memcg: reparent charges of children before processing parent

commit 4fb1a86f upstream.

Sometimes the cleanup after memcg hierarchy testing gets stuck in
mem_cgroup_reparent_charges(), unable to bring non-kmem usage down to 0.

There may turn out to be several causes, but a major cause is this: the
workitem to offline parent can get run before workitem to offline child;
parent's mem_cgroup_reparent_charges() circles around waiting for the
child's pages to be reparented to its lrus, but it's holding
cgroup_mutex which prevents the child from reaching its
mem_cgroup_reparent_charges().

Further testing showed that an ordered workqueue for cgroup_destroy_wq
is not always good enough: percpu_ref_kill_and_confirm's call_rcu_sched
stage on the way can mess up the order before reaching the workqueue.

Instead, when offlining a memcg, call mem_cgroup_reparent_charges() on
all its children (and grandchildren, in the correct order) to have their
charges reparented first.

[The version for 3.10.34 (or perhaps now 3.10.35) is this below.
Yes, more differences, and the old mem_cgroup_reparent_charges line
is intentionally left in for 3.10 whereas it was removed for 3.12+:
that's because the css/cgroup iterator changed in between, it used
not to supply the root of the subtree, but nowadays it does - Hugh]


Fixes: e5fca243 ("cgroup: use a dedicated workqueue for cgroup destruction")
Signed-off-by: Filipe Brandenburger <filbranden@google.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Tejun Heo <tj@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

mm/memcontrol.c

@@ -6326,9 +6326,23 @@ static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
 static void mem_cgroup_css_offline(struct cgroup *cont)
 {
 	struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
+	struct cgroup *iter;
 
 	mem_cgroup_invalidate_reclaim_iterators(memcg);
+
+	/*
+	 * This requires that offlining is serialized.  Right now that is
+	 * guaranteed because css_killed_work_fn() holds the cgroup_mutex.
+	 */
+	rcu_read_lock();
+	cgroup_for_each_descendant_post(iter, cont) {
+		rcu_read_unlock();
+		mem_cgroup_reparent_charges(mem_cgroup_from_cont(iter));
+		rcu_read_lock();
+	}
+	rcu_read_unlock();
 	mem_cgroup_reparent_charges(memcg);
+
 	mem_cgroup_destroy_all_caches(memcg);
 }