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 static 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;
76 INIT_HLIST_HEAD(&tb->owners);
77 hlist_add_head(&tb->node, &head->chain);
83 * Caller must hold hashbucket lock for this tb with local BH disabled
85 void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
87 if (hlist_empty(&tb->owners)) {
88 __hlist_del(&tb->node);
89 kmem_cache_free(cachep, tb);
93 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
94 const unsigned short snum)
96 inet_sk(sk)->inet_num = snum;
97 sk_add_bind_node(sk, &tb->owners);
98 inet_csk(sk)->icsk_bind_hash = tb;
102 * Get rid of any references to a local port held by the given sock.
104 static void __inet_put_port(struct sock *sk)
106 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
107 const int bhash = inet_bhashfn(sock_net(sk), inet_sk(sk)->inet_num,
108 hashinfo->bhash_size);
109 struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash];
110 struct inet_bind_bucket *tb;
112 spin_lock(&head->lock);
113 tb = inet_csk(sk)->icsk_bind_hash;
114 __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);
163 inet_csk_update_fastreuse(tb, child);
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,
174 const int dif, const int sdif, bool exact_dif)
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 || exact_dif) {
189 bool dev_match = (sk->sk_bound_dev_if == dif ||
190 sk->sk_bound_dev_if == sdif);
194 if (sk->sk_bound_dev_if)
197 if (READ_ONCE(sk->sk_incoming_cpu) == raw_smp_processor_id())
204 * Here are some nice properties to exploit here. The BSD API
205 * does not allow a listening sock to specify the remote port nor the
206 * remote address for the connection. So always assume those are both
207 * wildcarded during the search since they can never be otherwise.
210 /* called with rcu_read_lock() : No refcount taken on the socket */
211 struct sock *__inet_lookup_listener(struct net *net,
212 struct inet_hashinfo *hashinfo,
213 struct sk_buff *skb, int doff,
214 const __be32 saddr, __be16 sport,
215 const __be32 daddr, const unsigned short hnum,
216 const int dif, const int sdif)
218 unsigned int hash = inet_lhashfn(net, hnum);
219 struct inet_listen_hashbucket *ilb = &hashinfo->listening_hash[hash];
220 int score, hiscore = 0, matches = 0, reuseport = 0;
221 bool exact_dif = inet_exact_dif_match(net, skb);
222 struct sock *sk, *result = NULL;
223 struct hlist_nulls_node *node;
226 sk_nulls_for_each_rcu(sk, node, &ilb->nulls_head) {
227 score = compute_score(sk, net, hnum, daddr,
228 dif, sdif, exact_dif);
229 if (score > hiscore) {
230 reuseport = sk->sk_reuseport;
232 phash = inet_ehashfn(net, daddr, hnum,
234 result = reuseport_select_sock(sk, phash,
242 } else if (score == hiscore && reuseport) {
244 if (reciprocal_scale(phash, matches) == 0)
246 phash = next_pseudo_random32(phash);
251 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
253 /* All sockets share common refcount, but have different destructors */
254 void sock_gen_put(struct sock *sk)
256 if (!refcount_dec_and_test(&sk->sk_refcnt))
259 if (sk->sk_state == TCP_TIME_WAIT)
260 inet_twsk_free(inet_twsk(sk));
261 else if (sk->sk_state == TCP_NEW_SYN_RECV)
262 reqsk_free(inet_reqsk(sk));
266 EXPORT_SYMBOL_GPL(sock_gen_put);
268 void sock_edemux(struct sk_buff *skb)
270 sock_gen_put(skb->sk);
272 EXPORT_SYMBOL(sock_edemux);
274 struct sock *__inet_lookup_established(struct net *net,
275 struct inet_hashinfo *hashinfo,
276 const __be32 saddr, const __be16 sport,
277 const __be32 daddr, const u16 hnum,
278 const int dif, const int sdif)
280 INET_ADDR_COOKIE(acookie, saddr, daddr);
281 const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
283 const struct hlist_nulls_node *node;
284 /* Optimize here for direct hit, only listening connections can
285 * have wildcards anyways.
287 unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
288 unsigned int slot = hash & hashinfo->ehash_mask;
289 struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
292 sk_nulls_for_each_rcu(sk, node, &head->chain) {
293 if (sk->sk_hash != hash)
295 if (likely(INET_MATCH(sk, net, acookie,
296 saddr, daddr, ports, dif, sdif))) {
297 if (unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
299 if (unlikely(!INET_MATCH(sk, net, acookie,
309 * if the nulls value we got at the end of this lookup is
310 * not the expected one, we must restart lookup.
311 * We probably met an item that was moved to another chain.
313 if (get_nulls_value(node) != slot)
320 EXPORT_SYMBOL_GPL(__inet_lookup_established);
322 /* called with local bh disabled */
323 static int __inet_check_established(struct inet_timewait_death_row *death_row,
324 struct sock *sk, __u16 lport,
325 struct inet_timewait_sock **twp)
327 struct inet_hashinfo *hinfo = death_row->hashinfo;
328 struct inet_sock *inet = inet_sk(sk);
329 __be32 daddr = inet->inet_rcv_saddr;
330 __be32 saddr = inet->inet_daddr;
331 int dif = sk->sk_bound_dev_if;
332 struct net *net = sock_net(sk);
333 int sdif = l3mdev_master_ifindex_by_index(net, dif);
334 INET_ADDR_COOKIE(acookie, saddr, daddr);
335 const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
336 unsigned int hash = inet_ehashfn(net, daddr, lport,
337 saddr, inet->inet_dport);
338 struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
339 spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
341 const struct hlist_nulls_node *node;
342 struct inet_timewait_sock *tw = NULL;
346 sk_nulls_for_each(sk2, node, &head->chain) {
347 if (sk2->sk_hash != hash)
350 if (likely(INET_MATCH(sk2, net, acookie,
351 saddr, daddr, ports, dif, sdif))) {
352 if (sk2->sk_state == TCP_TIME_WAIT) {
354 if (twsk_unique(sk, sk2, twp))
361 /* Must record num and sport now. Otherwise we will see
362 * in hash table socket with a funny identity.
364 inet->inet_num = lport;
365 inet->inet_sport = htons(lport);
367 WARN_ON(!sk_unhashed(sk));
368 __sk_nulls_add_node_rcu(sk, &head->chain);
370 sk_nulls_del_node_init_rcu((struct sock *)tw);
371 __NET_INC_STATS(net, LINUX_MIB_TIMEWAITRECYCLED);
374 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
379 /* Silly. Should hash-dance instead... */
380 inet_twsk_deschedule_put(tw);
386 return -EADDRNOTAVAIL;
389 static u32 inet_sk_port_offset(const struct sock *sk)
391 const struct inet_sock *inet = inet_sk(sk);
393 return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
398 /* insert a socket into ehash, and eventually remove another one
399 * (The another one can be a SYN_RECV or TIMEWAIT
401 bool inet_ehash_insert(struct sock *sk, struct sock *osk)
403 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
404 struct hlist_nulls_head *list;
405 struct inet_ehash_bucket *head;
409 WARN_ON_ONCE(!sk_unhashed(sk));
411 sk->sk_hash = sk_ehashfn(sk);
412 head = inet_ehash_bucket(hashinfo, sk->sk_hash);
414 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
418 WARN_ON_ONCE(sk->sk_hash != osk->sk_hash);
419 ret = sk_nulls_del_node_init_rcu(osk);
422 __sk_nulls_add_node_rcu(sk, list);
427 bool inet_ehash_nolisten(struct sock *sk, struct sock *osk)
429 bool ok = inet_ehash_insert(sk, osk);
432 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
434 percpu_counter_inc(sk->sk_prot->orphan_count);
435 sk->sk_state = TCP_CLOSE;
436 sock_set_flag(sk, SOCK_DEAD);
437 inet_csk_destroy_sock(sk);
441 EXPORT_SYMBOL_GPL(inet_ehash_nolisten);
443 static int inet_reuseport_add_sock(struct sock *sk,
444 struct inet_listen_hashbucket *ilb)
446 struct inet_bind_bucket *tb = inet_csk(sk)->icsk_bind_hash;
447 const struct hlist_nulls_node *node;
449 kuid_t uid = sock_i_uid(sk);
451 sk_nulls_for_each_rcu(sk2, node, &ilb->nulls_head) {
453 sk2->sk_family == sk->sk_family &&
454 ipv6_only_sock(sk2) == ipv6_only_sock(sk) &&
455 sk2->sk_bound_dev_if == sk->sk_bound_dev_if &&
456 inet_csk(sk2)->icsk_bind_hash == tb &&
457 sk2->sk_reuseport && uid_eq(uid, sock_i_uid(sk2)) &&
458 inet_rcv_saddr_equal(sk, sk2, false))
459 return reuseport_add_sock(sk, sk2);
462 return reuseport_alloc(sk);
465 int __inet_hash(struct sock *sk, struct sock *osk)
467 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
468 struct inet_listen_hashbucket *ilb;
471 if (sk->sk_state != TCP_LISTEN) {
472 inet_ehash_nolisten(sk, osk);
475 WARN_ON(!sk_unhashed(sk));
476 ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
478 spin_lock(&ilb->lock);
479 if (sk->sk_reuseport) {
480 err = inet_reuseport_add_sock(sk, ilb);
484 if (IS_ENABLED(CONFIG_IPV6) && sk->sk_reuseport &&
485 sk->sk_family == AF_INET6)
486 __sk_nulls_add_node_tail_rcu(sk, &ilb->nulls_head);
488 __sk_nulls_add_node_rcu(sk, &ilb->nulls_head);
489 sock_set_flag(sk, SOCK_RCU_FREE);
490 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
492 spin_unlock(&ilb->lock);
496 EXPORT_SYMBOL(__inet_hash);
498 int inet_hash(struct sock *sk)
502 if (sk->sk_state != TCP_CLOSE) {
504 err = __inet_hash(sk, NULL);
510 EXPORT_SYMBOL_GPL(inet_hash);
512 void inet_unhash(struct sock *sk)
514 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
516 bool listener = false;
522 if (sk->sk_state == TCP_LISTEN) {
523 lock = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)].lock;
526 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
529 if (rcu_access_pointer(sk->sk_reuseport_cb))
530 reuseport_detach_sock(sk);
531 done = __sk_nulls_del_node_init_rcu(sk);
533 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
534 spin_unlock_bh(lock);
536 EXPORT_SYMBOL_GPL(inet_unhash);
538 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
539 struct sock *sk, u32 port_offset,
540 int (*check_established)(struct inet_timewait_death_row *,
541 struct sock *, __u16, struct inet_timewait_sock **))
543 struct inet_hashinfo *hinfo = death_row->hashinfo;
544 struct inet_timewait_sock *tw = NULL;
545 struct inet_bind_hashbucket *head;
546 int port = inet_sk(sk)->inet_num;
547 struct net *net = sock_net(sk);
548 struct inet_bind_bucket *tb;
549 u32 remaining, offset;
550 int ret, i, low, high;
554 head = &hinfo->bhash[inet_bhashfn(net, port,
556 tb = inet_csk(sk)->icsk_bind_hash;
557 spin_lock_bh(&head->lock);
558 if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
559 inet_ehash_nolisten(sk, NULL);
560 spin_unlock_bh(&head->lock);
563 spin_unlock(&head->lock);
564 /* No definite answer... Walk to established hash table */
565 ret = check_established(death_row, sk, port, NULL);
570 inet_get_local_port_range(net, &low, &high);
571 high++; /* [32768, 60999] -> [32768, 61000[ */
572 remaining = high - low;
573 if (likely(remaining > 1))
576 offset = (hint + port_offset) % remaining;
577 /* In first pass we try ports of @low parity.
578 * inet_csk_get_port() does the opposite choice.
583 for (i = 0; i < remaining; i += 2, port += 2) {
584 if (unlikely(port >= high))
586 if (inet_is_local_reserved_port(net, port))
588 head = &hinfo->bhash[inet_bhashfn(net, port,
590 spin_lock_bh(&head->lock);
592 /* Does not bother with rcv_saddr checks, because
593 * the established check is already unique enough.
595 inet_bind_bucket_for_each(tb, &head->chain) {
596 if (net_eq(ib_net(tb), net) && tb->port == port) {
597 if (tb->fastreuse >= 0 ||
598 tb->fastreuseport >= 0)
600 WARN_ON(hlist_empty(&tb->owners));
601 if (!check_established(death_row, sk,
608 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
611 spin_unlock_bh(&head->lock);
615 tb->fastreuseport = -1;
618 spin_unlock_bh(&head->lock);
623 if ((offset & 1) && remaining > 1)
624 goto other_parity_scan;
626 return -EADDRNOTAVAIL;
631 /* Head lock still held and bh's disabled */
632 inet_bind_hash(sk, tb, port);
633 if (sk_unhashed(sk)) {
634 inet_sk(sk)->inet_sport = htons(port);
635 inet_ehash_nolisten(sk, (struct sock *)tw);
638 inet_twsk_bind_unhash(tw, hinfo);
639 spin_unlock(&head->lock);
641 inet_twsk_deschedule_put(tw);
647 * Bind a port for a connect operation and hash it.
649 int inet_hash_connect(struct inet_timewait_death_row *death_row,
654 if (!inet_sk(sk)->inet_num)
655 port_offset = inet_sk_port_offset(sk);
656 return __inet_hash_connect(death_row, sk, port_offset,
657 __inet_check_established);
659 EXPORT_SYMBOL_GPL(inet_hash_connect);
661 void inet_hashinfo_init(struct inet_hashinfo *h)
665 for (i = 0; i < INET_LHTABLE_SIZE; i++) {
666 spin_lock_init(&h->listening_hash[i].lock);
667 INIT_HLIST_NULLS_HEAD(&h->listening_hash[i].nulls_head,
668 i + LISTENING_NULLS_BASE);
671 EXPORT_SYMBOL_GPL(inet_hashinfo_init);
673 int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
675 unsigned int locksz = sizeof(spinlock_t);
676 unsigned int i, nblocks = 1;
679 /* allocate 2 cache lines or at least one spinlock per cpu */
680 nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U);
681 nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
683 /* no more locks than number of hash buckets */
684 nblocks = min(nblocks, hashinfo->ehash_mask + 1);
686 hashinfo->ehash_locks = kvmalloc_array(nblocks, locksz, GFP_KERNEL);
687 if (!hashinfo->ehash_locks)
690 for (i = 0; i < nblocks; i++)
691 spin_lock_init(&hashinfo->ehash_locks[i]);
693 hashinfo->ehash_locks_mask = nblocks - 1;
696 EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);