memcg: fix endless loop caused by mem_cgroup_iter

Hugh has reported an endless loop when the hardlimit reclaim sees the
same group all the time.  This might happen when the reclaim races with
the memcg removal.

shrink_zone
                                                [rmdir root]
  mem_cgroup_iter(root, NULL, reclaim)
    // prev = NULL
    rcu_read_lock()
    mem_cgroup_iter_load
      last_visited = iter->last_visited   // gets root || NULL
      css_tryget(last_visited)            // failed
      last_visited = NULL                 [1]
    memcg = root = __mem_cgroup_iter_next(root, NULL)
    mem_cgroup_iter_update
      iter->last_visited = root;
    reclaim->generation = iter->generation

 mem_cgroup_iter(root, root, reclaim)
   // prev = root
   rcu_read_lock
    mem_cgroup_iter_load
      last_visited = iter->last_visited   // gets root
      css_tryget(last_visited)            // failed
    [1]

The issue seemed to be introduced by commit 5f5781619718 ("memcg: relax
memcg iter caching") which has replaced unconditional css_get/css_put by
css_tryget/css_put for the cached iterator.

This patch fixes the issue by skipping css_tryget on the root of the
tree walk in mem_cgroup_iter_load and symmetrically doesn't release it
in mem_cgroup_iter_update.

Signed-off-by: Michal Hocko <mhocko@suse.cz>
Reported-by: Hugh Dickins <hughd@google.com>
Tested-by: Hugh Dickins <hughd@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Greg Thelen <gthelen@google.com>
Cc: <stable@vger.kernel.org>	[3.10+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index c8336e8..da07784 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -1158,7 +1158,15 @@
 	if (iter->last_dead_count == *sequence) {
 		smp_rmb();
 		position = iter->last_visited;
-		if (position && !css_tryget(&position->css))
+
+		/*
+		 * We cannot take a reference to root because we might race
+		 * with root removal and returning NULL would end up in
+		 * an endless loop on the iterator user level when root
+		 * would be returned all the time.
+		 */
+		if (position && position != root &&
+				!css_tryget(&position->css))
 			position = NULL;
 	}
 	return position;
@@ -1167,9 +1175,11 @@
 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
 				   struct mem_cgroup *last_visited,
 				   struct mem_cgroup *new_position,
+				   struct mem_cgroup *root,
 				   int sequence)
 {
-	if (last_visited)
+	/* root reference counting symmetric to mem_cgroup_iter_load */
+	if (last_visited && last_visited != root)
 		css_put(&last_visited->css);
 	/*
 	 * We store the sequence count from the time @last_visited was
@@ -1244,7 +1254,8 @@
 		memcg = __mem_cgroup_iter_next(root, last_visited);
 
 		if (reclaim) {
-			mem_cgroup_iter_update(iter, last_visited, memcg, seq);
+			mem_cgroup_iter_update(iter, last_visited, memcg, root,
+					seq);
 
 			if (!memcg)
 				iter->generation++;