[PATCH v3 5/6] mm: flush memcg percpu stats and events before releasing

[Roman Gushchin]

Flush percpu stats and events data to corresponding before releasing
percpu memory.

Although per-cpu stats are never exactly precise, dropping them on
floor regularly may lead to an accumulation of an error. So, it's
safer to flush them before releasing.

To minimize the number of atomic updates, let's sum all stats/events
on all cpus locally, and then make a single update per entry.

Signed-off-by: Roman Gushchin <guro@xxxxxx>
Acked-by: Johannes Weiner <hannes@xxxxxxxxxxx>
---
 mm/memcontrol.c | 52 +++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 52 insertions(+)

diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index 1b5fe826d6d0..0f18bf2afea8 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -2119,6 +2119,56 @@ static void drain_all_stock(struct mem_cgroup *root_memcg)
 	mutex_unlock(&percpu_charge_mutex);
 }
 
+/*
+ * Flush all per-cpu stats and events into atomics.
+ * Try to minimize the number of atomic writes by gathering data from
+ * all cpus locally, and then make one atomic update.
+ * No locking is required, because no one has an access to
+ * the offlined percpu data.
+ */
+static void memcg_flush_offline_percpu(struct mem_cgroup *memcg)
+{
+	struct memcg_vmstats_percpu __percpu *vmstats_percpu;
+	struct lruvec_stat __percpu *lruvec_stat_cpu;
+	struct mem_cgroup_per_node *pn;
+	int cpu, i;
+	long x;
+
+	vmstats_percpu = memcg->vmstats_percpu_offlined;
+
+	for (i = 0; i < MEMCG_NR_STAT; i++) {
+		int nid;
+
+		x = 0;
+		for_each_possible_cpu(cpu)
+			x += per_cpu(vmstats_percpu->stat[i], cpu);
+		if (x)
+			atomic_long_add(x, &memcg->vmstats[i]);
+
+		if (i >= NR_VM_NODE_STAT_ITEMS)
+			continue;
+
+		for_each_node(nid) {
+			pn = mem_cgroup_nodeinfo(memcg, nid);
+			lruvec_stat_cpu = pn->lruvec_stat_cpu_offlined;
+
+			x = 0;
+			for_each_possible_cpu(cpu)
+				x += per_cpu(lruvec_stat_cpu->count[i], cpu);
+			if (x)
+				atomic_long_add(x, &pn->lruvec_stat[i]);
+		}
+	}
+
+	for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
+		x = 0;
+		for_each_possible_cpu(cpu)
+			x += per_cpu(vmstats_percpu->events[i], cpu);
+		if (x)
+			atomic_long_add(x, &memcg->vmevents[i]);
+	}
+}
+
 static int memcg_hotplug_cpu_dead(unsigned int cpu)
 {
 	struct memcg_vmstats_percpu __percpu *vmstats_percpu;
@@ -4618,6 +4668,8 @@ static void percpu_rcu_free(struct rcu_head *rcu)
 	struct mem_cgroup *memcg = container_of(rcu, struct mem_cgroup, rcu);
 	int node;
 
+	memcg_flush_offline_percpu(memcg);
+
 	for_each_node(node) {
 		struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
 
-- 
2.20.1