- 04 Oct, 2023 40 commits
-
-
SeongJae Park authored
DAMON sleeps for sampling interval after each sampling, and check if the aggregation interval and the ops update interval have passed using ktime_get_coarse_ts64() and baseline timestamps for the intervals. That design is for making the operations occur at deterministic timing regardless of the time that spend for each work. However, it turned out it is not that useful, and incur not-that-intuitive results. After all, timer functions, and especially sleep functions that DAMON uses to wait for specific timing, are not necessarily strictly accurate. It is legal design, so no problem. However, depending on such inaccuracies, the nr_accesses can be larger than aggregation interval divided by sampling interval. For example, with the default setting (5 ms sampling interval and 100 ms aggregation interval) we frequently show regions having nr_accesses larger than 20. Also, if the execution of a DAMOS scheme takes a long time, next aggregation could happen before enough number of samples are collected. This is not what usual users would intuitively expect. Since access check sampling is the smallest unit work of DAMON, using the number of passed sampling intervals as the DAMON-internal timer can easily avoid these problems. That is, convert aggregation and ops update intervals to numbers of sampling intervals that need to be passed before those operations be executed, count the number of passed sampling intervals, and invoke the operations as soon as the specific amount of sampling intervals passed. Make the change. Note that this could make a behavioral change to settings that using intervals that not aligned by the sampling interval. For example, if the sampling interval is 5 ms and the aggregation interval is 12 ms, DAMON effectively uses 15 ms as its aggregation interval, because it checks whether the aggregation interval after sleeping the sampling interval. This change will make DAMON to effectively use 10 ms as aggregation interval, since it uses 'aggregation interval / sampling interval * sampling interval' as the effective aggregation interval, and we don't use floating point types. Usual users would have used aligned intervals, so this behavioral change is not expected to make any meaningful impact, so just make this change. Link: https://lkml.kernel.org/r/20230914021523.60649-1-sj@kernel.orgSigned-off-by:
SeongJae Park <sj@kernel.org> Signed-off-by:
Andrew Morton <akpm@linux-foundation.org>
-
Zi Yan authored
__flush_dcache_pages() is called during hugetlb migration via migrate_pages() -> migrate_hugetlbs() -> unmap_and_move_huge_page() -> move_to_new_folio() -> flush_dcache_folio(). And with hugetlb and without sparsemem vmemmap, struct page is not guaranteed to be contiguous beyond a section. Use nth_page() instead. Without the fix, a wrong address might be used for data cache page flush. No bug is reported. The fix comes from code inspection. Link: https://lkml.kernel.org/r/20230913201248.452081-6-zi.yan@sent.com Fixes: 15fa3e8e ("mips: implement the new page table range API") Signed-off-by:
Zi Yan <ziy@nvidia.com> Cc: David Hildenbrand <david@redhat.com> Cc: Matthew Wilcox (Oracle) <willy@infradead.org> Cc: Mike Kravetz <mike.kravetz@oracle.com> Cc: Mike Rapoport (IBM) <rppt@kernel.org> Cc: Muchun Song <songmuchun@bytedance.com> Cc: Thomas Bogendoerfer <tsbogend@alpha.franken.de> Cc: <stable@vger.kernel.org> Signed-off-by:
-
Zi Yan authored
When dealing with hugetlb pages, struct page is not guaranteed to be contiguous on SPARSEMEM without VMEMMAP. Use nth_page() to handle it properly. Without the fix, a wrong subpage might be checked for HWPoison, causing wrong number of bytes of a page copied to user space. No bug is reported. The fix comes from code inspection. Link: https://lkml.kernel.org/r/20230913201248.452081-5-zi.yan@sent.com Fixes: 38c1ddbd ("hugetlbfs: improve read HWPOISON hugepage") Signed-off-by:
Muchun Song <songmuchun@bytedance.com> Cc: David Hildenbrand <david@redhat.com> Cc: Matthew Wilcox (Oracle) <willy@infradead.org> Cc: Mike Kravetz <mike.kravetz@oracle.com> Cc: Mike Rapoport (IBM) <rppt@kernel.org> Cc: Thomas Bogendoerfer <tsbogend@alpha.franken.de> Cc: <stable@vger.kernel.org> Signed-off-by:
-
Zi Yan authored
When dealing with hugetlb pages, manipulating struct page pointers directly can get to wrong struct page, since struct page is not guaranteed to be contiguous on SPARSEMEM without VMEMMAP. Use pfn calculation to handle it properly. Without the fix, a wrong number of page might be skipped. Since skip cannot be negative, scan_movable_page() will end early and might miss a movable page with -ENOENT. This might fail offline_pages(). No bug is reported. The fix comes from code inspection. Link: https://lkml.kernel.org/r/20230913201248.452081-4-zi.yan@sent.com Fixes: eeb0efd0 ("mm,memory_hotplug: fix scan_movable_pages() for gigantic hugepages") Signed-off-by:
David Hildenbrand <david@redhat.com> Cc: Matthew Wilcox (Oracle) <willy@infradead.org> Cc: Mike Kravetz <mike.kravetz@oracle.com> Cc: Mike Rapoport (IBM) <rppt@kernel.org> Cc: Thomas Bogendoerfer <tsbogend@alpha.franken.de> Cc: <stable@vger.kernel.org> Signed-off-by:
-
Zi Yan authored
When dealing with hugetlb pages, manipulating struct page pointers directly can get to wrong struct page, since struct page is not guaranteed to be contiguous on SPARSEMEM without VMEMMAP. Use nth_page() to handle it properly. A wrong or non-existing page might be tried to be grabbed, either leading to a non freeable page or kernel memory access errors. No bug is reported. It comes from code inspection. Link: https://lkml.kernel.org/r/20230913201248.452081-3-zi.yan@sent.com Fixes: 57a196a5 ("hugetlb: simplify hugetlb handling in follow_page_mask") Signed-off-by:
-
Zi Yan authored
Patch series "Use nth_page() in place of direct struct page manipulation", v3. On SPARSEMEM without VMEMMAP, struct page is not guaranteed to be contiguous, since each memory section's memmap might be allocated independently. hugetlb pages can go beyond a memory section size, thus direct struct page manipulation on hugetlb pages/subpages might give wrong struct page. Kernel provides nth_page() to do the manipulation properly. Use that whenever code can see hugetlb pages. This patch (of 5): When dealing with hugetlb pages, manipulating struct page pointers directly can get to wrong struct page, since struct page is not guaranteed to be contiguous on SPARSEMEM without VMEMMAP. Use nth_page() to handle it properly. Without the fix, page_kasan_tag_reset() could reset wrong page tags, causing a wrong kasan result. No related bug is reported. The fix comes from code inspection. Link: https://lkml.kernel.org/r/20230913201248.452081-1-zi.yan@sent.com Link: https://lkml.kernel.org/r/20230913201248.452081-2-zi.yan@sent.com Fixes: 2813b9c0 ("kasan, mm, arm64: tag non slab memory allocated via pagealloc") Signed-off-by:
-
Vlastimil Babka authored
This isolate_mode_t flag is effectively unused since 89f6c88a ("mm: __isolate_lru_page_prepare() in isolate_migratepages_block()") as sc->may_unmap is now checked directly (and only node_reclaim has a mode that sets it to 0). The last remaining place is mm_vmscan_lru_isolate tracepoint for the isolate_mode parameter. That one was mainly used to indicate the active/inactive mode, which the trace-vmscan-postprocess.pl script consumed, but that got silently broken. After fixing the script by the previous patch, it does not need the isolate_mode anymore. So just remove the parameter and with that the whole ISOLATE_UNMAPPED flag. Link: https://lkml.kernel.org/r/20230914131637.12204-4-vbabka@suse.czSigned-off-by:
Vlastimil Babka <vbabka@suse.cz> Cc: Hugh Dickins <hughd@google.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Vlastimil Babka <vbabka@suse.cz> Signed-off-by:
-
Vlastimil Babka authored
The script has fallen behind tracepoint changes for a while, fix it up. Most changes are mechanical (renames, removal of tracepoint parameters that are not used by the script). More notable change involves mm_vmscan_lru_isolate which is relying on the isolate_mode to determine if the inactive list is being scanned. However the parameter currently only indicates ISOLATE_UNMAPPED. We can use the lru parameter instead to determine which list is scanned, and stop checking isolate_mode. Link: https://lkml.kernel.org/r/20230914131637.12204-3-vbabka@suse.czSigned-off-by:
-
Matthew Wilcox (Oracle) authored
Inline it into __bread_gfp(). Link: https://lkml.kernel.org/r/20230914150011.843330-9-willy@infradead.orgSigned-off-by:
Matthew Wilcox (Oracle) <willy@infradead.org> Cc: Hui Zhu <teawater@antgroup.com> Signed-off-by:
-
Matthew Wilcox (Oracle) authored
Most of the callers of sb_getblk_gfp() already assumed that they were passing the entire GFP flags to use. Fix up the two callers that didn't, and remove the __GFP_NOFAIL from them since they both appear to correctly handle failure. Link: https://lkml.kernel.org/r/20230914150011.843330-8-willy@infradead.orgSigned-off-by:
-
Matthew Wilcox (Oracle) authored
Now that __getblk() is in the right place in the file, it is trivial to call it from sb_getblk(). Link: https://lkml.kernel.org/r/20230914150011.843330-7-willy@infradead.orgSigned-off-by:
-
Matthew Wilcox (Oracle) authored
Move these two functions up in the file for the benefit of the next patch, and pass in all of the GFP flags to use instead of the partial GFP flags used by __getblk_gfp(). Link: https://lkml.kernel.org/r/20230914150011.843330-6-willy@infradead.orgSigned-off-by:
-
Matthew Wilcox (Oracle) authored
__getblk() adds __GFP_NOFAIL, which is unnecessary for readahead; we're quite comfortable with the possibility that we may not get a bh back. Switch to bdev_getblk() which does not include __GFP_NOFAIL. Link: https://lkml.kernel.org/r/20230914150011.843330-5-willy@infradead.orgSigned-off-by:
-
Matthew Wilcox (Oracle) authored
sb_getblk_gfp adds __GFP_NOFAIL, which is unnecessary for readahead; we're quite comfortable with the possibility that we may not get a bh back. Switch to bdev_getblk() which does not include __GFP_NOFAIL. Link: https://lkml.kernel.org/r/20230914150011.843330-4-willy@infradead.orgSigned-off-by:
Hui Zhu <teawater@antgroup.com> Closes: https://lore.kernel.org/linux-fsdevel/20230811035705.3296-1-teawaterz@linux.alibaba.com/Signed-off-by:
-
Matthew Wilcox (Oracle) authored
grow_dev_page() is only called by grow_buffers(). grow_buffers() is only called by __getblk_slow() and __getblk_slow() is only called from __getblk_gfp(), so it is safe to move the GFP flags setting all the way up. With that done, add a new bdev_getblk() entry point that leaves the GFP flags the way the caller specified them. [willy@infradead.org: fix grow_dev_page() error handling] Link: https://lkml.kernel.org/r/ZRREEIwqiy5DijKB@casper.infradead.org Link: https://lkml.kernel.org/r/20230914150011.843330-3-willy@infradead.orgSigned-off-by:
-
Matthew Wilcox (Oracle) authored
Patch series "Add and use bdev_getblk()", v2. This patch series fixes a bug reported by Hui Zhu; see proposed patches v1 and v2: https://lore.kernel.org/linux-fsdevel/20230811035705.3296-1-teawaterz@linux.alibaba.com/ https://lore.kernel.org/linux-fsdevel/20230811071519.1094-1-teawaterz@linux.alibaba.com/ I decided to go in a rather different direction for this fix, and fix a related problem at the same time. I don't think there's any urgency to rush this into Linus' tree, nor have I marked it for stable. Reasonable people may disagree. This patch (of 8): Instead of creating entirely new flags, inherit them from grow_dev_page(). The other callers create the same flags that this function used to create. Link: https://lkml.kernel.org/r/20230914150011.843330-1-willy@infradead.org Link: https://lkml.kernel.org/r/20230914150011.843330-2-willy@infradead.orgSigned-off-by:
-
SeongJae Park authored
Document damos_before_apply tracepoint on the usage document. Link: https://lkml.kernel.org/r/20230913022050.2109-3-sj@kernel.orgSigned-off-by:
-
SeongJae Park authored
Patch series "mm/damon: add a tracepoint for damos apply target regions", v2. DAMON provides damon_aggregated tracepoint to let users record full monitoring results. Sometimes, users need to record monitoring results of specific pattern. DAMOS tried regions directory of DAMON sysfs interface allows it, but the interface is mainly designed for snapshots and therefore would be inefficient for such recording. Implement yet another tracepoint for efficient support of the usecase. This patch (of 2): DAMON provides damon_aggregated tracepoint, which exposes details of each region and its access monitoring results. It is useful for getting whole monitoring results, e.g., for recording purposes. For investigations of DAMOS, DAMON Sysfs interface provides DAMOS statistics and tried_regions directory. But, those provides only statistics and snapshots. If the scheme is frequently applied and if the user needs to know every detail of DAMOS behavior, the snapshot-based interface could be insufficient and expensive. As a last resort, userspace users need to record the all monitoring results via damon_aggregated tracepoint and simulate how DAMOS would worked. It is unnecessarily complicated. DAMON kernel API users, meanwhile, can do that easily via before_damos_apply() callback field of 'struct damon_callback', though. Add a tracepoint that will be called just after before_damos_apply() callback for more convenient investigations of DAMOS. The tracepoint exposes all details about each regions, similar to damon_aggregated tracepoint. Please note that DAMOS is currently not only for memory management but also for query-like efficient monitoring results retrievals (when 'stat' action is used). Until now, only statistics or snapshots were supported. Addition of this tracepoint allows efficient full recording of DAMOS-based filtered monitoring results. Link: https://lkml.kernel.org/r/20230913022050.2109-1-sj@kernel.org Link: https://lkml.kernel.org/r/20230913022050.2109-2-sj@kernel.orgSigned-off-by:
-
Kefeng Wang authored
Directly check the return of isolate_hugetlb() and folio_isolate_lru() to remove isolated variable, also setup err = -EBUSY in advance before isolation, and update err only when successfully queued for migration, which could help us to unify and simplify code a bit. Link: https://lkml.kernel.org/r/20230913095131.2426871-9-wangkefeng.wang@huawei.comSigned-off-by:
Kefeng Wang <wangkefeng.wang@huawei.com> Reviewed-by:
-
Kefeng Wang authored
There is some different between hugeTLB and THP behave when passed the address of a tail page, for THP, it will migrate the entire THP page, but for HugeTLB, it will return -EACCES, or -ENOENT before commit e66f17ff ("mm/hugetlb: take page table lock in follow_huge_pmd()"), -EACCES The page is mapped by multiple processes and can be moved only if MPOL_MF_MOVE_ALL is specified. -ENOENT The page is not present. But when check manual[1], both of the two errnos are not suitable, it is better to keep the same behave between hugetlb and THP when passed the address of a tail page, so let's just remove the PageHead() check for HugeTLB. [1] https://man7.org/linux/man-pages/man2/move_pages.2.html Link: https://lkml.kernel.org/r/20230913095131.2426871-8-wangkefeng.wang@huawei.comSigned-off-by:
Mike Kravetz <mike.kravetz@oracle.com> Acked-by:
-
Kefeng Wang authored
Use a folio in add_page_for_migration() to save compound_head() calls. Link: https://lkml.kernel.org/r/20230913095131.2426871-7-wangkefeng.wang@huawei.comSigned-off-by:
-
Kefeng Wang authored
Use __folio_test_movable(), no need to convert from folio to page again. Link: https://lkml.kernel.org/r/20230913095131.2426871-6-wangkefeng.wang@huawei.comSigned-off-by:
-
Kefeng Wang authored
At present, numa balance only support base page and PMD-mapped THP, but we will expand to support to migrate large folio/pte-mapped THP in the future, it is better to make migrate_misplaced_page() to take a folio instead of a page, and rename it to migrate_misplaced_folio(), it is a preparation, also this remove several compound_head() calls. Link: https://lkml.kernel.org/r/20230913095131.2426871-5-wangkefeng.wang@huawei.comSigned-off-by:
-
Kefeng Wang authored
Rename numamigrate_isolate_page() to numamigrate_isolate_folio(), then make it takes a folio and use folio API to save compound_head() calls. Link: https://lkml.kernel.org/r/20230913095131.2426871-4-wangkefeng.wang@huawei.comSigned-off-by:
-
Kefeng Wang authored
The check of THP mapped by multiple processes was introduced by commit 04fa5d6a ("mm: migrate: check page_count of THP before migrating") and refactor by commit 340ef390 ("mm: numa: cleanup flow of transhuge page migration"), which is out of date, since migrate_misplaced_page() is now using the standard migrate_pages() for small pages and THPs, the reference count checking is in folio_migrate_mapping(), so let's remove the special check for THP. Link: https://lkml.kernel.org/r/20230913095131.2426871-3-wangkefeng.wang@huawei.comSigned-off-by:
"Huang, Ying" <ying.huang@intel.com> Cc: David Hildenbrand <david@redhat.com> Cc: Hugh Dickins <hughd@google.com> Cc: Mike Kravetz <mike.kravetz@oracle.com> Cc: Zi Yan <ziy@nvidia.com> Signed-off-by:
-
Kefeng Wang authored
Patch series "mm: migrate: more folio conversion and unification", v3. Convert more migrate functions to use a folio, it is also a preparation for large folio migration support when balancing numa. This patch (of 8): The assert VM_BUG_ON_PAGE(order && !PageTransHuge(page), page) is not very useful, 1) for a tail/base page, order = 0, for a head page, the order > 0 && PageTransHuge() is true 2) there is a PageCompound() check and only base page is handled in do_numa_page(), and do_huge_pmd_numa_page() only handle PMD-mapped THP 3) even though the page is a tail page, isolate_lru_page() will post a warning, and fail to isolate the page 4) if large folio/pte-mapped THP migration supported in the future, we could migrate the entire folio if numa fault on a tail page so just remove the check. Link: https://lkml.kernel.org/r/20230913095131.2426871-1-wangkefeng.wang@huawei.com Link: https://lkml.kernel.org/r/20230913095131.2426871-2-wangkefeng.wang@huawei.comSigned-off-by: -
David Hildenbrand authored
Let's pass a folio; we are always mapping the entire thing. Link: https://lkml.kernel.org/r/20230913125113.313322-7-david@redhat.comSigned-off-by:
-
David Hildenbrand authored
Let's sanity-check PageAnonExclusive vs. mapcount in page_add_anon_rmap() and hugepage_add_anon_rmap() after setting PageAnonExclusive simply by re-reading the mapcounts. We can stop initializing the "first" variable in page_add_anon_rmap() and no longer need an atomic_inc_and_test() in hugepage_add_anon_rmap(). While at it, switch to VM_WARN_ON_FOLIO(). [david@redhat.com: update check for doubly-mapped page] Link: https://lkml.kernel.org/r/d8e5a093-2e22-c14b-7e64-6da280398d9f@redhat.com Link: https://lkml.kernel.org/r/20230913125113.313322-6-david@redhat.comSigned-off-by:
-
David Hildenbrand authored
If swapin code would ever decide to not use order-0 pages and supply a PTE-mapped large folio, we will have to change how we call __folio_set_anon() -- eventually with exclusive=false and an adjusted address. For now, let's add a VM_WARN_ON_FOLIO() with a comment about the situation. Link: https://lkml.kernel.org/r/20230913125113.313322-5-david@redhat.comSigned-off-by:
-
David Hildenbrand authored
Let's handle it in the caller; no need for the "first" check based on the mapcount. We really only end up with !anon pages in page_add_anon_rmap() via do_swap_page(), where we hold the folio lock. So races are not possible. Add a VM_WARN_ON_FOLIO() to make sure that we really hold the folio lock. In the future, we might want to let do_swap_page() use folio_add_new_anon_rmap() on new pages instead: however, we might have to pass then whether the folio is exclusive or not. So keep it in there for now. For hugetlb we never expect to have a non-anon page in hugepage_add_anon_rmap(). Remove that code, along with some other checks that are either not required or were checked in hugepage_add_new_anon_rmap() already. Link: https://lkml.kernel.org/r/20230913125113.313322-4-david@redhat.comSigned-off-by:
-
David Hildenbrand authored
Let's handle it in the caller. No need to pass the page. While at it, rename the function to __folio_set_anon() and pass "bool exclusive" instead of "int exclusive". Link: https://lkml.kernel.org/r/20230913125113.313322-3-david@redhat.comSigned-off-by:
-
David Hildenbrand authored
Patch series "Anon rmap cleanups". Some cleanups around rmap for anon pages. I'm working on more cleanups also around file rmap -- also to handle the "compound" parameter internally only and to let hugetlb use page_add_file_rmap(), but these changes make sense separately. This patch (of 6): That comment was added in commit 5dbe0af4 ("mm: fix kernel BUG at mm/rmap.c:1017!") to document why we can see vma->vm_end getting adjusted concurrently due to a VMA split. However, the optimized locking code was changed again in bf181b9f ("mm anon rmap: replace same_anon_vma linked list with an interval tree."). ... and later, the comment was changed in commit 0503ea8f ("mm/mmap: remove __vma_adjust()") to talk about "vma_merge" although the original issue was with VMA splitting. Let's just remove that comment. Nowadays, it's outdated, imprecise and confusing. Link: https://lkml.kernel.org/r/20230913125113.313322-1-david@redhat.com Link: https://lkml.kernel.org/r/20230913125113.313322-2-david@redhat.comSigned-off-by:
-
Xin Hao authored
At present, we support per-memcg reclaim strategy, however we do not know the number of transparent huge pages being reclaimed, as we know the transparent huge pages need to be splited before reclaim them, and they will bring some performance bottleneck effect. for example, when two memcg (A & B) are doing reclaim for anonymous pages at same time, and 'A' memcg is reclaiming a large number of transparent huge pages, we can better analyze that the performance bottleneck will be caused by 'A' memcg. therefore, in order to better analyze such problems, there add THP swap out info for per-memcg. [akpm@linux-foundation.orgL fix swap_writepage_fs(), per Johannes] Link: https://lkml.kernel.org/r/20230913213343.GB48476@cmpxchg.org Link: https://lkml.kernel.org/r/20230913164938.16918-1-vernhao@tencent.comSigned-off-by:
Xin Hao <vernhao@tencent.com> Suggested-by:
Johannes Weiner <hannes@cmpxchg.org> Acked-by:
-
liujinlong authored
When looking at the code in the memory part, I found that the purpose of the function prepare_scan_countis very different from the function name. It is easy to misunderstand when reading.The function prepare_scan_count mainly completes the assignment of the scan_control structure.Therefore, I suggest that the function name can be changed to prepare_scan_control, which is easier to understand. Link: https://lkml.kernel.org/r/20230912085923.27238-1-liujinlong@kylinos.cnSigned-off-by:
liujinlong <liujinlong@kylinos.cn> Signed-off-by:
-
Qi Zheng authored
Now there are no readers of shrinker_rwsem, so we can simply replace it with mutex lock. [akpm@linux-foundation.org: update the fix to alloc_shrinker_info()] Link: https://lkml.kernel.org/r/20230911094444.68966-46-zhengqi.arch@bytedance.comSigned-off-by:
Qi Zheng <zhengqi.arch@bytedance.com> Reviewed-by:
-
Qi Zheng authored
For now, reparent_shrinker_deferred() is the only holder of read lock of shrinker_rwsem. And it already holds the global cgroup_mutex, so it will not be called in parallel. Therefore, in order to convert shrinker_rwsem to shrinker_mutex later, here we change to hold the write lock of shrinker_rwsem to reparent. Link: https://lkml.kernel.org/r/20230911094444.68966-45-zhengqi.arch@bytedance.comSigned-off-by:
-
Qi Zheng authored
Like global slab shrink, this commit also uses refcount+RCU method to make memcg slab shrink lockless. Use the following script to do slab shrink stress test: ``` DIR="/root/shrinker/memcg/mnt" do_create() { mkdir -p /sys/fs/cgroup/memory/test echo 4G > /sys/fs/cgroup/memory/test/memory.limit_in_bytes for i in `seq 0 $1`; do mkdir -p /sys/fs/cgroup/memory/test/$i; echo $$ > /sys/fs/cgroup/memory/test/$i/cgroup.procs; mkdir -p $DIR/$i; done } do_mount() { for i in `seq $1 $2`; do mount -t tmpfs $i $DIR/$i; done } do_touch() { for i in `seq $1 $2`; do echo $$ > /sys/fs/cgroup/memory/test/$i/cgroup.procs; dd if=/dev/zero of=$DIR/$i/file$i bs=1M count=1 & done } case "$1" in touch) do_touch $2 $3 ;; test) do_create 4000 do_mount 0 4000 do_touch 0 3000 ;; *) exit 1 ;; esac ``` Save the above script, then run test and touch commands. Then we can use the following perf command to view hotspots: perf top -U -F 999 1) Before applying this patchset: 33.15% [kernel] [k] down_read_trylock 25.38% [kernel] [k] shrink_slab 21.75% [kernel] [k] up_read 4.45% [kernel] [k] _find_next_bit 2.27% [kernel] [k] do_shrink_slab 1.80% [kernel] [k] intel_idle_irq 1.79% [kernel] [k] shrink_lruvec 0.67% [kernel] [k] xas_descend 0.41% [kernel] [k] mem_cgroup_iter 0.40% [kernel] [k] shrink_node 0.38% [kernel] [k] list_lru_count_one 2) After applying this patchset: 64.56% [kernel] [k] shrink_slab 12.18% [kernel] [k] do_shrink_slab 3.30% [kernel] [k] __rcu_read_unlock 2.61% [kernel] [k] shrink_lruvec 2.49% [kernel] [k] __rcu_read_lock 1.93% [kernel] [k] intel_idle_irq 0.89% [kernel] [k] shrink_node 0.81% [kernel] [k] mem_cgroup_iter 0.77% [kernel] [k] mem_cgroup_calculate_protection 0.66% [kernel] [k] list_lru_count_one We can see that the first perf hotspot becomes shrink_slab, which is what we expect. Link: https://lkml.kernel.org/r/20230911094444.68966-44-zhengqi.arch@bytedance.comSigned-off-by: -
Qi Zheng authored
The shrinker_rwsem is a global read-write lock in shrinkers subsystem, which protects most operations such as slab shrink, registration and unregistration of shrinkers, etc. This can easily cause problems in the following cases. 1) When the memory pressure is high and there are many filesystems mounted or unmounted at the same time, slab shrink will be affected (down_read_trylock() failed). Such as the real workload mentioned by Kirill Tkhai: ``` One of the real workloads from my experience is start of an overcommitted node containing many starting containers after node crash (or many resuming containers after reboot for kernel update). In these cases memory pressure is huge, and the node goes round in long reclaim. ``` 2) If a shrinker is blocked (such as the case mentioned in [1]) and a writer comes in (such as mount a fs), then this writer will be blocked and cause all subsequent shrinker-related operations to be blocked. Even if there is no competitor when shrinking slab, there may still be a problem. The down_read_trylock() may become a perf hotspot with frequent calls to shrink_slab(). Because of the poor multicore scalability of atomic operations, this can lead to a significant drop in IPC (instructions per cycle). We used to implement the lockless slab shrink with SRCU [2], but then kernel test robot reported -88.8% regression in stress-ng.ramfs.ops_per_sec test case [3], so we reverted it [4]. This commit uses the refcount+RCU method [5] proposed by Dave Chinner to re-implement the lockless global slab shrink. The memcg slab shrink is handled in the subsequent patch. For now, all shrinker instances are converted to dynamically allocated and will be freed by call_rcu(). So we can use rcu_read_{lock,unlock}() to ensure that the shrinker instance is valid. And the shrinker instance will not be run again after unregistration. So the structure that records the pointer of shrinker instance can be safely freed without waiting for the RCU read-side critical section. In this way, while we implement the lockless slab shrink, we don't need to be blocked in unregister_shrinker(). The following are the test results: stress-ng --timeout 60 --times --verify --metrics-brief --ramfs 9 & 1) Before applying this patchset: setting to a 60 second run per stressor dispatching hogs: 9 ramfs stressor bogo ops real time usr time sys time bogo ops/s bogo ops/s (secs) (secs) (secs) (real time) (usr+sys time) ramfs 473062 60.00 8.00 279.13 7884.12 1647.59 for a 60.01s run time: 1440.34s available CPU time 7.99s user time ( 0.55%) 279.13s system time ( 19.38%) 287.12s total time ( 19.93%) load average: 7.12 2.99 1.15 successful run completed in 60.01s (1 min, 0.01 secs) 2) After applying this patchset: setting to a 60 second run per stressor dispatching hogs: 9 ramfs stressor bogo ops real time usr time sys time bogo ops/s bogo ops/s (secs) (secs) (secs) (real time) (usr+sys time) ramfs 477165 60.00 8.13 281.34 7952.55 1648.40 for a 60.01s run time: 1440.33s available CPU time 8.12s user time ( 0.56%) 281.34s system time ( 19.53%) 289.46s total time ( 20.10%) load average: 6.98 3.03 1.19 successful run completed in 60.01s (1 min, 0.01 secs) We can see that the ops/s has hardly changed. [1]. https://lore.kernel.org/lkml/20191129214541.3110-1-ptikhomirov@virtuozzo.com/ [2]. https://lore.kernel.org/lkml/20230313112819.38938-1-zhengqi.arch@bytedance.com/ [3]. https://lore.kernel.org/lkml/202305230837.db2c233f-yujie.liu@intel.com/ [4]. https://lore.kernel.org/all/20230609081518.3039120-1-qi.zheng@linux.dev/ [5]. https://lore.kernel.org/lkml/ZIJhou1d55d4H1s0@dread.disaster.area/ Link: https://lkml.kernel.org/r/20230911094444.68966-43-zhengqi.arch@bytedance.comSigned-off-by: -
Qi Zheng authored
With the new shrinker APIs, there is no action such as prealloc, so rename {prealloc|unregister}_memcg_shrinker() to shrinker_memcg_{alloc|remove}(), which corresponds to the idr_{alloc|remove}() inside the function. Link: https://lkml.kernel.org/r/20230911094444.68966-42-zhengqi.arch@bytedance.comSigned-off-by: -
Qi Zheng authored
Currently, we maintain two linear arrays per node per memcg, which are shrinker_info::map and shrinker_info::nr_deferred. And we need to resize them when the shrinker_nr_max is exceeded, that is, allocate a new array, and then copy the old array to the new array, and finally free the old array by RCU. For shrinker_info::map, we do set_bit() under the RCU lock, so we may set the value into the old map which is about to be freed. This may cause the value set to be lost. The current solution is not to copy the old map when resizing, but to set all the corresponding bits in the new map to 1. This solves the data loss problem, but bring the overhead of more pointless loops while doing memcg slab shrink. For shrinker_info::nr_deferred, we will only modify it under the read lock of shrinker_rwsem, so it will not run concurrently with the resizing. But after we make memcg slab shrink lockless, there will be the same data loss problem as shrinker_info::map, and we can't work around it like the map. For such resizable arrays, the most straightforward idea is to change it to xarray, like we did for list_lru [1]. We need to do xa_store() in the list_lru_add()-->set_shrinker_bit(), but this will cause memory allocation, and the list_lru_add() doesn't accept failure. A possible solution is to pre-allocate, but the location of pre-allocation is not well determined (such as deferred_split_shrinker case). Therefore, this commit chooses to introduce the following secondary array for shrinker_info::{map, nr_deferred}: +---------------+--------+--------+-----+ | shrinker_info | unit 0 | unit 1 | ... | (secondary array) +---------------+--------+--------+-----+ | v +---------------+-----+ | nr_deferred[] | map | (leaf array) +---------------+-----+ (shrinker_info_unit) The leaf array is never freed unless the memcg is destroyed. The secondary array will be resized every time the shrinker id exceeds shrinker_nr_max. So the shrinker_info_unit can be indexed from both the old and the new shrinker_info->unit[x]. Then even if we get the old secondary array under the RCU lock, the found map and nr_deferred are also true, so the updated nr_deferred and map will not be lost. [1]. https://lore.kernel.org/all/20220228122126.37293-13-songmuchun@bytedance.com/ [zhengqi.arch@bytedance.com: unlock the &shrinker_rwsem before the call to free_shrinker_info()] Link: https://lkml.kernel.org/r/20230928141517.12164-1-zhengqi.arch@bytedance.com Link: https://lkml.kernel.org/r/20230911094444.68966-41-zhengqi.arch@bytedance.comSigned-off-by:
-
