1 // SPDX-License-Identifier: GPL-2.0-only
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 * Implementation of the Transmission Control Protocol(TCP).
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
24 #include <net/busy_poll.h>
26 static bool tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win)
30 if (after(end_seq, s_win) && before(seq, e_win))
32 return seq == e_win && seq == end_seq;
35 static enum tcp_tw_status
36 tcp_timewait_check_oow_rate_limit(struct inet_timewait_sock *tw,
37 const struct sk_buff *skb, int mib_idx)
39 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
41 if (!tcp_oow_rate_limited(twsk_net(tw), skb, mib_idx,
42 &tcptw->tw_last_oow_ack_time)) {
43 /* Send ACK. Note, we do not put the bucket,
44 * it will be released by caller.
49 /* We are rate-limiting, so just release the tw sock and drop skb. */
51 return TCP_TW_SUCCESS;
54 static void twsk_rcv_nxt_update(struct tcp_timewait_sock *tcptw, u32 seq)
57 struct tcp_ao_info *ao;
59 ao = rcu_dereference(tcptw->ao_info);
60 if (unlikely(ao && seq < tcptw->tw_rcv_nxt))
61 WRITE_ONCE(ao->rcv_sne, ao->rcv_sne + 1);
63 tcptw->tw_rcv_nxt = seq;
67 * * Main purpose of TIME-WAIT state is to close connection gracefully,
68 * when one of ends sits in LAST-ACK or CLOSING retransmitting FIN
69 * (and, probably, tail of data) and one or more our ACKs are lost.
70 * * What is TIME-WAIT timeout? It is associated with maximal packet
71 * lifetime in the internet, which results in wrong conclusion, that
72 * it is set to catch "old duplicate segments" wandering out of their path.
73 * It is not quite correct. This timeout is calculated so that it exceeds
74 * maximal retransmission timeout enough to allow to lose one (or more)
75 * segments sent by peer and our ACKs. This time may be calculated from RTO.
76 * * When TIME-WAIT socket receives RST, it means that another end
77 * finally closed and we are allowed to kill TIME-WAIT too.
78 * * Second purpose of TIME-WAIT is catching old duplicate segments.
79 * Well, certainly it is pure paranoia, but if we load TIME-WAIT
80 * with this semantics, we MUST NOT kill TIME-WAIT state with RSTs.
81 * * If we invented some more clever way to catch duplicates
82 * (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs.
84 * The algorithm below is based on FORMAL INTERPRETATION of RFCs.
85 * When you compare it to RFCs, please, read section SEGMENT ARRIVES
86 * from the very beginning.
88 * NOTE. With recycling (and later with fin-wait-2) TW bucket
89 * is _not_ stateless. It means, that strictly speaking we must
90 * spinlock it. I do not want! Well, probability of misbehaviour
91 * is ridiculously low and, seems, we could use some mb() tricks
92 * to avoid misread sequence numbers, states etc. --ANK
94 * We don't need to initialize tmp_out.sack_ok as we don't use the results
97 tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb,
98 const struct tcphdr *th)
100 struct tcp_options_received tmp_opt;
101 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
102 bool paws_reject = false;
104 tmp_opt.saw_tstamp = 0;
105 if (th->doff > (sizeof(*th) >> 2) && tcptw->tw_ts_recent_stamp) {
106 tcp_parse_options(twsk_net(tw), skb, &tmp_opt, 0, NULL);
108 if (tmp_opt.saw_tstamp) {
109 if (tmp_opt.rcv_tsecr)
110 tmp_opt.rcv_tsecr -= tcptw->tw_ts_offset;
111 tmp_opt.ts_recent = tcptw->tw_ts_recent;
112 tmp_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp;
113 paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
117 if (tw->tw_substate == TCP_FIN_WAIT2) {
118 /* Just repeat all the checks of tcp_rcv_state_process() */
120 /* Out of window, send ACK */
122 !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
124 tcptw->tw_rcv_nxt + tcptw->tw_rcv_wnd))
125 return tcp_timewait_check_oow_rate_limit(
126 tw, skb, LINUX_MIB_TCPACKSKIPPEDFINWAIT2);
131 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt))
136 !after(TCP_SKB_CB(skb)->end_seq, tcptw->tw_rcv_nxt) ||
137 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) {
139 return TCP_TW_SUCCESS;
142 /* New data or FIN. If new data arrive after half-duplex close,
146 TCP_SKB_CB(skb)->end_seq != tcptw->tw_rcv_nxt + 1)
149 /* FIN arrived, enter true time-wait state. */
150 tw->tw_substate = TCP_TIME_WAIT;
151 twsk_rcv_nxt_update(tcptw, TCP_SKB_CB(skb)->end_seq);
153 if (tmp_opt.saw_tstamp) {
154 tcptw->tw_ts_recent_stamp = ktime_get_seconds();
155 tcptw->tw_ts_recent = tmp_opt.rcv_tsval;
158 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
163 * Now real TIME-WAIT state.
166 * "When a connection is [...] on TIME-WAIT state [...]
167 * [a TCP] MAY accept a new SYN from the remote TCP to
168 * reopen the connection directly, if it:
170 * (1) assigns its initial sequence number for the new
171 * connection to be larger than the largest sequence
172 * number it used on the previous connection incarnation,
175 * (2) returns to TIME-WAIT state if the SYN turns out
176 * to be an old duplicate".
180 (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt &&
181 (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) {
182 /* In window segment, it may be only reset or bare ack. */
185 /* This is TIME_WAIT assassination, in two flavors.
186 * Oh well... nobody has a sufficient solution to this
189 if (!READ_ONCE(twsk_net(tw)->ipv4.sysctl_tcp_rfc1337)) {
191 inet_twsk_deschedule_put(tw);
192 return TCP_TW_SUCCESS;
195 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
198 if (tmp_opt.saw_tstamp) {
199 tcptw->tw_ts_recent = tmp_opt.rcv_tsval;
200 tcptw->tw_ts_recent_stamp = ktime_get_seconds();
204 return TCP_TW_SUCCESS;
207 /* Out of window segment.
209 All the segments are ACKed immediately.
211 The only exception is new SYN. We accept it, if it is
212 not old duplicate and we are not in danger to be killed
213 by delayed old duplicates. RFC check is that it has
214 newer sequence number works at rates <40Mbit/sec.
215 However, if paws works, it is reliable AND even more,
216 we even may relax silly seq space cutoff.
218 RED-PEN: we violate main RFC requirement, if this SYN will appear
219 old duplicate (i.e. we receive RST in reply to SYN-ACK),
220 we must return socket to time-wait state. It is not good,
224 if (th->syn && !th->rst && !th->ack && !paws_reject &&
225 (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) ||
226 (tmp_opt.saw_tstamp &&
227 (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) {
228 u32 isn = tcptw->tw_snd_nxt + 65535 + 2;
231 TCP_SKB_CB(skb)->tcp_tw_isn = isn;
236 __NET_INC_STATS(twsk_net(tw), LINUX_MIB_PAWSESTABREJECTED);
239 /* In this case we must reset the TIMEWAIT timer.
241 * If it is ACKless SYN it may be both old duplicate
242 * and new good SYN with random sequence number <rcv_nxt.
243 * Do not reschedule in the last case.
245 if (paws_reject || th->ack)
246 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
248 return tcp_timewait_check_oow_rate_limit(
249 tw, skb, LINUX_MIB_TCPACKSKIPPEDTIMEWAIT);
252 return TCP_TW_SUCCESS;
254 EXPORT_SYMBOL(tcp_timewait_state_process);
256 static void tcp_time_wait_init(struct sock *sk, struct tcp_timewait_sock *tcptw)
258 #ifdef CONFIG_TCP_MD5SIG
259 const struct tcp_sock *tp = tcp_sk(sk);
260 struct tcp_md5sig_key *key;
263 * The timewait bucket does not have the key DB from the
264 * sock structure. We just make a quick copy of the
265 * md5 key being used (if indeed we are using one)
266 * so the timewait ack generating code has the key.
268 tcptw->tw_md5_key = NULL;
269 if (!static_branch_unlikely(&tcp_md5_needed.key))
272 key = tp->af_specific->md5_lookup(sk, sk);
274 tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC);
275 if (!tcptw->tw_md5_key)
277 if (!static_key_fast_inc_not_disabled(&tcp_md5_needed.key.key))
279 tcp_md5_add_sigpool();
284 kfree(tcptw->tw_md5_key);
285 tcptw->tw_md5_key = NULL;
290 * Move a socket to time-wait or dead fin-wait-2 state.
292 void tcp_time_wait(struct sock *sk, int state, int timeo)
294 const struct inet_connection_sock *icsk = inet_csk(sk);
295 struct tcp_sock *tp = tcp_sk(sk);
296 struct net *net = sock_net(sk);
297 struct inet_timewait_sock *tw;
299 tw = inet_twsk_alloc(sk, &net->ipv4.tcp_death_row, state);
302 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
303 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1);
305 tw->tw_transparent = inet_test_bit(TRANSPARENT, sk);
306 tw->tw_mark = sk->sk_mark;
307 tw->tw_priority = READ_ONCE(sk->sk_priority);
308 tw->tw_rcv_wscale = tp->rx_opt.rcv_wscale;
309 tcptw->tw_rcv_nxt = tp->rcv_nxt;
310 tcptw->tw_snd_nxt = tp->snd_nxt;
311 tcptw->tw_rcv_wnd = tcp_receive_window(tp);
312 tcptw->tw_ts_recent = tp->rx_opt.ts_recent;
313 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp;
314 tcptw->tw_ts_offset = tp->tsoffset;
315 tw->tw_usec_ts = tp->tcp_usec_ts;
316 tcptw->tw_last_oow_ack_time = 0;
317 tcptw->tw_tx_delay = tp->tcp_tx_delay;
318 tw->tw_txhash = sk->sk_txhash;
319 #if IS_ENABLED(CONFIG_IPV6)
320 if (tw->tw_family == PF_INET6) {
321 struct ipv6_pinfo *np = inet6_sk(sk);
323 tw->tw_v6_daddr = sk->sk_v6_daddr;
324 tw->tw_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
325 tw->tw_tclass = np->tclass;
326 tw->tw_flowlabel = be32_to_cpu(np->flow_label & IPV6_FLOWLABEL_MASK);
327 tw->tw_ipv6only = sk->sk_ipv6only;
331 tcp_time_wait_init(sk, tcptw);
332 tcp_ao_time_wait(tcptw, tp);
334 /* Get the TIME_WAIT timeout firing. */
338 if (state == TCP_TIME_WAIT)
339 timeo = TCP_TIMEWAIT_LEN;
341 /* tw_timer is pinned, so we need to make sure BH are disabled
342 * in following section, otherwise timer handler could run before
343 * we complete the initialization.
346 inet_twsk_schedule(tw, timeo);
348 * Note that access to tw after this point is illegal.
350 inet_twsk_hashdance(tw, sk, net->ipv4.tcp_death_row.hashinfo);
353 /* Sorry, if we're out of memory, just CLOSE this
354 * socket up. We've got bigger problems than
355 * non-graceful socket closings.
357 NET_INC_STATS(net, LINUX_MIB_TCPTIMEWAITOVERFLOW);
360 tcp_update_metrics(sk);
363 EXPORT_SYMBOL(tcp_time_wait);
365 #ifdef CONFIG_TCP_MD5SIG
366 static void tcp_md5_twsk_free_rcu(struct rcu_head *head)
368 struct tcp_md5sig_key *key;
370 key = container_of(head, struct tcp_md5sig_key, rcu);
372 static_branch_slow_dec_deferred(&tcp_md5_needed);
373 tcp_md5_release_sigpool();
377 void tcp_twsk_destructor(struct sock *sk)
379 #ifdef CONFIG_TCP_MD5SIG
380 if (static_branch_unlikely(&tcp_md5_needed.key)) {
381 struct tcp_timewait_sock *twsk = tcp_twsk(sk);
383 if (twsk->tw_md5_key)
384 call_rcu(&twsk->tw_md5_key->rcu, tcp_md5_twsk_free_rcu);
387 tcp_ao_destroy_sock(sk, true);
389 EXPORT_SYMBOL_GPL(tcp_twsk_destructor);
391 void tcp_twsk_purge(struct list_head *net_exit_list, int family)
393 bool purged_once = false;
396 list_for_each_entry(net, net_exit_list, exit_list) {
397 if (net->ipv4.tcp_death_row.hashinfo->pernet) {
398 /* Even if tw_refcount == 1, we must clean up kernel reqsk */
399 inet_twsk_purge(net->ipv4.tcp_death_row.hashinfo, family);
400 } else if (!purged_once) {
401 inet_twsk_purge(&tcp_hashinfo, family);
406 EXPORT_SYMBOL_GPL(tcp_twsk_purge);
408 /* Warning : This function is called without sk_listener being locked.
409 * Be sure to read socket fields once, as their value could change under us.
411 void tcp_openreq_init_rwin(struct request_sock *req,
412 const struct sock *sk_listener,
413 const struct dst_entry *dst)
415 struct inet_request_sock *ireq = inet_rsk(req);
416 const struct tcp_sock *tp = tcp_sk(sk_listener);
417 int full_space = tcp_full_space(sk_listener);
423 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
424 window_clamp = READ_ONCE(tp->window_clamp);
425 /* Set this up on the first call only */
426 req->rsk_window_clamp = window_clamp ? : dst_metric(dst, RTAX_WINDOW);
428 /* limit the window selection if the user enforce a smaller rx buffer */
429 if (sk_listener->sk_userlocks & SOCK_RCVBUF_LOCK &&
430 (req->rsk_window_clamp > full_space || req->rsk_window_clamp == 0))
431 req->rsk_window_clamp = full_space;
433 rcv_wnd = tcp_rwnd_init_bpf((struct sock *)req);
435 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
436 else if (full_space < rcv_wnd * mss)
437 full_space = rcv_wnd * mss;
439 /* tcp_full_space because it is guaranteed to be the first packet */
440 tcp_select_initial_window(sk_listener, full_space,
441 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
443 &req->rsk_window_clamp,
447 ireq->rcv_wscale = rcv_wscale;
449 EXPORT_SYMBOL(tcp_openreq_init_rwin);
451 static void tcp_ecn_openreq_child(struct tcp_sock *tp,
452 const struct request_sock *req)
454 tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0;
457 void tcp_ca_openreq_child(struct sock *sk, const struct dst_entry *dst)
459 struct inet_connection_sock *icsk = inet_csk(sk);
460 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
461 bool ca_got_dst = false;
463 if (ca_key != TCP_CA_UNSPEC) {
464 const struct tcp_congestion_ops *ca;
467 ca = tcp_ca_find_key(ca_key);
468 if (likely(ca && bpf_try_module_get(ca, ca->owner))) {
469 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
470 icsk->icsk_ca_ops = ca;
476 /* If no valid choice made yet, assign current system default ca. */
478 (!icsk->icsk_ca_setsockopt ||
479 !bpf_try_module_get(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner)))
480 tcp_assign_congestion_control(sk);
482 tcp_set_ca_state(sk, TCP_CA_Open);
484 EXPORT_SYMBOL_GPL(tcp_ca_openreq_child);
486 static void smc_check_reset_syn_req(const struct tcp_sock *oldtp,
487 struct request_sock *req,
488 struct tcp_sock *newtp)
490 #if IS_ENABLED(CONFIG_SMC)
491 struct inet_request_sock *ireq;
493 if (static_branch_unlikely(&tcp_have_smc)) {
494 ireq = inet_rsk(req);
495 if (oldtp->syn_smc && !ireq->smc_ok)
501 /* This is not only more efficient than what we used to do, it eliminates
502 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM
504 * Actually, we could lots of memory writes here. tp of listening
505 * socket contains all necessary default parameters.
507 struct sock *tcp_create_openreq_child(const struct sock *sk,
508 struct request_sock *req,
511 struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC);
512 const struct inet_request_sock *ireq = inet_rsk(req);
513 struct tcp_request_sock *treq = tcp_rsk(req);
514 struct inet_connection_sock *newicsk;
515 const struct tcp_sock *oldtp;
516 struct tcp_sock *newtp;
519 struct tcp_ao_key *ao_key;
525 newicsk = inet_csk(newsk);
526 newtp = tcp_sk(newsk);
529 smc_check_reset_syn_req(oldtp, req, newtp);
531 /* Now setup tcp_sock */
532 newtp->pred_flags = 0;
534 seq = treq->rcv_isn + 1;
535 newtp->rcv_wup = seq;
536 WRITE_ONCE(newtp->copied_seq, seq);
537 WRITE_ONCE(newtp->rcv_nxt, seq);
540 seq = treq->snt_isn + 1;
541 newtp->snd_sml = newtp->snd_una = seq;
542 WRITE_ONCE(newtp->snd_nxt, seq);
545 INIT_LIST_HEAD(&newtp->tsq_node);
546 INIT_LIST_HEAD(&newtp->tsorted_sent_queue);
548 tcp_init_wl(newtp, treq->rcv_isn);
550 minmax_reset(&newtp->rtt_min, tcp_jiffies32, ~0U);
551 newicsk->icsk_ack.lrcvtime = tcp_jiffies32;
553 newtp->lsndtime = tcp_jiffies32;
554 newsk->sk_txhash = READ_ONCE(treq->txhash);
555 newtp->total_retrans = req->num_retrans;
557 tcp_init_xmit_timers(newsk);
558 WRITE_ONCE(newtp->write_seq, newtp->pushed_seq = treq->snt_isn + 1);
560 if (sock_flag(newsk, SOCK_KEEPOPEN))
561 inet_csk_reset_keepalive_timer(newsk,
562 keepalive_time_when(newtp));
564 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok;
565 newtp->rx_opt.sack_ok = ireq->sack_ok;
566 newtp->window_clamp = req->rsk_window_clamp;
567 newtp->rcv_ssthresh = req->rsk_rcv_wnd;
568 newtp->rcv_wnd = req->rsk_rcv_wnd;
569 newtp->rx_opt.wscale_ok = ireq->wscale_ok;
570 if (newtp->rx_opt.wscale_ok) {
571 newtp->rx_opt.snd_wscale = ireq->snd_wscale;
572 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale;
574 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0;
575 newtp->window_clamp = min(newtp->window_clamp, 65535U);
577 newtp->snd_wnd = ntohs(tcp_hdr(skb)->window) << newtp->rx_opt.snd_wscale;
578 newtp->max_window = newtp->snd_wnd;
580 if (newtp->rx_opt.tstamp_ok) {
581 newtp->tcp_usec_ts = treq->req_usec_ts;
582 newtp->rx_opt.ts_recent = READ_ONCE(req->ts_recent);
583 newtp->rx_opt.ts_recent_stamp = ktime_get_seconds();
584 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
586 newtp->tcp_usec_ts = 0;
587 newtp->rx_opt.ts_recent_stamp = 0;
588 newtp->tcp_header_len = sizeof(struct tcphdr);
590 if (req->num_timeout) {
591 newtp->total_rto = req->num_timeout;
592 newtp->undo_marker = treq->snt_isn;
593 if (newtp->tcp_usec_ts) {
594 newtp->retrans_stamp = treq->snt_synack;
595 newtp->total_rto_time = (u32)(tcp_clock_us() -
596 newtp->retrans_stamp) / USEC_PER_MSEC;
598 newtp->retrans_stamp = div_u64(treq->snt_synack,
599 USEC_PER_SEC / TCP_TS_HZ);
600 newtp->total_rto_time = tcp_clock_ms() -
601 newtp->retrans_stamp;
603 newtp->total_rto_recoveries = 1;
605 newtp->tsoffset = treq->ts_off;
606 #ifdef CONFIG_TCP_MD5SIG
607 newtp->md5sig_info = NULL; /*XXX*/
610 newtp->ao_info = NULL;
611 ao_key = treq->af_specific->ao_lookup(sk, req,
612 tcp_rsk(req)->ao_keyid, -1);
614 newtp->tcp_header_len += tcp_ao_len_aligned(ao_key);
616 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len)
617 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len;
618 newtp->rx_opt.mss_clamp = req->mss;
619 tcp_ecn_openreq_child(newtp, req);
620 newtp->fastopen_req = NULL;
621 RCU_INIT_POINTER(newtp->fastopen_rsk, NULL);
623 newtp->bpf_chg_cc_inprogress = 0;
624 tcp_bpf_clone(sk, newsk);
626 __TCP_INC_STATS(sock_net(sk), TCP_MIB_PASSIVEOPENS);
630 EXPORT_SYMBOL(tcp_create_openreq_child);
633 * Process an incoming packet for SYN_RECV sockets represented as a
634 * request_sock. Normally sk is the listener socket but for TFO it
635 * points to the child socket.
637 * XXX (TFO) - The current impl contains a special check for ack
638 * validation and inside tcp_v4_reqsk_send_ack(). Can we do better?
640 * We don't need to initialize tmp_opt.sack_ok as we don't use the results
642 * Note: If @fastopen is true, this can be called from process context.
643 * Otherwise, this is from BH context.
646 struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb,
647 struct request_sock *req,
648 bool fastopen, bool *req_stolen)
650 struct tcp_options_received tmp_opt;
652 const struct tcphdr *th = tcp_hdr(skb);
653 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK);
654 bool paws_reject = false;
657 tmp_opt.saw_tstamp = 0;
658 if (th->doff > (sizeof(struct tcphdr)>>2)) {
659 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0, NULL);
661 if (tmp_opt.saw_tstamp) {
662 tmp_opt.ts_recent = READ_ONCE(req->ts_recent);
663 if (tmp_opt.rcv_tsecr)
664 tmp_opt.rcv_tsecr -= tcp_rsk(req)->ts_off;
665 /* We do not store true stamp, but it is not required,
666 * it can be estimated (approximately)
669 tmp_opt.ts_recent_stamp = ktime_get_seconds() - reqsk_timeout(req, TCP_RTO_MAX) / HZ;
670 paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
674 /* Check for pure retransmitted SYN. */
675 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn &&
676 flg == TCP_FLAG_SYN &&
679 * RFC793 draws (Incorrectly! It was fixed in RFC1122)
680 * this case on figure 6 and figure 8, but formal
681 * protocol description says NOTHING.
682 * To be more exact, it says that we should send ACK,
683 * because this segment (at least, if it has no data)
686 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT
687 * describe SYN-RECV state. All the description
688 * is wrong, we cannot believe to it and should
689 * rely only on common sense and implementation
692 * Enforce "SYN-ACK" according to figure 8, figure 6
693 * of RFC793, fixed by RFC1122.
695 * Note that even if there is new data in the SYN packet
696 * they will be thrown away too.
698 * Reset timer after retransmitting SYNACK, similar to
699 * the idea of fast retransmit in recovery.
701 if (!tcp_oow_rate_limited(sock_net(sk), skb,
702 LINUX_MIB_TCPACKSKIPPEDSYNRECV,
703 &tcp_rsk(req)->last_oow_ack_time) &&
705 !inet_rtx_syn_ack(sk, req)) {
706 unsigned long expires = jiffies;
708 expires += reqsk_timeout(req, TCP_RTO_MAX);
710 mod_timer_pending(&req->rsk_timer, expires);
712 req->rsk_timer.expires = expires;
717 /* Further reproduces section "SEGMENT ARRIVES"
718 for state SYN-RECEIVED of RFC793.
719 It is broken, however, it does not work only
720 when SYNs are crossed.
722 You would think that SYN crossing is impossible here, since
723 we should have a SYN_SENT socket (from connect()) on our end,
724 but this is not true if the crossed SYNs were sent to both
725 ends by a malicious third party. We must defend against this,
726 and to do that we first verify the ACK (as per RFC793, page
727 36) and reset if it is invalid. Is this a true full defense?
728 To convince ourselves, let us consider a way in which the ACK
729 test can still pass in this 'malicious crossed SYNs' case.
730 Malicious sender sends identical SYNs (and thus identical sequence
731 numbers) to both A and B:
736 By our good fortune, both A and B select the same initial
737 send sequence number of seven :-)
739 A: sends SYN|ACK, seq=7, ack_seq=8
740 B: sends SYN|ACK, seq=7, ack_seq=8
742 So we are now A eating this SYN|ACK, ACK test passes. So
743 does sequence test, SYN is truncated, and thus we consider
746 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this
747 bare ACK. Otherwise, we create an established connection. Both
748 ends (listening sockets) accept the new incoming connection and try
749 to talk to each other. 8-)
751 Note: This case is both harmless, and rare. Possibility is about the
752 same as us discovering intelligent life on another plant tomorrow.
754 But generally, we should (RFC lies!) to accept ACK
755 from SYNACK both here and in tcp_rcv_state_process().
756 tcp_rcv_state_process() does not, hence, we do not too.
758 Note that the case is absolutely generic:
759 we cannot optimize anything here without
760 violating protocol. All the checks must be made
761 before attempt to create socket.
764 /* RFC793 page 36: "If the connection is in any non-synchronized state ...
765 * and the incoming segment acknowledges something not yet
766 * sent (the segment carries an unacceptable ACK) ...
769 * Invalid ACK: reset will be sent by listening socket.
770 * Note that the ACK validity check for a Fast Open socket is done
771 * elsewhere and is checked directly against the child socket rather
772 * than req because user data may have been sent out.
774 if ((flg & TCP_FLAG_ACK) && !fastopen &&
775 (TCP_SKB_CB(skb)->ack_seq !=
776 tcp_rsk(req)->snt_isn + 1))
779 /* Also, it would be not so bad idea to check rcv_tsecr, which
780 * is essentially ACK extension and too early or too late values
781 * should cause reset in unsynchronized states.
784 /* RFC793: "first check sequence number". */
786 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
787 tcp_rsk(req)->rcv_nxt, tcp_rsk(req)->rcv_nxt + req->rsk_rcv_wnd)) {
788 /* Out of window: send ACK and drop. */
789 if (!(flg & TCP_FLAG_RST) &&
790 !tcp_oow_rate_limited(sock_net(sk), skb,
791 LINUX_MIB_TCPACKSKIPPEDSYNRECV,
792 &tcp_rsk(req)->last_oow_ack_time))
793 req->rsk_ops->send_ack(sk, skb, req);
795 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
799 /* In sequence, PAWS is OK. */
801 /* TODO: We probably should defer ts_recent change once
802 * we take ownership of @req.
804 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_nxt))
805 WRITE_ONCE(req->ts_recent, tmp_opt.rcv_tsval);
807 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) {
808 /* Truncate SYN, it is out of window starting
809 at tcp_rsk(req)->rcv_isn + 1. */
810 flg &= ~TCP_FLAG_SYN;
813 /* RFC793: "second check the RST bit" and
814 * "fourth, check the SYN bit"
816 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) {
817 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
818 goto embryonic_reset;
821 /* ACK sequence verified above, just make sure ACK is
822 * set. If ACK not set, just silently drop the packet.
824 * XXX (TFO) - if we ever allow "data after SYN", the
825 * following check needs to be removed.
827 if (!(flg & TCP_FLAG_ACK))
830 /* For Fast Open no more processing is needed (sk is the
836 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */
837 if (req->num_timeout < READ_ONCE(inet_csk(sk)->icsk_accept_queue.rskq_defer_accept) &&
838 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) {
839 inet_rsk(req)->acked = 1;
840 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP);
844 /* OK, ACK is valid, create big socket and
845 * feed this segment to it. It will repeat all
846 * the tests. THIS SEGMENT MUST MOVE SOCKET TO
847 * ESTABLISHED STATE. If it will be dropped after
848 * socket is created, wait for troubles.
850 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL,
853 goto listen_overflow;
855 if (own_req && rsk_drop_req(req)) {
856 reqsk_queue_removed(&inet_csk(req->rsk_listener)->icsk_accept_queue, req);
857 inet_csk_reqsk_queue_drop_and_put(req->rsk_listener, req);
861 sock_rps_save_rxhash(child, skb);
862 tcp_synack_rtt_meas(child, req);
863 *req_stolen = !own_req;
864 return inet_csk_complete_hashdance(sk, child, req, own_req);
867 if (sk != req->rsk_listener)
868 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMIGRATEREQFAILURE);
870 if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_abort_on_overflow)) {
871 inet_rsk(req)->acked = 1;
876 if (!(flg & TCP_FLAG_RST)) {
877 /* Received a bad SYN pkt - for TFO We try not to reset
878 * the local connection unless it's really necessary to
879 * avoid becoming vulnerable to outside attack aiming at
880 * resetting legit local connections.
882 req->rsk_ops->send_reset(sk, skb);
883 } else if (fastopen) { /* received a valid RST pkt */
884 reqsk_fastopen_remove(sk, req, true);
888 bool unlinked = inet_csk_reqsk_queue_drop(sk, req);
891 __NET_INC_STATS(sock_net(sk), LINUX_MIB_EMBRYONICRSTS);
892 *req_stolen = !unlinked;
896 EXPORT_SYMBOL(tcp_check_req);
899 * Queue segment on the new socket if the new socket is active,
900 * otherwise we just shortcircuit this and continue with
903 * For the vast majority of cases child->sk_state will be TCP_SYN_RECV
904 * when entering. But other states are possible due to a race condition
905 * where after __inet_lookup_established() fails but before the listener
906 * locked is obtained, other packets cause the same connection to
910 int tcp_child_process(struct sock *parent, struct sock *child,
912 __releases(&((child)->sk_lock.slock))
915 int state = child->sk_state;
917 /* record sk_napi_id and sk_rx_queue_mapping of child. */
918 sk_mark_napi_id_set(child, skb);
920 tcp_segs_in(tcp_sk(child), skb);
921 if (!sock_owned_by_user(child)) {
922 ret = tcp_rcv_state_process(child, skb);
923 /* Wakeup parent, send SIGIO */
924 if (state == TCP_SYN_RECV && child->sk_state != state)
925 parent->sk_data_ready(parent);
927 /* Alas, it is possible again, because we do lookup
928 * in main socket hash table and lock on listening
929 * socket does not protect us more.
931 __sk_add_backlog(child, skb);
934 bh_unlock_sock(child);
938 EXPORT_SYMBOL(tcp_child_process);