af_unix: Replace the big lock with small locks.

The hash table of AF_UNIX sockets is protected by the single lock.  This
patch replaces it with per-hash locks.

The effect is noticeable when we handle multiple sockets simultaneously.
Here is a test result on an EC2 c5.24xlarge instance.  It shows latency
(under 10us only) in unix_insert_unbound_socket() while 64 CPUs creating
1024 sockets for each in parallel.

  Without this patch:

     nsec          : count     distribution
        0          : 179      |                                        |
        500        : 3021     |*********                               |
        1000       : 6271     |*******************                     |
        1500       : 6318     |*******************                     |
        2000       : 5828     |*****************                       |
        2500       : 5124     |***************                         |
        3000       : 4426     |*************                           |
        3500       : 3672     |***********                             |
        4000       : 3138     |*********                               |
        4500       : 2811     |********                                |
        5000       : 2384     |*******                                 |
        5500       : 2023     |******                                  |
        6000       : 1954     |*****                                   |
        6500       : 1737     |*****                                   |
        7000       : 1749     |*****                                   |
        7500       : 1520     |****                                    |
        8000       : 1469     |****                                    |
        8500       : 1394     |****                                    |
        9000       : 1232     |***                                     |
        9500       : 1138     |***                                     |
        10000      : 994      |***                                     |

  With this patch:

     nsec          : count     distribution
        0          : 1634     |****                                    |
        500        : 13170    |****************************************|
        1000       : 13156    |*************************************** |
        1500       : 9010     |***************************             |
        2000       : 6363     |*******************                     |
        2500       : 4443     |*************                           |
        3000       : 3240     |*********                               |
        3500       : 2549     |*******                                 |
        4000       : 1872     |*****                                   |
        4500       : 1504     |****                                    |
        5000       : 1247     |***                                     |
        5500       : 1035     |***                                     |
        6000       : 889      |**                                      |
        6500       : 744      |**                                      |
        7000       : 634      |*                                       |
        7500       : 498      |*                                       |
        8000       : 433      |*                                       |
        8500       : 355      |*                                       |
        9000       : 336      |*                                       |
        9500       : 284      |                                        |
        10000      : 243      |                                        |

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.co.jp>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

include/net/af_unix.h

@@ -20,7 +20,7 @@ struct sock *unix_peer_get(struct sock *sk);
 #define UNIX_HASH_BITS	8
 
 extern unsigned int unix_tot_inflight;
-extern spinlock_t unix_table_lock;
+extern spinlock_t unix_table_locks[2 * UNIX_HASH_SIZE];
 extern struct hlist_head unix_socket_table[2 * UNIX_HASH_SIZE];
 
 struct unix_address {

net/unix/af_unix.c

@@ -117,14 +117,14 @@
 
 #include "scm.h"
 
+spinlock_t unix_table_locks[2 * UNIX_HASH_SIZE];
+EXPORT_SYMBOL_GPL(unix_table_locks);
 struct hlist_head unix_socket_table[2 * UNIX_HASH_SIZE];
 EXPORT_SYMBOL_GPL(unix_socket_table);
-DEFINE_SPINLOCK(unix_table_lock);
-EXPORT_SYMBOL_GPL(unix_table_lock);
 static atomic_long_t unix_nr_socks;
 
 /* SMP locking strategy:
- *    hash table is protected with spinlock unix_table_lock
+ *    hash table is protected with spinlock unix_table_locks
  *    each socket state is protected by separate spin lock.
  */
 
@@ -157,6 +157,25 @@ static unsigned int unix_abstract_hash(struct sockaddr_un *sunaddr,
 	return hash & (UNIX_HASH_SIZE - 1);
 }
 
+static void unix_table_double_lock(unsigned int hash1, unsigned int hash2)
+{
+	/* hash1 and hash2 is never the same because
+	 * one is between 0 and UNIX_HASH_SIZE - 1, and
+	 * another is between UNIX_HASH_SIZE and UNIX_HASH_SIZE * 2.
+	 */
+	if (hash1 > hash2)
+		swap(hash1, hash2);
+
+	spin_lock(&unix_table_locks[hash1]);
+	spin_lock_nested(&unix_table_locks[hash2], SINGLE_DEPTH_NESTING);
+}
+
+static void unix_table_double_unlock(unsigned int hash1, unsigned int hash2)
+{
+	spin_unlock(&unix_table_locks[hash1]);
+	spin_unlock(&unix_table_locks[hash2]);
+}
+
 #ifdef CONFIG_SECURITY_NETWORK
 static void unix_get_secdata(struct scm_cookie *scm, struct sk_buff *skb)
 {
@@ -298,16 +317,16 @@ static void __unix_set_addr_hash(struct sock *sk, struct unix_address *addr,
 
 static void unix_remove_socket(struct sock *sk)
 {
-	spin_lock(&unix_table_lock);
+	spin_lock(&unix_table_locks[sk->sk_hash]);
 	__unix_remove_socket(sk);
-	spin_unlock(&unix_table_lock);
+	spin_unlock(&unix_table_locks[sk->sk_hash]);
 }
 
 static void unix_insert_unbound_socket(struct sock *sk)
 {
-	spin_lock(&unix_table_lock);
+	spin_lock(&unix_table_locks[sk->sk_hash]);
 	__unix_insert_socket(sk);
-	spin_unlock(&unix_table_lock);
+	spin_unlock(&unix_table_locks[sk->sk_hash]);
 }
 
 static struct sock *__unix_find_socket_byname(struct net *net,
@@ -335,11 +354,11 @@ static inline struct sock *unix_find_socket_byname(struct net *net,
 {
 	struct sock *s;
 
-	spin_lock(&unix_table_lock);
+	spin_lock(&unix_table_locks[hash]);
 	s = __unix_find_socket_byname(net, sunname, len, hash);
 	if (s)
 		sock_hold(s);
-	spin_unlock(&unix_table_lock);
+	spin_unlock(&unix_table_locks[hash]);
 	return s;
 }
 
@@ -348,19 +367,18 @@ static struct sock *unix_find_socket_byinode(struct inode *i)
 	unsigned int hash = unix_bsd_hash(i);
 	struct sock *s;
 
-	spin_lock(&unix_table_lock);
+	spin_lock(&unix_table_locks[hash]);
 	sk_for_each(s, &unix_socket_table[hash]) {
 		struct dentry *dentry = unix_sk(s)->path.dentry;
 
 		if (dentry && d_backing_inode(dentry) == i) {
 			sock_hold(s);
-			goto found;
+			spin_unlock(&unix_table_locks[hash]);
+			return s;
 		}
 	}
-	s = NULL;
-found:
-	spin_unlock(&unix_table_lock);
-	return s;
+	spin_unlock(&unix_table_locks[hash]);
+	return NULL;
 }
 
 /* Support code for asymmetrically connected dgram sockets
@@ -1053,11 +1071,11 @@ static struct sock *unix_find_other(struct net *net,
 
 static int unix_autobind(struct sock *sk)
 {
+	unsigned int new_hash, old_hash = sk->sk_hash;
 	struct unix_sock *u = unix_sk(sk);
 	struct unix_address *addr;
 	unsigned int retries = 0;
 	static u32 ordernum = 1;
-	unsigned int new_hash;
 	int err;
 
 	err = mutex_lock_interruptible(&u->bindlock);
@@ -1081,12 +1099,13 @@ retry:
 		offsetof(struct sockaddr_un, sun_path) + 1;
 
 	new_hash = unix_abstract_hash(addr->name, addr->len, sk->sk_type);
-	spin_lock(&unix_table_lock);
+	unix_table_double_lock(old_hash, new_hash);
 	ordernum = (ordernum+1)&0xFFFFF;
 
 	if (__unix_find_socket_byname(sock_net(sk), addr->name, addr->len,
 				      new_hash)) {
-		spin_unlock(&unix_table_lock);
+		unix_table_double_unlock(old_hash, new_hash);
+
 		/*
 		 * __unix_find_socket_byname() may take long time if many names
 		 * are already in use.
@@ -1102,7 +1121,7 @@ retry:
 	}
 
 	__unix_set_addr_hash(sk, addr, new_hash);
-	spin_unlock(&unix_table_lock);
+	unix_table_double_unlock(old_hash, new_hash);
 	err = 0;
 
 out:	mutex_unlock(&u->bindlock);
@@ -1114,10 +1133,10 @@ static int unix_bind_bsd(struct sock *sk, struct sockaddr_un *sunaddr,
 {
 	umode_t mode = S_IFSOCK |
 	       (SOCK_INODE(sk->sk_socket)->i_mode & ~current_umask());
+	unsigned int new_hash, old_hash = sk->sk_hash;
 	struct unix_sock *u = unix_sk(sk);
 	struct user_namespace *ns; // barf...
 	struct unix_address *addr;
-	unsigned int new_hash;
 	struct dentry *dentry;
 	struct path parent;
 	int err;
@@ -1156,11 +1175,11 @@ static int unix_bind_bsd(struct sock *sk, struct sockaddr_un *sunaddr,
 		goto out_unlock;
 
 	new_hash = unix_bsd_hash(d_backing_inode(dentry));
-	spin_lock(&unix_table_lock);
+	unix_table_double_lock(old_hash, new_hash);
 	u->path.mnt = mntget(parent.mnt);
 	u->path.dentry = dget(dentry);
 	__unix_set_addr_hash(sk, addr, new_hash);
-	spin_unlock(&unix_table_lock);
+	unix_table_double_unlock(old_hash, new_hash);
 	mutex_unlock(&u->bindlock);
 	done_path_create(&parent, dentry);
 	return 0;
@@ -1181,9 +1200,9 @@ out:
 static int unix_bind_abstract(struct sock *sk, struct sockaddr_un *sunaddr,
 			      int addr_len)
 {
+	unsigned int new_hash, old_hash = sk->sk_hash;
 	struct unix_sock *u = unix_sk(sk);
 	struct unix_address *addr;
-	unsigned int new_hash;
 	int err;
 
 	addr = unix_create_addr(sunaddr, addr_len);
@@ -1200,19 +1219,19 @@ static int unix_bind_abstract(struct sock *sk, struct sockaddr_un *sunaddr,
 	}
 
 	new_hash = unix_abstract_hash(addr->name, addr->len, sk->sk_type);
-	spin_lock(&unix_table_lock);
+	unix_table_double_lock(old_hash, new_hash);
 
 	if (__unix_find_socket_byname(sock_net(sk), addr->name, addr->len,
 				      new_hash))
 		goto out_spin;
 
 	__unix_set_addr_hash(sk, addr, new_hash);
-	spin_unlock(&unix_table_lock);
+	unix_table_double_unlock(old_hash, new_hash);
 	mutex_unlock(&u->bindlock);
 	return 0;
 
 out_spin:
-	spin_unlock(&unix_table_lock);
+	unix_table_double_unlock(old_hash, new_hash);
 	err = -EADDRINUSE;
 out_mutex:
 	mutex_unlock(&u->bindlock);
@@ -1519,9 +1538,9 @@ restart:
 	 *
 	 * The contents of *(otheru->addr) and otheru->path
 	 * are seen fully set up here, since we have found
-	 * otheru in hash under unix_table_lock.  Insertion
+	 * otheru in hash under unix_table_locks.  Insertion
 	 * into the hash chain we'd found it in had been done
-	 * in an earlier critical area protected by unix_table_lock,
+	 * in an earlier critical area protected by unix_table_locks,
 	 * the same one where we'd set *(otheru->addr) contents,
 	 * as well as otheru->path and otheru->addr itself.
 	 *
@@ -3198,7 +3217,7 @@ static __poll_t unix_dgram_poll(struct file *file, struct socket *sock,
 #define BUCKET_SPACE (BITS_PER_LONG - (UNIX_HASH_BITS + 1) - 1)
 
 #define get_bucket(x) ((x) >> BUCKET_SPACE)
-#define get_offset(x) ((x) & ((1L << BUCKET_SPACE) - 1))
+#define get_offset(x) ((x) & ((1UL << BUCKET_SPACE) - 1))
 #define set_bucket_offset(b, o) ((b) << BUCKET_SPACE | (o))
 
 static struct sock *unix_from_bucket(struct seq_file *seq, loff_t *pos)
@@ -3222,7 +3241,7 @@ static struct sock *unix_next_socket(struct seq_file *seq,
 				     struct sock *sk,
 				     loff_t *pos)
 {
-	unsigned long bucket;
+	unsigned long bucket = get_bucket(*pos);
 
 	while (sk > (struct sock *)SEQ_START_TOKEN) {
 		sk = sk_next(sk);
@@ -3233,12 +3252,13 @@ static struct sock *unix_next_socket(struct seq_file *seq,
 	}
 
 	do {
+		spin_lock(&unix_table_locks[bucket]);
 		sk = unix_from_bucket(seq, pos);
 		if (sk)
 			return sk;
 
 next_bucket:
-		bucket = get_bucket(*pos) + 1;
+		spin_unlock(&unix_table_locks[bucket++]);
 		*pos = set_bucket_offset(bucket, 1);
 	} while (bucket < ARRAY_SIZE(unix_socket_table));
 
@@ -3246,10 +3266,7 @@ next_bucket:
 }
 
 static void *unix_seq_start(struct seq_file *seq, loff_t *pos)
-	__acquires(unix_table_lock)
 {
-	spin_lock(&unix_table_lock);
-
 	if (!*pos)
 		return SEQ_START_TOKEN;
 
@@ -3266,9 +3283,11 @@ static void *unix_seq_next(struct seq_file *seq, void *v, loff_t *pos)
 }
 
 static void unix_seq_stop(struct seq_file *seq, void *v)
-	__releases(unix_table_lock)
 {
-	spin_unlock(&unix_table_lock);
+	struct sock *sk = v;
+
+	if (sk)
+		spin_unlock(&unix_table_locks[sk->sk_hash]);
 }
 
 static int unix_seq_show(struct seq_file *seq, void *v)
@@ -3293,7 +3312,7 @@ static int unix_seq_show(struct seq_file *seq, void *v)
 			(s->sk_state == TCP_ESTABLISHED ? SS_CONNECTING : SS_DISCONNECTING),
 			sock_i_ino(s));
 
-		if (u->addr) {	// under unix_table_lock here
+		if (u->addr) {	// under unix_table_locks here
 			int i, len;
 			seq_putc(seq, ' ');
 
@@ -3452,10 +3471,13 @@ static void __init bpf_iter_register(void)
 
 static int __init af_unix_init(void)
 {
-	int rc = -1;
+	int i, rc = -1;
 
 	BUILD_BUG_ON(sizeof(struct unix_skb_parms) > sizeof_field(struct sk_buff, cb));
 
+	for (i = 0; i < 2 * UNIX_HASH_SIZE; i++)
+		spin_lock_init(&unix_table_locks[i]);
+
 	rc = proto_register(&unix_dgram_proto, 1);
 	if (rc != 0) {
 		pr_crit("%s: Cannot create unix_sock SLAB cache!\n", __func__);

net/unix/diag.c

@@ -13,7 +13,7 @@
 
 static int sk_diag_dump_name(struct sock *sk, struct sk_buff *nlskb)
 {
-	/* might or might not have unix_table_lock */
+	/* might or might not have unix_table_locks */
 	struct unix_address *addr = smp_load_acquire(&unix_sk(sk)->addr);
 
 	if (!addr)
@@ -204,13 +204,13 @@ static int unix_diag_dump(struct sk_buff *skb, struct netlink_callback *cb)
 	s_slot = cb->args[0];
 	num = s_num = cb->args[1];
 
-	spin_lock(&unix_table_lock);
 	for (slot = s_slot;
 	     slot < ARRAY_SIZE(unix_socket_table);
 	     s_num = 0, slot++) {
 		struct sock *sk;
 
 		num = 0;
+		spin_lock(&unix_table_locks[slot]);
 		sk_for_each(sk, &unix_socket_table[slot]) {
 			if (!net_eq(sock_net(sk), net))
 				continue;
@@ -221,14 +221,16 @@ static int unix_diag_dump(struct sk_buff *skb, struct netlink_callback *cb)
 			if (sk_diag_dump(sk, skb, req,
 					 NETLINK_CB(cb->skb).portid,
 					 cb->nlh->nlmsg_seq,
-					 NLM_F_MULTI) < 0)
+					 NLM_F_MULTI) < 0) {
+				spin_unlock(&unix_table_locks[slot]);
 				goto done;
+			}
 next:
 			num++;
 		}
+		spin_unlock(&unix_table_locks[slot]);
 	}
 done:
-	spin_unlock(&unix_table_lock);
 	cb->args[0] = slot;
 	cb->args[1] = num;
 
@@ -237,21 +239,19 @@ done:
 
 static struct sock *unix_lookup_by_ino(unsigned int ino)
 {
-	int i;
 	struct sock *sk;
+	int i;
 
-	spin_lock(&unix_table_lock);
 	for (i = 0; i < ARRAY_SIZE(unix_socket_table); i++) {
+		spin_lock(&unix_table_locks[i]);
 		sk_for_each(sk, &unix_socket_table[i])
 			if (ino == sock_i_ino(sk)) {
 				sock_hold(sk);
-				spin_unlock(&unix_table_lock);
-
+				spin_unlock(&unix_table_locks[i]);
 				return sk;
 			}
+		spin_unlock(&unix_table_locks[i]);
 	}
-
-	spin_unlock(&unix_table_lock);
 	return NULL;
 }