Skip to content

Commit

Permalink
af_unix: Replace the big lock with small locks.
Browse files Browse the repository at this point in the history
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>
  • Loading branch information
q2ven authored and kuba-moo committed Nov 27, 2021
1 parent e6b4b87 commit afd20b9
Show file tree
Hide file tree
Showing 3 changed files with 71 additions and 49 deletions.
2 changes: 1 addition & 1 deletion include/net/af_unix.h
Original file line number Diff line number Diff line change
Expand Up @@ -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 {
Expand Down
98 changes: 60 additions & 38 deletions net/unix/af_unix.c
Original file line number Diff line number Diff line change
Expand Up @@ -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.
*/

Expand Down Expand Up @@ -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)
{
Expand Down Expand Up @@ -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,
Expand Down Expand Up @@ -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;
}

Expand All @@ -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
Expand Down Expand Up @@ -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);
Expand All @@ -1081,12 +1099,13 @@ static int unix_autobind(struct sock *sk)
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.
Expand All @@ -1102,7 +1121,7 @@ static int unix_autobind(struct sock *sk)
}

__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);
Expand All @@ -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;
Expand Down Expand Up @@ -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;
Expand All @@ -1181,9 +1200,9 @@ static int unix_bind_bsd(struct sock *sk, struct sockaddr_un *sunaddr,
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);
Expand All @@ -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);
Expand Down Expand Up @@ -1519,9 +1538,9 @@ static int unix_stream_connect(struct socket *sock, struct sockaddr *uaddr,
*
* 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.
*
Expand Down Expand Up @@ -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)
Expand All @@ -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);
Expand All @@ -3233,23 +3252,21 @@ 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));

return NULL;
}

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;

Expand All @@ -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)
Expand All @@ -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, ' ');

Expand Down Expand Up @@ -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__);
Expand Down
20 changes: 10 additions & 10 deletions net/unix/diag.c
Original file line number Diff line number Diff line change
Expand Up @@ -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)
Expand Down Expand Up @@ -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;
Expand All @@ -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;

Expand All @@ -237,21 +239,19 @@ static int unix_diag_dump(struct sk_buff *skb, struct netlink_callback *cb)

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;
}

Expand Down

0 comments on commit afd20b9

Please sign in to comment.