[Devel] [PATCH RHEL7 COMMIT] mm/memcontrol: offload offline reclaim to background work.

[Konstantin Khorenko]

The commit is pushed to "branch-rh7-3.10.0-1062.12.1.vz7.145.x-ovz" and will appear at https://src.openvz.org/scm/ovz/vzkernel.git
after rh7-3.10.0-1062.12.1.vz7.145.1
------>
commit 6c8429f981c8bbdc4f8d6badb8e209230fb75ec5
Author: Andrey Ryabinin <aryabinin at virtuozzo.com>
Date:   Tue Mar 10 12:30:58 2020 +0300

    mm/memcontrol: offload offline reclaim to background work.
    
    Reclaiming memcg memory during offline may take significant
    amount of time. It's better to offload heavy work to decrease
    cgroup_mutex held time.
    
    https://jira.sw.ru/browse/PSBM-101639
    Signed-off-by: Andrey Ryabinin <aryabinin at virtuozzo.com>
    
    v2: typo fix: s/high_work/offline_reclaim_work/
---
 mm/memcontrol.c | 31 +++++++++++++++++++++----------
 1 file changed, 21 insertions(+), 10 deletions(-)

diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index 4846e7a9bf63b..7e61d1c66c753 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -316,6 +316,7 @@ struct mem_cgroup {
 	/* vmpressure notifications */
 	struct vmpressure vmpressure;
 	struct work_struct high_work;
+	struct work_struct offline_reclaim_work;
 
 	/*
 	 * the counter to account for kernel memory usage.
@@ -6520,6 +6521,22 @@ struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
 }
 EXPORT_SYMBOL(parent_mem_cgroup);
 
+static void offline_reclaim_func(struct work_struct *work)
+{
+	struct mem_cgroup *memcg;
+	int nr_retries = 5;
+
+	memcg = container_of(work, struct mem_cgroup, offline_reclaim_work);
+	lru_add_drain_all();
+
+	while (nr_retries && page_counter_read(&memcg->memory))
+		if (!try_to_free_mem_cgroup_pages(memcg, -1UL, GFP_KERNEL,
+						MEM_CGROUP_RECLAIM_NOSWAP))
+			nr_retries--;
+
+	lru_add_drain();
+}
+
 static void __init mem_cgroup_soft_limit_tree_init(void)
 {
 	struct mem_cgroup_tree_per_node *rtpn;
@@ -6569,6 +6586,7 @@ mem_cgroup_css_alloc(struct cgroup *cont)
 	INIT_LIST_HEAD(&memcg->oom_notify);
 	memcg->move_charge_at_immigrate = 0;
 	INIT_WORK(&memcg->high_work, high_work_func);
+	INIT_WORK(&memcg->offline_reclaim_work, offline_reclaim_func);
 	mutex_init(&memcg->thresholds_lock);
 	spin_lock_init(&memcg->move_lock);
 	vmpressure_init(&memcg->vmpressure);
@@ -6663,16 +6681,7 @@ static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
 
 static void mem_cgroup_free_all(struct mem_cgroup *memcg)
 {
-	int nr_retries = 5;
-
-	lru_add_drain_all();
-
-	while (nr_retries && page_counter_read(&memcg->memory))
-		if (!try_to_free_mem_cgroup_pages(memcg, -1UL, GFP_KERNEL,
-						MEM_CGROUP_RECLAIM_NOSWAP))
-			nr_retries--;
-
-	lru_add_drain();
+	schedule_work(&memcg->offline_reclaim_work);
 }
 
 static void mem_cgroup_css_offline(struct cgroup *cont)
@@ -6723,6 +6732,8 @@ static void mem_cgroup_css_free(struct cgroup *cont)
 {
 	struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
 
+	flush_work(&memcg->offline_reclaim_work);
+
 	/*
 	 * XXX: css_offline() would be where we should reparent all
 	 * memory to prepare the cgroup for destruction.  However,