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 static u32 inet_ehashfn(const struct net *net, const __be32 laddr,
32 const __u16 lport, const __be32 faddr,
35 static u32 inet_ehash_secret __read_mostly;
37 net_get_random_once(&inet_ehash_secret, sizeof(inet_ehash_secret));
39 return __inet_ehashfn(laddr, lport, faddr, fport,
40 inet_ehash_secret + net_hash_mix(net));
43 /* This function handles inet_sock, but also timewait and request sockets
46 static u32 sk_ehashfn(const struct sock *sk)
48 #if IS_ENABLED(CONFIG_IPV6)
49 if (sk->sk_family == AF_INET6 &&
50 !ipv6_addr_v4mapped(&sk->sk_v6_daddr))
51 return inet6_ehashfn(sock_net(sk),
52 &sk->sk_v6_rcv_saddr, sk->sk_num,
53 &sk->sk_v6_daddr, sk->sk_dport);
55 return inet_ehashfn(sock_net(sk),
56 sk->sk_rcv_saddr, sk->sk_num,
57 sk->sk_daddr, sk->sk_dport);
61 * Allocate and initialize a new local port bind bucket.
62 * The bindhash mutex for snum's hash chain must be held here.
64 struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep,
66 struct inet_bind_hashbucket *head,
67 const unsigned short snum,
70 struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
73 write_pnet(&tb->ib_net, net);
77 tb->fastreuseport = 0;
78 INIT_HLIST_HEAD(&tb->owners);
79 hlist_add_head(&tb->node, &head->chain);
85 * Caller must hold hashbucket lock for this tb with local BH disabled
87 void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
89 if (hlist_empty(&tb->owners)) {
90 __hlist_del(&tb->node);
91 kmem_cache_free(cachep, tb);
95 bool inet_bind_bucket_match(const struct inet_bind_bucket *tb, const struct net *net,
96 unsigned short port, int l3mdev)
98 return net_eq(ib_net(tb), net) && tb->port == port &&
102 static void inet_bind2_bucket_init(struct inet_bind2_bucket *tb,
104 struct inet_bind_hashbucket *head,
105 unsigned short port, int l3mdev,
106 const struct sock *sk)
108 write_pnet(&tb->ib_net, net);
111 #if IS_ENABLED(CONFIG_IPV6)
112 tb->family = sk->sk_family;
113 if (sk->sk_family == AF_INET6)
114 tb->v6_rcv_saddr = sk->sk_v6_rcv_saddr;
117 tb->rcv_saddr = sk->sk_rcv_saddr;
118 INIT_HLIST_HEAD(&tb->owners);
119 INIT_HLIST_HEAD(&tb->deathrow);
120 hlist_add_head(&tb->node, &head->chain);
123 struct inet_bind2_bucket *inet_bind2_bucket_create(struct kmem_cache *cachep,
125 struct inet_bind_hashbucket *head,
128 const struct sock *sk)
130 struct inet_bind2_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
133 inet_bind2_bucket_init(tb, net, head, port, l3mdev, sk);
138 /* Caller must hold hashbucket lock for this tb with local BH disabled */
139 void inet_bind2_bucket_destroy(struct kmem_cache *cachep, struct inet_bind2_bucket *tb)
141 if (hlist_empty(&tb->owners) && hlist_empty(&tb->deathrow)) {
142 __hlist_del(&tb->node);
143 kmem_cache_free(cachep, tb);
147 static bool inet_bind2_bucket_addr_match(const struct inet_bind2_bucket *tb2,
148 const struct sock *sk)
150 #if IS_ENABLED(CONFIG_IPV6)
151 if (sk->sk_family != tb2->family) {
152 if (sk->sk_family == AF_INET)
153 return ipv6_addr_v4mapped(&tb2->v6_rcv_saddr) &&
154 tb2->v6_rcv_saddr.s6_addr32[3] == sk->sk_rcv_saddr;
156 return ipv6_addr_v4mapped(&sk->sk_v6_rcv_saddr) &&
157 sk->sk_v6_rcv_saddr.s6_addr32[3] == tb2->rcv_saddr;
160 if (sk->sk_family == AF_INET6)
161 return ipv6_addr_equal(&tb2->v6_rcv_saddr,
162 &sk->sk_v6_rcv_saddr);
164 return tb2->rcv_saddr == sk->sk_rcv_saddr;
167 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
168 struct inet_bind2_bucket *tb2, unsigned short port)
170 inet_sk(sk)->inet_num = port;
171 sk_add_bind_node(sk, &tb->owners);
172 inet_csk(sk)->icsk_bind_hash = tb;
173 sk_add_bind2_node(sk, &tb2->owners);
174 inet_csk(sk)->icsk_bind2_hash = tb2;
178 * Get rid of any references to a local port held by the given sock.
180 static void __inet_put_port(struct sock *sk)
182 struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
183 struct inet_bind_hashbucket *head, *head2;
184 struct net *net = sock_net(sk);
185 struct inet_bind_bucket *tb;
188 bhash = inet_bhashfn(net, inet_sk(sk)->inet_num, hashinfo->bhash_size);
189 head = &hashinfo->bhash[bhash];
190 head2 = inet_bhashfn_portaddr(hashinfo, sk, net, inet_sk(sk)->inet_num);
192 spin_lock(&head->lock);
193 tb = inet_csk(sk)->icsk_bind_hash;
194 __sk_del_bind_node(sk);
195 inet_csk(sk)->icsk_bind_hash = NULL;
196 inet_sk(sk)->inet_num = 0;
197 inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
199 spin_lock(&head2->lock);
200 if (inet_csk(sk)->icsk_bind2_hash) {
201 struct inet_bind2_bucket *tb2 = inet_csk(sk)->icsk_bind2_hash;
203 __sk_del_bind2_node(sk);
204 inet_csk(sk)->icsk_bind2_hash = NULL;
205 inet_bind2_bucket_destroy(hashinfo->bind2_bucket_cachep, tb2);
207 spin_unlock(&head2->lock);
209 spin_unlock(&head->lock);
212 void inet_put_port(struct sock *sk)
218 EXPORT_SYMBOL(inet_put_port);
220 int __inet_inherit_port(const struct sock *sk, struct sock *child)
222 struct inet_hashinfo *table = tcp_or_dccp_get_hashinfo(sk);
223 unsigned short port = inet_sk(child)->inet_num;
224 struct inet_bind_hashbucket *head, *head2;
225 bool created_inet_bind_bucket = false;
226 struct net *net = sock_net(sk);
227 bool update_fastreuse = false;
228 struct inet_bind2_bucket *tb2;
229 struct inet_bind_bucket *tb;
232 bhash = inet_bhashfn(net, port, table->bhash_size);
233 head = &table->bhash[bhash];
234 head2 = inet_bhashfn_portaddr(table, child, net, port);
236 spin_lock(&head->lock);
237 spin_lock(&head2->lock);
238 tb = inet_csk(sk)->icsk_bind_hash;
239 tb2 = inet_csk(sk)->icsk_bind2_hash;
240 if (unlikely(!tb || !tb2)) {
241 spin_unlock(&head2->lock);
242 spin_unlock(&head->lock);
245 if (tb->port != port) {
246 l3mdev = inet_sk_bound_l3mdev(sk);
248 /* NOTE: using tproxy and redirecting skbs to a proxy
249 * on a different listener port breaks the assumption
250 * that the listener socket's icsk_bind_hash is the same
251 * as that of the child socket. We have to look up or
252 * create a new bind bucket for the child here. */
253 inet_bind_bucket_for_each(tb, &head->chain) {
254 if (inet_bind_bucket_match(tb, net, port, l3mdev))
258 tb = inet_bind_bucket_create(table->bind_bucket_cachep,
259 net, head, port, l3mdev);
261 spin_unlock(&head2->lock);
262 spin_unlock(&head->lock);
265 created_inet_bind_bucket = true;
267 update_fastreuse = true;
270 } else if (!inet_bind2_bucket_addr_match(tb2, child)) {
271 l3mdev = inet_sk_bound_l3mdev(sk);
274 tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, child);
276 tb2 = inet_bind2_bucket_create(table->bind2_bucket_cachep,
283 if (update_fastreuse)
284 inet_csk_update_fastreuse(tb, child);
285 inet_bind_hash(child, tb, tb2, port);
286 spin_unlock(&head2->lock);
287 spin_unlock(&head->lock);
292 if (created_inet_bind_bucket)
293 inet_bind_bucket_destroy(table->bind_bucket_cachep, tb);
294 spin_unlock(&head2->lock);
295 spin_unlock(&head->lock);
298 EXPORT_SYMBOL_GPL(__inet_inherit_port);
300 static struct inet_listen_hashbucket *
301 inet_lhash2_bucket_sk(struct inet_hashinfo *h, struct sock *sk)
305 #if IS_ENABLED(CONFIG_IPV6)
306 if (sk->sk_family == AF_INET6)
307 hash = ipv6_portaddr_hash(sock_net(sk),
308 &sk->sk_v6_rcv_saddr,
309 inet_sk(sk)->inet_num);
312 hash = ipv4_portaddr_hash(sock_net(sk),
313 inet_sk(sk)->inet_rcv_saddr,
314 inet_sk(sk)->inet_num);
315 return inet_lhash2_bucket(h, hash);
318 static inline int compute_score(struct sock *sk, struct net *net,
319 const unsigned short hnum, const __be32 daddr,
320 const int dif, const int sdif)
324 if (net_eq(sock_net(sk), net) && sk->sk_num == hnum &&
325 !ipv6_only_sock(sk)) {
326 if (sk->sk_rcv_saddr != daddr)
329 if (!inet_sk_bound_dev_eq(net, sk->sk_bound_dev_if, dif, sdif))
331 score = sk->sk_bound_dev_if ? 2 : 1;
333 if (sk->sk_family == PF_INET)
335 if (READ_ONCE(sk->sk_incoming_cpu) == raw_smp_processor_id())
341 static inline struct sock *lookup_reuseport(struct net *net, struct sock *sk,
342 struct sk_buff *skb, int doff,
343 __be32 saddr, __be16 sport,
344 __be32 daddr, unsigned short hnum)
346 struct sock *reuse_sk = NULL;
349 if (sk->sk_reuseport) {
350 phash = inet_ehashfn(net, daddr, hnum, saddr, sport);
351 reuse_sk = reuseport_select_sock(sk, phash, skb, doff);
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 = lookup_reuseport(net, sk, skb, doff,
379 saddr, sport, daddr, hnum);
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 = lookup_reuseport(net, sk, skb, doff, saddr, sport, daddr, hnum);
414 struct sock *__inet_lookup_listener(struct net *net,
415 struct inet_hashinfo *hashinfo,
416 struct sk_buff *skb, int doff,
417 const __be32 saddr, __be16 sport,
418 const __be32 daddr, const unsigned short hnum,
419 const int dif, const int sdif)
421 struct inet_listen_hashbucket *ilb2;
422 struct sock *result = NULL;
425 /* Lookup redirect from BPF */
426 if (static_branch_unlikely(&bpf_sk_lookup_enabled)) {
427 result = inet_lookup_run_bpf(net, hashinfo, skb, doff,
428 saddr, sport, daddr, hnum, dif);
433 hash2 = ipv4_portaddr_hash(net, daddr, hnum);
434 ilb2 = inet_lhash2_bucket(hashinfo, hash2);
436 result = inet_lhash2_lookup(net, ilb2, skb, doff,
437 saddr, sport, daddr, hnum,
442 /* Lookup lhash2 with INADDR_ANY */
443 hash2 = ipv4_portaddr_hash(net, htonl(INADDR_ANY), hnum);
444 ilb2 = inet_lhash2_bucket(hashinfo, hash2);
446 result = inet_lhash2_lookup(net, ilb2, skb, doff,
447 saddr, sport, htonl(INADDR_ANY), hnum,
454 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
456 /* All sockets share common refcount, but have different destructors */
457 void sock_gen_put(struct sock *sk)
459 if (!refcount_dec_and_test(&sk->sk_refcnt))
462 if (sk->sk_state == TCP_TIME_WAIT)
463 inet_twsk_free(inet_twsk(sk));
464 else if (sk->sk_state == TCP_NEW_SYN_RECV)
465 reqsk_free(inet_reqsk(sk));
469 EXPORT_SYMBOL_GPL(sock_gen_put);
471 void sock_edemux(struct sk_buff *skb)
473 sock_gen_put(skb->sk);
475 EXPORT_SYMBOL(sock_edemux);
477 struct sock *__inet_lookup_established(struct net *net,
478 struct inet_hashinfo *hashinfo,
479 const __be32 saddr, const __be16 sport,
480 const __be32 daddr, const u16 hnum,
481 const int dif, const int sdif)
483 INET_ADDR_COOKIE(acookie, saddr, daddr);
484 const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
486 const struct hlist_nulls_node *node;
487 /* Optimize here for direct hit, only listening connections can
488 * have wildcards anyways.
490 unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
491 unsigned int slot = hash & hashinfo->ehash_mask;
492 struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
495 sk_nulls_for_each_rcu(sk, node, &head->chain) {
496 if (sk->sk_hash != hash)
498 if (likely(inet_match(net, sk, acookie, ports, dif, sdif))) {
499 if (unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
501 if (unlikely(!inet_match(net, sk, acookie,
502 ports, dif, sdif))) {
510 * if the nulls value we got at the end of this lookup is
511 * not the expected one, we must restart lookup.
512 * We probably met an item that was moved to another chain.
514 if (get_nulls_value(node) != slot)
521 EXPORT_SYMBOL_GPL(__inet_lookup_established);
523 /* called with local bh disabled */
524 static int __inet_check_established(struct inet_timewait_death_row *death_row,
525 struct sock *sk, __u16 lport,
526 struct inet_timewait_sock **twp)
528 struct inet_hashinfo *hinfo = death_row->hashinfo;
529 struct inet_sock *inet = inet_sk(sk);
530 __be32 daddr = inet->inet_rcv_saddr;
531 __be32 saddr = inet->inet_daddr;
532 int dif = sk->sk_bound_dev_if;
533 struct net *net = sock_net(sk);
534 int sdif = l3mdev_master_ifindex_by_index(net, dif);
535 INET_ADDR_COOKIE(acookie, saddr, daddr);
536 const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
537 unsigned int hash = inet_ehashfn(net, daddr, lport,
538 saddr, inet->inet_dport);
539 struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
540 spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
542 const struct hlist_nulls_node *node;
543 struct inet_timewait_sock *tw = NULL;
547 sk_nulls_for_each(sk2, node, &head->chain) {
548 if (sk2->sk_hash != hash)
551 if (likely(inet_match(net, sk2, acookie, ports, dif, sdif))) {
552 if (sk2->sk_state == TCP_TIME_WAIT) {
554 if (twsk_unique(sk, sk2, twp))
561 /* Must record num and sport now. Otherwise we will see
562 * in hash table socket with a funny identity.
564 inet->inet_num = lport;
565 inet->inet_sport = htons(lport);
567 WARN_ON(!sk_unhashed(sk));
568 __sk_nulls_add_node_rcu(sk, &head->chain);
570 sk_nulls_del_node_init_rcu((struct sock *)tw);
571 __NET_INC_STATS(net, LINUX_MIB_TIMEWAITRECYCLED);
574 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
579 /* Silly. Should hash-dance instead... */
580 inet_twsk_deschedule_put(tw);
586 return -EADDRNOTAVAIL;
589 static u64 inet_sk_port_offset(const struct sock *sk)
591 const struct inet_sock *inet = inet_sk(sk);
593 return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
598 /* Searches for an exsiting socket in the ehash bucket list.
599 * Returns true if found, false otherwise.
601 static bool inet_ehash_lookup_by_sk(struct sock *sk,
602 struct hlist_nulls_head *list)
604 const __portpair ports = INET_COMBINED_PORTS(sk->sk_dport, sk->sk_num);
605 const int sdif = sk->sk_bound_dev_if;
606 const int dif = sk->sk_bound_dev_if;
607 const struct hlist_nulls_node *node;
608 struct net *net = sock_net(sk);
611 INET_ADDR_COOKIE(acookie, sk->sk_daddr, sk->sk_rcv_saddr);
613 sk_nulls_for_each_rcu(esk, node, list) {
614 if (esk->sk_hash != sk->sk_hash)
616 if (sk->sk_family == AF_INET) {
617 if (unlikely(inet_match(net, esk, acookie,
618 ports, dif, sdif))) {
622 #if IS_ENABLED(CONFIG_IPV6)
623 else if (sk->sk_family == AF_INET6) {
624 if (unlikely(inet6_match(net, esk,
626 &sk->sk_v6_rcv_saddr,
627 ports, dif, sdif))) {
636 /* Insert a socket into ehash, and eventually remove another one
637 * (The another one can be a SYN_RECV or TIMEWAIT)
638 * If an existing socket already exists, socket sk is not inserted,
639 * and sets found_dup_sk parameter to true.
641 bool inet_ehash_insert(struct sock *sk, struct sock *osk, bool *found_dup_sk)
643 struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
644 struct inet_ehash_bucket *head;
645 struct hlist_nulls_head *list;
649 WARN_ON_ONCE(!sk_unhashed(sk));
651 sk->sk_hash = sk_ehashfn(sk);
652 head = inet_ehash_bucket(hashinfo, sk->sk_hash);
654 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
658 WARN_ON_ONCE(sk->sk_hash != osk->sk_hash);
659 ret = sk_nulls_del_node_init_rcu(osk);
660 } else if (found_dup_sk) {
661 *found_dup_sk = inet_ehash_lookup_by_sk(sk, list);
667 __sk_nulls_add_node_rcu(sk, list);
674 bool inet_ehash_nolisten(struct sock *sk, struct sock *osk, bool *found_dup_sk)
676 bool ok = inet_ehash_insert(sk, osk, found_dup_sk);
679 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
681 this_cpu_inc(*sk->sk_prot->orphan_count);
682 inet_sk_set_state(sk, TCP_CLOSE);
683 sock_set_flag(sk, SOCK_DEAD);
684 inet_csk_destroy_sock(sk);
688 EXPORT_SYMBOL_GPL(inet_ehash_nolisten);
690 static int inet_reuseport_add_sock(struct sock *sk,
691 struct inet_listen_hashbucket *ilb)
693 struct inet_bind_bucket *tb = inet_csk(sk)->icsk_bind_hash;
694 const struct hlist_nulls_node *node;
696 kuid_t uid = sock_i_uid(sk);
698 sk_nulls_for_each_rcu(sk2, node, &ilb->nulls_head) {
700 sk2->sk_family == sk->sk_family &&
701 ipv6_only_sock(sk2) == ipv6_only_sock(sk) &&
702 sk2->sk_bound_dev_if == sk->sk_bound_dev_if &&
703 inet_csk(sk2)->icsk_bind_hash == tb &&
704 sk2->sk_reuseport && uid_eq(uid, sock_i_uid(sk2)) &&
705 inet_rcv_saddr_equal(sk, sk2, false))
706 return reuseport_add_sock(sk, sk2,
707 inet_rcv_saddr_any(sk));
710 return reuseport_alloc(sk, inet_rcv_saddr_any(sk));
713 int __inet_hash(struct sock *sk, struct sock *osk)
715 struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
716 struct inet_listen_hashbucket *ilb2;
719 if (sk->sk_state != TCP_LISTEN) {
721 inet_ehash_nolisten(sk, osk, NULL);
725 WARN_ON(!sk_unhashed(sk));
726 ilb2 = inet_lhash2_bucket_sk(hashinfo, sk);
728 spin_lock(&ilb2->lock);
729 if (sk->sk_reuseport) {
730 err = inet_reuseport_add_sock(sk, ilb2);
734 sock_set_flag(sk, SOCK_RCU_FREE);
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_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
742 spin_unlock(&ilb2->lock);
746 EXPORT_SYMBOL(__inet_hash);
748 int inet_hash(struct sock *sk)
752 if (sk->sk_state != TCP_CLOSE)
753 err = __inet_hash(sk, NULL);
757 EXPORT_SYMBOL_GPL(inet_hash);
759 void inet_unhash(struct sock *sk)
761 struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
766 if (sk->sk_state == TCP_LISTEN) {
767 struct inet_listen_hashbucket *ilb2;
769 ilb2 = inet_lhash2_bucket_sk(hashinfo, sk);
770 /* Don't disable bottom halves while acquiring the lock to
771 * avoid circular locking dependency on PREEMPT_RT.
773 spin_lock(&ilb2->lock);
774 if (sk_unhashed(sk)) {
775 spin_unlock(&ilb2->lock);
779 if (rcu_access_pointer(sk->sk_reuseport_cb))
780 reuseport_stop_listen_sock(sk);
782 __sk_nulls_del_node_init_rcu(sk);
783 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
784 spin_unlock(&ilb2->lock);
786 spinlock_t *lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
789 if (sk_unhashed(sk)) {
790 spin_unlock_bh(lock);
793 __sk_nulls_del_node_init_rcu(sk);
794 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
795 spin_unlock_bh(lock);
798 EXPORT_SYMBOL_GPL(inet_unhash);
800 static bool inet_bind2_bucket_match(const struct inet_bind2_bucket *tb,
801 const struct net *net, unsigned short port,
802 int l3mdev, const struct sock *sk)
804 if (!net_eq(ib2_net(tb), net) || tb->port != port ||
805 tb->l3mdev != l3mdev)
808 return inet_bind2_bucket_addr_match(tb, sk);
811 bool inet_bind2_bucket_match_addr_any(const struct inet_bind2_bucket *tb, const struct net *net,
812 unsigned short port, int l3mdev, const struct sock *sk)
814 if (!net_eq(ib2_net(tb), net) || tb->port != port ||
815 tb->l3mdev != l3mdev)
818 #if IS_ENABLED(CONFIG_IPV6)
819 if (sk->sk_family != tb->family) {
820 if (sk->sk_family == AF_INET)
821 return ipv6_addr_any(&tb->v6_rcv_saddr) ||
822 ipv6_addr_v4mapped_any(&tb->v6_rcv_saddr);
827 if (sk->sk_family == AF_INET6)
828 return ipv6_addr_any(&tb->v6_rcv_saddr);
830 return tb->rcv_saddr == 0;
833 /* The socket's bhash2 hashbucket spinlock must be held when this is called */
834 struct inet_bind2_bucket *
835 inet_bind2_bucket_find(const struct inet_bind_hashbucket *head, const struct net *net,
836 unsigned short port, int l3mdev, const struct sock *sk)
838 struct inet_bind2_bucket *bhash2 = NULL;
840 inet_bind_bucket_for_each(bhash2, &head->chain)
841 if (inet_bind2_bucket_match(bhash2, net, port, l3mdev, sk))
847 struct inet_bind_hashbucket *
848 inet_bhash2_addr_any_hashbucket(const struct sock *sk, const struct net *net, int port)
850 struct inet_hashinfo *hinfo = tcp_or_dccp_get_hashinfo(sk);
853 #if IS_ENABLED(CONFIG_IPV6)
854 if (sk->sk_family == AF_INET6)
855 hash = ipv6_portaddr_hash(net, &in6addr_any, port);
858 hash = ipv4_portaddr_hash(net, 0, port);
860 return &hinfo->bhash2[hash & (hinfo->bhash_size - 1)];
863 static void inet_update_saddr(struct sock *sk, void *saddr, int family)
865 if (family == AF_INET) {
866 inet_sk(sk)->inet_saddr = *(__be32 *)saddr;
867 sk_rcv_saddr_set(sk, inet_sk(sk)->inet_saddr);
869 #if IS_ENABLED(CONFIG_IPV6)
871 sk->sk_v6_rcv_saddr = *(struct in6_addr *)saddr;
876 static int __inet_bhash2_update_saddr(struct sock *sk, void *saddr, int family, bool reset)
878 struct inet_hashinfo *hinfo = tcp_or_dccp_get_hashinfo(sk);
879 struct inet_bind_hashbucket *head, *head2;
880 struct inet_bind2_bucket *tb2, *new_tb2;
881 int l3mdev = inet_sk_bound_l3mdev(sk);
882 int port = inet_sk(sk)->inet_num;
883 struct net *net = sock_net(sk);
886 if (!inet_csk(sk)->icsk_bind2_hash) {
887 /* Not bind()ed before. */
889 inet_reset_saddr(sk);
891 inet_update_saddr(sk, saddr, family);
896 /* Allocate a bind2 bucket ahead of time to avoid permanently putting
897 * the bhash2 table in an inconsistent state if a new tb2 bucket
900 new_tb2 = kmem_cache_alloc(hinfo->bind2_bucket_cachep, GFP_ATOMIC);
903 /* The (INADDR_ANY, port) bucket might have already
904 * been freed, then we cannot fixup icsk_bind2_hash,
905 * so we give up and unlink sk from bhash/bhash2 not
906 * to leave inconsistency in bhash2.
909 inet_reset_saddr(sk);
915 bhash = inet_bhashfn(net, port, hinfo->bhash_size);
916 head = &hinfo->bhash[bhash];
917 head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
919 /* If we change saddr locklessly, another thread
920 * iterating over bhash might see corrupted address.
922 spin_lock_bh(&head->lock);
924 spin_lock(&head2->lock);
925 __sk_del_bind2_node(sk);
926 inet_bind2_bucket_destroy(hinfo->bind2_bucket_cachep, inet_csk(sk)->icsk_bind2_hash);
927 spin_unlock(&head2->lock);
930 inet_reset_saddr(sk);
932 inet_update_saddr(sk, saddr, family);
934 head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
936 spin_lock(&head2->lock);
937 tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, sk);
940 inet_bind2_bucket_init(tb2, net, head2, port, l3mdev, sk);
942 sk_add_bind2_node(sk, &tb2->owners);
943 inet_csk(sk)->icsk_bind2_hash = tb2;
944 spin_unlock(&head2->lock);
946 spin_unlock_bh(&head->lock);
949 kmem_cache_free(hinfo->bind2_bucket_cachep, new_tb2);
954 int inet_bhash2_update_saddr(struct sock *sk, void *saddr, int family)
956 return __inet_bhash2_update_saddr(sk, saddr, family, false);
958 EXPORT_SYMBOL_GPL(inet_bhash2_update_saddr);
960 void inet_bhash2_reset_saddr(struct sock *sk)
962 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
963 __inet_bhash2_update_saddr(sk, NULL, 0, true);
965 EXPORT_SYMBOL_GPL(inet_bhash2_reset_saddr);
967 /* RFC 6056 3.3.4. Algorithm 4: Double-Hash Port Selection Algorithm
968 * Note that we use 32bit integers (vs RFC 'short integers')
969 * because 2^16 is not a multiple of num_ephemeral and this
970 * property might be used by clever attacker.
972 * RFC claims using TABLE_LENGTH=10 buckets gives an improvement, though
973 * attacks were since demonstrated, thus we use 65536 by default instead
974 * to really give more isolation and privacy, at the expense of 256kB
977 #define INET_TABLE_PERTURB_SIZE (1 << CONFIG_INET_TABLE_PERTURB_ORDER)
978 static u32 *table_perturb;
980 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
981 struct sock *sk, u64 port_offset,
982 int (*check_established)(struct inet_timewait_death_row *,
983 struct sock *, __u16, struct inet_timewait_sock **))
985 struct inet_hashinfo *hinfo = death_row->hashinfo;
986 struct inet_bind_hashbucket *head, *head2;
987 struct inet_timewait_sock *tw = NULL;
988 int port = inet_sk(sk)->inet_num;
989 struct net *net = sock_net(sk);
990 struct inet_bind2_bucket *tb2;
991 struct inet_bind_bucket *tb;
992 bool tb_created = false;
993 u32 remaining, offset;
994 int ret, i, low, high;
1000 ret = check_established(death_row, sk, port, NULL);
1005 l3mdev = inet_sk_bound_l3mdev(sk);
1007 inet_sk_get_local_port_range(sk, &low, &high);
1008 high++; /* [32768, 60999] -> [32768, 61000[ */
1009 remaining = high - low;
1010 if (likely(remaining > 1))
1013 get_random_sleepable_once(table_perturb,
1014 INET_TABLE_PERTURB_SIZE * sizeof(*table_perturb));
1015 index = port_offset & (INET_TABLE_PERTURB_SIZE - 1);
1017 offset = READ_ONCE(table_perturb[index]) + (port_offset >> 32);
1018 offset %= remaining;
1020 /* In first pass we try ports of @low parity.
1021 * inet_csk_get_port() does the opposite choice.
1025 port = low + offset;
1026 for (i = 0; i < remaining; i += 2, port += 2) {
1027 if (unlikely(port >= high))
1029 if (inet_is_local_reserved_port(net, port))
1031 head = &hinfo->bhash[inet_bhashfn(net, port,
1032 hinfo->bhash_size)];
1033 spin_lock_bh(&head->lock);
1035 /* Does not bother with rcv_saddr checks, because
1036 * the established check is already unique enough.
1038 inet_bind_bucket_for_each(tb, &head->chain) {
1039 if (inet_bind_bucket_match(tb, net, port, l3mdev)) {
1040 if (tb->fastreuse >= 0 ||
1041 tb->fastreuseport >= 0)
1043 WARN_ON(hlist_empty(&tb->owners));
1044 if (!check_established(death_row, sk,
1051 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
1052 net, head, port, l3mdev);
1054 spin_unlock_bh(&head->lock);
1059 tb->fastreuseport = -1;
1062 spin_unlock_bh(&head->lock);
1067 if ((offset & 1) && remaining > 1)
1068 goto other_parity_scan;
1070 return -EADDRNOTAVAIL;
1073 /* Find the corresponding tb2 bucket since we need to
1074 * add the socket to the bhash2 table as well
1076 head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
1077 spin_lock(&head2->lock);
1079 tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, sk);
1081 tb2 = inet_bind2_bucket_create(hinfo->bind2_bucket_cachep, net,
1082 head2, port, l3mdev, sk);
1087 /* Here we want to add a little bit of randomness to the next source
1088 * port that will be chosen. We use a max() with a random here so that
1089 * on low contention the randomness is maximal and on high contention
1090 * it may be inexistent.
1092 i = max_t(int, i, prandom_u32_max(8) * 2);
1093 WRITE_ONCE(table_perturb[index], READ_ONCE(table_perturb[index]) + i + 2);
1095 /* Head lock still held and bh's disabled */
1096 inet_bind_hash(sk, tb, tb2, port);
1098 if (sk_unhashed(sk)) {
1099 inet_sk(sk)->inet_sport = htons(port);
1100 inet_ehash_nolisten(sk, (struct sock *)tw, NULL);
1103 inet_twsk_bind_unhash(tw, hinfo);
1105 spin_unlock(&head2->lock);
1106 spin_unlock(&head->lock);
1109 inet_twsk_deschedule_put(tw);
1114 if (sk_hashed(sk)) {
1115 spinlock_t *lock = inet_ehash_lockp(hinfo, sk->sk_hash);
1117 sock_prot_inuse_add(net, sk->sk_prot, -1);
1120 __sk_nulls_del_node_init_rcu(sk);
1124 inet_sk(sk)->inet_sport = 0;
1125 inet_sk(sk)->inet_num = 0;
1128 inet_twsk_bind_unhash(tw, hinfo);
1131 spin_unlock(&head2->lock);
1133 inet_bind_bucket_destroy(hinfo->bind_bucket_cachep, tb);
1134 spin_unlock(&head->lock);
1137 inet_twsk_deschedule_put(tw);
1145 * Bind a port for a connect operation and hash it.
1147 int inet_hash_connect(struct inet_timewait_death_row *death_row,
1150 u64 port_offset = 0;
1152 if (!inet_sk(sk)->inet_num)
1153 port_offset = inet_sk_port_offset(sk);
1154 return __inet_hash_connect(death_row, sk, port_offset,
1155 __inet_check_established);
1157 EXPORT_SYMBOL_GPL(inet_hash_connect);
1159 static void init_hashinfo_lhash2(struct inet_hashinfo *h)
1163 for (i = 0; i <= h->lhash2_mask; i++) {
1164 spin_lock_init(&h->lhash2[i].lock);
1165 INIT_HLIST_NULLS_HEAD(&h->lhash2[i].nulls_head,
1166 i + LISTENING_NULLS_BASE);
1170 void __init inet_hashinfo2_init(struct inet_hashinfo *h, const char *name,
1171 unsigned long numentries, int scale,
1172 unsigned long low_limit,
1173 unsigned long high_limit)
1175 h->lhash2 = alloc_large_system_hash(name,
1184 init_hashinfo_lhash2(h);
1186 /* this one is used for source ports of outgoing connections */
1187 table_perturb = alloc_large_system_hash("Table-perturb",
1188 sizeof(*table_perturb),
1189 INET_TABLE_PERTURB_SIZE,
1191 INET_TABLE_PERTURB_SIZE,
1192 INET_TABLE_PERTURB_SIZE);
1195 int inet_hashinfo2_init_mod(struct inet_hashinfo *h)
1197 h->lhash2 = kmalloc_array(INET_LHTABLE_SIZE, sizeof(*h->lhash2), GFP_KERNEL);
1201 h->lhash2_mask = INET_LHTABLE_SIZE - 1;
1202 /* INET_LHTABLE_SIZE must be a power of 2 */
1203 BUG_ON(INET_LHTABLE_SIZE & h->lhash2_mask);
1205 init_hashinfo_lhash2(h);
1208 EXPORT_SYMBOL_GPL(inet_hashinfo2_init_mod);
1210 int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
1212 unsigned int locksz = sizeof(spinlock_t);
1213 unsigned int i, nblocks = 1;
1216 /* allocate 2 cache lines or at least one spinlock per cpu */
1217 nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U);
1218 nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
1220 /* no more locks than number of hash buckets */
1221 nblocks = min(nblocks, hashinfo->ehash_mask + 1);
1223 hashinfo->ehash_locks = kvmalloc_array(nblocks, locksz, GFP_KERNEL);
1224 if (!hashinfo->ehash_locks)
1227 for (i = 0; i < nblocks; i++)
1228 spin_lock_init(&hashinfo->ehash_locks[i]);
1230 hashinfo->ehash_locks_mask = nblocks - 1;
1233 EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);
1235 struct inet_hashinfo *inet_pernet_hashinfo_alloc(struct inet_hashinfo *hashinfo,
1236 unsigned int ehash_entries)
1238 struct inet_hashinfo *new_hashinfo;
1241 new_hashinfo = kmemdup(hashinfo, sizeof(*hashinfo), GFP_KERNEL);
1245 new_hashinfo->ehash = vmalloc_huge(ehash_entries * sizeof(struct inet_ehash_bucket),
1246 GFP_KERNEL_ACCOUNT);
1247 if (!new_hashinfo->ehash)
1250 new_hashinfo->ehash_mask = ehash_entries - 1;
1252 if (inet_ehash_locks_alloc(new_hashinfo))
1255 for (i = 0; i < ehash_entries; i++)
1256 INIT_HLIST_NULLS_HEAD(&new_hashinfo->ehash[i].chain, i);
1258 new_hashinfo->pernet = true;
1260 return new_hashinfo;
1263 vfree(new_hashinfo->ehash);
1265 kfree(new_hashinfo);
1269 EXPORT_SYMBOL_GPL(inet_pernet_hashinfo_alloc);
1271 void inet_pernet_hashinfo_free(struct inet_hashinfo *hashinfo)
1273 if (!hashinfo->pernet)
1276 inet_ehash_locks_free(hashinfo);
1277 vfree(hashinfo->ehash);
1280 EXPORT_SYMBOL_GPL(inet_pernet_hashinfo_free);