Skip to contentSkip to conversation

mm: thp: move deferred split queue to memcg's nodeinfo

PATCH In review Yang Shi 4 participants 8 messages Oct 1, 2019 – Oct 8, 2019

Subsystems (3)
lore
Share & tools
lore

Yang Shi

author
#0 lore raw

PATCH 1/1 mm: thp: move deferred split queue to memcg's nodeinfo

The commit 87eaceb3faa59b9b4d940ec9554ce251325d83fe ("mm: thp: make deferred split shrinker memcg aware") makes deferred split queue per memcg to resolve memcg pre-mature OOM problem. But, all nodes end up sharing the same queue instead of one queue per-node before the commit. It is not a big deal for memcg limit reclaim, but it may cause global kswapd shrink THPs from a different node.

And, 0-day testing reported -19.6% regression of stress-ng's madvise test [1]. I didn't see that much regression on my test box (24 threads, 48GB memory, 2 nodes), with the same test (stress-ng --timeout 1 --metrics-brief --sequential 72 --class vm --exclude spawn,exec), I saw average -3% (run the same test 10 times then calculate the average since the test itself may have most 15% variation according to my test) regression sometimes (not every time, sometimes I didn't see regression at all).

This might be caused by deferred split queue lock contention. With some configuration (i.e. just one root memcg) the lock contention my be worse than before (given 2 nodes, two locks are reduced to one lock).

So, moving deferred split queue to memcg's nodeinfo to make it NUMA aware again.

With this change stress-ng's madvise test shows average 4% improvement sometimes and I didn't see degradation anymore.

[1]: https://lore.kernel.org/lkml/20190930084604.GC17687@shao2-debian/

Cc Kirill A. Shutemov <[email protected]>
Cc Kirill Tkhai <[email protected]>
Cc Johannes Weiner <[email protected]>
Cc Michal Hocko <[email protected]>
Cc Hugh Dickins <[email protected]>
Cc Shakeel Butt <[email protected]>
Cc David Rientjes <[email protected]>
---
 include/linux/memcontrol.h |  8 ++++----
 mm/huge_memory.c           | 15 +++++++++------
 mm/memcontrol.c            | 29 +++++++++++++++++------------
 3 files changed, 30 insertions(+), 22 deletions(-)
include/linux/memcontrol.h +4 -4
index 9b60863..4b5c791 100644
@@ -137,6 +137,10 @@ struct mem_cgroup_per_node {
 	bool			congested;	/* memcg has many dirty pages */
 						/* backed by a congested BDI */
 
+#ifdef CONFIG_TRANSPARENT_HUGEPAGE
+	struct deferred_split deferred_split_queue;
+#endif
+
 	struct mem_cgroup	*memcg;		/* Back pointer, we cannot */
 						/* use container_of	   */
 };
@@ -330,10 +334,6 @@ struct mem_cgroup {
 	struct list_head event_list;
 	spinlock_t event_list_lock;
 
-#ifdef CONFIG_TRANSPARENT_HUGEPAGE
-	struct deferred_split deferred_split_queue;
-#endif
-
 	struct mem_cgroup_per_node *nodeinfo[0];
 	/* WARNING: nodeinfo must be the last member here */
 };
mm/huge_memory.c +9 -6
index c5cb6dc..3b78910 100644
@@ -500,10 +500,11 @@ pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 {
 	struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
-	struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
+	int nid = page_to_nid(page);
+	struct pglist_data *pgdat = NODE_DATA(nid);
 
 	if (memcg)
-		return &memcg->deferred_split_queue;
+		return &memcg->nodeinfo[nid]->deferred_split_queue;
 	else
 		return &pgdat->deferred_split_queue;
 }
@@ -2882,12 +2883,13 @@ void deferred_split_huge_page(struct page *page)
 static unsigned long deferred_split_count(struct shrinker *shrink,
 		struct shrink_control *sc)
 {
-	struct pglist_data *pgdata = NODE_DATA(sc->nid);
+	int nid = sc->nid;
+	struct pglist_data *pgdata = NODE_DATA(nid);
 	struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
 
 #ifdef CONFIG_MEMCG
 	if (sc->memcg)
-		ds_queue = &sc->memcg->deferred_split_queue;
+		ds_queue = &sc->memcg->nodeinfo[nid]->deferred_split_queue;
 #endif
 	return READ_ONCE(ds_queue->split_queue_len);
 }
@@ -2895,7 +2897,8 @@ static unsigned long deferred_split_count(struct shrinker *shrink,
 static unsigned long deferred_split_scan(struct shrinker *shrink,
 		struct shrink_control *sc)
 {
-	struct pglist_data *pgdata = NODE_DATA(sc->nid);
+	int nid = sc->nid;
+	struct pglist_data *pgdata = NODE_DATA(nid);
 	struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
 	unsigned long flags;
 	LIST_HEAD(list), *pos, *next;
@@ -2904,7 +2907,7 @@ static unsigned long deferred_split_scan(struct shrinker *shrink,
 
 #ifdef CONFIG_MEMCG
 	if (sc->memcg)
-		ds_queue = &sc->memcg->deferred_split_queue;
+		ds_queue = &sc->memcg->nodeinfo[nid]->deferred_split_queue;
 #endif
 
 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
mm/memcontrol.c +17 -12
index c313c49..19d4295 100644
@@ -4989,6 +4989,12 @@ static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
 	pn->on_tree = false;
 	pn->memcg = memcg;
 
+#ifdef CONFIG_TRANSPARENT_HUGEPAGE
+	spin_lock_init(&pn->deferred_split_queue.split_queue_lock);
+	INIT_LIST_HEAD(&pn->deferred_split_queue.split_queue);
+	pn->deferred_split_queue.split_queue_len = 0;
+#endif
+
 	memcg->nodeinfo[node] = pn;
 	return 0;
 }
@@ -5081,11 +5087,6 @@ static struct mem_cgroup *mem_cgroup_alloc(void)
 		memcg->cgwb_frn[i].done =
 			__WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
 #endif
-#ifdef CONFIG_TRANSPARENT_HUGEPAGE
-	spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
-	INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
-	memcg->deferred_split_queue.split_queue_len = 0;
-#endif
 	idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
 	return memcg;
 fail:
@@ -5419,6 +5420,8 @@ static int mem_cgroup_move_account(struct page *page,
 	unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
 	int ret;
 	bool anon;
+	struct deferred_split *ds_queue;
+	int nid = page_to_nid(page);
 
 	VM_BUG_ON(from == to);
 	VM_BUG_ON_PAGE(PageLRU(page), page);
@@ -5466,10 +5469,11 @@ static int mem_cgroup_move_account(struct page *page,
 
 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 	if (compound && !list_empty(page_deferred_list(page))) {
-		spin_lock(&from->deferred_split_queue.split_queue_lock);
+		ds_queue = &from->nodeinfo[nid]->deferred_split_queue;
+		spin_lock(&ds_queue->split_queue_lock);
 		list_del_init(page_deferred_list(page));
-		from->deferred_split_queue.split_queue_len--;
-		spin_unlock(&from->deferred_split_queue.split_queue_lock);
+		ds_queue->split_queue_len--;
+		spin_unlock(&ds_queue->split_queue_lock);
 	}
 #endif
 	/*
@@ -5483,11 +5487,12 @@ static int mem_cgroup_move_account(struct page *page,
 
 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 	if (compound && list_empty(page_deferred_list(page))) {
-		spin_lock(&to->deferred_split_queue.split_queue_lock);
+		ds_queue = &to->nodeinfo[nid]->deferred_split_queue;
+		spin_lock(&ds_queue->split_queue_lock);
 		list_add_tail(page_deferred_list(page),
-			      &to->deferred_split_queue.split_queue);
-		to->deferred_split_queue.split_queue_len++;
-		spin_unlock(&to->deferred_split_queue.split_queue_lock);
+			      &ds_queue->split_queue);
+		ds_queue->split_queue_len++;
+		spin_unlock(&ds_queue->split_queue_lock);
 	}
 #endif
 
Signature · 1 line
1.8.3.1
maintainer
#1 lore raw
Quoted 20 lines · Yang Shi

The commit 87eaceb3faa59b9b4d940ec9554ce251325d83fe ("mm: thp: make deferred split shrinker memcg aware") makes deferred split queue per memcg to resolve memcg pre-mature OOM problem. But, all nodes end up sharing the same queue instead of one queue per-node before the commit. It is not a big deal for memcg limit reclaim, but it may cause global kswapd shrink THPs from a different node.

And, 0-day testing reported -19.6% regression of stress-ng's madvise test [1]. I didn't see that much regression on my test box (24 threads, 48GB memory, 2 nodes), with the same test (stress-ng --timeout 1 --metrics-brief --sequential 72 --class vm --exclude spawn,exec), I saw average -3% (run the same test 10 times then calculate the average since the test itself may have most 15% variation according to my test) regression sometimes (not every time, sometimes I didn't see regression at all).

This might be caused by deferred split queue lock contention. With some configuration (i.e. just one root memcg) the lock contention my be worse than before (given 2 nodes, two locks are reduced to one lock).

Quoted 5 lines

So, moving deferred split queue to memcg's nodeinfo to make it NUMA aware again.

With this change stress-ng's madvise test shows average 4% improvement sometimes and I didn't see degradation anymore.

My concern about this getting more and more complex
(http://lkml.kernel.org/r/[email protected]) holds here even more. Can we step back and reconsider the whole thing please?

Quoted 167 lines

[1]: https://lore.kernel.org/lkml/20190930084604.GC17687@shao2-debian/

Cc Kirill A. Shutemov <[email protected]>
Cc Kirill Tkhai <[email protected]>
Cc Johannes Weiner <[email protected]>
Cc Michal Hocko <[email protected]>
Cc Hugh Dickins <[email protected]>
Cc Shakeel Butt <[email protected]>
Cc David Rientjes <[email protected]>
---
 include/linux/memcontrol.h |  8 ++++----
 mm/huge_memory.c           | 15 +++++++++------
 mm/memcontrol.c            | 29 +++++++++++++++++------------
 3 files changed, 30 insertions(+), 22 deletions(-)
include/linux/memcontrol.h +4 -4
index 9b60863..4b5c791 100644
@@ -137,6 +137,10 @@ struct mem_cgroup_per_node {
 	bool			congested;	/* memcg has many dirty pages */
 						/* backed by a congested BDI */
 
+#ifdef CONFIG_TRANSPARENT_HUGEPAGE
+	struct deferred_split deferred_split_queue;
+#endif
+
 	struct mem_cgroup	*memcg;		/* Back pointer, we cannot */
 						/* use container_of	   */
 };
@@ -330,10 +334,6 @@ struct mem_cgroup {
 	struct list_head event_list;
 	spinlock_t event_list_lock;
 
-#ifdef CONFIG_TRANSPARENT_HUGEPAGE
-	struct deferred_split deferred_split_queue;
-#endif
-
 	struct mem_cgroup_per_node *nodeinfo[0];
 	/* WARNING: nodeinfo must be the last member here */
 };
mm/huge_memory.c +9 -6
index c5cb6dc..3b78910 100644
@@ -500,10 +500,11 @@ pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 {
 	struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
-	struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
+	int nid = page_to_nid(page);
+	struct pglist_data *pgdat = NODE_DATA(nid);
 
 	if (memcg)
-		return &memcg->deferred_split_queue;
+		return &memcg->nodeinfo[nid]->deferred_split_queue;
 	else
 		return &pgdat->deferred_split_queue;
 }
@@ -2882,12 +2883,13 @@ void deferred_split_huge_page(struct page *page)
 static unsigned long deferred_split_count(struct shrinker *shrink,
 		struct shrink_control *sc)
 {
-	struct pglist_data *pgdata = NODE_DATA(sc->nid);
+	int nid = sc->nid;
+	struct pglist_data *pgdata = NODE_DATA(nid);
 	struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
 
 #ifdef CONFIG_MEMCG
 	if (sc->memcg)
-		ds_queue = &sc->memcg->deferred_split_queue;
+		ds_queue = &sc->memcg->nodeinfo[nid]->deferred_split_queue;
 #endif
 	return READ_ONCE(ds_queue->split_queue_len);
 }
@@ -2895,7 +2897,8 @@ static unsigned long deferred_split_count(struct shrinker *shrink,
 static unsigned long deferred_split_scan(struct shrinker *shrink,
 		struct shrink_control *sc)
 {
-	struct pglist_data *pgdata = NODE_DATA(sc->nid);
+	int nid = sc->nid;
+	struct pglist_data *pgdata = NODE_DATA(nid);
 	struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
 	unsigned long flags;
 	LIST_HEAD(list), *pos, *next;
@@ -2904,7 +2907,7 @@ static unsigned long deferred_split_scan(struct shrinker *shrink,
 
 #ifdef CONFIG_MEMCG
 	if (sc->memcg)
-		ds_queue = &sc->memcg->deferred_split_queue;
+		ds_queue = &sc->memcg->nodeinfo[nid]->deferred_split_queue;
 #endif
 
 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
mm/memcontrol.c +17 -12
index c313c49..19d4295 100644
@@ -4989,6 +4989,12 @@ static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
 	pn->on_tree = false;
 	pn->memcg = memcg;
 
+#ifdef CONFIG_TRANSPARENT_HUGEPAGE
+	spin_lock_init(&pn->deferred_split_queue.split_queue_lock);
+	INIT_LIST_HEAD(&pn->deferred_split_queue.split_queue);
+	pn->deferred_split_queue.split_queue_len = 0;
+#endif
+
 	memcg->nodeinfo[node] = pn;
 	return 0;
 }
@@ -5081,11 +5087,6 @@ static struct mem_cgroup *mem_cgroup_alloc(void)
 		memcg->cgwb_frn[i].done =
 			__WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
 #endif
-#ifdef CONFIG_TRANSPARENT_HUGEPAGE
-	spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
-	INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
-	memcg->deferred_split_queue.split_queue_len = 0;
-#endif
 	idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
 	return memcg;
 fail:
@@ -5419,6 +5420,8 @@ static int mem_cgroup_move_account(struct page *page,
 	unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
 	int ret;
 	bool anon;
+	struct deferred_split *ds_queue;
+	int nid = page_to_nid(page);
 
 	VM_BUG_ON(from == to);
 	VM_BUG_ON_PAGE(PageLRU(page), page);
@@ -5466,10 +5469,11 @@ static int mem_cgroup_move_account(struct page *page,
 
 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 	if (compound && !list_empty(page_deferred_list(page))) {
-		spin_lock(&from->deferred_split_queue.split_queue_lock);
+		ds_queue = &from->nodeinfo[nid]->deferred_split_queue;
+		spin_lock(&ds_queue->split_queue_lock);
 		list_del_init(page_deferred_list(page));
-		from->deferred_split_queue.split_queue_len--;
-		spin_unlock(&from->deferred_split_queue.split_queue_lock);
+		ds_queue->split_queue_len--;
+		spin_unlock(&ds_queue->split_queue_lock);
 	}
 #endif
 	/*
@@ -5483,11 +5487,12 @@ static int mem_cgroup_move_account(struct page *page,
 
 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 	if (compound && list_empty(page_deferred_list(page))) {
-		spin_lock(&to->deferred_split_queue.split_queue_lock);
+		ds_queue = &to->nodeinfo[nid]->deferred_split_queue;
+		spin_lock(&ds_queue->split_queue_lock);
 		list_add_tail(page_deferred_list(page),
-			      &to->deferred_split_queue.split_queue);
-		to->deferred_split_queue.split_queue_len++;
-		spin_unlock(&to->deferred_split_queue.split_queue_lock);
+			      &ds_queue->split_queue);
+		ds_queue->split_queue_len++;
+		spin_unlock(&ds_queue->split_queue_lock);
 	}
 #endif
 
Signature · 1 line
1.8.3.1
Signature · 2 lines
Michal Hocko
SUSE Labs
author
#2 lore raw
Quoted 21 lines · Michal Hocko
Quoted 20 lines · Yang Shi

The commit 87eaceb3faa59b9b4d940ec9554ce251325d83fe ("mm: thp: make deferred split shrinker memcg aware") makes deferred split queue per memcg to resolve memcg pre-mature OOM problem. But, all nodes end up sharing the same queue instead of one queue per-node before the commit. It is not a big deal for memcg limit reclaim, but it may cause global kswapd shrink THPs from a different node.

And, 0-day testing reported -19.6% regression of stress-ng's madvise test [1]. I didn't see that much regression on my test box (24 threads, 48GB memory, 2 nodes), with the same test (stress-ng --timeout 1 --metrics-brief --sequential 72 --class vm --exclude spawn,exec), I saw average -3% (run the same test 10 times then calculate the average since the test itself may have most 15% variation according to my test) regression sometimes (not every time, sometimes I didn't see regression at all).

This might be caused by deferred split queue lock contention. With some configuration (i.e. just one root memcg) the lock contention my be worse than before (given 2 nodes, two locks are reduced to one lock).

Quoted 8 lines
Quoted 5 lines

So, moving deferred split queue to memcg's nodeinfo to make it NUMA aware again.

With this change stress-ng's madvise test shows average 4% improvement sometimes and I didn't see degradation anymore.

My concern about this getting more and more complex
(http://lkml.kernel.org/r/[email protected]) holds here even more. Can we step back and reconsider the whole thing please?

I'm not intended to make this more complicated, but this patch is aimed to fix the NUMA awareness issue. It doesn't tune something further.

Quoted 168 lines
Quoted 168 lines

[1]: https://lore.kernel.org/lkml/20190930084604.GC17687@shao2-debian/

Cc Kirill A. Shutemov <[email protected]>
Cc Kirill Tkhai <[email protected]>
Cc Johannes Weiner <[email protected]>
Cc Michal Hocko <[email protected]>
Cc Hugh Dickins <[email protected]>
Cc Shakeel Butt <[email protected]>
Cc David Rientjes <[email protected]>
---
  include/linux/memcontrol.h |  8 ++++----
  mm/huge_memory.c           | 15 +++++++++------
  mm/memcontrol.c            | 29 +++++++++++++++++------------
  3 files changed, 30 insertions(+), 22 deletions(-)
include/linux/memcontrol.h +4 -4
index 9b60863..4b5c791 100644
@@ -137,6 +137,10 @@ struct mem_cgroup_per_node {
  	bool			congested;	/* memcg has many dirty pages */
  						/* backed by a congested BDI */
 
+#ifdef CONFIG_TRANSPARENT_HUGEPAGE
+	struct deferred_split deferred_split_queue;
+#endif
+
  	struct mem_cgroup	*memcg;		/* Back pointer, we cannot */
  						/* use container_of	   */
  };
@@ -330,10 +334,6 @@ struct mem_cgroup {
  	struct list_head event_list;
  	spinlock_t event_list_lock;
 
-#ifdef CONFIG_TRANSPARENT_HUGEPAGE
-	struct deferred_split deferred_split_queue;
-#endif
-
  	struct mem_cgroup_per_node *nodeinfo[0];
  	/* WARNING: nodeinfo must be the last member here */
  };
mm/huge_memory.c +9 -6
index c5cb6dc..3b78910 100644
@@ -500,10 +500,11 @@ pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
  static inline struct deferred_split *get_deferred_split_queue(struct page *page)
  {
  	struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
-	struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
+	int nid = page_to_nid(page);
+	struct pglist_data *pgdat = NODE_DATA(nid);
 
  	if (memcg)
-		return &memcg->deferred_split_queue;
+		return &memcg->nodeinfo[nid]->deferred_split_queue;
  	else
  		return &pgdat->deferred_split_queue;
  }
@@ -2882,12 +2883,13 @@ void deferred_split_huge_page(struct page *page)
  static unsigned long deferred_split_count(struct shrinker *shrink,
  		struct shrink_control *sc)
  {
-	struct pglist_data *pgdata = NODE_DATA(sc->nid);
+	int nid = sc->nid;
+	struct pglist_data *pgdata = NODE_DATA(nid);
  	struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
 
  #ifdef CONFIG_MEMCG
  	if (sc->memcg)
-		ds_queue = &sc->memcg->deferred_split_queue;
+		ds_queue = &sc->memcg->nodeinfo[nid]->deferred_split_queue;
  #endif
  	return READ_ONCE(ds_queue->split_queue_len);
  }
@@ -2895,7 +2897,8 @@ static unsigned long deferred_split_count(struct shrinker *shrink,
  static unsigned long deferred_split_scan(struct shrinker *shrink,
  		struct shrink_control *sc)
  {
-	struct pglist_data *pgdata = NODE_DATA(sc->nid);
+	int nid = sc->nid;
+	struct pglist_data *pgdata = NODE_DATA(nid);
  	struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
  	unsigned long flags;
  	LIST_HEAD(list), *pos, *next;
@@ -2904,7 +2907,7 @@ static unsigned long deferred_split_scan(struct shrinker *shrink,
 
  #ifdef CONFIG_MEMCG
  	if (sc->memcg)
-		ds_queue = &sc->memcg->deferred_split_queue;
+		ds_queue = &sc->memcg->nodeinfo[nid]->deferred_split_queue;
  #endif
 
  	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
mm/memcontrol.c +17 -12
index c313c49..19d4295 100644
@@ -4989,6 +4989,12 @@ static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
  	pn->on_tree = false;
  	pn->memcg = memcg;
 
+#ifdef CONFIG_TRANSPARENT_HUGEPAGE
+	spin_lock_init(&pn->deferred_split_queue.split_queue_lock);
+	INIT_LIST_HEAD(&pn->deferred_split_queue.split_queue);
+	pn->deferred_split_queue.split_queue_len = 0;
+#endif
+
  	memcg->nodeinfo[node] = pn;
  	return 0;
  }
@@ -5081,11 +5087,6 @@ static struct mem_cgroup *mem_cgroup_alloc(void)
  		memcg->cgwb_frn[i].done =
  			__WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
  #endif
-#ifdef CONFIG_TRANSPARENT_HUGEPAGE
-	spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
-	INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
-	memcg->deferred_split_queue.split_queue_len = 0;
-#endif
  	idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
  	return memcg;
  fail:
@@ -5419,6 +5420,8 @@ static int mem_cgroup_move_account(struct page *page,
  	unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
  	int ret;
  	bool anon;
+	struct deferred_split *ds_queue;
+	int nid = page_to_nid(page);
 
  	VM_BUG_ON(from == to);
  	VM_BUG_ON_PAGE(PageLRU(page), page);
@@ -5466,10 +5469,11 @@ static int mem_cgroup_move_account(struct page *page,
 
  #ifdef CONFIG_TRANSPARENT_HUGEPAGE
  	if (compound && !list_empty(page_deferred_list(page))) {
-		spin_lock(&from->deferred_split_queue.split_queue_lock);
+		ds_queue = &from->nodeinfo[nid]->deferred_split_queue;
+		spin_lock(&ds_queue->split_queue_lock);
  		list_del_init(page_deferred_list(page));
-		from->deferred_split_queue.split_queue_len--;
-		spin_unlock(&from->deferred_split_queue.split_queue_lock);
+		ds_queue->split_queue_len--;
+		spin_unlock(&ds_queue->split_queue_lock);
  	}
  #endif
  	/*
@@ -5483,11 +5487,12 @@ static int mem_cgroup_move_account(struct page *page,
 
  #ifdef CONFIG_TRANSPARENT_HUGEPAGE
  	if (compound && list_empty(page_deferred_list(page))) {
-		spin_lock(&to->deferred_split_queue.split_queue_lock);
+		ds_queue = &to->nodeinfo[nid]->deferred_split_queue;
+		spin_lock(&ds_queue->split_queue_lock);
  		list_add_tail(page_deferred_list(page),
-			      &to->deferred_split_queue.split_queue);
-		to->deferred_split_queue.split_queue_len++;
-		spin_unlock(&to->deferred_split_queue.split_queue_lock);
+			      &ds_queue->split_queue);
+		ds_queue->split_queue_len++;
+		spin_unlock(&ds_queue->split_queue_lock);
  	}
  #endif
 
Signature · 1 line
1.8.3.1
#3 lore raw
In reply to Michal Hocko
Quoted 24 lines · Michal Hocko
Quoted 23 lines · Yang Shi

The commit 87eaceb3faa59b9b4d940ec9554ce251325d83fe ("mm: thp: make deferred split shrinker memcg aware") makes deferred split queue per memcg to resolve memcg pre-mature OOM problem. But, all nodes end up sharing the same queue instead of one queue per-node before the commit. It is not a big deal for memcg limit reclaim, but it may cause global kswapd shrink THPs from a different node.

And, 0-day testing reported -19.6% regression of stress-ng's madvise test [1]. I didn't see that much regression on my test box (24 threads, 48GB memory, 2 nodes), with the same test (stress-ng --timeout 1 --metrics-brief --sequential 72 --class vm --exclude spawn,exec), I saw average -3% (run the same test 10 times then calculate the average since the test itself may have most 15% variation according to my test) regression sometimes (not every time, sometimes I didn't see regression at all).

This might be caused by deferred split queue lock contention. With some configuration (i.e. just one root memcg) the lock contention my be worse than before (given 2 nodes, two locks are reduced to one lock).

So, moving deferred split queue to memcg's nodeinfo to make it NUMA aware again.

Quoted 6 lines
Quoted 2 lines

With this change stress-ng's madvise test shows average 4% improvement sometimes and I didn't see degradation anymore.

My concern about this getting more and more complex
(http://lkml.kernel.org/r/[email protected]) holds here even more. Can we step back and reconsider the whole thing please?

What about freeing immediately after split via workqueue and also have a synchronous version called before going oom? Maybe there would be also other things that would benefit from this scheme instead of traditional reclaim and shrinkers?

maintainer
#4 lore raw
Quoted 28 lines · Vlastimil Babka
Quoted 27 lines · Michal Hocko
Quoted 25 lines · Yang Shi

The commit 87eaceb3faa59b9b4d940ec9554ce251325d83fe ("mm: thp: make deferred split shrinker memcg aware") makes deferred split queue per memcg to resolve memcg pre-mature OOM problem. But, all nodes end up sharing the same queue instead of one queue per-node before the commit. It is not a big deal for memcg limit reclaim, but it may cause global kswapd shrink THPs from a different node.

And, 0-day testing reported -19.6% regression of stress-ng's madvise test [1]. I didn't see that much regression on my test box (24 threads, 48GB memory, 2 nodes), with the same test (stress-ng --timeout 1 --metrics-brief --sequential 72 --class vm --exclude spawn,exec), I saw average -3% (run the same test 10 times then calculate the average since the test itself may have most 15% variation according to my test) regression sometimes (not every time, sometimes I didn't see regression at all).

This might be caused by deferred split queue lock contention. With some configuration (i.e. just one root memcg) the lock contention my be worse than before (given 2 nodes, two locks are reduced to one lock).

So, moving deferred split queue to memcg's nodeinfo to make it NUMA aware again.

With this change stress-ng's madvise test shows average 4% improvement sometimes and I didn't see degradation anymore.

Quoted 8 lines
Quoted 3 lines

My concern about this getting more and more complex
(http://lkml.kernel.org/r/[email protected]) holds here even more. Can we step back and reconsider the whole thing please?

What about freeing immediately after split via workqueue and also have a synchronous version called before going oom? Maybe there would be also other things that would benefit from this scheme instead of traditional reclaim and shrinkers?

That is exactly what we have discussed some time ago.

Signature · 2 lines
Michal Hocko
SUSE Labs
#5 lore raw
Quoted 33 lines · Michal Hocko
Quoted 32 lines · Vlastimil Babka
Quoted 30 lines · Michal Hocko
Quoted 20 lines · Yang Shi

The commit 87eaceb3faa59b9b4d940ec9554ce251325d83fe ("mm: thp: make deferred split shrinker memcg aware") makes deferred split queue per memcg to resolve memcg pre-mature OOM problem. But, all nodes end up sharing the same queue instead of one queue per-node before the commit. It is not a big deal for memcg limit reclaim, but it may cause global kswapd shrink THPs from a different node.

And, 0-day testing reported -19.6% regression of stress-ng's madvise test [1]. I didn't see that much regression on my test box (24 threads, 48GB memory, 2 nodes), with the same test (stress-ng --timeout 1 --metrics-brief --sequential 72 --class vm --exclude spawn,exec), I saw average -3% (run the same test 10 times then calculate the average since the test itself may have most 15% variation according to my test) regression sometimes (not every time, sometimes I didn't see regression at all).

This might be caused by deferred split queue lock contention. With some configuration (i.e. just one root memcg) the lock contention my be worse than before (given 2 nodes, two locks are reduced to one lock).

Quoted 5 lines

So, moving deferred split queue to memcg's nodeinfo to make it NUMA aware again.

With this change stress-ng's madvise test shows average 4% improvement sometimes and I didn't see degradation anymore.

My concern about this getting more and more complex
(http://lkml.kernel.org/r/[email protected]) holds here even more. Can we step back and reconsider the whole thing please?

Quoted 6 lines
Quoted 4 lines

What about freeing immediately after split via workqueue and also have a synchronous version called before going oom? Maybe there would be also other things that would benefit from this scheme instead of traditional reclaim and shrinkers?

That is exactly what we have discussed some time ago.

Yes, I've posted the patch:

http://lkml.kernel.org/r/20190827125911.boya23eowxhqmopa@box

But I still not sure that the approach is right. I expect it to trigger performance regressions. For system with pleanty of free memory, we will just pay split cost for nothing in many cases.

Signature · 1 line
 Kirill A. Shutemov
maintainer
#6 lore raw
Quoted 41 lines · Kirill A. Shutemov
Quoted 33 lines · Michal Hocko
Quoted 32 lines · Vlastimil Babka
Quoted 30 lines · Michal Hocko
Quoted 20 lines · Yang Shi

The commit 87eaceb3faa59b9b4d940ec9554ce251325d83fe ("mm: thp: make deferred split shrinker memcg aware") makes deferred split queue per memcg to resolve memcg pre-mature OOM problem. But, all nodes end up sharing the same queue instead of one queue per-node before the commit. It is not a big deal for memcg limit reclaim, but it may cause global kswapd shrink THPs from a different node.

And, 0-day testing reported -19.6% regression of stress-ng's madvise test [1]. I didn't see that much regression on my test box (24 threads, 48GB memory, 2 nodes), with the same test (stress-ng --timeout 1 --metrics-brief --sequential 72 --class vm --exclude spawn,exec), I saw average -3% (run the same test 10 times then calculate the average since the test itself may have most 15% variation according to my test) regression sometimes (not every time, sometimes I didn't see regression at all).

This might be caused by deferred split queue lock contention. With some configuration (i.e. just one root memcg) the lock contention my be worse than before (given 2 nodes, two locks are reduced to one lock).

Quoted 5 lines

So, moving deferred split queue to memcg's nodeinfo to make it NUMA aware again.

With this change stress-ng's madvise test shows average 4% improvement sometimes and I didn't see degradation anymore.

My concern about this getting more and more complex
(http://lkml.kernel.org/r/[email protected]) holds here even more. Can we step back and reconsider the whole thing please?

Quoted 6 lines
Quoted 4 lines

What about freeing immediately after split via workqueue and also have a synchronous version called before going oom? Maybe there would be also other things that would benefit from this scheme instead of traditional reclaim and shrinkers?

That is exactly what we have discussed some time ago.

Quoted 7 lines

Yes, I've posted the patch:

http://lkml.kernel.org/r/20190827125911.boya23eowxhqmopa@box

But I still not sure that the approach is right. I expect it to trigger performance regressions. For system with pleanty of free memory, we will just pay split cost for nothing in many cases.

I suspect it got lost in the email thread. Care to send as a separate RFC patch? We can put it to mm for a cycle or two to see how it behaves. The patch seems quite simple and straightforward from a very quick glance. It is a bit of a hack that it piggybacks on top of the shrinker code which should ideally go away if this approach works but that is a minor detail.

Signature · 2 lines
Michal Hocko
SUSE Labs
author
#7 lore raw
In reply to Kirill A. Shutemov
Quoted 36 lines · Kirill A. Shutemov
Quoted 35 lines · Michal Hocko
Quoted 34 lines · Vlastimil Babka
Quoted 21 lines · Michal Hocko
Quoted 20 lines · Yang Shi

The commit 87eaceb3faa59b9b4d940ec9554ce251325d83fe ("mm: thp: make deferred split shrinker memcg aware") makes deferred split queue per memcg to resolve memcg pre-mature OOM problem. But, all nodes end up sharing the same queue instead of one queue per-node before the commit. It is not a big deal for memcg limit reclaim, but it may cause global kswapd shrink THPs from a different node.

And, 0-day testing reported -19.6% regression of stress-ng's madvise test [1]. I didn't see that much regression on my test box (24 threads, 48GB memory, 2 nodes), with the same test (stress-ng --timeout 1 --metrics-brief --sequential 72 --class vm --exclude spawn,exec), I saw average -3% (run the same test 10 times then calculate the average since the test itself may have most 15% variation according to my test) regression sometimes (not every time, sometimes I didn't see regression at all).

This might be caused by deferred split queue lock contention. With some configuration (i.e. just one root memcg) the lock contention my be worse than before (given 2 nodes, two locks are reduced to one lock).

Quoted 8 lines
Quoted 5 lines

So, moving deferred split queue to memcg's nodeinfo to make it NUMA aware again.

With this change stress-ng's madvise test shows average 4% improvement sometimes and I didn't see degradation anymore.

My concern about this getting more and more complex
(http://lkml.kernel.org/r/[email protected]) holds here even more. Can we step back and reconsider the whole thing please?

What about freeing immediately after split via workqueue and also have a synchronous version called before going oom? Maybe there would be also other things that would benefit from this scheme instead of traditional reclaim and shrinkers?

Quoted 8 lines
Quoted 1 line

That is exactly what we have discussed some time ago.

Yes, I've posted the patch:

http://lkml.kernel.org/r/20190827125911.boya23eowxhqmopa@box

But I still not sure that the approach is right. I expect it to trigger performance regressions. For system with pleanty of free memory, we will just pay split cost for nothing in many cases.

This is exactly what I'm concerned about as well.

Quoted 1 line