mm: memcg/slab: fix use after free in obj_cgroup_charge

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.

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.

Link: https://lkml.kernel.org/r/20201028035013.99711-2-songmuchun@bytedance.com
Fixes: bf4f059 ("mm: memcg/slab: obj_cgroup API")
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Acked-by: Roman Gushchin <guro@fb.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Yafang Shao <laoar.shao@gmail.com>
Cc: Chris Down <chris@chrisdown.name>
Cc: Christian Brauner <christian.brauner@ubuntu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/memcontrol.c

@@ -3235,8 +3235,10 @@ 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 (unlikely(!css_tryget(&memcg->css)))
+		goto retry;
 	rcu_read_unlock();
 
 	nr_pages = size >> PAGE_SHIFT;