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 * Support for INET connection oriented protocols.
9 * Authors: See the TCP sources
12 #include <linux/module.h>
13 #include <linux/jhash.h>
15 #include <net/inet_connection_sock.h>
16 #include <net/inet_hashtables.h>
17 #include <net/inet_timewait_sock.h>
19 #include <net/route.h>
20 #include <net/tcp_states.h>
23 #include <net/sock_reuseport.h>
24 #include <net/addrconf.h>
26 #if IS_ENABLED(CONFIG_IPV6)
27 /* match_sk*_wildcard == true: IPV6_ADDR_ANY equals to any IPv6 addresses
28 * if IPv6 only, and any IPv4 addresses
30 * match_sk*_wildcard == false: addresses must be exactly the same, i.e.
31 * IPV6_ADDR_ANY only equals to IPV6_ADDR_ANY,
32 * and 0.0.0.0 equals to 0.0.0.0 only
34 static bool ipv6_rcv_saddr_equal(const struct in6_addr *sk1_rcv_saddr6,
35 const struct in6_addr *sk2_rcv_saddr6,
36 __be32 sk1_rcv_saddr, __be32 sk2_rcv_saddr,
37 bool sk1_ipv6only, bool sk2_ipv6only,
38 bool match_sk1_wildcard,
39 bool match_sk2_wildcard)
41 int addr_type = ipv6_addr_type(sk1_rcv_saddr6);
42 int addr_type2 = sk2_rcv_saddr6 ? ipv6_addr_type(sk2_rcv_saddr6) : IPV6_ADDR_MAPPED;
44 /* if both are mapped, treat as IPv4 */
45 if (addr_type == IPV6_ADDR_MAPPED && addr_type2 == IPV6_ADDR_MAPPED) {
47 if (sk1_rcv_saddr == sk2_rcv_saddr)
49 return (match_sk1_wildcard && !sk1_rcv_saddr) ||
50 (match_sk2_wildcard && !sk2_rcv_saddr);
55 if (addr_type == IPV6_ADDR_ANY && addr_type2 == IPV6_ADDR_ANY)
58 if (addr_type2 == IPV6_ADDR_ANY && match_sk2_wildcard &&
59 !(sk2_ipv6only && addr_type == IPV6_ADDR_MAPPED))
62 if (addr_type == IPV6_ADDR_ANY && match_sk1_wildcard &&
63 !(sk1_ipv6only && addr_type2 == IPV6_ADDR_MAPPED))
67 ipv6_addr_equal(sk1_rcv_saddr6, sk2_rcv_saddr6))
74 /* match_sk*_wildcard == true: 0.0.0.0 equals to any IPv4 addresses
75 * match_sk*_wildcard == false: addresses must be exactly the same, i.e.
76 * 0.0.0.0 only equals to 0.0.0.0
78 static bool ipv4_rcv_saddr_equal(__be32 sk1_rcv_saddr, __be32 sk2_rcv_saddr,
79 bool sk2_ipv6only, bool match_sk1_wildcard,
80 bool match_sk2_wildcard)
83 if (sk1_rcv_saddr == sk2_rcv_saddr)
85 return (match_sk1_wildcard && !sk1_rcv_saddr) ||
86 (match_sk2_wildcard && !sk2_rcv_saddr);
91 bool inet_rcv_saddr_equal(const struct sock *sk, const struct sock *sk2,
94 #if IS_ENABLED(CONFIG_IPV6)
95 if (sk->sk_family == AF_INET6)
96 return ipv6_rcv_saddr_equal(&sk->sk_v6_rcv_saddr,
105 return ipv4_rcv_saddr_equal(sk->sk_rcv_saddr, sk2->sk_rcv_saddr,
106 ipv6_only_sock(sk2), match_wildcard,
109 EXPORT_SYMBOL(inet_rcv_saddr_equal);
111 bool inet_rcv_saddr_any(const struct sock *sk)
113 #if IS_ENABLED(CONFIG_IPV6)
114 if (sk->sk_family == AF_INET6)
115 return ipv6_addr_any(&sk->sk_v6_rcv_saddr);
117 return !sk->sk_rcv_saddr;
120 static bool use_bhash2_on_bind(const struct sock *sk)
122 #if IS_ENABLED(CONFIG_IPV6)
125 if (sk->sk_family == AF_INET6) {
126 addr_type = ipv6_addr_type(&sk->sk_v6_rcv_saddr);
127 return addr_type != IPV6_ADDR_ANY &&
128 addr_type != IPV6_ADDR_MAPPED;
131 return sk->sk_rcv_saddr != htonl(INADDR_ANY);
134 static u32 get_bhash2_nulladdr_hash(const struct sock *sk, struct net *net,
137 #if IS_ENABLED(CONFIG_IPV6)
138 struct in6_addr nulladdr = {};
140 if (sk->sk_family == AF_INET6)
141 return ipv6_portaddr_hash(net, &nulladdr, port);
143 return ipv4_portaddr_hash(net, 0, port);
146 void inet_get_local_port_range(struct net *net, int *low, int *high)
151 seq = read_seqbegin(&net->ipv4.ip_local_ports.lock);
153 *low = net->ipv4.ip_local_ports.range[0];
154 *high = net->ipv4.ip_local_ports.range[1];
155 } while (read_seqretry(&net->ipv4.ip_local_ports.lock, seq));
157 EXPORT_SYMBOL(inet_get_local_port_range);
159 static bool bind_conflict_exist(const struct sock *sk, struct sock *sk2,
160 kuid_t sk_uid, bool relax,
161 bool reuseport_cb_ok, bool reuseport_ok)
168 bound_dev_if2 = READ_ONCE(sk2->sk_bound_dev_if);
170 if (!sk->sk_bound_dev_if || !bound_dev_if2 ||
171 sk->sk_bound_dev_if == bound_dev_if2) {
172 if (sk->sk_reuse && sk2->sk_reuse &&
173 sk2->sk_state != TCP_LISTEN) {
174 if (!relax || (!reuseport_ok && sk->sk_reuseport &&
175 sk2->sk_reuseport && reuseport_cb_ok &&
176 (sk2->sk_state == TCP_TIME_WAIT ||
177 uid_eq(sk_uid, sock_i_uid(sk2)))))
179 } else if (!reuseport_ok || !sk->sk_reuseport ||
180 !sk2->sk_reuseport || !reuseport_cb_ok ||
181 (sk2->sk_state != TCP_TIME_WAIT &&
182 !uid_eq(sk_uid, sock_i_uid(sk2)))) {
189 static bool check_bhash2_conflict(const struct sock *sk,
190 struct inet_bind2_bucket *tb2, kuid_t sk_uid,
191 bool relax, bool reuseport_cb_ok,
196 sk_for_each_bound_bhash2(sk2, &tb2->owners) {
197 if (sk->sk_family == AF_INET && ipv6_only_sock(sk2))
200 if (bind_conflict_exist(sk, sk2, sk_uid, relax,
201 reuseport_cb_ok, reuseport_ok))
207 /* This should be called only when the corresponding inet_bind_bucket spinlock
210 static int inet_csk_bind_conflict(const struct sock *sk, int port,
211 struct inet_bind_bucket *tb,
212 struct inet_bind2_bucket *tb2, /* may be null */
213 bool relax, bool reuseport_ok)
215 struct inet_hashinfo *hinfo = sk->sk_prot->h.hashinfo;
216 kuid_t uid = sock_i_uid((struct sock *)sk);
217 struct sock_reuseport *reuseport_cb;
218 struct inet_bind2_hashbucket *head2;
219 bool reuseport_cb_ok;
226 reuseport_cb = rcu_dereference(sk->sk_reuseport_cb);
227 /* paired with WRITE_ONCE() in __reuseport_(add|detach)_closed_sock */
228 reuseport_cb_ok = !reuseport_cb || READ_ONCE(reuseport_cb->num_closed_socks);
232 * Unlike other sk lookup places we do not check
233 * for sk_net here, since _all_ the socks listed
234 * in tb->owners and tb2->owners list belong
238 if (!use_bhash2_on_bind(sk)) {
239 sk_for_each_bound(sk2, &tb->owners)
240 if (bind_conflict_exist(sk, sk2, uid, relax,
241 reuseport_cb_ok, reuseport_ok) &&
242 inet_rcv_saddr_equal(sk, sk2, true))
248 if (tb2 && check_bhash2_conflict(sk, tb2, uid, relax, reuseport_cb_ok,
254 /* check there's no conflict with an existing IPV6_ADDR_ANY (if ipv6) or
255 * INADDR_ANY (if ipv4) socket.
257 hash = get_bhash2_nulladdr_hash(sk, net, port);
258 head2 = &hinfo->bhash2[hash & (hinfo->bhash_size - 1)];
260 l3mdev = inet_sk_bound_l3mdev(sk);
261 inet_bind_bucket_for_each(tb2, &head2->chain)
262 if (check_bind2_bucket_match_nulladdr(tb2, net, port, l3mdev, sk))
265 if (tb2 && check_bhash2_conflict(sk, tb2, uid, relax, reuseport_cb_ok,
273 * Find an open port number for the socket. Returns with the
274 * inet_bind_hashbucket lock held.
276 static struct inet_bind_hashbucket *
277 inet_csk_find_open_port(struct sock *sk, struct inet_bind_bucket **tb_ret,
278 struct inet_bind2_bucket **tb2_ret,
279 struct inet_bind2_hashbucket **head2_ret, int *port_ret)
281 struct inet_hashinfo *hinfo = sk->sk_prot->h.hashinfo;
282 struct inet_bind2_hashbucket *head2;
283 struct inet_bind_hashbucket *head;
284 struct net *net = sock_net(sk);
285 int i, low, high, attempt_half;
286 struct inet_bind2_bucket *tb2;
287 struct inet_bind_bucket *tb;
288 u32 remaining, offset;
293 l3mdev = inet_sk_bound_l3mdev(sk);
295 attempt_half = (sk->sk_reuse == SK_CAN_REUSE) ? 1 : 0;
297 inet_get_local_port_range(net, &low, &high);
298 high++; /* [32768, 60999] -> [32768, 61000[ */
302 int half = low + (((high - low) >> 2) << 1);
304 if (attempt_half == 1)
309 remaining = high - low;
310 if (likely(remaining > 1))
313 offset = prandom_u32() % remaining;
314 /* __inet_hash_connect() favors ports having @low parity
315 * We do the opposite to not pollute connect() users.
321 for (i = 0; i < remaining; i += 2, port += 2) {
322 if (unlikely(port >= high))
324 if (inet_is_local_reserved_port(net, port))
326 head = &hinfo->bhash[inet_bhashfn(net, port,
328 spin_lock_bh(&head->lock);
329 tb2 = inet_bind2_bucket_find(hinfo, net, port, l3mdev, sk,
331 inet_bind_bucket_for_each(tb, &head->chain)
332 if (check_bind_bucket_match(tb, net, port, l3mdev)) {
333 if (!inet_csk_bind_conflict(sk, port, tb, tb2,
341 spin_unlock_bh(&head->lock);
347 goto other_parity_scan;
349 if (attempt_half == 1) {
350 /* OK we now try the upper half of the range */
352 goto other_half_scan;
355 if (net->ipv4.sysctl_ip_autobind_reuse && !relax) {
356 /* We still have a chance to connect to different destinations */
358 goto ports_exhausted;
369 static inline int sk_reuseport_match(struct inet_bind_bucket *tb,
372 kuid_t uid = sock_i_uid(sk);
374 if (tb->fastreuseport <= 0)
376 if (!sk->sk_reuseport)
378 if (rcu_access_pointer(sk->sk_reuseport_cb))
380 if (!uid_eq(tb->fastuid, uid))
382 /* We only need to check the rcv_saddr if this tb was once marked
383 * without fastreuseport and then was reset, as we can only know that
384 * the fast_*rcv_saddr doesn't have any conflicts with the socks on the
387 if (tb->fastreuseport == FASTREUSEPORT_ANY)
389 #if IS_ENABLED(CONFIG_IPV6)
390 if (tb->fast_sk_family == AF_INET6)
391 return ipv6_rcv_saddr_equal(&tb->fast_v6_rcv_saddr,
396 ipv6_only_sock(sk), true, false);
398 return ipv4_rcv_saddr_equal(tb->fast_rcv_saddr, sk->sk_rcv_saddr,
399 ipv6_only_sock(sk), true, false);
402 void inet_csk_update_fastreuse(struct inet_bind_bucket *tb,
405 kuid_t uid = sock_i_uid(sk);
406 bool reuse = sk->sk_reuse && sk->sk_state != TCP_LISTEN;
408 if (hlist_empty(&tb->owners)) {
409 tb->fastreuse = reuse;
410 if (sk->sk_reuseport) {
411 tb->fastreuseport = FASTREUSEPORT_ANY;
413 tb->fast_rcv_saddr = sk->sk_rcv_saddr;
414 tb->fast_ipv6_only = ipv6_only_sock(sk);
415 tb->fast_sk_family = sk->sk_family;
416 #if IS_ENABLED(CONFIG_IPV6)
417 tb->fast_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
420 tb->fastreuseport = 0;
425 if (sk->sk_reuseport) {
426 /* We didn't match or we don't have fastreuseport set on
427 * the tb, but we have sk_reuseport set on this socket
428 * and we know that there are no bind conflicts with
429 * this socket in this tb, so reset our tb's reuseport
430 * settings so that any subsequent sockets that match
431 * our current socket will be put on the fast path.
433 * If we reset we need to set FASTREUSEPORT_STRICT so we
434 * do extra checking for all subsequent sk_reuseport
437 if (!sk_reuseport_match(tb, sk)) {
438 tb->fastreuseport = FASTREUSEPORT_STRICT;
440 tb->fast_rcv_saddr = sk->sk_rcv_saddr;
441 tb->fast_ipv6_only = ipv6_only_sock(sk);
442 tb->fast_sk_family = sk->sk_family;
443 #if IS_ENABLED(CONFIG_IPV6)
444 tb->fast_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
448 tb->fastreuseport = 0;
453 /* Obtain a reference to a local port for the given sock,
454 * if snum is zero it means select any available local port.
455 * We try to allocate an odd port (and leave even ports for connect())
457 int inet_csk_get_port(struct sock *sk, unsigned short snum)
459 bool reuse = sk->sk_reuse && sk->sk_state != TCP_LISTEN;
460 struct inet_hashinfo *hinfo = sk->sk_prot->h.hashinfo;
461 bool bhash_created = false, bhash2_created = false;
462 struct inet_bind2_bucket *tb2 = NULL;
463 struct inet_bind2_hashbucket *head2;
464 struct inet_bind_bucket *tb = NULL;
465 struct inet_bind_hashbucket *head;
466 struct net *net = sock_net(sk);
467 int ret = 1, port = snum;
468 bool found_port = false;
471 l3mdev = inet_sk_bound_l3mdev(sk);
474 head = inet_csk_find_open_port(sk, &tb, &tb2, &head2, &port);
481 head = &hinfo->bhash[inet_bhashfn(net, port,
483 spin_lock_bh(&head->lock);
484 inet_bind_bucket_for_each(tb, &head->chain)
485 if (check_bind_bucket_match(tb, net, port, l3mdev))
488 tb2 = inet_bind2_bucket_find(hinfo, net, port, l3mdev, sk,
493 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep, net,
497 bhash_created = true;
501 tb2 = inet_bind2_bucket_create(hinfo->bind2_bucket_cachep,
502 net, head2, port, l3mdev, sk);
505 bhash2_created = true;
508 /* If we had to find an open port, we already checked for conflicts */
509 if (!found_port && !hlist_empty(&tb->owners)) {
510 if (sk->sk_reuse == SK_FORCE_REUSE)
513 if ((tb->fastreuse > 0 && reuse) ||
514 sk_reuseport_match(tb, sk))
516 if (inet_csk_bind_conflict(sk, port, tb, tb2, true, true))
520 inet_csk_update_fastreuse(tb, sk);
522 if (!inet_csk(sk)->icsk_bind_hash)
523 inet_bind_hash(sk, tb, tb2, port);
524 WARN_ON(inet_csk(sk)->icsk_bind_hash != tb);
525 WARN_ON(inet_csk(sk)->icsk_bind2_hash != tb2);
531 inet_bind_bucket_destroy(hinfo->bind_bucket_cachep, tb);
533 inet_bind2_bucket_destroy(hinfo->bind2_bucket_cachep,
536 spin_unlock_bh(&head->lock);
539 EXPORT_SYMBOL_GPL(inet_csk_get_port);
542 * Wait for an incoming connection, avoid race conditions. This must be called
543 * with the socket locked.
545 static int inet_csk_wait_for_connect(struct sock *sk, long timeo)
547 struct inet_connection_sock *icsk = inet_csk(sk);
552 * True wake-one mechanism for incoming connections: only
553 * one process gets woken up, not the 'whole herd'.
554 * Since we do not 'race & poll' for established sockets
555 * anymore, the common case will execute the loop only once.
557 * Subtle issue: "add_wait_queue_exclusive()" will be added
558 * after any current non-exclusive waiters, and we know that
559 * it will always _stay_ after any new non-exclusive waiters
560 * because all non-exclusive waiters are added at the
561 * beginning of the wait-queue. As such, it's ok to "drop"
562 * our exclusiveness temporarily when we get woken up without
563 * having to remove and re-insert us on the wait queue.
566 prepare_to_wait_exclusive(sk_sleep(sk), &wait,
569 if (reqsk_queue_empty(&icsk->icsk_accept_queue))
570 timeo = schedule_timeout(timeo);
571 sched_annotate_sleep();
574 if (!reqsk_queue_empty(&icsk->icsk_accept_queue))
577 if (sk->sk_state != TCP_LISTEN)
579 err = sock_intr_errno(timeo);
580 if (signal_pending(current))
586 finish_wait(sk_sleep(sk), &wait);
591 * This will accept the next outstanding connection.
593 struct sock *inet_csk_accept(struct sock *sk, int flags, int *err, bool kern)
595 struct inet_connection_sock *icsk = inet_csk(sk);
596 struct request_sock_queue *queue = &icsk->icsk_accept_queue;
597 struct request_sock *req;
603 /* We need to make sure that this socket is listening,
604 * and that it has something pending.
607 if (sk->sk_state != TCP_LISTEN)
610 /* Find already established connection */
611 if (reqsk_queue_empty(queue)) {
612 long timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
614 /* If this is a non blocking socket don't sleep */
619 error = inet_csk_wait_for_connect(sk, timeo);
623 req = reqsk_queue_remove(queue, sk);
626 if (sk->sk_protocol == IPPROTO_TCP &&
627 tcp_rsk(req)->tfo_listener) {
628 spin_lock_bh(&queue->fastopenq.lock);
629 if (tcp_rsk(req)->tfo_listener) {
630 /* We are still waiting for the final ACK from 3WHS
631 * so can't free req now. Instead, we set req->sk to
632 * NULL to signify that the child socket is taken
633 * so reqsk_fastopen_remove() will free the req
634 * when 3WHS finishes (or is aborted).
639 spin_unlock_bh(&queue->fastopenq.lock);
644 if (newsk && mem_cgroup_sockets_enabled) {
647 /* atomically get the memory usage, set and charge the
652 /* The socket has not been accepted yet, no need to look at
653 * newsk->sk_wmem_queued.
655 amt = sk_mem_pages(newsk->sk_forward_alloc +
656 atomic_read(&newsk->sk_rmem_alloc));
657 mem_cgroup_sk_alloc(newsk);
658 if (newsk->sk_memcg && amt)
659 mem_cgroup_charge_skmem(newsk->sk_memcg, amt,
660 GFP_KERNEL | __GFP_NOFAIL);
673 EXPORT_SYMBOL(inet_csk_accept);
676 * Using different timers for retransmit, delayed acks and probes
677 * We may wish use just one timer maintaining a list of expire jiffies
680 void inet_csk_init_xmit_timers(struct sock *sk,
681 void (*retransmit_handler)(struct timer_list *t),
682 void (*delack_handler)(struct timer_list *t),
683 void (*keepalive_handler)(struct timer_list *t))
685 struct inet_connection_sock *icsk = inet_csk(sk);
687 timer_setup(&icsk->icsk_retransmit_timer, retransmit_handler, 0);
688 timer_setup(&icsk->icsk_delack_timer, delack_handler, 0);
689 timer_setup(&sk->sk_timer, keepalive_handler, 0);
690 icsk->icsk_pending = icsk->icsk_ack.pending = 0;
692 EXPORT_SYMBOL(inet_csk_init_xmit_timers);
694 void inet_csk_clear_xmit_timers(struct sock *sk)
696 struct inet_connection_sock *icsk = inet_csk(sk);
698 icsk->icsk_pending = icsk->icsk_ack.pending = 0;
700 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
701 sk_stop_timer(sk, &icsk->icsk_delack_timer);
702 sk_stop_timer(sk, &sk->sk_timer);
704 EXPORT_SYMBOL(inet_csk_clear_xmit_timers);
706 void inet_csk_delete_keepalive_timer(struct sock *sk)
708 sk_stop_timer(sk, &sk->sk_timer);
710 EXPORT_SYMBOL(inet_csk_delete_keepalive_timer);
712 void inet_csk_reset_keepalive_timer(struct sock *sk, unsigned long len)
714 sk_reset_timer(sk, &sk->sk_timer, jiffies + len);
716 EXPORT_SYMBOL(inet_csk_reset_keepalive_timer);
718 struct dst_entry *inet_csk_route_req(const struct sock *sk,
720 const struct request_sock *req)
722 const struct inet_request_sock *ireq = inet_rsk(req);
723 struct net *net = read_pnet(&ireq->ireq_net);
724 struct ip_options_rcu *opt;
728 opt = rcu_dereference(ireq->ireq_opt);
730 flowi4_init_output(fl4, ireq->ir_iif, ireq->ir_mark,
731 RT_CONN_FLAGS(sk), RT_SCOPE_UNIVERSE,
732 sk->sk_protocol, inet_sk_flowi_flags(sk),
733 (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr,
734 ireq->ir_loc_addr, ireq->ir_rmt_port,
735 htons(ireq->ir_num), sk->sk_uid);
736 security_req_classify_flow(req, flowi4_to_flowi_common(fl4));
737 rt = ip_route_output_flow(net, fl4, sk);
740 if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway)
749 __IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
752 EXPORT_SYMBOL_GPL(inet_csk_route_req);
754 struct dst_entry *inet_csk_route_child_sock(const struct sock *sk,
756 const struct request_sock *req)
758 const struct inet_request_sock *ireq = inet_rsk(req);
759 struct net *net = read_pnet(&ireq->ireq_net);
760 struct inet_sock *newinet = inet_sk(newsk);
761 struct ip_options_rcu *opt;
765 opt = rcu_dereference(ireq->ireq_opt);
766 fl4 = &newinet->cork.fl.u.ip4;
768 flowi4_init_output(fl4, ireq->ir_iif, ireq->ir_mark,
769 RT_CONN_FLAGS(sk), RT_SCOPE_UNIVERSE,
770 sk->sk_protocol, inet_sk_flowi_flags(sk),
771 (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr,
772 ireq->ir_loc_addr, ireq->ir_rmt_port,
773 htons(ireq->ir_num), sk->sk_uid);
774 security_req_classify_flow(req, flowi4_to_flowi_common(fl4));
775 rt = ip_route_output_flow(net, fl4, sk);
778 if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway)
785 __IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
788 EXPORT_SYMBOL_GPL(inet_csk_route_child_sock);
790 /* Decide when to expire the request and when to resend SYN-ACK */
791 static void syn_ack_recalc(struct request_sock *req,
792 const int max_syn_ack_retries,
793 const u8 rskq_defer_accept,
794 int *expire, int *resend)
796 if (!rskq_defer_accept) {
797 *expire = req->num_timeout >= max_syn_ack_retries;
801 *expire = req->num_timeout >= max_syn_ack_retries &&
802 (!inet_rsk(req)->acked || req->num_timeout >= rskq_defer_accept);
803 /* Do not resend while waiting for data after ACK,
804 * start to resend on end of deferring period to give
805 * last chance for data or ACK to create established socket.
807 *resend = !inet_rsk(req)->acked ||
808 req->num_timeout >= rskq_defer_accept - 1;
811 int inet_rtx_syn_ack(const struct sock *parent, struct request_sock *req)
813 int err = req->rsk_ops->rtx_syn_ack(parent, req);
819 EXPORT_SYMBOL(inet_rtx_syn_ack);
821 static struct request_sock *inet_reqsk_clone(struct request_sock *req,
824 struct sock *req_sk, *nreq_sk;
825 struct request_sock *nreq;
827 nreq = kmem_cache_alloc(req->rsk_ops->slab, GFP_ATOMIC | __GFP_NOWARN);
829 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMIGRATEREQFAILURE);
831 /* paired with refcount_inc_not_zero() in reuseport_migrate_sock() */
836 req_sk = req_to_sk(req);
837 nreq_sk = req_to_sk(nreq);
839 memcpy(nreq_sk, req_sk,
840 offsetof(struct sock, sk_dontcopy_begin));
841 memcpy(&nreq_sk->sk_dontcopy_end, &req_sk->sk_dontcopy_end,
842 req->rsk_ops->obj_size - offsetof(struct sock, sk_dontcopy_end));
844 sk_node_init(&nreq_sk->sk_node);
845 nreq_sk->sk_tx_queue_mapping = req_sk->sk_tx_queue_mapping;
846 #ifdef CONFIG_SOCK_RX_QUEUE_MAPPING
847 nreq_sk->sk_rx_queue_mapping = req_sk->sk_rx_queue_mapping;
849 nreq_sk->sk_incoming_cpu = req_sk->sk_incoming_cpu;
851 nreq->rsk_listener = sk;
853 /* We need not acquire fastopenq->lock
854 * because the child socket is locked in inet_csk_listen_stop().
856 if (sk->sk_protocol == IPPROTO_TCP && tcp_rsk(nreq)->tfo_listener)
857 rcu_assign_pointer(tcp_sk(nreq->sk)->fastopen_rsk, nreq);
862 static void reqsk_queue_migrated(struct request_sock_queue *queue,
863 const struct request_sock *req)
865 if (req->num_timeout == 0)
866 atomic_inc(&queue->young);
867 atomic_inc(&queue->qlen);
870 static void reqsk_migrate_reset(struct request_sock *req)
872 req->saved_syn = NULL;
873 #if IS_ENABLED(CONFIG_IPV6)
874 inet_rsk(req)->ipv6_opt = NULL;
875 inet_rsk(req)->pktopts = NULL;
877 inet_rsk(req)->ireq_opt = NULL;
881 /* return true if req was found in the ehash table */
882 static bool reqsk_queue_unlink(struct request_sock *req)
884 struct inet_hashinfo *hashinfo = req_to_sk(req)->sk_prot->h.hashinfo;
887 if (sk_hashed(req_to_sk(req))) {
888 spinlock_t *lock = inet_ehash_lockp(hashinfo, req->rsk_hash);
891 found = __sk_nulls_del_node_init_rcu(req_to_sk(req));
894 if (timer_pending(&req->rsk_timer) && del_timer_sync(&req->rsk_timer))
899 bool inet_csk_reqsk_queue_drop(struct sock *sk, struct request_sock *req)
901 bool unlinked = reqsk_queue_unlink(req);
904 reqsk_queue_removed(&inet_csk(sk)->icsk_accept_queue, req);
909 EXPORT_SYMBOL(inet_csk_reqsk_queue_drop);
911 void inet_csk_reqsk_queue_drop_and_put(struct sock *sk, struct request_sock *req)
913 inet_csk_reqsk_queue_drop(sk, req);
916 EXPORT_SYMBOL(inet_csk_reqsk_queue_drop_and_put);
918 static void reqsk_timer_handler(struct timer_list *t)
920 struct request_sock *req = from_timer(req, t, rsk_timer);
921 struct request_sock *nreq = NULL, *oreq = req;
922 struct sock *sk_listener = req->rsk_listener;
923 struct inet_connection_sock *icsk;
924 struct request_sock_queue *queue;
926 int max_syn_ack_retries, qlen, expire = 0, resend = 0;
928 if (inet_sk_state_load(sk_listener) != TCP_LISTEN) {
931 nsk = reuseport_migrate_sock(sk_listener, req_to_sk(req), NULL);
935 nreq = inet_reqsk_clone(req, nsk);
939 /* The new timer for the cloned req can decrease the 2
940 * by calling inet_csk_reqsk_queue_drop_and_put(), so
941 * hold another count to prevent use-after-free and
942 * call reqsk_put() just before return.
944 refcount_set(&nreq->rsk_refcnt, 2 + 1);
945 timer_setup(&nreq->rsk_timer, reqsk_timer_handler, TIMER_PINNED);
946 reqsk_queue_migrated(&inet_csk(nsk)->icsk_accept_queue, req);
952 icsk = inet_csk(sk_listener);
953 net = sock_net(sk_listener);
954 max_syn_ack_retries = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_synack_retries;
955 /* Normally all the openreqs are young and become mature
956 * (i.e. converted to established socket) for first timeout.
957 * If synack was not acknowledged for 1 second, it means
958 * one of the following things: synack was lost, ack was lost,
959 * rtt is high or nobody planned to ack (i.e. synflood).
960 * When server is a bit loaded, queue is populated with old
961 * open requests, reducing effective size of queue.
962 * When server is well loaded, queue size reduces to zero
963 * after several minutes of work. It is not synflood,
964 * it is normal operation. The solution is pruning
965 * too old entries overriding normal timeout, when
966 * situation becomes dangerous.
968 * Essentially, we reserve half of room for young
969 * embrions; and abort old ones without pity, if old
970 * ones are about to clog our table.
972 queue = &icsk->icsk_accept_queue;
973 qlen = reqsk_queue_len(queue);
974 if ((qlen << 1) > max(8U, READ_ONCE(sk_listener->sk_max_ack_backlog))) {
975 int young = reqsk_queue_len_young(queue) << 1;
977 while (max_syn_ack_retries > 2) {
980 max_syn_ack_retries--;
984 syn_ack_recalc(req, max_syn_ack_retries, READ_ONCE(queue->rskq_defer_accept),
986 req->rsk_ops->syn_ack_timeout(req);
989 !inet_rtx_syn_ack(sk_listener, req) ||
990 inet_rsk(req)->acked)) {
991 if (req->num_timeout++ == 0)
992 atomic_dec(&queue->young);
993 mod_timer(&req->rsk_timer, jiffies + reqsk_timeout(req, TCP_RTO_MAX));
998 if (!inet_ehash_insert(req_to_sk(nreq), req_to_sk(oreq), NULL)) {
1000 inet_csk_reqsk_queue_drop(sk_listener, nreq);
1004 __NET_INC_STATS(net, LINUX_MIB_TCPMIGRATEREQSUCCESS);
1005 reqsk_migrate_reset(oreq);
1006 reqsk_queue_removed(&inet_csk(oreq->rsk_listener)->icsk_accept_queue, oreq);
1013 /* Even if we can clone the req, we may need not retransmit any more
1014 * SYN+ACKs (nreq->num_timeout > max_syn_ack_retries, etc), or another
1015 * CPU may win the "own_req" race so that inet_ehash_insert() fails.
1018 __NET_INC_STATS(net, LINUX_MIB_TCPMIGRATEREQFAILURE);
1020 reqsk_migrate_reset(nreq);
1021 reqsk_queue_removed(queue, nreq);
1026 inet_csk_reqsk_queue_drop_and_put(oreq->rsk_listener, oreq);
1029 static void reqsk_queue_hash_req(struct request_sock *req,
1030 unsigned long timeout)
1032 timer_setup(&req->rsk_timer, reqsk_timer_handler, TIMER_PINNED);
1033 mod_timer(&req->rsk_timer, jiffies + timeout);
1035 inet_ehash_insert(req_to_sk(req), NULL, NULL);
1036 /* before letting lookups find us, make sure all req fields
1037 * are committed to memory and refcnt initialized.
1040 refcount_set(&req->rsk_refcnt, 2 + 1);
1043 void inet_csk_reqsk_queue_hash_add(struct sock *sk, struct request_sock *req,
1044 unsigned long timeout)
1046 reqsk_queue_hash_req(req, timeout);
1047 inet_csk_reqsk_queue_added(sk);
1049 EXPORT_SYMBOL_GPL(inet_csk_reqsk_queue_hash_add);
1051 static void inet_clone_ulp(const struct request_sock *req, struct sock *newsk,
1052 const gfp_t priority)
1054 struct inet_connection_sock *icsk = inet_csk(newsk);
1056 if (!icsk->icsk_ulp_ops)
1059 if (icsk->icsk_ulp_ops->clone)
1060 icsk->icsk_ulp_ops->clone(req, newsk, priority);
1064 * inet_csk_clone_lock - clone an inet socket, and lock its clone
1065 * @sk: the socket to clone
1066 * @req: request_sock
1067 * @priority: for allocation (%GFP_KERNEL, %GFP_ATOMIC, etc)
1069 * Caller must unlock socket even in error path (bh_unlock_sock(newsk))
1071 struct sock *inet_csk_clone_lock(const struct sock *sk,
1072 const struct request_sock *req,
1073 const gfp_t priority)
1075 struct sock *newsk = sk_clone_lock(sk, priority);
1078 struct inet_connection_sock *newicsk = inet_csk(newsk);
1080 inet_sk_set_state(newsk, TCP_SYN_RECV);
1081 newicsk->icsk_bind_hash = NULL;
1082 newicsk->icsk_bind2_hash = NULL;
1084 inet_sk(newsk)->inet_dport = inet_rsk(req)->ir_rmt_port;
1085 inet_sk(newsk)->inet_num = inet_rsk(req)->ir_num;
1086 inet_sk(newsk)->inet_sport = htons(inet_rsk(req)->ir_num);
1088 /* listeners have SOCK_RCU_FREE, not the children */
1089 sock_reset_flag(newsk, SOCK_RCU_FREE);
1091 inet_sk(newsk)->mc_list = NULL;
1093 newsk->sk_mark = inet_rsk(req)->ir_mark;
1094 atomic64_set(&newsk->sk_cookie,
1095 atomic64_read(&inet_rsk(req)->ir_cookie));
1097 newicsk->icsk_retransmits = 0;
1098 newicsk->icsk_backoff = 0;
1099 newicsk->icsk_probes_out = 0;
1100 newicsk->icsk_probes_tstamp = 0;
1102 /* Deinitialize accept_queue to trap illegal accesses. */
1103 memset(&newicsk->icsk_accept_queue, 0, sizeof(newicsk->icsk_accept_queue));
1105 inet_clone_ulp(req, newsk, priority);
1107 security_inet_csk_clone(newsk, req);
1111 EXPORT_SYMBOL_GPL(inet_csk_clone_lock);
1114 * At this point, there should be no process reference to this
1115 * socket, and thus no user references at all. Therefore we
1116 * can assume the socket waitqueue is inactive and nobody will
1117 * try to jump onto it.
1119 void inet_csk_destroy_sock(struct sock *sk)
1121 WARN_ON(sk->sk_state != TCP_CLOSE);
1122 WARN_ON(!sock_flag(sk, SOCK_DEAD));
1124 /* It cannot be in hash table! */
1125 WARN_ON(!sk_unhashed(sk));
1127 /* If it has not 0 inet_sk(sk)->inet_num, it must be bound */
1128 WARN_ON(inet_sk(sk)->inet_num && !inet_csk(sk)->icsk_bind_hash);
1130 sk->sk_prot->destroy(sk);
1132 sk_stream_kill_queues(sk);
1134 xfrm_sk_free_policy(sk);
1136 sk_refcnt_debug_release(sk);
1138 this_cpu_dec(*sk->sk_prot->orphan_count);
1142 EXPORT_SYMBOL(inet_csk_destroy_sock);
1144 /* This function allows to force a closure of a socket after the call to
1145 * tcp/dccp_create_openreq_child().
1147 void inet_csk_prepare_forced_close(struct sock *sk)
1148 __releases(&sk->sk_lock.slock)
1150 /* sk_clone_lock locked the socket and set refcnt to 2 */
1153 inet_csk_prepare_for_destroy_sock(sk);
1154 inet_sk(sk)->inet_num = 0;
1156 EXPORT_SYMBOL(inet_csk_prepare_forced_close);
1158 int inet_csk_listen_start(struct sock *sk)
1160 struct inet_connection_sock *icsk = inet_csk(sk);
1161 struct inet_sock *inet = inet_sk(sk);
1162 int err = -EADDRINUSE;
1164 reqsk_queue_alloc(&icsk->icsk_accept_queue);
1166 sk->sk_ack_backlog = 0;
1167 inet_csk_delack_init(sk);
1169 if (sk->sk_txrehash == SOCK_TXREHASH_DEFAULT)
1170 sk->sk_txrehash = READ_ONCE(sock_net(sk)->core.sysctl_txrehash);
1172 /* There is race window here: we announce ourselves listening,
1173 * but this transition is still not validated by get_port().
1174 * It is OK, because this socket enters to hash table only
1175 * after validation is complete.
1177 inet_sk_state_store(sk, TCP_LISTEN);
1178 if (!sk->sk_prot->get_port(sk, inet->inet_num)) {
1179 inet->inet_sport = htons(inet->inet_num);
1182 err = sk->sk_prot->hash(sk);
1188 inet_sk_set_state(sk, TCP_CLOSE);
1191 EXPORT_SYMBOL_GPL(inet_csk_listen_start);
1193 static void inet_child_forget(struct sock *sk, struct request_sock *req,
1196 sk->sk_prot->disconnect(child, O_NONBLOCK);
1200 this_cpu_inc(*sk->sk_prot->orphan_count);
1202 if (sk->sk_protocol == IPPROTO_TCP && tcp_rsk(req)->tfo_listener) {
1203 BUG_ON(rcu_access_pointer(tcp_sk(child)->fastopen_rsk) != req);
1204 BUG_ON(sk != req->rsk_listener);
1206 /* Paranoid, to prevent race condition if
1207 * an inbound pkt destined for child is
1208 * blocked by sock lock in tcp_v4_rcv().
1209 * Also to satisfy an assertion in
1210 * tcp_v4_destroy_sock().
1212 RCU_INIT_POINTER(tcp_sk(child)->fastopen_rsk, NULL);
1214 inet_csk_destroy_sock(child);
1217 struct sock *inet_csk_reqsk_queue_add(struct sock *sk,
1218 struct request_sock *req,
1221 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
1223 spin_lock(&queue->rskq_lock);
1224 if (unlikely(sk->sk_state != TCP_LISTEN)) {
1225 inet_child_forget(sk, req, child);
1229 req->dl_next = NULL;
1230 if (queue->rskq_accept_head == NULL)
1231 WRITE_ONCE(queue->rskq_accept_head, req);
1233 queue->rskq_accept_tail->dl_next = req;
1234 queue->rskq_accept_tail = req;
1235 sk_acceptq_added(sk);
1237 spin_unlock(&queue->rskq_lock);
1240 EXPORT_SYMBOL(inet_csk_reqsk_queue_add);
1242 struct sock *inet_csk_complete_hashdance(struct sock *sk, struct sock *child,
1243 struct request_sock *req, bool own_req)
1246 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
1247 reqsk_queue_removed(&inet_csk(req->rsk_listener)->icsk_accept_queue, req);
1249 if (sk != req->rsk_listener) {
1250 /* another listening sk has been selected,
1251 * migrate the req to it.
1253 struct request_sock *nreq;
1255 /* hold a refcnt for the nreq->rsk_listener
1256 * which is assigned in inet_reqsk_clone()
1259 nreq = inet_reqsk_clone(req, sk);
1261 inet_child_forget(sk, req, child);
1265 refcount_set(&nreq->rsk_refcnt, 1);
1266 if (inet_csk_reqsk_queue_add(sk, nreq, child)) {
1267 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMIGRATEREQSUCCESS);
1268 reqsk_migrate_reset(req);
1273 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMIGRATEREQFAILURE);
1274 reqsk_migrate_reset(nreq);
1276 } else if (inet_csk_reqsk_queue_add(sk, req, child)) {
1280 /* Too bad, another child took ownership of the request, undo. */
1282 bh_unlock_sock(child);
1286 EXPORT_SYMBOL(inet_csk_complete_hashdance);
1289 * This routine closes sockets which have been at least partially
1290 * opened, but not yet accepted.
1292 void inet_csk_listen_stop(struct sock *sk)
1294 struct inet_connection_sock *icsk = inet_csk(sk);
1295 struct request_sock_queue *queue = &icsk->icsk_accept_queue;
1296 struct request_sock *next, *req;
1298 /* Following specs, it would be better either to send FIN
1299 * (and enter FIN-WAIT-1, it is normal close)
1300 * or to send active reset (abort).
1301 * Certainly, it is pretty dangerous while synflood, but it is
1302 * bad justification for our negligence 8)
1303 * To be honest, we are not able to make either
1304 * of the variants now. --ANK
1306 while ((req = reqsk_queue_remove(queue, sk)) != NULL) {
1307 struct sock *child = req->sk, *nsk;
1308 struct request_sock *nreq;
1311 bh_lock_sock(child);
1312 WARN_ON(sock_owned_by_user(child));
1315 nsk = reuseport_migrate_sock(sk, child, NULL);
1317 nreq = inet_reqsk_clone(req, nsk);
1319 refcount_set(&nreq->rsk_refcnt, 1);
1321 if (inet_csk_reqsk_queue_add(nsk, nreq, child)) {
1322 __NET_INC_STATS(sock_net(nsk),
1323 LINUX_MIB_TCPMIGRATEREQSUCCESS);
1324 reqsk_migrate_reset(req);
1326 __NET_INC_STATS(sock_net(nsk),
1327 LINUX_MIB_TCPMIGRATEREQFAILURE);
1328 reqsk_migrate_reset(nreq);
1332 /* inet_csk_reqsk_queue_add() has already
1333 * called inet_child_forget() on failure case.
1335 goto skip_child_forget;
1339 inet_child_forget(sk, req, child);
1342 bh_unlock_sock(child);
1348 if (queue->fastopenq.rskq_rst_head) {
1349 /* Free all the reqs queued in rskq_rst_head. */
1350 spin_lock_bh(&queue->fastopenq.lock);
1351 req = queue->fastopenq.rskq_rst_head;
1352 queue->fastopenq.rskq_rst_head = NULL;
1353 spin_unlock_bh(&queue->fastopenq.lock);
1354 while (req != NULL) {
1355 next = req->dl_next;
1360 WARN_ON_ONCE(sk->sk_ack_backlog);
1362 EXPORT_SYMBOL_GPL(inet_csk_listen_stop);
1364 void inet_csk_addr2sockaddr(struct sock *sk, struct sockaddr *uaddr)
1366 struct sockaddr_in *sin = (struct sockaddr_in *)uaddr;
1367 const struct inet_sock *inet = inet_sk(sk);
1369 sin->sin_family = AF_INET;
1370 sin->sin_addr.s_addr = inet->inet_daddr;
1371 sin->sin_port = inet->inet_dport;
1373 EXPORT_SYMBOL_GPL(inet_csk_addr2sockaddr);
1375 static struct dst_entry *inet_csk_rebuild_route(struct sock *sk, struct flowi *fl)
1377 const struct inet_sock *inet = inet_sk(sk);
1378 const struct ip_options_rcu *inet_opt;
1379 __be32 daddr = inet->inet_daddr;
1384 inet_opt = rcu_dereference(inet->inet_opt);
1385 if (inet_opt && inet_opt->opt.srr)
1386 daddr = inet_opt->opt.faddr;
1388 rt = ip_route_output_ports(sock_net(sk), fl4, sk, daddr,
1389 inet->inet_saddr, inet->inet_dport,
1390 inet->inet_sport, sk->sk_protocol,
1391 RT_CONN_FLAGS(sk), sk->sk_bound_dev_if);
1395 sk_setup_caps(sk, &rt->dst);
1401 struct dst_entry *inet_csk_update_pmtu(struct sock *sk, u32 mtu)
1403 struct dst_entry *dst = __sk_dst_check(sk, 0);
1404 struct inet_sock *inet = inet_sk(sk);
1407 dst = inet_csk_rebuild_route(sk, &inet->cork.fl);
1411 dst->ops->update_pmtu(dst, sk, NULL, mtu, true);
1413 dst = __sk_dst_check(sk, 0);
1415 dst = inet_csk_rebuild_route(sk, &inet->cork.fl);
1419 EXPORT_SYMBOL_GPL(inet_csk_update_pmtu);