Kernel memory accounting container (v5)

include/linux/slab.h line 32

Quoted 6 lines · Pavel Emelyanov

include/linux/slab.h +1 -0

@@ -28,6 +28,7 @@
 #define SLAB_DESTROY_BY_RCU	0x00080000UL	/* Defer freeing slabs to RCU */
 #define SLAB_MEM_SPREAD		0x00100000UL	/* Spread some memory over cpuset */
 #define SLAB_TRACE		0x00200000UL	/* Trace allocations and frees */
+#define SLAB_NOTIFY		0x00400000UL	/* Notify major events */
 

You need to add SLAB_NOTIFY to the SLUB_NEVER_MERGE group. That way you can solve the merging issue.

Otherwise this looks okay now.

True, but we mark the slubs as notifyable at runtime, after they are merged. However, once someone decides to make his slab notifyable from the very beginning this makes sense, thanks.

Oh, great! Does this mean I can put Acked-by when sending this to Andrew?

Quoted 3 lines · Pavel Emelyanov

True, but we mark the slubs as notifyable at runtime, after they are merged. However, once someone decides to make his slab notifyable from the very beginning this makes sense, thanks.

This also makes sense if a device driver later creates a new slab. Adding the bit will switch off additional merges for that slab.

Quoted 1 line

Oh, great! Does this mean I can put Acked-by when sending this to Andrew?

Yes.

According to Christoph, there are already multiple people who want to control slab allocations and track memory for various reasons. So this is an introduction of such a hooks.

Currently, functions that are to call the notifiers are empty and marked as "weak". Thus, if there's only _one_ listener to these events, it can happily link with the vmlinux and handle the events with more than 10% of performance saved.

The events tracked are:
1. allocation of an object;
2. freeing of an object;
3. allocation of a new page for objects;
4. freeing this page.

More events can be added on demand.

The kmem cache marked with SLAB_NOTIFY flag will cause all the events above to generate notifications. By default no caches come with this flag.

The events are generated on slow paths only and as soon as the cache is marked as SLAB_NOTIFY, it will always use them for allocation.

Signed-off-by Pavel Emelyanov <[email protected]>

---

include/linux/slab.h +1 -0

index f3a8eec..68d8e65 100644
@@ -28,6 +28,7 @@
 #define SLAB_DESTROY_BY_RCU	0x00080000UL	/* Defer freeing slabs to RCU */
 #define SLAB_MEM_SPREAD		0x00100000UL	/* Spread some memory over cpuset */
 #define SLAB_TRACE		0x00200000UL	/* Trace allocations and frees */
+#define SLAB_NOTIFY		0x00400000UL	/* Notify major events */
 
 /* The following flags affect the page allocator grouping pages by mobility */
 #define SLAB_RECLAIM_ACCOUNT	0x00020000UL		/* Objects are reclaimable */

include/linux/slub_def.h +18 -0

index 40801e7..8cfd9ff 100644
@@ -200,4 +203,22 @@ static __always_inline void *kmalloc_nod
 }
 #endif
 
+struct slub_notify_struct {
+	struct kmem_cache *cachep;
+	void *objp;
+	gfp_t gfp;
+};
+
+enum {
+	SLUB_ON,
+	SLUB_OFF,
+	SLUB_ALLOC,
+	SLUB_FREE,
+	SLUB_NEWPAGE,
+	SLUB_FREEPAGE,
+};
+
+int slub_register_notifier(struct notifier_block *nb);
+void slub_unregister_notifier(struct notifier_block *nb);
+
 #endif /* _LINUX_SLUB_DEF_H */

mm/slub.c +42 -1

index 31d04a3..e066a0e 100644
@@ -1040,6 +1040,43 @@ static inline unsigned long kmem_cache_f
 }
 #define slub_debug 0
 #endif
+
+/*
+ * notifiers
+ */
+
+int __attribute__ ((weak)) slub_alloc_notify(struct kmem_cache *cachep,
+		void *obj, gfp_t gfp)
+{
+	return 0;
+}
+
+void __attribute__ ((weak)) slub_free_notify(struct kmem_cache *cachep,
+		void *obj)
+{
+}
+
+int __attribute__ ((weak)) slub_newpage_notify(struct kmem_cache *cachep,
+		struct page *pg, gfp_t gfp)
+{
+	return 0;
+}
+
+void __attribute__ ((weak)) slub_freepage_notify(struct kmem_cache *cachep,
+		struct page *pg)
+{
+}
+
+int __attribute__ ((weak)) slub_on_notify(struct kmem_cache *cachep)
+{
+	return 0;
+}
+
+int __attribute__ ((weak)) slub_off_notify(struct kmem_cache *cachep)
+{
+	return 0;
+}
+
 /*
  * Slab allocation and freeing
  */

@@ -1063,7 +1184,11 @@ static struct page *allocate_slab(struct
 		page = alloc_pages_node(node, flags, s->order);
 
 	if (!page)
-		return NULL;
+		goto out;
+
+	if ((s->flags & SLAB_NOTIFY) &&
+			slub_newpage_notify(s, page, flags) < 0)
+		goto out_free;
 
 	mod_zone_page_state(page_zone(page),
 		(s->flags & SLAB_RECLAIM_ACCOUNT) ?

mm/slub.c +17 -1

+out:
+	return NULL;
 }
 
 static void setup_object(struct kmem_cache *s, struct page *page,

@@ -1103,7 +1233,7 @@ static struct page *new_slab(struct kmem
 	page->slab = s;
 	page->flags |= 1 << PG_slab;
 	if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
-			SLAB_STORE_USER | SLAB_TRACE))
+			SLAB_STORE_USER | SLAB_TRACE | SLAB_NOTIFY))
 		SetSlabDebug(page);
 
 	start = page_address(page);

@@ -1158,6 +1288,9 @@ static void rcu_free_slab(struct rcu_hea
 
 static void free_slab(struct kmem_cache *s, struct page *page)
 {
+	if (s->flags & SLAB_NOTIFY)
+		slub_freepage_notify(s, page);
+
 	if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
 		/*
 		 * RCU free overloads the RCU head over the LRU

@@ -1548,9 +1681,16 @@ debug:
 	if (!alloc_debug_processing(s, c->page, object, addr))
 		goto another_slab;
 
+	if ((s->flags & SLAB_NOTIFY) &&
+			slub_alloc_notify(s, object, gfpflags) < 0) {
+		object = NULL;
+		goto out;
+	}
+
 	c->page->inuse++;
 	c->page->freelist = object[c->offset];
 	c->node = -1;
+out:
 	slab_unlock(c->page);
 	return object;
 }

@@ -1659,6 +1799,10 @@ slab_empty:
 debug:
 	if (!free_debug_processing(s, page, x, addr))
 		goto out_unlock;
+
+	if (s->flags & SLAB_NOTIFY)
+		slub_free_notify(s, x);
+
 	goto checks_ok;
 }
 

Quoted 4 lines · Pavel Emelyanov

According to Christoph, there are already multiple people who want to control slab allocations and track memory for various reasons. So this is an introduction of such a hooks.

Quoted 5 lines

Currently, functions that are to call the notifiers are empty and marked as "weak". Thus, if there's only _one_ listener to these events, it can happily link with the vmlinux and handle the events with more than 10% of performance saved.

Please check that weak objects work across platforms. Please see
http://www.ussg.iu.edu/hypermail/linux/kernel/0706.0/0593.html

Quoted 123 lines

The events tracked are:
1. allocation of an object;
2. freeing of an object;
3. allocation of a new page for objects;
4. freeing this page.

More events can be added on demand.

The kmem cache marked with SLAB_NOTIFY flag will cause all the events above to generate notifications. By default no caches come with this flag.

The events are generated on slow paths only and as soon as the cache is marked as SLAB_NOTIFY, it will always use them for allocation.

Signed-off-by Pavel Emelyanov <[email protected]>

---

include/linux/slab.h +1 -0

index f3a8eec..68d8e65 100644
@@ -28,6 +28,7 @@
 #define SLAB_DESTROY_BY_RCU	0x00080000UL	/* Defer freeing slabs to RCU */
 #define SLAB_MEM_SPREAD		0x00100000UL	/* Spread some memory over cpuset */
 #define SLAB_TRACE		0x00200000UL	/* Trace allocations and frees */
+#define SLAB_NOTIFY		0x00400000UL	/* Notify major events */
 
 /* The following flags affect the page allocator grouping pages by mobility */
 #define SLAB_RECLAIM_ACCOUNT	0x00020000UL		/* Objects are reclaimable */

include/linux/slub_def.h +18 -0

index 40801e7..8cfd9ff 100644
@@ -200,4 +203,22 @@ static __always_inline void *kmalloc_nod
 }
 #endif
 
+struct slub_notify_struct {
+	struct kmem_cache *cachep;
+	void *objp;
+	gfp_t gfp;
+};
+
+enum {
+	SLUB_ON,
+	SLUB_OFF,
+	SLUB_ALLOC,
+	SLUB_FREE,
+	SLUB_NEWPAGE,
+	SLUB_FREEPAGE,
+};
+
+int slub_register_notifier(struct notifier_block *nb);
+void slub_unregister_notifier(struct notifier_block *nb);
+
 #endif /* _LINUX_SLUB_DEF_H */

mm/slub.c +42 -1

index 31d04a3..e066a0e 100644
@@ -1040,6 +1040,43 @@ static inline unsigned long kmem_cache_f
 }
 #define slub_debug 0
 #endif
+
+/*
+ * notifiers
+ */
+
+int __attribute__ ((weak)) slub_alloc_notify(struct kmem_cache *cachep,
+		void *obj, gfp_t gfp)
+{
+	return 0;
+}
+
+void __attribute__ ((weak)) slub_free_notify(struct kmem_cache *cachep,
+		void *obj)
+{
+}
+
+int __attribute__ ((weak)) slub_newpage_notify(struct kmem_cache *cachep,
+		struct page *pg, gfp_t gfp)
+{
+	return 0;
+}
+
+void __attribute__ ((weak)) slub_freepage_notify(struct kmem_cache *cachep,
+		struct page *pg)
+{
+}
+
+int __attribute__ ((weak)) slub_on_notify(struct kmem_cache *cachep)
+{
+	return 0;
+}
+
+int __attribute__ ((weak)) slub_off_notify(struct kmem_cache *cachep)
+{
+	return 0;
+}
+
 /*
  * Slab allocation and freeing
  */

@@ -1063,7 +1184,11 @@ static struct page *allocate_slab(struct
 		page = alloc_pages_node(node, flags, s->order);
 
 	if (!page)
-		return NULL;
+		goto out;
+
+	if ((s->flags & SLAB_NOTIFY) &&
+			slub_newpage_notify(s, page, flags) < 0)
+		goto out_free;
 
 	mod_zone_page_state(page_zone(page),
 		(s->flags & SLAB_RECLAIM_ACCOUNT) ?

@@ -1071,6 +1196,11 @@ static struct page *allocate_slab(struct
 		pages);
 

mm/slub.c static struct page *allocate_slab(struct · line 1201

Quoted 4 lines

mm/slub.c +3 -0

 	return page;
+
+out_free:
+	__free_pages(page, s->order);

allocate_slab fails if sub_newpage_notify() fails? Sounds a bit

harsh, hard to review since the definition of the above function is not known.

Quoted 53 lines

mm/slub.c +17 -1

+out:
+	return NULL;
 }
 
 static void setup_object(struct kmem_cache *s, struct page *page,

@@ -1103,7 +1233,7 @@ static struct page *new_slab(struct kmem
 	page->slab = s;
 	page->flags |= 1 << PG_slab;
 	if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
-			SLAB_STORE_USER | SLAB_TRACE))
+			SLAB_STORE_USER | SLAB_TRACE | SLAB_NOTIFY))
 		SetSlabDebug(page);
 
 	start = page_address(page);

@@ -1158,6 +1288,9 @@ static void rcu_free_slab(struct rcu_hea
 
 static void free_slab(struct kmem_cache *s, struct page *page)
 {
+	if (s->flags & SLAB_NOTIFY)
+		slub_freepage_notify(s, page);
+
 	if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
 		/*
 		 * RCU free overloads the RCU head over the LRU

@@ -1548,9 +1681,16 @@ debug:
 	if (!alloc_debug_processing(s, c->page, object, addr))
 		goto another_slab;
 
+	if ((s->flags & SLAB_NOTIFY) &&
+			slub_alloc_notify(s, object, gfpflags) < 0) {
+		object = NULL;
+		goto out;
+	}
+
 	c->page->inuse++;
 	c->page->freelist = object[c->offset];
 	c->node = -1;
+out:
 	slab_unlock(c->page);
 	return object;
 }

@@ -1659,6 +1799,10 @@ slab_empty:
 debug:
 	if (!free_debug_processing(s, page, x, addr))
 		goto out_unlock;
+
+	if (s->flags & SLAB_NOTIFY)
+		slub_free_notify(s, x);
+
 	goto checks_ok;
 }
 

Quoted 2 lines

harsh, hard to review since the definition of the above function is not known.

What do you mean by "not known"? I is declared above :)

Quoted 4 lines · Pavel Emelyanov

Quoted 2 lines

harsh, hard to review since the definition of the above function is not known.

What do you mean by "not known"? I is declared above :)

I meant, we only know the stub weak functions. Does slub_*_notify fail only on fatal errors? If not, then the failure might be

unfair.

Quoted 3 lines · Balbir Singh

Quoted 2 lines · Pavel Emelyanov

Quoted 2 lines

harsh, hard to review since the definition of the above function is not known.

Quoted 4 lines

Quoted 2 lines

What do you mean by "not known"? I is declared above :)

I meant, we only know the stub weak functions. Does slub_*_notify fail only on fatal errors? If not, then the failure might be

It depends on the listener of the event when to fail. The slub notification mechanism is not the accounting at all it just sends notifications.

Quoted 2 lines

unfair.

If the user wants more than one listener for SLUB event, it can register them all as notifier_block-s so all of them will be notified.

This costs us about 10% of performance loss, in comparison with static linking.

The selected method of notification is srcu notifier blocks. This is selected because the "call" path, i.e. when the notification is done, is lockless and at the same time the handlers can sleep. Neither regular nor atomic notifiers provide such facilities.

Signed-off-by Pavel Emelyanov <[email protected]>

---

init/Kconfig +0 -0

index 684ccfb..e9acc29 100644

init/Kconfig +1 -0

+#endif
 	printk(KERN_INFO "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
 		" CPUs=%d, Nodes=%d\n",
 		caches, cache_line_size(),

Quoted 25 lines · Pavel Emelyanov

If the user wants more than one listener for SLUB event, it can register them all as notifier_block-s so all of them will be notified.

This costs us about 10% of performance loss, in comparison with static linking.

The selected method of notification is srcu notifier blocks. This is selected because the "call" path, i.e. when the notification is done, is lockless and at the same time the handlers can sleep. Neither regular nor atomic notifiers provide such facilities.

Signed-off-by Pavel Emelyanov <[email protected]>

---

init/Kconfig +0 -0

index 684ccfb..e9acc29 100644
@@ -593,6 +599,16 @@ config SLUB_DEBUG
 	  SLUB sysfs support. /sys/slab will not exist and there will be
 	  no support for cache validation etc.
 

init/Kconfig config SLUB_DEBUG · line 603

Quoted 2 lines

init/Kconfig +2 -0

+config SLUB_NOTIFY
+	default y

Should the default be on? Shouldn't it depend on KMEM?

Quoted 6 lines

init/Kconfig +6 -0

+	bool "Enable SLUB events generic notification"
+	depends on SLUB
+	help
+	  When Y, this option enables generic notifications of some major
+	  slub events. However, if you do know that there will be the
+	  only listener for them, you may say N here, so that callbacks

init/Kconfig config SLUB_DEBUG · line 613

Quoted 6 lines

init/Kconfig +4 -2

+	  will be called directly.
+
-
-
+#ifdef CONFIG_SLUB_NOTIFY
+	srcu_init_notifier_head(&slub_nb);

Can we get rid of the #ifdef CONFIG_SLUB_NOTIFY?

Quoted 5 lines

init/Kconfig +1 -0

+#endif
 	printk(KERN_INFO "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
 		" CPUs=%d, Nodes=%d\n",
 		caches, cache_line_size(),

Quoted 1 line

Should the default be on? Shouldn't it depend on KMEM?

Well... I think that is should be N by default and has noting to do with the KMEM :) Thanks for noticing.

Quoted 4 lines

mm/slub.c void __init kmem_cache_init(void) · line 2941

Quoted 2 lines

mm/slub.c +2 -0

+#ifdef CONFIG_SLUB_NOTIFY
+	srcu_init_notifier_head(&slub_nb);

Can we get rid of the #ifdef CONFIG_SLUB_NOTIFY?

I don't think this is really useful in the __init code :)

mm/slub.c static ssize_t defrag_ratio_store(struct · line 3962

Quoted 9 lines · Pavel Emelyanov

mm/slub.c +9 -0

+		for_each_node_state(n, N_NORMAL_MEMORY) {
+			struct kmem_cache_node *node;
+			struct page *pg;
+			
+			node = get_node(s, n);
+			spin_lock_irq(&node->list_lock);
+			list_for_each_entry(pg, &node->partial, lru)
+				SetSlabDebug(pg);
+			spin_unlock_irq(&node->list_lock);

Is it necessary to mark all the existing slabs with SLAB_DEBUG? Would it

not be sufficient to just mark the new ones coming up? The above operation could be a bit expensive.

Quoted 1 line

Is it necessary to mark all the existing slabs with SLAB_DEBUG? Would it

Yup. Otherwise we can never receive a single event e.g. if we make alloc/free in a loop, or similar, so that new slabs simply are not created.

Quoted 4 lines · Pavel Emelyanov

Quoted 1 line

Is it necessary to mark all the existing slabs with SLAB_DEBUG? Would it

Yup. Otherwise we can never receive a single event e.g. if we make alloc/free in a loop, or similar, so that new slabs simply are not created.

Right but on the other hand: If objects in these slabs are freed then your counters get decremented and may go negative.

The /sys/slab/<name>/cache_notify attribute controls whether the cache <name> is to be accounted or not.

By default no caches are accountable. Simply make

  # echo -n 1 > /sys/slab/<name>cache_notify
to turn notification of this cache on.

If we turn accounting on on some cache and this cache is merged with some other, this "other" will be notified as well. One can solve this by disabling of cache merging, i.e. booting with slub_nomerge option to the kernel.

Turning the notifications "on" causes all te subsequent allocations use the slow allocation path, so all the per-cpu caches are flushed and all the partial pages are marked as SlabDebug.

---

mm/slub.c +2 -0

index 31d04a3..e066a0e 100644
@@ -3779,6 +3932,60 @@ static ssize_t defrag_ratio_store(struct
 SLAB_ATTR(defrag_ratio);
 #endif
 
+static ssize_t cache_notify_show(struct kmem_cache *s, char *buf)
+{

mm/slub.c +4 -0

+		for_each_node_state(n, N_NORMAL_MEMORY) {
+			struct kmem_cache_node *node;
+			struct page *pg;
+			

mm/slub.c +22 -0

+		/*
+		 * TODO: make the notifications-off work
+		 */
+		if (any_slab_objects(s))
+			return -EBUSY;
+
+		s->flags &= ~SLAB_NOTIFY;
+		err = slub_off_notify(s);
+		if (err < 0) {
+			s->flags |= SLAB_NOTIFY;
+			return err;
+		}
+
+		return length;
+	}
+
+	return -EINVAL;
+}
+
+SLAB_ATTR(cache_notify);
+
 static struct attribute * slab_attrs[] = {
 	&slab_size_attr.attr,
 	&object_size_attr.attr,

@@ -3809,6 +4016,7 @@ static struct attribute * slab_attrs[] =
 #ifdef CONFIG_NUMA
 	&defrag_ratio_attr.attr,
 #endif
+	&cache_notify_attr.attr,
 	NULL
 };
 

Quoted 12 lines · Pavel Emelyanov

The /sys/slab/<name>/cache_notify attribute controls whether the cache <name> is to be accounted or not.

By default no caches are accountable. Simply make

  # echo -n 1 > /sys/slab/<name>cache_notify
to turn notification of this cache on.

If we turn accounting on on some cache and this cache is merged with some other, this "other" will be notified as well. One can solve this by disabling of cache merging, i.e. booting with slub_nomerge option to the kernel.

Quoted 5 lines

Turning the notifications "on" causes all te subsequent allocations use the slow allocation path, so all the per-cpu caches are flushed and all the partial pages are marked as SlabDebug.

Do we know of the overhead of slub notifications? It would be nice to know and probably add it to the help text in Kconfig.

Quoted 13 lines

---

mm/slub.c +2 -0

index 31d04a3..e066a0e 100644
@@ -3779,6 +3932,60 @@ static ssize_t defrag_ratio_store(struct
 SLAB_ATTR(defrag_ratio);
 #endif
 
+static ssize_t cache_notify_show(struct kmem_cache *s, char *buf)
+{

mm/slub.c static ssize_t defrag_ratio_store(struct · line 3946

Quoted 10 lines

mm/slub.c +10 -0

+	return sprintf(buf, "%d\n", !!(s->flags & SLAB_NOTIFY));
+}
+
+static ssize_t cache_notify_store(struct kmem_cache *s,
+		const char *buf, size_t length)
+{
+	int err;
+	int n;
+
+	if ((buf[0] == '1') && !(s->flags & SLAB_NOTIFY)) {

Won't this code break if I pass 10 as input in buf?

mm/slub.c static ssize_t defrag_ratio_store(struct · line 3953

Quoted 7 lines

mm/slub.c +7 -0

+		err = slub_on_notify(s);
+		if (err < 0)
+			return err;
+
+		s->flags |= SLAB_NOTIFY;
+
+		flush_all(s);

Sounds like an expensive operation, can we add a comment to indicate the same. Also, in the documentation, could we warn the user that turning notifications on/off can be expensive?

Quoted 4 lines

mm/slub.c +4 -0

+		for_each_node_state(n, N_NORMAL_MEMORY) {
+			struct kmem_cache_node *node;
+			struct page *pg;
+			

mm/slub.c static ssize_t defrag_ratio_store(struct · line 3967

Quoted 10 lines

mm/slub.c +10 -0

+			node = get_node(s, n);
+			spin_lock_irq(&node->list_lock);
+			list_for_each_entry(pg, &node->partial, lru)
+				SetSlabDebug(pg);
+			spin_unlock_irq(&node->list_lock);
+		}
+		return length;
+	}
+
+	if ((buf[0] == '0') && (s->flags & SLAB_NOTIFY)) {

What happens if buf is 01?

Quoted 33 lines

mm/slub.c +22 -0

+		/*
+		 * TODO: make the notifications-off work
+		 */
+		if (any_slab_objects(s))
+			return -EBUSY;
+
+		s->flags &= ~SLAB_NOTIFY;
+		err = slub_off_notify(s);
+		if (err < 0) {
+			s->flags |= SLAB_NOTIFY;
+			return err;
+		}
+
+		return length;
+	}
+
+	return -EINVAL;
+}
+
+SLAB_ATTR(cache_notify);
+
 static struct attribute * slab_attrs[] = {
 	&slab_size_attr.attr,
 	&object_size_attr.attr,

@@ -3809,6 +4016,7 @@ static struct attribute * slab_attrs[] =
 #ifdef CONFIG_NUMA
 	&defrag_ratio_attr.attr,
 #endif
+	&cache_notify_attr.attr,
 	NULL
 };
 

Quoted 8 lines

mm/slub.c +0 -0

index 31d04a3..e066a0e 100644
@@ -3779,6 +3932,60 @@ static ssize_t defrag_ratio_store(struct
 SLAB_ATTR(defrag_ratio);
 #endif
 

mm/slub.c static ssize_t defrag_ratio_store(struct · line 3946

Quoted 12 lines

mm/slub.c +12 -0

+static ssize_t cache_notify_show(struct kmem_cache *s, char *buf)
+{
+	return sprintf(buf, "%d\n", !!(s->flags & SLAB_NOTIFY));
+}
+
+static ssize_t cache_notify_store(struct kmem_cache *s,
+		const char *buf, size_t length)
+{
+	int err;
+	int n;
+
+	if ((buf[0] == '1') && !(s->flags & SLAB_NOTIFY)) {

Won't this code break if I pass 10 as input in buf?

Quoted 4 lines

mm/slub.c +4 -0

+		err = slub_on_notify(s);
+		if (err < 0)
+			return err;
+

mm/slub.c static ssize_t defrag_ratio_store(struct · line 3953

Quoted 3 lines

mm/slub.c +3 -0

+		s->flags |= SLAB_NOTIFY;
+
+		flush_all(s);

Sounds like an expensive operation, can we add a comment to indicate the same. Also, in the documentation, could we warn the user that turning notifications on/off can be expensive?

Quoted 4 lines

mm/slub.c +4 -0

+		for_each_node_state(n, N_NORMAL_MEMORY) {
+			struct kmem_cache_node *node;
+			struct page *pg;
+			

mm/slub.c static ssize_t defrag_ratio_store(struct · line 3967

Quoted 10 lines

mm/slub.c +10 -0

+			node = get_node(s, n);
+			spin_lock_irq(&node->list_lock);
+			list_for_each_entry(pg, &node->partial, lru)
+				SetSlabDebug(pg);
+			spin_unlock_irq(&node->list_lock);
+		}
+		return length;
+	}
+
+	if ((buf[0] == '0') && (s->flags & SLAB_NOTIFY)) {

What happens if buf is 01?

Quoted 33 lines

mm/slub.c +22 -0

+		/*
+		 * TODO: make the notifications-off work
+		 */
+		if (any_slab_objects(s))
+			return -EBUSY;
+
+		s->flags &= ~SLAB_NOTIFY;
+		err = slub_off_notify(s);
+		if (err < 0) {
+			s->flags |= SLAB_NOTIFY;
+			return err;
+		}
+
+		return length;
+	}
+
+	return -EINVAL;
+}
+
+SLAB_ATTR(cache_notify);
+
 static struct attribute * slab_attrs[] = {
 	&slab_size_attr.attr,
 	&object_size_attr.attr,

@@ -3809,6 +4016,7 @@ static struct attribute * slab_attrs[] =
 #ifdef CONFIG_NUMA
 	&defrag_ratio_attr.attr,
 #endif
+	&cache_notify_attr.attr,
 	NULL
 };
 

Quoted 3 lines

mm/slub.c static ssize_t defrag_ratio_store(struct · line 3946

Quoted 2 lines

mm/slub.c +1 -0

+	if ((buf[0] == '1') && !(s->flags & SLAB_NOTIFY)) {

Won't this code break if I pass 10 as input in buf?

I don't care about this :) And no /sys/slab/<xxx>/ file does. It's the problem of a user if he pass some shit and doesn't see the notifications turned on/off ;)

Quoted 6 lines

Quoted 2 lines

Review

Quoted 1 line

(quoted diff) +1 -0

+	if ((buf[0] == '1') && !(s->flags & SLAB_NOTIFY)) {

Won't this code break if I pass 10 as input in buf?

I don't care about this :) And no /sys/slab/<xxx>/ file does. It's the problem of a user if he pass some shit and doesn't see the notifications turned on/off ;)

Well, we don't want confused users? How does a user know what is a valid value to pass? Are you saying that we don't need

to care about user interface, if so I don't agree.

Quoted 6 lines · Balbir Singh

Quoted 6 lines

Quoted 2 lines

Review

Quoted 1 line

(quoted diff) +1 -0

+	if ((buf[0] == '1') && !(s->flags & SLAB_NOTIFY)) {

Won't this code break if I pass 10 as input in buf?

I don't care about this :) And no /sys/slab/<xxx>/ file does. It's the problem of a user if he pass some shit and doesn't see the notifications turned on/off ;)

Quoted 2 lines

Well, we don't want confused users? How does a user know what is a valid value to pass? Are you saying that we don't need

The rules are simple - 1 to turn them on and 0 to turn them off and nothing more. These values user sees when he "cat"s this file.

Quoted 1 line

to care about user interface, if so I don't agree.

That's the convention used in the /sys/slab/<x>/ files - the 1 means on the 0 means 0, all the others are not guaranteed :)

Quoted 9 lines · Pavel Emelyanov

Quoted 8 lines · Balbir Singh

Quoted 6 lines

Quoted 2 lines

Review

Quoted 1 line

(quoted diff) +1 -0

+	if ((buf[0] == '1') && !(s->flags & SLAB_NOTIFY)) {

Won't this code break if I pass 10 as input in buf?

I don't care about this :) And no /sys/slab/<xxx>/ file does. It's the problem of a user if he pass some shit and doesn't see the notifications turned on/off ;)

Well, we don't want confused users? How does a user know what is a valid value to pass? Are you saying that we don't need

Quoted 3 lines

The rules are simple - 1 to turn them on and 0 to turn them off and nothing more. These values user sees when he "cat"s this file.

I think correct error reporting is critical, if the user sets the value to 01 and that ends up disabling slab notifications, the handling is wrong.

Quoted 4 lines

Quoted 1 line

to care about user interface, if so I don't agree.

That's the convention used in the /sys/slab/<x>/ files - the 1 means on the 0 means 0, all the others are not guaranteed :)

Is this documented somewhere or is this interpreted from looking at the code of other file handlers?

Is this documented somewhere or is this interpreted from looking at the code of other file handlers?

Attach the controller to the control groups. This will work with the SLUB allocator only. However, if we need I can port this on SLAB (and maybe SLOB ;) ).

This setup is simple and stupid.

mm/kmemcontrol.c new file +9 -0

+static struct cftype kmem_files[] = {
+	{
+		.name = "usage",
+		.private = RES_USAGE,
+		.read = kmem_container_read,
+	},
+	{
+		.name = "limit",
+		.private = RES_LIMIT,

mm/kmemcontrol.c new file +13 -0

+static int kmem_populate(struct cgroup_subsys *ss, struct cgroup *cnt)
+{
+	return cgroup_add_files(cnt, ss, kmem_files, ARRAY_SIZE(kmem_files));
+}
+
+struct cgroup_subsys kmem_subsys = {
+	.name = "kmem",
+	.create = kmem_create,
+	.destroy  = kmem_destroy,
+	.populate = kmem_populate,
+	.subsys_id = kmem_subsys_id,
+	.early_init = 1,
+};

Quoted 6 lines · Pavel Emelyanov

Attach the controller to the control groups. This will work with the SLUB allocator only. However, if we need I can port this on SLAB (and maybe SLOB ;) ).

This setup is simple and stupid.

That makes it easier for me to review it:)

Quoted 15 lines

mm/kmemcontrol.c new file +15 -0

+static struct cftype kmem_files[] = {
+	{
+		.name = "usage",
+		.private = RES_USAGE,
+		.read = kmem_container_read,
+	},
+	{
+		.name = "limit",
+		.private = RES_LIMIT,
+		.write = kmem_container_write,
+		.read = kmem_container_read,
+	},
+	{
+		.name = "failcnt",
+		.private = RES_FAILCNT,

mm/kmemcontrol.c line 106

Quoted 4 lines

mm/kmemcontrol.c new file +4 -0

+		.read = kmem_container_read,
+	},
+};
+

Could we make the user interface similar to the one in the memory controller please! It would make it easier for users to configure and control both.

Quoted 14 lines

mm/kmemcontrol.c new file +13 -0

+static int kmem_populate(struct cgroup_subsys *ss, struct cgroup *cnt)
+{
+	return cgroup_add_files(cnt, ss, kmem_files, ARRAY_SIZE(kmem_files));
+}
+
+struct cgroup_subsys kmem_subsys = {
+	.name = "kmem",
+	.create = kmem_create,
+	.destroy  = kmem_destroy,
+	.populate = kmem_populate,
+	.subsys_id = kmem_subsys_id,
+	.early_init = 1,
+};

Quoted 3 lines

Could we make the user interface similar to the one in the memory controller please! It would make it easier for users to configure and control both.

Do you mean by merging the similar code or by making them look like the same? If the second, than can you point at least one difference except for "usage" vs "usage_in_bytes" (and similar) naming?

Quoted 7 lines · Pavel Emelyanov

Quoted 3 lines

Could we make the user interface similar to the one in the memory controller please! It would make it easier for users to configure and control both.

Do you mean by merging the similar code or by making them look like the same? If the second, than can you point at least one difference except for "usage" vs "usage_in_bytes" (and similar) naming?

I meant similar naming, it makes it is easier if the naming convention is the same.

Quoted 8 lines · Balbir Singh

Quoted 7 lines · Pavel Emelyanov

Quoted 3 lines

Could we make the user interface similar to the one in the memory controller please! It would make it easier for users to configure and control both.

Do you mean by merging the similar code or by making them look like the same? If the second, than can you point at least one difference except for "usage" vs "usage_in_bytes" (and similar) naming?

Quoted 2 lines

I meant similar naming, it makes it is easier if the naming convention is the same.

Ok. It looks like the naming in the memcontrol.c has changed after I copied it from there :)

BTW, after Andrew accept these patches I think I will spend some time thinking about how to consolidate the res_counter files.

Thanks,
Pavel

Quoted 11 lines · Pavel Emelyanov

Quoted 10 lines · Balbir Singh

Quoted 7 lines · Pavel Emelyanov

Quoted 3 lines

Could we make the user interface similar to the one in the memory controller please! It would make it easier for users to configure and control both.

Do you mean by merging the similar code or by making them look like the same? If the second, than can you point at least one difference except for "usage" vs "usage_in_bytes" (and similar) naming?

I meant similar naming, it makes it is easier if the naming convention is the same.

Quoted 6 lines

Ok. It looks like the naming in the memcontrol.c has changed after I copied it from there :)

BTW, after Andrew accept these patches I think I will spend some time thinking about how to consolidate the res_counter files.

Yes, we need to do that. Paul Menage wanted to pull in resource counters

into cgroups. Also, I'd like to be able to share res_counters between container groups or at-least support a hierarchy. We could discuss this at that point and/or send out an RFC to get more comments.

Quoted 2 lines

Thanks,
Pavel

Yes, we need to do that. Paul Menage wanted to pull in resource counters into cgroups.

Also, I'd like to be able to share res_counters between container groups or at-least support a hierarchy.

Quoted 2 lines · Balbir Singh

Yes, we need to do that. Paul Menage wanted to pull in resource counters into cgroups.

No, that's not what I wanted - I just wanted to try to cut down the boilerplate required to use resource counters. Ideally it would be as simple as including a call to some function such as "add_res_counter(...)" which would deal with setting up all the cgroup control files, provide handlers, etc. This also helps to keep all the APIs consistent.

See the patch that I sent last week entitled "Simplify memory controller and resource counter I/O" for a first step in that direction.

Quoted 2 lines

Also, I'd like to be able to share res_counters between container groups or at-least support a hierarchy.

A hierarchy sounds like a good idea.

For the sharing, there are two approaches - any subsystem that wants sharing has to implement it itself, or else implement sharing at the control groups level and make it (mostly) transparent to the subsystems. I think I prefer the latter and have been playing with some ways to do this.

Paul

Excellent, I prefer the later as well, but it would mean overheads for controllers not using the hierarchy.

a design such that parents<->children can effectively share resources, track them and do so recursively, that would be really nice.

Quoted 4 lines · Pavel Emelyanov

Quoted 3 lines

Could we make the user interface similar to the one in the memory controller please! It would make it easier for users to configure and control both.

Quoted 4 lines

Do you mean by merging the similar code or by making them look like the same? If the second, than can you point at least one difference except for "usage" vs "usage_in_bytes" (and similar) naming?

I meant similar naming, it makes it is easier if the naming convention is the same.

The struct page gets an extra pointer (just like it has with the RSS controller) and this pointer points to the array of the kmem_container pointers - one for each object stored on that page itself.

Thus the i'th object on the page is accounted to the container pointed by the i'th pointer on that array and when the object is freed we unaccount its size to this particular container, not the container current task belongs to.

This is done so, because the context objects are freed is most often not the same as the one this objects was allocated in (due to RCU and reference counters).

This controller disables the kmalloc caches accounting, since we cannot just track all the kmalloc calls, but we need to explicitly specify which kmalloc do we want to account for with (e.g.) additional GFP flag.

Signed-off-by Pavel Emelyanov <[email protected]>

---

include/linux/mm_types.h +6 -1

index f83bd17..8e1ef21 100644
@@ -83,9 +83,14 @@ struct page {
 	void *virtual;			/* Kernel virtual address (NULL if
 					   not kmapped, ie. highmem) */
 #endif /* WANT_PAGE_VIRTUAL */
+	union {
 #ifdef CONFIG_CGROUP_MEM_CONT
-	unsigned long page_cgroup;
+		unsigned long page_cgroup;
+#endif
+#ifdef CONFIG_CGROUP_KMEM
+		struct kmem_container **cgroups;
 #endif
+	};
 #ifdef CONFIG_PAGE_OWNER
 	int order;
 	unsigned int gfp_mask;

include/linux/slub_def.h +3 -0

index 40801e7..8cfd9ff 100644
@@ -69,6 +69,9 @@ struct kmem_cache {
 #endif
 };
 
+int slab_index(void *p, struct kmem_cache *s, void *addr);
+int is_kmalloc_cache(struct kmem_cache *s);
+
 /*
  * Kmalloc subsystem.
  */

mm/Makefile +1 -0

index 81232a1..f7ec4b7 100644
@@ -31,4 +31,5 @@ obj-$(CONFIG_MIGRATION) += migrate.o
 obj-$(CONFIG_SMP) += allocpercpu.o
 obj-$(CONFIG_QUICKLIST) += quicklist.o
 obj-$(CONFIG_CGROUP_MEM_CONT) += memcontrol.o
+obj-$(CONFIG_CGROUP_KMEM) += kmemcontrol.o
 

mm/kmemcontrol.c new file +43 -0

new file mode 100644
index 0000000..5698c0f
@@ -1,6 +1,9 @@
  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  * GNU General Public License for more details.
+ *
+ * Changelog:
+ *   2007 Pavel Emelyanov : Add slub accounting
  */
 
 #include <linux/mm.h>

@@ -120,3 +129,144 @@
 	.subsys_id = kmem_subsys_id,
 	.early_init = 1,
 };
+
+/*
+ * slub accounting
+ */
+
+int slub_newpage_notify(struct kmem_cache *s, struct page *pg, gfp_t flags)
+{
+	struct kmem_container **ptr;
+
+	ptr = kzalloc(s->objects * sizeof(struct kmem_container *), flags);
+	if (ptr == NULL)
+		return -ENOMEM;
+
+	pg->cgroups = ptr;
+	return 0;
+}
+
+void slub_freepage_notify(struct kmem_cache *s, struct page *pg)
+{
+	struct kmem_container **ptr;
+
+	ptr = pg->cgroups;
+	if (ptr == NULL)
+		return;
+
+	kfree(ptr);
+	pg->cgroups = NULL;
+}
+
+int slub_alloc_notify(struct kmem_cache *s, void *obj, gfp_t gfp)
+{
+	struct page *pg;
+	struct kmem_container *cnt;
+	struct kmem_container **obj_container;
+
+	pg = virt_to_head_page(obj);
+	obj_container = pg->cgroups;
+	if (unlikely(obj_container == NULL)) {
+		/*
+		 * turned on after some objects were allocated

mm/kmemcontrol.c new file +24 -0

+		goto err_locked;
+
+	css_get(&cnt->css);
+	rcu_read_unlock();
+	obj_container[slab_index(obj, s, page_address(pg))] = cnt;
+	return 0;
+
+err_locked:
+	rcu_read_unlock();
+err:
+	return -ENOMEM;
+}
+
+void slub_free_notify(struct kmem_cache *s, void *obj)
+{
+	struct page *pg;
+	struct kmem_container *cnt;
+	struct kmem_container **obj_container;
+
+	pg = virt_to_head_page(obj);
+	obj_container = pg->cgroups;
+	if (obj_container == NULL)
+		return;
+

mm/kmemcontrol.c new file +52 -0

+int slub_off_notify(struct kmem_cache *cachep)
+{
+	return 0;
+}
+
+#ifdef CONFIG_SLUB_NOTIFY
+static int kmem_notify(struct notifier_block *nb, unsigned long cmd, void *arg)
+{
+	int ret;
+	struct slub_notify_struct *ns;
+
+	ns = (struct slub_notify_struct *)arg;
+
+	switch (cmd) {
+	case SLUB_ALLOC:
+		ret = slub_alloc_notify(ns->cachep, ns->objp, ns->gfp);
+		break;
+	case SLUB_FREE:
+		ret = 0;
+		slub_free_notify(ns->cachep, ns->objp);
+		break;
+	case SLUB_NEWPAGE:
+		ret = slub_newpage_notify(ns->cachep, ns->objp, ns->gfp);
+		break;
+	case SLUB_FREEPAGE:
+		ret = 0;
+		slub_freepage_notify(ns->cachep, ns->objp);
+		break;
+	case SLUB_ON:
+		ret = slub_on_notify(ns->cachep);
+		break;
+	case SLUB_OFF:
+		ret = slub_off_notify(ns->cachep);
+		break;
+	default:
+		return NOTIFY_DONE;
+	}
+
+	return (ret < 0) ? notifier_from_errno(ret) : NOTIFY_OK;
+}
+
+static struct notifier_block kmem_block = {
+	.notifier_call = kmem_notify,
+};
+
+static int kmem_subsys_register(void)
+{
+	return slub_register_notifier(&kmem_block);
+}
+
+__initcall(kmem_subsys_register);
+#endif

mm/slub.c +9 -1

index 31d04a3..e066a0e 100644
@@ -327,7 +327,7 @@ static inline void set_freepointer(struc
 	for (__p = (__free); __p; __p = get_freepointer((__s), __p))
 
 /* Determine object index from a given position */
-static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
+inline int slab_index(void *p, struct kmem_cache *s, void *addr)
 {
 	return (p - addr) / s->size;
 }

@@ -2360,6 +2504,14 @@ EXPORT_SYMBOL(kmem_cache_destroy);
 struct kmem_cache kmalloc_caches[PAGE_SHIFT] __cacheline_aligned;
 EXPORT_SYMBOL(kmalloc_caches);
 
+int is_kmalloc_cache(struct kmem_cache *s)
+{
+	int km_idx;
+
+	km_idx = s - kmalloc_caches;
+	return km_idx >= 0 && km_idx < ARRAY_SIZE(kmalloc_caches);
+}
+

Quoted 15 lines

mm/kmemcontrol.c new file +15 -0

+int slub_alloc_notify(struct kmem_cache *s, void *obj, gfp_t gfp)
+{
+	struct page *pg;
+	struct kmem_container *cnt;
+	struct kmem_container **obj_container;
+
+	pg = virt_to_head_page(obj);
+	obj_container = pg->cgroups;
+	if (unlikely(obj_container == NULL)) {
+		/*
+		 * turned on after some objects were allocated
+		 */
+		if (slub_newpage_notify(s, pg, GFP_ATOMIC) < 0)
+			goto err;
+

mm/kmemcontrol.c line 181

Quoted 6 lines

mm/kmemcontrol.c new file +6 -0

+		obj_container = pg->cgroups;
+	}
+
+	rcu_read_lock();
+	cnt = task_kmem_container(current);
+	if (res_counter_charge(&cnt->res, s->size))

Is s->size measure in pages or bytes? I suspect it is bytes.

Quoted 29 lines

mm/kmemcontrol.c new file +29 -0

+		goto err_locked;
+
+	css_get(&cnt->css);
+	rcu_read_unlock();
+	obj_container[slab_index(obj, s, page_address(pg))] = cnt;
+	return 0;
+
+err_locked:
+	rcu_read_unlock();
+err:
+	return -ENOMEM;
+}
+
+void slub_free_notify(struct kmem_cache *s, void *obj)
+{
+	struct page *pg;
+	struct kmem_container *cnt;
+	struct kmem_container **obj_container;
+
+	pg = virt_to_head_page(obj);
+	obj_container = pg->cgroups;
+	if (obj_container == NULL)
+		return;
+
+	obj_container += slab_index(obj, s, page_address(pg));
+	cnt = *obj_container;
+	if (cnt == NULL)
+		return;
+

mm/kmemcontrol.c line 215

Quoted 5 lines

mm/kmemcontrol.c new file +5 -0

+	res_counter_uncharge(&cnt->res, s->size);
+	*obj_container = NULL;
+	css_put(&cnt->css);
+}
+

Quick check, slub_free_notify() and slab_alloc_notify() are called from serialized contexts, right?

mm/kmemcontrol.c line 220

Quoted 5 lines

mm/kmemcontrol.c new file +5 -0

+int slub_on_notify(struct kmem_cache *cachep)
+{
+	return (is_kmalloc_cache(cachep) ? -EINVAL : 0);
+}
+

I know you've mentioned several times in the comments that kmalloc slab's cannot be accounted for, could you please add a big comment here as well.

Also, looks like if I turn on notification for kmalloc

slab's, the listener fails, which will cause all
allocations to fail?

Quoted 4 lines

Quoted 1 line

+

Quick check, slub_free_notify() and slab_alloc_notify() are called from serialized contexts, right?

Yup.

Quoted 4 lines · Pavel Emelyanov

Quoted 3 lines

Quoted 1 line

+

Quick check, slub_free_notify() and slab_alloc_notify() are called from serialized contexts, right?

Yup.

How is it serialized?

Quoted 3 lines · Christoph Lameter

Quoted 1 line · Pavel Emelyanov

Quoted 1 line

Quoted 1 line

+

Quoted 5 lines

Quoted 3 lines

Quoted 2 lines

Quick check, slub_free_notify() and slab_alloc_notify() are called from serialized contexts, right?

Yup.

How is it serialized?

They are booth called from __slab_alloc()/__slab_free() from under the slab_lock(page).

Quoted 6 lines · Pavel Emelyanov

Quoted 5 lines · Christoph Lameter

Quoted 3 lines · Pavel Emelyanov

Quoted 3 lines

Quoted 1 line

+

Quick check, slub_free_notify() and slab_alloc_notify() are called from serialized contexts, right?

Quoted 4 lines

Quoted 2 lines

Quoted 1 line

Yup.

How is it serialized?

They are booth called from __slab_alloc()/__slab_free() from under the slab_lock(page).

This means they are serialized per slab. Which means you can guarantee that multiple of these callbacks are not done at the same time for the same object. Is that what you need?

Changes since v.4:
* make SLAB_NOTIFY caches mark pages as SlabDebug. That makes the interesting paths simpler (thanks to Christoph);
* the change above caused appropriate changes in "turn notifications on" path - all available pages must become SlabDebug and page's freelists must be flushed;
* added two more events - "on" and "off" to make kmalloc caches disabling more gracefully;
* turning notifications "off" is marked as "TODO". Right now it's hard w/o massive rework of slub.c in respect to full slabs handling.

Changes since v.3:
* moved alloc/free notification into slow path and make "notify-able" caches walk this path always;
* introduced some optimization for the case, when there's only one listener for SLUB events (saves more that 10% of performance);
* ported on 2.6.23-rc6-mm1 tree.

Changes since v.2:
* introduced generic notifiers for slub. right now there are only events, needed by accounting, but this set can be extended in the future;
* moved the controller core into separate file, so that its extension and/or porting on slAb will look more logical;
* fixed this message :).

Changes since v.1:
* fixed Paul's comment about subsystem registration;
* return ERR_PTR from ->create callback, not NULL;
* make container-to-object assignment in rcu-safe section;
* make turning accounting on and off with "1" and "0".

Quoted 29 lines · Pavel Emelyanov

Changes since v.4:
* make SLAB_NOTIFY caches mark pages as SlabDebug. That makes the interesting paths simpler (thanks to Christoph);
* the change above caused appropriate changes in "turn notifications on" path - all available pages must become SlabDebug and page's freelists must be flushed;
* added two more events - "on" and "off" to make kmalloc caches disabling more gracefully;
* turning notifications "off" is marked as "TODO". Right now it's hard w/o massive rework of slub.c in respect to full slabs handling.

Changes since v.3:
* moved alloc/free notification into slow path and make "notify-able" caches walk this path always;
* introduced some optimization for the case, when there's only one listener for SLUB events (saves more that 10% of performance);
* ported on 2.6.23-rc6-mm1 tree.

Changes since v.2:
* introduced generic notifiers for slub. right now there are only events, needed by accounting, but this set can be extended in the future;
* moved the controller core into separate file, so that its extension and/or porting on slAb will look more logical;
* fixed this message :).

Quoted 6 lines

Changes since v.1:
* fixed Paul's comment about subsystem registration;
* return ERR_PTR from ->create callback, not NULL;
* make container-to-object assignment in rcu-safe section;
* make turning accounting on and off with "1" and "0".

Hi, Pavel,

Overall, the patches look good, except for the comments noted in the other patch postings. I am going to test these patches and see how they hold out.

Changes since v.4:
* make SLAB_NOTIFY caches mark pages as SlabDebug. That makes the interesting paths simpler (thanks to Christoph);
* the change above caused appropriate changes in "turn notifications on" path - all available pages must become SlabDebug and page's freelists must be flushed;
* added two more events - "on" and "off" to make kmalloc caches disabling more gracefully;
* turning notifications "off" is marked as "TODO". Right now it's hard w/o massive rework of slub.c in respect to full slabs handling.

Changes since v.3:
* moved alloc/free notification into slow path and make "notify-able" caches walk this path always;
* introduced some optimization for the case, when there's only one listener for SLUB events (saves more that 10% of performance);
* ported on 2.6.23-rc6-mm1 tree.

Changes since v.2:
* introduced generic notifiers for slub. right now there are only events, needed by accounting, but this set can be extended in the future;
* moved the controller core into separate file, so that its extension and/or porting on slAb will look more logical;
* fixed this message :).

Changes since v.1:
* fixed Paul's comment about subsystem registration;
* return ERR_PTR from ->create callback, not NULL;
* make container-to-object assignment in rcu-safe section;
* make turning accounting on and off with "1" and "0".

============================================================

Long time ago we decided to start memory control with the user memory container. Now this container in -mm tree and I think we can start with the kmem one.

First of all - why do we need this kind of control. The major "pros" is that kernel memory control protects the system from DoS attacks by processes that live in container. As our experience shows many exploits simply do not work in the container with limited kernel memory.

I can split the kernel memory container into 4 parts:

1. kmalloc-ed objects control
2. vmalloc-ed objects control
3. buddy allocated pages control
4. kmem_cache_alloc-ed objects control

the control of first tree types of objects has one peculiarity:
one need to explicitly point out which allocations he wants to account and this becomes not-configurable and is to be discussed.

On the other hands such objects as anon_vma-s, file-s, sighangds, vfsmounts, etc are created by user request always and should always be accounted. Fortunately they are allocated from their own caches and thus the whole kmem cache can be accountable.

This is exactly what this patchset does - it adds the ability to account for the total size of kmem-cache-allocated objects from specified kmem caches.

This is based on the SLUB allocator, Paul's control groups and the resource counters I made for RSS controller and which are in -mm tree already.

To play with it, one need to mount the container file system with -o kmem and then mark some caches as accountable via /sys/slab/<cache_name>/cache_notify.

As I have already told kmalloc caches cannot be accounted easily so turning the accounting on for them will fail with -EINVAL.

Turning the accounting off is possible only if the cache has no objects. This is done so because turning accounting off implies marking of all the slabs in the cache as not-debug, but due to full-pages in slub are not stored in any lists (usually) this is impossible to do so, however this is in todo list.

Thanks,
Pavel

_______________________________________________
Devel mailing list
[email protected]
https://openvz.org/mailman/listinfo/devel

Hi Pavel,

One question about the general design of this - have you tested an approach where rather than tagging each object within the cache with the cgroup that allocated it, you instead have (inside the cache code) a separate cache structure for each cgroup? So the space overheads would go from having a per-object overhead (one pointer per object?) to having a "wastage" overhead (on average half a slab per cgroup). And the time overhead would be the time required to lookup the relevant cache for a cgroup at the start of the allocation operation, and the relevant cache for an object (from its struct page) at deallocation, rather then the time required to update the per-object housekeeping pointer.

Each cache would need to be assigned a unique ID, used as an index into a per-cgroup lookup table of localized caches. (This could almost be regarded as a form of kmem_cache namespace).

It seems to me that this alternative approach would be a lower memory overhead for people who have the kernel memory controller compiled in but aren't using it, or are only using a few groups.

Paul

Such a lookup would require a hastable or something similar. We already have such a bad experience (with OpenVZ RSS fractions accounting for example). Hash lookups imply the CPU caches screwup and hurt the performance. See also the comment below.

I thought the same some time ago and tried to make a per-beancounter kmem caches. The result was awful - the memory waste was much larger than in the case of pointer-per-object approach.

Quoted 4 lines · Pavel Emelyanov

Such a lookup would require a hastable or something similar. We already have such a bad experience (with OpenVZ RSS fractions accounting for example). Hash lookups imply the CPU caches screwup and hurt the performance. See also the comment below.

I think you could do it with an array lookup if you assigned an index to each cache as it was created, and used that as an offset into a per-cgroup array.

Quoted 3 lines

I thought the same some time ago and tried to make a per-beancounter kmem caches. The result was awful - the memory waste was much larger than in the case of pointer-per-object approach.

OK, fair enough.

Was this with a large number of bean counters? I imagine that with a small number, the waste might be rather more reasonable.

Paul