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_reschedule(tw, tw->tw_timeout);
167 inet_twsk_reschedule(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_reschedule(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_reschedule(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 /* Get the TIME_WAIT timeout firing. */
330 tw->tw_timeout = rto;
332 tw->tw_timeout = TCP_TIMEWAIT_LEN;
333 if (state == TCP_TIME_WAIT)
334 timeo = TCP_TIMEWAIT_LEN;
337 inet_twsk_schedule(tw, timeo);
338 /* Linkage updates. */
339 __inet_twsk_hashdance(tw, sk, &tcp_hashinfo);
342 /* Sorry, if we're out of memory, just CLOSE this
343 * socket up. We've got bigger problems than
344 * non-graceful socket closings.
346 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPTIMEWAITOVERFLOW);
349 tcp_update_metrics(sk);
353 void tcp_twsk_destructor(struct sock *sk)
355 #ifdef CONFIG_TCP_MD5SIG
356 struct tcp_timewait_sock *twsk = tcp_twsk(sk);
358 if (twsk->tw_md5_key)
359 kfree_rcu(twsk->tw_md5_key, rcu);
362 EXPORT_SYMBOL_GPL(tcp_twsk_destructor);
364 /* Warning : This function is called without sk_listener being locked.
365 * Be sure to read socket fields once, as their value could change under us.
367 void tcp_openreq_init_rwin(struct request_sock *req,
368 const struct sock *sk_listener,
369 const struct dst_entry *dst)
371 struct inet_request_sock *ireq = inet_rsk(req);
372 const struct tcp_sock *tp = tcp_sk(sk_listener);
373 u16 user_mss = READ_ONCE(tp->rx_opt.user_mss);
374 int full_space = tcp_full_space(sk_listener);
375 int mss = dst_metric_advmss(dst);
379 if (user_mss && user_mss < mss)
382 window_clamp = READ_ONCE(tp->window_clamp);
383 /* Set this up on the first call only */
384 req->rsk_window_clamp = window_clamp ? : dst_metric(dst, RTAX_WINDOW);
386 /* limit the window selection if the user enforce a smaller rx buffer */
387 if (sk_listener->sk_userlocks & SOCK_RCVBUF_LOCK &&
388 (req->rsk_window_clamp > full_space || req->rsk_window_clamp == 0))
389 req->rsk_window_clamp = full_space;
391 /* tcp_full_space because it is guaranteed to be the first packet */
392 tcp_select_initial_window(full_space,
393 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
395 &req->rsk_window_clamp,
398 dst_metric(dst, RTAX_INITRWND));
399 ireq->rcv_wscale = rcv_wscale;
401 EXPORT_SYMBOL(tcp_openreq_init_rwin);
403 static void tcp_ecn_openreq_child(struct tcp_sock *tp,
404 const struct request_sock *req)
406 tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0;
409 void tcp_ca_openreq_child(struct sock *sk, const struct dst_entry *dst)
411 struct inet_connection_sock *icsk = inet_csk(sk);
412 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
413 bool ca_got_dst = false;
415 if (ca_key != TCP_CA_UNSPEC) {
416 const struct tcp_congestion_ops *ca;
419 ca = tcp_ca_find_key(ca_key);
420 if (likely(ca && try_module_get(ca->owner))) {
421 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
422 icsk->icsk_ca_ops = ca;
428 /* If no valid choice made yet, assign current system default ca. */
430 (!icsk->icsk_ca_setsockopt ||
431 !try_module_get(icsk->icsk_ca_ops->owner)))
432 tcp_assign_congestion_control(sk);
434 tcp_set_ca_state(sk, TCP_CA_Open);
436 EXPORT_SYMBOL_GPL(tcp_ca_openreq_child);
438 /* This is not only more efficient than what we used to do, it eliminates
439 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM
441 * Actually, we could lots of memory writes here. tp of listening
442 * socket contains all necessary default parameters.
444 struct sock *tcp_create_openreq_child(const struct sock *sk,
445 struct request_sock *req,
448 struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC);
451 const struct inet_request_sock *ireq = inet_rsk(req);
452 struct tcp_request_sock *treq = tcp_rsk(req);
453 struct inet_connection_sock *newicsk = inet_csk(newsk);
454 struct tcp_sock *newtp = tcp_sk(newsk);
456 /* Now setup tcp_sock */
457 newtp->pred_flags = 0;
459 newtp->rcv_wup = newtp->copied_seq =
460 newtp->rcv_nxt = treq->rcv_isn + 1;
463 newtp->snd_sml = newtp->snd_una =
464 newtp->snd_nxt = newtp->snd_up = treq->snt_isn + 1;
466 tcp_prequeue_init(newtp);
467 INIT_LIST_HEAD(&newtp->tsq_node);
469 tcp_init_wl(newtp, treq->rcv_isn);
472 newtp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
473 newtp->rtt_min[0].rtt = ~0U;
474 newicsk->icsk_rto = TCP_TIMEOUT_INIT;
476 newtp->packets_out = 0;
477 newtp->retrans_out = 0;
478 newtp->sacked_out = 0;
479 newtp->fackets_out = 0;
480 newtp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
481 tcp_enable_early_retrans(newtp);
482 newtp->tlp_high_seq = 0;
483 newtp->lsndtime = treq->snt_synack.stamp_jiffies;
484 newsk->sk_txhash = treq->txhash;
485 newtp->last_oow_ack_time = 0;
486 newtp->total_retrans = req->num_retrans;
488 /* So many TCP implementations out there (incorrectly) count the
489 * initial SYN frame in their delayed-ACK and congestion control
490 * algorithms that we must have the following bandaid to talk
491 * efficiently to them. -DaveM
493 newtp->snd_cwnd = TCP_INIT_CWND;
494 newtp->snd_cwnd_cnt = 0;
496 tcp_init_xmit_timers(newsk);
497 __skb_queue_head_init(&newtp->out_of_order_queue);
498 newtp->write_seq = newtp->pushed_seq = treq->snt_isn + 1;
500 newtp->rx_opt.saw_tstamp = 0;
502 newtp->rx_opt.dsack = 0;
503 newtp->rx_opt.num_sacks = 0;
507 if (sock_flag(newsk, SOCK_KEEPOPEN))
508 inet_csk_reset_keepalive_timer(newsk,
509 keepalive_time_when(newtp));
511 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok;
512 if ((newtp->rx_opt.sack_ok = ireq->sack_ok) != 0) {
514 tcp_enable_fack(newtp);
516 newtp->window_clamp = req->rsk_window_clamp;
517 newtp->rcv_ssthresh = req->rsk_rcv_wnd;
518 newtp->rcv_wnd = req->rsk_rcv_wnd;
519 newtp->rx_opt.wscale_ok = ireq->wscale_ok;
520 if (newtp->rx_opt.wscale_ok) {
521 newtp->rx_opt.snd_wscale = ireq->snd_wscale;
522 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale;
524 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0;
525 newtp->window_clamp = min(newtp->window_clamp, 65535U);
527 newtp->snd_wnd = (ntohs(tcp_hdr(skb)->window) <<
528 newtp->rx_opt.snd_wscale);
529 newtp->max_window = newtp->snd_wnd;
531 if (newtp->rx_opt.tstamp_ok) {
532 newtp->rx_opt.ts_recent = req->ts_recent;
533 newtp->rx_opt.ts_recent_stamp = get_seconds();
534 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
536 newtp->rx_opt.ts_recent_stamp = 0;
537 newtp->tcp_header_len = sizeof(struct tcphdr);
540 #ifdef CONFIG_TCP_MD5SIG
541 newtp->md5sig_info = NULL; /*XXX*/
542 if (newtp->af_specific->md5_lookup(sk, newsk))
543 newtp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
545 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len)
546 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len;
547 newtp->rx_opt.mss_clamp = req->mss;
548 tcp_ecn_openreq_child(newtp, req);
549 newtp->fastopen_rsk = NULL;
550 newtp->syn_data_acked = 0;
551 newtp->rack.mstamp.v64 = 0;
552 newtp->rack.advanced = 0;
554 newtp->saved_syn = req->saved_syn;
555 req->saved_syn = NULL;
557 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_PASSIVEOPENS);
561 EXPORT_SYMBOL(tcp_create_openreq_child);
564 * Process an incoming packet for SYN_RECV sockets represented as a
565 * request_sock. Normally sk is the listener socket but for TFO it
566 * points to the child socket.
568 * XXX (TFO) - The current impl contains a special check for ack
569 * validation and inside tcp_v4_reqsk_send_ack(). Can we do better?
571 * We don't need to initialize tmp_opt.sack_ok as we don't use the results
574 struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb,
575 struct request_sock *req,
578 struct tcp_options_received tmp_opt;
580 const struct tcphdr *th = tcp_hdr(skb);
581 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK);
582 bool paws_reject = false;
584 tmp_opt.saw_tstamp = 0;
585 if (th->doff > (sizeof(struct tcphdr)>>2)) {
586 tcp_parse_options(skb, &tmp_opt, 0, NULL);
588 if (tmp_opt.saw_tstamp) {
589 tmp_opt.ts_recent = req->ts_recent;
590 /* We do not store true stamp, but it is not required,
591 * it can be estimated (approximately)
594 tmp_opt.ts_recent_stamp = get_seconds() - ((TCP_TIMEOUT_INIT/HZ)<<req->num_timeout);
595 paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
599 /* Check for pure retransmitted SYN. */
600 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn &&
601 flg == TCP_FLAG_SYN &&
604 * RFC793 draws (Incorrectly! It was fixed in RFC1122)
605 * this case on figure 6 and figure 8, but formal
606 * protocol description says NOTHING.
607 * To be more exact, it says that we should send ACK,
608 * because this segment (at least, if it has no data)
611 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT
612 * describe SYN-RECV state. All the description
613 * is wrong, we cannot believe to it and should
614 * rely only on common sense and implementation
617 * Enforce "SYN-ACK" according to figure 8, figure 6
618 * of RFC793, fixed by RFC1122.
620 * Note that even if there is new data in the SYN packet
621 * they will be thrown away too.
623 * Reset timer after retransmitting SYNACK, similar to
624 * the idea of fast retransmit in recovery.
626 if (!tcp_oow_rate_limited(sock_net(sk), skb,
627 LINUX_MIB_TCPACKSKIPPEDSYNRECV,
628 &tcp_rsk(req)->last_oow_ack_time) &&
630 !inet_rtx_syn_ack(sk, req)) {
631 unsigned long expires = jiffies;
633 expires += min(TCP_TIMEOUT_INIT << req->num_timeout,
636 mod_timer_pending(&req->rsk_timer, expires);
638 req->rsk_timer.expires = expires;
643 /* Further reproduces section "SEGMENT ARRIVES"
644 for state SYN-RECEIVED of RFC793.
645 It is broken, however, it does not work only
646 when SYNs are crossed.
648 You would think that SYN crossing is impossible here, since
649 we should have a SYN_SENT socket (from connect()) on our end,
650 but this is not true if the crossed SYNs were sent to both
651 ends by a malicious third party. We must defend against this,
652 and to do that we first verify the ACK (as per RFC793, page
653 36) and reset if it is invalid. Is this a true full defense?
654 To convince ourselves, let us consider a way in which the ACK
655 test can still pass in this 'malicious crossed SYNs' case.
656 Malicious sender sends identical SYNs (and thus identical sequence
657 numbers) to both A and B:
662 By our good fortune, both A and B select the same initial
663 send sequence number of seven :-)
665 A: sends SYN|ACK, seq=7, ack_seq=8
666 B: sends SYN|ACK, seq=7, ack_seq=8
668 So we are now A eating this SYN|ACK, ACK test passes. So
669 does sequence test, SYN is truncated, and thus we consider
672 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this
673 bare ACK. Otherwise, we create an established connection. Both
674 ends (listening sockets) accept the new incoming connection and try
675 to talk to each other. 8-)
677 Note: This case is both harmless, and rare. Possibility is about the
678 same as us discovering intelligent life on another plant tomorrow.
680 But generally, we should (RFC lies!) to accept ACK
681 from SYNACK both here and in tcp_rcv_state_process().
682 tcp_rcv_state_process() does not, hence, we do not too.
684 Note that the case is absolutely generic:
685 we cannot optimize anything here without
686 violating protocol. All the checks must be made
687 before attempt to create socket.
690 /* RFC793 page 36: "If the connection is in any non-synchronized state ...
691 * and the incoming segment acknowledges something not yet
692 * sent (the segment carries an unacceptable ACK) ...
695 * Invalid ACK: reset will be sent by listening socket.
696 * Note that the ACK validity check for a Fast Open socket is done
697 * elsewhere and is checked directly against the child socket rather
698 * than req because user data may have been sent out.
700 if ((flg & TCP_FLAG_ACK) && !fastopen &&
701 (TCP_SKB_CB(skb)->ack_seq !=
702 tcp_rsk(req)->snt_isn + 1))
705 /* Also, it would be not so bad idea to check rcv_tsecr, which
706 * is essentially ACK extension and too early or too late values
707 * should cause reset in unsynchronized states.
710 /* RFC793: "first check sequence number". */
712 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
713 tcp_rsk(req)->rcv_nxt, tcp_rsk(req)->rcv_nxt + req->rsk_rcv_wnd)) {
714 /* Out of window: send ACK and drop. */
715 if (!(flg & TCP_FLAG_RST))
716 req->rsk_ops->send_ack(sk, skb, req);
718 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
722 /* In sequence, PAWS is OK. */
724 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_nxt))
725 req->ts_recent = tmp_opt.rcv_tsval;
727 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) {
728 /* Truncate SYN, it is out of window starting
729 at tcp_rsk(req)->rcv_isn + 1. */
730 flg &= ~TCP_FLAG_SYN;
733 /* RFC793: "second check the RST bit" and
734 * "fourth, check the SYN bit"
736 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) {
737 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
738 goto embryonic_reset;
741 /* ACK sequence verified above, just make sure ACK is
742 * set. If ACK not set, just silently drop the packet.
744 * XXX (TFO) - if we ever allow "data after SYN", the
745 * following check needs to be removed.
747 if (!(flg & TCP_FLAG_ACK))
750 /* For Fast Open no more processing is needed (sk is the
756 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */
757 if (req->num_timeout < inet_csk(sk)->icsk_accept_queue.rskq_defer_accept &&
758 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) {
759 inet_rsk(req)->acked = 1;
760 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP);
764 /* OK, ACK is valid, create big socket and
765 * feed this segment to it. It will repeat all
766 * the tests. THIS SEGMENT MUST MOVE SOCKET TO
767 * ESTABLISHED STATE. If it will be dropped after
768 * socket is created, wait for troubles.
770 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL);
772 goto listen_overflow;
774 sock_rps_save_rxhash(child, skb);
775 tcp_synack_rtt_meas(child, req);
776 inet_csk_reqsk_queue_drop(sk, req);
777 inet_csk_reqsk_queue_add(sk, req, child);
778 /* Warning: caller must not call reqsk_put(req);
779 * child stole last reference on it.
784 if (!sysctl_tcp_abort_on_overflow) {
785 inet_rsk(req)->acked = 1;
790 if (!(flg & TCP_FLAG_RST)) {
791 /* Received a bad SYN pkt - for TFO We try not to reset
792 * the local connection unless it's really necessary to
793 * avoid becoming vulnerable to outside attack aiming at
794 * resetting legit local connections.
796 req->rsk_ops->send_reset(sk, skb);
797 } else if (fastopen) { /* received a valid RST pkt */
798 reqsk_fastopen_remove(sk, req, true);
802 inet_csk_reqsk_queue_drop(sk, req);
803 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_EMBRYONICRSTS);
807 EXPORT_SYMBOL(tcp_check_req);
810 * Queue segment on the new socket if the new socket is active,
811 * otherwise we just shortcircuit this and continue with
814 * For the vast majority of cases child->sk_state will be TCP_SYN_RECV
815 * when entering. But other states are possible due to a race condition
816 * where after __inet_lookup_established() fails but before the listener
817 * locked is obtained, other packets cause the same connection to
821 int tcp_child_process(struct sock *parent, struct sock *child,
825 int state = child->sk_state;
827 if (!sock_owned_by_user(child)) {
828 ret = tcp_rcv_state_process(child, skb);
829 /* Wakeup parent, send SIGIO */
830 if (state == TCP_SYN_RECV && child->sk_state != state)
831 parent->sk_data_ready(parent);
833 /* Alas, it is possible again, because we do lookup
834 * in main socket hash table and lock on listening
835 * socket does not protect us more.
837 __sk_add_backlog(child, skb);
840 bh_unlock_sock(child);
844 EXPORT_SYMBOL(tcp_child_process);