Re: [PATCH 1/5] mm: memcg/slab: Fix return child memcg objcg for root memcg

[Roman Gushchin <guro@fb.com>]

On Tue, Oct 27, 2020 at 10:42:36AM -0700, Roman Gushchin wrote:
> On Tue, Oct 27, 2020 at 04:02:52PM +0800, Muchun Song wrote:
> > Consider the following memcg hierarchy.
> > 
> >                     root
> >                    /    \
> >                   A      B
> > 
> > If we get the objcg of memcg A failed, the get_obj_cgroup_from_current
> > can return the wrong objcg for the root memcg.
> > 
> > Fixes: bf4f059954dc ("mm: memcg/slab: obj_cgroup API")
> > Signed-off-by: Muchun Song <songmuchun@bytedance.com>
> > ---
> >  mm/memcontrol.c | 5 ++++-
> >  1 file changed, 4 insertions(+), 1 deletion(-)
> 
> Hi Muchun,
> 
> thank you for the patchset!
> 
> A generic note: it's usually not a good idea to group fixes with non-fixes
> in one patchset if fixes are planned to be backported to stable.
> 
> > 
> > diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > index 1337775b04f3..fcbd79c5023e 100644
> > --- a/mm/memcontrol.c
> > +++ b/mm/memcontrol.c
> > @@ -2945,7 +2945,7 @@ struct mem_cgroup *mem_cgroup_from_obj(void *p)
> >  
> >  __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
> >  {
> > -	struct obj_cgroup *objcg = NULL;
> > +	struct obj_cgroup *objcg;
> >  	struct mem_cgroup *memcg;
> >  
> >  	if (memcg_kmem_bypass())
> > @@ -2964,6 +2964,9 @@ __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
> >  	}
> >  	rcu_read_unlock();
> >  
> > +	if (memcg == root_mem_cgroup)
> > +		return NULL;
> > +
> >  	return objcg;
> >  }
> 
> I agree, it's a bug. Good catch, thanks!
> 
> I would prefer a slightly different fix though. Because we're going to change how
> the root memory cgroup is handled (an rfc version posted here:
> https://lkml.org/lkml/2020/10/21/612), it's better to not use a comparison
> with the root_mem_cgroup and handle the obj_cgroup_tryget() return code instead.
> Something like this:
> 
> @@ -2973,6 +2973,7 @@ __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
>  		objcg = rcu_dereference(memcg->objcg);
>  		if (objcg && obj_cgroup_tryget(objcg))
>  			break;
> +		objcg = NULL;
>  	}
>  	rcu_read_unlock();
> 
> Or a more explicit:
>    if (obj_cgroup_tryget(objcg)) {
>      ...
>    } else {
>      ...
>    }
> 
> Also, please, add:
> Cc: stable@vger.kernel.org

After a second thought, it's also a highly theoretical race. In order to have
obj_cgroup_tryget() failed we need to enter get_obj_cgroup_from_current()
racing with cgroup offlining (which is already tricky) and then have
memcg_reparent_objcgs() finish in between reading memcg->objcg and bumping
the reference counter.

So I'm not sure we need a stable@ backport, sorry for the confusion.

Thanks!