel.com
Link: https://lkml.kernel.org/r/1627970362-61305-2-git-send-email-feng.tang@intel.com
Co-developed-by: Ben Widawsky <ben.widawsky@intel.com>
Signed-off-by: Ben Widawsky <ben.widawsky@intel.com>
Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com>
Signed-off-by: Feng Tang <feng.tang@intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Huang Ying <ying.huang@intel.com>b
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-09-02 15:00:06 -07:00
|
|
|
|
|
|
|
|
|
2021-06-30 18:51:10 -07:00
|
|
|
|
mm/mempolicy: don't handle MPOL_LOCAL like a fake MPOL_PREFERRED policy
MPOL_LOCAL policy has been setup as a real policy, but it is still handled
like a faked POL_PREFERRED policy with one internal MPOL_F_LOCAL flag bit
set, and there are many places having to judge the real 'prefer' or the
'local' policy, which are quite confusing.
In current code, there are 4 cases that MPOL_LOCAL are used:
1. user specifies 'local' policy
2. user specifies 'prefer' policy, but with empty nodemask
3. system 'default' policy is used
4. 'prefer' policy + valid 'preferred' node with MPOL_F_STATIC_NODES
flag set, and when it is 'rebind' to a nodemask which doesn't contains
the 'preferred' node, it will perform as 'local' policy
So make 'local' a real policy instead of a fake 'prefer' one, and kill
MPOL_F_LOCAL bit, which can greatly reduce the confusion for code reading.
For case 4, the logic of mpol_rebind_preferred() is confusing, as Michal
Hocko pointed out:
: I do believe that rebinding preferred policy is just bogus and it should
: be dropped altogether on the ground that a preference is a mere hint from
: userspace where to start the allocation. Unless I am missing something
: cpusets will be always authoritative for the final placement. The
: preferred node just acts as a starting point and it should be really
: preserved when cpusets changes. Otherwise we have a very subtle behavior
: corner cases.
So dump all the tricky transformation between 'prefer' and 'local', and
just record the new nodemask of rebinding.
[feng.tang@intel.com: fix a problem in mpol_set_nodemask(), per Michal Hocko]
Link: https://lkml.kernel.org/r/1622560492-1294-3-git-send-email-feng.tang@intel.com
[feng.tang@intel.com: refine code and comments of mpol_set_nodemask(), per Michal]
Link: https://lkml.kernel.org/r/20210603081807.GE56979@shbuild999.sh.intel.com
Link: https://lkml.kernel.org/r/1622469956-82897-3-git-send-email-feng.tang@intel.com
Signed-off-by: Feng Tang <feng.tang@intel.com>
Suggested-by: Michal Hocko <mhocko@suse.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ben Widawsky <ben.widawsky@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-30 18:51:00 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-10-25 14:16:30 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/numa_balancing: allow migrate on protnone reference with MPOL_PREFERRED_MANY policy
commit bda420b98505 ("numa balancing: migrate on fault among multiple
bound nodes") added support for migrate on protnone reference with
MPOL_BIND memory policy. This allowed numa fault migration when the
executing node is part of the policy mask for MPOL_BIND. This patch
extends migration support to MPOL_PREFERRED_MANY policy.
Currently, we cannot specify MPOL_PREFERRED_MANY with the mempolicy flag
MPOL_F_NUMA_BALANCING. This causes issues when we want to use
NUMA_BALANCING_MEMORY_TIERING. To effectively use the slow memory tier,
the kernel should not allocate pages from the slower memory tier via
allocation control zonelist fallback. Instead, we should move cold pages
from the faster memory node via memory demotion. For a page allocation,
kswapd is only woken up after we try to allocate pages from all nodes in
the allocation zone list. This implies that, without using memory
policies, we will end up allocating hot pages in the slower memory tier.
MPOL_PREFERRED_MANY was added by commit b27abaccf8e8 ("mm/mempolicy: add
MPOL_PREFERRED_MANY for multiple preferred nodes") to allow better
allocation control when we have memory tiers in the system. With
MPOL_PREFERRED_MANY, the user can use a policy node mask consisting only
of faster memory nodes. When we fail to allocate pages from the faster
memory node, kswapd would be woken up, allowing demotion of cold pages to
slower memory nodes.
With the current kernel, such usage of memory policies implies we can't do
page promotion from a slower memory tier to a faster memory tier using
numa fault. This patch fixes this issue.
For MPOL_PREFERRED_MANY, if the executing node is in the policy node mask,
we allow numa migration to the executing nodes. If the executing node is
not in the policy node mask, we do not allow numa migration.
Example:
On a 2-sockets system, NUMA node N0, N1 and N2 are in socket 0,
N3 in socket 1. N0, N1 and N3 have fast memory and CPU, while
N2 has slow memory and no CPU. For a workload, we may use
MPOL_PREFERRED_MANY with nodemask N0 and N1 set because the workload
runs on CPUs of socket 0 at most times. Then, even if the workload
runs on CPUs of N3 occasionally, we will not try to migrate the workload
pages from N2 to N3 because users may want to avoid cross-socket access
as much as possible in the long term.
In below table, Process is the Process executing node and
Curr Loc Pgs is the numa node where page present(folio node)
===========================================================
Process Policy Curr Loc Pgs Observation
-----------------------------------------------------------
N0 N0 N1 N1 Pages Migrated from N1 to N0
N0 N0 N1 N2 Pages Migrated from N2 to N0
N0 N0 N1 N3 Pages Migrated from N3 to N0
N3 N0 N1 N0 Pages NOT Migrated to N3
N3 N0 N1 N1 Pages NOT Migrated to N3
N3 N0 N1 N2 Pages NOT Migrated to N3
------------------------------------------------------------
Link: https://lkml.kernel.org/r/158acc57319129aa46d50fd64c9330f3e7c7b4bf.1711373653.git.donettom@linux.ibm.com
Link: https://lkml.kernel.org/r/369d6a58758396335fd1176d97bbca4e7730d75a.1709909210.git.donettom@linux.ibm.com
Signed-off-by: Aneesh Kumar K.V (IBM) <aneesh.kumar@kernel.org>
Signed-off-by: Donet Tom <donettom@linux.ibm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Feng Tang <feng.tang@intel.com>
Cc: Huang, Ying <ying.huang@intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rik van Riel <riel@surriel.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-08 09:15:38 -06:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
numa balancing: migrate on fault among multiple bound nodes
Now, NUMA balancing can only optimize the page placement among the NUMA
nodes if the default memory policy is used. Because the memory policy
specified explicitly should take precedence. But this seems too strict in
some situations. For example, on a system with 4 NUMA nodes, if the
memory of an application is bound to the node 0 and 1, NUMA balancing can
potentially migrate the pages between the node 0 and 1 to reduce
cross-node accessing without breaking the explicit memory binding policy.
So in this patch, we add MPOL_F_NUMA_BALANCING mode flag to
set_mempolicy() when mode is MPOL_BIND. With the flag specified, NUMA
balancing will be enabled within the thread to optimize the page placement
within the constrains of the specified memory binding policy. With the
newly added flag, the NUMA balancing control mechanism becomes,
- sysctl knob numa_balancing can enable/disable the NUMA balancing
globally.
- even if sysctl numa_balancing is enabled, the NUMA balancing will be
disabled for the memory areas or applications with the explicit
memory policy by default.
- MPOL_F_NUMA_BALANCING can be used to enable the NUMA balancing for
the applications when specifying the explicit memory policy
(MPOL_BIND).
Various page placement optimization based on the NUMA balancing can be
done with these flags. As the first step, in this patch, if the memory of
the application is bound to multiple nodes (MPOL_BIND), and in the hint
page fault handler the accessing node are in the policy nodemask, the page
will be tried to be migrated to the accessing node to reduce the
cross-node accessing.
If the newly added MPOL_F_NUMA_BALANCING flag is specified by an
application on an old kernel version without its support, set_mempolicy()
will return -1 and errno will be set to EINVAL. The application can use
this behavior to run on both old and new kernel versions.
And if the MPOL_F_NUMA_BALANCING flag is specified for the mode other than
MPOL_BIND, set_mempolicy() will return -1 and errno will be set to EINVAL
as before. Because we don't support optimization based on the NUMA
balancing for these modes.
In the previous version of the patch, we tried to reuse MPOL_MF_LAZY for
mbind(). But that flag is tied to MPOL_MF_MOVE.*, so it seems not a good
API/ABI for the purpose of the patch.
And because it's not clear whether it's necessary to enable NUMA balancing
for a specific memory area inside an application, so we only add the flag
at the thread level (set_mempolicy()) instead of the memory area level
(mbind()). We can do that when it become necessary.
To test the patch, we run a test case as follows on a 4-node machine with
192 GB memory (48 GB per node).
1. Change pmbench memory accessing benchmark to call set_mempolicy()
to bind its memory to node 1 and 3 and enable NUMA balancing. Some
related code snippets are as follows,
#include <numaif.h>
#include <numa.h>
struct bitmask *bmp;
int ret;
bmp = numa_parse_nodestring("1,3");
ret = set_mempolicy(MPOL_BIND | MPOL_F_NUMA_BALANCING,
bmp->maskp, bmp->size + 1);
/* If MPOL_F_NUMA_BALANCING isn't supported, fall back to MPOL_BIND */
if (ret < 0 && errno == EINVAL)
ret = set_mempolicy(MPOL_BIND, bmp->maskp, bmp->size + 1);
if (ret < 0) {
perror("Failed to call set_mempolicy");
exit(-1);
}
2. Run a memory eater on node 3 to use 40 GB memory before running pmbench.
3. Run pmbench with 64 processes, the working-set size of each process
is 640 MB, so the total working-set size is 64 * 640 MB = 40 GB. The
CPU and the memory (as in step 1.) of all pmbench processes is bound
to node 1 and 3. So, after CPU usage is balanced, some pmbench
processes run on the CPUs of the node 3 will access the memory of
the node 1.
4. After the pmbench processes run for 100 seconds, kill the memory
eater. Now it's possible for some pmbench processes to migrate
their pages from node 1 to node 3 to reduce cross-node accessing.
Test results show that, with the patch, the pages can be migrated from
node 1 to node 3 after killing the memory eater, and the pmbench score
can increase about 17.5%.
Link: https://lkml.kernel.org/r/20210120061235.148637-2-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-24 12:09:43 -08:00
|
|
|
|
mm/numa_balancing: allow migrate on protnone reference with MPOL_PREFERRED_MANY policy
commit bda420b98505 ("numa balancing: migrate on fault among multiple
bound nodes") added support for migrate on protnone reference with
MPOL_BIND memory policy. This allowed numa fault migration when the
executing node is part of the policy mask for MPOL_BIND. This patch
extends migration support to MPOL_PREFERRED_MANY policy.
Currently, we cannot specify MPOL_PREFERRED_MANY with the mempolicy flag
MPOL_F_NUMA_BALANCING. This causes issues when we want to use
NUMA_BALANCING_MEMORY_TIERING. To effectively use the slow memory tier,
the kernel should not allocate pages from the slower memory tier via
allocation control zonelist fallback. Instead, we should move cold pages
from the faster memory node via memory demotion. For a page allocation,
kswapd is only woken up after we try to allocate pages from all nodes in
the allocation zone list. This implies that, without using memory
policies, we will end up allocating hot pages in the slower memory tier.
MPOL_PREFERRED_MANY was added by commit b27abaccf8e8 ("mm/mempolicy: add
MPOL_PREFERRED_MANY for multiple preferred nodes") to allow better
allocation control when we have memory tiers in the system. With
MPOL_PREFERRED_MANY, the user can use a policy node mask consisting only
of faster memory nodes. When we fail to allocate pages from the faster
memory node, kswapd would be woken up, allowing demotion of cold pages to
slower memory nodes.
With the current kernel, such usage of memory policies implies we can't do
page promotion from a slower memory tier to a faster memory tier using
numa fault. This patch fixes this issue.
For MPOL_PREFERRED_MANY, if the executing node is in the policy node mask,
we allow numa migration to the executing nodes. If the executing node is
not in the policy node mask, we do not allow numa migration.
Example:
On a 2-sockets system, NUMA node N0, N1 and N2 are in socket 0,
N3 in socket 1. N0, N1 and N3 have fast memory and CPU, while
N2 has slow memory and no CPU. For a workload, we may use
MPOL_PREFERRED_MANY with nodemask N0 and N1 set because the workload
runs on CPUs of socket 0 at most times. Then, even if the workload
runs on CPUs of N3 occasionally, we will not try to migrate the workload
pages from N2 to N3 because users may want to avoid cross-socket access
as much as possible in the long term.
In below table, Process is the Process executing node and
Curr Loc Pgs is the numa node where page present(folio node)
===========================================================
Process Policy Curr Loc Pgs Observation
-----------------------------------------------------------
N0 N0 N1 N1 Pages Migrated from N1 to N0
N0 N0 N1 N2 Pages Migrated from N2 to N0
N0 N0 N1 N3 Pages Migrated from N3 to N0
N3 N0 N1 N0 Pages NOT Migrated to N3
N3 N0 N1 N1 Pages NOT Migrated to N3
N3 N0 N1 N2 Pages NOT Migrated to N3
------------------------------------------------------------
Link: https://lkml.kernel.org/r/158acc57319129aa46d50fd64c9330f3e7c7b4bf.1711373653.git.donettom@linux.ibm.com
Link: https://lkml.kernel.org/r/369d6a58758396335fd1176d97bbca4e7730d75a.1709909210.git.donettom@linux.ibm.com
Signed-off-by: Aneesh Kumar K.V (IBM) <aneesh.kumar@kernel.org>
Signed-off-by: Donet Tom <donettom@linux.ibm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Feng Tang <feng.tang@intel.com>
Cc: Huang, Ying <ying.huang@intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rik van Riel <riel@surriel.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-08 09:15:38 -06:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-30 18:51:10 -07:00
|
|
|
|
numa balancing: migrate on fault among multiple bound nodes
Now, NUMA balancing can only optimize the page placement among the NUMA
nodes if the default memory policy is used. Because the memory policy
specified explicitly should take precedence. But this seems too strict in
some situations. For example, on a system with 4 NUMA nodes, if the
memory of an application is bound to the node 0 and 1, NUMA balancing can
potentially migrate the pages between the node 0 and 1 to reduce
cross-node accessing without breaking the explicit memory binding policy.
So in this patch, we add MPOL_F_NUMA_BALANCING mode flag to
set_mempolicy() when mode is MPOL_BIND. With the flag specified, NUMA
balancing will be enabled within the thread to optimize the page placement
within the constrains of the specified memory binding policy. With the
newly added flag, the NUMA balancing control mechanism becomes,
- sysctl knob numa_balancing can enable/disable the NUMA balancing
globally.
- even if sysctl numa_balancing is enabled, the NUMA balancing will be
disabled for the memory areas or applications with the explicit
memory policy by default.
- MPOL_F_NUMA_BALANCING can be used to enable the NUMA balancing for
the applications when specifying the explicit memory policy
(MPOL_BIND).
Various page placement optimization based on the NUMA balancing can be
done with these flags. As the first step, in this patch, if the memory of
the application is bound to multiple nodes (MPOL_BIND), and in the hint
page fault handler the accessing node are in the policy nodemask, the page
will be tried to be migrated to the accessing node to reduce the
cross-node accessing.
If the newly added MPOL_F_NUMA_BALANCING flag is specified by an
application on an old kernel version without its support, set_mempolicy()
will return -1 and errno will be set to EINVAL. The application can use
this behavior to run on both old and new kernel versions.
And if the MPOL_F_NUMA_BALANCING flag is specified for the mode other than
MPOL_BIND, set_mempolicy() will return -1 and errno will be set to EINVAL
as before. Because we don't support optimization based on the NUMA
balancing for these modes.
In the previous version of the patch, we tried to reuse MPOL_MF_LAZY for
mbind(). But that flag is tied to MPOL_MF_MOVE.*, so it seems not a good
API/ABI for the purpose of the patch.
And because it's not clear whether it's necessary to enable NUMA balancing
for a specific memory area inside an application, so we only add the flag
at the thread level (set_mempolicy()) instead of the memory area level
(mbind()). We can do that when it become necessary.
To test the patch, we run a test case as follows on a 4-node machine with
192 GB memory (48 GB per node).
1. Change pmbench memory accessing benchmark to call set_mempolicy()
to bind its memory to node 1 and 3 and enable NUMA balancing. Some
related code snippets are as follows,
#include <numaif.h>
#include <numa.h>
struct bitmask *bmp;
int ret;
bmp = numa_parse_nodestring("1,3");
ret = set_mempolicy(MPOL_BIND | MPOL_F_NUMA_BALANCING,
bmp->maskp, bmp->size + 1);
/* If MPOL_F_NUMA_BALANCING isn't supported, fall back to MPOL_BIND */
if (ret < 0 && errno == EINVAL)
ret = set_mempolicy(MPOL_BIND, bmp->maskp, bmp->size + 1);
if (ret < 0) {
perror("Failed to call set_mempolicy");
exit(-1);
}
2. Run a memory eater on node 3 to use 40 GB memory before running pmbench.
3. Run pmbench with 64 processes, the working-set size of each process
is 640 MB, so the total working-set size is 64 * 640 MB = 40 GB. The
CPU and the memory (as in step 1.) of all pmbench processes is bound
to node 1 and 3. So, after CPU usage is balanced, some pmbench
processes run on the CPUs of the node 3 will access the memory of
the node 1.
4. After the pmbench processes run for 100 seconds, kill the memory
eater. Now it's possible for some pmbench processes to migrate
their pages from node 1 to node 3 to reduce cross-node accessing.
Test results show that, with the patch, the pages can be migrated from
node 1 to node 3 after killing the memory eater, and the pmbench score
can increase about 17.5%.
Link: https://lkml.kernel.org/r/20210120061235.148637-2-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-24 12:09:43 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-05-19 17:14:10 -07:00
|
|
|
|
2012-10-25 14:16:30 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-30 18:51:10 -07:00
|
|
|
|
2012-10-25 14:16:30 +02:00
|
|
|
|
2016-05-19 17:14:10 -07:00
|
|
|
|
mm/mempolicy: use numa_node_id() instead of cpu_to_node()
Patch series "Allow migrate on protnone reference with MPOL_PREFERRED_MANY
policy:, v4.
This patchset is to optimize the cross-socket memory access with
MPOL_PREFERRED_MANY policy.
To test this patch we ran the following test on a 3 node system.
Node 0 - 2GB - Tier 1
Node 1 - 11GB - Tier 1
Node 6 - 10GB - Tier 2
Below changes are made to memcached to set the memory policy,
It select Node0 and Node1 as preferred nodes.
#include <numaif.h>
#include <numa.h>
unsigned long nodemask;
int ret;
nodemask = 0x03;
ret = set_mempolicy(MPOL_PREFERRED_MANY | MPOL_F_NUMA_BALANCING,
&nodemask, 10);
/* If MPOL_F_NUMA_BALANCING isn't supported,
* fall back to MPOL_PREFERRED_MANY */
if (ret < 0 && errno == EINVAL){
printf("set mem policy normal\n");
ret = set_mempolicy(MPOL_PREFERRED_MANY, &nodemask, 10);
}
if (ret < 0) {
perror("Failed to call set_mempolicy");
exit(-1);
}
Test Procedure:
===============
1. Make sure memory tiring and demotion are enabled.
2. Start memcached.
# ./memcached -b 100000 -m 204800 -u root -c 1000000 -t 7
-d -s "/tmp/memcached.sock"
3. Run memtier_benchmark to store 3200000 keys.
#./memtier_benchmark -S "/tmp/memcached.sock" --protocol=memcache_binary
--threads=1 --pipeline=1 --ratio=1:0 --key-pattern=S:S --key-minimum=1
--key-maximum=3200000 -n allkeys -c 1 -R -x 1 -d 1024
4. Start a memory eater on node 0 and 1. This will demote all memcached
pages to node 6.
5. Make sure all the memcached pages got demoted to lower tier by reading
/proc/<memcaced PID>/numa_maps.
# cat /proc/2771/numa_maps
---
default anon=1009 dirty=1009 active=0 N6=1009 kernelpagesize_kB=64
default anon=1009 dirty=1009 active=0 N6=1009 kernelpagesize_kB=64
---
6. Kill memory eater.
7. Read the pgpromote_success counter.
8. Start reading the keys by running memtier_benchmark.
#./memtier_benchmark -S "/tmp/memcached.sock" --protocol=memcache_binary
--pipeline=1 --distinct-client-seed --ratio=0:3 --key-pattern=R:R
--key-minimum=1 --key-maximum=3200000 -n allkeys
--threads=64 -c 1 -R -x 6
9. Read the pgpromote_success counter.
Test Results:
=============
Without Patch
------------------
1. pgpromote_success before test
Node 0: pgpromote_success 11
Node 1: pgpromote_success 140974
pgpromote_success after test
Node 0: pgpromote_success 11
Node 1: pgpromote_success 140974
2. Memtier-benchmark result.
AGGREGATED AVERAGE RESULTS (6 runs)
==================================================================
Type Ops/sec Hits/sec Misses/sec Avg. Latency p50 Latency
------------------------------------------------------------------
Sets 0.00 --- --- --- ---
Gets 305792.03 305791.93 0.10 0.18949 0.16700
Waits 0.00 --- --- --- ---
Totals 305792.03 305791.93 0.10 0.18949 0.16700
======================================
p99 Latency p99.9 Latency KB/sec
-------------------------------------
--- --- 0.00
0.44700 1.71100 11542.69
--- --- ---
0.44700 1.71100 11542.69
With Patch
---------------
1. pgpromote_success before test
Node 0: pgpromote_success 5
Node 1: pgpromote_success 89386
pgpromote_success after test
Node 0: pgpromote_success 57895
Node 1: pgpromote_success 141463
2. Memtier-benchmark result.
AGGREGATED AVERAGE RESULTS (6 runs)
====================================================================
Type Ops/sec Hits/sec Misses/sec Avg. Latency p50 Latency
--------------------------------------------------------------------
Sets 0.00 --- --- --- ---
Gets 521942.24 521942.07 0.17 0.11459 0.10300
Waits 0.00 --- --- --- ---
Totals 521942.24 521942.07 0.17 0.11459 0.10300
=======================================
p99 Latency p99.9 Latency KB/sec
---------------------------------------
--- --- 0.00
0.23100 0.31900 19701.68
--- --- ---
0.23100 0.31900 19701.68
Test Result Analysis:
=====================
1. With patch we could observe pages are getting promoted.
2. Memtier-benchmark results shows that, with the patch,
performance has increased more than 50%.
Ops/sec without fix - 305792.03
Ops/sec with fix - 521942.24
This patch (of 2):
Instead of using 'cpu_to_node()', we use 'numa_node_id()', which is
quicker. smp_processor_id is guaranteed to be stable in the
'mpol_misplaced()' function because it is called with ptl held.
lockdep_assert_held was added to ensure that.
No functional change in this patch.
[donettom@linux.ibm.com: add "* @vmf: structure describing the fault" comment]
Link: https://lkml.kernel.org/r/d8b993ea9dccfac0bc3ed61d3a81f4ac5f376e46.1711002865.git.donettom@linux.ibm.com
Link: https://lkml.kernel.org/r/cover.1711373653.git.donettom@linux.ibm.com
Link: https://lkml.kernel.org/r/6059f034f436734b472d066db69676fb3a459864.1711373653.git.donettom@linux.ibm.com
Link: https://lkml.kernel.org/r/cover.1709909210.git.donettom@linux.ibm.com
Link: https://lkml.kernel.org/r/744646531af02cc687cde8ae788fb1779e99d02c.1709909210.git.donettom@linux.ibm.com
Signed-off-by: Aneesh Kumar K.V (IBM) <aneesh.kumar@kernel.org>
Signed-off-by: Donet Tom <donettom@linux.ibm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Feng Tang <feng.tang@intel.com>
Cc: Huang, Ying <ying.huang@intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rik van Riel <riel@surriel.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-08 09:15:37 -06:00
|
|
|
|
2012-10-25 14:16:30 +02:00
|
|
|
|
2021-06-30 18:51:10 -07:00
|
|
|
|
2024-07-29 14:47:17 +05:30
|
|
|
|
2012-10-25 14:16:30 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-11-02 18:19:13 +00:00
|
|
|
|
2023-09-21 15:44:16 +08:00
|
|
|
|
2012-11-12 09:17:07 +00:00
|
|
|
|
2013-10-07 11:29:20 +01:00
|
|
|
|
2012-11-02 18:19:13 +00:00
|
|
|
|
2023-09-21 15:44:17 +08:00
|
|
|
|
2023-09-21 15:44:16 +08:00
|
|
|
|
2013-10-07 11:29:39 +01:00
|
|
|
|
2012-11-12 09:17:07 +00:00
|
|
|
|
|
|
|
|
|
2012-10-25 14:16:30 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-09-01 16:15:07 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-10-08 16:29:19 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2012-12-21 23:10:25 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-10-16 01:26:26 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
mempolicy: remove mempolicy sharing
Dave Jones' system call fuzz testing tool "trinity" triggered the
following bug error with slab debugging enabled
=============================================================================
BUG numa_policy (Not tainted): Poison overwritten
-----------------------------------------------------------------------------
INFO: 0xffff880146498250-0xffff880146498250. First byte 0x6a instead of 0x6b
INFO: Allocated in mpol_new+0xa3/0x140 age=46310 cpu=6 pid=32154
__slab_alloc+0x3d3/0x445
kmem_cache_alloc+0x29d/0x2b0
mpol_new+0xa3/0x140
sys_mbind+0x142/0x620
system_call_fastpath+0x16/0x1b
INFO: Freed in __mpol_put+0x27/0x30 age=46268 cpu=6 pid=32154
__slab_free+0x2e/0x1de
kmem_cache_free+0x25a/0x260
__mpol_put+0x27/0x30
remove_vma+0x68/0x90
exit_mmap+0x118/0x140
mmput+0x73/0x110
exit_mm+0x108/0x130
do_exit+0x162/0xb90
do_group_exit+0x4f/0xc0
sys_exit_group+0x17/0x20
system_call_fastpath+0x16/0x1b
INFO: Slab 0xffffea0005192600 objects=27 used=27 fp=0x (null) flags=0x20000000004080
INFO: Object 0xffff880146498250 @offset=592 fp=0xffff88014649b9d0
The problem is that the structure is being prematurely freed due to a
reference count imbalance. In the following case mbind(addr, len) should
replace the memory policies of both vma1 and vma2 and thus they will
become to share the same mempolicy and the new mempolicy will have the
MPOL_F_SHARED flag.
+-------------------+-------------------+
| vma1 | vma2(shmem) |
+-------------------+-------------------+
| |
addr addr+len
alloc_pages_vma() uses get_vma_policy() and mpol_cond_put() pair for
maintaining the mempolicy reference count. The current rule is that
get_vma_policy() only increments refcount for shmem VMA and
mpol_conf_put() only decrements refcount if the policy has
MPOL_F_SHARED.
In above case, vma1 is not shmem vma and vma->policy has MPOL_F_SHARED!
The reference count will be decreased even though was not increased
whenever alloc_page_vma() is called. This has been broken since commit
[52cd3b07: mempolicy: rework mempolicy Reference Counting] in 2008.
There is another serious bug with the sharing of memory policies.
Currently, mempolicy rebind logic (it is called from cpuset rebinding)
ignores a refcount of mempolicy and override it forcibly. Thus, any
mempolicy sharing may cause mempolicy corruption. The bug was
introduced by commit [68860ec1: cpusets: automatic numa mempolicy
rebinding].
Ideally, the shared policy handling would be rewritten to either
properly handle COW of the policy structures or at least reference count
MPOL_F_SHARED based exclusively on information within the policy.
However, this patch takes the easier approach of disabling any policy
sharing between VMAs. Each new range allocated with sp_alloc will
allocate a new policy, set the reference count to 1 and drop the
reference count of the old policy. This increases the memory footprint
but is not expected to be a major problem as mbind() is unlikely to be
used for fine-grained ranges. It is also inefficient because it means
we allocate a new policy even in cases where mbind_range() could use the
new_policy passed to it. However, it is more straight-forward and the
change should be invisible to the user.
[mgorman@suse.de: Edited changelog]
Reported-by: Dave Jones <davej@redhat.com>,
Cc: Christoph Lameter <cl@linux.com>,
Reviewed-by: Christoph Lameter <cl@linux.com>
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Josh Boyer <jwboyer@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-10-08 16:29:16 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
mempolicy: remove mempolicy sharing
Dave Jones' system call fuzz testing tool "trinity" triggered the
following bug error with slab debugging enabled
=============================================================================
BUG numa_policy (Not tainted): Poison overwritten
-----------------------------------------------------------------------------
INFO: 0xffff880146498250-0xffff880146498250. First byte 0x6a instead of 0x6b
INFO: Allocated in mpol_new+0xa3/0x140 age=46310 cpu=6 pid=32154
__slab_alloc+0x3d3/0x445
kmem_cache_alloc+0x29d/0x2b0
mpol_new+0xa3/0x140
sys_mbind+0x142/0x620
system_call_fastpath+0x16/0x1b
INFO: Freed in __mpol_put+0x27/0x30 age=46268 cpu=6 pid=32154
__slab_free+0x2e/0x1de
kmem_cache_free+0x25a/0x260
__mpol_put+0x27/0x30
remove_vma+0x68/0x90
exit_mmap+0x118/0x140
mmput+0x73/0x110
exit_mm+0x108/0x130
do_exit+0x162/0xb90
do_group_exit+0x4f/0xc0
sys_exit_group+0x17/0x20
system_call_fastpath+0x16/0x1b
INFO: Slab 0xffffea0005192600 objects=27 used=27 fp=0x (null) flags=0x20000000004080
INFO: Object 0xffff880146498250 @offset=592 fp=0xffff88014649b9d0
The problem is that the structure is being prematurely freed due to a
reference count imbalance. In the following case mbind(addr, len) should
replace the memory policies of both vma1 and vma2 and thus they will
become to share the same mempolicy and the new mempolicy will have the
MPOL_F_SHARED flag.
+-------------------+-------------------+
| vma1 | vma2(shmem) |
+-------------------+-------------------+
| |
addr addr+len
alloc_pages_vma() uses get_vma_policy() and mpol_cond_put() pair for
maintaining the mempolicy reference count. The current rule is that
get_vma_policy() only increments refcount for shmem VMA and
mpol_conf_put() only decrements refcount if the policy has
MPOL_F_SHARED.
In above case, vma1 is not shmem vma and vma->policy has MPOL_F_SHARED!
The reference count will be decreased even though was not increased
whenever alloc_page_vma() is called. This has been broken since commit
[52cd3b07: mempolicy: rework mempolicy Reference Counting] in 2008.
There is another serious bug with the sharing of memory policies.
Currently, mempolicy rebind logic (it is called from cpuset rebinding)
ignores a refcount of mempolicy and override it forcibly. Thus, any
mempolicy sharing may cause mempolicy corruption. The bug was
introduced by commit [68860ec1: cpusets: automatic numa mempolicy
rebinding].
Ideally, the shared policy handling would be rewritten to either
properly handle COW of the policy structures or at least reference count
MPOL_F_SHARED based exclusively on information within the policy.
However, this patch takes the easier approach of disabling any policy
sharing between VMAs. Each new range allocated with sp_alloc will
allocate a new policy, set the reference count to 1 and drop the
reference count of the old policy. This increases the memory footprint
but is not expected to be a major problem as mbind() is unlikely to be
used for fine-grained ranges. It is also inefficient because it means
we allocate a new policy even in cases where mbind_range() could use the
new_policy passed to it. However, it is more straight-forward and the
change should be invisible to the user.
[mgorman@suse.de: Edited changelog]
Reported-by: Dave Jones <davej@redhat.com>,
Cc: Christoph Lameter <cl@linux.com>,
Reviewed-by: Christoph Lameter <cl@linux.com>
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Josh Boyer <jwboyer@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-10-08 16:29:16 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
mempolicy: remove mempolicy sharing
Dave Jones' system call fuzz testing tool "trinity" triggered the
following bug error with slab debugging enabled
=============================================================================
BUG numa_policy (Not tainted): Poison overwritten
-----------------------------------------------------------------------------
INFO: 0xffff880146498250-0xffff880146498250. First byte 0x6a instead of 0x6b
INFO: Allocated in mpol_new+0xa3/0x140 age=46310 cpu=6 pid=32154
__slab_alloc+0x3d3/0x445
kmem_cache_alloc+0x29d/0x2b0
mpol_new+0xa3/0x140
sys_mbind+0x142/0x620
system_call_fastpath+0x16/0x1b
INFO: Freed in __mpol_put+0x27/0x30 age=46268 cpu=6 pid=32154
__slab_free+0x2e/0x1de
kmem_cache_free+0x25a/0x260
__mpol_put+0x27/0x30
remove_vma+0x68/0x90
exit_mmap+0x118/0x140
mmput+0x73/0x110
exit_mm+0x108/0x130
do_exit+0x162/0xb90
do_group_exit+0x4f/0xc0
sys_exit_group+0x17/0x20
system_call_fastpath+0x16/0x1b
INFO: Slab 0xffffea0005192600 objects=27 used=27 fp=0x (null) flags=0x20000000004080
INFO: Object 0xffff880146498250 @offset=592 fp=0xffff88014649b9d0
The problem is that the structure is being prematurely freed due to a
reference count imbalance. In the following case mbind(addr, len) should
replace the memory policies of both vma1 and vma2 and thus they will
become to share the same mempolicy and the new mempolicy will have the
MPOL_F_SHARED flag.
+-------------------+-------------------+
| vma1 | vma2(shmem) |
+-------------------+-------------------+
| |
addr addr+len
alloc_pages_vma() uses get_vma_policy() and mpol_cond_put() pair for
maintaining the mempolicy reference count. The current rule is that
get_vma_policy() only increments refcount for shmem VMA and
mpol_conf_put() only decrements refcount if the policy has
MPOL_F_SHARED.
In above case, vma1 is not shmem vma and vma->policy has MPOL_F_SHARED!
The reference count will be decreased even though was not increased
whenever alloc_page_vma() is called. This has been broken since commit
[52cd3b07: mempolicy: rework mempolicy Reference Counting] in 2008.
There is another serious bug with the sharing of memory policies.
Currently, mempolicy rebind logic (it is called from cpuset rebinding)
ignores a refcount of mempolicy and override it forcibly. Thus, any
mempolicy sharing may cause mempolicy corruption. The bug was
introduced by commit [68860ec1: cpusets: automatic numa mempolicy
rebinding].
Ideally, the shared policy handling would be rewritten to either
properly handle COW of the policy structures or at least reference count
MPOL_F_SHARED based exclusively on information within the policy.
However, this patch takes the easier approach of disabling any policy
sharing between VMAs. Each new range allocated with sp_alloc will
allocate a new policy, set the reference count to 1 and drop the
reference count of the old policy. This increases the memory footprint
but is not expected to be a major problem as mbind() is unlikely to be
used for fine-grained ranges. It is also inefficient because it means
we allocate a new policy even in cases where mbind_range() could use the
new_policy passed to it. However, it is more straight-forward and the
change should be invisible to the user.
[mgorman@suse.de: Edited changelog]
Reported-by: Dave Jones <davej@redhat.com>,
Cc: Christoph Lameter <cl@linux.com>,
Reviewed-by: Christoph Lameter <cl@linux.com>
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Josh Boyer <jwboyer@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-10-08 16:29:16 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-12-21 23:10:25 +00:00
|
|
|
|
mempolicy: remove mempolicy sharing
Dave Jones' system call fuzz testing tool "trinity" triggered the
following bug error with slab debugging enabled
=============================================================================
BUG numa_policy (Not tainted): Poison overwritten
-----------------------------------------------------------------------------
INFO: 0xffff880146498250-0xffff880146498250. First byte 0x6a instead of 0x6b
INFO: Allocated in mpol_new+0xa3/0x140 age=46310 cpu=6 pid=32154
__slab_alloc+0x3d3/0x445
kmem_cache_alloc+0x29d/0x2b0
mpol_new+0xa3/0x140
sys_mbind+0x142/0x620
system_call_fastpath+0x16/0x1b
INFO: Freed in __mpol_put+0x27/0x30 age=46268 cpu=6 pid=32154
__slab_free+0x2e/0x1de
kmem_cache_free+0x25a/0x260
__mpol_put+0x27/0x30
remove_vma+0x68/0x90
exit_mmap+0x118/0x140
mmput+0x73/0x110
exit_mm+0x108/0x130
do_exit+0x162/0xb90
do_group_exit+0x4f/0xc0
sys_exit_group+0x17/0x20
system_call_fastpath+0x16/0x1b
INFO: Slab 0xffffea0005192600 objects=27 used=27 fp=0x (null) flags=0x20000000004080
INFO: Object 0xffff880146498250 @offset=592 fp=0xffff88014649b9d0
The problem is that the structure is being prematurely freed due to a
reference count imbalance. In the following case mbind(addr, len) should
replace the memory policies of both vma1 and vma2 and thus they will
become to share the same mempolicy and the new mempolicy will have the
MPOL_F_SHARED flag.
+-------------------+-------------------+
| vma1 | vma2(shmem) |
+-------------------+-------------------+
| |
addr addr+len
alloc_pages_vma() uses get_vma_policy() and mpol_cond_put() pair for
maintaining the mempolicy reference count. The current rule is that
get_vma_policy() only increments refcount for shmem VMA and
mpol_conf_put() only decrements refcount if the policy has
MPOL_F_SHARED.
In above case, vma1 is not shmem vma and vma->policy has MPOL_F_SHARED!
The reference count will be decreased even though was not increased
whenever alloc_page_vma() is called. This has been broken since commit
[52cd3b07: mempolicy: rework mempolicy Reference Counting] in 2008.
There is another serious bug with the sharing of memory policies.
Currently, mempolicy rebind logic (it is called from cpuset rebinding)
ignores a refcount of mempolicy and override it forcibly. Thus, any
mempolicy sharing may cause mempolicy corruption. The bug was
introduced by commit [68860ec1: cpusets: automatic numa mempolicy
rebinding].
Ideally, the shared policy handling would be rewritten to either
properly handle COW of the policy structures or at least reference count
MPOL_F_SHARED based exclusively on information within the policy.
However, this patch takes the easier approach of disabling any policy
sharing between VMAs. Each new range allocated with sp_alloc will
allocate a new policy, set the reference count to 1 and drop the
reference count of the old policy. This increases the memory footprint
but is not expected to be a major problem as mbind() is unlikely to be
used for fine-grained ranges. It is also inefficient because it means
we allocate a new policy even in cases where mbind_range() could use the
new_policy passed to it. However, it is more straight-forward and the
change should be invisible to the user.
[mgorman@suse.de: Edited changelog]
Reported-by: Dave Jones <davej@redhat.com>,
Cc: Christoph Lameter <cl@linux.com>,
Reviewed-by: Christoph Lameter <cl@linux.com>
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Josh Boyer <jwboyer@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-10-08 16:29:16 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-03 02:21:34 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2012-10-08 16:29:17 -07:00
|
|
|
|
2012-12-21 23:10:25 +00:00
|
|
|
|
|
|
|
|
|
2012-10-08 16:29:17 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2012-12-21 23:10:25 +00:00
|
|
|
|
2016-01-14 15:18:36 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-12-21 23:10:25 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-03-08 12:43:29 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2013-03-08 12:43:28 -08:00
|
|
|
|
2012-12-21 23:10:25 +00:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-01-14 15:18:36 -08:00
|
|
|
|
2012-12-21 23:10:25 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-10-08 16:29:17 -07:00
|
|
|
|
2012-12-21 23:10:25 +00:00
|
|
|
|
|
|
|
|
|
2016-01-14 15:18:36 -08:00
|
|
|
|
2012-12-21 23:10:25 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: fix mpol_new leak in shared_policy_replace
If mpol_new is allocated but not used in restart loop, mpol_new will be
freed via mpol_put before returning to the caller. But refcnt is not
initialized yet, so mpol_put could not do the right things and might
leak the unused mpol_new. This would happen if mempolicy was updated on
the shared shmem file while the sp->lock has been dropped during the
memory allocation.
This issue could be triggered easily with the below code snippet if
there are many processes doing the below work at the same time:
shmid = shmget((key_t)5566, 1024 * PAGE_SIZE, 0666|IPC_CREAT);
shm = shmat(shmid, 0, 0);
loop many times {
mbind(shm, 1024 * PAGE_SIZE, MPOL_LOCAL, mask, maxnode, 0);
mbind(shm + 128 * PAGE_SIZE, 128 * PAGE_SIZE, MPOL_DEFAULT, mask,
maxnode, 0);
}
Link: https://lkml.kernel.org/r/20220329111416.27954-1-linmiaohe@huawei.com
Fixes: 42288fe366c4 ("mm: mempolicy: Convert shared_policy mutex to spinlock")
Signed-off-by: Miaohe Lin <linmiaohe@huawei.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: <stable@vger.kernel.org> [3.8]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-04-08 13:09:07 -07:00
|
|
|
|
2012-12-21 23:10:25 +00:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
mempolicy: use struct mempolicy pointer in shmem_sb_info
This patch replaces the mempolicy mode, mode_flags, and nodemask in the
shmem_sb_info struct with a struct mempolicy pointer, initialized to NULL.
This removes dependency on the details of mempolicy from shmem.c and hugetlbfs
inode.c and simplifies the interfaces.
mpol_parse_str() in mempolicy.c is changed to return, via a pointer to a
pointer arg, a struct mempolicy pointer on success. For MPOL_DEFAULT, the
returned pointer is NULL. Further, mpol_parse_str() now takes a 'no_context'
argument that causes the input nodemask to be stored in the w.user_nodemask of
the created mempolicy for use when the mempolicy is installed in a tmpfs inode
shared policy tree. At that time, any cpuset contextualization is applied to
the original input nodemask. This preserves the previous behavior where the
input nodemask was stored in the superblock. We can think of the returned
mempolicy as "context free".
Because mpol_parse_str() is now calling mpol_new(), we can remove from
mpol_to_str() the semantic checks that mpol_new() already performs.
Add 'no_context' parameter to mpol_to_str() to specify that it should format
the nodemask in w.user_nodemask for 'bind' and 'interleave' policies.
Change mpol_shared_policy_init() to take a pointer to a "context free" struct
mempolicy and to create a new, "contextualized" mempolicy using the mode,
mode_flags and user_nodemask from the input mempolicy.
Note: we know that the mempolicy passed to mpol_to_str() or
mpol_shared_policy_init() from a tmpfs superblock is "context free". This
is currently the only instance thereof. However, if we found more uses for
this concept, and introduced any ambiguity as to whether a mempolicy was
context free or not, we could add another internal mode flag to identify
context free mempolicies. Then, we could remove the 'no_context' argument
from mpol_to_str().
Added shmem_get_sbmpol() to return a reference counted superblock mempolicy,
if one exists, to pass to mpol_shared_policy_init(). We must add the
reference under the sb stat_lock to prevent races with replacement of the mpol
by remount. This reference is removed in mpol_shared_policy_init().
[akpm@linux-foundation.org: build fix]
[akpm@linux-foundation.org: another build fix]
[akpm@linux-foundation.org: yet another build fix]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:26 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: make set_mempolicy(MPOL_INTERLEAV) N_HIGH_MEMORY aware
At first, init_task's mems_allowed is initialized as this.
init_task->mems_allowed == node_state[N_POSSIBLE]
And cpuset's top_cpuset mask is initialized as this
top_cpuset->mems_allowed = node_state[N_HIGH_MEMORY]
Before 2.6.29:
policy's mems_allowed is initialized as this.
1. update tasks->mems_allowed by its cpuset->mems_allowed.
2. policy->mems_allowed = nodes_and(tasks->mems_allowed, user's mask)
Updating task's mems_allowed in reference to top_cpuset's one.
cpuset's mems_allowed is aware of N_HIGH_MEMORY, always.
In 2.6.30: After commit 58568d2a8215cb6f55caf2332017d7bdff954e1c
("cpuset,mm: update tasks' mems_allowed in time"), policy's mems_allowed
is initialized as this.
1. policy->mems_allowd = nodes_and(task->mems_allowed, user's mask)
Here, if task is in top_cpuset, task->mems_allowed is not updated from
init's one. Assume user excutes command as #numactrl --interleave=all
,....
policy->mems_allowd = nodes_and(N_POSSIBLE, ALL_SET_MASK)
Then, policy's mems_allowd can includes a possible node, which has no pgdat.
MPOL's INTERLEAVE just scans nodemask of task->mems_allowd and access this
directly.
NODE_DATA(nid)->zonelist even if NODE_DATA(nid)==NULL
Then, what's we need is making policy->mems_allowed be aware of
N_HIGH_MEMORY. This patch does that. But to do so, extra nodemask will
be on statck. Because I know cpumask has a new interface of
CPUMASK_ALLOC(), I added it to node.
This patch stands on old behavior. But I feel this fix itself is just a
Band-Aid. But to do fundametal fix, we have to take care of memory
hotplug and it takes time. (task->mems_allowd should be N_HIGH_MEMORY, I
think.)
mpol_set_nodemask() should be aware of N_HIGH_MEMORY and policy's nodemask
should be includes only online nodes.
In old behavior, this is guaranteed by frequent reference to cpuset's
code. Now, most of them are removed and mempolicy has to check it by
itself.
To do check, a few nodemask_t will be used for calculating nodemask. But,
size of nodemask_t can be big and it's not good to allocate them on stack.
Now, cpumask_t has CPUMASK_ALLOC/FREE an easy code for get scratch area.
NODEMASK_ALLOC/FREE shoudl be there.
[akpm@linux-foundation.org: cleanups & tweaks]
Tested-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-08-06 15:07:33 -07:00
|
|
|
|
mempolicy: use struct mempolicy pointer in shmem_sb_info
This patch replaces the mempolicy mode, mode_flags, and nodemask in the
shmem_sb_info struct with a struct mempolicy pointer, initialized to NULL.
This removes dependency on the details of mempolicy from shmem.c and hugetlbfs
inode.c and simplifies the interfaces.
mpol_parse_str() in mempolicy.c is changed to return, via a pointer to a
pointer arg, a struct mempolicy pointer on success. For MPOL_DEFAULT, the
returned pointer is NULL. Further, mpol_parse_str() now takes a 'no_context'
argument that causes the input nodemask to be stored in the w.user_nodemask of
the created mempolicy for use when the mempolicy is installed in a tmpfs inode
shared policy tree. At that time, any cpuset contextualization is applied to
the original input nodemask. This preserves the previous behavior where the
input nodemask was stored in the superblock. We can think of the returned
mempolicy as "context free".
Because mpol_parse_str() is now calling mpol_new(), we can remove from
mpol_to_str() the semantic checks that mpol_new() already performs.
Add 'no_context' parameter to mpol_to_str() to specify that it should format
the nodemask in w.user_nodemask for 'bind' and 'interleave' policies.
Change mpol_shared_policy_init() to take a pointer to a "context free" struct
mempolicy and to create a new, "contextualized" mempolicy using the mode,
mode_flags and user_nodemask from the input mempolicy.
Note: we know that the mempolicy passed to mpol_to_str() or
mpol_shared_policy_init() from a tmpfs superblock is "context free". This
is currently the only instance thereof. However, if we found more uses for
this concept, and introduced any ambiguity as to whether a mempolicy was
context free or not, we could add another internal mode flag to identify
context free mempolicies. Then, we could remove the 'no_context' argument
from mpol_to_str().
Added shmem_get_sbmpol() to return a reference counted superblock mempolicy,
if one exists, to pass to mpol_shared_policy_init(). We must add the
reference under the sb stat_lock to prevent races with replacement of the mpol
by remount. This reference is removed in mpol_shared_policy_init().
[akpm@linux-foundation.org: build fix]
[akpm@linux-foundation.org: another build fix]
[akpm@linux-foundation.org: yet another build fix]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:26 -07:00
|
|
|
|
|
|
|
|
|
2006-01-14 13:20:48 -08:00
|
|
|
|
cpuset,mm: update tasks' mems_allowed in time
Fix allocating page cache/slab object on the unallowed node when memory
spread is set by updating tasks' mems_allowed after its cpuset's mems is
changed.
In order to update tasks' mems_allowed in time, we must modify the code of
memory policy. Because the memory policy is applied in the process's
context originally. After applying this patch, one task directly
manipulates anothers mems_allowed, and we use alloc_lock in the
task_struct to protect mems_allowed and memory policy of the task.
But in the fast path, we didn't use lock to protect them, because adding a
lock may lead to performance regression. But if we don't add a lock,the
task might see no nodes when changing cpuset's mems_allowed to some
non-overlapping set. In order to avoid it, we set all new allowed nodes,
then clear newly disallowed ones.
[lee.schermerhorn@hp.com:
The rework of mpol_new() to extract the adjusting of the node mask to
apply cpuset and mpol flags "context" breaks set_mempolicy() and mbind()
with MPOL_PREFERRED and a NULL nodemask--i.e., explicit local
allocation. Fix this by adding the check for MPOL_PREFERRED and empty
node mask to mpol_new_mpolicy().
Remove the now unneeded 'nodes = NULL' from mpol_new().
Note that mpol_new_mempolicy() is always called with a non-NULL
'nodes' parameter now that it has been removed from mpol_new().
Therefore, we don't need to test nodes for NULL before testing it for
'empty'. However, just to be extra paranoid, add a VM_BUG_ON() to
verify this assumption.]
[lee.schermerhorn@hp.com:
I don't think the function name 'mpol_new_mempolicy' is descriptive
enough to differentiate it from mpol_new().
This function applies cpuset set context, usually constraining nodes
to those allowed by the cpuset. However, when the 'RELATIVE_NODES flag
is set, it also translates the nodes. So I settled on
'mpol_set_nodemask()', because the comment block for mpol_new() mentions
that we need to call this function to "set nodes".
Some additional minor line length, whitespace and typo cleanup.]
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 15:31:49 -07:00
|
|
|
|
|
|
|
|
|
mempolicy: use struct mempolicy pointer in shmem_sb_info
This patch replaces the mempolicy mode, mode_flags, and nodemask in the
shmem_sb_info struct with a struct mempolicy pointer, initialized to NULL.
This removes dependency on the details of mempolicy from shmem.c and hugetlbfs
inode.c and simplifies the interfaces.
mpol_parse_str() in mempolicy.c is changed to return, via a pointer to a
pointer arg, a struct mempolicy pointer on success. For MPOL_DEFAULT, the
returned pointer is NULL. Further, mpol_parse_str() now takes a 'no_context'
argument that causes the input nodemask to be stored in the w.user_nodemask of
the created mempolicy for use when the mempolicy is installed in a tmpfs inode
shared policy tree. At that time, any cpuset contextualization is applied to
the original input nodemask. This preserves the previous behavior where the
input nodemask was stored in the superblock. We can think of the returned
mempolicy as "context free".
Because mpol_parse_str() is now calling mpol_new(), we can remove from
mpol_to_str() the semantic checks that mpol_new() already performs.
Add 'no_context' parameter to mpol_to_str() to specify that it should format
the nodemask in w.user_nodemask for 'bind' and 'interleave' policies.
Change mpol_shared_policy_init() to take a pointer to a "context free" struct
mempolicy and to create a new, "contextualized" mempolicy using the mode,
mode_flags and user_nodemask from the input mempolicy.
Note: we know that the mempolicy passed to mpol_to_str() or
mpol_shared_policy_init() from a tmpfs superblock is "context free". This
is currently the only instance thereof. However, if we found more uses for
this concept, and introduced any ambiguity as to whether a mempolicy was
context free or not, we could add another internal mode flag to identify
context free mempolicies. Then, we could remove the 'no_context' argument
from mpol_to_str().
Added shmem_get_sbmpol() to return a reference counted superblock mempolicy,
if one exists, to pass to mpol_shared_policy_init(). We must add the
reference under the sb stat_lock to prevent races with replacement of the mpol
by remount. This reference is removed in mpol_shared_policy_init().
[akpm@linux-foundation.org: build fix]
[akpm@linux-foundation.org: another build fix]
[akpm@linux-foundation.org: yet another build fix]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:26 -07:00
|
|
|
|
2016-01-14 15:18:36 -08:00
|
|
|
|
2006-01-14 13:20:48 -08:00
|
|
|
|
mempolicy: use struct mempolicy pointer in shmem_sb_info
This patch replaces the mempolicy mode, mode_flags, and nodemask in the
shmem_sb_info struct with a struct mempolicy pointer, initialized to NULL.
This removes dependency on the details of mempolicy from shmem.c and hugetlbfs
inode.c and simplifies the interfaces.
mpol_parse_str() in mempolicy.c is changed to return, via a pointer to a
pointer arg, a struct mempolicy pointer on success. For MPOL_DEFAULT, the
returned pointer is NULL. Further, mpol_parse_str() now takes a 'no_context'
argument that causes the input nodemask to be stored in the w.user_nodemask of
the created mempolicy for use when the mempolicy is installed in a tmpfs inode
shared policy tree. At that time, any cpuset contextualization is applied to
the original input nodemask. This preserves the previous behavior where the
input nodemask was stored in the superblock. We can think of the returned
mempolicy as "context free".
Because mpol_parse_str() is now calling mpol_new(), we can remove from
mpol_to_str() the semantic checks that mpol_new() already performs.
Add 'no_context' parameter to mpol_to_str() to specify that it should format
the nodemask in w.user_nodemask for 'bind' and 'interleave' policies.
Change mpol_shared_policy_init() to take a pointer to a "context free" struct
mempolicy and to create a new, "contextualized" mempolicy using the mode,
mode_flags and user_nodemask from the input mempolicy.
Note: we know that the mempolicy passed to mpol_to_str() or
mpol_shared_policy_init() from a tmpfs superblock is "context free". This
is currently the only instance thereof. However, if we found more uses for
this concept, and introduced any ambiguity as to whether a mempolicy was
context free or not, we could add another internal mode flag to identify
context free mempolicies. Then, we could remove the 'no_context' argument
from mpol_to_str().
Added shmem_get_sbmpol() to return a reference counted superblock mempolicy,
if one exists, to pass to mpol_shared_policy_init(). We must add the
reference under the sb stat_lock to prevent races with replacement of the mpol
by remount. This reference is removed in mpol_shared_policy_init().
[akpm@linux-foundation.org: build fix]
[akpm@linux-foundation.org: another build fix]
[akpm@linux-foundation.org: yet another build fix]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:26 -07:00
|
|
|
|
2023-10-03 02:22:59 -07:00
|
|
|
|
|
|
|
|
|
mm: make set_mempolicy(MPOL_INTERLEAV) N_HIGH_MEMORY aware
At first, init_task's mems_allowed is initialized as this.
init_task->mems_allowed == node_state[N_POSSIBLE]
And cpuset's top_cpuset mask is initialized as this
top_cpuset->mems_allowed = node_state[N_HIGH_MEMORY]
Before 2.6.29:
policy's mems_allowed is initialized as this.
1. update tasks->mems_allowed by its cpuset->mems_allowed.
2. policy->mems_allowed = nodes_and(tasks->mems_allowed, user's mask)
Updating task's mems_allowed in reference to top_cpuset's one.
cpuset's mems_allowed is aware of N_HIGH_MEMORY, always.
In 2.6.30: After commit 58568d2a8215cb6f55caf2332017d7bdff954e1c
("cpuset,mm: update tasks' mems_allowed in time"), policy's mems_allowed
is initialized as this.
1. policy->mems_allowd = nodes_and(task->mems_allowed, user's mask)
Here, if task is in top_cpuset, task->mems_allowed is not updated from
init's one. Assume user excutes command as #numactrl --interleave=all
,....
policy->mems_allowd = nodes_and(N_POSSIBLE, ALL_SET_MASK)
Then, policy's mems_allowd can includes a possible node, which has no pgdat.
MPOL's INTERLEAVE just scans nodemask of task->mems_allowd and access this
directly.
NODE_DATA(nid)->zonelist even if NODE_DATA(nid)==NULL
Then, what's we need is making policy->mems_allowed be aware of
N_HIGH_MEMORY. This patch does that. But to do so, extra nodemask will
be on statck. Because I know cpumask has a new interface of
CPUMASK_ALLOC(), I added it to node.
This patch stands on old behavior. But I feel this fix itself is just a
Band-Aid. But to do fundametal fix, we have to take care of memory
hotplug and it takes time. (task->mems_allowd should be N_HIGH_MEMORY, I
think.)
mpol_set_nodemask() should be aware of N_HIGH_MEMORY and policy's nodemask
should be includes only online nodes.
In old behavior, this is guaranteed by frequent reference to cpuset's
code. Now, most of them are removed and mempolicy has to check it by
itself.
To do check, a few nodemask_t will be used for calculating nodemask. But,
size of nodemask_t can be big and it's not good to allocate them on stack.
Now, cpumask_t has CPUMASK_ALLOC/FREE an easy code for get scratch area.
NODEMASK_ALLOC/FREE shoudl be there.
[akpm@linux-foundation.org: cleanups & tweaks]
Tested-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-08-06 15:07:33 -07:00
|
|
|
|
2006-01-14 13:20:48 -08:00
|
|
|
|
mm: make set_mempolicy(MPOL_INTERLEAV) N_HIGH_MEMORY aware
At first, init_task's mems_allowed is initialized as this.
init_task->mems_allowed == node_state[N_POSSIBLE]
And cpuset's top_cpuset mask is initialized as this
top_cpuset->mems_allowed = node_state[N_HIGH_MEMORY]
Before 2.6.29:
policy's mems_allowed is initialized as this.
1. update tasks->mems_allowed by its cpuset->mems_allowed.
2. policy->mems_allowed = nodes_and(tasks->mems_allowed, user's mask)
Updating task's mems_allowed in reference to top_cpuset's one.
cpuset's mems_allowed is aware of N_HIGH_MEMORY, always.
In 2.6.30: After commit 58568d2a8215cb6f55caf2332017d7bdff954e1c
("cpuset,mm: update tasks' mems_allowed in time"), policy's mems_allowed
is initialized as this.
1. policy->mems_allowd = nodes_and(task->mems_allowed, user's mask)
Here, if task is in top_cpuset, task->mems_allowed is not updated from
init's one. Assume user excutes command as #numactrl --interleave=all
,....
policy->mems_allowd = nodes_and(N_POSSIBLE, ALL_SET_MASK)
Then, policy's mems_allowd can includes a possible node, which has no pgdat.
MPOL's INTERLEAVE just scans nodemask of task->mems_allowd and access this
directly.
NODE_DATA(nid)->zonelist even if NODE_DATA(nid)==NULL
Then, what's we need is making policy->mems_allowed be aware of
N_HIGH_MEMORY. This patch does that. But to do so, extra nodemask will
be on statck. Because I know cpumask has a new interface of
CPUMASK_ALLOC(), I added it to node.
This patch stands on old behavior. But I feel this fix itself is just a
Band-Aid. But to do fundametal fix, we have to take care of memory
hotplug and it takes time. (task->mems_allowd should be N_HIGH_MEMORY, I
think.)
mpol_set_nodemask() should be aware of N_HIGH_MEMORY and policy's nodemask
should be includes only online nodes.
In old behavior, this is guaranteed by frequent reference to cpuset's
code. Now, most of them are removed and mempolicy has to check it by
itself.
To do check, a few nodemask_t will be used for calculating nodemask. But,
size of nodemask_t can be big and it's not good to allocate them on stack.
Now, cpumask_t has CPUMASK_ALLOC/FREE an easy code for get scratch area.
NODEMASK_ALLOC/FREE shoudl be there.
[akpm@linux-foundation.org: cleanups & tweaks]
Tested-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-08-06 15:07:33 -07:00
|
|
|
|
2010-06-29 15:05:30 -07:00
|
|
|
|
2023-10-03 02:22:59 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-05-25 23:42:58 -07:00
|
|
|
|
cpuset,mm: update tasks' mems_allowed in time
Fix allocating page cache/slab object on the unallowed node when memory
spread is set by updating tasks' mems_allowed after its cpuset's mems is
changed.
In order to update tasks' mems_allowed in time, we must modify the code of
memory policy. Because the memory policy is applied in the process's
context originally. After applying this patch, one task directly
manipulates anothers mems_allowed, and we use alloc_lock in the
task_struct to protect mems_allowed and memory policy of the task.
But in the fast path, we didn't use lock to protect them, because adding a
lock may lead to performance regression. But if we don't add a lock,the
task might see no nodes when changing cpuset's mems_allowed to some
non-overlapping set. In order to avoid it, we set all new allowed nodes,
then clear newly disallowed ones.
[lee.schermerhorn@hp.com:
The rework of mpol_new() to extract the adjusting of the node mask to
apply cpuset and mpol flags "context" breaks set_mempolicy() and mbind()
with MPOL_PREFERRED and a NULL nodemask--i.e., explicit local
allocation. Fix this by adding the check for MPOL_PREFERRED and empty
node mask to mpol_new_mpolicy().
Remove the now unneeded 'nodes = NULL' from mpol_new().
Note that mpol_new_mempolicy() is always called with a non-NULL
'nodes' parameter now that it has been removed from mpol_new().
Therefore, we don't need to test nodes for NULL before testing it for
'empty'. However, just to be extra paranoid, add a VM_BUG_ON() to
verify this assumption.]
[lee.schermerhorn@hp.com:
I don't think the function name 'mpol_new_mempolicy' is descriptive
enough to differentiate it from mpol_new().
This function applies cpuset set context, usually constraining nodes
to those allowed by the cpuset. However, when the 'RELATIVE_NODES flag
is set, it also translates the nodes. So I settled on
'mpol_set_nodemask()', because the comment block for mpol_new() mentions
that we need to call this function to "set nodes".
Some additional minor line length, whitespace and typo cleanup.]
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 15:31:49 -07:00
|
|
|
|
|
|
|
|
|
2023-10-03 02:22:59 -07:00
|
|
|
|
cpuset,mm: update tasks' mems_allowed in time
Fix allocating page cache/slab object on the unallowed node when memory
spread is set by updating tasks' mems_allowed after its cpuset's mems is
changed.
In order to update tasks' mems_allowed in time, we must modify the code of
memory policy. Because the memory policy is applied in the process's
context originally. After applying this patch, one task directly
manipulates anothers mems_allowed, and we use alloc_lock in the
task_struct to protect mems_allowed and memory policy of the task.
But in the fast path, we didn't use lock to protect them, because adding a
lock may lead to performance regression. But if we don't add a lock,the
task might see no nodes when changing cpuset's mems_allowed to some
non-overlapping set. In order to avoid it, we set all new allowed nodes,
then clear newly disallowed ones.
[lee.schermerhorn@hp.com:
The rework of mpol_new() to extract the adjusting of the node mask to
apply cpuset and mpol flags "context" breaks set_mempolicy() and mbind()
with MPOL_PREFERRED and a NULL nodemask--i.e., explicit local
allocation. Fix this by adding the check for MPOL_PREFERRED and empty
node mask to mpol_new_mpolicy().
Remove the now unneeded 'nodes = NULL' from mpol_new().
Note that mpol_new_mempolicy() is always called with a non-NULL
'nodes' parameter now that it has been removed from mpol_new().
Therefore, we don't need to test nodes for NULL before testing it for
'empty'. However, just to be extra paranoid, add a VM_BUG_ON() to
verify this assumption.]
[lee.schermerhorn@hp.com:
I don't think the function name 'mpol_new_mempolicy' is descriptive
enough to differentiate it from mpol_new().
This function applies cpuset set context, usually constraining nodes
to those allowed by the cpuset. However, when the 'RELATIVE_NODES flag
is set, it also translates the nodes. So I settled on
'mpol_set_nodemask()', because the comment block for mpol_new() mentions
that we need to call this function to "set nodes".
Some additional minor line length, whitespace and typo cleanup.]
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 15:31:49 -07:00
|
|
|
|
2010-05-24 14:32:04 -07:00
|
|
|
|
2023-10-03 02:22:59 -07:00
|
|
|
|
2006-01-14 13:20:48 -08:00
|
|
|
|
2023-10-03 02:22:59 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-05-25 23:42:58 -07:00
|
|
|
|
mm: make set_mempolicy(MPOL_INTERLEAV) N_HIGH_MEMORY aware
At first, init_task's mems_allowed is initialized as this.
init_task->mems_allowed == node_state[N_POSSIBLE]
And cpuset's top_cpuset mask is initialized as this
top_cpuset->mems_allowed = node_state[N_HIGH_MEMORY]
Before 2.6.29:
policy's mems_allowed is initialized as this.
1. update tasks->mems_allowed by its cpuset->mems_allowed.
2. policy->mems_allowed = nodes_and(tasks->mems_allowed, user's mask)
Updating task's mems_allowed in reference to top_cpuset's one.
cpuset's mems_allowed is aware of N_HIGH_MEMORY, always.
In 2.6.30: After commit 58568d2a8215cb6f55caf2332017d7bdff954e1c
("cpuset,mm: update tasks' mems_allowed in time"), policy's mems_allowed
is initialized as this.
1. policy->mems_allowd = nodes_and(task->mems_allowed, user's mask)
Here, if task is in top_cpuset, task->mems_allowed is not updated from
init's one. Assume user excutes command as #numactrl --interleave=all
,....
policy->mems_allowd = nodes_and(N_POSSIBLE, ALL_SET_MASK)
Then, policy's mems_allowd can includes a possible node, which has no pgdat.
MPOL's INTERLEAVE just scans nodemask of task->mems_allowd and access this
directly.
NODE_DATA(nid)->zonelist even if NODE_DATA(nid)==NULL
Then, what's we need is making policy->mems_allowed be aware of
N_HIGH_MEMORY. This patch does that. But to do so, extra nodemask will
be on statck. Because I know cpumask has a new interface of
CPUMASK_ALLOC(), I added it to node.
This patch stands on old behavior. But I feel this fix itself is just a
Band-Aid. But to do fundametal fix, we have to take care of memory
hotplug and it takes time. (task->mems_allowd should be N_HIGH_MEMORY, I
think.)
mpol_set_nodemask() should be aware of N_HIGH_MEMORY and policy's nodemask
should be includes only online nodes.
In old behavior, this is guaranteed by frequent reference to cpuset's
code. Now, most of them are removed and mempolicy has to check it by
itself.
To do check, a few nodemask_t will be used for calculating nodemask. But,
size of nodemask_t can be big and it's not good to allocate them on stack.
Now, cpumask_t has CPUMASK_ALLOC/FREE an easy code for get scratch area.
NODEMASK_ALLOC/FREE shoudl be there.
[akpm@linux-foundation.org: cleanups & tweaks]
Tested-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Miao Xie <miaox@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Paul Menage <menage@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: David Rientjes <rientjes@google.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-08-06 15:07:33 -07:00
|
|
|
|
2010-06-29 15:05:30 -07:00
|
|
|
|
|
|
|
|
|
2006-01-14 13:20:48 -08:00
|
|
|
|
|
|
|
|
|
2025-08-27 17:52:45 +00:00
|
|
|
|
2006-01-14 13:20:48 -08:00
|
|
|
|
2023-10-03 02:20:14 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-03 02:20:14 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-03 02:20:14 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2012-10-08 16:29:19 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2025-08-27 17:52:45 +00:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2023-10-03 02:20:14 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-03 02:20:14 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2023-10-03 02:20:14 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-03 02:20:14 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2023-10-03 02:20:14 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2025-08-27 17:52:45 +00:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2012-11-22 11:16:36 +00:00
|
|
|
|
2014-01-29 14:05:42 -08:00
|
|
|
|
2012-11-22 11:16:36 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-01-29 14:05:42 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-05-14 15:17:09 -07:00
|
|
|
|
2016-03-17 14:19:47 -07:00
|
|
|
|
2014-01-29 14:05:42 -08:00
|
|
|
|
2012-11-22 11:16:36 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-01-29 14:05:42 -08:00
|
|
|
|
2012-11-22 11:16:36 +00:00
|
|
|
|
|
|
|
|
|
2014-01-29 14:05:42 -08:00
|
|
|
|
2012-11-22 11:16:36 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-01-29 14:05:43 -08:00
|
|
|
|
2012-11-22 11:16:36 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2007-07-15 23:38:15 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2007-07-20 10:11:58 +09:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-07-20 10:11:58 +09:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2012-11-02 18:19:13 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-30 18:51:10 -07:00
|
|
|
|
2012-11-02 18:19:13 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-07-15 23:38:15 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-12-12 13:51:33 -08:00
|
|
|
|
2007-07-15 23:38:15 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2007-07-15 23:38:15 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-04-28 02:12:25 -07:00
|
|
|
|
2014-06-06 14:38:30 -07:00
|
|
|
|
2012-11-22 11:16:36 +00:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2005-10-29 18:16:59 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2008-04-28 02:12:25 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
[PATCH] cpusets: automatic numa mempolicy rebinding
This patch automatically updates a tasks NUMA mempolicy when its cpuset
memory placement changes. It does so within the context of the task,
without any need to support low level external mempolicy manipulation.
If a system is not using cpusets, or if running on a system with just the
root (all-encompassing) cpuset, then this remap is a no-op. Only when a
task is moved between cpusets, or a cpusets memory placement is changed
does the following apply. Otherwise, the main routine below,
rebind_policy() is not even called.
When mixing cpusets, scheduler affinity, and NUMA mempolicies, the
essential role of cpusets is to place jobs (several related tasks) on a set
of CPUs and Memory Nodes, the essential role of sched_setaffinity is to
manage a jobs processor placement within its allowed cpuset, and the
essential role of NUMA mempolicy (mbind, set_mempolicy) is to manage a jobs
memory placement within its allowed cpuset.
However, CPU affinity and NUMA memory placement are managed within the
kernel using absolute system wide numbering, not cpuset relative numbering.
This is ok until a job is migrated to a different cpuset, or what's the
same, a jobs cpuset is moved to different CPUs and Memory Nodes.
Then the CPU affinity and NUMA memory placement of the tasks in the job
need to be updated, to preserve their cpuset-relative position. This can
be done for CPU affinity using sched_setaffinity() from user code, as one
task can modify anothers CPU affinity. This cannot be done from an
external task for NUMA memory placement, as that can only be modified in
the context of the task using it.
However, it easy enough to remap a tasks NUMA mempolicy automatically when
a task is migrated, using the existing cpuset mechanism to trigger a
refresh of a tasks memory placement after its cpuset has changed. All that
is needed is the old and new nodemask, and notice to the task that it needs
to rebind its mempolicy. The tasks mems_allowed has the old mask, the
tasks cpuset has the new mask, and the existing
cpuset_update_current_mems_allowed() mechanism provides the notice. The
bitmap/cpumask/nodemask remap operators provide the cpuset relative
calculations.
This patch leaves open a couple of issues:
1) Updating vma and shmfs/tmpfs/hugetlbfs memory policies:
These mempolicies may reference nodes outside of those allowed to
the current task by its cpuset. Tasks are migrated as part of jobs,
which reside on what might be several cpusets in a subtree. When such
a job is migrated, all NUMA memory policy references to nodes within
that cpuset subtree should be translated, and references to any nodes
outside that subtree should be left untouched. A future patch will
provide the cpuset mechanism needed to mark such subtrees. With that
patch, we will be able to correctly migrate these other memory policies
across a job migration.
2) Updating cpuset, affinity and memory policies in user space:
This is harder. Any placement state stored in user space using
system-wide numbering will be invalidated across a migration. More
work will be required to provide user code with a migration-safe means
to manage its cpuset relative placement, while preserving the current
API's that pass system wide numbers, not cpuset relative numbers across
the kernel-user boundary.
Signed-off-by: Paul Jackson <pj@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-10-30 15:02:36 -08:00
|
|
|
|
mempolicy: rework shmem mpol parsing and display
mm/shmem.c currently contains functions to parse and display memory policy
strings for the tmpfs 'mpol' mount option. Move this to mm/mempolicy.c with
the rest of the mempolicy support. With subsequent patches, we'll be able to
remove knowledge of the details [mode, flags, policy, ...] completely from
shmem.c
1) replace shmem_parse_mpol() in mm/shmem.c with mpol_parse_str() in
mm/mempolicy.c. Rework to use the policy_types[] array [used by
mpol_to_str()] to look up mode by name.
2) use mpol_to_str() to format policy for shmem_show_mpol(). mpol_to_str()
expects a pointer to a struct mempolicy, so temporarily construct one.
This will be replaced with a reference to a struct mempolicy in the tmpfs
superblock in a subsequent patch.
NOTE 1: I changed mpol_to_str() to use a colon ':' rather than an equal
sign '=' as the nodemask delimiter to match mpol_parse_str() and the
tmpfs/shmem mpol mount option formatting that now uses mpol_to_str(). This
is a user visible change to numa_maps, but then the addition of the mode
flags already changed the display. It makes sense to me to have the mounts
and numa_maps display the policy in the same format. However, if anyone
objects strongly, I can pass the desired nodemask delimeter as an arg to
mpol_to_str().
Note 2: Like show_numa_map(), I don't check the return code from
mpol_to_str(). I do use a longer buffer than the one provided by
show_numa_map(), which seems to have sufficed so far.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-05-24 14:32:04 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
2012-10-25 14:16:29 +02:00
|
|
|
|
mm/mempolicy: add MPOL_PREFERRED_MANY for multiple preferred nodes
Patch series "Introduce multi-preference mempolicy", v7.
This patch series introduces the concept of the MPOL_PREFERRED_MANY
mempolicy. This mempolicy mode can be used with either the
set_mempolicy(2) or mbind(2) interfaces. Like the MPOL_PREFERRED
interface, it allows an application to set a preference for nodes which
will fulfil memory allocation requests. Unlike the MPOL_PREFERRED mode,
it takes a set of nodes. Like the MPOL_BIND interface, it works over a
set of nodes. Unlike MPOL_BIND, it will not cause a SIGSEGV or invoke the
OOM killer if those preferred nodes are not available.
Along with these patches are patches for libnuma, numactl, numademo, and
memhog. They still need some polish, but can be found here:
https://gitlab.com/bwidawsk/numactl/-/tree/prefer-many It allows new
usage: `numactl -P 0,3,4`
The goal of the new mode is to enable some use-cases when using tiered memory
usage models which I've lovingly named.
1a. The Hare - The interconnect is fast enough to meet bandwidth and
latency requirements allowing preference to be given to all nodes with
"fast" memory.
1b. The Indiscriminate Hare - An application knows it wants fast
memory (or perhaps slow memory), but doesn't care which node it runs
on. The application can prefer a set of nodes and then xpu bind to
the local node (cpu, accelerator, etc). This reverses the nodes are
chosen today where the kernel attempts to use local memory to the CPU
whenever possible. This will attempt to use the local accelerator to
the memory.
2. The Tortoise - The administrator (or the application itself) is
aware it only needs slow memory, and so can prefer that.
Much of this is almost achievable with the bind interface, but the bind
interface suffers from an inability to fallback to another set of nodes if
binding fails to all nodes in the nodemask.
Like MPOL_BIND a nodemask is given. Inherently this removes ordering from the
preference.
> /* Set first two nodes as preferred in an 8 node system. */
> const unsigned long nodes = 0x3
> set_mempolicy(MPOL_PREFER_MANY, &nodes, 8);
> /* Mimic interleave policy, but have fallback *.
> const unsigned long nodes = 0xaa
> set_mempolicy(MPOL_PREFER_MANY, &nodes, 8);
Some internal discussion took place around the interface. There are two
alternatives which we have discussed, plus one I stuck in:
1. Ordered list of nodes. Currently it's believed that the added
complexity is nod needed for expected usecases.
2. A flag for bind to allow falling back to other nodes. This
confuses the notion of binding and is less flexible than the current
solution.
3. Create flags or new modes that helps with some ordering. This
offers both a friendlier API as well as a solution for more customized
usage. It's unknown if it's worth the complexity to support this.
Here is sample code for how this might work:
> // Prefer specific nodes for some something wacky
> set_mempolicy(MPOL_PREFER_MANY, 0x17c, 1024);
>
> // Default
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_SOCKET, NULL, 0);
> // which is the same as
> set_mempolicy(MPOL_DEFAULT, NULL, 0);
>
> // The Hare
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE, NULL, 0);
>
> // The Tortoise
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE_REV, NULL, 0);
>
> // Prefer the fast memory of the first two sockets
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE, -1, 2);
>
This patch (of 5):
The NUMA APIs currently allow passing in a "preferred node" as a single
bit set in a nodemask. If more than one bit it set, bits after the first
are ignored.
This single node is generally OK for location-based NUMA where memory
being allocated will eventually be operated on by a single CPU. However,
in systems with multiple memory types, folks want to target a *type* of
memory instead of a location. For instance, someone might want some
high-bandwidth memory but do not care about the CPU next to which it is
allocated. Or, they want a cheap, high capacity allocation and want to
target all NUMA nodes which have persistent memory in volatile mode. In
both of these cases, the application wants to target a *set* of nodes, but
does not want strict MPOL_BIND behavior as that could lead to OOM killer
or SIGSEGV.
So add MPOL_PREFERRED_MANY policy to support the multiple preferred nodes
requirement. This is not a pie-in-the-sky dream for an API. This was a
response to a specific ask of more than one group at Intel. Specifically:
1. There are existing libraries that target memory types such as
https://github.com/memkind/memkind. These are known to suffer from
SIGSEGV's when memory is low on targeted memory "kinds" that span more
than one node. The MCDRAM on a Xeon Phi in "Cluster on Die" mode is an
example of this.
2. Volatile-use persistent memory users want to have a memory policy
which is targeted at either "cheap and slow" (PMEM) or "expensive and
fast" (DRAM). However, they do not want to experience allocation
failures when the targeted type is unavailable.
3. Allocate-then-run. Generally, we let the process scheduler decide
on which physical CPU to run a task. That location provides a default
allocation policy, and memory availability is not generally considered
when placing tasks. For situations where memory is valuable and
constrained, some users want to allocate memory first, *then* allocate
close compute resources to the allocation. This is the reverse of the
normal (CPU) model. Accelerators such as GPUs that operate on
core-mm-managed memory are interested in this model.
A check is added in sanitize_mpol_flags() to not permit 'prefer_many'
policy to be used for now, and will be removed in later patch after all
implementations for 'prefer_many' are ready, as suggested by Michal Hocko.
[mhocko@kernel.org: suggest to refine policy_node/policy_nodemask handling]
Link: https://lkml.kernel.org/r/1627970362-61305-1-git-send-email-feng.tang@intel.com
Link: https://lore.kernel.org/r/20200630212517.308045-4-ben.widawsky@intel.com
Link: https://lkml.kernel.org/r/1627970362-61305-2-git-send-email-feng.tang@intel.com
Co-developed-by: Ben Widawsky <ben.widawsky@intel.com>
Signed-off-by: Ben Widawsky <ben.widawsky@intel.com>
Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com>
Signed-off-by: Feng Tang <feng.tang@intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Huang Ying <ying.huang@intel.com>b
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-09-02 15:00:06 -07:00
|
|
|
|
2010-05-24 14:32:04 -07:00
|
|
|
|
2006-01-08 01:01:02 -08:00
|
|
|
|
mempolicy: rework shmem mpol parsing and display
mm/shmem.c currently contains functions to parse and display memory policy
strings for the tmpfs 'mpol' mount option. Move this to mm/mempolicy.c with
the rest of the mempolicy support. With subsequent patches, we'll be able to
remove knowledge of the details [mode, flags, policy, ...] completely from
shmem.c
1) replace shmem_parse_mpol() in mm/shmem.c with mpol_parse_str() in
mm/mempolicy.c. Rework to use the policy_types[] array [used by
mpol_to_str()] to look up mode by name.
2) use mpol_to_str() to format policy for shmem_show_mpol(). mpol_to_str()
expects a pointer to a struct mempolicy, so temporarily construct one.
This will be replaced with a reference to a struct mempolicy in the tmpfs
superblock in a subsequent patch.
NOTE 1: I changed mpol_to_str() to use a colon ':' rather than an equal
sign '=' as the nodemask delimiter to match mpol_parse_str() and the
tmpfs/shmem mpol mount option formatting that now uses mpol_to_str(). This
is a user visible change to numa_maps, but then the addition of the mode
flags already changed the display. It makes sense to me to have the mounts
and numa_maps display the policy in the same format. However, if anyone
objects strongly, I can pass the desired nodemask delimeter as an arg to
mpol_to_str().
Note 2: Like show_numa_map(), I don't check the return code from
mpol_to_str(). I do use a longer buffer than the one provided by
show_numa_map(), which seems to have sufficed so far.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:23 -07:00
|
|
|
|
|
|
|
|
|
tmpfs mempolicy: fix /proc/mounts corrupting memory
Recently I suggested using "mount -o remount,mpol=local /tmp" in NUMA
mempolicy testing. Very nasty. Reading /proc/mounts, /proc/pid/mounts
or /proc/pid/mountinfo may then corrupt one bit of kernel memory, often
in a page table (causing "Bad swap" or "Bad page map" warning or "Bad
pagetable" oops), sometimes in a vm_area_struct or rbnode or somewhere
worse. "mpol=prefer" and "mpol=prefer:Node" are equally toxic.
Recent NUMA enhancements are not to blame: this dates back to 2.6.35,
when commit e17f74af351c "mempolicy: don't call mpol_set_nodemask() when
no_context" skipped mpol_parse_str()'s call to mpol_set_nodemask(),
which used to initialize v.preferred_node, or set MPOL_F_LOCAL in flags.
With slab poisoning, you can then rely on mpol_to_str() to set the bit
for node 0x6b6b, probably in the next page above the caller's stack.
mpol_parse_str() is only called from shmem_parse_options(): no_context
is always true, so call it unused for now, and remove !no_context code.
Set v.nodes or v.preferred_node or MPOL_F_LOCAL as mpol_to_str() might
expect. Then mpol_to_str() can ignore its no_context argument also,
the mpol being appropriately initialized whether contextualized or not.
Rename its no_context unused too, and let subsequent patch remove them
(that's not needed for stable backporting, which would involve rejects).
I don't understand why MPOL_LOCAL is described as a pseudo-policy:
it's a reasonable policy which suffers from a confusing implementation
in terms of MPOL_PREFERRED with MPOL_F_LOCAL. I believe this would be
much more robust if MPOL_LOCAL were recognized in switch statements
throughout, MPOL_F_LOCAL deleted, and MPOL_PREFERRED use the (possibly
empty) nodes mask like everyone else, instead of its preferred_node
variant (I presume an optimization from the days before MPOL_LOCAL).
But that would take me too long to get right and fully tested.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: stable@vger.kernel.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-01-02 02:01:33 -08:00
|
|
|
|
mempolicy: rework shmem mpol parsing and display
mm/shmem.c currently contains functions to parse and display memory policy
strings for the tmpfs 'mpol' mount option. Move this to mm/mempolicy.c with
the rest of the mempolicy support. With subsequent patches, we'll be able to
remove knowledge of the details [mode, flags, policy, ...] completely from
shmem.c
1) replace shmem_parse_mpol() in mm/shmem.c with mpol_parse_str() in
mm/mempolicy.c. Rework to use the policy_types[] array [used by
mpol_to_str()] to look up mode by name.
2) use mpol_to_str() to format policy for shmem_show_mpol(). mpol_to_str()
expects a pointer to a struct mempolicy, so temporarily construct one.
This will be replaced with a reference to a struct mempolicy in the tmpfs
superblock in a subsequent patch.
NOTE 1: I changed mpol_to_str() to use a colon ':' rather than an equal
sign '=' as the nodemask delimiter to match mpol_parse_str() and the
tmpfs/shmem mpol mount option formatting that now uses mpol_to_str(). This
is a user visible change to numa_maps, but then the addition of the mode
flags already changed the display. It makes sense to me to have the mounts
and numa_maps display the policy in the same format. However, if anyone
objects strongly, I can pass the desired nodemask delimeter as an arg to
mpol_to_str().
Note 2: Like show_numa_map(), I don't check the return code from
mpol_to_str(). I do use a longer buffer than the one provided by
show_numa_map(), which seems to have sufficed so far.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:23 -07:00
|
|
|
|
mempolicy: use struct mempolicy pointer in shmem_sb_info
This patch replaces the mempolicy mode, mode_flags, and nodemask in the
shmem_sb_info struct with a struct mempolicy pointer, initialized to NULL.
This removes dependency on the details of mempolicy from shmem.c and hugetlbfs
inode.c and simplifies the interfaces.
mpol_parse_str() in mempolicy.c is changed to return, via a pointer to a
pointer arg, a struct mempolicy pointer on success. For MPOL_DEFAULT, the
returned pointer is NULL. Further, mpol_parse_str() now takes a 'no_context'
argument that causes the input nodemask to be stored in the w.user_nodemask of
the created mempolicy for use when the mempolicy is installed in a tmpfs inode
shared policy tree. At that time, any cpuset contextualization is applied to
the original input nodemask. This preserves the previous behavior where the
input nodemask was stored in the superblock. We can think of the returned
mempolicy as "context free".
Because mpol_parse_str() is now calling mpol_new(), we can remove from
mpol_to_str() the semantic checks that mpol_new() already performs.
Add 'no_context' parameter to mpol_to_str() to specify that it should format
the nodemask in w.user_nodemask for 'bind' and 'interleave' policies.
Change mpol_shared_policy_init() to take a pointer to a "context free" struct
mempolicy and to create a new, "contextualized" mempolicy using the mode,
mode_flags and user_nodemask from the input mempolicy.
Note: we know that the mempolicy passed to mpol_to_str() or
mpol_shared_policy_init() from a tmpfs superblock is "context free". This
is currently the only instance thereof. However, if we found more uses for
this concept, and introduced any ambiguity as to whether a mempolicy was
context free or not, we could add another internal mode flag to identify
context free mempolicies. Then, we could remove the 'no_context' argument
from mpol_to_str().
Added shmem_get_sbmpol() to return a reference counted superblock mempolicy,
if one exists, to pass to mpol_shared_policy_init(). We must add the
reference under the sb stat_lock to prevent races with replacement of the mpol
by remount. This reference is removed in mpol_shared_policy_init().
[akpm@linux-foundation.org: build fix]
[akpm@linux-foundation.org: another build fix]
[akpm@linux-foundation.org: yet another build fix]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:26 -07:00
|
|
|
|
mempolicy: rework shmem mpol parsing and display
mm/shmem.c currently contains functions to parse and display memory policy
strings for the tmpfs 'mpol' mount option. Move this to mm/mempolicy.c with
the rest of the mempolicy support. With subsequent patches, we'll be able to
remove knowledge of the details [mode, flags, policy, ...] completely from
shmem.c
1) replace shmem_parse_mpol() in mm/shmem.c with mpol_parse_str() in
mm/mempolicy.c. Rework to use the policy_types[] array [used by
mpol_to_str()] to look up mode by name.
2) use mpol_to_str() to format policy for shmem_show_mpol(). mpol_to_str()
expects a pointer to a struct mempolicy, so temporarily construct one.
This will be replaced with a reference to a struct mempolicy in the tmpfs
superblock in a subsequent patch.
NOTE 1: I changed mpol_to_str() to use a colon ':' rather than an equal
sign '=' as the nodemask delimiter to match mpol_parse_str() and the
tmpfs/shmem mpol mount option formatting that now uses mpol_to_str(). This
is a user visible change to numa_maps, but then the addition of the mode
flags already changed the display. It makes sense to me to have the mounts
and numa_maps display the policy in the same format. However, if anyone
objects strongly, I can pass the desired nodemask delimeter as an arg to
mpol_to_str().
Note 2: Like show_numa_map(), I don't check the return code from
mpol_to_str(). I do use a longer buffer than the one provided by
show_numa_map(), which seems to have sufficed so far.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-01-14 14:08:24 -08:00
|
|
|
|
mempolicy: rework shmem mpol parsing and display
mm/shmem.c currently contains functions to parse and display memory policy
strings for the tmpfs 'mpol' mount option. Move this to mm/mempolicy.c with
the rest of the mempolicy support. With subsequent patches, we'll be able to
remove knowledge of the details [mode, flags, policy, ...] completely from
shmem.c
1) replace shmem_parse_mpol() in mm/shmem.c with mpol_parse_str() in
mm/mempolicy.c. Rework to use the policy_types[] array [used by
mpol_to_str()] to look up mode by name.
2) use mpol_to_str() to format policy for shmem_show_mpol(). mpol_to_str()
expects a pointer to a struct mempolicy, so temporarily construct one.
This will be replaced with a reference to a struct mempolicy in the tmpfs
superblock in a subsequent patch.
NOTE 1: I changed mpol_to_str() to use a colon ':' rather than an equal
sign '=' as the nodemask delimiter to match mpol_parse_str() and the
tmpfs/shmem mpol mount option formatting that now uses mpol_to_str(). This
is a user visible change to numa_maps, but then the addition of the mode
flags already changed the display. It makes sense to me to have the mounts
and numa_maps display the policy in the same format. However, if anyone
objects strongly, I can pass the desired nodemask delimeter as an arg to
mpol_to_str().
Note 2: Like show_numa_map(), I don't check the return code from
mpol_to_str(). I do use a longer buffer than the one provided by
show_numa_map(), which seems to have sufficed so far.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:23 -07:00
|
|
|
|
2013-01-02 02:04:23 -08:00
|
|
|
|
mempolicy: rework shmem mpol parsing and display
mm/shmem.c currently contains functions to parse and display memory policy
strings for the tmpfs 'mpol' mount option. Move this to mm/mempolicy.c with
the rest of the mempolicy support. With subsequent patches, we'll be able to
remove knowledge of the details [mode, flags, policy, ...] completely from
shmem.c
1) replace shmem_parse_mpol() in mm/shmem.c with mpol_parse_str() in
mm/mempolicy.c. Rework to use the policy_types[] array [used by
mpol_to_str()] to look up mode by name.
2) use mpol_to_str() to format policy for shmem_show_mpol(). mpol_to_str()
expects a pointer to a struct mempolicy, so temporarily construct one.
This will be replaced with a reference to a struct mempolicy in the tmpfs
superblock in a subsequent patch.
NOTE 1: I changed mpol_to_str() to use a colon ':' rather than an equal
sign '=' as the nodemask delimiter to match mpol_parse_str() and the
tmpfs/shmem mpol mount option formatting that now uses mpol_to_str(). This
is a user visible change to numa_maps, but then the addition of the mode
flags already changed the display. It makes sense to me to have the mounts
and numa_maps display the policy in the same format. However, if anyone
objects strongly, I can pass the desired nodemask delimeter as an arg to
mpol_to_str().
Note 2: Like show_numa_map(), I don't check the return code from
mpol_to_str(). I do use a longer buffer than the one provided by
show_numa_map(), which seems to have sufficed so far.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:23 -07:00
|
|
|
|
mempolicy: use struct mempolicy pointer in shmem_sb_info
This patch replaces the mempolicy mode, mode_flags, and nodemask in the
shmem_sb_info struct with a struct mempolicy pointer, initialized to NULL.
This removes dependency on the details of mempolicy from shmem.c and hugetlbfs
inode.c and simplifies the interfaces.
mpol_parse_str() in mempolicy.c is changed to return, via a pointer to a
pointer arg, a struct mempolicy pointer on success. For MPOL_DEFAULT, the
returned pointer is NULL. Further, mpol_parse_str() now takes a 'no_context'
argument that causes the input nodemask to be stored in the w.user_nodemask of
the created mempolicy for use when the mempolicy is installed in a tmpfs inode
shared policy tree. At that time, any cpuset contextualization is applied to
the original input nodemask. This preserves the previous behavior where the
input nodemask was stored in the superblock. We can think of the returned
mempolicy as "context free".
Because mpol_parse_str() is now calling mpol_new(), we can remove from
mpol_to_str() the semantic checks that mpol_new() already performs.
Add 'no_context' parameter to mpol_to_str() to specify that it should format
the nodemask in w.user_nodemask for 'bind' and 'interleave' policies.
Change mpol_shared_policy_init() to take a pointer to a "context free" struct
mempolicy and to create a new, "contextualized" mempolicy using the mode,
mode_flags and user_nodemask from the input mempolicy.
Note: we know that the mempolicy passed to mpol_to_str() or
mpol_shared_policy_init() from a tmpfs superblock is "context free". This
is currently the only instance thereof. However, if we found more uses for
this concept, and introduced any ambiguity as to whether a mempolicy was
context free or not, we could add another internal mode flag to identify
context free mempolicies. Then, we could remove the 'no_context' argument
from mpol_to_str().
Added shmem_get_sbmpol() to return a reference counted superblock mempolicy,
if one exists, to pass to mpol_shared_policy_init(). We must add the
reference under the sb stat_lock to prevent races with replacement of the mpol
by remount. This reference is removed in mpol_shared_policy_init().
[akpm@linux-foundation.org: build fix]
[akpm@linux-foundation.org: another build fix]
[akpm@linux-foundation.org: yet another build fix]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:26 -07:00
|
|
|
|
tmpfs mempolicy: fix /proc/mounts corrupting memory
Recently I suggested using "mount -o remount,mpol=local /tmp" in NUMA
mempolicy testing. Very nasty. Reading /proc/mounts, /proc/pid/mounts
or /proc/pid/mountinfo may then corrupt one bit of kernel memory, often
in a page table (causing "Bad swap" or "Bad page map" warning or "Bad
pagetable" oops), sometimes in a vm_area_struct or rbnode or somewhere
worse. "mpol=prefer" and "mpol=prefer:Node" are equally toxic.
Recent NUMA enhancements are not to blame: this dates back to 2.6.35,
when commit e17f74af351c "mempolicy: don't call mpol_set_nodemask() when
no_context" skipped mpol_parse_str()'s call to mpol_set_nodemask(),
which used to initialize v.preferred_node, or set MPOL_F_LOCAL in flags.
With slab poisoning, you can then rely on mpol_to_str() to set the bit
for node 0x6b6b, probably in the next page above the caller's stack.
mpol_parse_str() is only called from shmem_parse_options(): no_context
is always true, so call it unused for now, and remove !no_context code.
Set v.nodes or v.preferred_node or MPOL_F_LOCAL as mpol_to_str() might
expect. Then mpol_to_str() can ignore its no_context argument also,
the mpol being appropriately initialized whether contextualized or not.
Rename its no_context unused too, and let subsequent patch remove them
(that's not needed for stable backporting, which would involve rejects).
I don't understand why MPOL_LOCAL is described as a pseudo-policy:
it's a reasonable policy which suffers from a confusing implementation
in terms of MPOL_PREFERRED with MPOL_F_LOCAL. I believe this would be
much more robust if MPOL_LOCAL were recognized in switch statements
throughout, MPOL_F_LOCAL deleted, and MPOL_PREFERRED use the (possibly
empty) nodes mask like everyone else, instead of its preferred_node
variant (I presume an optimization from the days before MPOL_LOCAL).
But that would take me too long to get right and fully tested.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: stable@vger.kernel.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-01-02 02:01:33 -08:00
|
|
|
|
mempolicy: use struct mempolicy pointer in shmem_sb_info
This patch replaces the mempolicy mode, mode_flags, and nodemask in the
shmem_sb_info struct with a struct mempolicy pointer, initialized to NULL.
This removes dependency on the details of mempolicy from shmem.c and hugetlbfs
inode.c and simplifies the interfaces.
mpol_parse_str() in mempolicy.c is changed to return, via a pointer to a
pointer arg, a struct mempolicy pointer on success. For MPOL_DEFAULT, the
returned pointer is NULL. Further, mpol_parse_str() now takes a 'no_context'
argument that causes the input nodemask to be stored in the w.user_nodemask of
the created mempolicy for use when the mempolicy is installed in a tmpfs inode
shared policy tree. At that time, any cpuset contextualization is applied to
the original input nodemask. This preserves the previous behavior where the
input nodemask was stored in the superblock. We can think of the returned
mempolicy as "context free".
Because mpol_parse_str() is now calling mpol_new(), we can remove from
mpol_to_str() the semantic checks that mpol_new() already performs.
Add 'no_context' parameter to mpol_to_str() to specify that it should format
the nodemask in w.user_nodemask for 'bind' and 'interleave' policies.
Change mpol_shared_policy_init() to take a pointer to a "context free" struct
mempolicy and to create a new, "contextualized" mempolicy using the mode,
mode_flags and user_nodemask from the input mempolicy.
Note: we know that the mempolicy passed to mpol_to_str() or
mpol_shared_policy_init() from a tmpfs superblock is "context free". This
is currently the only instance thereof. However, if we found more uses for
this concept, and introduced any ambiguity as to whether a mempolicy was
context free or not, we could add another internal mode flag to identify
context free mempolicies. Then, we could remove the 'no_context' argument
from mpol_to_str().
Added shmem_get_sbmpol() to return a reference counted superblock mempolicy,
if one exists, to pass to mpol_shared_policy_init(). We must add the
reference under the sb stat_lock to prevent races with replacement of the mpol
by remount. This reference is removed in mpol_shared_policy_init().
[akpm@linux-foundation.org: build fix]
[akpm@linux-foundation.org: another build fix]
[akpm@linux-foundation.org: yet another build fix]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:26 -07:00
|
|
|
|
mempolicy: rework shmem mpol parsing and display
mm/shmem.c currently contains functions to parse and display memory policy
strings for the tmpfs 'mpol' mount option. Move this to mm/mempolicy.c with
the rest of the mempolicy support. With subsequent patches, we'll be able to
remove knowledge of the details [mode, flags, policy, ...] completely from
shmem.c
1) replace shmem_parse_mpol() in mm/shmem.c with mpol_parse_str() in
mm/mempolicy.c. Rework to use the policy_types[] array [used by
mpol_to_str()] to look up mode by name.
2) use mpol_to_str() to format policy for shmem_show_mpol(). mpol_to_str()
expects a pointer to a struct mempolicy, so temporarily construct one.
This will be replaced with a reference to a struct mempolicy in the tmpfs
superblock in a subsequent patch.
NOTE 1: I changed mpol_to_str() to use a colon ':' rather than an equal
sign '=' as the nodemask delimiter to match mpol_parse_str() and the
tmpfs/shmem mpol mount option formatting that now uses mpol_to_str(). This
is a user visible change to numa_maps, but then the addition of the mode
flags already changed the display. It makes sense to me to have the mounts
and numa_maps display the policy in the same format. However, if anyone
objects strongly, I can pass the desired nodemask delimeter as an arg to
mpol_to_str().
Note 2: Like show_numa_map(), I don't check the return code from
mpol_to_str(). I do use a longer buffer than the one provided by
show_numa_map(), which seems to have sufficed so far.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:23 -07:00
|
|
|
|
|
|
|
|
|
2018-10-26 15:06:57 -07:00
|
|
|
|
mempolicy: rework shmem mpol parsing and display
mm/shmem.c currently contains functions to parse and display memory policy
strings for the tmpfs 'mpol' mount option. Move this to mm/mempolicy.c with
the rest of the mempolicy support. With subsequent patches, we'll be able to
remove knowledge of the details [mode, flags, policy, ...] completely from
shmem.c
1) replace shmem_parse_mpol() in mm/shmem.c with mpol_parse_str() in
mm/mempolicy.c. Rework to use the policy_types[] array [used by
mpol_to_str()] to look up mode by name.
2) use mpol_to_str() to format policy for shmem_show_mpol(). mpol_to_str()
expects a pointer to a struct mempolicy, so temporarily construct one.
This will be replaced with a reference to a struct mempolicy in the tmpfs
superblock in a subsequent patch.
NOTE 1: I changed mpol_to_str() to use a colon ':' rather than an equal
sign '=' as the nodemask delimiter to match mpol_parse_str() and the
tmpfs/shmem mpol mount option formatting that now uses mpol_to_str(). This
is a user visible change to numa_maps, but then the addition of the mode
flags already changed the display. It makes sense to me to have the mounts
and numa_maps display the policy in the same format. However, if anyone
objects strongly, I can pass the desired nodemask delimeter as an arg to
mpol_to_str().
Note 2: Like show_numa_map(), I don't check the return code from
mpol_to_str(). I do use a longer buffer than the one provided by
show_numa_map(), which seems to have sufficed so far.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:23 -07:00
|
|
|
|
2020-01-30 22:11:07 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: rework shmem mpol parsing and display
mm/shmem.c currently contains functions to parse and display memory policy
strings for the tmpfs 'mpol' mount option. Move this to mm/mempolicy.c with
the rest of the mempolicy support. With subsequent patches, we'll be able to
remove knowledge of the details [mode, flags, policy, ...] completely from
shmem.c
1) replace shmem_parse_mpol() in mm/shmem.c with mpol_parse_str() in
mm/mempolicy.c. Rework to use the policy_types[] array [used by
mpol_to_str()] to look up mode by name.
2) use mpol_to_str() to format policy for shmem_show_mpol(). mpol_to_str()
expects a pointer to a struct mempolicy, so temporarily construct one.
This will be replaced with a reference to a struct mempolicy in the tmpfs
superblock in a subsequent patch.
NOTE 1: I changed mpol_to_str() to use a colon ':' rather than an equal
sign '=' as the nodemask delimiter to match mpol_parse_str() and the
tmpfs/shmem mpol mount option formatting that now uses mpol_to_str(). This
is a user visible change to numa_maps, but then the addition of the mode
flags already changed the display. It makes sense to me to have the mounts
and numa_maps display the policy in the same format. However, if anyone
objects strongly, I can pass the desired nodemask delimeter as an arg to
mpol_to_str().
Note 2: Like show_numa_map(), I don't check the return code from
mpol_to_str(). I do use a longer buffer than the one provided by
show_numa_map(), which seems to have sufficed so far.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: use struct mempolicy pointer in shmem_sb_info
This patch replaces the mempolicy mode, mode_flags, and nodemask in the
shmem_sb_info struct with a struct mempolicy pointer, initialized to NULL.
This removes dependency on the details of mempolicy from shmem.c and hugetlbfs
inode.c and simplifies the interfaces.
mpol_parse_str() in mempolicy.c is changed to return, via a pointer to a
pointer arg, a struct mempolicy pointer on success. For MPOL_DEFAULT, the
returned pointer is NULL. Further, mpol_parse_str() now takes a 'no_context'
argument that causes the input nodemask to be stored in the w.user_nodemask of
the created mempolicy for use when the mempolicy is installed in a tmpfs inode
shared policy tree. At that time, any cpuset contextualization is applied to
the original input nodemask. This preserves the previous behavior where the
input nodemask was stored in the superblock. We can think of the returned
mempolicy as "context free".
Because mpol_parse_str() is now calling mpol_new(), we can remove from
mpol_to_str() the semantic checks that mpol_new() already performs.
Add 'no_context' parameter to mpol_to_str() to specify that it should format
the nodemask in w.user_nodemask for 'bind' and 'interleave' policies.
Change mpol_shared_policy_init() to take a pointer to a "context free" struct
mempolicy and to create a new, "contextualized" mempolicy using the mode,
mode_flags and user_nodemask from the input mempolicy.
Note: we know that the mempolicy passed to mpol_to_str() or
mpol_shared_policy_init() from a tmpfs superblock is "context free". This
is currently the only instance thereof. However, if we found more uses for
this concept, and introduced any ambiguity as to whether a mempolicy was
context free or not, we could add another internal mode flag to identify
context free mempolicies. Then, we could remove the 'no_context' argument
from mpol_to_str().
Added shmem_get_sbmpol() to return a reference counted superblock mempolicy,
if one exists, to pass to mpol_shared_policy_init(). We must add the
reference under the sb stat_lock to prevent races with replacement of the mpol
by remount. This reference is removed in mpol_shared_policy_init().
[akpm@linux-foundation.org: build fix]
[akpm@linux-foundation.org: another build fix]
[akpm@linux-foundation.org: yet another build fix]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:26 -07:00
|
|
|
|
mempolicy: rework shmem mpol parsing and display
mm/shmem.c currently contains functions to parse and display memory policy
strings for the tmpfs 'mpol' mount option. Move this to mm/mempolicy.c with
the rest of the mempolicy support. With subsequent patches, we'll be able to
remove knowledge of the details [mode, flags, policy, ...] completely from
shmem.c
1) replace shmem_parse_mpol() in mm/shmem.c with mpol_parse_str() in
mm/mempolicy.c. Rework to use the policy_types[] array [used by
mpol_to_str()] to look up mode by name.
2) use mpol_to_str() to format policy for shmem_show_mpol(). mpol_to_str()
expects a pointer to a struct mempolicy, so temporarily construct one.
This will be replaced with a reference to a struct mempolicy in the tmpfs
superblock in a subsequent patch.
NOTE 1: I changed mpol_to_str() to use a colon ':' rather than an equal
sign '=' as the nodemask delimiter to match mpol_parse_str() and the
tmpfs/shmem mpol mount option formatting that now uses mpol_to_str(). This
is a user visible change to numa_maps, but then the addition of the mode
flags already changed the display. It makes sense to me to have the mounts
and numa_maps display the policy in the same format. However, if anyone
objects strongly, I can pass the desired nodemask delimeter as an arg to
mpol_to_str().
Note 2: Like show_numa_map(), I don't check the return code from
mpol_to_str(). I do use a longer buffer than the one provided by
show_numa_map(), which seems to have sufficed so far.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:23 -07:00
|
|
|
|
2012-12-12 13:51:33 -08:00
|
|
|
|
mempolicy: rework shmem mpol parsing and display
mm/shmem.c currently contains functions to parse and display memory policy
strings for the tmpfs 'mpol' mount option. Move this to mm/mempolicy.c with
the rest of the mempolicy support. With subsequent patches, we'll be able to
remove knowledge of the details [mode, flags, policy, ...] completely from
shmem.c
1) replace shmem_parse_mpol() in mm/shmem.c with mpol_parse_str() in
mm/mempolicy.c. Rework to use the policy_types[] array [used by
mpol_to_str()] to look up mode by name.
2) use mpol_to_str() to format policy for shmem_show_mpol(). mpol_to_str()
expects a pointer to a struct mempolicy, so temporarily construct one.
This will be replaced with a reference to a struct mempolicy in the tmpfs
superblock in a subsequent patch.
NOTE 1: I changed mpol_to_str() to use a colon ':' rather than an equal
sign '=' as the nodemask delimiter to match mpol_parse_str() and the
tmpfs/shmem mpol mount option formatting that now uses mpol_to_str(). This
is a user visible change to numa_maps, but then the addition of the mode
flags already changed the display. It makes sense to me to have the mounts
and numa_maps display the policy in the same format. However, if anyone
objects strongly, I can pass the desired nodemask delimeter as an arg to
mpol_to_str().
Note 2: Like show_numa_map(), I don't check the return code from
mpol_to_str(). I do use a longer buffer than the one provided by
show_numa_map(), which seems to have sufficed so far.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:23 -07:00
|
|
|
|
mempolicy: use struct mempolicy pointer in shmem_sb_info
This patch replaces the mempolicy mode, mode_flags, and nodemask in the
shmem_sb_info struct with a struct mempolicy pointer, initialized to NULL.
This removes dependency on the details of mempolicy from shmem.c and hugetlbfs
inode.c and simplifies the interfaces.
mpol_parse_str() in mempolicy.c is changed to return, via a pointer to a
pointer arg, a struct mempolicy pointer on success. For MPOL_DEFAULT, the
returned pointer is NULL. Further, mpol_parse_str() now takes a 'no_context'
argument that causes the input nodemask to be stored in the w.user_nodemask of
the created mempolicy for use when the mempolicy is installed in a tmpfs inode
shared policy tree. At that time, any cpuset contextualization is applied to
the original input nodemask. This preserves the previous behavior where the
input nodemask was stored in the superblock. We can think of the returned
mempolicy as "context free".
Because mpol_parse_str() is now calling mpol_new(), we can remove from
mpol_to_str() the semantic checks that mpol_new() already performs.
Add 'no_context' parameter to mpol_to_str() to specify that it should format
the nodemask in w.user_nodemask for 'bind' and 'interleave' policies.
Change mpol_shared_policy_init() to take a pointer to a "context free" struct
mempolicy and to create a new, "contextualized" mempolicy using the mode,
mode_flags and user_nodemask from the input mempolicy.
Note: we know that the mempolicy passed to mpol_to_str() or
mpol_shared_policy_init() from a tmpfs superblock is "context free". This
is currently the only instance thereof. However, if we found more uses for
this concept, and introduced any ambiguity as to whether a mempolicy was
context free or not, we could add another internal mode flag to identify
context free mempolicies. Then, we could remove the 'no_context' argument
from mpol_to_str().
Added shmem_get_sbmpol() to return a reference counted superblock mempolicy,
if one exists, to pass to mpol_shared_policy_init(). We must add the
reference under the sb stat_lock to prevent races with replacement of the mpol
by remount. This reference is removed in mpol_shared_policy_init().
[akpm@linux-foundation.org: build fix]
[akpm@linux-foundation.org: another build fix]
[akpm@linux-foundation.org: yet another build fix]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:26 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-10-26 15:06:57 -07:00
|
|
|
|
|
|
|
|
|
mempolicy: rework shmem mpol parsing and display
mm/shmem.c currently contains functions to parse and display memory policy
strings for the tmpfs 'mpol' mount option. Move this to mm/mempolicy.c with
the rest of the mempolicy support. With subsequent patches, we'll be able to
remove knowledge of the details [mode, flags, policy, ...] completely from
shmem.c
1) replace shmem_parse_mpol() in mm/shmem.c with mpol_parse_str() in
mm/mempolicy.c. Rework to use the policy_types[] array [used by
mpol_to_str()] to look up mode by name.
2) use mpol_to_str() to format policy for shmem_show_mpol(). mpol_to_str()
expects a pointer to a struct mempolicy, so temporarily construct one.
This will be replaced with a reference to a struct mempolicy in the tmpfs
superblock in a subsequent patch.
NOTE 1: I changed mpol_to_str() to use a colon ':' rather than an equal
sign '=' as the nodemask delimiter to match mpol_parse_str() and the
tmpfs/shmem mpol mount option formatting that now uses mpol_to_str(). This
is a user visible change to numa_maps, but then the addition of the mode
flags already changed the display. It makes sense to me to have the mounts
and numa_maps display the policy in the same format. However, if anyone
objects strongly, I can pass the desired nodemask delimeter as an arg to
mpol_to_str().
Note 2: Like show_numa_map(), I don't check the return code from
mpol_to_str(). I do use a longer buffer than the one provided by
show_numa_map(), which seems to have sufficed so far.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:23 -07:00
|
|
|
|
|
|
|
|
|
mempolicy: use struct mempolicy pointer in shmem_sb_info
This patch replaces the mempolicy mode, mode_flags, and nodemask in the
shmem_sb_info struct with a struct mempolicy pointer, initialized to NULL.
This removes dependency on the details of mempolicy from shmem.c and hugetlbfs
inode.c and simplifies the interfaces.
mpol_parse_str() in mempolicy.c is changed to return, via a pointer to a
pointer arg, a struct mempolicy pointer on success. For MPOL_DEFAULT, the
returned pointer is NULL. Further, mpol_parse_str() now takes a 'no_context'
argument that causes the input nodemask to be stored in the w.user_nodemask of
the created mempolicy for use when the mempolicy is installed in a tmpfs inode
shared policy tree. At that time, any cpuset contextualization is applied to
the original input nodemask. This preserves the previous behavior where the
input nodemask was stored in the superblock. We can think of the returned
mempolicy as "context free".
Because mpol_parse_str() is now calling mpol_new(), we can remove from
mpol_to_str() the semantic checks that mpol_new() already performs.
Add 'no_context' parameter to mpol_to_str() to specify that it should format
the nodemask in w.user_nodemask for 'bind' and 'interleave' policies.
Change mpol_shared_policy_init() to take a pointer to a "context free" struct
mempolicy and to create a new, "contextualized" mempolicy using the mode,
mode_flags and user_nodemask from the input mempolicy.
Note: we know that the mempolicy passed to mpol_to_str() or
mpol_shared_policy_init() from a tmpfs superblock is "context free". This
is currently the only instance thereof. However, if we found more uses for
this concept, and introduced any ambiguity as to whether a mempolicy was
context free or not, we could add another internal mode flag to identify
context free mempolicies. Then, we could remove the 'no_context' argument
from mpol_to_str().
Added shmem_get_sbmpol() to return a reference counted superblock mempolicy,
if one exists, to pass to mpol_shared_policy_init(). We must add the
reference under the sb stat_lock to prevent races with replacement of the mpol
by remount. This reference is removed in mpol_shared_policy_init().
[akpm@linux-foundation.org: build fix]
[akpm@linux-foundation.org: another build fix]
[akpm@linux-foundation.org: yet another build fix]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:26 -07:00
|
|
|
|
mempolicy: rework shmem mpol parsing and display
mm/shmem.c currently contains functions to parse and display memory policy
strings for the tmpfs 'mpol' mount option. Move this to mm/mempolicy.c with
the rest of the mempolicy support. With subsequent patches, we'll be able to
remove knowledge of the details [mode, flags, policy, ...] completely from
shmem.c
1) replace shmem_parse_mpol() in mm/shmem.c with mpol_parse_str() in
mm/mempolicy.c. Rework to use the policy_types[] array [used by
mpol_to_str()] to look up mode by name.
2) use mpol_to_str() to format policy for shmem_show_mpol(). mpol_to_str()
expects a pointer to a struct mempolicy, so temporarily construct one.
This will be replaced with a reference to a struct mempolicy in the tmpfs
superblock in a subsequent patch.
NOTE 1: I changed mpol_to_str() to use a colon ':' rather than an equal
sign '=' as the nodemask delimiter to match mpol_parse_str() and the
tmpfs/shmem mpol mount option formatting that now uses mpol_to_str(). This
is a user visible change to numa_maps, but then the addition of the mode
flags already changed the display. It makes sense to me to have the mounts
and numa_maps display the policy in the same format. However, if anyone
objects strongly, I can pass the desired nodemask delimeter as an arg to
mpol_to_str().
Note 2: Like show_numa_map(), I don't check the return code from
mpol_to_str(). I do use a longer buffer than the one provided by
show_numa_map(), which seems to have sufficed so far.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:23 -07:00
|
|
|
|
mempolicy: use struct mempolicy pointer in shmem_sb_info
This patch replaces the mempolicy mode, mode_flags, and nodemask in the
shmem_sb_info struct with a struct mempolicy pointer, initialized to NULL.
This removes dependency on the details of mempolicy from shmem.c and hugetlbfs
inode.c and simplifies the interfaces.
mpol_parse_str() in mempolicy.c is changed to return, via a pointer to a
pointer arg, a struct mempolicy pointer on success. For MPOL_DEFAULT, the
returned pointer is NULL. Further, mpol_parse_str() now takes a 'no_context'
argument that causes the input nodemask to be stored in the w.user_nodemask of
the created mempolicy for use when the mempolicy is installed in a tmpfs inode
shared policy tree. At that time, any cpuset contextualization is applied to
the original input nodemask. This preserves the previous behavior where the
input nodemask was stored in the superblock. We can think of the returned
mempolicy as "context free".
Because mpol_parse_str() is now calling mpol_new(), we can remove from
mpol_to_str() the semantic checks that mpol_new() already performs.
Add 'no_context' parameter to mpol_to_str() to specify that it should format
the nodemask in w.user_nodemask for 'bind' and 'interleave' policies.
Change mpol_shared_policy_init() to take a pointer to a "context free" struct
mempolicy and to create a new, "contextualized" mempolicy using the mode,
mode_flags and user_nodemask from the input mempolicy.
Note: we know that the mempolicy passed to mpol_to_str() or
mpol_shared_policy_init() from a tmpfs superblock is "context free". This
is currently the only instance thereof. However, if we found more uses for
this concept, and introduced any ambiguity as to whether a mempolicy was
context free or not, we could add another internal mode flag to identify
context free mempolicies. Then, we could remove the 'no_context' argument
from mpol_to_str().
Added shmem_get_sbmpol() to return a reference counted superblock mempolicy,
if one exists, to pass to mpol_shared_policy_init(). We must add the
reference under the sb stat_lock to prevent races with replacement of the mpol
by remount. This reference is removed in mpol_shared_policy_init().
[akpm@linux-foundation.org: build fix]
[akpm@linux-foundation.org: another build fix]
[akpm@linux-foundation.org: yet another build fix]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:26 -07:00
|
|
|
|
2020-04-01 21:10:58 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: use struct mempolicy pointer in shmem_sb_info
This patch replaces the mempolicy mode, mode_flags, and nodemask in the
shmem_sb_info struct with a struct mempolicy pointer, initialized to NULL.
This removes dependency on the details of mempolicy from shmem.c and hugetlbfs
inode.c and simplifies the interfaces.
mpol_parse_str() in mempolicy.c is changed to return, via a pointer to a
pointer arg, a struct mempolicy pointer on success. For MPOL_DEFAULT, the
returned pointer is NULL. Further, mpol_parse_str() now takes a 'no_context'
argument that causes the input nodemask to be stored in the w.user_nodemask of
the created mempolicy for use when the mempolicy is installed in a tmpfs inode
shared policy tree. At that time, any cpuset contextualization is applied to
the original input nodemask. This preserves the previous behavior where the
input nodemask was stored in the superblock. We can think of the returned
mempolicy as "context free".
Because mpol_parse_str() is now calling mpol_new(), we can remove from
mpol_to_str() the semantic checks that mpol_new() already performs.
Add 'no_context' parameter to mpol_to_str() to specify that it should format
the nodemask in w.user_nodemask for 'bind' and 'interleave' policies.
Change mpol_shared_policy_init() to take a pointer to a "context free" struct
mempolicy and to create a new, "contextualized" mempolicy using the mode,
mode_flags and user_nodemask from the input mempolicy.
Note: we know that the mempolicy passed to mpol_to_str() or
mpol_shared_policy_init() from a tmpfs superblock is "context free". This
is currently the only instance thereof. However, if we found more uses for
this concept, and introduced any ambiguity as to whether a mempolicy was
context free or not, we could add another internal mode flag to identify
context free mempolicies. Then, we could remove the 'no_context' argument
from mpol_to_str().
Added shmem_get_sbmpol() to return a reference counted superblock mempolicy,
if one exists, to pass to mpol_shared_policy_init(). We must add the
reference under the sb stat_lock to prevent races with replacement of the mpol
by remount. This reference is removed in mpol_shared_policy_init().
[akpm@linux-foundation.org: build fix]
[akpm@linux-foundation.org: another build fix]
[akpm@linux-foundation.org: yet another build fix]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:26 -07:00
|
|
|
|
mempolicy: rework shmem mpol parsing and display
mm/shmem.c currently contains functions to parse and display memory policy
strings for the tmpfs 'mpol' mount option. Move this to mm/mempolicy.c with
the rest of the mempolicy support. With subsequent patches, we'll be able to
remove knowledge of the details [mode, flags, policy, ...] completely from
shmem.c
1) replace shmem_parse_mpol() in mm/shmem.c with mpol_parse_str() in
mm/mempolicy.c. Rework to use the policy_types[] array [used by
mpol_to_str()] to look up mode by name.
2) use mpol_to_str() to format policy for shmem_show_mpol(). mpol_to_str()
expects a pointer to a struct mempolicy, so temporarily construct one.
This will be replaced with a reference to a struct mempolicy in the tmpfs
superblock in a subsequent patch.
NOTE 1: I changed mpol_to_str() to use a colon ':' rather than an equal
sign '=' as the nodemask delimiter to match mpol_parse_str() and the
tmpfs/shmem mpol mount option formatting that now uses mpol_to_str(). This
is a user visible change to numa_maps, but then the addition of the mode
flags already changed the display. It makes sense to me to have the mounts
and numa_maps display the policy in the same format. However, if anyone
objects strongly, I can pass the desired nodemask delimeter as an arg to
mpol_to_str().
Note 2: Like show_numa_map(), I don't check the return code from
mpol_to_str(). I do use a longer buffer than the one provided by
show_numa_map(), which seems to have sufficed so far.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-03-23 13:35:32 -07:00
|
|
|
|
|
|
|
|
|
2020-04-01 21:10:58 -07:00
|
|
|
|
|
|
|
|
|
mempolicy: rework shmem mpol parsing and display
mm/shmem.c currently contains functions to parse and display memory policy
strings for the tmpfs 'mpol' mount option. Move this to mm/mempolicy.c with
the rest of the mempolicy support. With subsequent patches, we'll be able to
remove knowledge of the details [mode, flags, policy, ...] completely from
shmem.c
1) replace shmem_parse_mpol() in mm/shmem.c with mpol_parse_str() in
mm/mempolicy.c. Rework to use the policy_types[] array [used by
mpol_to_str()] to look up mode by name.
2) use mpol_to_str() to format policy for shmem_show_mpol(). mpol_to_str()
expects a pointer to a struct mempolicy, so temporarily construct one.
This will be replaced with a reference to a struct mempolicy in the tmpfs
superblock in a subsequent patch.
NOTE 1: I changed mpol_to_str() to use a colon ':' rather than an equal
sign '=' as the nodemask delimiter to match mpol_parse_str() and the
tmpfs/shmem mpol mount option formatting that now uses mpol_to_str(). This
is a user visible change to numa_maps, but then the addition of the mode
flags already changed the display. It makes sense to me to have the mounts
and numa_maps display the policy in the same format. However, if anyone
objects strongly, I can pass the desired nodemask delimeter as an arg to
mpol_to_str().
Note 2: Like show_numa_map(), I don't check the return code from
mpol_to_str(). I do use a longer buffer than the one provided by
show_numa_map(), which seems to have sufficed so far.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
mempolicy: rework shmem mpol parsing and display
mm/shmem.c currently contains functions to parse and display memory policy
strings for the tmpfs 'mpol' mount option. Move this to mm/mempolicy.c with
the rest of the mempolicy support. With subsequent patches, we'll be able to
remove knowledge of the details [mode, flags, policy, ...] completely from
shmem.c
1) replace shmem_parse_mpol() in mm/shmem.c with mpol_parse_str() in
mm/mempolicy.c. Rework to use the policy_types[] array [used by
mpol_to_str()] to look up mode by name.
2) use mpol_to_str() to format policy for shmem_show_mpol(). mpol_to_str()
expects a pointer to a struct mempolicy, so temporarily construct one.
This will be replaced with a reference to a struct mempolicy in the tmpfs
superblock in a subsequent patch.
NOTE 1: I changed mpol_to_str() to use a colon ':' rather than an equal
sign '=' as the nodemask delimiter to match mpol_parse_str() and the
tmpfs/shmem mpol mount option formatting that now uses mpol_to_str(). This
is a user visible change to numa_maps, but then the addition of the mode
flags already changed the display. It makes sense to me to have the mounts
and numa_maps display the policy in the same format. However, if anyone
objects strongly, I can pass the desired nodemask delimeter as an arg to
mpol_to_str().
Note 2: Like show_numa_map(), I don't check the return code from
mpol_to_str(). I do use a longer buffer than the one provided by
show_numa_map(), which seems to have sufficed so far.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-12-12 13:51:33 -08:00
|
|
|
|
2008-04-28 02:13:24 -07:00
|
|
|
|
mempolicy: use struct mempolicy pointer in shmem_sb_info
This patch replaces the mempolicy mode, mode_flags, and nodemask in the
shmem_sb_info struct with a struct mempolicy pointer, initialized to NULL.
This removes dependency on the details of mempolicy from shmem.c and hugetlbfs
inode.c and simplifies the interfaces.
mpol_parse_str() in mempolicy.c is changed to return, via a pointer to a
pointer arg, a struct mempolicy pointer on success. For MPOL_DEFAULT, the
returned pointer is NULL. Further, mpol_parse_str() now takes a 'no_context'
argument that causes the input nodemask to be stored in the w.user_nodemask of
the created mempolicy for use when the mempolicy is installed in a tmpfs inode
shared policy tree. At that time, any cpuset contextualization is applied to
the original input nodemask. This preserves the previous behavior where the
input nodemask was stored in the superblock. We can think of the returned
mempolicy as "context free".
Because mpol_parse_str() is now calling mpol_new(), we can remove from
mpol_to_str() the semantic checks that mpol_new() already performs.
Add 'no_context' parameter to mpol_to_str() to specify that it should format
the nodemask in w.user_nodemask for 'bind' and 'interleave' policies.
Change mpol_shared_policy_init() to take a pointer to a "context free" struct
mempolicy and to create a new, "contextualized" mempolicy using the mode,
mode_flags and user_nodemask from the input mempolicy.
Note: we know that the mempolicy passed to mpol_to_str() or
mpol_shared_policy_init() from a tmpfs superblock is "context free". This
is currently the only instance thereof. However, if we found more uses for
this concept, and introduced any ambiguity as to whether a mempolicy was
context free or not, we could add another internal mode flag to identify
context free mempolicies. Then, we could remove the 'no_context' argument
from mpol_to_str().
Added shmem_get_sbmpol() to return a reference counted superblock mempolicy,
if one exists, to pass to mpol_shared_policy_init(). We must add the
reference under the sb stat_lock to prevent races with replacement of the mpol
by remount. This reference is removed in mpol_shared_policy_init().
[akpm@linux-foundation.org: build fix]
[akpm@linux-foundation.org: another build fix]
[akpm@linux-foundation.org: yet another build fix]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:26 -07:00
|
|
|
|
2008-04-28 02:13:24 -07:00
|
|
|
|
mempolicy: use struct mempolicy pointer in shmem_sb_info
This patch replaces the mempolicy mode, mode_flags, and nodemask in the
shmem_sb_info struct with a struct mempolicy pointer, initialized to NULL.
This removes dependency on the details of mempolicy from shmem.c and hugetlbfs
inode.c and simplifies the interfaces.
mpol_parse_str() in mempolicy.c is changed to return, via a pointer to a
pointer arg, a struct mempolicy pointer on success. For MPOL_DEFAULT, the
returned pointer is NULL. Further, mpol_parse_str() now takes a 'no_context'
argument that causes the input nodemask to be stored in the w.user_nodemask of
the created mempolicy for use when the mempolicy is installed in a tmpfs inode
shared policy tree. At that time, any cpuset contextualization is applied to
the original input nodemask. This preserves the previous behavior where the
input nodemask was stored in the superblock. We can think of the returned
mempolicy as "context free".
Because mpol_parse_str() is now calling mpol_new(), we can remove from
mpol_to_str() the semantic checks that mpol_new() already performs.
Add 'no_context' parameter to mpol_to_str() to specify that it should format
the nodemask in w.user_nodemask for 'bind' and 'interleave' policies.
Change mpol_shared_policy_init() to take a pointer to a "context free" struct
mempolicy and to create a new, "contextualized" mempolicy using the mode,
mode_flags and user_nodemask from the input mempolicy.
Note: we know that the mempolicy passed to mpol_to_str() or
mpol_shared_policy_init() from a tmpfs superblock is "context free". This
is currently the only instance thereof. However, if we found more uses for
this concept, and introduced any ambiguity as to whether a mempolicy was
context free or not, we could add another internal mode flag to identify
context free mempolicies. Then, we could remove the 'no_context' argument
from mpol_to_str().
Added shmem_get_sbmpol() to return a reference counted superblock mempolicy,
if one exists, to pass to mpol_shared_policy_init(). We must add the
reference under the sb stat_lock to prevent races with replacement of the mpol
by remount. This reference is removed in mpol_shared_policy_init().
[akpm@linux-foundation.org: build fix]
[akpm@linux-foundation.org: another build fix]
[akpm@linux-foundation.org: yet another build fix]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:26 -07:00
|
|
|
|
2008-04-28 02:13:24 -07:00
|
|
|
|
mempolicy: use struct mempolicy pointer in shmem_sb_info
This patch replaces the mempolicy mode, mode_flags, and nodemask in the
shmem_sb_info struct with a struct mempolicy pointer, initialized to NULL.
This removes dependency on the details of mempolicy from shmem.c and hugetlbfs
inode.c and simplifies the interfaces.
mpol_parse_str() in mempolicy.c is changed to return, via a pointer to a
pointer arg, a struct mempolicy pointer on success. For MPOL_DEFAULT, the
returned pointer is NULL. Further, mpol_parse_str() now takes a 'no_context'
argument that causes the input nodemask to be stored in the w.user_nodemask of
the created mempolicy for use when the mempolicy is installed in a tmpfs inode
shared policy tree. At that time, any cpuset contextualization is applied to
the original input nodemask. This preserves the previous behavior where the
input nodemask was stored in the superblock. We can think of the returned
mempolicy as "context free".
Because mpol_parse_str() is now calling mpol_new(), we can remove from
mpol_to_str() the semantic checks that mpol_new() already performs.
Add 'no_context' parameter to mpol_to_str() to specify that it should format
the nodemask in w.user_nodemask for 'bind' and 'interleave' policies.
Change mpol_shared_policy_init() to take a pointer to a "context free" struct
mempolicy and to create a new, "contextualized" mempolicy using the mode,
mode_flags and user_nodemask from the input mempolicy.
Note: we know that the mempolicy passed to mpol_to_str() or
mpol_shared_policy_init() from a tmpfs superblock is "context free". This
is currently the only instance thereof. However, if we found more uses for
this concept, and introduced any ambiguity as to whether a mempolicy was
context free or not, we could add another internal mode flag to identify
context free mempolicies. Then, we could remove the 'no_context' argument
from mpol_to_str().
Added shmem_get_sbmpol() to return a reference counted superblock mempolicy,
if one exists, to pass to mpol_shared_policy_init(). We must add the
reference under the sb stat_lock to prevent races with replacement of the mpol
by remount. This reference is removed in mpol_shared_policy_init().
[akpm@linux-foundation.org: build fix]
[akpm@linux-foundation.org: another build fix]
[akpm@linux-foundation.org: yet another build fix]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:26 -07:00
|
|
|
|
2008-04-28 02:13:24 -07:00
|
|
|
|
|
|
|
|
|
2010-03-23 13:35:28 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: add MPOL_PREFERRED_MANY for multiple preferred nodes
Patch series "Introduce multi-preference mempolicy", v7.
This patch series introduces the concept of the MPOL_PREFERRED_MANY
mempolicy. This mempolicy mode can be used with either the
set_mempolicy(2) or mbind(2) interfaces. Like the MPOL_PREFERRED
interface, it allows an application to set a preference for nodes which
will fulfil memory allocation requests. Unlike the MPOL_PREFERRED mode,
it takes a set of nodes. Like the MPOL_BIND interface, it works over a
set of nodes. Unlike MPOL_BIND, it will not cause a SIGSEGV or invoke the
OOM killer if those preferred nodes are not available.
Along with these patches are patches for libnuma, numactl, numademo, and
memhog. They still need some polish, but can be found here:
https://gitlab.com/bwidawsk/numactl/-/tree/prefer-many It allows new
usage: `numactl -P 0,3,4`
The goal of the new mode is to enable some use-cases when using tiered memory
usage models which I've lovingly named.
1a. The Hare - The interconnect is fast enough to meet bandwidth and
latency requirements allowing preference to be given to all nodes with
"fast" memory.
1b. The Indiscriminate Hare - An application knows it wants fast
memory (or perhaps slow memory), but doesn't care which node it runs
on. The application can prefer a set of nodes and then xpu bind to
the local node (cpu, accelerator, etc). This reverses the nodes are
chosen today where the kernel attempts to use local memory to the CPU
whenever possible. This will attempt to use the local accelerator to
the memory.
2. The Tortoise - The administrator (or the application itself) is
aware it only needs slow memory, and so can prefer that.
Much of this is almost achievable with the bind interface, but the bind
interface suffers from an inability to fallback to another set of nodes if
binding fails to all nodes in the nodemask.
Like MPOL_BIND a nodemask is given. Inherently this removes ordering from the
preference.
> /* Set first two nodes as preferred in an 8 node system. */
> const unsigned long nodes = 0x3
> set_mempolicy(MPOL_PREFER_MANY, &nodes, 8);
> /* Mimic interleave policy, but have fallback *.
> const unsigned long nodes = 0xaa
> set_mempolicy(MPOL_PREFER_MANY, &nodes, 8);
Some internal discussion took place around the interface. There are two
alternatives which we have discussed, plus one I stuck in:
1. Ordered list of nodes. Currently it's believed that the added
complexity is nod needed for expected usecases.
2. A flag for bind to allow falling back to other nodes. This
confuses the notion of binding and is less flexible than the current
solution.
3. Create flags or new modes that helps with some ordering. This
offers both a friendlier API as well as a solution for more customized
usage. It's unknown if it's worth the complexity to support this.
Here is sample code for how this might work:
> // Prefer specific nodes for some something wacky
> set_mempolicy(MPOL_PREFER_MANY, 0x17c, 1024);
>
> // Default
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_SOCKET, NULL, 0);
> // which is the same as
> set_mempolicy(MPOL_DEFAULT, NULL, 0);
>
> // The Hare
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE, NULL, 0);
>
> // The Tortoise
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE_REV, NULL, 0);
>
> // Prefer the fast memory of the first two sockets
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE, -1, 2);
>
This patch (of 5):
The NUMA APIs currently allow passing in a "preferred node" as a single
bit set in a nodemask. If more than one bit it set, bits after the first
are ignored.
This single node is generally OK for location-based NUMA where memory
being allocated will eventually be operated on by a single CPU. However,
in systems with multiple memory types, folks want to target a *type* of
memory instead of a location. For instance, someone might want some
high-bandwidth memory but do not care about the CPU next to which it is
allocated. Or, they want a cheap, high capacity allocation and want to
target all NUMA nodes which have persistent memory in volatile mode. In
both of these cases, the application wants to target a *set* of nodes, but
does not want strict MPOL_BIND behavior as that could lead to OOM killer
or SIGSEGV.
So add MPOL_PREFERRED_MANY policy to support the multiple preferred nodes
requirement. This is not a pie-in-the-sky dream for an API. This was a
response to a specific ask of more than one group at Intel. Specifically:
1. There are existing libraries that target memory types such as
https://github.com/memkind/memkind. These are known to suffer from
SIGSEGV's when memory is low on targeted memory "kinds" that span more
than one node. The MCDRAM on a Xeon Phi in "Cluster on Die" mode is an
example of this.
2. Volatile-use persistent memory users want to have a memory policy
which is targeted at either "cheap and slow" (PMEM) or "expensive and
fast" (DRAM). However, they do not want to experience allocation
failures when the targeted type is unavailable.
3. Allocate-then-run. Generally, we let the process scheduler decide
on which physical CPU to run a task. That location provides a default
allocation policy, and memory availability is not generally considered
when placing tasks. For situations where memory is valuable and
constrained, some users want to allocate memory first, *then* allocate
close compute resources to the allocation. This is the reverse of the
normal (CPU) model. Accelerators such as GPUs that operate on
core-mm-managed memory are interested in this model.
A check is added in sanitize_mpol_flags() to not permit 'prefer_many'
policy to be used for now, and will be removed in later patch after all
implementations for 'prefer_many' are ready, as suggested by Michal Hocko.
[mhocko@kernel.org: suggest to refine policy_node/policy_nodemask handling]
Link: https://lkml.kernel.org/r/1627970362-61305-1-git-send-email-feng.tang@intel.com
Link: https://lore.kernel.org/r/20200630212517.308045-4-ben.widawsky@intel.com
Link: https://lkml.kernel.org/r/1627970362-61305-2-git-send-email-feng.tang@intel.com
Co-developed-by: Ben Widawsky <ben.widawsky@intel.com>
Signed-off-by: Ben Widawsky <ben.widawsky@intel.com>
Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com>
Signed-off-by: Feng Tang <feng.tang@intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Huang Ying <ying.huang@intel.com>b
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-09-02 15:00:06 -07:00
|
|
|
|
2010-03-23 13:35:30 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: rework shmem mpol parsing and display
mm/shmem.c currently contains functions to parse and display memory policy
strings for the tmpfs 'mpol' mount option. Move this to mm/mempolicy.c with
the rest of the mempolicy support. With subsequent patches, we'll be able to
remove knowledge of the details [mode, flags, policy, ...] completely from
shmem.c
1) replace shmem_parse_mpol() in mm/shmem.c with mpol_parse_str() in
mm/mempolicy.c. Rework to use the policy_types[] array [used by
mpol_to_str()] to look up mode by name.
2) use mpol_to_str() to format policy for shmem_show_mpol(). mpol_to_str()
expects a pointer to a struct mempolicy, so temporarily construct one.
This will be replaced with a reference to a struct mempolicy in the tmpfs
superblock in a subsequent patch.
NOTE 1: I changed mpol_to_str() to use a colon ':' rather than an equal
sign '=' as the nodemask delimiter to match mpol_parse_str() and the
tmpfs/shmem mpol mount option formatting that now uses mpol_to_str(). This
is a user visible change to numa_maps, but then the addition of the mode
flags already changed the display. It makes sense to me to have the mounts
and numa_maps display the policy in the same format. However, if anyone
objects strongly, I can pass the desired nodemask delimeter as an arg to
mpol_to_str().
Note 2: Like show_numa_map(), I don't check the return code from
mpol_to_str(). I do use a longer buffer than the one provided by
show_numa_map(), which seems to have sufficed so far.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:23 -07:00
|
|
|
|
|
|
|
|
|
mempolicy: use struct mempolicy pointer in shmem_sb_info
This patch replaces the mempolicy mode, mode_flags, and nodemask in the
shmem_sb_info struct with a struct mempolicy pointer, initialized to NULL.
This removes dependency on the details of mempolicy from shmem.c and hugetlbfs
inode.c and simplifies the interfaces.
mpol_parse_str() in mempolicy.c is changed to return, via a pointer to a
pointer arg, a struct mempolicy pointer on success. For MPOL_DEFAULT, the
returned pointer is NULL. Further, mpol_parse_str() now takes a 'no_context'
argument that causes the input nodemask to be stored in the w.user_nodemask of
the created mempolicy for use when the mempolicy is installed in a tmpfs inode
shared policy tree. At that time, any cpuset contextualization is applied to
the original input nodemask. This preserves the previous behavior where the
input nodemask was stored in the superblock. We can think of the returned
mempolicy as "context free".
Because mpol_parse_str() is now calling mpol_new(), we can remove from
mpol_to_str() the semantic checks that mpol_new() already performs.
Add 'no_context' parameter to mpol_to_str() to specify that it should format
the nodemask in w.user_nodemask for 'bind' and 'interleave' policies.
Change mpol_shared_policy_init() to take a pointer to a "context free" struct
mempolicy and to create a new, "contextualized" mempolicy using the mode,
mode_flags and user_nodemask from the input mempolicy.
Note: we know that the mempolicy passed to mpol_to_str() or
mpol_shared_policy_init() from a tmpfs superblock is "context free". This
is currently the only instance thereof. However, if we found more uses for
this concept, and introduced any ambiguity as to whether a mempolicy was
context free or not, we could add another internal mode flag to identify
context free mempolicies. Then, we could remove the 'no_context' argument
from mpol_to_str().
Added shmem_get_sbmpol() to return a reference counted superblock mempolicy,
if one exists, to pass to mpol_shared_policy_init(). We must add the
reference under the sb stat_lock to prevent races with replacement of the mpol
by remount. This reference is removed in mpol_shared_policy_init().
[akpm@linux-foundation.org: build fix]
[akpm@linux-foundation.org: another build fix]
[akpm@linux-foundation.org: yet another build fix]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:26 -07:00
|
|
|
|
mempolicy: rework shmem mpol parsing and display
mm/shmem.c currently contains functions to parse and display memory policy
strings for the tmpfs 'mpol' mount option. Move this to mm/mempolicy.c with
the rest of the mempolicy support. With subsequent patches, we'll be able to
remove knowledge of the details [mode, flags, policy, ...] completely from
shmem.c
1) replace shmem_parse_mpol() in mm/shmem.c with mpol_parse_str() in
mm/mempolicy.c. Rework to use the policy_types[] array [used by
mpol_to_str()] to look up mode by name.
2) use mpol_to_str() to format policy for shmem_show_mpol(). mpol_to_str()
expects a pointer to a struct mempolicy, so temporarily construct one.
This will be replaced with a reference to a struct mempolicy in the tmpfs
superblock in a subsequent patch.
NOTE 1: I changed mpol_to_str() to use a colon ':' rather than an equal
sign '=' as the nodemask delimiter to match mpol_parse_str() and the
tmpfs/shmem mpol mount option formatting that now uses mpol_to_str(). This
is a user visible change to numa_maps, but then the addition of the mode
flags already changed the display. It makes sense to me to have the mounts
and numa_maps display the policy in the same format. However, if anyone
objects strongly, I can pass the desired nodemask delimeter as an arg to
mpol_to_str().
Note 2: Like show_numa_map(), I don't check the return code from
mpol_to_str(). I do use a longer buffer than the one provided by
show_numa_map(), which seems to have sufficed so far.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: use struct mempolicy pointer in shmem_sb_info
This patch replaces the mempolicy mode, mode_flags, and nodemask in the
shmem_sb_info struct with a struct mempolicy pointer, initialized to NULL.
This removes dependency on the details of mempolicy from shmem.c and hugetlbfs
inode.c and simplifies the interfaces.
mpol_parse_str() in mempolicy.c is changed to return, via a pointer to a
pointer arg, a struct mempolicy pointer on success. For MPOL_DEFAULT, the
returned pointer is NULL. Further, mpol_parse_str() now takes a 'no_context'
argument that causes the input nodemask to be stored in the w.user_nodemask of
the created mempolicy for use when the mempolicy is installed in a tmpfs inode
shared policy tree. At that time, any cpuset contextualization is applied to
the original input nodemask. This preserves the previous behavior where the
input nodemask was stored in the superblock. We can think of the returned
mempolicy as "context free".
Because mpol_parse_str() is now calling mpol_new(), we can remove from
mpol_to_str() the semantic checks that mpol_new() already performs.
Add 'no_context' parameter to mpol_to_str() to specify that it should format
the nodemask in w.user_nodemask for 'bind' and 'interleave' policies.
Change mpol_shared_policy_init() to take a pointer to a "context free" struct
mempolicy and to create a new, "contextualized" mempolicy using the mode,
mode_flags and user_nodemask from the input mempolicy.
Note: we know that the mempolicy passed to mpol_to_str() or
mpol_shared_policy_init() from a tmpfs superblock is "context free". This
is currently the only instance thereof. However, if we found more uses for
this concept, and introduced any ambiguity as to whether a mempolicy was
context free or not, we could add another internal mode flag to identify
context free mempolicies. Then, we could remove the 'no_context' argument
from mpol_to_str().
Added shmem_get_sbmpol() to return a reference counted superblock mempolicy,
if one exists, to pass to mpol_shared_policy_init(). We must add the
reference under the sb stat_lock to prevent races with replacement of the mpol
by remount. This reference is removed in mpol_shared_policy_init().
[akpm@linux-foundation.org: build fix]
[akpm@linux-foundation.org: another build fix]
[akpm@linux-foundation.org: yet another build fix]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:26 -07:00
|
|
|
|
mempolicy: rework shmem mpol parsing and display
mm/shmem.c currently contains functions to parse and display memory policy
strings for the tmpfs 'mpol' mount option. Move this to mm/mempolicy.c with
the rest of the mempolicy support. With subsequent patches, we'll be able to
remove knowledge of the details [mode, flags, policy, ...] completely from
shmem.c
1) replace shmem_parse_mpol() in mm/shmem.c with mpol_parse_str() in
mm/mempolicy.c. Rework to use the policy_types[] array [used by
mpol_to_str()] to look up mode by name.
2) use mpol_to_str() to format policy for shmem_show_mpol(). mpol_to_str()
expects a pointer to a struct mempolicy, so temporarily construct one.
This will be replaced with a reference to a struct mempolicy in the tmpfs
superblock in a subsequent patch.
NOTE 1: I changed mpol_to_str() to use a colon ':' rather than an equal
sign '=' as the nodemask delimiter to match mpol_parse_str() and the
tmpfs/shmem mpol mount option formatting that now uses mpol_to_str(). This
is a user visible change to numa_maps, but then the addition of the mode
flags already changed the display. It makes sense to me to have the mounts
and numa_maps display the policy in the same format. However, if anyone
objects strongly, I can pass the desired nodemask delimeter as an arg to
mpol_to_str().
Note 2: Like show_numa_map(), I don't check the return code from
mpol_to_str(). I do use a longer buffer than the one provided by
show_numa_map(), which seems to have sufficed so far.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:23 -07:00
|
|
|
|
mempolicy: use struct mempolicy pointer in shmem_sb_info
This patch replaces the mempolicy mode, mode_flags, and nodemask in the
shmem_sb_info struct with a struct mempolicy pointer, initialized to NULL.
This removes dependency on the details of mempolicy from shmem.c and hugetlbfs
inode.c and simplifies the interfaces.
mpol_parse_str() in mempolicy.c is changed to return, via a pointer to a
pointer arg, a struct mempolicy pointer on success. For MPOL_DEFAULT, the
returned pointer is NULL. Further, mpol_parse_str() now takes a 'no_context'
argument that causes the input nodemask to be stored in the w.user_nodemask of
the created mempolicy for use when the mempolicy is installed in a tmpfs inode
shared policy tree. At that time, any cpuset contextualization is applied to
the original input nodemask. This preserves the previous behavior where the
input nodemask was stored in the superblock. We can think of the returned
mempolicy as "context free".
Because mpol_parse_str() is now calling mpol_new(), we can remove from
mpol_to_str() the semantic checks that mpol_new() already performs.
Add 'no_context' parameter to mpol_to_str() to specify that it should format
the nodemask in w.user_nodemask for 'bind' and 'interleave' policies.
Change mpol_shared_policy_init() to take a pointer to a "context free" struct
mempolicy and to create a new, "contextualized" mempolicy using the mode,
mode_flags and user_nodemask from the input mempolicy.
Note: we know that the mempolicy passed to mpol_to_str() or
mpol_shared_policy_init() from a tmpfs superblock is "context free". This
is currently the only instance thereof. However, if we found more uses for
this concept, and introduced any ambiguity as to whether a mempolicy was
context free or not, we could add another internal mode flag to identify
context free mempolicies. Then, we could remove the 'no_context' argument
from mpol_to_str().
Added shmem_get_sbmpol() to return a reference counted superblock mempolicy,
if one exists, to pass to mpol_shared_policy_init(). We must add the
reference under the sb stat_lock to prevent races with replacement of the mpol
by remount. This reference is removed in mpol_shared_policy_init().
[akpm@linux-foundation.org: build fix]
[akpm@linux-foundation.org: another build fix]
[akpm@linux-foundation.org: yet another build fix]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:26 -07:00
|
|
|
|
mempolicy: rework shmem mpol parsing and display
mm/shmem.c currently contains functions to parse and display memory policy
strings for the tmpfs 'mpol' mount option. Move this to mm/mempolicy.c with
the rest of the mempolicy support. With subsequent patches, we'll be able to
remove knowledge of the details [mode, flags, policy, ...] completely from
shmem.c
1) replace shmem_parse_mpol() in mm/shmem.c with mpol_parse_str() in
mm/mempolicy.c. Rework to use the policy_types[] array [used by
mpol_to_str()] to look up mode by name.
2) use mpol_to_str() to format policy for shmem_show_mpol(). mpol_to_str()
expects a pointer to a struct mempolicy, so temporarily construct one.
This will be replaced with a reference to a struct mempolicy in the tmpfs
superblock in a subsequent patch.
NOTE 1: I changed mpol_to_str() to use a colon ':' rather than an equal
sign '=' as the nodemask delimiter to match mpol_parse_str() and the
tmpfs/shmem mpol mount option formatting that now uses mpol_to_str(). This
is a user visible change to numa_maps, but then the addition of the mode
flags already changed the display. It makes sense to me to have the mounts
and numa_maps display the policy in the same format. However, if anyone
objects strongly, I can pass the desired nodemask delimeter as an arg to
mpol_to_str().
Note 2: Like show_numa_map(), I don't check the return code from
mpol_to_str(). I do use a longer buffer than the one provided by
show_numa_map(), which seems to have sufficed so far.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:23 -07:00
|
|
|
|
2010-03-23 13:35:32 -07:00
|
|
|
|
mempolicy: rework shmem mpol parsing and display
mm/shmem.c currently contains functions to parse and display memory policy
strings for the tmpfs 'mpol' mount option. Move this to mm/mempolicy.c with
the rest of the mempolicy support. With subsequent patches, we'll be able to
remove knowledge of the details [mode, flags, policy, ...] completely from
shmem.c
1) replace shmem_parse_mpol() in mm/shmem.c with mpol_parse_str() in
mm/mempolicy.c. Rework to use the policy_types[] array [used by
mpol_to_str()] to look up mode by name.
2) use mpol_to_str() to format policy for shmem_show_mpol(). mpol_to_str()
expects a pointer to a struct mempolicy, so temporarily construct one.
This will be replaced with a reference to a struct mempolicy in the tmpfs
superblock in a subsequent patch.
NOTE 1: I changed mpol_to_str() to use a colon ':' rather than an equal
sign '=' as the nodemask delimiter to match mpol_parse_str() and the
tmpfs/shmem mpol mount option formatting that now uses mpol_to_str(). This
is a user visible change to numa_maps, but then the addition of the mode
flags already changed the display. It makes sense to me to have the mounts
and numa_maps display the policy in the same format. However, if anyone
objects strongly, I can pass the desired nodemask delimeter as an arg to
mpol_to_str().
Note 2: Like show_numa_map(), I don't check the return code from
mpol_to_str(). I do use a longer buffer than the one provided by
show_numa_map(), which seems to have sufficed so far.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:23 -07:00
|
|
|
|
mempolicy: use struct mempolicy pointer in shmem_sb_info
This patch replaces the mempolicy mode, mode_flags, and nodemask in the
shmem_sb_info struct with a struct mempolicy pointer, initialized to NULL.
This removes dependency on the details of mempolicy from shmem.c and hugetlbfs
inode.c and simplifies the interfaces.
mpol_parse_str() in mempolicy.c is changed to return, via a pointer to a
pointer arg, a struct mempolicy pointer on success. For MPOL_DEFAULT, the
returned pointer is NULL. Further, mpol_parse_str() now takes a 'no_context'
argument that causes the input nodemask to be stored in the w.user_nodemask of
the created mempolicy for use when the mempolicy is installed in a tmpfs inode
shared policy tree. At that time, any cpuset contextualization is applied to
the original input nodemask. This preserves the previous behavior where the
input nodemask was stored in the superblock. We can think of the returned
mempolicy as "context free".
Because mpol_parse_str() is now calling mpol_new(), we can remove from
mpol_to_str() the semantic checks that mpol_new() already performs.
Add 'no_context' parameter to mpol_to_str() to specify that it should format
the nodemask in w.user_nodemask for 'bind' and 'interleave' policies.
Change mpol_shared_policy_init() to take a pointer to a "context free" struct
mempolicy and to create a new, "contextualized" mempolicy using the mode,
mode_flags and user_nodemask from the input mempolicy.
Note: we know that the mempolicy passed to mpol_to_str() or
mpol_shared_policy_init() from a tmpfs superblock is "context free". This
is currently the only instance thereof. However, if we found more uses for
this concept, and introduced any ambiguity as to whether a mempolicy was
context free or not, we could add another internal mode flag to identify
context free mempolicies. Then, we could remove the 'no_context' argument
from mpol_to_str().
Added shmem_get_sbmpol() to return a reference counted superblock mempolicy,
if one exists, to pass to mpol_shared_policy_init(). We must add the
reference under the sb stat_lock to prevent races with replacement of the mpol
by remount. This reference is removed in mpol_shared_policy_init().
[akpm@linux-foundation.org: build fix]
[akpm@linux-foundation.org: another build fix]
[akpm@linux-foundation.org: yet another build fix]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:26 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-03-23 13:35:32 -07:00
|
|
|
|
|
|
|
|
|
tmpfs mempolicy: fix /proc/mounts corrupting memory
Recently I suggested using "mount -o remount,mpol=local /tmp" in NUMA
mempolicy testing. Very nasty. Reading /proc/mounts, /proc/pid/mounts
or /proc/pid/mountinfo may then corrupt one bit of kernel memory, often
in a page table (causing "Bad swap" or "Bad page map" warning or "Bad
pagetable" oops), sometimes in a vm_area_struct or rbnode or somewhere
worse. "mpol=prefer" and "mpol=prefer:Node" are equally toxic.
Recent NUMA enhancements are not to blame: this dates back to 2.6.35,
when commit e17f74af351c "mempolicy: don't call mpol_set_nodemask() when
no_context" skipped mpol_parse_str()'s call to mpol_set_nodemask(),
which used to initialize v.preferred_node, or set MPOL_F_LOCAL in flags.
With slab poisoning, you can then rely on mpol_to_str() to set the bit
for node 0x6b6b, probably in the next page above the caller's stack.
mpol_parse_str() is only called from shmem_parse_options(): no_context
is always true, so call it unused for now, and remove !no_context code.
Set v.nodes or v.preferred_node or MPOL_F_LOCAL as mpol_to_str() might
expect. Then mpol_to_str() can ignore its no_context argument also,
the mpol being appropriately initialized whether contextualized or not.
Rename its no_context unused too, and let subsequent patch remove them
(that's not needed for stable backporting, which would involve rejects).
I don't understand why MPOL_LOCAL is described as a pseudo-policy:
it's a reasonable policy which suffers from a confusing implementation
in terms of MPOL_PREFERRED with MPOL_F_LOCAL. I believe this would be
much more robust if MPOL_LOCAL were recognized in switch statements
throughout, MPOL_F_LOCAL deleted, and MPOL_PREFERRED use the (possibly
empty) nodes mask like everyone else, instead of its preferred_node
variant (I presume an optimization from the days before MPOL_LOCAL).
But that would take me too long to get right and fully tested.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: stable@vger.kernel.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-01-02 02:01:33 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-30 18:51:10 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: don't handle MPOL_LOCAL like a fake MPOL_PREFERRED policy
MPOL_LOCAL policy has been setup as a real policy, but it is still handled
like a faked POL_PREFERRED policy with one internal MPOL_F_LOCAL flag bit
set, and there are many places having to judge the real 'prefer' or the
'local' policy, which are quite confusing.
In current code, there are 4 cases that MPOL_LOCAL are used:
1. user specifies 'local' policy
2. user specifies 'prefer' policy, but with empty nodemask
3. system 'default' policy is used
4. 'prefer' policy + valid 'preferred' node with MPOL_F_STATIC_NODES
flag set, and when it is 'rebind' to a nodemask which doesn't contains
the 'preferred' node, it will perform as 'local' policy
So make 'local' a real policy instead of a fake 'prefer' one, and kill
MPOL_F_LOCAL bit, which can greatly reduce the confusion for code reading.
For case 4, the logic of mpol_rebind_preferred() is confusing, as Michal
Hocko pointed out:
: I do believe that rebinding preferred policy is just bogus and it should
: be dropped altogether on the ground that a preference is a mere hint from
: userspace where to start the allocation. Unless I am missing something
: cpusets will be always authoritative for the final placement. The
: preferred node just acts as a starting point and it should be really
: preserved when cpusets changes. Otherwise we have a very subtle behavior
: corner cases.
So dump all the tricky transformation between 'prefer' and 'local', and
just record the new nodemask of rebinding.
[feng.tang@intel.com: fix a problem in mpol_set_nodemask(), per Michal Hocko]
Link: https://lkml.kernel.org/r/1622560492-1294-3-git-send-email-feng.tang@intel.com
[feng.tang@intel.com: refine code and comments of mpol_set_nodemask(), per Michal]
Link: https://lkml.kernel.org/r/20210603081807.GE56979@shbuild999.sh.intel.com
Link: https://lkml.kernel.org/r/1622469956-82897-3-git-send-email-feng.tang@intel.com
Signed-off-by: Feng Tang <feng.tang@intel.com>
Suggested-by: Michal Hocko <mhocko@suse.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ben Widawsky <ben.widawsky@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-30 18:51:00 -07:00
|
|
|
|
2021-06-30 18:51:10 -07:00
|
|
|
|
tmpfs mempolicy: fix /proc/mounts corrupting memory
Recently I suggested using "mount -o remount,mpol=local /tmp" in NUMA
mempolicy testing. Very nasty. Reading /proc/mounts, /proc/pid/mounts
or /proc/pid/mountinfo may then corrupt one bit of kernel memory, often
in a page table (causing "Bad swap" or "Bad page map" warning or "Bad
pagetable" oops), sometimes in a vm_area_struct or rbnode or somewhere
worse. "mpol=prefer" and "mpol=prefer:Node" are equally toxic.
Recent NUMA enhancements are not to blame: this dates back to 2.6.35,
when commit e17f74af351c "mempolicy: don't call mpol_set_nodemask() when
no_context" skipped mpol_parse_str()'s call to mpol_set_nodemask(),
which used to initialize v.preferred_node, or set MPOL_F_LOCAL in flags.
With slab poisoning, you can then rely on mpol_to_str() to set the bit
for node 0x6b6b, probably in the next page above the caller's stack.
mpol_parse_str() is only called from shmem_parse_options(): no_context
is always true, so call it unused for now, and remove !no_context code.
Set v.nodes or v.preferred_node or MPOL_F_LOCAL as mpol_to_str() might
expect. Then mpol_to_str() can ignore its no_context argument also,
the mpol being appropriately initialized whether contextualized or not.
Rename its no_context unused too, and let subsequent patch remove them
(that's not needed for stable backporting, which would involve rejects).
I don't understand why MPOL_LOCAL is described as a pseudo-policy:
it's a reasonable policy which suffers from a confusing implementation
in terms of MPOL_PREFERRED with MPOL_F_LOCAL. I believe this would be
much more robust if MPOL_LOCAL were recognized in switch statements
throughout, MPOL_F_LOCAL deleted, and MPOL_PREFERRED use the (possibly
empty) nodes mask like everyone else, instead of its preferred_node
variant (I presume an optimization from the days before MPOL_LOCAL).
But that would take me too long to get right and fully tested.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: stable@vger.kernel.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-01-02 02:01:33 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-03-23 13:35:32 -07:00
|
|
|
|
mempolicy: use struct mempolicy pointer in shmem_sb_info
This patch replaces the mempolicy mode, mode_flags, and nodemask in the
shmem_sb_info struct with a struct mempolicy pointer, initialized to NULL.
This removes dependency on the details of mempolicy from shmem.c and hugetlbfs
inode.c and simplifies the interfaces.
mpol_parse_str() in mempolicy.c is changed to return, via a pointer to a
pointer arg, a struct mempolicy pointer on success. For MPOL_DEFAULT, the
returned pointer is NULL. Further, mpol_parse_str() now takes a 'no_context'
argument that causes the input nodemask to be stored in the w.user_nodemask of
the created mempolicy for use when the mempolicy is installed in a tmpfs inode
shared policy tree. At that time, any cpuset contextualization is applied to
the original input nodemask. This preserves the previous behavior where the
input nodemask was stored in the superblock. We can think of the returned
mempolicy as "context free".
Because mpol_parse_str() is now calling mpol_new(), we can remove from
mpol_to_str() the semantic checks that mpol_new() already performs.
Add 'no_context' parameter to mpol_to_str() to specify that it should format
the nodemask in w.user_nodemask for 'bind' and 'interleave' policies.
Change mpol_shared_policy_init() to take a pointer to a "context free" struct
mempolicy and to create a new, "contextualized" mempolicy using the mode,
mode_flags and user_nodemask from the input mempolicy.
Note: we know that the mempolicy passed to mpol_to_str() or
mpol_shared_policy_init() from a tmpfs superblock is "context free". This
is currently the only instance thereof. However, if we found more uses for
this concept, and introduced any ambiguity as to whether a mempolicy was
context free or not, we could add another internal mode flag to identify
context free mempolicies. Then, we could remove the 'no_context' argument
from mpol_to_str().
Added shmem_get_sbmpol() to return a reference counted superblock mempolicy,
if one exists, to pass to mpol_shared_policy_init(). We must add the
reference under the sb stat_lock to prevent races with replacement of the mpol
by remount. This reference is removed in mpol_shared_policy_init().
[akpm@linux-foundation.org: build fix]
[akpm@linux-foundation.org: another build fix]
[akpm@linux-foundation.org: yet another build fix]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:26 -07:00
|
|
|
|
mempolicy: rework shmem mpol parsing and display
mm/shmem.c currently contains functions to parse and display memory policy
strings for the tmpfs 'mpol' mount option. Move this to mm/mempolicy.c with
the rest of the mempolicy support. With subsequent patches, we'll be able to
remove knowledge of the details [mode, flags, policy, ...] completely from
shmem.c
1) replace shmem_parse_mpol() in mm/shmem.c with mpol_parse_str() in
mm/mempolicy.c. Rework to use the policy_types[] array [used by
mpol_to_str()] to look up mode by name.
2) use mpol_to_str() to format policy for shmem_show_mpol(). mpol_to_str()
expects a pointer to a struct mempolicy, so temporarily construct one.
This will be replaced with a reference to a struct mempolicy in the tmpfs
superblock in a subsequent patch.
NOTE 1: I changed mpol_to_str() to use a colon ':' rather than an equal
sign '=' as the nodemask delimiter to match mpol_parse_str() and the
tmpfs/shmem mpol mount option formatting that now uses mpol_to_str(). This
is a user visible change to numa_maps, but then the addition of the mode
flags already changed the display. It makes sense to me to have the mounts
and numa_maps display the policy in the same format. However, if anyone
objects strongly, I can pass the desired nodemask delimeter as an arg to
mpol_to_str().
Note 2: Like show_numa_map(), I don't check the return code from
mpol_to_str(). I do use a longer buffer than the one provided by
show_numa_map(), which seems to have sufficed so far.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: use struct mempolicy pointer in shmem_sb_info
This patch replaces the mempolicy mode, mode_flags, and nodemask in the
shmem_sb_info struct with a struct mempolicy pointer, initialized to NULL.
This removes dependency on the details of mempolicy from shmem.c and hugetlbfs
inode.c and simplifies the interfaces.
mpol_parse_str() in mempolicy.c is changed to return, via a pointer to a
pointer arg, a struct mempolicy pointer on success. For MPOL_DEFAULT, the
returned pointer is NULL. Further, mpol_parse_str() now takes a 'no_context'
argument that causes the input nodemask to be stored in the w.user_nodemask of
the created mempolicy for use when the mempolicy is installed in a tmpfs inode
shared policy tree. At that time, any cpuset contextualization is applied to
the original input nodemask. This preserves the previous behavior where the
input nodemask was stored in the superblock. We can think of the returned
mempolicy as "context free".
Because mpol_parse_str() is now calling mpol_new(), we can remove from
mpol_to_str() the semantic checks that mpol_new() already performs.
Add 'no_context' parameter to mpol_to_str() to specify that it should format
the nodemask in w.user_nodemask for 'bind' and 'interleave' policies.
Change mpol_shared_policy_init() to take a pointer to a "context free" struct
mempolicy and to create a new, "contextualized" mempolicy using the mode,
mode_flags and user_nodemask from the input mempolicy.
Note: we know that the mempolicy passed to mpol_to_str() or
mpol_shared_policy_init() from a tmpfs superblock is "context free". This
is currently the only instance thereof. However, if we found more uses for
this concept, and introduced any ambiguity as to whether a mempolicy was
context free or not, we could add another internal mode flag to identify
context free mempolicies. Then, we could remove the 'no_context' argument
from mpol_to_str().
Added shmem_get_sbmpol() to return a reference counted superblock mempolicy,
if one exists, to pass to mpol_shared_policy_init(). We must add the
reference under the sb stat_lock to prevent races with replacement of the mpol
by remount. This reference is removed in mpol_shared_policy_init().
[akpm@linux-foundation.org: build fix]
[akpm@linux-foundation.org: another build fix]
[akpm@linux-foundation.org: yet another build fix]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:26 -07:00
|
|
|
|
|
|
|
|
|
mempolicy: rework shmem mpol parsing and display
mm/shmem.c currently contains functions to parse and display memory policy
strings for the tmpfs 'mpol' mount option. Move this to mm/mempolicy.c with
the rest of the mempolicy support. With subsequent patches, we'll be able to
remove knowledge of the details [mode, flags, policy, ...] completely from
shmem.c
1) replace shmem_parse_mpol() in mm/shmem.c with mpol_parse_str() in
mm/mempolicy.c. Rework to use the policy_types[] array [used by
mpol_to_str()] to look up mode by name.
2) use mpol_to_str() to format policy for shmem_show_mpol(). mpol_to_str()
expects a pointer to a struct mempolicy, so temporarily construct one.
This will be replaced with a reference to a struct mempolicy in the tmpfs
superblock in a subsequent patch.
NOTE 1: I changed mpol_to_str() to use a colon ':' rather than an equal
sign '=' as the nodemask delimiter to match mpol_parse_str() and the
tmpfs/shmem mpol mount option formatting that now uses mpol_to_str(). This
is a user visible change to numa_maps, but then the addition of the mode
flags already changed the display. It makes sense to me to have the mounts
and numa_maps display the policy in the same format. However, if anyone
objects strongly, I can pass the desired nodemask delimeter as an arg to
mpol_to_str().
Note 2: Like show_numa_map(), I don't check the return code from
mpol_to_str(). I do use a longer buffer than the one provided by
show_numa_map(), which seems to have sufficed so far.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: use struct mempolicy pointer in shmem_sb_info
This patch replaces the mempolicy mode, mode_flags, and nodemask in the
shmem_sb_info struct with a struct mempolicy pointer, initialized to NULL.
This removes dependency on the details of mempolicy from shmem.c and hugetlbfs
inode.c and simplifies the interfaces.
mpol_parse_str() in mempolicy.c is changed to return, via a pointer to a
pointer arg, a struct mempolicy pointer on success. For MPOL_DEFAULT, the
returned pointer is NULL. Further, mpol_parse_str() now takes a 'no_context'
argument that causes the input nodemask to be stored in the w.user_nodemask of
the created mempolicy for use when the mempolicy is installed in a tmpfs inode
shared policy tree. At that time, any cpuset contextualization is applied to
the original input nodemask. This preserves the previous behavior where the
input nodemask was stored in the superblock. We can think of the returned
mempolicy as "context free".
Because mpol_parse_str() is now calling mpol_new(), we can remove from
mpol_to_str() the semantic checks that mpol_new() already performs.
Add 'no_context' parameter to mpol_to_str() to specify that it should format
the nodemask in w.user_nodemask for 'bind' and 'interleave' policies.
Change mpol_shared_policy_init() to take a pointer to a "context free" struct
mempolicy and to create a new, "contextualized" mempolicy using the mode,
mode_flags and user_nodemask from the input mempolicy.
Note: we know that the mempolicy passed to mpol_to_str() or
mpol_shared_policy_init() from a tmpfs superblock is "context free". This
is currently the only instance thereof. However, if we found more uses for
this concept, and introduced any ambiguity as to whether a mempolicy was
context free or not, we could add another internal mode flag to identify
context free mempolicies. Then, we could remove the 'no_context' argument
from mpol_to_str().
Added shmem_get_sbmpol() to return a reference counted superblock mempolicy,
if one exists, to pass to mpol_shared_policy_init(). We must add the
reference under the sb stat_lock to prevent races with replacement of the mpol
by remount. This reference is removed in mpol_shared_policy_init().
[akpm@linux-foundation.org: build fix]
[akpm@linux-foundation.org: another build fix]
[akpm@linux-foundation.org: yet another build fix]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:26 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-11-12 15:07:28 -08:00
|
|
|
|
2024-07-08 08:56:32 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-01-08 01:01:02 -08:00
|
|
|
|
2013-11-12 15:07:28 -08:00
|
|
|
|
2006-01-08 01:01:02 -08:00
|
|
|
|
|
|
|
|
|
2013-11-12 15:07:28 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-01-08 01:01:02 -08:00
|
|
|
|
2024-07-08 08:56:32 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: use MPOL_PREFERRED for system-wide default policy
Currently, when one specifies MPOL_DEFAULT via a NUMA memory policy API
[set_mempolicy(), mbind() and internal versions], the kernel simply installs a
NULL struct mempolicy pointer in the appropriate context: task policy, vma
policy, or shared policy. This causes any use of that policy to "fall back"
to the next most specific policy scope.
The only use of MPOL_DEFAULT to mean "local allocation" is in the system
default policy. This requires extra checks/cases for MPOL_DEFAULT in many
mempolicy.c functions.
There is another, "preferred" way to specify local allocation via the APIs.
That is using the MPOL_PREFERRED policy mode with an empty nodemask.
Internally, the empty nodemask gets converted to a preferred_node id of '-1'.
All internal usage of MPOL_PREFERRED will convert the '-1' to the id of the
node local to the cpu where the allocation occurs.
System default policy, except during boot, is hard-coded to "local
allocation". By using the MPOL_PREFERRED mode with a negative value of
preferred node for system default policy, MPOL_DEFAULT will never occur in the
'policy' member of a struct mempolicy. Thus, we can remove all checks for
MPOL_DEFAULT when converting policy to a node id/zonelist in the allocation
paths.
In slab_node() return local node id when policy pointer is NULL. No need to
set a pol value to take the switch default. Replace switch default with
BUG()--i.e., shouldn't happen.
With this patch MPOL_DEFAULT is only used in the APIs, including internal
calls to do_set_mempolicy() and in the display of policy in
/proc/<pid>/numa_maps. It always means "fall back" to the the next most
specific policy scope. This simplifies the description of memory policies
quite a bit, with no visible change in behavior.
get_mempolicy() continues to return MPOL_DEFAULT and an empty nodemask when
the requested policy [task or vma/shared] is NULL. These are the values one
would supply via set_mempolicy() or mbind() to achieve that condition--default
behavior.
This patch updates Documentation to reflect this change.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:18 -07:00
|
|
|
|
2013-11-12 15:07:28 -08:00
|
|
|
|
|
|
|
|
|
mempolicy: use MPOL_PREFERRED for system-wide default policy
Currently, when one specifies MPOL_DEFAULT via a NUMA memory policy API
[set_mempolicy(), mbind() and internal versions], the kernel simply installs a
NULL struct mempolicy pointer in the appropriate context: task policy, vma
policy, or shared policy. This causes any use of that policy to "fall back"
to the next most specific policy scope.
The only use of MPOL_DEFAULT to mean "local allocation" is in the system
default policy. This requires extra checks/cases for MPOL_DEFAULT in many
mempolicy.c functions.
There is another, "preferred" way to specify local allocation via the APIs.
That is using the MPOL_PREFERRED policy mode with an empty nodemask.
Internally, the empty nodemask gets converted to a preferred_node id of '-1'.
All internal usage of MPOL_PREFERRED will convert the '-1' to the id of the
node local to the cpu where the allocation occurs.
System default policy, except during boot, is hard-coded to "local
allocation". By using the MPOL_PREFERRED mode with a negative value of
preferred node for system default policy, MPOL_DEFAULT will never occur in the
'policy' member of a struct mempolicy. Thus, we can remove all checks for
MPOL_DEFAULT when converting policy to a node id/zonelist in the allocation
paths.
In slab_node() return local node id when policy pointer is NULL. No need to
set a pol value to take the switch default. Replace switch default with
BUG()--i.e., shouldn't happen.
With this patch MPOL_DEFAULT is only used in the APIs, including internal
calls to do_set_mempolicy() and in the display of policy in
/proc/<pid>/numa_maps. It always means "fall back" to the the next most
specific policy scope. This simplifies the description of memory policies
quite a bit, with no visible change in behavior.
get_mempolicy() continues to return MPOL_DEFAULT and an empty nodemask when
the requested policy [task or vma/shared] is NULL. These are the values one
would supply via set_mempolicy() or mbind() to achieve that condition--default
behavior.
This patch updates Documentation to reflect this change.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:18 -07:00
|
|
|
|
2006-01-08 01:01:02 -08:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: don't handle MPOL_LOCAL like a fake MPOL_PREFERRED policy
MPOL_LOCAL policy has been setup as a real policy, but it is still handled
like a faked POL_PREFERRED policy with one internal MPOL_F_LOCAL flag bit
set, and there are many places having to judge the real 'prefer' or the
'local' policy, which are quite confusing.
In current code, there are 4 cases that MPOL_LOCAL are used:
1. user specifies 'local' policy
2. user specifies 'prefer' policy, but with empty nodemask
3. system 'default' policy is used
4. 'prefer' policy + valid 'preferred' node with MPOL_F_STATIC_NODES
flag set, and when it is 'rebind' to a nodemask which doesn't contains
the 'preferred' node, it will perform as 'local' policy
So make 'local' a real policy instead of a fake 'prefer' one, and kill
MPOL_F_LOCAL bit, which can greatly reduce the confusion for code reading.
For case 4, the logic of mpol_rebind_preferred() is confusing, as Michal
Hocko pointed out:
: I do believe that rebinding preferred policy is just bogus and it should
: be dropped altogether on the ground that a preference is a mere hint from
: userspace where to start the allocation. Unless I am missing something
: cpusets will be always authoritative for the final placement. The
: preferred node just acts as a starting point and it should be really
: preserved when cpusets changes. Otherwise we have a very subtle behavior
: corner cases.
So dump all the tricky transformation between 'prefer' and 'local', and
just record the new nodemask of rebinding.
[feng.tang@intel.com: fix a problem in mpol_set_nodemask(), per Michal Hocko]
Link: https://lkml.kernel.org/r/1622560492-1294-3-git-send-email-feng.tang@intel.com
[feng.tang@intel.com: refine code and comments of mpol_set_nodemask(), per Michal]
Link: https://lkml.kernel.org/r/20210603081807.GE56979@shbuild999.sh.intel.com
Link: https://lkml.kernel.org/r/1622469956-82897-3-git-send-email-feng.tang@intel.com
Signed-off-by: Feng Tang <feng.tang@intel.com>
Suggested-by: Michal Hocko <mhocko@suse.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ben Widawsky <ben.widawsky@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-30 18:51:00 -07:00
|
|
|
|
2006-01-08 01:01:02 -08:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: add MPOL_PREFERRED_MANY for multiple preferred nodes
Patch series "Introduce multi-preference mempolicy", v7.
This patch series introduces the concept of the MPOL_PREFERRED_MANY
mempolicy. This mempolicy mode can be used with either the
set_mempolicy(2) or mbind(2) interfaces. Like the MPOL_PREFERRED
interface, it allows an application to set a preference for nodes which
will fulfil memory allocation requests. Unlike the MPOL_PREFERRED mode,
it takes a set of nodes. Like the MPOL_BIND interface, it works over a
set of nodes. Unlike MPOL_BIND, it will not cause a SIGSEGV or invoke the
OOM killer if those preferred nodes are not available.
Along with these patches are patches for libnuma, numactl, numademo, and
memhog. They still need some polish, but can be found here:
https://gitlab.com/bwidawsk/numactl/-/tree/prefer-many It allows new
usage: `numactl -P 0,3,4`
The goal of the new mode is to enable some use-cases when using tiered memory
usage models which I've lovingly named.
1a. The Hare - The interconnect is fast enough to meet bandwidth and
latency requirements allowing preference to be given to all nodes with
"fast" memory.
1b. The Indiscriminate Hare - An application knows it wants fast
memory (or perhaps slow memory), but doesn't care which node it runs
on. The application can prefer a set of nodes and then xpu bind to
the local node (cpu, accelerator, etc). This reverses the nodes are
chosen today where the kernel attempts to use local memory to the CPU
whenever possible. This will attempt to use the local accelerator to
the memory.
2. The Tortoise - The administrator (or the application itself) is
aware it only needs slow memory, and so can prefer that.
Much of this is almost achievable with the bind interface, but the bind
interface suffers from an inability to fallback to another set of nodes if
binding fails to all nodes in the nodemask.
Like MPOL_BIND a nodemask is given. Inherently this removes ordering from the
preference.
> /* Set first two nodes as preferred in an 8 node system. */
> const unsigned long nodes = 0x3
> set_mempolicy(MPOL_PREFER_MANY, &nodes, 8);
> /* Mimic interleave policy, but have fallback *.
> const unsigned long nodes = 0xaa
> set_mempolicy(MPOL_PREFER_MANY, &nodes, 8);
Some internal discussion took place around the interface. There are two
alternatives which we have discussed, plus one I stuck in:
1. Ordered list of nodes. Currently it's believed that the added
complexity is nod needed for expected usecases.
2. A flag for bind to allow falling back to other nodes. This
confuses the notion of binding and is less flexible than the current
solution.
3. Create flags or new modes that helps with some ordering. This
offers both a friendlier API as well as a solution for more customized
usage. It's unknown if it's worth the complexity to support this.
Here is sample code for how this might work:
> // Prefer specific nodes for some something wacky
> set_mempolicy(MPOL_PREFER_MANY, 0x17c, 1024);
>
> // Default
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_SOCKET, NULL, 0);
> // which is the same as
> set_mempolicy(MPOL_DEFAULT, NULL, 0);
>
> // The Hare
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE, NULL, 0);
>
> // The Tortoise
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE_REV, NULL, 0);
>
> // Prefer the fast memory of the first two sockets
> set_mempolicy(MPOL_PREFER_MANY | MPOL_F_PREFER_ORDER_TYPE, -1, 2);
>
This patch (of 5):
The NUMA APIs currently allow passing in a "preferred node" as a single
bit set in a nodemask. If more than one bit it set, bits after the first
are ignored.
This single node is generally OK for location-based NUMA where memory
being allocated will eventually be operated on by a single CPU. However,
in systems with multiple memory types, folks want to target a *type* of
memory instead of a location. For instance, someone might want some
high-bandwidth memory but do not care about the CPU next to which it is
allocated. Or, they want a cheap, high capacity allocation and want to
target all NUMA nodes which have persistent memory in volatile mode. In
both of these cases, the application wants to target a *set* of nodes, but
does not want strict MPOL_BIND behavior as that could lead to OOM killer
or SIGSEGV.
So add MPOL_PREFERRED_MANY policy to support the multiple preferred nodes
requirement. This is not a pie-in-the-sky dream for an API. This was a
response to a specific ask of more than one group at Intel. Specifically:
1. There are existing libraries that target memory types such as
https://github.com/memkind/memkind. These are known to suffer from
SIGSEGV's when memory is low on targeted memory "kinds" that span more
than one node. The MCDRAM on a Xeon Phi in "Cluster on Die" mode is an
example of this.
2. Volatile-use persistent memory users want to have a memory policy
which is targeted at either "cheap and slow" (PMEM) or "expensive and
fast" (DRAM). However, they do not want to experience allocation
failures when the targeted type is unavailable.
3. Allocate-then-run. Generally, we let the process scheduler decide
on which physical CPU to run a task. That location provides a default
allocation policy, and memory availability is not generally considered
when placing tasks. For situations where memory is valuable and
constrained, some users want to allocate memory first, *then* allocate
close compute resources to the allocation. This is the reverse of the
normal (CPU) model. Accelerators such as GPUs that operate on
core-mm-managed memory are interested in this model.
A check is added in sanitize_mpol_flags() to not permit 'prefer_many'
policy to be used for now, and will be removed in later patch after all
implementations for 'prefer_many' are ready, as suggested by Michal Hocko.
[mhocko@kernel.org: suggest to refine policy_node/policy_nodemask handling]
Link: https://lkml.kernel.org/r/1627970362-61305-1-git-send-email-feng.tang@intel.com
Link: https://lore.kernel.org/r/20200630212517.308045-4-ben.widawsky@intel.com
Link: https://lkml.kernel.org/r/1627970362-61305-2-git-send-email-feng.tang@intel.com
Co-developed-by: Ben Widawsky <ben.widawsky@intel.com>
Signed-off-by: Ben Widawsky <ben.widawsky@intel.com>
Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com>
Signed-off-by: Feng Tang <feng.tang@intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Huang Ying <ying.huang@intel.com>b
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-09-02 15:00:06 -07:00
|
|
|
|
2006-01-08 01:01:02 -08:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted interleaving
When a system has multiple NUMA nodes and it becomes bandwidth hungry,
using the current MPOL_INTERLEAVE could be an wise option.
However, if those NUMA nodes consist of different types of memory such as
socket-attached DRAM and CXL/PCIe attached DRAM, the round-robin based
interleave policy does not optimally distribute data to make use of their
different bandwidth characteristics.
Instead, interleave is more effective when the allocation policy follows
each NUMA nodes' bandwidth weight rather than a simple 1:1 distribution.
This patch introduces a new memory policy, MPOL_WEIGHTED_INTERLEAVE,
enabling weighted interleave between NUMA nodes. Weighted interleave
allows for proportional distribution of memory across multiple numa nodes,
preferably apportioned to match the bandwidth of each node.
For example, if a system has 1 CPU node (0), and 2 memory nodes (0,1),
with bandwidth of (100GB/s, 50GB/s) respectively, the appropriate weight
distribution is (2:1).
Weights for each node can be assigned via the new sysfs extension:
/sys/kernel/mm/mempolicy/weighted_interleave/
For now, the default value of all nodes will be `1`, which matches the
behavior of standard 1:1 round-robin interleave. An extension will be
added in the future to allow default values to be registered at kernel and
device bringup time.
The policy allocates a number of pages equal to the set weights. For
example, if the weights are (2,1), then 2 pages will be allocated on node0
for every 1 page allocated on node1.
The new flag MPOL_WEIGHTED_INTERLEAVE can be used in set_mempolicy(2)
and mbind(2).
Some high level notes about the pieces of weighted interleave:
current->il_prev:
Tracks the node previously allocated from.
current->il_weight:
The active weight of the current node (current->il_prev)
When this reaches 0, current->il_prev is set to the next node
and current->il_weight is set to the next weight.
weighted_interleave_nodes:
Counts the number of allocations as they occur, and applies the
weight for the current node. When the weight reaches 0, switch
to the next node. Operates only on task->mempolicy.
weighted_interleave_nid:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the node based on the given index.
Operates on VMA policies.
bulk_array_weighted_interleave:
Gets the total weight of the nodemask as well as each individual
node weight, then calculates the number of "interleave rounds" as
well as any delta ("partial round"). Calculates the number of
pages for each node and allocates them.
If a node was scheduled for interleave via interleave_nodes, the
current weight will be allocated first.
Operates only on the task->mempolicy.
One piece of complexity is the interaction between a recent refactor which
split the logic to acquire the "ilx" (interleave index) of an allocation
and the actually application of the interleave. If a call to
alloc_pages_mpol() were made with a weighted-interleave policy and ilx set
to NO_INTERLEAVE_INDEX, weighted_interleave_nodes() would operate on a VMA
policy - violating the description above.
An inspection of all callers of alloc_pages_mpol() shows that all external
callers set ilx to `0`, an index value, or will call get_vma_policy() to
acquire the ilx.
For example, mm/shmem.c may call into alloc_pages_mpol. The call stacks
all set (pgoff_t ilx) or end up in `get_vma_policy()`. This enforces the
`weighted_interleave_nodes()` and `weighted_interleave_nid()` policy
requirements (task/vma respectively).
Link: https://lkml.kernel.org/r/20240202170238.90004-4-gregory.price@memverge.com
Suggested-by: Hasan Al Maruf <Hasan.Maruf@amd.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Co-developed-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Co-developed-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Signed-off-by: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:37 -05:00
|
|
|
|
2021-06-30 18:51:10 -07:00
|
|
|
|
2006-01-08 01:01:02 -08:00
|
|
|
|
|
|
|
|
|
2013-11-12 15:07:28 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-01-08 01:01:02 -08:00
|
|
|
|
|
|
|
|
|
2013-11-21 14:32:06 -08:00
|
|
|
|
2006-01-08 01:01:02 -08:00
|
|
|
|
2008-04-28 02:13:21 -07:00
|
|
|
|
2013-11-12 15:07:28 -08:00
|
|
|
|
mempolicy: add MPOL_F_STATIC_NODES flag
Add an optional mempolicy mode flag, MPOL_F_STATIC_NODES, that suppresses the
node remap when the policy is rebound.
Adds another member to struct mempolicy, nodemask_t user_nodemask, as part of
a union with cpuset_mems_allowed:
struct mempolicy {
...
union {
nodemask_t cpuset_mems_allowed;
nodemask_t user_nodemask;
} w;
}
that stores the the nodemask that the user passed when he or she created the
mempolicy via set_mempolicy() or mbind(). When using MPOL_F_STATIC_NODES,
which is passed with any mempolicy mode, the user's passed nodemask
intersected with the VMA or task's allowed nodes is always used when
determining the preferred node, setting the MPOL_BIND zonelist, or creating
the interleave nodemask. This happens whenever the policy is rebound,
including when a task's cpuset assignment changes or the cpuset's mems are
changed.
This creates an interesting side-effect in that it allows the mempolicy
"intent" to lie dormant and uneffected until it has access to the node(s) that
it desires. For example, if you currently ask for an interleaved policy over
a set of nodes that you do not have access to, the mempolicy is not created
and the task continues to use the previous policy. With this change, however,
it is possible to create the same mempolicy; it is only effected when access
to nodes in the nodemask is acquired.
It is also possible to mount tmpfs with the static nodemask behavior when
specifying a node or nodemask. To do this, simply add "=static" immediately
following the mempolicy mode at mount time:
mount -o remount mpol=interleave=static:1-3
Also removes mpol_check_policy() and folds its logic into mpol_new() since it
is now obsoleted. The unused vma_mpol_equal() is also removed.
Cc: Paul Jackson <pj@sgi.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:12:27 -07:00
|
|
|
|
mempolicy: clean-up mpol-to-str() mempolicy formatting
mpol-to-str() formats memory policies into printable strings. Currently this
is only used to display "numa_maps". A subsequent patch will use
mpol_to_str() for formatting tmpfs [shmem] mpol mount options, allowing us to
remove essentially duplicate code in mm/shmem.c. This patch cleans up
mpol_to_str() generally and in preparation for that patch.
1) show_numa_maps() is not checking the return code from mpol_to_str().
There's not a lot we can do in this context if mpol_to_str() did return the
error [insufficient space in buffer]. Proposed "solution": just check,
under DEBUG_VM, that callers are providing sufficient buffer space for the
policy, flags, and a few nodes. This way, we'll get some display.
show_numa_maps() is providing a 50-byte buffer, so it won't trip this
check. 50-bytes should be sufficient unless one has a large number of
nodes in a very sparse nodemask.
2) The display of the new mode flags ["static" & "relative"] was set up to
display multiple flags, separated by a "bar" '|'. However, this support is
incomplete--e.g., need_bar was never incremented; and currently, these two
flags are mutually exclusive. So remove the "bar" support, for now, and
only display one flag.
3) Use snprint() to format flags, so as not to overflow the buffer. Not
that it's ever happed, AFAIK.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:22 -07:00
|
|
|
|
2024-07-08 08:56:32 +01:00
|
|
|
|
mempolicy: clean-up mpol-to-str() mempolicy formatting
mpol-to-str() formats memory policies into printable strings. Currently this
is only used to display "numa_maps". A subsequent patch will use
mpol_to_str() for formatting tmpfs [shmem] mpol mount options, allowing us to
remove essentially duplicate code in mm/shmem.c. This patch cleans up
mpol_to_str() generally and in preparation for that patch.
1) show_numa_maps() is not checking the return code from mpol_to_str().
There's not a lot we can do in this context if mpol_to_str() did return the
error [insufficient space in buffer]. Proposed "solution": just check,
under DEBUG_VM, that callers are providing sufficient buffer space for the
policy, flags, and a few nodes. This way, we'll get some display.
show_numa_maps() is providing a 50-byte buffer, so it won't trip this
check. 50-bytes should be sufficient unless one has a large number of
nodes in a very sparse nodemask.
2) The display of the new mode flags ["static" & "relative"] was set up to
display multiple flags, separated by a "bar" '|'. However, this support is
incomplete--e.g., need_bar was never incremented; and currently, these two
flags are mutually exclusive. So remove the "bar" support, for now, and
only display one flag.
3) Use snprint() to format flags, so as not to overflow the buffer. Not
that it's ever happed, AFAIK.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:22 -07:00
|
|
|
|
mempolicy: add MPOL_F_STATIC_NODES flag
Add an optional mempolicy mode flag, MPOL_F_STATIC_NODES, that suppresses the
node remap when the policy is rebound.
Adds another member to struct mempolicy, nodemask_t user_nodemask, as part of
a union with cpuset_mems_allowed:
struct mempolicy {
...
union {
nodemask_t cpuset_mems_allowed;
nodemask_t user_nodemask;
} w;
}
that stores the the nodemask that the user passed when he or she created the
mempolicy via set_mempolicy() or mbind(). When using MPOL_F_STATIC_NODES,
which is passed with any mempolicy mode, the user's passed nodemask
intersected with the VMA or task's allowed nodes is always used when
determining the preferred node, setting the MPOL_BIND zonelist, or creating
the interleave nodemask. This happens whenever the policy is rebound,
including when a task's cpuset assignment changes or the cpuset's mems are
changed.
This creates an interesting side-effect in that it allows the mempolicy
"intent" to lie dormant and uneffected until it has access to the node(s) that
it desires. For example, if you currently ask for an interleaved policy over
a set of nodes that you do not have access to, the mempolicy is not created
and the task continues to use the previous policy. With this change, however,
it is possible to create the same mempolicy; it is only effected when access
to nodes in the nodemask is acquired.
It is also possible to mount tmpfs with the static nodemask behavior when
specifying a node or nodemask. To do this, simply add "=static" immediately
following the mempolicy mode at mount time:
mount -o remount mpol=interleave=static:1-3
Also removes mpol_check_policy() and folds its logic into mpol_new() since it
is now obsoleted. The unused vma_mpol_equal() is also removed.
Cc: Paul Jackson <pj@sgi.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:12:27 -07:00
|
|
|
|
mempolicy: clean-up mpol-to-str() mempolicy formatting
mpol-to-str() formats memory policies into printable strings. Currently this
is only used to display "numa_maps". A subsequent patch will use
mpol_to_str() for formatting tmpfs [shmem] mpol mount options, allowing us to
remove essentially duplicate code in mm/shmem.c. This patch cleans up
mpol_to_str() generally and in preparation for that patch.
1) show_numa_maps() is not checking the return code from mpol_to_str().
There's not a lot we can do in this context if mpol_to_str() did return the
error [insufficient space in buffer]. Proposed "solution": just check,
under DEBUG_VM, that callers are providing sufficient buffer space for the
policy, flags, and a few nodes. This way, we'll get some display.
show_numa_maps() is providing a 50-byte buffer, so it won't trip this
check. 50-bytes should be sufficient unless one has a large number of
nodes in a very sparse nodemask.
2) The display of the new mode flags ["static" & "relative"] was set up to
display multiple flags, separated by a "bar" '|'. However, this support is
incomplete--e.g., need_bar was never incremented; and currently, these two
flags are mutually exclusive. So remove the "bar" support, for now, and
only display one flag.
3) Use snprint() to format flags, so as not to overflow the buffer. Not
that it's ever happed, AFAIK.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:13:22 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-07-08 08:56:32 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mempolicy: add MPOL_F_STATIC_NODES flag
Add an optional mempolicy mode flag, MPOL_F_STATIC_NODES, that suppresses the
node remap when the policy is rebound.
Adds another member to struct mempolicy, nodemask_t user_nodemask, as part of
a union with cpuset_mems_allowed:
struct mempolicy {
...
union {
nodemask_t cpuset_mems_allowed;
nodemask_t user_nodemask;
} w;
}
that stores the the nodemask that the user passed when he or she created the
mempolicy via set_mempolicy() or mbind(). When using MPOL_F_STATIC_NODES,
which is passed with any mempolicy mode, the user's passed nodemask
intersected with the VMA or task's allowed nodes is always used when
determining the preferred node, setting the MPOL_BIND zonelist, or creating
the interleave nodemask. This happens whenever the policy is rebound,
including when a task's cpuset assignment changes or the cpuset's mems are
changed.
This creates an interesting side-effect in that it allows the mempolicy
"intent" to lie dormant and uneffected until it has access to the node(s) that
it desires. For example, if you currently ask for an interleaved policy over
a set of nodes that you do not have access to, the mempolicy is not created
and the task continues to use the previous policy. With this change, however,
it is possible to create the same mempolicy; it is only effected when access
to nodes in the nodemask is acquired.
It is also possible to mount tmpfs with the static nodemask behavior when
specifying a node or nodemask. To do this, simply add "=static" immediately
following the mempolicy mode at mount time:
mount -o remount mpol=interleave=static:1-3
Also removes mpol_check_policy() and folds its logic into mpol_new() since it
is now obsoleted. The unused vma_mpol_equal() is also removed.
Cc: Paul Jackson <pj@sgi.com>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 02:12:27 -07:00
|
|
|
|
|
|
|
|
|
2015-02-13 14:38:02 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-01-08 01:01:02 -08:00
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-04-17 16:28:36 +09:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: support memory hotplug in weighted interleave
The weighted interleave policy distributes page allocations across
multiple NUMA nodes based on their performance weight, thereby improving
memory bandwidth utilization. The weight values for each node are
configured through sysfs.
Previously, sysfs entries for configuring weighted interleave were created
for all possible nodes (N_POSSIBLE) at initialization, including nodes
that might not have memory. However, not all nodes in N_POSSIBLE are
usable at runtime, as some may remain memoryless or offline. This led to
sysfs entries being created for unusable nodes, causing potential
misconfiguration issues.
To address this issue, this patch modifies the sysfs creation logic to:
1) Limit sysfs entries to nodes that are online and have memory, avoiding
the creation of sysfs entries for nodes that cannot be used.
2) Support memory hotplug by dynamically adding and removing sysfs entries
based on whether a node transitions into or out of the N_MEMORY state.
Additionally, the patch ensures that sysfs attributes are properly managed
when nodes go offline, preventing stale or redundant entries from
persisting in the system.
By making these changes, the weighted interleave policy now manages its
sysfs entries more efficiently, ensuring that only relevant nodes are
considered for interleaving, and dynamically adapting to memory hotplug
events.
[dan.carpenter@linaro.org: fix error code in sysfs_wi_node_add()]
Link: https://lkml.kernel.org/r/aBjL7Bwc0QBzgajK@stanley.mountain
Link: https://lkml.kernel.org/r/20250417072839.711-4-rakie.kim@sk.com
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Dan Carpenter <dan.carpenter@linaro.org>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:37 +09:00
|
|
|
|
2025-04-17 16:28:36 +09:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: Weighted Interleave Auto-tuning
On machines with multiple memory nodes, interleaving page allocations
across nodes allows for better utilization of each node's bandwidth.
Previous work by Gregory Price [1] introduced weighted interleave, which
allowed for pages to be allocated across nodes according to user-set
ratios.
Ideally, these weights should be proportional to their bandwidth, so that
under bandwidth pressure, each node uses its maximal efficient bandwidth
and prevents latency from increasing exponentially.
Previously, weighted interleave's default weights were just 1s -- which
would be equivalent to the (unweighted) interleave mempolicy, which goes
through the nodes in a round-robin fashion, ignoring bandwidth
information.
This patch has two main goals: First, it makes weighted interleave easier
to use for users who wish to relieve bandwidth pressure when using nodes
with varying bandwidth (CXL). By providing a set of "real" default
weights that just work out of the box, users who might not have the
capability (or wish to) perform experimentation to find the most optimal
weights for their system can still take advantage of bandwidth-informed
weighted interleave.
Second, it allows for weighted interleave to dynamically adjust to
hotplugged memory with new bandwidth information. Instead of manually
updating node weights every time new bandwidth information is reported or
taken off, weighted interleave adjusts and provides a new set of default
weights for weighted interleave to use when there is a change in bandwidth
information.
To meet these goals, this patch introduces an auto-configuration mode for
the interleave weights that provides a reasonable set of default weights,
calculated using bandwidth data reported by the system. In auto mode,
weights are dynamically adjusted based on whatever the current bandwidth
information reports (and responds to hotplug events).
This patch still supports users manually writing weights into the nodeN
sysfs interface by entering into manual mode. When a user enters manual
mode, the system stops dynamically updating any of the node weights, even
during hotplug events that shift the optimal weight distribution.
A new sysfs interface "auto" is introduced, which allows users to switch
between the auto (writing 1 or Y) and manual (writing 0 or N) modes. The
system also automatically enters manual mode when a nodeN interface is
manually written to.
There is one functional change that this patch makes to the existing
weighted_interleave ABI: previously, writing 0 directly to a nodeN
interface was said to reset the weight to the system default. Before this
patch, the default for all weights were 1, which meant that writing 0 and
1 were functionally equivalent. With this patch, writing 0 is invalid.
Link: https://lkml.kernel.org/r/20250520141236.2987309-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: wordsmithing changes, simplification, fixes]
Link: https://lkml.kernel.org/r/20250511025840.2410154-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: remove auto_kobj_attr field from struct sysfs_wi_group]
Link: https://lkml.kernel.org/r/20250512142511.3959833-1-joshua.hahnjy@gmail.com
https://lore.kernel.org/linux-mm/20240202170238.90004-1-gregory.price@memverge.com/ [1]
Link: https://lkml.kernel.org/r/20250505182328.4148265-1-joshua.hahnjy@gmail.com
Co-developed-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Yunjeong Mun <yunjeong.mun@sk.com>
Suggested-by: Oscar Salvador <osalvador@suse.de>
Suggested-by: Ying Huang <ying.huang@linux.alibaba.com>
Suggested-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Reviewed-by: Honggyu Kim <honggyu.kim@sk.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-05 11:23:28 -07:00
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: Weighted Interleave Auto-tuning
On machines with multiple memory nodes, interleaving page allocations
across nodes allows for better utilization of each node's bandwidth.
Previous work by Gregory Price [1] introduced weighted interleave, which
allowed for pages to be allocated across nodes according to user-set
ratios.
Ideally, these weights should be proportional to their bandwidth, so that
under bandwidth pressure, each node uses its maximal efficient bandwidth
and prevents latency from increasing exponentially.
Previously, weighted interleave's default weights were just 1s -- which
would be equivalent to the (unweighted) interleave mempolicy, which goes
through the nodes in a round-robin fashion, ignoring bandwidth
information.
This patch has two main goals: First, it makes weighted interleave easier
to use for users who wish to relieve bandwidth pressure when using nodes
with varying bandwidth (CXL). By providing a set of "real" default
weights that just work out of the box, users who might not have the
capability (or wish to) perform experimentation to find the most optimal
weights for their system can still take advantage of bandwidth-informed
weighted interleave.
Second, it allows for weighted interleave to dynamically adjust to
hotplugged memory with new bandwidth information. Instead of manually
updating node weights every time new bandwidth information is reported or
taken off, weighted interleave adjusts and provides a new set of default
weights for weighted interleave to use when there is a change in bandwidth
information.
To meet these goals, this patch introduces an auto-configuration mode for
the interleave weights that provides a reasonable set of default weights,
calculated using bandwidth data reported by the system. In auto mode,
weights are dynamically adjusted based on whatever the current bandwidth
information reports (and responds to hotplug events).
This patch still supports users manually writing weights into the nodeN
sysfs interface by entering into manual mode. When a user enters manual
mode, the system stops dynamically updating any of the node weights, even
during hotplug events that shift the optimal weight distribution.
A new sysfs interface "auto" is introduced, which allows users to switch
between the auto (writing 1 or Y) and manual (writing 0 or N) modes. The
system also automatically enters manual mode when a nodeN interface is
manually written to.
There is one functional change that this patch makes to the existing
weighted_interleave ABI: previously, writing 0 directly to a nodeN
interface was said to reset the weight to the system default. Before this
patch, the default for all weights were 1, which meant that writing 0 and
1 were functionally equivalent. With this patch, writing 0 is invalid.
Link: https://lkml.kernel.org/r/20250520141236.2987309-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: wordsmithing changes, simplification, fixes]
Link: https://lkml.kernel.org/r/20250511025840.2410154-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: remove auto_kobj_attr field from struct sysfs_wi_group]
Link: https://lkml.kernel.org/r/20250512142511.3959833-1-joshua.hahnjy@gmail.com
https://lore.kernel.org/linux-mm/20240202170238.90004-1-gregory.price@memverge.com/ [1]
Link: https://lkml.kernel.org/r/20250505182328.4148265-1-joshua.hahnjy@gmail.com
Co-developed-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Yunjeong Mun <yunjeong.mun@sk.com>
Suggested-by: Oscar Salvador <osalvador@suse.de>
Suggested-by: Ying Huang <ying.huang@linux.alibaba.com>
Suggested-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Reviewed-by: Honggyu Kim <honggyu.kim@sk.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-05 11:23:28 -07:00
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: Weighted Interleave Auto-tuning
On machines with multiple memory nodes, interleaving page allocations
across nodes allows for better utilization of each node's bandwidth.
Previous work by Gregory Price [1] introduced weighted interleave, which
allowed for pages to be allocated across nodes according to user-set
ratios.
Ideally, these weights should be proportional to their bandwidth, so that
under bandwidth pressure, each node uses its maximal efficient bandwidth
and prevents latency from increasing exponentially.
Previously, weighted interleave's default weights were just 1s -- which
would be equivalent to the (unweighted) interleave mempolicy, which goes
through the nodes in a round-robin fashion, ignoring bandwidth
information.
This patch has two main goals: First, it makes weighted interleave easier
to use for users who wish to relieve bandwidth pressure when using nodes
with varying bandwidth (CXL). By providing a set of "real" default
weights that just work out of the box, users who might not have the
capability (or wish to) perform experimentation to find the most optimal
weights for their system can still take advantage of bandwidth-informed
weighted interleave.
Second, it allows for weighted interleave to dynamically adjust to
hotplugged memory with new bandwidth information. Instead of manually
updating node weights every time new bandwidth information is reported or
taken off, weighted interleave adjusts and provides a new set of default
weights for weighted interleave to use when there is a change in bandwidth
information.
To meet these goals, this patch introduces an auto-configuration mode for
the interleave weights that provides a reasonable set of default weights,
calculated using bandwidth data reported by the system. In auto mode,
weights are dynamically adjusted based on whatever the current bandwidth
information reports (and responds to hotplug events).
This patch still supports users manually writing weights into the nodeN
sysfs interface by entering into manual mode. When a user enters manual
mode, the system stops dynamically updating any of the node weights, even
during hotplug events that shift the optimal weight distribution.
A new sysfs interface "auto" is introduced, which allows users to switch
between the auto (writing 1 or Y) and manual (writing 0 or N) modes. The
system also automatically enters manual mode when a nodeN interface is
manually written to.
There is one functional change that this patch makes to the existing
weighted_interleave ABI: previously, writing 0 directly to a nodeN
interface was said to reset the weight to the system default. Before this
patch, the default for all weights were 1, which meant that writing 0 and
1 were functionally equivalent. With this patch, writing 0 is invalid.
Link: https://lkml.kernel.org/r/20250520141236.2987309-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: wordsmithing changes, simplification, fixes]
Link: https://lkml.kernel.org/r/20250511025840.2410154-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: remove auto_kobj_attr field from struct sysfs_wi_group]
Link: https://lkml.kernel.org/r/20250512142511.3959833-1-joshua.hahnjy@gmail.com
https://lore.kernel.org/linux-mm/20240202170238.90004-1-gregory.price@memverge.com/ [1]
Link: https://lkml.kernel.org/r/20250505182328.4148265-1-joshua.hahnjy@gmail.com
Co-developed-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Yunjeong Mun <yunjeong.mun@sk.com>
Suggested-by: Oscar Salvador <osalvador@suse.de>
Suggested-by: Ying Huang <ying.huang@linux.alibaba.com>
Suggested-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Reviewed-by: Honggyu Kim <honggyu.kim@sk.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-05 11:23:28 -07:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
|
|
|
|
|
2026-02-21 20:03:00 -08:00
|
|
|
|
mm/mempolicy: Weighted Interleave Auto-tuning
On machines with multiple memory nodes, interleaving page allocations
across nodes allows for better utilization of each node's bandwidth.
Previous work by Gregory Price [1] introduced weighted interleave, which
allowed for pages to be allocated across nodes according to user-set
ratios.
Ideally, these weights should be proportional to their bandwidth, so that
under bandwidth pressure, each node uses its maximal efficient bandwidth
and prevents latency from increasing exponentially.
Previously, weighted interleave's default weights were just 1s -- which
would be equivalent to the (unweighted) interleave mempolicy, which goes
through the nodes in a round-robin fashion, ignoring bandwidth
information.
This patch has two main goals: First, it makes weighted interleave easier
to use for users who wish to relieve bandwidth pressure when using nodes
with varying bandwidth (CXL). By providing a set of "real" default
weights that just work out of the box, users who might not have the
capability (or wish to) perform experimentation to find the most optimal
weights for their system can still take advantage of bandwidth-informed
weighted interleave.
Second, it allows for weighted interleave to dynamically adjust to
hotplugged memory with new bandwidth information. Instead of manually
updating node weights every time new bandwidth information is reported or
taken off, weighted interleave adjusts and provides a new set of default
weights for weighted interleave to use when there is a change in bandwidth
information.
To meet these goals, this patch introduces an auto-configuration mode for
the interleave weights that provides a reasonable set of default weights,
calculated using bandwidth data reported by the system. In auto mode,
weights are dynamically adjusted based on whatever the current bandwidth
information reports (and responds to hotplug events).
This patch still supports users manually writing weights into the nodeN
sysfs interface by entering into manual mode. When a user enters manual
mode, the system stops dynamically updating any of the node weights, even
during hotplug events that shift the optimal weight distribution.
A new sysfs interface "auto" is introduced, which allows users to switch
between the auto (writing 1 or Y) and manual (writing 0 or N) modes. The
system also automatically enters manual mode when a nodeN interface is
manually written to.
There is one functional change that this patch makes to the existing
weighted_interleave ABI: previously, writing 0 directly to a nodeN
interface was said to reset the weight to the system default. Before this
patch, the default for all weights were 1, which meant that writing 0 and
1 were functionally equivalent. With this patch, writing 0 is invalid.
Link: https://lkml.kernel.org/r/20250520141236.2987309-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: wordsmithing changes, simplification, fixes]
Link: https://lkml.kernel.org/r/20250511025840.2410154-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: remove auto_kobj_attr field from struct sysfs_wi_group]
Link: https://lkml.kernel.org/r/20250512142511.3959833-1-joshua.hahnjy@gmail.com
https://lore.kernel.org/linux-mm/20240202170238.90004-1-gregory.price@memverge.com/ [1]
Link: https://lkml.kernel.org/r/20250505182328.4148265-1-joshua.hahnjy@gmail.com
Co-developed-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Yunjeong Mun <yunjeong.mun@sk.com>
Suggested-by: Oscar Salvador <osalvador@suse.de>
Suggested-by: Ying Huang <ying.huang@linux.alibaba.com>
Suggested-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Reviewed-by: Honggyu Kim <honggyu.kim@sk.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-05 11:23:28 -07:00
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: Weighted Interleave Auto-tuning
On machines with multiple memory nodes, interleaving page allocations
across nodes allows for better utilization of each node's bandwidth.
Previous work by Gregory Price [1] introduced weighted interleave, which
allowed for pages to be allocated across nodes according to user-set
ratios.
Ideally, these weights should be proportional to their bandwidth, so that
under bandwidth pressure, each node uses its maximal efficient bandwidth
and prevents latency from increasing exponentially.
Previously, weighted interleave's default weights were just 1s -- which
would be equivalent to the (unweighted) interleave mempolicy, which goes
through the nodes in a round-robin fashion, ignoring bandwidth
information.
This patch has two main goals: First, it makes weighted interleave easier
to use for users who wish to relieve bandwidth pressure when using nodes
with varying bandwidth (CXL). By providing a set of "real" default
weights that just work out of the box, users who might not have the
capability (or wish to) perform experimentation to find the most optimal
weights for their system can still take advantage of bandwidth-informed
weighted interleave.
Second, it allows for weighted interleave to dynamically adjust to
hotplugged memory with new bandwidth information. Instead of manually
updating node weights every time new bandwidth information is reported or
taken off, weighted interleave adjusts and provides a new set of default
weights for weighted interleave to use when there is a change in bandwidth
information.
To meet these goals, this patch introduces an auto-configuration mode for
the interleave weights that provides a reasonable set of default weights,
calculated using bandwidth data reported by the system. In auto mode,
weights are dynamically adjusted based on whatever the current bandwidth
information reports (and responds to hotplug events).
This patch still supports users manually writing weights into the nodeN
sysfs interface by entering into manual mode. When a user enters manual
mode, the system stops dynamically updating any of the node weights, even
during hotplug events that shift the optimal weight distribution.
A new sysfs interface "auto" is introduced, which allows users to switch
between the auto (writing 1 or Y) and manual (writing 0 or N) modes. The
system also automatically enters manual mode when a nodeN interface is
manually written to.
There is one functional change that this patch makes to the existing
weighted_interleave ABI: previously, writing 0 directly to a nodeN
interface was said to reset the weight to the system default. Before this
patch, the default for all weights were 1, which meant that writing 0 and
1 were functionally equivalent. With this patch, writing 0 is invalid.
Link: https://lkml.kernel.org/r/20250520141236.2987309-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: wordsmithing changes, simplification, fixes]
Link: https://lkml.kernel.org/r/20250511025840.2410154-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: remove auto_kobj_attr field from struct sysfs_wi_group]
Link: https://lkml.kernel.org/r/20250512142511.3959833-1-joshua.hahnjy@gmail.com
https://lore.kernel.org/linux-mm/20240202170238.90004-1-gregory.price@memverge.com/ [1]
Link: https://lkml.kernel.org/r/20250505182328.4148265-1-joshua.hahnjy@gmail.com
Co-developed-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Yunjeong Mun <yunjeong.mun@sk.com>
Suggested-by: Oscar Salvador <osalvador@suse.de>
Suggested-by: Ying Huang <ying.huang@linux.alibaba.com>
Suggested-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Reviewed-by: Honggyu Kim <honggyu.kim@sk.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-05 11:23:28 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-02-21 20:03:00 -08:00
|
|
|
|
mm/mempolicy: Weighted Interleave Auto-tuning
On machines with multiple memory nodes, interleaving page allocations
across nodes allows for better utilization of each node's bandwidth.
Previous work by Gregory Price [1] introduced weighted interleave, which
allowed for pages to be allocated across nodes according to user-set
ratios.
Ideally, these weights should be proportional to their bandwidth, so that
under bandwidth pressure, each node uses its maximal efficient bandwidth
and prevents latency from increasing exponentially.
Previously, weighted interleave's default weights were just 1s -- which
would be equivalent to the (unweighted) interleave mempolicy, which goes
through the nodes in a round-robin fashion, ignoring bandwidth
information.
This patch has two main goals: First, it makes weighted interleave easier
to use for users who wish to relieve bandwidth pressure when using nodes
with varying bandwidth (CXL). By providing a set of "real" default
weights that just work out of the box, users who might not have the
capability (or wish to) perform experimentation to find the most optimal
weights for their system can still take advantage of bandwidth-informed
weighted interleave.
Second, it allows for weighted interleave to dynamically adjust to
hotplugged memory with new bandwidth information. Instead of manually
updating node weights every time new bandwidth information is reported or
taken off, weighted interleave adjusts and provides a new set of default
weights for weighted interleave to use when there is a change in bandwidth
information.
To meet these goals, this patch introduces an auto-configuration mode for
the interleave weights that provides a reasonable set of default weights,
calculated using bandwidth data reported by the system. In auto mode,
weights are dynamically adjusted based on whatever the current bandwidth
information reports (and responds to hotplug events).
This patch still supports users manually writing weights into the nodeN
sysfs interface by entering into manual mode. When a user enters manual
mode, the system stops dynamically updating any of the node weights, even
during hotplug events that shift the optimal weight distribution.
A new sysfs interface "auto" is introduced, which allows users to switch
between the auto (writing 1 or Y) and manual (writing 0 or N) modes. The
system also automatically enters manual mode when a nodeN interface is
manually written to.
There is one functional change that this patch makes to the existing
weighted_interleave ABI: previously, writing 0 directly to a nodeN
interface was said to reset the weight to the system default. Before this
patch, the default for all weights were 1, which meant that writing 0 and
1 were functionally equivalent. With this patch, writing 0 is invalid.
Link: https://lkml.kernel.org/r/20250520141236.2987309-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: wordsmithing changes, simplification, fixes]
Link: https://lkml.kernel.org/r/20250511025840.2410154-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: remove auto_kobj_attr field from struct sysfs_wi_group]
Link: https://lkml.kernel.org/r/20250512142511.3959833-1-joshua.hahnjy@gmail.com
https://lore.kernel.org/linux-mm/20240202170238.90004-1-gregory.price@memverge.com/ [1]
Link: https://lkml.kernel.org/r/20250505182328.4148265-1-joshua.hahnjy@gmail.com
Co-developed-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Yunjeong Mun <yunjeong.mun@sk.com>
Suggested-by: Oscar Salvador <osalvador@suse.de>
Suggested-by: Ying Huang <ying.huang@linux.alibaba.com>
Suggested-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Reviewed-by: Honggyu Kim <honggyu.kim@sk.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-05 11:23:28 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-01 08:57:02 +08:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: Weighted Interleave Auto-tuning
On machines with multiple memory nodes, interleaving page allocations
across nodes allows for better utilization of each node's bandwidth.
Previous work by Gregory Price [1] introduced weighted interleave, which
allowed for pages to be allocated across nodes according to user-set
ratios.
Ideally, these weights should be proportional to their bandwidth, so that
under bandwidth pressure, each node uses its maximal efficient bandwidth
and prevents latency from increasing exponentially.
Previously, weighted interleave's default weights were just 1s -- which
would be equivalent to the (unweighted) interleave mempolicy, which goes
through the nodes in a round-robin fashion, ignoring bandwidth
information.
This patch has two main goals: First, it makes weighted interleave easier
to use for users who wish to relieve bandwidth pressure when using nodes
with varying bandwidth (CXL). By providing a set of "real" default
weights that just work out of the box, users who might not have the
capability (or wish to) perform experimentation to find the most optimal
weights for their system can still take advantage of bandwidth-informed
weighted interleave.
Second, it allows for weighted interleave to dynamically adjust to
hotplugged memory with new bandwidth information. Instead of manually
updating node weights every time new bandwidth information is reported or
taken off, weighted interleave adjusts and provides a new set of default
weights for weighted interleave to use when there is a change in bandwidth
information.
To meet these goals, this patch introduces an auto-configuration mode for
the interleave weights that provides a reasonable set of default weights,
calculated using bandwidth data reported by the system. In auto mode,
weights are dynamically adjusted based on whatever the current bandwidth
information reports (and responds to hotplug events).
This patch still supports users manually writing weights into the nodeN
sysfs interface by entering into manual mode. When a user enters manual
mode, the system stops dynamically updating any of the node weights, even
during hotplug events that shift the optimal weight distribution.
A new sysfs interface "auto" is introduced, which allows users to switch
between the auto (writing 1 or Y) and manual (writing 0 or N) modes. The
system also automatically enters manual mode when a nodeN interface is
manually written to.
There is one functional change that this patch makes to the existing
weighted_interleave ABI: previously, writing 0 directly to a nodeN
interface was said to reset the weight to the system default. Before this
patch, the default for all weights were 1, which meant that writing 0 and
1 were functionally equivalent. With this patch, writing 0 is invalid.
Link: https://lkml.kernel.org/r/20250520141236.2987309-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: wordsmithing changes, simplification, fixes]
Link: https://lkml.kernel.org/r/20250511025840.2410154-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: remove auto_kobj_attr field from struct sysfs_wi_group]
Link: https://lkml.kernel.org/r/20250512142511.3959833-1-joshua.hahnjy@gmail.com
https://lore.kernel.org/linux-mm/20240202170238.90004-1-gregory.price@memverge.com/ [1]
Link: https://lkml.kernel.org/r/20250505182328.4148265-1-joshua.hahnjy@gmail.com
Co-developed-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Yunjeong Mun <yunjeong.mun@sk.com>
Suggested-by: Oscar Salvador <osalvador@suse.de>
Suggested-by: Ying Huang <ying.huang@linux.alibaba.com>
Suggested-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Reviewed-by: Honggyu Kim <honggyu.kim@sk.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-05 11:23:28 -07:00
|
|
|
|
2026-04-01 08:57:02 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: Weighted Interleave Auto-tuning
On machines with multiple memory nodes, interleaving page allocations
across nodes allows for better utilization of each node's bandwidth.
Previous work by Gregory Price [1] introduced weighted interleave, which
allowed for pages to be allocated across nodes according to user-set
ratios.
Ideally, these weights should be proportional to their bandwidth, so that
under bandwidth pressure, each node uses its maximal efficient bandwidth
and prevents latency from increasing exponentially.
Previously, weighted interleave's default weights were just 1s -- which
would be equivalent to the (unweighted) interleave mempolicy, which goes
through the nodes in a round-robin fashion, ignoring bandwidth
information.
This patch has two main goals: First, it makes weighted interleave easier
to use for users who wish to relieve bandwidth pressure when using nodes
with varying bandwidth (CXL). By providing a set of "real" default
weights that just work out of the box, users who might not have the
capability (or wish to) perform experimentation to find the most optimal
weights for their system can still take advantage of bandwidth-informed
weighted interleave.
Second, it allows for weighted interleave to dynamically adjust to
hotplugged memory with new bandwidth information. Instead of manually
updating node weights every time new bandwidth information is reported or
taken off, weighted interleave adjusts and provides a new set of default
weights for weighted interleave to use when there is a change in bandwidth
information.
To meet these goals, this patch introduces an auto-configuration mode for
the interleave weights that provides a reasonable set of default weights,
calculated using bandwidth data reported by the system. In auto mode,
weights are dynamically adjusted based on whatever the current bandwidth
information reports (and responds to hotplug events).
This patch still supports users manually writing weights into the nodeN
sysfs interface by entering into manual mode. When a user enters manual
mode, the system stops dynamically updating any of the node weights, even
during hotplug events that shift the optimal weight distribution.
A new sysfs interface "auto" is introduced, which allows users to switch
between the auto (writing 1 or Y) and manual (writing 0 or N) modes. The
system also automatically enters manual mode when a nodeN interface is
manually written to.
There is one functional change that this patch makes to the existing
weighted_interleave ABI: previously, writing 0 directly to a nodeN
interface was said to reset the weight to the system default. Before this
patch, the default for all weights were 1, which meant that writing 0 and
1 were functionally equivalent. With this patch, writing 0 is invalid.
Link: https://lkml.kernel.org/r/20250520141236.2987309-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: wordsmithing changes, simplification, fixes]
Link: https://lkml.kernel.org/r/20250511025840.2410154-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: remove auto_kobj_attr field from struct sysfs_wi_group]
Link: https://lkml.kernel.org/r/20250512142511.3959833-1-joshua.hahnjy@gmail.com
https://lore.kernel.org/linux-mm/20240202170238.90004-1-gregory.price@memverge.com/ [1]
Link: https://lkml.kernel.org/r/20250505182328.4148265-1-joshua.hahnjy@gmail.com
Co-developed-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Yunjeong Mun <yunjeong.mun@sk.com>
Suggested-by: Oscar Salvador <osalvador@suse.de>
Suggested-by: Ying Huang <ying.huang@linux.alibaba.com>
Suggested-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Reviewed-by: Honggyu Kim <honggyu.kim@sk.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-05 11:23:28 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-04-17 16:28:36 +09:00
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
mm/mempolicy: support memory hotplug in weighted interleave
The weighted interleave policy distributes page allocations across
multiple NUMA nodes based on their performance weight, thereby improving
memory bandwidth utilization. The weight values for each node are
configured through sysfs.
Previously, sysfs entries for configuring weighted interleave were created
for all possible nodes (N_POSSIBLE) at initialization, including nodes
that might not have memory. However, not all nodes in N_POSSIBLE are
usable at runtime, as some may remain memoryless or offline. This led to
sysfs entries being created for unusable nodes, causing potential
misconfiguration issues.
To address this issue, this patch modifies the sysfs creation logic to:
1) Limit sysfs entries to nodes that are online and have memory, avoiding
the creation of sysfs entries for nodes that cannot be used.
2) Support memory hotplug by dynamically adding and removing sysfs entries
based on whether a node transitions into or out of the N_MEMORY state.
Additionally, the patch ensures that sysfs attributes are properly managed
when nodes go offline, preventing stale or redundant entries from
persisting in the system.
By making these changes, the weighted interleave policy now manages its
sysfs entries more efficiently, ensuring that only relevant nodes are
considered for interleaving, and dynamically adapting to memory hotplug
events.
[dan.carpenter@linaro.org: fix error code in sysfs_wi_node_add()]
Link: https://lkml.kernel.org/r/aBjL7Bwc0QBzgajK@stanley.mountain
Link: https://lkml.kernel.org/r/20250417072839.711-4-rakie.kim@sk.com
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Dan Carpenter <dan.carpenter@linaro.org>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:37 +09:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
2025-04-17 16:28:36 +09:00
|
|
|
|
mm/mempolicy: support memory hotplug in weighted interleave
The weighted interleave policy distributes page allocations across
multiple NUMA nodes based on their performance weight, thereby improving
memory bandwidth utilization. The weight values for each node are
configured through sysfs.
Previously, sysfs entries for configuring weighted interleave were created
for all possible nodes (N_POSSIBLE) at initialization, including nodes
that might not have memory. However, not all nodes in N_POSSIBLE are
usable at runtime, as some may remain memoryless or offline. This led to
sysfs entries being created for unusable nodes, causing potential
misconfiguration issues.
To address this issue, this patch modifies the sysfs creation logic to:
1) Limit sysfs entries to nodes that are online and have memory, avoiding
the creation of sysfs entries for nodes that cannot be used.
2) Support memory hotplug by dynamically adding and removing sysfs entries
based on whether a node transitions into or out of the N_MEMORY state.
Additionally, the patch ensures that sysfs attributes are properly managed
when nodes go offline, preventing stale or redundant entries from
persisting in the system.
By making these changes, the weighted interleave policy now manages its
sysfs entries more efficiently, ensuring that only relevant nodes are
considered for interleaving, and dynamically adapting to memory hotplug
events.
[dan.carpenter@linaro.org: fix error code in sysfs_wi_node_add()]
Link: https://lkml.kernel.org/r/aBjL7Bwc0QBzgajK@stanley.mountain
Link: https://lkml.kernel.org/r/20250417072839.711-4-rakie.kim@sk.com
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Dan Carpenter <dan.carpenter@linaro.org>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:37 +09:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
|
|
|
|
|
2025-04-17 16:28:36 +09:00
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
mm/mempolicy: fix memory leaks in weighted interleave sysfs
Patch series "Enhance sysfs handling for memory hotplug in weighted
interleave", v9.
The following patch series enhances the weighted interleave policy in the
memory management subsystem by improving sysfs handling, fixing memory
leaks, and introducing dynamic sysfs updates for memory hotplug support.
This patch (of 3):
Memory leaks occurred when removing sysfs attributes for weighted
interleave. Improper kobject deallocation led to unreleased memory when
initialization failed or when nodes were removed.
The risk of leak is low because it only appears to trigger if setup
fails. Setup only fails due to -ENOMEM which is unlikely to happen
from a late_initcall() when memory pressure is low.
This patch resolves the issue by replacing unnecessary `kfree()` calls
with proper `kobject_del()` and `kobject_put()` sequences, ensuring
correct teardown and preventing memory leaks.
By explicitly calling `kobject_del()` before `kobject_put()`, the release
function is now invoked safely, and internal sysfs state is correctly
cleaned up. This guarantees that the memory associated with the kobject
is fully released and avoids resource leaks, thereby improving system
stability.
Additionally, sysfs_remove_file() is no longer called from the release
function to avoid accessing invalid sysfs state after kobject_del(). All
attribute removals are now done before kobject_del(), preventing WARN_ON()
in kernfs and ensuring safe and consistent cleanup of sysfs entries.
Link: https://lkml.kernel.org/r/20250417072839.711-1-rakie.kim@sk.com
Link: https://lkml.kernel.org/r/20250417072839.711-2-rakie.kim@sk.com
Fixes: dce41f5ae253 ("mm/mempolicy: implement the sysfs-based weighted_interleave interface")
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Honggyu Kim <honggyu.kim@sk.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Yunjeong Mun <yunjeong.mun@sk.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:35 +09:00
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
mm/mempolicy: fix memory leaks in weighted interleave sysfs
Patch series "Enhance sysfs handling for memory hotplug in weighted
interleave", v9.
The following patch series enhances the weighted interleave policy in the
memory management subsystem by improving sysfs handling, fixing memory
leaks, and introducing dynamic sysfs updates for memory hotplug support.
This patch (of 3):
Memory leaks occurred when removing sysfs attributes for weighted
interleave. Improper kobject deallocation led to unreleased memory when
initialization failed or when nodes were removed.
The risk of leak is low because it only appears to trigger if setup
fails. Setup only fails due to -ENOMEM which is unlikely to happen
from a late_initcall() when memory pressure is low.
This patch resolves the issue by replacing unnecessary `kfree()` calls
with proper `kobject_del()` and `kobject_put()` sequences, ensuring
correct teardown and preventing memory leaks.
By explicitly calling `kobject_del()` before `kobject_put()`, the release
function is now invoked safely, and internal sysfs state is correctly
cleaned up. This guarantees that the memory associated with the kobject
is fully released and avoids resource leaks, thereby improving system
stability.
Additionally, sysfs_remove_file() is no longer called from the release
function to avoid accessing invalid sysfs state after kobject_del(). All
attribute removals are now done before kobject_del(), preventing WARN_ON()
in kernfs and ensuring safe and consistent cleanup of sysfs entries.
Link: https://lkml.kernel.org/r/20250417072839.711-1-rakie.kim@sk.com
Link: https://lkml.kernel.org/r/20250417072839.711-2-rakie.kim@sk.com
Fixes: dce41f5ae253 ("mm/mempolicy: implement the sysfs-based weighted_interleave interface")
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Honggyu Kim <honggyu.kim@sk.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Yunjeong Mun <yunjeong.mun@sk.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:35 +09:00
|
|
|
|
2025-04-17 16:28:36 +09:00
|
|
|
|
mm/mempolicy: fix memory leaks in weighted interleave sysfs
Patch series "Enhance sysfs handling for memory hotplug in weighted
interleave", v9.
The following patch series enhances the weighted interleave policy in the
memory management subsystem by improving sysfs handling, fixing memory
leaks, and introducing dynamic sysfs updates for memory hotplug support.
This patch (of 3):
Memory leaks occurred when removing sysfs attributes for weighted
interleave. Improper kobject deallocation led to unreleased memory when
initialization failed or when nodes were removed.
The risk of leak is low because it only appears to trigger if setup
fails. Setup only fails due to -ENOMEM which is unlikely to happen
from a late_initcall() when memory pressure is low.
This patch resolves the issue by replacing unnecessary `kfree()` calls
with proper `kobject_del()` and `kobject_put()` sequences, ensuring
correct teardown and preventing memory leaks.
By explicitly calling `kobject_del()` before `kobject_put()`, the release
function is now invoked safely, and internal sysfs state is correctly
cleaned up. This guarantees that the memory associated with the kobject
is fully released and avoids resource leaks, thereby improving system
stability.
Additionally, sysfs_remove_file() is no longer called from the release
function to avoid accessing invalid sysfs state after kobject_del(). All
attribute removals are now done before kobject_del(), preventing WARN_ON()
in kernfs and ensuring safe and consistent cleanup of sysfs entries.
Link: https://lkml.kernel.org/r/20250417072839.711-1-rakie.kim@sk.com
Link: https://lkml.kernel.org/r/20250417072839.711-2-rakie.kim@sk.com
Fixes: dce41f5ae253 ("mm/mempolicy: implement the sysfs-based weighted_interleave interface")
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Honggyu Kim <honggyu.kim@sk.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Yunjeong Mun <yunjeong.mun@sk.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:35 +09:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: Weighted Interleave Auto-tuning
On machines with multiple memory nodes, interleaving page allocations
across nodes allows for better utilization of each node's bandwidth.
Previous work by Gregory Price [1] introduced weighted interleave, which
allowed for pages to be allocated across nodes according to user-set
ratios.
Ideally, these weights should be proportional to their bandwidth, so that
under bandwidth pressure, each node uses its maximal efficient bandwidth
and prevents latency from increasing exponentially.
Previously, weighted interleave's default weights were just 1s -- which
would be equivalent to the (unweighted) interleave mempolicy, which goes
through the nodes in a round-robin fashion, ignoring bandwidth
information.
This patch has two main goals: First, it makes weighted interleave easier
to use for users who wish to relieve bandwidth pressure when using nodes
with varying bandwidth (CXL). By providing a set of "real" default
weights that just work out of the box, users who might not have the
capability (or wish to) perform experimentation to find the most optimal
weights for their system can still take advantage of bandwidth-informed
weighted interleave.
Second, it allows for weighted interleave to dynamically adjust to
hotplugged memory with new bandwidth information. Instead of manually
updating node weights every time new bandwidth information is reported or
taken off, weighted interleave adjusts and provides a new set of default
weights for weighted interleave to use when there is a change in bandwidth
information.
To meet these goals, this patch introduces an auto-configuration mode for
the interleave weights that provides a reasonable set of default weights,
calculated using bandwidth data reported by the system. In auto mode,
weights are dynamically adjusted based on whatever the current bandwidth
information reports (and responds to hotplug events).
This patch still supports users manually writing weights into the nodeN
sysfs interface by entering into manual mode. When a user enters manual
mode, the system stops dynamically updating any of the node weights, even
during hotplug events that shift the optimal weight distribution.
A new sysfs interface "auto" is introduced, which allows users to switch
between the auto (writing 1 or Y) and manual (writing 0 or N) modes. The
system also automatically enters manual mode when a nodeN interface is
manually written to.
There is one functional change that this patch makes to the existing
weighted_interleave ABI: previously, writing 0 directly to a nodeN
interface was said to reset the weight to the system default. Before this
patch, the default for all weights were 1, which meant that writing 0 and
1 were functionally equivalent. With this patch, writing 0 is invalid.
Link: https://lkml.kernel.org/r/20250520141236.2987309-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: wordsmithing changes, simplification, fixes]
Link: https://lkml.kernel.org/r/20250511025840.2410154-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: remove auto_kobj_attr field from struct sysfs_wi_group]
Link: https://lkml.kernel.org/r/20250512142511.3959833-1-joshua.hahnjy@gmail.com
https://lore.kernel.org/linux-mm/20240202170238.90004-1-gregory.price@memverge.com/ [1]
Link: https://lkml.kernel.org/r/20250505182328.4148265-1-joshua.hahnjy@gmail.com
Co-developed-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Yunjeong Mun <yunjeong.mun@sk.com>
Suggested-by: Oscar Salvador <osalvador@suse.de>
Suggested-by: Ying Huang <ying.huang@linux.alibaba.com>
Suggested-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Reviewed-by: Honggyu Kim <honggyu.kim@sk.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-05 11:23:28 -07:00
|
|
|
|
mm/mempolicy: fix memory leaks in weighted interleave sysfs
Patch series "Enhance sysfs handling for memory hotplug in weighted
interleave", v9.
The following patch series enhances the weighted interleave policy in the
memory management subsystem by improving sysfs handling, fixing memory
leaks, and introducing dynamic sysfs updates for memory hotplug support.
This patch (of 3):
Memory leaks occurred when removing sysfs attributes for weighted
interleave. Improper kobject deallocation led to unreleased memory when
initialization failed or when nodes were removed.
The risk of leak is low because it only appears to trigger if setup
fails. Setup only fails due to -ENOMEM which is unlikely to happen
from a late_initcall() when memory pressure is low.
This patch resolves the issue by replacing unnecessary `kfree()` calls
with proper `kobject_del()` and `kobject_put()` sequences, ensuring
correct teardown and preventing memory leaks.
By explicitly calling `kobject_del()` before `kobject_put()`, the release
function is now invoked safely, and internal sysfs state is correctly
cleaned up. This guarantees that the memory associated with the kobject
is fully released and avoids resource leaks, thereby improving system
stability.
Additionally, sysfs_remove_file() is no longer called from the release
function to avoid accessing invalid sysfs state after kobject_del(). All
attribute removals are now done before kobject_del(), preventing WARN_ON()
in kernfs and ensuring safe and consistent cleanup of sysfs entries.
Link: https://lkml.kernel.org/r/20250417072839.711-1-rakie.kim@sk.com
Link: https://lkml.kernel.org/r/20250417072839.711-2-rakie.kim@sk.com
Fixes: dce41f5ae253 ("mm/mempolicy: implement the sysfs-based weighted_interleave interface")
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Honggyu Kim <honggyu.kim@sk.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Yunjeong Mun <yunjeong.mun@sk.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:35 +09:00
|
|
|
|
mm/mempolicy: Weighted Interleave Auto-tuning
On machines with multiple memory nodes, interleaving page allocations
across nodes allows for better utilization of each node's bandwidth.
Previous work by Gregory Price [1] introduced weighted interleave, which
allowed for pages to be allocated across nodes according to user-set
ratios.
Ideally, these weights should be proportional to their bandwidth, so that
under bandwidth pressure, each node uses its maximal efficient bandwidth
and prevents latency from increasing exponentially.
Previously, weighted interleave's default weights were just 1s -- which
would be equivalent to the (unweighted) interleave mempolicy, which goes
through the nodes in a round-robin fashion, ignoring bandwidth
information.
This patch has two main goals: First, it makes weighted interleave easier
to use for users who wish to relieve bandwidth pressure when using nodes
with varying bandwidth (CXL). By providing a set of "real" default
weights that just work out of the box, users who might not have the
capability (or wish to) perform experimentation to find the most optimal
weights for their system can still take advantage of bandwidth-informed
weighted interleave.
Second, it allows for weighted interleave to dynamically adjust to
hotplugged memory with new bandwidth information. Instead of manually
updating node weights every time new bandwidth information is reported or
taken off, weighted interleave adjusts and provides a new set of default
weights for weighted interleave to use when there is a change in bandwidth
information.
To meet these goals, this patch introduces an auto-configuration mode for
the interleave weights that provides a reasonable set of default weights,
calculated using bandwidth data reported by the system. In auto mode,
weights are dynamically adjusted based on whatever the current bandwidth
information reports (and responds to hotplug events).
This patch still supports users manually writing weights into the nodeN
sysfs interface by entering into manual mode. When a user enters manual
mode, the system stops dynamically updating any of the node weights, even
during hotplug events that shift the optimal weight distribution.
A new sysfs interface "auto" is introduced, which allows users to switch
between the auto (writing 1 or Y) and manual (writing 0 or N) modes. The
system also automatically enters manual mode when a nodeN interface is
manually written to.
There is one functional change that this patch makes to the existing
weighted_interleave ABI: previously, writing 0 directly to a nodeN
interface was said to reset the weight to the system default. Before this
patch, the default for all weights were 1, which meant that writing 0 and
1 were functionally equivalent. With this patch, writing 0 is invalid.
Link: https://lkml.kernel.org/r/20250520141236.2987309-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: wordsmithing changes, simplification, fixes]
Link: https://lkml.kernel.org/r/20250511025840.2410154-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: remove auto_kobj_attr field from struct sysfs_wi_group]
Link: https://lkml.kernel.org/r/20250512142511.3959833-1-joshua.hahnjy@gmail.com
https://lore.kernel.org/linux-mm/20240202170238.90004-1-gregory.price@memverge.com/ [1]
Link: https://lkml.kernel.org/r/20250505182328.4148265-1-joshua.hahnjy@gmail.com
Co-developed-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Yunjeong Mun <yunjeong.mun@sk.com>
Suggested-by: Oscar Salvador <osalvador@suse.de>
Suggested-by: Ying Huang <ying.huang@linux.alibaba.com>
Suggested-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Reviewed-by: Honggyu Kim <honggyu.kim@sk.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-05 11:23:28 -07:00
|
|
|
|
mm/mempolicy: fix memory leaks in weighted interleave sysfs
Patch series "Enhance sysfs handling for memory hotplug in weighted
interleave", v9.
The following patch series enhances the weighted interleave policy in the
memory management subsystem by improving sysfs handling, fixing memory
leaks, and introducing dynamic sysfs updates for memory hotplug support.
This patch (of 3):
Memory leaks occurred when removing sysfs attributes for weighted
interleave. Improper kobject deallocation led to unreleased memory when
initialization failed or when nodes were removed.
The risk of leak is low because it only appears to trigger if setup
fails. Setup only fails due to -ENOMEM which is unlikely to happen
from a late_initcall() when memory pressure is low.
This patch resolves the issue by replacing unnecessary `kfree()` calls
with proper `kobject_del()` and `kobject_put()` sequences, ensuring
correct teardown and preventing memory leaks.
By explicitly calling `kobject_del()` before `kobject_put()`, the release
function is now invoked safely, and internal sysfs state is correctly
cleaned up. This guarantees that the memory associated with the kobject
is fully released and avoids resource leaks, thereby improving system
stability.
Additionally, sysfs_remove_file() is no longer called from the release
function to avoid accessing invalid sysfs state after kobject_del(). All
attribute removals are now done before kobject_del(), preventing WARN_ON()
in kernfs and ensuring safe and consistent cleanup of sysfs entries.
Link: https://lkml.kernel.org/r/20250417072839.711-1-rakie.kim@sk.com
Link: https://lkml.kernel.org/r/20250417072839.711-2-rakie.kim@sk.com
Fixes: dce41f5ae253 ("mm/mempolicy: implement the sysfs-based weighted_interleave interface")
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Honggyu Kim <honggyu.kim@sk.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Yunjeong Mun <yunjeong.mun@sk.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:35 +09:00
|
|
|
|
mm/mempolicy: Weighted Interleave Auto-tuning
On machines with multiple memory nodes, interleaving page allocations
across nodes allows for better utilization of each node's bandwidth.
Previous work by Gregory Price [1] introduced weighted interleave, which
allowed for pages to be allocated across nodes according to user-set
ratios.
Ideally, these weights should be proportional to their bandwidth, so that
under bandwidth pressure, each node uses its maximal efficient bandwidth
and prevents latency from increasing exponentially.
Previously, weighted interleave's default weights were just 1s -- which
would be equivalent to the (unweighted) interleave mempolicy, which goes
through the nodes in a round-robin fashion, ignoring bandwidth
information.
This patch has two main goals: First, it makes weighted interleave easier
to use for users who wish to relieve bandwidth pressure when using nodes
with varying bandwidth (CXL). By providing a set of "real" default
weights that just work out of the box, users who might not have the
capability (or wish to) perform experimentation to find the most optimal
weights for their system can still take advantage of bandwidth-informed
weighted interleave.
Second, it allows for weighted interleave to dynamically adjust to
hotplugged memory with new bandwidth information. Instead of manually
updating node weights every time new bandwidth information is reported or
taken off, weighted interleave adjusts and provides a new set of default
weights for weighted interleave to use when there is a change in bandwidth
information.
To meet these goals, this patch introduces an auto-configuration mode for
the interleave weights that provides a reasonable set of default weights,
calculated using bandwidth data reported by the system. In auto mode,
weights are dynamically adjusted based on whatever the current bandwidth
information reports (and responds to hotplug events).
This patch still supports users manually writing weights into the nodeN
sysfs interface by entering into manual mode. When a user enters manual
mode, the system stops dynamically updating any of the node weights, even
during hotplug events that shift the optimal weight distribution.
A new sysfs interface "auto" is introduced, which allows users to switch
between the auto (writing 1 or Y) and manual (writing 0 or N) modes. The
system also automatically enters manual mode when a nodeN interface is
manually written to.
There is one functional change that this patch makes to the existing
weighted_interleave ABI: previously, writing 0 directly to a nodeN
interface was said to reset the weight to the system default. Before this
patch, the default for all weights were 1, which meant that writing 0 and
1 were functionally equivalent. With this patch, writing 0 is invalid.
Link: https://lkml.kernel.org/r/20250520141236.2987309-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: wordsmithing changes, simplification, fixes]
Link: https://lkml.kernel.org/r/20250511025840.2410154-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: remove auto_kobj_attr field from struct sysfs_wi_group]
Link: https://lkml.kernel.org/r/20250512142511.3959833-1-joshua.hahnjy@gmail.com
https://lore.kernel.org/linux-mm/20240202170238.90004-1-gregory.price@memverge.com/ [1]
Link: https://lkml.kernel.org/r/20250505182328.4148265-1-joshua.hahnjy@gmail.com
Co-developed-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Yunjeong Mun <yunjeong.mun@sk.com>
Suggested-by: Oscar Salvador <osalvador@suse.de>
Suggested-by: Ying Huang <ying.huang@linux.alibaba.com>
Suggested-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Reviewed-by: Honggyu Kim <honggyu.kim@sk.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-05 11:23:28 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-06-02 09:23:40 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: fix memory leaks in weighted interleave sysfs
Patch series "Enhance sysfs handling for memory hotplug in weighted
interleave", v9.
The following patch series enhances the weighted interleave policy in the
memory management subsystem by improving sysfs handling, fixing memory
leaks, and introducing dynamic sysfs updates for memory hotplug support.
This patch (of 3):
Memory leaks occurred when removing sysfs attributes for weighted
interleave. Improper kobject deallocation led to unreleased memory when
initialization failed or when nodes were removed.
The risk of leak is low because it only appears to trigger if setup
fails. Setup only fails due to -ENOMEM which is unlikely to happen
from a late_initcall() when memory pressure is low.
This patch resolves the issue by replacing unnecessary `kfree()` calls
with proper `kobject_del()` and `kobject_put()` sequences, ensuring
correct teardown and preventing memory leaks.
By explicitly calling `kobject_del()` before `kobject_put()`, the release
function is now invoked safely, and internal sysfs state is correctly
cleaned up. This guarantees that the memory associated with the kobject
is fully released and avoids resource leaks, thereby improving system
stability.
Additionally, sysfs_remove_file() is no longer called from the release
function to avoid accessing invalid sysfs state after kobject_del(). All
attribute removals are now done before kobject_del(), preventing WARN_ON()
in kernfs and ensuring safe and consistent cleanup of sysfs entries.
Link: https://lkml.kernel.org/r/20250417072839.711-1-rakie.kim@sk.com
Link: https://lkml.kernel.org/r/20250417072839.711-2-rakie.kim@sk.com
Fixes: dce41f5ae253 ("mm/mempolicy: implement the sysfs-based weighted_interleave interface")
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Honggyu Kim <honggyu.kim@sk.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Yunjeong Mun <yunjeong.mun@sk.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:35 +09:00
|
|
|
|
|
|
|
|
|
2026-04-07 07:14:14 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: Weighted Interleave Auto-tuning
On machines with multiple memory nodes, interleaving page allocations
across nodes allows for better utilization of each node's bandwidth.
Previous work by Gregory Price [1] introduced weighted interleave, which
allowed for pages to be allocated across nodes according to user-set
ratios.
Ideally, these weights should be proportional to their bandwidth, so that
under bandwidth pressure, each node uses its maximal efficient bandwidth
and prevents latency from increasing exponentially.
Previously, weighted interleave's default weights were just 1s -- which
would be equivalent to the (unweighted) interleave mempolicy, which goes
through the nodes in a round-robin fashion, ignoring bandwidth
information.
This patch has two main goals: First, it makes weighted interleave easier
to use for users who wish to relieve bandwidth pressure when using nodes
with varying bandwidth (CXL). By providing a set of "real" default
weights that just work out of the box, users who might not have the
capability (or wish to) perform experimentation to find the most optimal
weights for their system can still take advantage of bandwidth-informed
weighted interleave.
Second, it allows for weighted interleave to dynamically adjust to
hotplugged memory with new bandwidth information. Instead of manually
updating node weights every time new bandwidth information is reported or
taken off, weighted interleave adjusts and provides a new set of default
weights for weighted interleave to use when there is a change in bandwidth
information.
To meet these goals, this patch introduces an auto-configuration mode for
the interleave weights that provides a reasonable set of default weights,
calculated using bandwidth data reported by the system. In auto mode,
weights are dynamically adjusted based on whatever the current bandwidth
information reports (and responds to hotplug events).
This patch still supports users manually writing weights into the nodeN
sysfs interface by entering into manual mode. When a user enters manual
mode, the system stops dynamically updating any of the node weights, even
during hotplug events that shift the optimal weight distribution.
A new sysfs interface "auto" is introduced, which allows users to switch
between the auto (writing 1 or Y) and manual (writing 0 or N) modes. The
system also automatically enters manual mode when a nodeN interface is
manually written to.
There is one functional change that this patch makes to the existing
weighted_interleave ABI: previously, writing 0 directly to a nodeN
interface was said to reset the weight to the system default. Before this
patch, the default for all weights were 1, which meant that writing 0 and
1 were functionally equivalent. With this patch, writing 0 is invalid.
Link: https://lkml.kernel.org/r/20250520141236.2987309-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: wordsmithing changes, simplification, fixes]
Link: https://lkml.kernel.org/r/20250511025840.2410154-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: remove auto_kobj_attr field from struct sysfs_wi_group]
Link: https://lkml.kernel.org/r/20250512142511.3959833-1-joshua.hahnjy@gmail.com
https://lore.kernel.org/linux-mm/20240202170238.90004-1-gregory.price@memverge.com/ [1]
Link: https://lkml.kernel.org/r/20250505182328.4148265-1-joshua.hahnjy@gmail.com
Co-developed-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Yunjeong Mun <yunjeong.mun@sk.com>
Suggested-by: Oscar Salvador <osalvador@suse.de>
Suggested-by: Ying Huang <ying.huang@linux.alibaba.com>
Suggested-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Reviewed-by: Honggyu Kim <honggyu.kim@sk.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-05 11:23:28 -07:00
|
|
|
|
2025-04-17 16:28:36 +09:00
|
|
|
|
mm/mempolicy: Weighted Interleave Auto-tuning
On machines with multiple memory nodes, interleaving page allocations
across nodes allows for better utilization of each node's bandwidth.
Previous work by Gregory Price [1] introduced weighted interleave, which
allowed for pages to be allocated across nodes according to user-set
ratios.
Ideally, these weights should be proportional to their bandwidth, so that
under bandwidth pressure, each node uses its maximal efficient bandwidth
and prevents latency from increasing exponentially.
Previously, weighted interleave's default weights were just 1s -- which
would be equivalent to the (unweighted) interleave mempolicy, which goes
through the nodes in a round-robin fashion, ignoring bandwidth
information.
This patch has two main goals: First, it makes weighted interleave easier
to use for users who wish to relieve bandwidth pressure when using nodes
with varying bandwidth (CXL). By providing a set of "real" default
weights that just work out of the box, users who might not have the
capability (or wish to) perform experimentation to find the most optimal
weights for their system can still take advantage of bandwidth-informed
weighted interleave.
Second, it allows for weighted interleave to dynamically adjust to
hotplugged memory with new bandwidth information. Instead of manually
updating node weights every time new bandwidth information is reported or
taken off, weighted interleave adjusts and provides a new set of default
weights for weighted interleave to use when there is a change in bandwidth
information.
To meet these goals, this patch introduces an auto-configuration mode for
the interleave weights that provides a reasonable set of default weights,
calculated using bandwidth data reported by the system. In auto mode,
weights are dynamically adjusted based on whatever the current bandwidth
information reports (and responds to hotplug events).
This patch still supports users manually writing weights into the nodeN
sysfs interface by entering into manual mode. When a user enters manual
mode, the system stops dynamically updating any of the node weights, even
during hotplug events that shift the optimal weight distribution.
A new sysfs interface "auto" is introduced, which allows users to switch
between the auto (writing 1 or Y) and manual (writing 0 or N) modes. The
system also automatically enters manual mode when a nodeN interface is
manually written to.
There is one functional change that this patch makes to the existing
weighted_interleave ABI: previously, writing 0 directly to a nodeN
interface was said to reset the weight to the system default. Before this
patch, the default for all weights were 1, which meant that writing 0 and
1 were functionally equivalent. With this patch, writing 0 is invalid.
Link: https://lkml.kernel.org/r/20250520141236.2987309-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: wordsmithing changes, simplification, fixes]
Link: https://lkml.kernel.org/r/20250511025840.2410154-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: remove auto_kobj_attr field from struct sysfs_wi_group]
Link: https://lkml.kernel.org/r/20250512142511.3959833-1-joshua.hahnjy@gmail.com
https://lore.kernel.org/linux-mm/20240202170238.90004-1-gregory.price@memverge.com/ [1]
Link: https://lkml.kernel.org/r/20250505182328.4148265-1-joshua.hahnjy@gmail.com
Co-developed-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Yunjeong Mun <yunjeong.mun@sk.com>
Suggested-by: Oscar Salvador <osalvador@suse.de>
Suggested-by: Ying Huang <ying.huang@linux.alibaba.com>
Suggested-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Reviewed-by: Honggyu Kim <honggyu.kim@sk.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-05 11:23:28 -07:00
|
|
|
|
2025-04-17 16:28:36 +09:00
|
|
|
|
mm/mempolicy: Weighted Interleave Auto-tuning
On machines with multiple memory nodes, interleaving page allocations
across nodes allows for better utilization of each node's bandwidth.
Previous work by Gregory Price [1] introduced weighted interleave, which
allowed for pages to be allocated across nodes according to user-set
ratios.
Ideally, these weights should be proportional to their bandwidth, so that
under bandwidth pressure, each node uses its maximal efficient bandwidth
and prevents latency from increasing exponentially.
Previously, weighted interleave's default weights were just 1s -- which
would be equivalent to the (unweighted) interleave mempolicy, which goes
through the nodes in a round-robin fashion, ignoring bandwidth
information.
This patch has two main goals: First, it makes weighted interleave easier
to use for users who wish to relieve bandwidth pressure when using nodes
with varying bandwidth (CXL). By providing a set of "real" default
weights that just work out of the box, users who might not have the
capability (or wish to) perform experimentation to find the most optimal
weights for their system can still take advantage of bandwidth-informed
weighted interleave.
Second, it allows for weighted interleave to dynamically adjust to
hotplugged memory with new bandwidth information. Instead of manually
updating node weights every time new bandwidth information is reported or
taken off, weighted interleave adjusts and provides a new set of default
weights for weighted interleave to use when there is a change in bandwidth
information.
To meet these goals, this patch introduces an auto-configuration mode for
the interleave weights that provides a reasonable set of default weights,
calculated using bandwidth data reported by the system. In auto mode,
weights are dynamically adjusted based on whatever the current bandwidth
information reports (and responds to hotplug events).
This patch still supports users manually writing weights into the nodeN
sysfs interface by entering into manual mode. When a user enters manual
mode, the system stops dynamically updating any of the node weights, even
during hotplug events that shift the optimal weight distribution.
A new sysfs interface "auto" is introduced, which allows users to switch
between the auto (writing 1 or Y) and manual (writing 0 or N) modes. The
system also automatically enters manual mode when a nodeN interface is
manually written to.
There is one functional change that this patch makes to the existing
weighted_interleave ABI: previously, writing 0 directly to a nodeN
interface was said to reset the weight to the system default. Before this
patch, the default for all weights were 1, which meant that writing 0 and
1 were functionally equivalent. With this patch, writing 0 is invalid.
Link: https://lkml.kernel.org/r/20250520141236.2987309-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: wordsmithing changes, simplification, fixes]
Link: https://lkml.kernel.org/r/20250511025840.2410154-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: remove auto_kobj_attr field from struct sysfs_wi_group]
Link: https://lkml.kernel.org/r/20250512142511.3959833-1-joshua.hahnjy@gmail.com
https://lore.kernel.org/linux-mm/20240202170238.90004-1-gregory.price@memverge.com/ [1]
Link: https://lkml.kernel.org/r/20250505182328.4148265-1-joshua.hahnjy@gmail.com
Co-developed-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Yunjeong Mun <yunjeong.mun@sk.com>
Suggested-by: Oscar Salvador <osalvador@suse.de>
Suggested-by: Ying Huang <ying.huang@linux.alibaba.com>
Suggested-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Reviewed-by: Honggyu Kim <honggyu.kim@sk.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-05 11:23:28 -07:00
|
|
|
|
mm/mempolicy: fix memory leaks in weighted interleave sysfs
Patch series "Enhance sysfs handling for memory hotplug in weighted
interleave", v9.
The following patch series enhances the weighted interleave policy in the
memory management subsystem by improving sysfs handling, fixing memory
leaks, and introducing dynamic sysfs updates for memory hotplug support.
This patch (of 3):
Memory leaks occurred when removing sysfs attributes for weighted
interleave. Improper kobject deallocation led to unreleased memory when
initialization failed or when nodes were removed.
The risk of leak is low because it only appears to trigger if setup
fails. Setup only fails due to -ENOMEM which is unlikely to happen
from a late_initcall() when memory pressure is low.
This patch resolves the issue by replacing unnecessary `kfree()` calls
with proper `kobject_del()` and `kobject_put()` sequences, ensuring
correct teardown and preventing memory leaks.
By explicitly calling `kobject_del()` before `kobject_put()`, the release
function is now invoked safely, and internal sysfs state is correctly
cleaned up. This guarantees that the memory associated with the kobject
is fully released and avoids resource leaks, thereby improving system
stability.
Additionally, sysfs_remove_file() is no longer called from the release
function to avoid accessing invalid sysfs state after kobject_del(). All
attribute removals are now done before kobject_del(), preventing WARN_ON()
in kernfs and ensuring safe and consistent cleanup of sysfs entries.
Link: https://lkml.kernel.org/r/20250417072839.711-1-rakie.kim@sk.com
Link: https://lkml.kernel.org/r/20250417072839.711-2-rakie.kim@sk.com
Fixes: dce41f5ae253 ("mm/mempolicy: implement the sysfs-based weighted_interleave interface")
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Honggyu Kim <honggyu.kim@sk.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Yunjeong Mun <yunjeong.mun@sk.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:35 +09:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-04-17 16:28:36 +09:00
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: fix memory leaks in weighted interleave sysfs
Patch series "Enhance sysfs handling for memory hotplug in weighted
interleave", v9.
The following patch series enhances the weighted interleave policy in the
memory management subsystem by improving sysfs handling, fixing memory
leaks, and introducing dynamic sysfs updates for memory hotplug support.
This patch (of 3):
Memory leaks occurred when removing sysfs attributes for weighted
interleave. Improper kobject deallocation led to unreleased memory when
initialization failed or when nodes were removed.
The risk of leak is low because it only appears to trigger if setup
fails. Setup only fails due to -ENOMEM which is unlikely to happen
from a late_initcall() when memory pressure is low.
This patch resolves the issue by replacing unnecessary `kfree()` calls
with proper `kobject_del()` and `kobject_put()` sequences, ensuring
correct teardown and preventing memory leaks.
By explicitly calling `kobject_del()` before `kobject_put()`, the release
function is now invoked safely, and internal sysfs state is correctly
cleaned up. This guarantees that the memory associated with the kobject
is fully released and avoids resource leaks, thereby improving system
stability.
Additionally, sysfs_remove_file() is no longer called from the release
function to avoid accessing invalid sysfs state after kobject_del(). All
attribute removals are now done before kobject_del(), preventing WARN_ON()
in kernfs and ensuring safe and consistent cleanup of sysfs entries.
Link: https://lkml.kernel.org/r/20250417072839.711-1-rakie.kim@sk.com
Link: https://lkml.kernel.org/r/20250417072839.711-2-rakie.kim@sk.com
Fixes: dce41f5ae253 ("mm/mempolicy: implement the sysfs-based weighted_interleave interface")
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Honggyu Kim <honggyu.kim@sk.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Yunjeong Mun <yunjeong.mun@sk.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:35 +09:00
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
|
|
|
|
|
2025-04-17 16:28:36 +09:00
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
mm/mempolicy: support memory hotplug in weighted interleave
The weighted interleave policy distributes page allocations across
multiple NUMA nodes based on their performance weight, thereby improving
memory bandwidth utilization. The weight values for each node are
configured through sysfs.
Previously, sysfs entries for configuring weighted interleave were created
for all possible nodes (N_POSSIBLE) at initialization, including nodes
that might not have memory. However, not all nodes in N_POSSIBLE are
usable at runtime, as some may remain memoryless or offline. This led to
sysfs entries being created for unusable nodes, causing potential
misconfiguration issues.
To address this issue, this patch modifies the sysfs creation logic to:
1) Limit sysfs entries to nodes that are online and have memory, avoiding
the creation of sysfs entries for nodes that cannot be used.
2) Support memory hotplug by dynamically adding and removing sysfs entries
based on whether a node transitions into or out of the N_MEMORY state.
Additionally, the patch ensures that sysfs attributes are properly managed
when nodes go offline, preventing stale or redundant entries from
persisting in the system.
By making these changes, the weighted interleave policy now manages its
sysfs entries more efficiently, ensuring that only relevant nodes are
considered for interleaving, and dynamically adapting to memory hotplug
events.
[dan.carpenter@linaro.org: fix error code in sysfs_wi_node_add()]
Link: https://lkml.kernel.org/r/aBjL7Bwc0QBzgajK@stanley.mountain
Link: https://lkml.kernel.org/r/20250417072839.711-4-rakie.kim@sk.com
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Dan Carpenter <dan.carpenter@linaro.org>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:37 +09:00
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
mm/mempolicy: support memory hotplug in weighted interleave
The weighted interleave policy distributes page allocations across
multiple NUMA nodes based on their performance weight, thereby improving
memory bandwidth utilization. The weight values for each node are
configured through sysfs.
Previously, sysfs entries for configuring weighted interleave were created
for all possible nodes (N_POSSIBLE) at initialization, including nodes
that might not have memory. However, not all nodes in N_POSSIBLE are
usable at runtime, as some may remain memoryless or offline. This led to
sysfs entries being created for unusable nodes, causing potential
misconfiguration issues.
To address this issue, this patch modifies the sysfs creation logic to:
1) Limit sysfs entries to nodes that are online and have memory, avoiding
the creation of sysfs entries for nodes that cannot be used.
2) Support memory hotplug by dynamically adding and removing sysfs entries
based on whether a node transitions into or out of the N_MEMORY state.
Additionally, the patch ensures that sysfs attributes are properly managed
when nodes go offline, preventing stale or redundant entries from
persisting in the system.
By making these changes, the weighted interleave policy now manages its
sysfs entries more efficiently, ensuring that only relevant nodes are
considered for interleaving, and dynamically adapting to memory hotplug
events.
[dan.carpenter@linaro.org: fix error code in sysfs_wi_node_add()]
Link: https://lkml.kernel.org/r/aBjL7Bwc0QBzgajK@stanley.mountain
Link: https://lkml.kernel.org/r/20250417072839.711-4-rakie.kim@sk.com
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Dan Carpenter <dan.carpenter@linaro.org>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:37 +09:00
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
mm/mempolicy: support memory hotplug in weighted interleave
The weighted interleave policy distributes page allocations across
multiple NUMA nodes based on their performance weight, thereby improving
memory bandwidth utilization. The weight values for each node are
configured through sysfs.
Previously, sysfs entries for configuring weighted interleave were created
for all possible nodes (N_POSSIBLE) at initialization, including nodes
that might not have memory. However, not all nodes in N_POSSIBLE are
usable at runtime, as some may remain memoryless or offline. This led to
sysfs entries being created for unusable nodes, causing potential
misconfiguration issues.
To address this issue, this patch modifies the sysfs creation logic to:
1) Limit sysfs entries to nodes that are online and have memory, avoiding
the creation of sysfs entries for nodes that cannot be used.
2) Support memory hotplug by dynamically adding and removing sysfs entries
based on whether a node transitions into or out of the N_MEMORY state.
Additionally, the patch ensures that sysfs attributes are properly managed
when nodes go offline, preventing stale or redundant entries from
persisting in the system.
By making these changes, the weighted interleave policy now manages its
sysfs entries more efficiently, ensuring that only relevant nodes are
considered for interleaving, and dynamically adapting to memory hotplug
events.
[dan.carpenter@linaro.org: fix error code in sysfs_wi_node_add()]
Link: https://lkml.kernel.org/r/aBjL7Bwc0QBzgajK@stanley.mountain
Link: https://lkml.kernel.org/r/20250417072839.711-4-rakie.kim@sk.com
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Dan Carpenter <dan.carpenter@linaro.org>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:37 +09:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-02-21 16:37:42 -08:00
|
|
|
|
mm/mempolicy: support memory hotplug in weighted interleave
The weighted interleave policy distributes page allocations across
multiple NUMA nodes based on their performance weight, thereby improving
memory bandwidth utilization. The weight values for each node are
configured through sysfs.
Previously, sysfs entries for configuring weighted interleave were created
for all possible nodes (N_POSSIBLE) at initialization, including nodes
that might not have memory. However, not all nodes in N_POSSIBLE are
usable at runtime, as some may remain memoryless or offline. This led to
sysfs entries being created for unusable nodes, causing potential
misconfiguration issues.
To address this issue, this patch modifies the sysfs creation logic to:
1) Limit sysfs entries to nodes that are online and have memory, avoiding
the creation of sysfs entries for nodes that cannot be used.
2) Support memory hotplug by dynamically adding and removing sysfs entries
based on whether a node transitions into or out of the N_MEMORY state.
Additionally, the patch ensures that sysfs attributes are properly managed
when nodes go offline, preventing stale or redundant entries from
persisting in the system.
By making these changes, the weighted interleave policy now manages its
sysfs entries more efficiently, ensuring that only relevant nodes are
considered for interleaving, and dynamically adapting to memory hotplug
events.
[dan.carpenter@linaro.org: fix error code in sysfs_wi_node_add()]
Link: https://lkml.kernel.org/r/aBjL7Bwc0QBzgajK@stanley.mountain
Link: https://lkml.kernel.org/r/20250417072839.711-4-rakie.kim@sk.com
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Dan Carpenter <dan.carpenter@linaro.org>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:37 +09:00
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: support memory hotplug in weighted interleave
The weighted interleave policy distributes page allocations across
multiple NUMA nodes based on their performance weight, thereby improving
memory bandwidth utilization. The weight values for each node are
configured through sysfs.
Previously, sysfs entries for configuring weighted interleave were created
for all possible nodes (N_POSSIBLE) at initialization, including nodes
that might not have memory. However, not all nodes in N_POSSIBLE are
usable at runtime, as some may remain memoryless or offline. This led to
sysfs entries being created for unusable nodes, causing potential
misconfiguration issues.
To address this issue, this patch modifies the sysfs creation logic to:
1) Limit sysfs entries to nodes that are online and have memory, avoiding
the creation of sysfs entries for nodes that cannot be used.
2) Support memory hotplug by dynamically adding and removing sysfs entries
based on whether a node transitions into or out of the N_MEMORY state.
Additionally, the patch ensures that sysfs attributes are properly managed
when nodes go offline, preventing stale or redundant entries from
persisting in the system.
By making these changes, the weighted interleave policy now manages its
sysfs entries more efficiently, ensuring that only relevant nodes are
considered for interleaving, and dynamically adapting to memory hotplug
events.
[dan.carpenter@linaro.org: fix error code in sysfs_wi_node_add()]
Link: https://lkml.kernel.org/r/aBjL7Bwc0QBzgajK@stanley.mountain
Link: https://lkml.kernel.org/r/20250417072839.711-4-rakie.kim@sk.com
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Dan Carpenter <dan.carpenter@linaro.org>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:37 +09:00
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: support memory hotplug in weighted interleave
The weighted interleave policy distributes page allocations across
multiple NUMA nodes based on their performance weight, thereby improving
memory bandwidth utilization. The weight values for each node are
configured through sysfs.
Previously, sysfs entries for configuring weighted interleave were created
for all possible nodes (N_POSSIBLE) at initialization, including nodes
that might not have memory. However, not all nodes in N_POSSIBLE are
usable at runtime, as some may remain memoryless or offline. This led to
sysfs entries being created for unusable nodes, causing potential
misconfiguration issues.
To address this issue, this patch modifies the sysfs creation logic to:
1) Limit sysfs entries to nodes that are online and have memory, avoiding
the creation of sysfs entries for nodes that cannot be used.
2) Support memory hotplug by dynamically adding and removing sysfs entries
based on whether a node transitions into or out of the N_MEMORY state.
Additionally, the patch ensures that sysfs attributes are properly managed
when nodes go offline, preventing stale or redundant entries from
persisting in the system.
By making these changes, the weighted interleave policy now manages its
sysfs entries more efficiently, ensuring that only relevant nodes are
considered for interleaving, and dynamically adapting to memory hotplug
events.
[dan.carpenter@linaro.org: fix error code in sysfs_wi_node_add()]
Link: https://lkml.kernel.org/r/aBjL7Bwc0QBzgajK@stanley.mountain
Link: https://lkml.kernel.org/r/20250417072839.711-4-rakie.kim@sk.com
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Dan Carpenter <dan.carpenter@linaro.org>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:37 +09:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
mm/mempolicy: support memory hotplug in weighted interleave
The weighted interleave policy distributes page allocations across
multiple NUMA nodes based on their performance weight, thereby improving
memory bandwidth utilization. The weight values for each node are
configured through sysfs.
Previously, sysfs entries for configuring weighted interleave were created
for all possible nodes (N_POSSIBLE) at initialization, including nodes
that might not have memory. However, not all nodes in N_POSSIBLE are
usable at runtime, as some may remain memoryless or offline. This led to
sysfs entries being created for unusable nodes, causing potential
misconfiguration issues.
To address this issue, this patch modifies the sysfs creation logic to:
1) Limit sysfs entries to nodes that are online and have memory, avoiding
the creation of sysfs entries for nodes that cannot be used.
2) Support memory hotplug by dynamically adding and removing sysfs entries
based on whether a node transitions into or out of the N_MEMORY state.
Additionally, the patch ensures that sysfs attributes are properly managed
when nodes go offline, preventing stale or redundant entries from
persisting in the system.
By making these changes, the weighted interleave policy now manages its
sysfs entries more efficiently, ensuring that only relevant nodes are
considered for interleaving, and dynamically adapting to memory hotplug
events.
[dan.carpenter@linaro.org: fix error code in sysfs_wi_node_add()]
Link: https://lkml.kernel.org/r/aBjL7Bwc0QBzgajK@stanley.mountain
Link: https://lkml.kernel.org/r/20250417072839.711-4-rakie.kim@sk.com
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Dan Carpenter <dan.carpenter@linaro.org>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:37 +09:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: support memory hotplug in weighted interleave
The weighted interleave policy distributes page allocations across
multiple NUMA nodes based on their performance weight, thereby improving
memory bandwidth utilization. The weight values for each node are
configured through sysfs.
Previously, sysfs entries for configuring weighted interleave were created
for all possible nodes (N_POSSIBLE) at initialization, including nodes
that might not have memory. However, not all nodes in N_POSSIBLE are
usable at runtime, as some may remain memoryless or offline. This led to
sysfs entries being created for unusable nodes, causing potential
misconfiguration issues.
To address this issue, this patch modifies the sysfs creation logic to:
1) Limit sysfs entries to nodes that are online and have memory, avoiding
the creation of sysfs entries for nodes that cannot be used.
2) Support memory hotplug by dynamically adding and removing sysfs entries
based on whether a node transitions into or out of the N_MEMORY state.
Additionally, the patch ensures that sysfs attributes are properly managed
when nodes go offline, preventing stale or redundant entries from
persisting in the system.
By making these changes, the weighted interleave policy now manages its
sysfs entries more efficiently, ensuring that only relevant nodes are
considered for interleaving, and dynamically adapting to memory hotplug
events.
[dan.carpenter@linaro.org: fix error code in sysfs_wi_node_add()]
Link: https://lkml.kernel.org/r/aBjL7Bwc0QBzgajK@stanley.mountain
Link: https://lkml.kernel.org/r/20250417072839.711-4-rakie.kim@sk.com
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Dan Carpenter <dan.carpenter@linaro.org>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:37 +09:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
mm/mempolicy: support memory hotplug in weighted interleave
The weighted interleave policy distributes page allocations across
multiple NUMA nodes based on their performance weight, thereby improving
memory bandwidth utilization. The weight values for each node are
configured through sysfs.
Previously, sysfs entries for configuring weighted interleave were created
for all possible nodes (N_POSSIBLE) at initialization, including nodes
that might not have memory. However, not all nodes in N_POSSIBLE are
usable at runtime, as some may remain memoryless or offline. This led to
sysfs entries being created for unusable nodes, causing potential
misconfiguration issues.
To address this issue, this patch modifies the sysfs creation logic to:
1) Limit sysfs entries to nodes that are online and have memory, avoiding
the creation of sysfs entries for nodes that cannot be used.
2) Support memory hotplug by dynamically adding and removing sysfs entries
based on whether a node transitions into or out of the N_MEMORY state.
Additionally, the patch ensures that sysfs attributes are properly managed
when nodes go offline, preventing stale or redundant entries from
persisting in the system.
By making these changes, the weighted interleave policy now manages its
sysfs entries more efficiently, ensuring that only relevant nodes are
considered for interleaving, and dynamically adapting to memory hotplug
events.
[dan.carpenter@linaro.org: fix error code in sysfs_wi_node_add()]
Link: https://lkml.kernel.org/r/aBjL7Bwc0QBzgajK@stanley.mountain
Link: https://lkml.kernel.org/r/20250417072839.711-4-rakie.kim@sk.com
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Dan Carpenter <dan.carpenter@linaro.org>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:37 +09:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-06-16 15:51:52 +02:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: support memory hotplug in weighted interleave
The weighted interleave policy distributes page allocations across
multiple NUMA nodes based on their performance weight, thereby improving
memory bandwidth utilization. The weight values for each node are
configured through sysfs.
Previously, sysfs entries for configuring weighted interleave were created
for all possible nodes (N_POSSIBLE) at initialization, including nodes
that might not have memory. However, not all nodes in N_POSSIBLE are
usable at runtime, as some may remain memoryless or offline. This led to
sysfs entries being created for unusable nodes, causing potential
misconfiguration issues.
To address this issue, this patch modifies the sysfs creation logic to:
1) Limit sysfs entries to nodes that are online and have memory, avoiding
the creation of sysfs entries for nodes that cannot be used.
2) Support memory hotplug by dynamically adding and removing sysfs entries
based on whether a node transitions into or out of the N_MEMORY state.
Additionally, the patch ensures that sysfs attributes are properly managed
when nodes go offline, preventing stale or redundant entries from
persisting in the system.
By making these changes, the weighted interleave policy now manages its
sysfs entries more efficiently, ensuring that only relevant nodes are
considered for interleaving, and dynamically adapting to memory hotplug
events.
[dan.carpenter@linaro.org: fix error code in sysfs_wi_node_add()]
Link: https://lkml.kernel.org/r/aBjL7Bwc0QBzgajK@stanley.mountain
Link: https://lkml.kernel.org/r/20250417072839.711-4-rakie.kim@sk.com
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Dan Carpenter <dan.carpenter@linaro.org>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:37 +09:00
|
|
|
|
|
|
|
|
|
2025-06-16 15:51:52 +02:00
|
|
|
|
mm/mempolicy: support memory hotplug in weighted interleave
The weighted interleave policy distributes page allocations across
multiple NUMA nodes based on their performance weight, thereby improving
memory bandwidth utilization. The weight values for each node are
configured through sysfs.
Previously, sysfs entries for configuring weighted interleave were created
for all possible nodes (N_POSSIBLE) at initialization, including nodes
that might not have memory. However, not all nodes in N_POSSIBLE are
usable at runtime, as some may remain memoryless or offline. This led to
sysfs entries being created for unusable nodes, causing potential
misconfiguration issues.
To address this issue, this patch modifies the sysfs creation logic to:
1) Limit sysfs entries to nodes that are online and have memory, avoiding
the creation of sysfs entries for nodes that cannot be used.
2) Support memory hotplug by dynamically adding and removing sysfs entries
based on whether a node transitions into or out of the N_MEMORY state.
Additionally, the patch ensures that sysfs attributes are properly managed
when nodes go offline, preventing stale or redundant entries from
persisting in the system.
By making these changes, the weighted interleave policy now manages its
sysfs entries more efficiently, ensuring that only relevant nodes are
considered for interleaving, and dynamically adapting to memory hotplug
events.
[dan.carpenter@linaro.org: fix error code in sysfs_wi_node_add()]
Link: https://lkml.kernel.org/r/aBjL7Bwc0QBzgajK@stanley.mountain
Link: https://lkml.kernel.org/r/20250417072839.711-4-rakie.kim@sk.com
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Dan Carpenter <dan.carpenter@linaro.org>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:37 +09:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-06-16 15:51:52 +02:00
|
|
|
|
mm/mempolicy: support memory hotplug in weighted interleave
The weighted interleave policy distributes page allocations across
multiple NUMA nodes based on their performance weight, thereby improving
memory bandwidth utilization. The weight values for each node are
configured through sysfs.
Previously, sysfs entries for configuring weighted interleave were created
for all possible nodes (N_POSSIBLE) at initialization, including nodes
that might not have memory. However, not all nodes in N_POSSIBLE are
usable at runtime, as some may remain memoryless or offline. This led to
sysfs entries being created for unusable nodes, causing potential
misconfiguration issues.
To address this issue, this patch modifies the sysfs creation logic to:
1) Limit sysfs entries to nodes that are online and have memory, avoiding
the creation of sysfs entries for nodes that cannot be used.
2) Support memory hotplug by dynamically adding and removing sysfs entries
based on whether a node transitions into or out of the N_MEMORY state.
Additionally, the patch ensures that sysfs attributes are properly managed
when nodes go offline, preventing stale or redundant entries from
persisting in the system.
By making these changes, the weighted interleave policy now manages its
sysfs entries more efficiently, ensuring that only relevant nodes are
considered for interleaving, and dynamically adapting to memory hotplug
events.
[dan.carpenter@linaro.org: fix error code in sysfs_wi_node_add()]
Link: https://lkml.kernel.org/r/aBjL7Bwc0QBzgajK@stanley.mountain
Link: https://lkml.kernel.org/r/20250417072839.711-4-rakie.kim@sk.com
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Dan Carpenter <dan.carpenter@linaro.org>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:37 +09:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
|
|
|
|
|
2025-04-17 16:28:36 +09:00
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-02-21 17:06:51 -08:00
|
|
|
|
2025-04-17 16:28:36 +09:00
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
mm/mempolicy: support memory hotplug in weighted interleave
The weighted interleave policy distributes page allocations across
multiple NUMA nodes based on their performance weight, thereby improving
memory bandwidth utilization. The weight values for each node are
configured through sysfs.
Previously, sysfs entries for configuring weighted interleave were created
for all possible nodes (N_POSSIBLE) at initialization, including nodes
that might not have memory. However, not all nodes in N_POSSIBLE are
usable at runtime, as some may remain memoryless or offline. This led to
sysfs entries being created for unusable nodes, causing potential
misconfiguration issues.
To address this issue, this patch modifies the sysfs creation logic to:
1) Limit sysfs entries to nodes that are online and have memory, avoiding
the creation of sysfs entries for nodes that cannot be used.
2) Support memory hotplug by dynamically adding and removing sysfs entries
based on whether a node transitions into or out of the N_MEMORY state.
Additionally, the patch ensures that sysfs attributes are properly managed
when nodes go offline, preventing stale or redundant entries from
persisting in the system.
By making these changes, the weighted interleave policy now manages its
sysfs entries more efficiently, ensuring that only relevant nodes are
considered for interleaving, and dynamically adapting to memory hotplug
events.
[dan.carpenter@linaro.org: fix error code in sysfs_wi_node_add()]
Link: https://lkml.kernel.org/r/aBjL7Bwc0QBzgajK@stanley.mountain
Link: https://lkml.kernel.org/r/20250417072839.711-4-rakie.kim@sk.com
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Dan Carpenter <dan.carpenter@linaro.org>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:37 +09:00
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
2025-04-17 16:28:36 +09:00
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
mm/mempolicy: fix memory leaks in weighted interleave sysfs
Patch series "Enhance sysfs handling for memory hotplug in weighted
interleave", v9.
The following patch series enhances the weighted interleave policy in the
memory management subsystem by improving sysfs handling, fixing memory
leaks, and introducing dynamic sysfs updates for memory hotplug support.
This patch (of 3):
Memory leaks occurred when removing sysfs attributes for weighted
interleave. Improper kobject deallocation led to unreleased memory when
initialization failed or when nodes were removed.
The risk of leak is low because it only appears to trigger if setup
fails. Setup only fails due to -ENOMEM which is unlikely to happen
from a late_initcall() when memory pressure is low.
This patch resolves the issue by replacing unnecessary `kfree()` calls
with proper `kobject_del()` and `kobject_put()` sequences, ensuring
correct teardown and preventing memory leaks.
By explicitly calling `kobject_del()` before `kobject_put()`, the release
function is now invoked safely, and internal sysfs state is correctly
cleaned up. This guarantees that the memory associated with the kobject
is fully released and avoids resource leaks, thereby improving system
stability.
Additionally, sysfs_remove_file() is no longer called from the release
function to avoid accessing invalid sysfs state after kobject_del(). All
attribute removals are now done before kobject_del(), preventing WARN_ON()
in kernfs and ensuring safe and consistent cleanup of sysfs entries.
Link: https://lkml.kernel.org/r/20250417072839.711-1-rakie.kim@sk.com
Link: https://lkml.kernel.org/r/20250417072839.711-2-rakie.kim@sk.com
Fixes: dce41f5ae253 ("mm/mempolicy: implement the sysfs-based weighted_interleave interface")
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Honggyu Kim <honggyu.kim@sk.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Yunjeong Mun <yunjeong.mun@sk.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:35 +09:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
mm/mempolicy: Weighted Interleave Auto-tuning
On machines with multiple memory nodes, interleaving page allocations
across nodes allows for better utilization of each node's bandwidth.
Previous work by Gregory Price [1] introduced weighted interleave, which
allowed for pages to be allocated across nodes according to user-set
ratios.
Ideally, these weights should be proportional to their bandwidth, so that
under bandwidth pressure, each node uses its maximal efficient bandwidth
and prevents latency from increasing exponentially.
Previously, weighted interleave's default weights were just 1s -- which
would be equivalent to the (unweighted) interleave mempolicy, which goes
through the nodes in a round-robin fashion, ignoring bandwidth
information.
This patch has two main goals: First, it makes weighted interleave easier
to use for users who wish to relieve bandwidth pressure when using nodes
with varying bandwidth (CXL). By providing a set of "real" default
weights that just work out of the box, users who might not have the
capability (or wish to) perform experimentation to find the most optimal
weights for their system can still take advantage of bandwidth-informed
weighted interleave.
Second, it allows for weighted interleave to dynamically adjust to
hotplugged memory with new bandwidth information. Instead of manually
updating node weights every time new bandwidth information is reported or
taken off, weighted interleave adjusts and provides a new set of default
weights for weighted interleave to use when there is a change in bandwidth
information.
To meet these goals, this patch introduces an auto-configuration mode for
the interleave weights that provides a reasonable set of default weights,
calculated using bandwidth data reported by the system. In auto mode,
weights are dynamically adjusted based on whatever the current bandwidth
information reports (and responds to hotplug events).
This patch still supports users manually writing weights into the nodeN
sysfs interface by entering into manual mode. When a user enters manual
mode, the system stops dynamically updating any of the node weights, even
during hotplug events that shift the optimal weight distribution.
A new sysfs interface "auto" is introduced, which allows users to switch
between the auto (writing 1 or Y) and manual (writing 0 or N) modes. The
system also automatically enters manual mode when a nodeN interface is
manually written to.
There is one functional change that this patch makes to the existing
weighted_interleave ABI: previously, writing 0 directly to a nodeN
interface was said to reset the weight to the system default. Before this
patch, the default for all weights were 1, which meant that writing 0 and
1 were functionally equivalent. With this patch, writing 0 is invalid.
Link: https://lkml.kernel.org/r/20250520141236.2987309-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: wordsmithing changes, simplification, fixes]
Link: https://lkml.kernel.org/r/20250511025840.2410154-1-joshua.hahnjy@gmail.com
[joshua.hahnjy@gmail.com: remove auto_kobj_attr field from struct sysfs_wi_group]
Link: https://lkml.kernel.org/r/20250512142511.3959833-1-joshua.hahnjy@gmail.com
https://lore.kernel.org/linux-mm/20240202170238.90004-1-gregory.price@memverge.com/ [1]
Link: https://lkml.kernel.org/r/20250505182328.4148265-1-joshua.hahnjy@gmail.com
Co-developed-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Suggested-by: Yunjeong Mun <yunjeong.mun@sk.com>
Suggested-by: Oscar Salvador <osalvador@suse.de>
Suggested-by: Ying Huang <ying.huang@linux.alibaba.com>
Suggested-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>
Reviewed-by: Huang Ying <ying.huang@linux.alibaba.com>
Reviewed-by: Honggyu Kim <honggyu.kim@sk.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-05 11:23:28 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: support memory hotplug in weighted interleave
The weighted interleave policy distributes page allocations across
multiple NUMA nodes based on their performance weight, thereby improving
memory bandwidth utilization. The weight values for each node are
configured through sysfs.
Previously, sysfs entries for configuring weighted interleave were created
for all possible nodes (N_POSSIBLE) at initialization, including nodes
that might not have memory. However, not all nodes in N_POSSIBLE are
usable at runtime, as some may remain memoryless or offline. This led to
sysfs entries being created for unusable nodes, causing potential
misconfiguration issues.
To address this issue, this patch modifies the sysfs creation logic to:
1) Limit sysfs entries to nodes that are online and have memory, avoiding
the creation of sysfs entries for nodes that cannot be used.
2) Support memory hotplug by dynamically adding and removing sysfs entries
based on whether a node transitions into or out of the N_MEMORY state.
Additionally, the patch ensures that sysfs attributes are properly managed
when nodes go offline, preventing stale or redundant entries from
persisting in the system.
By making these changes, the weighted interleave policy now manages its
sysfs entries more efficiently, ensuring that only relevant nodes are
considered for interleaving, and dynamically adapting to memory hotplug
events.
[dan.carpenter@linaro.org: fix error code in sysfs_wi_node_add()]
Link: https://lkml.kernel.org/r/aBjL7Bwc0QBzgajK@stanley.mountain
Link: https://lkml.kernel.org/r/20250417072839.711-4-rakie.kim@sk.com
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Dan Carpenter <dan.carpenter@linaro.org>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:37 +09:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-04-17 16:28:36 +09:00
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
mm/mempolicy: support memory hotplug in weighted interleave
The weighted interleave policy distributes page allocations across
multiple NUMA nodes based on their performance weight, thereby improving
memory bandwidth utilization. The weight values for each node are
configured through sysfs.
Previously, sysfs entries for configuring weighted interleave were created
for all possible nodes (N_POSSIBLE) at initialization, including nodes
that might not have memory. However, not all nodes in N_POSSIBLE are
usable at runtime, as some may remain memoryless or offline. This led to
sysfs entries being created for unusable nodes, causing potential
misconfiguration issues.
To address this issue, this patch modifies the sysfs creation logic to:
1) Limit sysfs entries to nodes that are online and have memory, avoiding
the creation of sysfs entries for nodes that cannot be used.
2) Support memory hotplug by dynamically adding and removing sysfs entries
based on whether a node transitions into or out of the N_MEMORY state.
Additionally, the patch ensures that sysfs attributes are properly managed
when nodes go offline, preventing stale or redundant entries from
persisting in the system.
By making these changes, the weighted interleave policy now manages its
sysfs entries more efficiently, ensuring that only relevant nodes are
considered for interleaving, and dynamically adapting to memory hotplug
events.
[dan.carpenter@linaro.org: fix error code in sysfs_wi_node_add()]
Link: https://lkml.kernel.org/r/aBjL7Bwc0QBzgajK@stanley.mountain
Link: https://lkml.kernel.org/r/20250417072839.711-4-rakie.kim@sk.com
Co-developed-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Yunjeong Mun <yunjeong.mun@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Dan Carpenter <dan.carpenter@linaro.org>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:37 +09:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: fix memory leaks in weighted interleave sysfs
Patch series "Enhance sysfs handling for memory hotplug in weighted
interleave", v9.
The following patch series enhances the weighted interleave policy in the
memory management subsystem by improving sysfs handling, fixing memory
leaks, and introducing dynamic sysfs updates for memory hotplug support.
This patch (of 3):
Memory leaks occurred when removing sysfs attributes for weighted
interleave. Improper kobject deallocation led to unreleased memory when
initialization failed or when nodes were removed.
The risk of leak is low because it only appears to trigger if setup
fails. Setup only fails due to -ENOMEM which is unlikely to happen
from a late_initcall() when memory pressure is low.
This patch resolves the issue by replacing unnecessary `kfree()` calls
with proper `kobject_del()` and `kobject_put()` sequences, ensuring
correct teardown and preventing memory leaks.
By explicitly calling `kobject_del()` before `kobject_put()`, the release
function is now invoked safely, and internal sysfs state is correctly
cleaned up. This guarantees that the memory associated with the kobject
is fully released and avoids resource leaks, thereby improving system
stability.
Additionally, sysfs_remove_file() is no longer called from the release
function to avoid accessing invalid sysfs state after kobject_del(). All
attribute removals are now done before kobject_del(), preventing WARN_ON()
in kernfs and ensuring safe and consistent cleanup of sysfs entries.
Link: https://lkml.kernel.org/r/20250417072839.711-1-rakie.kim@sk.com
Link: https://lkml.kernel.org/r/20250417072839.711-2-rakie.kim@sk.com
Fixes: dce41f5ae253 ("mm/mempolicy: implement the sysfs-based weighted_interleave interface")
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Honggyu Kim <honggyu.kim@sk.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Yunjeong Mun <yunjeong.mun@sk.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:35 +09:00
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: fix memory leaks in weighted interleave sysfs
Patch series "Enhance sysfs handling for memory hotplug in weighted
interleave", v9.
The following patch series enhances the weighted interleave policy in the
memory management subsystem by improving sysfs handling, fixing memory
leaks, and introducing dynamic sysfs updates for memory hotplug support.
This patch (of 3):
Memory leaks occurred when removing sysfs attributes for weighted
interleave. Improper kobject deallocation led to unreleased memory when
initialization failed or when nodes were removed.
The risk of leak is low because it only appears to trigger if setup
fails. Setup only fails due to -ENOMEM which is unlikely to happen
from a late_initcall() when memory pressure is low.
This patch resolves the issue by replacing unnecessary `kfree()` calls
with proper `kobject_del()` and `kobject_put()` sequences, ensuring
correct teardown and preventing memory leaks.
By explicitly calling `kobject_del()` before `kobject_put()`, the release
function is now invoked safely, and internal sysfs state is correctly
cleaned up. This guarantees that the memory associated with the kobject
is fully released and avoids resource leaks, thereby improving system
stability.
Additionally, sysfs_remove_file() is no longer called from the release
function to avoid accessing invalid sysfs state after kobject_del(). All
attribute removals are now done before kobject_del(), preventing WARN_ON()
in kernfs and ensuring safe and consistent cleanup of sysfs entries.
Link: https://lkml.kernel.org/r/20250417072839.711-1-rakie.kim@sk.com
Link: https://lkml.kernel.org/r/20250417072839.711-2-rakie.kim@sk.com
Fixes: dce41f5ae253 ("mm/mempolicy: implement the sysfs-based weighted_interleave interface")
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Honggyu Kim <honggyu.kim@sk.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Yunjeong Mun <yunjeong.mun@sk.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:35 +09:00
|
|
|
|
2025-06-16 15:51:52 +02:00
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
mm/mempolicy: fix memory leaks in weighted interleave sysfs
Patch series "Enhance sysfs handling for memory hotplug in weighted
interleave", v9.
The following patch series enhances the weighted interleave policy in the
memory management subsystem by improving sysfs handling, fixing memory
leaks, and introducing dynamic sysfs updates for memory hotplug support.
This patch (of 3):
Memory leaks occurred when removing sysfs attributes for weighted
interleave. Improper kobject deallocation led to unreleased memory when
initialization failed or when nodes were removed.
The risk of leak is low because it only appears to trigger if setup
fails. Setup only fails due to -ENOMEM which is unlikely to happen
from a late_initcall() when memory pressure is low.
This patch resolves the issue by replacing unnecessary `kfree()` calls
with proper `kobject_del()` and `kobject_put()` sequences, ensuring
correct teardown and preventing memory leaks.
By explicitly calling `kobject_del()` before `kobject_put()`, the release
function is now invoked safely, and internal sysfs state is correctly
cleaned up. This guarantees that the memory associated with the kobject
is fully released and avoids resource leaks, thereby improving system
stability.
Additionally, sysfs_remove_file() is no longer called from the release
function to avoid accessing invalid sysfs state after kobject_del(). All
attribute removals are now done before kobject_del(), preventing WARN_ON()
in kernfs and ensuring safe and consistent cleanup of sysfs entries.
Link: https://lkml.kernel.org/r/20250417072839.711-1-rakie.kim@sk.com
Link: https://lkml.kernel.org/r/20250417072839.711-2-rakie.kim@sk.com
Fixes: dce41f5ae253 ("mm/mempolicy: implement the sysfs-based weighted_interleave interface")
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Honggyu Kim <honggyu.kim@sk.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Yunjeong Mun <yunjeong.mun@sk.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:35 +09:00
|
|
|
|
|
|
|
|
|
2025-04-17 16:28:36 +09:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: fix memory leaks in weighted interleave sysfs
Patch series "Enhance sysfs handling for memory hotplug in weighted
interleave", v9.
The following patch series enhances the weighted interleave policy in the
memory management subsystem by improving sysfs handling, fixing memory
leaks, and introducing dynamic sysfs updates for memory hotplug support.
This patch (of 3):
Memory leaks occurred when removing sysfs attributes for weighted
interleave. Improper kobject deallocation led to unreleased memory when
initialization failed or when nodes were removed.
The risk of leak is low because it only appears to trigger if setup
fails. Setup only fails due to -ENOMEM which is unlikely to happen
from a late_initcall() when memory pressure is low.
This patch resolves the issue by replacing unnecessary `kfree()` calls
with proper `kobject_del()` and `kobject_put()` sequences, ensuring
correct teardown and preventing memory leaks.
By explicitly calling `kobject_del()` before `kobject_put()`, the release
function is now invoked safely, and internal sysfs state is correctly
cleaned up. This guarantees that the memory associated with the kobject
is fully released and avoids resource leaks, thereby improving system
stability.
Additionally, sysfs_remove_file() is no longer called from the release
function to avoid accessing invalid sysfs state after kobject_del(). All
attribute removals are now done before kobject_del(), preventing WARN_ON()
in kernfs and ensuring safe and consistent cleanup of sysfs entries.
Link: https://lkml.kernel.org/r/20250417072839.711-1-rakie.kim@sk.com
Link: https://lkml.kernel.org/r/20250417072839.711-2-rakie.kim@sk.com
Fixes: dce41f5ae253 ("mm/mempolicy: implement the sysfs-based weighted_interleave interface")
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Honggyu Kim <honggyu.kim@sk.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Yunjeong Mun <yunjeong.mun@sk.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:35 +09:00
|
|
|
|
2025-04-17 16:28:36 +09:00
|
|
|
|
mm/mempolicy: fix memory leaks in weighted interleave sysfs
Patch series "Enhance sysfs handling for memory hotplug in weighted
interleave", v9.
The following patch series enhances the weighted interleave policy in the
memory management subsystem by improving sysfs handling, fixing memory
leaks, and introducing dynamic sysfs updates for memory hotplug support.
This patch (of 3):
Memory leaks occurred when removing sysfs attributes for weighted
interleave. Improper kobject deallocation led to unreleased memory when
initialization failed or when nodes were removed.
The risk of leak is low because it only appears to trigger if setup
fails. Setup only fails due to -ENOMEM which is unlikely to happen
from a late_initcall() when memory pressure is low.
This patch resolves the issue by replacing unnecessary `kfree()` calls
with proper `kobject_del()` and `kobject_put()` sequences, ensuring
correct teardown and preventing memory leaks.
By explicitly calling `kobject_del()` before `kobject_put()`, the release
function is now invoked safely, and internal sysfs state is correctly
cleaned up. This guarantees that the memory associated with the kobject
is fully released and avoids resource leaks, thereby improving system
stability.
Additionally, sysfs_remove_file() is no longer called from the release
function to avoid accessing invalid sysfs state after kobject_del(). All
attribute removals are now done before kobject_del(), preventing WARN_ON()
in kernfs and ensuring safe and consistent cleanup of sysfs entries.
Link: https://lkml.kernel.org/r/20250417072839.711-1-rakie.kim@sk.com
Link: https://lkml.kernel.org/r/20250417072839.711-2-rakie.kim@sk.com
Fixes: dce41f5ae253 ("mm/mempolicy: implement the sysfs-based weighted_interleave interface")
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Honggyu Kim <honggyu.kim@sk.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Yunjeong Mun <yunjeong.mun@sk.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:35 +09:00
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: fix memory leaks in weighted interleave sysfs
Patch series "Enhance sysfs handling for memory hotplug in weighted
interleave", v9.
The following patch series enhances the weighted interleave policy in the
memory management subsystem by improving sysfs handling, fixing memory
leaks, and introducing dynamic sysfs updates for memory hotplug support.
This patch (of 3):
Memory leaks occurred when removing sysfs attributes for weighted
interleave. Improper kobject deallocation led to unreleased memory when
initialization failed or when nodes were removed.
The risk of leak is low because it only appears to trigger if setup
fails. Setup only fails due to -ENOMEM which is unlikely to happen
from a late_initcall() when memory pressure is low.
This patch resolves the issue by replacing unnecessary `kfree()` calls
with proper `kobject_del()` and `kobject_put()` sequences, ensuring
correct teardown and preventing memory leaks.
By explicitly calling `kobject_del()` before `kobject_put()`, the release
function is now invoked safely, and internal sysfs state is correctly
cleaned up. This guarantees that the memory associated with the kobject
is fully released and avoids resource leaks, thereby improving system
stability.
Additionally, sysfs_remove_file() is no longer called from the release
function to avoid accessing invalid sysfs state after kobject_del(). All
attribute removals are now done before kobject_del(), preventing WARN_ON()
in kernfs and ensuring safe and consistent cleanup of sysfs entries.
Link: https://lkml.kernel.org/r/20250417072839.711-1-rakie.kim@sk.com
Link: https://lkml.kernel.org/r/20250417072839.711-2-rakie.kim@sk.com
Fixes: dce41f5ae253 ("mm/mempolicy: implement the sysfs-based weighted_interleave interface")
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Honggyu Kim <honggyu.kim@sk.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Yunjeong Mun <yunjeong.mun@sk.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:35 +09:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: fix memory leaks in weighted interleave sysfs
Patch series "Enhance sysfs handling for memory hotplug in weighted
interleave", v9.
The following patch series enhances the weighted interleave policy in the
memory management subsystem by improving sysfs handling, fixing memory
leaks, and introducing dynamic sysfs updates for memory hotplug support.
This patch (of 3):
Memory leaks occurred when removing sysfs attributes for weighted
interleave. Improper kobject deallocation led to unreleased memory when
initialization failed or when nodes were removed.
The risk of leak is low because it only appears to trigger if setup
fails. Setup only fails due to -ENOMEM which is unlikely to happen
from a late_initcall() when memory pressure is low.
This patch resolves the issue by replacing unnecessary `kfree()` calls
with proper `kobject_del()` and `kobject_put()` sequences, ensuring
correct teardown and preventing memory leaks.
By explicitly calling `kobject_del()` before `kobject_put()`, the release
function is now invoked safely, and internal sysfs state is correctly
cleaned up. This guarantees that the memory associated with the kobject
is fully released and avoids resource leaks, thereby improving system
stability.
Additionally, sysfs_remove_file() is no longer called from the release
function to avoid accessing invalid sysfs state after kobject_del(). All
attribute removals are now done before kobject_del(), preventing WARN_ON()
in kernfs and ensuring safe and consistent cleanup of sysfs entries.
Link: https://lkml.kernel.org/r/20250417072839.711-1-rakie.kim@sk.com
Link: https://lkml.kernel.org/r/20250417072839.711-2-rakie.kim@sk.com
Fixes: dce41f5ae253 ("mm/mempolicy: implement the sysfs-based weighted_interleave interface")
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Honggyu Kim <honggyu.kim@sk.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Yunjeong Mun <yunjeong.mun@sk.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:35 +09:00
|
|
|
|
|
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
mm/mempolicy: fix memory leaks in weighted interleave sysfs
Patch series "Enhance sysfs handling for memory hotplug in weighted
interleave", v9.
The following patch series enhances the weighted interleave policy in the
memory management subsystem by improving sysfs handling, fixing memory
leaks, and introducing dynamic sysfs updates for memory hotplug support.
This patch (of 3):
Memory leaks occurred when removing sysfs attributes for weighted
interleave. Improper kobject deallocation led to unreleased memory when
initialization failed or when nodes were removed.
The risk of leak is low because it only appears to trigger if setup
fails. Setup only fails due to -ENOMEM which is unlikely to happen
from a late_initcall() when memory pressure is low.
This patch resolves the issue by replacing unnecessary `kfree()` calls
with proper `kobject_del()` and `kobject_put()` sequences, ensuring
correct teardown and preventing memory leaks.
By explicitly calling `kobject_del()` before `kobject_put()`, the release
function is now invoked safely, and internal sysfs state is correctly
cleaned up. This guarantees that the memory associated with the kobject
is fully released and avoids resource leaks, thereby improving system
stability.
Additionally, sysfs_remove_file() is no longer called from the release
function to avoid accessing invalid sysfs state after kobject_del(). All
attribute removals are now done before kobject_del(), preventing WARN_ON()
in kernfs and ensuring safe and consistent cleanup of sysfs entries.
Link: https://lkml.kernel.org/r/20250417072839.711-1-rakie.kim@sk.com
Link: https://lkml.kernel.org/r/20250417072839.711-2-rakie.kim@sk.com
Fixes: dce41f5ae253 ("mm/mempolicy: implement the sysfs-based weighted_interleave interface")
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Reviewed-by: Gregory Price <gourry@gourry.net>
Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Honggyu Kim <honggyu.kim@sk.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Yunjeong Mun <yunjeong.mun@sk.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-04-17 16:28:35 +09:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/mempolicy: implement the sysfs-based weighted_interleave interface
Patch series "mm/mempolicy: weighted interleave mempolicy and sysfs
extension", v5.
Weighted interleave is a new interleave policy intended to make use of
heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin distribution of
memory across all nodes in a nodemask, while weighted interleave
distributes memory across nodes according to a provided weight. (Weight =
# of page allocations per round)
Weighted interleave is intended to reduce average latency when bandwidth
is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available bandwidth in
a heterogeneous hardware environment (different hardware provides
different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly and then
exponentially. By keeping bandwidth usage distributed according to
available bandwidth, we therefore can reduce the average latency of a
cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using the
"maximum sustained bandwidth" of each hardware component in a heterogenous
memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted means
VMA policies were set (mbind2). We see weighted interleave is not always
beneficial when applied globally, but is always beneficial when applied to
bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <ravis.opensrc@micron.com>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <gregory.price@memverge.com>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <hyeongtak.ji@sk.com>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality but
did not adjust tests for weighting. Basically the weights were set to 1,
which is the default, and it should behave the same as MPOL_INTERLEAVE if
logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
This patch (of 4):
This patch provides a way to set interleave weight information under sysfs
at /sys/kernel/mm/mempolicy/weighted_interleave/nodeN
The sysfs structure is designed as follows.
$ tree /sys/kernel/mm/mempolicy/
/sys/kernel/mm/mempolicy/ [1]
└── weighted_interleave [2]
├── node0 [3]
└── node1
Each file above can be explained as follows.
[1] mm/mempolicy: configuration interface for mempolicy subsystem
[2] weighted_interleave/: config interface for weighted interleave policy
[3] weighted_interleave/nodeN: weight for nodeN
If a node value is set to `0`, the system-default value will be used.
As of this patch, the system-default for all nodes is always 1.
Link: https://lkml.kernel.org/r/20240202170238.90004-1-gregory.price@memverge.com
Link: https://lkml.kernel.org/r/20240202170238.90004-2-gregory.price@memverge.com
Suggested-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>
Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Co-developed-by: Gregory Price <gregory.price@memverge.com>
Signed-off-by: Gregory Price <gregory.price@memverge.com>
Co-developed-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hasan Al Maruf <Hasan.Maruf@amd.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Srinivasulu Thanneeru <sthanneeru.opensrc@micron.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-02-02 12:02:35 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|