Re: [PATCH v3] slab: Add check for memcg_data != OBJEXTS_ALLOC_FAIL in folio_memcg_kmem

[Vlastimil Babka]

On 10/14/25 22:14, Shakeel Butt wrote:
> On Tue, Oct 14, 2025 at 09:12:43AM -0700, Suren Baghdasaryan wrote:
>> On Tue, Oct 14, 2025 at 8:28 AM Hao Ge <hao.ge@xxxxxxxxx> wrote:
>> >
>> > From: Hao Ge <gehao@xxxxxxxxxx>
>> >
>> > Since OBJEXTS_ALLOC_FAIL and MEMCG_DATA_OBJEXTS currently share
>> > the same bit position, we cannot determine whether memcg_data still
>> > points to the slabobj_ext vector simply by checking
>> > folio->memcg_data & MEMCG_DATA_OBJEXTS.
>> >
>> > If obj_exts allocation failed, slab->obj_exts is set to OBJEXTS_ALLOC_FAIL,
>> > and during the release of the associated folio, the BUG check is triggered
>> > because it was mistakenly assumed that a valid folio->memcg_data
>> > was not cleared before freeing the folio.
>> >
>> > So let's check for memcg_data != OBJEXTS_ALLOC_FAIL in folio_memcg_kmem.
>> >
>> > Fixes: 7612833192d5 ("slab: Reuse first bit for OBJEXTS_ALLOC_FAIL")
>> > Suggested-by: Harry Yoo <harry.yoo@xxxxxxxxxx>
>> > Signed-off-by: Hao Ge <gehao@xxxxxxxxxx>
>> 
>> Reviewed-by: Suren Baghdasaryan <surenb@xxxxxxxxxx>
>> 
>> nit: I think it would be helpful if the changelog explained why we
>> need the additional check. We can have the same bit set in two
>> different situations:
>> 1. object extension vector allocation failure;
>> 2. memcg_data pointing to a valid mem_cgroup.
>> To distinguish between them, we need to check not only the bit itself
>> but also the rest of this field. If the rest is NULL, we have case 1,
>> otherwise case 2.
> 
> With Suren's suggestion, you can add:
> 
> Acked-by: Shakeel Butt <shakeel.butt@xxxxxxxxx>

Thanks, I added Suren's suggestion and pushed to slab/for-next-fixes:

https://git.kernel.org/pub/scm/linux/kernel/git/vbabka/slab.git/commit/?h=slab/for-next-fixes&id=711c435c89e59ee32bf8bb1c0d875a07931da5a8

Resisted the impulse to change the single VM_BUG_ON_FOLIO to
VM_WARN_ON_ONCE_FOLIO because we're still going to do that systematically,
right?