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())
336 INDIRECT_CALLABLE_DECLARE(inet_ehashfn_t udp_ehashfn);
338 struct sock *inet_lookup_reuseport(struct net *net, struct sock *sk,
339 struct sk_buff *skb, int doff,
340 __be32 saddr, __be16 sport,
341 __be32 daddr, unsigned short hnum,
342 inet_ehashfn_t *ehashfn)
344 struct sock *reuse_sk = NULL;
347 if (sk->sk_reuseport) {
348 phash = INDIRECT_CALL_2(ehashfn, udp_ehashfn, inet_ehashfn,
349 net, daddr, hnum, saddr, sport);
350 reuse_sk = reuseport_select_sock(sk, phash, skb, doff);
354 EXPORT_SYMBOL_GPL(inet_lookup_reuseport);
357 * Here are some nice properties to exploit here. The BSD API
358 * does not allow a listening sock to specify the remote port nor the
359 * remote address for the connection. So always assume those are both
360 * wildcarded during the search since they can never be otherwise.
363 /* called with rcu_read_lock() : No refcount taken on the socket */
364 static struct sock *inet_lhash2_lookup(struct net *net,
365 struct inet_listen_hashbucket *ilb2,
366 struct sk_buff *skb, int doff,
367 const __be32 saddr, __be16 sport,
368 const __be32 daddr, const unsigned short hnum,
369 const int dif, const int sdif)
371 struct sock *sk, *result = NULL;
372 struct hlist_nulls_node *node;
373 int score, hiscore = 0;
375 sk_nulls_for_each_rcu(sk, node, &ilb2->nulls_head) {
376 score = compute_score(sk, net, hnum, daddr, dif, sdif);
377 if (score > hiscore) {
378 result = inet_lookup_reuseport(net, sk, skb, doff,
379 saddr, sport, daddr, hnum, inet_ehashfn);
391 static inline struct sock *inet_lookup_run_bpf(struct net *net,
392 struct inet_hashinfo *hashinfo,
393 struct sk_buff *skb, int doff,
394 __be32 saddr, __be16 sport,
395 __be32 daddr, u16 hnum, const int dif)
397 struct sock *sk, *reuse_sk;
400 if (hashinfo != net->ipv4.tcp_death_row.hashinfo)
401 return NULL; /* only TCP is supported */
403 no_reuseport = bpf_sk_lookup_run_v4(net, IPPROTO_TCP, saddr, sport,
404 daddr, hnum, dif, &sk);
405 if (no_reuseport || IS_ERR_OR_NULL(sk))
408 reuse_sk = inet_lookup_reuseport(net, sk, skb, doff, saddr, sport, daddr, hnum,
415 struct sock *__inet_lookup_listener(struct net *net,
416 struct inet_hashinfo *hashinfo,
417 struct sk_buff *skb, int doff,
418 const __be32 saddr, __be16 sport,
419 const __be32 daddr, const unsigned short hnum,
420 const int dif, const int sdif)
422 struct inet_listen_hashbucket *ilb2;
423 struct sock *result = NULL;
426 /* Lookup redirect from BPF */
427 if (static_branch_unlikely(&bpf_sk_lookup_enabled)) {
428 result = inet_lookup_run_bpf(net, hashinfo, skb, doff,
429 saddr, sport, daddr, hnum, dif);
434 hash2 = ipv4_portaddr_hash(net, daddr, hnum);
435 ilb2 = inet_lhash2_bucket(hashinfo, hash2);
437 result = inet_lhash2_lookup(net, ilb2, skb, doff,
438 saddr, sport, daddr, hnum,
443 /* Lookup lhash2 with INADDR_ANY */
444 hash2 = ipv4_portaddr_hash(net, htonl(INADDR_ANY), hnum);
445 ilb2 = inet_lhash2_bucket(hashinfo, hash2);
447 result = inet_lhash2_lookup(net, ilb2, skb, doff,
448 saddr, sport, htonl(INADDR_ANY), hnum,
455 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
457 /* All sockets share common refcount, but have different destructors */
458 void sock_gen_put(struct sock *sk)
460 if (!refcount_dec_and_test(&sk->sk_refcnt))
463 if (sk->sk_state == TCP_TIME_WAIT)
464 inet_twsk_free(inet_twsk(sk));
465 else if (sk->sk_state == TCP_NEW_SYN_RECV)
466 reqsk_free(inet_reqsk(sk));
470 EXPORT_SYMBOL_GPL(sock_gen_put);
472 void sock_edemux(struct sk_buff *skb)
474 sock_gen_put(skb->sk);
476 EXPORT_SYMBOL(sock_edemux);
478 struct sock *__inet_lookup_established(struct net *net,
479 struct inet_hashinfo *hashinfo,
480 const __be32 saddr, const __be16 sport,
481 const __be32 daddr, const u16 hnum,
482 const int dif, const int sdif)
484 INET_ADDR_COOKIE(acookie, saddr, daddr);
485 const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
487 const struct hlist_nulls_node *node;
488 /* Optimize here for direct hit, only listening connections can
489 * have wildcards anyways.
491 unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
492 unsigned int slot = hash & hashinfo->ehash_mask;
493 struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
496 sk_nulls_for_each_rcu(sk, node, &head->chain) {
497 if (sk->sk_hash != hash)
499 if (likely(inet_match(net, sk, acookie, ports, dif, sdif))) {
500 if (unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
502 if (unlikely(!inet_match(net, sk, acookie,
503 ports, dif, sdif))) {
511 * if the nulls value we got at the end of this lookup is
512 * not the expected one, we must restart lookup.
513 * We probably met an item that was moved to another chain.
515 if (get_nulls_value(node) != slot)
522 EXPORT_SYMBOL_GPL(__inet_lookup_established);
524 /* called with local bh disabled */
525 static int __inet_check_established(struct inet_timewait_death_row *death_row,
526 struct sock *sk, __u16 lport,
527 struct inet_timewait_sock **twp)
529 struct inet_hashinfo *hinfo = death_row->hashinfo;
530 struct inet_sock *inet = inet_sk(sk);
531 __be32 daddr = inet->inet_rcv_saddr;
532 __be32 saddr = inet->inet_daddr;
533 int dif = sk->sk_bound_dev_if;
534 struct net *net = sock_net(sk);
535 int sdif = l3mdev_master_ifindex_by_index(net, dif);
536 INET_ADDR_COOKIE(acookie, saddr, daddr);
537 const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
538 unsigned int hash = inet_ehashfn(net, daddr, lport,
539 saddr, inet->inet_dport);
540 struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
541 spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
543 const struct hlist_nulls_node *node;
544 struct inet_timewait_sock *tw = NULL;
548 sk_nulls_for_each(sk2, node, &head->chain) {
549 if (sk2->sk_hash != hash)
552 if (likely(inet_match(net, sk2, acookie, ports, dif, sdif))) {
553 if (sk2->sk_state == TCP_TIME_WAIT) {
555 if (twsk_unique(sk, sk2, twp))
562 /* Must record num and sport now. Otherwise we will see
563 * in hash table socket with a funny identity.
565 inet->inet_num = lport;
566 inet->inet_sport = htons(lport);
568 WARN_ON(!sk_unhashed(sk));
569 __sk_nulls_add_node_rcu(sk, &head->chain);
571 sk_nulls_del_node_init_rcu((struct sock *)tw);
572 __NET_INC_STATS(net, LINUX_MIB_TIMEWAITRECYCLED);
575 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
580 /* Silly. Should hash-dance instead... */
581 inet_twsk_deschedule_put(tw);
587 return -EADDRNOTAVAIL;
590 static u64 inet_sk_port_offset(const struct sock *sk)
592 const struct inet_sock *inet = inet_sk(sk);
594 return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
599 /* Searches for an exsiting socket in the ehash bucket list.
600 * Returns true if found, false otherwise.
602 static bool inet_ehash_lookup_by_sk(struct sock *sk,
603 struct hlist_nulls_head *list)
605 const __portpair ports = INET_COMBINED_PORTS(sk->sk_dport, sk->sk_num);
606 const int sdif = sk->sk_bound_dev_if;
607 const int dif = sk->sk_bound_dev_if;
608 const struct hlist_nulls_node *node;
609 struct net *net = sock_net(sk);
612 INET_ADDR_COOKIE(acookie, sk->sk_daddr, sk->sk_rcv_saddr);
614 sk_nulls_for_each_rcu(esk, node, list) {
615 if (esk->sk_hash != sk->sk_hash)
617 if (sk->sk_family == AF_INET) {
618 if (unlikely(inet_match(net, esk, acookie,
619 ports, dif, sdif))) {
623 #if IS_ENABLED(CONFIG_IPV6)
624 else if (sk->sk_family == AF_INET6) {
625 if (unlikely(inet6_match(net, esk,
627 &sk->sk_v6_rcv_saddr,
628 ports, dif, sdif))) {
637 /* Insert a socket into ehash, and eventually remove another one
638 * (The another one can be a SYN_RECV or TIMEWAIT)
639 * If an existing socket already exists, socket sk is not inserted,
640 * and sets found_dup_sk parameter to true.
642 bool inet_ehash_insert(struct sock *sk, struct sock *osk, bool *found_dup_sk)
644 struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
645 struct inet_ehash_bucket *head;
646 struct hlist_nulls_head *list;
650 WARN_ON_ONCE(!sk_unhashed(sk));
652 sk->sk_hash = sk_ehashfn(sk);
653 head = inet_ehash_bucket(hashinfo, sk->sk_hash);
655 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
659 WARN_ON_ONCE(sk->sk_hash != osk->sk_hash);
660 ret = sk_nulls_del_node_init_rcu(osk);
661 } else if (found_dup_sk) {
662 *found_dup_sk = inet_ehash_lookup_by_sk(sk, list);
668 __sk_nulls_add_node_rcu(sk, list);
675 bool inet_ehash_nolisten(struct sock *sk, struct sock *osk, bool *found_dup_sk)
677 bool ok = inet_ehash_insert(sk, osk, found_dup_sk);
680 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
682 this_cpu_inc(*sk->sk_prot->orphan_count);
683 inet_sk_set_state(sk, TCP_CLOSE);
684 sock_set_flag(sk, SOCK_DEAD);
685 inet_csk_destroy_sock(sk);
689 EXPORT_SYMBOL_GPL(inet_ehash_nolisten);
691 static int inet_reuseport_add_sock(struct sock *sk,
692 struct inet_listen_hashbucket *ilb)
694 struct inet_bind_bucket *tb = inet_csk(sk)->icsk_bind_hash;
695 const struct hlist_nulls_node *node;
697 kuid_t uid = sock_i_uid(sk);
699 sk_nulls_for_each_rcu(sk2, node, &ilb->nulls_head) {
701 sk2->sk_family == sk->sk_family &&
702 ipv6_only_sock(sk2) == ipv6_only_sock(sk) &&
703 sk2->sk_bound_dev_if == sk->sk_bound_dev_if &&
704 inet_csk(sk2)->icsk_bind_hash == tb &&
705 sk2->sk_reuseport && uid_eq(uid, sock_i_uid(sk2)) &&
706 inet_rcv_saddr_equal(sk, sk2, false))
707 return reuseport_add_sock(sk, sk2,
708 inet_rcv_saddr_any(sk));
711 return reuseport_alloc(sk, inet_rcv_saddr_any(sk));
714 int __inet_hash(struct sock *sk, struct sock *osk)
716 struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
717 struct inet_listen_hashbucket *ilb2;
720 if (sk->sk_state != TCP_LISTEN) {
722 inet_ehash_nolisten(sk, osk, NULL);
726 WARN_ON(!sk_unhashed(sk));
727 ilb2 = inet_lhash2_bucket_sk(hashinfo, sk);
729 spin_lock(&ilb2->lock);
730 if (sk->sk_reuseport) {
731 err = inet_reuseport_add_sock(sk, ilb2);
735 if (IS_ENABLED(CONFIG_IPV6) && sk->sk_reuseport &&
736 sk->sk_family == AF_INET6)
737 __sk_nulls_add_node_tail_rcu(sk, &ilb2->nulls_head);
739 __sk_nulls_add_node_rcu(sk, &ilb2->nulls_head);
740 sock_set_flag(sk, SOCK_RCU_FREE);
741 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
743 spin_unlock(&ilb2->lock);
747 EXPORT_SYMBOL(__inet_hash);
749 int inet_hash(struct sock *sk)
753 if (sk->sk_state != TCP_CLOSE)
754 err = __inet_hash(sk, NULL);
758 EXPORT_SYMBOL_GPL(inet_hash);
760 void inet_unhash(struct sock *sk)
762 struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
767 if (sk->sk_state == TCP_LISTEN) {
768 struct inet_listen_hashbucket *ilb2;
770 ilb2 = inet_lhash2_bucket_sk(hashinfo, sk);
771 /* Don't disable bottom halves while acquiring the lock to
772 * avoid circular locking dependency on PREEMPT_RT.
774 spin_lock(&ilb2->lock);
775 if (sk_unhashed(sk)) {
776 spin_unlock(&ilb2->lock);
780 if (rcu_access_pointer(sk->sk_reuseport_cb))
781 reuseport_stop_listen_sock(sk);
783 __sk_nulls_del_node_init_rcu(sk);
784 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
785 spin_unlock(&ilb2->lock);
787 spinlock_t *lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
790 if (sk_unhashed(sk)) {
791 spin_unlock_bh(lock);
794 __sk_nulls_del_node_init_rcu(sk);
795 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
796 spin_unlock_bh(lock);
799 EXPORT_SYMBOL_GPL(inet_unhash);
801 static bool inet_bind2_bucket_match(const struct inet_bind2_bucket *tb,
802 const struct net *net, unsigned short port,
803 int l3mdev, const struct sock *sk)
805 #if IS_ENABLED(CONFIG_IPV6)
806 if (sk->sk_family != tb->family)
809 if (sk->sk_family == AF_INET6)
810 return net_eq(ib2_net(tb), net) && tb->port == port &&
811 tb->l3mdev == l3mdev &&
812 ipv6_addr_equal(&tb->v6_rcv_saddr, &sk->sk_v6_rcv_saddr);
815 return net_eq(ib2_net(tb), net) && tb->port == port &&
816 tb->l3mdev == l3mdev && tb->rcv_saddr == sk->sk_rcv_saddr;
819 bool inet_bind2_bucket_match_addr_any(const struct inet_bind2_bucket *tb, const struct net *net,
820 unsigned short port, int l3mdev, const struct sock *sk)
822 #if IS_ENABLED(CONFIG_IPV6)
823 if (sk->sk_family != tb->family) {
824 if (sk->sk_family == AF_INET)
825 return net_eq(ib2_net(tb), net) && tb->port == port &&
826 tb->l3mdev == l3mdev &&
827 ipv6_addr_any(&tb->v6_rcv_saddr);
832 if (sk->sk_family == AF_INET6)
833 return net_eq(ib2_net(tb), net) && tb->port == port &&
834 tb->l3mdev == l3mdev &&
835 ipv6_addr_any(&tb->v6_rcv_saddr);
838 return net_eq(ib2_net(tb), net) && tb->port == port &&
839 tb->l3mdev == l3mdev && tb->rcv_saddr == 0;
842 /* The socket's bhash2 hashbucket spinlock must be held when this is called */
843 struct inet_bind2_bucket *
844 inet_bind2_bucket_find(const struct inet_bind_hashbucket *head, const struct net *net,
845 unsigned short port, int l3mdev, const struct sock *sk)
847 struct inet_bind2_bucket *bhash2 = NULL;
849 inet_bind_bucket_for_each(bhash2, &head->chain)
850 if (inet_bind2_bucket_match(bhash2, net, port, l3mdev, sk))
856 struct inet_bind_hashbucket *
857 inet_bhash2_addr_any_hashbucket(const struct sock *sk, const struct net *net, int port)
859 struct inet_hashinfo *hinfo = tcp_or_dccp_get_hashinfo(sk);
862 #if IS_ENABLED(CONFIG_IPV6)
863 if (sk->sk_family == AF_INET6)
864 hash = ipv6_portaddr_hash(net, &in6addr_any, port);
867 hash = ipv4_portaddr_hash(net, 0, port);
869 return &hinfo->bhash2[hash & (hinfo->bhash_size - 1)];
872 static void inet_update_saddr(struct sock *sk, void *saddr, int family)
874 if (family == AF_INET) {
875 inet_sk(sk)->inet_saddr = *(__be32 *)saddr;
876 sk_rcv_saddr_set(sk, inet_sk(sk)->inet_saddr);
878 #if IS_ENABLED(CONFIG_IPV6)
880 sk->sk_v6_rcv_saddr = *(struct in6_addr *)saddr;
885 static int __inet_bhash2_update_saddr(struct sock *sk, void *saddr, int family, bool reset)
887 struct inet_hashinfo *hinfo = tcp_or_dccp_get_hashinfo(sk);
888 struct inet_bind_hashbucket *head, *head2;
889 struct inet_bind2_bucket *tb2, *new_tb2;
890 int l3mdev = inet_sk_bound_l3mdev(sk);
891 int port = inet_sk(sk)->inet_num;
892 struct net *net = sock_net(sk);
895 if (!inet_csk(sk)->icsk_bind2_hash) {
896 /* Not bind()ed before. */
898 inet_reset_saddr(sk);
900 inet_update_saddr(sk, saddr, family);
905 /* Allocate a bind2 bucket ahead of time to avoid permanently putting
906 * the bhash2 table in an inconsistent state if a new tb2 bucket
909 new_tb2 = kmem_cache_alloc(hinfo->bind2_bucket_cachep, GFP_ATOMIC);
912 /* The (INADDR_ANY, port) bucket might have already
913 * been freed, then we cannot fixup icsk_bind2_hash,
914 * so we give up and unlink sk from bhash/bhash2 not
915 * to leave inconsistency in bhash2.
918 inet_reset_saddr(sk);
924 bhash = inet_bhashfn(net, port, hinfo->bhash_size);
925 head = &hinfo->bhash[bhash];
926 head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
928 /* If we change saddr locklessly, another thread
929 * iterating over bhash might see corrupted address.
931 spin_lock_bh(&head->lock);
933 spin_lock(&head2->lock);
934 __sk_del_bind2_node(sk);
935 inet_bind2_bucket_destroy(hinfo->bind2_bucket_cachep, inet_csk(sk)->icsk_bind2_hash);
936 spin_unlock(&head2->lock);
939 inet_reset_saddr(sk);
941 inet_update_saddr(sk, saddr, family);
943 head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
945 spin_lock(&head2->lock);
946 tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, sk);
949 inet_bind2_bucket_init(tb2, net, head2, port, l3mdev, sk);
951 sk_add_bind2_node(sk, &tb2->owners);
952 inet_csk(sk)->icsk_bind2_hash = tb2;
953 spin_unlock(&head2->lock);
955 spin_unlock_bh(&head->lock);
958 kmem_cache_free(hinfo->bind2_bucket_cachep, new_tb2);
963 int inet_bhash2_update_saddr(struct sock *sk, void *saddr, int family)
965 return __inet_bhash2_update_saddr(sk, saddr, family, false);
967 EXPORT_SYMBOL_GPL(inet_bhash2_update_saddr);
969 void inet_bhash2_reset_saddr(struct sock *sk)
971 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
972 __inet_bhash2_update_saddr(sk, NULL, 0, true);
974 EXPORT_SYMBOL_GPL(inet_bhash2_reset_saddr);
976 /* RFC 6056 3.3.4. Algorithm 4: Double-Hash Port Selection Algorithm
977 * Note that we use 32bit integers (vs RFC 'short integers')
978 * because 2^16 is not a multiple of num_ephemeral and this
979 * property might be used by clever attacker.
981 * RFC claims using TABLE_LENGTH=10 buckets gives an improvement, though
982 * attacks were since demonstrated, thus we use 65536 by default instead
983 * to really give more isolation and privacy, at the expense of 256kB
986 #define INET_TABLE_PERTURB_SIZE (1 << CONFIG_INET_TABLE_PERTURB_ORDER)
987 static u32 *table_perturb;
989 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
990 struct sock *sk, u64 port_offset,
991 int (*check_established)(struct inet_timewait_death_row *,
992 struct sock *, __u16, struct inet_timewait_sock **))
994 struct inet_hashinfo *hinfo = death_row->hashinfo;
995 struct inet_bind_hashbucket *head, *head2;
996 struct inet_timewait_sock *tw = NULL;
997 int port = inet_sk(sk)->inet_num;
998 struct net *net = sock_net(sk);
999 struct inet_bind2_bucket *tb2;
1000 struct inet_bind_bucket *tb;
1001 bool tb_created = false;
1002 u32 remaining, offset;
1003 int ret, i, low, high;
1009 ret = check_established(death_row, sk, port, NULL);
1014 l3mdev = inet_sk_bound_l3mdev(sk);
1016 inet_sk_get_local_port_range(sk, &low, &high);
1017 high++; /* [32768, 60999] -> [32768, 61000[ */
1018 remaining = high - low;
1019 if (likely(remaining > 1))
1022 get_random_sleepable_once(table_perturb,
1023 INET_TABLE_PERTURB_SIZE * sizeof(*table_perturb));
1024 index = port_offset & (INET_TABLE_PERTURB_SIZE - 1);
1026 offset = READ_ONCE(table_perturb[index]) + (port_offset >> 32);
1027 offset %= remaining;
1029 /* In first pass we try ports of @low parity.
1030 * inet_csk_get_port() does the opposite choice.
1034 port = low + offset;
1035 for (i = 0; i < remaining; i += 2, port += 2) {
1036 if (unlikely(port >= high))
1038 if (inet_is_local_reserved_port(net, port))
1040 head = &hinfo->bhash[inet_bhashfn(net, port,
1041 hinfo->bhash_size)];
1042 spin_lock_bh(&head->lock);
1044 /* Does not bother with rcv_saddr checks, because
1045 * the established check is already unique enough.
1047 inet_bind_bucket_for_each(tb, &head->chain) {
1048 if (inet_bind_bucket_match(tb, net, port, l3mdev)) {
1049 if (tb->fastreuse >= 0 ||
1050 tb->fastreuseport >= 0)
1052 WARN_ON(hlist_empty(&tb->owners));
1053 if (!check_established(death_row, sk,
1060 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
1061 net, head, port, l3mdev);
1063 spin_unlock_bh(&head->lock);
1068 tb->fastreuseport = -1;
1071 spin_unlock_bh(&head->lock);
1076 if ((offset & 1) && remaining > 1)
1077 goto other_parity_scan;
1079 return -EADDRNOTAVAIL;
1082 /* Find the corresponding tb2 bucket since we need to
1083 * add the socket to the bhash2 table as well
1085 head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
1086 spin_lock(&head2->lock);
1088 tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, sk);
1090 tb2 = inet_bind2_bucket_create(hinfo->bind2_bucket_cachep, net,
1091 head2, port, l3mdev, sk);
1096 /* Here we want to add a little bit of randomness to the next source
1097 * port that will be chosen. We use a max() with a random here so that
1098 * on low contention the randomness is maximal and on high contention
1099 * it may be inexistent.
1101 i = max_t(int, i, get_random_u32_below(8) * 2);
1102 WRITE_ONCE(table_perturb[index], READ_ONCE(table_perturb[index]) + i + 2);
1104 /* Head lock still held and bh's disabled */
1105 inet_bind_hash(sk, tb, tb2, port);
1107 if (sk_unhashed(sk)) {
1108 inet_sk(sk)->inet_sport = htons(port);
1109 inet_ehash_nolisten(sk, (struct sock *)tw, NULL);
1112 inet_twsk_bind_unhash(tw, hinfo);
1114 spin_unlock(&head2->lock);
1115 spin_unlock(&head->lock);
1118 inet_twsk_deschedule_put(tw);
1123 spin_unlock(&head2->lock);
1125 inet_bind_bucket_destroy(hinfo->bind_bucket_cachep, tb);
1126 spin_unlock_bh(&head->lock);
1131 * Bind a port for a connect operation and hash it.
1133 int inet_hash_connect(struct inet_timewait_death_row *death_row,
1136 u64 port_offset = 0;
1138 if (!inet_sk(sk)->inet_num)
1139 port_offset = inet_sk_port_offset(sk);
1140 return __inet_hash_connect(death_row, sk, port_offset,
1141 __inet_check_established);
1143 EXPORT_SYMBOL_GPL(inet_hash_connect);
1145 static void init_hashinfo_lhash2(struct inet_hashinfo *h)
1149 for (i = 0; i <= h->lhash2_mask; i++) {
1150 spin_lock_init(&h->lhash2[i].lock);
1151 INIT_HLIST_NULLS_HEAD(&h->lhash2[i].nulls_head,
1152 i + LISTENING_NULLS_BASE);
1156 void __init inet_hashinfo2_init(struct inet_hashinfo *h, const char *name,
1157 unsigned long numentries, int scale,
1158 unsigned long low_limit,
1159 unsigned long high_limit)
1161 h->lhash2 = alloc_large_system_hash(name,
1170 init_hashinfo_lhash2(h);
1172 /* this one is used for source ports of outgoing connections */
1173 table_perturb = alloc_large_system_hash("Table-perturb",
1174 sizeof(*table_perturb),
1175 INET_TABLE_PERTURB_SIZE,
1177 INET_TABLE_PERTURB_SIZE,
1178 INET_TABLE_PERTURB_SIZE);
1181 int inet_hashinfo2_init_mod(struct inet_hashinfo *h)
1183 h->lhash2 = kmalloc_array(INET_LHTABLE_SIZE, sizeof(*h->lhash2), GFP_KERNEL);
1187 h->lhash2_mask = INET_LHTABLE_SIZE - 1;
1188 /* INET_LHTABLE_SIZE must be a power of 2 */
1189 BUG_ON(INET_LHTABLE_SIZE & h->lhash2_mask);
1191 init_hashinfo_lhash2(h);
1194 EXPORT_SYMBOL_GPL(inet_hashinfo2_init_mod);
1196 int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
1198 unsigned int locksz = sizeof(spinlock_t);
1199 unsigned int i, nblocks = 1;
1202 /* allocate 2 cache lines or at least one spinlock per cpu */
1203 nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U);
1204 nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
1206 /* no more locks than number of hash buckets */
1207 nblocks = min(nblocks, hashinfo->ehash_mask + 1);
1209 hashinfo->ehash_locks = kvmalloc_array(nblocks, locksz, GFP_KERNEL);
1210 if (!hashinfo->ehash_locks)
1213 for (i = 0; i < nblocks; i++)
1214 spin_lock_init(&hashinfo->ehash_locks[i]);
1216 hashinfo->ehash_locks_mask = nblocks - 1;
1219 EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);
1221 struct inet_hashinfo *inet_pernet_hashinfo_alloc(struct inet_hashinfo *hashinfo,
1222 unsigned int ehash_entries)
1224 struct inet_hashinfo *new_hashinfo;
1227 new_hashinfo = kmemdup(hashinfo, sizeof(*hashinfo), GFP_KERNEL);
1231 new_hashinfo->ehash = vmalloc_huge(ehash_entries * sizeof(struct inet_ehash_bucket),
1232 GFP_KERNEL_ACCOUNT);
1233 if (!new_hashinfo->ehash)
1236 new_hashinfo->ehash_mask = ehash_entries - 1;
1238 if (inet_ehash_locks_alloc(new_hashinfo))
1241 for (i = 0; i < ehash_entries; i++)
1242 INIT_HLIST_NULLS_HEAD(&new_hashinfo->ehash[i].chain, i);
1244 new_hashinfo->pernet = true;
1246 return new_hashinfo;
1249 vfree(new_hashinfo->ehash);
1251 kfree(new_hashinfo);
1255 EXPORT_SYMBOL_GPL(inet_pernet_hashinfo_alloc);
1257 void inet_pernet_hashinfo_free(struct inet_hashinfo *hashinfo)
1259 if (!hashinfo->pernet)
1262 inet_ehash_locks_free(hashinfo);
1263 vfree(hashinfo->ehash);
1266 EXPORT_SYMBOL_GPL(inet_pernet_hashinfo_free);