2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
22 #include <linux/module.h>
23 #include <linux/slab.h>
24 #include <linux/sysctl.h>
25 #include <linux/workqueue.h>
27 #include <net/inet_common.h>
30 int sysctl_tcp_syncookies __read_mostly = 1;
31 EXPORT_SYMBOL(sysctl_tcp_syncookies);
33 int sysctl_tcp_abort_on_overflow __read_mostly;
35 struct inet_timewait_death_row tcp_death_row = {
36 .sysctl_max_tw_buckets = NR_FILE * 2,
37 .hashinfo = &tcp_hashinfo,
39 EXPORT_SYMBOL_GPL(tcp_death_row);
41 static bool tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win)
45 if (after(end_seq, s_win) && before(seq, e_win))
47 return seq == e_win && seq == end_seq;
50 static enum tcp_tw_status
51 tcp_timewait_check_oow_rate_limit(struct inet_timewait_sock *tw,
52 const struct sk_buff *skb, int mib_idx)
54 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
56 if (!tcp_oow_rate_limited(twsk_net(tw), skb, mib_idx,
57 &tcptw->tw_last_oow_ack_time)) {
58 /* Send ACK. Note, we do not put the bucket,
59 * it will be released by caller.
64 /* We are rate-limiting, so just release the tw sock and drop skb. */
66 return TCP_TW_SUCCESS;
70 * * Main purpose of TIME-WAIT state is to close connection gracefully,
71 * when one of ends sits in LAST-ACK or CLOSING retransmitting FIN
72 * (and, probably, tail of data) and one or more our ACKs are lost.
73 * * What is TIME-WAIT timeout? It is associated with maximal packet
74 * lifetime in the internet, which results in wrong conclusion, that
75 * it is set to catch "old duplicate segments" wandering out of their path.
76 * It is not quite correct. This timeout is calculated so that it exceeds
77 * maximal retransmission timeout enough to allow to lose one (or more)
78 * segments sent by peer and our ACKs. This time may be calculated from RTO.
79 * * When TIME-WAIT socket receives RST, it means that another end
80 * finally closed and we are allowed to kill TIME-WAIT too.
81 * * Second purpose of TIME-WAIT is catching old duplicate segments.
82 * Well, certainly it is pure paranoia, but if we load TIME-WAIT
83 * with this semantics, we MUST NOT kill TIME-WAIT state with RSTs.
84 * * If we invented some more clever way to catch duplicates
85 * (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs.
87 * The algorithm below is based on FORMAL INTERPRETATION of RFCs.
88 * When you compare it to RFCs, please, read section SEGMENT ARRIVES
89 * from the very beginning.
91 * NOTE. With recycling (and later with fin-wait-2) TW bucket
92 * is _not_ stateless. It means, that strictly speaking we must
93 * spinlock it. I do not want! Well, probability of misbehaviour
94 * is ridiculously low and, seems, we could use some mb() tricks
95 * to avoid misread sequence numbers, states etc. --ANK
97 * We don't need to initialize tmp_out.sack_ok as we don't use the results
100 tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb,
101 const struct tcphdr *th)
103 struct tcp_options_received tmp_opt;
104 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
105 bool paws_reject = false;
107 tmp_opt.saw_tstamp = 0;
108 if (th->doff > (sizeof(*th) >> 2) && tcptw->tw_ts_recent_stamp) {
109 tcp_parse_options(skb, &tmp_opt, 0, NULL);
111 if (tmp_opt.saw_tstamp) {
112 tmp_opt.rcv_tsecr -= tcptw->tw_ts_offset;
113 tmp_opt.ts_recent = tcptw->tw_ts_recent;
114 tmp_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp;
115 paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
119 if (tw->tw_substate == TCP_FIN_WAIT2) {
120 /* Just repeat all the checks of tcp_rcv_state_process() */
122 /* Out of window, send ACK */
124 !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
126 tcptw->tw_rcv_nxt + tcptw->tw_rcv_wnd))
127 return tcp_timewait_check_oow_rate_limit(
128 tw, skb, LINUX_MIB_TCPACKSKIPPEDFINWAIT2);
133 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt))
138 !after(TCP_SKB_CB(skb)->end_seq, tcptw->tw_rcv_nxt) ||
139 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) {
141 return TCP_TW_SUCCESS;
144 /* New data or FIN. If new data arrive after half-duplex close,
148 TCP_SKB_CB(skb)->end_seq != tcptw->tw_rcv_nxt + 1) {
150 inet_twsk_deschedule_put(tw);
154 /* FIN arrived, enter true time-wait state. */
155 tw->tw_substate = TCP_TIME_WAIT;
156 tcptw->tw_rcv_nxt = TCP_SKB_CB(skb)->end_seq;
157 if (tmp_opt.saw_tstamp) {
158 tcptw->tw_ts_recent_stamp = get_seconds();
159 tcptw->tw_ts_recent = tmp_opt.rcv_tsval;
162 if (tcp_death_row.sysctl_tw_recycle &&
163 tcptw->tw_ts_recent_stamp &&
164 tcp_tw_remember_stamp(tw))
165 inet_twsk_schedule(tw, tw->tw_timeout);
167 inet_twsk_schedule(tw, TCP_TIMEWAIT_LEN);
172 * Now real TIME-WAIT state.
175 * "When a connection is [...] on TIME-WAIT state [...]
176 * [a TCP] MAY accept a new SYN from the remote TCP to
177 * reopen the connection directly, if it:
179 * (1) assigns its initial sequence number for the new
180 * connection to be larger than the largest sequence
181 * number it used on the previous connection incarnation,
184 * (2) returns to TIME-WAIT state if the SYN turns out
185 * to be an old duplicate".
189 (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt &&
190 (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) {
191 /* In window segment, it may be only reset or bare ack. */
194 /* This is TIME_WAIT assassination, in two flavors.
195 * Oh well... nobody has a sufficient solution to this
198 if (sysctl_tcp_rfc1337 == 0) {
200 inet_twsk_deschedule_put(tw);
201 return TCP_TW_SUCCESS;
204 inet_twsk_schedule(tw, TCP_TIMEWAIT_LEN);
206 if (tmp_opt.saw_tstamp) {
207 tcptw->tw_ts_recent = tmp_opt.rcv_tsval;
208 tcptw->tw_ts_recent_stamp = get_seconds();
212 return TCP_TW_SUCCESS;
215 /* Out of window segment.
217 All the segments are ACKed immediately.
219 The only exception is new SYN. We accept it, if it is
220 not old duplicate and we are not in danger to be killed
221 by delayed old duplicates. RFC check is that it has
222 newer sequence number works at rates <40Mbit/sec.
223 However, if paws works, it is reliable AND even more,
224 we even may relax silly seq space cutoff.
226 RED-PEN: we violate main RFC requirement, if this SYN will appear
227 old duplicate (i.e. we receive RST in reply to SYN-ACK),
228 we must return socket to time-wait state. It is not good,
232 if (th->syn && !th->rst && !th->ack && !paws_reject &&
233 (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) ||
234 (tmp_opt.saw_tstamp &&
235 (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) {
236 u32 isn = tcptw->tw_snd_nxt + 65535 + 2;
239 TCP_SKB_CB(skb)->tcp_tw_isn = isn;
244 NET_INC_STATS_BH(twsk_net(tw), LINUX_MIB_PAWSESTABREJECTED);
247 /* In this case we must reset the TIMEWAIT timer.
249 * If it is ACKless SYN it may be both old duplicate
250 * and new good SYN with random sequence number <rcv_nxt.
251 * Do not reschedule in the last case.
253 if (paws_reject || th->ack)
254 inet_twsk_schedule(tw, TCP_TIMEWAIT_LEN);
256 return tcp_timewait_check_oow_rate_limit(
257 tw, skb, LINUX_MIB_TCPACKSKIPPEDTIMEWAIT);
260 return TCP_TW_SUCCESS;
262 EXPORT_SYMBOL(tcp_timewait_state_process);
265 * Move a socket to time-wait or dead fin-wait-2 state.
267 void tcp_time_wait(struct sock *sk, int state, int timeo)
269 const struct inet_connection_sock *icsk = inet_csk(sk);
270 const struct tcp_sock *tp = tcp_sk(sk);
271 struct inet_timewait_sock *tw;
272 bool recycle_ok = false;
274 if (tcp_death_row.sysctl_tw_recycle && tp->rx_opt.ts_recent_stamp)
275 recycle_ok = tcp_remember_stamp(sk);
277 tw = inet_twsk_alloc(sk, &tcp_death_row, state);
280 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
281 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1);
282 struct inet_sock *inet = inet_sk(sk);
284 tw->tw_transparent = inet->transparent;
285 tw->tw_rcv_wscale = tp->rx_opt.rcv_wscale;
286 tcptw->tw_rcv_nxt = tp->rcv_nxt;
287 tcptw->tw_snd_nxt = tp->snd_nxt;
288 tcptw->tw_rcv_wnd = tcp_receive_window(tp);
289 tcptw->tw_ts_recent = tp->rx_opt.ts_recent;
290 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp;
291 tcptw->tw_ts_offset = tp->tsoffset;
292 tcptw->tw_last_oow_ack_time = 0;
294 #if IS_ENABLED(CONFIG_IPV6)
295 if (tw->tw_family == PF_INET6) {
296 struct ipv6_pinfo *np = inet6_sk(sk);
298 tw->tw_v6_daddr = sk->sk_v6_daddr;
299 tw->tw_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
300 tw->tw_tclass = np->tclass;
301 tw->tw_flowlabel = be32_to_cpu(np->flow_label & IPV6_FLOWLABEL_MASK);
302 tw->tw_ipv6only = sk->sk_ipv6only;
306 #ifdef CONFIG_TCP_MD5SIG
308 * The timewait bucket does not have the key DB from the
309 * sock structure. We just make a quick copy of the
310 * md5 key being used (if indeed we are using one)
311 * so the timewait ack generating code has the key.
314 struct tcp_md5sig_key *key;
315 tcptw->tw_md5_key = NULL;
316 key = tp->af_specific->md5_lookup(sk, sk);
318 tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC);
319 if (tcptw->tw_md5_key && !tcp_alloc_md5sig_pool())
325 /* Linkage updates. */
326 __inet_twsk_hashdance(tw, sk, &tcp_hashinfo);
328 /* Get the TIME_WAIT timeout firing. */
333 tw->tw_timeout = rto;
335 tw->tw_timeout = TCP_TIMEWAIT_LEN;
336 if (state == TCP_TIME_WAIT)
337 timeo = TCP_TIMEWAIT_LEN;
340 inet_twsk_schedule(tw, timeo);
343 /* Sorry, if we're out of memory, just CLOSE this
344 * socket up. We've got bigger problems than
345 * non-graceful socket closings.
347 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPTIMEWAITOVERFLOW);
350 tcp_update_metrics(sk);
354 void tcp_twsk_destructor(struct sock *sk)
356 #ifdef CONFIG_TCP_MD5SIG
357 struct tcp_timewait_sock *twsk = tcp_twsk(sk);
359 if (twsk->tw_md5_key)
360 kfree_rcu(twsk->tw_md5_key, rcu);
363 EXPORT_SYMBOL_GPL(tcp_twsk_destructor);
365 void tcp_openreq_init_rwin(struct request_sock *req,
366 struct sock *sk, struct dst_entry *dst)
368 struct inet_request_sock *ireq = inet_rsk(req);
369 struct tcp_sock *tp = tcp_sk(sk);
371 int mss = dst_metric_advmss(dst);
373 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
374 mss = tp->rx_opt.user_mss;
376 /* Set this up on the first call only */
377 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
379 /* limit the window selection if the user enforce a smaller rx buffer */
380 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
381 (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
382 req->window_clamp = tcp_full_space(sk);
384 /* tcp_full_space because it is guaranteed to be the first packet */
385 tcp_select_initial_window(tcp_full_space(sk),
386 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
391 dst_metric(dst, RTAX_INITRWND));
392 ireq->rcv_wscale = rcv_wscale;
394 EXPORT_SYMBOL(tcp_openreq_init_rwin);
396 static void tcp_ecn_openreq_child(struct tcp_sock *tp,
397 const struct request_sock *req)
399 tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0;
402 void tcp_ca_openreq_child(struct sock *sk, const struct dst_entry *dst)
404 struct inet_connection_sock *icsk = inet_csk(sk);
405 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
406 bool ca_got_dst = false;
408 if (ca_key != TCP_CA_UNSPEC) {
409 const struct tcp_congestion_ops *ca;
412 ca = tcp_ca_find_key(ca_key);
413 if (likely(ca && try_module_get(ca->owner))) {
414 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
415 icsk->icsk_ca_ops = ca;
421 /* If no valid choice made yet, assign current system default ca. */
423 (!icsk->icsk_ca_setsockopt ||
424 !try_module_get(icsk->icsk_ca_ops->owner)))
425 tcp_assign_congestion_control(sk);
427 tcp_set_ca_state(sk, TCP_CA_Open);
429 EXPORT_SYMBOL_GPL(tcp_ca_openreq_child);
431 /* This is not only more efficient than what we used to do, it eliminates
432 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM
434 * Actually, we could lots of memory writes here. tp of listening
435 * socket contains all necessary default parameters.
437 struct sock *tcp_create_openreq_child(struct sock *sk, struct request_sock *req, struct sk_buff *skb)
439 struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC);
442 const struct inet_request_sock *ireq = inet_rsk(req);
443 struct tcp_request_sock *treq = tcp_rsk(req);
444 struct inet_connection_sock *newicsk = inet_csk(newsk);
445 struct tcp_sock *newtp = tcp_sk(newsk);
447 /* Now setup tcp_sock */
448 newtp->pred_flags = 0;
450 newtp->rcv_wup = newtp->copied_seq =
451 newtp->rcv_nxt = treq->rcv_isn + 1;
454 newtp->snd_sml = newtp->snd_una =
455 newtp->snd_nxt = newtp->snd_up = treq->snt_isn + 1;
457 tcp_prequeue_init(newtp);
458 INIT_LIST_HEAD(&newtp->tsq_node);
460 tcp_init_wl(newtp, treq->rcv_isn);
463 newtp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
464 newicsk->icsk_rto = TCP_TIMEOUT_INIT;
466 newtp->packets_out = 0;
467 newtp->retrans_out = 0;
468 newtp->sacked_out = 0;
469 newtp->fackets_out = 0;
470 newtp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
471 tcp_enable_early_retrans(newtp);
472 newtp->tlp_high_seq = 0;
473 newtp->lsndtime = treq->snt_synack;
474 newtp->last_oow_ack_time = 0;
475 newtp->total_retrans = req->num_retrans;
477 /* So many TCP implementations out there (incorrectly) count the
478 * initial SYN frame in their delayed-ACK and congestion control
479 * algorithms that we must have the following bandaid to talk
480 * efficiently to them. -DaveM
482 newtp->snd_cwnd = TCP_INIT_CWND;
483 newtp->snd_cwnd_cnt = 0;
485 tcp_init_xmit_timers(newsk);
486 __skb_queue_head_init(&newtp->out_of_order_queue);
487 newtp->write_seq = newtp->pushed_seq = treq->snt_isn + 1;
489 newtp->rx_opt.saw_tstamp = 0;
491 newtp->rx_opt.dsack = 0;
492 newtp->rx_opt.num_sacks = 0;
496 if (sock_flag(newsk, SOCK_KEEPOPEN))
497 inet_csk_reset_keepalive_timer(newsk,
498 keepalive_time_when(newtp));
500 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok;
501 if ((newtp->rx_opt.sack_ok = ireq->sack_ok) != 0) {
503 tcp_enable_fack(newtp);
505 newtp->window_clamp = req->window_clamp;
506 newtp->rcv_ssthresh = req->rcv_wnd;
507 newtp->rcv_wnd = req->rcv_wnd;
508 newtp->rx_opt.wscale_ok = ireq->wscale_ok;
509 if (newtp->rx_opt.wscale_ok) {
510 newtp->rx_opt.snd_wscale = ireq->snd_wscale;
511 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale;
513 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0;
514 newtp->window_clamp = min(newtp->window_clamp, 65535U);
516 newtp->snd_wnd = (ntohs(tcp_hdr(skb)->window) <<
517 newtp->rx_opt.snd_wscale);
518 newtp->max_window = newtp->snd_wnd;
520 if (newtp->rx_opt.tstamp_ok) {
521 newtp->rx_opt.ts_recent = req->ts_recent;
522 newtp->rx_opt.ts_recent_stamp = get_seconds();
523 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
525 newtp->rx_opt.ts_recent_stamp = 0;
526 newtp->tcp_header_len = sizeof(struct tcphdr);
529 #ifdef CONFIG_TCP_MD5SIG
530 newtp->md5sig_info = NULL; /*XXX*/
531 if (newtp->af_specific->md5_lookup(sk, newsk))
532 newtp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
534 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len)
535 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len;
536 newtp->rx_opt.mss_clamp = req->mss;
537 tcp_ecn_openreq_child(newtp, req);
538 newtp->fastopen_rsk = NULL;
539 newtp->syn_data_acked = 0;
541 newtp->saved_syn = req->saved_syn;
542 req->saved_syn = NULL;
544 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_PASSIVEOPENS);
548 EXPORT_SYMBOL(tcp_create_openreq_child);
551 * Process an incoming packet for SYN_RECV sockets represented as a
552 * request_sock. Normally sk is the listener socket but for TFO it
553 * points to the child socket.
555 * XXX (TFO) - The current impl contains a special check for ack
556 * validation and inside tcp_v4_reqsk_send_ack(). Can we do better?
558 * We don't need to initialize tmp_opt.sack_ok as we don't use the results
561 struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb,
562 struct request_sock *req,
565 struct tcp_options_received tmp_opt;
567 const struct tcphdr *th = tcp_hdr(skb);
568 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK);
569 bool paws_reject = false;
571 BUG_ON(fastopen == (sk->sk_state == TCP_LISTEN));
573 tmp_opt.saw_tstamp = 0;
574 if (th->doff > (sizeof(struct tcphdr)>>2)) {
575 tcp_parse_options(skb, &tmp_opt, 0, NULL);
577 if (tmp_opt.saw_tstamp) {
578 tmp_opt.ts_recent = req->ts_recent;
579 /* We do not store true stamp, but it is not required,
580 * it can be estimated (approximately)
583 tmp_opt.ts_recent_stamp = get_seconds() - ((TCP_TIMEOUT_INIT/HZ)<<req->num_timeout);
584 paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
588 /* Check for pure retransmitted SYN. */
589 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn &&
590 flg == TCP_FLAG_SYN &&
593 * RFC793 draws (Incorrectly! It was fixed in RFC1122)
594 * this case on figure 6 and figure 8, but formal
595 * protocol description says NOTHING.
596 * To be more exact, it says that we should send ACK,
597 * because this segment (at least, if it has no data)
600 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT
601 * describe SYN-RECV state. All the description
602 * is wrong, we cannot believe to it and should
603 * rely only on common sense and implementation
606 * Enforce "SYN-ACK" according to figure 8, figure 6
607 * of RFC793, fixed by RFC1122.
609 * Note that even if there is new data in the SYN packet
610 * they will be thrown away too.
612 * Reset timer after retransmitting SYNACK, similar to
613 * the idea of fast retransmit in recovery.
615 if (!tcp_oow_rate_limited(sock_net(sk), skb,
616 LINUX_MIB_TCPACKSKIPPEDSYNRECV,
617 &tcp_rsk(req)->last_oow_ack_time) &&
619 !inet_rtx_syn_ack(sk, req)) {
620 unsigned long expires = jiffies;
622 expires += min(TCP_TIMEOUT_INIT << req->num_timeout,
625 mod_timer_pending(&req->rsk_timer, expires);
627 req->rsk_timer.expires = expires;
632 /* Further reproduces section "SEGMENT ARRIVES"
633 for state SYN-RECEIVED of RFC793.
634 It is broken, however, it does not work only
635 when SYNs are crossed.
637 You would think that SYN crossing is impossible here, since
638 we should have a SYN_SENT socket (from connect()) on our end,
639 but this is not true if the crossed SYNs were sent to both
640 ends by a malicious third party. We must defend against this,
641 and to do that we first verify the ACK (as per RFC793, page
642 36) and reset if it is invalid. Is this a true full defense?
643 To convince ourselves, let us consider a way in which the ACK
644 test can still pass in this 'malicious crossed SYNs' case.
645 Malicious sender sends identical SYNs (and thus identical sequence
646 numbers) to both A and B:
651 By our good fortune, both A and B select the same initial
652 send sequence number of seven :-)
654 A: sends SYN|ACK, seq=7, ack_seq=8
655 B: sends SYN|ACK, seq=7, ack_seq=8
657 So we are now A eating this SYN|ACK, ACK test passes. So
658 does sequence test, SYN is truncated, and thus we consider
661 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this
662 bare ACK. Otherwise, we create an established connection. Both
663 ends (listening sockets) accept the new incoming connection and try
664 to talk to each other. 8-)
666 Note: This case is both harmless, and rare. Possibility is about the
667 same as us discovering intelligent life on another plant tomorrow.
669 But generally, we should (RFC lies!) to accept ACK
670 from SYNACK both here and in tcp_rcv_state_process().
671 tcp_rcv_state_process() does not, hence, we do not too.
673 Note that the case is absolutely generic:
674 we cannot optimize anything here without
675 violating protocol. All the checks must be made
676 before attempt to create socket.
679 /* RFC793 page 36: "If the connection is in any non-synchronized state ...
680 * and the incoming segment acknowledges something not yet
681 * sent (the segment carries an unacceptable ACK) ...
684 * Invalid ACK: reset will be sent by listening socket.
685 * Note that the ACK validity check for a Fast Open socket is done
686 * elsewhere and is checked directly against the child socket rather
687 * than req because user data may have been sent out.
689 if ((flg & TCP_FLAG_ACK) && !fastopen &&
690 (TCP_SKB_CB(skb)->ack_seq !=
691 tcp_rsk(req)->snt_isn + 1))
694 /* Also, it would be not so bad idea to check rcv_tsecr, which
695 * is essentially ACK extension and too early or too late values
696 * should cause reset in unsynchronized states.
699 /* RFC793: "first check sequence number". */
701 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
702 tcp_rsk(req)->rcv_nxt, tcp_rsk(req)->rcv_nxt + req->rcv_wnd)) {
703 /* Out of window: send ACK and drop. */
704 if (!(flg & TCP_FLAG_RST))
705 req->rsk_ops->send_ack(sk, skb, req);
707 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
711 /* In sequence, PAWS is OK. */
713 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_nxt))
714 req->ts_recent = tmp_opt.rcv_tsval;
716 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) {
717 /* Truncate SYN, it is out of window starting
718 at tcp_rsk(req)->rcv_isn + 1. */
719 flg &= ~TCP_FLAG_SYN;
722 /* RFC793: "second check the RST bit" and
723 * "fourth, check the SYN bit"
725 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) {
726 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
727 goto embryonic_reset;
730 /* ACK sequence verified above, just make sure ACK is
731 * set. If ACK not set, just silently drop the packet.
733 * XXX (TFO) - if we ever allow "data after SYN", the
734 * following check needs to be removed.
736 if (!(flg & TCP_FLAG_ACK))
739 /* For Fast Open no more processing is needed (sk is the
745 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */
746 if (req->num_timeout < inet_csk(sk)->icsk_accept_queue.rskq_defer_accept &&
747 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) {
748 inet_rsk(req)->acked = 1;
749 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP);
753 /* OK, ACK is valid, create big socket and
754 * feed this segment to it. It will repeat all
755 * the tests. THIS SEGMENT MUST MOVE SOCKET TO
756 * ESTABLISHED STATE. If it will be dropped after
757 * socket is created, wait for troubles.
759 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL);
761 goto listen_overflow;
763 inet_csk_reqsk_queue_drop(sk, req);
764 inet_csk_reqsk_queue_add(sk, req, child);
765 /* Warning: caller must not call reqsk_put(req);
766 * child stole last reference on it.
771 if (!sysctl_tcp_abort_on_overflow) {
772 inet_rsk(req)->acked = 1;
777 if (!(flg & TCP_FLAG_RST)) {
778 /* Received a bad SYN pkt - for TFO We try not to reset
779 * the local connection unless it's really necessary to
780 * avoid becoming vulnerable to outside attack aiming at
781 * resetting legit local connections.
783 req->rsk_ops->send_reset(sk, skb);
784 } else if (fastopen) { /* received a valid RST pkt */
785 reqsk_fastopen_remove(sk, req, true);
789 inet_csk_reqsk_queue_drop(sk, req);
790 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_EMBRYONICRSTS);
794 EXPORT_SYMBOL(tcp_check_req);
797 * Queue segment on the new socket if the new socket is active,
798 * otherwise we just shortcircuit this and continue with
801 * For the vast majority of cases child->sk_state will be TCP_SYN_RECV
802 * when entering. But other states are possible due to a race condition
803 * where after __inet_lookup_established() fails but before the listener
804 * locked is obtained, other packets cause the same connection to
808 int tcp_child_process(struct sock *parent, struct sock *child,
812 int state = child->sk_state;
814 if (!sock_owned_by_user(child)) {
815 ret = tcp_rcv_state_process(child, skb, tcp_hdr(skb),
817 /* Wakeup parent, send SIGIO */
818 if (state == TCP_SYN_RECV && child->sk_state != state)
819 parent->sk_data_ready(parent);
821 /* Alas, it is possible again, because we do lookup
822 * in main socket hash table and lock on listening
823 * socket does not protect us more.
825 __sk_add_backlog(child, skb);
828 bh_unlock_sock(child);
832 EXPORT_SYMBOL(tcp_child_process);