[Devel] [PATCH rh7] mm/memcg: Don't stuck in mem_cgroup_reparent_charges() forever

[Andrey Ryabinin]

mem_cgroup_reparent_charges() supposed bring down all non-kmem charges.
Sometimes this doesn't happen due to some bug in memcg accounting or
leak, etc. When this happens, mem_cgroup_reparent_charges() loops
forever under the global cgroup_mutex, causing a lot of troubles for the
whole system. Instead of endless loop, make several retries, WARN() and
break the loop if reparenting was unsuccessful.

https://jira.sw.ru/browse/PSBM-86092
Signed-off-by: Andrey Ryabinin <aryabinin at virtuozzo.com>
---
 mm/memcontrol.c | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index 59cf47972f9e..755c09e050a7 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -4010,6 +4010,7 @@ static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
  */
 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
 {
+	int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
 	int node, zid;
 
 	do {
@@ -4030,6 +4031,11 @@ static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
 		memcg_oom_recover(memcg);
 		cond_resched();
 
+		if (WARN(!--nr_retries, "memory %ld > kmem %ld\n",
+				page_counter_read(&memcg->memory),
+				page_counter_read(&memcg->kmem)))
+			break;
+
 		/*
 		 * Kernel memory may not necessarily be trackable to a specific
 		 * process. So they are not migrated, and therefore we can't
@@ -4044,6 +4050,8 @@ static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
 		 */
 	} while (page_counter_read(&memcg->memory) -
 		 page_counter_read(&memcg->kmem) > 0);
+
+	WARN_ON(!nr_retries);
 }
 
 /*
-- 
2.16.4