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.
Fixes: bf4f059954dc ("mm: memcg/slab: obj_cgroup API")
Signed-off-by: Muchun Song <[email protected]>
Acked-by: Roman Gushchin <[email protected]>
---
changelog in v2:
1. Add unlikely and update the commit log suggested by Roman.
mm/memcontrol.c | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)
diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index 8c8b4c3ed5a0..d9cdf899c6fc 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -3221,8 +3221,10 @@ 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;
--
2.20.1
