mm: memcontrol: use special workqueue for creating per-memcg caches

synchronize_sched() is a heavy operation and calling it per each cache owned by a memory cgroup being destroyed may take quite some time. What is worse, it's currently called under the slab_mutex, stalling all works doing cache creation/destruction.

Actually, there isn't much point in calling synchronize_sched() for each cache - it's enough to call it just once - after setting cpu_partial for all caches and before shrinking them. This way, we can also move it out of the slab_mutex, which we have to hold for iterating over the slab cache list.

Link https://bugzilla.kernel.org/show_bug.cgi?id=172991

Signed-off-by Vladimir Davydov <[email protected]>

Reported-by Doug Smythies <[email protected]>

Cc Christoph Lameter <[email protected]>

Cc David Rientjes <[email protected]>

Cc Johannes Weiner <[email protected]>

Cc Joonsoo Kim <[email protected]>

Cc Michal Hocko <[email protected]>

Cc Pekka Enberg <[email protected]>

Creating a lot of cgroups at the same time might stall all worker threads with kmem cache creation works, because kmem cache creation is done with the slab_mutex held. To prevent that from happening, let's use a special workqueue for kmem cache creation with max in-flight work items equal to 1.

Link https://bugzilla.kernel.org/show_bug.cgi?id=172981

Signed-off-by Vladimir Davydov <[email protected]>

Reported-by Doug Smythies <[email protected]>

Cc Christoph Lameter <[email protected]>

Cc David Rientjes <[email protected]>

Cc Johannes Weiner <[email protected]>

Cc Joonsoo Kim <[email protected]>

Cc Michal Hocko <[email protected]>

Cc Pekka Enberg <[email protected]>

---
 mm/memcontrol.c | 15 ++++++++++++++-
 1 file changed, 14 insertions(+), 1 deletion(-)

mm/memcontrol.c +14 -1

index 4be518d4e68a..c1efe59e3a20 100644
@@ -2175,6 +2175,8 @@ struct memcg_kmem_cache_create_work {
 	struct work_struct work;
 };
 
+static struct workqueue_struct *memcg_kmem_cache_create_wq;
+
 static void memcg_kmem_cache_create_func(struct work_struct *w)
 {
 	struct memcg_kmem_cache_create_work *cw =

@@ -2206,7 +2208,7 @@ static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
 	cw->cachep = cachep;
 	INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
 
-	schedule_work(&cw->work);
+	queue_work(memcg_kmem_cache_create_wq, &cw->work);
 }
 
 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,

@@ -5794,6 +5796,17 @@ static int __init mem_cgroup_init(void)
 {
 	int cpu, node;
 
+#ifndef CONFIG_SLOB
+	/*
+	 * Kmem cache creation is mostly done with the slab_mutex held,
+	 * so use a special workqueue to avoid stalling all worker
+	 * threads in case lots of cgroups are created simultaneously.
+	 */
+	memcg_kmem_cache_create_wq =
+		alloc_workqueue("memcg_kmem_cache_create", 0, 1);
+	BUG_ON(!memcg_kmem_cache_create_wq);
+#endif
+
 	hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
 
 	for_each_possible_cpu(cpu)

Signature · 1 line

2.1.4

Quoted 6 lines · Vladimir Davydov

Creating a lot of cgroups at the same time might stall all worker threads with kmem cache creation works, because kmem cache creation is done with the slab_mutex held. To prevent that from happening, let's use a special workqueue for kmem cache creation with max in-flight work items equal to 1.

Quoted 1 line

Link https://bugzilla.kernel.org/show_bug.cgi?id=172981

This looks like a regression but I am not really sure I understand what has caused it. We had the WQ based cache creation since kmem was introduced more or less. So is it 801faf0db894 ("mm/slab: lockless decision to grow cache") which was pointed by bisection that changed the timing resp. relaxed the cache creation to the point that would allow this runaway?

This would be really useful for the stable backport consideration.

Also, if I understand the fix correctly, now we do limit the number of workers to 1 thread. Is this really what we want? Wouldn't it be possible that few memcgs could starve others fromm having their cache created? What would be the result, missed charges?

Quoted 7 lines · Michal Hocko

Quoted 6 lines · Vladimir Davydov

Creating a lot of cgroups at the same time might stall all worker threads with kmem cache creation works, because kmem cache creation is done with the slab_mutex held. To prevent that from happening, let's use a special workqueue for kmem cache creation with max in-flight work items equal to 1.

Quoted 8 lines

Quoted 1 line

Link https://bugzilla.kernel.org/show_bug.cgi?id=172981

This looks like a regression but I am not really sure I understand what has caused it. We had the WQ based cache creation since kmem was introduced more or less. So is it 801faf0db894 ("mm/slab: lockless decision to grow cache") which was pointed by bisection that changed the timing resp. relaxed the cache creation to the point that would allow this runaway?

It is in case of SLAB. For SLUB the issue was caused by commit 81ae6d03952c ("mm/slub.c: replace kick_all_cpus_sync() with synchronize_sched() in kmem_cache_shrink()").

Quoted 7 lines

This would be really useful for the stable backport consideration.

Also, if I understand the fix correctly, now we do limit the number of workers to 1 thread. Is this really what we want? Wouldn't it be possible that few memcgs could starve others fromm having their cache created? What would be the result, missed charges?

Now kmem caches are created in FIFO order, i.e. if one memcg called kmem_cache_alloc on a non-existent cache before another, it will be served first.

Since the number of caches that can be created by a single memcg is obviously limited,

I don't see any possibility of starvation.

Actually, this patch doesn't introduce any functional changes regarding the order in which kmem caches are created, as the work function holds the global slab_mutex during its whole runtime anyway. We only avoid creating a thread per each work by making the queue single-threaded.

Quoted 17 lines · Vladimir Davydov

Quoted 7 lines · Michal Hocko

Quoted 6 lines · Vladimir Davydov

Creating a lot of cgroups at the same time might stall all worker threads with kmem cache creation works, because kmem cache creation is done with the slab_mutex held. To prevent that from happening, let's use a special workqueue for kmem cache creation with max in-flight work items equal to 1.

Quoted 8 lines

Quoted 1 line

Link https://bugzilla.kernel.org/show_bug.cgi?id=172981

This looks like a regression but I am not really sure I understand what has caused it. We had the WQ based cache creation since kmem was introduced more or less. So is it 801faf0db894 ("mm/slab: lockless decision to grow cache") which was pointed by bisection that changed the timing resp. relaxed the cache creation to the point that would allow this runaway?

Quoted 3 lines

It is in case of SLAB. For SLUB the issue was caused by commit 81ae6d03952c ("mm/slub.c: replace kick_all_cpus_sync() with synchronize_sched() in kmem_cache_shrink()").

OK, thanks for the confirmation. This would be useful in the changelog imho.

Quoted 8 lines

Quoted 7 lines

This would be really useful for the stable backport consideration.

Also, if I understand the fix correctly, now we do limit the number of workers to 1 thread. Is this really what we want? Wouldn't it be possible that few memcgs could starve others fromm having their cache created? What would be the result, missed charges?

Quoted 3 lines

Now kmem caches are created in FIFO order, i.e. if one memcg called kmem_cache_alloc on a non-existent cache before another, it will be served first.

I do not see where this FIFO is guaranteed. __memcg_schedule_kmem_cache_create doesn't seem to be using ordered WQ.

Quoted 1 line

From 10f5f126800912c6a4b78a8b615138c1322694ad Mon Sep 17 00:00:00 2001

From: Vladimir Davydov <[email protected]>
Date: Sat, 1 Oct 2016 16:39:09 +0300
Subject: [PATCH] mm: memcontrol: use special workqueue for creating per-memcg
caches

Creating a lot of cgroups at the same time might stall all worker threads with kmem cache creation works, because kmem cache creation is done with the slab_mutex held. The problem was amplified by commits 801faf0db894 ("mm/slab: lockless decision to grow cache") in case of SLAB and 81ae6d03952c ("mm/slub.c: replace kick_all_cpus_sync() with synchronize_sched() in kmem_cache_shrink()") in case of SLUB, which increased the maximal time the slab_mutex can be held.

To prevent that from happening, let's use a special ordered single threaded workqueue for kmem cache creation. This shouldn't introduce any functional changes regarding how kmem caches are created, as the work function holds the global slab_mutex during its whole runtime anyway, making it impossible to run more than one work at a time. By using a single threaded workqueue, we just avoid creating a thread per each work. Ordering is required to avoid a situation when a cgroup's work is put off indefinitely because there are other cgroups to serve, in other words to guarantee fairness.

Link https://bugzilla.kernel.org/show_bug.cgi?id=172981

Signed-off-by Vladimir Davydov <[email protected]>

Reported-by Doug Smythies <[email protected]>

Cc Christoph Lameter <[email protected]>

Cc David Rientjes <[email protected]>

Cc Johannes Weiner <[email protected]>

Cc Joonsoo Kim <[email protected]>

Cc Michal Hocko <[email protected]>

Cc Pekka Enberg <[email protected]>

mm/memcontrol.c +14 -1

index 4be518d4e68a..8d753d87ca37 100644
@@ -2175,6 +2175,8 @@ struct memcg_kmem_cache_create_work {
 	struct work_struct work;
 };
 
+static struct workqueue_struct *memcg_kmem_cache_create_wq;
+
 static void memcg_kmem_cache_create_func(struct work_struct *w)
 {
 	struct memcg_kmem_cache_create_work *cw =

@@ -2206,7 +2208,7 @@ static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
 	cw->cachep = cachep;
 	INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
 
-	schedule_work(&cw->work);
+	queue_work(memcg_kmem_cache_create_wq, &cw->work);
 }
 
 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,

@@ -5794,6 +5796,17 @@ static int __init mem_cgroup_init(void)
 {
 	int cpu, node;
 
+#ifndef CONFIG_SLOB
+	/*
+	 * Kmem cache creation is mostly done with the slab_mutex held,
+	 * so use a special workqueue to avoid stalling all worker
+	 * threads in case lots of cgroups are created simultaneously.
+	 */
+	memcg_kmem_cache_create_wq =
+		alloc_ordered_workqueue("memcg_kmem_cache_create", 0);
+	BUG_ON(!memcg_kmem_cache_create_wq);
+#endif
+
 	hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
 
 	for_each_possible_cpu(cpu)

Creating a lot of cgroups at the same time might stall all worker threads with kmem cache creation works, because kmem cache creation is done with the slab_mutex held. The problem was amplified by commits 801faf0db894 ("mm/slab: lockless decision to grow cache") in case of SLAB and 81ae6d03952c ("mm/slub.c: replace kick_all_cpus_sync() with synchronize_sched() in kmem_cache_shrink()") in case of SLUB, which increased the maximal time the slab_mutex can be held.

To prevent that from happening, let's use a special ordered single threaded workqueue for kmem cache creation. This shouldn't introduce any functional changes regarding how kmem caches are created, as the work function holds the global slab_mutex during its whole runtime anyway, making it impossible to run more than one work at a time. By using a single threaded workqueue, we just avoid creating a thread per each work. Ordering is required to avoid a situation when a cgroup's work is put off indefinitely because there are other cgroups to serve, in other words to guarantee fairness.

Quoted 8 lines · On Tue, 4 Oct 2016 16:14:17 +0300 Vladimir Davydov <[email protected]>…

Creating a lot of cgroups at the same time might stall all worker threads with kmem cache creation works, because kmem cache creation is done with the slab_mutex held. The problem was amplified by commits 801faf0db894 ("mm/slab: lockless decision to grow cache") in case of SLAB and 81ae6d03952c ("mm/slub.c: replace kick_all_cpus_sync() with synchronize_sched() in kmem_cache_shrink()") in case of SLUB, which increased the maximal time the slab_mutex can be held.

Quoted 9 lines

To prevent that from happening, let's use a special ordered single threaded workqueue for kmem cache creation. This shouldn't introduce any functional changes regarding how kmem caches are created, as the work function holds the global slab_mutex during its whole runtime anyway, making it impossible to run more than one work at a time. By using a single threaded workqueue, we just avoid creating a thread per each work. Ordering is required to avoid a situation when a cgroup's work is put off indefinitely because there are other cgroups to serve, in other words to guarantee fairness.

I'm having trouble working out the urgency of this patch?