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>
26 #include <linux/static_key.h>
28 #include <net/inet_common.h>
30 #include <net/busy_poll.h>
32 static bool tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win)
36 if (after(end_seq, s_win) && before(seq, e_win))
38 return seq == e_win && seq == end_seq;
41 static enum tcp_tw_status
42 tcp_timewait_check_oow_rate_limit(struct inet_timewait_sock *tw,
43 const struct sk_buff *skb, int mib_idx)
45 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
47 if (!tcp_oow_rate_limited(twsk_net(tw), skb, mib_idx,
48 &tcptw->tw_last_oow_ack_time)) {
49 /* Send ACK. Note, we do not put the bucket,
50 * it will be released by caller.
55 /* We are rate-limiting, so just release the tw sock and drop skb. */
57 return TCP_TW_SUCCESS;
61 * * Main purpose of TIME-WAIT state is to close connection gracefully,
62 * when one of ends sits in LAST-ACK or CLOSING retransmitting FIN
63 * (and, probably, tail of data) and one or more our ACKs are lost.
64 * * What is TIME-WAIT timeout? It is associated with maximal packet
65 * lifetime in the internet, which results in wrong conclusion, that
66 * it is set to catch "old duplicate segments" wandering out of their path.
67 * It is not quite correct. This timeout is calculated so that it exceeds
68 * maximal retransmission timeout enough to allow to lose one (or more)
69 * segments sent by peer and our ACKs. This time may be calculated from RTO.
70 * * When TIME-WAIT socket receives RST, it means that another end
71 * finally closed and we are allowed to kill TIME-WAIT too.
72 * * Second purpose of TIME-WAIT is catching old duplicate segments.
73 * Well, certainly it is pure paranoia, but if we load TIME-WAIT
74 * with this semantics, we MUST NOT kill TIME-WAIT state with RSTs.
75 * * If we invented some more clever way to catch duplicates
76 * (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs.
78 * The algorithm below is based on FORMAL INTERPRETATION of RFCs.
79 * When you compare it to RFCs, please, read section SEGMENT ARRIVES
80 * from the very beginning.
82 * NOTE. With recycling (and later with fin-wait-2) TW bucket
83 * is _not_ stateless. It means, that strictly speaking we must
84 * spinlock it. I do not want! Well, probability of misbehaviour
85 * is ridiculously low and, seems, we could use some mb() tricks
86 * to avoid misread sequence numbers, states etc. --ANK
88 * We don't need to initialize tmp_out.sack_ok as we don't use the results
91 tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb,
92 const struct tcphdr *th)
94 struct tcp_options_received tmp_opt;
95 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
96 bool paws_reject = false;
98 tmp_opt.saw_tstamp = 0;
99 if (th->doff > (sizeof(*th) >> 2) && tcptw->tw_ts_recent_stamp) {
100 tcp_parse_options(twsk_net(tw), skb, &tmp_opt, 0, NULL);
102 if (tmp_opt.saw_tstamp) {
103 if (tmp_opt.rcv_tsecr)
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))
119 return tcp_timewait_check_oow_rate_limit(
120 tw, skb, LINUX_MIB_TCPACKSKIPPEDFINWAIT2);
125 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt))
130 !after(TCP_SKB_CB(skb)->end_seq, tcptw->tw_rcv_nxt) ||
131 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) {
133 return TCP_TW_SUCCESS;
136 /* New data or FIN. If new data arrive after half-duplex close,
140 TCP_SKB_CB(skb)->end_seq != tcptw->tw_rcv_nxt + 1)
143 /* FIN arrived, enter true time-wait state. */
144 tw->tw_substate = TCP_TIME_WAIT;
145 tcptw->tw_rcv_nxt = TCP_SKB_CB(skb)->end_seq;
146 if (tmp_opt.saw_tstamp) {
147 tcptw->tw_ts_recent_stamp = ktime_get_seconds();
148 tcptw->tw_ts_recent = tmp_opt.rcv_tsval;
151 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
156 * Now real TIME-WAIT state.
159 * "When a connection is [...] on TIME-WAIT state [...]
160 * [a TCP] MAY accept a new SYN from the remote TCP to
161 * reopen the connection directly, if it:
163 * (1) assigns its initial sequence number for the new
164 * connection to be larger than the largest sequence
165 * number it used on the previous connection incarnation,
168 * (2) returns to TIME-WAIT state if the SYN turns out
169 * to be an old duplicate".
173 (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt &&
174 (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) {
175 /* In window segment, it may be only reset or bare ack. */
178 /* This is TIME_WAIT assassination, in two flavors.
179 * Oh well... nobody has a sufficient solution to this
182 if (twsk_net(tw)->ipv4.sysctl_tcp_rfc1337 == 0) {
184 inet_twsk_deschedule_put(tw);
185 return TCP_TW_SUCCESS;
188 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
191 if (tmp_opt.saw_tstamp) {
192 tcptw->tw_ts_recent = tmp_opt.rcv_tsval;
193 tcptw->tw_ts_recent_stamp = ktime_get_seconds();
197 return TCP_TW_SUCCESS;
200 /* Out of window segment.
202 All the segments are ACKed immediately.
204 The only exception is new SYN. We accept it, if it is
205 not old duplicate and we are not in danger to be killed
206 by delayed old duplicates. RFC check is that it has
207 newer sequence number works at rates <40Mbit/sec.
208 However, if paws works, it is reliable AND even more,
209 we even may relax silly seq space cutoff.
211 RED-PEN: we violate main RFC requirement, if this SYN will appear
212 old duplicate (i.e. we receive RST in reply to SYN-ACK),
213 we must return socket to time-wait state. It is not good,
217 if (th->syn && !th->rst && !th->ack && !paws_reject &&
218 (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) ||
219 (tmp_opt.saw_tstamp &&
220 (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) {
221 u32 isn = tcptw->tw_snd_nxt + 65535 + 2;
224 TCP_SKB_CB(skb)->tcp_tw_isn = isn;
229 __NET_INC_STATS(twsk_net(tw), LINUX_MIB_PAWSESTABREJECTED);
232 /* In this case we must reset the TIMEWAIT timer.
234 * If it is ACKless SYN it may be both old duplicate
235 * and new good SYN with random sequence number <rcv_nxt.
236 * Do not reschedule in the last case.
238 if (paws_reject || th->ack)
239 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
241 return tcp_timewait_check_oow_rate_limit(
242 tw, skb, LINUX_MIB_TCPACKSKIPPEDTIMEWAIT);
245 return TCP_TW_SUCCESS;
247 EXPORT_SYMBOL(tcp_timewait_state_process);
250 * Move a socket to time-wait or dead fin-wait-2 state.
252 void tcp_time_wait(struct sock *sk, int state, int timeo)
254 const struct inet_connection_sock *icsk = inet_csk(sk);
255 const struct tcp_sock *tp = tcp_sk(sk);
256 struct inet_timewait_sock *tw;
257 struct inet_timewait_death_row *tcp_death_row = &sock_net(sk)->ipv4.tcp_death_row;
259 tw = inet_twsk_alloc(sk, tcp_death_row, state);
262 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
263 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1);
264 struct inet_sock *inet = inet_sk(sk);
266 tw->tw_transparent = inet->transparent;
267 tw->tw_mark = sk->sk_mark;
268 tw->tw_rcv_wscale = tp->rx_opt.rcv_wscale;
269 tcptw->tw_rcv_nxt = tp->rcv_nxt;
270 tcptw->tw_snd_nxt = tp->snd_nxt;
271 tcptw->tw_rcv_wnd = tcp_receive_window(tp);
272 tcptw->tw_ts_recent = tp->rx_opt.ts_recent;
273 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp;
274 tcptw->tw_ts_offset = tp->tsoffset;
275 tcptw->tw_last_oow_ack_time = 0;
277 #if IS_ENABLED(CONFIG_IPV6)
278 if (tw->tw_family == PF_INET6) {
279 struct ipv6_pinfo *np = inet6_sk(sk);
281 tw->tw_v6_daddr = sk->sk_v6_daddr;
282 tw->tw_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
283 tw->tw_tclass = np->tclass;
284 tw->tw_flowlabel = be32_to_cpu(np->flow_label & IPV6_FLOWLABEL_MASK);
285 tw->tw_ipv6only = sk->sk_ipv6only;
289 #ifdef CONFIG_TCP_MD5SIG
291 * The timewait bucket does not have the key DB from the
292 * sock structure. We just make a quick copy of the
293 * md5 key being used (if indeed we are using one)
294 * so the timewait ack generating code has the key.
297 struct tcp_md5sig_key *key;
298 tcptw->tw_md5_key = NULL;
299 key = tp->af_specific->md5_lookup(sk, sk);
301 tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC);
302 BUG_ON(tcptw->tw_md5_key && !tcp_alloc_md5sig_pool());
307 /* Get the TIME_WAIT timeout firing. */
311 if (state == TCP_TIME_WAIT)
312 timeo = TCP_TIMEWAIT_LEN;
314 /* tw_timer is pinned, so we need to make sure BH are disabled
315 * in following section, otherwise timer handler could run before
316 * we complete the initialization.
319 inet_twsk_schedule(tw, timeo);
321 * Note that access to tw after this point is illegal.
323 inet_twsk_hashdance(tw, sk, &tcp_hashinfo);
326 /* Sorry, if we're out of memory, just CLOSE this
327 * socket up. We've got bigger problems than
328 * non-graceful socket closings.
330 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEWAITOVERFLOW);
333 tcp_update_metrics(sk);
336 EXPORT_SYMBOL(tcp_time_wait);
338 void tcp_twsk_destructor(struct sock *sk)
340 #ifdef CONFIG_TCP_MD5SIG
341 struct tcp_timewait_sock *twsk = tcp_twsk(sk);
343 if (twsk->tw_md5_key)
344 kfree_rcu(twsk->tw_md5_key, rcu);
347 EXPORT_SYMBOL_GPL(tcp_twsk_destructor);
349 /* Warning : This function is called without sk_listener being locked.
350 * Be sure to read socket fields once, as their value could change under us.
352 void tcp_openreq_init_rwin(struct request_sock *req,
353 const struct sock *sk_listener,
354 const struct dst_entry *dst)
356 struct inet_request_sock *ireq = inet_rsk(req);
357 const struct tcp_sock *tp = tcp_sk(sk_listener);
358 int full_space = tcp_full_space(sk_listener);
364 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
365 window_clamp = READ_ONCE(tp->window_clamp);
366 /* Set this up on the first call only */
367 req->rsk_window_clamp = window_clamp ? : dst_metric(dst, RTAX_WINDOW);
369 /* limit the window selection if the user enforce a smaller rx buffer */
370 if (sk_listener->sk_userlocks & SOCK_RCVBUF_LOCK &&
371 (req->rsk_window_clamp > full_space || req->rsk_window_clamp == 0))
372 req->rsk_window_clamp = full_space;
374 rcv_wnd = tcp_rwnd_init_bpf((struct sock *)req);
376 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
377 else if (full_space < rcv_wnd * mss)
378 full_space = rcv_wnd * mss;
380 /* tcp_full_space because it is guaranteed to be the first packet */
381 tcp_select_initial_window(sk_listener, full_space,
382 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
384 &req->rsk_window_clamp,
388 ireq->rcv_wscale = rcv_wscale;
390 EXPORT_SYMBOL(tcp_openreq_init_rwin);
392 static void tcp_ecn_openreq_child(struct tcp_sock *tp,
393 const struct request_sock *req)
395 tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0;
398 void tcp_ca_openreq_child(struct sock *sk, const struct dst_entry *dst)
400 struct inet_connection_sock *icsk = inet_csk(sk);
401 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
402 bool ca_got_dst = false;
404 if (ca_key != TCP_CA_UNSPEC) {
405 const struct tcp_congestion_ops *ca;
408 ca = tcp_ca_find_key(ca_key);
409 if (likely(ca && try_module_get(ca->owner))) {
410 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
411 icsk->icsk_ca_ops = ca;
417 /* If no valid choice made yet, assign current system default ca. */
419 (!icsk->icsk_ca_setsockopt ||
420 !try_module_get(icsk->icsk_ca_ops->owner)))
421 tcp_assign_congestion_control(sk);
423 tcp_set_ca_state(sk, TCP_CA_Open);
425 EXPORT_SYMBOL_GPL(tcp_ca_openreq_child);
427 static void smc_check_reset_syn_req(struct tcp_sock *oldtp,
428 struct request_sock *req,
429 struct tcp_sock *newtp)
431 #if IS_ENABLED(CONFIG_SMC)
432 struct inet_request_sock *ireq;
434 if (static_branch_unlikely(&tcp_have_smc)) {
435 ireq = inet_rsk(req);
436 if (oldtp->syn_smc && !ireq->smc_ok)
442 /* This is not only more efficient than what we used to do, it eliminates
443 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM
445 * Actually, we could lots of memory writes here. tp of listening
446 * socket contains all necessary default parameters.
448 struct sock *tcp_create_openreq_child(const struct sock *sk,
449 struct request_sock *req,
452 struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC);
453 const struct inet_request_sock *ireq = inet_rsk(req);
454 struct tcp_request_sock *treq = tcp_rsk(req);
455 struct inet_connection_sock *newicsk;
456 struct tcp_sock *oldtp, *newtp;
462 newicsk = inet_csk(newsk);
463 newtp = tcp_sk(newsk);
466 smc_check_reset_syn_req(oldtp, req, newtp);
468 /* Now setup tcp_sock */
469 newtp->pred_flags = 0;
471 seq = treq->rcv_isn + 1;
472 newtp->rcv_wup = seq;
473 WRITE_ONCE(newtp->copied_seq, seq);
474 WRITE_ONCE(newtp->rcv_nxt, seq);
477 newtp->snd_sml = newtp->snd_una =
478 newtp->snd_nxt = newtp->snd_up = treq->snt_isn + 1;
480 INIT_LIST_HEAD(&newtp->tsq_node);
481 INIT_LIST_HEAD(&newtp->tsorted_sent_queue);
483 tcp_init_wl(newtp, treq->rcv_isn);
486 newtp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
487 minmax_reset(&newtp->rtt_min, tcp_jiffies32, ~0U);
488 newicsk->icsk_rto = TCP_TIMEOUT_INIT;
489 newicsk->icsk_ack.lrcvtime = tcp_jiffies32;
491 newtp->packets_out = 0;
492 newtp->retrans_out = 0;
493 newtp->sacked_out = 0;
494 newtp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
495 newtp->tlp_high_seq = 0;
496 newtp->lsndtime = tcp_jiffies32;
497 newsk->sk_txhash = treq->txhash;
498 newtp->last_oow_ack_time = 0;
499 newtp->total_retrans = req->num_retrans;
501 /* So many TCP implementations out there (incorrectly) count the
502 * initial SYN frame in their delayed-ACK and congestion control
503 * algorithms that we must have the following bandaid to talk
504 * efficiently to them. -DaveM
506 newtp->snd_cwnd = TCP_INIT_CWND;
507 newtp->snd_cwnd_cnt = 0;
509 /* There's a bubble in the pipe until at least the first ACK. */
510 newtp->app_limited = ~0U;
512 tcp_init_xmit_timers(newsk);
513 WRITE_ONCE(newtp->write_seq, newtp->pushed_seq = treq->snt_isn + 1);
515 newtp->rx_opt.saw_tstamp = 0;
517 newtp->rx_opt.dsack = 0;
518 newtp->rx_opt.num_sacks = 0;
522 if (sock_flag(newsk, SOCK_KEEPOPEN))
523 inet_csk_reset_keepalive_timer(newsk,
524 keepalive_time_when(newtp));
526 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok;
527 newtp->rx_opt.sack_ok = ireq->sack_ok;
528 newtp->window_clamp = req->rsk_window_clamp;
529 newtp->rcv_ssthresh = req->rsk_rcv_wnd;
530 newtp->rcv_wnd = req->rsk_rcv_wnd;
531 newtp->rx_opt.wscale_ok = ireq->wscale_ok;
532 if (newtp->rx_opt.wscale_ok) {
533 newtp->rx_opt.snd_wscale = ireq->snd_wscale;
534 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale;
536 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0;
537 newtp->window_clamp = min(newtp->window_clamp, 65535U);
539 newtp->snd_wnd = ntohs(tcp_hdr(skb)->window) << newtp->rx_opt.snd_wscale;
540 newtp->max_window = newtp->snd_wnd;
542 if (newtp->rx_opt.tstamp_ok) {
543 newtp->rx_opt.ts_recent = req->ts_recent;
544 newtp->rx_opt.ts_recent_stamp = ktime_get_seconds();
545 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
547 newtp->rx_opt.ts_recent_stamp = 0;
548 newtp->tcp_header_len = sizeof(struct tcphdr);
550 newtp->tsoffset = treq->ts_off;
551 #ifdef CONFIG_TCP_MD5SIG
552 newtp->md5sig_info = NULL; /*XXX*/
553 if (newtp->af_specific->md5_lookup(sk, newsk))
554 newtp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
556 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len)
557 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len;
558 newtp->rx_opt.mss_clamp = req->mss;
559 tcp_ecn_openreq_child(newtp, req);
560 newtp->fastopen_req = NULL;
561 newtp->fastopen_rsk = NULL;
562 newtp->syn_data_acked = 0;
563 newtp->rack.mstamp = 0;
564 newtp->rack.advanced = 0;
565 newtp->rack.reo_wnd_steps = 1;
566 newtp->rack.last_delivered = 0;
567 newtp->rack.reo_wnd_persist = 0;
568 newtp->rack.dsack_seen = 0;
570 __TCP_INC_STATS(sock_net(sk), TCP_MIB_PASSIVEOPENS);
574 EXPORT_SYMBOL(tcp_create_openreq_child);
577 * Process an incoming packet for SYN_RECV sockets represented as a
578 * request_sock. Normally sk is the listener socket but for TFO it
579 * points to the child socket.
581 * XXX (TFO) - The current impl contains a special check for ack
582 * validation and inside tcp_v4_reqsk_send_ack(). Can we do better?
584 * We don't need to initialize tmp_opt.sack_ok as we don't use the results
587 struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb,
588 struct request_sock *req,
589 bool fastopen, bool *req_stolen)
591 struct tcp_options_received tmp_opt;
593 const struct tcphdr *th = tcp_hdr(skb);
594 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK);
595 bool paws_reject = false;
598 tmp_opt.saw_tstamp = 0;
599 if (th->doff > (sizeof(struct tcphdr)>>2)) {
600 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0, NULL);
602 if (tmp_opt.saw_tstamp) {
603 tmp_opt.ts_recent = req->ts_recent;
604 if (tmp_opt.rcv_tsecr)
605 tmp_opt.rcv_tsecr -= tcp_rsk(req)->ts_off;
606 /* We do not store true stamp, but it is not required,
607 * it can be estimated (approximately)
610 tmp_opt.ts_recent_stamp = ktime_get_seconds() - ((TCP_TIMEOUT_INIT/HZ)<<req->num_timeout);
611 paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
615 /* Check for pure retransmitted SYN. */
616 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn &&
617 flg == TCP_FLAG_SYN &&
620 * RFC793 draws (Incorrectly! It was fixed in RFC1122)
621 * this case on figure 6 and figure 8, but formal
622 * protocol description says NOTHING.
623 * To be more exact, it says that we should send ACK,
624 * because this segment (at least, if it has no data)
627 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT
628 * describe SYN-RECV state. All the description
629 * is wrong, we cannot believe to it and should
630 * rely only on common sense and implementation
633 * Enforce "SYN-ACK" according to figure 8, figure 6
634 * of RFC793, fixed by RFC1122.
636 * Note that even if there is new data in the SYN packet
637 * they will be thrown away too.
639 * Reset timer after retransmitting SYNACK, similar to
640 * the idea of fast retransmit in recovery.
642 if (!tcp_oow_rate_limited(sock_net(sk), skb,
643 LINUX_MIB_TCPACKSKIPPEDSYNRECV,
644 &tcp_rsk(req)->last_oow_ack_time) &&
646 !inet_rtx_syn_ack(sk, req)) {
647 unsigned long expires = jiffies;
649 expires += min(TCP_TIMEOUT_INIT << req->num_timeout,
652 mod_timer_pending(&req->rsk_timer, expires);
654 req->rsk_timer.expires = expires;
659 /* Further reproduces section "SEGMENT ARRIVES"
660 for state SYN-RECEIVED of RFC793.
661 It is broken, however, it does not work only
662 when SYNs are crossed.
664 You would think that SYN crossing is impossible here, since
665 we should have a SYN_SENT socket (from connect()) on our end,
666 but this is not true if the crossed SYNs were sent to both
667 ends by a malicious third party. We must defend against this,
668 and to do that we first verify the ACK (as per RFC793, page
669 36) and reset if it is invalid. Is this a true full defense?
670 To convince ourselves, let us consider a way in which the ACK
671 test can still pass in this 'malicious crossed SYNs' case.
672 Malicious sender sends identical SYNs (and thus identical sequence
673 numbers) to both A and B:
678 By our good fortune, both A and B select the same initial
679 send sequence number of seven :-)
681 A: sends SYN|ACK, seq=7, ack_seq=8
682 B: sends SYN|ACK, seq=7, ack_seq=8
684 So we are now A eating this SYN|ACK, ACK test passes. So
685 does sequence test, SYN is truncated, and thus we consider
688 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this
689 bare ACK. Otherwise, we create an established connection. Both
690 ends (listening sockets) accept the new incoming connection and try
691 to talk to each other. 8-)
693 Note: This case is both harmless, and rare. Possibility is about the
694 same as us discovering intelligent life on another plant tomorrow.
696 But generally, we should (RFC lies!) to accept ACK
697 from SYNACK both here and in tcp_rcv_state_process().
698 tcp_rcv_state_process() does not, hence, we do not too.
700 Note that the case is absolutely generic:
701 we cannot optimize anything here without
702 violating protocol. All the checks must be made
703 before attempt to create socket.
706 /* RFC793 page 36: "If the connection is in any non-synchronized state ...
707 * and the incoming segment acknowledges something not yet
708 * sent (the segment carries an unacceptable ACK) ...
711 * Invalid ACK: reset will be sent by listening socket.
712 * Note that the ACK validity check for a Fast Open socket is done
713 * elsewhere and is checked directly against the child socket rather
714 * than req because user data may have been sent out.
716 if ((flg & TCP_FLAG_ACK) && !fastopen &&
717 (TCP_SKB_CB(skb)->ack_seq !=
718 tcp_rsk(req)->snt_isn + 1))
721 /* Also, it would be not so bad idea to check rcv_tsecr, which
722 * is essentially ACK extension and too early or too late values
723 * should cause reset in unsynchronized states.
726 /* RFC793: "first check sequence number". */
728 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
729 tcp_rsk(req)->rcv_nxt, tcp_rsk(req)->rcv_nxt + req->rsk_rcv_wnd)) {
730 /* Out of window: send ACK and drop. */
731 if (!(flg & TCP_FLAG_RST) &&
732 !tcp_oow_rate_limited(sock_net(sk), skb,
733 LINUX_MIB_TCPACKSKIPPEDSYNRECV,
734 &tcp_rsk(req)->last_oow_ack_time))
735 req->rsk_ops->send_ack(sk, skb, req);
737 __NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
741 /* In sequence, PAWS is OK. */
743 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_nxt))
744 req->ts_recent = tmp_opt.rcv_tsval;
746 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) {
747 /* Truncate SYN, it is out of window starting
748 at tcp_rsk(req)->rcv_isn + 1. */
749 flg &= ~TCP_FLAG_SYN;
752 /* RFC793: "second check the RST bit" and
753 * "fourth, check the SYN bit"
755 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) {
756 __TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
757 goto embryonic_reset;
760 /* ACK sequence verified above, just make sure ACK is
761 * set. If ACK not set, just silently drop the packet.
763 * XXX (TFO) - if we ever allow "data after SYN", the
764 * following check needs to be removed.
766 if (!(flg & TCP_FLAG_ACK))
769 /* For Fast Open no more processing is needed (sk is the
775 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */
776 if (req->num_timeout < inet_csk(sk)->icsk_accept_queue.rskq_defer_accept &&
777 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) {
778 inet_rsk(req)->acked = 1;
779 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP);
783 /* OK, ACK is valid, create big socket and
784 * feed this segment to it. It will repeat all
785 * the tests. THIS SEGMENT MUST MOVE SOCKET TO
786 * ESTABLISHED STATE. If it will be dropped after
787 * socket is created, wait for troubles.
789 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL,
792 goto listen_overflow;
794 sock_rps_save_rxhash(child, skb);
795 tcp_synack_rtt_meas(child, req);
796 *req_stolen = !own_req;
797 return inet_csk_complete_hashdance(sk, child, req, own_req);
800 if (!sock_net(sk)->ipv4.sysctl_tcp_abort_on_overflow) {
801 inet_rsk(req)->acked = 1;
806 if (!(flg & TCP_FLAG_RST)) {
807 /* Received a bad SYN pkt - for TFO We try not to reset
808 * the local connection unless it's really necessary to
809 * avoid becoming vulnerable to outside attack aiming at
810 * resetting legit local connections.
812 req->rsk_ops->send_reset(sk, skb);
813 } else if (fastopen) { /* received a valid RST pkt */
814 reqsk_fastopen_remove(sk, req, true);
818 bool unlinked = inet_csk_reqsk_queue_drop(sk, req);
821 __NET_INC_STATS(sock_net(sk), LINUX_MIB_EMBRYONICRSTS);
822 *req_stolen = !unlinked;
826 EXPORT_SYMBOL(tcp_check_req);
829 * Queue segment on the new socket if the new socket is active,
830 * otherwise we just shortcircuit this and continue with
833 * For the vast majority of cases child->sk_state will be TCP_SYN_RECV
834 * when entering. But other states are possible due to a race condition
835 * where after __inet_lookup_established() fails but before the listener
836 * locked is obtained, other packets cause the same connection to
840 int tcp_child_process(struct sock *parent, struct sock *child,
844 int state = child->sk_state;
846 /* record NAPI ID of child */
847 sk_mark_napi_id(child, skb);
849 tcp_segs_in(tcp_sk(child), skb);
850 if (!sock_owned_by_user(child)) {
851 ret = tcp_rcv_state_process(child, skb);
852 /* Wakeup parent, send SIGIO */
853 if (state == TCP_SYN_RECV && child->sk_state != state)
854 parent->sk_data_ready(parent);
856 /* Alas, it is possible again, because we do lookup
857 * in main socket hash table and lock on listening
858 * socket does not protect us more.
860 __sk_add_backlog(child, skb);
863 bh_unlock_sock(child);
867 EXPORT_SYMBOL(tcp_child_process);