[PATCH] mm: memcontrol: don't batch updates of local VM stats and events

[PATCH] mm: memcontrol: don't batch updates of local VM stats and events

[Johannes Weiner]

The kernel test robot noticed a 26% will-it-scale pagefault regression
from commit 42a300353577 ("mm: memcontrol: fix recursive statistics
correctness & scalabilty"). This appears to be caused by bouncing the
additional cachelines from the new hierarchical statistics counters.

We can fix this by getting rid of the batched local counters instead.

Originally, there were *only* group-local counters, and they were
fully maintained per cpu. A reader of a stats file high up in the
cgroup tree would have to walk the entire subtree and collect each
level's per-cpu counters to get the recursive view. This was
prohibitively expensive, and so we switched to per-cpu batched updates
of the local counters during a983b5ebee57 ("mm: memcontrol: fix
excessive complexity in memory.stat reporting"), reducing the
complexity from nr_subgroups * nr_cpus to nr_subgroups.

With growing machines and cgroup trees, the tree walk itself became
too expensive for monitoring top-level groups, and this is when the
culprit patch added hierarchy counters on each cgroup level. When the
per-cpu batch size would be reached, both the local and the hierarchy
counters would get batch-updated from the per-cpu delta simultaneously.

This makes local and hierarchical counter reads blazingly fast, but it
unfortunately makes the write-side too cache line intense.

Since local counter reads were never a problem - we only centralized
them to accelerate the hierarchy walk - and use of the local counters
are becoming rarer due to replacement with hierarchical views (ongoing
rework in the page reclaim and workingset code), we can make those
local counters unbatched per-cpu counters again.

The scheme will then be as such:

   when a memcg statistic changes, the writer will:
   - update the local counter (per-cpu)
   - update the batch counter (per-cpu). If the batch is full:
   - spill the batch into the group's atomic_t
   - spill the batch into all ancestors' atomic_ts
   - empty out the batch counter (per-cpu)

   when a local memcg counter is read, the reader will:
   - collect the local counter from all cpus

   when a hiearchy memcg counter is read, the reader will:
   - read the atomic_t

We might be able to simplify this further and make the recursive
counters unbatched per-cpu counters as well (batch upward propagation,
but leave per-cpu collection to the readers), but that will require a
more in-depth analysis and testing of all the callsites. Deal with the
immediate regression for now.

Fixes: 42a300353577 ("mm: memcontrol: fix recursive statistics correctness & scalabilty")
Reported-by: kernel test robot <rong.a.chen@intel.com>
Tested-by: kernel test robot <rong.a.chen@intel.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
---
 include/linux/memcontrol.h | 26 ++++++++++++++++--------
 mm/memcontrol.c            | 41 ++++++++++++++++++++++++++------------
 2 files changed, 46 insertions(+), 21 deletions(-)

diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h
index bc74d6a4407c..2d23ae7bd36d 100644
--- a/include/linux/memcontrol.h
+++ b/include/linux/memcontrol.h
@@ -126,9 +126,12 @@ struct memcg_shrinker_map {
 struct mem_cgroup_per_node {
 	struct lruvec		lruvec;
 
+	/* Legacy local VM stats */
+	struct lruvec_stat __percpu *lruvec_stat_local;
+
+	/* Subtree VM stats (batched updates) */
 	struct lruvec_stat __percpu *lruvec_stat_cpu;
 	atomic_long_t		lruvec_stat[NR_VM_NODE_STAT_ITEMS];
-	atomic_long_t		lruvec_stat_local[NR_VM_NODE_STAT_ITEMS];
 
 	unsigned long		lru_zone_size[MAX_NR_ZONES][NR_LRU_LISTS];
 
@@ -274,17 +277,18 @@ struct mem_cgroup {
 	atomic_t		moving_account;
 	struct task_struct	*move_lock_task;
 
-	/* memory.stat */
+	/* Legacy local VM stats and events */
+	struct memcg_vmstats_percpu __percpu *vmstats_local;
+
+	/* Subtree VM stats and events (batched updates) */
 	struct memcg_vmstats_percpu __percpu *vmstats_percpu;
 
 	MEMCG_PADDING(_pad2_);
 
 	atomic_long_t		vmstats[MEMCG_NR_STAT];
-	atomic_long_t		vmstats_local[MEMCG_NR_STAT];
-
 	atomic_long_t		vmevents[NR_VM_EVENT_ITEMS];
-	atomic_long_t		vmevents_local[NR_VM_EVENT_ITEMS];
 
+	/* memory.events */
 	atomic_long_t		memory_events[MEMCG_NR_MEMORY_EVENTS];
 
 	unsigned long		socket_pressure;
@@ -576,7 +580,11 @@ static inline unsigned long memcg_page_state(struct mem_cgroup *memcg, int idx)
 static inline unsigned long memcg_page_state_local(struct mem_cgroup *memcg,
 						   int idx)
 {
-	long x = atomic_long_read(&memcg->vmstats_local[idx]);
+	long x = 0;
+	int cpu;
+
+	for_each_possible_cpu(cpu)
+		x += per_cpu(memcg->vmstats_local->stat[idx], cpu);
 #ifdef CONFIG_SMP
 	if (x < 0)
 		x = 0;
@@ -650,13 +658,15 @@ static inline unsigned long lruvec_page_state_local(struct lruvec *lruvec,
 						    enum node_stat_item idx)
 {
 	struct mem_cgroup_per_node *pn;
-	long x;
+	long x = 0;
+	int cpu;
 
 	if (mem_cgroup_disabled())
 		return node_page_state(lruvec_pgdat(lruvec), idx);
 
 	pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
-	x = atomic_long_read(&pn->lruvec_stat_local[idx]);
+	for_each_possible_cpu(cpu)
+		x += per_cpu(pn->lruvec_stat_local->count[idx], cpu);
 #ifdef CONFIG_SMP
 	if (x < 0)
 		x = 0;
diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index e50a2db5b4ff..8d42e5c7bf37 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -700,11 +700,12 @@ void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
 	if (mem_cgroup_disabled())
 		return;
 
+	__this_cpu_add(memcg->vmstats_local->stat[idx], val);
+
 	x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
 	if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
 		struct mem_cgroup *mi;
 
-		atomic_long_add(x, &memcg->vmstats_local[idx]);
 		for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
 			atomic_long_add(x, &mi->vmstats[idx]);
 		x = 0;
@@ -754,11 +755,12 @@ void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
 	__mod_memcg_state(memcg, idx, val);
 
 	/* Update lruvec */
+	__this_cpu_add(pn->lruvec_stat_local->count[idx], val);
+
 	x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
 	if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
 		struct mem_cgroup_per_node *pi;
 
-		atomic_long_add(x, &pn->lruvec_stat_local[idx]);
 		for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
 			atomic_long_add(x, &pi->lruvec_stat[idx]);
 		x = 0;
@@ -780,11 +782,12 @@ void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
 	if (mem_cgroup_disabled())
 		return;
 
+	__this_cpu_add(memcg->vmstats_local->events[idx], count);
+
 	x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
 	if (unlikely(x > MEMCG_CHARGE_BATCH)) {
 		struct mem_cgroup *mi;
 
-		atomic_long_add(x, &memcg->vmevents_local[idx]);
 		for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
 			atomic_long_add(x, &mi->vmevents[idx]);
 		x = 0;
@@ -799,7 +802,12 @@ static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
 
 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
 {
-	return atomic_long_read(&memcg->vmevents_local[event]);
+	long x = 0;
+	int cpu;
+
+	for_each_possible_cpu(cpu)
+		x += per_cpu(memcg->vmstats_local->events[event], cpu);
+	return x;
 }
 
 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
@@ -2200,11 +2208,9 @@ static int memcg_hotplug_cpu_dead(unsigned int cpu)
 			long x;
 
 			x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
-			if (x) {
-				atomic_long_add(x, &memcg->vmstats_local[i]);
+			if (x)
 				for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
 					atomic_long_add(x, &memcg->vmstats[i]);
-			}
 
 			if (i >= NR_VM_NODE_STAT_ITEMS)
 				continue;
@@ -2214,12 +2220,10 @@ static int memcg_hotplug_cpu_dead(unsigned int cpu)
 
 				pn = mem_cgroup_nodeinfo(memcg, nid);
 				x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
-				if (x) {
-					atomic_long_add(x, &pn->lruvec_stat_local[i]);
+				if (x)
 					do {
 						atomic_long_add(x, &pn->lruvec_stat[i]);
 					} while ((pn = parent_nodeinfo(pn, nid)));
-				}
 			}
 		}
 
@@ -2227,11 +2231,9 @@ static int memcg_hotplug_cpu_dead(unsigned int cpu)
 			long x;
 
 			x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
-			if (x) {
-				atomic_long_add(x, &memcg->vmevents_local[i]);
+			if (x)
 				for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
 					atomic_long_add(x, &memcg->vmevents[i]);
-			}
 		}
 	}
 
@@ -4492,8 +4494,15 @@ static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
 	if (!pn)
 		return 1;
 
+	pn->lruvec_stat_local = alloc_percpu(struct lruvec_stat);
+	if (!pn->lruvec_stat_local) {
+		kfree(pn);
+		return 1;
+	}
+
 	pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
 	if (!pn->lruvec_stat_cpu) {
+		free_percpu(pn->lruvec_stat_local);
 		kfree(pn);
 		return 1;
 	}
@@ -4515,6 +4524,7 @@ static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
 		return;
 
 	free_percpu(pn->lruvec_stat_cpu);
+	free_percpu(pn->lruvec_stat_local);
 	kfree(pn);
 }
 
@@ -4525,6 +4535,7 @@ static void __mem_cgroup_free(struct mem_cgroup *memcg)
 	for_each_node(node)
 		free_mem_cgroup_per_node_info(memcg, node);
 	free_percpu(memcg->vmstats_percpu);
+	free_percpu(memcg->vmstats_local);
 	kfree(memcg);
 }
 
@@ -4553,6 +4564,10 @@ static struct mem_cgroup *mem_cgroup_alloc(void)
 	if (memcg->id.id < 0)
 		goto fail;
 
+	memcg->vmstats_local = alloc_percpu(struct memcg_vmstats_percpu);
+	if (!memcg->vmstats_local)
+		goto fail;
+
 	memcg->vmstats_percpu = alloc_percpu(struct memcg_vmstats_percpu);
 	if (!memcg->vmstats_percpu)
 		goto fail;
-- 
2.21.0

Re: [PATCH] mm: memcontrol: don't batch updates of local VM stats and events

[Shakeel Butt]

On Tue, May 21, 2019 at 8:16 AM Johannes Weiner <hannes@cmpxchg.org> wrote:
>
> The kernel test robot noticed a 26% will-it-scale pagefault regression
> from commit 42a300353577 ("mm: memcontrol: fix recursive statistics
> correctness & scalabilty"). This appears to be caused by bouncing the
> additional cachelines from the new hierarchical statistics counters.
>
> We can fix this by getting rid of the batched local counters instead.
>
> Originally, there were *only* group-local counters, and they were
> fully maintained per cpu. A reader of a stats file high up in the
> cgroup tree would have to walk the entire subtree and collect each
> level's per-cpu counters to get the recursive view. This was
> prohibitively expensive, and so we switched to per-cpu batched updates
> of the local counters during a983b5ebee57 ("mm: memcontrol: fix
> excessive complexity in memory.stat reporting"), reducing the
> complexity from nr_subgroups * nr_cpus to nr_subgroups.
>
> With growing machines and cgroup trees, the tree walk itself became
> too expensive for monitoring top-level groups, and this is when the
> culprit patch added hierarchy counters on each cgroup level. When the
> per-cpu batch size would be reached, both the local and the hierarchy
> counters would get batch-updated from the per-cpu delta simultaneously.
>
> This makes local and hierarchical counter reads blazingly fast, but it
> unfortunately makes the write-side too cache line intense.
>
> Since local counter reads were never a problem - we only centralized
> them to accelerate the hierarchy walk - and use of the local counters
> are becoming rarer due to replacement with hierarchical views (ongoing
> rework in the page reclaim and workingset code), we can make those
> local counters unbatched per-cpu counters again.
>
> The scheme will then be as such:
>
>    when a memcg statistic changes, the writer will:
>    - update the local counter (per-cpu)
>    - update the batch counter (per-cpu). If the batch is full:
>    - spill the batch into the group's atomic_t
>    - spill the batch into all ancestors' atomic_ts
>    - empty out the batch counter (per-cpu)
>
>    when a local memcg counter is read, the reader will:
>    - collect the local counter from all cpus
>
>    when a hiearchy memcg counter is read, the reader will:
>    - read the atomic_t
>
> We might be able to simplify this further and make the recursive
> counters unbatched per-cpu counters as well (batch upward propagation,
> but leave per-cpu collection to the readers), but that will require a
> more in-depth analysis and testing of all the callsites. Deal with the
> immediate regression for now.
>
> Fixes: 42a300353577 ("mm: memcontrol: fix recursive statistics correctness & scalabilty")
> Reported-by: kernel test robot <rong.a.chen@intel.com>
> Tested-by: kernel test robot <rong.a.chen@intel.com>
> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
> ---
>  include/linux/memcontrol.h | 26 ++++++++++++++++--------
>  mm/memcontrol.c            | 41 ++++++++++++++++++++++++++------------
>  2 files changed, 46 insertions(+), 21 deletions(-)
>
> diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h
> index bc74d6a4407c..2d23ae7bd36d 100644
> --- a/include/linux/memcontrol.h
> +++ b/include/linux/memcontrol.h
> @@ -126,9 +126,12 @@ struct memcg_shrinker_map {
>  struct mem_cgroup_per_node {
>         struct lruvec           lruvec;
>
> +       /* Legacy local VM stats */
> +       struct lruvec_stat __percpu *lruvec_stat_local;
> +
> +       /* Subtree VM stats (batched updates) */
>         struct lruvec_stat __percpu *lruvec_stat_cpu;
>         atomic_long_t           lruvec_stat[NR_VM_NODE_STAT_ITEMS];
> -       atomic_long_t           lruvec_stat_local[NR_VM_NODE_STAT_ITEMS];
>
>         unsigned long           lru_zone_size[MAX_NR_ZONES][NR_LRU_LISTS];
>
> @@ -274,17 +277,18 @@ struct mem_cgroup {
>         atomic_t                moving_account;
>         struct task_struct      *move_lock_task;
>
> -       /* memory.stat */
> +       /* Legacy local VM stats and events */
> +       struct memcg_vmstats_percpu __percpu *vmstats_local;
> +
> +       /* Subtree VM stats and events (batched updates) */
>         struct memcg_vmstats_percpu __percpu *vmstats_percpu;
>
>         MEMCG_PADDING(_pad2_);
>
>         atomic_long_t           vmstats[MEMCG_NR_STAT];
> -       atomic_long_t           vmstats_local[MEMCG_NR_STAT];
> -
>         atomic_long_t           vmevents[NR_VM_EVENT_ITEMS];
> -       atomic_long_t           vmevents_local[NR_VM_EVENT_ITEMS];
>
> +       /* memory.events */
>         atomic_long_t           memory_events[MEMCG_NR_MEMORY_EVENTS];
>
>         unsigned long           socket_pressure;
> @@ -576,7 +580,11 @@ static inline unsigned long memcg_page_state(struct mem_cgroup *memcg, int idx)
>  static inline unsigned long memcg_page_state_local(struct mem_cgroup *memcg,
>                                                    int idx)
>  {
> -       long x = atomic_long_read(&memcg->vmstats_local[idx]);
> +       long x = 0;
> +       int cpu;
> +
> +       for_each_possible_cpu(cpu)
> +               x += per_cpu(memcg->vmstats_local->stat[idx], cpu);
>  #ifdef CONFIG_SMP
>         if (x < 0)
>                 x = 0;
> @@ -650,13 +658,15 @@ static inline unsigned long lruvec_page_state_local(struct lruvec *lruvec,
>                                                     enum node_stat_item idx)
>  {
>         struct mem_cgroup_per_node *pn;
> -       long x;
> +       long x = 0;
> +       int cpu;
>
>         if (mem_cgroup_disabled())
>                 return node_page_state(lruvec_pgdat(lruvec), idx);
>
>         pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
> -       x = atomic_long_read(&pn->lruvec_stat_local[idx]);
> +       for_each_possible_cpu(cpu)
> +               x += per_cpu(pn->lruvec_stat_local->count[idx], cpu);
>  #ifdef CONFIG_SMP
>         if (x < 0)
>                 x = 0;
> diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> index e50a2db5b4ff..8d42e5c7bf37 100644
> --- a/mm/memcontrol.c
> +++ b/mm/memcontrol.c
> @@ -700,11 +700,12 @@ void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
>         if (mem_cgroup_disabled())
>                 return;
>
> +       __this_cpu_add(memcg->vmstats_local->stat[idx], val);
> +
>         x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
>         if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
>                 struct mem_cgroup *mi;
>
> -               atomic_long_add(x, &memcg->vmstats_local[idx]);

I was suspecting the following for-loop+atomic-add for the regression.
Why the above atomic-add is the culprit?

>                 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
>                         atomic_long_add(x, &mi->vmstats[idx]);
>                 x = 0;
> @@ -754,11 +755,12 @@ void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
>         __mod_memcg_state(memcg, idx, val);
>
>         /* Update lruvec */
> +       __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
> +
>         x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
>         if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
>                 struct mem_cgroup_per_node *pi;
>
> -               atomic_long_add(x, &pn->lruvec_stat_local[idx]);
>                 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
>                         atomic_long_add(x, &pi->lruvec_stat[idx]);
>                 x = 0;
> @@ -780,11 +782,12 @@ void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
>         if (mem_cgroup_disabled())
>                 return;
>
> +       __this_cpu_add(memcg->vmstats_local->events[idx], count);
> +
>         x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
>         if (unlikely(x > MEMCG_CHARGE_BATCH)) {
>                 struct mem_cgroup *mi;
>
> -               atomic_long_add(x, &memcg->vmevents_local[idx]);
>                 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
>                         atomic_long_add(x, &mi->vmevents[idx]);
>                 x = 0;
> @@ -799,7 +802,12 @@ static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
>
>  static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
>  {
> -       return atomic_long_read(&memcg->vmevents_local[event]);
> +       long x = 0;
> +       int cpu;
> +
> +       for_each_possible_cpu(cpu)
> +               x += per_cpu(memcg->vmstats_local->events[event], cpu);
> +       return x;
>  }
>
>  static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
> @@ -2200,11 +2208,9 @@ static int memcg_hotplug_cpu_dead(unsigned int cpu)
>                         long x;
>
>                         x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
> -                       if (x) {
> -                               atomic_long_add(x, &memcg->vmstats_local[i]);
> +                       if (x)
>                                 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
>                                         atomic_long_add(x, &memcg->vmstats[i]);
> -                       }
>
>                         if (i >= NR_VM_NODE_STAT_ITEMS)
>                                 continue;
> @@ -2214,12 +2220,10 @@ static int memcg_hotplug_cpu_dead(unsigned int cpu)
>
>                                 pn = mem_cgroup_nodeinfo(memcg, nid);
>                                 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
> -                               if (x) {
> -                                       atomic_long_add(x, &pn->lruvec_stat_local[i]);
> +                               if (x)
>                                         do {
>                                                 atomic_long_add(x, &pn->lruvec_stat[i]);
>                                         } while ((pn = parent_nodeinfo(pn, nid)));
> -                               }
>                         }
>                 }
>
> @@ -2227,11 +2231,9 @@ static int memcg_hotplug_cpu_dead(unsigned int cpu)
>                         long x;
>
>                         x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
> -                       if (x) {
> -                               atomic_long_add(x, &memcg->vmevents_local[i]);
> +                       if (x)
>                                 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
>                                         atomic_long_add(x, &memcg->vmevents[i]);
> -                       }
>                 }
>         }
>
> @@ -4492,8 +4494,15 @@ static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
>         if (!pn)
>                 return 1;
>
> +       pn->lruvec_stat_local = alloc_percpu(struct lruvec_stat);
> +       if (!pn->lruvec_stat_local) {
> +               kfree(pn);
> +               return 1;
> +       }
> +
>         pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
>         if (!pn->lruvec_stat_cpu) {
> +               free_percpu(pn->lruvec_stat_local);
>                 kfree(pn);
>                 return 1;
>         }
> @@ -4515,6 +4524,7 @@ static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
>                 return;
>
>         free_percpu(pn->lruvec_stat_cpu);
> +       free_percpu(pn->lruvec_stat_local);
>         kfree(pn);
>  }
>
> @@ -4525,6 +4535,7 @@ static void __mem_cgroup_free(struct mem_cgroup *memcg)
>         for_each_node(node)
>                 free_mem_cgroup_per_node_info(memcg, node);
>         free_percpu(memcg->vmstats_percpu);
> +       free_percpu(memcg->vmstats_local);
>         kfree(memcg);
>  }
>
> @@ -4553,6 +4564,10 @@ static struct mem_cgroup *mem_cgroup_alloc(void)
>         if (memcg->id.id < 0)
>                 goto fail;
>
> +       memcg->vmstats_local = alloc_percpu(struct memcg_vmstats_percpu);
> +       if (!memcg->vmstats_local)
> +               goto fail;
> +
>         memcg->vmstats_percpu = alloc_percpu(struct memcg_vmstats_percpu);
>         if (!memcg->vmstats_percpu)
>                 goto fail;
> --
> 2.21.0
>

Re: [PATCH] mm: memcontrol: don't batch updates of local VM stats and events

[Linus Torvalds]

On Tue, May 28, 2019 at 9:00 AM Shakeel Butt <shakeelb@google.com> wrote:
>
> I was suspecting the following for-loop+atomic-add for the regression.

If I read the kernel test robot reports correctly, Johannes' fix patch
does fix the regression (well - mostly. The original reported
regression was 26%, and with Johannes' fix patch it was 3% - so still
a slight performance regression, but not nearly as bad).

> Why the above atomic-add is the culprit?

I think the problem with that one is that it's cross-cpu statistics,
so you end up with lots of cacheline bounces on the local counts when
you have lots of load.

But yes, the recursive updates still do show a small regression,
probably because there's still some overhead from the looping up in
the hierarchy. You still get *those* cacheline bounces, but now they
are limited to the upper hierarchies that only get updated at batch
time.

Johannes? Am I reading this right?

                   Linus

Re: [PATCH] mm: memcontrol: don't batch updates of local VM stats and events

[Johannes Weiner]

On Tue, May 28, 2019 at 10:37:15AM -0700, Linus Torvalds wrote:
> On Tue, May 28, 2019 at 9:00 AM Shakeel Butt <shakeelb@google.com> wrote:
> >
> > I was suspecting the following for-loop+atomic-add for the regression.
> 
> If I read the kernel test robot reports correctly, Johannes' fix patch
> does fix the regression (well - mostly. The original reported
> regression was 26%, and with Johannes' fix patch it was 3% - so still
> a slight performance regression, but not nearly as bad).
> 
> > Why the above atomic-add is the culprit?
> 
> I think the problem with that one is that it's cross-cpu statistics,
> so you end up with lots of cacheline bounces on the local counts when
> you have lots of load.

In this case, that's true for both of them. The workload runs at the
root cgroup level, so per definition the local and the recursive
counters at that level are identical and written to at the same
rate. Adding the new counter obviously caused the regression, but
they're contributing equally to the cost, and we could
remove/per-cpuify either of them for the fix.

So why did I unshare the old counter instead of the new one? Well, the
old counter *used* to be unshared for the longest time, and was only
made into a shared one to make recursive aggregation cheaper - before
there was a dedicated recursive counter. But now that we have that
recursive counter, there isn't much reason to keep the local counter
shared and bounce it around on updates.

Essentially, this fix-up is a revert of a983b5ebee57 ("mm: memcontrol:
fix excessive complexity in memory.stat reporting") since the problem
described in that patch is now solved from the other end.

> But yes, the recursive updates still do show a small regression,
> probably because there's still some overhead from the looping up in
> the hierarchy. You still get *those* cacheline bounces, but now they
> are limited to the upper hierarchies that only get updated at batch
> time.

Right, I reduce the *shared* data back to how it was before the patch,
but it still adds a second (per-cpu) counter that needs to get bumped,
and the loop adds a branch as well.

But while I would expect that to show up in a case like will-it-scale,
I'd be surprised if the remaining difference would be noticeable for
real workloads that actually work with the memory they allocate.