[PATCH v1 0/3] bpf: hash: use per-bucket spinlock

[PATCH v1 0/3] bpf: hash: use per-bucket spinlock

[Ming Lei]

Hi,

This patchset tries to optimize ebpf hash map, and follows
the idea:

	Both htab_map_update_elem() and htab_map_delete_elem()
	can be called from eBPF program, and they may be in kernel
	hot path, it isn't efficient to use a per-hashtable lock
	in this two helpers, so this patch converts the lock into
	per-bucket spinlock.

With this patchset, looks the performance penalty from eBPF
decreased a lot, see the following test:

	1) run 'tools/biolatency' of bcc before running block test;

	2) run fio to test block throught over /dev/nullb0,
	(randread, 16jobs, libaio, 4k bs) and the test box
	is one 24cores(dual sockets) VM server:
	- without patchset:  607K IOPS
	- with this patchset: 1184K IOPS
	- without running eBPF prog: 1492K IOPS

TODO:
	- remove the per-hashtable atomic counter

V1:
	- fix the wrong 3/3 patch 


 kernel/bpf/hashtab.c | 60 +++++++++++++++++++++++++++++++---------------------
 1 file changed, 36 insertions(+), 24 deletions(-)

Thanks,

[PATCH v1 1/3] bpf: hash: use atomic count

[Ming Lei]

Preparing for removing global per-hashtable lock, so
the counter need to be defined as aotmic_t first.

Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Ming Lei <tom.leiming@gmail.com>
---
 kernel/bpf/hashtab.c | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)

diff --git a/kernel/bpf/hashtab.c b/kernel/bpf/hashtab.c
index 34777b3..2615388 100644
--- a/kernel/bpf/hashtab.c
+++ b/kernel/bpf/hashtab.c
@@ -18,7 +18,7 @@ struct bpf_htab {
 	struct bpf_map map;
 	struct hlist_head *buckets;
 	raw_spinlock_t lock;
-	u32 count;	/* number of elements in this hashtable */
+	atomic_t count;	/* number of elements in this hashtable */
 	u32 n_buckets;	/* number of hash buckets */
 	u32 elem_size;	/* size of each element in bytes */
 };
@@ -106,7 +106,7 @@ static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
 		INIT_HLIST_HEAD(&htab->buckets[i]);
 
 	raw_spin_lock_init(&htab->lock);
-	htab->count = 0;
+	atomic_set(&htab->count, 0);
 
 	return &htab->map;
 
@@ -256,7 +256,7 @@ static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
 
 	l_old = lookup_elem_raw(head, l_new->hash, key, key_size);
 
-	if (!l_old && unlikely(htab->count >= map->max_entries)) {
+	if (!l_old && unlikely(atomic_read(&htab->count) >= map->max_entries)) {
 		/* if elem with this 'key' doesn't exist and we've reached
 		 * max_entries limit, fail insertion of new elem
 		 */
@@ -284,7 +284,7 @@ static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
 		hlist_del_rcu(&l_old->hash_node);
 		kfree_rcu(l_old, rcu);
 	} else {
-		htab->count++;
+		atomic_inc(&htab->count);
 	}
 	raw_spin_unlock_irqrestore(&htab->lock, flags);
 
@@ -319,7 +319,7 @@ static int htab_map_delete_elem(struct bpf_map *map, void *key)
 
 	if (l) {
 		hlist_del_rcu(&l->hash_node);
-		htab->count--;
+		atomic_dec(&htab->count);
 		kfree_rcu(l, rcu);
 		ret = 0;
 	}
@@ -339,7 +339,7 @@ static void delete_all_elements(struct bpf_htab *htab)
 
 		hlist_for_each_entry_safe(l, n, head, hash_node) {
 			hlist_del_rcu(&l->hash_node);
-			htab->count--;
+			atomic_dec(&htab->count);
 			kfree(l);
 		}
 	}
-- 
1.9.1

[PATCH v1 2/3] bpf: hash: move select_bucket() out of htab's spinlock

[Ming Lei]

The spinlock is just used for protecting the per-bucket
hlist, so it isn't needed for selecting bucket.

Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Ming Lei <tom.leiming@gmail.com>
---
 kernel/bpf/hashtab.c | 6 ++----
 1 file changed, 2 insertions(+), 4 deletions(-)

diff --git a/kernel/bpf/hashtab.c b/kernel/bpf/hashtab.c
index 2615388..d857fcb 100644
--- a/kernel/bpf/hashtab.c
+++ b/kernel/bpf/hashtab.c
@@ -248,12 +248,11 @@ static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
 	memcpy(l_new->key + round_up(key_size, 8), value, map->value_size);
 
 	l_new->hash = htab_map_hash(l_new->key, key_size);
+	head = select_bucket(htab, l_new->hash);
 
 	/* bpf_map_update_elem() can be called in_irq() */
 	raw_spin_lock_irqsave(&htab->lock, flags);
 
-	head = select_bucket(htab, l_new->hash);
-
 	l_old = lookup_elem_raw(head, l_new->hash, key, key_size);
 
 	if (!l_old && unlikely(atomic_read(&htab->count) >= map->max_entries)) {
@@ -310,11 +309,10 @@ static int htab_map_delete_elem(struct bpf_map *map, void *key)
 	key_size = map->key_size;
 
 	hash = htab_map_hash(key, key_size);
+	head = select_bucket(htab, hash);
 
 	raw_spin_lock_irqsave(&htab->lock, flags);
 
-	head = select_bucket(htab, hash);
-
 	l = lookup_elem_raw(head, hash, key, key_size);
 
 	if (l) {
-- 
1.9.1

[PATCH v1 3/3] bpf: hash: use per-bucket spinlock

[Ming Lei]

Both htab_map_update_elem() and htab_map_delete_elem() can be
called from eBPF program, and they may be in kernel hot path,
so it isn't efficient to use a per-hashtable lock in this two
helpers.

The per-hashtable spinlock is used for protecting bucket's
hlist, and per-bucket lock is just enough. This patch converts
the per-hashtable lock into per-bucket spinlock, so that
contention can be decreased a lot.

Signed-off-by: Ming Lei <tom.leiming@gmail.com>
---
 kernel/bpf/hashtab.c | 46 ++++++++++++++++++++++++++++++----------------
 1 file changed, 30 insertions(+), 16 deletions(-)

diff --git a/kernel/bpf/hashtab.c b/kernel/bpf/hashtab.c
index d857fcb..67222a9 100644
--- a/kernel/bpf/hashtab.c
+++ b/kernel/bpf/hashtab.c
@@ -14,10 +14,14 @@
 #include <linux/filter.h>
 #include <linux/vmalloc.h>
 
+struct bucket {
+	struct hlist_head head;
+	raw_spinlock_t lock;
+};
+
 struct bpf_htab {
 	struct bpf_map map;
-	struct hlist_head *buckets;
-	raw_spinlock_t lock;
+	struct bucket *buckets;
 	atomic_t count;	/* number of elements in this hashtable */
 	u32 n_buckets;	/* number of hash buckets */
 	u32 elem_size;	/* size of each element in bytes */
@@ -88,24 +92,25 @@ static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
 		/* make sure page count doesn't overflow */
 		goto free_htab;
 
-	htab->map.pages = round_up(htab->n_buckets * sizeof(struct hlist_head) +
+	htab->map.pages = round_up(htab->n_buckets * sizeof(struct bucket) +
 				   htab->elem_size * htab->map.max_entries,
 				   PAGE_SIZE) >> PAGE_SHIFT;
 
 	err = -ENOMEM;
-	htab->buckets = kmalloc_array(htab->n_buckets, sizeof(struct hlist_head),
+	htab->buckets = kmalloc_array(htab->n_buckets, sizeof(struct bucket),
 				      GFP_USER | __GFP_NOWARN);
 
 	if (!htab->buckets) {
-		htab->buckets = vmalloc(htab->n_buckets * sizeof(struct hlist_head));
+		htab->buckets = vmalloc(htab->n_buckets * sizeof(struct bucket));
 		if (!htab->buckets)
 			goto free_htab;
 	}
 
-	for (i = 0; i < htab->n_buckets; i++)
-		INIT_HLIST_HEAD(&htab->buckets[i]);
+	for (i = 0; i < htab->n_buckets; i++) {
+		INIT_HLIST_HEAD(&htab->buckets[i].head);
+		raw_spin_lock_init(&htab->buckets[i].lock);
+	}
 
-	raw_spin_lock_init(&htab->lock);
 	atomic_set(&htab->count, 0);
 
 	return &htab->map;
@@ -120,11 +125,16 @@ static inline u32 htab_map_hash(const void *key, u32 key_len)
 	return jhash(key, key_len, 0);
 }
 
-static inline struct hlist_head *select_bucket(struct bpf_htab *htab, u32 hash)
+static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
 {
 	return &htab->buckets[hash & (htab->n_buckets - 1)];
 }
 
+static inline struct hlist_head *select_bucket(struct bpf_htab *htab, u32 hash)
+{
+	return &__select_bucket(htab, hash)->head;
+}
+
 static struct htab_elem *lookup_elem_raw(struct hlist_head *head, u32 hash,
 					 void *key, u32 key_size)
 {
@@ -227,6 +237,7 @@ static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 	struct htab_elem *l_new, *l_old;
 	struct hlist_head *head;
+	struct bucket *b;
 	unsigned long flags;
 	u32 key_size;
 	int ret;
@@ -248,10 +259,11 @@ static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
 	memcpy(l_new->key + round_up(key_size, 8), value, map->value_size);
 
 	l_new->hash = htab_map_hash(l_new->key, key_size);
-	head = select_bucket(htab, l_new->hash);
+	b = __select_bucket(htab, l_new->hash);
+	head = &b->head;
 
 	/* bpf_map_update_elem() can be called in_irq() */
-	raw_spin_lock_irqsave(&htab->lock, flags);
+	raw_spin_lock_irqsave(&b->lock, flags);
 
 	l_old = lookup_elem_raw(head, l_new->hash, key, key_size);
 
@@ -285,11 +297,11 @@ static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
 	} else {
 		atomic_inc(&htab->count);
 	}
-	raw_spin_unlock_irqrestore(&htab->lock, flags);
+	raw_spin_unlock_irqrestore(&b->lock, flags);
 
 	return 0;
 err:
-	raw_spin_unlock_irqrestore(&htab->lock, flags);
+	raw_spin_unlock_irqrestore(&b->lock, flags);
 	kfree(l_new);
 	return ret;
 }
@@ -299,6 +311,7 @@ static int htab_map_delete_elem(struct bpf_map *map, void *key)
 {
 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 	struct hlist_head *head;
+	struct bucket *b;
 	struct htab_elem *l;
 	unsigned long flags;
 	u32 hash, key_size;
@@ -309,9 +322,10 @@ static int htab_map_delete_elem(struct bpf_map *map, void *key)
 	key_size = map->key_size;
 
 	hash = htab_map_hash(key, key_size);
-	head = select_bucket(htab, hash);
+	b = __select_bucket(htab, hash);
+	head = &b->head;
 
-	raw_spin_lock_irqsave(&htab->lock, flags);
+	raw_spin_lock_irqsave(&b->lock, flags);
 
 	l = lookup_elem_raw(head, hash, key, key_size);
 
@@ -322,7 +336,7 @@ static int htab_map_delete_elem(struct bpf_map *map, void *key)
 		ret = 0;
 	}
 
-	raw_spin_unlock_irqrestore(&htab->lock, flags);
+	raw_spin_unlock_irqrestore(&b->lock, flags);
 	return ret;
 }
 
-- 
1.9.1

Re: [PATCH v1 3/3] bpf: hash: use per-bucket spinlock

[Daniel Borkmann]

On 12/28/2015 01:55 PM, Ming Lei wrote:
> Both htab_map_update_elem() and htab_map_delete_elem() can be
> called from eBPF program, and they may be in kernel hot path,
> so it isn't efficient to use a per-hashtable lock in this two
> helpers.
>
> The per-hashtable spinlock is used for protecting bucket's
> hlist, and per-bucket lock is just enough. This patch converts
> the per-hashtable lock into per-bucket spinlock, so that
> contention can be decreased a lot.
>
> Signed-off-by: Ming Lei <tom.leiming@gmail.com>
> ---
>   kernel/bpf/hashtab.c | 46 ++++++++++++++++++++++++++++++----------------
>   1 file changed, 30 insertions(+), 16 deletions(-)
>
> diff --git a/kernel/bpf/hashtab.c b/kernel/bpf/hashtab.c
> index d857fcb..67222a9 100644
> --- a/kernel/bpf/hashtab.c
> +++ b/kernel/bpf/hashtab.c
> @@ -14,10 +14,14 @@
>   #include <linux/filter.h>
>   #include <linux/vmalloc.h>
>
> +struct bucket {
> +	struct hlist_head head;
> +	raw_spinlock_t lock;
> +};
> +
>   struct bpf_htab {
>   	struct bpf_map map;
> -	struct hlist_head *buckets;
> -	raw_spinlock_t lock;
> +	struct bucket *buckets;
>   	atomic_t count;	/* number of elements in this hashtable */
>   	u32 n_buckets;	/* number of hash buckets */
>   	u32 elem_size;	/* size of each element in bytes */
> @@ -88,24 +92,25 @@ static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
>   		/* make sure page count doesn't overflow */
>   		goto free_htab;

When adapting memory accounting and allocation sizes below where you replace
sizeof(struct hlist_head) with sizeof(struct bucket), is there a reason why
you don't update the overflow checks along with it?

[...]
         /* prevent zero size kmalloc and check for u32 overflow */
         if (htab->n_buckets == 0 ||
             htab->n_buckets > U32_MAX / sizeof(struct hlist_head))
                 goto free_htab;

         if ((u64) htab->n_buckets * sizeof(struct hlist_head) +
             (u64) htab->elem_size * htab->map.max_entries >=
             U32_MAX - PAGE_SIZE)
                 /* make sure page count doesn't overflow */
                 goto free_htab;
[...]

> -	htab->map.pages = round_up(htab->n_buckets * sizeof(struct hlist_head) +
> +	htab->map.pages = round_up(htab->n_buckets * sizeof(struct bucket) +
>   				   htab->elem_size * htab->map.max_entries,
>   				   PAGE_SIZE) >> PAGE_SHIFT;
>
>   	err = -ENOMEM;
> -	htab->buckets = kmalloc_array(htab->n_buckets, sizeof(struct hlist_head),
> +	htab->buckets = kmalloc_array(htab->n_buckets, sizeof(struct bucket),
>   				      GFP_USER | __GFP_NOWARN);
>
>   	if (!htab->buckets) {
> -		htab->buckets = vmalloc(htab->n_buckets * sizeof(struct hlist_head));
> +		htab->buckets = vmalloc(htab->n_buckets * sizeof(struct bucket));
>   		if (!htab->buckets)
>   			goto free_htab;
>   	}

[...]

Thanks,
Daniel