On Tue, Oct 27, 2020 at 04:02:53PM +0800, Muchun Song wrote:
> The rcu_read_lock/unlock only can guarantee that the memcg will
> not be freed, but it cannot guarantee the success of css_get to
> memcg. This can be happened when we reparent the memcg. So using
> css_tryget instead of css_get.
Hm, I really doubt that it can be reproduced in the real life, but
I think you're formally correct.
If the whole process of a cgroup offlining is completed between reading
a objcg->memcg pointer and bumping the css reference on another CPU,
and there are exactly 0 external references to this memory cgroup
(how we get to the obj_cgroup_charge() then?), css_get() can change
the ref counter from 0 back to 1.
Is this a good description of the problem?
If so, I think it's worth a more detailed description and mentioning that
it's not a real-life bug. It also probably doesn't deserve a stable@ backport.
>
> Fixes: bf4f059954dc ("mm: memcg/slab: obj_cgroup API")
> Signed-off-by: Muchun Song <songmuchun@xxxxxxxxxxxxx>
> ---
> mm/memcontrol.c | 4 +++-
> 1 file changed, 3 insertions(+), 1 deletion(-)
>
> diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> index fcbd79c5023e..c0c27bee23ad 100644
> --- a/mm/memcontrol.c
> +++ b/mm/memcontrol.c
> @@ -3223,8 +3223,10 @@ int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
> * independently later.
> */
> rcu_read_lock();
> +retry:
> memcg = obj_cgroup_memcg(objcg);
> - css_get(&memcg->css);
> + if (!css_tryget(&memcg->css))
if (unlikely(!css_tryget(&memcg->css)))
^^^^^^^^
maybe?
> + goto retry;
Otherwise the fix looks good to me.
Please, feel free to add
Acked-by: Roman Gushchin <guro@xxxxxx>
after extending the commit description.
Thanks!
