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/addrconf.h>
24 #include <net/inet_connection_sock.h>
25 #include <net/inet_hashtables.h>
26 #include <net/secure_seq.h>
29 #include <net/sock_reuseport.h>
31 static u32 inet_ehashfn(const struct net *net, const __be32 laddr,
32 const __u16 lport, const __be32 faddr,
35 static u32 inet_ehash_secret __read_mostly;
37 net_get_random_once(&inet_ehash_secret, sizeof(inet_ehash_secret));
39 return __inet_ehashfn(laddr, lport, faddr, fport,
40 inet_ehash_secret + net_hash_mix(net));
43 /* This function handles inet_sock, but also timewait and request sockets
46 u32 sk_ehashfn(const struct sock *sk)
48 #if IS_ENABLED(CONFIG_IPV6)
49 if (sk->sk_family == AF_INET6 &&
50 !ipv6_addr_v4mapped(&sk->sk_v6_daddr))
51 return inet6_ehashfn(sock_net(sk),
52 &sk->sk_v6_rcv_saddr, sk->sk_num,
53 &sk->sk_v6_daddr, sk->sk_dport);
55 return inet_ehashfn(sock_net(sk),
56 sk->sk_rcv_saddr, sk->sk_num,
57 sk->sk_daddr, sk->sk_dport);
61 * Allocate and initialize a new local port bind bucket.
62 * The bindhash mutex for snum's hash chain must be held here.
64 struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep,
66 struct inet_bind_hashbucket *head,
67 const unsigned short snum)
69 struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
72 write_pnet(&tb->ib_net, net);
75 tb->fastreuseport = 0;
77 INIT_HLIST_HEAD(&tb->owners);
78 hlist_add_head(&tb->node, &head->chain);
84 * Caller must hold hashbucket lock for this tb with local BH disabled
86 void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
88 if (hlist_empty(&tb->owners)) {
89 __hlist_del(&tb->node);
90 kmem_cache_free(cachep, tb);
94 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
95 const unsigned short snum)
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 spin_lock(&head->lock);
115 tb = inet_csk(sk)->icsk_bind_hash;
116 __sk_del_bind_node(sk);
118 inet_csk(sk)->icsk_bind_hash = NULL;
119 inet_sk(sk)->inet_num = 0;
120 inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
121 spin_unlock(&head->lock);
124 void inet_put_port(struct sock *sk)
130 EXPORT_SYMBOL(inet_put_port);
132 int __inet_inherit_port(const struct sock *sk, struct sock *child)
134 struct inet_hashinfo *table = sk->sk_prot->h.hashinfo;
135 unsigned short port = inet_sk(child)->inet_num;
136 const int bhash = inet_bhashfn(sock_net(sk), port,
138 struct inet_bind_hashbucket *head = &table->bhash[bhash];
139 struct inet_bind_bucket *tb;
141 spin_lock(&head->lock);
142 tb = inet_csk(sk)->icsk_bind_hash;
144 spin_unlock(&head->lock);
147 if (tb->port != port) {
148 /* NOTE: using tproxy and redirecting skbs to a proxy
149 * on a different listener port breaks the assumption
150 * that the listener socket's icsk_bind_hash is the same
151 * as that of the child socket. We have to look up or
152 * create a new bind bucket for the child here. */
153 inet_bind_bucket_for_each(tb, &head->chain) {
154 if (net_eq(ib_net(tb), sock_net(sk)) &&
159 tb = inet_bind_bucket_create(table->bind_bucket_cachep,
160 sock_net(sk), head, port);
162 spin_unlock(&head->lock);
166 inet_csk_update_fastreuse(tb, child);
168 inet_bind_hash(child, tb, port);
169 spin_unlock(&head->lock);
173 EXPORT_SYMBOL_GPL(__inet_inherit_port);
175 static inline int compute_score(struct sock *sk, struct net *net,
176 const unsigned short hnum, const __be32 daddr,
177 const int dif, bool exact_dif)
180 struct inet_sock *inet = inet_sk(sk);
182 if (net_eq(sock_net(sk), net) && inet->inet_num == hnum &&
183 !ipv6_only_sock(sk)) {
184 __be32 rcv_saddr = inet->inet_rcv_saddr;
185 score = sk->sk_family == PF_INET ? 2 : 1;
187 if (rcv_saddr != daddr)
191 if (sk->sk_bound_dev_if || exact_dif) {
192 if (sk->sk_bound_dev_if != dif)
196 if (sk->sk_incoming_cpu == raw_smp_processor_id())
203 * Here are some nice properties to exploit here. The BSD API
204 * does not allow a listening sock to specify the remote port nor the
205 * remote address for the connection. So always assume those are both
206 * wildcarded during the search since they can never be otherwise.
209 /* called with rcu_read_lock() : No refcount taken on the socket */
210 struct sock *__inet_lookup_listener(struct net *net,
211 struct inet_hashinfo *hashinfo,
212 struct sk_buff *skb, int doff,
213 const __be32 saddr, __be16 sport,
214 const __be32 daddr, const unsigned short hnum,
217 unsigned int hash = inet_lhashfn(net, hnum);
218 struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash];
219 int score, hiscore = 0, matches = 0, reuseport = 0;
220 bool exact_dif = inet_exact_dif_match(net, skb);
221 struct sock *sk, *result = NULL;
222 struct hlist_nulls_node *node;
225 sk_nulls_for_each_rcu(sk, node, &ilb->nulls_head) {
226 score = compute_score(sk, net, hnum, daddr, dif, exact_dif);
227 if (score > hiscore) {
228 reuseport = sk->sk_reuseport;
230 phash = inet_ehashfn(net, daddr, hnum,
232 result = reuseport_select_sock(sk, phash,
240 } else if (score == hiscore && reuseport) {
242 if (reciprocal_scale(phash, matches) == 0)
244 phash = next_pseudo_random32(phash);
249 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
251 /* All sockets share common refcount, but have different destructors */
252 void sock_gen_put(struct sock *sk)
254 if (!atomic_dec_and_test(&sk->sk_refcnt))
257 if (sk->sk_state == TCP_TIME_WAIT)
258 inet_twsk_free(inet_twsk(sk));
259 else if (sk->sk_state == TCP_NEW_SYN_RECV)
260 reqsk_free(inet_reqsk(sk));
264 EXPORT_SYMBOL_GPL(sock_gen_put);
266 void sock_edemux(struct sk_buff *skb)
268 sock_gen_put(skb->sk);
270 EXPORT_SYMBOL(sock_edemux);
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];
290 sk_nulls_for_each_rcu(sk, node, &head->chain) {
291 if (sk->sk_hash != hash)
293 if (likely(INET_MATCH(sk, net, acookie,
294 saddr, daddr, ports, dif))) {
295 if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt)))
297 if (unlikely(!INET_MATCH(sk, net, acookie,
298 saddr, daddr, ports, dif))) {
306 * if the nulls value we got at the end of this lookup is
307 * not the expected one, we must restart lookup.
308 * We probably met an item that was moved to another chain.
310 if (get_nulls_value(node) != slot)
317 EXPORT_SYMBOL_GPL(__inet_lookup_established);
319 /* called with local bh disabled */
320 static int __inet_check_established(struct inet_timewait_death_row *death_row,
321 struct sock *sk, __u16 lport,
322 struct inet_timewait_sock **twp)
324 struct inet_hashinfo *hinfo = death_row->hashinfo;
325 struct inet_sock *inet = inet_sk(sk);
326 __be32 daddr = inet->inet_rcv_saddr;
327 __be32 saddr = inet->inet_daddr;
328 int dif = sk->sk_bound_dev_if;
329 INET_ADDR_COOKIE(acookie, saddr, daddr);
330 const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
331 struct net *net = sock_net(sk);
332 unsigned int hash = inet_ehashfn(net, daddr, lport,
333 saddr, inet->inet_dport);
334 struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
335 spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
337 const struct hlist_nulls_node *node;
338 struct inet_timewait_sock *tw = NULL;
342 sk_nulls_for_each(sk2, node, &head->chain) {
343 if (sk2->sk_hash != hash)
346 if (likely(INET_MATCH(sk2, net, acookie,
347 saddr, daddr, ports, dif))) {
348 if (sk2->sk_state == TCP_TIME_WAIT) {
350 if (twsk_unique(sk, sk2, twp))
357 /* Must record num and sport now. Otherwise we will see
358 * in hash table socket with a funny identity.
360 inet->inet_num = lport;
361 inet->inet_sport = htons(lport);
363 WARN_ON(!sk_unhashed(sk));
364 __sk_nulls_add_node_rcu(sk, &head->chain);
366 sk_nulls_del_node_init_rcu((struct sock *)tw);
367 __NET_INC_STATS(net, LINUX_MIB_TIMEWAITRECYCLED);
370 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
375 /* Silly. Should hash-dance instead... */
376 inet_twsk_deschedule_put(tw);
382 return -EADDRNOTAVAIL;
385 static u32 inet_sk_port_offset(const struct sock *sk)
387 const struct inet_sock *inet = inet_sk(sk);
389 return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
394 /* insert a socket into ehash, and eventually remove another one
395 * (The another one can be a SYN_RECV or TIMEWAIT
397 bool inet_ehash_insert(struct sock *sk, struct sock *osk)
399 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
400 struct hlist_nulls_head *list;
401 struct inet_ehash_bucket *head;
405 WARN_ON_ONCE(!sk_unhashed(sk));
407 sk->sk_hash = sk_ehashfn(sk);
408 head = inet_ehash_bucket(hashinfo, sk->sk_hash);
410 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
414 WARN_ON_ONCE(sk->sk_hash != osk->sk_hash);
415 ret = sk_nulls_del_node_init_rcu(osk);
418 __sk_nulls_add_node_rcu(sk, list);
423 bool inet_ehash_nolisten(struct sock *sk, struct sock *osk)
425 bool ok = inet_ehash_insert(sk, osk);
428 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
430 percpu_counter_inc(sk->sk_prot->orphan_count);
431 sk->sk_state = TCP_CLOSE;
432 sock_set_flag(sk, SOCK_DEAD);
433 inet_csk_destroy_sock(sk);
437 EXPORT_SYMBOL_GPL(inet_ehash_nolisten);
439 static int inet_reuseport_add_sock(struct sock *sk,
440 struct inet_listen_hashbucket *ilb,
441 int (*saddr_same)(const struct sock *sk1,
442 const struct sock *sk2,
443 bool match_wildcard))
445 struct inet_bind_bucket *tb = inet_csk(sk)->icsk_bind_hash;
446 const struct hlist_nulls_node *node;
448 kuid_t uid = sock_i_uid(sk);
450 sk_nulls_for_each_rcu(sk2, node, &ilb->nulls_head) {
452 sk2->sk_family == sk->sk_family &&
453 ipv6_only_sock(sk2) == ipv6_only_sock(sk) &&
454 sk2->sk_bound_dev_if == sk->sk_bound_dev_if &&
455 inet_csk(sk2)->icsk_bind_hash == tb &&
456 sk2->sk_reuseport && uid_eq(uid, sock_i_uid(sk2)) &&
457 saddr_same(sk, sk2, false))
458 return reuseport_add_sock(sk, sk2);
461 return reuseport_alloc(sk);
464 int __inet_hash(struct sock *sk, struct sock *osk,
465 int (*saddr_same)(const struct sock *sk1,
466 const struct sock *sk2,
467 bool match_wildcard))
469 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
470 struct inet_listen_hashbucket *ilb;
473 if (sk->sk_state != TCP_LISTEN) {
474 inet_ehash_nolisten(sk, osk);
477 WARN_ON(!sk_unhashed(sk));
478 ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
480 spin_lock(&ilb->lock);
481 if (sk->sk_reuseport) {
482 err = inet_reuseport_add_sock(sk, ilb, saddr_same);
486 if (IS_ENABLED(CONFIG_IPV6) && sk->sk_reuseport &&
487 sk->sk_family == AF_INET6)
488 __sk_nulls_add_node_tail_rcu(sk, &ilb->nulls_head);
490 __sk_nulls_add_node_rcu(sk, &ilb->nulls_head);
491 sock_set_flag(sk, SOCK_RCU_FREE);
492 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
494 spin_unlock(&ilb->lock);
498 EXPORT_SYMBOL(__inet_hash);
500 int inet_hash(struct sock *sk)
504 if (sk->sk_state != TCP_CLOSE) {
506 err = __inet_hash(sk, NULL, ipv4_rcv_saddr_equal);
512 EXPORT_SYMBOL_GPL(inet_hash);
514 void inet_unhash(struct sock *sk)
516 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
518 bool listener = false;
524 if (sk->sk_state == TCP_LISTEN) {
525 lock = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)].lock;
528 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
531 if (rcu_access_pointer(sk->sk_reuseport_cb))
532 reuseport_detach_sock(sk);
533 done = __sk_nulls_del_node_init_rcu(sk);
535 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
536 spin_unlock_bh(lock);
538 EXPORT_SYMBOL_GPL(inet_unhash);
540 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
541 struct sock *sk, u32 port_offset,
542 int (*check_established)(struct inet_timewait_death_row *,
543 struct sock *, __u16, struct inet_timewait_sock **))
545 struct inet_hashinfo *hinfo = death_row->hashinfo;
546 struct inet_timewait_sock *tw = NULL;
547 struct inet_bind_hashbucket *head;
548 int port = inet_sk(sk)->inet_num;
549 struct net *net = sock_net(sk);
550 struct inet_bind_bucket *tb;
551 u32 remaining, offset;
552 int ret, i, low, high;
556 head = &hinfo->bhash[inet_bhashfn(net, port,
558 tb = inet_csk(sk)->icsk_bind_hash;
559 spin_lock_bh(&head->lock);
560 if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
561 inet_ehash_nolisten(sk, NULL);
562 spin_unlock_bh(&head->lock);
565 spin_unlock(&head->lock);
566 /* No definite answer... Walk to established hash table */
567 ret = check_established(death_row, sk, port, NULL);
572 inet_get_local_port_range(net, &low, &high);
573 high++; /* [32768, 60999] -> [32768, 61000[ */
574 remaining = high - low;
575 if (likely(remaining > 1))
578 offset = (hint + port_offset) % remaining;
579 /* In first pass we try ports of @low parity.
580 * inet_csk_get_port() does the opposite choice.
585 for (i = 0; i < remaining; i += 2, port += 2) {
586 if (unlikely(port >= high))
588 if (inet_is_local_reserved_port(net, port))
590 head = &hinfo->bhash[inet_bhashfn(net, port,
592 spin_lock_bh(&head->lock);
594 /* Does not bother with rcv_saddr checks, because
595 * the established check is already unique enough.
597 inet_bind_bucket_for_each(tb, &head->chain) {
598 if (net_eq(ib_net(tb), net) && tb->port == port) {
599 if (tb->fastreuse >= 0 ||
600 tb->fastreuseport >= 0)
602 WARN_ON(hlist_empty(&tb->owners));
603 if (!check_established(death_row, sk,
610 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
613 spin_unlock_bh(&head->lock);
617 tb->fastreuseport = -1;
620 spin_unlock_bh(&head->lock);
625 if ((offset & 1) && remaining > 1)
626 goto other_parity_scan;
628 return -EADDRNOTAVAIL;
633 /* Head lock still held and bh's disabled */
634 inet_bind_hash(sk, tb, port);
635 if (sk_unhashed(sk)) {
636 inet_sk(sk)->inet_sport = htons(port);
637 inet_ehash_nolisten(sk, (struct sock *)tw);
640 inet_twsk_bind_unhash(tw, hinfo);
641 spin_unlock(&head->lock);
643 inet_twsk_deschedule_put(tw);
649 * Bind a port for a connect operation and hash it.
651 int inet_hash_connect(struct inet_timewait_death_row *death_row,
656 if (!inet_sk(sk)->inet_num)
657 port_offset = inet_sk_port_offset(sk);
658 return __inet_hash_connect(death_row, sk, port_offset,
659 __inet_check_established);
661 EXPORT_SYMBOL_GPL(inet_hash_connect);
663 void inet_hashinfo_init(struct inet_hashinfo *h)
667 for (i = 0; i < INET_LHTABLE_SIZE; i++) {
668 spin_lock_init(&h->listening_hash[i].lock);
669 INIT_HLIST_NULLS_HEAD(&h->listening_hash[i].nulls_head,
670 i + LISTENING_NULLS_BASE);
673 EXPORT_SYMBOL_GPL(inet_hashinfo_init);
675 int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
677 unsigned int locksz = sizeof(spinlock_t);
678 unsigned int i, nblocks = 1;
681 /* allocate 2 cache lines or at least one spinlock per cpu */
682 nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U);
683 nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
685 /* no more locks than number of hash buckets */
686 nblocks = min(nblocks, hashinfo->ehash_mask + 1);
688 hashinfo->ehash_locks = kmalloc_array(nblocks, locksz,
689 GFP_KERNEL | __GFP_NOWARN);
690 if (!hashinfo->ehash_locks)
691 hashinfo->ehash_locks = vmalloc(nblocks * locksz);
693 if (!hashinfo->ehash_locks)
696 for (i = 0; i < nblocks; i++)
697 spin_lock_init(&hashinfo->ehash_locks[i]);
699 hashinfo->ehash_locks_mask = nblocks - 1;
702 EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);