Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
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.
6 *
7 * Implementation of the Transmission Control Protocol(TCP).
8 *
9 * Authors: Ross Biro
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>
20 */
21
22#include <net/tcp.h>
23#include <net/xfrm.h>
24#include <net/busy_poll.h>
25#include <net/rstreason.h>
26
27static bool tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win)
28{
29 if (seq == s_win)
30 return true;
31 if (after(end_seq, s_win) && before(seq, e_win))
32 return true;
33 return seq == e_win && seq == end_seq;
34}
35
36static enum tcp_tw_status
37tcp_timewait_check_oow_rate_limit(struct inet_timewait_sock *tw,
38 const struct sk_buff *skb, int mib_idx)
39{
40 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
41
42 if (!tcp_oow_rate_limited(twsk_net(tw), skb, mib_idx,
43 &tcptw->tw_last_oow_ack_time)) {
44 /* Send ACK. Note, we do not put the bucket,
45 * it will be released by caller.
46 */
47 return TCP_TW_ACK;
48 }
49
50 /* We are rate-limiting, so just release the tw sock and drop skb. */
51 inet_twsk_put(tw);
52 return TCP_TW_SUCCESS;
53}
54
55static void twsk_rcv_nxt_update(struct tcp_timewait_sock *tcptw, u32 seq,
56 u32 rcv_nxt)
57{
58#ifdef CONFIG_TCP_AO
59 struct tcp_ao_info *ao;
60
61 ao = rcu_dereference(tcptw->ao_info);
62 if (unlikely(ao && seq < rcv_nxt))
63 WRITE_ONCE(ao->rcv_sne, ao->rcv_sne + 1);
64#endif
65 WRITE_ONCE(tcptw->tw_rcv_nxt, seq);
66}
67
68/*
69 * * Main purpose of TIME-WAIT state is to close connection gracefully,
70 * when one of ends sits in LAST-ACK or CLOSING retransmitting FIN
71 * (and, probably, tail of data) and one or more our ACKs are lost.
72 * * What is TIME-WAIT timeout? It is associated with maximal packet
73 * lifetime in the internet, which results in wrong conclusion, that
74 * it is set to catch "old duplicate segments" wandering out of their path.
75 * It is not quite correct. This timeout is calculated so that it exceeds
76 * maximal retransmission timeout enough to allow to lose one (or more)
77 * segments sent by peer and our ACKs. This time may be calculated from RTO.
78 * * When TIME-WAIT socket receives RST, it means that another end
79 * finally closed and we are allowed to kill TIME-WAIT too.
80 * * Second purpose of TIME-WAIT is catching old duplicate segments.
81 * Well, certainly it is pure paranoia, but if we load TIME-WAIT
82 * with this semantics, we MUST NOT kill TIME-WAIT state with RSTs.
83 * * If we invented some more clever way to catch duplicates
84 * (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs.
85 *
86 * The algorithm below is based on FORMAL INTERPRETATION of RFCs.
87 * When you compare it to RFCs, please, read section SEGMENT ARRIVES
88 * from the very beginning.
89 *
90 * NOTE. With recycling (and later with fin-wait-2) TW bucket
91 * is _not_ stateless. It means, that strictly speaking we must
92 * spinlock it. I do not want! Well, probability of misbehaviour
93 * is ridiculously low and, seems, we could use some mb() tricks
94 * to avoid misread sequence numbers, states etc. --ANK
95 *
96 * We don't need to initialize tmp_out.sack_ok as we don't use the results
97 */
98enum tcp_tw_status
99tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb,
100 const struct tcphdr *th, u32 *tw_isn)
101{
102 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
103 u32 rcv_nxt = READ_ONCE(tcptw->tw_rcv_nxt);
104 struct tcp_options_received tmp_opt;
105 bool paws_reject = false;
106 int ts_recent_stamp;
107
108 tmp_opt.saw_tstamp = 0;
109 ts_recent_stamp = READ_ONCE(tcptw->tw_ts_recent_stamp);
110 if (th->doff > (sizeof(*th) >> 2) && ts_recent_stamp) {
111 tcp_parse_options(twsk_net(tw), skb, &tmp_opt, 0, NULL);
112
113 if (tmp_opt.saw_tstamp) {
114 if (tmp_opt.rcv_tsecr)
115 tmp_opt.rcv_tsecr -= tcptw->tw_ts_offset;
116 tmp_opt.ts_recent = READ_ONCE(tcptw->tw_ts_recent);
117 tmp_opt.ts_recent_stamp = ts_recent_stamp;
118 paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
119 }
120 }
121
122 if (READ_ONCE(tw->tw_substate) == TCP_FIN_WAIT2) {
123 /* Just repeat all the checks of tcp_rcv_state_process() */
124
125 /* Out of window, send ACK */
126 if (paws_reject ||
127 !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
128 rcv_nxt,
129 rcv_nxt + tcptw->tw_rcv_wnd))
130 return tcp_timewait_check_oow_rate_limit(
131 tw, skb, LINUX_MIB_TCPACKSKIPPEDFINWAIT2);
132
133 if (th->rst)
134 goto kill;
135
136 if (th->syn && !before(TCP_SKB_CB(skb)->seq, rcv_nxt))
137 return TCP_TW_RST;
138
139 /* Dup ACK? */
140 if (!th->ack ||
141 !after(TCP_SKB_CB(skb)->end_seq, rcv_nxt) ||
142 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) {
143 inet_twsk_put(tw);
144 return TCP_TW_SUCCESS;
145 }
146
147 /* New data or FIN. If new data arrive after half-duplex close,
148 * reset.
149 */
150 if (!th->fin ||
151 TCP_SKB_CB(skb)->end_seq != rcv_nxt + 1)
152 return TCP_TW_RST;
153
154 /* FIN arrived, enter true time-wait state. */
155 WRITE_ONCE(tw->tw_substate, TCP_TIME_WAIT);
156 twsk_rcv_nxt_update(tcptw, TCP_SKB_CB(skb)->end_seq,
157 rcv_nxt);
158
159 if (tmp_opt.saw_tstamp) {
160 WRITE_ONCE(tcptw->tw_ts_recent_stamp,
161 ktime_get_seconds());
162 WRITE_ONCE(tcptw->tw_ts_recent,
163 tmp_opt.rcv_tsval);
164 }
165
166 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
167 return TCP_TW_ACK;
168 }
169
170 /*
171 * Now real TIME-WAIT state.
172 *
173 * RFC 1122:
174 * "When a connection is [...] on TIME-WAIT state [...]
175 * [a TCP] MAY accept a new SYN from the remote TCP to
176 * reopen the connection directly, if it:
177 *
178 * (1) assigns its initial sequence number for the new
179 * connection to be larger than the largest sequence
180 * number it used on the previous connection incarnation,
181 * and
182 *
183 * (2) returns to TIME-WAIT state if the SYN turns out
184 * to be an old duplicate".
185 */
186
187 if (!paws_reject &&
188 (TCP_SKB_CB(skb)->seq == rcv_nxt &&
189 (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) {
190 /* In window segment, it may be only reset or bare ack. */
191
192 if (th->rst) {
193 /* This is TIME_WAIT assassination, in two flavors.
194 * Oh well... nobody has a sufficient solution to this
195 * protocol bug yet.
196 */
197 if (!READ_ONCE(twsk_net(tw)->ipv4.sysctl_tcp_rfc1337)) {
198kill:
199 inet_twsk_deschedule_put(tw);
200 return TCP_TW_SUCCESS;
201 }
202 } else {
203 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
204 }
205
206 if (tmp_opt.saw_tstamp) {
207 WRITE_ONCE(tcptw->tw_ts_recent,
208 tmp_opt.rcv_tsval);
209 WRITE_ONCE(tcptw->tw_ts_recent_stamp,
210 ktime_get_seconds());
211 }
212
213 inet_twsk_put(tw);
214 return TCP_TW_SUCCESS;
215 }
216
217 /* Out of window segment.
218
219 All the segments are ACKed immediately.
220
221 The only exception is new SYN. We accept it, if it is
222 not old duplicate and we are not in danger to be killed
223 by delayed old duplicates. RFC check is that it has
224 newer sequence number works at rates <40Mbit/sec.
225 However, if paws works, it is reliable AND even more,
226 we even may relax silly seq space cutoff.
227
228 RED-PEN: we violate main RFC requirement, if this SYN will appear
229 old duplicate (i.e. we receive RST in reply to SYN-ACK),
230 we must return socket to time-wait state. It is not good,
231 but not fatal yet.
232 */
233
234 if (th->syn && !th->rst && !th->ack && !paws_reject &&
235 (after(TCP_SKB_CB(skb)->seq, rcv_nxt) ||
236 (tmp_opt.saw_tstamp &&
237 (s32)(READ_ONCE(tcptw->tw_ts_recent) - tmp_opt.rcv_tsval) < 0))) {
238 u32 isn = tcptw->tw_snd_nxt + 65535 + 2;
239 if (isn == 0)
240 isn++;
241 *tw_isn = isn;
242 return TCP_TW_SYN;
243 }
244
245 if (paws_reject)
246 __NET_INC_STATS(twsk_net(tw), LINUX_MIB_PAWSESTABREJECTED);
247
248 if (!th->rst) {
249 /* In this case we must reset the TIMEWAIT timer.
250 *
251 * If it is ACKless SYN it may be both old duplicate
252 * and new good SYN with random sequence number <rcv_nxt.
253 * Do not reschedule in the last case.
254 */
255 if (paws_reject || th->ack)
256 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
257
258 return tcp_timewait_check_oow_rate_limit(
259 tw, skb, LINUX_MIB_TCPACKSKIPPEDTIMEWAIT);
260 }
261 inet_twsk_put(tw);
262 return TCP_TW_SUCCESS;
263}
264EXPORT_SYMBOL(tcp_timewait_state_process);
265
266static void tcp_time_wait_init(struct sock *sk, struct tcp_timewait_sock *tcptw)
267{
268#ifdef CONFIG_TCP_MD5SIG
269 const struct tcp_sock *tp = tcp_sk(sk);
270 struct tcp_md5sig_key *key;
271
272 /*
273 * The timewait bucket does not have the key DB from the
274 * sock structure. We just make a quick copy of the
275 * md5 key being used (if indeed we are using one)
276 * so the timewait ack generating code has the key.
277 */
278 tcptw->tw_md5_key = NULL;
279 if (!static_branch_unlikely(&tcp_md5_needed.key))
280 return;
281
282 key = tp->af_specific->md5_lookup(sk, sk);
283 if (key) {
284 tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC);
285 if (!tcptw->tw_md5_key)
286 return;
287 if (!static_key_fast_inc_not_disabled(&tcp_md5_needed.key.key))
288 goto out_free;
289 tcp_md5_add_sigpool();
290 }
291 return;
292out_free:
293 WARN_ON_ONCE(1);
294 kfree(tcptw->tw_md5_key);
295 tcptw->tw_md5_key = NULL;
296#endif
297}
298
299/*
300 * Move a socket to time-wait or dead fin-wait-2 state.
301 */
302void tcp_time_wait(struct sock *sk, int state, int timeo)
303{
304 const struct inet_connection_sock *icsk = inet_csk(sk);
305 struct tcp_sock *tp = tcp_sk(sk);
306 struct net *net = sock_net(sk);
307 struct inet_timewait_sock *tw;
308
309 tw = inet_twsk_alloc(sk, &net->ipv4.tcp_death_row, state);
310
311 if (tw) {
312 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
313 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1);
314
315 tw->tw_transparent = inet_test_bit(TRANSPARENT, sk);
316 tw->tw_mark = sk->sk_mark;
317 tw->tw_priority = READ_ONCE(sk->sk_priority);
318 tw->tw_rcv_wscale = tp->rx_opt.rcv_wscale;
319 tcptw->tw_rcv_nxt = tp->rcv_nxt;
320 tcptw->tw_snd_nxt = tp->snd_nxt;
321 tcptw->tw_rcv_wnd = tcp_receive_window(tp);
322 tcptw->tw_ts_recent = tp->rx_opt.ts_recent;
323 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp;
324 tcptw->tw_ts_offset = tp->tsoffset;
325 tw->tw_usec_ts = tp->tcp_usec_ts;
326 tcptw->tw_last_oow_ack_time = 0;
327 tcptw->tw_tx_delay = tp->tcp_tx_delay;
328 tw->tw_txhash = sk->sk_txhash;
329 tw->tw_tx_queue_mapping = sk->sk_tx_queue_mapping;
330#ifdef CONFIG_SOCK_RX_QUEUE_MAPPING
331 tw->tw_rx_queue_mapping = sk->sk_rx_queue_mapping;
332#endif
333#if IS_ENABLED(CONFIG_IPV6)
334 if (tw->tw_family == PF_INET6) {
335 struct ipv6_pinfo *np = inet6_sk(sk);
336
337 tw->tw_v6_daddr = sk->sk_v6_daddr;
338 tw->tw_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
339 tw->tw_tclass = np->tclass;
340 tw->tw_flowlabel = be32_to_cpu(np->flow_label & IPV6_FLOWLABEL_MASK);
341 tw->tw_ipv6only = sk->sk_ipv6only;
342 }
343#endif
344
345 tcp_time_wait_init(sk, tcptw);
346 tcp_ao_time_wait(tcptw, tp);
347
348 /* Get the TIME_WAIT timeout firing. */
349 if (timeo < rto)
350 timeo = rto;
351
352 if (state == TCP_TIME_WAIT)
353 timeo = TCP_TIMEWAIT_LEN;
354
355 /* Linkage updates.
356 * Note that access to tw after this point is illegal.
357 */
358 inet_twsk_hashdance_schedule(tw, sk, net->ipv4.tcp_death_row.hashinfo, timeo);
359 } else {
360 /* Sorry, if we're out of memory, just CLOSE this
361 * socket up. We've got bigger problems than
362 * non-graceful socket closings.
363 */
364 NET_INC_STATS(net, LINUX_MIB_TCPTIMEWAITOVERFLOW);
365 }
366
367 tcp_update_metrics(sk);
368 tcp_done(sk);
369}
370EXPORT_SYMBOL(tcp_time_wait);
371
372#ifdef CONFIG_TCP_MD5SIG
373static void tcp_md5_twsk_free_rcu(struct rcu_head *head)
374{
375 struct tcp_md5sig_key *key;
376
377 key = container_of(head, struct tcp_md5sig_key, rcu);
378 kfree(key);
379 static_branch_slow_dec_deferred(&tcp_md5_needed);
380 tcp_md5_release_sigpool();
381}
382#endif
383
384void tcp_twsk_destructor(struct sock *sk)
385{
386#ifdef CONFIG_TCP_MD5SIG
387 if (static_branch_unlikely(&tcp_md5_needed.key)) {
388 struct tcp_timewait_sock *twsk = tcp_twsk(sk);
389
390 if (twsk->tw_md5_key)
391 call_rcu(&twsk->tw_md5_key->rcu, tcp_md5_twsk_free_rcu);
392 }
393#endif
394 tcp_ao_destroy_sock(sk, true);
395}
396EXPORT_SYMBOL_GPL(tcp_twsk_destructor);
397
398void tcp_twsk_purge(struct list_head *net_exit_list)
399{
400 bool purged_once = false;
401 struct net *net;
402
403 list_for_each_entry(net, net_exit_list, exit_list) {
404 if (net->ipv4.tcp_death_row.hashinfo->pernet) {
405 /* Even if tw_refcount == 1, we must clean up kernel reqsk */
406 inet_twsk_purge(net->ipv4.tcp_death_row.hashinfo);
407 } else if (!purged_once) {
408 inet_twsk_purge(&tcp_hashinfo);
409 purged_once = true;
410 }
411 }
412}
413
414/* Warning : This function is called without sk_listener being locked.
415 * Be sure to read socket fields once, as their value could change under us.
416 */
417void tcp_openreq_init_rwin(struct request_sock *req,
418 const struct sock *sk_listener,
419 const struct dst_entry *dst)
420{
421 struct inet_request_sock *ireq = inet_rsk(req);
422 const struct tcp_sock *tp = tcp_sk(sk_listener);
423 int full_space = tcp_full_space(sk_listener);
424 u32 window_clamp;
425 __u8 rcv_wscale;
426 u32 rcv_wnd;
427 int mss;
428
429 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
430 window_clamp = READ_ONCE(tp->window_clamp);
431 /* Set this up on the first call only */
432 req->rsk_window_clamp = window_clamp ? : dst_metric(dst, RTAX_WINDOW);
433
434 /* limit the window selection if the user enforce a smaller rx buffer */
435 if (sk_listener->sk_userlocks & SOCK_RCVBUF_LOCK &&
436 (req->rsk_window_clamp > full_space || req->rsk_window_clamp == 0))
437 req->rsk_window_clamp = full_space;
438
439 rcv_wnd = tcp_rwnd_init_bpf((struct sock *)req);
440 if (rcv_wnd == 0)
441 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
442 else if (full_space < rcv_wnd * mss)
443 full_space = rcv_wnd * mss;
444
445 /* tcp_full_space because it is guaranteed to be the first packet */
446 tcp_select_initial_window(sk_listener, full_space,
447 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
448 &req->rsk_rcv_wnd,
449 &req->rsk_window_clamp,
450 ireq->wscale_ok,
451 &rcv_wscale,
452 rcv_wnd);
453 ireq->rcv_wscale = rcv_wscale;
454}
455EXPORT_SYMBOL(tcp_openreq_init_rwin);
456
457static void tcp_ecn_openreq_child(struct tcp_sock *tp,
458 const struct request_sock *req)
459{
460 tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0;
461}
462
463void tcp_ca_openreq_child(struct sock *sk, const struct dst_entry *dst)
464{
465 struct inet_connection_sock *icsk = inet_csk(sk);
466 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
467 bool ca_got_dst = false;
468
469 if (ca_key != TCP_CA_UNSPEC) {
470 const struct tcp_congestion_ops *ca;
471
472 rcu_read_lock();
473 ca = tcp_ca_find_key(ca_key);
474 if (likely(ca && bpf_try_module_get(ca, ca->owner))) {
475 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
476 icsk->icsk_ca_ops = ca;
477 ca_got_dst = true;
478 }
479 rcu_read_unlock();
480 }
481
482 /* If no valid choice made yet, assign current system default ca. */
483 if (!ca_got_dst &&
484 (!icsk->icsk_ca_setsockopt ||
485 !bpf_try_module_get(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner)))
486 tcp_assign_congestion_control(sk);
487
488 tcp_set_ca_state(sk, TCP_CA_Open);
489}
490EXPORT_SYMBOL_GPL(tcp_ca_openreq_child);
491
492static void smc_check_reset_syn_req(const struct tcp_sock *oldtp,
493 struct request_sock *req,
494 struct tcp_sock *newtp)
495{
496#if IS_ENABLED(CONFIG_SMC)
497 struct inet_request_sock *ireq;
498
499 if (static_branch_unlikely(&tcp_have_smc)) {
500 ireq = inet_rsk(req);
501 if (oldtp->syn_smc && !ireq->smc_ok)
502 newtp->syn_smc = 0;
503 }
504#endif
505}
506
507/* This is not only more efficient than what we used to do, it eliminates
508 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM
509 *
510 * Actually, we could lots of memory writes here. tp of listening
511 * socket contains all necessary default parameters.
512 */
513struct sock *tcp_create_openreq_child(const struct sock *sk,
514 struct request_sock *req,
515 struct sk_buff *skb)
516{
517 struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC);
518 const struct inet_request_sock *ireq = inet_rsk(req);
519 struct tcp_request_sock *treq = tcp_rsk(req);
520 struct inet_connection_sock *newicsk;
521 const struct tcp_sock *oldtp;
522 struct tcp_sock *newtp;
523 u32 seq;
524
525 if (!newsk)
526 return NULL;
527
528 newicsk = inet_csk(newsk);
529 newtp = tcp_sk(newsk);
530 oldtp = tcp_sk(sk);
531
532 smc_check_reset_syn_req(oldtp, req, newtp);
533
534 /* Now setup tcp_sock */
535 newtp->pred_flags = 0;
536
537 seq = treq->rcv_isn + 1;
538 newtp->rcv_wup = seq;
539 WRITE_ONCE(newtp->copied_seq, seq);
540 WRITE_ONCE(newtp->rcv_nxt, seq);
541 newtp->segs_in = 1;
542
543 seq = treq->snt_isn + 1;
544 newtp->snd_sml = newtp->snd_una = seq;
545 WRITE_ONCE(newtp->snd_nxt, seq);
546 newtp->snd_up = seq;
547
548 INIT_LIST_HEAD(&newtp->tsq_node);
549 INIT_LIST_HEAD(&newtp->tsorted_sent_queue);
550
551 tcp_init_wl(newtp, treq->rcv_isn);
552
553 minmax_reset(&newtp->rtt_min, tcp_jiffies32, ~0U);
554 newicsk->icsk_ack.lrcvtime = tcp_jiffies32;
555
556 newtp->lsndtime = tcp_jiffies32;
557 newsk->sk_txhash = READ_ONCE(treq->txhash);
558 newtp->total_retrans = req->num_retrans;
559
560 tcp_init_xmit_timers(newsk);
561 WRITE_ONCE(newtp->write_seq, newtp->pushed_seq = treq->snt_isn + 1);
562
563 if (sock_flag(newsk, SOCK_KEEPOPEN))
564 inet_csk_reset_keepalive_timer(newsk,
565 keepalive_time_when(newtp));
566
567 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok;
568 newtp->rx_opt.sack_ok = ireq->sack_ok;
569 newtp->window_clamp = req->rsk_window_clamp;
570 newtp->rcv_ssthresh = req->rsk_rcv_wnd;
571 newtp->rcv_wnd = req->rsk_rcv_wnd;
572 newtp->rx_opt.wscale_ok = ireq->wscale_ok;
573 if (newtp->rx_opt.wscale_ok) {
574 newtp->rx_opt.snd_wscale = ireq->snd_wscale;
575 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale;
576 } else {
577 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0;
578 newtp->window_clamp = min(newtp->window_clamp, 65535U);
579 }
580 newtp->snd_wnd = ntohs(tcp_hdr(skb)->window) << newtp->rx_opt.snd_wscale;
581 newtp->max_window = newtp->snd_wnd;
582
583 if (newtp->rx_opt.tstamp_ok) {
584 newtp->tcp_usec_ts = treq->req_usec_ts;
585 newtp->rx_opt.ts_recent = READ_ONCE(req->ts_recent);
586 newtp->rx_opt.ts_recent_stamp = ktime_get_seconds();
587 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
588 } else {
589 newtp->tcp_usec_ts = 0;
590 newtp->rx_opt.ts_recent_stamp = 0;
591 newtp->tcp_header_len = sizeof(struct tcphdr);
592 }
593 if (req->num_timeout) {
594 newtp->total_rto = req->num_timeout;
595 newtp->undo_marker = treq->snt_isn;
596 if (newtp->tcp_usec_ts) {
597 newtp->retrans_stamp = treq->snt_synack;
598 newtp->total_rto_time = (u32)(tcp_clock_us() -
599 newtp->retrans_stamp) / USEC_PER_MSEC;
600 } else {
601 newtp->retrans_stamp = div_u64(treq->snt_synack,
602 USEC_PER_SEC / TCP_TS_HZ);
603 newtp->total_rto_time = tcp_clock_ms() -
604 newtp->retrans_stamp;
605 }
606 newtp->total_rto_recoveries = 1;
607 }
608 newtp->tsoffset = treq->ts_off;
609#ifdef CONFIG_TCP_MD5SIG
610 newtp->md5sig_info = NULL; /*XXX*/
611#endif
612#ifdef CONFIG_TCP_AO
613 newtp->ao_info = NULL;
614
615 if (tcp_rsk_used_ao(req)) {
616 struct tcp_ao_key *ao_key;
617
618 ao_key = treq->af_specific->ao_lookup(sk, req, tcp_rsk(req)->ao_keyid, -1);
619 if (ao_key)
620 newtp->tcp_header_len += tcp_ao_len_aligned(ao_key);
621 }
622 #endif
623 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len)
624 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len;
625 newtp->rx_opt.mss_clamp = req->mss;
626 tcp_ecn_openreq_child(newtp, req);
627 newtp->fastopen_req = NULL;
628 RCU_INIT_POINTER(newtp->fastopen_rsk, NULL);
629
630 newtp->bpf_chg_cc_inprogress = 0;
631 tcp_bpf_clone(sk, newsk);
632
633 __TCP_INC_STATS(sock_net(sk), TCP_MIB_PASSIVEOPENS);
634
635 xa_init_flags(&newsk->sk_user_frags, XA_FLAGS_ALLOC1);
636
637 return newsk;
638}
639EXPORT_SYMBOL(tcp_create_openreq_child);
640
641/*
642 * Process an incoming packet for SYN_RECV sockets represented as a
643 * request_sock. Normally sk is the listener socket but for TFO it
644 * points to the child socket.
645 *
646 * XXX (TFO) - The current impl contains a special check for ack
647 * validation and inside tcp_v4_reqsk_send_ack(). Can we do better?
648 *
649 * We don't need to initialize tmp_opt.sack_ok as we don't use the results
650 *
651 * Note: If @fastopen is true, this can be called from process context.
652 * Otherwise, this is from BH context.
653 */
654
655struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb,
656 struct request_sock *req,
657 bool fastopen, bool *req_stolen)
658{
659 struct tcp_options_received tmp_opt;
660 struct sock *child;
661 const struct tcphdr *th = tcp_hdr(skb);
662 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK);
663 bool paws_reject = false;
664 bool own_req;
665
666 tmp_opt.saw_tstamp = 0;
667 if (th->doff > (sizeof(struct tcphdr)>>2)) {
668 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0, NULL);
669
670 if (tmp_opt.saw_tstamp) {
671 tmp_opt.ts_recent = READ_ONCE(req->ts_recent);
672 if (tmp_opt.rcv_tsecr)
673 tmp_opt.rcv_tsecr -= tcp_rsk(req)->ts_off;
674 /* We do not store true stamp, but it is not required,
675 * it can be estimated (approximately)
676 * from another data.
677 */
678 tmp_opt.ts_recent_stamp = ktime_get_seconds() - reqsk_timeout(req, TCP_RTO_MAX) / HZ;
679 paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
680 }
681 }
682
683 /* Check for pure retransmitted SYN. */
684 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn &&
685 flg == TCP_FLAG_SYN &&
686 !paws_reject) {
687 /*
688 * RFC793 draws (Incorrectly! It was fixed in RFC1122)
689 * this case on figure 6 and figure 8, but formal
690 * protocol description says NOTHING.
691 * To be more exact, it says that we should send ACK,
692 * because this segment (at least, if it has no data)
693 * is out of window.
694 *
695 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT
696 * describe SYN-RECV state. All the description
697 * is wrong, we cannot believe to it and should
698 * rely only on common sense and implementation
699 * experience.
700 *
701 * Enforce "SYN-ACK" according to figure 8, figure 6
702 * of RFC793, fixed by RFC1122.
703 *
704 * Note that even if there is new data in the SYN packet
705 * they will be thrown away too.
706 *
707 * Reset timer after retransmitting SYNACK, similar to
708 * the idea of fast retransmit in recovery.
709 */
710 if (!tcp_oow_rate_limited(sock_net(sk), skb,
711 LINUX_MIB_TCPACKSKIPPEDSYNRECV,
712 &tcp_rsk(req)->last_oow_ack_time) &&
713
714 !inet_rtx_syn_ack(sk, req)) {
715 unsigned long expires = jiffies;
716
717 expires += reqsk_timeout(req, TCP_RTO_MAX);
718 if (!fastopen)
719 mod_timer_pending(&req->rsk_timer, expires);
720 else
721 req->rsk_timer.expires = expires;
722 }
723 return NULL;
724 }
725
726 /* Further reproduces section "SEGMENT ARRIVES"
727 for state SYN-RECEIVED of RFC793.
728 It is broken, however, it does not work only
729 when SYNs are crossed.
730
731 You would think that SYN crossing is impossible here, since
732 we should have a SYN_SENT socket (from connect()) on our end,
733 but this is not true if the crossed SYNs were sent to both
734 ends by a malicious third party. We must defend against this,
735 and to do that we first verify the ACK (as per RFC793, page
736 36) and reset if it is invalid. Is this a true full defense?
737 To convince ourselves, let us consider a way in which the ACK
738 test can still pass in this 'malicious crossed SYNs' case.
739 Malicious sender sends identical SYNs (and thus identical sequence
740 numbers) to both A and B:
741
742 A: gets SYN, seq=7
743 B: gets SYN, seq=7
744
745 By our good fortune, both A and B select the same initial
746 send sequence number of seven :-)
747
748 A: sends SYN|ACK, seq=7, ack_seq=8
749 B: sends SYN|ACK, seq=7, ack_seq=8
750
751 So we are now A eating this SYN|ACK, ACK test passes. So
752 does sequence test, SYN is truncated, and thus we consider
753 it a bare ACK.
754
755 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this
756 bare ACK. Otherwise, we create an established connection. Both
757 ends (listening sockets) accept the new incoming connection and try
758 to talk to each other. 8-)
759
760 Note: This case is both harmless, and rare. Possibility is about the
761 same as us discovering intelligent life on another plant tomorrow.
762
763 But generally, we should (RFC lies!) to accept ACK
764 from SYNACK both here and in tcp_rcv_state_process().
765 tcp_rcv_state_process() does not, hence, we do not too.
766
767 Note that the case is absolutely generic:
768 we cannot optimize anything here without
769 violating protocol. All the checks must be made
770 before attempt to create socket.
771 */
772
773 /* RFC793 page 36: "If the connection is in any non-synchronized state ...
774 * and the incoming segment acknowledges something not yet
775 * sent (the segment carries an unacceptable ACK) ...
776 * a reset is sent."
777 *
778 * Invalid ACK: reset will be sent by listening socket.
779 * Note that the ACK validity check for a Fast Open socket is done
780 * elsewhere and is checked directly against the child socket rather
781 * than req because user data may have been sent out.
782 */
783 if ((flg & TCP_FLAG_ACK) && !fastopen &&
784 (TCP_SKB_CB(skb)->ack_seq !=
785 tcp_rsk(req)->snt_isn + 1))
786 return sk;
787
788 /* Also, it would be not so bad idea to check rcv_tsecr, which
789 * is essentially ACK extension and too early or too late values
790 * should cause reset in unsynchronized states.
791 */
792
793 /* RFC793: "first check sequence number". */
794
795 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq,
796 TCP_SKB_CB(skb)->end_seq,
797 tcp_rsk(req)->rcv_nxt,
798 tcp_rsk(req)->rcv_nxt +
799 tcp_synack_window(req))) {
800 /* Out of window: send ACK and drop. */
801 if (!(flg & TCP_FLAG_RST) &&
802 !tcp_oow_rate_limited(sock_net(sk), skb,
803 LINUX_MIB_TCPACKSKIPPEDSYNRECV,
804 &tcp_rsk(req)->last_oow_ack_time))
805 req->rsk_ops->send_ack(sk, skb, req);
806 if (paws_reject)
807 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
808 return NULL;
809 }
810
811 /* In sequence, PAWS is OK. */
812
813 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) {
814 /* Truncate SYN, it is out of window starting
815 at tcp_rsk(req)->rcv_isn + 1. */
816 flg &= ~TCP_FLAG_SYN;
817 }
818
819 /* RFC793: "second check the RST bit" and
820 * "fourth, check the SYN bit"
821 */
822 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) {
823 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
824 goto embryonic_reset;
825 }
826
827 /* ACK sequence verified above, just make sure ACK is
828 * set. If ACK not set, just silently drop the packet.
829 *
830 * XXX (TFO) - if we ever allow "data after SYN", the
831 * following check needs to be removed.
832 */
833 if (!(flg & TCP_FLAG_ACK))
834 return NULL;
835
836 /* For Fast Open no more processing is needed (sk is the
837 * child socket).
838 */
839 if (fastopen)
840 return sk;
841
842 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */
843 if (req->num_timeout < READ_ONCE(inet_csk(sk)->icsk_accept_queue.rskq_defer_accept) &&
844 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) {
845 inet_rsk(req)->acked = 1;
846 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP);
847 return NULL;
848 }
849
850 /* OK, ACK is valid, create big socket and
851 * feed this segment to it. It will repeat all
852 * the tests. THIS SEGMENT MUST MOVE SOCKET TO
853 * ESTABLISHED STATE. If it will be dropped after
854 * socket is created, wait for troubles.
855 */
856 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL,
857 req, &own_req);
858 if (!child)
859 goto listen_overflow;
860
861 if (own_req && tmp_opt.saw_tstamp &&
862 !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_nxt))
863 tcp_sk(child)->rx_opt.ts_recent = tmp_opt.rcv_tsval;
864
865 if (own_req && rsk_drop_req(req)) {
866 reqsk_queue_removed(&inet_csk(req->rsk_listener)->icsk_accept_queue, req);
867 inet_csk_reqsk_queue_drop_and_put(req->rsk_listener, req);
868 return child;
869 }
870
871 sock_rps_save_rxhash(child, skb);
872 tcp_synack_rtt_meas(child, req);
873 *req_stolen = !own_req;
874 return inet_csk_complete_hashdance(sk, child, req, own_req);
875
876listen_overflow:
877 if (sk != req->rsk_listener)
878 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMIGRATEREQFAILURE);
879
880 if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_abort_on_overflow)) {
881 inet_rsk(req)->acked = 1;
882 return NULL;
883 }
884
885embryonic_reset:
886 if (!(flg & TCP_FLAG_RST)) {
887 /* Received a bad SYN pkt - for TFO We try not to reset
888 * the local connection unless it's really necessary to
889 * avoid becoming vulnerable to outside attack aiming at
890 * resetting legit local connections.
891 */
892 req->rsk_ops->send_reset(sk, skb, SK_RST_REASON_INVALID_SYN);
893 } else if (fastopen) { /* received a valid RST pkt */
894 reqsk_fastopen_remove(sk, req, true);
895 tcp_reset(sk, skb);
896 }
897 if (!fastopen) {
898 bool unlinked = inet_csk_reqsk_queue_drop(sk, req);
899
900 if (unlinked)
901 __NET_INC_STATS(sock_net(sk), LINUX_MIB_EMBRYONICRSTS);
902 *req_stolen = !unlinked;
903 }
904 return NULL;
905}
906EXPORT_SYMBOL(tcp_check_req);
907
908/*
909 * Queue segment on the new socket if the new socket is active,
910 * otherwise we just shortcircuit this and continue with
911 * the new socket.
912 *
913 * For the vast majority of cases child->sk_state will be TCP_SYN_RECV
914 * when entering. But other states are possible due to a race condition
915 * where after __inet_lookup_established() fails but before the listener
916 * locked is obtained, other packets cause the same connection to
917 * be created.
918 */
919
920enum skb_drop_reason tcp_child_process(struct sock *parent, struct sock *child,
921 struct sk_buff *skb)
922 __releases(&((child)->sk_lock.slock))
923{
924 enum skb_drop_reason reason = SKB_NOT_DROPPED_YET;
925 int state = child->sk_state;
926
927 /* record sk_napi_id and sk_rx_queue_mapping of child. */
928 sk_mark_napi_id_set(child, skb);
929
930 tcp_segs_in(tcp_sk(child), skb);
931 if (!sock_owned_by_user(child)) {
932 reason = tcp_rcv_state_process(child, skb);
933 /* Wakeup parent, send SIGIO */
934 if (state == TCP_SYN_RECV && child->sk_state != state)
935 parent->sk_data_ready(parent);
936 } else {
937 /* Alas, it is possible again, because we do lookup
938 * in main socket hash table and lock on listening
939 * socket does not protect us more.
940 */
941 __sk_add_backlog(child, skb);
942 }
943
944 bh_unlock_sock(child);
945 sock_put(child);
946 return reason;
947}
948EXPORT_SYMBOL(tcp_child_process);
1/*
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.
5 *
6 * Implementation of the Transmission Control Protocol(TCP).
7 *
8 * Authors: Ross Biro
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>
19 */
20
21#include <linux/mm.h>
22#include <linux/module.h>
23#include <linux/slab.h>
24#include <linux/sysctl.h>
25#include <linux/workqueue.h>
26#include <net/tcp.h>
27#include <net/inet_common.h>
28#include <net/xfrm.h>
29
30int sysctl_tcp_syncookies __read_mostly = 1;
31EXPORT_SYMBOL(sysctl_tcp_syncookies);
32
33int sysctl_tcp_abort_on_overflow __read_mostly;
34
35struct 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 */
45
46 .twcal_hand = -1,
47 .twcal_timer = TIMER_INITIALIZER(inet_twdr_twcal_tick, 0,
48 (unsigned long)&tcp_death_row),
49};
50EXPORT_SYMBOL_GPL(tcp_death_row);
51
52static bool tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win)
53{
54 if (seq == s_win)
55 return true;
56 if (after(end_seq, s_win) && before(seq, e_win))
57 return true;
58 return seq == e_win && seq == end_seq;
59}
60
61/*
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.
78 *
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.
82 *
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
88 *
89 * We don't need to initialize tmp_out.sack_ok as we don't use the results
90 */
91enum tcp_tw_status
92tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb,
93 const struct tcphdr *th)
94{
95 struct tcp_options_received tmp_opt;
96 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
97 bool paws_reject = false;
98
99 tmp_opt.saw_tstamp = 0;
100 if (th->doff > (sizeof(*th) >> 2) && tcptw->tw_ts_recent_stamp) {
101 tcp_parse_options(skb, &tmp_opt, 0, NULL);
102
103 if (tmp_opt.saw_tstamp) {
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);
108 }
109 }
110
111 if (tw->tw_substate == TCP_FIN_WAIT2) {
112 /* Just repeat all the checks of tcp_rcv_state_process() */
113
114 /* Out of window, send ACK */
115 if (paws_reject ||
116 !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
117 tcptw->tw_rcv_nxt,
118 tcptw->tw_rcv_nxt + tcptw->tw_rcv_wnd))
119 return TCP_TW_ACK;
120
121 if (th->rst)
122 goto kill;
123
124 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt))
125 goto kill_with_rst;
126
127 /* Dup ACK? */
128 if (!th->ack ||
129 !after(TCP_SKB_CB(skb)->end_seq, tcptw->tw_rcv_nxt) ||
130 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) {
131 inet_twsk_put(tw);
132 return TCP_TW_SUCCESS;
133 }
134
135 /* New data or FIN. If new data arrive after half-duplex close,
136 * reset.
137 */
138 if (!th->fin ||
139 TCP_SKB_CB(skb)->end_seq != tcptw->tw_rcv_nxt + 1) {
140kill_with_rst:
141 inet_twsk_deschedule(tw, &tcp_death_row);
142 inet_twsk_put(tw);
143 return TCP_TW_RST;
144 }
145
146 /* FIN arrived, enter true time-wait state. */
147 tw->tw_substate = TCP_TIME_WAIT;
148 tcptw->tw_rcv_nxt = TCP_SKB_CB(skb)->end_seq;
149 if (tmp_opt.saw_tstamp) {
150 tcptw->tw_ts_recent_stamp = get_seconds();
151 tcptw->tw_ts_recent = tmp_opt.rcv_tsval;
152 }
153
154 if (tcp_death_row.sysctl_tw_recycle &&
155 tcptw->tw_ts_recent_stamp &&
156 tcp_tw_remember_stamp(tw))
157 inet_twsk_schedule(tw, &tcp_death_row, tw->tw_timeout,
158 TCP_TIMEWAIT_LEN);
159 else
160 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN,
161 TCP_TIMEWAIT_LEN);
162 return TCP_TW_ACK;
163 }
164
165 /*
166 * Now real TIME-WAIT state.
167 *
168 * RFC 1122:
169 * "When a connection is [...] on TIME-WAIT state [...]
170 * [a TCP] MAY accept a new SYN from the remote TCP to
171 * reopen the connection directly, if it:
172 *
173 * (1) assigns its initial sequence number for the new
174 * connection to be larger than the largest sequence
175 * number it used on the previous connection incarnation,
176 * and
177 *
178 * (2) returns to TIME-WAIT state if the SYN turns out
179 * to be an old duplicate".
180 */
181
182 if (!paws_reject &&
183 (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt &&
184 (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) {
185 /* In window segment, it may be only reset or bare ack. */
186
187 if (th->rst) {
188 /* This is TIME_WAIT assassination, in two flavors.
189 * Oh well... nobody has a sufficient solution to this
190 * protocol bug yet.
191 */
192 if (sysctl_tcp_rfc1337 == 0) {
193kill:
194 inet_twsk_deschedule(tw, &tcp_death_row);
195 inet_twsk_put(tw);
196 return TCP_TW_SUCCESS;
197 }
198 }
199 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN,
200 TCP_TIMEWAIT_LEN);
201
202 if (tmp_opt.saw_tstamp) {
203 tcptw->tw_ts_recent = tmp_opt.rcv_tsval;
204 tcptw->tw_ts_recent_stamp = get_seconds();
205 }
206
207 inet_twsk_put(tw);
208 return TCP_TW_SUCCESS;
209 }
210
211 /* Out of window segment.
212
213 All the segments are ACKed immediately.
214
215 The only exception is new SYN. We accept it, if it is
216 not old duplicate and we are not in danger to be killed
217 by delayed old duplicates. RFC check is that it has
218 newer sequence number works at rates <40Mbit/sec.
219 However, if paws works, it is reliable AND even more,
220 we even may relax silly seq space cutoff.
221
222 RED-PEN: we violate main RFC requirement, if this SYN will appear
223 old duplicate (i.e. we receive RST in reply to SYN-ACK),
224 we must return socket to time-wait state. It is not good,
225 but not fatal yet.
226 */
227
228 if (th->syn && !th->rst && !th->ack && !paws_reject &&
229 (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) ||
230 (tmp_opt.saw_tstamp &&
231 (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) {
232 u32 isn = tcptw->tw_snd_nxt + 65535 + 2;
233 if (isn == 0)
234 isn++;
235 TCP_SKB_CB(skb)->when = isn;
236 return TCP_TW_SYN;
237 }
238
239 if (paws_reject)
240 NET_INC_STATS_BH(twsk_net(tw), LINUX_MIB_PAWSESTABREJECTED);
241
242 if (!th->rst) {
243 /* In this case we must reset the TIMEWAIT timer.
244 *
245 * If it is ACKless SYN it may be both old duplicate
246 * and new good SYN with random sequence number <rcv_nxt.
247 * Do not reschedule in the last case.
248 */
249 if (paws_reject || th->ack)
250 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN,
251 TCP_TIMEWAIT_LEN);
252
253 /* Send ACK. Note, we do not put the bucket,
254 * it will be released by caller.
255 */
256 return TCP_TW_ACK;
257 }
258 inet_twsk_put(tw);
259 return TCP_TW_SUCCESS;
260}
261EXPORT_SYMBOL(tcp_timewait_state_process);
262
263/*
264 * Move a socket to time-wait or dead fin-wait-2 state.
265 */
266void tcp_time_wait(struct sock *sk, int state, int timeo)
267{
268 struct inet_timewait_sock *tw = NULL;
269 const struct inet_connection_sock *icsk = inet_csk(sk);
270 const struct tcp_sock *tp = tcp_sk(sk);
271 bool recycle_ok = false;
272
273 if (tcp_death_row.sysctl_tw_recycle && tp->rx_opt.ts_recent_stamp)
274 recycle_ok = tcp_remember_stamp(sk);
275
276 if (tcp_death_row.tw_count < tcp_death_row.sysctl_max_tw_buckets)
277 tw = inet_twsk_alloc(sk, state);
278
279 if (tw != NULL) {
280 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
281 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1);
282 struct inet_sock *inet = inet_sk(sk);
283
284 tw->tw_transparent = inet->transparent;
285 tw->tw_rcv_wscale = tp->rx_opt.rcv_wscale;
286 tcptw->tw_rcv_nxt = tp->rcv_nxt;
287 tcptw->tw_snd_nxt = tp->snd_nxt;
288 tcptw->tw_rcv_wnd = tcp_receive_window(tp);
289 tcptw->tw_ts_recent = tp->rx_opt.ts_recent;
290 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp;
291 tcptw->tw_ts_offset = tp->tsoffset;
292
293#if IS_ENABLED(CONFIG_IPV6)
294 if (tw->tw_family == PF_INET6) {
295 struct ipv6_pinfo *np = inet6_sk(sk);
296
297 tw->tw_v6_daddr = sk->sk_v6_daddr;
298 tw->tw_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
299 tw->tw_tclass = np->tclass;
300 tw->tw_flowlabel = np->flow_label >> 12;
301 tw->tw_ipv6only = np->ipv6only;
302 }
303#endif
304
305#ifdef CONFIG_TCP_MD5SIG
306 /*
307 * The timewait bucket does not have the key DB from the
308 * sock structure. We just make a quick copy of the
309 * md5 key being used (if indeed we are using one)
310 * so the timewait ack generating code has the key.
311 */
312 do {
313 struct tcp_md5sig_key *key;
314 tcptw->tw_md5_key = NULL;
315 key = tp->af_specific->md5_lookup(sk, sk);
316 if (key != NULL) {
317 tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC);
318 if (tcptw->tw_md5_key && !tcp_alloc_md5sig_pool())
319 BUG();
320 }
321 } while (0);
322#endif
323
324 /* Linkage updates. */
325 __inet_twsk_hashdance(tw, sk, &tcp_hashinfo);
326
327 /* Get the TIME_WAIT timeout firing. */
328 if (timeo < rto)
329 timeo = rto;
330
331 if (recycle_ok) {
332 tw->tw_timeout = rto;
333 } else {
334 tw->tw_timeout = TCP_TIMEWAIT_LEN;
335 if (state == TCP_TIME_WAIT)
336 timeo = TCP_TIMEWAIT_LEN;
337 }
338
339 inet_twsk_schedule(tw, &tcp_death_row, timeo,
340 TCP_TIMEWAIT_LEN);
341 inet_twsk_put(tw);
342 } else {
343 /* Sorry, if we're out of memory, just CLOSE this
344 * socket up. We've got bigger problems than
345 * non-graceful socket closings.
346 */
347 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPTIMEWAITOVERFLOW);
348 }
349
350 tcp_update_metrics(sk);
351 tcp_done(sk);
352}
353
354void tcp_twsk_destructor(struct sock *sk)
355{
356#ifdef CONFIG_TCP_MD5SIG
357 struct tcp_timewait_sock *twsk = tcp_twsk(sk);
358
359 if (twsk->tw_md5_key)
360 kfree_rcu(twsk->tw_md5_key, rcu);
361#endif
362}
363EXPORT_SYMBOL_GPL(tcp_twsk_destructor);
364
365static inline void TCP_ECN_openreq_child(struct tcp_sock *tp,
366 struct request_sock *req)
367{
368 tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0;
369}
370
371/* This is not only more efficient than what we used to do, it eliminates
372 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM
373 *
374 * Actually, we could lots of memory writes here. tp of listening
375 * socket contains all necessary default parameters.
376 */
377struct sock *tcp_create_openreq_child(struct sock *sk, struct request_sock *req, struct sk_buff *skb)
378{
379 struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC);
380
381 if (newsk != NULL) {
382 const struct inet_request_sock *ireq = inet_rsk(req);
383 struct tcp_request_sock *treq = tcp_rsk(req);
384 struct inet_connection_sock *newicsk = inet_csk(newsk);
385 struct tcp_sock *newtp = tcp_sk(newsk);
386
387 /* Now setup tcp_sock */
388 newtp->pred_flags = 0;
389
390 newtp->rcv_wup = newtp->copied_seq =
391 newtp->rcv_nxt = treq->rcv_isn + 1;
392
393 newtp->snd_sml = newtp->snd_una =
394 newtp->snd_nxt = newtp->snd_up = treq->snt_isn + 1;
395
396 tcp_prequeue_init(newtp);
397 INIT_LIST_HEAD(&newtp->tsq_node);
398
399 tcp_init_wl(newtp, treq->rcv_isn);
400
401 newtp->srtt_us = 0;
402 newtp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
403 newicsk->icsk_rto = TCP_TIMEOUT_INIT;
404
405 newtp->packets_out = 0;
406 newtp->retrans_out = 0;
407 newtp->sacked_out = 0;
408 newtp->fackets_out = 0;
409 newtp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
410 tcp_enable_early_retrans(newtp);
411 newtp->tlp_high_seq = 0;
412 newtp->lsndtime = treq->snt_synack;
413 newtp->total_retrans = req->num_retrans;
414
415 /* So many TCP implementations out there (incorrectly) count the
416 * initial SYN frame in their delayed-ACK and congestion control
417 * algorithms that we must have the following bandaid to talk
418 * efficiently to them. -DaveM
419 */
420 newtp->snd_cwnd = TCP_INIT_CWND;
421 newtp->snd_cwnd_cnt = 0;
422
423 if (newicsk->icsk_ca_ops != &tcp_init_congestion_ops &&
424 !try_module_get(newicsk->icsk_ca_ops->owner))
425 newicsk->icsk_ca_ops = &tcp_init_congestion_ops;
426
427 tcp_set_ca_state(newsk, TCP_CA_Open);
428 tcp_init_xmit_timers(newsk);
429 __skb_queue_head_init(&newtp->out_of_order_queue);
430 newtp->write_seq = newtp->pushed_seq = treq->snt_isn + 1;
431
432 newtp->rx_opt.saw_tstamp = 0;
433
434 newtp->rx_opt.dsack = 0;
435 newtp->rx_opt.num_sacks = 0;
436
437 newtp->urg_data = 0;
438
439 if (sock_flag(newsk, SOCK_KEEPOPEN))
440 inet_csk_reset_keepalive_timer(newsk,
441 keepalive_time_when(newtp));
442
443 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok;
444 if ((newtp->rx_opt.sack_ok = ireq->sack_ok) != 0) {
445 if (sysctl_tcp_fack)
446 tcp_enable_fack(newtp);
447 }
448 newtp->window_clamp = req->window_clamp;
449 newtp->rcv_ssthresh = req->rcv_wnd;
450 newtp->rcv_wnd = req->rcv_wnd;
451 newtp->rx_opt.wscale_ok = ireq->wscale_ok;
452 if (newtp->rx_opt.wscale_ok) {
453 newtp->rx_opt.snd_wscale = ireq->snd_wscale;
454 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale;
455 } else {
456 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0;
457 newtp->window_clamp = min(newtp->window_clamp, 65535U);
458 }
459 newtp->snd_wnd = (ntohs(tcp_hdr(skb)->window) <<
460 newtp->rx_opt.snd_wscale);
461 newtp->max_window = newtp->snd_wnd;
462
463 if (newtp->rx_opt.tstamp_ok) {
464 newtp->rx_opt.ts_recent = req->ts_recent;
465 newtp->rx_opt.ts_recent_stamp = get_seconds();
466 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
467 } else {
468 newtp->rx_opt.ts_recent_stamp = 0;
469 newtp->tcp_header_len = sizeof(struct tcphdr);
470 }
471 newtp->tsoffset = 0;
472#ifdef CONFIG_TCP_MD5SIG
473 newtp->md5sig_info = NULL; /*XXX*/
474 if (newtp->af_specific->md5_lookup(sk, newsk))
475 newtp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
476#endif
477 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len)
478 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len;
479 newtp->rx_opt.mss_clamp = req->mss;
480 TCP_ECN_openreq_child(newtp, req);
481 newtp->fastopen_rsk = NULL;
482 newtp->syn_data_acked = 0;
483
484 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_PASSIVEOPENS);
485 }
486 return newsk;
487}
488EXPORT_SYMBOL(tcp_create_openreq_child);
489
490/*
491 * Process an incoming packet for SYN_RECV sockets represented as a
492 * request_sock. Normally sk is the listener socket but for TFO it
493 * points to the child socket.
494 *
495 * XXX (TFO) - The current impl contains a special check for ack
496 * validation and inside tcp_v4_reqsk_send_ack(). Can we do better?
497 *
498 * We don't need to initialize tmp_opt.sack_ok as we don't use the results
499 */
500
501struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb,
502 struct request_sock *req,
503 struct request_sock **prev,
504 bool fastopen)
505{
506 struct tcp_options_received tmp_opt;
507 struct sock *child;
508 const struct tcphdr *th = tcp_hdr(skb);
509 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK);
510 bool paws_reject = false;
511
512 BUG_ON(fastopen == (sk->sk_state == TCP_LISTEN));
513
514 tmp_opt.saw_tstamp = 0;
515 if (th->doff > (sizeof(struct tcphdr)>>2)) {
516 tcp_parse_options(skb, &tmp_opt, 0, NULL);
517
518 if (tmp_opt.saw_tstamp) {
519 tmp_opt.ts_recent = req->ts_recent;
520 /* We do not store true stamp, but it is not required,
521 * it can be estimated (approximately)
522 * from another data.
523 */
524 tmp_opt.ts_recent_stamp = get_seconds() - ((TCP_TIMEOUT_INIT/HZ)<<req->num_timeout);
525 paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
526 }
527 }
528
529 /* Check for pure retransmitted SYN. */
530 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn &&
531 flg == TCP_FLAG_SYN &&
532 !paws_reject) {
533 /*
534 * RFC793 draws (Incorrectly! It was fixed in RFC1122)
535 * this case on figure 6 and figure 8, but formal
536 * protocol description says NOTHING.
537 * To be more exact, it says that we should send ACK,
538 * because this segment (at least, if it has no data)
539 * is out of window.
540 *
541 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT
542 * describe SYN-RECV state. All the description
543 * is wrong, we cannot believe to it and should
544 * rely only on common sense and implementation
545 * experience.
546 *
547 * Enforce "SYN-ACK" according to figure 8, figure 6
548 * of RFC793, fixed by RFC1122.
549 *
550 * Note that even if there is new data in the SYN packet
551 * they will be thrown away too.
552 *
553 * Reset timer after retransmitting SYNACK, similar to
554 * the idea of fast retransmit in recovery.
555 */
556 if (!inet_rtx_syn_ack(sk, req))
557 req->expires = min(TCP_TIMEOUT_INIT << req->num_timeout,
558 TCP_RTO_MAX) + jiffies;
559 return NULL;
560 }
561
562 /* Further reproduces section "SEGMENT ARRIVES"
563 for state SYN-RECEIVED of RFC793.
564 It is broken, however, it does not work only
565 when SYNs are crossed.
566
567 You would think that SYN crossing is impossible here, since
568 we should have a SYN_SENT socket (from connect()) on our end,
569 but this is not true if the crossed SYNs were sent to both
570 ends by a malicious third party. We must defend against this,
571 and to do that we first verify the ACK (as per RFC793, page
572 36) and reset if it is invalid. Is this a true full defense?
573 To convince ourselves, let us consider a way in which the ACK
574 test can still pass in this 'malicious crossed SYNs' case.
575 Malicious sender sends identical SYNs (and thus identical sequence
576 numbers) to both A and B:
577
578 A: gets SYN, seq=7
579 B: gets SYN, seq=7
580
581 By our good fortune, both A and B select the same initial
582 send sequence number of seven :-)
583
584 A: sends SYN|ACK, seq=7, ack_seq=8
585 B: sends SYN|ACK, seq=7, ack_seq=8
586
587 So we are now A eating this SYN|ACK, ACK test passes. So
588 does sequence test, SYN is truncated, and thus we consider
589 it a bare ACK.
590
591 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this
592 bare ACK. Otherwise, we create an established connection. Both
593 ends (listening sockets) accept the new incoming connection and try
594 to talk to each other. 8-)
595
596 Note: This case is both harmless, and rare. Possibility is about the
597 same as us discovering intelligent life on another plant tomorrow.
598
599 But generally, we should (RFC lies!) to accept ACK
600 from SYNACK both here and in tcp_rcv_state_process().
601 tcp_rcv_state_process() does not, hence, we do not too.
602
603 Note that the case is absolutely generic:
604 we cannot optimize anything here without
605 violating protocol. All the checks must be made
606 before attempt to create socket.
607 */
608
609 /* RFC793 page 36: "If the connection is in any non-synchronized state ...
610 * and the incoming segment acknowledges something not yet
611 * sent (the segment carries an unacceptable ACK) ...
612 * a reset is sent."
613 *
614 * Invalid ACK: reset will be sent by listening socket.
615 * Note that the ACK validity check for a Fast Open socket is done
616 * elsewhere and is checked directly against the child socket rather
617 * than req because user data may have been sent out.
618 */
619 if ((flg & TCP_FLAG_ACK) && !fastopen &&
620 (TCP_SKB_CB(skb)->ack_seq !=
621 tcp_rsk(req)->snt_isn + 1))
622 return sk;
623
624 /* Also, it would be not so bad idea to check rcv_tsecr, which
625 * is essentially ACK extension and too early or too late values
626 * should cause reset in unsynchronized states.
627 */
628
629 /* RFC793: "first check sequence number". */
630
631 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
632 tcp_rsk(req)->rcv_nxt, tcp_rsk(req)->rcv_nxt + req->rcv_wnd)) {
633 /* Out of window: send ACK and drop. */
634 if (!(flg & TCP_FLAG_RST))
635 req->rsk_ops->send_ack(sk, skb, req);
636 if (paws_reject)
637 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
638 return NULL;
639 }
640
641 /* In sequence, PAWS is OK. */
642
643 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_nxt))
644 req->ts_recent = tmp_opt.rcv_tsval;
645
646 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) {
647 /* Truncate SYN, it is out of window starting
648 at tcp_rsk(req)->rcv_isn + 1. */
649 flg &= ~TCP_FLAG_SYN;
650 }
651
652 /* RFC793: "second check the RST bit" and
653 * "fourth, check the SYN bit"
654 */
655 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) {
656 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
657 goto embryonic_reset;
658 }
659
660 /* ACK sequence verified above, just make sure ACK is
661 * set. If ACK not set, just silently drop the packet.
662 *
663 * XXX (TFO) - if we ever allow "data after SYN", the
664 * following check needs to be removed.
665 */
666 if (!(flg & TCP_FLAG_ACK))
667 return NULL;
668
669 /* For Fast Open no more processing is needed (sk is the
670 * child socket).
671 */
672 if (fastopen)
673 return sk;
674
675 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */
676 if (req->num_timeout < inet_csk(sk)->icsk_accept_queue.rskq_defer_accept &&
677 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) {
678 inet_rsk(req)->acked = 1;
679 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP);
680 return NULL;
681 }
682
683 /* OK, ACK is valid, create big socket and
684 * feed this segment to it. It will repeat all
685 * the tests. THIS SEGMENT MUST MOVE SOCKET TO
686 * ESTABLISHED STATE. If it will be dropped after
687 * socket is created, wait for troubles.
688 */
689 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL);
690 if (child == NULL)
691 goto listen_overflow;
692
693 inet_csk_reqsk_queue_unlink(sk, req, prev);
694 inet_csk_reqsk_queue_removed(sk, req);
695
696 inet_csk_reqsk_queue_add(sk, req, child);
697 return child;
698
699listen_overflow:
700 if (!sysctl_tcp_abort_on_overflow) {
701 inet_rsk(req)->acked = 1;
702 return NULL;
703 }
704
705embryonic_reset:
706 if (!(flg & TCP_FLAG_RST)) {
707 /* Received a bad SYN pkt - for TFO We try not to reset
708 * the local connection unless it's really necessary to
709 * avoid becoming vulnerable to outside attack aiming at
710 * resetting legit local connections.
711 */
712 req->rsk_ops->send_reset(sk, skb);
713 } else if (fastopen) { /* received a valid RST pkt */
714 reqsk_fastopen_remove(sk, req, true);
715 tcp_reset(sk);
716 }
717 if (!fastopen) {
718 inet_csk_reqsk_queue_drop(sk, req, prev);
719 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_EMBRYONICRSTS);
720 }
721 return NULL;
722}
723EXPORT_SYMBOL(tcp_check_req);
724
725/*
726 * Queue segment on the new socket if the new socket is active,
727 * otherwise we just shortcircuit this and continue with
728 * the new socket.
729 *
730 * For the vast majority of cases child->sk_state will be TCP_SYN_RECV
731 * when entering. But other states are possible due to a race condition
732 * where after __inet_lookup_established() fails but before the listener
733 * locked is obtained, other packets cause the same connection to
734 * be created.
735 */
736
737int tcp_child_process(struct sock *parent, struct sock *child,
738 struct sk_buff *skb)
739{
740 int ret = 0;
741 int state = child->sk_state;
742
743 if (!sock_owned_by_user(child)) {
744 ret = tcp_rcv_state_process(child, skb, tcp_hdr(skb),
745 skb->len);
746 /* Wakeup parent, send SIGIO */
747 if (state == TCP_SYN_RECV && child->sk_state != state)
748 parent->sk_data_ready(parent);
749 } else {
750 /* Alas, it is possible again, because we do lookup
751 * in main socket hash table and lock on listening
752 * socket does not protect us more.
753 */
754 __sk_add_backlog(child, skb);
755 }
756
757 bh_unlock_sock(child);
758 sock_put(child);
759 return ret;
760}
761EXPORT_SYMBOL(tcp_child_process);