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 int 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 1;
76 }
77
78 return 0;
79}
80
81static int 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 1;
98 }
99 return 0;
100}
101
102static __inline__ int tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win)
103{
104 if (seq == s_win)
105 return 1;
106 if (after(end_seq, s_win) && before(seq, e_win))
107 return 1;
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 u8 *hash_location;
145 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
146 int paws_reject = 0;
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 int recycle_ok = 0;
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 defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
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 ipv6_addr_copy(&tw6->tw_v6_daddr, &np->daddr);
347 ipv6_addr_copy(&tw6->tw_v6_rcv_saddr, &np->rcv_saddr);
348 tw->tw_ipv6only = np->ipv6only;
349 }
350#endif
351
352#ifdef CONFIG_TCP_MD5SIG
353 /*
354 * The timewait bucket does not have the key DB from the
355 * sock structure. We just make a quick copy of the
356 * md5 key being used (if indeed we are using one)
357 * so the timewait ack generating code has the key.
358 */
359 do {
360 struct tcp_md5sig_key *key;
361 memset(tcptw->tw_md5_key, 0, sizeof(tcptw->tw_md5_key));
362 tcptw->tw_md5_keylen = 0;
363 key = tp->af_specific->md5_lookup(sk, sk);
364 if (key != NULL) {
365 memcpy(&tcptw->tw_md5_key, key->key, key->keylen);
366 tcptw->tw_md5_keylen = key->keylen;
367 if (tcp_alloc_md5sig_pool(sk) == NULL)
368 BUG();
369 }
370 } while (0);
371#endif
372
373 /* Linkage updates. */
374 __inet_twsk_hashdance(tw, sk, &tcp_hashinfo);
375
376 /* Get the TIME_WAIT timeout firing. */
377 if (timeo < rto)
378 timeo = rto;
379
380 if (recycle_ok) {
381 tw->tw_timeout = rto;
382 } else {
383 tw->tw_timeout = TCP_TIMEWAIT_LEN;
384 if (state == TCP_TIME_WAIT)
385 timeo = TCP_TIMEWAIT_LEN;
386 }
387
388 inet_twsk_schedule(tw, &tcp_death_row, timeo,
389 TCP_TIMEWAIT_LEN);
390 inet_twsk_put(tw);
391 } else {
392 /* Sorry, if we're out of memory, just CLOSE this
393 * socket up. We've got bigger problems than
394 * non-graceful socket closings.
395 */
396 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPTIMEWAITOVERFLOW);
397 }
398
399 tcp_update_metrics(sk);
400 tcp_done(sk);
401}
402
403void tcp_twsk_destructor(struct sock *sk)
404{
405#ifdef CONFIG_TCP_MD5SIG
406 struct tcp_timewait_sock *twsk = tcp_twsk(sk);
407 if (twsk->tw_md5_keylen)
408 tcp_free_md5sig_pool();
409#endif
410}
411EXPORT_SYMBOL_GPL(tcp_twsk_destructor);
412
413static inline void TCP_ECN_openreq_child(struct tcp_sock *tp,
414 struct request_sock *req)
415{
416 tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0;
417}
418
419/* This is not only more efficient than what we used to do, it eliminates
420 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM
421 *
422 * Actually, we could lots of memory writes here. tp of listening
423 * socket contains all necessary default parameters.
424 */
425struct sock *tcp_create_openreq_child(struct sock *sk, struct request_sock *req, struct sk_buff *skb)
426{
427 struct sock *newsk = inet_csk_clone(sk, req, GFP_ATOMIC);
428
429 if (newsk != NULL) {
430 const struct inet_request_sock *ireq = inet_rsk(req);
431 struct tcp_request_sock *treq = tcp_rsk(req);
432 struct inet_connection_sock *newicsk = inet_csk(newsk);
433 struct tcp_sock *newtp = tcp_sk(newsk);
434 struct tcp_sock *oldtp = tcp_sk(sk);
435 struct tcp_cookie_values *oldcvp = oldtp->cookie_values;
436
437 /* TCP Cookie Transactions require space for the cookie pair,
438 * as it differs for each connection. There is no need to
439 * copy any s_data_payload stored at the original socket.
440 * Failure will prevent resuming the connection.
441 *
442 * Presumed copied, in order of appearance:
443 * cookie_in_always, cookie_out_never
444 */
445 if (oldcvp != NULL) {
446 struct tcp_cookie_values *newcvp =
447 kzalloc(sizeof(*newtp->cookie_values),
448 GFP_ATOMIC);
449
450 if (newcvp != NULL) {
451 kref_init(&newcvp->kref);
452 newcvp->cookie_desired =
453 oldcvp->cookie_desired;
454 newtp->cookie_values = newcvp;
455 } else {
456 /* Not Yet Implemented */
457 newtp->cookie_values = NULL;
458 }
459 }
460
461 /* Now setup tcp_sock */
462 newtp->pred_flags = 0;
463
464 newtp->rcv_wup = newtp->copied_seq =
465 newtp->rcv_nxt = treq->rcv_isn + 1;
466
467 newtp->snd_sml = newtp->snd_una =
468 newtp->snd_nxt = newtp->snd_up =
469 treq->snt_isn + 1 + tcp_s_data_size(oldtp);
470
471 tcp_prequeue_init(newtp);
472
473 tcp_init_wl(newtp, treq->rcv_isn);
474
475 newtp->srtt = 0;
476 newtp->mdev = TCP_TIMEOUT_INIT;
477 newicsk->icsk_rto = TCP_TIMEOUT_INIT;
478
479 newtp->packets_out = 0;
480 newtp->retrans_out = 0;
481 newtp->sacked_out = 0;
482 newtp->fackets_out = 0;
483 newtp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
484
485 /* So many TCP implementations out there (incorrectly) count the
486 * initial SYN frame in their delayed-ACK and congestion control
487 * algorithms that we must have the following bandaid to talk
488 * efficiently to them. -DaveM
489 */
490 newtp->snd_cwnd = TCP_INIT_CWND;
491 newtp->snd_cwnd_cnt = 0;
492 newtp->bytes_acked = 0;
493
494 newtp->frto_counter = 0;
495 newtp->frto_highmark = 0;
496
497 newicsk->icsk_ca_ops = &tcp_init_congestion_ops;
498
499 tcp_set_ca_state(newsk, TCP_CA_Open);
500 tcp_init_xmit_timers(newsk);
501 skb_queue_head_init(&newtp->out_of_order_queue);
502 newtp->write_seq = newtp->pushed_seq =
503 treq->snt_isn + 1 + tcp_s_data_size(oldtp);
504
505 newtp->rx_opt.saw_tstamp = 0;
506
507 newtp->rx_opt.dsack = 0;
508 newtp->rx_opt.num_sacks = 0;
509
510 newtp->urg_data = 0;
511
512 if (sock_flag(newsk, SOCK_KEEPOPEN))
513 inet_csk_reset_keepalive_timer(newsk,
514 keepalive_time_when(newtp));
515
516 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok;
517 if ((newtp->rx_opt.sack_ok = ireq->sack_ok) != 0) {
518 if (sysctl_tcp_fack)
519 tcp_enable_fack(newtp);
520 }
521 newtp->window_clamp = req->window_clamp;
522 newtp->rcv_ssthresh = req->rcv_wnd;
523 newtp->rcv_wnd = req->rcv_wnd;
524 newtp->rx_opt.wscale_ok = ireq->wscale_ok;
525 if (newtp->rx_opt.wscale_ok) {
526 newtp->rx_opt.snd_wscale = ireq->snd_wscale;
527 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale;
528 } else {
529 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0;
530 newtp->window_clamp = min(newtp->window_clamp, 65535U);
531 }
532 newtp->snd_wnd = (ntohs(tcp_hdr(skb)->window) <<
533 newtp->rx_opt.snd_wscale);
534 newtp->max_window = newtp->snd_wnd;
535
536 if (newtp->rx_opt.tstamp_ok) {
537 newtp->rx_opt.ts_recent = req->ts_recent;
538 newtp->rx_opt.ts_recent_stamp = get_seconds();
539 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
540 } else {
541 newtp->rx_opt.ts_recent_stamp = 0;
542 newtp->tcp_header_len = sizeof(struct tcphdr);
543 }
544#ifdef CONFIG_TCP_MD5SIG
545 newtp->md5sig_info = NULL; /*XXX*/
546 if (newtp->af_specific->md5_lookup(sk, newsk))
547 newtp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
548#endif
549 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len)
550 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len;
551 newtp->rx_opt.mss_clamp = req->mss;
552 TCP_ECN_openreq_child(newtp, req);
553
554 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_PASSIVEOPENS);
555 }
556 return newsk;
557}
558EXPORT_SYMBOL(tcp_create_openreq_child);
559
560/*
561 * Process an incoming packet for SYN_RECV sockets represented
562 * as a request_sock.
563 */
564
565struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb,
566 struct request_sock *req,
567 struct request_sock **prev)
568{
569 struct tcp_options_received tmp_opt;
570 u8 *hash_location;
571 struct sock *child;
572 const struct tcphdr *th = tcp_hdr(skb);
573 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK);
574 int paws_reject = 0;
575
576 tmp_opt.saw_tstamp = 0;
577 if (th->doff > (sizeof(struct tcphdr)>>2)) {
578 tcp_parse_options(skb, &tmp_opt, &hash_location, 0);
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->retrans);
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 req->rsk_ops->rtx_syn_ack(sk, req, NULL);
613 return NULL;
614 }
615
616 /* Further reproduces section "SEGMENT ARRIVES"
617 for state SYN-RECEIVED of RFC793.
618 It is broken, however, it does not work only
619 when SYNs are crossed.
620
621 You would think that SYN crossing is impossible here, since
622 we should have a SYN_SENT socket (from connect()) on our end,
623 but this is not true if the crossed SYNs were sent to both
624 ends by a malicious third party. We must defend against this,
625 and to do that we first verify the ACK (as per RFC793, page
626 36) and reset if it is invalid. Is this a true full defense?
627 To convince ourselves, let us consider a way in which the ACK
628 test can still pass in this 'malicious crossed SYNs' case.
629 Malicious sender sends identical SYNs (and thus identical sequence
630 numbers) to both A and B:
631
632 A: gets SYN, seq=7
633 B: gets SYN, seq=7
634
635 By our good fortune, both A and B select the same initial
636 send sequence number of seven :-)
637
638 A: sends SYN|ACK, seq=7, ack_seq=8
639 B: sends SYN|ACK, seq=7, ack_seq=8
640
641 So we are now A eating this SYN|ACK, ACK test passes. So
642 does sequence test, SYN is truncated, and thus we consider
643 it a bare ACK.
644
645 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this
646 bare ACK. Otherwise, we create an established connection. Both
647 ends (listening sockets) accept the new incoming connection and try
648 to talk to each other. 8-)
649
650 Note: This case is both harmless, and rare. Possibility is about the
651 same as us discovering intelligent life on another plant tomorrow.
652
653 But generally, we should (RFC lies!) to accept ACK
654 from SYNACK both here and in tcp_rcv_state_process().
655 tcp_rcv_state_process() does not, hence, we do not too.
656
657 Note that the case is absolutely generic:
658 we cannot optimize anything here without
659 violating protocol. All the checks must be made
660 before attempt to create socket.
661 */
662
663 /* RFC793 page 36: "If the connection is in any non-synchronized state ...
664 * and the incoming segment acknowledges something not yet
665 * sent (the segment carries an unacceptable ACK) ...
666 * a reset is sent."
667 *
668 * Invalid ACK: reset will be sent by listening socket
669 */
670 if ((flg & TCP_FLAG_ACK) &&
671 (TCP_SKB_CB(skb)->ack_seq !=
672 tcp_rsk(req)->snt_isn + 1 + tcp_s_data_size(tcp_sk(sk))))
673 return sk;
674
675 /* Also, it would be not so bad idea to check rcv_tsecr, which
676 * is essentially ACK extension and too early or too late values
677 * should cause reset in unsynchronized states.
678 */
679
680 /* RFC793: "first check sequence number". */
681
682 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
683 tcp_rsk(req)->rcv_isn + 1, tcp_rsk(req)->rcv_isn + 1 + req->rcv_wnd)) {
684 /* Out of window: send ACK and drop. */
685 if (!(flg & TCP_FLAG_RST))
686 req->rsk_ops->send_ack(sk, skb, req);
687 if (paws_reject)
688 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
689 return NULL;
690 }
691
692 /* In sequence, PAWS is OK. */
693
694 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_isn + 1))
695 req->ts_recent = tmp_opt.rcv_tsval;
696
697 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) {
698 /* Truncate SYN, it is out of window starting
699 at tcp_rsk(req)->rcv_isn + 1. */
700 flg &= ~TCP_FLAG_SYN;
701 }
702
703 /* RFC793: "second check the RST bit" and
704 * "fourth, check the SYN bit"
705 */
706 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) {
707 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
708 goto embryonic_reset;
709 }
710
711 /* ACK sequence verified above, just make sure ACK is
712 * set. If ACK not set, just silently drop the packet.
713 */
714 if (!(flg & TCP_FLAG_ACK))
715 return NULL;
716
717 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */
718 if (req->retrans < inet_csk(sk)->icsk_accept_queue.rskq_defer_accept &&
719 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) {
720 inet_rsk(req)->acked = 1;
721 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP);
722 return NULL;
723 }
724 if (tmp_opt.saw_tstamp && tmp_opt.rcv_tsecr)
725 tcp_rsk(req)->snt_synack = tmp_opt.rcv_tsecr;
726 else if (req->retrans) /* don't take RTT sample if retrans && ~TS */
727 tcp_rsk(req)->snt_synack = 0;
728
729 /* OK, ACK is valid, create big socket and
730 * feed this segment to it. It will repeat all
731 * the tests. THIS SEGMENT MUST MOVE SOCKET TO
732 * ESTABLISHED STATE. If it will be dropped after
733 * socket is created, wait for troubles.
734 */
735 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL);
736 if (child == NULL)
737 goto listen_overflow;
738
739 inet_csk_reqsk_queue_unlink(sk, req, prev);
740 inet_csk_reqsk_queue_removed(sk, req);
741
742 inet_csk_reqsk_queue_add(sk, req, child);
743 return child;
744
745listen_overflow:
746 if (!sysctl_tcp_abort_on_overflow) {
747 inet_rsk(req)->acked = 1;
748 return NULL;
749 }
750
751embryonic_reset:
752 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_EMBRYONICRSTS);
753 if (!(flg & TCP_FLAG_RST))
754 req->rsk_ops->send_reset(sk, skb);
755
756 inet_csk_reqsk_queue_drop(sk, req, prev);
757 return NULL;
758}
759EXPORT_SYMBOL(tcp_check_req);
760
761/*
762 * Queue segment on the new socket if the new socket is active,
763 * otherwise we just shortcircuit this and continue with
764 * the new socket.
765 */
766
767int tcp_child_process(struct sock *parent, struct sock *child,
768 struct sk_buff *skb)
769{
770 int ret = 0;
771 int state = child->sk_state;
772
773 if (!sock_owned_by_user(child)) {
774 ret = tcp_rcv_state_process(child, skb, tcp_hdr(skb),
775 skb->len);
776 /* Wakeup parent, send SIGIO */
777 if (state == TCP_SYN_RECV && child->sk_state != state)
778 parent->sk_data_ready(parent, 0);
779 } else {
780 /* Alas, it is possible again, because we do lookup
781 * in main socket hash table and lock on listening
782 * socket does not protect us more.
783 */
784 __sk_add_backlog(child, skb);
785 }
786
787 bh_unlock_sock(child);
788 sock_put(child);
789 return ret;
790}
791EXPORT_SYMBOL(tcp_child_process);
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
26static bool tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win)
27{
28 if (seq == s_win)
29 return true;
30 if (after(end_seq, s_win) && before(seq, e_win))
31 return true;
32 return seq == e_win && seq == end_seq;
33}
34
35static enum tcp_tw_status
36tcp_timewait_check_oow_rate_limit(struct inet_timewait_sock *tw,
37 const struct sk_buff *skb, int mib_idx)
38{
39 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
40
41 if (!tcp_oow_rate_limited(twsk_net(tw), skb, mib_idx,
42 &tcptw->tw_last_oow_ack_time)) {
43 /* Send ACK. Note, we do not put the bucket,
44 * it will be released by caller.
45 */
46 return TCP_TW_ACK;
47 }
48
49 /* We are rate-limiting, so just release the tw sock and drop skb. */
50 inet_twsk_put(tw);
51 return TCP_TW_SUCCESS;
52}
53
54/*
55 * * Main purpose of TIME-WAIT state is to close connection gracefully,
56 * when one of ends sits in LAST-ACK or CLOSING retransmitting FIN
57 * (and, probably, tail of data) and one or more our ACKs are lost.
58 * * What is TIME-WAIT timeout? It is associated with maximal packet
59 * lifetime in the internet, which results in wrong conclusion, that
60 * it is set to catch "old duplicate segments" wandering out of their path.
61 * It is not quite correct. This timeout is calculated so that it exceeds
62 * maximal retransmission timeout enough to allow to lose one (or more)
63 * segments sent by peer and our ACKs. This time may be calculated from RTO.
64 * * When TIME-WAIT socket receives RST, it means that another end
65 * finally closed and we are allowed to kill TIME-WAIT too.
66 * * Second purpose of TIME-WAIT is catching old duplicate segments.
67 * Well, certainly it is pure paranoia, but if we load TIME-WAIT
68 * with this semantics, we MUST NOT kill TIME-WAIT state with RSTs.
69 * * If we invented some more clever way to catch duplicates
70 * (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs.
71 *
72 * The algorithm below is based on FORMAL INTERPRETATION of RFCs.
73 * When you compare it to RFCs, please, read section SEGMENT ARRIVES
74 * from the very beginning.
75 *
76 * NOTE. With recycling (and later with fin-wait-2) TW bucket
77 * is _not_ stateless. It means, that strictly speaking we must
78 * spinlock it. I do not want! Well, probability of misbehaviour
79 * is ridiculously low and, seems, we could use some mb() tricks
80 * to avoid misread sequence numbers, states etc. --ANK
81 *
82 * We don't need to initialize tmp_out.sack_ok as we don't use the results
83 */
84enum tcp_tw_status
85tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb,
86 const struct tcphdr *th)
87{
88 struct tcp_options_received tmp_opt;
89 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
90 bool paws_reject = false;
91
92 tmp_opt.saw_tstamp = 0;
93 if (th->doff > (sizeof(*th) >> 2) && tcptw->tw_ts_recent_stamp) {
94 tcp_parse_options(twsk_net(tw), skb, &tmp_opt, 0, NULL);
95
96 if (tmp_opt.saw_tstamp) {
97 if (tmp_opt.rcv_tsecr)
98 tmp_opt.rcv_tsecr -= tcptw->tw_ts_offset;
99 tmp_opt.ts_recent = tcptw->tw_ts_recent;
100 tmp_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp;
101 paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
102 }
103 }
104
105 if (tw->tw_substate == TCP_FIN_WAIT2) {
106 /* Just repeat all the checks of tcp_rcv_state_process() */
107
108 /* Out of window, send ACK */
109 if (paws_reject ||
110 !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
111 tcptw->tw_rcv_nxt,
112 tcptw->tw_rcv_nxt + tcptw->tw_rcv_wnd))
113 return tcp_timewait_check_oow_rate_limit(
114 tw, skb, LINUX_MIB_TCPACKSKIPPEDFINWAIT2);
115
116 if (th->rst)
117 goto kill;
118
119 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt))
120 return TCP_TW_RST;
121
122 /* Dup ACK? */
123 if (!th->ack ||
124 !after(TCP_SKB_CB(skb)->end_seq, tcptw->tw_rcv_nxt) ||
125 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) {
126 inet_twsk_put(tw);
127 return TCP_TW_SUCCESS;
128 }
129
130 /* New data or FIN. If new data arrive after half-duplex close,
131 * reset.
132 */
133 if (!th->fin ||
134 TCP_SKB_CB(skb)->end_seq != tcptw->tw_rcv_nxt + 1)
135 return TCP_TW_RST;
136
137 /* FIN arrived, enter true time-wait state. */
138 tw->tw_substate = TCP_TIME_WAIT;
139 tcptw->tw_rcv_nxt = TCP_SKB_CB(skb)->end_seq;
140 if (tmp_opt.saw_tstamp) {
141 tcptw->tw_ts_recent_stamp = ktime_get_seconds();
142 tcptw->tw_ts_recent = tmp_opt.rcv_tsval;
143 }
144
145 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
146 return TCP_TW_ACK;
147 }
148
149 /*
150 * Now real TIME-WAIT state.
151 *
152 * RFC 1122:
153 * "When a connection is [...] on TIME-WAIT state [...]
154 * [a TCP] MAY accept a new SYN from the remote TCP to
155 * reopen the connection directly, if it:
156 *
157 * (1) assigns its initial sequence number for the new
158 * connection to be larger than the largest sequence
159 * number it used on the previous connection incarnation,
160 * and
161 *
162 * (2) returns to TIME-WAIT state if the SYN turns out
163 * to be an old duplicate".
164 */
165
166 if (!paws_reject &&
167 (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt &&
168 (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) {
169 /* In window segment, it may be only reset or bare ack. */
170
171 if (th->rst) {
172 /* This is TIME_WAIT assassination, in two flavors.
173 * Oh well... nobody has a sufficient solution to this
174 * protocol bug yet.
175 */
176 if (!READ_ONCE(twsk_net(tw)->ipv4.sysctl_tcp_rfc1337)) {
177kill:
178 inet_twsk_deschedule_put(tw);
179 return TCP_TW_SUCCESS;
180 }
181 } else {
182 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
183 }
184
185 if (tmp_opt.saw_tstamp) {
186 tcptw->tw_ts_recent = tmp_opt.rcv_tsval;
187 tcptw->tw_ts_recent_stamp = ktime_get_seconds();
188 }
189
190 inet_twsk_put(tw);
191 return TCP_TW_SUCCESS;
192 }
193
194 /* Out of window segment.
195
196 All the segments are ACKed immediately.
197
198 The only exception is new SYN. We accept it, if it is
199 not old duplicate and we are not in danger to be killed
200 by delayed old duplicates. RFC check is that it has
201 newer sequence number works at rates <40Mbit/sec.
202 However, if paws works, it is reliable AND even more,
203 we even may relax silly seq space cutoff.
204
205 RED-PEN: we violate main RFC requirement, if this SYN will appear
206 old duplicate (i.e. we receive RST in reply to SYN-ACK),
207 we must return socket to time-wait state. It is not good,
208 but not fatal yet.
209 */
210
211 if (th->syn && !th->rst && !th->ack && !paws_reject &&
212 (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) ||
213 (tmp_opt.saw_tstamp &&
214 (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) {
215 u32 isn = tcptw->tw_snd_nxt + 65535 + 2;
216 if (isn == 0)
217 isn++;
218 TCP_SKB_CB(skb)->tcp_tw_isn = isn;
219 return TCP_TW_SYN;
220 }
221
222 if (paws_reject)
223 __NET_INC_STATS(twsk_net(tw), LINUX_MIB_PAWSESTABREJECTED);
224
225 if (!th->rst) {
226 /* In this case we must reset the TIMEWAIT timer.
227 *
228 * If it is ACKless SYN it may be both old duplicate
229 * and new good SYN with random sequence number <rcv_nxt.
230 * Do not reschedule in the last case.
231 */
232 if (paws_reject || th->ack)
233 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
234
235 return tcp_timewait_check_oow_rate_limit(
236 tw, skb, LINUX_MIB_TCPACKSKIPPEDTIMEWAIT);
237 }
238 inet_twsk_put(tw);
239 return TCP_TW_SUCCESS;
240}
241EXPORT_SYMBOL(tcp_timewait_state_process);
242
243static void tcp_time_wait_init(struct sock *sk, struct tcp_timewait_sock *tcptw)
244{
245#ifdef CONFIG_TCP_MD5SIG
246 const struct tcp_sock *tp = tcp_sk(sk);
247 struct tcp_md5sig_key *key;
248
249 /*
250 * The timewait bucket does not have the key DB from the
251 * sock structure. We just make a quick copy of the
252 * md5 key being used (if indeed we are using one)
253 * so the timewait ack generating code has the key.
254 */
255 tcptw->tw_md5_key = NULL;
256 if (!static_branch_unlikely(&tcp_md5_needed.key))
257 return;
258
259 key = tp->af_specific->md5_lookup(sk, sk);
260 if (key) {
261 tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC);
262 if (!tcptw->tw_md5_key)
263 return;
264 if (!tcp_alloc_md5sig_pool())
265 goto out_free;
266 if (!static_key_fast_inc_not_disabled(&tcp_md5_needed.key.key))
267 goto out_free;
268 }
269 return;
270out_free:
271 WARN_ON_ONCE(1);
272 kfree(tcptw->tw_md5_key);
273 tcptw->tw_md5_key = NULL;
274#endif
275}
276
277/*
278 * Move a socket to time-wait or dead fin-wait-2 state.
279 */
280void tcp_time_wait(struct sock *sk, int state, int timeo)
281{
282 const struct inet_connection_sock *icsk = inet_csk(sk);
283 const struct tcp_sock *tp = tcp_sk(sk);
284 struct net *net = sock_net(sk);
285 struct inet_timewait_sock *tw;
286
287 tw = inet_twsk_alloc(sk, &net->ipv4.tcp_death_row, state);
288
289 if (tw) {
290 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
291 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1);
292 struct inet_sock *inet = inet_sk(sk);
293
294 tw->tw_transparent = inet->transparent;
295 tw->tw_mark = sk->sk_mark;
296 tw->tw_priority = sk->sk_priority;
297 tw->tw_rcv_wscale = tp->rx_opt.rcv_wscale;
298 tcptw->tw_rcv_nxt = tp->rcv_nxt;
299 tcptw->tw_snd_nxt = tp->snd_nxt;
300 tcptw->tw_rcv_wnd = tcp_receive_window(tp);
301 tcptw->tw_ts_recent = tp->rx_opt.ts_recent;
302 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp;
303 tcptw->tw_ts_offset = tp->tsoffset;
304 tcptw->tw_last_oow_ack_time = 0;
305 tcptw->tw_tx_delay = tp->tcp_tx_delay;
306#if IS_ENABLED(CONFIG_IPV6)
307 if (tw->tw_family == PF_INET6) {
308 struct ipv6_pinfo *np = inet6_sk(sk);
309
310 tw->tw_v6_daddr = sk->sk_v6_daddr;
311 tw->tw_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
312 tw->tw_tclass = np->tclass;
313 tw->tw_flowlabel = be32_to_cpu(np->flow_label & IPV6_FLOWLABEL_MASK);
314 tw->tw_txhash = sk->sk_txhash;
315 tw->tw_ipv6only = sk->sk_ipv6only;
316 }
317#endif
318
319 tcp_time_wait_init(sk, tcptw);
320
321 /* Get the TIME_WAIT timeout firing. */
322 if (timeo < rto)
323 timeo = rto;
324
325 if (state == TCP_TIME_WAIT)
326 timeo = TCP_TIMEWAIT_LEN;
327
328 /* tw_timer is pinned, so we need to make sure BH are disabled
329 * in following section, otherwise timer handler could run before
330 * we complete the initialization.
331 */
332 local_bh_disable();
333 inet_twsk_schedule(tw, timeo);
334 /* Linkage updates.
335 * Note that access to tw after this point is illegal.
336 */
337 inet_twsk_hashdance(tw, sk, net->ipv4.tcp_death_row.hashinfo);
338 local_bh_enable();
339 } else {
340 /* Sorry, if we're out of memory, just CLOSE this
341 * socket up. We've got bigger problems than
342 * non-graceful socket closings.
343 */
344 NET_INC_STATS(net, LINUX_MIB_TCPTIMEWAITOVERFLOW);
345 }
346
347 tcp_update_metrics(sk);
348 tcp_done(sk);
349}
350EXPORT_SYMBOL(tcp_time_wait);
351
352void tcp_twsk_destructor(struct sock *sk)
353{
354#ifdef CONFIG_TCP_MD5SIG
355 if (static_branch_unlikely(&tcp_md5_needed.key)) {
356 struct tcp_timewait_sock *twsk = tcp_twsk(sk);
357
358 if (twsk->tw_md5_key) {
359 kfree_rcu(twsk->tw_md5_key, rcu);
360 static_branch_slow_dec_deferred(&tcp_md5_needed);
361 }
362 }
363#endif
364}
365EXPORT_SYMBOL_GPL(tcp_twsk_destructor);
366
367void tcp_twsk_purge(struct list_head *net_exit_list, int family)
368{
369 bool purged_once = false;
370 struct net *net;
371
372 list_for_each_entry(net, net_exit_list, exit_list) {
373 if (net->ipv4.tcp_death_row.hashinfo->pernet) {
374 /* Even if tw_refcount == 1, we must clean up kernel reqsk */
375 inet_twsk_purge(net->ipv4.tcp_death_row.hashinfo, family);
376 } else if (!purged_once) {
377 /* The last refcount is decremented in tcp_sk_exit_batch() */
378 if (refcount_read(&net->ipv4.tcp_death_row.tw_refcount) == 1)
379 continue;
380
381 inet_twsk_purge(&tcp_hashinfo, family);
382 purged_once = true;
383 }
384 }
385}
386EXPORT_SYMBOL_GPL(tcp_twsk_purge);
387
388/* Warning : This function is called without sk_listener being locked.
389 * Be sure to read socket fields once, as their value could change under us.
390 */
391void tcp_openreq_init_rwin(struct request_sock *req,
392 const struct sock *sk_listener,
393 const struct dst_entry *dst)
394{
395 struct inet_request_sock *ireq = inet_rsk(req);
396 const struct tcp_sock *tp = tcp_sk(sk_listener);
397 int full_space = tcp_full_space(sk_listener);
398 u32 window_clamp;
399 __u8 rcv_wscale;
400 u32 rcv_wnd;
401 int mss;
402
403 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
404 window_clamp = READ_ONCE(tp->window_clamp);
405 /* Set this up on the first call only */
406 req->rsk_window_clamp = window_clamp ? : dst_metric(dst, RTAX_WINDOW);
407
408 /* limit the window selection if the user enforce a smaller rx buffer */
409 if (sk_listener->sk_userlocks & SOCK_RCVBUF_LOCK &&
410 (req->rsk_window_clamp > full_space || req->rsk_window_clamp == 0))
411 req->rsk_window_clamp = full_space;
412
413 rcv_wnd = tcp_rwnd_init_bpf((struct sock *)req);
414 if (rcv_wnd == 0)
415 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
416 else if (full_space < rcv_wnd * mss)
417 full_space = rcv_wnd * mss;
418
419 /* tcp_full_space because it is guaranteed to be the first packet */
420 tcp_select_initial_window(sk_listener, full_space,
421 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
422 &req->rsk_rcv_wnd,
423 &req->rsk_window_clamp,
424 ireq->wscale_ok,
425 &rcv_wscale,
426 rcv_wnd);
427 ireq->rcv_wscale = rcv_wscale;
428}
429EXPORT_SYMBOL(tcp_openreq_init_rwin);
430
431static void tcp_ecn_openreq_child(struct tcp_sock *tp,
432 const struct request_sock *req)
433{
434 tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0;
435}
436
437void tcp_ca_openreq_child(struct sock *sk, const struct dst_entry *dst)
438{
439 struct inet_connection_sock *icsk = inet_csk(sk);
440 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
441 bool ca_got_dst = false;
442
443 if (ca_key != TCP_CA_UNSPEC) {
444 const struct tcp_congestion_ops *ca;
445
446 rcu_read_lock();
447 ca = tcp_ca_find_key(ca_key);
448 if (likely(ca && bpf_try_module_get(ca, ca->owner))) {
449 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
450 icsk->icsk_ca_ops = ca;
451 ca_got_dst = true;
452 }
453 rcu_read_unlock();
454 }
455
456 /* If no valid choice made yet, assign current system default ca. */
457 if (!ca_got_dst &&
458 (!icsk->icsk_ca_setsockopt ||
459 !bpf_try_module_get(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner)))
460 tcp_assign_congestion_control(sk);
461
462 tcp_set_ca_state(sk, TCP_CA_Open);
463}
464EXPORT_SYMBOL_GPL(tcp_ca_openreq_child);
465
466static void smc_check_reset_syn_req(struct tcp_sock *oldtp,
467 struct request_sock *req,
468 struct tcp_sock *newtp)
469{
470#if IS_ENABLED(CONFIG_SMC)
471 struct inet_request_sock *ireq;
472
473 if (static_branch_unlikely(&tcp_have_smc)) {
474 ireq = inet_rsk(req);
475 if (oldtp->syn_smc && !ireq->smc_ok)
476 newtp->syn_smc = 0;
477 }
478#endif
479}
480
481/* This is not only more efficient than what we used to do, it eliminates
482 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM
483 *
484 * Actually, we could lots of memory writes here. tp of listening
485 * socket contains all necessary default parameters.
486 */
487struct sock *tcp_create_openreq_child(const struct sock *sk,
488 struct request_sock *req,
489 struct sk_buff *skb)
490{
491 struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC);
492 const struct inet_request_sock *ireq = inet_rsk(req);
493 struct tcp_request_sock *treq = tcp_rsk(req);
494 struct inet_connection_sock *newicsk;
495 struct tcp_sock *oldtp, *newtp;
496 u32 seq;
497
498 if (!newsk)
499 return NULL;
500
501 newicsk = inet_csk(newsk);
502 newtp = tcp_sk(newsk);
503 oldtp = tcp_sk(sk);
504
505 smc_check_reset_syn_req(oldtp, req, newtp);
506
507 /* Now setup tcp_sock */
508 newtp->pred_flags = 0;
509
510 seq = treq->rcv_isn + 1;
511 newtp->rcv_wup = seq;
512 WRITE_ONCE(newtp->copied_seq, seq);
513 WRITE_ONCE(newtp->rcv_nxt, seq);
514 newtp->segs_in = 1;
515
516 seq = treq->snt_isn + 1;
517 newtp->snd_sml = newtp->snd_una = seq;
518 WRITE_ONCE(newtp->snd_nxt, seq);
519 newtp->snd_up = seq;
520
521 INIT_LIST_HEAD(&newtp->tsq_node);
522 INIT_LIST_HEAD(&newtp->tsorted_sent_queue);
523
524 tcp_init_wl(newtp, treq->rcv_isn);
525
526 minmax_reset(&newtp->rtt_min, tcp_jiffies32, ~0U);
527 newicsk->icsk_ack.lrcvtime = tcp_jiffies32;
528
529 newtp->lsndtime = tcp_jiffies32;
530 newsk->sk_txhash = treq->txhash;
531 newtp->total_retrans = req->num_retrans;
532
533 tcp_init_xmit_timers(newsk);
534 WRITE_ONCE(newtp->write_seq, newtp->pushed_seq = treq->snt_isn + 1);
535
536 if (sock_flag(newsk, SOCK_KEEPOPEN))
537 inet_csk_reset_keepalive_timer(newsk,
538 keepalive_time_when(newtp));
539
540 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok;
541 newtp->rx_opt.sack_ok = ireq->sack_ok;
542 newtp->window_clamp = req->rsk_window_clamp;
543 newtp->rcv_ssthresh = req->rsk_rcv_wnd;
544 newtp->rcv_wnd = req->rsk_rcv_wnd;
545 newtp->rx_opt.wscale_ok = ireq->wscale_ok;
546 if (newtp->rx_opt.wscale_ok) {
547 newtp->rx_opt.snd_wscale = ireq->snd_wscale;
548 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale;
549 } else {
550 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0;
551 newtp->window_clamp = min(newtp->window_clamp, 65535U);
552 }
553 newtp->snd_wnd = ntohs(tcp_hdr(skb)->window) << newtp->rx_opt.snd_wscale;
554 newtp->max_window = newtp->snd_wnd;
555
556 if (newtp->rx_opt.tstamp_ok) {
557 newtp->rx_opt.ts_recent = req->ts_recent;
558 newtp->rx_opt.ts_recent_stamp = ktime_get_seconds();
559 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
560 } else {
561 newtp->rx_opt.ts_recent_stamp = 0;
562 newtp->tcp_header_len = sizeof(struct tcphdr);
563 }
564 if (req->num_timeout) {
565 newtp->undo_marker = treq->snt_isn;
566 newtp->retrans_stamp = div_u64(treq->snt_synack,
567 USEC_PER_SEC / TCP_TS_HZ);
568 }
569 newtp->tsoffset = treq->ts_off;
570#ifdef CONFIG_TCP_MD5SIG
571 newtp->md5sig_info = NULL; /*XXX*/
572 if (treq->af_specific->req_md5_lookup(sk, req_to_sk(req)))
573 newtp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
574#endif
575 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len)
576 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len;
577 newtp->rx_opt.mss_clamp = req->mss;
578 tcp_ecn_openreq_child(newtp, req);
579 newtp->fastopen_req = NULL;
580 RCU_INIT_POINTER(newtp->fastopen_rsk, NULL);
581
582 newtp->bpf_chg_cc_inprogress = 0;
583 tcp_bpf_clone(sk, newsk);
584
585 __TCP_INC_STATS(sock_net(sk), TCP_MIB_PASSIVEOPENS);
586
587 return newsk;
588}
589EXPORT_SYMBOL(tcp_create_openreq_child);
590
591/*
592 * Process an incoming packet for SYN_RECV sockets represented as a
593 * request_sock. Normally sk is the listener socket but for TFO it
594 * points to the child socket.
595 *
596 * XXX (TFO) - The current impl contains a special check for ack
597 * validation and inside tcp_v4_reqsk_send_ack(). Can we do better?
598 *
599 * We don't need to initialize tmp_opt.sack_ok as we don't use the results
600 */
601
602struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb,
603 struct request_sock *req,
604 bool fastopen, bool *req_stolen)
605{
606 struct tcp_options_received tmp_opt;
607 struct sock *child;
608 const struct tcphdr *th = tcp_hdr(skb);
609 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK);
610 bool paws_reject = false;
611 bool own_req;
612
613 tmp_opt.saw_tstamp = 0;
614 if (th->doff > (sizeof(struct tcphdr)>>2)) {
615 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0, NULL);
616
617 if (tmp_opt.saw_tstamp) {
618 tmp_opt.ts_recent = req->ts_recent;
619 if (tmp_opt.rcv_tsecr)
620 tmp_opt.rcv_tsecr -= tcp_rsk(req)->ts_off;
621 /* We do not store true stamp, but it is not required,
622 * it can be estimated (approximately)
623 * from another data.
624 */
625 tmp_opt.ts_recent_stamp = ktime_get_seconds() - reqsk_timeout(req, TCP_RTO_MAX) / HZ;
626 paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
627 }
628 }
629
630 /* Check for pure retransmitted SYN. */
631 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn &&
632 flg == TCP_FLAG_SYN &&
633 !paws_reject) {
634 /*
635 * RFC793 draws (Incorrectly! It was fixed in RFC1122)
636 * this case on figure 6 and figure 8, but formal
637 * protocol description says NOTHING.
638 * To be more exact, it says that we should send ACK,
639 * because this segment (at least, if it has no data)
640 * is out of window.
641 *
642 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT
643 * describe SYN-RECV state. All the description
644 * is wrong, we cannot believe to it and should
645 * rely only on common sense and implementation
646 * experience.
647 *
648 * Enforce "SYN-ACK" according to figure 8, figure 6
649 * of RFC793, fixed by RFC1122.
650 *
651 * Note that even if there is new data in the SYN packet
652 * they will be thrown away too.
653 *
654 * Reset timer after retransmitting SYNACK, similar to
655 * the idea of fast retransmit in recovery.
656 */
657 if (!tcp_oow_rate_limited(sock_net(sk), skb,
658 LINUX_MIB_TCPACKSKIPPEDSYNRECV,
659 &tcp_rsk(req)->last_oow_ack_time) &&
660
661 !inet_rtx_syn_ack(sk, req)) {
662 unsigned long expires = jiffies;
663
664 expires += reqsk_timeout(req, TCP_RTO_MAX);
665 if (!fastopen)
666 mod_timer_pending(&req->rsk_timer, expires);
667 else
668 req->rsk_timer.expires = expires;
669 }
670 return NULL;
671 }
672
673 /* Further reproduces section "SEGMENT ARRIVES"
674 for state SYN-RECEIVED of RFC793.
675 It is broken, however, it does not work only
676 when SYNs are crossed.
677
678 You would think that SYN crossing is impossible here, since
679 we should have a SYN_SENT socket (from connect()) on our end,
680 but this is not true if the crossed SYNs were sent to both
681 ends by a malicious third party. We must defend against this,
682 and to do that we first verify the ACK (as per RFC793, page
683 36) and reset if it is invalid. Is this a true full defense?
684 To convince ourselves, let us consider a way in which the ACK
685 test can still pass in this 'malicious crossed SYNs' case.
686 Malicious sender sends identical SYNs (and thus identical sequence
687 numbers) to both A and B:
688
689 A: gets SYN, seq=7
690 B: gets SYN, seq=7
691
692 By our good fortune, both A and B select the same initial
693 send sequence number of seven :-)
694
695 A: sends SYN|ACK, seq=7, ack_seq=8
696 B: sends SYN|ACK, seq=7, ack_seq=8
697
698 So we are now A eating this SYN|ACK, ACK test passes. So
699 does sequence test, SYN is truncated, and thus we consider
700 it a bare ACK.
701
702 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this
703 bare ACK. Otherwise, we create an established connection. Both
704 ends (listening sockets) accept the new incoming connection and try
705 to talk to each other. 8-)
706
707 Note: This case is both harmless, and rare. Possibility is about the
708 same as us discovering intelligent life on another plant tomorrow.
709
710 But generally, we should (RFC lies!) to accept ACK
711 from SYNACK both here and in tcp_rcv_state_process().
712 tcp_rcv_state_process() does not, hence, we do not too.
713
714 Note that the case is absolutely generic:
715 we cannot optimize anything here without
716 violating protocol. All the checks must be made
717 before attempt to create socket.
718 */
719
720 /* RFC793 page 36: "If the connection is in any non-synchronized state ...
721 * and the incoming segment acknowledges something not yet
722 * sent (the segment carries an unacceptable ACK) ...
723 * a reset is sent."
724 *
725 * Invalid ACK: reset will be sent by listening socket.
726 * Note that the ACK validity check for a Fast Open socket is done
727 * elsewhere and is checked directly against the child socket rather
728 * than req because user data may have been sent out.
729 */
730 if ((flg & TCP_FLAG_ACK) && !fastopen &&
731 (TCP_SKB_CB(skb)->ack_seq !=
732 tcp_rsk(req)->snt_isn + 1))
733 return sk;
734
735 /* Also, it would be not so bad idea to check rcv_tsecr, which
736 * is essentially ACK extension and too early or too late values
737 * should cause reset in unsynchronized states.
738 */
739
740 /* RFC793: "first check sequence number". */
741
742 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
743 tcp_rsk(req)->rcv_nxt, tcp_rsk(req)->rcv_nxt + req->rsk_rcv_wnd)) {
744 /* Out of window: send ACK and drop. */
745 if (!(flg & TCP_FLAG_RST) &&
746 !tcp_oow_rate_limited(sock_net(sk), skb,
747 LINUX_MIB_TCPACKSKIPPEDSYNRECV,
748 &tcp_rsk(req)->last_oow_ack_time))
749 req->rsk_ops->send_ack(sk, skb, req);
750 if (paws_reject)
751 __NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
752 return NULL;
753 }
754
755 /* In sequence, PAWS is OK. */
756
757 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_nxt))
758 req->ts_recent = tmp_opt.rcv_tsval;
759
760 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) {
761 /* Truncate SYN, it is out of window starting
762 at tcp_rsk(req)->rcv_isn + 1. */
763 flg &= ~TCP_FLAG_SYN;
764 }
765
766 /* RFC793: "second check the RST bit" and
767 * "fourth, check the SYN bit"
768 */
769 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) {
770 __TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
771 goto embryonic_reset;
772 }
773
774 /* ACK sequence verified above, just make sure ACK is
775 * set. If ACK not set, just silently drop the packet.
776 *
777 * XXX (TFO) - if we ever allow "data after SYN", the
778 * following check needs to be removed.
779 */
780 if (!(flg & TCP_FLAG_ACK))
781 return NULL;
782
783 /* For Fast Open no more processing is needed (sk is the
784 * child socket).
785 */
786 if (fastopen)
787 return sk;
788
789 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */
790 if (req->num_timeout < inet_csk(sk)->icsk_accept_queue.rskq_defer_accept &&
791 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) {
792 inet_rsk(req)->acked = 1;
793 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP);
794 return NULL;
795 }
796
797 /* OK, ACK is valid, create big socket and
798 * feed this segment to it. It will repeat all
799 * the tests. THIS SEGMENT MUST MOVE SOCKET TO
800 * ESTABLISHED STATE. If it will be dropped after
801 * socket is created, wait for troubles.
802 */
803 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL,
804 req, &own_req);
805 if (!child)
806 goto listen_overflow;
807
808 if (own_req && rsk_drop_req(req)) {
809 reqsk_queue_removed(&inet_csk(req->rsk_listener)->icsk_accept_queue, req);
810 inet_csk_reqsk_queue_drop_and_put(req->rsk_listener, req);
811 return child;
812 }
813
814 sock_rps_save_rxhash(child, skb);
815 tcp_synack_rtt_meas(child, req);
816 *req_stolen = !own_req;
817 return inet_csk_complete_hashdance(sk, child, req, own_req);
818
819listen_overflow:
820 if (sk != req->rsk_listener)
821 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMIGRATEREQFAILURE);
822
823 if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_abort_on_overflow)) {
824 inet_rsk(req)->acked = 1;
825 return NULL;
826 }
827
828embryonic_reset:
829 if (!(flg & TCP_FLAG_RST)) {
830 /* Received a bad SYN pkt - for TFO We try not to reset
831 * the local connection unless it's really necessary to
832 * avoid becoming vulnerable to outside attack aiming at
833 * resetting legit local connections.
834 */
835 req->rsk_ops->send_reset(sk, skb);
836 } else if (fastopen) { /* received a valid RST pkt */
837 reqsk_fastopen_remove(sk, req, true);
838 tcp_reset(sk, skb);
839 }
840 if (!fastopen) {
841 bool unlinked = inet_csk_reqsk_queue_drop(sk, req);
842
843 if (unlinked)
844 __NET_INC_STATS(sock_net(sk), LINUX_MIB_EMBRYONICRSTS);
845 *req_stolen = !unlinked;
846 }
847 return NULL;
848}
849EXPORT_SYMBOL(tcp_check_req);
850
851/*
852 * Queue segment on the new socket if the new socket is active,
853 * otherwise we just shortcircuit this and continue with
854 * the new socket.
855 *
856 * For the vast majority of cases child->sk_state will be TCP_SYN_RECV
857 * when entering. But other states are possible due to a race condition
858 * where after __inet_lookup_established() fails but before the listener
859 * locked is obtained, other packets cause the same connection to
860 * be created.
861 */
862
863int tcp_child_process(struct sock *parent, struct sock *child,
864 struct sk_buff *skb)
865 __releases(&((child)->sk_lock.slock))
866{
867 int ret = 0;
868 int state = child->sk_state;
869
870 /* record sk_napi_id and sk_rx_queue_mapping of child. */
871 sk_mark_napi_id_set(child, skb);
872
873 tcp_segs_in(tcp_sk(child), skb);
874 if (!sock_owned_by_user(child)) {
875 ret = tcp_rcv_state_process(child, skb);
876 /* Wakeup parent, send SIGIO */
877 if (state == TCP_SYN_RECV && child->sk_state != state)
878 parent->sk_data_ready(parent);
879 } else {
880 /* Alas, it is possible again, because we do lookup
881 * in main socket hash table and lock on listening
882 * socket does not protect us more.
883 */
884 __sk_add_backlog(child, skb);
885 }
886
887 bh_unlock_sock(child);
888 sock_put(child);
889 return ret;
890}
891EXPORT_SYMBOL(tcp_child_process);