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 .period = TCP_TIMEWAIT_LEN / INET_TWDR_TWKILL_SLOTS,
38 .death_lock = __SPIN_LOCK_UNLOCKED(tcp_death_row.death_lock),
39 .hashinfo = &tcp_hashinfo,
40 .tw_timer = TIMER_INITIALIZER(inet_twdr_hangman, 0,
41 (unsigned long)&tcp_death_row),
42 .twkill_work = __WORK_INITIALIZER(tcp_death_row.twkill_work,
43 inet_twdr_twkill_work),
44 /* Short-time timewait calendar */
47 .twcal_timer = TIMER_INITIALIZER(inet_twdr_twcal_tick, 0,
48 (unsigned long)&tcp_death_row),
50 EXPORT_SYMBOL_GPL(tcp_death_row);
52 static bool tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win)
56 if (after(end_seq, s_win) && before(seq, e_win))
58 return seq == e_win && seq == end_seq;
62 * * Main purpose of TIME-WAIT state is to close connection gracefully,
63 * when one of ends sits in LAST-ACK or CLOSING retransmitting FIN
64 * (and, probably, tail of data) and one or more our ACKs are lost.
65 * * What is TIME-WAIT timeout? It is associated with maximal packet
66 * lifetime in the internet, which results in wrong conclusion, that
67 * it is set to catch "old duplicate segments" wandering out of their path.
68 * It is not quite correct. This timeout is calculated so that it exceeds
69 * maximal retransmission timeout enough to allow to lose one (or more)
70 * segments sent by peer and our ACKs. This time may be calculated from RTO.
71 * * When TIME-WAIT socket receives RST, it means that another end
72 * finally closed and we are allowed to kill TIME-WAIT too.
73 * * Second purpose of TIME-WAIT is catching old duplicate segments.
74 * Well, certainly it is pure paranoia, but if we load TIME-WAIT
75 * with this semantics, we MUST NOT kill TIME-WAIT state with RSTs.
76 * * If we invented some more clever way to catch duplicates
77 * (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs.
79 * The algorithm below is based on FORMAL INTERPRETATION of RFCs.
80 * When you compare it to RFCs, please, read section SEGMENT ARRIVES
81 * from the very beginning.
83 * NOTE. With recycling (and later with fin-wait-2) TW bucket
84 * is _not_ stateless. It means, that strictly speaking we must
85 * spinlock it. I do not want! Well, probability of misbehaviour
86 * is ridiculously low and, seems, we could use some mb() tricks
87 * to avoid misread sequence numbers, states etc. --ANK
89 * We don't need to initialize tmp_out.sack_ok as we don't use the results
92 tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb,
93 const struct tcphdr *th)
95 struct tcp_options_received tmp_opt;
96 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
97 bool paws_reject = false;
99 tmp_opt.saw_tstamp = 0;
100 if (th->doff > (sizeof(*th) >> 2) && tcptw->tw_ts_recent_stamp) {
101 tcp_parse_options(skb, &tmp_opt, 0, NULL);
103 if (tmp_opt.saw_tstamp) {
104 tmp_opt.rcv_tsecr -= tcptw->tw_ts_offset;
105 tmp_opt.ts_recent = tcptw->tw_ts_recent;
106 tmp_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp;
107 paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
111 if (tw->tw_substate == TCP_FIN_WAIT2) {
112 /* Just repeat all the checks of tcp_rcv_state_process() */
114 /* Out of window, send ACK */
116 !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
118 tcptw->tw_rcv_nxt + tcptw->tw_rcv_wnd))
124 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt))
129 !after(TCP_SKB_CB(skb)->end_seq, tcptw->tw_rcv_nxt) ||
130 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) {
132 return TCP_TW_SUCCESS;
135 /* New data or FIN. If new data arrive after half-duplex close,
139 TCP_SKB_CB(skb)->end_seq != tcptw->tw_rcv_nxt + 1) {
141 inet_twsk_deschedule(tw, &tcp_death_row);
146 /* FIN arrived, enter true time-wait state. */
147 tw->tw_substate = TCP_TIME_WAIT;
148 tcptw->tw_rcv_nxt = TCP_SKB_CB(skb)->end_seq;
149 if (tmp_opt.saw_tstamp) {
150 tcptw->tw_ts_recent_stamp = get_seconds();
151 tcptw->tw_ts_recent = tmp_opt.rcv_tsval;
154 if (tcp_death_row.sysctl_tw_recycle &&
155 tcptw->tw_ts_recent_stamp &&
156 tcp_tw_remember_stamp(tw))
157 inet_twsk_schedule(tw, &tcp_death_row, tw->tw_timeout,
160 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN,
166 * Now real TIME-WAIT state.
169 * "When a connection is [...] on TIME-WAIT state [...]
170 * [a TCP] MAY accept a new SYN from the remote TCP to
171 * reopen the connection directly, if it:
173 * (1) assigns its initial sequence number for the new
174 * connection to be larger than the largest sequence
175 * number it used on the previous connection incarnation,
178 * (2) returns to TIME-WAIT state if the SYN turns out
179 * to be an old duplicate".
183 (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt &&
184 (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) {
185 /* In window segment, it may be only reset or bare ack. */
188 /* This is TIME_WAIT assassination, in two flavors.
189 * Oh well... nobody has a sufficient solution to this
192 if (sysctl_tcp_rfc1337 == 0) {
194 inet_twsk_deschedule(tw, &tcp_death_row);
196 return TCP_TW_SUCCESS;
199 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN,
202 if (tmp_opt.saw_tstamp) {
203 tcptw->tw_ts_recent = tmp_opt.rcv_tsval;
204 tcptw->tw_ts_recent_stamp = get_seconds();
208 return TCP_TW_SUCCESS;
211 /* Out of window segment.
213 All the segments are ACKed immediately.
215 The only exception is new SYN. We accept it, if it is
216 not old duplicate and we are not in danger to be killed
217 by delayed old duplicates. RFC check is that it has
218 newer sequence number works at rates <40Mbit/sec.
219 However, if paws works, it is reliable AND even more,
220 we even may relax silly seq space cutoff.
222 RED-PEN: we violate main RFC requirement, if this SYN will appear
223 old duplicate (i.e. we receive RST in reply to SYN-ACK),
224 we must return socket to time-wait state. It is not good,
228 if (th->syn && !th->rst && !th->ack && !paws_reject &&
229 (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) ||
230 (tmp_opt.saw_tstamp &&
231 (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) {
232 u32 isn = tcptw->tw_snd_nxt + 65535 + 2;
235 TCP_SKB_CB(skb)->tcp_tw_isn = isn;
240 NET_INC_STATS_BH(twsk_net(tw), LINUX_MIB_PAWSESTABREJECTED);
243 /* In this case we must reset the TIMEWAIT timer.
245 * If it is ACKless SYN it may be both old duplicate
246 * and new good SYN with random sequence number <rcv_nxt.
247 * Do not reschedule in the last case.
249 if (paws_reject || th->ack)
250 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN,
253 /* Send ACK. Note, we do not put the bucket,
254 * it will be released by caller.
259 return TCP_TW_SUCCESS;
261 EXPORT_SYMBOL(tcp_timewait_state_process);
264 * Move a socket to time-wait or dead fin-wait-2 state.
266 void tcp_time_wait(struct sock *sk, int state, int timeo)
268 struct inet_timewait_sock *tw = NULL;
269 const struct inet_connection_sock *icsk = inet_csk(sk);
270 const struct tcp_sock *tp = tcp_sk(sk);
271 bool recycle_ok = false;
273 if (tcp_death_row.sysctl_tw_recycle && tp->rx_opt.ts_recent_stamp)
274 recycle_ok = tcp_remember_stamp(sk);
276 if (tcp_death_row.tw_count < tcp_death_row.sysctl_max_tw_buckets)
277 tw = inet_twsk_alloc(sk, 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;
293 #if IS_ENABLED(CONFIG_IPV6)
294 if (tw->tw_family == PF_INET6) {
295 struct ipv6_pinfo *np = inet6_sk(sk);
297 tw->tw_v6_daddr = sk->sk_v6_daddr;
298 tw->tw_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
299 tw->tw_tclass = np->tclass;
300 tw->tw_flowlabel = np->flow_label >> 12;
301 tw->tw_ipv6only = sk->sk_ipv6only;
305 #ifdef CONFIG_TCP_MD5SIG
307 * The timewait bucket does not have the key DB from the
308 * sock structure. We just make a quick copy of the
309 * md5 key being used (if indeed we are using one)
310 * so the timewait ack generating code has the key.
313 struct tcp_md5sig_key *key;
314 tcptw->tw_md5_key = NULL;
315 key = tp->af_specific->md5_lookup(sk, sk);
317 tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC);
318 if (tcptw->tw_md5_key && !tcp_alloc_md5sig_pool())
324 /* Linkage updates. */
325 __inet_twsk_hashdance(tw, sk, &tcp_hashinfo);
327 /* Get the TIME_WAIT timeout firing. */
332 tw->tw_timeout = rto;
334 tw->tw_timeout = TCP_TIMEWAIT_LEN;
335 if (state == TCP_TIME_WAIT)
336 timeo = TCP_TIMEWAIT_LEN;
339 inet_twsk_schedule(tw, &tcp_death_row, 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 (!ca_got_dst && !try_module_get(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 /* This is not only more efficient than what we used to do, it eliminates
429 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM
431 * Actually, we could lots of memory writes here. tp of listening
432 * socket contains all necessary default parameters.
434 struct sock *tcp_create_openreq_child(struct sock *sk, struct request_sock *req, struct sk_buff *skb)
436 struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC);
439 const struct inet_request_sock *ireq = inet_rsk(req);
440 struct tcp_request_sock *treq = tcp_rsk(req);
441 struct inet_connection_sock *newicsk = inet_csk(newsk);
442 struct tcp_sock *newtp = tcp_sk(newsk);
444 /* Now setup tcp_sock */
445 newtp->pred_flags = 0;
447 newtp->rcv_wup = newtp->copied_seq =
448 newtp->rcv_nxt = treq->rcv_isn + 1;
450 newtp->snd_sml = newtp->snd_una =
451 newtp->snd_nxt = newtp->snd_up = treq->snt_isn + 1;
453 tcp_prequeue_init(newtp);
454 INIT_LIST_HEAD(&newtp->tsq_node);
456 tcp_init_wl(newtp, treq->rcv_isn);
459 newtp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
460 newicsk->icsk_rto = TCP_TIMEOUT_INIT;
462 newtp->packets_out = 0;
463 newtp->retrans_out = 0;
464 newtp->sacked_out = 0;
465 newtp->fackets_out = 0;
466 newtp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
467 tcp_enable_early_retrans(newtp);
468 newtp->tlp_high_seq = 0;
469 newtp->lsndtime = treq->snt_synack;
470 newtp->last_oow_ack_time = 0;
471 newtp->total_retrans = req->num_retrans;
473 /* So many TCP implementations out there (incorrectly) count the
474 * initial SYN frame in their delayed-ACK and congestion control
475 * algorithms that we must have the following bandaid to talk
476 * efficiently to them. -DaveM
478 newtp->snd_cwnd = TCP_INIT_CWND;
479 newtp->snd_cwnd_cnt = 0;
481 tcp_init_xmit_timers(newsk);
482 __skb_queue_head_init(&newtp->out_of_order_queue);
483 newtp->write_seq = newtp->pushed_seq = treq->snt_isn + 1;
485 newtp->rx_opt.saw_tstamp = 0;
487 newtp->rx_opt.dsack = 0;
488 newtp->rx_opt.num_sacks = 0;
492 if (sock_flag(newsk, SOCK_KEEPOPEN))
493 inet_csk_reset_keepalive_timer(newsk,
494 keepalive_time_when(newtp));
496 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok;
497 if ((newtp->rx_opt.sack_ok = ireq->sack_ok) != 0) {
499 tcp_enable_fack(newtp);
501 newtp->window_clamp = req->window_clamp;
502 newtp->rcv_ssthresh = req->rcv_wnd;
503 newtp->rcv_wnd = req->rcv_wnd;
504 newtp->rx_opt.wscale_ok = ireq->wscale_ok;
505 if (newtp->rx_opt.wscale_ok) {
506 newtp->rx_opt.snd_wscale = ireq->snd_wscale;
507 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale;
509 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0;
510 newtp->window_clamp = min(newtp->window_clamp, 65535U);
512 newtp->snd_wnd = (ntohs(tcp_hdr(skb)->window) <<
513 newtp->rx_opt.snd_wscale);
514 newtp->max_window = newtp->snd_wnd;
516 if (newtp->rx_opt.tstamp_ok) {
517 newtp->rx_opt.ts_recent = req->ts_recent;
518 newtp->rx_opt.ts_recent_stamp = get_seconds();
519 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
521 newtp->rx_opt.ts_recent_stamp = 0;
522 newtp->tcp_header_len = sizeof(struct tcphdr);
525 #ifdef CONFIG_TCP_MD5SIG
526 newtp->md5sig_info = NULL; /*XXX*/
527 if (newtp->af_specific->md5_lookup(sk, newsk))
528 newtp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
530 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len)
531 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len;
532 newtp->rx_opt.mss_clamp = req->mss;
533 tcp_ecn_openreq_child(newtp, req);
534 newtp->fastopen_rsk = NULL;
535 newtp->syn_data_acked = 0;
537 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_PASSIVEOPENS);
541 EXPORT_SYMBOL(tcp_create_openreq_child);
544 * Process an incoming packet for SYN_RECV sockets represented as a
545 * request_sock. Normally sk is the listener socket but for TFO it
546 * points to the child socket.
548 * XXX (TFO) - The current impl contains a special check for ack
549 * validation and inside tcp_v4_reqsk_send_ack(). Can we do better?
551 * We don't need to initialize tmp_opt.sack_ok as we don't use the results
554 struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb,
555 struct request_sock *req,
556 struct request_sock **prev,
559 struct tcp_options_received tmp_opt;
561 const struct tcphdr *th = tcp_hdr(skb);
562 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK);
563 bool paws_reject = false;
565 BUG_ON(fastopen == (sk->sk_state == TCP_LISTEN));
567 tmp_opt.saw_tstamp = 0;
568 if (th->doff > (sizeof(struct tcphdr)>>2)) {
569 tcp_parse_options(skb, &tmp_opt, 0, NULL);
571 if (tmp_opt.saw_tstamp) {
572 tmp_opt.ts_recent = req->ts_recent;
573 /* We do not store true stamp, but it is not required,
574 * it can be estimated (approximately)
577 tmp_opt.ts_recent_stamp = get_seconds() - ((TCP_TIMEOUT_INIT/HZ)<<req->num_timeout);
578 paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
582 /* Check for pure retransmitted SYN. */
583 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn &&
584 flg == TCP_FLAG_SYN &&
587 * RFC793 draws (Incorrectly! It was fixed in RFC1122)
588 * this case on figure 6 and figure 8, but formal
589 * protocol description says NOTHING.
590 * To be more exact, it says that we should send ACK,
591 * because this segment (at least, if it has no data)
594 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT
595 * describe SYN-RECV state. All the description
596 * is wrong, we cannot believe to it and should
597 * rely only on common sense and implementation
600 * Enforce "SYN-ACK" according to figure 8, figure 6
601 * of RFC793, fixed by RFC1122.
603 * Note that even if there is new data in the SYN packet
604 * they will be thrown away too.
606 * Reset timer after retransmitting SYNACK, similar to
607 * the idea of fast retransmit in recovery.
609 if (!tcp_oow_rate_limited(sock_net(sk), skb,
610 LINUX_MIB_TCPACKSKIPPEDSYNRECV,
611 &tcp_rsk(req)->last_oow_ack_time) &&
613 !inet_rtx_syn_ack(sk, req))
614 req->expires = min(TCP_TIMEOUT_INIT << req->num_timeout,
615 TCP_RTO_MAX) + jiffies;
619 /* Further reproduces section "SEGMENT ARRIVES"
620 for state SYN-RECEIVED of RFC793.
621 It is broken, however, it does not work only
622 when SYNs are crossed.
624 You would think that SYN crossing is impossible here, since
625 we should have a SYN_SENT socket (from connect()) on our end,
626 but this is not true if the crossed SYNs were sent to both
627 ends by a malicious third party. We must defend against this,
628 and to do that we first verify the ACK (as per RFC793, page
629 36) and reset if it is invalid. Is this a true full defense?
630 To convince ourselves, let us consider a way in which the ACK
631 test can still pass in this 'malicious crossed SYNs' case.
632 Malicious sender sends identical SYNs (and thus identical sequence
633 numbers) to both A and B:
638 By our good fortune, both A and B select the same initial
639 send sequence number of seven :-)
641 A: sends SYN|ACK, seq=7, ack_seq=8
642 B: sends SYN|ACK, seq=7, ack_seq=8
644 So we are now A eating this SYN|ACK, ACK test passes. So
645 does sequence test, SYN is truncated, and thus we consider
648 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this
649 bare ACK. Otherwise, we create an established connection. Both
650 ends (listening sockets) accept the new incoming connection and try
651 to talk to each other. 8-)
653 Note: This case is both harmless, and rare. Possibility is about the
654 same as us discovering intelligent life on another plant tomorrow.
656 But generally, we should (RFC lies!) to accept ACK
657 from SYNACK both here and in tcp_rcv_state_process().
658 tcp_rcv_state_process() does not, hence, we do not too.
660 Note that the case is absolutely generic:
661 we cannot optimize anything here without
662 violating protocol. All the checks must be made
663 before attempt to create socket.
666 /* RFC793 page 36: "If the connection is in any non-synchronized state ...
667 * and the incoming segment acknowledges something not yet
668 * sent (the segment carries an unacceptable ACK) ...
671 * Invalid ACK: reset will be sent by listening socket.
672 * Note that the ACK validity check for a Fast Open socket is done
673 * elsewhere and is checked directly against the child socket rather
674 * than req because user data may have been sent out.
676 if ((flg & TCP_FLAG_ACK) && !fastopen &&
677 (TCP_SKB_CB(skb)->ack_seq !=
678 tcp_rsk(req)->snt_isn + 1))
681 /* Also, it would be not so bad idea to check rcv_tsecr, which
682 * is essentially ACK extension and too early or too late values
683 * should cause reset in unsynchronized states.
686 /* RFC793: "first check sequence number". */
688 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
689 tcp_rsk(req)->rcv_nxt, tcp_rsk(req)->rcv_nxt + req->rcv_wnd)) {
690 /* Out of window: send ACK and drop. */
691 if (!(flg & TCP_FLAG_RST))
692 req->rsk_ops->send_ack(sk, skb, req);
694 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
698 /* In sequence, PAWS is OK. */
700 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_nxt))
701 req->ts_recent = tmp_opt.rcv_tsval;
703 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) {
704 /* Truncate SYN, it is out of window starting
705 at tcp_rsk(req)->rcv_isn + 1. */
706 flg &= ~TCP_FLAG_SYN;
709 /* RFC793: "second check the RST bit" and
710 * "fourth, check the SYN bit"
712 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) {
713 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
714 goto embryonic_reset;
717 /* ACK sequence verified above, just make sure ACK is
718 * set. If ACK not set, just silently drop the packet.
720 * XXX (TFO) - if we ever allow "data after SYN", the
721 * following check needs to be removed.
723 if (!(flg & TCP_FLAG_ACK))
726 /* For Fast Open no more processing is needed (sk is the
732 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */
733 if (req->num_timeout < inet_csk(sk)->icsk_accept_queue.rskq_defer_accept &&
734 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) {
735 inet_rsk(req)->acked = 1;
736 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP);
740 /* OK, ACK is valid, create big socket and
741 * feed this segment to it. It will repeat all
742 * the tests. THIS SEGMENT MUST MOVE SOCKET TO
743 * ESTABLISHED STATE. If it will be dropped after
744 * socket is created, wait for troubles.
746 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL);
748 goto listen_overflow;
750 inet_csk_reqsk_queue_unlink(sk, req, prev);
751 inet_csk_reqsk_queue_removed(sk, req);
753 inet_csk_reqsk_queue_add(sk, req, child);
757 if (!sysctl_tcp_abort_on_overflow) {
758 inet_rsk(req)->acked = 1;
763 if (!(flg & TCP_FLAG_RST)) {
764 /* Received a bad SYN pkt - for TFO We try not to reset
765 * the local connection unless it's really necessary to
766 * avoid becoming vulnerable to outside attack aiming at
767 * resetting legit local connections.
769 req->rsk_ops->send_reset(sk, skb);
770 } else if (fastopen) { /* received a valid RST pkt */
771 reqsk_fastopen_remove(sk, req, true);
775 inet_csk_reqsk_queue_drop(sk, req, prev);
776 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_EMBRYONICRSTS);
780 EXPORT_SYMBOL(tcp_check_req);
783 * Queue segment on the new socket if the new socket is active,
784 * otherwise we just shortcircuit this and continue with
787 * For the vast majority of cases child->sk_state will be TCP_SYN_RECV
788 * when entering. But other states are possible due to a race condition
789 * where after __inet_lookup_established() fails but before the listener
790 * locked is obtained, other packets cause the same connection to
794 int tcp_child_process(struct sock *parent, struct sock *child,
798 int state = child->sk_state;
800 if (!sock_owned_by_user(child)) {
801 ret = tcp_rcv_state_process(child, skb, tcp_hdr(skb),
803 /* Wakeup parent, send SIGIO */
804 if (state == TCP_SYN_RECV && child->sk_state != state)
805 parent->sk_data_ready(parent);
807 /* Alas, it is possible again, because we do lookup
808 * in main socket hash table and lock on listening
809 * socket does not protect us more.
811 __sk_add_backlog(child, skb);
814 bh_unlock_sock(child);
818 EXPORT_SYMBOL(tcp_child_process);