[PATCH] mm/slab: initialize slab->stride early to avoid memory ordering issues

[PATCH] mm/slab: initialize slab->stride early to avoid memory ordering issues

[Harry Yoo]

When alloc_slab_obj_exts() is called later in time (instead of at slab
allocation & initialization step), slab->stride and slab->obj_exts are
set when the slab is already accessible by multiple CPUs.

The current implementation does not enforce memory ordering between
slab->stride and slab->obj_exts. However, for correctness, slab->stride
must be visible before slab->obj_exts, otherwise concurrent readers
may observe slab->obj_exts as non-zero while stride is still stale,
leading to incorrect reference counting of object cgroups.

There has been a bug report [1] that showed symptoms of incorrect
reference counting of object cgroups, which could be triggered by
this memory ordering issue.

Fix this by unconditionally initializing slab->stride in
alloc_slab_obj_exts_early(), before the need_slab_obj_exts() check.
In case of SLAB_OBJ_EXT_IN_OBJ, it is overridden in the same function.

This ensures stride is set before the slab becomes visible to
other CPUs via the per-node partial slab list (protected by spinlock
with acquire/release semantics), preventing them from observing
inconsistent stride value.

Thanks to Shakeel Butt for pointing out this issue [2].

Fixes: 7a8e71bc619d ("mm/slab: use stride to access slabobj_ext")
Reported-by: Venkat Rao Bagalkote <venkat88@linux.ibm.com>
Closes: https://lore.kernel.org/lkml/ca241daa-e7e7-4604-a48d-de91ec9184a5@linux.ibm.com [1]
Link: https://lore.kernel.org/linux-mm/aZu9G9mVIVzSm6Ft@hyeyoo [2]
Signed-off-by: Harry Yoo <harry.yoo@oracle.com>
---

I tested this patch, but I could not confirm that this actually fixes
the issue reported by [1]. It would be nice if Venkat could help
confirm; but perhaps it's challenging to reliably reproduce...

Since this logically makes sense, it would be worth fix it anyway.

 mm/slub.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/mm/slub.c b/mm/slub.c
index 18c30872d196..afa98065d74f 100644
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -2196,7 +2196,6 @@ int alloc_slab_obj_exts(struct slab *slab, struct kmem_cache *s,
 retry:
 	old_exts = READ_ONCE(slab->obj_exts);
 	handle_failed_objexts_alloc(old_exts, vec, objects);
-	slab_set_stride(slab, sizeof(struct slabobj_ext));
 
 	if (new_slab) {
 		/*
@@ -2272,6 +2271,9 @@ static void alloc_slab_obj_exts_early(struct kmem_cache *s, struct slab *slab)
 	void *addr;
 	unsigned long obj_exts;
 
+	/* Initialize stride early to avoid memory ordering issues */
+	slab_set_stride(slab, sizeof(struct slabobj_ext));
+
 	if (!need_slab_obj_exts(s))
 		return;
 
@@ -2288,7 +2290,6 @@ static void alloc_slab_obj_exts_early(struct kmem_cache *s, struct slab *slab)
 		obj_exts |= MEMCG_DATA_OBJEXTS;
 #endif
 		slab->obj_exts = obj_exts;
-		slab_set_stride(slab, sizeof(struct slabobj_ext));
 	} else if (s->flags & SLAB_OBJ_EXT_IN_OBJ) {
 		unsigned int offset = obj_exts_offset_in_object(s);
 
-- 
2.43.0

Re: [PATCH] mm/slab: initialize slab->stride early to avoid memory ordering issues

[Harry Yoo]

On Mon, Feb 23, 2026 at 04:58:09PM +0900, Harry Yoo wrote:
> When alloc_slab_obj_exts() is called later in time (instead of at slab
> allocation & initialization step), slab->stride and slab->obj_exts are
> set when the slab is already accessible by multiple CPUs.
> 
> The current implementation does not enforce memory ordering between
> slab->stride and slab->obj_exts. However, for correctness, slab->stride
> must be visible before slab->obj_exts, otherwise concurrent readers
> may observe slab->obj_exts as non-zero while stride is still stale,
> leading to incorrect reference counting of object cgroups.
> 
> There has been a bug report [1] that showed symptoms of incorrect
> reference counting of object cgroups, which could be triggered by
> this memory ordering issue.
> 
> Fix this by unconditionally initializing slab->stride in
> alloc_slab_obj_exts_early(), before the need_slab_obj_exts() check.
> In case of SLAB_OBJ_EXT_IN_OBJ, it is overridden in the same function.
> 
> This ensures stride is set before the slab becomes visible to
> other CPUs via the per-node partial slab list (protected by spinlock
> with acquire/release semantics), preventing them from observing
> inconsistent stride value.
> 
> Thanks to Shakeel Butt for pointing out this issue [2].
> 
> Fixes: 7a8e71bc619d ("mm/slab: use stride to access slabobj_ext")
> Reported-by: Venkat Rao Bagalkote <venkat88@linux.ibm.com>
> Closes: https://lore.kernel.org/lkml/ca241daa-e7e7-4604-a48d-de91ec9184a5@linux.ibm.com [1]
> Link: https://lore.kernel.org/linux-mm/aZu9G9mVIVzSm6Ft@hyeyoo [2]
> Signed-off-by: Harry Yoo <harry.yoo@oracle.com>
> ---

Vlastimil, could you please update the changelog when applying this
to the tree? I think this also explains [3] (thanks for raising it
off-list, Vlastimil!):

When alloc_slab_obj_exts() is called later (instead of during slab
allocation and initialization), slab->stride and slab->obj_exts are
updated after the slab is already accessible by multiple CPUs.

The current implementation does not enforce memory ordering between
slab->stride and slab->obj_exts. For correctness, slab->stride must be
visible before slab->obj_exts. Otherwise, concurrent readers may observe
slab->obj_exts as non-zero while stride is still stale.

With stale slab->stride, slab_obj_ext() could return the wrong obj_ext.
This could cause two problems:

  - obj_cgroup_put() is called on the wrong objcg, leading to
    a use-after-free due to incorrect reference counting [1] by
    decrementing the reference count more than it was incremented.

  - refill_obj_stock() is called on the wrong objcg, leading to
    a page_counter overflow [2] by uncharging more memory than charged.

Fix this by unconditionally initializing slab->stride in
alloc_slab_obj_exts_early(), before the need_slab_obj_exts() check.
In the case of SLAB_OBJ_EXT_IN_OBJ, it is overridden in the function.

This ensures updates to slab->stride become visible before the slab
can be accessed by other CPUs via the per-node partial slab list
(protected by spinlock with acquire/release semantics).

Thanks to Shakeel Butt for pointing out this issue [3].

Fixes: 7a8e71bc619d ("mm/slab: use stride to access slabobj_ext")
Reported-by: Venkat Rao Bagalkote <venkat88@linux.ibm.com>
Closes: https://lore.kernel.org/lkml/ca241daa-e7e7-4604-a48d-de91ec9184a5@linux.ibm.com [1]
Closes: https://lore.kernel.org/all/ddff7c7d-c0c3-4780-808f-9a83268bbf0c@linux.ibm.com [2]
Link: https://lore.kernel.org/linux-mm/aZu9G9mVIVzSm6Ft@hyeyoo [3]
Signed-off-by: Harry Yoo <harry.yoo@oracle.com>

-- 
Cheers,
Harry / Hyeonggon