Re: [PATCH] memcg: fix slab accounting in refill_obj_stock() trylock path

[Hao Li]

On Fri, Feb 27, 2026 at 08:46:18AM +0100, Vlastimil Babka wrote:
> On 2/27/26 02:01, Hao Li wrote:
> > On Thu, Feb 26, 2026 at 02:44:02PM +0100, Vlastimil Babka wrote:
> >> On 2/26/26 14:39, Shakeel Butt wrote:
> >> > On Thu, Feb 26, 2026 at 07:51:37PM +0800, Hao Li wrote:
> >> >> In the trylock path of refill_obj_stock(), mod_objcg_mlstate() should
> >> >> use the real alloc/free bytes (i.e., nr_acct) for accounting, rather
> >> >> than nr_bytes.
> >> >> 
> >> >> Fixes: 200577f69f29 ("memcg: objcg stock trylock without irq disabling")
> >> >> Cc: stable@xxxxxxxxxxxxxxx
> >> >> Signed-off-by: Hao Li <hao.li@xxxxxxxxx>
> >> > 
> >> > Thanks for the fix.
> >> > 
> >> > Acked-by: Shakeel Butt <shakeel.butt@xxxxxxxxx>
> >> 
> >> What are the user-visible effects of the bug?
> > 
> > The user-visible impact is that the NR_SLAB_RECLAIMABLE_B and
> > NR_SLAB_UNRECLAIMABLE_B stats can end up being incorrect.
> > 
> > For example, if a user allocates a 6144-byte object, then before this fix
> > refill_obj_stock() calls mod_objcg_mlstate(..., nr_bytes=2048), even though it
> > should account for 6144 bytes (i.e., nr_acct).
> > 
> > When the user later frees the same object with kfree(), refill_obj_stock() calls
> > mod_objcg_mlstate(..., nr_bytes=6144). This ends up adding 6144 to the stats,
> > but it should be applying -6144 (i.e., nr_acct) since the object is being
> > freed.
> 
> Thanks, I'm sure Andrew will amend the changelog with those useful details.

Got it. Thanks.

> 
> Weird that we went since 6.16 with nobody noticing the stats were off - it
> sounds they could get really way off?

Indeed, it does seem a bit unbelievable. I suspect the conditions required for
this issue to occur are quite strict: a process context first hold the
obj_stock.lock, then get interrupted by an IRQ, and the IRQ path also reach
refill_obj_stock and then hit the local_trylock-failed path.

Therefore, I think a small amount of data distortion might be hard to observe.

-- 
Thanks,
Hao