2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
22 #include <linux/module.h>
23 #include <linux/slab.h>
24 #include <linux/sysctl.h>
25 #include <linux/workqueue.h>
27 #include <net/inet_common.h>
30 int sysctl_tcp_syncookies __read_mostly = 1;
31 EXPORT_SYMBOL(sysctl_tcp_syncookies);
33 int sysctl_tcp_abort_on_overflow __read_mostly;
35 struct inet_timewait_death_row tcp_death_row = {
36 .sysctl_max_tw_buckets = NR_FILE * 2,
37 .period = TCP_TIMEWAIT_LEN / INET_TWDR_TWKILL_SLOTS,
38 .death_lock = __SPIN_LOCK_UNLOCKED(tcp_death_row.death_lock),
39 .hashinfo = &tcp_hashinfo,
40 .tw_timer = TIMER_INITIALIZER(inet_twdr_hangman, 0,
41 (unsigned long)&tcp_death_row),
42 .twkill_work = __WORK_INITIALIZER(tcp_death_row.twkill_work,
43 inet_twdr_twkill_work),
44 /* Short-time timewait calendar */
47 .twcal_timer = TIMER_INITIALIZER(inet_twdr_twcal_tick, 0,
48 (unsigned long)&tcp_death_row),
50 EXPORT_SYMBOL_GPL(tcp_death_row);
52 static bool tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win)
56 if (after(end_seq, s_win) && before(seq, e_win))
58 return seq == e_win && seq == end_seq;
61 static enum tcp_tw_status
62 tcp_timewait_check_oow_rate_limit(struct inet_timewait_sock *tw,
63 const struct sk_buff *skb, int mib_idx)
65 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
67 if (!tcp_oow_rate_limited(twsk_net(tw), skb, mib_idx,
68 &tcptw->tw_last_oow_ack_time)) {
69 /* Send ACK. Note, we do not put the bucket,
70 * it will be released by caller.
75 /* We are rate-limiting, so just release the tw sock and drop skb. */
77 return TCP_TW_SUCCESS;
81 * * Main purpose of TIME-WAIT state is to close connection gracefully,
82 * when one of ends sits in LAST-ACK or CLOSING retransmitting FIN
83 * (and, probably, tail of data) and one or more our ACKs are lost.
84 * * What is TIME-WAIT timeout? It is associated with maximal packet
85 * lifetime in the internet, which results in wrong conclusion, that
86 * it is set to catch "old duplicate segments" wandering out of their path.
87 * It is not quite correct. This timeout is calculated so that it exceeds
88 * maximal retransmission timeout enough to allow to lose one (or more)
89 * segments sent by peer and our ACKs. This time may be calculated from RTO.
90 * * When TIME-WAIT socket receives RST, it means that another end
91 * finally closed and we are allowed to kill TIME-WAIT too.
92 * * Second purpose of TIME-WAIT is catching old duplicate segments.
93 * Well, certainly it is pure paranoia, but if we load TIME-WAIT
94 * with this semantics, we MUST NOT kill TIME-WAIT state with RSTs.
95 * * If we invented some more clever way to catch duplicates
96 * (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs.
98 * The algorithm below is based on FORMAL INTERPRETATION of RFCs.
99 * When you compare it to RFCs, please, read section SEGMENT ARRIVES
100 * from the very beginning.
102 * NOTE. With recycling (and later with fin-wait-2) TW bucket
103 * is _not_ stateless. It means, that strictly speaking we must
104 * spinlock it. I do not want! Well, probability of misbehaviour
105 * is ridiculously low and, seems, we could use some mb() tricks
106 * to avoid misread sequence numbers, states etc. --ANK
108 * We don't need to initialize tmp_out.sack_ok as we don't use the results
111 tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb,
112 const struct tcphdr *th)
114 struct tcp_options_received tmp_opt;
115 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
116 bool paws_reject = false;
118 tmp_opt.saw_tstamp = 0;
119 if (th->doff > (sizeof(*th) >> 2) && tcptw->tw_ts_recent_stamp) {
120 tcp_parse_options(skb, &tmp_opt, 0, NULL);
122 if (tmp_opt.saw_tstamp) {
123 tmp_opt.rcv_tsecr -= tcptw->tw_ts_offset;
124 tmp_opt.ts_recent = tcptw->tw_ts_recent;
125 tmp_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp;
126 paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
130 if (tw->tw_substate == TCP_FIN_WAIT2) {
131 /* Just repeat all the checks of tcp_rcv_state_process() */
133 /* Out of window, send ACK */
135 !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
137 tcptw->tw_rcv_nxt + tcptw->tw_rcv_wnd))
138 return tcp_timewait_check_oow_rate_limit(
139 tw, skb, LINUX_MIB_TCPACKSKIPPEDFINWAIT2);
144 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt))
149 !after(TCP_SKB_CB(skb)->end_seq, tcptw->tw_rcv_nxt) ||
150 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) {
152 return TCP_TW_SUCCESS;
155 /* New data or FIN. If new data arrive after half-duplex close,
159 TCP_SKB_CB(skb)->end_seq != tcptw->tw_rcv_nxt + 1) {
161 inet_twsk_deschedule(tw, &tcp_death_row);
166 /* FIN arrived, enter true time-wait state. */
167 tw->tw_substate = TCP_TIME_WAIT;
168 tcptw->tw_rcv_nxt = TCP_SKB_CB(skb)->end_seq;
169 if (tmp_opt.saw_tstamp) {
170 tcptw->tw_ts_recent_stamp = get_seconds();
171 tcptw->tw_ts_recent = tmp_opt.rcv_tsval;
174 if (tcp_death_row.sysctl_tw_recycle &&
175 tcptw->tw_ts_recent_stamp &&
176 tcp_tw_remember_stamp(tw))
177 inet_twsk_schedule(tw, &tcp_death_row, tw->tw_timeout,
180 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN,
186 * Now real TIME-WAIT state.
189 * "When a connection is [...] on TIME-WAIT state [...]
190 * [a TCP] MAY accept a new SYN from the remote TCP to
191 * reopen the connection directly, if it:
193 * (1) assigns its initial sequence number for the new
194 * connection to be larger than the largest sequence
195 * number it used on the previous connection incarnation,
198 * (2) returns to TIME-WAIT state if the SYN turns out
199 * to be an old duplicate".
203 (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt &&
204 (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) {
205 /* In window segment, it may be only reset or bare ack. */
208 /* This is TIME_WAIT assassination, in two flavors.
209 * Oh well... nobody has a sufficient solution to this
212 if (sysctl_tcp_rfc1337 == 0) {
214 inet_twsk_deschedule(tw, &tcp_death_row);
216 return TCP_TW_SUCCESS;
219 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN,
222 if (tmp_opt.saw_tstamp) {
223 tcptw->tw_ts_recent = tmp_opt.rcv_tsval;
224 tcptw->tw_ts_recent_stamp = get_seconds();
228 return TCP_TW_SUCCESS;
231 /* Out of window segment.
233 All the segments are ACKed immediately.
235 The only exception is new SYN. We accept it, if it is
236 not old duplicate and we are not in danger to be killed
237 by delayed old duplicates. RFC check is that it has
238 newer sequence number works at rates <40Mbit/sec.
239 However, if paws works, it is reliable AND even more,
240 we even may relax silly seq space cutoff.
242 RED-PEN: we violate main RFC requirement, if this SYN will appear
243 old duplicate (i.e. we receive RST in reply to SYN-ACK),
244 we must return socket to time-wait state. It is not good,
248 if (th->syn && !th->rst && !th->ack && !paws_reject &&
249 (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) ||
250 (tmp_opt.saw_tstamp &&
251 (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) {
252 u32 isn = tcptw->tw_snd_nxt + 65535 + 2;
255 TCP_SKB_CB(skb)->tcp_tw_isn = isn;
260 NET_INC_STATS_BH(twsk_net(tw), LINUX_MIB_PAWSESTABREJECTED);
263 /* In this case we must reset the TIMEWAIT timer.
265 * If it is ACKless SYN it may be both old duplicate
266 * and new good SYN with random sequence number <rcv_nxt.
267 * Do not reschedule in the last case.
269 if (paws_reject || th->ack)
270 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN,
273 return tcp_timewait_check_oow_rate_limit(
274 tw, skb, LINUX_MIB_TCPACKSKIPPEDTIMEWAIT);
277 return TCP_TW_SUCCESS;
279 EXPORT_SYMBOL(tcp_timewait_state_process);
282 * Move a socket to time-wait or dead fin-wait-2 state.
284 void tcp_time_wait(struct sock *sk, int state, int timeo)
286 struct inet_timewait_sock *tw = NULL;
287 const struct inet_connection_sock *icsk = inet_csk(sk);
288 const struct tcp_sock *tp = tcp_sk(sk);
289 bool recycle_ok = false;
291 if (tcp_death_row.sysctl_tw_recycle && tp->rx_opt.ts_recent_stamp)
292 recycle_ok = tcp_remember_stamp(sk);
294 if (tcp_death_row.tw_count < tcp_death_row.sysctl_max_tw_buckets)
295 tw = inet_twsk_alloc(sk, state);
298 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
299 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1);
300 struct inet_sock *inet = inet_sk(sk);
302 tw->tw_transparent = inet->transparent;
303 tw->tw_rcv_wscale = tp->rx_opt.rcv_wscale;
304 tcptw->tw_rcv_nxt = tp->rcv_nxt;
305 tcptw->tw_snd_nxt = tp->snd_nxt;
306 tcptw->tw_rcv_wnd = tcp_receive_window(tp);
307 tcptw->tw_ts_recent = tp->rx_opt.ts_recent;
308 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp;
309 tcptw->tw_ts_offset = tp->tsoffset;
310 tcptw->tw_last_oow_ack_time = 0;
312 #if IS_ENABLED(CONFIG_IPV6)
313 if (tw->tw_family == PF_INET6) {
314 struct ipv6_pinfo *np = inet6_sk(sk);
316 tw->tw_v6_daddr = sk->sk_v6_daddr;
317 tw->tw_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
318 tw->tw_tclass = np->tclass;
319 tw->tw_flowlabel = np->flow_label >> 12;
320 tw->tw_ipv6only = sk->sk_ipv6only;
324 #ifdef CONFIG_TCP_MD5SIG
326 * The timewait bucket does not have the key DB from the
327 * sock structure. We just make a quick copy of the
328 * md5 key being used (if indeed we are using one)
329 * so the timewait ack generating code has the key.
332 struct tcp_md5sig_key *key;
333 tcptw->tw_md5_key = NULL;
334 key = tp->af_specific->md5_lookup(sk, sk);
336 tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC);
337 if (tcptw->tw_md5_key && !tcp_alloc_md5sig_pool())
343 /* Linkage updates. */
344 __inet_twsk_hashdance(tw, sk, &tcp_hashinfo);
346 /* Get the TIME_WAIT timeout firing. */
351 tw->tw_timeout = rto;
353 tw->tw_timeout = TCP_TIMEWAIT_LEN;
354 if (state == TCP_TIME_WAIT)
355 timeo = TCP_TIMEWAIT_LEN;
358 inet_twsk_schedule(tw, &tcp_death_row, timeo,
362 /* Sorry, if we're out of memory, just CLOSE this
363 * socket up. We've got bigger problems than
364 * non-graceful socket closings.
366 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPTIMEWAITOVERFLOW);
369 tcp_update_metrics(sk);
373 void tcp_twsk_destructor(struct sock *sk)
375 #ifdef CONFIG_TCP_MD5SIG
376 struct tcp_timewait_sock *twsk = tcp_twsk(sk);
378 if (twsk->tw_md5_key)
379 kfree_rcu(twsk->tw_md5_key, rcu);
382 EXPORT_SYMBOL_GPL(tcp_twsk_destructor);
384 void tcp_openreq_init_rwin(struct request_sock *req,
385 struct sock *sk, struct dst_entry *dst)
387 struct inet_request_sock *ireq = inet_rsk(req);
388 struct tcp_sock *tp = tcp_sk(sk);
390 int mss = dst_metric_advmss(dst);
392 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
393 mss = tp->rx_opt.user_mss;
395 /* Set this up on the first call only */
396 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
398 /* limit the window selection if the user enforce a smaller rx buffer */
399 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
400 (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
401 req->window_clamp = tcp_full_space(sk);
403 /* tcp_full_space because it is guaranteed to be the first packet */
404 tcp_select_initial_window(tcp_full_space(sk),
405 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
410 dst_metric(dst, RTAX_INITRWND));
411 ireq->rcv_wscale = rcv_wscale;
413 EXPORT_SYMBOL(tcp_openreq_init_rwin);
415 static void tcp_ecn_openreq_child(struct tcp_sock *tp,
416 const struct request_sock *req)
418 tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0;
421 void tcp_ca_openreq_child(struct sock *sk, const struct dst_entry *dst)
423 struct inet_connection_sock *icsk = inet_csk(sk);
424 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
425 bool ca_got_dst = false;
427 if (ca_key != TCP_CA_UNSPEC) {
428 const struct tcp_congestion_ops *ca;
431 ca = tcp_ca_find_key(ca_key);
432 if (likely(ca && try_module_get(ca->owner))) {
433 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
434 icsk->icsk_ca_ops = ca;
440 if (!ca_got_dst && !try_module_get(icsk->icsk_ca_ops->owner))
441 tcp_assign_congestion_control(sk);
443 tcp_set_ca_state(sk, TCP_CA_Open);
445 EXPORT_SYMBOL_GPL(tcp_ca_openreq_child);
447 /* This is not only more efficient than what we used to do, it eliminates
448 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM
450 * Actually, we could lots of memory writes here. tp of listening
451 * socket contains all necessary default parameters.
453 struct sock *tcp_create_openreq_child(struct sock *sk, struct request_sock *req, struct sk_buff *skb)
455 struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC);
458 const struct inet_request_sock *ireq = inet_rsk(req);
459 struct tcp_request_sock *treq = tcp_rsk(req);
460 struct inet_connection_sock *newicsk = inet_csk(newsk);
461 struct tcp_sock *newtp = tcp_sk(newsk);
463 /* Now setup tcp_sock */
464 newtp->pred_flags = 0;
466 newtp->rcv_wup = newtp->copied_seq =
467 newtp->rcv_nxt = treq->rcv_isn + 1;
469 newtp->snd_sml = newtp->snd_una =
470 newtp->snd_nxt = newtp->snd_up = treq->snt_isn + 1;
472 tcp_prequeue_init(newtp);
473 INIT_LIST_HEAD(&newtp->tsq_node);
475 tcp_init_wl(newtp, treq->rcv_isn);
478 newtp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
479 newicsk->icsk_rto = TCP_TIMEOUT_INIT;
481 newtp->packets_out = 0;
482 newtp->retrans_out = 0;
483 newtp->sacked_out = 0;
484 newtp->fackets_out = 0;
485 newtp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
486 tcp_enable_early_retrans(newtp);
487 newtp->tlp_high_seq = 0;
488 newtp->lsndtime = treq->snt_synack;
489 newtp->last_oow_ack_time = 0;
490 newtp->total_retrans = req->num_retrans;
492 /* So many TCP implementations out there (incorrectly) count the
493 * initial SYN frame in their delayed-ACK and congestion control
494 * algorithms that we must have the following bandaid to talk
495 * efficiently to them. -DaveM
497 newtp->snd_cwnd = TCP_INIT_CWND;
498 newtp->snd_cwnd_cnt = 0;
500 tcp_init_xmit_timers(newsk);
501 __skb_queue_head_init(&newtp->out_of_order_queue);
502 newtp->write_seq = newtp->pushed_seq = treq->snt_isn + 1;
504 newtp->rx_opt.saw_tstamp = 0;
506 newtp->rx_opt.dsack = 0;
507 newtp->rx_opt.num_sacks = 0;
511 if (sock_flag(newsk, SOCK_KEEPOPEN))
512 inet_csk_reset_keepalive_timer(newsk,
513 keepalive_time_when(newtp));
515 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok;
516 if ((newtp->rx_opt.sack_ok = ireq->sack_ok) != 0) {
518 tcp_enable_fack(newtp);
520 newtp->window_clamp = req->window_clamp;
521 newtp->rcv_ssthresh = req->rcv_wnd;
522 newtp->rcv_wnd = req->rcv_wnd;
523 newtp->rx_opt.wscale_ok = ireq->wscale_ok;
524 if (newtp->rx_opt.wscale_ok) {
525 newtp->rx_opt.snd_wscale = ireq->snd_wscale;
526 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale;
528 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0;
529 newtp->window_clamp = min(newtp->window_clamp, 65535U);
531 newtp->snd_wnd = (ntohs(tcp_hdr(skb)->window) <<
532 newtp->rx_opt.snd_wscale);
533 newtp->max_window = newtp->snd_wnd;
535 if (newtp->rx_opt.tstamp_ok) {
536 newtp->rx_opt.ts_recent = req->ts_recent;
537 newtp->rx_opt.ts_recent_stamp = get_seconds();
538 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
540 newtp->rx_opt.ts_recent_stamp = 0;
541 newtp->tcp_header_len = sizeof(struct tcphdr);
544 #ifdef CONFIG_TCP_MD5SIG
545 newtp->md5sig_info = NULL; /*XXX*/
546 if (newtp->af_specific->md5_lookup(sk, newsk))
547 newtp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
549 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len)
550 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len;
551 newtp->rx_opt.mss_clamp = req->mss;
552 tcp_ecn_openreq_child(newtp, req);
553 newtp->fastopen_rsk = NULL;
554 newtp->syn_data_acked = 0;
556 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_PASSIVEOPENS);
560 EXPORT_SYMBOL(tcp_create_openreq_child);
563 * Process an incoming packet for SYN_RECV sockets represented as a
564 * request_sock. Normally sk is the listener socket but for TFO it
565 * points to the child socket.
567 * XXX (TFO) - The current impl contains a special check for ack
568 * validation and inside tcp_v4_reqsk_send_ack(). Can we do better?
570 * We don't need to initialize tmp_opt.sack_ok as we don't use the results
573 struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb,
574 struct request_sock *req,
575 struct request_sock **prev,
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 BUG_ON(fastopen == (sk->sk_state == TCP_LISTEN));
586 tmp_opt.saw_tstamp = 0;
587 if (th->doff > (sizeof(struct tcphdr)>>2)) {
588 tcp_parse_options(skb, &tmp_opt, 0, NULL);
590 if (tmp_opt.saw_tstamp) {
591 tmp_opt.ts_recent = req->ts_recent;
592 /* We do not store true stamp, but it is not required,
593 * it can be estimated (approximately)
596 tmp_opt.ts_recent_stamp = get_seconds() - ((TCP_TIMEOUT_INIT/HZ)<<req->num_timeout);
597 paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
601 /* Check for pure retransmitted SYN. */
602 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn &&
603 flg == TCP_FLAG_SYN &&
606 * RFC793 draws (Incorrectly! It was fixed in RFC1122)
607 * this case on figure 6 and figure 8, but formal
608 * protocol description says NOTHING.
609 * To be more exact, it says that we should send ACK,
610 * because this segment (at least, if it has no data)
613 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT
614 * describe SYN-RECV state. All the description
615 * is wrong, we cannot believe to it and should
616 * rely only on common sense and implementation
619 * Enforce "SYN-ACK" according to figure 8, figure 6
620 * of RFC793, fixed by RFC1122.
622 * Note that even if there is new data in the SYN packet
623 * they will be thrown away too.
625 * Reset timer after retransmitting SYNACK, similar to
626 * the idea of fast retransmit in recovery.
628 if (!tcp_oow_rate_limited(sock_net(sk), skb,
629 LINUX_MIB_TCPACKSKIPPEDSYNRECV,
630 &tcp_rsk(req)->last_oow_ack_time) &&
632 !inet_rtx_syn_ack(sk, req))
633 req->expires = min(TCP_TIMEOUT_INIT << req->num_timeout,
634 TCP_RTO_MAX) + jiffies;
638 /* Further reproduces section "SEGMENT ARRIVES"
639 for state SYN-RECEIVED of RFC793.
640 It is broken, however, it does not work only
641 when SYNs are crossed.
643 You would think that SYN crossing is impossible here, since
644 we should have a SYN_SENT socket (from connect()) on our end,
645 but this is not true if the crossed SYNs were sent to both
646 ends by a malicious third party. We must defend against this,
647 and to do that we first verify the ACK (as per RFC793, page
648 36) and reset if it is invalid. Is this a true full defense?
649 To convince ourselves, let us consider a way in which the ACK
650 test can still pass in this 'malicious crossed SYNs' case.
651 Malicious sender sends identical SYNs (and thus identical sequence
652 numbers) to both A and B:
657 By our good fortune, both A and B select the same initial
658 send sequence number of seven :-)
660 A: sends SYN|ACK, seq=7, ack_seq=8
661 B: sends SYN|ACK, seq=7, ack_seq=8
663 So we are now A eating this SYN|ACK, ACK test passes. So
664 does sequence test, SYN is truncated, and thus we consider
667 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this
668 bare ACK. Otherwise, we create an established connection. Both
669 ends (listening sockets) accept the new incoming connection and try
670 to talk to each other. 8-)
672 Note: This case is both harmless, and rare. Possibility is about the
673 same as us discovering intelligent life on another plant tomorrow.
675 But generally, we should (RFC lies!) to accept ACK
676 from SYNACK both here and in tcp_rcv_state_process().
677 tcp_rcv_state_process() does not, hence, we do not too.
679 Note that the case is absolutely generic:
680 we cannot optimize anything here without
681 violating protocol. All the checks must be made
682 before attempt to create socket.
685 /* RFC793 page 36: "If the connection is in any non-synchronized state ...
686 * and the incoming segment acknowledges something not yet
687 * sent (the segment carries an unacceptable ACK) ...
690 * Invalid ACK: reset will be sent by listening socket.
691 * Note that the ACK validity check for a Fast Open socket is done
692 * elsewhere and is checked directly against the child socket rather
693 * than req because user data may have been sent out.
695 if ((flg & TCP_FLAG_ACK) && !fastopen &&
696 (TCP_SKB_CB(skb)->ack_seq !=
697 tcp_rsk(req)->snt_isn + 1))
700 /* Also, it would be not so bad idea to check rcv_tsecr, which
701 * is essentially ACK extension and too early or too late values
702 * should cause reset in unsynchronized states.
705 /* RFC793: "first check sequence number". */
707 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
708 tcp_rsk(req)->rcv_nxt, tcp_rsk(req)->rcv_nxt + req->rcv_wnd)) {
709 /* Out of window: send ACK and drop. */
710 if (!(flg & TCP_FLAG_RST))
711 req->rsk_ops->send_ack(sk, skb, req);
713 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
717 /* In sequence, PAWS is OK. */
719 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_nxt))
720 req->ts_recent = tmp_opt.rcv_tsval;
722 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) {
723 /* Truncate SYN, it is out of window starting
724 at tcp_rsk(req)->rcv_isn + 1. */
725 flg &= ~TCP_FLAG_SYN;
728 /* RFC793: "second check the RST bit" and
729 * "fourth, check the SYN bit"
731 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) {
732 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
733 goto embryonic_reset;
736 /* ACK sequence verified above, just make sure ACK is
737 * set. If ACK not set, just silently drop the packet.
739 * XXX (TFO) - if we ever allow "data after SYN", the
740 * following check needs to be removed.
742 if (!(flg & TCP_FLAG_ACK))
745 /* For Fast Open no more processing is needed (sk is the
751 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */
752 if (req->num_timeout < inet_csk(sk)->icsk_accept_queue.rskq_defer_accept &&
753 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) {
754 inet_rsk(req)->acked = 1;
755 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP);
759 /* OK, ACK is valid, create big socket and
760 * feed this segment to it. It will repeat all
761 * the tests. THIS SEGMENT MUST MOVE SOCKET TO
762 * ESTABLISHED STATE. If it will be dropped after
763 * socket is created, wait for troubles.
765 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL);
767 goto listen_overflow;
769 inet_csk_reqsk_queue_unlink(sk, req, prev);
770 inet_csk_reqsk_queue_removed(sk, req);
772 inet_csk_reqsk_queue_add(sk, req, child);
776 if (!sysctl_tcp_abort_on_overflow) {
777 inet_rsk(req)->acked = 1;
782 if (!(flg & TCP_FLAG_RST)) {
783 /* Received a bad SYN pkt - for TFO We try not to reset
784 * the local connection unless it's really necessary to
785 * avoid becoming vulnerable to outside attack aiming at
786 * resetting legit local connections.
788 req->rsk_ops->send_reset(sk, skb);
789 } else if (fastopen) { /* received a valid RST pkt */
790 reqsk_fastopen_remove(sk, req, true);
794 inet_csk_reqsk_queue_drop(sk, req, prev);
795 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_EMBRYONICRSTS);
799 EXPORT_SYMBOL(tcp_check_req);
802 * Queue segment on the new socket if the new socket is active,
803 * otherwise we just shortcircuit this and continue with
806 * For the vast majority of cases child->sk_state will be TCP_SYN_RECV
807 * when entering. But other states are possible due to a race condition
808 * where after __inet_lookup_established() fails but before the listener
809 * locked is obtained, other packets cause the same connection to
813 int tcp_child_process(struct sock *parent, struct sock *child,
817 int state = child->sk_state;
819 if (!sock_owned_by_user(child)) {
820 ret = tcp_rcv_state_process(child, skb, tcp_hdr(skb),
822 /* Wakeup parent, send SIGIO */
823 if (state == TCP_SYN_RECV && child->sk_state != state)
824 parent->sk_data_ready(parent);
826 /* Alas, it is possible again, because we do lookup
827 * in main socket hash table and lock on listening
828 * socket does not protect us more.
830 __sk_add_backlog(child, skb);
833 bh_unlock_sock(child);
837 EXPORT_SYMBOL(tcp_child_process);