Re: [External] Re: [PATCH v2] mm: memcg/slab: fix memory leak at non-root kmem_cache destroy

[Roman Gushchin <guro@fb.com>]

On Thu, Jul 16, 2020 at 01:07:02PM +0800, Muchun Song wrote:
> On Thu, Jul 16, 2020 at 1:54 AM Roman Gushchin <guro@fb.com> wrote:
> >
> > On Thu, Jul 16, 2020 at 12:50:22AM +0800, Muchun Song wrote:
> > > If the kmem_cache refcount is greater than one, we should not
> > > mark the root kmem_cache as dying. If we mark the root kmem_cache
> > > dying incorrectly, the non-root kmem_cache can never be destroyed.
> > > It resulted in memory leak when memcg was destroyed. We can use the
> > > following steps to reproduce.
> > >
> > >   1) Use kmem_cache_create() to create a new kmem_cache named A.
> > >   2) Coincidentally, the kmem_cache A is an alias for kmem_cache B,
> > >      so the refcount of B is just increased.
> > >   3) Use kmem_cache_destroy() to destroy the kmem_cache A, just
> > >      decrease the B's refcount but mark the B as dying.
> > >   4) Create a new memory cgroup and alloc memory from the kmem_cache
> > >      B. It leads to create a non-root kmem_cache for allocating memory.
> > >   5) When destroy the memory cgroup created in the step 4), the
> > >      non-root kmem_cache can never be destroyed.
> > >
> > > If we repeat steps 4) and 5), this will cause a lot of memory leak.
> > > So only when refcount reach zero, we mark the root kmem_cache as dying.
> > >
> > > Fixes: 92ee383f6daa ("mm: fix race between kmem_cache destroy, create and deactivate")
> > > Signed-off-by: Muchun Song <songmuchun@bytedance.com>
> > > Reviewed-by: Shakeel Butt <shakeelb@google.com>
> > > ---
> > >
> > > changelog in v2:
> > >  1) Fix a confusing typo in the commit log.
> >
> > Ok, now I see the problem. Thank you for fixing the commit log!
> >
> > >  2) Remove flush_memcg_workqueue() for !CONFIG_MEMCG_KMEM.
> > >  3) Introduce a new helper memcg_set_kmem_cache_dying() to fix a race
> > >     condition between flush_memcg_workqueue() and slab_unmergeable().
> > >
> > >  mm/slab_common.c | 54 +++++++++++++++++++++++++++++++++++++++++++++++-------
> > >  1 file changed, 47 insertions(+), 7 deletions(-)
> > >
> > > diff --git a/mm/slab_common.c b/mm/slab_common.c
> > > index 8c1ffbf7de45..c4958116e3fd 100644
> > > --- a/mm/slab_common.c
> > > +++ b/mm/slab_common.c
> > > @@ -258,6 +258,11 @@ static void memcg_unlink_cache(struct kmem_cache *s)
> > >               list_del(&s->memcg_params.kmem_caches_node);
> > >       }
> > >  }
> > > +
> > > +static inline bool memcg_kmem_cache_dying(struct kmem_cache *s)
> > > +{
> > > +     return is_root_cache(s) && s->memcg_params.dying;
> > > +}
> > >  #else
> > >  static inline int init_memcg_params(struct kmem_cache *s,
> > >                                   struct kmem_cache *root_cache)
> > > @@ -272,6 +277,11 @@ static inline void destroy_memcg_params(struct kmem_cache *s)
> > >  static inline void memcg_unlink_cache(struct kmem_cache *s)
> > >  {
> > >  }
> > > +
> > > +static inline bool memcg_kmem_cache_dying(struct kmem_cache *s)
> > > +{
> > > +     return false;
> > > +}
> > >  #endif /* CONFIG_MEMCG_KMEM */
> > >
> > >  /*
> > > @@ -326,6 +336,13 @@ int slab_unmergeable(struct kmem_cache *s)
> > >       if (s->refcount < 0)
> > >               return 1;
> > >
> > > +     /*
> > > +      * If the kmem_cache is dying. We should also skip this
> > > +      * kmem_cache.
> > > +      */
> > > +     if (memcg_kmem_cache_dying(s))
> > > +             return 1;
> > > +
> > >       return 0;
> > >  }
> > >
> > > @@ -886,12 +903,15 @@ static int shutdown_memcg_caches(struct kmem_cache *s)
> > >       return 0;
> > >  }
> > >
> > > -static void flush_memcg_workqueue(struct kmem_cache *s)
> > > +static void memcg_set_kmem_cache_dying(struct kmem_cache *s)
> > >  {
> > >       spin_lock_irq(&memcg_kmem_wq_lock);
> > >       s->memcg_params.dying = true;
> > >       spin_unlock_irq(&memcg_kmem_wq_lock);
> > > +}
> > >
> > > +static void flush_memcg_workqueue(struct kmem_cache *s)
> > > +{
> > >       /*
> > >        * SLAB and SLUB deactivate the kmem_caches through call_rcu. Make
> > >        * sure all registered rcu callbacks have been invoked.
> > > @@ -923,10 +943,6 @@ static inline int shutdown_memcg_caches(struct kmem_cache *s)
> > >  {
> > >       return 0;
> > >  }
> > > -
> > > -static inline void flush_memcg_workqueue(struct kmem_cache *s)
> > > -{
> > > -}
> > >  #endif /* CONFIG_MEMCG_KMEM */
> > >
> > >  void slab_kmem_cache_release(struct kmem_cache *s)
> > > @@ -944,8 +960,6 @@ void kmem_cache_destroy(struct kmem_cache *s)
> > >       if (unlikely(!s))
> > >               return;
> > >
> > > -     flush_memcg_workqueue(s);
> > > -
> > >       get_online_cpus();
> > >       get_online_mems();
> > >
> > > @@ -955,6 +969,32 @@ void kmem_cache_destroy(struct kmem_cache *s)
> > >       if (s->refcount)
> > >               goto out_unlock;
> > >
> > > +#ifdef CONFIG_MEMCG_KMEM
> > > +     memcg_set_kmem_cache_dying(s);
> > > +
> > > +     mutex_unlock(&slab_mutex);
> >
> > Hm, but in theory s->refcount can be increased here?
> 
> I have tried my best to read all the codes that operate on s->refcount.
> There is only one place which increases the s->refcount, it is the
> __kmem_cache_alias(). If the kmem cache is dying, the slab_unmergeable()
> can never return true for the dying kmem cache because it is the same slab_mutex
> protection, so I think that there is not a problem, right?

So the problem is that you're checking s->refcount under slab_mutex,
then you're releasing the mutex and then set the dying bit. But nothing prevents
someone else to clone the kmem_cache on a different CPU as soon as you've released
the slab_mutex.

However, I think there is a simple solution for this problem.
Can you, please, check the following patch?

--
diff --git a/mm/slab_common.c b/mm/slab_common.c
index 37d48a56431d..d6661e7af01e 100644
--- a/mm/slab_common.c
+++ b/mm/slab_common.c
@@ -326,6 +326,11 @@ int slab_unmergeable(struct kmem_cache *s)
 	if (s->refcount < 0)
 		return 1;
 
+#ifdef CONFIG_MEMCG_KMEM
+	if (s->memcg_params.dying)
+		return 1;
+#endif
+
 	return 0;
 }
 
@@ -888,9 +893,20 @@ static int shutdown_memcg_caches(struct kmem_cache *s)
 
 static void flush_memcg_workqueue(struct kmem_cache *s)
 {
-	spin_lock_irq(&memcg_kmem_wq_lock);
-	s->memcg_params.dying = true;
-	spin_unlock_irq(&memcg_kmem_wq_lock);
+	bool skip = false;
+
+	mutex_lock(&slab_mutex);
+	if (s->refcount == 1) {
+		spin_lock_irq(&memcg_kmem_wq_lock);
+		s->memcg_params.dying = true;
+		spin_unlock_irq(&memcg_kmem_wq_lock);
+	} else {
+		skip = true;
+	}
+	mutex_unlock(&slab_mutex);
+
+	if (skip)
+		return;
 
 	/*
 	 * SLAB and SLUB deactivate the kmem_caches through call_rcu. Make
-- 
2.26.2


--
> 
> > So it doesn't solve the problem completely, but makes it less probable, right?
> >
> > I wonder if it's possible to (additionally) protect s->refcount with a
> > memcg_kmem_wq_lock, so that we can check it in the context of flush_memcg_workqueue()?
> >
> > > +
> > > +     put_online_mems();
> > > +     put_online_cpus();
> > > +
> > > +     flush_memcg_workqueue(s);
> > > +
> > > +     get_online_cpus();
> > > +     get_online_mems();
> > > +
> > > +     mutex_lock(&slab_mutex);
> > > +
> > > +     if (WARN(s->refcount,
> > > +              "kmem_cache_destroy %s: Slab cache is still referenced\n",
> > > +              s->name)) {
> > > +             /*
> > > +              * Reset the dying flag setted by memcg_set_kmem_cache_dying().
> > > +              */
> > > +             s->memcg_params.dying = false;
> > > +             goto out_unlock;
> > > +     }
> > > +#endif
> > > +
> > >       err = shutdown_memcg_caches(s);
> > >       if (!err)
> > >               err = shutdown_cache(s);
> > > --
> > > 2.11.0
> > >
> >
> > Other than the problem above your patch looks really good to me, however we should
> > be really careful here, as it should in theory be back-ported to a big number
> > of old stable kernels. And because it's (hopefully) fixed in 5.9, it's a backport-only
> > patch.
> >
> > So I wonder if we can mitigate the problem by disabling cache sharing for some
> > specific kmem_caches instead? Like for all with SLAB_ACCOUNT and maybe for all except
> > a hard-coded list (if kmem accounting is enabled). Do you mind sharing any details
> > on how this problem reveals itself in the real life?
> 
> One day I debug another problem, coincidentally I executed the
> following command:
> 
>      cat /sys/kernel/debug/memcg_slabinfo | grep deact
> 
> And I found an interesting thing, a lot of non-root kmem_cache's
> active_objs is zero.
> But why are they not destroyed? So I launched a detailed
> investigation. Finally I found
> out that the reason is that we mark the root kmem_cache as dying
> incorrectly. I have
> applied this patch on our server for several weeks. And this problem
> has disappeared.

Got it, it's a great catch!

Thank you!

Roman