Skip to contentSkip to conversation

Allowing more than 64k connections and heavily optimize bind(0) time.

PATCH Evgeniy Polyakov 3 participants 10 messages Dec 24, 2008 – Jan 12, 2009

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

Hi.

With simple extension to the binding mechanism, which allows to bind more than 64k sockets (or smaller amount, depending on sysctl parameters), we have to traverse the whole bind hash table to find out empty bucket. And while it is not a problem for example for 32k connections, bind() completion time grows exponentially (since after each successful binding we have to traverse one bucket more to find empty one) even if we start each time from random offset inside the hash table.

So, when hash table is full, and we want to add another socket, we have to traverse the whole table no matter what, so effectivelly this will be the worst case performance and it will be constant.

Attached picture shows bind() time depending on number of already bound sockets.

Green area corresponds to the usual binding to zero port process, which turns on kernel port selection as described above. Red area is the bind process, when number of reuse-bound sockets is not limited by 64k (or sysctl parameters). The same exponential growth (hidden by the green area) before number of ports reaches sysctl limit.

At this time bind hash table has exactly one reuse-enbaled socket in a bucket, but it is possible that they have different addresses. Actually kernel selects the first port to try randomly, so at the beginning bind will take roughly constant time, but with time number of port to check after random start will increase. And that will have exponential growth, but because of above random selection, not every next port selection will necessary take longer time than previous. So we have to consider the area below in the graph (if you could zoom it, you could find, that there are many different times placed there), so area can hide another.

Blue area corresponds to the port selection optimization.

This is rather simple design approach: hashtable now maintains (unprecise and racely updated) number of currently bound sockets, and when number of such sockets becomes greater than predefined value (I use maximum port range defined by sysctls), we stop traversing the whole bind hash table and just stop at first matching bucket after random start. Above limit roughly corresponds to the case, when bind hash table is full and we turned on mechanism of allowing to bind more reuse-enabled sockets, so it does not change behaviour of other sockets.

As I just have been told, it results in this strange behaviour:

$ grep -n :60013 netstat.res
33101:tcp        0      0 local1:60013       remote:80       ESTABLISHED
33105:tcp        0      0 local1:60013       remote:80       ESTABLISHED
52084:tcp        0      0 local2:60013       remote:80       ESTABLISHED
52085:tcp        0      0 local2:60013       remote:80       ESTABLISHED
58249:tcp        0      0 local3:60013       remote:80       ESTABLISHED

it is yet to resolve what it is and how much harm it brings :)

include/net/inet_hashtables.h +2 -1
index 5cc182f..6389aea 100644
@@ -80,6 +80,7 @@ struct inet_bind_bucket {
 	struct net		*ib_net;
 	unsigned short		port;
 	signed short		fastreuse;
+	int			num_owners;
 	struct hlist_node	node;
 	struct hlist_head	owners;
 };
@@ -114,7 +115,7 @@ struct inet_hashinfo {
 	struct inet_bind_hashbucket	*bhash;
 
 	unsigned int			bhash_size;
-	/* Note : 4 bytes padding on 64 bit arches */
+	int				bsockets;
 
 	/* All sockets in TCP_LISTEN state will be in here.  This is the only
 	 * table where wildcard'd TCP sockets can exist.  Hash function here
net/ipv4/inet_connection_sock.c +31 -6
index bd1278a..333a30e 100644
@@ -93,24 +93,40 @@ int inet_csk_get_port(struct sock *sk, unsigned short snum)
 	struct inet_bind_hashbucket *head;
 	struct hlist_node *node;
 	struct inet_bind_bucket *tb;
-	int ret;
+	int ret, attempts = 5;
 	struct net *net = sock_net(sk);
 
 	local_bh_disable();
 	if (!snum) {
 		int remaining, rover, low, high;
+ 		int smallest_size, smallest_rover;
 
+again:
 		inet_get_local_port_range(&low, &high);
 		remaining = (high - low) + 1;
-		rover = net_random() % remaining + low;
+ 		smallest_rover = rover = net_random() % remaining + low;
+ 		smallest_size = ~0;
 
 		do {
 			head = &hashinfo->bhash[inet_bhashfn(net, rover,
 					hashinfo->bhash_size)];
 			spin_lock(&head->lock);
 			inet_bind_bucket_for_each(tb, node, &head->chain)
-				if (tb->ib_net == net && tb->port == rover)
+				if (tb->ib_net == net && tb->port == rover) {
+ 					if (tb->fastreuse > 0 &&
+ 					    sk->sk_reuse &&
+ 					    sk->sk_state != TCP_LISTEN &&
+ 					    (tb->num_owners < smallest_size || smallest_size == ~0)) {
+ 						smallest_size = tb->num_owners;
+ 						smallest_rover = rover;
+						if (hashinfo->bsockets > (high - low) + 1) {
+							spin_unlock(&head->lock);
+							snum = smallest_rover;
+							goto have_snum;
+						}
+ 					}
 					goto next;
+				}
 			break;
 		next:
 			spin_unlock(&head->lock);
@@ -125,14 +141,19 @@ int inet_csk_get_port(struct sock *sk, unsigned short snum)
 		 * the top level, not from the 'break;' statement.
 		 */
 		ret = 1;
-		if (remaining <= 0)
+		if (remaining <= 0) {
+			if (smallest_size != ~0) {
+				snum = smallest_rover;
+				goto have_snum;
+			}
 			goto fail;
-
+		}
 		/* OK, here is the one we will use.  HEAD is
 		 * non-NULL and we hold it's mutex.
 		 */
 		snum = rover;
 	} else {
+have_snum:
 		head = &hashinfo->bhash[inet_bhashfn(net, snum,
 				hashinfo->bhash_size)];
 		spin_lock(&head->lock);
@@ -149,8 +170,12 @@ tb_found:
 			goto success;
 		} else {
 			ret = 1;
-			if (inet_csk(sk)->icsk_af_ops->bind_conflict(sk, tb))
+			if (inet_csk(sk)->icsk_af_ops->bind_conflict(sk, tb)) {
+				printk("Got a bind conflict for port %d, sk_reuse: %d.\n", snum, sk->sk_reuse);
+				if (sk->sk_reuse && sk->sk_state != TCP_LISTEN && --attempts >= 0)
+					goto again;
 				goto fail_unlock;
+			}
 		}
 	}
 tb_not_found:
net/ipv4/inet_hashtables.c +10 -1
index 4498190..399ec17 100644
@@ -38,6 +38,7 @@ struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep,
 		tb->ib_net       = hold_net(net);
 		tb->port      = snum;
 		tb->fastreuse = 0;
+		tb->num_owners = 0;
 		INIT_HLIST_HEAD(&tb->owners);
 		hlist_add_head(&tb->node, &head->chain);
 	}
@@ -59,8 +60,13 @@ void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket
 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
 		    const unsigned short snum)
 {
+	struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
+
+	hashinfo->bsockets++;
+
 	inet_sk(sk)->num = snum;
 	sk_add_bind_node(sk, &tb->owners);
+	tb->num_owners++;
 	inet_csk(sk)->icsk_bind_hash = tb;
 }
 
@@ -74,10 +80,13 @@ static void __inet_put_port(struct sock *sk)
 			hashinfo->bhash_size);
 	struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash];
 	struct inet_bind_bucket *tb;
+	
+	hashinfo->bsockets--;
 
 	spin_lock(&head->lock);
 	tb = inet_csk(sk)->icsk_bind_hash;
 	__sk_del_bind_node(sk);
+	tb->num_owners--;
 	inet_csk(sk)->icsk_bind_hash = NULL;
 	inet_sk(sk)->num = 0;
 	inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
@@ -450,9 +459,9 @@ int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 			 */
 			inet_bind_bucket_for_each(tb, node, &head->chain) {
 				if (tb->ib_net == net && tb->port == port) {
-					WARN_ON(hlist_empty(&tb->owners));
 					if (tb->fastreuse >= 0)
 						goto next_port;
+					WARN_ON(hlist_empty(&tb->owners));
 					if (!check_established(death_row, sk,
 								port, &tw))
 						goto ok;
net/ipv4/tcp_ipv4.c +2 -0
index 5c8fa7f..3e70134 100644
@@ -101,6 +101,8 @@ struct inet_hashinfo __cacheline_aligned tcp_hashinfo = {
 	.lhash_lock  = __RW_LOCK_UNLOCKED(tcp_hashinfo.lhash_lock),
 	.lhash_users = ATOMIC_INIT(0),
 	.lhash_wait  = __WAIT_QUEUE_HEAD_INITIALIZER(tcp_hashinfo.lhash_wait),
+
+	.bsockets    = 0,
 };
 
 static inline __u32 tcp_v4_init_sequence(struct sk_buff *skb)
Signature · 1 line
	Evgeniy Polyakov
#1 lore raw
Quoted 7 lines · Evgeniy Polyakov

As I just have been told, it results in this strange behaviour:

$ grep -n :60013 netstat.res
33101:tcp        0      0 local1:60013       remote:80       ESTABLISHED
33105:tcp        0      0 local1:60013       remote:80       ESTABLISHED
52084:tcp        0      0 local2:60013       remote:80       ESTABLISHED
52085:tcp        0      0 local2:60013       remote:80       ESTABLISHED
58249:tcp        0      0 local3:60013       remote:80       ESTABLISHED

I suppose this issue with the multiple sockets bound to the same local address and port is now fixed. Likely problem is in the way my patch broke the old code, which assumed that table may contain only single reuse socket added via bind() with 0 port.

In the updated version bind() checks that selected bucket contains fast reuse sockets already, and in that case runs the whole bind conflict check, which involves address and port, otherwise socket is just added into the table.

Patch was not yet heavily tested though, but it passed several trivial lots-of-binds test application runs.
I will update patch if production testing reveals some problems.

include/net/inet_hashtables.h +2 -1
index 5cc182f..6389aea 100644
@@ -80,6 +80,7 @@ struct inet_bind_bucket {
 	struct net		*ib_net;
 	unsigned short		port;
 	signed short		fastreuse;
+	int			num_owners;
 	struct hlist_node	node;
 	struct hlist_head	owners;
 };
@@ -114,7 +115,7 @@ struct inet_hashinfo {
 	struct inet_bind_hashbucket	*bhash;
 
 	unsigned int			bhash_size;
-	/* Note : 4 bytes padding on 64 bit arches */
+	int				bsockets;
 
 	/* All sockets in TCP_LISTEN state will be in here.  This is the only
 	 * table where wildcard'd TCP sockets can exist.  Hash function here
net/ipv4/inet_connection_sock.c +32 -7
index bd1278a..0813b3d 100644
@@ -93,24 +93,40 @@ int inet_csk_get_port(struct sock *sk, unsigned short snum)
 	struct inet_bind_hashbucket *head;
 	struct hlist_node *node;
 	struct inet_bind_bucket *tb;
-	int ret;
+	int ret, attempts = 5;
 	struct net *net = sock_net(sk);
+ 	int smallest_size = -1, smallest_rover;
 
 	local_bh_disable();
 	if (!snum) {
 		int remaining, rover, low, high;
 
+again:
 		inet_get_local_port_range(&low, &high);
 		remaining = (high - low) + 1;
-		rover = net_random() % remaining + low;
+ 		smallest_rover = rover = net_random() % remaining + low;
 
+		smallest_size = -1;
 		do {
 			head = &hashinfo->bhash[inet_bhashfn(net, rover,
 					hashinfo->bhash_size)];
 			spin_lock(&head->lock);
 			inet_bind_bucket_for_each(tb, node, &head->chain)
-				if (tb->ib_net == net && tb->port == rover)
+				if (tb->ib_net == net && tb->port == rover) {
+ 					if (tb->fastreuse > 0 &&
+ 					    sk->sk_reuse &&
+ 					    sk->sk_state != TCP_LISTEN &&
+ 					    (tb->num_owners < smallest_size || smallest_size == -1)) {
+ 						smallest_size = tb->num_owners;
+ 						smallest_rover = rover;
+						if (hashinfo->bsockets > (high - low) + 1) {
+							spin_unlock(&head->lock);
+							snum = smallest_rover;
+							goto have_snum;
+						}
+ 					}
 					goto next;
+				}
 			break;
 		next:
 			spin_unlock(&head->lock);
@@ -125,14 +141,19 @@ int inet_csk_get_port(struct sock *sk, unsigned short snum)
 		 * the top level, not from the 'break;' statement.
 		 */
 		ret = 1;
-		if (remaining <= 0)
+		if (remaining <= 0) {
+			if (smallest_size != -1) {
+				snum = smallest_rover;
+				goto have_snum;
+			}
 			goto fail;
-
+		}
 		/* OK, here is the one we will use.  HEAD is
 		 * non-NULL and we hold it's mutex.
 		 */
 		snum = rover;
 	} else {
+have_snum:
 		head = &hashinfo->bhash[inet_bhashfn(net, snum,
 				hashinfo->bhash_size)];
 		spin_lock(&head->lock);
@@ -145,12 +166,16 @@ int inet_csk_get_port(struct sock *sk, unsigned short snum)
 tb_found:
 	if (!hlist_empty(&tb->owners)) {
 		if (tb->fastreuse > 0 &&
-		    sk->sk_reuse && sk->sk_state != TCP_LISTEN) {
+		    sk->sk_reuse && sk->sk_state != TCP_LISTEN &&
+		    smallest_size == -1) {
 			goto success;
 		} else {
 			ret = 1;
-			if (inet_csk(sk)->icsk_af_ops->bind_conflict(sk, tb))
+			if (inet_csk(sk)->icsk_af_ops->bind_conflict(sk, tb)) {
+				if (sk->sk_reuse && sk->sk_state != TCP_LISTEN && --attempts >= 0)
+					goto again;
 				goto fail_unlock;
+			}
 		}
 	}
 tb_not_found:
net/ipv4/inet_hashtables.c +10 -1
index 4498190..399ec17 100644
@@ -38,6 +38,7 @@ struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep,
 		tb->ib_net       = hold_net(net);
 		tb->port      = snum;
 		tb->fastreuse = 0;
+		tb->num_owners = 0;
 		INIT_HLIST_HEAD(&tb->owners);
 		hlist_add_head(&tb->node, &head->chain);
 	}
@@ -59,8 +60,13 @@ void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket
 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
 		    const unsigned short snum)
 {
+	struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
+
+	hashinfo->bsockets++;
+
 	inet_sk(sk)->num = snum;
 	sk_add_bind_node(sk, &tb->owners);
+	tb->num_owners++;
 	inet_csk(sk)->icsk_bind_hash = tb;
 }
 
@@ -74,10 +80,13 @@ static void __inet_put_port(struct sock *sk)
 			hashinfo->bhash_size);
 	struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash];
 	struct inet_bind_bucket *tb;
+	
+	hashinfo->bsockets--;
 
 	spin_lock(&head->lock);
 	tb = inet_csk(sk)->icsk_bind_hash;
 	__sk_del_bind_node(sk);
+	tb->num_owners--;
 	inet_csk(sk)->icsk_bind_hash = NULL;
 	inet_sk(sk)->num = 0;
 	inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
@@ -450,9 +459,9 @@ int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 			 */
 			inet_bind_bucket_for_each(tb, node, &head->chain) {
 				if (tb->ib_net == net && tb->port == port) {
-					WARN_ON(hlist_empty(&tb->owners));
 					if (tb->fastreuse >= 0)
 						goto next_port;
+					WARN_ON(hlist_empty(&tb->owners));
 					if (!check_established(death_row, sk,
 								port, &tw))
 						goto ok;
net/ipv4/tcp_ipv4.c +2 -0
index 5c8fa7f..3e70134 100644
@@ -101,6 +101,8 @@ struct inet_hashinfo __cacheline_aligned tcp_hashinfo = {
 	.lhash_lock  = __RW_LOCK_UNLOCKED(tcp_hashinfo.lhash_lock),
 	.lhash_users = ATOMIC_INIT(0),
 	.lhash_wait  = __WAIT_QUEUE_HEAD_INITIALIZER(tcp_hashinfo.lhash_wait),
+
+	.bsockets    = 0,
 };
 
 static inline __u32 tcp_v4_init_sequence(struct sk_buff *skb)
Signature · 1 line
	Evgeniy Polyakov
#2 lore raw

Ping :)

I'm getting out from the vacations tomorrow and will try to more extensively test the patch (QA group found a bug in ipv6 side, but that was with the .24 backported patch, current one does not have that codepath iirc). But I would like to know if this worth it.

Quoted 199 lines · Evgeniy Polyakov
Quoted 7 lines · Evgeniy Polyakov

As I just have been told, it results in this strange behaviour:

$ grep -n :60013 netstat.res
33101:tcp        0      0 local1:60013       remote:80       ESTABLISHED
33105:tcp        0      0 local1:60013       remote:80       ESTABLISHED
52084:tcp        0      0 local2:60013       remote:80       ESTABLISHED
52085:tcp        0      0 local2:60013       remote:80       ESTABLISHED
58249:tcp        0      0 local3:60013       remote:80       ESTABLISHED

I suppose this issue with the multiple sockets bound to the same local address and port is now fixed. Likely problem is in the way my patch broke the old code, which assumed that table may contain only single reuse socket added via bind() with 0 port.

In the updated version bind() checks that selected bucket contains fast reuse sockets already, and in that case runs the whole bind conflict check, which involves address and port, otherwise socket is just added into the table.

Patch was not yet heavily tested though, but it passed several trivial lots-of-binds test application runs.
I will update patch if production testing reveals some problems.

include/net/inet_hashtables.h +2 -1
index 5cc182f..6389aea 100644
@@ -80,6 +80,7 @@ struct inet_bind_bucket {
 	struct net		*ib_net;
 	unsigned short		port;
 	signed short		fastreuse;
+	int			num_owners;
 	struct hlist_node	node;
 	struct hlist_head	owners;
 };
@@ -114,7 +115,7 @@ struct inet_hashinfo {
 	struct inet_bind_hashbucket	*bhash;
 
 	unsigned int			bhash_size;
-	/* Note : 4 bytes padding on 64 bit arches */
+	int				bsockets;
 
 	/* All sockets in TCP_LISTEN state will be in here.  This is the only
 	 * table where wildcard'd TCP sockets can exist.  Hash function here
net/ipv4/inet_connection_sock.c +32 -7
index bd1278a..0813b3d 100644
@@ -93,24 +93,40 @@ int inet_csk_get_port(struct sock *sk, unsigned short snum)
 	struct inet_bind_hashbucket *head;
 	struct hlist_node *node;
 	struct inet_bind_bucket *tb;
-	int ret;
+	int ret, attempts = 5;
 	struct net *net = sock_net(sk);
+ 	int smallest_size = -1, smallest_rover;
 
 	local_bh_disable();
 	if (!snum) {
 		int remaining, rover, low, high;
 
+again:
 		inet_get_local_port_range(&low, &high);
 		remaining = (high - low) + 1;
-		rover = net_random() % remaining + low;
+ 		smallest_rover = rover = net_random() % remaining + low;
 
+		smallest_size = -1;
 		do {
 			head = &hashinfo->bhash[inet_bhashfn(net, rover,
 					hashinfo->bhash_size)];
 			spin_lock(&head->lock);
 			inet_bind_bucket_for_each(tb, node, &head->chain)
-				if (tb->ib_net == net && tb->port == rover)
+				if (tb->ib_net == net && tb->port == rover) {
+ 					if (tb->fastreuse > 0 &&
+ 					    sk->sk_reuse &&
+ 					    sk->sk_state != TCP_LISTEN &&
+ 					    (tb->num_owners < smallest_size || smallest_size == -1)) {
+ 						smallest_size = tb->num_owners;
+ 						smallest_rover = rover;
+						if (hashinfo->bsockets > (high - low) + 1) {
+							spin_unlock(&head->lock);
+							snum = smallest_rover;
+							goto have_snum;
+						}
+ 					}
 					goto next;
+				}
 			break;
 		next:
 			spin_unlock(&head->lock);
@@ -125,14 +141,19 @@ int inet_csk_get_port(struct sock *sk, unsigned short snum)
 		 * the top level, not from the 'break;' statement.
 		 */
 		ret = 1;
-		if (remaining <= 0)
+		if (remaining <= 0) {
+			if (smallest_size != -1) {
+				snum = smallest_rover;
+				goto have_snum;
+			}
 			goto fail;
-
+		}
 		/* OK, here is the one we will use.  HEAD is
 		 * non-NULL and we hold it's mutex.
 		 */
 		snum = rover;
 	} else {
+have_snum:
 		head = &hashinfo->bhash[inet_bhashfn(net, snum,
 				hashinfo->bhash_size)];
 		spin_lock(&head->lock);
@@ -145,12 +166,16 @@ int inet_csk_get_port(struct sock *sk, unsigned short snum)
 tb_found:
 	if (!hlist_empty(&tb->owners)) {
 		if (tb->fastreuse > 0 &&
-		    sk->sk_reuse && sk->sk_state != TCP_LISTEN) {
+		    sk->sk_reuse && sk->sk_state != TCP_LISTEN &&
+		    smallest_size == -1) {
 			goto success;
 		} else {
 			ret = 1;
-			if (inet_csk(sk)->icsk_af_ops->bind_conflict(sk, tb))
+			if (inet_csk(sk)->icsk_af_ops->bind_conflict(sk, tb)) {
+				if (sk->sk_reuse && sk->sk_state != TCP_LISTEN && --attempts >= 0)
+					goto again;
 				goto fail_unlock;
+			}
 		}
 	}
 tb_not_found:
net/ipv4/inet_hashtables.c +10 -1
index 4498190..399ec17 100644
@@ -38,6 +38,7 @@ struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep,
 		tb->ib_net       = hold_net(net);
 		tb->port      = snum;
 		tb->fastreuse = 0;
+		tb->num_owners = 0;
 		INIT_HLIST_HEAD(&tb->owners);
 		hlist_add_head(&tb->node, &head->chain);
 	}
@@ -59,8 +60,13 @@ void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket
 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
 		    const unsigned short snum)
 {
+	struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
+
+	hashinfo->bsockets++;
+
 	inet_sk(sk)->num = snum;
 	sk_add_bind_node(sk, &tb->owners);
+	tb->num_owners++;
 	inet_csk(sk)->icsk_bind_hash = tb;
 }
 
@@ -74,10 +80,13 @@ static void __inet_put_port(struct sock *sk)
 			hashinfo->bhash_size);
 	struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash];
 	struct inet_bind_bucket *tb;
+	
+	hashinfo->bsockets--;
 
 	spin_lock(&head->lock);
 	tb = inet_csk(sk)->icsk_bind_hash;
 	__sk_del_bind_node(sk);
+	tb->num_owners--;
 	inet_csk(sk)->icsk_bind_hash = NULL;
 	inet_sk(sk)->num = 0;
 	inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
@@ -450,9 +459,9 @@ int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 			 */
 			inet_bind_bucket_for_each(tb, node, &head->chain) {
 				if (tb->ib_net == net && tb->port == port) {
-					WARN_ON(hlist_empty(&tb->owners));
 					if (tb->fastreuse >= 0)
 						goto next_port;
+					WARN_ON(hlist_empty(&tb->owners));
 					if (!check_established(death_row, sk,
 								port, &tw))
 						goto ok;
net/ipv4/tcp_ipv4.c +2 -0
index 5c8fa7f..3e70134 100644
@@ -101,6 +101,8 @@ struct inet_hashinfo __cacheline_aligned tcp_hashinfo = {
 	.lhash_lock  = __RW_LOCK_UNLOCKED(tcp_hashinfo.lhash_lock),
 	.lhash_users = ATOMIC_INIT(0),
 	.lhash_wait  = __WAIT_QUEUE_HEAD_INITIALIZER(tcp_hashinfo.lhash_wait),
+
+	.bsockets    = 0,
 };
 
 static inline __u32 tcp_v4_init_sequence(struct sk_buff *skb)
Signature · 1 line
	Evgeniy Polyakov
maintainer
#3 lore raw

From: Evgeniy Polyakov <[email protected]>
Date: Sat, 10 Jan 2009 13:59:31 +0300

Quoted 6 lines

Ping :)

I'm getting out from the vacations tomorrow and will try to more extensively test the patch (QA group found a bug in ipv6 side, but that was with the .24 backported patch, current one does not have that codepath iirc). But I would like to know if this worth it.

I'll look at it at some point but it won't make this merge window, sorry.

#4 lore raw
Quoted 7 lines · David Miller
Quoted 4 lines

I'm getting out from the vacations tomorrow and will try to more extensively test the patch (QA group found a bug in ipv6 side, but that was with the .24 backported patch, current one does not have that codepath iirc). But I would like to know if this worth it.

I'll look at it at some point but it won't make this merge window, sorry.

Ok, please drop me a not and if things are good, merge the patch into the appropriate tree for the next window when time permits.

Signature · 1 line
	Evgeniy Polyakov
#5 lore raw

I did test on loaded squid.

49878 connections established ,1421/sec
It is not peak time yet, passed 24 hours testing.

I can't compare load, because it is real life load and always vary, but thing i can say, it doesn't crash :-)

Quoted 1 line · On Sunday 11 January 2009 14:52:06 Evgeniy Polyakov wrote:

On Sat, Jan 10, 2009 at 11:20:19PM -0800, David Miller ([email protected])

Quoted 10 lines · wrote:
Quoted 7 lines
Quoted 4 lines

I'm getting out from the vacations tomorrow and will try to more extensively test the patch (QA group found a bug in ipv6 side, but that was with the .24 backported patch, current one does not have that codepath iirc). But I would like to know if this worth it.

I'll look at it at some point but it won't make this merge window, sorry.

Ok, please drop me a not and if things are good, merge the patch into the appropriate tree for the next window when time permits.

#6 lore raw
Quoted 7 lines · Denys Fedoryschenko

I did test on loaded squid.

49878 connections established ,1421/sec
It is not peak time yet, passed 24 hours testing.

I can't compare load, because it is real life load and always vary, but thing i can say, it doesn't crash :-)

Thanks a lot for giving this a try Denys!

Signature · 1 line
	Evgeniy Polyakov
maintainer Tested-by Denys Fedoryschenko
#7 lore raw
In reply to Denys Fedoryschenko

From: Denys Fedoryschenko <[email protected]>
Date: Sun, 11 Jan 2009 15:03:10 +0200

Quoted 7 lines

I did test on loaded squid.

49878 connections established ,1421/sec
It is not peak time yet, passed 24 hours testing.

I can't compare load, because it is real life load and always vary, but thing i can say, it doesn't crash :-)

Ok, I applied Evegniy's patch, but it didn't even come WITHIN A MILE of applying to the current tree. Eric Dumazet's RCU hash table changes and net namespace macro modifications destroyed all of the context and content in these patches. It was one huge reject, even with dumb patch.

I fixed it up and applied it to what will become net-next-2.6, but that's only because I'm in a good mood and I made Evgeniy wait a long time already.

The zero compile time initializer of ->bsockets to zero was superfluous, so I dropped it entirely.

Here is what I committed.

inet: Allowing more than 64k connections and heavily optimize bind(0) time.

With simple extension to the binding mechanism, which allows to bind more than 64k sockets (or smaller amount, depending on sysctl parameters), we have to traverse the whole bind hash table to find out empty bucket. And while it is not a problem for example for 32k connections, bind() completion time grows exponentially (since after each successful binding we have to traverse one bucket more to find empty one) even if we start each time from random offset inside the hash table.

So, when hash table is full, and we want to add another socket, we have to traverse the whole table no matter what, so effectivelly this will be the worst case performance and it will be constant.

Attached picture shows bind() time depending on number of already bound sockets.

Green area corresponds to the usual binding to zero port process, which turns on kernel port selection as described above. Red area is the bind process, when number of reuse-bound sockets is not limited by 64k (or sysctl parameters). The same exponential growth (hidden by the green area) before number of ports reaches sysctl limit.

At this time bind hash table has exactly one reuse-enbaled socket in a bucket, but it is possible that they have different addresses. Actually kernel selects the first port to try randomly, so at the beginning bind will take roughly constant time, but with time number of port to check after random start will increase. And that will have exponential growth, but because of above random selection, not every next port selection will necessary take longer time than previous. So we have to consider the area below in the graph (if you could zoom it, you could find, that there are many different times placed there), so area can hide another.

Blue area corresponds to the port selection optimization.

This is rather simple design approach: hashtable now maintains (unprecise and racely updated) number of currently bound sockets, and when number of such sockets becomes greater than predefined value (I use maximum port range defined by sysctls), we stop traversing the whole bind hash table and just stop at first matching bucket after random start. Above limit roughly corresponds to the case, when bind hash table is full and we turned on mechanism of allowing to bind more reuse-enabled sockets, so it does not change behaviour of other sockets.

As I just have been told, it results in this strange behaviour:

$ grep -n :60013 netstat.res
33101:tcp        0      0 local1:60013       remote:80       ESTABLISHED
33105:tcp        0      0 local1:60013       remote:80       ESTABLISHED
52084:tcp        0      0 local2:60013       remote:80       ESTABLISHED
52085:tcp        0      0 local2:60013       remote:80       ESTABLISHED
58249:tcp        0      0 local3:60013       remote:80       ESTABLISHED

it is yet to resolve what it is and how much harm it brings :)

---
 include/net/inet_hashtables.h   |    3 ++-
 net/ipv4/inet_connection_sock.c |   39 ++++++++++++++++++++++++++++++++-------
 net/ipv4/inet_hashtables.c      |   11 ++++++++++-
 3 files changed, 44 insertions(+), 9 deletions(-)
include/net/inet_hashtables.h +2 -1
index f44bb5c..cdc08c1 100644
@@ -82,6 +82,7 @@ struct inet_bind_bucket {
 #endif
 	unsigned short		port;
 	signed short		fastreuse;
+	int			num_owners;
 	struct hlist_node	node;
 	struct hlist_head	owners;
 };
@@ -133,7 +134,7 @@ struct inet_hashinfo {
 	struct inet_bind_hashbucket	*bhash;
 
 	unsigned int			bhash_size;
-	/* Note : 4 bytes padding on 64 bit arches */
+	int				bsockets;
 
 	struct kmem_cache		*bind_bucket_cachep;
 
net/ipv4/inet_connection_sock.c +32 -7
index f26ab38..5061dbc 100644
@@ -93,24 +93,40 @@ int inet_csk_get_port(struct sock *sk, unsigned short snum)
 	struct inet_bind_hashbucket *head;
 	struct hlist_node *node;
 	struct inet_bind_bucket *tb;
-	int ret;
+	int ret, attempts = 5;
 	struct net *net = sock_net(sk);
+	int smallest_size = -1, smallest_rover;
 
 	local_bh_disable();
 	if (!snum) {
 		int remaining, rover, low, high;
 
+again:
 		inet_get_local_port_range(&low, &high);
 		remaining = (high - low) + 1;
-		rover = net_random() % remaining + low;
+		smallest_rover = rover = net_random() % remaining + low;
 
+		smallest_size = -1;
 		do {
 			head = &hashinfo->bhash[inet_bhashfn(net, rover,
 					hashinfo->bhash_size)];
 			spin_lock(&head->lock);
 			inet_bind_bucket_for_each(tb, node, &head->chain)
-				if (ib_net(tb) == net && tb->port == rover)
+				if (ib_net(tb) == net && tb->port == rover) {
+					if (tb->fastreuse > 0 &&
+					    sk->sk_reuse &&
+					    sk->sk_state != TCP_LISTEN &&
+					    (tb->num_owners < smallest_size || smallest_size == -1)) {
+						smallest_size = tb->num_owners;
+						smallest_rover = rover;
+						if (hashinfo->bsockets > (high - low) + 1) {
+							spin_unlock(&head->lock);
+							snum = smallest_rover;
+							goto have_snum;
+						}
+					}
 					goto next;
+				}
 			break;
 		next:
 			spin_unlock(&head->lock);
@@ -125,14 +141,19 @@ int inet_csk_get_port(struct sock *sk, unsigned short snum)
 		 * the top level, not from the 'break;' statement.
 		 */
 		ret = 1;
-		if (remaining <= 0)
+		if (remaining <= 0) {
+			if (smallest_size != -1) {
+				snum = smallest_rover;
+				goto have_snum;
+			}
 			goto fail;
-
+		}
 		/* OK, here is the one we will use.  HEAD is
 		 * non-NULL and we hold it's mutex.
 		 */
 		snum = rover;
 	} else {
+have_snum:
 		head = &hashinfo->bhash[inet_bhashfn(net, snum,
 				hashinfo->bhash_size)];
 		spin_lock(&head->lock);
@@ -145,12 +166,16 @@ int inet_csk_get_port(struct sock *sk, unsigned short snum)
 tb_found:
 	if (!hlist_empty(&tb->owners)) {
 		if (tb->fastreuse > 0 &&
-		    sk->sk_reuse && sk->sk_state != TCP_LISTEN) {
+		    sk->sk_reuse && sk->sk_state != TCP_LISTEN &&
+		    smallest_size == -1) {
 			goto success;
 		} else {
 			ret = 1;
-			if (inet_csk(sk)->icsk_af_ops->bind_conflict(sk, tb))
+			if (inet_csk(sk)->icsk_af_ops->bind_conflict(sk, tb)) {
+				if (sk->sk_reuse && sk->sk_state != TCP_LISTEN && --attempts >= 0)
+					goto again;
 				goto fail_unlock;
+			}
 		}
 	}
 tb_not_found:
net/ipv4/inet_hashtables.c +10 -1
index 6a1045d..d7b6178 100644
@@ -38,6 +38,7 @@ struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep,
 		write_pnet(&tb->ib_net, hold_net(net));
 		tb->port      = snum;
 		tb->fastreuse = 0;
+		tb->num_owners = 0;
 		INIT_HLIST_HEAD(&tb->owners);
 		hlist_add_head(&tb->node, &head->chain);
 	}
@@ -59,8 +60,13 @@ void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket
 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
 		    const unsigned short snum)
 {
+	struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
+
+	hashinfo->bsockets++;
+
 	inet_sk(sk)->num = snum;
 	sk_add_bind_node(sk, &tb->owners);
+	tb->num_owners++;
 	inet_csk(sk)->icsk_bind_hash = tb;
 }
 
@@ -75,9 +81,12 @@ static void __inet_put_port(struct sock *sk)
 	struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash];
 	struct inet_bind_bucket *tb;
 
+	hashinfo->bsockets--;
+
 	spin_lock(&head->lock);
 	tb = inet_csk(sk)->icsk_bind_hash;
 	__sk_del_bind_node(sk);
+	tb->num_owners--;
 	inet_csk(sk)->icsk_bind_hash = NULL;
 	inet_sk(sk)->num = 0;
 	inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
@@ -444,9 +453,9 @@ int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 			 */
 			inet_bind_bucket_for_each(tb, node, &head->chain) {
 				if (ib_net(tb) == net && tb->port == port) {
-					WARN_ON(hlist_empty(&tb->owners));
 					if (tb->fastreuse >= 0)
 						goto next_port;
+					WARN_ON(hlist_empty(&tb->owners));
 					if (!check_established(death_row, sk,
 								port, &tw))
 						goto ok;
Signature · 1 line
1.6.1.15.g159c88
#8 lore raw
Quoted 5 lines · David Miller

Ok, I applied Evegniy's patch, but it didn't even come WITHIN A MILE of applying to the current tree. Eric Dumazet's RCU hash table changes and net namespace macro modifications destroyed all of the context and content in these patches. It was one huge reject, even with dumb patch.

Thank you David.

Quoted 9 lines

As I just have been told, it results in this strange behaviour:

$ grep -n :60013 netstat.res
33101:tcp        0      0 local1:60013       remote:80       ESTABLISHED
33105:tcp        0      0 local1:60013       remote:80       ESTABLISHED
52084:tcp        0      0 local2:60013       remote:80       ESTABLISHED
52085:tcp        0      0 local2:60013       remote:80       ESTABLISHED
58249:tcp        0      0 local3:60013       remote:80       ESTABLISHED

it is yet to resolve what it is and how much harm it brings :)

It should be already resolved in the patch :)

Signature · 1 line
	Evgeniy Polyakov
#9 lore raw
In reply to Evgeniy Polyakov
Quoted 8 lines · Evgeniy Polyakov

With simple extension to the binding mechanism, which allows to bind more than 64k sockets (or smaller amount, depending on sysctl parameters), we have to traverse the whole bind hash table to find out empty bucket. And while it is not a problem for example for 32k connections, bind() completion time grows exponentially (since after each successful binding we have to traverse one bucket more to find empty one) even if we start each time from random offset inside the hash table.

Quoted 6 lines

So, when hash table is full, and we want to add another socket, we have to traverse the whole table no matter what, so effectivelly this will be the worst case performance and it will be constant.

Attached picture shows bind() time depending on number of already bound sockets.

Really attached.

Signature · 1 line
	Evgeniy Polyakov