[PATCH] mm: switch deferred split shrinker to list_lru fix

[Johannes Weiner]

Lance Yang points out a weirdness in the list_lru code with
cgroup.memory=nokmem: in this mode, list_lru collapses to a shared
per-node list that holds the folios, but __list_lru_add() still sets
the shrinker bit on the owning memcg.

Usually this is fine, because the generic shrinker code ignores these
random bits when !memcg_kmem_online(). But the THP shrinker still has
the SHRINKER_NONSLAB flag set, which specifically declares an
independence from kmem. As a result, the shrinker fires twice per
reclaim cycle: one during the regular root cgroup scan, and then one
more time triggered from whichever memcg got the shrinker bit.

Drop the flag, since it's no longer true. The deferred_split shrinker
then behaves like every other list_lru-backed shrinker under nokmem,
including the non-kmem ones (zswap, workingset shadow_nodes): skipped
from memcg-internal reclaim, driven by global reclaim only.

This needs proper cleaning up on the shrinker and list_lru side, but
that's scope for a follow-up series. Just make it consistent now.

Reported-by: Lance Yang <lance.yang@xxxxxxxxx>
Signed-off-by: Johannes Weiner <hannes@xxxxxxxxxxx>
---
 mm/huge_memory.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/mm/huge_memory.c b/mm/huge_memory.c
index 72f6caf0fec6..aef495891f8c 100644
--- a/mm/huge_memory.c
+++ b/mm/huge_memory.c
@@ -956,8 +956,7 @@ 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_NONSLAB,
+						 SHRINKER_MEMCG_AWARE,
 						 "thp-deferred_split");
 	if (!deferred_split_shrinker)
 		return -ENOMEM;
-- 
2.54.0