[INET]: Remove per bucket rwlock in tcp/dccp ehash table.
Eric Dumazet [Wed, 7 Nov 2007 10:40:20 +0000 (02:40 -0800)]
As done two years ago on IP route cache table (commit
22c047ccbc68fa8f3fa57f0e8f906479a062c426) , we can avoid using one
lock per hash bucket for the huge TCP/DCCP hash tables.

On a typical x86_64 platform, this saves about 2MB or 4MB of ram, for
litle performance differences. (we hit a different cache line for the
rwlock, but then the bucket cache line have a better sharing factor
among cpus, since we dirty it less often). For netstat or ss commands
that want a full scan of hash table, we perform fewer memory accesses.

Using a 'small' table of hashed rwlocks should be more than enough to
provide correct SMP concurrency between different buckets, without
using too much memory. Sizing of this table depends on
num_possible_cpus() and various CONFIG settings.

This patch provides some locking abstraction that may ease a future
work using a different model for TCP/DCCP table.

Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Acked-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

include/net/inet_hashtables.h
net/dccp/proto.c
net/ipv4/inet_diag.c
net/ipv4/inet_hashtables.c
net/ipv4/inet_timewait_sock.c
net/ipv4/tcp.c
net/ipv4/tcp_ipv4.c
net/ipv6/inet6_hashtables.c

index 4427dcd..8461cda 100644 (file)
@@ -37,7 +37,6 @@
  * I'll experiment with dynamic table growth later.
  */
 struct inet_ehash_bucket {
-       rwlock_t          lock;
        struct hlist_head chain;
        struct hlist_head twchain;
 };
@@ -100,6 +99,9 @@ struct inet_hashinfo {
         * TIME_WAIT sockets use a separate chain (twchain).
         */
        struct inet_ehash_bucket        *ehash;
+       rwlock_t                        *ehash_locks;
+       unsigned int                    ehash_size;
+       unsigned int                    ehash_locks_mask;
 
        /* Ok, let's try this, I give up, we do need a local binding
         * TCP hash as well as the others for fast bind/connect.
@@ -107,7 +109,7 @@ struct inet_hashinfo {
        struct inet_bind_hashbucket     *bhash;
 
        unsigned int                    bhash_size;
-       unsigned int                    ehash_size;
+       /* Note : 4 bytes padding on 64 bit arches */
 
        /* 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
@@ -134,6 +136,62 @@ static inline struct inet_ehash_bucket *inet_ehash_bucket(
        return &hashinfo->ehash[hash & (hashinfo->ehash_size - 1)];
 }
 
+static inline rwlock_t *inet_ehash_lockp(
+       struct inet_hashinfo *hashinfo,
+       unsigned int hash)
+{
+       return &hashinfo->ehash_locks[hash & hashinfo->ehash_locks_mask];
+}
+
+static inline int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
+{
+       unsigned int i, size = 256;
+#if defined(CONFIG_PROVE_LOCKING)
+       unsigned int nr_pcpus = 2;
+#else
+       unsigned int nr_pcpus = num_possible_cpus();
+#endif
+       if (nr_pcpus >= 4)
+               size = 512;
+       if (nr_pcpus >= 8)
+               size = 1024;
+       if (nr_pcpus >= 16)
+               size = 2048;
+       if (nr_pcpus >= 32)
+               size = 4096;
+       if (sizeof(rwlock_t) != 0) {
+#ifdef CONFIG_NUMA
+               if (size * sizeof(rwlock_t) > PAGE_SIZE)
+                       hashinfo->ehash_locks = vmalloc(size * sizeof(rwlock_t));
+               else
+#endif
+               hashinfo->ehash_locks = kmalloc(size * sizeof(rwlock_t),
+                                               GFP_KERNEL);
+               if (!hashinfo->ehash_locks)
+                       return ENOMEM;
+               for (i = 0; i < size; i++)
+                       rwlock_init(&hashinfo->ehash_locks[i]);
+       }
+       hashinfo->ehash_locks_mask = size - 1;
+       return 0;
+}
+
+static inline void inet_ehash_locks_free(struct inet_hashinfo *hashinfo)
+{
+       if (hashinfo->ehash_locks) {
+#ifdef CONFIG_NUMA
+               unsigned int size = (hashinfo->ehash_locks_mask + 1) *
+                                                       sizeof(rwlock_t);
+               if (size > PAGE_SIZE)
+                       vfree(hashinfo->ehash_locks);
+               else
+#else
+               kfree(hashinfo->ehash_locks);
+#endif
+               hashinfo->ehash_locks = NULL;
+       }
+}
+
 extern struct inet_bind_bucket *
                    inet_bind_bucket_create(struct kmem_cache *cachep,
                                            struct inet_bind_hashbucket *head,
@@ -222,7 +280,7 @@ static inline void __inet_hash(struct inet_hashinfo *hashinfo,
                sk->sk_hash = inet_sk_ehashfn(sk);
                head = inet_ehash_bucket(hashinfo, sk->sk_hash);
                list = &head->chain;
-               lock = &head->lock;
+               lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
                write_lock(lock);
        }
        __sk_add_node(sk, list);
@@ -253,7 +311,7 @@ static inline void inet_unhash(struct inet_hashinfo *hashinfo, struct sock *sk)
                inet_listen_wlock(hashinfo);
                lock = &hashinfo->lhash_lock;
        } else {
-               lock = &inet_ehash_bucket(hashinfo, sk->sk_hash)->lock;
+               lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
                write_lock_bh(lock);
        }
 
@@ -354,9 +412,10 @@ static inline struct sock *
         */
        unsigned int hash = inet_ehashfn(daddr, hnum, saddr, sport);
        struct inet_ehash_bucket *head = inet_ehash_bucket(hashinfo, hash);
+       rwlock_t *lock = inet_ehash_lockp(hashinfo, hash);
 
        prefetch(head->chain.first);
-       read_lock(&head->lock);
+       read_lock(lock);
        sk_for_each(sk, node, &head->chain) {
                if (INET_MATCH(sk, hash, acookie, saddr, daddr, ports, dif))
                        goto hit; /* You sunk my battleship! */
@@ -369,7 +428,7 @@ static inline struct sock *
        }
        sk = NULL;
 out:
-       read_unlock(&head->lock);
+       read_unlock(lock);
        return sk;
 hit:
        sock_hold(sk);
index d849739..7a3bea9 100644 (file)
@@ -1072,11 +1072,13 @@ static int __init dccp_init(void)
        }
 
        for (i = 0; i < dccp_hashinfo.ehash_size; i++) {
-               rwlock_init(&dccp_hashinfo.ehash[i].lock);
                INIT_HLIST_HEAD(&dccp_hashinfo.ehash[i].chain);
                INIT_HLIST_HEAD(&dccp_hashinfo.ehash[i].twchain);
        }
 
+       if (inet_ehash_locks_alloc(&dccp_hashinfo))
+                       goto out_free_dccp_ehash;
+
        bhash_order = ehash_order;
 
        do {
@@ -1091,7 +1093,7 @@ static int __init dccp_init(void)
 
        if (!dccp_hashinfo.bhash) {
                DCCP_CRIT("Failed to allocate DCCP bind hash table");
-               goto out_free_dccp_ehash;
+               goto out_free_dccp_locks;
        }
 
        for (i = 0; i < dccp_hashinfo.bhash_size; i++) {
@@ -1121,6 +1123,8 @@ out_free_dccp_mib:
 out_free_dccp_bhash:
        free_pages((unsigned long)dccp_hashinfo.bhash, bhash_order);
        dccp_hashinfo.bhash = NULL;
+out_free_dccp_locks:
+       inet_ehash_locks_free(&dccp_hashinfo);
 out_free_dccp_ehash:
        free_pages((unsigned long)dccp_hashinfo.ehash, ehash_order);
        dccp_hashinfo.ehash = NULL;
@@ -1139,6 +1143,7 @@ static void __exit dccp_fini(void)
        free_pages((unsigned long)dccp_hashinfo.ehash,
                   get_order(dccp_hashinfo.ehash_size *
                             sizeof(struct inet_ehash_bucket)));
+       inet_ehash_locks_free(&dccp_hashinfo);
        kmem_cache_destroy(dccp_hashinfo.bind_bucket_cachep);
        dccp_ackvec_exit();
        dccp_sysctl_exit();
index dc429b6..b017073 100644 (file)
@@ -747,13 +747,14 @@ skip_listen_ht:
 
        for (i = s_i; i < hashinfo->ehash_size; i++) {
                struct inet_ehash_bucket *head = &hashinfo->ehash[i];
+               rwlock_t *lock = inet_ehash_lockp(hashinfo, i);
                struct sock *sk;
                struct hlist_node *node;
 
                if (i > s_i)
                        s_num = 0;
 
-               read_lock_bh(&head->lock);
+               read_lock_bh(lock);
                num = 0;
                sk_for_each(sk, node, &head->chain) {
                        struct inet_sock *inet = inet_sk(sk);
@@ -769,7 +770,7 @@ skip_listen_ht:
                            r->id.idiag_dport)
                                goto next_normal;
                        if (inet_csk_diag_dump(sk, skb, cb) < 0) {
-                               read_unlock_bh(&head->lock);
+                               read_unlock_bh(lock);
                                goto done;
                        }
 next_normal:
@@ -791,14 +792,14 @@ next_normal:
                                    r->id.idiag_dport)
                                        goto next_dying;
                                if (inet_twsk_diag_dump(tw, skb, cb) < 0) {
-                                       read_unlock_bh(&head->lock);
+                                       read_unlock_bh(lock);
                                        goto done;
                                }
 next_dying:
                                ++num;
                        }
                }
-               read_unlock_bh(&head->lock);
+               read_unlock_bh(lock);
        }
 
 done:
index 16eecc7..67704da 100644 (file)
@@ -204,12 +204,13 @@ static int __inet_check_established(struct inet_timewait_death_row *death_row,
        const __portpair ports = INET_COMBINED_PORTS(inet->dport, lport);
        unsigned int hash = inet_ehashfn(daddr, lport, saddr, inet->dport);
        struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
+       rwlock_t *lock = inet_ehash_lockp(hinfo, hash);
        struct sock *sk2;
        const struct hlist_node *node;
        struct inet_timewait_sock *tw;
 
        prefetch(head->chain.first);
-       write_lock(&head->lock);
+       write_lock(lock);
 
        /* Check TIME-WAIT sockets first. */
        sk_for_each(sk2, node, &head->twchain) {
@@ -239,7 +240,7 @@ unique:
        BUG_TRAP(sk_unhashed(sk));
        __sk_add_node(sk, &head->chain);
        sock_prot_inc_use(sk->sk_prot);
-       write_unlock(&head->lock);
+       write_unlock(lock);
 
        if (twp) {
                *twp = tw;
@@ -255,7 +256,7 @@ unique:
        return 0;
 
 not_unique:
-       write_unlock(&head->lock);
+       write_unlock(lock);
        return -EADDRNOTAVAIL;
 }
 
index 4e189e2..a60b99e 100644 (file)
@@ -20,16 +20,16 @@ static void __inet_twsk_kill(struct inet_timewait_sock *tw,
        struct inet_bind_hashbucket *bhead;
        struct inet_bind_bucket *tb;
        /* Unlink from established hashes. */
-       struct inet_ehash_bucket *ehead = inet_ehash_bucket(hashinfo, tw->tw_hash);
+       rwlock_t *lock = inet_ehash_lockp(hashinfo, tw->tw_hash);
 
-       write_lock(&ehead->lock);
+       write_lock(lock);
        if (hlist_unhashed(&tw->tw_node)) {
-               write_unlock(&ehead->lock);
+               write_unlock(lock);
                return;
        }
        __hlist_del(&tw->tw_node);
        sk_node_init(&tw->tw_node);
-       write_unlock(&ehead->lock);
+       write_unlock(lock);
 
        /* Disassociate with bind bucket. */
        bhead = &hashinfo->bhash[inet_bhashfn(tw->tw_num, hashinfo->bhash_size)];
@@ -59,6 +59,7 @@ void __inet_twsk_hashdance(struct inet_timewait_sock *tw, struct sock *sk,
        const struct inet_sock *inet = inet_sk(sk);
        const struct inet_connection_sock *icsk = inet_csk(sk);
        struct inet_ehash_bucket *ehead = inet_ehash_bucket(hashinfo, sk->sk_hash);
+       rwlock_t *lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
        struct inet_bind_hashbucket *bhead;
        /* Step 1: Put TW into bind hash. Original socket stays there too.
           Note, that any socket with inet->num != 0 MUST be bound in
@@ -71,7 +72,7 @@ void __inet_twsk_hashdance(struct inet_timewait_sock *tw, struct sock *sk,
        inet_twsk_add_bind_node(tw, &tw->tw_tb->owners);
        spin_unlock(&bhead->lock);
 
-       write_lock(&ehead->lock);
+       write_lock(lock);
 
        /* Step 2: Remove SK from established hash. */
        if (__sk_del_node_init(sk))
@@ -81,7 +82,7 @@ void __inet_twsk_hashdance(struct inet_timewait_sock *tw, struct sock *sk,
        inet_twsk_add_node(tw, &ehead->twchain);
        atomic_inc(&tw->tw_refcnt);
 
-       write_unlock(&ehead->lock);
+       write_unlock(lock);
 }
 
 EXPORT_SYMBOL_GPL(__inet_twsk_hashdance);
index c64072b..8e65182 100644 (file)
@@ -2456,11 +2456,11 @@ void __init tcp_init(void)
                                        thash_entries ? 0 : 512 * 1024);
        tcp_hashinfo.ehash_size = 1 << tcp_hashinfo.ehash_size;
        for (i = 0; i < tcp_hashinfo.ehash_size; i++) {
-               rwlock_init(&tcp_hashinfo.ehash[i].lock);
                INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].chain);
                INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].twchain);
        }
-
+       if (inet_ehash_locks_alloc(&tcp_hashinfo))
+               panic("TCP: failed to alloc ehash_locks");
        tcp_hashinfo.bhash =
                alloc_large_system_hash("TCP bind",
                                        sizeof(struct inet_bind_hashbucket),
index e9127cd..e566f3c 100644 (file)
@@ -2049,8 +2049,9 @@ static void *established_get_first(struct seq_file *seq)
                struct sock *sk;
                struct hlist_node *node;
                struct inet_timewait_sock *tw;
+               rwlock_t *lock = inet_ehash_lockp(&tcp_hashinfo, st->bucket);
 
-               read_lock_bh(&tcp_hashinfo.ehash[st->bucket].lock);
+               read_lock_bh(lock);
                sk_for_each(sk, node, &tcp_hashinfo.ehash[st->bucket].chain) {
                        if (sk->sk_family != st->family) {
                                continue;
@@ -2067,7 +2068,7 @@ static void *established_get_first(struct seq_file *seq)
                        rc = tw;
                        goto out;
                }
-               read_unlock_bh(&tcp_hashinfo.ehash[st->bucket].lock);
+               read_unlock_bh(lock);
                st->state = TCP_SEQ_STATE_ESTABLISHED;
        }
 out:
@@ -2094,11 +2095,11 @@ get_tw:
                        cur = tw;
                        goto out;
                }
-               read_unlock_bh(&tcp_hashinfo.ehash[st->bucket].lock);
+               read_unlock_bh(inet_ehash_lockp(&tcp_hashinfo, st->bucket));
                st->state = TCP_SEQ_STATE_ESTABLISHED;
 
                if (++st->bucket < tcp_hashinfo.ehash_size) {
-                       read_lock_bh(&tcp_hashinfo.ehash[st->bucket].lock);
+                       read_lock_bh(inet_ehash_lockp(&tcp_hashinfo, st->bucket));
                        sk = sk_head(&tcp_hashinfo.ehash[st->bucket].chain);
                } else {
                        cur = NULL;
@@ -2206,7 +2207,7 @@ static void tcp_seq_stop(struct seq_file *seq, void *v)
        case TCP_SEQ_STATE_TIME_WAIT:
        case TCP_SEQ_STATE_ESTABLISHED:
                if (v)
-                       read_unlock_bh(&tcp_hashinfo.ehash[st->bucket].lock);
+                       read_unlock_bh(inet_ehash_lockp(&tcp_hashinfo, st->bucket));
                break;
        }
 }
index d6f1026..adc73ad 100644 (file)
@@ -37,9 +37,8 @@ void __inet6_hash(struct inet_hashinfo *hashinfo,
        } else {
                unsigned int hash;
                sk->sk_hash = hash = inet6_sk_ehashfn(sk);
-               hash &= (hashinfo->ehash_size - 1);
-               list = &hashinfo->ehash[hash].chain;
-               lock = &hashinfo->ehash[hash].lock;
+               list = &inet_ehash_bucket(hashinfo, hash)->chain;
+               lock = inet_ehash_lockp(hashinfo, hash);
                write_lock(lock);
        }
 
@@ -70,9 +69,10 @@ struct sock *__inet6_lookup_established(struct inet_hashinfo *hashinfo,
         */
        unsigned int hash = inet6_ehashfn(daddr, hnum, saddr, sport);
        struct inet_ehash_bucket *head = inet_ehash_bucket(hashinfo, hash);
+       rwlock_t *lock = inet_ehash_lockp(hashinfo, hash);
 
        prefetch(head->chain.first);
-       read_lock(&head->lock);
+       read_lock(lock);
        sk_for_each(sk, node, &head->chain) {
                /* For IPV6 do the cheaper port and family tests first. */
                if (INET6_MATCH(sk, hash, saddr, daddr, ports, dif))
@@ -92,12 +92,12 @@ struct sock *__inet6_lookup_established(struct inet_hashinfo *hashinfo,
                                goto hit;
                }
        }
-       read_unlock(&head->lock);
+       read_unlock(lock);
        return NULL;
 
 hit:
        sock_hold(sk);
-       read_unlock(&head->lock);
+       read_unlock(lock);
        return sk;
 }
 EXPORT_SYMBOL(__inet6_lookup_established);
@@ -175,12 +175,13 @@ static int __inet6_check_established(struct inet_timewait_death_row *death_row,
        const unsigned int hash = inet6_ehashfn(daddr, lport, saddr,
                                                inet->dport);
        struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
+       rwlock_t *lock = inet_ehash_lockp(hinfo, hash);
        struct sock *sk2;
        const struct hlist_node *node;
        struct inet_timewait_sock *tw;
 
        prefetch(head->chain.first);
-       write_lock(&head->lock);
+       write_lock(lock);
 
        /* Check TIME-WAIT sockets first. */
        sk_for_each(sk2, node, &head->twchain) {
@@ -216,7 +217,7 @@ unique:
        __sk_add_node(sk, &head->chain);
        sk->sk_hash = hash;
        sock_prot_inc_use(sk->sk_prot);
-       write_unlock(&head->lock);
+       write_unlock(lock);
 
        if (twp != NULL) {
                *twp = tw;
@@ -231,7 +232,7 @@ unique:
        return 0;
 
 not_unique:
-       write_unlock(&head->lock);
+       write_unlock(lock);
        return -EADDRNOTAVAIL;
 }