1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * Generic INET transport hashtables
9 * Authors: Lotsa people, from code originally in tcp
12 #include <linux/module.h>
13 #include <linux/random.h>
14 #include <linux/sched.h>
15 #include <linux/slab.h>
16 #include <linux/wait.h>
17 #include <linux/vmalloc.h>
18 #include <linux/memblock.h>
20 #include <net/addrconf.h>
21 #include <net/inet_connection_sock.h>
22 #include <net/inet_hashtables.h>
23 #include <net/secure_seq.h>
26 #include <net/sock_reuseport.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 static 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,
67 struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
70 write_pnet(&tb->ib_net, net);
74 tb->fastreuseport = 0;
75 INIT_HLIST_HEAD(&tb->owners);
76 hlist_add_head(&tb->node, &head->chain);
82 * Caller must hold hashbucket lock for this tb with local BH disabled
84 void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
86 if (hlist_empty(&tb->owners)) {
87 __hlist_del(&tb->node);
88 kmem_cache_free(cachep, tb);
92 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
93 const unsigned short snum)
95 inet_sk(sk)->inet_num = snum;
96 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);
114 inet_csk(sk)->icsk_bind_hash = NULL;
115 inet_sk(sk)->inet_num = 0;
116 inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
117 spin_unlock(&head->lock);
120 void inet_put_port(struct sock *sk)
126 EXPORT_SYMBOL(inet_put_port);
128 int __inet_inherit_port(const struct sock *sk, struct sock *child)
130 struct inet_hashinfo *table = sk->sk_prot->h.hashinfo;
131 unsigned short port = inet_sk(child)->inet_num;
132 const int bhash = inet_bhashfn(sock_net(sk), port,
134 struct inet_bind_hashbucket *head = &table->bhash[bhash];
135 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 l3mdev = inet_sk_bound_l3mdev(sk);
147 /* NOTE: using tproxy and redirecting skbs to a proxy
148 * on a different listener port breaks the assumption
149 * that the listener socket's icsk_bind_hash is the same
150 * as that of the child socket. We have to look up or
151 * create a new bind bucket for the child here. */
152 inet_bind_bucket_for_each(tb, &head->chain) {
153 if (net_eq(ib_net(tb), sock_net(sk)) &&
154 tb->l3mdev == l3mdev && tb->port == port)
158 tb = inet_bind_bucket_create(table->bind_bucket_cachep,
159 sock_net(sk), head, port,
162 spin_unlock(&head->lock);
167 inet_bind_hash(child, tb, port);
168 spin_unlock(&head->lock);
172 EXPORT_SYMBOL_GPL(__inet_inherit_port);
174 static struct inet_listen_hashbucket *
175 inet_lhash2_bucket_sk(struct inet_hashinfo *h, struct sock *sk)
179 #if IS_ENABLED(CONFIG_IPV6)
180 if (sk->sk_family == AF_INET6)
181 hash = ipv6_portaddr_hash(sock_net(sk),
182 &sk->sk_v6_rcv_saddr,
183 inet_sk(sk)->inet_num);
186 hash = ipv4_portaddr_hash(sock_net(sk),
187 inet_sk(sk)->inet_rcv_saddr,
188 inet_sk(sk)->inet_num);
189 return inet_lhash2_bucket(h, hash);
192 static void inet_hash2(struct inet_hashinfo *h, struct sock *sk)
194 struct inet_listen_hashbucket *ilb2;
199 ilb2 = inet_lhash2_bucket_sk(h, sk);
201 spin_lock(&ilb2->lock);
202 if (sk->sk_reuseport && sk->sk_family == AF_INET6)
203 hlist_add_tail_rcu(&inet_csk(sk)->icsk_listen_portaddr_node,
206 hlist_add_head_rcu(&inet_csk(sk)->icsk_listen_portaddr_node,
209 spin_unlock(&ilb2->lock);
212 static void inet_unhash2(struct inet_hashinfo *h, struct sock *sk)
214 struct inet_listen_hashbucket *ilb2;
217 WARN_ON_ONCE(hlist_unhashed(&inet_csk(sk)->icsk_listen_portaddr_node)))
220 ilb2 = inet_lhash2_bucket_sk(h, sk);
222 spin_lock(&ilb2->lock);
223 hlist_del_init_rcu(&inet_csk(sk)->icsk_listen_portaddr_node);
225 spin_unlock(&ilb2->lock);
228 static inline int compute_score(struct sock *sk, struct net *net,
229 const unsigned short hnum, const __be32 daddr,
230 const int dif, const int sdif, bool exact_dif)
234 if (net_eq(sock_net(sk), net) && sk->sk_num == hnum &&
235 !ipv6_only_sock(sk)) {
236 if (sk->sk_rcv_saddr != daddr)
239 if (!inet_sk_bound_dev_eq(net, sk->sk_bound_dev_if, dif, sdif))
242 score = sk->sk_family == PF_INET ? 2 : 1;
243 if (sk->sk_incoming_cpu == raw_smp_processor_id())
250 * Here are some nice properties to exploit here. The BSD API
251 * does not allow a listening sock to specify the remote port nor the
252 * remote address for the connection. So always assume those are both
253 * wildcarded during the search since they can never be otherwise.
256 /* called with rcu_read_lock() : No refcount taken on the socket */
257 static struct sock *inet_lhash2_lookup(struct net *net,
258 struct inet_listen_hashbucket *ilb2,
259 struct sk_buff *skb, int doff,
260 const __be32 saddr, __be16 sport,
261 const __be32 daddr, const unsigned short hnum,
262 const int dif, const int sdif)
264 bool exact_dif = inet_exact_dif_match(net, skb);
265 struct inet_connection_sock *icsk;
266 struct sock *sk, *result = NULL;
267 int score, hiscore = 0;
270 inet_lhash2_for_each_icsk_rcu(icsk, &ilb2->head) {
271 sk = (struct sock *)icsk;
272 score = compute_score(sk, net, hnum, daddr,
273 dif, sdif, exact_dif);
274 if (score > hiscore) {
275 if (sk->sk_reuseport) {
276 phash = inet_ehashfn(net, daddr, hnum,
278 result = reuseport_select_sock(sk, phash,
291 struct sock *__inet_lookup_listener(struct net *net,
292 struct inet_hashinfo *hashinfo,
293 struct sk_buff *skb, int doff,
294 const __be32 saddr, __be16 sport,
295 const __be32 daddr, const unsigned short hnum,
296 const int dif, const int sdif)
298 struct inet_listen_hashbucket *ilb2;
299 struct sock *result = NULL;
302 hash2 = ipv4_portaddr_hash(net, daddr, hnum);
303 ilb2 = inet_lhash2_bucket(hashinfo, hash2);
305 result = inet_lhash2_lookup(net, ilb2, skb, doff,
306 saddr, sport, daddr, hnum,
311 /* Lookup lhash2 with INADDR_ANY */
312 hash2 = ipv4_portaddr_hash(net, htonl(INADDR_ANY), hnum);
313 ilb2 = inet_lhash2_bucket(hashinfo, hash2);
315 result = inet_lhash2_lookup(net, ilb2, skb, doff,
316 saddr, sport, htonl(INADDR_ANY), hnum,
323 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
325 /* All sockets share common refcount, but have different destructors */
326 void sock_gen_put(struct sock *sk)
328 if (!refcount_dec_and_test(&sk->sk_refcnt))
331 if (sk->sk_state == TCP_TIME_WAIT)
332 inet_twsk_free(inet_twsk(sk));
333 else if (sk->sk_state == TCP_NEW_SYN_RECV)
334 reqsk_free(inet_reqsk(sk));
338 EXPORT_SYMBOL_GPL(sock_gen_put);
340 void sock_edemux(struct sk_buff *skb)
342 sock_gen_put(skb->sk);
344 EXPORT_SYMBOL(sock_edemux);
346 struct sock *__inet_lookup_established(struct net *net,
347 struct inet_hashinfo *hashinfo,
348 const __be32 saddr, const __be16 sport,
349 const __be32 daddr, const u16 hnum,
350 const int dif, const int sdif)
352 INET_ADDR_COOKIE(acookie, saddr, daddr);
353 const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
355 const struct hlist_nulls_node *node;
356 /* Optimize here for direct hit, only listening connections can
357 * have wildcards anyways.
359 unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
360 unsigned int slot = hash & hashinfo->ehash_mask;
361 struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
364 sk_nulls_for_each_rcu(sk, node, &head->chain) {
365 if (sk->sk_hash != hash)
367 if (likely(INET_MATCH(sk, net, acookie,
368 saddr, daddr, ports, dif, sdif))) {
369 if (unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
371 if (unlikely(!INET_MATCH(sk, net, acookie,
381 * if the nulls value we got at the end of this lookup is
382 * not the expected one, we must restart lookup.
383 * We probably met an item that was moved to another chain.
385 if (get_nulls_value(node) != slot)
392 EXPORT_SYMBOL_GPL(__inet_lookup_established);
394 /* called with local bh disabled */
395 static int __inet_check_established(struct inet_timewait_death_row *death_row,
396 struct sock *sk, __u16 lport,
397 struct inet_timewait_sock **twp)
399 struct inet_hashinfo *hinfo = death_row->hashinfo;
400 struct inet_sock *inet = inet_sk(sk);
401 __be32 daddr = inet->inet_rcv_saddr;
402 __be32 saddr = inet->inet_daddr;
403 int dif = sk->sk_bound_dev_if;
404 struct net *net = sock_net(sk);
405 int sdif = l3mdev_master_ifindex_by_index(net, dif);
406 INET_ADDR_COOKIE(acookie, saddr, daddr);
407 const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
408 unsigned int hash = inet_ehashfn(net, daddr, lport,
409 saddr, inet->inet_dport);
410 struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
411 spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
413 const struct hlist_nulls_node *node;
414 struct inet_timewait_sock *tw = NULL;
418 sk_nulls_for_each(sk2, node, &head->chain) {
419 if (sk2->sk_hash != hash)
422 if (likely(INET_MATCH(sk2, net, acookie,
423 saddr, daddr, ports, dif, sdif))) {
424 if (sk2->sk_state == TCP_TIME_WAIT) {
426 if (twsk_unique(sk, sk2, twp))
433 /* Must record num and sport now. Otherwise we will see
434 * in hash table socket with a funny identity.
436 inet->inet_num = lport;
437 inet->inet_sport = htons(lport);
439 WARN_ON(!sk_unhashed(sk));
440 __sk_nulls_add_node_rcu(sk, &head->chain);
442 sk_nulls_del_node_init_rcu((struct sock *)tw);
443 __NET_INC_STATS(net, LINUX_MIB_TIMEWAITRECYCLED);
446 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
451 /* Silly. Should hash-dance instead... */
452 inet_twsk_deschedule_put(tw);
458 return -EADDRNOTAVAIL;
461 static u32 inet_sk_port_offset(const struct sock *sk)
463 const struct inet_sock *inet = inet_sk(sk);
465 return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
470 /* insert a socket into ehash, and eventually remove another one
471 * (The another one can be a SYN_RECV or TIMEWAIT
473 bool inet_ehash_insert(struct sock *sk, struct sock *osk)
475 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
476 struct hlist_nulls_head *list;
477 struct inet_ehash_bucket *head;
481 WARN_ON_ONCE(!sk_unhashed(sk));
483 sk->sk_hash = sk_ehashfn(sk);
484 head = inet_ehash_bucket(hashinfo, sk->sk_hash);
486 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
490 WARN_ON_ONCE(sk->sk_hash != osk->sk_hash);
491 ret = sk_nulls_del_node_init_rcu(osk);
494 __sk_nulls_add_node_rcu(sk, list);
499 bool inet_ehash_nolisten(struct sock *sk, struct sock *osk)
501 bool ok = inet_ehash_insert(sk, osk);
504 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
506 percpu_counter_inc(sk->sk_prot->orphan_count);
507 inet_sk_set_state(sk, TCP_CLOSE);
508 sock_set_flag(sk, SOCK_DEAD);
509 inet_csk_destroy_sock(sk);
513 EXPORT_SYMBOL_GPL(inet_ehash_nolisten);
515 static int inet_reuseport_add_sock(struct sock *sk,
516 struct inet_listen_hashbucket *ilb)
518 struct inet_bind_bucket *tb = inet_csk(sk)->icsk_bind_hash;
520 kuid_t uid = sock_i_uid(sk);
522 sk_for_each_rcu(sk2, &ilb->head) {
524 sk2->sk_family == sk->sk_family &&
525 ipv6_only_sock(sk2) == ipv6_only_sock(sk) &&
526 sk2->sk_bound_dev_if == sk->sk_bound_dev_if &&
527 inet_csk(sk2)->icsk_bind_hash == tb &&
528 sk2->sk_reuseport && uid_eq(uid, sock_i_uid(sk2)) &&
529 inet_rcv_saddr_equal(sk, sk2, false))
530 return reuseport_add_sock(sk, sk2,
531 inet_rcv_saddr_any(sk));
534 return reuseport_alloc(sk, inet_rcv_saddr_any(sk));
537 int __inet_hash(struct sock *sk, struct sock *osk)
539 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
540 struct inet_listen_hashbucket *ilb;
543 if (sk->sk_state != TCP_LISTEN) {
544 inet_ehash_nolisten(sk, osk);
547 WARN_ON(!sk_unhashed(sk));
548 ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
550 spin_lock(&ilb->lock);
551 if (sk->sk_reuseport) {
552 err = inet_reuseport_add_sock(sk, ilb);
556 if (IS_ENABLED(CONFIG_IPV6) && sk->sk_reuseport &&
557 sk->sk_family == AF_INET6)
558 hlist_add_tail_rcu(&sk->sk_node, &ilb->head);
560 hlist_add_head_rcu(&sk->sk_node, &ilb->head);
561 inet_hash2(hashinfo, sk);
563 sock_set_flag(sk, SOCK_RCU_FREE);
564 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
566 spin_unlock(&ilb->lock);
570 EXPORT_SYMBOL(__inet_hash);
572 int inet_hash(struct sock *sk)
576 if (sk->sk_state != TCP_CLOSE) {
578 err = __inet_hash(sk, NULL);
584 EXPORT_SYMBOL_GPL(inet_hash);
586 void inet_unhash(struct sock *sk)
588 struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
589 struct inet_listen_hashbucket *ilb = NULL;
595 if (sk->sk_state == TCP_LISTEN) {
596 ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
599 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
605 if (rcu_access_pointer(sk->sk_reuseport_cb))
606 reuseport_detach_sock(sk);
608 inet_unhash2(hashinfo, sk);
609 __sk_del_node_init(sk);
612 __sk_nulls_del_node_init_rcu(sk);
614 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
616 spin_unlock_bh(lock);
618 EXPORT_SYMBOL_GPL(inet_unhash);
620 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
621 struct sock *sk, u32 port_offset,
622 int (*check_established)(struct inet_timewait_death_row *,
623 struct sock *, __u16, struct inet_timewait_sock **))
625 struct inet_hashinfo *hinfo = death_row->hashinfo;
626 struct inet_timewait_sock *tw = NULL;
627 struct inet_bind_hashbucket *head;
628 int port = inet_sk(sk)->inet_num;
629 struct net *net = sock_net(sk);
630 struct inet_bind_bucket *tb;
631 u32 remaining, offset;
632 int ret, i, low, high;
637 head = &hinfo->bhash[inet_bhashfn(net, port,
639 tb = inet_csk(sk)->icsk_bind_hash;
640 spin_lock_bh(&head->lock);
641 if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
642 inet_ehash_nolisten(sk, NULL);
643 spin_unlock_bh(&head->lock);
646 spin_unlock(&head->lock);
647 /* No definite answer... Walk to established hash table */
648 ret = check_established(death_row, sk, port, NULL);
653 l3mdev = inet_sk_bound_l3mdev(sk);
655 inet_get_local_port_range(net, &low, &high);
656 high++; /* [32768, 60999] -> [32768, 61000[ */
657 remaining = high - low;
658 if (likely(remaining > 1))
661 offset = (hint + port_offset) % remaining;
662 /* In first pass we try ports of @low parity.
663 * inet_csk_get_port() does the opposite choice.
668 for (i = 0; i < remaining; i += 2, port += 2) {
669 if (unlikely(port >= high))
671 if (inet_is_local_reserved_port(net, port))
673 head = &hinfo->bhash[inet_bhashfn(net, port,
675 spin_lock_bh(&head->lock);
677 /* Does not bother with rcv_saddr checks, because
678 * the established check is already unique enough.
680 inet_bind_bucket_for_each(tb, &head->chain) {
681 if (net_eq(ib_net(tb), net) && tb->l3mdev == l3mdev &&
683 if (tb->fastreuse >= 0 ||
684 tb->fastreuseport >= 0)
686 WARN_ON(hlist_empty(&tb->owners));
687 if (!check_established(death_row, sk,
694 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
695 net, head, port, l3mdev);
697 spin_unlock_bh(&head->lock);
701 tb->fastreuseport = -1;
704 spin_unlock_bh(&head->lock);
709 if ((offset & 1) && remaining > 1)
710 goto other_parity_scan;
712 return -EADDRNOTAVAIL;
717 /* Head lock still held and bh's disabled */
718 inet_bind_hash(sk, tb, port);
719 if (sk_unhashed(sk)) {
720 inet_sk(sk)->inet_sport = htons(port);
721 inet_ehash_nolisten(sk, (struct sock *)tw);
724 inet_twsk_bind_unhash(tw, hinfo);
725 spin_unlock(&head->lock);
727 inet_twsk_deschedule_put(tw);
733 * Bind a port for a connect operation and hash it.
735 int inet_hash_connect(struct inet_timewait_death_row *death_row,
740 if (!inet_sk(sk)->inet_num)
741 port_offset = inet_sk_port_offset(sk);
742 return __inet_hash_connect(death_row, sk, port_offset,
743 __inet_check_established);
745 EXPORT_SYMBOL_GPL(inet_hash_connect);
747 void inet_hashinfo_init(struct inet_hashinfo *h)
751 for (i = 0; i < INET_LHTABLE_SIZE; i++) {
752 spin_lock_init(&h->listening_hash[i].lock);
753 INIT_HLIST_HEAD(&h->listening_hash[i].head);
754 h->listening_hash[i].count = 0;
759 EXPORT_SYMBOL_GPL(inet_hashinfo_init);
761 static void init_hashinfo_lhash2(struct inet_hashinfo *h)
765 for (i = 0; i <= h->lhash2_mask; i++) {
766 spin_lock_init(&h->lhash2[i].lock);
767 INIT_HLIST_HEAD(&h->lhash2[i].head);
768 h->lhash2[i].count = 0;
772 void __init inet_hashinfo2_init(struct inet_hashinfo *h, const char *name,
773 unsigned long numentries, int scale,
774 unsigned long low_limit,
775 unsigned long high_limit)
777 h->lhash2 = alloc_large_system_hash(name,
786 init_hashinfo_lhash2(h);
789 int inet_hashinfo2_init_mod(struct inet_hashinfo *h)
791 h->lhash2 = kmalloc_array(INET_LHTABLE_SIZE, sizeof(*h->lhash2), GFP_KERNEL);
795 h->lhash2_mask = INET_LHTABLE_SIZE - 1;
796 /* INET_LHTABLE_SIZE must be a power of 2 */
797 BUG_ON(INET_LHTABLE_SIZE & h->lhash2_mask);
799 init_hashinfo_lhash2(h);
802 EXPORT_SYMBOL_GPL(inet_hashinfo2_init_mod);
804 int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
806 unsigned int locksz = sizeof(spinlock_t);
807 unsigned int i, nblocks = 1;
810 /* allocate 2 cache lines or at least one spinlock per cpu */
811 nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U);
812 nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
814 /* no more locks than number of hash buckets */
815 nblocks = min(nblocks, hashinfo->ehash_mask + 1);
817 hashinfo->ehash_locks = kvmalloc_array(nblocks, locksz, GFP_KERNEL);
818 if (!hashinfo->ehash_locks)
821 for (i = 0; i < nblocks; i++)
822 spin_lock_init(&hashinfo->ehash_locks[i]);
824 hashinfo->ehash_locks_mask = nblocks - 1;
827 EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);