Re: [PATCH] slab: Fix obj_ext is mistakenly considered NULL due to race condition

[Harry Yoo <harry.yoo@oracle.com>]

On Thu, Oct 23, 2025 at 04:46:42PM +0800, Hao Ge wrote:
> Hi Harry
> 
> 
> On 2025/10/23 16:23, Hao Ge wrote:
> > Hi Harry
> > 
> > 
> > On 2025/10/23 15:50, Harry Yoo wrote:
> > > On Thu, Oct 23, 2025 at 11:11:56AM +0800, Hao Ge wrote:
> > > > Hi Harry
> > > > 
> > > > 
> > > > On 2025/10/23 10:24, Harry Yoo wrote:
> > > > > On Thu, Oct 23, 2025 at 09:21:17AM +0800, Hao Ge wrote:
> > > > > > From: Hao Ge <gehao@kylinos.cn>
> > > > > > 
> > > > > > If two competing threads enter alloc_slab_obj_exts(), and the
> > > > > > thread that failed to allocate the object extension vector exits
> > > > > > after the one that succeeded, it will mistakenly assume slab->obj_ext
> > > > > > is still empty due to its own allocation failure. This
> > > > > > will then trigger
> > > > > > warnings enforced by CONFIG_MEM_ALLOC_PROFILING_DEBUG checks in
> > > > > > the subsequent free path.
> > > > > > 
> > > > > > Therefore, let's add an additional check when
> > > > > > alloc_slab_obj_exts fails.
> > > > > > 
> > > > > > Signed-off-by: Hao Ge <gehao@kylinos.cn>
> > > > > > ---
> > > > > >    mm/slub.c | 9 ++++++---
> > > > > >    1 file changed, 6 insertions(+), 3 deletions(-)
> > > > > > 
> > > > > > diff --git a/mm/slub.c b/mm/slub.c
> > > > > > index d4403341c9df..42276f0cc920 100644
> > > > > > --- a/mm/slub.c
> > > > > > +++ b/mm/slub.c
> > > > > > @@ -2227,9 +2227,12 @@ prepare_slab_obj_exts_hook(struct
> > > > > > kmem_cache *s, gfp_t flags, void *p)
> > > > > >        slab = virt_to_slab(p);
> > > > > >        if (!slab_obj_exts(slab) &&
> > > > > >            alloc_slab_obj_exts(slab, s, flags, false)) {
> > > > > > -        pr_warn_once("%s, %s: Failed to create slab
> > > > > > extension vector!\n",
> > > > > > -                 __func__, s->name);
> > > > > > -        return NULL;
> > > > > > +        /* Recheck if a racing thread has successfully
> > > > > > allocated slab->obj_exts. */
> > > > > > +        if (!slab_obj_exts(slab)) {
> > > > > > +            pr_warn_once("%s, %s: Failed to create slab
> > > > > > extension vector!\n",
> > > > > > +                     __func__, s->name);
> > > > > > +            return NULL;
> > > > > > +        }
> > > > > >        }
> > > > > Maybe this patch is a bit paranoid... since if
> > > > > mark_failed_objexts_alloc()
> > > > > win cmpxchg() and then someone else allocates the object
> > > > > extension vector,
> > > > > the warning will still be printed anyway.
> > > Oh, just to be clear I was talking about the other warning:
> > > pr_warn_once("%s, %s: Failed to create slab extension vector!",
> > > __func__, s->name);
> > > 
> > > > The process that successfully allocates slab_exts will call
> > > > handle_failed_objexts_alloc, setting ref->ct = CODETAG_EMPTY
> > > > to prevent the warning from being triggered.
> > > But yeah I see what you mean.
> > > 
> > > As you mentioned, if the process that failed to allocate the vector wins
> > > cmpxchg(), later process that successfully allocate the vector would
> > > call set_codetag_empty(), so no warning.
> > > 
> > > But if the process that allocates the vector wins cmpxchg(),
> > > then it won't call set_codetag_empty(), so the process
> > > that was trying to set OBJEXTS_ALLOC_FAIL now needs to set the tag.
> > 
> > Yes, the case I'm encountering is exactly this one.
> > 
> > > 
> > > > > But anyway, I think there is a better way to do this:
> > > What do you think about the diff I suggested below, though?
> > 
> > Sorry for the delayed response earlier; I was trying to deduce all
> > possible scenarios.
> > 
> > It makes sense to me, and I will submit the V2 version based on this
> > suggestion.
> > 
> > Thank you for your help.
> > 
> > > 
> > > > > diff --git a/mm/slub.c b/mm/slub.c
> > > > > index dd4c85ea1038..d08d7580349d 100644
> > > > > --- a/mm/slub.c
> > > > > +++ b/mm/slub.c
> > > > > @@ -2052,9 +2052,9 @@ static inline void
> > > > > mark_objexts_empty(struct slabobj_ext *obj_exts)
> > > > >        }
> > > > >    }
> > > > > -static inline void mark_failed_objexts_alloc(struct slab *slab)
> > > > > +static inline bool mark_failed_objexts_alloc(struct slab *slab)
> > > > >    {
> > > > > -    cmpxchg(&slab->obj_exts, 0, OBJEXTS_ALLOC_FAIL);
> > > > > +    return cmpxchg(&slab->obj_exts, 0, OBJEXTS_ALLOC_FAIL) == 0;
> > > > >    }
> > > > >    static inline void handle_failed_objexts_alloc(unsigned
> > > > > long obj_exts,
> > > > > @@ -2076,7 +2076,7 @@ static inline void
> > > > > handle_failed_objexts_alloc(unsigned long obj_exts,
> > > > >    #else /* CONFIG_MEM_ALLOC_PROFILING_DEBUG */
> > > > >    static inline void mark_objexts_empty(struct slabobj_ext
> > > > > *obj_exts) {}
> > > > > -static inline void mark_failed_objexts_alloc(struct slab *slab) {}
> > > > > +static inline bool mark_failed_objexts_alloc(struct slab
> > > > > *slab) { return true; }
> 
> Maybe it returns false here.
>
> When CONFIG_MEM_ALLOC_PROFILING_DEBUG is not enabled,
> 
> The following condition will never be executed:
> 
> if (!mark_failed_objexts_alloc(slab) && slab_obj_exts(slab))

Good point. But without CONFIG_MEM_ALLOC_PROFILING_DEBUG, we don't know
if someone else successfully allocated the vector or not (unlike, with
CONFIG_MEM_ALLOC_PROFILING_DEBUG enabled, we know that when we lose
cmpxchg()). We cannot "fix" the case where a process fails to allocate
the vector but another allocates the vector.

So I'm not sure if checking slab_obj_exts() once more is worth it in
this case, but I'm fine with either way.

> if another process that allocates the vector, we will lose one count.

By "one count" you mean skipping accounting the object in memory
profiling, right?

> > > > >    static inline void handle_failed_objexts_alloc(unsigned
> > > > > long obj_exts,
> > > > >                struct slabobj_ext *vec, unsigned int objects) {}
> > > > > @@ -2125,7 +2125,9 @@ int alloc_slab_obj_exts(struct slab
> > > > > *slab, struct kmem_cache *s,
> > > > >        }
> > > > >        if (!vec) {
> > > > >            /* Mark vectors which failed to allocate */
> > > > > -        mark_failed_objexts_alloc(slab);
> > > > > +        if (!mark_failed_objexts_alloc(slab) &&
> > > > > +            slab_obj_exts(slab))
> > > > > +            return 0;
> > > > >            return -ENOMEM;
> > > > >        }
> > > > > 

-- 
Cheers,
Harry / Hyeonggon