Re: [PATCH bpf-next 05/15] bpf: Fix incorrect mem_cgroup_put

[Yafang Shao <laoar.shao@gmail.com>]

On Fri, Aug 12, 2022 at 1:34 PM Muchun Song <muchun.song@linux.dev> wrote:
>
>
>
> > On Aug 12, 2022, at 08:27, Yafang Shao <laoar.shao@gmail.com> wrote:
> >
> > On Thu, Aug 11, 2022 at 11:47 PM Shakeel Butt <shakeelb@google.com> wrote:
> >>
> >> On Thu, Aug 11, 2022 at 10:49:13AM +0800, Yafang Shao wrote:
> >>> On Thu, Aug 11, 2022 at 1:07 AM Shakeel Butt <shakeelb@google.com> wrote:
> >>>>
> >>>> On Wed, Aug 10, 2022 at 03:18:30PM +0000, Yafang Shao wrote:
> >>>>> The memcg may be the root_mem_cgroup, in which case we shouldn't put it.
> >>>>
> >>>> No, it is ok to put root_mem_cgroup. css_put already handles the root
> >>>> cgroups.
> >>>>
> >>>
> >>> Ah, this commit log doesn't describe the issue clearly. I should improve it.
> >>> The issue is that in bpf_map_get_memcg() it doesn't get the objcg if
> >>> map->objcg is NULL (that can happen if the map belongs to the root
> >>> memcg), so we shouldn't put the objcg if map->objcg is NULL neither in
> >>> bpf_map_put_memcg().
> >>
> >> Sorry I am still not understanding. We are not 'getting' objcg in
> >> bpf_map_get_memcg(). We are 'getting' memcg from map->objcg and if that
> >> is NULL the function is returning root memcg and putting root memcg is
> >> totally fine.
> >
> > When the map belongs to root_mem_cgroup, the map->objcg is NULL, right ?
> > See also bpf_map_save_memcg() and it describes clearly in the comment -
> >
> > static void bpf_map_save_memcg(struct bpf_map *map)
> > {
> >        /* Currently if a map is created by a process belonging to the root
> >         * memory cgroup, get_obj_cgroup_from_current() will return NULL.
> >         * So we have to check map->objcg for being NULL each time it's
> >         * being used.
> >         */
> >        map->objcg = get_obj_cgroup_from_current();
> > }
> >
> > So for the root_mem_cgroup case, bpf_map_get_memcg() will return
> > root_mem_cgroup directly without getting its css, right ? See below,
> >
> > static struct mem_cgroup *bpf_map_get_memcg(const struct bpf_map *map)
> > {
> >
> >        if (map->objcg)
> >                return get_mem_cgroup_from_objcg(map->objcg);
> >
> >        return root_mem_cgroup;   // without css_get(&memcg->css);
> > }
> >
> > But it will put the css unconditionally.  See below,
> >
> > memcg = bpf_map_get_memcg(map);
> > ...
> > mem_cgroup_put(memcg);
> >
> > So we should put it *conditionally* as well.
>
> Hi,
>
> No. We could put root_mem_cgroup unconditionally since the root css
> is treated as no reference css. See css_put().
>
> static inline void css_put(struct cgroup_subsys_state *css)
> {
>         // The root memcg’s css has been set with CSS_NO_REF.
>         if (!(css->flags & CSS_NO_REF))
>                 percpu_ref_put(&css->refcnt);
> }

Indeed. I missed that.
The call stack is so deep that I didn't look into it :(
Thanks for the information.

-- 
Regards
Yafang