* [PATCH 0/1] mm/shmem: don't use the memcg of an uncharged folio
@ 2026-11-19 19:36 Chris Mason
2026-11-19 19:36 ` [PATCH 1/1] " Chris Mason
0 siblings, 1 reply; 3+ messages in thread
From: Chris Mason @ 2026-11-19 19:36 UTC (permalink / raw)
To: hughd, david, baolin.wang, akpm, qi.zheng, qinyuntan,
shakeel.butt, linux-mm, linux-kernel
Hi everyone,
This is relatively niche, but actually new with 7.4-rc1, and I would expect it
to trigger under heavy load with huge=always on tmpfs. We'd need to run out of
room in the tmpfs or reclaim memory while huge folios are migrated.
Qinyun Tan's pending series fixes the oops:
[PATCH v4 RESEND 2/5] mm: memcontrol: drop kmemcg_id and use
mem_cgroup_id() for list_lru indexing
https://lore.kernel.org/all/20261105092233.3558950-3-qinyuntan@linux.alibaba.com/
It makes memcg_list_lru_allocated() accept a NULL memcg, but I think we
can still race and corrupt lists instead:
shmem_unused_huge_requeue() __migrate_folio()
=========================== =================
folio_memcg_list_lru_alloc()
folio_memcg() is NULL
no lists are allocated
mem_cgroup_migrate()
commit_charge(newfolio)
get_obj_cgroup_from_folio()
returns the objcg of memcg M
list_lru_add(M)
if M has no lists, the inode
goes on the parent's list
[ later, M gets lists of its own ]
list_lru_del(M)
locks M's list
unlinks from the parent's list
I haven't tried to reproduce on top of the kmemcg_id series, so that's
all just guessing from the code.
This is on top of v7.4-rc3.
Chris Mason (1):
mm/shmem: don't use the memcg of an uncharged folio
mm/shmem.c | 29 ++++++++++++++++++++++-------
1 file changed, 22 insertions(+), 7 deletions(-)
--
2.55.0
^ permalink raw reply [flat|nested] 3+ messages in thread* [PATCH 1/1] mm/shmem: don't use the memcg of an uncharged folio 2026-11-19 19:36 [PATCH 0/1] mm/shmem: don't use the memcg of an uncharged folio Chris Mason @ 2026-11-19 19:36 ` Chris Mason 2026-11-20 7:10 ` Qi Zheng 0 siblings, 1 reply; 3+ messages in thread From: Chris Mason @ 2026-11-19 19:36 UTC (permalink / raw) To: hughd, david, baolin.wang, akpm, qi.zheng, qinyuntan, shakeel.butt, linux-mm, linux-kernel The shrinker for unused huge tmpfs folios can dereference a NULL memcg if it races with the migration of the folio at the end of a file. folio_memcg_list_lru_alloc() assumes that folio_memcg() is never NULL. But it is NULL until the folio is charged, and a migration target can be in the page cache before charging. shmem_unused_huge_shrink() looks the folio up without the folio lock, so we can race with migration: __migrate_folio() shmem_unused_huge_shrink() ================= ========================== __folio_migrate_mapping() xas_store(newfolio) filemap_get_entry() returns newfolio folio_trylock() fails shmem_unused_huge_requeue() folio_memcg_list_lru_alloc() folio_memcg() is NULL memcg->kmemcg_id folio_migrate_flags() mem_cgroup_migrate() commit_charge(newfolio) Fix it by keeping the original objcg around while the inode is isolated. Fixes: 743e7d025d2f ("mm: shmem: make unused huge shrinker memcg aware") Signed-off-by: Chris Mason <mason@kernel.org> --- mm/shmem.c | 29 ++++++++++++++++++++++------- 1 file changed, 22 insertions(+), 7 deletions(-) diff --git a/mm/shmem.c b/mm/shmem.c index 10cef89b00b0..4465a43d6f74 100644 --- a/mm/shmem.c +++ b/mm/shmem.c @@ -758,6 +758,13 @@ shmem_unused_huge_alloc_lru(struct shmem_sb_info *sbinfo, struct folio *folio, { int ret; + /* + * A migration target is in the page cache before it is charged. Once + * charged, its objcg can't change while we hold a folio reference. + */ + if (!mem_cgroup_disabled() && !folio_memcg_charged(folio)) + return ERR_PTR(-EAGAIN); + ret = folio_memcg_list_lru_alloc(folio, &sbinfo->shrinklist, gfp); if (ret) return ERR_PTR(ret); @@ -887,12 +894,11 @@ static enum lru_status shmem_unused_huge_isolate(struct list_head *item, return LRU_REMOVED; } + /* keep shrinklist_objcg, requeue may need it */ list_lru_isolate(lru, item); - objcg = shmem_get_and_clear_objcg(info); set_shmem_unused_huge_isolated(info); list_add_tail(&info->shrinklist, &scan->list); spin_unlock(&info->lock); - obj_cgroup_put(objcg); return LRU_REMOVED; } @@ -928,32 +934,41 @@ static bool is_shmem_unused_huge_match(struct folio *folio, static void shmem_unused_huge_drop(struct inode *inode) { struct shmem_inode_info *info = SHMEM_I(inode); + struct obj_cgroup *objcg; spin_lock(&info->lock); list_del_init(&info->shrinklist); + objcg = shmem_get_and_clear_objcg(info); spin_unlock(&info->lock); + + obj_cgroup_put(objcg); } static void shmem_unused_huge_requeue(struct inode *inode, struct folio *folio) { struct shmem_inode_info *info = SHMEM_I(inode); struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb); - struct obj_cgroup *objcg; + struct obj_cgroup *objcg, *old_objcg; int nid = folio_nid(folio); objcg = shmem_unused_huge_alloc_lru(sbinfo, folio, GFP_NOWAIT); - if (IS_ERR(objcg)) { - shmem_unused_huge_drop(inode); - return; - } spin_lock(&info->lock); + old_objcg = shmem_get_and_clear_objcg(info); + /* The lists for old_objcg exist, the inode was queued under it */ + if (IS_ERR(objcg)) { + objcg = old_objcg; + old_objcg = NULL; + } + /* Requeue the inode to shrinklist */ list_del_init(&info->shrinklist); shmem_unused_huge_lru_add(sbinfo, &info->shrinklist, nid, objcg); info->shrinklist_objcg = objcg; info->shrinklist_nid = nid; spin_unlock(&info->lock); + + obj_cgroup_put(old_objcg); } static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo, -- 2.55.0 ^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: [PATCH 1/1] mm/shmem: don't use the memcg of an uncharged folio 2026-11-19 19:36 ` [PATCH 1/1] " Chris Mason @ 2026-11-20 7:10 ` Qi Zheng 0 siblings, 0 replies; 3+ messages in thread From: Qi Zheng @ 2026-11-20 7:10 UTC (permalink / raw) To: Chris Mason, hughd, david, baolin.wang, akpm, qinyuntan, shakeel.butt, linux-mm, linux-kernel On 11/20/26 3:36 AM, Chris Mason wrote: > The shrinker for unused huge tmpfs folios can dereference a NULL memcg > if it races with the migration of the folio at the end of a file. > > folio_memcg_list_lru_alloc() assumes that folio_memcg() is never NULL. > But it is NULL until the folio is charged, and a migration target can be > in the page cache before charging. > > shmem_unused_huge_shrink() looks the folio up without the folio lock, > so we can race with migration: > > __migrate_folio() shmem_unused_huge_shrink() > ================= ========================== > __folio_migrate_mapping() > xas_store(newfolio) > filemap_get_entry() > returns newfolio > folio_trylock() fails > shmem_unused_huge_requeue() > folio_memcg_list_lru_alloc() > folio_memcg() is NULL > memcg->kmemcg_id > folio_migrate_flags() > mem_cgroup_migrate() > commit_charge(newfolio) Thanks for catching this! Some nits below. > > Fix it by keeping the original objcg around while the inode is isolated. > > Fixes: 743e7d025d2f ("mm: shmem: make unused huge shrinker memcg aware") > Signed-off-by: Chris Mason <mason@kernel.org> > --- > mm/shmem.c | 29 ++++++++++++++++++++++------- > 1 file changed, 22 insertions(+), 7 deletions(-) > > diff --git a/mm/shmem.c b/mm/shmem.c > index 10cef89b00b0..4465a43d6f74 100644 > --- a/mm/shmem.c > +++ b/mm/shmem.c > @@ -758,6 +758,13 @@ shmem_unused_huge_alloc_lru(struct shmem_sb_info *sbinfo, struct folio *folio, > { > int ret; > > + /* > + * A migration target is in the page cache before it is charged. Once > + * charged, its objcg can't change while we hold a folio reference. > + */ > + if (!mem_cgroup_disabled() && !folio_memcg_charged(folio)) > + return ERR_PTR(-EAGAIN); > + > ret = folio_memcg_list_lru_alloc(folio, &sbinfo->shrinklist, gfp); > if (ret) > return ERR_PTR(ret); > @@ -887,12 +894,11 @@ static enum lru_status shmem_unused_huge_isolate(struct list_head *item, > return LRU_REMOVED; > } > > + /* keep shrinklist_objcg, requeue may need it */ > list_lru_isolate(lru, item); > - objcg = shmem_get_and_clear_objcg(info); > set_shmem_unused_huge_isolated(info); > list_add_tail(&info->shrinklist, &scan->list); > spin_unlock(&info->lock); > - obj_cgroup_put(objcg); > > return LRU_REMOVED; > } > @@ -928,32 +934,41 @@ static bool is_shmem_unused_huge_match(struct folio *folio, > static void shmem_unused_huge_drop(struct inode *inode) > { > struct shmem_inode_info *info = SHMEM_I(inode); > + struct obj_cgroup *objcg; > > spin_lock(&info->lock); > list_del_init(&info->shrinklist); > + objcg = shmem_get_and_clear_objcg(info); > spin_unlock(&info->lock); > + > + obj_cgroup_put(objcg); > } > > static void shmem_unused_huge_requeue(struct inode *inode, struct folio *folio) > { > struct shmem_inode_info *info = SHMEM_I(inode); > struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb); > - struct obj_cgroup *objcg; > + struct obj_cgroup *objcg, *old_objcg; > int nid = folio_nid(folio); > > objcg = shmem_unused_huge_alloc_lru(sbinfo, folio, GFP_NOWAIT); > - if (IS_ERR(objcg)) { > - shmem_unused_huge_drop(inode); > - return; > - } Maybe it's better to just return early in the failure path. How about this: +static void shmem_unused_huge_requeue_locked(struct shmem_inode_info *info, + struct shmem_sb_info *sbinfo, int nid) +{ + lockdep_assert_held(&info->lock); + + list_del_init(&info->shrinklist); + shmem_unused_huge_lru_add(sbinfo, &info->shrinklist, nid, + info->shrinklist_objcg); + info->shrinklist_nid = nid; } static void shmem_unused_huge_requeue(struct inode *inode, struct folio *folio) @@ -943,17 +975,18 @@ static void shmem_unused_huge_requeue(struct inode *inode, struct folio *folio) objcg = shmem_unused_huge_alloc_lru(sbinfo, folio, GFP_NOWAIT); if (IS_ERR(objcg)) { - shmem_unused_huge_drop(inode); + spin_lock(&info->lock); + shmem_unused_huge_requeue_locked(info, sbinfo, nid); + spin_unlock(&info->lock); return; } spin_lock(&info->lock); - /* Requeue the inode to shrinklist */ - list_del_init(&info->shrinklist); - shmem_unused_huge_lru_add(sbinfo, &info->shrinklist, nid, objcg); - info->shrinklist_objcg = objcg; - info->shrinklist_nid = nid; + swap(info->shrinklist_objcg, objcg); + shmem_unused_huge_requeue_locked(info, sbinfo, nid); spin_unlock(&info->lock); + + obj_cgroup_put(objcg); } Thanks, Qi > > spin_lock(&info->lock); > + old_objcg = shmem_get_and_clear_objcg(info); > + /* The lists for old_objcg exist, the inode was queued under it */ > + if (IS_ERR(objcg)) { > + objcg = old_objcg; > + old_objcg = NULL; > + } > + > /* Requeue the inode to shrinklist */ > list_del_init(&info->shrinklist); > shmem_unused_huge_lru_add(sbinfo, &info->shrinklist, nid, objcg); > info->shrinklist_objcg = objcg; > info->shrinklist_nid = nid; > spin_unlock(&info->lock); > + > + obj_cgroup_put(old_objcg); > } > > static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo, ^ permalink raw reply [flat|nested] 3+ messages in thread
end of thread, other threads:[~2026-11-20 7:10 UTC | newest] Thread overview: 3+ messages (download: mbox.gz / follow: Atom feed) -- links below jump to the message on this page -- 2026-11-19 19:36 [PATCH 0/1] mm/shmem: don't use the memcg of an uncharged folio Chris Mason 2026-11-19 19:36 ` [PATCH 1/1] " Chris Mason 2026-11-20 7:10 ` Qi Zheng
This is a public inbox, see mirroring instructions for how to clone and mirror all data and code used for this inbox
all inboxes | Powered by JetHome®