Re: [PATCH] mm: memcg: fix use after free in mem_cgroup_iter()

[Shakeel Butt]

On Sun, Jul 22, 2018 at 11:44 PM Michal Hocko <mhocko@xxxxxxxxxx> wrote:
>
> On Thu 19-07-18 09:23:10, Shakeel Butt wrote:
> > On Thu, Jul 19, 2018 at 3:43 AM Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> > >
> > > [CC Andrew]
> > >
> > > On Thu 19-07-18 18:06:47, Jing Xia wrote:
> > > > It was reported that a kernel crash happened in mem_cgroup_iter(),
> > > > which can be triggered if the legacy cgroup-v1 non-hierarchical
> > > > mode is used.
> > > >
> > > > Unable to handle kernel paging request at virtual address 6b6b6b6b6b6b8f
> > > > ......
> > > > Call trace:
> > > >   mem_cgroup_iter+0x2e0/0x6d4
> > > >   shrink_zone+0x8c/0x324
> > > >   balance_pgdat+0x450/0x640
> > > >   kswapd+0x130/0x4b8
> > > >   kthread+0xe8/0xfc
> > > >   ret_from_fork+0x10/0x20
> > > >
> > > >   mem_cgroup_iter():
> > > >       ......
> > > >       if (css_tryget(css))    <-- crash here
> > > >           break;
> > > >       ......
> > > >
> > > > The crashing reason is that mem_cgroup_iter() uses the memcg object
> > > > whose pointer is stored in iter->position, which has been freed before
> > > > and filled with POISON_FREE(0x6b).
> > > >
> > > > And the root cause of the use-after-free issue is that
> > > > invalidate_reclaim_iterators() fails to reset the value of
> > > > iter->position to NULL when the css of the memcg is released in non-
> > > > hierarchical mode.
> > >
> > > Well, spotted!
> > >
> > > I suspect
> > > Fixes: 6df38689e0e9 ("mm: memcontrol: fix possible memcg leak due to interrupted reclaim")
> > >
> > > but maybe it goes further into past. I also suggest
> > > Cc: stable
> > >
> > > even though the non-hierarchical mode is strongly discouraged.
> >
> > Why not set root_mem_cgroup's use_hierarchy to true by default on
> > init? If someone wants non-hierarchical mode, they can explicitly set
> > it to false.
>
> We do not change defaults under users feet usually.

Then how non-hierarchical mode is being discouraged currently? I don't
see any comments in the docs.

Shakeel