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;
55 * * Main purpose of TIME-WAIT state is to close connection gracefully,
56 * when one of ends sits in LAST-ACK or CLOSING retransmitting FIN
57 * (and, probably, tail of data) and one or more our ACKs are lost.
58 * * What is TIME-WAIT timeout? It is associated with maximal packet
59 * lifetime in the internet, which results in wrong conclusion, that
60 * it is set to catch "old duplicate segments" wandering out of their path.
61 * It is not quite correct. This timeout is calculated so that it exceeds
62 * maximal retransmission timeout enough to allow to lose one (or more)
63 * segments sent by peer and our ACKs. This time may be calculated from RTO.
64 * * When TIME-WAIT socket receives RST, it means that another end
65 * finally closed and we are allowed to kill TIME-WAIT too.
66 * * Second purpose of TIME-WAIT is catching old duplicate segments.
67 * Well, certainly it is pure paranoia, but if we load TIME-WAIT
68 * with this semantics, we MUST NOT kill TIME-WAIT state with RSTs.
69 * * If we invented some more clever way to catch duplicates
70 * (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs.
72 * The algorithm below is based on FORMAL INTERPRETATION of RFCs.
73 * When you compare it to RFCs, please, read section SEGMENT ARRIVES
74 * from the very beginning.
76 * NOTE. With recycling (and later with fin-wait-2) TW bucket
77 * is _not_ stateless. It means, that strictly speaking we must
78 * spinlock it. I do not want! Well, probability of misbehaviour
79 * is ridiculously low and, seems, we could use some mb() tricks
80 * to avoid misread sequence numbers, states etc. --ANK
82 * We don't need to initialize tmp_out.sack_ok as we don't use the results
85 tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb,
86 const struct tcphdr *th)
88 struct tcp_options_received tmp_opt;
89 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
90 bool paws_reject = false;
92 tmp_opt.saw_tstamp = 0;
93 if (th->doff > (sizeof(*th) >> 2) && tcptw->tw_ts_recent_stamp) {
94 tcp_parse_options(twsk_net(tw), skb, &tmp_opt, 0, NULL);
96 if (tmp_opt.saw_tstamp) {
97 if (tmp_opt.rcv_tsecr)
98 tmp_opt.rcv_tsecr -= tcptw->tw_ts_offset;
99 tmp_opt.ts_recent = tcptw->tw_ts_recent;
100 tmp_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp;
101 paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
105 if (tw->tw_substate == TCP_FIN_WAIT2) {
106 /* Just repeat all the checks of tcp_rcv_state_process() */
108 /* Out of window, send ACK */
110 !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
112 tcptw->tw_rcv_nxt + tcptw->tw_rcv_wnd))
113 return tcp_timewait_check_oow_rate_limit(
114 tw, skb, LINUX_MIB_TCPACKSKIPPEDFINWAIT2);
119 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt))
124 !after(TCP_SKB_CB(skb)->end_seq, tcptw->tw_rcv_nxt) ||
125 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) {
127 return TCP_TW_SUCCESS;
130 /* New data or FIN. If new data arrive after half-duplex close,
134 TCP_SKB_CB(skb)->end_seq != tcptw->tw_rcv_nxt + 1)
137 /* FIN arrived, enter true time-wait state. */
138 tw->tw_substate = TCP_TIME_WAIT;
139 tcptw->tw_rcv_nxt = TCP_SKB_CB(skb)->end_seq;
140 if (tmp_opt.saw_tstamp) {
141 tcptw->tw_ts_recent_stamp = ktime_get_seconds();
142 tcptw->tw_ts_recent = tmp_opt.rcv_tsval;
145 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
150 * Now real TIME-WAIT state.
153 * "When a connection is [...] on TIME-WAIT state [...]
154 * [a TCP] MAY accept a new SYN from the remote TCP to
155 * reopen the connection directly, if it:
157 * (1) assigns its initial sequence number for the new
158 * connection to be larger than the largest sequence
159 * number it used on the previous connection incarnation,
162 * (2) returns to TIME-WAIT state if the SYN turns out
163 * to be an old duplicate".
167 (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt &&
168 (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) {
169 /* In window segment, it may be only reset or bare ack. */
172 /* This is TIME_WAIT assassination, in two flavors.
173 * Oh well... nobody has a sufficient solution to this
176 if (twsk_net(tw)->ipv4.sysctl_tcp_rfc1337 == 0) {
178 inet_twsk_deschedule_put(tw);
179 return TCP_TW_SUCCESS;
182 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
185 if (tmp_opt.saw_tstamp) {
186 tcptw->tw_ts_recent = tmp_opt.rcv_tsval;
187 tcptw->tw_ts_recent_stamp = ktime_get_seconds();
191 return TCP_TW_SUCCESS;
194 /* Out of window segment.
196 All the segments are ACKed immediately.
198 The only exception is new SYN. We accept it, if it is
199 not old duplicate and we are not in danger to be killed
200 by delayed old duplicates. RFC check is that it has
201 newer sequence number works at rates <40Mbit/sec.
202 However, if paws works, it is reliable AND even more,
203 we even may relax silly seq space cutoff.
205 RED-PEN: we violate main RFC requirement, if this SYN will appear
206 old duplicate (i.e. we receive RST in reply to SYN-ACK),
207 we must return socket to time-wait state. It is not good,
211 if (th->syn && !th->rst && !th->ack && !paws_reject &&
212 (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) ||
213 (tmp_opt.saw_tstamp &&
214 (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) {
215 u32 isn = tcptw->tw_snd_nxt + 65535 + 2;
218 TCP_SKB_CB(skb)->tcp_tw_isn = isn;
223 __NET_INC_STATS(twsk_net(tw), LINUX_MIB_PAWSESTABREJECTED);
226 /* In this case we must reset the TIMEWAIT timer.
228 * If it is ACKless SYN it may be both old duplicate
229 * and new good SYN with random sequence number <rcv_nxt.
230 * Do not reschedule in the last case.
232 if (paws_reject || th->ack)
233 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
235 return tcp_timewait_check_oow_rate_limit(
236 tw, skb, LINUX_MIB_TCPACKSKIPPEDTIMEWAIT);
239 return TCP_TW_SUCCESS;
241 EXPORT_SYMBOL(tcp_timewait_state_process);
244 * Move a socket to time-wait or dead fin-wait-2 state.
246 void tcp_time_wait(struct sock *sk, int state, int timeo)
248 const struct inet_connection_sock *icsk = inet_csk(sk);
249 const struct tcp_sock *tp = tcp_sk(sk);
250 struct inet_timewait_sock *tw;
251 struct inet_timewait_death_row *tcp_death_row = &sock_net(sk)->ipv4.tcp_death_row;
253 tw = inet_twsk_alloc(sk, tcp_death_row, state);
256 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
257 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1);
258 struct inet_sock *inet = inet_sk(sk);
260 tw->tw_transparent = inet->transparent;
261 tw->tw_mark = sk->sk_mark;
262 tw->tw_priority = sk->sk_priority;
263 tw->tw_rcv_wscale = tp->rx_opt.rcv_wscale;
264 tcptw->tw_rcv_nxt = tp->rcv_nxt;
265 tcptw->tw_snd_nxt = tp->snd_nxt;
266 tcptw->tw_rcv_wnd = tcp_receive_window(tp);
267 tcptw->tw_ts_recent = tp->rx_opt.ts_recent;
268 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp;
269 tcptw->tw_ts_offset = tp->tsoffset;
270 tcptw->tw_last_oow_ack_time = 0;
271 tcptw->tw_tx_delay = tp->tcp_tx_delay;
272 #if IS_ENABLED(CONFIG_IPV6)
273 if (tw->tw_family == PF_INET6) {
274 struct ipv6_pinfo *np = inet6_sk(sk);
276 tw->tw_v6_daddr = sk->sk_v6_daddr;
277 tw->tw_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
278 tw->tw_tclass = np->tclass;
279 tw->tw_flowlabel = be32_to_cpu(np->flow_label & IPV6_FLOWLABEL_MASK);
280 tw->tw_txhash = sk->sk_txhash;
281 tw->tw_ipv6only = sk->sk_ipv6only;
285 #ifdef CONFIG_TCP_MD5SIG
287 * The timewait bucket does not have the key DB from the
288 * sock structure. We just make a quick copy of the
289 * md5 key being used (if indeed we are using one)
290 * so the timewait ack generating code has the key.
293 tcptw->tw_md5_key = NULL;
294 if (static_branch_unlikely(&tcp_md5_needed)) {
295 struct tcp_md5sig_key *key;
297 key = tp->af_specific->md5_lookup(sk, sk);
299 tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC);
300 BUG_ON(tcptw->tw_md5_key && !tcp_alloc_md5sig_pool());
306 /* Get the TIME_WAIT timeout firing. */
310 if (state == TCP_TIME_WAIT)
311 timeo = TCP_TIMEWAIT_LEN;
313 /* tw_timer is pinned, so we need to make sure BH are disabled
314 * in following section, otherwise timer handler could run before
315 * we complete the initialization.
318 inet_twsk_schedule(tw, timeo);
320 * Note that access to tw after this point is illegal.
322 inet_twsk_hashdance(tw, sk, &tcp_hashinfo);
325 /* Sorry, if we're out of memory, just CLOSE this
326 * socket up. We've got bigger problems than
327 * non-graceful socket closings.
329 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEWAITOVERFLOW);
332 tcp_update_metrics(sk);
335 EXPORT_SYMBOL(tcp_time_wait);
337 void tcp_twsk_destructor(struct sock *sk)
339 #ifdef CONFIG_TCP_MD5SIG
340 if (static_branch_unlikely(&tcp_md5_needed)) {
341 struct tcp_timewait_sock *twsk = tcp_twsk(sk);
343 if (twsk->tw_md5_key)
344 kfree_rcu(twsk->tw_md5_key, rcu);
348 EXPORT_SYMBOL_GPL(tcp_twsk_destructor);
350 /* Warning : This function is called without sk_listener being locked.
351 * Be sure to read socket fields once, as their value could change under us.
353 void tcp_openreq_init_rwin(struct request_sock *req,
354 const struct sock *sk_listener,
355 const struct dst_entry *dst)
357 struct inet_request_sock *ireq = inet_rsk(req);
358 const struct tcp_sock *tp = tcp_sk(sk_listener);
359 int full_space = tcp_full_space(sk_listener);
365 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
366 window_clamp = READ_ONCE(tp->window_clamp);
367 /* Set this up on the first call only */
368 req->rsk_window_clamp = window_clamp ? : dst_metric(dst, RTAX_WINDOW);
370 /* limit the window selection if the user enforce a smaller rx buffer */
371 if (sk_listener->sk_userlocks & SOCK_RCVBUF_LOCK &&
372 (req->rsk_window_clamp > full_space || req->rsk_window_clamp == 0))
373 req->rsk_window_clamp = full_space;
375 rcv_wnd = tcp_rwnd_init_bpf((struct sock *)req);
377 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
378 else if (full_space < rcv_wnd * mss)
379 full_space = rcv_wnd * mss;
381 /* tcp_full_space because it is guaranteed to be the first packet */
382 tcp_select_initial_window(sk_listener, full_space,
383 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
385 &req->rsk_window_clamp,
389 ireq->rcv_wscale = rcv_wscale;
391 EXPORT_SYMBOL(tcp_openreq_init_rwin);
393 static void tcp_ecn_openreq_child(struct tcp_sock *tp,
394 const struct request_sock *req)
396 tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0;
399 void tcp_ca_openreq_child(struct sock *sk, const struct dst_entry *dst)
401 struct inet_connection_sock *icsk = inet_csk(sk);
402 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
403 bool ca_got_dst = false;
405 if (ca_key != TCP_CA_UNSPEC) {
406 const struct tcp_congestion_ops *ca;
409 ca = tcp_ca_find_key(ca_key);
410 if (likely(ca && bpf_try_module_get(ca, ca->owner))) {
411 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
412 icsk->icsk_ca_ops = ca;
418 /* If no valid choice made yet, assign current system default ca. */
420 (!icsk->icsk_ca_setsockopt ||
421 !bpf_try_module_get(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner)))
422 tcp_assign_congestion_control(sk);
424 tcp_set_ca_state(sk, TCP_CA_Open);
426 EXPORT_SYMBOL_GPL(tcp_ca_openreq_child);
428 static void smc_check_reset_syn_req(struct tcp_sock *oldtp,
429 struct request_sock *req,
430 struct tcp_sock *newtp)
432 #if IS_ENABLED(CONFIG_SMC)
433 struct inet_request_sock *ireq;
435 if (static_branch_unlikely(&tcp_have_smc)) {
436 ireq = inet_rsk(req);
437 if (oldtp->syn_smc && !ireq->smc_ok)
443 /* This is not only more efficient than what we used to do, it eliminates
444 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM
446 * Actually, we could lots of memory writes here. tp of listening
447 * socket contains all necessary default parameters.
449 struct sock *tcp_create_openreq_child(const struct sock *sk,
450 struct request_sock *req,
453 struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC);
454 const struct inet_request_sock *ireq = inet_rsk(req);
455 struct tcp_request_sock *treq = tcp_rsk(req);
456 struct inet_connection_sock *newicsk;
457 struct tcp_sock *oldtp, *newtp;
463 newicsk = inet_csk(newsk);
464 newtp = tcp_sk(newsk);
467 smc_check_reset_syn_req(oldtp, req, newtp);
469 /* Now setup tcp_sock */
470 newtp->pred_flags = 0;
472 seq = treq->rcv_isn + 1;
473 newtp->rcv_wup = seq;
474 WRITE_ONCE(newtp->copied_seq, seq);
475 WRITE_ONCE(newtp->rcv_nxt, seq);
478 seq = treq->snt_isn + 1;
479 newtp->snd_sml = newtp->snd_una = seq;
480 WRITE_ONCE(newtp->snd_nxt, seq);
483 INIT_LIST_HEAD(&newtp->tsq_node);
484 INIT_LIST_HEAD(&newtp->tsorted_sent_queue);
486 tcp_init_wl(newtp, treq->rcv_isn);
488 minmax_reset(&newtp->rtt_min, tcp_jiffies32, ~0U);
489 newicsk->icsk_ack.lrcvtime = tcp_jiffies32;
491 newtp->lsndtime = tcp_jiffies32;
492 newsk->sk_txhash = treq->txhash;
493 newtp->total_retrans = req->num_retrans;
495 tcp_init_xmit_timers(newsk);
496 WRITE_ONCE(newtp->write_seq, newtp->pushed_seq = treq->snt_isn + 1);
498 if (sock_flag(newsk, SOCK_KEEPOPEN))
499 inet_csk_reset_keepalive_timer(newsk,
500 keepalive_time_when(newtp));
502 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok;
503 newtp->rx_opt.sack_ok = ireq->sack_ok;
504 newtp->window_clamp = req->rsk_window_clamp;
505 newtp->rcv_ssthresh = req->rsk_rcv_wnd;
506 newtp->rcv_wnd = req->rsk_rcv_wnd;
507 newtp->rx_opt.wscale_ok = ireq->wscale_ok;
508 if (newtp->rx_opt.wscale_ok) {
509 newtp->rx_opt.snd_wscale = ireq->snd_wscale;
510 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale;
512 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0;
513 newtp->window_clamp = min(newtp->window_clamp, 65535U);
515 newtp->snd_wnd = ntohs(tcp_hdr(skb)->window) << newtp->rx_opt.snd_wscale;
516 newtp->max_window = newtp->snd_wnd;
518 if (newtp->rx_opt.tstamp_ok) {
519 newtp->rx_opt.ts_recent = req->ts_recent;
520 newtp->rx_opt.ts_recent_stamp = ktime_get_seconds();
521 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
523 newtp->rx_opt.ts_recent_stamp = 0;
524 newtp->tcp_header_len = sizeof(struct tcphdr);
526 if (req->num_timeout) {
527 newtp->undo_marker = treq->snt_isn;
528 newtp->retrans_stamp = div_u64(treq->snt_synack,
529 USEC_PER_SEC / TCP_TS_HZ);
531 newtp->tsoffset = treq->ts_off;
532 #ifdef CONFIG_TCP_MD5SIG
533 newtp->md5sig_info = NULL; /*XXX*/
534 if (newtp->af_specific->md5_lookup(sk, newsk))
535 newtp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
537 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len)
538 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len;
539 newtp->rx_opt.mss_clamp = req->mss;
540 tcp_ecn_openreq_child(newtp, req);
541 newtp->fastopen_req = NULL;
542 RCU_INIT_POINTER(newtp->fastopen_rsk, NULL);
544 tcp_bpf_clone(sk, newsk);
546 __TCP_INC_STATS(sock_net(sk), TCP_MIB_PASSIVEOPENS);
550 EXPORT_SYMBOL(tcp_create_openreq_child);
553 * Process an incoming packet for SYN_RECV sockets represented as a
554 * request_sock. Normally sk is the listener socket but for TFO it
555 * points to the child socket.
557 * XXX (TFO) - The current impl contains a special check for ack
558 * validation and inside tcp_v4_reqsk_send_ack(). Can we do better?
560 * We don't need to initialize tmp_opt.sack_ok as we don't use the results
563 struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb,
564 struct request_sock *req,
565 bool fastopen, bool *req_stolen)
567 struct tcp_options_received tmp_opt;
569 const struct tcphdr *th = tcp_hdr(skb);
570 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK);
571 bool paws_reject = false;
574 tmp_opt.saw_tstamp = 0;
575 if (th->doff > (sizeof(struct tcphdr)>>2)) {
576 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0, NULL);
578 if (tmp_opt.saw_tstamp) {
579 tmp_opt.ts_recent = req->ts_recent;
580 if (tmp_opt.rcv_tsecr)
581 tmp_opt.rcv_tsecr -= tcp_rsk(req)->ts_off;
582 /* We do not store true stamp, but it is not required,
583 * it can be estimated (approximately)
586 tmp_opt.ts_recent_stamp = ktime_get_seconds() - ((TCP_TIMEOUT_INIT/HZ)<<req->num_timeout);
587 paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
591 /* Check for pure retransmitted SYN. */
592 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn &&
593 flg == TCP_FLAG_SYN &&
596 * RFC793 draws (Incorrectly! It was fixed in RFC1122)
597 * this case on figure 6 and figure 8, but formal
598 * protocol description says NOTHING.
599 * To be more exact, it says that we should send ACK,
600 * because this segment (at least, if it has no data)
603 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT
604 * describe SYN-RECV state. All the description
605 * is wrong, we cannot believe to it and should
606 * rely only on common sense and implementation
609 * Enforce "SYN-ACK" according to figure 8, figure 6
610 * of RFC793, fixed by RFC1122.
612 * Note that even if there is new data in the SYN packet
613 * they will be thrown away too.
615 * Reset timer after retransmitting SYNACK, similar to
616 * the idea of fast retransmit in recovery.
618 if (!tcp_oow_rate_limited(sock_net(sk), skb,
619 LINUX_MIB_TCPACKSKIPPEDSYNRECV,
620 &tcp_rsk(req)->last_oow_ack_time) &&
622 !inet_rtx_syn_ack(sk, req)) {
623 unsigned long expires = jiffies;
625 expires += min(TCP_TIMEOUT_INIT << req->num_timeout,
628 mod_timer_pending(&req->rsk_timer, expires);
630 req->rsk_timer.expires = expires;
635 /* Further reproduces section "SEGMENT ARRIVES"
636 for state SYN-RECEIVED of RFC793.
637 It is broken, however, it does not work only
638 when SYNs are crossed.
640 You would think that SYN crossing is impossible here, since
641 we should have a SYN_SENT socket (from connect()) on our end,
642 but this is not true if the crossed SYNs were sent to both
643 ends by a malicious third party. We must defend against this,
644 and to do that we first verify the ACK (as per RFC793, page
645 36) and reset if it is invalid. Is this a true full defense?
646 To convince ourselves, let us consider a way in which the ACK
647 test can still pass in this 'malicious crossed SYNs' case.
648 Malicious sender sends identical SYNs (and thus identical sequence
649 numbers) to both A and B:
654 By our good fortune, both A and B select the same initial
655 send sequence number of seven :-)
657 A: sends SYN|ACK, seq=7, ack_seq=8
658 B: sends SYN|ACK, seq=7, ack_seq=8
660 So we are now A eating this SYN|ACK, ACK test passes. So
661 does sequence test, SYN is truncated, and thus we consider
664 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this
665 bare ACK. Otherwise, we create an established connection. Both
666 ends (listening sockets) accept the new incoming connection and try
667 to talk to each other. 8-)
669 Note: This case is both harmless, and rare. Possibility is about the
670 same as us discovering intelligent life on another plant tomorrow.
672 But generally, we should (RFC lies!) to accept ACK
673 from SYNACK both here and in tcp_rcv_state_process().
674 tcp_rcv_state_process() does not, hence, we do not too.
676 Note that the case is absolutely generic:
677 we cannot optimize anything here without
678 violating protocol. All the checks must be made
679 before attempt to create socket.
682 /* RFC793 page 36: "If the connection is in any non-synchronized state ...
683 * and the incoming segment acknowledges something not yet
684 * sent (the segment carries an unacceptable ACK) ...
687 * Invalid ACK: reset will be sent by listening socket.
688 * Note that the ACK validity check for a Fast Open socket is done
689 * elsewhere and is checked directly against the child socket rather
690 * than req because user data may have been sent out.
692 if ((flg & TCP_FLAG_ACK) && !fastopen &&
693 (TCP_SKB_CB(skb)->ack_seq !=
694 tcp_rsk(req)->snt_isn + 1))
697 /* Also, it would be not so bad idea to check rcv_tsecr, which
698 * is essentially ACK extension and too early or too late values
699 * should cause reset in unsynchronized states.
702 /* RFC793: "first check sequence number". */
704 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
705 tcp_rsk(req)->rcv_nxt, tcp_rsk(req)->rcv_nxt + req->rsk_rcv_wnd)) {
706 /* Out of window: send ACK and drop. */
707 if (!(flg & TCP_FLAG_RST) &&
708 !tcp_oow_rate_limited(sock_net(sk), skb,
709 LINUX_MIB_TCPACKSKIPPEDSYNRECV,
710 &tcp_rsk(req)->last_oow_ack_time))
711 req->rsk_ops->send_ack(sk, skb, req);
713 __NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
717 /* In sequence, PAWS is OK. */
719 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_nxt))
720 req->ts_recent = tmp_opt.rcv_tsval;
722 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) {
723 /* Truncate SYN, it is out of window starting
724 at tcp_rsk(req)->rcv_isn + 1. */
725 flg &= ~TCP_FLAG_SYN;
728 /* RFC793: "second check the RST bit" and
729 * "fourth, check the SYN bit"
731 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) {
732 __TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
733 goto embryonic_reset;
736 /* ACK sequence verified above, just make sure ACK is
737 * set. If ACK not set, just silently drop the packet.
739 * XXX (TFO) - if we ever allow "data after SYN", the
740 * following check needs to be removed.
742 if (!(flg & TCP_FLAG_ACK))
745 /* For Fast Open no more processing is needed (sk is the
751 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */
752 if (req->num_timeout < inet_csk(sk)->icsk_accept_queue.rskq_defer_accept &&
753 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) {
754 inet_rsk(req)->acked = 1;
755 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP);
759 /* OK, ACK is valid, create big socket and
760 * feed this segment to it. It will repeat all
761 * the tests. THIS SEGMENT MUST MOVE SOCKET TO
762 * ESTABLISHED STATE. If it will be dropped after
763 * socket is created, wait for troubles.
765 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL,
768 goto listen_overflow;
770 if (own_req && rsk_drop_req(req)) {
771 reqsk_queue_removed(&inet_csk(req->rsk_listener)->icsk_accept_queue, req);
772 inet_csk_reqsk_queue_drop_and_put(req->rsk_listener, req);
776 sock_rps_save_rxhash(child, skb);
777 tcp_synack_rtt_meas(child, req);
778 *req_stolen = !own_req;
779 return inet_csk_complete_hashdance(sk, child, req, own_req);
782 if (sk != req->rsk_listener)
783 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMIGRATEREQFAILURE);
785 if (!sock_net(sk)->ipv4.sysctl_tcp_abort_on_overflow) {
786 inet_rsk(req)->acked = 1;
791 if (!(flg & TCP_FLAG_RST)) {
792 /* Received a bad SYN pkt - for TFO We try not to reset
793 * the local connection unless it's really necessary to
794 * avoid becoming vulnerable to outside attack aiming at
795 * resetting legit local connections.
797 req->rsk_ops->send_reset(sk, skb);
798 } else if (fastopen) { /* received a valid RST pkt */
799 reqsk_fastopen_remove(sk, req, true);
803 bool unlinked = inet_csk_reqsk_queue_drop(sk, req);
806 __NET_INC_STATS(sock_net(sk), LINUX_MIB_EMBRYONICRSTS);
807 *req_stolen = !unlinked;
811 EXPORT_SYMBOL(tcp_check_req);
814 * Queue segment on the new socket if the new socket is active,
815 * otherwise we just shortcircuit this and continue with
818 * For the vast majority of cases child->sk_state will be TCP_SYN_RECV
819 * when entering. But other states are possible due to a race condition
820 * where after __inet_lookup_established() fails but before the listener
821 * locked is obtained, other packets cause the same connection to
825 int tcp_child_process(struct sock *parent, struct sock *child,
827 __releases(&((child)->sk_lock.slock))
830 int state = child->sk_state;
832 /* record sk_napi_id and sk_rx_queue_mapping of child. */
833 sk_mark_napi_id_set(child, skb);
835 tcp_segs_in(tcp_sk(child), skb);
836 if (!sock_owned_by_user(child)) {
837 ret = tcp_rcv_state_process(child, skb);
838 /* Wakeup parent, send SIGIO */
839 if (state == TCP_SYN_RECV && child->sk_state != state)
840 parent->sk_data_ready(parent);
842 /* Alas, it is possible again, because we do lookup
843 * in main socket hash table and lock on listening
844 * socket does not protect us more.
846 __sk_add_backlog(child, skb);
849 bh_unlock_sock(child);
853 EXPORT_SYMBOL(tcp_child_process);