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>
23 #include <linux/module.h>
24 #include <linux/slab.h>
25 #include <linux/sysctl.h>
26 #include <linux/workqueue.h>
27 #include <linux/static_key.h>
29 #include <net/inet_common.h>
31 #include <net/busy_poll.h>
33 static bool tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win)
37 if (after(end_seq, s_win) && before(seq, e_win))
39 return seq == e_win && seq == end_seq;
42 static enum tcp_tw_status
43 tcp_timewait_check_oow_rate_limit(struct inet_timewait_sock *tw,
44 const struct sk_buff *skb, int mib_idx)
46 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
48 if (!tcp_oow_rate_limited(twsk_net(tw), skb, mib_idx,
49 &tcptw->tw_last_oow_ack_time)) {
50 /* Send ACK. Note, we do not put the bucket,
51 * it will be released by caller.
56 /* We are rate-limiting, so just release the tw sock and drop skb. */
58 return TCP_TW_SUCCESS;
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(twsk_net(tw), skb, &tmp_opt, 0, NULL);
103 if (tmp_opt.saw_tstamp) {
104 if (tmp_opt.rcv_tsecr)
105 tmp_opt.rcv_tsecr -= tcptw->tw_ts_offset;
106 tmp_opt.ts_recent = tcptw->tw_ts_recent;
107 tmp_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp;
108 paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
112 if (tw->tw_substate == TCP_FIN_WAIT2) {
113 /* Just repeat all the checks of tcp_rcv_state_process() */
115 /* Out of window, send ACK */
117 !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
119 tcptw->tw_rcv_nxt + tcptw->tw_rcv_wnd))
120 return tcp_timewait_check_oow_rate_limit(
121 tw, skb, LINUX_MIB_TCPACKSKIPPEDFINWAIT2);
126 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt))
131 !after(TCP_SKB_CB(skb)->end_seq, tcptw->tw_rcv_nxt) ||
132 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) {
134 return TCP_TW_SUCCESS;
137 /* New data or FIN. If new data arrive after half-duplex close,
141 TCP_SKB_CB(skb)->end_seq != tcptw->tw_rcv_nxt + 1)
144 /* FIN arrived, enter true time-wait state. */
145 tw->tw_substate = TCP_TIME_WAIT;
146 tcptw->tw_rcv_nxt = TCP_SKB_CB(skb)->end_seq;
147 if (tmp_opt.saw_tstamp) {
148 tcptw->tw_ts_recent_stamp = ktime_get_seconds();
149 tcptw->tw_ts_recent = tmp_opt.rcv_tsval;
152 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
157 * Now real TIME-WAIT state.
160 * "When a connection is [...] on TIME-WAIT state [...]
161 * [a TCP] MAY accept a new SYN from the remote TCP to
162 * reopen the connection directly, if it:
164 * (1) assigns its initial sequence number for the new
165 * connection to be larger than the largest sequence
166 * number it used on the previous connection incarnation,
169 * (2) returns to TIME-WAIT state if the SYN turns out
170 * to be an old duplicate".
174 (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt &&
175 (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) {
176 /* In window segment, it may be only reset or bare ack. */
179 /* This is TIME_WAIT assassination, in two flavors.
180 * Oh well... nobody has a sufficient solution to this
183 if (twsk_net(tw)->ipv4.sysctl_tcp_rfc1337 == 0) {
185 inet_twsk_deschedule_put(tw);
186 return TCP_TW_SUCCESS;
189 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
192 if (tmp_opt.saw_tstamp) {
193 tcptw->tw_ts_recent = tmp_opt.rcv_tsval;
194 tcptw->tw_ts_recent_stamp = ktime_get_seconds();
198 return TCP_TW_SUCCESS;
201 /* Out of window segment.
203 All the segments are ACKed immediately.
205 The only exception is new SYN. We accept it, if it is
206 not old duplicate and we are not in danger to be killed
207 by delayed old duplicates. RFC check is that it has
208 newer sequence number works at rates <40Mbit/sec.
209 However, if paws works, it is reliable AND even more,
210 we even may relax silly seq space cutoff.
212 RED-PEN: we violate main RFC requirement, if this SYN will appear
213 old duplicate (i.e. we receive RST in reply to SYN-ACK),
214 we must return socket to time-wait state. It is not good,
218 if (th->syn && !th->rst && !th->ack && !paws_reject &&
219 (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) ||
220 (tmp_opt.saw_tstamp &&
221 (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) {
222 u32 isn = tcptw->tw_snd_nxt + 65535 + 2;
225 TCP_SKB_CB(skb)->tcp_tw_isn = isn;
230 __NET_INC_STATS(twsk_net(tw), LINUX_MIB_PAWSESTABREJECTED);
233 /* In this case we must reset the TIMEWAIT timer.
235 * If it is ACKless SYN it may be both old duplicate
236 * and new good SYN with random sequence number <rcv_nxt.
237 * Do not reschedule in the last case.
239 if (paws_reject || th->ack)
240 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
242 return tcp_timewait_check_oow_rate_limit(
243 tw, skb, LINUX_MIB_TCPACKSKIPPEDTIMEWAIT);
246 return TCP_TW_SUCCESS;
248 EXPORT_SYMBOL(tcp_timewait_state_process);
251 * Move a socket to time-wait or dead fin-wait-2 state.
253 void tcp_time_wait(struct sock *sk, int state, int timeo)
255 const struct inet_connection_sock *icsk = inet_csk(sk);
256 const struct tcp_sock *tp = tcp_sk(sk);
257 struct inet_timewait_sock *tw;
258 struct inet_timewait_death_row *tcp_death_row = &sock_net(sk)->ipv4.tcp_death_row;
260 tw = inet_twsk_alloc(sk, tcp_death_row, state);
263 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
264 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1);
265 struct inet_sock *inet = inet_sk(sk);
267 tw->tw_transparent = inet->transparent;
268 tw->tw_mark = sk->sk_mark;
269 tw->tw_rcv_wscale = tp->rx_opt.rcv_wscale;
270 tcptw->tw_rcv_nxt = tp->rcv_nxt;
271 tcptw->tw_snd_nxt = tp->snd_nxt;
272 tcptw->tw_rcv_wnd = tcp_receive_window(tp);
273 tcptw->tw_ts_recent = tp->rx_opt.ts_recent;
274 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp;
275 tcptw->tw_ts_offset = tp->tsoffset;
276 tcptw->tw_last_oow_ack_time = 0;
277 tcptw->tw_tx_delay = tp->tcp_tx_delay;
278 #if IS_ENABLED(CONFIG_IPV6)
279 if (tw->tw_family == PF_INET6) {
280 struct ipv6_pinfo *np = inet6_sk(sk);
282 tw->tw_v6_daddr = sk->sk_v6_daddr;
283 tw->tw_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
284 tw->tw_tclass = np->tclass;
285 tw->tw_flowlabel = be32_to_cpu(np->flow_label & IPV6_FLOWLABEL_MASK);
286 tw->tw_txhash = sk->sk_txhash;
287 tw->tw_ipv6only = sk->sk_ipv6only;
291 #ifdef CONFIG_TCP_MD5SIG
293 * The timewait bucket does not have the key DB from the
294 * sock structure. We just make a quick copy of the
295 * md5 key being used (if indeed we are using one)
296 * so the timewait ack generating code has the key.
299 tcptw->tw_md5_key = NULL;
300 if (static_branch_unlikely(&tcp_md5_needed)) {
301 struct tcp_md5sig_key *key;
303 key = tp->af_specific->md5_lookup(sk, sk);
305 tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC);
306 BUG_ON(tcptw->tw_md5_key && !tcp_alloc_md5sig_pool());
312 /* Get the TIME_WAIT timeout firing. */
316 if (state == TCP_TIME_WAIT)
317 timeo = TCP_TIMEWAIT_LEN;
319 /* tw_timer is pinned, so we need to make sure BH are disabled
320 * in following section, otherwise timer handler could run before
321 * we complete the initialization.
324 inet_twsk_schedule(tw, timeo);
326 * Note that access to tw after this point is illegal.
328 inet_twsk_hashdance(tw, sk, &tcp_hashinfo);
331 /* Sorry, if we're out of memory, just CLOSE this
332 * socket up. We've got bigger problems than
333 * non-graceful socket closings.
335 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEWAITOVERFLOW);
338 tcp_update_metrics(sk);
341 EXPORT_SYMBOL(tcp_time_wait);
343 void tcp_twsk_destructor(struct sock *sk)
345 #ifdef CONFIG_TCP_MD5SIG
346 if (static_branch_unlikely(&tcp_md5_needed)) {
347 struct tcp_timewait_sock *twsk = tcp_twsk(sk);
349 if (twsk->tw_md5_key)
350 kfree_rcu(twsk->tw_md5_key, rcu);
354 EXPORT_SYMBOL_GPL(tcp_twsk_destructor);
356 /* Warning : This function is called without sk_listener being locked.
357 * Be sure to read socket fields once, as their value could change under us.
359 void tcp_openreq_init_rwin(struct request_sock *req,
360 const struct sock *sk_listener,
361 const struct dst_entry *dst)
363 struct inet_request_sock *ireq = inet_rsk(req);
364 const struct tcp_sock *tp = tcp_sk(sk_listener);
365 int full_space = tcp_full_space(sk_listener);
371 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
372 window_clamp = READ_ONCE(tp->window_clamp);
373 /* Set this up on the first call only */
374 req->rsk_window_clamp = window_clamp ? : dst_metric(dst, RTAX_WINDOW);
376 /* limit the window selection if the user enforce a smaller rx buffer */
377 if (sk_listener->sk_userlocks & SOCK_RCVBUF_LOCK &&
378 (req->rsk_window_clamp > full_space || req->rsk_window_clamp == 0))
379 req->rsk_window_clamp = full_space;
381 rcv_wnd = tcp_rwnd_init_bpf((struct sock *)req);
383 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
384 else if (full_space < rcv_wnd * mss)
385 full_space = rcv_wnd * mss;
387 /* tcp_full_space because it is guaranteed to be the first packet */
388 tcp_select_initial_window(sk_listener, full_space,
389 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
391 &req->rsk_window_clamp,
395 ireq->rcv_wscale = rcv_wscale;
397 EXPORT_SYMBOL(tcp_openreq_init_rwin);
399 static void tcp_ecn_openreq_child(struct tcp_sock *tp,
400 const struct request_sock *req)
402 tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0;
405 void tcp_ca_openreq_child(struct sock *sk, const struct dst_entry *dst)
407 struct inet_connection_sock *icsk = inet_csk(sk);
408 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
409 bool ca_got_dst = false;
411 if (ca_key != TCP_CA_UNSPEC) {
412 const struct tcp_congestion_ops *ca;
415 ca = tcp_ca_find_key(ca_key);
416 if (likely(ca && try_module_get(ca->owner))) {
417 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
418 icsk->icsk_ca_ops = ca;
424 /* If no valid choice made yet, assign current system default ca. */
426 (!icsk->icsk_ca_setsockopt ||
427 !try_module_get(icsk->icsk_ca_ops->owner)))
428 tcp_assign_congestion_control(sk);
430 tcp_set_ca_state(sk, TCP_CA_Open);
432 EXPORT_SYMBOL_GPL(tcp_ca_openreq_child);
434 static void smc_check_reset_syn_req(struct tcp_sock *oldtp,
435 struct request_sock *req,
436 struct tcp_sock *newtp)
438 #if IS_ENABLED(CONFIG_SMC)
439 struct inet_request_sock *ireq;
441 if (static_branch_unlikely(&tcp_have_smc)) {
442 ireq = inet_rsk(req);
443 if (oldtp->syn_smc && !ireq->smc_ok)
449 /* This is not only more efficient than what we used to do, it eliminates
450 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM
452 * Actually, we could lots of memory writes here. tp of listening
453 * socket contains all necessary default parameters.
455 struct sock *tcp_create_openreq_child(const struct sock *sk,
456 struct request_sock *req,
459 struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC);
460 const struct inet_request_sock *ireq = inet_rsk(req);
461 struct tcp_request_sock *treq = tcp_rsk(req);
462 struct inet_connection_sock *newicsk;
463 struct tcp_sock *oldtp, *newtp;
468 newicsk = inet_csk(newsk);
469 newtp = tcp_sk(newsk);
472 smc_check_reset_syn_req(oldtp, req, newtp);
474 /* Now setup tcp_sock */
475 newtp->pred_flags = 0;
477 newtp->rcv_wup = newtp->copied_seq =
478 newtp->rcv_nxt = treq->rcv_isn + 1;
481 newtp->snd_sml = newtp->snd_una =
482 newtp->snd_nxt = newtp->snd_up = treq->snt_isn + 1;
484 INIT_LIST_HEAD(&newtp->tsq_node);
485 INIT_LIST_HEAD(&newtp->tsorted_sent_queue);
487 tcp_init_wl(newtp, treq->rcv_isn);
489 minmax_reset(&newtp->rtt_min, tcp_jiffies32, ~0U);
490 newicsk->icsk_ack.lrcvtime = tcp_jiffies32;
492 newtp->lsndtime = tcp_jiffies32;
493 newsk->sk_txhash = treq->txhash;
494 newtp->total_retrans = req->num_retrans;
496 tcp_init_xmit_timers(newsk);
497 newtp->write_seq = newtp->pushed_seq = treq->snt_isn + 1;
499 if (sock_flag(newsk, SOCK_KEEPOPEN))
500 inet_csk_reset_keepalive_timer(newsk,
501 keepalive_time_when(newtp));
503 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok;
504 newtp->rx_opt.sack_ok = ireq->sack_ok;
505 newtp->window_clamp = req->rsk_window_clamp;
506 newtp->rcv_ssthresh = req->rsk_rcv_wnd;
507 newtp->rcv_wnd = req->rsk_rcv_wnd;
508 newtp->rx_opt.wscale_ok = ireq->wscale_ok;
509 if (newtp->rx_opt.wscale_ok) {
510 newtp->rx_opt.snd_wscale = ireq->snd_wscale;
511 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale;
513 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0;
514 newtp->window_clamp = min(newtp->window_clamp, 65535U);
516 newtp->snd_wnd = ntohs(tcp_hdr(skb)->window) << newtp->rx_opt.snd_wscale;
517 newtp->max_window = newtp->snd_wnd;
519 if (newtp->rx_opt.tstamp_ok) {
520 newtp->rx_opt.ts_recent = req->ts_recent;
521 newtp->rx_opt.ts_recent_stamp = ktime_get_seconds();
522 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
524 newtp->rx_opt.ts_recent_stamp = 0;
525 newtp->tcp_header_len = sizeof(struct tcphdr);
527 if (req->num_timeout) {
528 newtp->undo_marker = treq->snt_isn;
529 newtp->retrans_stamp = div_u64(treq->snt_synack,
530 USEC_PER_SEC / TCP_TS_HZ);
532 newtp->tsoffset = treq->ts_off;
533 #ifdef CONFIG_TCP_MD5SIG
534 newtp->md5sig_info = NULL; /*XXX*/
535 if (newtp->af_specific->md5_lookup(sk, newsk))
536 newtp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
538 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len)
539 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len;
540 newtp->rx_opt.mss_clamp = req->mss;
541 tcp_ecn_openreq_child(newtp, req);
542 newtp->fastopen_req = NULL;
543 newtp->fastopen_rsk = NULL;
545 __TCP_INC_STATS(sock_net(sk), TCP_MIB_PASSIVEOPENS);
549 EXPORT_SYMBOL(tcp_create_openreq_child);
552 * Process an incoming packet for SYN_RECV sockets represented as a
553 * request_sock. Normally sk is the listener socket but for TFO it
554 * points to the child socket.
556 * XXX (TFO) - The current impl contains a special check for ack
557 * validation and inside tcp_v4_reqsk_send_ack(). Can we do better?
559 * We don't need to initialize tmp_opt.sack_ok as we don't use the results
562 struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb,
563 struct request_sock *req,
564 bool fastopen, bool *req_stolen)
566 struct tcp_options_received tmp_opt;
568 const struct tcphdr *th = tcp_hdr(skb);
569 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK);
570 bool paws_reject = false;
573 tmp_opt.saw_tstamp = 0;
574 if (th->doff > (sizeof(struct tcphdr)>>2)) {
575 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0, NULL);
577 if (tmp_opt.saw_tstamp) {
578 tmp_opt.ts_recent = req->ts_recent;
579 if (tmp_opt.rcv_tsecr)
580 tmp_opt.rcv_tsecr -= tcp_rsk(req)->ts_off;
581 /* We do not store true stamp, but it is not required,
582 * it can be estimated (approximately)
585 tmp_opt.ts_recent_stamp = ktime_get_seconds() - ((TCP_TIMEOUT_INIT/HZ)<<req->num_timeout);
586 paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
590 /* Check for pure retransmitted SYN. */
591 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn &&
592 flg == TCP_FLAG_SYN &&
595 * RFC793 draws (Incorrectly! It was fixed in RFC1122)
596 * this case on figure 6 and figure 8, but formal
597 * protocol description says NOTHING.
598 * To be more exact, it says that we should send ACK,
599 * because this segment (at least, if it has no data)
602 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT
603 * describe SYN-RECV state. All the description
604 * is wrong, we cannot believe to it and should
605 * rely only on common sense and implementation
608 * Enforce "SYN-ACK" according to figure 8, figure 6
609 * of RFC793, fixed by RFC1122.
611 * Note that even if there is new data in the SYN packet
612 * they will be thrown away too.
614 * Reset timer after retransmitting SYNACK, similar to
615 * the idea of fast retransmit in recovery.
617 if (!tcp_oow_rate_limited(sock_net(sk), skb,
618 LINUX_MIB_TCPACKSKIPPEDSYNRECV,
619 &tcp_rsk(req)->last_oow_ack_time) &&
621 !inet_rtx_syn_ack(sk, req)) {
622 unsigned long expires = jiffies;
624 expires += min(TCP_TIMEOUT_INIT << req->num_timeout,
627 mod_timer_pending(&req->rsk_timer, expires);
629 req->rsk_timer.expires = expires;
634 /* Further reproduces section "SEGMENT ARRIVES"
635 for state SYN-RECEIVED of RFC793.
636 It is broken, however, it does not work only
637 when SYNs are crossed.
639 You would think that SYN crossing is impossible here, since
640 we should have a SYN_SENT socket (from connect()) on our end,
641 but this is not true if the crossed SYNs were sent to both
642 ends by a malicious third party. We must defend against this,
643 and to do that we first verify the ACK (as per RFC793, page
644 36) and reset if it is invalid. Is this a true full defense?
645 To convince ourselves, let us consider a way in which the ACK
646 test can still pass in this 'malicious crossed SYNs' case.
647 Malicious sender sends identical SYNs (and thus identical sequence
648 numbers) to both A and B:
653 By our good fortune, both A and B select the same initial
654 send sequence number of seven :-)
656 A: sends SYN|ACK, seq=7, ack_seq=8
657 B: sends SYN|ACK, seq=7, ack_seq=8
659 So we are now A eating this SYN|ACK, ACK test passes. So
660 does sequence test, SYN is truncated, and thus we consider
663 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this
664 bare ACK. Otherwise, we create an established connection. Both
665 ends (listening sockets) accept the new incoming connection and try
666 to talk to each other. 8-)
668 Note: This case is both harmless, and rare. Possibility is about the
669 same as us discovering intelligent life on another plant tomorrow.
671 But generally, we should (RFC lies!) to accept ACK
672 from SYNACK both here and in tcp_rcv_state_process().
673 tcp_rcv_state_process() does not, hence, we do not too.
675 Note that the case is absolutely generic:
676 we cannot optimize anything here without
677 violating protocol. All the checks must be made
678 before attempt to create socket.
681 /* RFC793 page 36: "If the connection is in any non-synchronized state ...
682 * and the incoming segment acknowledges something not yet
683 * sent (the segment carries an unacceptable ACK) ...
686 * Invalid ACK: reset will be sent by listening socket.
687 * Note that the ACK validity check for a Fast Open socket is done
688 * elsewhere and is checked directly against the child socket rather
689 * than req because user data may have been sent out.
691 if ((flg & TCP_FLAG_ACK) && !fastopen &&
692 (TCP_SKB_CB(skb)->ack_seq !=
693 tcp_rsk(req)->snt_isn + 1))
696 /* Also, it would be not so bad idea to check rcv_tsecr, which
697 * is essentially ACK extension and too early or too late values
698 * should cause reset in unsynchronized states.
701 /* RFC793: "first check sequence number". */
703 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
704 tcp_rsk(req)->rcv_nxt, tcp_rsk(req)->rcv_nxt + req->rsk_rcv_wnd)) {
705 /* Out of window: send ACK and drop. */
706 if (!(flg & TCP_FLAG_RST) &&
707 !tcp_oow_rate_limited(sock_net(sk), skb,
708 LINUX_MIB_TCPACKSKIPPEDSYNRECV,
709 &tcp_rsk(req)->last_oow_ack_time))
710 req->rsk_ops->send_ack(sk, skb, req);
712 __NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
716 /* In sequence, PAWS is OK. */
718 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_nxt))
719 req->ts_recent = tmp_opt.rcv_tsval;
721 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) {
722 /* Truncate SYN, it is out of window starting
723 at tcp_rsk(req)->rcv_isn + 1. */
724 flg &= ~TCP_FLAG_SYN;
727 /* RFC793: "second check the RST bit" and
728 * "fourth, check the SYN bit"
730 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) {
731 __TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
732 goto embryonic_reset;
735 /* ACK sequence verified above, just make sure ACK is
736 * set. If ACK not set, just silently drop the packet.
738 * XXX (TFO) - if we ever allow "data after SYN", the
739 * following check needs to be removed.
741 if (!(flg & TCP_FLAG_ACK))
744 /* For Fast Open no more processing is needed (sk is the
750 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */
751 if (req->num_timeout < inet_csk(sk)->icsk_accept_queue.rskq_defer_accept &&
752 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) {
753 inet_rsk(req)->acked = 1;
754 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP);
758 /* OK, ACK is valid, create big socket and
759 * feed this segment to it. It will repeat all
760 * the tests. THIS SEGMENT MUST MOVE SOCKET TO
761 * ESTABLISHED STATE. If it will be dropped after
762 * socket is created, wait for troubles.
764 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL,
767 goto listen_overflow;
769 sock_rps_save_rxhash(child, skb);
770 tcp_synack_rtt_meas(child, req);
771 *req_stolen = !own_req;
772 return inet_csk_complete_hashdance(sk, child, req, own_req);
775 if (!sock_net(sk)->ipv4.sysctl_tcp_abort_on_overflow) {
776 inet_rsk(req)->acked = 1;
781 if (!(flg & TCP_FLAG_RST)) {
782 /* Received a bad SYN pkt - for TFO We try not to reset
783 * the local connection unless it's really necessary to
784 * avoid becoming vulnerable to outside attack aiming at
785 * resetting legit local connections.
787 req->rsk_ops->send_reset(sk, skb);
788 } else if (fastopen) { /* received a valid RST pkt */
789 reqsk_fastopen_remove(sk, req, true);
793 inet_csk_reqsk_queue_drop(sk, req);
794 __NET_INC_STATS(sock_net(sk), LINUX_MIB_EMBRYONICRSTS);
798 EXPORT_SYMBOL(tcp_check_req);
801 * Queue segment on the new socket if the new socket is active,
802 * otherwise we just shortcircuit this and continue with
805 * For the vast majority of cases child->sk_state will be TCP_SYN_RECV
806 * when entering. But other states are possible due to a race condition
807 * where after __inet_lookup_established() fails but before the listener
808 * locked is obtained, other packets cause the same connection to
812 int tcp_child_process(struct sock *parent, struct sock *child,
816 int state = child->sk_state;
818 /* record NAPI ID of child */
819 sk_mark_napi_id(child, skb);
821 tcp_segs_in(tcp_sk(child), skb);
822 if (!sock_owned_by_user(child)) {
823 ret = tcp_rcv_state_process(child, skb);
824 /* Wakeup parent, send SIGIO */
825 if (state == TCP_SYN_RECV && child->sk_state != state)
826 parent->sk_data_ready(parent);
828 /* Alas, it is possible again, because we do lookup
829 * in main socket hash table and lock on listening
830 * socket does not protect us more.
832 __sk_add_backlog(child, skb);
835 bh_unlock_sock(child);
839 EXPORT_SYMBOL(tcp_child_process);