[PATCH v4 14/16] memcg: Convert mem_cgroup_from_obj_folio() to mem_cgroup_from_obj_slab()

[PATCH v4 14/16] memcg: Convert mem_cgroup_from_obj_folio() to mem_cgroup_from_obj_slab()

[Matthew Wilcox (Oracle)]

In preparation for splitting struct slab from struct page and struct
folio, convert the pointer to a slab rather than a folio.  This means
we can end up passing a NULL slab pointer to mem_cgroup_from_obj_slab()
if the pointer is not to a page allocated to slab, and we handle that
appropriately by returning NULL.

Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: cgroups@vger.kernel.org
---
 mm/memcontrol.c | 36 +++++++++++++-----------------------
 1 file changed, 13 insertions(+), 23 deletions(-)

diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index 025da46d9959..b239d8ad511a 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -2589,38 +2589,28 @@ static inline void mod_objcg_mlstate(struct obj_cgroup *objcg,
 }
 
 static __always_inline
-struct mem_cgroup *mem_cgroup_from_obj_folio(struct folio *folio, void *p)
+struct mem_cgroup *mem_cgroup_from_obj_slab(struct slab *slab, void *p)
 {
 	/*
 	 * Slab objects are accounted individually, not per-page.
 	 * Memcg membership data for each individual object is saved in
 	 * slab->obj_exts.
 	 */
-	if (folio_test_slab(folio)) {
-		struct slabobj_ext *obj_exts;
-		struct slab *slab;
-		unsigned int off;
-
-		slab = folio_slab(folio);
-		obj_exts = slab_obj_exts(slab);
-		if (!obj_exts)
-			return NULL;
+	struct slabobj_ext *obj_exts;
+	unsigned int off;
 
-		off = obj_to_index(slab->slab_cache, slab, p);
-		if (obj_exts[off].objcg)
-			return obj_cgroup_memcg(obj_exts[off].objcg);
+	if (!slab)
+		return NULL;
 
+	obj_exts = slab_obj_exts(slab);
+	if (!obj_exts)
 		return NULL;
-	}
 
-	/*
-	 * folio_memcg_check() is used here, because in theory we can encounter
-	 * a folio where the slab flag has been cleared already, but
-	 * slab->obj_exts has not been freed yet
-	 * folio_memcg_check() will guarantee that a proper memory
-	 * cgroup pointer or NULL will be returned.
-	 */
-	return folio_memcg_check(folio);
+	off = obj_to_index(slab->slab_cache, slab, p);
+	if (obj_exts[off].objcg)
+		return obj_cgroup_memcg(obj_exts[off].objcg);
+
+	return NULL;
 }
 
 /*
@@ -2637,7 +2627,7 @@ struct mem_cgroup *mem_cgroup_from_slab_obj(void *p)
 	if (mem_cgroup_disabled())
 		return NULL;
 
-	return mem_cgroup_from_obj_folio(virt_to_folio(p), p);
+	return mem_cgroup_from_obj_slab(virt_to_slab(p), p);
 }
 
 static struct obj_cgroup *__get_obj_cgroup_from_memcg(struct mem_cgroup *memcg)
-- 
2.47.2

Re: [PATCH v4 14/16] memcg: Convert mem_cgroup_from_obj_folio() to mem_cgroup_from_obj_slab()

[Johannes Weiner]

On Thu, Nov 13, 2025 at 12:09:28AM +0000, Matthew Wilcox (Oracle) wrote:
> -	/*
> -	 * folio_memcg_check() is used here, because in theory we can encounter
> -	 * a folio where the slab flag has been cleared already, but
> -	 * slab->obj_exts has not been freed yet
> -	 * folio_memcg_check() will guarantee that a proper memory
> -	 * cgroup pointer or NULL will be returned.
> -	 */
> -	return folio_memcg_check(folio);
> +	off = obj_to_index(slab->slab_cache, slab, p);
> +	if (obj_exts[off].objcg)
> +		return obj_cgroup_memcg(obj_exts[off].objcg);
> +
> +	return NULL;
>  }
>  
>  /*
> @@ -2637,7 +2627,7 @@ struct mem_cgroup *mem_cgroup_from_slab_obj(void *p)
>  	if (mem_cgroup_disabled())
>  		return NULL;
>  
> -	return mem_cgroup_from_obj_folio(virt_to_folio(p), p);
> +	return mem_cgroup_from_obj_slab(virt_to_slab(p), p);

The name undoubtedly sucks, but there is a comment above this function
that this can be used on non-slab kernel pages as well.

E.g. !vmap kernel stack pages -> mod_lruvec_kmem_state -> mem_cgroup_from_obj_slab

How about:

	if ((slab = virt_to_slap(p)))
		return mem_cgroup_from_obj_slab(slab, p);
	return folio_memcg_check(virt_to_folio(p), p);

Re: [PATCH v4 14/16] memcg: Convert mem_cgroup_from_obj_folio() to mem_cgroup_from_obj_slab()

[Vlastimil Babka]

On 11/13/25 17:14, Johannes Weiner wrote:
> On Thu, Nov 13, 2025 at 12:09:28AM +0000, Matthew Wilcox (Oracle) wrote:
>> -	/*
>> -	 * folio_memcg_check() is used here, because in theory we can encounter
>> -	 * a folio where the slab flag has been cleared already, but
>> -	 * slab->obj_exts has not been freed yet
>> -	 * folio_memcg_check() will guarantee that a proper memory
>> -	 * cgroup pointer or NULL will be returned.
>> -	 */

I've only noticed this comment and thought it describes something impossible
to happen (maybe it was possible in the past) so seemed fine to delete it.

>> -	return folio_memcg_check(folio);
>> +	off = obj_to_index(slab->slab_cache, slab, p);
>> +	if (obj_exts[off].objcg)
>> +		return obj_cgroup_memcg(obj_exts[off].objcg);
>> +
>> +	return NULL;
>>  }
>>  
>>  /*
>> @@ -2637,7 +2627,7 @@ struct mem_cgroup *mem_cgroup_from_slab_obj(void *p)
>>  	if (mem_cgroup_disabled())
>>  		return NULL;
>>  
>> -	return mem_cgroup_from_obj_folio(virt_to_folio(p), p);
>> +	return mem_cgroup_from_obj_slab(virt_to_slab(p), p);
> 
> The name undoubtedly sucks, but there is a comment above this function
> that this can be used on non-slab kernel pages as well.

Didn't notice this one.

> E.g. !vmap kernel stack pages -> mod_lruvec_kmem_state -> mem_cgroup_from_obj_slab
> 
> How about:
> 
> 	if ((slab = virt_to_slap(p)))
> 		return mem_cgroup_from_obj_slab(slab, p);
> 	return folio_memcg_check(virt_to_folio(p), p);

page_memcg_check() maybe instead? we shouldn't get a tail page here, no?

Re: [PATCH v4 14/16] memcg: Convert mem_cgroup_from_obj_folio() to mem_cgroup_from_obj_slab()

[Shakeel Butt]

On Thu, Nov 13, 2025 at 05:28:59PM +0100, Vlastimil Babka wrote:
> > E.g. !vmap kernel stack pages -> mod_lruvec_kmem_state -> mem_cgroup_from_obj_slab
> > 
> > How about:
> > 
> > 	if ((slab = virt_to_slap(p)))
> > 		return mem_cgroup_from_obj_slab(slab, p);
> > 	return folio_memcg_check(virt_to_folio(p), p);
> 
> page_memcg_check() maybe instead? we shouldn't get a tail page here, no?

Do you mean page_memcg_check(virt_to_page(p), p)? But virt_to_page(p)
can return tail page, right?

Re: [PATCH v4 14/16] memcg: Convert mem_cgroup_from_obj_folio() to mem_cgroup_from_obj_slab()

[Matthew Wilcox]

On Thu, Nov 13, 2025 at 11:42:01AM -0800, Shakeel Butt wrote:
> On Thu, Nov 13, 2025 at 05:28:59PM +0100, Vlastimil Babka wrote:
> > > E.g. !vmap kernel stack pages -> mod_lruvec_kmem_state -> mem_cgroup_from_obj_slab
> > > 
> > > How about:
> > > 
> > > 	if ((slab = virt_to_slap(p)))
> > > 		return mem_cgroup_from_obj_slab(slab, p);
> > > 	return folio_memcg_check(virt_to_folio(p), p);
> > 
> > page_memcg_check() maybe instead? we shouldn't get a tail page here, no?
> 
> Do you mean page_memcg_check(virt_to_page(p), p)? But virt_to_page(p)
> can return tail page, right?

Only if it's legitimate to call mod_lruvec_kmem_state() with "a pointer
to somewhere inside the object" rather than "a pointer to the object".

For example, it's legitimate to call copy_to_user() with a pointer
somewhere inside the object, so the usercopy code has to handle that
case.  But it's only legitimate to call kfree() with a pointer that's
at the start of an object, so kfree() can cheerfully BUG_ON()
PageTail(virt_to_page(p)).

Re: [PATCH v4 14/16] memcg: Convert mem_cgroup_from_obj_folio() to mem_cgroup_from_obj_slab()

[Shakeel Butt]

On Thu, Nov 13, 2025 at 08:33:50PM +0000, Matthew Wilcox wrote:
> On Thu, Nov 13, 2025 at 11:42:01AM -0800, Shakeel Butt wrote:
> > On Thu, Nov 13, 2025 at 05:28:59PM +0100, Vlastimil Babka wrote:
> > > > E.g. !vmap kernel stack pages -> mod_lruvec_kmem_state -> mem_cgroup_from_obj_slab
> > > > 
> > > > How about:
> > > > 
> > > > 	if ((slab = virt_to_slap(p)))
> > > > 		return mem_cgroup_from_obj_slab(slab, p);
> > > > 	return folio_memcg_check(virt_to_folio(p), p);
> > > 
> > > page_memcg_check() maybe instead? we shouldn't get a tail page here, no?
> > 
> > Do you mean page_memcg_check(virt_to_page(p), p)? But virt_to_page(p)
> > can return tail page, right?
> 
> Only if it's legitimate to call mod_lruvec_kmem_state() with "a pointer
> to somewhere inside the object" rather than "a pointer to the object".
> 
> For example, it's legitimate to call copy_to_user() with a pointer
> somewhere inside the object, so the usercopy code has to handle that
> case.  But it's only legitimate to call kfree() with a pointer that's
> at the start of an object, so kfree() can cheerfully BUG_ON()
> PageTail(virt_to_page(p)).

Yes it makes sense. Though for slab objects we are allowing to pass any
address within the object (address of list_head for list_lru) to
mem_cgroup_from_slab_obj() (and thus mod_lruvec_kmem_state()) but for
normal kernel memory, if we use virt_to_page() then we will restrict to
the starting address. Anyways just noticed this pecularity, not
suggesting to do anything differently.

Re: [PATCH v4 14/16] memcg: Convert mem_cgroup_from_obj_folio() to mem_cgroup_from_obj_slab()

[Matthew Wilcox]

On Thu, Nov 13, 2025 at 11:14:24AM -0500, Johannes Weiner wrote:
> On Thu, Nov 13, 2025 at 12:09:28AM +0000, Matthew Wilcox (Oracle) wrote:
> > -	/*
> > -	 * folio_memcg_check() is used here, because in theory we can encounter
> > -	 * a folio where the slab flag has been cleared already, but
> > -	 * slab->obj_exts has not been freed yet
> > -	 * folio_memcg_check() will guarantee that a proper memory
> > -	 * cgroup pointer or NULL will be returned.
> > -	 */
> > -	return folio_memcg_check(folio);
> > +	off = obj_to_index(slab->slab_cache, slab, p);
> > +	if (obj_exts[off].objcg)
> > +		return obj_cgroup_memcg(obj_exts[off].objcg);
> > +
> > +	return NULL;
> >  }
> >  
> >  /*
> > @@ -2637,7 +2627,7 @@ struct mem_cgroup *mem_cgroup_from_slab_obj(void *p)
> >  	if (mem_cgroup_disabled())
> >  		return NULL;
> >  
> > -	return mem_cgroup_from_obj_folio(virt_to_folio(p), p);
> > +	return mem_cgroup_from_obj_slab(virt_to_slab(p), p);
> 
> The name undoubtedly sucks, but there is a comment above this function
> that this can be used on non-slab kernel pages as well.

Oh, I see.  Usercopy calls this kind of thing 'heap object', so
perhaps eventually rename this function to mem_cgroup_from_heap_obj()?

> E.g. !vmap kernel stack pages -> mod_lruvec_kmem_state -> mem_cgroup_from_obj_slab

That actually seems to be the only user ... and I wasn't testing on a
!VMAP_STACK build, so I wouldn't've caught it.

> How about:
> 
> 	if ((slab = virt_to_slap(p)))
> 		return mem_cgroup_from_obj_slab(slab, p);
> 	return folio_memcg_check(virt_to_folio(p), p);

Mild updates, here's my counteroffer:

commit 6ca8243530e4
Author: Matthew Wilcox (Oracle) <willy@infradead.org>
Date:   Thu Nov 13 11:30:59 2025 -0500

    fix-memcg-slab

diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index 597673e272a9..a2e6f409c5e8 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -2599,9 +2599,6 @@ struct mem_cgroup *mem_cgroup_from_obj_slab(struct slab *slab, void *p)
 	struct slabobj_ext *obj_exts;
 	unsigned int off;
 
-	if (!slab)
-		return NULL;
-
 	obj_exts = slab_obj_exts(slab);
 	if (!obj_exts)
 		return NULL;
@@ -2624,10 +2621,15 @@ struct mem_cgroup *mem_cgroup_from_obj_slab(struct slab *slab, void *p)
  */
 struct mem_cgroup *mem_cgroup_from_slab_obj(void *p)
 {
+	struct slab *slab;
+
 	if (mem_cgroup_disabled())
 		return NULL;
 
-	return mem_cgroup_from_obj_slab(virt_to_slab(p), p);
+	slab = virt_to_slab(p);
+	if (slab)
+		return mem_cgroup_from_obj_slab(slab, p);
+	return folio_memcg_check(virt_to_folio(p));
 }
 
 static struct obj_cgroup *__get_obj_cgroup_from_memcg(struct mem_cgroup *memcg)

Re: [PATCH v4 14/16] memcg: Convert mem_cgroup_from_obj_folio() to mem_cgroup_from_obj_slab()

[Johannes Weiner]

On Thu, Nov 13, 2025 at 04:39:41PM +0000, Matthew Wilcox wrote:
> --- a/mm/memcontrol.c
> +++ b/mm/memcontrol.c
> @@ -2599,9 +2599,6 @@ struct mem_cgroup *mem_cgroup_from_obj_slab(struct slab *slab, void *p)
>  	struct slabobj_ext *obj_exts;
>  	unsigned int off;
>  
> -	if (!slab)
> -		return NULL;
> -
>  	obj_exts = slab_obj_exts(slab);
>  	if (!obj_exts)
>  		return NULL;
> @@ -2624,10 +2621,15 @@ struct mem_cgroup *mem_cgroup_from_obj_slab(struct slab *slab, void *p)
>   */
>  struct mem_cgroup *mem_cgroup_from_slab_obj(void *p)
>  {
> +	struct slab *slab;
> +
>  	if (mem_cgroup_disabled())
>  		return NULL;
>  
> -	return mem_cgroup_from_obj_slab(virt_to_slab(p), p);
> +	slab = virt_to_slab(p);
> +	if (slab)
> +		return mem_cgroup_from_obj_slab(slab, p);
> +	return folio_memcg_check(virt_to_folio(p));

Looks good to me, thanks!

With that folded in, for the combined patch:

Acked-by: Johannes Weiner <hannes@cmpxchg.org>

Re: [PATCH v4 14/16] memcg: Convert mem_cgroup_from_obj_folio() to mem_cgroup_from_obj_slab()

[Vlastimil Babka]

On 11/13/25 20:16, Johannes Weiner wrote:
> On Thu, Nov 13, 2025 at 04:39:41PM +0000, Matthew Wilcox wrote:
>> --- a/mm/memcontrol.c
>> +++ b/mm/memcontrol.c
>> @@ -2599,9 +2599,6 @@ struct mem_cgroup *mem_cgroup_from_obj_slab(struct slab *slab, void *p)
>>  	struct slabobj_ext *obj_exts;
>>  	unsigned int off;
>>  
>> -	if (!slab)
>> -		return NULL;
>> -
>>  	obj_exts = slab_obj_exts(slab);
>>  	if (!obj_exts)
>>  		return NULL;
>> @@ -2624,10 +2621,15 @@ struct mem_cgroup *mem_cgroup_from_obj_slab(struct slab *slab, void *p)
>>   */
>>  struct mem_cgroup *mem_cgroup_from_slab_obj(void *p)
>>  {
>> +	struct slab *slab;
>> +
>>  	if (mem_cgroup_disabled())
>>  		return NULL;
>>  
>> -	return mem_cgroup_from_obj_slab(virt_to_slab(p), p);
>> +	slab = virt_to_slab(p);
>> +	if (slab)
>> +		return mem_cgroup_from_obj_slab(slab, p);
>> +	return folio_memcg_check(virt_to_folio(p));
> 
> Looks good to me, thanks!
> 
> With that folded in, for the combined patch:
> 
> Acked-by: Johannes Weiner <hannes@cmpxchg.org>

Thanks, folded.

Re: [PATCH v4 14/16] memcg: Convert mem_cgroup_from_obj_folio() to mem_cgroup_from_obj_slab()

[Harry Yoo]

On Thu, Nov 13, 2025 at 04:39:41PM +0000, Matthew Wilcox wrote:
> On Thu, Nov 13, 2025 at 11:14:24AM -0500, Johannes Weiner wrote:
> > On Thu, Nov 13, 2025 at 12:09:28AM +0000, Matthew Wilcox (Oracle) wrote:
> > > -	/*
> > > -	 * folio_memcg_check() is used here, because in theory we can encounter
> > > -	 * a folio where the slab flag has been cleared already, but
> > > -	 * slab->obj_exts has not been freed yet
> > > -	 * folio_memcg_check() will guarantee that a proper memory
> > > -	 * cgroup pointer or NULL will be returned.
> > > -	 */
> > > -	return folio_memcg_check(folio);
> > > +	off = obj_to_index(slab->slab_cache, slab, p);
> > > +	if (obj_exts[off].objcg)
> > > +		return obj_cgroup_memcg(obj_exts[off].objcg);
> > > +
> > > +	return NULL;
> > >  }
> > >  
> > >  /*
> > > @@ -2637,7 +2627,7 @@ struct mem_cgroup *mem_cgroup_from_slab_obj(void *p)
> > >  	if (mem_cgroup_disabled())
> > >  		return NULL;
> > >  
> > > -	return mem_cgroup_from_obj_folio(virt_to_folio(p), p);
> > > +	return mem_cgroup_from_obj_slab(virt_to_slab(p), p);
> > 
> > The name undoubtedly sucks, but there is a comment above this function
> > that this can be used on non-slab kernel pages as well.
> 
> Oh, I see.  Usercopy calls this kind of thing 'heap object', so
> perhaps eventually rename this function to mem_cgroup_from_heap_obj()?
> 
> > E.g. !vmap kernel stack pages -> mod_lruvec_kmem_state -> mem_cgroup_from_obj_slab
> 
> That actually seems to be the only user ... and I wasn't testing on a
> !VMAP_STACK build, so I wouldn't've caught it.
> 
> > How about:
> > 
> > 	if ((slab = virt_to_slap(p)))
> > 		return mem_cgroup_from_obj_slab(slab, p);
> > 	return folio_memcg_check(virt_to_folio(p), p);
> 
> Mild updates, here's my counteroffer:
> 
> commit 6ca8243530e4
> Author: Matthew Wilcox (Oracle) <willy@infradead.org>
> Date:   Thu Nov 13 11:30:59 2025 -0500
> 
>     fix-memcg-slab

With this folded,
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>

-- 
Cheers,
Harry / Hyeonggon

> diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> index 597673e272a9..a2e6f409c5e8 100644
> --- a/mm/memcontrol.c
> +++ b/mm/memcontrol.c
> @@ -2599,9 +2599,6 @@ struct mem_cgroup *mem_cgroup_from_obj_slab(struct slab *slab, void *p)
>  	struct slabobj_ext *obj_exts;
>  	unsigned int off;
>  
> -	if (!slab)
> -		return NULL;
> -
>  	obj_exts = slab_obj_exts(slab);
>  	if (!obj_exts)
>  		return NULL;
> @@ -2624,10 +2621,15 @@ struct mem_cgroup *mem_cgroup_from_obj_slab(struct slab *slab, void *p)
>   */
>  struct mem_cgroup *mem_cgroup_from_slab_obj(void *p)
>  {
> +	struct slab *slab;
> +
>  	if (mem_cgroup_disabled())
>  		return NULL;
>  
> -	return mem_cgroup_from_obj_slab(virt_to_slab(p), p);
> +	slab = virt_to_slab(p);
> +	if (slab)
> +		return mem_cgroup_from_obj_slab(slab, p);
> +	return folio_memcg_check(virt_to_folio(p));
>  }
>  
>  static struct obj_cgroup *__get_obj_cgroup_from_memcg(struct mem_cgroup *memcg)