1 // SPDX-License-Identifier: GPL-2.0-only
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * Implementation of the Transmission Control Protocol(TCP).
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
24 #include <net/busy_poll.h>
26 static bool tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win)
30 if (after(end_seq, s_win) && before(seq, e_win))
32 return seq == e_win && seq == end_seq;
35 static enum tcp_tw_status
36 tcp_timewait_check_oow_rate_limit(struct inet_timewait_sock *tw,
37 const struct sk_buff *skb, int mib_idx)
39 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
41 if (!tcp_oow_rate_limited(twsk_net(tw), skb, mib_idx,
42 &tcptw->tw_last_oow_ack_time)) {
43 /* Send ACK. Note, we do not put the bucket,
44 * it will be released by caller.
49 /* We are rate-limiting, so just release the tw sock and drop skb. */
51 return TCP_TW_SUCCESS;
55 * * Main purpose of TIME-WAIT state is to close connection gracefully,
56 * when one of ends sits in LAST-ACK or CLOSING retransmitting FIN
57 * (and, probably, tail of data) and one or more our ACKs are lost.
58 * * What is TIME-WAIT timeout? It is associated with maximal packet
59 * lifetime in the internet, which results in wrong conclusion, that
60 * it is set to catch "old duplicate segments" wandering out of their path.
61 * It is not quite correct. This timeout is calculated so that it exceeds
62 * maximal retransmission timeout enough to allow to lose one (or more)
63 * segments sent by peer and our ACKs. This time may be calculated from RTO.
64 * * When TIME-WAIT socket receives RST, it means that another end
65 * finally closed and we are allowed to kill TIME-WAIT too.
66 * * Second purpose of TIME-WAIT is catching old duplicate segments.
67 * Well, certainly it is pure paranoia, but if we load TIME-WAIT
68 * with this semantics, we MUST NOT kill TIME-WAIT state with RSTs.
69 * * If we invented some more clever way to catch duplicates
70 * (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs.
72 * The algorithm below is based on FORMAL INTERPRETATION of RFCs.
73 * When you compare it to RFCs, please, read section SEGMENT ARRIVES
74 * from the very beginning.
76 * NOTE. With recycling (and later with fin-wait-2) TW bucket
77 * is _not_ stateless. It means, that strictly speaking we must
78 * spinlock it. I do not want! Well, probability of misbehaviour
79 * is ridiculously low and, seems, we could use some mb() tricks
80 * to avoid misread sequence numbers, states etc. --ANK
82 * We don't need to initialize tmp_out.sack_ok as we don't use the results
85 tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb,
86 const struct tcphdr *th)
88 struct tcp_options_received tmp_opt;
89 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
90 bool paws_reject = false;
92 tmp_opt.saw_tstamp = 0;
93 if (th->doff > (sizeof(*th) >> 2) && tcptw->tw_ts_recent_stamp) {
94 tcp_parse_options(twsk_net(tw), skb, &tmp_opt, 0, NULL);
96 if (tmp_opt.saw_tstamp) {
97 if (tmp_opt.rcv_tsecr)
98 tmp_opt.rcv_tsecr -= tcptw->tw_ts_offset;
99 tmp_opt.ts_recent = tcptw->tw_ts_recent;
100 tmp_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp;
101 paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
105 if (tw->tw_substate == TCP_FIN_WAIT2) {
106 /* Just repeat all the checks of tcp_rcv_state_process() */
108 /* Out of window, send ACK */
110 !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
112 tcptw->tw_rcv_nxt + tcptw->tw_rcv_wnd))
113 return tcp_timewait_check_oow_rate_limit(
114 tw, skb, LINUX_MIB_TCPACKSKIPPEDFINWAIT2);
119 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt))
124 !after(TCP_SKB_CB(skb)->end_seq, tcptw->tw_rcv_nxt) ||
125 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) {
127 return TCP_TW_SUCCESS;
130 /* New data or FIN. If new data arrive after half-duplex close,
134 TCP_SKB_CB(skb)->end_seq != tcptw->tw_rcv_nxt + 1)
137 /* FIN arrived, enter true time-wait state. */
138 tw->tw_substate = TCP_TIME_WAIT;
139 tcptw->tw_rcv_nxt = TCP_SKB_CB(skb)->end_seq;
140 if (tmp_opt.saw_tstamp) {
141 tcptw->tw_ts_recent_stamp = ktime_get_seconds();
142 tcptw->tw_ts_recent = tmp_opt.rcv_tsval;
145 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
150 * Now real TIME-WAIT state.
153 * "When a connection is [...] on TIME-WAIT state [...]
154 * [a TCP] MAY accept a new SYN from the remote TCP to
155 * reopen the connection directly, if it:
157 * (1) assigns its initial sequence number for the new
158 * connection to be larger than the largest sequence
159 * number it used on the previous connection incarnation,
162 * (2) returns to TIME-WAIT state if the SYN turns out
163 * to be an old duplicate".
167 (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt &&
168 (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) {
169 /* In window segment, it may be only reset or bare ack. */
172 /* This is TIME_WAIT assassination, in two flavors.
173 * Oh well... nobody has a sufficient solution to this
176 if (!READ_ONCE(twsk_net(tw)->ipv4.sysctl_tcp_rfc1337)) {
178 inet_twsk_deschedule_put(tw);
179 return TCP_TW_SUCCESS;
182 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
185 if (tmp_opt.saw_tstamp) {
186 tcptw->tw_ts_recent = tmp_opt.rcv_tsval;
187 tcptw->tw_ts_recent_stamp = ktime_get_seconds();
191 return TCP_TW_SUCCESS;
194 /* Out of window segment.
196 All the segments are ACKed immediately.
198 The only exception is new SYN. We accept it, if it is
199 not old duplicate and we are not in danger to be killed
200 by delayed old duplicates. RFC check is that it has
201 newer sequence number works at rates <40Mbit/sec.
202 However, if paws works, it is reliable AND even more,
203 we even may relax silly seq space cutoff.
205 RED-PEN: we violate main RFC requirement, if this SYN will appear
206 old duplicate (i.e. we receive RST in reply to SYN-ACK),
207 we must return socket to time-wait state. It is not good,
211 if (th->syn && !th->rst && !th->ack && !paws_reject &&
212 (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) ||
213 (tmp_opt.saw_tstamp &&
214 (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) {
215 u32 isn = tcptw->tw_snd_nxt + 65535 + 2;
218 TCP_SKB_CB(skb)->tcp_tw_isn = isn;
223 __NET_INC_STATS(twsk_net(tw), LINUX_MIB_PAWSESTABREJECTED);
226 /* In this case we must reset the TIMEWAIT timer.
228 * If it is ACKless SYN it may be both old duplicate
229 * and new good SYN with random sequence number <rcv_nxt.
230 * Do not reschedule in the last case.
232 if (paws_reject || th->ack)
233 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
235 return tcp_timewait_check_oow_rate_limit(
236 tw, skb, LINUX_MIB_TCPACKSKIPPEDTIMEWAIT);
239 return TCP_TW_SUCCESS;
241 EXPORT_SYMBOL(tcp_timewait_state_process);
243 static void tcp_time_wait_init(struct sock *sk, struct tcp_timewait_sock *tcptw)
245 #ifdef CONFIG_TCP_MD5SIG
246 const struct tcp_sock *tp = tcp_sk(sk);
247 struct tcp_md5sig_key *key;
250 * The timewait bucket does not have the key DB from the
251 * sock structure. We just make a quick copy of the
252 * md5 key being used (if indeed we are using one)
253 * so the timewait ack generating code has the key.
255 tcptw->tw_md5_key = NULL;
256 if (!static_branch_unlikely(&tcp_md5_needed.key))
259 key = tp->af_specific->md5_lookup(sk, sk);
261 tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC);
262 if (!tcptw->tw_md5_key)
264 if (!tcp_alloc_md5sig_pool())
266 if (!static_key_fast_inc_not_disabled(&tcp_md5_needed.key.key))
272 kfree(tcptw->tw_md5_key);
273 tcptw->tw_md5_key = NULL;
278 * Move a socket to time-wait or dead fin-wait-2 state.
280 void tcp_time_wait(struct sock *sk, int state, int timeo)
282 const struct inet_connection_sock *icsk = inet_csk(sk);
283 const struct tcp_sock *tp = tcp_sk(sk);
284 struct net *net = sock_net(sk);
285 struct inet_timewait_sock *tw;
287 tw = inet_twsk_alloc(sk, &net->ipv4.tcp_death_row, state);
290 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
291 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1);
292 struct inet_sock *inet = inet_sk(sk);
294 tw->tw_transparent = inet->transparent;
295 tw->tw_mark = sk->sk_mark;
296 tw->tw_priority = sk->sk_priority;
297 tw->tw_rcv_wscale = tp->rx_opt.rcv_wscale;
298 tcptw->tw_rcv_nxt = tp->rcv_nxt;
299 tcptw->tw_snd_nxt = tp->snd_nxt;
300 tcptw->tw_rcv_wnd = tcp_receive_window(tp);
301 tcptw->tw_ts_recent = tp->rx_opt.ts_recent;
302 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp;
303 tcptw->tw_ts_offset = tp->tsoffset;
304 tcptw->tw_last_oow_ack_time = 0;
305 tcptw->tw_tx_delay = tp->tcp_tx_delay;
306 tw->tw_txhash = sk->sk_txhash;
307 #if IS_ENABLED(CONFIG_IPV6)
308 if (tw->tw_family == PF_INET6) {
309 struct ipv6_pinfo *np = inet6_sk(sk);
311 tw->tw_v6_daddr = sk->sk_v6_daddr;
312 tw->tw_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
313 tw->tw_tclass = np->tclass;
314 tw->tw_flowlabel = be32_to_cpu(np->flow_label & IPV6_FLOWLABEL_MASK);
315 tw->tw_ipv6only = sk->sk_ipv6only;
319 tcp_time_wait_init(sk, tcptw);
321 /* Get the TIME_WAIT timeout firing. */
325 if (state == TCP_TIME_WAIT)
326 timeo = TCP_TIMEWAIT_LEN;
328 /* tw_timer is pinned, so we need to make sure BH are disabled
329 * in following section, otherwise timer handler could run before
330 * we complete the initialization.
333 inet_twsk_schedule(tw, timeo);
335 * Note that access to tw after this point is illegal.
337 inet_twsk_hashdance(tw, sk, net->ipv4.tcp_death_row.hashinfo);
340 /* Sorry, if we're out of memory, just CLOSE this
341 * socket up. We've got bigger problems than
342 * non-graceful socket closings.
344 NET_INC_STATS(net, LINUX_MIB_TCPTIMEWAITOVERFLOW);
347 tcp_update_metrics(sk);
350 EXPORT_SYMBOL(tcp_time_wait);
352 void tcp_twsk_destructor(struct sock *sk)
354 #ifdef CONFIG_TCP_MD5SIG
355 if (static_branch_unlikely(&tcp_md5_needed.key)) {
356 struct tcp_timewait_sock *twsk = tcp_twsk(sk);
358 if (twsk->tw_md5_key) {
359 kfree_rcu(twsk->tw_md5_key, rcu);
360 static_branch_slow_dec_deferred(&tcp_md5_needed);
365 EXPORT_SYMBOL_GPL(tcp_twsk_destructor);
367 void tcp_twsk_purge(struct list_head *net_exit_list, int family)
369 bool purged_once = false;
372 list_for_each_entry(net, net_exit_list, exit_list) {
373 if (net->ipv4.tcp_death_row.hashinfo->pernet) {
374 /* Even if tw_refcount == 1, we must clean up kernel reqsk */
375 inet_twsk_purge(net->ipv4.tcp_death_row.hashinfo, family);
376 } else if (!purged_once) {
377 /* The last refcount is decremented in tcp_sk_exit_batch() */
378 if (refcount_read(&net->ipv4.tcp_death_row.tw_refcount) == 1)
381 inet_twsk_purge(&tcp_hashinfo, family);
386 EXPORT_SYMBOL_GPL(tcp_twsk_purge);
388 /* Warning : This function is called without sk_listener being locked.
389 * Be sure to read socket fields once, as their value could change under us.
391 void tcp_openreq_init_rwin(struct request_sock *req,
392 const struct sock *sk_listener,
393 const struct dst_entry *dst)
395 struct inet_request_sock *ireq = inet_rsk(req);
396 const struct tcp_sock *tp = tcp_sk(sk_listener);
397 int full_space = tcp_full_space(sk_listener);
403 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
404 window_clamp = READ_ONCE(tp->window_clamp);
405 /* Set this up on the first call only */
406 req->rsk_window_clamp = window_clamp ? : dst_metric(dst, RTAX_WINDOW);
408 /* limit the window selection if the user enforce a smaller rx buffer */
409 if (sk_listener->sk_userlocks & SOCK_RCVBUF_LOCK &&
410 (req->rsk_window_clamp > full_space || req->rsk_window_clamp == 0))
411 req->rsk_window_clamp = full_space;
413 rcv_wnd = tcp_rwnd_init_bpf((struct sock *)req);
415 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
416 else if (full_space < rcv_wnd * mss)
417 full_space = rcv_wnd * mss;
419 /* tcp_full_space because it is guaranteed to be the first packet */
420 tcp_select_initial_window(sk_listener, full_space,
421 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
423 &req->rsk_window_clamp,
427 ireq->rcv_wscale = rcv_wscale;
429 EXPORT_SYMBOL(tcp_openreq_init_rwin);
431 static void tcp_ecn_openreq_child(struct tcp_sock *tp,
432 const struct request_sock *req)
434 tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0;
437 void tcp_ca_openreq_child(struct sock *sk, const struct dst_entry *dst)
439 struct inet_connection_sock *icsk = inet_csk(sk);
440 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
441 bool ca_got_dst = false;
443 if (ca_key != TCP_CA_UNSPEC) {
444 const struct tcp_congestion_ops *ca;
447 ca = tcp_ca_find_key(ca_key);
448 if (likely(ca && bpf_try_module_get(ca, ca->owner))) {
449 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
450 icsk->icsk_ca_ops = ca;
456 /* If no valid choice made yet, assign current system default ca. */
458 (!icsk->icsk_ca_setsockopt ||
459 !bpf_try_module_get(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner)))
460 tcp_assign_congestion_control(sk);
462 tcp_set_ca_state(sk, TCP_CA_Open);
464 EXPORT_SYMBOL_GPL(tcp_ca_openreq_child);
466 static void smc_check_reset_syn_req(const struct tcp_sock *oldtp,
467 struct request_sock *req,
468 struct tcp_sock *newtp)
470 #if IS_ENABLED(CONFIG_SMC)
471 struct inet_request_sock *ireq;
473 if (static_branch_unlikely(&tcp_have_smc)) {
474 ireq = inet_rsk(req);
475 if (oldtp->syn_smc && !ireq->smc_ok)
481 /* This is not only more efficient than what we used to do, it eliminates
482 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM
484 * Actually, we could lots of memory writes here. tp of listening
485 * socket contains all necessary default parameters.
487 struct sock *tcp_create_openreq_child(const struct sock *sk,
488 struct request_sock *req,
491 struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC);
492 const struct inet_request_sock *ireq = inet_rsk(req);
493 struct tcp_request_sock *treq = tcp_rsk(req);
494 struct inet_connection_sock *newicsk;
495 const struct tcp_sock *oldtp;
496 struct tcp_sock *newtp;
502 newicsk = inet_csk(newsk);
503 newtp = tcp_sk(newsk);
506 smc_check_reset_syn_req(oldtp, req, newtp);
508 /* Now setup tcp_sock */
509 newtp->pred_flags = 0;
511 seq = treq->rcv_isn + 1;
512 newtp->rcv_wup = seq;
513 WRITE_ONCE(newtp->copied_seq, seq);
514 WRITE_ONCE(newtp->rcv_nxt, seq);
517 seq = treq->snt_isn + 1;
518 newtp->snd_sml = newtp->snd_una = seq;
519 WRITE_ONCE(newtp->snd_nxt, seq);
522 INIT_LIST_HEAD(&newtp->tsq_node);
523 INIT_LIST_HEAD(&newtp->tsorted_sent_queue);
525 tcp_init_wl(newtp, treq->rcv_isn);
527 minmax_reset(&newtp->rtt_min, tcp_jiffies32, ~0U);
528 newicsk->icsk_ack.lrcvtime = tcp_jiffies32;
530 newtp->lsndtime = tcp_jiffies32;
531 newsk->sk_txhash = READ_ONCE(treq->txhash);
532 newtp->total_retrans = req->num_retrans;
534 tcp_init_xmit_timers(newsk);
535 WRITE_ONCE(newtp->write_seq, newtp->pushed_seq = treq->snt_isn + 1);
537 if (sock_flag(newsk, SOCK_KEEPOPEN))
538 inet_csk_reset_keepalive_timer(newsk,
539 keepalive_time_when(newtp));
541 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok;
542 newtp->rx_opt.sack_ok = ireq->sack_ok;
543 newtp->window_clamp = req->rsk_window_clamp;
544 newtp->rcv_ssthresh = req->rsk_rcv_wnd;
545 newtp->rcv_wnd = req->rsk_rcv_wnd;
546 newtp->rx_opt.wscale_ok = ireq->wscale_ok;
547 if (newtp->rx_opt.wscale_ok) {
548 newtp->rx_opt.snd_wscale = ireq->snd_wscale;
549 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale;
551 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0;
552 newtp->window_clamp = min(newtp->window_clamp, 65535U);
554 newtp->snd_wnd = ntohs(tcp_hdr(skb)->window) << newtp->rx_opt.snd_wscale;
555 newtp->max_window = newtp->snd_wnd;
557 if (newtp->rx_opt.tstamp_ok) {
558 newtp->rx_opt.ts_recent = req->ts_recent;
559 newtp->rx_opt.ts_recent_stamp = ktime_get_seconds();
560 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
562 newtp->rx_opt.ts_recent_stamp = 0;
563 newtp->tcp_header_len = sizeof(struct tcphdr);
565 if (req->num_timeout) {
566 newtp->undo_marker = treq->snt_isn;
567 newtp->retrans_stamp = div_u64(treq->snt_synack,
568 USEC_PER_SEC / TCP_TS_HZ);
570 newtp->tsoffset = treq->ts_off;
571 #ifdef CONFIG_TCP_MD5SIG
572 newtp->md5sig_info = NULL; /*XXX*/
573 if (treq->af_specific->req_md5_lookup(sk, req_to_sk(req)))
574 newtp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
576 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len)
577 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len;
578 newtp->rx_opt.mss_clamp = req->mss;
579 tcp_ecn_openreq_child(newtp, req);
580 newtp->fastopen_req = NULL;
581 RCU_INIT_POINTER(newtp->fastopen_rsk, NULL);
583 newtp->bpf_chg_cc_inprogress = 0;
584 tcp_bpf_clone(sk, newsk);
586 __TCP_INC_STATS(sock_net(sk), TCP_MIB_PASSIVEOPENS);
590 EXPORT_SYMBOL(tcp_create_openreq_child);
593 * Process an incoming packet for SYN_RECV sockets represented as a
594 * request_sock. Normally sk is the listener socket but for TFO it
595 * points to the child socket.
597 * XXX (TFO) - The current impl contains a special check for ack
598 * validation and inside tcp_v4_reqsk_send_ack(). Can we do better?
600 * We don't need to initialize tmp_opt.sack_ok as we don't use the results
602 * Note: If @fastopen is true, this can be called from process context.
603 * Otherwise, this is from BH context.
606 struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb,
607 struct request_sock *req,
608 bool fastopen, bool *req_stolen)
610 struct tcp_options_received tmp_opt;
612 const struct tcphdr *th = tcp_hdr(skb);
613 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK);
614 bool paws_reject = false;
617 tmp_opt.saw_tstamp = 0;
618 if (th->doff > (sizeof(struct tcphdr)>>2)) {
619 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0, NULL);
621 if (tmp_opt.saw_tstamp) {
622 tmp_opt.ts_recent = req->ts_recent;
623 if (tmp_opt.rcv_tsecr)
624 tmp_opt.rcv_tsecr -= tcp_rsk(req)->ts_off;
625 /* We do not store true stamp, but it is not required,
626 * it can be estimated (approximately)
629 tmp_opt.ts_recent_stamp = ktime_get_seconds() - reqsk_timeout(req, TCP_RTO_MAX) / HZ;
630 paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
634 /* Check for pure retransmitted SYN. */
635 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn &&
636 flg == TCP_FLAG_SYN &&
639 * RFC793 draws (Incorrectly! It was fixed in RFC1122)
640 * this case on figure 6 and figure 8, but formal
641 * protocol description says NOTHING.
642 * To be more exact, it says that we should send ACK,
643 * because this segment (at least, if it has no data)
646 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT
647 * describe SYN-RECV state. All the description
648 * is wrong, we cannot believe to it and should
649 * rely only on common sense and implementation
652 * Enforce "SYN-ACK" according to figure 8, figure 6
653 * of RFC793, fixed by RFC1122.
655 * Note that even if there is new data in the SYN packet
656 * they will be thrown away too.
658 * Reset timer after retransmitting SYNACK, similar to
659 * the idea of fast retransmit in recovery.
661 if (!tcp_oow_rate_limited(sock_net(sk), skb,
662 LINUX_MIB_TCPACKSKIPPEDSYNRECV,
663 &tcp_rsk(req)->last_oow_ack_time) &&
665 !inet_rtx_syn_ack(sk, req)) {
666 unsigned long expires = jiffies;
668 expires += reqsk_timeout(req, TCP_RTO_MAX);
670 mod_timer_pending(&req->rsk_timer, expires);
672 req->rsk_timer.expires = expires;
677 /* Further reproduces section "SEGMENT ARRIVES"
678 for state SYN-RECEIVED of RFC793.
679 It is broken, however, it does not work only
680 when SYNs are crossed.
682 You would think that SYN crossing is impossible here, since
683 we should have a SYN_SENT socket (from connect()) on our end,
684 but this is not true if the crossed SYNs were sent to both
685 ends by a malicious third party. We must defend against this,
686 and to do that we first verify the ACK (as per RFC793, page
687 36) and reset if it is invalid. Is this a true full defense?
688 To convince ourselves, let us consider a way in which the ACK
689 test can still pass in this 'malicious crossed SYNs' case.
690 Malicious sender sends identical SYNs (and thus identical sequence
691 numbers) to both A and B:
696 By our good fortune, both A and B select the same initial
697 send sequence number of seven :-)
699 A: sends SYN|ACK, seq=7, ack_seq=8
700 B: sends SYN|ACK, seq=7, ack_seq=8
702 So we are now A eating this SYN|ACK, ACK test passes. So
703 does sequence test, SYN is truncated, and thus we consider
706 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this
707 bare ACK. Otherwise, we create an established connection. Both
708 ends (listening sockets) accept the new incoming connection and try
709 to talk to each other. 8-)
711 Note: This case is both harmless, and rare. Possibility is about the
712 same as us discovering intelligent life on another plant tomorrow.
714 But generally, we should (RFC lies!) to accept ACK
715 from SYNACK both here and in tcp_rcv_state_process().
716 tcp_rcv_state_process() does not, hence, we do not too.
718 Note that the case is absolutely generic:
719 we cannot optimize anything here without
720 violating protocol. All the checks must be made
721 before attempt to create socket.
724 /* RFC793 page 36: "If the connection is in any non-synchronized state ...
725 * and the incoming segment acknowledges something not yet
726 * sent (the segment carries an unacceptable ACK) ...
729 * Invalid ACK: reset will be sent by listening socket.
730 * Note that the ACK validity check for a Fast Open socket is done
731 * elsewhere and is checked directly against the child socket rather
732 * than req because user data may have been sent out.
734 if ((flg & TCP_FLAG_ACK) && !fastopen &&
735 (TCP_SKB_CB(skb)->ack_seq !=
736 tcp_rsk(req)->snt_isn + 1))
739 /* Also, it would be not so bad idea to check rcv_tsecr, which
740 * is essentially ACK extension and too early or too late values
741 * should cause reset in unsynchronized states.
744 /* RFC793: "first check sequence number". */
746 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
747 tcp_rsk(req)->rcv_nxt, tcp_rsk(req)->rcv_nxt + req->rsk_rcv_wnd)) {
748 /* Out of window: send ACK and drop. */
749 if (!(flg & TCP_FLAG_RST) &&
750 !tcp_oow_rate_limited(sock_net(sk), skb,
751 LINUX_MIB_TCPACKSKIPPEDSYNRECV,
752 &tcp_rsk(req)->last_oow_ack_time))
753 req->rsk_ops->send_ack(sk, skb, req);
755 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
759 /* In sequence, PAWS is OK. */
761 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_nxt))
762 req->ts_recent = tmp_opt.rcv_tsval;
764 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) {
765 /* Truncate SYN, it is out of window starting
766 at tcp_rsk(req)->rcv_isn + 1. */
767 flg &= ~TCP_FLAG_SYN;
770 /* RFC793: "second check the RST bit" and
771 * "fourth, check the SYN bit"
773 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) {
774 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
775 goto embryonic_reset;
778 /* ACK sequence verified above, just make sure ACK is
779 * set. If ACK not set, just silently drop the packet.
781 * XXX (TFO) - if we ever allow "data after SYN", the
782 * following check needs to be removed.
784 if (!(flg & TCP_FLAG_ACK))
787 /* For Fast Open no more processing is needed (sk is the
793 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */
794 if (req->num_timeout < inet_csk(sk)->icsk_accept_queue.rskq_defer_accept &&
795 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) {
796 inet_rsk(req)->acked = 1;
797 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP);
801 /* OK, ACK is valid, create big socket and
802 * feed this segment to it. It will repeat all
803 * the tests. THIS SEGMENT MUST MOVE SOCKET TO
804 * ESTABLISHED STATE. If it will be dropped after
805 * socket is created, wait for troubles.
807 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL,
810 goto listen_overflow;
812 if (own_req && rsk_drop_req(req)) {
813 reqsk_queue_removed(&inet_csk(req->rsk_listener)->icsk_accept_queue, req);
814 inet_csk_reqsk_queue_drop_and_put(req->rsk_listener, req);
818 sock_rps_save_rxhash(child, skb);
819 tcp_synack_rtt_meas(child, req);
820 *req_stolen = !own_req;
821 return inet_csk_complete_hashdance(sk, child, req, own_req);
824 if (sk != req->rsk_listener)
825 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMIGRATEREQFAILURE);
827 if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_abort_on_overflow)) {
828 inet_rsk(req)->acked = 1;
833 if (!(flg & TCP_FLAG_RST)) {
834 /* Received a bad SYN pkt - for TFO We try not to reset
835 * the local connection unless it's really necessary to
836 * avoid becoming vulnerable to outside attack aiming at
837 * resetting legit local connections.
839 req->rsk_ops->send_reset(sk, skb);
840 } else if (fastopen) { /* received a valid RST pkt */
841 reqsk_fastopen_remove(sk, req, true);
845 bool unlinked = inet_csk_reqsk_queue_drop(sk, req);
848 __NET_INC_STATS(sock_net(sk), LINUX_MIB_EMBRYONICRSTS);
849 *req_stolen = !unlinked;
853 EXPORT_SYMBOL(tcp_check_req);
856 * Queue segment on the new socket if the new socket is active,
857 * otherwise we just shortcircuit this and continue with
860 * For the vast majority of cases child->sk_state will be TCP_SYN_RECV
861 * when entering. But other states are possible due to a race condition
862 * where after __inet_lookup_established() fails but before the listener
863 * locked is obtained, other packets cause the same connection to
867 int tcp_child_process(struct sock *parent, struct sock *child,
869 __releases(&((child)->sk_lock.slock))
872 int state = child->sk_state;
874 /* record sk_napi_id and sk_rx_queue_mapping of child. */
875 sk_mark_napi_id_set(child, skb);
877 tcp_segs_in(tcp_sk(child), skb);
878 if (!sock_owned_by_user(child)) {
879 ret = tcp_rcv_state_process(child, skb);
880 /* Wakeup parent, send SIGIO */
881 if (state == TCP_SYN_RECV && child->sk_state != state)
882 parent->sk_data_ready(parent);
884 /* Alas, it is possible again, because we do lookup
885 * in main socket hash table and lock on listening
886 * socket does not protect us more.
888 __sk_add_backlog(child, skb);
891 bh_unlock_sock(child);
895 EXPORT_SYMBOL(tcp_child_process);