Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1376456 > unrolled thread

[PATCH v2 00/11] mm/slab: reduce lock contention in alloc path

Back to article view | Back to linux.kernel

Contents

  [PATCH v2 00/11] mm/slab: reduce lock contention in alloc path js1304@gmail.com - 2016-04-12 07:00 +0200
    [PATCH v2 05/11] mm/slab: clean-up kmem_cache_node setup js1304@gmail.com - 2016-04-12 07:00 +0200
    [PATCH v2 04/11] mm/slab: factor out kmem_cache_node initialization code js1304@gmail.com - 2016-04-12 07:00 +0200
      Re: [PATCH v2 04/11] mm/slab: factor out kmem_cache_node initialization  code Christoph Lameter <cl@linux.com> - 2016-04-12 19:00 +0200

#1376456 — [PATCH v2 00/11] mm/slab: reduce lock contention in alloc path

From: Joonsoo Kim <iamjoonsoo.kim@lge.com>

Major changes from v1
o hold node lock instead of slab_mutex in kmem_cache_shrink()
o fix suspend-to-ram issue reported by Nishanth
o use synchronize_sched() instead of kick_all_cpus_sync()

While processing concurrent allocation, SLAB could be contended
a lot because it did a lots of work with holding a lock. This
patchset try to reduce the number of critical section to reduce
lock contention. Major changes are lockless decision to allocate
more slab and lockless cpu cache refill from the newly allocated slab.

Below is the result of concurrent allocation/free in slab allocation
benchmark made by Christoph a long time ago. I make the output simpler.
The number shows cycle count during alloc/free respectively so less
is better.

* Before
Kmalloc N*alloc N*free(32): Average=365/806
Kmalloc N*alloc N*free(64): Average=452/690
Kmalloc N*alloc N*free(128): Average=736/886
Kmalloc N*alloc N*free(256): Average=1167/985
Kmalloc N*alloc N*free(512): Average=2088/1125
Kmalloc N*alloc N*free(1024): Average=4115/1184
Kmalloc N*alloc N*free(2048): Average=8451/1748
Kmalloc N*alloc N*free(4096): Average=16024/2048

* After
Kmalloc N*alloc N*free(32): Average=344/792
Kmalloc N*alloc N*free(64): Average=347/882
Kmalloc N*alloc N*free(128): Average=390/959
Kmalloc N*alloc N*free(256): Average=393/1067
Kmalloc N*alloc N*free(512): Average=683/1229
Kmalloc N*alloc N*free(1024): Average=1295/1325
Kmalloc N*alloc N*free(2048): Average=2513/1664
Kmalloc N*alloc N*free(4096): Average=4742/2172

It shows that performance improves greatly (roughly more than 50%)
for the object class whose size is more than 128 bytes.

Joonsoo Kim (11):
  mm/slab: fix the theoretical race by holding proper lock
  mm/slab: remove BAD_ALIEN_MAGIC again
  mm/slab: drain the free slab as much as possible
  mm/slab: factor out kmem_cache_node initialization code
  mm/slab: clean-up kmem_cache_node setup
  mm/slab: don't keep free slabs if free_objects exceeds free_limit
  mm/slab: racy access/modify the slab color
  mm/slab: make cache_grow() handle the page allocated on arbitrary node
  mm/slab: separate cache_grow() to two parts
  mm/slab: refill cpu cache through a new slab without holding a node
    lock
  mm/slab: lockless decision to grow cache

 mm/slab.c | 562 +++++++++++++++++++++++++++++++++-----------------------------
 1 file changed, 295 insertions(+), 267 deletions(-)

-- 
1.9.1

[toc] | [next] | [standalone]

#1376457 — [PATCH v2 05/11] mm/slab: clean-up kmem_cache_node setup

From: Joonsoo Kim <iamjoonsoo.kim@lge.com>

There are mostly same code for setting up kmem_cache_node either in
cpuup_prepare() or alloc_kmem_cache_node().  Factor out and clean-up them.

v2
o Rename setup_kmem_cache_node_node to setup_kmem_cache_nodes
o Fix suspend-to-ram issue reported by Nishanth

Tested-by: Nishanth Menon <nm@ti.com>
Tested-by: Jon Hunter <jonathanh@nvidia.com>
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
---
 mm/slab.c | 168 +++++++++++++++++++++++++-------------------------------------
 1 file changed, 68 insertions(+), 100 deletions(-)

diff --git a/mm/slab.c b/mm/slab.c
index 49af685..27cb390 100644
--- a/mm/slab.c
+++ b/mm/slab.c
@@ -898,6 +898,63 @@ static int init_cache_node_node(int node)
 	return 0;
 }
 
+static int setup_kmem_cache_node(struct kmem_cache *cachep,
+				int node, gfp_t gfp, bool force_change)
+{
+	int ret = -ENOMEM;
+	struct kmem_cache_node *n;
+	struct array_cache *old_shared = NULL;
+	struct array_cache *new_shared = NULL;
+	struct alien_cache **new_alien = NULL;
+	LIST_HEAD(list);
+
+	if (use_alien_caches) {
+		new_alien = alloc_alien_cache(node, cachep->limit, gfp);
+		if (!new_alien)
+			goto fail;
+	}
+
+	if (cachep->shared) {
+		new_shared = alloc_arraycache(node,
+			cachep->shared * cachep->batchcount, 0xbaadf00d, gfp);
+		if (!new_shared)
+			goto fail;
+	}
+
+	ret = init_cache_node(cachep, node, gfp);
+	if (ret)
+		goto fail;
+
+	n = get_node(cachep, node);
+	spin_lock_irq(&n->list_lock);
+	if (n->shared && force_change) {
+		free_block(cachep, n->shared->entry,
+				n->shared->avail, node, &list);
+		n->shared->avail = 0;
+	}
+
+	if (!n->shared || force_change) {
+		old_shared = n->shared;
+		n->shared = new_shared;
+		new_shared = NULL;
+	}
+
+	if (!n->alien) {
+		n->alien = new_alien;
+		new_alien = NULL;
+	}
+
+	spin_unlock_irq(&n->list_lock);
+	slabs_destroy(cachep, &list);
+
+fail:
+	kfree(old_shared);
+	kfree(new_shared);
+	free_alien_cache(new_alien);
+
+	return ret;
+}
+
 static void cpuup_canceled(long cpu)
 {
 	struct kmem_cache *cachep;
@@ -969,7 +1026,6 @@ free_slab:
 static int cpuup_prepare(long cpu)
 {
 	struct kmem_cache *cachep;
-	struct kmem_cache_node *n = NULL;
 	int node = cpu_to_mem(cpu);
 	int err;
 
@@ -988,44 +1044,9 @@ static int cpuup_prepare(long cpu)
 	 * array caches
 	 */
 	list_for_each_entry(cachep, &slab_caches, list) {
-		struct array_cache *shared = NULL;
-		struct alien_cache **alien = NULL;
-
-		if (cachep->shared) {
-			shared = alloc_arraycache(node,
-				cachep->shared * cachep->batchcount,
-				0xbaadf00d, GFP_KERNEL);
-			if (!shared)
-				goto bad;
-		}
-		if (use_alien_caches) {
-			alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
-			if (!alien) {
-				kfree(shared);
-				goto bad;
-			}
-		}
-		n = get_node(cachep, node);
-		BUG_ON(!n);
-
-		spin_lock_irq(&n->list_lock);
-		if (!n->shared) {
-			/*
-			 * We are serialised from CPU_DEAD or
-			 * CPU_UP_CANCELLED by the cpucontrol lock
-			 */
-			n->shared = shared;
-			shared = NULL;
-		}
-#ifdef CONFIG_NUMA
-		if (!n->alien) {
-			n->alien = alien;
-			alien = NULL;
-		}
-#endif
-		spin_unlock_irq(&n->list_lock);
-		kfree(shared);
-		free_alien_cache(alien);
+		err = setup_kmem_cache_node(cachep, node, GFP_KERNEL, false);
+		if (err)
+			goto bad;
 	}
 
 	return 0;
@@ -3676,72 +3697,19 @@ EXPORT_SYMBOL(kfree);
 /*
  * This initializes kmem_cache_node or resizes various caches for all nodes.
  */
-static int alloc_kmem_cache_node(struct kmem_cache *cachep, gfp_t gfp)
+static int setup_kmem_cache_nodes(struct kmem_cache *cachep, gfp_t gfp)
 {
+	int ret;
 	int node;
 	struct kmem_cache_node *n;
-	struct array_cache *new_shared;
-	struct alien_cache **new_alien = NULL;
 
 	for_each_online_node(node) {
-
-		if (use_alien_caches) {
-			new_alien = alloc_alien_cache(node, cachep->limit, gfp);
-			if (!new_alien)
-				goto fail;
-		}
-
-		new_shared = NULL;
-		if (cachep->shared) {
-			new_shared = alloc_arraycache(node,
-				cachep->shared*cachep->batchcount,
-					0xbaadf00d, gfp);
-			if (!new_shared) {
-				free_alien_cache(new_alien);
-				goto fail;
-			}
-		}
-
-		n = get_node(cachep, node);
-		if (n) {
-			struct array_cache *shared = n->shared;
-			LIST_HEAD(list);
-
-			spin_lock_irq(&n->list_lock);
-
-			if (shared)
-				free_block(cachep, shared->entry,
-						shared->avail, node, &list);
-
-			n->shared = new_shared;
-			if (!n->alien) {
-				n->alien = new_alien;
-				new_alien = NULL;
-			}
-			n->free_limit = (1 + nr_cpus_node(node)) *
-					cachep->batchcount + cachep->num;
-			spin_unlock_irq(&n->list_lock);
-			slabs_destroy(cachep, &list);
-			kfree(shared);
-			free_alien_cache(new_alien);
-			continue;
-		}
-		n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
-		if (!n) {
-			free_alien_cache(new_alien);
-			kfree(new_shared);
+		ret = setup_kmem_cache_node(cachep, node, gfp, true);
+		if (ret)
 			goto fail;
-		}
 
-		kmem_cache_node_init(n);
-		n->next_reap = jiffies + REAPTIMEOUT_NODE +
-				((unsigned long)cachep) % REAPTIMEOUT_NODE;
-		n->shared = new_shared;
-		n->alien = new_alien;
-		n->free_limit = (1 + nr_cpus_node(node)) *
-					cachep->batchcount + cachep->num;
-		cachep->node[node] = n;
 	}
+
 	return 0;
 
 fail:
@@ -3783,7 +3751,7 @@ static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
 	cachep->shared = shared;
 
 	if (!prev)
-		goto alloc_node;
+		goto setup_node;
 
 	for_each_online_cpu(cpu) {
 		LIST_HEAD(list);
@@ -3800,8 +3768,8 @@ static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
 	}
 	free_percpu(prev);
 
-alloc_node:
-	return alloc_kmem_cache_node(cachep, gfp);
+setup_node:
+	return setup_kmem_cache_nodes(cachep, gfp);
 }
 
 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
-- 
1.9.1

[toc] | [prev] | [next] | [standalone]

#1376458 — [PATCH v2 04/11] mm/slab: factor out kmem_cache_node initialization code

From: Joonsoo Kim <iamjoonsoo.kim@lge.com>

It can be reused on other place, so factor out it.  Following patch will
use it.

Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
---
 mm/slab.c | 68 ++++++++++++++++++++++++++++++++++++---------------------------
 1 file changed, 39 insertions(+), 29 deletions(-)

diff --git a/mm/slab.c b/mm/slab.c
index 5451929..49af685 100644
--- a/mm/slab.c
+++ b/mm/slab.c
@@ -841,6 +841,40 @@ static inline gfp_t gfp_exact_node(gfp_t flags)
 }
 #endif
 
+static int init_cache_node(struct kmem_cache *cachep, int node, gfp_t gfp)
+{
+	struct kmem_cache_node *n;
+
+	/*
+	 * Set up the kmem_cache_node for cpu before we can
+	 * begin anything. Make sure some other cpu on this
+	 * node has not already allocated this
+	 */
+	n = get_node(cachep, node);
+	if (n)
+		return 0;
+
+	n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
+	if (!n)
+		return -ENOMEM;
+
+	kmem_cache_node_init(n);
+	n->next_reap = jiffies + REAPTIMEOUT_NODE +
+		    ((unsigned long)cachep) % REAPTIMEOUT_NODE;
+
+	n->free_limit =
+		(1 + nr_cpus_node(node)) * cachep->batchcount + cachep->num;
+
+	/*
+	 * The kmem_cache_nodes don't come and go as CPUs
+	 * come and go.  slab_mutex is sufficient
+	 * protection here.
+	 */
+	cachep->node[node] = n;
+
+	return 0;
+}
+
 /*
  * Allocates and initializes node for a node on each slab cache, used for
  * either memory or cpu hotplug.  If memory is being hot-added, the kmem_cache_node
@@ -852,39 +886,15 @@ static inline gfp_t gfp_exact_node(gfp_t flags)
  */
 static int init_cache_node_node(int node)
 {
+	int ret;
 	struct kmem_cache *cachep;
-	struct kmem_cache_node *n;
-	const size_t memsize = sizeof(struct kmem_cache_node);
 
 	list_for_each_entry(cachep, &slab_caches, list) {
-		/*
-		 * Set up the kmem_cache_node for cpu before we can
-		 * begin anything. Make sure some other cpu on this
-		 * node has not already allocated this
-		 */
-		n = get_node(cachep, node);
-		if (!n) {
-			n = kmalloc_node(memsize, GFP_KERNEL, node);
-			if (!n)
-				return -ENOMEM;
-			kmem_cache_node_init(n);
-			n->next_reap = jiffies + REAPTIMEOUT_NODE +
-			    ((unsigned long)cachep) % REAPTIMEOUT_NODE;
-
-			/*
-			 * The kmem_cache_nodes don't come and go as CPUs
-			 * come and go.  slab_mutex is sufficient
-			 * protection here.
-			 */
-			cachep->node[node] = n;
-		}
-
-		spin_lock_irq(&n->list_lock);
-		n->free_limit =
-			(1 + nr_cpus_node(node)) *
-			cachep->batchcount + cachep->num;
-		spin_unlock_irq(&n->list_lock);
+		ret = init_cache_node(cachep, node, GFP_KERNEL);
+		if (ret)
+			return ret;
 	}
+
 	return 0;
 }
 
-- 
1.9.1

[toc] | [prev] | [next] | [standalone]

#1377023 — Re: [PATCH v2 04/11] mm/slab: factor out kmem_cache_node initialization code

Acked-by: Christoph Lameter <cl@linux.com>

[toc] | [prev] | [standalone]

Back to top | Article view | linux.kernel

csiph-web