[PATCH mm-unstable v3 4/5] mm: restart if multiple traversals raced

[Kinsey Ho <kinseyho@xxxxxxxxxx>]

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.

Signed-off-by: Kinsey Ho <kinseyho@xxxxxxxxxx>
---
 include/linux/memcontrol.h |  4 ++--
 mm/memcontrol.c            | 22 ++++++++++++++--------
 2 files changed, 16 insertions(+), 10 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 51b194a4c375..33bd379c738b 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -986,7 +986,7 @@ 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 cgroup_subsys_state *css;
 	struct mem_cgroup *memcg = NULL;
 	struct mem_cgroup *pos = NULL;
 
@@ -999,18 +999,20 @@ struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
 	rcu_read_lock();
 restart:
 	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 +1020,7 @@ struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
 		pos = prev;
 	}
 
-	if (pos)
-		css = &pos->css;
+	css = pos ? &pos->css : NULL;
 
 	for (;;) {
 		css = css_next_descendant_pre(css, &root->css);
@@ -1033,21 +1034,26 @@ struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
 		 * 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
-- 
2.46.0.295.g3b9ea8a38a-goog