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(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;
140 if (tb->port != port) {
141 /* NOTE: using tproxy and redirecting skbs to a proxy
142 * on a different listener port breaks the assumption
143 * that the listener socket's icsk_bind_hash is the same
144 * as that of the child socket. We have to look up or
145 * create a new bind bucket for the child here. */
146 inet_bind_bucket_for_each(tb, &head->chain) {
147 if (net_eq(ib_net(tb), sock_net(sk)) &&
152 tb = inet_bind_bucket_create(table->bind_bucket_cachep,
153 sock_net(sk), head, port);
155 spin_unlock(&head->lock);
160 inet_bind_hash(child, tb, port);
161 spin_unlock(&head->lock);
165 EXPORT_SYMBOL_GPL(__inet_inherit_port);
167 static inline int compute_score(struct sock *sk, struct net *net,
168 const unsigned short hnum, const __be32 daddr,
172 struct inet_sock *inet = inet_sk(sk);
174 if (net_eq(sock_net(sk), net) && inet->inet_num == hnum &&
175 !ipv6_only_sock(sk)) {
176 __be32 rcv_saddr = inet->inet_rcv_saddr;
177 score = sk->sk_family == PF_INET ? 2 : 1;
179 if (rcv_saddr != daddr)
183 if (sk->sk_bound_dev_if) {
184 if (sk->sk_bound_dev_if != dif)
193 * Don't inline this cruft. Here are some nice properties to exploit here. The
194 * BSD API does not allow a listening sock to specify the remote port nor the
195 * remote address for the connection. So always assume those are both
196 * wildcarded during the search since they can never be otherwise.
200 struct sock *__inet_lookup_listener(struct net *net,
201 struct inet_hashinfo *hashinfo,
202 const __be32 saddr, __be16 sport,
203 const __be32 daddr, const unsigned short hnum,
206 struct sock *sk, *result;
207 struct hlist_nulls_node *node;
208 unsigned int hash = inet_lhashfn(net, hnum);
209 struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash];
210 int score, hiscore, matches = 0, reuseport = 0;
217 sk_nulls_for_each_rcu(sk, node, &ilb->head) {
218 score = compute_score(sk, net, hnum, daddr, dif);
219 if (score > hiscore) {
222 reuseport = sk->sk_reuseport;
224 phash = inet_ehashfn(net, daddr, hnum,
228 } else if (score == hiscore && reuseport) {
230 if (reciprocal_scale(phash, matches) == 0)
232 phash = next_pseudo_random32(phash);
236 * if the nulls value we got at the end of this lookup is
237 * not the expected one, we must restart lookup.
238 * We probably met an item that was moved to another chain.
240 if (get_nulls_value(node) != hash + LISTENING_NULLS_BASE)
243 if (unlikely(!atomic_inc_not_zero(&result->sk_refcnt)))
245 else if (unlikely(compute_score(result, net, hnum, daddr,
254 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
256 /* All sockets share common refcount, but have different destructors */
257 void sock_gen_put(struct sock *sk)
259 if (!atomic_dec_and_test(&sk->sk_refcnt))
262 if (sk->sk_state == TCP_TIME_WAIT)
263 inet_twsk_free(inet_twsk(sk));
264 else if (sk->sk_state == TCP_NEW_SYN_RECV)
265 reqsk_free(inet_reqsk(sk));
269 EXPORT_SYMBOL_GPL(sock_gen_put);
271 void sock_edemux(struct sk_buff *skb)
273 sock_gen_put(skb->sk);
275 EXPORT_SYMBOL(sock_edemux);
277 struct sock *__inet_lookup_established(struct net *net,
278 struct inet_hashinfo *hashinfo,
279 const __be32 saddr, const __be16 sport,
280 const __be32 daddr, const u16 hnum,
283 INET_ADDR_COOKIE(acookie, saddr, daddr);
284 const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
286 const struct hlist_nulls_node *node;
287 /* Optimize here for direct hit, only listening connections can
288 * have wildcards anyways.
290 unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
291 unsigned int slot = hash & hashinfo->ehash_mask;
292 struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
296 sk_nulls_for_each_rcu(sk, node, &head->chain) {
297 if (sk->sk_hash != hash)
299 if (likely(INET_MATCH(sk, net, acookie,
300 saddr, daddr, ports, dif))) {
301 if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt)))
303 if (unlikely(!INET_MATCH(sk, net, acookie,
304 saddr, daddr, ports, dif))) {
312 * if the nulls value we got at the end of this lookup is
313 * not the expected one, we must restart lookup.
314 * We probably met an item that was moved to another chain.
316 if (get_nulls_value(node) != slot)
324 EXPORT_SYMBOL_GPL(__inet_lookup_established);
326 /* called with local bh disabled */
327 static int __inet_check_established(struct inet_timewait_death_row *death_row,
328 struct sock *sk, __u16 lport,
329 struct inet_timewait_sock **twp)
331 struct inet_hashinfo *hinfo = death_row->hashinfo;
332 struct inet_sock *inet = inet_sk(sk);
333 __be32 daddr = inet->inet_rcv_saddr;
334 __be32 saddr = inet->inet_daddr;
335 int dif = sk->sk_bound_dev_if;
336 INET_ADDR_COOKIE(acookie, saddr, daddr);
337 const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
338 struct net *net = sock_net(sk);
339 unsigned int hash = inet_ehashfn(net, daddr, lport,
340 saddr, inet->inet_dport);
341 struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
342 spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
344 const struct hlist_nulls_node *node;
345 struct inet_timewait_sock *tw = NULL;
350 sk_nulls_for_each(sk2, node, &head->chain) {
351 if (sk2->sk_hash != hash)
354 if (likely(INET_MATCH(sk2, net, acookie,
355 saddr, daddr, ports, dif))) {
356 if (sk2->sk_state == TCP_TIME_WAIT) {
358 if (twsk_unique(sk, sk2, twp))
365 /* Must record num and sport now. Otherwise we will see
366 * in hash table socket with a funny identity.
368 inet->inet_num = lport;
369 inet->inet_sport = htons(lport);
371 WARN_ON(!sk_unhashed(sk));
372 __sk_nulls_add_node_rcu(sk, &head->chain);
374 twrefcnt = inet_twsk_unhash(tw);
375 NET_INC_STATS_BH(net, LINUX_MIB_TIMEWAITRECYCLED);
380 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
385 /* Silly. Should hash-dance instead... */
386 inet_twsk_deschedule(tw);
394 return -EADDRNOTAVAIL;
397 static u32 inet_sk_port_offset(const struct sock *sk)
399 const struct inet_sock *inet = inet_sk(sk);
401 return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
406 int __inet_hash_nolisten(struct sock *sk, struct inet_timewait_sock *tw)
408 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
409 struct hlist_nulls_head *list;
410 struct inet_ehash_bucket *head;
414 WARN_ON(!sk_unhashed(sk));
416 sk->sk_hash = sk_ehashfn(sk);
417 head = inet_ehash_bucket(hashinfo, sk->sk_hash);
419 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
422 __sk_nulls_add_node_rcu(sk, list);
424 WARN_ON(sk->sk_hash != tw->tw_hash);
425 twrefcnt = inet_twsk_unhash(tw);
428 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
431 EXPORT_SYMBOL_GPL(__inet_hash_nolisten);
433 int __inet_hash(struct sock *sk, struct inet_timewait_sock *tw)
435 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
436 struct inet_listen_hashbucket *ilb;
438 if (sk->sk_state != TCP_LISTEN)
439 return __inet_hash_nolisten(sk, tw);
441 WARN_ON(!sk_unhashed(sk));
442 ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
444 spin_lock(&ilb->lock);
445 __sk_nulls_add_node_rcu(sk, &ilb->head);
446 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
447 spin_unlock(&ilb->lock);
450 EXPORT_SYMBOL(__inet_hash);
452 void inet_hash(struct sock *sk)
454 if (sk->sk_state != TCP_CLOSE) {
456 __inet_hash(sk, NULL);
460 EXPORT_SYMBOL_GPL(inet_hash);
462 void inet_unhash(struct sock *sk)
464 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
471 if (sk->sk_state == TCP_LISTEN)
472 lock = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)].lock;
474 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
477 done = __sk_nulls_del_node_init_rcu(sk);
479 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
480 spin_unlock_bh(lock);
482 EXPORT_SYMBOL_GPL(inet_unhash);
484 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
485 struct sock *sk, u32 port_offset,
486 int (*check_established)(struct inet_timewait_death_row *,
487 struct sock *, __u16, struct inet_timewait_sock **))
489 struct inet_hashinfo *hinfo = death_row->hashinfo;
490 const unsigned short snum = inet_sk(sk)->inet_num;
491 struct inet_bind_hashbucket *head;
492 struct inet_bind_bucket *tb;
494 struct net *net = sock_net(sk);
498 int i, remaining, low, high, port;
500 u32 offset = hint + port_offset;
501 struct inet_timewait_sock *tw = NULL;
503 inet_get_local_port_range(net, &low, &high);
504 remaining = (high - low) + 1;
506 /* By starting with offset being an even number,
507 * we tend to leave about 50% of ports for other uses,
513 for (i = 0; i < remaining; i++) {
514 port = low + (i + offset) % remaining;
515 if (inet_is_local_reserved_port(net, port))
517 head = &hinfo->bhash[inet_bhashfn(net, port,
519 spin_lock(&head->lock);
521 /* Does not bother with rcv_saddr checks,
522 * because the established check is already
525 inet_bind_bucket_for_each(tb, &head->chain) {
526 if (net_eq(ib_net(tb), net) &&
528 if (tb->fastreuse >= 0 ||
529 tb->fastreuseport >= 0)
531 WARN_ON(hlist_empty(&tb->owners));
532 if (!check_established(death_row, sk,
539 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
542 spin_unlock(&head->lock);
546 tb->fastreuseport = -1;
550 spin_unlock(&head->lock);
554 return -EADDRNOTAVAIL;
557 hint += (i + 2) & ~1;
559 /* Head lock still held and bh's disabled */
560 inet_bind_hash(sk, tb, port);
561 if (sk_unhashed(sk)) {
562 inet_sk(sk)->inet_sport = htons(port);
563 twrefcnt += __inet_hash_nolisten(sk, tw);
566 twrefcnt += inet_twsk_bind_unhash(tw, hinfo);
567 spin_unlock(&head->lock);
570 inet_twsk_deschedule(tw);
581 head = &hinfo->bhash[inet_bhashfn(net, snum, hinfo->bhash_size)];
582 tb = inet_csk(sk)->icsk_bind_hash;
583 spin_lock_bh(&head->lock);
584 if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
585 __inet_hash_nolisten(sk, NULL);
586 spin_unlock_bh(&head->lock);
589 spin_unlock(&head->lock);
590 /* No definite answer... Walk to established hash table */
591 ret = check_established(death_row, sk, snum, NULL);
599 * Bind a port for a connect operation and hash it.
601 int inet_hash_connect(struct inet_timewait_death_row *death_row,
606 if (!inet_sk(sk)->inet_num)
607 port_offset = inet_sk_port_offset(sk);
608 return __inet_hash_connect(death_row, sk, port_offset,
609 __inet_check_established);
611 EXPORT_SYMBOL_GPL(inet_hash_connect);
613 void inet_hashinfo_init(struct inet_hashinfo *h)
617 for (i = 0; i < INET_LHTABLE_SIZE; i++) {
618 spin_lock_init(&h->listening_hash[i].lock);
619 INIT_HLIST_NULLS_HEAD(&h->listening_hash[i].head,
620 i + LISTENING_NULLS_BASE);
623 EXPORT_SYMBOL_GPL(inet_hashinfo_init);
625 int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
627 unsigned int i, nblocks = 1;
629 if (sizeof(spinlock_t) != 0) {
630 /* allocate 2 cache lines or at least one spinlock per cpu */
631 nblocks = max_t(unsigned int,
632 2 * L1_CACHE_BYTES / sizeof(spinlock_t),
634 nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
636 /* no more locks than number of hash buckets */
637 nblocks = min(nblocks, hashinfo->ehash_mask + 1);
639 hashinfo->ehash_locks = kmalloc_array(nblocks, sizeof(spinlock_t),
640 GFP_KERNEL | __GFP_NOWARN);
641 if (!hashinfo->ehash_locks)
642 hashinfo->ehash_locks = vmalloc(nblocks * sizeof(spinlock_t));
644 if (!hashinfo->ehash_locks)
647 for (i = 0; i < nblocks; i++)
648 spin_lock_init(&hashinfo->ehash_locks[i]);
650 hashinfo->ehash_locks_mask = nblocks - 1;
653 EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);