Loading...
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
52/* VJ's idea. Save last timestamp seen from this destination
53 * and hold it at least for normal timewait interval to use for duplicate
54 * segment detection in subsequent connections, before they enter synchronized
55 * state.
56 */
57
58static bool tcp_remember_stamp(struct sock *sk)
59{
60 const struct inet_connection_sock *icsk = inet_csk(sk);
61 struct tcp_sock *tp = tcp_sk(sk);
62 struct inet_peer *peer;
63 bool release_it;
64
65 peer = icsk->icsk_af_ops->get_peer(sk, &release_it);
66 if (peer) {
67 if ((s32)(peer->tcp_ts - tp->rx_opt.ts_recent) <= 0 ||
68 ((u32)get_seconds() - peer->tcp_ts_stamp > TCP_PAWS_MSL &&
69 peer->tcp_ts_stamp <= (u32)tp->rx_opt.ts_recent_stamp)) {
70 peer->tcp_ts_stamp = (u32)tp->rx_opt.ts_recent_stamp;
71 peer->tcp_ts = tp->rx_opt.ts_recent;
72 }
73 if (release_it)
74 inet_putpeer(peer);
75 return true;
76 }
77
78 return false;
79}
80
81static bool tcp_tw_remember_stamp(struct inet_timewait_sock *tw)
82{
83 struct sock *sk = (struct sock *) tw;
84 struct inet_peer *peer;
85
86 peer = twsk_getpeer(sk);
87 if (peer) {
88 const struct tcp_timewait_sock *tcptw = tcp_twsk(sk);
89
90 if ((s32)(peer->tcp_ts - tcptw->tw_ts_recent) <= 0 ||
91 ((u32)get_seconds() - peer->tcp_ts_stamp > TCP_PAWS_MSL &&
92 peer->tcp_ts_stamp <= (u32)tcptw->tw_ts_recent_stamp)) {
93 peer->tcp_ts_stamp = (u32)tcptw->tw_ts_recent_stamp;
94 peer->tcp_ts = tcptw->tw_ts_recent;
95 }
96 inet_putpeer(peer);
97 return true;
98 }
99 return false;
100}
101
102static bool tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win)
103{
104 if (seq == s_win)
105 return true;
106 if (after(end_seq, s_win) && before(seq, e_win))
107 return true;
108 return seq == e_win && seq == end_seq;
109}
110
111/*
112 * * Main purpose of TIME-WAIT state is to close connection gracefully,
113 * when one of ends sits in LAST-ACK or CLOSING retransmitting FIN
114 * (and, probably, tail of data) and one or more our ACKs are lost.
115 * * What is TIME-WAIT timeout? It is associated with maximal packet
116 * lifetime in the internet, which results in wrong conclusion, that
117 * it is set to catch "old duplicate segments" wandering out of their path.
118 * It is not quite correct. This timeout is calculated so that it exceeds
119 * maximal retransmission timeout enough to allow to lose one (or more)
120 * segments sent by peer and our ACKs. This time may be calculated from RTO.
121 * * When TIME-WAIT socket receives RST, it means that another end
122 * finally closed and we are allowed to kill TIME-WAIT too.
123 * * Second purpose of TIME-WAIT is catching old duplicate segments.
124 * Well, certainly it is pure paranoia, but if we load TIME-WAIT
125 * with this semantics, we MUST NOT kill TIME-WAIT state with RSTs.
126 * * If we invented some more clever way to catch duplicates
127 * (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs.
128 *
129 * The algorithm below is based on FORMAL INTERPRETATION of RFCs.
130 * When you compare it to RFCs, please, read section SEGMENT ARRIVES
131 * from the very beginning.
132 *
133 * NOTE. With recycling (and later with fin-wait-2) TW bucket
134 * is _not_ stateless. It means, that strictly speaking we must
135 * spinlock it. I do not want! Well, probability of misbehaviour
136 * is ridiculously low and, seems, we could use some mb() tricks
137 * to avoid misread sequence numbers, states etc. --ANK
138 */
139enum tcp_tw_status
140tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb,
141 const struct tcphdr *th)
142{
143 struct tcp_options_received tmp_opt;
144 const u8 *hash_location;
145 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
146 bool paws_reject = false;
147
148 tmp_opt.saw_tstamp = 0;
149 if (th->doff > (sizeof(*th) >> 2) && tcptw->tw_ts_recent_stamp) {
150 tcp_parse_options(skb, &tmp_opt, &hash_location, 0);
151
152 if (tmp_opt.saw_tstamp) {
153 tmp_opt.ts_recent = tcptw->tw_ts_recent;
154 tmp_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp;
155 paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
156 }
157 }
158
159 if (tw->tw_substate == TCP_FIN_WAIT2) {
160 /* Just repeat all the checks of tcp_rcv_state_process() */
161
162 /* Out of window, send ACK */
163 if (paws_reject ||
164 !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
165 tcptw->tw_rcv_nxt,
166 tcptw->tw_rcv_nxt + tcptw->tw_rcv_wnd))
167 return TCP_TW_ACK;
168
169 if (th->rst)
170 goto kill;
171
172 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt))
173 goto kill_with_rst;
174
175 /* Dup ACK? */
176 if (!th->ack ||
177 !after(TCP_SKB_CB(skb)->end_seq, tcptw->tw_rcv_nxt) ||
178 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) {
179 inet_twsk_put(tw);
180 return TCP_TW_SUCCESS;
181 }
182
183 /* New data or FIN. If new data arrive after half-duplex close,
184 * reset.
185 */
186 if (!th->fin ||
187 TCP_SKB_CB(skb)->end_seq != tcptw->tw_rcv_nxt + 1) {
188kill_with_rst:
189 inet_twsk_deschedule(tw, &tcp_death_row);
190 inet_twsk_put(tw);
191 return TCP_TW_RST;
192 }
193
194 /* FIN arrived, enter true time-wait state. */
195 tw->tw_substate = TCP_TIME_WAIT;
196 tcptw->tw_rcv_nxt = TCP_SKB_CB(skb)->end_seq;
197 if (tmp_opt.saw_tstamp) {
198 tcptw->tw_ts_recent_stamp = get_seconds();
199 tcptw->tw_ts_recent = tmp_opt.rcv_tsval;
200 }
201
202 if (tcp_death_row.sysctl_tw_recycle &&
203 tcptw->tw_ts_recent_stamp &&
204 tcp_tw_remember_stamp(tw))
205 inet_twsk_schedule(tw, &tcp_death_row, tw->tw_timeout,
206 TCP_TIMEWAIT_LEN);
207 else
208 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN,
209 TCP_TIMEWAIT_LEN);
210 return TCP_TW_ACK;
211 }
212
213 /*
214 * Now real TIME-WAIT state.
215 *
216 * RFC 1122:
217 * "When a connection is [...] on TIME-WAIT state [...]
218 * [a TCP] MAY accept a new SYN from the remote TCP to
219 * reopen the connection directly, if it:
220 *
221 * (1) assigns its initial sequence number for the new
222 * connection to be larger than the largest sequence
223 * number it used on the previous connection incarnation,
224 * and
225 *
226 * (2) returns to TIME-WAIT state if the SYN turns out
227 * to be an old duplicate".
228 */
229
230 if (!paws_reject &&
231 (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt &&
232 (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) {
233 /* In window segment, it may be only reset or bare ack. */
234
235 if (th->rst) {
236 /* This is TIME_WAIT assassination, in two flavors.
237 * Oh well... nobody has a sufficient solution to this
238 * protocol bug yet.
239 */
240 if (sysctl_tcp_rfc1337 == 0) {
241kill:
242 inet_twsk_deschedule(tw, &tcp_death_row);
243 inet_twsk_put(tw);
244 return TCP_TW_SUCCESS;
245 }
246 }
247 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN,
248 TCP_TIMEWAIT_LEN);
249
250 if (tmp_opt.saw_tstamp) {
251 tcptw->tw_ts_recent = tmp_opt.rcv_tsval;
252 tcptw->tw_ts_recent_stamp = get_seconds();
253 }
254
255 inet_twsk_put(tw);
256 return TCP_TW_SUCCESS;
257 }
258
259 /* Out of window segment.
260
261 All the segments are ACKed immediately.
262
263 The only exception is new SYN. We accept it, if it is
264 not old duplicate and we are not in danger to be killed
265 by delayed old duplicates. RFC check is that it has
266 newer sequence number works at rates <40Mbit/sec.
267 However, if paws works, it is reliable AND even more,
268 we even may relax silly seq space cutoff.
269
270 RED-PEN: we violate main RFC requirement, if this SYN will appear
271 old duplicate (i.e. we receive RST in reply to SYN-ACK),
272 we must return socket to time-wait state. It is not good,
273 but not fatal yet.
274 */
275
276 if (th->syn && !th->rst && !th->ack && !paws_reject &&
277 (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) ||
278 (tmp_opt.saw_tstamp &&
279 (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) {
280 u32 isn = tcptw->tw_snd_nxt + 65535 + 2;
281 if (isn == 0)
282 isn++;
283 TCP_SKB_CB(skb)->when = isn;
284 return TCP_TW_SYN;
285 }
286
287 if (paws_reject)
288 NET_INC_STATS_BH(twsk_net(tw), LINUX_MIB_PAWSESTABREJECTED);
289
290 if (!th->rst) {
291 /* In this case we must reset the TIMEWAIT timer.
292 *
293 * If it is ACKless SYN it may be both old duplicate
294 * and new good SYN with random sequence number <rcv_nxt.
295 * Do not reschedule in the last case.
296 */
297 if (paws_reject || th->ack)
298 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN,
299 TCP_TIMEWAIT_LEN);
300
301 /* Send ACK. Note, we do not put the bucket,
302 * it will be released by caller.
303 */
304 return TCP_TW_ACK;
305 }
306 inet_twsk_put(tw);
307 return TCP_TW_SUCCESS;
308}
309EXPORT_SYMBOL(tcp_timewait_state_process);
310
311/*
312 * Move a socket to time-wait or dead fin-wait-2 state.
313 */
314void tcp_time_wait(struct sock *sk, int state, int timeo)
315{
316 struct inet_timewait_sock *tw = NULL;
317 const struct inet_connection_sock *icsk = inet_csk(sk);
318 const struct tcp_sock *tp = tcp_sk(sk);
319 bool recycle_ok = false;
320
321 if (tcp_death_row.sysctl_tw_recycle && tp->rx_opt.ts_recent_stamp)
322 recycle_ok = tcp_remember_stamp(sk);
323
324 if (tcp_death_row.tw_count < tcp_death_row.sysctl_max_tw_buckets)
325 tw = inet_twsk_alloc(sk, state);
326
327 if (tw != NULL) {
328 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
329 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1);
330
331 tw->tw_transparent = inet_sk(sk)->transparent;
332 tw->tw_rcv_wscale = tp->rx_opt.rcv_wscale;
333 tcptw->tw_rcv_nxt = tp->rcv_nxt;
334 tcptw->tw_snd_nxt = tp->snd_nxt;
335 tcptw->tw_rcv_wnd = tcp_receive_window(tp);
336 tcptw->tw_ts_recent = tp->rx_opt.ts_recent;
337 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp;
338
339#if IS_ENABLED(CONFIG_IPV6)
340 if (tw->tw_family == PF_INET6) {
341 struct ipv6_pinfo *np = inet6_sk(sk);
342 struct inet6_timewait_sock *tw6;
343
344 tw->tw_ipv6_offset = inet6_tw_offset(sk->sk_prot);
345 tw6 = inet6_twsk((struct sock *)tw);
346 tw6->tw_v6_daddr = np->daddr;
347 tw6->tw_v6_rcv_saddr = np->rcv_saddr;
348 tw->tw_tclass = np->tclass;
349 tw->tw_ipv6only = np->ipv6only;
350 }
351#endif
352
353#ifdef CONFIG_TCP_MD5SIG
354 /*
355 * The timewait bucket does not have the key DB from the
356 * sock structure. We just make a quick copy of the
357 * md5 key being used (if indeed we are using one)
358 * so the timewait ack generating code has the key.
359 */
360 do {
361 struct tcp_md5sig_key *key;
362 tcptw->tw_md5_key = NULL;
363 key = tp->af_specific->md5_lookup(sk, sk);
364 if (key != NULL) {
365 tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC);
366 if (tcptw->tw_md5_key && tcp_alloc_md5sig_pool(sk) == NULL)
367 BUG();
368 }
369 } while (0);
370#endif
371
372 /* Linkage updates. */
373 __inet_twsk_hashdance(tw, sk, &tcp_hashinfo);
374
375 /* Get the TIME_WAIT timeout firing. */
376 if (timeo < rto)
377 timeo = rto;
378
379 if (recycle_ok) {
380 tw->tw_timeout = rto;
381 } else {
382 tw->tw_timeout = TCP_TIMEWAIT_LEN;
383 if (state == TCP_TIME_WAIT)
384 timeo = TCP_TIMEWAIT_LEN;
385 }
386
387 inet_twsk_schedule(tw, &tcp_death_row, timeo,
388 TCP_TIMEWAIT_LEN);
389 inet_twsk_put(tw);
390 } else {
391 /* Sorry, if we're out of memory, just CLOSE this
392 * socket up. We've got bigger problems than
393 * non-graceful socket closings.
394 */
395 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPTIMEWAITOVERFLOW);
396 }
397
398 tcp_update_metrics(sk);
399 tcp_done(sk);
400}
401
402void tcp_twsk_destructor(struct sock *sk)
403{
404#ifdef CONFIG_TCP_MD5SIG
405 struct tcp_timewait_sock *twsk = tcp_twsk(sk);
406 if (twsk->tw_md5_key) {
407 tcp_free_md5sig_pool();
408 kfree_rcu(twsk->tw_md5_key, rcu);
409 }
410#endif
411}
412EXPORT_SYMBOL_GPL(tcp_twsk_destructor);
413
414static inline void TCP_ECN_openreq_child(struct tcp_sock *tp,
415 struct request_sock *req)
416{
417 tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0;
418}
419
420/* This is not only more efficient than what we used to do, it eliminates
421 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM
422 *
423 * Actually, we could lots of memory writes here. tp of listening
424 * socket contains all necessary default parameters.
425 */
426struct sock *tcp_create_openreq_child(struct sock *sk, struct request_sock *req, struct sk_buff *skb)
427{
428 struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC);
429
430 if (newsk != NULL) {
431 const struct inet_request_sock *ireq = inet_rsk(req);
432 struct tcp_request_sock *treq = tcp_rsk(req);
433 struct inet_connection_sock *newicsk = inet_csk(newsk);
434 struct tcp_sock *newtp = tcp_sk(newsk);
435 struct tcp_sock *oldtp = tcp_sk(sk);
436 struct tcp_cookie_values *oldcvp = oldtp->cookie_values;
437
438 /* TCP Cookie Transactions require space for the cookie pair,
439 * as it differs for each connection. There is no need to
440 * copy any s_data_payload stored at the original socket.
441 * Failure will prevent resuming the connection.
442 *
443 * Presumed copied, in order of appearance:
444 * cookie_in_always, cookie_out_never
445 */
446 if (oldcvp != NULL) {
447 struct tcp_cookie_values *newcvp =
448 kzalloc(sizeof(*newtp->cookie_values),
449 GFP_ATOMIC);
450
451 if (newcvp != NULL) {
452 kref_init(&newcvp->kref);
453 newcvp->cookie_desired =
454 oldcvp->cookie_desired;
455 newtp->cookie_values = newcvp;
456 } else {
457 /* Not Yet Implemented */
458 newtp->cookie_values = NULL;
459 }
460 }
461
462 /* Now setup tcp_sock */
463 newtp->pred_flags = 0;
464
465 newtp->rcv_wup = newtp->copied_seq =
466 newtp->rcv_nxt = treq->rcv_isn + 1;
467
468 newtp->snd_sml = newtp->snd_una =
469 newtp->snd_nxt = newtp->snd_up =
470 treq->snt_isn + 1 + tcp_s_data_size(oldtp);
471
472 tcp_prequeue_init(newtp);
473
474 tcp_init_wl(newtp, treq->rcv_isn);
475
476 newtp->srtt = 0;
477 newtp->mdev = TCP_TIMEOUT_INIT;
478 newicsk->icsk_rto = TCP_TIMEOUT_INIT;
479
480 newtp->packets_out = 0;
481 newtp->retrans_out = 0;
482 newtp->sacked_out = 0;
483 newtp->fackets_out = 0;
484 newtp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
485 tcp_enable_early_retrans(newtp);
486
487 /* So many TCP implementations out there (incorrectly) count the
488 * initial SYN frame in their delayed-ACK and congestion control
489 * algorithms that we must have the following bandaid to talk
490 * efficiently to them. -DaveM
491 */
492 newtp->snd_cwnd = TCP_INIT_CWND;
493 newtp->snd_cwnd_cnt = 0;
494 newtp->bytes_acked = 0;
495
496 newtp->frto_counter = 0;
497 newtp->frto_highmark = 0;
498
499 if (newicsk->icsk_ca_ops != &tcp_init_congestion_ops &&
500 !try_module_get(newicsk->icsk_ca_ops->owner))
501 newicsk->icsk_ca_ops = &tcp_init_congestion_ops;
502
503 tcp_set_ca_state(newsk, TCP_CA_Open);
504 tcp_init_xmit_timers(newsk);
505 skb_queue_head_init(&newtp->out_of_order_queue);
506 newtp->write_seq = newtp->pushed_seq =
507 treq->snt_isn + 1 + tcp_s_data_size(oldtp);
508
509 newtp->rx_opt.saw_tstamp = 0;
510
511 newtp->rx_opt.dsack = 0;
512 newtp->rx_opt.num_sacks = 0;
513
514 newtp->urg_data = 0;
515
516 if (sock_flag(newsk, SOCK_KEEPOPEN))
517 inet_csk_reset_keepalive_timer(newsk,
518 keepalive_time_when(newtp));
519
520 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok;
521 if ((newtp->rx_opt.sack_ok = ireq->sack_ok) != 0) {
522 if (sysctl_tcp_fack)
523 tcp_enable_fack(newtp);
524 }
525 newtp->window_clamp = req->window_clamp;
526 newtp->rcv_ssthresh = req->rcv_wnd;
527 newtp->rcv_wnd = req->rcv_wnd;
528 newtp->rx_opt.wscale_ok = ireq->wscale_ok;
529 if (newtp->rx_opt.wscale_ok) {
530 newtp->rx_opt.snd_wscale = ireq->snd_wscale;
531 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale;
532 } else {
533 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0;
534 newtp->window_clamp = min(newtp->window_clamp, 65535U);
535 }
536 newtp->snd_wnd = (ntohs(tcp_hdr(skb)->window) <<
537 newtp->rx_opt.snd_wscale);
538 newtp->max_window = newtp->snd_wnd;
539
540 if (newtp->rx_opt.tstamp_ok) {
541 newtp->rx_opt.ts_recent = req->ts_recent;
542 newtp->rx_opt.ts_recent_stamp = get_seconds();
543 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
544 } else {
545 newtp->rx_opt.ts_recent_stamp = 0;
546 newtp->tcp_header_len = sizeof(struct tcphdr);
547 }
548#ifdef CONFIG_TCP_MD5SIG
549 newtp->md5sig_info = NULL; /*XXX*/
550 if (newtp->af_specific->md5_lookup(sk, newsk))
551 newtp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
552#endif
553 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len)
554 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len;
555 newtp->rx_opt.mss_clamp = req->mss;
556 TCP_ECN_openreq_child(newtp, req);
557
558 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_PASSIVEOPENS);
559 }
560 return newsk;
561}
562EXPORT_SYMBOL(tcp_create_openreq_child);
563
564/*
565 * Process an incoming packet for SYN_RECV sockets represented
566 * as a request_sock.
567 */
568
569struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb,
570 struct request_sock *req,
571 struct request_sock **prev)
572{
573 struct tcp_options_received tmp_opt;
574 const u8 *hash_location;
575 struct sock *child;
576 const struct tcphdr *th = tcp_hdr(skb);
577 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK);
578 bool paws_reject = false;
579
580 tmp_opt.saw_tstamp = 0;
581 if (th->doff > (sizeof(struct tcphdr)>>2)) {
582 tcp_parse_options(skb, &tmp_opt, &hash_location, 0);
583
584 if (tmp_opt.saw_tstamp) {
585 tmp_opt.ts_recent = req->ts_recent;
586 /* We do not store true stamp, but it is not required,
587 * it can be estimated (approximately)
588 * from another data.
589 */
590 tmp_opt.ts_recent_stamp = get_seconds() - ((TCP_TIMEOUT_INIT/HZ)<<req->retrans);
591 paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
592 }
593 }
594
595 /* Check for pure retransmitted SYN. */
596 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn &&
597 flg == TCP_FLAG_SYN &&
598 !paws_reject) {
599 /*
600 * RFC793 draws (Incorrectly! It was fixed in RFC1122)
601 * this case on figure 6 and figure 8, but formal
602 * protocol description says NOTHING.
603 * To be more exact, it says that we should send ACK,
604 * because this segment (at least, if it has no data)
605 * is out of window.
606 *
607 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT
608 * describe SYN-RECV state. All the description
609 * is wrong, we cannot believe to it and should
610 * rely only on common sense and implementation
611 * experience.
612 *
613 * Enforce "SYN-ACK" according to figure 8, figure 6
614 * of RFC793, fixed by RFC1122.
615 */
616 req->rsk_ops->rtx_syn_ack(sk, req, NULL);
617 return NULL;
618 }
619
620 /* Further reproduces section "SEGMENT ARRIVES"
621 for state SYN-RECEIVED of RFC793.
622 It is broken, however, it does not work only
623 when SYNs are crossed.
624
625 You would think that SYN crossing is impossible here, since
626 we should have a SYN_SENT socket (from connect()) on our end,
627 but this is not true if the crossed SYNs were sent to both
628 ends by a malicious third party. We must defend against this,
629 and to do that we first verify the ACK (as per RFC793, page
630 36) and reset if it is invalid. Is this a true full defense?
631 To convince ourselves, let us consider a way in which the ACK
632 test can still pass in this 'malicious crossed SYNs' case.
633 Malicious sender sends identical SYNs (and thus identical sequence
634 numbers) to both A and B:
635
636 A: gets SYN, seq=7
637 B: gets SYN, seq=7
638
639 By our good fortune, both A and B select the same initial
640 send sequence number of seven :-)
641
642 A: sends SYN|ACK, seq=7, ack_seq=8
643 B: sends SYN|ACK, seq=7, ack_seq=8
644
645 So we are now A eating this SYN|ACK, ACK test passes. So
646 does sequence test, SYN is truncated, and thus we consider
647 it a bare ACK.
648
649 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this
650 bare ACK. Otherwise, we create an established connection. Both
651 ends (listening sockets) accept the new incoming connection and try
652 to talk to each other. 8-)
653
654 Note: This case is both harmless, and rare. Possibility is about the
655 same as us discovering intelligent life on another plant tomorrow.
656
657 But generally, we should (RFC lies!) to accept ACK
658 from SYNACK both here and in tcp_rcv_state_process().
659 tcp_rcv_state_process() does not, hence, we do not too.
660
661 Note that the case is absolutely generic:
662 we cannot optimize anything here without
663 violating protocol. All the checks must be made
664 before attempt to create socket.
665 */
666
667 /* RFC793 page 36: "If the connection is in any non-synchronized state ...
668 * and the incoming segment acknowledges something not yet
669 * sent (the segment carries an unacceptable ACK) ...
670 * a reset is sent."
671 *
672 * Invalid ACK: reset will be sent by listening socket
673 */
674 if ((flg & TCP_FLAG_ACK) &&
675 (TCP_SKB_CB(skb)->ack_seq !=
676 tcp_rsk(req)->snt_isn + 1 + tcp_s_data_size(tcp_sk(sk))))
677 return sk;
678
679 /* Also, it would be not so bad idea to check rcv_tsecr, which
680 * is essentially ACK extension and too early or too late values
681 * should cause reset in unsynchronized states.
682 */
683
684 /* RFC793: "first check sequence number". */
685
686 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
687 tcp_rsk(req)->rcv_isn + 1, tcp_rsk(req)->rcv_isn + 1 + req->rcv_wnd)) {
688 /* Out of window: send ACK and drop. */
689 if (!(flg & TCP_FLAG_RST))
690 req->rsk_ops->send_ack(sk, skb, req);
691 if (paws_reject)
692 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
693 return NULL;
694 }
695
696 /* In sequence, PAWS is OK. */
697
698 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_isn + 1))
699 req->ts_recent = tmp_opt.rcv_tsval;
700
701 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) {
702 /* Truncate SYN, it is out of window starting
703 at tcp_rsk(req)->rcv_isn + 1. */
704 flg &= ~TCP_FLAG_SYN;
705 }
706
707 /* RFC793: "second check the RST bit" and
708 * "fourth, check the SYN bit"
709 */
710 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) {
711 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
712 goto embryonic_reset;
713 }
714
715 /* ACK sequence verified above, just make sure ACK is
716 * set. If ACK not set, just silently drop the packet.
717 */
718 if (!(flg & TCP_FLAG_ACK))
719 return NULL;
720
721 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */
722 if (req->retrans < inet_csk(sk)->icsk_accept_queue.rskq_defer_accept &&
723 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) {
724 inet_rsk(req)->acked = 1;
725 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP);
726 return NULL;
727 }
728 if (tmp_opt.saw_tstamp && tmp_opt.rcv_tsecr)
729 tcp_rsk(req)->snt_synack = tmp_opt.rcv_tsecr;
730 else if (req->retrans) /* don't take RTT sample if retrans && ~TS */
731 tcp_rsk(req)->snt_synack = 0;
732
733 /* OK, ACK is valid, create big socket and
734 * feed this segment to it. It will repeat all
735 * the tests. THIS SEGMENT MUST MOVE SOCKET TO
736 * ESTABLISHED STATE. If it will be dropped after
737 * socket is created, wait for troubles.
738 */
739 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL);
740 if (child == NULL)
741 goto listen_overflow;
742
743 inet_csk_reqsk_queue_unlink(sk, req, prev);
744 inet_csk_reqsk_queue_removed(sk, req);
745
746 inet_csk_reqsk_queue_add(sk, req, child);
747 return child;
748
749listen_overflow:
750 if (!sysctl_tcp_abort_on_overflow) {
751 inet_rsk(req)->acked = 1;
752 return NULL;
753 }
754
755embryonic_reset:
756 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_EMBRYONICRSTS);
757 if (!(flg & TCP_FLAG_RST))
758 req->rsk_ops->send_reset(sk, skb);
759
760 inet_csk_reqsk_queue_drop(sk, req, prev);
761 return NULL;
762}
763EXPORT_SYMBOL(tcp_check_req);
764
765/*
766 * Queue segment on the new socket if the new socket is active,
767 * otherwise we just shortcircuit this and continue with
768 * the new socket.
769 */
770
771int tcp_child_process(struct sock *parent, struct sock *child,
772 struct sk_buff *skb)
773{
774 int ret = 0;
775 int state = child->sk_state;
776
777 if (!sock_owned_by_user(child)) {
778 ret = tcp_rcv_state_process(child, skb, tcp_hdr(skb),
779 skb->len);
780 /* Wakeup parent, send SIGIO */
781 if (state == TCP_SYN_RECV && child->sk_state != state)
782 parent->sk_data_ready(parent, 0);
783 } else {
784 /* Alas, it is possible again, because we do lookup
785 * in main socket hash table and lock on listening
786 * socket does not protect us more.
787 */
788 __sk_add_backlog(child, skb);
789 }
790
791 bh_unlock_sock(child);
792 sock_put(child);
793 return ret;
794}
795EXPORT_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_abort_on_overflow __read_mostly;
31
32struct inet_timewait_death_row tcp_death_row = {
33 .sysctl_max_tw_buckets = NR_FILE * 2,
34 .hashinfo = &tcp_hashinfo,
35};
36EXPORT_SYMBOL_GPL(tcp_death_row);
37
38static bool tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win)
39{
40 if (seq == s_win)
41 return true;
42 if (after(end_seq, s_win) && before(seq, e_win))
43 return true;
44 return seq == e_win && seq == end_seq;
45}
46
47static enum tcp_tw_status
48tcp_timewait_check_oow_rate_limit(struct inet_timewait_sock *tw,
49 const struct sk_buff *skb, int mib_idx)
50{
51 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
52
53 if (!tcp_oow_rate_limited(twsk_net(tw), skb, mib_idx,
54 &tcptw->tw_last_oow_ack_time)) {
55 /* Send ACK. Note, we do not put the bucket,
56 * it will be released by caller.
57 */
58 return TCP_TW_ACK;
59 }
60
61 /* We are rate-limiting, so just release the tw sock and drop skb. */
62 inet_twsk_put(tw);
63 return TCP_TW_SUCCESS;
64}
65
66/*
67 * * Main purpose of TIME-WAIT state is to close connection gracefully,
68 * when one of ends sits in LAST-ACK or CLOSING retransmitting FIN
69 * (and, probably, tail of data) and one or more our ACKs are lost.
70 * * What is TIME-WAIT timeout? It is associated with maximal packet
71 * lifetime in the internet, which results in wrong conclusion, that
72 * it is set to catch "old duplicate segments" wandering out of their path.
73 * It is not quite correct. This timeout is calculated so that it exceeds
74 * maximal retransmission timeout enough to allow to lose one (or more)
75 * segments sent by peer and our ACKs. This time may be calculated from RTO.
76 * * When TIME-WAIT socket receives RST, it means that another end
77 * finally closed and we are allowed to kill TIME-WAIT too.
78 * * Second purpose of TIME-WAIT is catching old duplicate segments.
79 * Well, certainly it is pure paranoia, but if we load TIME-WAIT
80 * with this semantics, we MUST NOT kill TIME-WAIT state with RSTs.
81 * * If we invented some more clever way to catch duplicates
82 * (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs.
83 *
84 * The algorithm below is based on FORMAL INTERPRETATION of RFCs.
85 * When you compare it to RFCs, please, read section SEGMENT ARRIVES
86 * from the very beginning.
87 *
88 * NOTE. With recycling (and later with fin-wait-2) TW bucket
89 * is _not_ stateless. It means, that strictly speaking we must
90 * spinlock it. I do not want! Well, probability of misbehaviour
91 * is ridiculously low and, seems, we could use some mb() tricks
92 * to avoid misread sequence numbers, states etc. --ANK
93 *
94 * We don't need to initialize tmp_out.sack_ok as we don't use the results
95 */
96enum tcp_tw_status
97tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb,
98 const struct tcphdr *th)
99{
100 struct tcp_options_received tmp_opt;
101 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
102 bool paws_reject = false;
103
104 tmp_opt.saw_tstamp = 0;
105 if (th->doff > (sizeof(*th) >> 2) && tcptw->tw_ts_recent_stamp) {
106 tcp_parse_options(skb, &tmp_opt, 0, NULL);
107
108 if (tmp_opt.saw_tstamp) {
109 tmp_opt.rcv_tsecr -= tcptw->tw_ts_offset;
110 tmp_opt.ts_recent = tcptw->tw_ts_recent;
111 tmp_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp;
112 paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
113 }
114 }
115
116 if (tw->tw_substate == TCP_FIN_WAIT2) {
117 /* Just repeat all the checks of tcp_rcv_state_process() */
118
119 /* Out of window, send ACK */
120 if (paws_reject ||
121 !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
122 tcptw->tw_rcv_nxt,
123 tcptw->tw_rcv_nxt + tcptw->tw_rcv_wnd))
124 return tcp_timewait_check_oow_rate_limit(
125 tw, skb, LINUX_MIB_TCPACKSKIPPEDFINWAIT2);
126
127 if (th->rst)
128 goto kill;
129
130 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt))
131 return TCP_TW_RST;
132
133 /* Dup ACK? */
134 if (!th->ack ||
135 !after(TCP_SKB_CB(skb)->end_seq, tcptw->tw_rcv_nxt) ||
136 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) {
137 inet_twsk_put(tw);
138 return TCP_TW_SUCCESS;
139 }
140
141 /* New data or FIN. If new data arrive after half-duplex close,
142 * reset.
143 */
144 if (!th->fin ||
145 TCP_SKB_CB(skb)->end_seq != tcptw->tw_rcv_nxt + 1)
146 return TCP_TW_RST;
147
148 /* FIN arrived, enter true time-wait state. */
149 tw->tw_substate = TCP_TIME_WAIT;
150 tcptw->tw_rcv_nxt = TCP_SKB_CB(skb)->end_seq;
151 if (tmp_opt.saw_tstamp) {
152 tcptw->tw_ts_recent_stamp = get_seconds();
153 tcptw->tw_ts_recent = tmp_opt.rcv_tsval;
154 }
155
156 if (tcp_death_row.sysctl_tw_recycle &&
157 tcptw->tw_ts_recent_stamp &&
158 tcp_tw_remember_stamp(tw))
159 inet_twsk_reschedule(tw, tw->tw_timeout);
160 else
161 inet_twsk_reschedule(tw, 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_put(tw);
195 return TCP_TW_SUCCESS;
196 }
197 }
198 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
199
200 if (tmp_opt.saw_tstamp) {
201 tcptw->tw_ts_recent = tmp_opt.rcv_tsval;
202 tcptw->tw_ts_recent_stamp = get_seconds();
203 }
204
205 inet_twsk_put(tw);
206 return TCP_TW_SUCCESS;
207 }
208
209 /* Out of window segment.
210
211 All the segments are ACKed immediately.
212
213 The only exception is new SYN. We accept it, if it is
214 not old duplicate and we are not in danger to be killed
215 by delayed old duplicates. RFC check is that it has
216 newer sequence number works at rates <40Mbit/sec.
217 However, if paws works, it is reliable AND even more,
218 we even may relax silly seq space cutoff.
219
220 RED-PEN: we violate main RFC requirement, if this SYN will appear
221 old duplicate (i.e. we receive RST in reply to SYN-ACK),
222 we must return socket to time-wait state. It is not good,
223 but not fatal yet.
224 */
225
226 if (th->syn && !th->rst && !th->ack && !paws_reject &&
227 (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) ||
228 (tmp_opt.saw_tstamp &&
229 (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) {
230 u32 isn = tcptw->tw_snd_nxt + 65535 + 2;
231 if (isn == 0)
232 isn++;
233 TCP_SKB_CB(skb)->tcp_tw_isn = isn;
234 return TCP_TW_SYN;
235 }
236
237 if (paws_reject)
238 NET_INC_STATS_BH(twsk_net(tw), LINUX_MIB_PAWSESTABREJECTED);
239
240 if (!th->rst) {
241 /* In this case we must reset the TIMEWAIT timer.
242 *
243 * If it is ACKless SYN it may be both old duplicate
244 * and new good SYN with random sequence number <rcv_nxt.
245 * Do not reschedule in the last case.
246 */
247 if (paws_reject || th->ack)
248 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
249
250 return tcp_timewait_check_oow_rate_limit(
251 tw, skb, LINUX_MIB_TCPACKSKIPPEDTIMEWAIT);
252 }
253 inet_twsk_put(tw);
254 return TCP_TW_SUCCESS;
255}
256EXPORT_SYMBOL(tcp_timewait_state_process);
257
258/*
259 * Move a socket to time-wait or dead fin-wait-2 state.
260 */
261void tcp_time_wait(struct sock *sk, int state, int timeo)
262{
263 const struct inet_connection_sock *icsk = inet_csk(sk);
264 const struct tcp_sock *tp = tcp_sk(sk);
265 struct inet_timewait_sock *tw;
266 bool recycle_ok = false;
267
268 if (tcp_death_row.sysctl_tw_recycle && tp->rx_opt.ts_recent_stamp)
269 recycle_ok = tcp_remember_stamp(sk);
270
271 tw = inet_twsk_alloc(sk, &tcp_death_row, state);
272
273 if (tw) {
274 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
275 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1);
276 struct inet_sock *inet = inet_sk(sk);
277
278 tw->tw_transparent = inet->transparent;
279 tw->tw_rcv_wscale = tp->rx_opt.rcv_wscale;
280 tcptw->tw_rcv_nxt = tp->rcv_nxt;
281 tcptw->tw_snd_nxt = tp->snd_nxt;
282 tcptw->tw_rcv_wnd = tcp_receive_window(tp);
283 tcptw->tw_ts_recent = tp->rx_opt.ts_recent;
284 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp;
285 tcptw->tw_ts_offset = tp->tsoffset;
286 tcptw->tw_last_oow_ack_time = 0;
287
288#if IS_ENABLED(CONFIG_IPV6)
289 if (tw->tw_family == PF_INET6) {
290 struct ipv6_pinfo *np = inet6_sk(sk);
291
292 tw->tw_v6_daddr = sk->sk_v6_daddr;
293 tw->tw_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
294 tw->tw_tclass = np->tclass;
295 tw->tw_flowlabel = be32_to_cpu(np->flow_label & IPV6_FLOWLABEL_MASK);
296 tw->tw_ipv6only = sk->sk_ipv6only;
297 }
298#endif
299
300#ifdef CONFIG_TCP_MD5SIG
301 /*
302 * The timewait bucket does not have the key DB from the
303 * sock structure. We just make a quick copy of the
304 * md5 key being used (if indeed we are using one)
305 * so the timewait ack generating code has the key.
306 */
307 do {
308 struct tcp_md5sig_key *key;
309 tcptw->tw_md5_key = NULL;
310 key = tp->af_specific->md5_lookup(sk, sk);
311 if (key) {
312 tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC);
313 if (tcptw->tw_md5_key && !tcp_alloc_md5sig_pool())
314 BUG();
315 }
316 } while (0);
317#endif
318
319 /* Get the TIME_WAIT timeout firing. */
320 if (timeo < rto)
321 timeo = rto;
322
323 if (recycle_ok) {
324 tw->tw_timeout = rto;
325 } else {
326 tw->tw_timeout = TCP_TIMEWAIT_LEN;
327 if (state == TCP_TIME_WAIT)
328 timeo = TCP_TIMEWAIT_LEN;
329 }
330
331 inet_twsk_schedule(tw, timeo);
332 /* Linkage updates. */
333 __inet_twsk_hashdance(tw, sk, &tcp_hashinfo);
334 inet_twsk_put(tw);
335 } else {
336 /* Sorry, if we're out of memory, just CLOSE this
337 * socket up. We've got bigger problems than
338 * non-graceful socket closings.
339 */
340 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPTIMEWAITOVERFLOW);
341 }
342
343 tcp_update_metrics(sk);
344 tcp_done(sk);
345}
346
347void tcp_twsk_destructor(struct sock *sk)
348{
349#ifdef CONFIG_TCP_MD5SIG
350 struct tcp_timewait_sock *twsk = tcp_twsk(sk);
351
352 if (twsk->tw_md5_key)
353 kfree_rcu(twsk->tw_md5_key, rcu);
354#endif
355}
356EXPORT_SYMBOL_GPL(tcp_twsk_destructor);
357
358/* Warning : This function is called without sk_listener being locked.
359 * Be sure to read socket fields once, as their value could change under us.
360 */
361void tcp_openreq_init_rwin(struct request_sock *req,
362 const struct sock *sk_listener,
363 const struct dst_entry *dst)
364{
365 struct inet_request_sock *ireq = inet_rsk(req);
366 const struct tcp_sock *tp = tcp_sk(sk_listener);
367 u16 user_mss = READ_ONCE(tp->rx_opt.user_mss);
368 int full_space = tcp_full_space(sk_listener);
369 int mss = dst_metric_advmss(dst);
370 u32 window_clamp;
371 __u8 rcv_wscale;
372
373 if (user_mss && user_mss < mss)
374 mss = user_mss;
375
376 window_clamp = READ_ONCE(tp->window_clamp);
377 /* Set this up on the first call only */
378 req->rsk_window_clamp = window_clamp ? : dst_metric(dst, RTAX_WINDOW);
379
380 /* limit the window selection if the user enforce a smaller rx buffer */
381 if (sk_listener->sk_userlocks & SOCK_RCVBUF_LOCK &&
382 (req->rsk_window_clamp > full_space || req->rsk_window_clamp == 0))
383 req->rsk_window_clamp = full_space;
384
385 /* tcp_full_space because it is guaranteed to be the first packet */
386 tcp_select_initial_window(full_space,
387 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
388 &req->rsk_rcv_wnd,
389 &req->rsk_window_clamp,
390 ireq->wscale_ok,
391 &rcv_wscale,
392 dst_metric(dst, RTAX_INITRWND));
393 ireq->rcv_wscale = rcv_wscale;
394}
395EXPORT_SYMBOL(tcp_openreq_init_rwin);
396
397static void tcp_ecn_openreq_child(struct tcp_sock *tp,
398 const struct request_sock *req)
399{
400 tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0;
401}
402
403void tcp_ca_openreq_child(struct sock *sk, const struct dst_entry *dst)
404{
405 struct inet_connection_sock *icsk = inet_csk(sk);
406 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
407 bool ca_got_dst = false;
408
409 if (ca_key != TCP_CA_UNSPEC) {
410 const struct tcp_congestion_ops *ca;
411
412 rcu_read_lock();
413 ca = tcp_ca_find_key(ca_key);
414 if (likely(ca && try_module_get(ca->owner))) {
415 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
416 icsk->icsk_ca_ops = ca;
417 ca_got_dst = true;
418 }
419 rcu_read_unlock();
420 }
421
422 /* If no valid choice made yet, assign current system default ca. */
423 if (!ca_got_dst &&
424 (!icsk->icsk_ca_setsockopt ||
425 !try_module_get(icsk->icsk_ca_ops->owner)))
426 tcp_assign_congestion_control(sk);
427
428 tcp_set_ca_state(sk, TCP_CA_Open);
429}
430EXPORT_SYMBOL_GPL(tcp_ca_openreq_child);
431
432/* This is not only more efficient than what we used to do, it eliminates
433 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM
434 *
435 * Actually, we could lots of memory writes here. tp of listening
436 * socket contains all necessary default parameters.
437 */
438struct sock *tcp_create_openreq_child(const struct sock *sk,
439 struct request_sock *req,
440 struct sk_buff *skb)
441{
442 struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC);
443
444 if (newsk) {
445 const struct inet_request_sock *ireq = inet_rsk(req);
446 struct tcp_request_sock *treq = tcp_rsk(req);
447 struct inet_connection_sock *newicsk = inet_csk(newsk);
448 struct tcp_sock *newtp = tcp_sk(newsk);
449
450 /* Now setup tcp_sock */
451 newtp->pred_flags = 0;
452
453 newtp->rcv_wup = newtp->copied_seq =
454 newtp->rcv_nxt = treq->rcv_isn + 1;
455 newtp->segs_in = 1;
456
457 newtp->snd_sml = newtp->snd_una =
458 newtp->snd_nxt = newtp->snd_up = treq->snt_isn + 1;
459
460 tcp_prequeue_init(newtp);
461 INIT_LIST_HEAD(&newtp->tsq_node);
462
463 tcp_init_wl(newtp, treq->rcv_isn);
464
465 newtp->srtt_us = 0;
466 newtp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
467 newtp->rtt_min[0].rtt = ~0U;
468 newicsk->icsk_rto = TCP_TIMEOUT_INIT;
469
470 newtp->packets_out = 0;
471 newtp->retrans_out = 0;
472 newtp->sacked_out = 0;
473 newtp->fackets_out = 0;
474 newtp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
475 tcp_enable_early_retrans(newtp);
476 newtp->tlp_high_seq = 0;
477 newtp->lsndtime = treq->snt_synack.stamp_jiffies;
478 newsk->sk_txhash = treq->txhash;
479 newtp->last_oow_ack_time = 0;
480 newtp->total_retrans = req->num_retrans;
481
482 /* So many TCP implementations out there (incorrectly) count the
483 * initial SYN frame in their delayed-ACK and congestion control
484 * algorithms that we must have the following bandaid to talk
485 * efficiently to them. -DaveM
486 */
487 newtp->snd_cwnd = TCP_INIT_CWND;
488 newtp->snd_cwnd_cnt = 0;
489
490 tcp_init_xmit_timers(newsk);
491 __skb_queue_head_init(&newtp->out_of_order_queue);
492 newtp->write_seq = newtp->pushed_seq = treq->snt_isn + 1;
493
494 newtp->rx_opt.saw_tstamp = 0;
495
496 newtp->rx_opt.dsack = 0;
497 newtp->rx_opt.num_sacks = 0;
498
499 newtp->urg_data = 0;
500
501 if (sock_flag(newsk, SOCK_KEEPOPEN))
502 inet_csk_reset_keepalive_timer(newsk,
503 keepalive_time_when(newtp));
504
505 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok;
506 if ((newtp->rx_opt.sack_ok = ireq->sack_ok) != 0) {
507 if (sysctl_tcp_fack)
508 tcp_enable_fack(newtp);
509 }
510 newtp->window_clamp = req->rsk_window_clamp;
511 newtp->rcv_ssthresh = req->rsk_rcv_wnd;
512 newtp->rcv_wnd = req->rsk_rcv_wnd;
513 newtp->rx_opt.wscale_ok = ireq->wscale_ok;
514 if (newtp->rx_opt.wscale_ok) {
515 newtp->rx_opt.snd_wscale = ireq->snd_wscale;
516 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale;
517 } else {
518 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0;
519 newtp->window_clamp = min(newtp->window_clamp, 65535U);
520 }
521 newtp->snd_wnd = (ntohs(tcp_hdr(skb)->window) <<
522 newtp->rx_opt.snd_wscale);
523 newtp->max_window = newtp->snd_wnd;
524
525 if (newtp->rx_opt.tstamp_ok) {
526 newtp->rx_opt.ts_recent = req->ts_recent;
527 newtp->rx_opt.ts_recent_stamp = get_seconds();
528 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
529 } else {
530 newtp->rx_opt.ts_recent_stamp = 0;
531 newtp->tcp_header_len = sizeof(struct tcphdr);
532 }
533 newtp->tsoffset = 0;
534#ifdef CONFIG_TCP_MD5SIG
535 newtp->md5sig_info = NULL; /*XXX*/
536 if (newtp->af_specific->md5_lookup(sk, newsk))
537 newtp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
538#endif
539 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len)
540 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len;
541 newtp->rx_opt.mss_clamp = req->mss;
542 tcp_ecn_openreq_child(newtp, req);
543 newtp->fastopen_rsk = NULL;
544 newtp->syn_data_acked = 0;
545 newtp->rack.mstamp.v64 = 0;
546 newtp->rack.advanced = 0;
547
548 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_PASSIVEOPENS);
549 }
550 return newsk;
551}
552EXPORT_SYMBOL(tcp_create_openreq_child);
553
554/*
555 * Process an incoming packet for SYN_RECV sockets represented as a
556 * request_sock. Normally sk is the listener socket but for TFO it
557 * points to the child socket.
558 *
559 * XXX (TFO) - The current impl contains a special check for ack
560 * validation and inside tcp_v4_reqsk_send_ack(). Can we do better?
561 *
562 * We don't need to initialize tmp_opt.sack_ok as we don't use the results
563 */
564
565struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb,
566 struct request_sock *req,
567 bool fastopen)
568{
569 struct tcp_options_received tmp_opt;
570 struct sock *child;
571 const struct tcphdr *th = tcp_hdr(skb);
572 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK);
573 bool paws_reject = false;
574 bool own_req;
575
576 tmp_opt.saw_tstamp = 0;
577 if (th->doff > (sizeof(struct tcphdr)>>2)) {
578 tcp_parse_options(skb, &tmp_opt, 0, NULL);
579
580 if (tmp_opt.saw_tstamp) {
581 tmp_opt.ts_recent = req->ts_recent;
582 /* We do not store true stamp, but it is not required,
583 * it can be estimated (approximately)
584 * from another data.
585 */
586 tmp_opt.ts_recent_stamp = get_seconds() - ((TCP_TIMEOUT_INIT/HZ)<<req->num_timeout);
587 paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
588 }
589 }
590
591 /* Check for pure retransmitted SYN. */
592 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn &&
593 flg == TCP_FLAG_SYN &&
594 !paws_reject) {
595 /*
596 * RFC793 draws (Incorrectly! It was fixed in RFC1122)
597 * this case on figure 6 and figure 8, but formal
598 * protocol description says NOTHING.
599 * To be more exact, it says that we should send ACK,
600 * because this segment (at least, if it has no data)
601 * is out of window.
602 *
603 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT
604 * describe SYN-RECV state. All the description
605 * is wrong, we cannot believe to it and should
606 * rely only on common sense and implementation
607 * experience.
608 *
609 * Enforce "SYN-ACK" according to figure 8, figure 6
610 * of RFC793, fixed by RFC1122.
611 *
612 * Note that even if there is new data in the SYN packet
613 * they will be thrown away too.
614 *
615 * Reset timer after retransmitting SYNACK, similar to
616 * the idea of fast retransmit in recovery.
617 */
618 if (!tcp_oow_rate_limited(sock_net(sk), skb,
619 LINUX_MIB_TCPACKSKIPPEDSYNRECV,
620 &tcp_rsk(req)->last_oow_ack_time) &&
621
622 !inet_rtx_syn_ack(sk, req)) {
623 unsigned long expires = jiffies;
624
625 expires += min(TCP_TIMEOUT_INIT << req->num_timeout,
626 TCP_RTO_MAX);
627 if (!fastopen)
628 mod_timer_pending(&req->rsk_timer, expires);
629 else
630 req->rsk_timer.expires = expires;
631 }
632 return NULL;
633 }
634
635 /* Further reproduces section "SEGMENT ARRIVES"
636 for state SYN-RECEIVED of RFC793.
637 It is broken, however, it does not work only
638 when SYNs are crossed.
639
640 You would think that SYN crossing is impossible here, since
641 we should have a SYN_SENT socket (from connect()) on our end,
642 but this is not true if the crossed SYNs were sent to both
643 ends by a malicious third party. We must defend against this,
644 and to do that we first verify the ACK (as per RFC793, page
645 36) and reset if it is invalid. Is this a true full defense?
646 To convince ourselves, let us consider a way in which the ACK
647 test can still pass in this 'malicious crossed SYNs' case.
648 Malicious sender sends identical SYNs (and thus identical sequence
649 numbers) to both A and B:
650
651 A: gets SYN, seq=7
652 B: gets SYN, seq=7
653
654 By our good fortune, both A and B select the same initial
655 send sequence number of seven :-)
656
657 A: sends SYN|ACK, seq=7, ack_seq=8
658 B: sends SYN|ACK, seq=7, ack_seq=8
659
660 So we are now A eating this SYN|ACK, ACK test passes. So
661 does sequence test, SYN is truncated, and thus we consider
662 it a bare ACK.
663
664 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this
665 bare ACK. Otherwise, we create an established connection. Both
666 ends (listening sockets) accept the new incoming connection and try
667 to talk to each other. 8-)
668
669 Note: This case is both harmless, and rare. Possibility is about the
670 same as us discovering intelligent life on another plant tomorrow.
671
672 But generally, we should (RFC lies!) to accept ACK
673 from SYNACK both here and in tcp_rcv_state_process().
674 tcp_rcv_state_process() does not, hence, we do not too.
675
676 Note that the case is absolutely generic:
677 we cannot optimize anything here without
678 violating protocol. All the checks must be made
679 before attempt to create socket.
680 */
681
682 /* RFC793 page 36: "If the connection is in any non-synchronized state ...
683 * and the incoming segment acknowledges something not yet
684 * sent (the segment carries an unacceptable ACK) ...
685 * a reset is sent."
686 *
687 * Invalid ACK: reset will be sent by listening socket.
688 * Note that the ACK validity check for a Fast Open socket is done
689 * elsewhere and is checked directly against the child socket rather
690 * than req because user data may have been sent out.
691 */
692 if ((flg & TCP_FLAG_ACK) && !fastopen &&
693 (TCP_SKB_CB(skb)->ack_seq !=
694 tcp_rsk(req)->snt_isn + 1))
695 return sk;
696
697 /* Also, it would be not so bad idea to check rcv_tsecr, which
698 * is essentially ACK extension and too early or too late values
699 * should cause reset in unsynchronized states.
700 */
701
702 /* RFC793: "first check sequence number". */
703
704 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
705 tcp_rsk(req)->rcv_nxt, tcp_rsk(req)->rcv_nxt + req->rsk_rcv_wnd)) {
706 /* Out of window: send ACK and drop. */
707 if (!(flg & TCP_FLAG_RST))
708 req->rsk_ops->send_ack(sk, skb, req);
709 if (paws_reject)
710 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
711 return NULL;
712 }
713
714 /* In sequence, PAWS is OK. */
715
716 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_nxt))
717 req->ts_recent = tmp_opt.rcv_tsval;
718
719 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) {
720 /* Truncate SYN, it is out of window starting
721 at tcp_rsk(req)->rcv_isn + 1. */
722 flg &= ~TCP_FLAG_SYN;
723 }
724
725 /* RFC793: "second check the RST bit" and
726 * "fourth, check the SYN bit"
727 */
728 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) {
729 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
730 goto embryonic_reset;
731 }
732
733 /* ACK sequence verified above, just make sure ACK is
734 * set. If ACK not set, just silently drop the packet.
735 *
736 * XXX (TFO) - if we ever allow "data after SYN", the
737 * following check needs to be removed.
738 */
739 if (!(flg & TCP_FLAG_ACK))
740 return NULL;
741
742 /* For Fast Open no more processing is needed (sk is the
743 * child socket).
744 */
745 if (fastopen)
746 return sk;
747
748 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */
749 if (req->num_timeout < inet_csk(sk)->icsk_accept_queue.rskq_defer_accept &&
750 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) {
751 inet_rsk(req)->acked = 1;
752 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP);
753 return NULL;
754 }
755
756 /* OK, ACK is valid, create big socket and
757 * feed this segment to it. It will repeat all
758 * the tests. THIS SEGMENT MUST MOVE SOCKET TO
759 * ESTABLISHED STATE. If it will be dropped after
760 * socket is created, wait for troubles.
761 */
762 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL,
763 req, &own_req);
764 if (!child)
765 goto listen_overflow;
766
767 sock_rps_save_rxhash(child, skb);
768 tcp_synack_rtt_meas(child, req);
769 return inet_csk_complete_hashdance(sk, child, req, own_req);
770
771listen_overflow:
772 if (!sysctl_tcp_abort_on_overflow) {
773 inet_rsk(req)->acked = 1;
774 return NULL;
775 }
776
777embryonic_reset:
778 if (!(flg & TCP_FLAG_RST)) {
779 /* Received a bad SYN pkt - for TFO We try not to reset
780 * the local connection unless it's really necessary to
781 * avoid becoming vulnerable to outside attack aiming at
782 * resetting legit local connections.
783 */
784 req->rsk_ops->send_reset(sk, skb);
785 } else if (fastopen) { /* received a valid RST pkt */
786 reqsk_fastopen_remove(sk, req, true);
787 tcp_reset(sk);
788 }
789 if (!fastopen) {
790 inet_csk_reqsk_queue_drop(sk, req);
791 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_EMBRYONICRSTS);
792 }
793 return NULL;
794}
795EXPORT_SYMBOL(tcp_check_req);
796
797/*
798 * Queue segment on the new socket if the new socket is active,
799 * otherwise we just shortcircuit this and continue with
800 * the new socket.
801 *
802 * For the vast majority of cases child->sk_state will be TCP_SYN_RECV
803 * when entering. But other states are possible due to a race condition
804 * where after __inet_lookup_established() fails but before the listener
805 * locked is obtained, other packets cause the same connection to
806 * be created.
807 */
808
809int tcp_child_process(struct sock *parent, struct sock *child,
810 struct sk_buff *skb)
811{
812 int ret = 0;
813 int state = child->sk_state;
814
815 tcp_segs_in(tcp_sk(child), skb);
816 if (!sock_owned_by_user(child)) {
817 ret = tcp_rcv_state_process(child, skb);
818 /* Wakeup parent, send SIGIO */
819 if (state == TCP_SYN_RECV && child->sk_state != state)
820 parent->sk_data_ready(parent);
821 } else {
822 /* Alas, it is possible again, because we do lookup
823 * in main socket hash table and lock on listening
824 * socket does not protect us more.
825 */
826 __sk_add_backlog(child, skb);
827 }
828
829 bh_unlock_sock(child);
830 sock_put(child);
831 return ret;
832}
833EXPORT_SYMBOL(tcp_child_process);