[PATCH 4/5] mm: release per-node memcg percpu data prematurely

[Roman Gushchin]

Similar to memcg-level statistics, per-node data isn't expected
to be hot after cgroup removal. Switching over to atomics and
prematurely releasing percpu data helps to reduce the memory
footprint of dying cgroups.

Signed-off-by: Roman Gushchin <guro@xxxxxx>
---
 include/linux/memcontrol.h |  1 +
 mm/memcontrol.c            | 24 +++++++++++++++++++++++-
 2 files changed, 24 insertions(+), 1 deletion(-)

diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h
index 569337514230..f296693d102b 100644
--- a/include/linux/memcontrol.h
+++ b/include/linux/memcontrol.h
@@ -127,6 +127,7 @@ struct mem_cgroup_per_node {
 	struct lruvec		lruvec;
 
 	struct lruvec_stat __rcu /* __percpu */ *lruvec_stat_cpu;
+	struct lruvec_stat __percpu *lruvec_stat_cpu_offlined;
 	atomic_long_t		lruvec_stat[NR_VM_NODE_STAT_ITEMS];
 
 	unsigned long		lru_zone_size[MAX_NR_ZONES][NR_LRU_LISTS];
diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index 8c55954e6f23..18e863890392 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -4459,7 +4459,7 @@ static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
 	if (!pn)
 		return;
 
-	free_percpu(pn->lruvec_stat_cpu);
+	WARN_ON_ONCE(pn->lruvec_stat_cpu != NULL);
 	kfree(pn);
 }
 
@@ -4615,7 +4615,17 @@ static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
 static void mem_cgroup_free_percpu(struct rcu_head *rcu)
 {
 	struct mem_cgroup *memcg = container_of(rcu, struct mem_cgroup, rcu);
+	int node;
+
+	for_each_node(node) {
+		struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
 
+		if (!pn)
+			continue;
+
+		free_percpu(pn->lruvec_stat_cpu_offlined);
+		WARN_ON_ONCE(pn->lruvec_stat_cpu != NULL);
+	}
 	free_percpu(memcg->vmstats_percpu_offlined);
 	WARN_ON_ONCE(memcg->vmstats_percpu);
 
@@ -4624,6 +4634,18 @@ static void mem_cgroup_free_percpu(struct rcu_head *rcu)
 
 static void mem_cgroup_offline_percpu(struct mem_cgroup *memcg)
 {
+	int node;
+
+	for_each_node(node) {
+		struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
+
+		if (!pn)
+			continue;
+
+		pn->lruvec_stat_cpu_offlined = (struct lruvec_stat __percpu *)
+			rcu_dereference(pn->lruvec_stat_cpu);
+		rcu_assign_pointer(pn->lruvec_stat_cpu, NULL);
+	}
 	memcg->vmstats_percpu_offlined = (struct memcg_vmstats_percpu __percpu*)
 		rcu_dereference(memcg->vmstats_percpu);
 	rcu_assign_pointer(memcg->vmstats_percpu, NULL);
-- 
2.20.1