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 #if IS_ENABLED(CONFIG_IPV6)
24 #include <net/inet6_hashtables.h>
26 #include <net/secure_seq.h>
29 #include <net/sock_reuseport.h>
31 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));
42 EXPORT_SYMBOL_GPL(inet_ehashfn);
44 /* This function handles inet_sock, but also timewait and request sockets
47 static u32 sk_ehashfn(const struct sock *sk)
49 #if IS_ENABLED(CONFIG_IPV6)
50 if (sk->sk_family == AF_INET6 &&
51 !ipv6_addr_v4mapped(&sk->sk_v6_daddr))
52 return inet6_ehashfn(sock_net(sk),
53 &sk->sk_v6_rcv_saddr, sk->sk_num,
54 &sk->sk_v6_daddr, sk->sk_dport);
56 return inet_ehashfn(sock_net(sk),
57 sk->sk_rcv_saddr, sk->sk_num,
58 sk->sk_daddr, sk->sk_dport);
62 * Allocate and initialize a new local port bind bucket.
63 * The bindhash mutex for snum's hash chain must be held here.
65 struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep,
67 struct inet_bind_hashbucket *head,
68 const unsigned short snum,
71 struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
74 write_pnet(&tb->ib_net, net);
78 tb->fastreuseport = 0;
79 INIT_HLIST_HEAD(&tb->owners);
80 hlist_add_head(&tb->node, &head->chain);
86 * Caller must hold hashbucket lock for this tb with local BH disabled
88 void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
90 if (hlist_empty(&tb->owners)) {
91 __hlist_del(&tb->node);
92 kmem_cache_free(cachep, tb);
96 bool inet_bind_bucket_match(const struct inet_bind_bucket *tb, const struct net *net,
97 unsigned short port, int l3mdev)
99 return net_eq(ib_net(tb), net) && tb->port == port &&
100 tb->l3mdev == l3mdev;
103 static void inet_bind2_bucket_init(struct inet_bind2_bucket *tb,
105 struct inet_bind_hashbucket *head,
106 unsigned short port, int l3mdev,
107 const struct sock *sk)
109 write_pnet(&tb->ib_net, net);
112 #if IS_ENABLED(CONFIG_IPV6)
113 tb->family = sk->sk_family;
114 if (sk->sk_family == AF_INET6)
115 tb->v6_rcv_saddr = sk->sk_v6_rcv_saddr;
118 tb->rcv_saddr = sk->sk_rcv_saddr;
119 INIT_HLIST_HEAD(&tb->owners);
120 INIT_HLIST_HEAD(&tb->deathrow);
121 hlist_add_head(&tb->node, &head->chain);
124 struct inet_bind2_bucket *inet_bind2_bucket_create(struct kmem_cache *cachep,
126 struct inet_bind_hashbucket *head,
129 const struct sock *sk)
131 struct inet_bind2_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
134 inet_bind2_bucket_init(tb, net, head, port, l3mdev, sk);
139 /* Caller must hold hashbucket lock for this tb with local BH disabled */
140 void inet_bind2_bucket_destroy(struct kmem_cache *cachep, struct inet_bind2_bucket *tb)
142 if (hlist_empty(&tb->owners) && hlist_empty(&tb->deathrow)) {
143 __hlist_del(&tb->node);
144 kmem_cache_free(cachep, tb);
148 static bool inet_bind2_bucket_addr_match(const struct inet_bind2_bucket *tb2,
149 const struct sock *sk)
151 #if IS_ENABLED(CONFIG_IPV6)
152 if (sk->sk_family != tb2->family)
155 if (sk->sk_family == AF_INET6)
156 return ipv6_addr_equal(&tb2->v6_rcv_saddr,
157 &sk->sk_v6_rcv_saddr);
159 return tb2->rcv_saddr == sk->sk_rcv_saddr;
162 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
163 struct inet_bind2_bucket *tb2, unsigned short port)
165 inet_sk(sk)->inet_num = port;
166 sk_add_bind_node(sk, &tb->owners);
167 inet_csk(sk)->icsk_bind_hash = tb;
168 sk_add_bind2_node(sk, &tb2->owners);
169 inet_csk(sk)->icsk_bind2_hash = tb2;
173 * Get rid of any references to a local port held by the given sock.
175 static void __inet_put_port(struct sock *sk)
177 struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
178 struct inet_bind_hashbucket *head, *head2;
179 struct net *net = sock_net(sk);
180 struct inet_bind_bucket *tb;
183 bhash = inet_bhashfn(net, inet_sk(sk)->inet_num, hashinfo->bhash_size);
184 head = &hashinfo->bhash[bhash];
185 head2 = inet_bhashfn_portaddr(hashinfo, sk, net, inet_sk(sk)->inet_num);
187 spin_lock(&head->lock);
188 tb = inet_csk(sk)->icsk_bind_hash;
189 __sk_del_bind_node(sk);
190 inet_csk(sk)->icsk_bind_hash = NULL;
191 inet_sk(sk)->inet_num = 0;
192 inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
194 spin_lock(&head2->lock);
195 if (inet_csk(sk)->icsk_bind2_hash) {
196 struct inet_bind2_bucket *tb2 = inet_csk(sk)->icsk_bind2_hash;
198 __sk_del_bind2_node(sk);
199 inet_csk(sk)->icsk_bind2_hash = NULL;
200 inet_bind2_bucket_destroy(hashinfo->bind2_bucket_cachep, tb2);
202 spin_unlock(&head2->lock);
204 spin_unlock(&head->lock);
207 void inet_put_port(struct sock *sk)
213 EXPORT_SYMBOL(inet_put_port);
215 int __inet_inherit_port(const struct sock *sk, struct sock *child)
217 struct inet_hashinfo *table = tcp_or_dccp_get_hashinfo(sk);
218 unsigned short port = inet_sk(child)->inet_num;
219 struct inet_bind_hashbucket *head, *head2;
220 bool created_inet_bind_bucket = false;
221 struct net *net = sock_net(sk);
222 bool update_fastreuse = false;
223 struct inet_bind2_bucket *tb2;
224 struct inet_bind_bucket *tb;
227 bhash = inet_bhashfn(net, port, table->bhash_size);
228 head = &table->bhash[bhash];
229 head2 = inet_bhashfn_portaddr(table, child, net, port);
231 spin_lock(&head->lock);
232 spin_lock(&head2->lock);
233 tb = inet_csk(sk)->icsk_bind_hash;
234 tb2 = inet_csk(sk)->icsk_bind2_hash;
235 if (unlikely(!tb || !tb2)) {
236 spin_unlock(&head2->lock);
237 spin_unlock(&head->lock);
240 if (tb->port != port) {
241 l3mdev = inet_sk_bound_l3mdev(sk);
243 /* NOTE: using tproxy and redirecting skbs to a proxy
244 * on a different listener port breaks the assumption
245 * that the listener socket's icsk_bind_hash is the same
246 * as that of the child socket. We have to look up or
247 * create a new bind bucket for the child here. */
248 inet_bind_bucket_for_each(tb, &head->chain) {
249 if (inet_bind_bucket_match(tb, net, port, l3mdev))
253 tb = inet_bind_bucket_create(table->bind_bucket_cachep,
254 net, head, port, l3mdev);
256 spin_unlock(&head2->lock);
257 spin_unlock(&head->lock);
260 created_inet_bind_bucket = true;
262 update_fastreuse = true;
265 } else if (!inet_bind2_bucket_addr_match(tb2, child)) {
266 l3mdev = inet_sk_bound_l3mdev(sk);
269 tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, child);
271 tb2 = inet_bind2_bucket_create(table->bind2_bucket_cachep,
278 if (update_fastreuse)
279 inet_csk_update_fastreuse(tb, child);
280 inet_bind_hash(child, tb, tb2, port);
281 spin_unlock(&head2->lock);
282 spin_unlock(&head->lock);
287 if (created_inet_bind_bucket)
288 inet_bind_bucket_destroy(table->bind_bucket_cachep, tb);
289 spin_unlock(&head2->lock);
290 spin_unlock(&head->lock);
293 EXPORT_SYMBOL_GPL(__inet_inherit_port);
295 static struct inet_listen_hashbucket *
296 inet_lhash2_bucket_sk(struct inet_hashinfo *h, struct sock *sk)
300 #if IS_ENABLED(CONFIG_IPV6)
301 if (sk->sk_family == AF_INET6)
302 hash = ipv6_portaddr_hash(sock_net(sk),
303 &sk->sk_v6_rcv_saddr,
304 inet_sk(sk)->inet_num);
307 hash = ipv4_portaddr_hash(sock_net(sk),
308 inet_sk(sk)->inet_rcv_saddr,
309 inet_sk(sk)->inet_num);
310 return inet_lhash2_bucket(h, hash);
313 static inline int compute_score(struct sock *sk, struct net *net,
314 const unsigned short hnum, const __be32 daddr,
315 const int dif, const int sdif)
319 if (net_eq(sock_net(sk), net) && sk->sk_num == hnum &&
320 !ipv6_only_sock(sk)) {
321 if (sk->sk_rcv_saddr != daddr)
324 if (!inet_sk_bound_dev_eq(net, sk->sk_bound_dev_if, dif, sdif))
326 score = sk->sk_bound_dev_if ? 2 : 1;
328 if (sk->sk_family == PF_INET)
330 if (READ_ONCE(sk->sk_incoming_cpu) == raw_smp_processor_id())
337 * inet_lookup_reuseport() - execute reuseport logic on AF_INET socket if necessary.
338 * @net: network namespace.
339 * @sk: AF_INET socket, must be in TCP_LISTEN state for TCP or TCP_CLOSE for UDP.
340 * @skb: context for a potential SK_REUSEPORT program.
341 * @doff: header offset.
342 * @saddr: source address.
343 * @sport: source port.
344 * @daddr: destination address.
345 * @hnum: destination port in host byte order.
346 * @ehashfn: hash function used to generate the fallback hash.
348 * Return: NULL if sk doesn't have SO_REUSEPORT set, otherwise a pointer to
349 * the selected sock or an error.
351 struct sock *inet_lookup_reuseport(struct net *net, struct sock *sk,
352 struct sk_buff *skb, int doff,
353 __be32 saddr, __be16 sport,
354 __be32 daddr, unsigned short hnum,
355 inet_ehashfn_t *ehashfn)
357 struct sock *reuse_sk = NULL;
360 if (sk->sk_reuseport) {
361 phash = INDIRECT_CALL_2(ehashfn, udp_ehashfn, inet_ehashfn,
362 net, daddr, hnum, saddr, sport);
363 reuse_sk = reuseport_select_sock(sk, phash, skb, doff);
367 EXPORT_SYMBOL_GPL(inet_lookup_reuseport);
370 * Here are some nice properties to exploit here. The BSD API
371 * does not allow a listening sock to specify the remote port nor the
372 * remote address for the connection. So always assume those are both
373 * wildcarded during the search since they can never be otherwise.
376 /* called with rcu_read_lock() : No refcount taken on the socket */
377 static struct sock *inet_lhash2_lookup(struct net *net,
378 struct inet_listen_hashbucket *ilb2,
379 struct sk_buff *skb, int doff,
380 const __be32 saddr, __be16 sport,
381 const __be32 daddr, const unsigned short hnum,
382 const int dif, const int sdif)
384 struct sock *sk, *result = NULL;
385 struct hlist_nulls_node *node;
386 int score, hiscore = 0;
388 sk_nulls_for_each_rcu(sk, node, &ilb2->nulls_head) {
389 score = compute_score(sk, net, hnum, daddr, dif, sdif);
390 if (score > hiscore) {
391 result = inet_lookup_reuseport(net, sk, skb, doff,
392 saddr, sport, daddr, hnum, inet_ehashfn);
404 struct sock *inet_lookup_run_sk_lookup(struct net *net,
406 struct sk_buff *skb, int doff,
407 __be32 saddr, __be16 sport,
408 __be32 daddr, u16 hnum, const int dif,
409 inet_ehashfn_t *ehashfn)
411 struct sock *sk, *reuse_sk;
414 no_reuseport = bpf_sk_lookup_run_v4(net, protocol, saddr, sport,
415 daddr, hnum, dif, &sk);
416 if (no_reuseport || IS_ERR_OR_NULL(sk))
419 reuse_sk = inet_lookup_reuseport(net, sk, skb, doff, saddr, sport, daddr, hnum,
426 struct sock *__inet_lookup_listener(struct net *net,
427 struct inet_hashinfo *hashinfo,
428 struct sk_buff *skb, int doff,
429 const __be32 saddr, __be16 sport,
430 const __be32 daddr, const unsigned short hnum,
431 const int dif, const int sdif)
433 struct inet_listen_hashbucket *ilb2;
434 struct sock *result = NULL;
437 /* Lookup redirect from BPF */
438 if (static_branch_unlikely(&bpf_sk_lookup_enabled) &&
439 hashinfo == net->ipv4.tcp_death_row.hashinfo) {
440 result = inet_lookup_run_sk_lookup(net, IPPROTO_TCP, skb, doff,
441 saddr, sport, daddr, hnum, dif,
447 hash2 = ipv4_portaddr_hash(net, daddr, hnum);
448 ilb2 = inet_lhash2_bucket(hashinfo, hash2);
450 result = inet_lhash2_lookup(net, ilb2, skb, doff,
451 saddr, sport, daddr, hnum,
456 /* Lookup lhash2 with INADDR_ANY */
457 hash2 = ipv4_portaddr_hash(net, htonl(INADDR_ANY), hnum);
458 ilb2 = inet_lhash2_bucket(hashinfo, hash2);
460 result = inet_lhash2_lookup(net, ilb2, skb, doff,
461 saddr, sport, htonl(INADDR_ANY), hnum,
468 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
470 /* All sockets share common refcount, but have different destructors */
471 void sock_gen_put(struct sock *sk)
473 if (!refcount_dec_and_test(&sk->sk_refcnt))
476 if (sk->sk_state == TCP_TIME_WAIT)
477 inet_twsk_free(inet_twsk(sk));
478 else if (sk->sk_state == TCP_NEW_SYN_RECV)
479 reqsk_free(inet_reqsk(sk));
483 EXPORT_SYMBOL_GPL(sock_gen_put);
485 void sock_edemux(struct sk_buff *skb)
487 sock_gen_put(skb->sk);
489 EXPORT_SYMBOL(sock_edemux);
491 struct sock *__inet_lookup_established(struct net *net,
492 struct inet_hashinfo *hashinfo,
493 const __be32 saddr, const __be16 sport,
494 const __be32 daddr, const u16 hnum,
495 const int dif, const int sdif)
497 INET_ADDR_COOKIE(acookie, saddr, daddr);
498 const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
500 const struct hlist_nulls_node *node;
501 /* Optimize here for direct hit, only listening connections can
502 * have wildcards anyways.
504 unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
505 unsigned int slot = hash & hashinfo->ehash_mask;
506 struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
509 sk_nulls_for_each_rcu(sk, node, &head->chain) {
510 if (sk->sk_hash != hash)
512 if (likely(inet_match(net, sk, acookie, ports, dif, sdif))) {
513 if (unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
515 if (unlikely(!inet_match(net, sk, acookie,
516 ports, dif, sdif))) {
524 * if the nulls value we got at the end of this lookup is
525 * not the expected one, we must restart lookup.
526 * We probably met an item that was moved to another chain.
528 if (get_nulls_value(node) != slot)
535 EXPORT_SYMBOL_GPL(__inet_lookup_established);
537 /* called with local bh disabled */
538 static int __inet_check_established(struct inet_timewait_death_row *death_row,
539 struct sock *sk, __u16 lport,
540 struct inet_timewait_sock **twp)
542 struct inet_hashinfo *hinfo = death_row->hashinfo;
543 struct inet_sock *inet = inet_sk(sk);
544 __be32 daddr = inet->inet_rcv_saddr;
545 __be32 saddr = inet->inet_daddr;
546 int dif = sk->sk_bound_dev_if;
547 struct net *net = sock_net(sk);
548 int sdif = l3mdev_master_ifindex_by_index(net, dif);
549 INET_ADDR_COOKIE(acookie, saddr, daddr);
550 const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
551 unsigned int hash = inet_ehashfn(net, daddr, lport,
552 saddr, inet->inet_dport);
553 struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
554 spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
556 const struct hlist_nulls_node *node;
557 struct inet_timewait_sock *tw = NULL;
561 sk_nulls_for_each(sk2, node, &head->chain) {
562 if (sk2->sk_hash != hash)
565 if (likely(inet_match(net, sk2, acookie, ports, dif, sdif))) {
566 if (sk2->sk_state == TCP_TIME_WAIT) {
568 if (twsk_unique(sk, sk2, twp))
575 /* Must record num and sport now. Otherwise we will see
576 * in hash table socket with a funny identity.
578 inet->inet_num = lport;
579 inet->inet_sport = htons(lport);
581 WARN_ON(!sk_unhashed(sk));
582 __sk_nulls_add_node_rcu(sk, &head->chain);
584 sk_nulls_del_node_init_rcu((struct sock *)tw);
585 __NET_INC_STATS(net, LINUX_MIB_TIMEWAITRECYCLED);
588 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
593 /* Silly. Should hash-dance instead... */
594 inet_twsk_deschedule_put(tw);
600 return -EADDRNOTAVAIL;
603 static u64 inet_sk_port_offset(const struct sock *sk)
605 const struct inet_sock *inet = inet_sk(sk);
607 return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
612 /* Searches for an exsiting socket in the ehash bucket list.
613 * Returns true if found, false otherwise.
615 static bool inet_ehash_lookup_by_sk(struct sock *sk,
616 struct hlist_nulls_head *list)
618 const __portpair ports = INET_COMBINED_PORTS(sk->sk_dport, sk->sk_num);
619 const int sdif = sk->sk_bound_dev_if;
620 const int dif = sk->sk_bound_dev_if;
621 const struct hlist_nulls_node *node;
622 struct net *net = sock_net(sk);
625 INET_ADDR_COOKIE(acookie, sk->sk_daddr, sk->sk_rcv_saddr);
627 sk_nulls_for_each_rcu(esk, node, list) {
628 if (esk->sk_hash != sk->sk_hash)
630 if (sk->sk_family == AF_INET) {
631 if (unlikely(inet_match(net, esk, acookie,
632 ports, dif, sdif))) {
636 #if IS_ENABLED(CONFIG_IPV6)
637 else if (sk->sk_family == AF_INET6) {
638 if (unlikely(inet6_match(net, esk,
640 &sk->sk_v6_rcv_saddr,
641 ports, dif, sdif))) {
650 /* Insert a socket into ehash, and eventually remove another one
651 * (The another one can be a SYN_RECV or TIMEWAIT)
652 * If an existing socket already exists, socket sk is not inserted,
653 * and sets found_dup_sk parameter to true.
655 bool inet_ehash_insert(struct sock *sk, struct sock *osk, bool *found_dup_sk)
657 struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
658 struct inet_ehash_bucket *head;
659 struct hlist_nulls_head *list;
663 WARN_ON_ONCE(!sk_unhashed(sk));
665 sk->sk_hash = sk_ehashfn(sk);
666 head = inet_ehash_bucket(hashinfo, sk->sk_hash);
668 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
672 WARN_ON_ONCE(sk->sk_hash != osk->sk_hash);
673 ret = sk_nulls_del_node_init_rcu(osk);
674 } else if (found_dup_sk) {
675 *found_dup_sk = inet_ehash_lookup_by_sk(sk, list);
681 __sk_nulls_add_node_rcu(sk, list);
688 bool inet_ehash_nolisten(struct sock *sk, struct sock *osk, bool *found_dup_sk)
690 bool ok = inet_ehash_insert(sk, osk, found_dup_sk);
693 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
695 this_cpu_inc(*sk->sk_prot->orphan_count);
696 inet_sk_set_state(sk, TCP_CLOSE);
697 sock_set_flag(sk, SOCK_DEAD);
698 inet_csk_destroy_sock(sk);
702 EXPORT_SYMBOL_GPL(inet_ehash_nolisten);
704 static int inet_reuseport_add_sock(struct sock *sk,
705 struct inet_listen_hashbucket *ilb)
707 struct inet_bind_bucket *tb = inet_csk(sk)->icsk_bind_hash;
708 const struct hlist_nulls_node *node;
710 kuid_t uid = sock_i_uid(sk);
712 sk_nulls_for_each_rcu(sk2, node, &ilb->nulls_head) {
714 sk2->sk_family == sk->sk_family &&
715 ipv6_only_sock(sk2) == ipv6_only_sock(sk) &&
716 sk2->sk_bound_dev_if == sk->sk_bound_dev_if &&
717 inet_csk(sk2)->icsk_bind_hash == tb &&
718 sk2->sk_reuseport && uid_eq(uid, sock_i_uid(sk2)) &&
719 inet_rcv_saddr_equal(sk, sk2, false))
720 return reuseport_add_sock(sk, sk2,
721 inet_rcv_saddr_any(sk));
724 return reuseport_alloc(sk, inet_rcv_saddr_any(sk));
727 int __inet_hash(struct sock *sk, struct sock *osk)
729 struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
730 struct inet_listen_hashbucket *ilb2;
733 if (sk->sk_state != TCP_LISTEN) {
735 inet_ehash_nolisten(sk, osk, NULL);
739 WARN_ON(!sk_unhashed(sk));
740 ilb2 = inet_lhash2_bucket_sk(hashinfo, sk);
742 spin_lock(&ilb2->lock);
743 if (sk->sk_reuseport) {
744 err = inet_reuseport_add_sock(sk, ilb2);
748 if (IS_ENABLED(CONFIG_IPV6) && sk->sk_reuseport &&
749 sk->sk_family == AF_INET6)
750 __sk_nulls_add_node_tail_rcu(sk, &ilb2->nulls_head);
752 __sk_nulls_add_node_rcu(sk, &ilb2->nulls_head);
753 sock_set_flag(sk, SOCK_RCU_FREE);
754 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
756 spin_unlock(&ilb2->lock);
760 EXPORT_SYMBOL(__inet_hash);
762 int inet_hash(struct sock *sk)
766 if (sk->sk_state != TCP_CLOSE)
767 err = __inet_hash(sk, NULL);
771 EXPORT_SYMBOL_GPL(inet_hash);
773 void inet_unhash(struct sock *sk)
775 struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
780 if (sk->sk_state == TCP_LISTEN) {
781 struct inet_listen_hashbucket *ilb2;
783 ilb2 = inet_lhash2_bucket_sk(hashinfo, sk);
784 /* Don't disable bottom halves while acquiring the lock to
785 * avoid circular locking dependency on PREEMPT_RT.
787 spin_lock(&ilb2->lock);
788 if (sk_unhashed(sk)) {
789 spin_unlock(&ilb2->lock);
793 if (rcu_access_pointer(sk->sk_reuseport_cb))
794 reuseport_stop_listen_sock(sk);
796 __sk_nulls_del_node_init_rcu(sk);
797 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
798 spin_unlock(&ilb2->lock);
800 spinlock_t *lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
803 if (sk_unhashed(sk)) {
804 spin_unlock_bh(lock);
807 __sk_nulls_del_node_init_rcu(sk);
808 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
809 spin_unlock_bh(lock);
812 EXPORT_SYMBOL_GPL(inet_unhash);
814 static bool inet_bind2_bucket_match(const struct inet_bind2_bucket *tb,
815 const struct net *net, unsigned short port,
816 int l3mdev, const struct sock *sk)
818 if (!net_eq(ib2_net(tb), net) || tb->port != port ||
819 tb->l3mdev != l3mdev)
822 #if IS_ENABLED(CONFIG_IPV6)
823 if (sk->sk_family != tb->family)
826 if (sk->sk_family == AF_INET6)
827 return ipv6_addr_equal(&tb->v6_rcv_saddr, &sk->sk_v6_rcv_saddr);
829 return tb->rcv_saddr == sk->sk_rcv_saddr;
832 bool inet_bind2_bucket_match_addr_any(const struct inet_bind2_bucket *tb, const struct net *net,
833 unsigned short port, int l3mdev, const struct sock *sk)
835 if (!net_eq(ib2_net(tb), net) || tb->port != port ||
836 tb->l3mdev != l3mdev)
839 #if IS_ENABLED(CONFIG_IPV6)
840 if (sk->sk_family != tb->family) {
841 if (sk->sk_family == AF_INET)
842 return ipv6_addr_any(&tb->v6_rcv_saddr);
847 if (sk->sk_family == AF_INET6)
848 return ipv6_addr_any(&tb->v6_rcv_saddr);
850 return tb->rcv_saddr == 0;
853 /* The socket's bhash2 hashbucket spinlock must be held when this is called */
854 struct inet_bind2_bucket *
855 inet_bind2_bucket_find(const struct inet_bind_hashbucket *head, const struct net *net,
856 unsigned short port, int l3mdev, const struct sock *sk)
858 struct inet_bind2_bucket *bhash2 = NULL;
860 inet_bind_bucket_for_each(bhash2, &head->chain)
861 if (inet_bind2_bucket_match(bhash2, net, port, l3mdev, sk))
867 struct inet_bind_hashbucket *
868 inet_bhash2_addr_any_hashbucket(const struct sock *sk, const struct net *net, int port)
870 struct inet_hashinfo *hinfo = tcp_or_dccp_get_hashinfo(sk);
873 #if IS_ENABLED(CONFIG_IPV6)
874 if (sk->sk_family == AF_INET6)
875 hash = ipv6_portaddr_hash(net, &in6addr_any, port);
878 hash = ipv4_portaddr_hash(net, 0, port);
880 return &hinfo->bhash2[hash & (hinfo->bhash_size - 1)];
883 static void inet_update_saddr(struct sock *sk, void *saddr, int family)
885 if (family == AF_INET) {
886 inet_sk(sk)->inet_saddr = *(__be32 *)saddr;
887 sk_rcv_saddr_set(sk, inet_sk(sk)->inet_saddr);
889 #if IS_ENABLED(CONFIG_IPV6)
891 sk->sk_v6_rcv_saddr = *(struct in6_addr *)saddr;
896 static int __inet_bhash2_update_saddr(struct sock *sk, void *saddr, int family, bool reset)
898 struct inet_hashinfo *hinfo = tcp_or_dccp_get_hashinfo(sk);
899 struct inet_bind_hashbucket *head, *head2;
900 struct inet_bind2_bucket *tb2, *new_tb2;
901 int l3mdev = inet_sk_bound_l3mdev(sk);
902 int port = inet_sk(sk)->inet_num;
903 struct net *net = sock_net(sk);
906 if (!inet_csk(sk)->icsk_bind2_hash) {
907 /* Not bind()ed before. */
909 inet_reset_saddr(sk);
911 inet_update_saddr(sk, saddr, family);
916 /* Allocate a bind2 bucket ahead of time to avoid permanently putting
917 * the bhash2 table in an inconsistent state if a new tb2 bucket
920 new_tb2 = kmem_cache_alloc(hinfo->bind2_bucket_cachep, GFP_ATOMIC);
923 /* The (INADDR_ANY, port) bucket might have already
924 * been freed, then we cannot fixup icsk_bind2_hash,
925 * so we give up and unlink sk from bhash/bhash2 not
926 * to leave inconsistency in bhash2.
929 inet_reset_saddr(sk);
935 bhash = inet_bhashfn(net, port, hinfo->bhash_size);
936 head = &hinfo->bhash[bhash];
937 head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
939 /* If we change saddr locklessly, another thread
940 * iterating over bhash might see corrupted address.
942 spin_lock_bh(&head->lock);
944 spin_lock(&head2->lock);
945 __sk_del_bind2_node(sk);
946 inet_bind2_bucket_destroy(hinfo->bind2_bucket_cachep, inet_csk(sk)->icsk_bind2_hash);
947 spin_unlock(&head2->lock);
950 inet_reset_saddr(sk);
952 inet_update_saddr(sk, saddr, family);
954 head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
956 spin_lock(&head2->lock);
957 tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, sk);
960 inet_bind2_bucket_init(tb2, net, head2, port, l3mdev, sk);
962 sk_add_bind2_node(sk, &tb2->owners);
963 inet_csk(sk)->icsk_bind2_hash = tb2;
964 spin_unlock(&head2->lock);
966 spin_unlock_bh(&head->lock);
969 kmem_cache_free(hinfo->bind2_bucket_cachep, new_tb2);
974 int inet_bhash2_update_saddr(struct sock *sk, void *saddr, int family)
976 return __inet_bhash2_update_saddr(sk, saddr, family, false);
978 EXPORT_SYMBOL_GPL(inet_bhash2_update_saddr);
980 void inet_bhash2_reset_saddr(struct sock *sk)
982 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
983 __inet_bhash2_update_saddr(sk, NULL, 0, true);
985 EXPORT_SYMBOL_GPL(inet_bhash2_reset_saddr);
987 /* RFC 6056 3.3.4. Algorithm 4: Double-Hash Port Selection Algorithm
988 * Note that we use 32bit integers (vs RFC 'short integers')
989 * because 2^16 is not a multiple of num_ephemeral and this
990 * property might be used by clever attacker.
992 * RFC claims using TABLE_LENGTH=10 buckets gives an improvement, though
993 * attacks were since demonstrated, thus we use 65536 by default instead
994 * to really give more isolation and privacy, at the expense of 256kB
997 #define INET_TABLE_PERTURB_SIZE (1 << CONFIG_INET_TABLE_PERTURB_ORDER)
998 static u32 *table_perturb;
1000 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
1001 struct sock *sk, u64 port_offset,
1002 int (*check_established)(struct inet_timewait_death_row *,
1003 struct sock *, __u16, struct inet_timewait_sock **))
1005 struct inet_hashinfo *hinfo = death_row->hashinfo;
1006 struct inet_bind_hashbucket *head, *head2;
1007 struct inet_timewait_sock *tw = NULL;
1008 int port = inet_sk(sk)->inet_num;
1009 struct net *net = sock_net(sk);
1010 struct inet_bind2_bucket *tb2;
1011 struct inet_bind_bucket *tb;
1012 bool tb_created = false;
1013 u32 remaining, offset;
1014 int ret, i, low, high;
1020 ret = check_established(death_row, sk, port, NULL);
1025 l3mdev = inet_sk_bound_l3mdev(sk);
1027 inet_sk_get_local_port_range(sk, &low, &high);
1028 high++; /* [32768, 60999] -> [32768, 61000[ */
1029 remaining = high - low;
1030 if (likely(remaining > 1))
1033 get_random_sleepable_once(table_perturb,
1034 INET_TABLE_PERTURB_SIZE * sizeof(*table_perturb));
1035 index = port_offset & (INET_TABLE_PERTURB_SIZE - 1);
1037 offset = READ_ONCE(table_perturb[index]) + (port_offset >> 32);
1038 offset %= remaining;
1040 /* In first pass we try ports of @low parity.
1041 * inet_csk_get_port() does the opposite choice.
1045 port = low + offset;
1046 for (i = 0; i < remaining; i += 2, port += 2) {
1047 if (unlikely(port >= high))
1049 if (inet_is_local_reserved_port(net, port))
1051 head = &hinfo->bhash[inet_bhashfn(net, port,
1052 hinfo->bhash_size)];
1053 spin_lock_bh(&head->lock);
1055 /* Does not bother with rcv_saddr checks, because
1056 * the established check is already unique enough.
1058 inet_bind_bucket_for_each(tb, &head->chain) {
1059 if (inet_bind_bucket_match(tb, net, port, l3mdev)) {
1060 if (tb->fastreuse >= 0 ||
1061 tb->fastreuseport >= 0)
1063 WARN_ON(hlist_empty(&tb->owners));
1064 if (!check_established(death_row, sk,
1071 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
1072 net, head, port, l3mdev);
1074 spin_unlock_bh(&head->lock);
1079 tb->fastreuseport = -1;
1082 spin_unlock_bh(&head->lock);
1087 if ((offset & 1) && remaining > 1)
1088 goto other_parity_scan;
1090 return -EADDRNOTAVAIL;
1093 /* Find the corresponding tb2 bucket since we need to
1094 * add the socket to the bhash2 table as well
1096 head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
1097 spin_lock(&head2->lock);
1099 tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, sk);
1101 tb2 = inet_bind2_bucket_create(hinfo->bind2_bucket_cachep, net,
1102 head2, port, l3mdev, sk);
1107 /* Here we want to add a little bit of randomness to the next source
1108 * port that will be chosen. We use a max() with a random here so that
1109 * on low contention the randomness is maximal and on high contention
1110 * it may be inexistent.
1112 i = max_t(int, i, get_random_u32_below(8) * 2);
1113 WRITE_ONCE(table_perturb[index], READ_ONCE(table_perturb[index]) + i + 2);
1115 /* Head lock still held and bh's disabled */
1116 inet_bind_hash(sk, tb, tb2, port);
1118 if (sk_unhashed(sk)) {
1119 inet_sk(sk)->inet_sport = htons(port);
1120 inet_ehash_nolisten(sk, (struct sock *)tw, NULL);
1123 inet_twsk_bind_unhash(tw, hinfo);
1125 spin_unlock(&head2->lock);
1126 spin_unlock(&head->lock);
1129 inet_twsk_deschedule_put(tw);
1134 spin_unlock(&head2->lock);
1136 inet_bind_bucket_destroy(hinfo->bind_bucket_cachep, tb);
1137 spin_unlock_bh(&head->lock);
1142 * Bind a port for a connect operation and hash it.
1144 int inet_hash_connect(struct inet_timewait_death_row *death_row,
1147 u64 port_offset = 0;
1149 if (!inet_sk(sk)->inet_num)
1150 port_offset = inet_sk_port_offset(sk);
1151 return __inet_hash_connect(death_row, sk, port_offset,
1152 __inet_check_established);
1154 EXPORT_SYMBOL_GPL(inet_hash_connect);
1156 static void init_hashinfo_lhash2(struct inet_hashinfo *h)
1160 for (i = 0; i <= h->lhash2_mask; i++) {
1161 spin_lock_init(&h->lhash2[i].lock);
1162 INIT_HLIST_NULLS_HEAD(&h->lhash2[i].nulls_head,
1163 i + LISTENING_NULLS_BASE);
1167 void __init inet_hashinfo2_init(struct inet_hashinfo *h, const char *name,
1168 unsigned long numentries, int scale,
1169 unsigned long low_limit,
1170 unsigned long high_limit)
1172 h->lhash2 = alloc_large_system_hash(name,
1181 init_hashinfo_lhash2(h);
1183 /* this one is used for source ports of outgoing connections */
1184 table_perturb = alloc_large_system_hash("Table-perturb",
1185 sizeof(*table_perturb),
1186 INET_TABLE_PERTURB_SIZE,
1188 INET_TABLE_PERTURB_SIZE,
1189 INET_TABLE_PERTURB_SIZE);
1192 int inet_hashinfo2_init_mod(struct inet_hashinfo *h)
1194 h->lhash2 = kmalloc_array(INET_LHTABLE_SIZE, sizeof(*h->lhash2), GFP_KERNEL);
1198 h->lhash2_mask = INET_LHTABLE_SIZE - 1;
1199 /* INET_LHTABLE_SIZE must be a power of 2 */
1200 BUG_ON(INET_LHTABLE_SIZE & h->lhash2_mask);
1202 init_hashinfo_lhash2(h);
1205 EXPORT_SYMBOL_GPL(inet_hashinfo2_init_mod);
1207 int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
1209 unsigned int locksz = sizeof(spinlock_t);
1210 unsigned int i, nblocks = 1;
1213 /* allocate 2 cache lines or at least one spinlock per cpu */
1214 nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U);
1215 nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
1217 /* no more locks than number of hash buckets */
1218 nblocks = min(nblocks, hashinfo->ehash_mask + 1);
1220 hashinfo->ehash_locks = kvmalloc_array(nblocks, locksz, GFP_KERNEL);
1221 if (!hashinfo->ehash_locks)
1224 for (i = 0; i < nblocks; i++)
1225 spin_lock_init(&hashinfo->ehash_locks[i]);
1227 hashinfo->ehash_locks_mask = nblocks - 1;
1230 EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);
1232 struct inet_hashinfo *inet_pernet_hashinfo_alloc(struct inet_hashinfo *hashinfo,
1233 unsigned int ehash_entries)
1235 struct inet_hashinfo *new_hashinfo;
1238 new_hashinfo = kmemdup(hashinfo, sizeof(*hashinfo), GFP_KERNEL);
1242 new_hashinfo->ehash = vmalloc_huge(ehash_entries * sizeof(struct inet_ehash_bucket),
1243 GFP_KERNEL_ACCOUNT);
1244 if (!new_hashinfo->ehash)
1247 new_hashinfo->ehash_mask = ehash_entries - 1;
1249 if (inet_ehash_locks_alloc(new_hashinfo))
1252 for (i = 0; i < ehash_entries; i++)
1253 INIT_HLIST_NULLS_HEAD(&new_hashinfo->ehash[i].chain, i);
1255 new_hashinfo->pernet = true;
1257 return new_hashinfo;
1260 vfree(new_hashinfo->ehash);
1262 kfree(new_hashinfo);
1266 EXPORT_SYMBOL_GPL(inet_pernet_hashinfo_alloc);
1268 void inet_pernet_hashinfo_free(struct inet_hashinfo *hashinfo)
1270 if (!hashinfo->pernet)
1273 inet_ehash_locks_free(hashinfo);
1274 vfree(hashinfo->ehash);
1277 EXPORT_SYMBOL_GPL(inet_pernet_hashinfo_free);