mm: restart if multiple traversals raced

Currently, if multiple reclaimers raced on the same position, the
reclaimers which detect the race will still reclaim from the same memcg. 
Instead, the reclaimers which detect the race should move on to the next
memcg in the hierarchy.

So, in the case where multiple traversals race, jump back to the start of
the mem_cgroup_iter() function to find the next memcg in the hierarchy to
reclaim from.

Link: https://lkml.kernel.org/r/[email protected]
Reported-by: [email protected]
Closes: https://lore.kernel.org/[email protected]/
Signed-off-by: Kinsey Ho <[email protected]>
Reviewed-by: T.J. Mercier <[email protected]>
Cc: Johannes Weiner <[email protected]>
Cc: Michal Hocko <[email protected]>
Cc: Michal Koutný <[email protected]>
Cc: Muchun Song <[email protected]>
Cc: Roman Gushchin <[email protected]>
Cc: Shakeel Butt <[email protected]>
Cc: Tejun Heo <[email protected]>
Cc: Yosry Ahmed <[email protected]>
Cc: Zefan Li <[email protected]>
Cc: Hugh Dickins <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>

2 files changed, 19 insertions, 11 deletions

diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h
index fe05fdb92779..2ef94c74847d 100644
--- a/include/linux/memcontrol.h
+++ b/include/linux/memcontrol.h
@@ -57,7 +57,7 @@ enum memcg_memory_event {
 
 struct mem_cgroup_reclaim_cookie {
 	pg_data_t *pgdat;
-	unsigned int generation;
+	int generation;
 };
 
 #ifdef CONFIG_MEMCG
@@ -78,7 +78,7 @@ struct lruvec_stats;
 struct mem_cgroup_reclaim_iter {
 	struct mem_cgroup *position;
 	/* scan generation, increased every round-trip */
-	unsigned int generation;
+	atomic_t generation;
 };
 
 /*
diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index 38e3251ed482..36cbc7730745 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -986,8 +986,8 @@ struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
 				   struct mem_cgroup_reclaim_cookie *reclaim)
 {
 	struct mem_cgroup_reclaim_iter *iter;
-	struct cgroup_subsys_state *css = NULL;
-	struct mem_cgroup *memcg = NULL;
+	struct cgroup_subsys_state *css;
+	struct mem_cgroup *memcg;
 	struct mem_cgroup *pos = NULL;
 
 	if (mem_cgroup_disabled())
@@ -998,19 +998,23 @@ struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
 
 	rcu_read_lock();
 restart:
+	memcg = NULL;
+
 	if (reclaim) {
+		int gen;
 		struct mem_cgroup_per_node *mz;
 
 		mz = root->nodeinfo[reclaim->pgdat->node_id];
 		iter = &mz->iter;
+		gen = atomic_read(&iter->generation);
 
 		/*
 		 * On start, join the current reclaim iteration cycle.
 		 * Exit when a concurrent walker completes it.
 		 */
 		if (!prev)
-			reclaim->generation = iter->generation;
-		else if (reclaim->generation != iter->generation)
+			reclaim->generation = gen;
+		else if (reclaim->generation != gen)
 			goto out_unlock;
 
 		pos = READ_ONCE(iter->position);
@@ -1018,8 +1022,7 @@ restart:
 		pos = prev;
 	}
 
-	if (pos)
-		css = &pos->css;
+	css = pos ? &pos->css : NULL;
 
 	for (;;) {
 		css = css_next_descendant_pre(css, &root->css);
@@ -1033,21 +1036,26 @@ restart:
 		 * and kicking, and don't take an extra reference.
 		 */
 		if (css == &root->css || css_tryget(css)) {
-			memcg = mem_cgroup_from_css(css);
 			break;
 		}
 	}
 
+	memcg = mem_cgroup_from_css(css);
+
 	if (reclaim) {
 		/*
 		 * The position could have already been updated by a competing
 		 * thread, so check that the value hasn't changed since we read
 		 * it to avoid reclaiming from the same cgroup twice.
 		 */
-		(void)cmpxchg(&iter->position, pos, memcg);
+		if (cmpxchg(&iter->position, pos, memcg) != pos) {
+			if (css && css != &root->css)
+				css_put(css);
+			goto restart;
+		}
 
 		if (!memcg) {
-			iter->generation++;
+			atomic_inc(&iter->generation);
 
 			/*
 			 * Reclaimers share the hierarchy walk, and a