On Mon, Oct 7, 2024 at 6:57 PM Jingxiang Zeng
<jingxiangzeng.cas@xxxxxxxxx> wrote:
>
> From: Jingxiang Zeng <linuszeng@xxxxxxxxxxx>
>
> Commit 14aa8b2d5c2e ("mm/mglru: don't sync disk for each aging cycle")
> removed the opportunity to wake up flushers during the MGLRU page
> reclamation process can lead to an increased likelihood of triggering OOM
> when encountering many dirty pages during reclamation on MGLRU.
>
> This leads to premature OOM if there are too many dirty pages in cgroup:
> Killed
>
> dd invoked oom-killer: gfp_mask=0x101cca(GFP_HIGHUSER_MOVABLE|__GFP_WRITE),
> order=0, oom_score_adj=0
>
> Call Trace:
> <TASK>
> dump_stack_lvl+0x5f/0x80
> dump_stack+0x14/0x20
> dump_header+0x46/0x1b0
> oom_kill_process+0x104/0x220
> out_of_memory+0x112/0x5a0
> mem_cgroup_out_of_memory+0x13b/0x150
> try_charge_memcg+0x44f/0x5c0
> charge_memcg+0x34/0x50
> __mem_cgroup_charge+0x31/0x90
> filemap_add_folio+0x4b/0xf0
> __filemap_get_folio+0x1a4/0x5b0
> ? srso_return_thunk+0x5/0x5f
> ? __block_commit_write+0x82/0xb0
> ext4_da_write_begin+0xe5/0x270
> generic_perform_write+0x134/0x2b0
> ext4_buffered_write_iter+0x57/0xd0
> ext4_file_write_iter+0x76/0x7d0
> ? selinux_file_permission+0x119/0x150
> ? srso_return_thunk+0x5/0x5f
> ? srso_return_thunk+0x5/0x5f
> vfs_write+0x30c/0x440
> ksys_write+0x65/0xe0
> __x64_sys_write+0x1e/0x30
> x64_sys_call+0x11c2/0x1d50
> do_syscall_64+0x47/0x110
> entry_SYSCALL_64_after_hwframe+0x76/0x7e
>
> memory: usage 308224kB, limit 308224kB, failcnt 2589
> swap: usage 0kB, limit 9007199254740988kB, failcnt 0
>
> ...
> file_dirty 303247360
> file_writeback 0
> ...
>
> oom-kill:constraint=CONSTRAINT_MEMCG,nodemask=(null),cpuset=test,
> mems_allowed=0,oom_memcg=/test,task_memcg=/test,task=dd,pid=4404,uid=0
> Memory cgroup out of memory: Killed process 4404 (dd) total-vm:10512kB,
> anon-rss:1152kB, file-rss:1824kB, shmem-rss:0kB, UID:0 pgtables:76kB
> oom_score_adj:0
>
> The flusher wake up was removed to decrease SSD wearing, but if we are
> seeing all dirty folios at the tail of an LRU, not waking up the flusher
> could lead to thrashing easily. So wake it up when a mem cgroups is about
> to OOM due to dirty caches.
>
> ---
> Changes from v3:
> - Avoid taking lock and reduce overhead on folio isolation by
> checking the right flags and rework wake up condition, fixing the
> performance regression reported by Chris Li.
> [Chris Li, Kairui Song]
> - Move the wake up check to try_to_shrink_lruvec to cover kswapd
> case as well, and update comments. [Kairui Song]
> - Link to v3: https://lore.kernel.org/all/20240924121358.30685-1-jingxiangzeng.cas@xxxxxxxxx/
> Changes from v2:
> - Acquire the lock before calling the folio_check_dirty_writeback
> function. [Wei Xu, Jingxiang Zeng]
> - Link to v2: https://lore.kernel.org/all/20240913084506.3606292-1-jingxiangzeng.cas@xxxxxxxxx/
> Changes from v1:
> - Add code to count the number of unqueued_dirty in the sort_folio
> function. [Wei Xu, Jingxiang Zeng]
> - Link to v1: https://lore.kernel.org/all/20240829102543.189453-1-jingxiangzeng.cas@xxxxxxxxx/
> ---
>
> Fixes: 14aa8b2d5c2e ("mm/mglru: don't sync disk for each aging cycle")
> Signed-off-by: Zeng Jingxiang <linuszeng@xxxxxxxxxxx>
> Signed-off-by: Kairui Song <kasong@xxxxxxxxxxx>
> Cc: T.J. Mercier <tjmercier@xxxxxxxxxx>
> Cc: Wei Xu <weixugc@xxxxxxxxxx>
> Cc: Yu Zhao <yuzhao@xxxxxxxxxx>
> ---
> mm/vmscan.c | 19 ++++++++++++++++---
> 1 file changed, 16 insertions(+), 3 deletions(-)
>
> diff --git a/mm/vmscan.c b/mm/vmscan.c
> index dc7a285b256b..2a5c2fe81467 100644
> --- a/mm/vmscan.c
> +++ b/mm/vmscan.c
> @@ -4291,6 +4291,7 @@ static bool sort_folio(struct lruvec *lruvec, struct folio *folio, struct scan_c
> int tier_idx)
> {
> bool success;
> + bool dirty, writeback;
> int gen = folio_lru_gen(folio);
> int type = folio_is_file_lru(folio);
> int zone = folio_zonenum(folio);
> @@ -4336,9 +4337,14 @@ static bool sort_folio(struct lruvec *lruvec, struct folio *folio, struct scan_c
> return true;
> }
>
> + dirty = folio_test_dirty(folio);
> + writeback = folio_test_writeback(folio);
> + if (type == LRU_GEN_FILE && dirty && !writeback)
> + sc->nr.unqueued_dirty += delta;
> +
This sounds good. BTW, when shrink_folio_list() in evict_folios()
returns, we should add stat.nr_unqueued_dirty to sc->nr.unqueued_dirty
there as well.
> /* waiting for writeback */
> - if (folio_test_locked(folio) || folio_test_writeback(folio) ||
> - (type == LRU_GEN_FILE && folio_test_dirty(folio))) {
> + if (folio_test_locked(folio) || writeback ||
> + (type == LRU_GEN_FILE && dirty)) {
> gen = folio_inc_gen(lruvec, folio, true);
> list_move(&folio->lru, &lrugen->folios[gen][type][zone]);
> return true;
> @@ -4454,7 +4460,7 @@ static int scan_folios(struct lruvec *lruvec, struct scan_control *sc,
> trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, MAX_LRU_BATCH,
> scanned, skipped, isolated,
> type ? LRU_INACTIVE_FILE : LRU_INACTIVE_ANON);
> -
> + sc->nr.taken += scanned;
I think we should only include file pages (in sort_folio) and isolated
pages into sc->nr.taken, instead of all scanned pages. For example, if
there are only unevictable and unqueued dirty pages, we would still
like to wake up the flusher threads, but because nr.taken counts
unevictable pages as well, the wakeup condition in
try_to_shrink_lruvec() won't be met.
> /*
> * There might not be eligible folios due to reclaim_idx. Check the
> * remaining to prevent livelock if it's not making progress.
> @@ -4796,6 +4802,13 @@ static bool try_to_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
> cond_resched();
> }
>
> + /*
> + * If too many file cache in the coldest generation can't be evicted
> + * due to being dirty, wake up the flusher.
> + */
> + if (sc->nr.unqueued_dirty && sc->nr.unqueued_dirty == sc->nr.taken)
> + wakeup_flusher_threads(WB_REASON_VMSCAN);
> +
try_to_shrink_lruvec() can be called from shrink_node() for global
reclaim as well. We need to reset sc->nr before calling
lru_gen_shrink_node() there. MGLRU didn't need that because it didn't
use sc->nr until this change.
> /* whether this lruvec should be rotated */
> return nr_to_scan < 0;
> }
> --
> 2.43.5
>
