On a system booted with cgroup.memory=nokmem, the deferred split shrinker is demoted to a non-memcg-aware one. As a result, partially unmapped THPs are only split under global reclaim; memcg (limit-induced) reclaim never splits them, so a cgroup under memory pressure keeps its underused THPs intact. The shrinker has carried SHRINKER_NONSLAB since commit 0a432dcbeb32 ("mm: shrinker: make shrinker not depend on memcg kmem") precisely so it would keep working with kmem accounting disabled. Commit fafaeceb89a5 ("mm: switch deferred split shrinker to list_lru") dropped it knowingly: under nokmem the list_lru collapses into a shared per-node list while __list_lru_add() still sets the shrinker bit on whichever memcg happens to add the first item, and with the flag set the shrinker would run twice per reclaim cycle [1]. NONSLAB-backed lrus keep their per-memcg lists under nokmem now, so the shrinker bit identifies the owning memcg again. Restore the flag. This was noticed by Michal during review of the patch "mm/list_lru: don't copy stale shrinker id from non-memcg-aware shrinkers" [2]. [1] https://lore.kernel.org/all/ah9PGv12mqai84ES@cmpxchg.org/ [2] https://lore.kernel.org/lkml/697713c4-0857-485b-aba7-c74f37a3c8b4@linux.alibaba.com/ Fixes: fafaeceb89a5 ("mm: switch deferred split shrinker to list_lru") Suggested-by: Michal Koutný Signed-off-by: Qinyun Tan Acked-by: Zi Yan Acked-by: David Hildenbrand (Arm) Acked-by: Usama Arif Reviewed-by: Baolin Wang Reviewed-by: Johannes Weiner --- mm/huge_memory.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/mm/huge_memory.c b/mm/huge_memory.c index c5d11147b69ae..87c9fc48b7cfd 100644 --- a/mm/huge_memory.c +++ b/mm/huge_memory.c @@ -1064,7 +1064,8 @@ int folio_memcg_alloc_deferred(struct folio *folio) static int __init thp_shrinker_init(void) { deferred_split_shrinker = shrinker_alloc(SHRINKER_NUMA_AWARE | - SHRINKER_MEMCG_AWARE, + SHRINKER_MEMCG_AWARE | + SHRINKER_NONSLAB, "thp-deferred_split"); if (!deferred_split_shrinker) return -ENOMEM; -- 2.43.7