2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Generic INET transport hashtables
8 * Authors: Lotsa people, from code originally in tcp
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public License
12 * as published by the Free Software Foundation; either version
13 * 2 of the License, or (at your option) any later version.
16 #include <linux/module.h>
17 #include <linux/random.h>
18 #include <linux/sched.h>
19 #include <linux/slab.h>
20 #include <linux/wait.h>
22 #include <net/inet_connection_sock.h>
23 #include <net/inet_hashtables.h>
24 #include <net/secure_seq.h>
27 static unsigned int inet_ehashfn(struct net *net, const __be32 laddr,
28 const __u16 lport, const __be32 faddr,
31 static u32 inet_ehash_secret __read_mostly;
33 net_get_random_once(&inet_ehash_secret, sizeof(inet_ehash_secret));
35 return __inet_ehashfn(laddr, lport, faddr, fport,
36 inet_ehash_secret + net_hash_mix(net));
40 static unsigned int inet_sk_ehashfn(const struct sock *sk)
42 const struct inet_sock *inet = inet_sk(sk);
43 const __be32 laddr = inet->inet_rcv_saddr;
44 const __u16 lport = inet->inet_num;
45 const __be32 faddr = inet->inet_daddr;
46 const __be16 fport = inet->inet_dport;
47 struct net *net = sock_net(sk);
49 return inet_ehashfn(net, laddr, lport, faddr, fport);
53 * Allocate and initialize a new local port bind bucket.
54 * The bindhash mutex for snum's hash chain must be held here.
56 struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep,
58 struct inet_bind_hashbucket *head,
59 const unsigned short snum)
61 struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
64 write_pnet(&tb->ib_net, net);
67 tb->fastreuseport = 0;
69 INIT_HLIST_HEAD(&tb->owners);
70 hlist_add_head(&tb->node, &head->chain);
76 * Caller must hold hashbucket lock for this tb with local BH disabled
78 void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
80 if (hlist_empty(&tb->owners)) {
81 __hlist_del(&tb->node);
82 kmem_cache_free(cachep, tb);
86 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
87 const unsigned short snum)
89 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
91 atomic_inc(&hashinfo->bsockets);
93 inet_sk(sk)->inet_num = snum;
94 sk_add_bind_node(sk, &tb->owners);
96 inet_csk(sk)->icsk_bind_hash = tb;
100 * Get rid of any references to a local port held by the given sock.
102 static void __inet_put_port(struct sock *sk)
104 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
105 const int bhash = inet_bhashfn(sock_net(sk), inet_sk(sk)->inet_num,
106 hashinfo->bhash_size);
107 struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash];
108 struct inet_bind_bucket *tb;
110 atomic_dec(&hashinfo->bsockets);
112 spin_lock(&head->lock);
113 tb = inet_csk(sk)->icsk_bind_hash;
114 __sk_del_bind_node(sk);
116 inet_csk(sk)->icsk_bind_hash = NULL;
117 inet_sk(sk)->inet_num = 0;
118 inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
119 spin_unlock(&head->lock);
122 void inet_put_port(struct sock *sk)
128 EXPORT_SYMBOL(inet_put_port);
130 int __inet_inherit_port(struct sock *sk, struct sock *child)
132 struct inet_hashinfo *table = sk->sk_prot->h.hashinfo;
133 unsigned short port = inet_sk(child)->inet_num;
134 const int bhash = inet_bhashfn(sock_net(sk), port,
136 struct inet_bind_hashbucket *head = &table->bhash[bhash];
137 struct inet_bind_bucket *tb;
139 spin_lock(&head->lock);
140 tb = inet_csk(sk)->icsk_bind_hash;
141 if (tb->port != port) {
142 /* NOTE: using tproxy and redirecting skbs to a proxy
143 * on a different listener port breaks the assumption
144 * that the listener socket's icsk_bind_hash is the same
145 * as that of the child socket. We have to look up or
146 * create a new bind bucket for the child here. */
147 inet_bind_bucket_for_each(tb, &head->chain) {
148 if (net_eq(ib_net(tb), sock_net(sk)) &&
153 tb = inet_bind_bucket_create(table->bind_bucket_cachep,
154 sock_net(sk), head, port);
156 spin_unlock(&head->lock);
161 inet_bind_hash(child, tb, port);
162 spin_unlock(&head->lock);
166 EXPORT_SYMBOL_GPL(__inet_inherit_port);
168 static inline int compute_score(struct sock *sk, struct net *net,
169 const unsigned short hnum, const __be32 daddr,
173 struct inet_sock *inet = inet_sk(sk);
175 if (net_eq(sock_net(sk), net) && inet->inet_num == hnum &&
176 !ipv6_only_sock(sk)) {
177 __be32 rcv_saddr = inet->inet_rcv_saddr;
178 score = sk->sk_family == PF_INET ? 2 : 1;
180 if (rcv_saddr != daddr)
184 if (sk->sk_bound_dev_if) {
185 if (sk->sk_bound_dev_if != dif)
194 * Don't inline this cruft. Here are some nice properties to exploit here. The
195 * BSD API does not allow a listening sock to specify the remote port nor the
196 * remote address for the connection. So always assume those are both
197 * wildcarded during the search since they can never be otherwise.
201 struct sock *__inet_lookup_listener(struct net *net,
202 struct inet_hashinfo *hashinfo,
203 const __be32 saddr, __be16 sport,
204 const __be32 daddr, const unsigned short hnum,
207 struct sock *sk, *result;
208 struct hlist_nulls_node *node;
209 unsigned int hash = inet_lhashfn(net, hnum);
210 struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash];
211 int score, hiscore, matches = 0, reuseport = 0;
218 sk_nulls_for_each_rcu(sk, node, &ilb->head) {
219 score = compute_score(sk, net, hnum, daddr, dif);
220 if (score > hiscore) {
223 reuseport = sk->sk_reuseport;
225 phash = inet_ehashfn(net, daddr, hnum,
229 } else if (score == hiscore && reuseport) {
231 if (reciprocal_scale(phash, matches) == 0)
233 phash = next_pseudo_random32(phash);
237 * if the nulls value we got at the end of this lookup is
238 * not the expected one, we must restart lookup.
239 * We probably met an item that was moved to another chain.
241 if (get_nulls_value(node) != hash + LISTENING_NULLS_BASE)
244 if (unlikely(!atomic_inc_not_zero(&result->sk_refcnt)))
246 else if (unlikely(compute_score(result, net, hnum, daddr,
255 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
257 /* All sockets share common refcount, but have different destructors */
258 void sock_gen_put(struct sock *sk)
260 if (!atomic_dec_and_test(&sk->sk_refcnt))
263 if (sk->sk_state == TCP_TIME_WAIT)
264 inet_twsk_free(inet_twsk(sk));
265 else if (sk->sk_state == TCP_NEW_SYN_RECV)
266 reqsk_free(inet_reqsk(sk));
270 EXPORT_SYMBOL_GPL(sock_gen_put);
272 struct sock *__inet_lookup_established(struct net *net,
273 struct inet_hashinfo *hashinfo,
274 const __be32 saddr, const __be16 sport,
275 const __be32 daddr, const u16 hnum,
278 INET_ADDR_COOKIE(acookie, saddr, daddr);
279 const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
281 const struct hlist_nulls_node *node;
282 /* Optimize here for direct hit, only listening connections can
283 * have wildcards anyways.
285 unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
286 unsigned int slot = hash & hashinfo->ehash_mask;
287 struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
291 sk_nulls_for_each_rcu(sk, node, &head->chain) {
292 if (sk->sk_hash != hash)
294 if (likely(INET_MATCH(sk, net, acookie,
295 saddr, daddr, ports, dif))) {
296 if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt)))
298 if (unlikely(!INET_MATCH(sk, net, acookie,
299 saddr, daddr, ports, dif))) {
307 * if the nulls value we got at the end of this lookup is
308 * not the expected one, we must restart lookup.
309 * We probably met an item that was moved to another chain.
311 if (get_nulls_value(node) != slot)
319 EXPORT_SYMBOL_GPL(__inet_lookup_established);
321 /* called with local bh disabled */
322 static int __inet_check_established(struct inet_timewait_death_row *death_row,
323 struct sock *sk, __u16 lport,
324 struct inet_timewait_sock **twp)
326 struct inet_hashinfo *hinfo = death_row->hashinfo;
327 struct inet_sock *inet = inet_sk(sk);
328 __be32 daddr = inet->inet_rcv_saddr;
329 __be32 saddr = inet->inet_daddr;
330 int dif = sk->sk_bound_dev_if;
331 INET_ADDR_COOKIE(acookie, saddr, daddr);
332 const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
333 struct net *net = sock_net(sk);
334 unsigned int hash = inet_ehashfn(net, daddr, lport,
335 saddr, inet->inet_dport);
336 struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
337 spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
339 const struct hlist_nulls_node *node;
340 struct inet_timewait_sock *tw = NULL;
345 sk_nulls_for_each(sk2, node, &head->chain) {
346 if (sk2->sk_hash != hash)
349 if (likely(INET_MATCH(sk2, net, acookie,
350 saddr, daddr, ports, dif))) {
351 if (sk2->sk_state == TCP_TIME_WAIT) {
353 if (twsk_unique(sk, sk2, twp))
360 /* Must record num and sport now. Otherwise we will see
361 * in hash table socket with a funny identity.
363 inet->inet_num = lport;
364 inet->inet_sport = htons(lport);
366 WARN_ON(!sk_unhashed(sk));
367 __sk_nulls_add_node_rcu(sk, &head->chain);
369 twrefcnt = inet_twsk_unhash(tw);
370 NET_INC_STATS_BH(net, LINUX_MIB_TIMEWAITRECYCLED);
375 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
380 /* Silly. Should hash-dance instead... */
381 inet_twsk_deschedule(tw, death_row);
389 return -EADDRNOTAVAIL;
392 static inline u32 inet_sk_port_offset(const struct sock *sk)
394 const struct inet_sock *inet = inet_sk(sk);
395 return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
400 int __inet_hash_nolisten(struct sock *sk, struct inet_timewait_sock *tw)
402 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
403 struct hlist_nulls_head *list;
405 struct inet_ehash_bucket *head;
408 WARN_ON(!sk_unhashed(sk));
410 sk->sk_hash = inet_sk_ehashfn(sk);
411 head = inet_ehash_bucket(hashinfo, sk->sk_hash);
413 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
416 __sk_nulls_add_node_rcu(sk, list);
418 WARN_ON(sk->sk_hash != tw->tw_hash);
419 twrefcnt = inet_twsk_unhash(tw);
422 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
425 EXPORT_SYMBOL_GPL(__inet_hash_nolisten);
427 static void __inet_hash(struct sock *sk)
429 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
430 struct inet_listen_hashbucket *ilb;
432 if (sk->sk_state != TCP_LISTEN) {
433 __inet_hash_nolisten(sk, NULL);
437 WARN_ON(!sk_unhashed(sk));
438 ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
440 spin_lock(&ilb->lock);
441 __sk_nulls_add_node_rcu(sk, &ilb->head);
442 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
443 spin_unlock(&ilb->lock);
446 void inet_hash(struct sock *sk)
448 if (sk->sk_state != TCP_CLOSE) {
454 EXPORT_SYMBOL_GPL(inet_hash);
456 void inet_unhash(struct sock *sk)
458 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
465 if (sk->sk_state == TCP_LISTEN)
466 lock = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)].lock;
468 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
471 done = __sk_nulls_del_node_init_rcu(sk);
473 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
474 spin_unlock_bh(lock);
476 EXPORT_SYMBOL_GPL(inet_unhash);
478 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
479 struct sock *sk, u32 port_offset,
480 int (*check_established)(struct inet_timewait_death_row *,
481 struct sock *, __u16, struct inet_timewait_sock **),
482 int (*hash)(struct sock *sk, struct inet_timewait_sock *twp))
484 struct inet_hashinfo *hinfo = death_row->hashinfo;
485 const unsigned short snum = inet_sk(sk)->inet_num;
486 struct inet_bind_hashbucket *head;
487 struct inet_bind_bucket *tb;
489 struct net *net = sock_net(sk);
493 int i, remaining, low, high, port;
495 u32 offset = hint + port_offset;
496 struct inet_timewait_sock *tw = NULL;
498 inet_get_local_port_range(net, &low, &high);
499 remaining = (high - low) + 1;
502 for (i = 1; i <= remaining; i++) {
503 port = low + (i + offset) % remaining;
504 if (inet_is_local_reserved_port(net, port))
506 head = &hinfo->bhash[inet_bhashfn(net, port,
508 spin_lock(&head->lock);
510 /* Does not bother with rcv_saddr checks,
511 * because the established check is already
514 inet_bind_bucket_for_each(tb, &head->chain) {
515 if (net_eq(ib_net(tb), net) &&
517 if (tb->fastreuse >= 0 ||
518 tb->fastreuseport >= 0)
520 WARN_ON(hlist_empty(&tb->owners));
521 if (!check_established(death_row, sk,
528 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
531 spin_unlock(&head->lock);
535 tb->fastreuseport = -1;
539 spin_unlock(&head->lock);
543 return -EADDRNOTAVAIL;
548 /* Head lock still held and bh's disabled */
549 inet_bind_hash(sk, tb, port);
550 if (sk_unhashed(sk)) {
551 inet_sk(sk)->inet_sport = htons(port);
552 twrefcnt += hash(sk, tw);
555 twrefcnt += inet_twsk_bind_unhash(tw, hinfo);
556 spin_unlock(&head->lock);
559 inet_twsk_deschedule(tw, death_row);
570 head = &hinfo->bhash[inet_bhashfn(net, snum, hinfo->bhash_size)];
571 tb = inet_csk(sk)->icsk_bind_hash;
572 spin_lock_bh(&head->lock);
573 if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
575 spin_unlock_bh(&head->lock);
578 spin_unlock(&head->lock);
579 /* No definite answer... Walk to established hash table */
580 ret = check_established(death_row, sk, snum, NULL);
588 * Bind a port for a connect operation and hash it.
590 int inet_hash_connect(struct inet_timewait_death_row *death_row,
593 return __inet_hash_connect(death_row, sk, inet_sk_port_offset(sk),
594 __inet_check_established, __inet_hash_nolisten);
596 EXPORT_SYMBOL_GPL(inet_hash_connect);
598 void inet_hashinfo_init(struct inet_hashinfo *h)
602 atomic_set(&h->bsockets, 0);
603 for (i = 0; i < INET_LHTABLE_SIZE; i++) {
604 spin_lock_init(&h->listening_hash[i].lock);
605 INIT_HLIST_NULLS_HEAD(&h->listening_hash[i].head,
606 i + LISTENING_NULLS_BASE);
609 EXPORT_SYMBOL_GPL(inet_hashinfo_init);