Pss is the sum of the sizes of clean and dirty private pages, and the
proportional sizes of clean and dirty shared pages:
Private = Private_Dirty + Private_Clean
Shared_Proportional = Shared_Dirty_Proportional + Shared_Clean_Proportional
Pss = Private + Shared_Proportional
The Shared*Proportional fields are not present in smaps, so it is not
always possible to determine how much of the Pss is from dirty pages and
how much is from clean pages. This information can be useful for
measuring memory usage for the purpose of optimisation, since clean pages
can usually be discarded by the kernel immediately while dirty pages
cannot.
The smaps routines in the kernel already have access to this data, so add
a Pss_Dirty to show it to userspace. Pss_Clean is not added since it can
be calculated from Pss and Pss_Dirty.
Link: https://lkml.kernel.org/r/20220620081251.2928103-1-vincent.whitchurch@axis.com
Signed-off-by: Vincent Whitchurch <vincent.whitchurch@axis.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Passing index to pte_offset_map_lock() directly so the below calculation
can be avoided. Rename orig_pte to ptep as it's not changed. Also use
helper is_swap_pte() to improve the readability. No functional change
intended.
[akpm@linux-foundation.org: reduce scope of `ptep']
Link: https://lkml.kernel.org/r/20220618090527.37843-1-linmiaohe@huawei.com
Signed-off-by: Miaohe Lin <linmiaohe@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
commit 641844f561 ("mm/hugetlb: introduce minimum hugepage order") fixed
a static checker warning and introduced a global variable minimum_order to
fix the warning. However, the local variable in
dissolve_free_huge_pages() can be initialized to
huge_page_order(&default_hstate) to fix the warning.
So remove minimum_order to simplify the code.
Link: https://lkml.kernel.org/r/20220616033846.96937-1-songmuchun@bytedance.com
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
For now, the feature of hugetlb_free_vmemmap is not compatible with the
feature of memory_hotplug.memmap_on_memory, and hugetlb_free_vmemmap takes
precedence over memory_hotplug.memmap_on_memory. However, someone wants
to make memory_hotplug.memmap_on_memory takes precedence over
hugetlb_free_vmemmap since memmap_on_memory makes it more likely to
succeed memory hotplug in close-to-OOM situations. So the decision of
making hugetlb_free_vmemmap take precedence is not wise and elegant.
The proper approach is to have hugetlb_vmemmap.c do the check whether the
section which the HugeTLB pages belong to can be optimized. If the
section's vmemmap pages are allocated from the added memory block itself,
hugetlb_free_vmemmap should refuse to optimize the vmemmap, otherwise, do
the optimization. Then both kernel parameters are compatible. So this
patch introduces VmemmapSelfHosted to mask any non-optimizable vmemmap
pages. The hugetlb_vmemmap can use this flag to detect if a vmemmap page
can be optimized.
[songmuchun@bytedance.com: walk vmemmap page tables to avoid false-positive]
Link: https://lkml.kernel.org/r/20220620110616.12056-3-songmuchun@bytedance.com
Link: https://lkml.kernel.org/r/20220617135650.74901-3-songmuchun@bytedance.com
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Co-developed-by: Oscar Salvador <osalvador@suse.de>
Signed-off-by: Oscar Salvador <osalvador@suse.de>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Paul E. McKenney <paulmck@kernel.org>
Cc: Xiongchun Duan <duanxiongchun@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "make hugetlb_optimize_vmemmap compatible with
memmap_on_memory", v3.
This series makes hugetlb_optimize_vmemmap compatible with
memmap_on_memory.
This patch (of 2):
We are almost running out of section flags, only one bit is available in
the worst case (powerpc with 256k pages). However, there are still some
free bits (in ->section_mem_map) on other architectures (e.g. x86_64 has
10 bits available, arm64 has 8 bits available with worst case of 64K
pages). We have hard coded those numbers in code, it is inconvenient to
use those bits on other architectures except powerpc. So transfer those
section flags to enumeration to make it easy to add new section flags in
the future. Also, move SECTION_TAINT_ZONE_DEVICE into the scope of
CONFIG_ZONE_DEVICE to save a bit on non-zone-device case.
[songmuchun@bytedance.com: replace enum with defines per David]
Link: https://lkml.kernel.org/r/20220620110616.12056-2-songmuchun@bytedance.com
Link: https://lkml.kernel.org/r/20220617135650.74901-1-songmuchun@bytedance.com
Link: https://lkml.kernel.org/r/20220617135650.74901-2-songmuchun@bytedance.com
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Reviewed-by: David Hildenbrand <david@redhat.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Paul E. McKenney <paulmck@kernel.org>
Cc: Xiongchun Duan <duanxiongchun@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
The only caller already has a folio, so push the folio->page conversion
down a level.
Link: https://lkml.kernel.org/r/20220617175020.717127-21-willy@infradead.org
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
All callers now have a folio, so push the folio->page conversion
down to this function.
[akpm@linux-foundation.org: uninline destroy_large_folio() to fix build issue]
Link: https://lkml.kernel.org/r/20220617175020.717127-20-willy@infradead.org
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
All the callers now have a folio. Saves several calls to compound_head,
totalling 502 bytes of text.
Link: https://lkml.kernel.org/r/20220617175020.717127-19-willy@infradead.org
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
All the callers now have a folio, so pass it in. This doesn't
save any text, but it does save a call to compound_head() as
folio_test_hugetlb() does not contain a call like PageHuge() does.
Link: https://lkml.kernel.org/r/20220617175020.717127-18-willy@infradead.org
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Pages linked through the LRU list cannot be tail pages as ->compound_head
is in a union with one of the words of the list_head, and they cannot
be ZONE_DEVICE pages as ->pgmap is in a union with the same word.
Saves 60 bytes of text by removing a call to page_is_fake_head().
Link: https://lkml.kernel.org/r/20220617175020.717127-15-willy@infradead.org
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
This function was already calling compound_head(), but now it can
cache the result of calling compound_head() and avoid calling it again.
Saves 299 bytes of text by avoiding various calls to compound_page()
and avoiding checks of PageTail.
Link: https://lkml.kernel.org/r/20220617175020.717127-14-willy@infradead.org
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Save a few calls to compound_head by converting the passed page to
a folio. Reduces kernel text size by 74 bytes.
Link: https://lkml.kernel.org/r/20220617175020.717127-13-willy@infradead.org
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Do the per-cpu dereferencing of the fbatches once which saves 14 bytes
of text and several percpu relocations.
Link: https://lkml.kernel.org/r/20220617175020.717127-12-willy@infradead.org
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
The function is too long, so pull this complicated conditional out into
cpu_needs_drain(). This ends up shrinking the text by 14 bytes,
by allowing GCC to cache the result of calling per_cpu() instead of
relocating each lookup individually.
Link: https://lkml.kernel.org/r/20220617175020.717127-11-willy@infradead.org
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
No change to generated code, but this struct no longer contains any
pagevecs, and not all the folio batches it contains are lru.
Link: https://lkml.kernel.org/r/20220617175020.717127-10-willy@infradead.org
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Rename it to just 'activate', saving 696 bytes of text from removals
of compound_page() and the pagevec_lru_move_fn() infrastructure.
Inline need_activate_page_drain() into its only caller.
Link: https://lkml.kernel.org/r/20220617175020.717127-9-willy@infradead.org
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Using folios instead of pages removes several calls to compound_head(),
shrinking the kernel by 1089 bytes of text.
Link: https://lkml.kernel.org/r/20220617175020.717127-8-willy@infradead.org
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Using folios instead of pages shrinks deactivate_page() and
lru_deactivate_fn() by 778 bytes between them.
Link: https://lkml.kernel.org/r/20220617175020.717127-7-willy@infradead.org
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Use a folio throughout lru_deactivate_file_fn(), removing many hidden
calls to compound_head(). Shrinks the kernel by 864 bytes of text.
Link: https://lkml.kernel.org/r/20220617175020.717127-6-willy@infradead.org
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
When adding folios to the LRU for the first time, the LRU flag will
already be clear, so skip the test-and-clear part of moving from one
LRU to another.
Removes 285 bytes from kernel text, mostly due to removing
__pagevec_lru_add().
Link: https://lkml.kernel.org/r/20220617175020.717127-5-willy@infradead.org
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
__pagevec_lru_add has no callers outside swap.c, so make it static,
and move it to a more logical position in the file.
Link: https://lkml.kernel.org/r/20220617175020.717127-4-willy@infradead.org
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Start converting the LRU from pagevecs to folio_batches.
Combine the functionality of pagevec_add_and_need_flush() with
pagevec_lru_move_fn() in the new folio_batch_add_and_move().
Convert the lru_rotate pagevec to a folio_batch.
Adds 223 bytes total to kernel text, because we're duplicating
infrastructure. This will be more than made up for in future patches.
Link: https://lkml.kernel.org/r/20220617175020.717127-3-willy@infradead.org
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "Convert the swap code to be more folio-based".
There's still more to do with the swap code, but this reaps a lot of the
folio benefit. More than 4kB of kernel text saved (with the UEK7 kernel
config). I don't know how much that's going to translate into CPU
savings, but some of those compound_head() calls are on every page free,
so it should be noticable. It might even be noticable just from an
I-cache consumption perspective.
This patch (of 22):
This is just a wrapper around release_pages() for now. Place the
prototype in mm.h along with folio_put() and folio_put_refs().
Link: https://lkml.kernel.org/r/20220617175020.717127-1-willy@infradead.org
Link: https://lkml.kernel.org/r/20220617175020.717127-2-willy@infradead.org
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Remove a few hidden calls to compound_head, saving 76 bytes of text.
Link: https://lkml.kernel.org/r/20220617154248.700416-6-willy@infradead.org
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Remove a few hidden calls to compound_head, saving 411 bytes of text.
Link: https://lkml.kernel.org/r/20220617154248.700416-5-willy@infradead.org
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Remove a few hidden calls to compound_head, saving 387 bytes of text on
my test configuration.
Link: https://lkml.kernel.org/r/20220617154248.700416-4-willy@infradead.org
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Remove a few hidden calls to compound_head, saving 279 bytes of text.
Link: https://lkml.kernel.org/r/20220617154248.700416-3-willy@infradead.org
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "nvert much of vmscan to folios"
vmscan always operates on folios since it puts the pages on the LRU list.
Switching all of these functions from pages to folios saves 1483 bytes of
text from removing all the baggage around calling compound_page() and
similar functions.
This patch (of 5):
This is a straightforward conversion which removes several hidden calls
to compound_head, saving 330 bytes of kernel text.
Link: https://lkml.kernel.org/r/20220617154248.700416-1-willy@infradead.org
Link: https://lkml.kernel.org/r/20220617154248.700416-2-willy@infradead.org
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Similar to our MM_CP_DIRTY_ACCT handling for shared, writable mappings, we
can try mapping anonymous pages in a private writable mapping writable if
they are exclusive, the PTE is already dirty, and no special handling
applies. Mapping the anonymous page writable is essentially the same
thing the write fault handler would do in this case.
Special handling is required for uffd-wp and softdirty tracking, so take
care of that properly. Also, leave PROT_NONE handling alone for now; in
the future, we could similarly extend the logic in do_numa_page() or use
pte_mk_savedwrite() here.
While this improves mprotect(PROT_READ)+mprotect(PROT_READ|PROT_WRITE)
performance, it should also be a valuable optimization for uffd-wp, when
un-protecting.
This has been previously suggested by Peter Collingbourne in [1], relevant
in the context of the Scudo memory allocator, before we had
PageAnonExclusive.
This commit doesn't add the same handling for PMDs (i.e., anonymous THP,
anonymous hugetlb); benchmark results from Andrea indicate that there are
minor performance gains, so it's might still be valuable to streamline
that logic for all anonymous pages in the future.
As we now also set MM_CP_DIRTY_ACCT for private mappings, let's rename it
to MM_CP_TRY_CHANGE_WRITABLE, to make it clearer what's actually
happening.
Micro-benchmark courtesy of Andrea:
===
#define _GNU_SOURCE
#include <sys/mman.h>
#include <stdlib.h>
#include <string.h>
#include <stdio.h>
#include <unistd.h>
#define SIZE (1024*1024*1024)
int main(int argc, char *argv[])
{
char *p;
if (posix_memalign((void **)&p, sysconf(_SC_PAGESIZE)*512, SIZE))
perror("posix_memalign"), exit(1);
if (madvise(p, SIZE, argc > 1 ? MADV_HUGEPAGE : MADV_NOHUGEPAGE))
perror("madvise");
explicit_bzero(p, SIZE);
for (int loops = 0; loops < 40; loops++) {
if (mprotect(p, SIZE, PROT_READ))
perror("mprotect"), exit(1);
if (mprotect(p, SIZE, PROT_READ|PROT_WRITE))
perror("mprotect"), exit(1);
explicit_bzero(p, SIZE);
}
}
===
Results on my Ryzen 9 3900X:
Stock 10 runs (lower is better): AVG 6.398s, STDEV 0.043
Patched 10 runs (lower is better): AVG 3.780s, STDEV 0.026
===
[1] https://lkml.kernel.org/r/20210429214801.2583336-1-pcc@google.com
Link: https://lkml.kernel.org/r/20220614093629.76309-1-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Suggested-by: Peter Collingbourne <pcc@google.com>
Acked-by: Peter Xu <peterx@redhat.com>
Cc: Nadav Amit <nadav.amit@gmail.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
On Android this test is getting stuck in an infinite loop due to
indeterminate behavior:
The local variables steps and signalled were being reset to 1 and 0
respectively after every jump back to sigsetjmp by siglongjmp in the
signal handler. The test was incrementing them and expecting them to
retain their incremented values. The documentation for siglongjmp says:
All accessible objects have values as of the time sigsetjmp() was called,
except that the values of objects of automatic storage duration which are
local to the function containing the invocation of the corresponding
sigsetjmp() which do not have volatile-qualified type and which are
changed between the sigsetjmp() invocation and siglongjmp() call are
indeterminate.
Tagging steps and signalled with volatile enabled the test to pass.
Link: https://lkml.kernel.org/r/20220613233321.431282-1-edliaw@google.com
Signed-off-by: Edward Liaw <edliaw@google.com>
Reviewed-by: Axel Rasmussen <axelrasmussen@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Peter Xu <peterx@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
This commit documents the usage of DAMON_LRU_SORT for admins.
Link: https://lkml.kernel.org/r/20220613192301.8817-10-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Users can do data access-aware LRU-lists sorting using 'LRU_PRIO' and
'LRU_DEPRIO' DAMOS actions. However, finding best parameters including
the hotness/coldness thresholds, CPU quota, and watermarks could be
challenging for some users. To make the scheme easy to be used without
complex tuning for common situations, this commit implements a static
kernel module called 'DAMON_LRU_SORT' using the 'LRU_PRIO' and
'LRU_DEPRIO' DAMOS actions.
It proactively sorts LRU-lists using DAMON with conservatively chosen
default values of the parameters. That is, the module under its default
parameters will make no harm for common situations but provide some level
of efficiency improvements for systems having clear hot/cold access
pattern under a level of memory pressure while consuming only a limited
small portion of CPU time.
Link: https://lkml.kernel.org/r/20220613192301.8817-9-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
This commit documents the 'LRU_DEPRIO' scheme action for DAMON sysfs
interface.`
Link: https://lkml.kernel.org/r/20220613192301.8817-8-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
This commit adds a new DAMON-based operation scheme action called
'LRU_DEPRIO' for physical address space. The action deprioritizes pages
in the memory area of the target access pattern on their LRU lists. This
is hence supposed to be used for rarely accessed (cold) memory regions so
that cold pages could be more likely reclaimed first under memory
pressure. Internally, it simply calls 'lru_deactivate()'.
Using this with 'LRU_PRIO' action for hot pages, users can proactively
sort LRU lists based on the access pattern. That is, it can make the LRU
lists somewhat more trustworthy source of access temperature. As a
result, efficiency of LRU-lists based mechanisms including the reclamation
target selection could be improved.
Link: https://lkml.kernel.org/r/20220613192301.8817-7-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
This commit documents the 'lru_prio' scheme action for DAMON sysfs
interface.
Link: https://lkml.kernel.org/r/20220613192301.8817-6-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
This commit adds a new DAMOS action called 'LRU_PRIO' for the physical
address space. The action prioritizes pages in the memory regions of the
user-specified target access pattern on their LRU lists. This is hence
supposed to be used for frequently accessed (hot) memory regions so that
hot pages could be more likely protected under memory pressure.
Internally, it simply calls 'mark_page_accessed()'.
Link: https://lkml.kernel.org/r/20220613192301.8817-5-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
This commit moves code for 'DAMOS_PAGEOUT' handling of the physical
address space monitoring operations set to a separate function so that its
caller, 'damon_pa_apply_scheme()', can be more easily extended for
additional DAMOS actions later.
Link: https://lkml.kernel.org/r/20220613192301.8817-4-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "Extend DAMOS for Proactive LRU-lists Sorting".
Introduction
============
In short, this patchset 1) extends DAMON-based Operation Schemes (DAMOS)
for low overhead data access pattern based LRU-lists sorting, and 2)
implements a static kernel module for easy use of conservatively-tuned
version of that using the extended DAMOS capability.
Background
----------
As page-granularity access checking overhead could be significant on huge
systems, LRU lists are normally not proactively sorted but partially and
reactively sorted for special events including specific user requests,
system calls and memory pressure. As a result, LRU lists are sometimes
not so perfectly prepared to be used as a trustworthy access pattern
source for some situations including reclamation target pages selection
under sudden memory pressure.
DAMON-based Proactive LRU-lists Sorting
---------------------------------------
Because DAMON can identify access patterns of best-effort accuracy while
inducing only user-specified range of overhead, using DAMON for Proactive
LRU-lists Sorting (PLRUS) could be helpful for this situation. The idea
is quite simple. Find hot pages and cold pages using DAMON, and
prioritize hot pages while deprioritizing cold pages on their LRU-lists.
This patchset extends DAMON to support such schemes by introducing a
couple of new DAMOS actions for prioritizing and deprioritizing memory
regions of specific access patterns on their LRU-lists. In detail, this
patchset simply uses 'mark_page_accessed()' and 'deactivate_page()'
functions for prioritization and deprioritization of pages on their LRU
lists, respectively.
To make the scheme easy to use without complex tuning for common
situations, this patchset further implements a static kernel module called
'DAMON_LRU_SORT' using the extended DAMOS functionality. It proactively
sorts LRU-lists using DAMON with conservatively chosen default
hotness/coldness thresholds and small CPU usage quota limit. That is, the
module under its default parameters will make no harm for common situation
but provide some level of benefit for systems having clear hot/cold access
pattern under only memory pressure while consuming only limited small
portion of CPU time.
Related Works
-------------
Proactive reclamation is well known to be helpful for reducing non-optimal
reclamation target selection caused performance drops. However, proactive
reclamation is not a best option for some cases, because it could incur
additional I/O. For an example, it could be prohitive for systems using
storage devices that total number of writes is limited, or cloud block
storages that charges every I/O.
Some proactive reclamation approaches[1,2] induce a level of memory
pressure using memcg files or swappiness while monitoring PSI. As
reclamation target selection is still relying on the original LRU-lists
mechanism, using DAMON-based proactive reclamation before inducing the
proactive reclamation could allow more memory saving with same level of
performance overhead, or less performance overhead with same level of
memory saving.
[1] https://blogs.oracle.com/linux/post/anticipating-your-memory-needs
[2] https://www.pdl.cmu.edu/ftp/NVM/tmo_asplos22.pdf
Evaluation
==========
In short, PLRUS achieves 10% memory PSI (some) reduction, 14% major page
faults reduction, and 3.74% speedup under memory pressure.
Setup
-----
To show the effect of PLRUS, I run PARSEC3 and SPLASH-2X benchmarks under
below variant systems and measure a few metrics including the runtime of
each workload, number of system-wide major page faults, and system-wide
memory PSI (some).
- orig: v5.18-rc4 based mm-unstable kernel + this patchset, but no DAMON scheme
applied.
- mprs: Same to 'orig' but artificial memory pressure is induced.
- plrus: Same to 'mprs' but a radically tuned PLRUS scheme is applied to the
entire physical address space of the system.
For the artificial memory pressure, I set 'memory.limit_in_bytes' to 75%
of the running workload's peak RSS, wait 1 second, remove the pressure by
setting it to 200% of the peak RSS, wait 10 seconds, and repeat the
procedure until the workload finishes[1]. I use zram based swap device.
The tests are automated[2].
[1] https://github.com/awslabs/damon-tests/blob/next/perf/runners/back/0009_memcg_pressure.sh
[2] https://github.com/awslabs/damon-tests/blob/next/perf/full_once_config.sh
Radically Tuned PLRUS
---------------------
To show effect of PLRUS on the PARSEC3/SPLASH-2X workloads which runs for
no long time, we use radically tuned version of PLRUS. The version asks
DAMON to do the proactive LRU-lists sorting as below.
1. Find any memory regions shown some accesses (approximately >=20 accesses per
100 sampling) and prioritize pages of the regions on their LRU lists using
up to 2% CPU time. Under the CPU time limit, prioritize regions having
higher access frequency and kept the access frequency longer first.
2. Find any memory regions shown no access for at least >=5 seconds and
deprioritize pages of the rgions on their LRU lists using up to 2% CPU time.
Under the CPU time limit, deprioritize regions that not accessed for longer
time first.
Results
-------
I repeat the tests 25 times and calculate average of the measured numbers.
The results are as below:
metric orig mprs plrus plrus/mprs
runtime_seconds 190.06 292.83 281.87 0.96
pgmajfaults 852.55 8769420.00 7525040.00 0.86
memory_psi_some_us 106911.00 6943420.00 6220920.00 0.90
The first row is for legend. The first cell shows the metric that the
following cells of the row shows. Second, third, and fourth cells show
the metrics under the configs shown at the first row of the cell, and the
fifth cell shows the metric under 'plrus' divided by the metric under
'mprs'. Second row shows the averaged runtime of the workloads in
seconds. Third row shows the number of system-wide major page faults
while the test was ongoing. Fourth row shows the system-wide memory
pressure stall for some processes in microseconds while the test was
ongoing.
In short, PLRUS achieves 10% memory PSI (some) reduction, 14% major page
faults reduction, and 3.74% speedup under memory pressure. We also
confirmed the CPU usage of kdamond was 2.61% of single CPU, which is below
4% as expected.
Sequence of Patches
===================
The first and second patch cleans up DAMON debugfs interface and
DAMOS_PAGEOUT handling code of physical address space monitoring
operations implementation for easier extension of the code.
The thrid and fourth patches implement a new DAMOS action called
'lru_prio', which prioritizes pages under memory regions which have a
user-specified access pattern, and document it, respectively. The fifth
and sixth patches implement yet another new DAMOS action called
'lru_deprio', which deprioritizes pages under memory regions which have a
user-specified access pattern, and document it, respectively.
The seventh patch implements a static kernel module called
'damon_lru_sort', which utilizes the DAMON-based proactive LRU-lists
sorting under conservatively chosen default parameter. Finally, the
eighth patch documents 'damon_lru_sort'.
This patch (of 8):
DAMON debugfs interface assumes users will write 'damos_action' value
directly to the 'schemes' file. This makes adding new 'damos_action' in
the middle of its definition breaks the backward compatibility of DAMON
debugfs interface, as values of some 'damos_action' could be changed. To
mitigate the situation, this commit adds mappings between the user inputs
and 'damos_action' value and makes DAMON debugfs code uses those.
Link: https://lkml.kernel.org/r/20220613192301.8817-1-sj@kernel.org
Link: https://lkml.kernel.org/r/20220613192301.8817-2-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
swap_cache_info are not statistics that could be easily used to tune
system performance because they are not easily accessile. Also they can't
provide really useful info when OOM occurs. Remove these statistics can
also help mitigate unneeded global swap_cache_info cacheline contention.
Link: https://lkml.kernel.org/r/20220608144031.829-4-linmiaohe@huawei.com
Signed-off-by: Miaohe Lin <linmiaohe@huawei.com>
Suggested-by: David Hildenbrand <david@redhat.com>
Reviewed-by: David Hildenbrand <david@redhat.com>
Reviewed-by: Muchun Song <songmuchun@bytedance.com>
Acked-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
si->inuse_pages could still be accessed concurrently now. The plain reads
outside si->lock critical section, i.e. swap_show and si_swapinfo, which
results in data races. READ_ONCE and WRITE_ONCE is used to fix such data
races. Note these data races should be ok because they're just used for
showing swap info.
[linmiaohe@huawei.com: use WRITE_ONCE to pair with READ_ONCE]
Link: https://lkml.kernel.org/r/20220625093346.48894-2-linmiaohe@huawei.com
Link: https://lkml.kernel.org/r/20220608144031.829-3-linmiaohe@huawei.com
Signed-off-by: Miaohe Lin <linmiaohe@huawei.com>
Reviewed-by: David Hildenbrand <david@redhat.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
A get_random_bytes() function can cause a high contention if it is called
across CPUs simultaneously. Because it shares one lock per all CPUs:
<snip>
class name con-bounces contentions waittime-min waittime-max waittime-total waittime-avg acq-bounces acquisitions holdtime-min holdtime-max holdtime-total holdtime-avg
&crng->lock: 663145 665886 0.05 8.85 261966.66 0.39 7188152 13731279 0.04 11.89 2181582.30 0.16
-----------
&crng->lock 307835 [<00000000acba59cd>] _extract_crng+0x48/0x90
&crng->lock 358051 [<00000000f0075abc>] _crng_backtrack_protect+0x32/0x90
-----------
&crng->lock 234241 [<00000000f0075abc>] _crng_backtrack_protect+0x32/0x90
&crng->lock 431645 [<00000000acba59cd>] _extract_crng+0x48/0x90
<snip>
Switch from the get_random_bytes() to prandom_u32() that does not have any
internal contention when a random value is needed for the tests.
The reason is to minimize CPU cycles introduced by the test-suite itself
from the vmalloc performance metrics.
Link: https://lkml.kernel.org/r/20220607093449.3100-6-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Nicholas Piggin <npiggin@gmail.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sony.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
__find_vmap_area() finds a "vmap_area" based on passed address. It scan
the specific "vmap_area_root" rb-tree. Extend the function with one extra
argument, so any tree can be specified where the search has to be done.
There is no functional change as a result of this patch.
Link: https://lkml.kernel.org/r/20220607093449.3100-5-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Baoquan He <bhe@redhat.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Nicholas Piggin <npiggin@gmail.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sony.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
A vmap_area can travel between different places. For example
attached/detached to/from different rb-trees. In order to prevent fancy
bugs, initialize a VA's list node after it is removed from the list, so it
pairs with VA's rb_node which is also initialized.
There is no functional change as a result of this patch.
Link: https://lkml.kernel.org/r/20220607093449.3100-4-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Baoquan He <bhe@redhat.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Nicholas Piggin <npiggin@gmail.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sony.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
It implies that __alloc_vmap_area() allocates only from the global vmap
space, therefore a list-head and rb-tree, which represent a free vmap
space, are not passed as parameters to this function and are accessed
directly from this function.
Extend the __alloc_vmap_area() and other dependent functions to have a
possibility to allocate from different trees making an interface common
and not specific.
There is no functional change as a result of this patch.
Link: https://lkml.kernel.org/r/20220607093449.3100-3-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Baoquan He <bhe@redhat.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Nicholas Piggin <npiggin@gmail.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sony.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "Reduce a vmalloc internal lock contention preparation work".
This small serias is preparation work to implement per-cpu vmalloc
allocation in order to reduce a high internal lock contention. This
series does not introduce any functional changes, it is only about
preparation.
This patch (of 5):
Currently link_va() and unlik_va(), in order to figure out a tree type,
compares a passed root value with a global free_vmap_area_root variable to
distinguish the augmented rb-tree from a regular one. It is hard coded
since such functions can manipulate only with specific
"free_vmap_area_root" tree that represents a global free vmap space.
Make it common by introducing "_augment" versions of both internal
functions, so it is possible to deal with different trees.
There is no functional change as a result of this patch.
Link: https://lkml.kernel.org/r/20220607093449.3100-1-urezki@gmail.com
Link: https://lkml.kernel.org/r/20220607093449.3100-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Baoquan He <bhe@redhat.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Nicholas Piggin <npiggin@gmail.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sony.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Vincent Whitchurch
Baolin Wang
Miaohe Lin
Muchun Song
Matthew Wilcox (Oracle)
David Hildenbrand
Edward Liaw
SeongJae Park
Uladzislau Rezki (Sony)