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->bhash2);
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->bhash2)) {
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 *tb2,
105 struct inet_bind_hashbucket *head,
106 struct inet_bind_bucket *tb,
107 const struct sock *sk)
109 write_pnet(&tb2->ib_net, net);
110 tb2->l3mdev = tb->l3mdev;
111 tb2->port = tb->port;
112 #if IS_ENABLED(CONFIG_IPV6)
113 BUILD_BUG_ON(USHRT_MAX < (IPV6_ADDR_ANY | IPV6_ADDR_MAPPED));
114 if (sk->sk_family == AF_INET6) {
115 tb2->addr_type = ipv6_addr_type(&sk->sk_v6_rcv_saddr);
116 tb2->v6_rcv_saddr = sk->sk_v6_rcv_saddr;
118 tb2->addr_type = IPV6_ADDR_MAPPED;
119 ipv6_addr_set_v4mapped(sk->sk_rcv_saddr, &tb2->v6_rcv_saddr);
122 tb2->rcv_saddr = sk->sk_rcv_saddr;
124 INIT_HLIST_HEAD(&tb2->owners);
125 hlist_add_head(&tb2->node, &head->chain);
126 hlist_add_head(&tb2->bhash_node, &tb->bhash2);
129 struct inet_bind2_bucket *inet_bind2_bucket_create(struct kmem_cache *cachep,
131 struct inet_bind_hashbucket *head,
132 struct inet_bind_bucket *tb,
133 const struct sock *sk)
135 struct inet_bind2_bucket *tb2 = kmem_cache_alloc(cachep, GFP_ATOMIC);
138 inet_bind2_bucket_init(tb2, net, head, tb, sk);
143 /* Caller must hold hashbucket lock for this tb with local BH disabled */
144 void inet_bind2_bucket_destroy(struct kmem_cache *cachep, struct inet_bind2_bucket *tb)
146 if (hlist_empty(&tb->owners)) {
147 __hlist_del(&tb->node);
148 __hlist_del(&tb->bhash_node);
149 kmem_cache_free(cachep, tb);
153 static bool inet_bind2_bucket_addr_match(const struct inet_bind2_bucket *tb2,
154 const struct sock *sk)
156 #if IS_ENABLED(CONFIG_IPV6)
157 if (sk->sk_family == AF_INET6)
158 return ipv6_addr_equal(&tb2->v6_rcv_saddr, &sk->sk_v6_rcv_saddr);
160 if (tb2->addr_type != IPV6_ADDR_MAPPED)
163 return tb2->rcv_saddr == sk->sk_rcv_saddr;
166 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
167 struct inet_bind2_bucket *tb2, unsigned short port)
169 inet_sk(sk)->inet_num = port;
170 inet_csk(sk)->icsk_bind_hash = tb;
171 inet_csk(sk)->icsk_bind2_hash = tb2;
172 sk_add_bind_node(sk, &tb2->owners);
176 * Get rid of any references to a local port held by the given sock.
178 static void __inet_put_port(struct sock *sk)
180 struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
181 struct inet_bind_hashbucket *head, *head2;
182 struct net *net = sock_net(sk);
183 struct inet_bind_bucket *tb;
186 bhash = inet_bhashfn(net, inet_sk(sk)->inet_num, hashinfo->bhash_size);
187 head = &hashinfo->bhash[bhash];
188 head2 = inet_bhashfn_portaddr(hashinfo, sk, net, inet_sk(sk)->inet_num);
190 spin_lock(&head->lock);
191 tb = inet_csk(sk)->icsk_bind_hash;
192 inet_csk(sk)->icsk_bind_hash = NULL;
193 inet_sk(sk)->inet_num = 0;
195 spin_lock(&head2->lock);
196 if (inet_csk(sk)->icsk_bind2_hash) {
197 struct inet_bind2_bucket *tb2 = inet_csk(sk)->icsk_bind2_hash;
199 __sk_del_bind_node(sk);
200 inet_csk(sk)->icsk_bind2_hash = NULL;
201 inet_bind2_bucket_destroy(hashinfo->bind2_bucket_cachep, tb2);
203 spin_unlock(&head2->lock);
205 inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
206 spin_unlock(&head->lock);
209 void inet_put_port(struct sock *sk)
215 EXPORT_SYMBOL(inet_put_port);
217 int __inet_inherit_port(const struct sock *sk, struct sock *child)
219 struct inet_hashinfo *table = tcp_or_dccp_get_hashinfo(sk);
220 unsigned short port = inet_sk(child)->inet_num;
221 struct inet_bind_hashbucket *head, *head2;
222 bool created_inet_bind_bucket = false;
223 struct net *net = sock_net(sk);
224 bool update_fastreuse = false;
225 struct inet_bind2_bucket *tb2;
226 struct inet_bind_bucket *tb;
229 bhash = inet_bhashfn(net, port, table->bhash_size);
230 head = &table->bhash[bhash];
231 head2 = inet_bhashfn_portaddr(table, child, net, port);
233 spin_lock(&head->lock);
234 spin_lock(&head2->lock);
235 tb = inet_csk(sk)->icsk_bind_hash;
236 tb2 = inet_csk(sk)->icsk_bind2_hash;
237 if (unlikely(!tb || !tb2)) {
238 spin_unlock(&head2->lock);
239 spin_unlock(&head->lock);
242 if (tb->port != port) {
243 l3mdev = inet_sk_bound_l3mdev(sk);
245 /* NOTE: using tproxy and redirecting skbs to a proxy
246 * on a different listener port breaks the assumption
247 * that the listener socket's icsk_bind_hash is the same
248 * as that of the child socket. We have to look up or
249 * create a new bind bucket for the child here. */
250 inet_bind_bucket_for_each(tb, &head->chain) {
251 if (inet_bind_bucket_match(tb, net, port, l3mdev))
255 tb = inet_bind_bucket_create(table->bind_bucket_cachep,
256 net, head, port, l3mdev);
258 spin_unlock(&head2->lock);
259 spin_unlock(&head->lock);
262 created_inet_bind_bucket = true;
264 update_fastreuse = true;
267 } else if (!inet_bind2_bucket_addr_match(tb2, child)) {
268 l3mdev = inet_sk_bound_l3mdev(sk);
271 tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, child);
273 tb2 = inet_bind2_bucket_create(table->bind2_bucket_cachep,
274 net, head2, tb, child);
279 if (update_fastreuse)
280 inet_csk_update_fastreuse(tb, child);
281 inet_bind_hash(child, tb, tb2, port);
282 spin_unlock(&head2->lock);
283 spin_unlock(&head->lock);
288 if (created_inet_bind_bucket)
289 inet_bind_bucket_destroy(table->bind_bucket_cachep, tb);
290 spin_unlock(&head2->lock);
291 spin_unlock(&head->lock);
294 EXPORT_SYMBOL_GPL(__inet_inherit_port);
296 static struct inet_listen_hashbucket *
297 inet_lhash2_bucket_sk(struct inet_hashinfo *h, struct sock *sk)
301 #if IS_ENABLED(CONFIG_IPV6)
302 if (sk->sk_family == AF_INET6)
303 hash = ipv6_portaddr_hash(sock_net(sk),
304 &sk->sk_v6_rcv_saddr,
305 inet_sk(sk)->inet_num);
308 hash = ipv4_portaddr_hash(sock_net(sk),
309 inet_sk(sk)->inet_rcv_saddr,
310 inet_sk(sk)->inet_num);
311 return inet_lhash2_bucket(h, hash);
314 static inline int compute_score(struct sock *sk, struct net *net,
315 const unsigned short hnum, const __be32 daddr,
316 const int dif, const int sdif)
320 if (net_eq(sock_net(sk), net) && sk->sk_num == hnum &&
321 !ipv6_only_sock(sk)) {
322 if (sk->sk_rcv_saddr != daddr)
325 if (!inet_sk_bound_dev_eq(net, sk->sk_bound_dev_if, dif, sdif))
327 score = sk->sk_bound_dev_if ? 2 : 1;
329 if (sk->sk_family == PF_INET)
331 if (READ_ONCE(sk->sk_incoming_cpu) == raw_smp_processor_id())
338 * inet_lookup_reuseport() - execute reuseport logic on AF_INET socket if necessary.
339 * @net: network namespace.
340 * @sk: AF_INET socket, must be in TCP_LISTEN state for TCP or TCP_CLOSE for UDP.
341 * @skb: context for a potential SK_REUSEPORT program.
342 * @doff: header offset.
343 * @saddr: source address.
344 * @sport: source port.
345 * @daddr: destination address.
346 * @hnum: destination port in host byte order.
347 * @ehashfn: hash function used to generate the fallback hash.
349 * Return: NULL if sk doesn't have SO_REUSEPORT set, otherwise a pointer to
350 * the selected sock or an error.
352 struct sock *inet_lookup_reuseport(struct net *net, struct sock *sk,
353 struct sk_buff *skb, int doff,
354 __be32 saddr, __be16 sport,
355 __be32 daddr, unsigned short hnum,
356 inet_ehashfn_t *ehashfn)
358 struct sock *reuse_sk = NULL;
361 if (sk->sk_reuseport) {
362 phash = INDIRECT_CALL_2(ehashfn, udp_ehashfn, inet_ehashfn,
363 net, daddr, hnum, saddr, sport);
364 reuse_sk = reuseport_select_sock(sk, phash, skb, doff);
368 EXPORT_SYMBOL_GPL(inet_lookup_reuseport);
371 * Here are some nice properties to exploit here. The BSD API
372 * does not allow a listening sock to specify the remote port nor the
373 * remote address for the connection. So always assume those are both
374 * wildcarded during the search since they can never be otherwise.
377 /* called with rcu_read_lock() : No refcount taken on the socket */
378 static struct sock *inet_lhash2_lookup(struct net *net,
379 struct inet_listen_hashbucket *ilb2,
380 struct sk_buff *skb, int doff,
381 const __be32 saddr, __be16 sport,
382 const __be32 daddr, const unsigned short hnum,
383 const int dif, const int sdif)
385 struct sock *sk, *result = NULL;
386 struct hlist_nulls_node *node;
387 int score, hiscore = 0;
389 sk_nulls_for_each_rcu(sk, node, &ilb2->nulls_head) {
390 score = compute_score(sk, net, hnum, daddr, dif, sdif);
391 if (score > hiscore) {
392 result = inet_lookup_reuseport(net, sk, skb, doff,
393 saddr, sport, daddr, hnum, inet_ehashfn);
405 struct sock *inet_lookup_run_sk_lookup(struct net *net,
407 struct sk_buff *skb, int doff,
408 __be32 saddr, __be16 sport,
409 __be32 daddr, u16 hnum, const int dif,
410 inet_ehashfn_t *ehashfn)
412 struct sock *sk, *reuse_sk;
415 no_reuseport = bpf_sk_lookup_run_v4(net, protocol, saddr, sport,
416 daddr, hnum, dif, &sk);
417 if (no_reuseport || IS_ERR_OR_NULL(sk))
420 reuse_sk = inet_lookup_reuseport(net, sk, skb, doff, saddr, sport, daddr, hnum,
427 struct sock *__inet_lookup_listener(struct net *net,
428 struct inet_hashinfo *hashinfo,
429 struct sk_buff *skb, int doff,
430 const __be32 saddr, __be16 sport,
431 const __be32 daddr, const unsigned short hnum,
432 const int dif, const int sdif)
434 struct inet_listen_hashbucket *ilb2;
435 struct sock *result = NULL;
438 /* Lookup redirect from BPF */
439 if (static_branch_unlikely(&bpf_sk_lookup_enabled) &&
440 hashinfo == net->ipv4.tcp_death_row.hashinfo) {
441 result = inet_lookup_run_sk_lookup(net, IPPROTO_TCP, skb, doff,
442 saddr, sport, daddr, hnum, dif,
448 hash2 = ipv4_portaddr_hash(net, daddr, hnum);
449 ilb2 = inet_lhash2_bucket(hashinfo, hash2);
451 result = inet_lhash2_lookup(net, ilb2, skb, doff,
452 saddr, sport, daddr, hnum,
457 /* Lookup lhash2 with INADDR_ANY */
458 hash2 = ipv4_portaddr_hash(net, htonl(INADDR_ANY), hnum);
459 ilb2 = inet_lhash2_bucket(hashinfo, hash2);
461 result = inet_lhash2_lookup(net, ilb2, skb, doff,
462 saddr, sport, htonl(INADDR_ANY), hnum,
469 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
471 /* All sockets share common refcount, but have different destructors */
472 void sock_gen_put(struct sock *sk)
474 if (!refcount_dec_and_test(&sk->sk_refcnt))
477 if (sk->sk_state == TCP_TIME_WAIT)
478 inet_twsk_free(inet_twsk(sk));
479 else if (sk->sk_state == TCP_NEW_SYN_RECV)
480 reqsk_free(inet_reqsk(sk));
484 EXPORT_SYMBOL_GPL(sock_gen_put);
486 void sock_edemux(struct sk_buff *skb)
488 sock_gen_put(skb->sk);
490 EXPORT_SYMBOL(sock_edemux);
492 struct sock *__inet_lookup_established(struct net *net,
493 struct inet_hashinfo *hashinfo,
494 const __be32 saddr, const __be16 sport,
495 const __be32 daddr, const u16 hnum,
496 const int dif, const int sdif)
498 INET_ADDR_COOKIE(acookie, saddr, daddr);
499 const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
501 const struct hlist_nulls_node *node;
502 /* Optimize here for direct hit, only listening connections can
503 * have wildcards anyways.
505 unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
506 unsigned int slot = hash & hashinfo->ehash_mask;
507 struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
510 sk_nulls_for_each_rcu(sk, node, &head->chain) {
511 if (sk->sk_hash != hash)
513 if (likely(inet_match(net, sk, acookie, ports, dif, sdif))) {
514 if (unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
516 if (unlikely(!inet_match(net, sk, acookie,
517 ports, dif, sdif))) {
525 * if the nulls value we got at the end of this lookup is
526 * not the expected one, we must restart lookup.
527 * We probably met an item that was moved to another chain.
529 if (get_nulls_value(node) != slot)
536 EXPORT_SYMBOL_GPL(__inet_lookup_established);
538 /* called with local bh disabled */
539 static int __inet_check_established(struct inet_timewait_death_row *death_row,
540 struct sock *sk, __u16 lport,
541 struct inet_timewait_sock **twp)
543 struct inet_hashinfo *hinfo = death_row->hashinfo;
544 struct inet_sock *inet = inet_sk(sk);
545 __be32 daddr = inet->inet_rcv_saddr;
546 __be32 saddr = inet->inet_daddr;
547 int dif = sk->sk_bound_dev_if;
548 struct net *net = sock_net(sk);
549 int sdif = l3mdev_master_ifindex_by_index(net, dif);
550 INET_ADDR_COOKIE(acookie, saddr, daddr);
551 const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
552 unsigned int hash = inet_ehashfn(net, daddr, lport,
553 saddr, inet->inet_dport);
554 struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
555 spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
557 const struct hlist_nulls_node *node;
558 struct inet_timewait_sock *tw = NULL;
562 sk_nulls_for_each(sk2, node, &head->chain) {
563 if (sk2->sk_hash != hash)
566 if (likely(inet_match(net, sk2, acookie, ports, dif, sdif))) {
567 if (sk2->sk_state == TCP_TIME_WAIT) {
569 if (twsk_unique(sk, sk2, twp))
576 /* Must record num and sport now. Otherwise we will see
577 * in hash table socket with a funny identity.
579 inet->inet_num = lport;
580 inet->inet_sport = htons(lport);
582 WARN_ON(!sk_unhashed(sk));
583 __sk_nulls_add_node_rcu(sk, &head->chain);
585 sk_nulls_del_node_init_rcu((struct sock *)tw);
586 __NET_INC_STATS(net, LINUX_MIB_TIMEWAITRECYCLED);
589 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
594 /* Silly. Should hash-dance instead... */
595 inet_twsk_deschedule_put(tw);
601 return -EADDRNOTAVAIL;
604 static u64 inet_sk_port_offset(const struct sock *sk)
606 const struct inet_sock *inet = inet_sk(sk);
608 return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
613 /* Searches for an exsiting socket in the ehash bucket list.
614 * Returns true if found, false otherwise.
616 static bool inet_ehash_lookup_by_sk(struct sock *sk,
617 struct hlist_nulls_head *list)
619 const __portpair ports = INET_COMBINED_PORTS(sk->sk_dport, sk->sk_num);
620 const int sdif = sk->sk_bound_dev_if;
621 const int dif = sk->sk_bound_dev_if;
622 const struct hlist_nulls_node *node;
623 struct net *net = sock_net(sk);
626 INET_ADDR_COOKIE(acookie, sk->sk_daddr, sk->sk_rcv_saddr);
628 sk_nulls_for_each_rcu(esk, node, list) {
629 if (esk->sk_hash != sk->sk_hash)
631 if (sk->sk_family == AF_INET) {
632 if (unlikely(inet_match(net, esk, acookie,
633 ports, dif, sdif))) {
637 #if IS_ENABLED(CONFIG_IPV6)
638 else if (sk->sk_family == AF_INET6) {
639 if (unlikely(inet6_match(net, esk,
641 &sk->sk_v6_rcv_saddr,
642 ports, dif, sdif))) {
651 /* Insert a socket into ehash, and eventually remove another one
652 * (The another one can be a SYN_RECV or TIMEWAIT)
653 * If an existing socket already exists, socket sk is not inserted,
654 * and sets found_dup_sk parameter to true.
656 bool inet_ehash_insert(struct sock *sk, struct sock *osk, bool *found_dup_sk)
658 struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
659 struct inet_ehash_bucket *head;
660 struct hlist_nulls_head *list;
664 WARN_ON_ONCE(!sk_unhashed(sk));
666 sk->sk_hash = sk_ehashfn(sk);
667 head = inet_ehash_bucket(hashinfo, sk->sk_hash);
669 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
673 WARN_ON_ONCE(sk->sk_hash != osk->sk_hash);
674 ret = sk_nulls_del_node_init_rcu(osk);
675 } else if (found_dup_sk) {
676 *found_dup_sk = inet_ehash_lookup_by_sk(sk, list);
682 __sk_nulls_add_node_rcu(sk, list);
689 bool inet_ehash_nolisten(struct sock *sk, struct sock *osk, bool *found_dup_sk)
691 bool ok = inet_ehash_insert(sk, osk, found_dup_sk);
694 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
696 this_cpu_inc(*sk->sk_prot->orphan_count);
697 inet_sk_set_state(sk, TCP_CLOSE);
698 sock_set_flag(sk, SOCK_DEAD);
699 inet_csk_destroy_sock(sk);
703 EXPORT_SYMBOL_GPL(inet_ehash_nolisten);
705 static int inet_reuseport_add_sock(struct sock *sk,
706 struct inet_listen_hashbucket *ilb)
708 struct inet_bind_bucket *tb = inet_csk(sk)->icsk_bind_hash;
709 const struct hlist_nulls_node *node;
711 kuid_t uid = sock_i_uid(sk);
713 sk_nulls_for_each_rcu(sk2, node, &ilb->nulls_head) {
715 sk2->sk_family == sk->sk_family &&
716 ipv6_only_sock(sk2) == ipv6_only_sock(sk) &&
717 sk2->sk_bound_dev_if == sk->sk_bound_dev_if &&
718 inet_csk(sk2)->icsk_bind_hash == tb &&
719 sk2->sk_reuseport && uid_eq(uid, sock_i_uid(sk2)) &&
720 inet_rcv_saddr_equal(sk, sk2, false))
721 return reuseport_add_sock(sk, sk2,
722 inet_rcv_saddr_any(sk));
725 return reuseport_alloc(sk, inet_rcv_saddr_any(sk));
728 int __inet_hash(struct sock *sk, struct sock *osk)
730 struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
731 struct inet_listen_hashbucket *ilb2;
734 if (sk->sk_state != TCP_LISTEN) {
736 inet_ehash_nolisten(sk, osk, NULL);
740 WARN_ON(!sk_unhashed(sk));
741 ilb2 = inet_lhash2_bucket_sk(hashinfo, sk);
743 spin_lock(&ilb2->lock);
744 if (sk->sk_reuseport) {
745 err = inet_reuseport_add_sock(sk, ilb2);
749 sock_set_flag(sk, SOCK_RCU_FREE);
750 if (IS_ENABLED(CONFIG_IPV6) && sk->sk_reuseport &&
751 sk->sk_family == AF_INET6)
752 __sk_nulls_add_node_tail_rcu(sk, &ilb2->nulls_head);
754 __sk_nulls_add_node_rcu(sk, &ilb2->nulls_head);
755 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
757 spin_unlock(&ilb2->lock);
761 EXPORT_SYMBOL(__inet_hash);
763 int inet_hash(struct sock *sk)
767 if (sk->sk_state != TCP_CLOSE)
768 err = __inet_hash(sk, NULL);
772 EXPORT_SYMBOL_GPL(inet_hash);
774 void inet_unhash(struct sock *sk)
776 struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
781 if (sk->sk_state == TCP_LISTEN) {
782 struct inet_listen_hashbucket *ilb2;
784 ilb2 = inet_lhash2_bucket_sk(hashinfo, sk);
785 /* Don't disable bottom halves while acquiring the lock to
786 * avoid circular locking dependency on PREEMPT_RT.
788 spin_lock(&ilb2->lock);
789 if (sk_unhashed(sk)) {
790 spin_unlock(&ilb2->lock);
794 if (rcu_access_pointer(sk->sk_reuseport_cb))
795 reuseport_stop_listen_sock(sk);
797 __sk_nulls_del_node_init_rcu(sk);
798 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
799 spin_unlock(&ilb2->lock);
801 spinlock_t *lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
804 if (sk_unhashed(sk)) {
805 spin_unlock_bh(lock);
808 __sk_nulls_del_node_init_rcu(sk);
809 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
810 spin_unlock_bh(lock);
813 EXPORT_SYMBOL_GPL(inet_unhash);
815 static bool inet_bind2_bucket_match(const struct inet_bind2_bucket *tb,
816 const struct net *net, unsigned short port,
817 int l3mdev, const struct sock *sk)
819 if (!net_eq(ib2_net(tb), net) || tb->port != port ||
820 tb->l3mdev != l3mdev)
823 return inet_bind2_bucket_addr_match(tb, sk);
826 bool inet_bind2_bucket_match_addr_any(const struct inet_bind2_bucket *tb, const struct net *net,
827 unsigned short port, int l3mdev, const struct sock *sk)
829 if (!net_eq(ib2_net(tb), net) || tb->port != port ||
830 tb->l3mdev != l3mdev)
833 #if IS_ENABLED(CONFIG_IPV6)
834 if (tb->addr_type == IPV6_ADDR_ANY)
837 if (tb->addr_type != IPV6_ADDR_MAPPED)
840 if (sk->sk_family == AF_INET6 &&
841 !ipv6_addr_v4mapped(&sk->sk_v6_rcv_saddr))
844 return tb->rcv_saddr == 0;
847 /* The socket's bhash2 hashbucket spinlock must be held when this is called */
848 struct inet_bind2_bucket *
849 inet_bind2_bucket_find(const struct inet_bind_hashbucket *head, const struct net *net,
850 unsigned short port, int l3mdev, const struct sock *sk)
852 struct inet_bind2_bucket *bhash2 = NULL;
854 inet_bind_bucket_for_each(bhash2, &head->chain)
855 if (inet_bind2_bucket_match(bhash2, net, port, l3mdev, sk))
861 struct inet_bind_hashbucket *
862 inet_bhash2_addr_any_hashbucket(const struct sock *sk, const struct net *net, int port)
864 struct inet_hashinfo *hinfo = tcp_or_dccp_get_hashinfo(sk);
867 #if IS_ENABLED(CONFIG_IPV6)
868 if (sk->sk_family == AF_INET6)
869 hash = ipv6_portaddr_hash(net, &in6addr_any, port);
872 hash = ipv4_portaddr_hash(net, 0, port);
874 return &hinfo->bhash2[hash & (hinfo->bhash_size - 1)];
877 static void inet_update_saddr(struct sock *sk, void *saddr, int family)
879 if (family == AF_INET) {
880 inet_sk(sk)->inet_saddr = *(__be32 *)saddr;
881 sk_rcv_saddr_set(sk, inet_sk(sk)->inet_saddr);
883 #if IS_ENABLED(CONFIG_IPV6)
885 sk->sk_v6_rcv_saddr = *(struct in6_addr *)saddr;
890 static int __inet_bhash2_update_saddr(struct sock *sk, void *saddr, int family, bool reset)
892 struct inet_hashinfo *hinfo = tcp_or_dccp_get_hashinfo(sk);
893 struct inet_bind_hashbucket *head, *head2;
894 struct inet_bind2_bucket *tb2, *new_tb2;
895 int l3mdev = inet_sk_bound_l3mdev(sk);
896 int port = inet_sk(sk)->inet_num;
897 struct net *net = sock_net(sk);
900 if (!inet_csk(sk)->icsk_bind2_hash) {
901 /* Not bind()ed before. */
903 inet_reset_saddr(sk);
905 inet_update_saddr(sk, saddr, family);
910 /* Allocate a bind2 bucket ahead of time to avoid permanently putting
911 * the bhash2 table in an inconsistent state if a new tb2 bucket
914 new_tb2 = kmem_cache_alloc(hinfo->bind2_bucket_cachep, GFP_ATOMIC);
917 /* The (INADDR_ANY, port) bucket might have already
918 * been freed, then we cannot fixup icsk_bind2_hash,
919 * so we give up and unlink sk from bhash/bhash2 not
920 * to leave inconsistency in bhash2.
923 inet_reset_saddr(sk);
929 bhash = inet_bhashfn(net, port, hinfo->bhash_size);
930 head = &hinfo->bhash[bhash];
931 head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
933 /* If we change saddr locklessly, another thread
934 * iterating over bhash might see corrupted address.
936 spin_lock_bh(&head->lock);
938 spin_lock(&head2->lock);
939 __sk_del_bind_node(sk);
940 inet_bind2_bucket_destroy(hinfo->bind2_bucket_cachep, inet_csk(sk)->icsk_bind2_hash);
941 spin_unlock(&head2->lock);
944 inet_reset_saddr(sk);
946 inet_update_saddr(sk, saddr, family);
948 head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
950 spin_lock(&head2->lock);
951 tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, sk);
954 inet_bind2_bucket_init(tb2, net, head2, inet_csk(sk)->icsk_bind_hash, sk);
956 inet_csk(sk)->icsk_bind2_hash = tb2;
957 sk_add_bind_node(sk, &tb2->owners);
958 spin_unlock(&head2->lock);
960 spin_unlock_bh(&head->lock);
963 kmem_cache_free(hinfo->bind2_bucket_cachep, new_tb2);
968 int inet_bhash2_update_saddr(struct sock *sk, void *saddr, int family)
970 return __inet_bhash2_update_saddr(sk, saddr, family, false);
972 EXPORT_SYMBOL_GPL(inet_bhash2_update_saddr);
974 void inet_bhash2_reset_saddr(struct sock *sk)
976 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
977 __inet_bhash2_update_saddr(sk, NULL, 0, true);
979 EXPORT_SYMBOL_GPL(inet_bhash2_reset_saddr);
981 /* RFC 6056 3.3.4. Algorithm 4: Double-Hash Port Selection Algorithm
982 * Note that we use 32bit integers (vs RFC 'short integers')
983 * because 2^16 is not a multiple of num_ephemeral and this
984 * property might be used by clever attacker.
986 * RFC claims using TABLE_LENGTH=10 buckets gives an improvement, though
987 * attacks were since demonstrated, thus we use 65536 by default instead
988 * to really give more isolation and privacy, at the expense of 256kB
991 #define INET_TABLE_PERTURB_SIZE (1 << CONFIG_INET_TABLE_PERTURB_ORDER)
992 static u32 *table_perturb;
994 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
995 struct sock *sk, u64 port_offset,
996 int (*check_established)(struct inet_timewait_death_row *,
997 struct sock *, __u16, struct inet_timewait_sock **))
999 struct inet_hashinfo *hinfo = death_row->hashinfo;
1000 struct inet_bind_hashbucket *head, *head2;
1001 struct inet_timewait_sock *tw = NULL;
1002 int port = inet_sk(sk)->inet_num;
1003 struct net *net = sock_net(sk);
1004 struct inet_bind2_bucket *tb2;
1005 struct inet_bind_bucket *tb;
1006 bool tb_created = false;
1007 u32 remaining, offset;
1008 int ret, i, low, high;
1015 ret = check_established(death_row, sk, port, NULL);
1020 l3mdev = inet_sk_bound_l3mdev(sk);
1022 local_ports = inet_sk_get_local_port_range(sk, &low, &high);
1023 step = local_ports ? 1 : 2;
1025 high++; /* [32768, 60999] -> [32768, 61000[ */
1026 remaining = high - low;
1027 if (!local_ports && remaining > 1)
1030 get_random_sleepable_once(table_perturb,
1031 INET_TABLE_PERTURB_SIZE * sizeof(*table_perturb));
1032 index = port_offset & (INET_TABLE_PERTURB_SIZE - 1);
1034 offset = READ_ONCE(table_perturb[index]) + (port_offset >> 32);
1035 offset %= remaining;
1037 /* In first pass we try ports of @low parity.
1038 * inet_csk_get_port() does the opposite choice.
1043 port = low + offset;
1044 for (i = 0; i < remaining; i += step, port += step) {
1045 if (unlikely(port >= high))
1047 if (inet_is_local_reserved_port(net, port))
1049 head = &hinfo->bhash[inet_bhashfn(net, port,
1050 hinfo->bhash_size)];
1051 spin_lock_bh(&head->lock);
1053 /* Does not bother with rcv_saddr checks, because
1054 * the established check is already unique enough.
1056 inet_bind_bucket_for_each(tb, &head->chain) {
1057 if (inet_bind_bucket_match(tb, net, port, l3mdev)) {
1058 if (tb->fastreuse >= 0 ||
1059 tb->fastreuseport >= 0)
1061 WARN_ON(hlist_empty(&tb->bhash2));
1062 if (!check_established(death_row, sk,
1069 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
1070 net, head, port, l3mdev);
1072 spin_unlock_bh(&head->lock);
1077 tb->fastreuseport = -1;
1080 spin_unlock_bh(&head->lock);
1086 if ((offset & 1) && remaining > 1)
1087 goto other_parity_scan;
1089 return -EADDRNOTAVAIL;
1092 /* Find the corresponding tb2 bucket since we need to
1093 * add the socket to the bhash2 table as well
1095 head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
1096 spin_lock(&head2->lock);
1098 tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, sk);
1100 tb2 = inet_bind2_bucket_create(hinfo->bind2_bucket_cachep, net,
1106 /* Here we want to add a little bit of randomness to the next source
1107 * port that will be chosen. We use a max() with a random here so that
1108 * on low contention the randomness is maximal and on high contention
1109 * it may be inexistent.
1111 i = max_t(int, i, get_random_u32_below(8) * step);
1112 WRITE_ONCE(table_perturb[index], READ_ONCE(table_perturb[index]) + i + step);
1114 /* Head lock still held and bh's disabled */
1115 inet_bind_hash(sk, tb, tb2, port);
1117 if (sk_unhashed(sk)) {
1118 inet_sk(sk)->inet_sport = htons(port);
1119 inet_ehash_nolisten(sk, (struct sock *)tw, NULL);
1122 inet_twsk_bind_unhash(tw, hinfo);
1124 spin_unlock(&head2->lock);
1125 spin_unlock(&head->lock);
1128 inet_twsk_deschedule_put(tw);
1133 if (sk_hashed(sk)) {
1134 spinlock_t *lock = inet_ehash_lockp(hinfo, sk->sk_hash);
1136 sock_prot_inuse_add(net, sk->sk_prot, -1);
1139 __sk_nulls_del_node_init_rcu(sk);
1143 inet_sk(sk)->inet_sport = 0;
1144 inet_sk(sk)->inet_num = 0;
1147 inet_twsk_bind_unhash(tw, hinfo);
1150 spin_unlock(&head2->lock);
1152 inet_bind_bucket_destroy(hinfo->bind_bucket_cachep, tb);
1153 spin_unlock(&head->lock);
1156 inet_twsk_deschedule_put(tw);
1164 * Bind a port for a connect operation and hash it.
1166 int inet_hash_connect(struct inet_timewait_death_row *death_row,
1169 u64 port_offset = 0;
1171 if (!inet_sk(sk)->inet_num)
1172 port_offset = inet_sk_port_offset(sk);
1173 return __inet_hash_connect(death_row, sk, port_offset,
1174 __inet_check_established);
1176 EXPORT_SYMBOL_GPL(inet_hash_connect);
1178 static void init_hashinfo_lhash2(struct inet_hashinfo *h)
1182 for (i = 0; i <= h->lhash2_mask; i++) {
1183 spin_lock_init(&h->lhash2[i].lock);
1184 INIT_HLIST_NULLS_HEAD(&h->lhash2[i].nulls_head,
1185 i + LISTENING_NULLS_BASE);
1189 void __init inet_hashinfo2_init(struct inet_hashinfo *h, const char *name,
1190 unsigned long numentries, int scale,
1191 unsigned long low_limit,
1192 unsigned long high_limit)
1194 h->lhash2 = alloc_large_system_hash(name,
1203 init_hashinfo_lhash2(h);
1205 /* this one is used for source ports of outgoing connections */
1206 table_perturb = alloc_large_system_hash("Table-perturb",
1207 sizeof(*table_perturb),
1208 INET_TABLE_PERTURB_SIZE,
1210 INET_TABLE_PERTURB_SIZE,
1211 INET_TABLE_PERTURB_SIZE);
1214 int inet_hashinfo2_init_mod(struct inet_hashinfo *h)
1216 h->lhash2 = kmalloc_array(INET_LHTABLE_SIZE, sizeof(*h->lhash2), GFP_KERNEL);
1220 h->lhash2_mask = INET_LHTABLE_SIZE - 1;
1221 /* INET_LHTABLE_SIZE must be a power of 2 */
1222 BUG_ON(INET_LHTABLE_SIZE & h->lhash2_mask);
1224 init_hashinfo_lhash2(h);
1227 EXPORT_SYMBOL_GPL(inet_hashinfo2_init_mod);
1229 int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
1231 unsigned int locksz = sizeof(spinlock_t);
1232 unsigned int i, nblocks = 1;
1235 /* allocate 2 cache lines or at least one spinlock per cpu */
1236 nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U);
1237 nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
1239 /* no more locks than number of hash buckets */
1240 nblocks = min(nblocks, hashinfo->ehash_mask + 1);
1242 hashinfo->ehash_locks = kvmalloc_array(nblocks, locksz, GFP_KERNEL);
1243 if (!hashinfo->ehash_locks)
1246 for (i = 0; i < nblocks; i++)
1247 spin_lock_init(&hashinfo->ehash_locks[i]);
1249 hashinfo->ehash_locks_mask = nblocks - 1;
1252 EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);
1254 struct inet_hashinfo *inet_pernet_hashinfo_alloc(struct inet_hashinfo *hashinfo,
1255 unsigned int ehash_entries)
1257 struct inet_hashinfo *new_hashinfo;
1260 new_hashinfo = kmemdup(hashinfo, sizeof(*hashinfo), GFP_KERNEL);
1264 new_hashinfo->ehash = vmalloc_huge(ehash_entries * sizeof(struct inet_ehash_bucket),
1265 GFP_KERNEL_ACCOUNT);
1266 if (!new_hashinfo->ehash)
1269 new_hashinfo->ehash_mask = ehash_entries - 1;
1271 if (inet_ehash_locks_alloc(new_hashinfo))
1274 for (i = 0; i < ehash_entries; i++)
1275 INIT_HLIST_NULLS_HEAD(&new_hashinfo->ehash[i].chain, i);
1277 new_hashinfo->pernet = true;
1279 return new_hashinfo;
1282 vfree(new_hashinfo->ehash);
1284 kfree(new_hashinfo);
1288 EXPORT_SYMBOL_GPL(inet_pernet_hashinfo_alloc);
1290 void inet_pernet_hashinfo_free(struct inet_hashinfo *hashinfo)
1292 if (!hashinfo->pernet)
1295 inet_ehash_locks_free(hashinfo);
1296 vfree(hashinfo->ehash);
1299 EXPORT_SYMBOL_GPL(inet_pernet_hashinfo_free);