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>
21 #include <linux/vmalloc.h>
23 #include <net/inet_connection_sock.h>
24 #include <net/inet_hashtables.h>
25 #include <net/secure_seq.h>
28 static u32 inet_ehashfn(const struct net *net, const __be32 laddr,
29 const __u16 lport, const __be32 faddr,
32 static u32 inet_ehash_secret __read_mostly;
34 net_get_random_once(&inet_ehash_secret, sizeof(inet_ehash_secret));
36 return __inet_ehashfn(laddr, lport, faddr, fport,
37 inet_ehash_secret + net_hash_mix(net));
40 /* This function handles inet_sock, but also timewait and request sockets
43 u32 sk_ehashfn(const struct sock *sk)
45 #if IS_ENABLED(CONFIG_IPV6)
46 if (sk->sk_family == AF_INET6 &&
47 !ipv6_addr_v4mapped(&sk->sk_v6_daddr))
48 return inet6_ehashfn(sock_net(sk),
49 &sk->sk_v6_rcv_saddr, sk->sk_num,
50 &sk->sk_v6_daddr, sk->sk_dport);
52 return inet_ehashfn(sock_net(sk),
53 sk->sk_rcv_saddr, sk->sk_num,
54 sk->sk_daddr, sk->sk_dport);
58 * Allocate and initialize a new local port bind bucket.
59 * The bindhash mutex for snum's hash chain must be held here.
61 struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep,
63 struct inet_bind_hashbucket *head,
64 const unsigned short snum)
66 struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
69 write_pnet(&tb->ib_net, net);
72 tb->fastreuseport = 0;
74 INIT_HLIST_HEAD(&tb->owners);
75 hlist_add_head(&tb->node, &head->chain);
81 * Caller must hold hashbucket lock for this tb with local BH disabled
83 void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
85 if (hlist_empty(&tb->owners)) {
86 __hlist_del(&tb->node);
87 kmem_cache_free(cachep, tb);
91 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
92 const unsigned short snum)
94 inet_sk(sk)->inet_num = snum;
95 sk_add_bind_node(sk, &tb->owners);
97 inet_csk(sk)->icsk_bind_hash = tb;
101 * Get rid of any references to a local port held by the given sock.
103 static void __inet_put_port(struct sock *sk)
105 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
106 const int bhash = inet_bhashfn(sock_net(sk), inet_sk(sk)->inet_num,
107 hashinfo->bhash_size);
108 struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash];
109 struct inet_bind_bucket *tb;
111 spin_lock(&head->lock);
112 tb = inet_csk(sk)->icsk_bind_hash;
113 __sk_del_bind_node(sk);
115 inet_csk(sk)->icsk_bind_hash = NULL;
116 inet_sk(sk)->inet_num = 0;
117 inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
118 spin_unlock(&head->lock);
121 void inet_put_port(struct sock *sk)
127 EXPORT_SYMBOL(inet_put_port);
129 int __inet_inherit_port(const struct sock *sk, struct sock *child)
131 struct inet_hashinfo *table = sk->sk_prot->h.hashinfo;
132 unsigned short port = inet_sk(child)->inet_num;
133 const int bhash = inet_bhashfn(sock_net(sk), port,
135 struct inet_bind_hashbucket *head = &table->bhash[bhash];
136 struct inet_bind_bucket *tb;
138 spin_lock(&head->lock);
139 tb = inet_csk(sk)->icsk_bind_hash;
141 spin_unlock(&head->lock);
144 if (tb->port != port) {
145 /* NOTE: using tproxy and redirecting skbs to a proxy
146 * on a different listener port breaks the assumption
147 * that the listener socket's icsk_bind_hash is the same
148 * as that of the child socket. We have to look up or
149 * create a new bind bucket for the child here. */
150 inet_bind_bucket_for_each(tb, &head->chain) {
151 if (net_eq(ib_net(tb), sock_net(sk)) &&
156 tb = inet_bind_bucket_create(table->bind_bucket_cachep,
157 sock_net(sk), head, port);
159 spin_unlock(&head->lock);
164 inet_bind_hash(child, tb, port);
165 spin_unlock(&head->lock);
169 EXPORT_SYMBOL_GPL(__inet_inherit_port);
171 static inline int compute_score(struct sock *sk, struct net *net,
172 const unsigned short hnum, const __be32 daddr,
176 struct inet_sock *inet = inet_sk(sk);
178 if (net_eq(sock_net(sk), net) && inet->inet_num == hnum &&
179 !ipv6_only_sock(sk)) {
180 __be32 rcv_saddr = inet->inet_rcv_saddr;
181 score = sk->sk_family == PF_INET ? 2 : 1;
183 if (rcv_saddr != daddr)
187 if (sk->sk_bound_dev_if) {
188 if (sk->sk_bound_dev_if != dif)
192 if (sk->sk_incoming_cpu == raw_smp_processor_id())
199 * Don't inline this cruft. Here are some nice properties to exploit here. The
200 * BSD API does not allow a listening sock to specify the remote port nor the
201 * remote address for the connection. So always assume those are both
202 * wildcarded during the search since they can never be otherwise.
206 struct sock *__inet_lookup_listener(struct net *net,
207 struct inet_hashinfo *hashinfo,
208 const __be32 saddr, __be16 sport,
209 const __be32 daddr, const unsigned short hnum,
212 struct sock *sk, *result;
213 struct hlist_nulls_node *node;
214 unsigned int hash = inet_lhashfn(net, hnum);
215 struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash];
216 int score, hiscore, matches = 0, reuseport = 0;
223 sk_nulls_for_each_rcu(sk, node, &ilb->head) {
224 score = compute_score(sk, net, hnum, daddr, dif);
225 if (score > hiscore) {
228 reuseport = sk->sk_reuseport;
230 phash = inet_ehashfn(net, daddr, hnum,
234 } else if (score == hiscore && reuseport) {
236 if (reciprocal_scale(phash, matches) == 0)
238 phash = next_pseudo_random32(phash);
242 * if the nulls value we got at the end of this lookup is
243 * not the expected one, we must restart lookup.
244 * We probably met an item that was moved to another chain.
246 if (get_nulls_value(node) != hash + LISTENING_NULLS_BASE)
249 if (unlikely(!atomic_inc_not_zero(&result->sk_refcnt)))
251 else if (unlikely(compute_score(result, net, hnum, daddr,
260 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
262 /* All sockets share common refcount, but have different destructors */
263 void sock_gen_put(struct sock *sk)
265 if (!atomic_dec_and_test(&sk->sk_refcnt))
268 if (sk->sk_state == TCP_TIME_WAIT)
269 inet_twsk_free(inet_twsk(sk));
270 else if (sk->sk_state == TCP_NEW_SYN_RECV)
271 reqsk_free(inet_reqsk(sk));
275 EXPORT_SYMBOL_GPL(sock_gen_put);
277 void sock_edemux(struct sk_buff *skb)
279 sock_gen_put(skb->sk);
281 EXPORT_SYMBOL(sock_edemux);
283 struct sock *__inet_lookup_established(struct net *net,
284 struct inet_hashinfo *hashinfo,
285 const __be32 saddr, const __be16 sport,
286 const __be32 daddr, const u16 hnum,
289 INET_ADDR_COOKIE(acookie, saddr, daddr);
290 const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
292 const struct hlist_nulls_node *node;
293 /* Optimize here for direct hit, only listening connections can
294 * have wildcards anyways.
296 unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
297 unsigned int slot = hash & hashinfo->ehash_mask;
298 struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
302 sk_nulls_for_each_rcu(sk, node, &head->chain) {
303 if (sk->sk_hash != hash)
305 if (likely(INET_MATCH(sk, net, acookie,
306 saddr, daddr, ports, dif))) {
307 if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt)))
309 if (unlikely(!INET_MATCH(sk, net, acookie,
310 saddr, daddr, ports, dif))) {
318 * if the nulls value we got at the end of this lookup is
319 * not the expected one, we must restart lookup.
320 * We probably met an item that was moved to another chain.
322 if (get_nulls_value(node) != slot)
330 EXPORT_SYMBOL_GPL(__inet_lookup_established);
332 /* called with local bh disabled */
333 static int __inet_check_established(struct inet_timewait_death_row *death_row,
334 struct sock *sk, __u16 lport,
335 struct inet_timewait_sock **twp)
337 struct inet_hashinfo *hinfo = death_row->hashinfo;
338 struct inet_sock *inet = inet_sk(sk);
339 __be32 daddr = inet->inet_rcv_saddr;
340 __be32 saddr = inet->inet_daddr;
341 int dif = sk->sk_bound_dev_if;
342 INET_ADDR_COOKIE(acookie, saddr, daddr);
343 const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
344 struct net *net = sock_net(sk);
345 unsigned int hash = inet_ehashfn(net, daddr, lport,
346 saddr, inet->inet_dport);
347 struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
348 spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
350 const struct hlist_nulls_node *node;
351 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 sk_nulls_del_node_init_rcu((struct sock *)tw);
380 NET_INC_STATS_BH(net, LINUX_MIB_TIMEWAITRECYCLED);
383 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
388 /* Silly. Should hash-dance instead... */
389 inet_twsk_deschedule_put(tw);
395 return -EADDRNOTAVAIL;
398 static u32 inet_sk_port_offset(const struct sock *sk)
400 const struct inet_sock *inet = inet_sk(sk);
402 return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
407 /* insert a socket into ehash, and eventually remove another one
408 * (The another one can be a SYN_RECV or TIMEWAIT
410 int inet_ehash_insert(struct sock *sk, struct sock *osk)
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_ONCE(!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 != osk->sk_hash);
429 sk_nulls_del_node_init_rcu(osk);
435 void __inet_hash_nolisten(struct sock *sk, struct sock *osk)
437 inet_ehash_insert(sk, osk);
438 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
440 EXPORT_SYMBOL_GPL(__inet_hash_nolisten);
442 void __inet_hash(struct sock *sk, struct sock *osk)
444 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
445 struct inet_listen_hashbucket *ilb;
447 if (sk->sk_state != TCP_LISTEN)
448 return __inet_hash_nolisten(sk, osk);
450 WARN_ON(!sk_unhashed(sk));
451 ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
453 spin_lock(&ilb->lock);
454 __sk_nulls_add_node_rcu(sk, &ilb->head);
455 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
456 spin_unlock(&ilb->lock);
458 EXPORT_SYMBOL(__inet_hash);
460 void inet_hash(struct sock *sk)
462 if (sk->sk_state != TCP_CLOSE) {
464 __inet_hash(sk, NULL);
468 EXPORT_SYMBOL_GPL(inet_hash);
470 void inet_unhash(struct sock *sk)
472 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
479 if (sk->sk_state == TCP_LISTEN)
480 lock = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)].lock;
482 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
485 done = __sk_nulls_del_node_init_rcu(sk);
487 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
488 spin_unlock_bh(lock);
490 EXPORT_SYMBOL_GPL(inet_unhash);
492 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
493 struct sock *sk, u32 port_offset,
494 int (*check_established)(struct inet_timewait_death_row *,
495 struct sock *, __u16, struct inet_timewait_sock **))
497 struct inet_hashinfo *hinfo = death_row->hashinfo;
498 const unsigned short snum = inet_sk(sk)->inet_num;
499 struct inet_bind_hashbucket *head;
500 struct inet_bind_bucket *tb;
502 struct net *net = sock_net(sk);
505 int i, remaining, low, high, port;
507 u32 offset = hint + port_offset;
508 struct inet_timewait_sock *tw = NULL;
510 inet_get_local_port_range(net, &low, &high);
511 remaining = (high - low) + 1;
513 /* By starting with offset being an even number,
514 * we tend to leave about 50% of ports for other uses,
520 for (i = 0; i < remaining; i++) {
521 port = low + (i + offset) % remaining;
522 if (inet_is_local_reserved_port(net, port))
524 head = &hinfo->bhash[inet_bhashfn(net, port,
526 spin_lock(&head->lock);
528 /* Does not bother with rcv_saddr checks,
529 * because the established check is already
532 inet_bind_bucket_for_each(tb, &head->chain) {
533 if (net_eq(ib_net(tb), net) &&
535 if (tb->fastreuse >= 0 ||
536 tb->fastreuseport >= 0)
538 WARN_ON(hlist_empty(&tb->owners));
539 if (!check_established(death_row, sk,
546 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
549 spin_unlock(&head->lock);
553 tb->fastreuseport = -1;
557 spin_unlock(&head->lock);
561 return -EADDRNOTAVAIL;
564 hint += (i + 2) & ~1;
566 /* Head lock still held and bh's disabled */
567 inet_bind_hash(sk, tb, port);
568 if (sk_unhashed(sk)) {
569 inet_sk(sk)->inet_sport = htons(port);
570 __inet_hash_nolisten(sk, (struct sock *)tw);
573 inet_twsk_bind_unhash(tw, hinfo);
574 spin_unlock(&head->lock);
577 inet_twsk_deschedule_put(tw);
583 head = &hinfo->bhash[inet_bhashfn(net, snum, hinfo->bhash_size)];
584 tb = inet_csk(sk)->icsk_bind_hash;
585 spin_lock_bh(&head->lock);
586 if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
587 __inet_hash_nolisten(sk, NULL);
588 spin_unlock_bh(&head->lock);
591 spin_unlock(&head->lock);
592 /* No definite answer... Walk to established hash table */
593 ret = check_established(death_row, sk, snum, NULL);
601 * Bind a port for a connect operation and hash it.
603 int inet_hash_connect(struct inet_timewait_death_row *death_row,
608 if (!inet_sk(sk)->inet_num)
609 port_offset = inet_sk_port_offset(sk);
610 return __inet_hash_connect(death_row, sk, port_offset,
611 __inet_check_established);
613 EXPORT_SYMBOL_GPL(inet_hash_connect);
615 void inet_hashinfo_init(struct inet_hashinfo *h)
619 for (i = 0; i < INET_LHTABLE_SIZE; i++) {
620 spin_lock_init(&h->listening_hash[i].lock);
621 INIT_HLIST_NULLS_HEAD(&h->listening_hash[i].head,
622 i + LISTENING_NULLS_BASE);
625 EXPORT_SYMBOL_GPL(inet_hashinfo_init);
627 int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
629 unsigned int locksz = sizeof(spinlock_t);
630 unsigned int i, nblocks = 1;
633 /* allocate 2 cache lines or at least one spinlock per cpu */
634 nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U);
635 nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
637 /* no more locks than number of hash buckets */
638 nblocks = min(nblocks, hashinfo->ehash_mask + 1);
640 hashinfo->ehash_locks = kmalloc_array(nblocks, locksz,
641 GFP_KERNEL | __GFP_NOWARN);
642 if (!hashinfo->ehash_locks)
643 hashinfo->ehash_locks = vmalloc(nblocks * locksz);
645 if (!hashinfo->ehash_locks)
648 for (i = 0; i < nblocks; i++)
649 spin_lock_init(&hashinfo->ehash_locks[i]);
651 hashinfo->ehash_locks_mask = nblocks - 1;
654 EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);