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 u32 inet_ehashfn(const 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));
39 /* This function handles inet_sock, but also timewait and request sockets
42 u32 sk_ehashfn(const struct sock *sk)
44 #if IS_ENABLED(CONFIG_IPV6)
45 if (sk->sk_family == AF_INET6 &&
46 !ipv6_addr_v4mapped(&sk->sk_v6_daddr))
47 return inet6_ehashfn(sock_net(sk),
48 &sk->sk_v6_rcv_saddr, sk->sk_num,
49 &sk->sk_v6_daddr, sk->sk_dport);
51 return inet_ehashfn(sock_net(sk),
52 sk->sk_rcv_saddr, sk->sk_num,
53 sk->sk_daddr, sk->sk_dport);
57 * Allocate and initialize a new local port bind bucket.
58 * The bindhash mutex for snum's hash chain must be held here.
60 struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep,
62 struct inet_bind_hashbucket *head,
63 const unsigned short snum)
65 struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
68 write_pnet(&tb->ib_net, net);
71 tb->fastreuseport = 0;
73 INIT_HLIST_HEAD(&tb->owners);
74 hlist_add_head(&tb->node, &head->chain);
80 * Caller must hold hashbucket lock for this tb with local BH disabled
82 void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
84 if (hlist_empty(&tb->owners)) {
85 __hlist_del(&tb->node);
86 kmem_cache_free(cachep, tb);
90 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
91 const unsigned short snum)
93 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
95 atomic_inc(&hashinfo->bsockets);
97 inet_sk(sk)->inet_num = snum;
98 sk_add_bind_node(sk, &tb->owners);
100 inet_csk(sk)->icsk_bind_hash = tb;
104 * Get rid of any references to a local port held by the given sock.
106 static void __inet_put_port(struct sock *sk)
108 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
109 const int bhash = inet_bhashfn(sock_net(sk), inet_sk(sk)->inet_num,
110 hashinfo->bhash_size);
111 struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash];
112 struct inet_bind_bucket *tb;
114 atomic_dec(&hashinfo->bsockets);
116 spin_lock(&head->lock);
117 tb = inet_csk(sk)->icsk_bind_hash;
118 __sk_del_bind_node(sk);
120 inet_csk(sk)->icsk_bind_hash = NULL;
121 inet_sk(sk)->inet_num = 0;
122 inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
123 spin_unlock(&head->lock);
126 void inet_put_port(struct sock *sk)
132 EXPORT_SYMBOL(inet_put_port);
134 int __inet_inherit_port(struct sock *sk, struct sock *child)
136 struct inet_hashinfo *table = sk->sk_prot->h.hashinfo;
137 unsigned short port = inet_sk(child)->inet_num;
138 const int bhash = inet_bhashfn(sock_net(sk), port,
140 struct inet_bind_hashbucket *head = &table->bhash[bhash];
141 struct inet_bind_bucket *tb;
143 spin_lock(&head->lock);
144 tb = inet_csk(sk)->icsk_bind_hash;
145 if (tb->port != port) {
146 /* NOTE: using tproxy and redirecting skbs to a proxy
147 * on a different listener port breaks the assumption
148 * that the listener socket's icsk_bind_hash is the same
149 * as that of the child socket. We have to look up or
150 * create a new bind bucket for the child here. */
151 inet_bind_bucket_for_each(tb, &head->chain) {
152 if (net_eq(ib_net(tb), sock_net(sk)) &&
157 tb = inet_bind_bucket_create(table->bind_bucket_cachep,
158 sock_net(sk), head, port);
160 spin_unlock(&head->lock);
165 inet_bind_hash(child, tb, port);
166 spin_unlock(&head->lock);
170 EXPORT_SYMBOL_GPL(__inet_inherit_port);
172 static inline int compute_score(struct sock *sk, struct net *net,
173 const unsigned short hnum, const __be32 daddr,
177 struct inet_sock *inet = inet_sk(sk);
179 if (net_eq(sock_net(sk), net) && inet->inet_num == hnum &&
180 !ipv6_only_sock(sk)) {
181 __be32 rcv_saddr = inet->inet_rcv_saddr;
182 score = sk->sk_family == PF_INET ? 2 : 1;
184 if (rcv_saddr != daddr)
188 if (sk->sk_bound_dev_if) {
189 if (sk->sk_bound_dev_if != dif)
198 * Don't inline this cruft. Here are some nice properties to exploit here. The
199 * BSD API does not allow a listening sock to specify the remote port nor the
200 * remote address for the connection. So always assume those are both
201 * wildcarded during the search since they can never be otherwise.
205 struct sock *__inet_lookup_listener(struct net *net,
206 struct inet_hashinfo *hashinfo,
207 const __be32 saddr, __be16 sport,
208 const __be32 daddr, const unsigned short hnum,
211 struct sock *sk, *result;
212 struct hlist_nulls_node *node;
213 unsigned int hash = inet_lhashfn(net, hnum);
214 struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash];
215 int score, hiscore, matches = 0, reuseport = 0;
222 sk_nulls_for_each_rcu(sk, node, &ilb->head) {
223 score = compute_score(sk, net, hnum, daddr, dif);
224 if (score > hiscore) {
227 reuseport = sk->sk_reuseport;
229 phash = inet_ehashfn(net, daddr, hnum,
233 } else if (score == hiscore && reuseport) {
235 if (reciprocal_scale(phash, matches) == 0)
237 phash = next_pseudo_random32(phash);
241 * if the nulls value we got at the end of this lookup is
242 * not the expected one, we must restart lookup.
243 * We probably met an item that was moved to another chain.
245 if (get_nulls_value(node) != hash + LISTENING_NULLS_BASE)
248 if (unlikely(!atomic_inc_not_zero(&result->sk_refcnt)))
250 else if (unlikely(compute_score(result, net, hnum, daddr,
259 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
261 /* All sockets share common refcount, but have different destructors */
262 void sock_gen_put(struct sock *sk)
264 if (!atomic_dec_and_test(&sk->sk_refcnt))
267 if (sk->sk_state == TCP_TIME_WAIT)
268 inet_twsk_free(inet_twsk(sk));
269 else if (sk->sk_state == TCP_NEW_SYN_RECV)
270 reqsk_free(inet_reqsk(sk));
274 EXPORT_SYMBOL_GPL(sock_gen_put);
276 void sock_edemux(struct sk_buff *skb)
278 sock_gen_put(skb->sk);
280 EXPORT_SYMBOL(sock_edemux);
282 struct sock *__inet_lookup_established(struct net *net,
283 struct inet_hashinfo *hashinfo,
284 const __be32 saddr, const __be16 sport,
285 const __be32 daddr, const u16 hnum,
288 INET_ADDR_COOKIE(acookie, saddr, daddr);
289 const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
291 const struct hlist_nulls_node *node;
292 /* Optimize here for direct hit, only listening connections can
293 * have wildcards anyways.
295 unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
296 unsigned int slot = hash & hashinfo->ehash_mask;
297 struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
301 sk_nulls_for_each_rcu(sk, node, &head->chain) {
302 if (sk->sk_hash != hash)
304 if (likely(INET_MATCH(sk, net, acookie,
305 saddr, daddr, ports, dif))) {
306 if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt)))
308 if (unlikely(!INET_MATCH(sk, net, acookie,
309 saddr, daddr, ports, dif))) {
317 * if the nulls value we got at the end of this lookup is
318 * not the expected one, we must restart lookup.
319 * We probably met an item that was moved to another chain.
321 if (get_nulls_value(node) != slot)
329 EXPORT_SYMBOL_GPL(__inet_lookup_established);
331 /* called with local bh disabled */
332 static int __inet_check_established(struct inet_timewait_death_row *death_row,
333 struct sock *sk, __u16 lport,
334 struct inet_timewait_sock **twp)
336 struct inet_hashinfo *hinfo = death_row->hashinfo;
337 struct inet_sock *inet = inet_sk(sk);
338 __be32 daddr = inet->inet_rcv_saddr;
339 __be32 saddr = inet->inet_daddr;
340 int dif = sk->sk_bound_dev_if;
341 INET_ADDR_COOKIE(acookie, saddr, daddr);
342 const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
343 struct net *net = sock_net(sk);
344 unsigned int hash = inet_ehashfn(net, daddr, lport,
345 saddr, inet->inet_dport);
346 struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
347 spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
349 const struct hlist_nulls_node *node;
350 struct inet_timewait_sock *tw = NULL;
355 sk_nulls_for_each(sk2, node, &head->chain) {
356 if (sk2->sk_hash != hash)
359 if (likely(INET_MATCH(sk2, net, acookie,
360 saddr, daddr, ports, dif))) {
361 if (sk2->sk_state == TCP_TIME_WAIT) {
363 if (twsk_unique(sk, sk2, twp))
370 /* Must record num and sport now. Otherwise we will see
371 * in hash table socket with a funny identity.
373 inet->inet_num = lport;
374 inet->inet_sport = htons(lport);
376 WARN_ON(!sk_unhashed(sk));
377 __sk_nulls_add_node_rcu(sk, &head->chain);
379 twrefcnt = inet_twsk_unhash(tw);
380 NET_INC_STATS_BH(net, LINUX_MIB_TIMEWAITRECYCLED);
385 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
390 /* Silly. Should hash-dance instead... */
391 inet_twsk_deschedule(tw, death_row);
399 return -EADDRNOTAVAIL;
402 static inline u32 inet_sk_port_offset(const struct sock *sk)
404 const struct inet_sock *inet = inet_sk(sk);
405 return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
410 int __inet_hash_nolisten(struct sock *sk, struct inet_timewait_sock *tw)
412 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
413 struct hlist_nulls_head *list;
414 struct inet_ehash_bucket *head;
418 WARN_ON(!sk_unhashed(sk));
420 sk->sk_hash = sk_ehashfn(sk);
421 head = inet_ehash_bucket(hashinfo, sk->sk_hash);
423 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
426 __sk_nulls_add_node_rcu(sk, list);
428 WARN_ON(sk->sk_hash != tw->tw_hash);
429 twrefcnt = inet_twsk_unhash(tw);
432 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
435 EXPORT_SYMBOL_GPL(__inet_hash_nolisten);
437 int __inet_hash(struct sock *sk, struct inet_timewait_sock *tw)
439 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
440 struct inet_listen_hashbucket *ilb;
442 if (sk->sk_state != TCP_LISTEN)
443 return __inet_hash_nolisten(sk, tw);
445 WARN_ON(!sk_unhashed(sk));
446 ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
448 spin_lock(&ilb->lock);
449 __sk_nulls_add_node_rcu(sk, &ilb->head);
450 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
451 spin_unlock(&ilb->lock);
454 EXPORT_SYMBOL(__inet_hash);
456 void inet_hash(struct sock *sk)
458 if (sk->sk_state != TCP_CLOSE) {
460 __inet_hash(sk, NULL);
464 EXPORT_SYMBOL_GPL(inet_hash);
466 void inet_unhash(struct sock *sk)
468 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
475 if (sk->sk_state == TCP_LISTEN)
476 lock = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)].lock;
478 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
481 done = __sk_nulls_del_node_init_rcu(sk);
483 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
484 spin_unlock_bh(lock);
486 EXPORT_SYMBOL_GPL(inet_unhash);
488 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
489 struct sock *sk, u32 port_offset,
490 int (*check_established)(struct inet_timewait_death_row *,
491 struct sock *, __u16, struct inet_timewait_sock **),
492 int (*hash)(struct sock *sk, struct inet_timewait_sock *twp))
494 struct inet_hashinfo *hinfo = death_row->hashinfo;
495 const unsigned short snum = inet_sk(sk)->inet_num;
496 struct inet_bind_hashbucket *head;
497 struct inet_bind_bucket *tb;
499 struct net *net = sock_net(sk);
503 int i, remaining, low, high, port;
505 u32 offset = hint + port_offset;
506 struct inet_timewait_sock *tw = NULL;
508 inet_get_local_port_range(net, &low, &high);
509 remaining = (high - low) + 1;
512 for (i = 1; i <= remaining; i++) {
513 port = low + (i + offset) % remaining;
514 if (inet_is_local_reserved_port(net, port))
516 head = &hinfo->bhash[inet_bhashfn(net, port,
518 spin_lock(&head->lock);
520 /* Does not bother with rcv_saddr checks,
521 * because the established check is already
524 inet_bind_bucket_for_each(tb, &head->chain) {
525 if (net_eq(ib_net(tb), net) &&
527 if (tb->fastreuse >= 0 ||
528 tb->fastreuseport >= 0)
530 WARN_ON(hlist_empty(&tb->owners));
531 if (!check_established(death_row, sk,
538 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
541 spin_unlock(&head->lock);
545 tb->fastreuseport = -1;
549 spin_unlock(&head->lock);
553 return -EADDRNOTAVAIL;
558 /* Head lock still held and bh's disabled */
559 inet_bind_hash(sk, tb, port);
560 if (sk_unhashed(sk)) {
561 inet_sk(sk)->inet_sport = htons(port);
562 twrefcnt += hash(sk, tw);
565 twrefcnt += inet_twsk_bind_unhash(tw, hinfo);
566 spin_unlock(&head->lock);
569 inet_twsk_deschedule(tw, death_row);
580 head = &hinfo->bhash[inet_bhashfn(net, snum, hinfo->bhash_size)];
581 tb = inet_csk(sk)->icsk_bind_hash;
582 spin_lock_bh(&head->lock);
583 if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
585 spin_unlock_bh(&head->lock);
588 spin_unlock(&head->lock);
589 /* No definite answer... Walk to established hash table */
590 ret = check_established(death_row, sk, snum, NULL);
598 * Bind a port for a connect operation and hash it.
600 int inet_hash_connect(struct inet_timewait_death_row *death_row,
603 return __inet_hash_connect(death_row, sk, inet_sk_port_offset(sk),
604 __inet_check_established, __inet_hash_nolisten);
606 EXPORT_SYMBOL_GPL(inet_hash_connect);
608 void inet_hashinfo_init(struct inet_hashinfo *h)
612 atomic_set(&h->bsockets, 0);
613 for (i = 0; i < INET_LHTABLE_SIZE; i++) {
614 spin_lock_init(&h->listening_hash[i].lock);
615 INIT_HLIST_NULLS_HEAD(&h->listening_hash[i].head,
616 i + LISTENING_NULLS_BASE);
619 EXPORT_SYMBOL_GPL(inet_hashinfo_init);