[PATCH] mm: memcg/slab: fix percpu slab vmstats flushing

[PATCH] mm: memcg/slab: fix percpu slab vmstats flushing

[Roman Gushchin]

Currently slab percpu vmstats are flushed twice: during the memcg
offlining and just before freeing the memcg structure.

Each time percpu counters are summed, added to the atomic counterparts
and propagated up by the cgroup tree.

The problem is that percpu counters are not zeroed after the first
flushing. So every cached percpu value is summed twice. It creates
a small error (up to 32 pages per cpu, but usually less) which
accumulates on parent cgroup level. After creating and destroying
of thousands of child cgroups, slab counter on parent level can
be way off the real value.

For now, let's just stop flushing slab counters on memcg offlining.
It can't be done correctly without scheduling a work on each cpu:
reading and zeroing it during css offlining can race with an
asynchronous update, which doesn't expect values to be changed
underneath.

With this change, slab counters on parent level will become eventually
consistent. Once all dying children are gone, values are correct.
And if not, the error is capped by 32 * NR_CPUS pages per dying
cgroup.

It's not perfect, as slab are reparented, so any updates after
the reparenting will happen on the parent level. It means that
if a slab page was allocated, a counter on child level was bumped,
then the page was reparented and freed, the annihilation of positive
and negative counter values will not happen until the child cgroup is
released. It makes slab counters different from others, and it might
want us to implement flushing in a correct form again.
But it's also a question of performance: scheduling a work on each
cpu isn't free, and it's an open question if the benefit of having
more accurate counters is worth it.

We might also consider flushing all counters on offlining, not only
slab counters.

So let's fix the main problem now: make the slab counters eventually
consistent, so at least the error won't grow with uptime (or more
precisely the number of created and destroyed cgroups). And think
about the accuracy of counters separately.

Signed-off-by: Roman Gushchin <guro@fb.com>
Fixes: bee07b33db78 ("mm: memcontrol: flush percpu slab vmstats on kmem offlining")
Cc: stable@vger.kernel.org
---
 include/linux/mmzone.h |  5 ++---
 mm/memcontrol.c        | 37 +++++++++----------------------------
 2 files changed, 11 insertions(+), 31 deletions(-)

diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
index 89d8ff06c9ce..5334ad8fc7bd 100644
--- a/include/linux/mmzone.h
+++ b/include/linux/mmzone.h
@@ -215,9 +215,8 @@ enum node_stat_item {
 	NR_INACTIVE_FILE,	/*  "     "     "   "       "         */
 	NR_ACTIVE_FILE,		/*  "     "     "   "       "         */
 	NR_UNEVICTABLE,		/*  "     "     "   "       "         */
-	NR_SLAB_RECLAIMABLE,	/* Please do not reorder this item */
-	NR_SLAB_UNRECLAIMABLE,	/* and this one without looking at
-				 * memcg_flush_percpu_vmstats() first. */
+	NR_SLAB_RECLAIMABLE,
+	NR_SLAB_UNRECLAIMABLE,
 	NR_ISOLATED_ANON,	/* Temporary isolated pages from anon lru */
 	NR_ISOLATED_FILE,	/* Temporary isolated pages from file lru */
 	WORKINGSET_NODES,
diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index 601405b207fb..3165db39827a 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -3287,49 +3287,34 @@ static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
 	}
 }
 
-static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg, bool slab_only)
+static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg)
 {
-	unsigned long stat[MEMCG_NR_STAT];
+	unsigned long stat[MEMCG_NR_STAT] = {0};
 	struct mem_cgroup *mi;
 	int node, cpu, i;
-	int min_idx, max_idx;
-
-	if (slab_only) {
-		min_idx = NR_SLAB_RECLAIMABLE;
-		max_idx = NR_SLAB_UNRECLAIMABLE;
-	} else {
-		min_idx = 0;
-		max_idx = MEMCG_NR_STAT;
-	}
-
-	for (i = min_idx; i < max_idx; i++)
-		stat[i] = 0;
 
 	for_each_online_cpu(cpu)
-		for (i = min_idx; i < max_idx; i++)
+		for (i = 0; i < MEMCG_NR_STAT; i++)
 			stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
 
 	for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
-		for (i = min_idx; i < max_idx; i++)
+		for (i = 0; i < MEMCG_NR_STAT; i++)
 			atomic_long_add(stat[i], &mi->vmstats[i]);
 
-	if (!slab_only)
-		max_idx = NR_VM_NODE_STAT_ITEMS;
-
 	for_each_node(node) {
 		struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
 		struct mem_cgroup_per_node *pi;
 
-		for (i = min_idx; i < max_idx; i++)
+		for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
 			stat[i] = 0;
 
 		for_each_online_cpu(cpu)
-			for (i = min_idx; i < max_idx; i++)
+			for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
 				stat[i] += per_cpu(
 					pn->lruvec_stat_cpu->count[i], cpu);
 
 		for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
-			for (i = min_idx; i < max_idx; i++)
+			for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
 				atomic_long_add(stat[i], &pi->lruvec_stat[i]);
 	}
 }
@@ -3403,13 +3388,9 @@ static void memcg_offline_kmem(struct mem_cgroup *memcg)
 		parent = root_mem_cgroup;
 
 	/*
-	 * Deactivate and reparent kmem_caches. Then flush percpu
-	 * slab statistics to have precise values at the parent and
-	 * all ancestor levels. It's required to keep slab stats
-	 * accurate after the reparenting of kmem_caches.
+	 * Deactivate and reparent kmem_caches.
 	 */
 	memcg_deactivate_kmem_caches(memcg, parent);
-	memcg_flush_percpu_vmstats(memcg, true);
 
 	kmemcg_id = memcg->kmemcg_id;
 	BUG_ON(kmemcg_id < 0);
@@ -4913,7 +4894,7 @@ static void mem_cgroup_free(struct mem_cgroup *memcg)
 	 * Flush percpu vmstats and vmevents to guarantee the value correctness
 	 * on parent's and all ancestor levels.
 	 */
-	memcg_flush_percpu_vmstats(memcg, false);
+	memcg_flush_percpu_vmstats(memcg);
 	memcg_flush_percpu_vmevents(memcg);
 	__mem_cgroup_free(memcg);
 }
-- 
2.17.1

Re: [PATCH] mm: memcg/slab: fix percpu slab vmstats flushing

[Johannes Weiner]

On Wed, Dec 18, 2019 at 03:05:01PM -0800, Roman Gushchin wrote:
> Currently slab percpu vmstats are flushed twice: during the memcg
> offlining and just before freeing the memcg structure.

Please explain here why this double flushing is done. You allude to it
below in how it goes wrong, but it'd be better to describe the intent
when describing the current implementation, to be clear about the
trade offs we are making with this patch.

> Each time percpu counters are summed, added to the atomic counterparts
> and propagated up by the cgroup tree.
> 
> The problem is that percpu counters are not zeroed after the first
> flushing. So every cached percpu value is summed twice. It creates
> a small error (up to 32 pages per cpu, but usually less) which
> accumulates on parent cgroup level. After creating and destroying
> of thousands of child cgroups, slab counter on parent level can
> be way off the real value.
> 
> For now, let's just stop flushing slab counters on memcg offlining.
> It can't be done correctly without scheduling a work on each cpu:
> reading and zeroing it during css offlining can race with an
> asynchronous update, which doesn't expect values to be changed
> underneath.
> 
> With this change, slab counters on parent level will become eventually
> consistent. Once all dying children are gone, values are correct.
> And if not, the error is capped by 32 * NR_CPUS pages per dying
> cgroup.
> 
> It's not perfect, as slab are reparented, so any updates after
> the reparenting will happen on the parent level. It means that
> if a slab page was allocated, a counter on child level was bumped,
> then the page was reparented and freed, the annihilation of positive
> and negative counter values will not happen until the child cgroup is
> released. It makes slab counters different from others, and it might
> want us to implement flushing in a correct form again.
> But it's also a question of performance: scheduling a work on each
> cpu isn't free, and it's an open question if the benefit of having
> more accurate counters is worth it.
> 
> We might also consider flushing all counters on offlining, not only
> slab counters.
> 
> So let's fix the main problem now: make the slab counters eventually
> consistent, so at least the error won't grow with uptime (or more
> precisely the number of created and destroyed cgroups). And think
> about the accuracy of counters separately.
> 
> Signed-off-by: Roman Gushchin <guro@fb.com>
> Fixes: bee07b33db78 ("mm: memcontrol: flush percpu slab vmstats on kmem offlining")
> Cc: stable@vger.kernel.org

Other than that, the change looks reasonable to me.

Acked-by: Johannes Weiner <hannes@cmpxchg.org>