Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
6 *
7 * Implementation of the Transmission Control Protocol(TCP).
8 *
9 * Authors: Ross Biro
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
20 */
21
22#include <net/tcp.h>
23#include <net/xfrm.h>
24#include <net/busy_poll.h>
25
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);
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
54static void twsk_rcv_nxt_update(struct tcp_timewait_sock *tcptw, u32 seq)
55{
56#ifdef CONFIG_TCP_AO
57 struct tcp_ao_info *ao;
58
59 ao = rcu_dereference(tcptw->ao_info);
60 if (unlikely(ao && seq < tcptw->tw_rcv_nxt))
61 WRITE_ONCE(ao->rcv_sne, ao->rcv_sne + 1);
62#endif
63 tcptw->tw_rcv_nxt = seq;
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(twsk_net(tw), skb, &tmp_opt, 0, NULL);
107
108 if (tmp_opt.saw_tstamp) {
109 if (tmp_opt.rcv_tsecr)
110 tmp_opt.rcv_tsecr -= tcptw->tw_ts_offset;
111 tmp_opt.ts_recent = tcptw->tw_ts_recent;
112 tmp_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp;
113 paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
114 }
115 }
116
117 if (tw->tw_substate == TCP_FIN_WAIT2) {
118 /* Just repeat all the checks of tcp_rcv_state_process() */
119
120 /* Out of window, send ACK */
121 if (paws_reject ||
122 !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
123 tcptw->tw_rcv_nxt,
124 tcptw->tw_rcv_nxt + tcptw->tw_rcv_wnd))
125 return tcp_timewait_check_oow_rate_limit(
126 tw, skb, LINUX_MIB_TCPACKSKIPPEDFINWAIT2);
127
128 if (th->rst)
129 goto kill;
130
131 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt))
132 return TCP_TW_RST;
133
134 /* Dup ACK? */
135 if (!th->ack ||
136 !after(TCP_SKB_CB(skb)->end_seq, tcptw->tw_rcv_nxt) ||
137 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) {
138 inet_twsk_put(tw);
139 return TCP_TW_SUCCESS;
140 }
141
142 /* New data or FIN. If new data arrive after half-duplex close,
143 * reset.
144 */
145 if (!th->fin ||
146 TCP_SKB_CB(skb)->end_seq != tcptw->tw_rcv_nxt + 1)
147 return TCP_TW_RST;
148
149 /* FIN arrived, enter true time-wait state. */
150 tw->tw_substate = TCP_TIME_WAIT;
151 twsk_rcv_nxt_update(tcptw, TCP_SKB_CB(skb)->end_seq);
152
153 if (tmp_opt.saw_tstamp) {
154 tcptw->tw_ts_recent_stamp = ktime_get_seconds();
155 tcptw->tw_ts_recent = tmp_opt.rcv_tsval;
156 }
157
158 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
159 return TCP_TW_ACK;
160 }
161
162 /*
163 * Now real TIME-WAIT state.
164 *
165 * RFC 1122:
166 * "When a connection is [...] on TIME-WAIT state [...]
167 * [a TCP] MAY accept a new SYN from the remote TCP to
168 * reopen the connection directly, if it:
169 *
170 * (1) assigns its initial sequence number for the new
171 * connection to be larger than the largest sequence
172 * number it used on the previous connection incarnation,
173 * and
174 *
175 * (2) returns to TIME-WAIT state if the SYN turns out
176 * to be an old duplicate".
177 */
178
179 if (!paws_reject &&
180 (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt &&
181 (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) {
182 /* In window segment, it may be only reset or bare ack. */
183
184 if (th->rst) {
185 /* This is TIME_WAIT assassination, in two flavors.
186 * Oh well... nobody has a sufficient solution to this
187 * protocol bug yet.
188 */
189 if (!READ_ONCE(twsk_net(tw)->ipv4.sysctl_tcp_rfc1337)) {
190kill:
191 inet_twsk_deschedule_put(tw);
192 return TCP_TW_SUCCESS;
193 }
194 } else {
195 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
196 }
197
198 if (tmp_opt.saw_tstamp) {
199 tcptw->tw_ts_recent = tmp_opt.rcv_tsval;
200 tcptw->tw_ts_recent_stamp = ktime_get_seconds();
201 }
202
203 inet_twsk_put(tw);
204 return TCP_TW_SUCCESS;
205 }
206
207 /* Out of window segment.
208
209 All the segments are ACKed immediately.
210
211 The only exception is new SYN. We accept it, if it is
212 not old duplicate and we are not in danger to be killed
213 by delayed old duplicates. RFC check is that it has
214 newer sequence number works at rates <40Mbit/sec.
215 However, if paws works, it is reliable AND even more,
216 we even may relax silly seq space cutoff.
217
218 RED-PEN: we violate main RFC requirement, if this SYN will appear
219 old duplicate (i.e. we receive RST in reply to SYN-ACK),
220 we must return socket to time-wait state. It is not good,
221 but not fatal yet.
222 */
223
224 if (th->syn && !th->rst && !th->ack && !paws_reject &&
225 (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) ||
226 (tmp_opt.saw_tstamp &&
227 (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) {
228 u32 isn = tcptw->tw_snd_nxt + 65535 + 2;
229 if (isn == 0)
230 isn++;
231 TCP_SKB_CB(skb)->tcp_tw_isn = isn;
232 return TCP_TW_SYN;
233 }
234
235 if (paws_reject)
236 __NET_INC_STATS(twsk_net(tw), LINUX_MIB_PAWSESTABREJECTED);
237
238 if (!th->rst) {
239 /* In this case we must reset the TIMEWAIT timer.
240 *
241 * If it is ACKless SYN it may be both old duplicate
242 * and new good SYN with random sequence number <rcv_nxt.
243 * Do not reschedule in the last case.
244 */
245 if (paws_reject || th->ack)
246 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
247
248 return tcp_timewait_check_oow_rate_limit(
249 tw, skb, LINUX_MIB_TCPACKSKIPPEDTIMEWAIT);
250 }
251 inet_twsk_put(tw);
252 return TCP_TW_SUCCESS;
253}
254EXPORT_SYMBOL(tcp_timewait_state_process);
255
256static void tcp_time_wait_init(struct sock *sk, struct tcp_timewait_sock *tcptw)
257{
258#ifdef CONFIG_TCP_MD5SIG
259 const struct tcp_sock *tp = tcp_sk(sk);
260 struct tcp_md5sig_key *key;
261
262 /*
263 * The timewait bucket does not have the key DB from the
264 * sock structure. We just make a quick copy of the
265 * md5 key being used (if indeed we are using one)
266 * so the timewait ack generating code has the key.
267 */
268 tcptw->tw_md5_key = NULL;
269 if (!static_branch_unlikely(&tcp_md5_needed.key))
270 return;
271
272 key = tp->af_specific->md5_lookup(sk, sk);
273 if (key) {
274 tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC);
275 if (!tcptw->tw_md5_key)
276 return;
277 if (!static_key_fast_inc_not_disabled(&tcp_md5_needed.key.key))
278 goto out_free;
279 tcp_md5_add_sigpool();
280 }
281 return;
282out_free:
283 WARN_ON_ONCE(1);
284 kfree(tcptw->tw_md5_key);
285 tcptw->tw_md5_key = NULL;
286#endif
287}
288
289/*
290 * Move a socket to time-wait or dead fin-wait-2 state.
291 */
292void tcp_time_wait(struct sock *sk, int state, int timeo)
293{
294 const struct inet_connection_sock *icsk = inet_csk(sk);
295 struct tcp_sock *tp = tcp_sk(sk);
296 struct net *net = sock_net(sk);
297 struct inet_timewait_sock *tw;
298
299 tw = inet_twsk_alloc(sk, &net->ipv4.tcp_death_row, state);
300
301 if (tw) {
302 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
303 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1);
304
305 tw->tw_transparent = inet_test_bit(TRANSPARENT, sk);
306 tw->tw_mark = sk->sk_mark;
307 tw->tw_priority = READ_ONCE(sk->sk_priority);
308 tw->tw_rcv_wscale = tp->rx_opt.rcv_wscale;
309 tcptw->tw_rcv_nxt = tp->rcv_nxt;
310 tcptw->tw_snd_nxt = tp->snd_nxt;
311 tcptw->tw_rcv_wnd = tcp_receive_window(tp);
312 tcptw->tw_ts_recent = tp->rx_opt.ts_recent;
313 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp;
314 tcptw->tw_ts_offset = tp->tsoffset;
315 tw->tw_usec_ts = tp->tcp_usec_ts;
316 tcptw->tw_last_oow_ack_time = 0;
317 tcptw->tw_tx_delay = tp->tcp_tx_delay;
318 tw->tw_txhash = sk->sk_txhash;
319#if IS_ENABLED(CONFIG_IPV6)
320 if (tw->tw_family == PF_INET6) {
321 struct ipv6_pinfo *np = inet6_sk(sk);
322
323 tw->tw_v6_daddr = sk->sk_v6_daddr;
324 tw->tw_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
325 tw->tw_tclass = np->tclass;
326 tw->tw_flowlabel = be32_to_cpu(np->flow_label & IPV6_FLOWLABEL_MASK);
327 tw->tw_ipv6only = sk->sk_ipv6only;
328 }
329#endif
330
331 tcp_time_wait_init(sk, tcptw);
332 tcp_ao_time_wait(tcptw, tp);
333
334 /* Get the TIME_WAIT timeout firing. */
335 if (timeo < rto)
336 timeo = rto;
337
338 if (state == TCP_TIME_WAIT)
339 timeo = TCP_TIMEWAIT_LEN;
340
341 /* tw_timer is pinned, so we need to make sure BH are disabled
342 * in following section, otherwise timer handler could run before
343 * we complete the initialization.
344 */
345 local_bh_disable();
346 inet_twsk_schedule(tw, timeo);
347 /* Linkage updates.
348 * Note that access to tw after this point is illegal.
349 */
350 inet_twsk_hashdance(tw, sk, net->ipv4.tcp_death_row.hashinfo);
351 local_bh_enable();
352 } else {
353 /* Sorry, if we're out of memory, just CLOSE this
354 * socket up. We've got bigger problems than
355 * non-graceful socket closings.
356 */
357 NET_INC_STATS(net, LINUX_MIB_TCPTIMEWAITOVERFLOW);
358 }
359
360 tcp_update_metrics(sk);
361 tcp_done(sk);
362}
363EXPORT_SYMBOL(tcp_time_wait);
364
365#ifdef CONFIG_TCP_MD5SIG
366static void tcp_md5_twsk_free_rcu(struct rcu_head *head)
367{
368 struct tcp_md5sig_key *key;
369
370 key = container_of(head, struct tcp_md5sig_key, rcu);
371 kfree(key);
372 static_branch_slow_dec_deferred(&tcp_md5_needed);
373 tcp_md5_release_sigpool();
374}
375#endif
376
377void tcp_twsk_destructor(struct sock *sk)
378{
379#ifdef CONFIG_TCP_MD5SIG
380 if (static_branch_unlikely(&tcp_md5_needed.key)) {
381 struct tcp_timewait_sock *twsk = tcp_twsk(sk);
382
383 if (twsk->tw_md5_key)
384 call_rcu(&twsk->tw_md5_key->rcu, tcp_md5_twsk_free_rcu);
385 }
386#endif
387 tcp_ao_destroy_sock(sk, true);
388}
389EXPORT_SYMBOL_GPL(tcp_twsk_destructor);
390
391void tcp_twsk_purge(struct list_head *net_exit_list, int family)
392{
393 bool purged_once = false;
394 struct net *net;
395
396 list_for_each_entry(net, net_exit_list, exit_list) {
397 if (net->ipv4.tcp_death_row.hashinfo->pernet) {
398 /* Even if tw_refcount == 1, we must clean up kernel reqsk */
399 inet_twsk_purge(net->ipv4.tcp_death_row.hashinfo, family);
400 } else if (!purged_once) {
401 /* The last refcount is decremented in tcp_sk_exit_batch() */
402 if (refcount_read(&net->ipv4.tcp_death_row.tw_refcount) == 1)
403 continue;
404
405 inet_twsk_purge(&tcp_hashinfo, family);
406 purged_once = true;
407 }
408 }
409}
410EXPORT_SYMBOL_GPL(tcp_twsk_purge);
411
412/* Warning : This function is called without sk_listener being locked.
413 * Be sure to read socket fields once, as their value could change under us.
414 */
415void tcp_openreq_init_rwin(struct request_sock *req,
416 const struct sock *sk_listener,
417 const struct dst_entry *dst)
418{
419 struct inet_request_sock *ireq = inet_rsk(req);
420 const struct tcp_sock *tp = tcp_sk(sk_listener);
421 int full_space = tcp_full_space(sk_listener);
422 u32 window_clamp;
423 __u8 rcv_wscale;
424 u32 rcv_wnd;
425 int mss;
426
427 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
428 window_clamp = READ_ONCE(tp->window_clamp);
429 /* Set this up on the first call only */
430 req->rsk_window_clamp = window_clamp ? : dst_metric(dst, RTAX_WINDOW);
431
432 /* limit the window selection if the user enforce a smaller rx buffer */
433 if (sk_listener->sk_userlocks & SOCK_RCVBUF_LOCK &&
434 (req->rsk_window_clamp > full_space || req->rsk_window_clamp == 0))
435 req->rsk_window_clamp = full_space;
436
437 rcv_wnd = tcp_rwnd_init_bpf((struct sock *)req);
438 if (rcv_wnd == 0)
439 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
440 else if (full_space < rcv_wnd * mss)
441 full_space = rcv_wnd * mss;
442
443 /* tcp_full_space because it is guaranteed to be the first packet */
444 tcp_select_initial_window(sk_listener, full_space,
445 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
446 &req->rsk_rcv_wnd,
447 &req->rsk_window_clamp,
448 ireq->wscale_ok,
449 &rcv_wscale,
450 rcv_wnd);
451 ireq->rcv_wscale = rcv_wscale;
452}
453EXPORT_SYMBOL(tcp_openreq_init_rwin);
454
455static void tcp_ecn_openreq_child(struct tcp_sock *tp,
456 const struct request_sock *req)
457{
458 tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0;
459}
460
461void tcp_ca_openreq_child(struct sock *sk, const struct dst_entry *dst)
462{
463 struct inet_connection_sock *icsk = inet_csk(sk);
464 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
465 bool ca_got_dst = false;
466
467 if (ca_key != TCP_CA_UNSPEC) {
468 const struct tcp_congestion_ops *ca;
469
470 rcu_read_lock();
471 ca = tcp_ca_find_key(ca_key);
472 if (likely(ca && bpf_try_module_get(ca, ca->owner))) {
473 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
474 icsk->icsk_ca_ops = ca;
475 ca_got_dst = true;
476 }
477 rcu_read_unlock();
478 }
479
480 /* If no valid choice made yet, assign current system default ca. */
481 if (!ca_got_dst &&
482 (!icsk->icsk_ca_setsockopt ||
483 !bpf_try_module_get(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner)))
484 tcp_assign_congestion_control(sk);
485
486 tcp_set_ca_state(sk, TCP_CA_Open);
487}
488EXPORT_SYMBOL_GPL(tcp_ca_openreq_child);
489
490static void smc_check_reset_syn_req(const struct tcp_sock *oldtp,
491 struct request_sock *req,
492 struct tcp_sock *newtp)
493{
494#if IS_ENABLED(CONFIG_SMC)
495 struct inet_request_sock *ireq;
496
497 if (static_branch_unlikely(&tcp_have_smc)) {
498 ireq = inet_rsk(req);
499 if (oldtp->syn_smc && !ireq->smc_ok)
500 newtp->syn_smc = 0;
501 }
502#endif
503}
504
505/* This is not only more efficient than what we used to do, it eliminates
506 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM
507 *
508 * Actually, we could lots of memory writes here. tp of listening
509 * socket contains all necessary default parameters.
510 */
511struct sock *tcp_create_openreq_child(const struct sock *sk,
512 struct request_sock *req,
513 struct sk_buff *skb)
514{
515 struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC);
516 const struct inet_request_sock *ireq = inet_rsk(req);
517 struct tcp_request_sock *treq = tcp_rsk(req);
518 struct inet_connection_sock *newicsk;
519 const struct tcp_sock *oldtp;
520 struct tcp_sock *newtp;
521 u32 seq;
522#ifdef CONFIG_TCP_AO
523 struct tcp_ao_key *ao_key;
524#endif
525
526 if (!newsk)
527 return NULL;
528
529 newicsk = inet_csk(newsk);
530 newtp = tcp_sk(newsk);
531 oldtp = tcp_sk(sk);
532
533 smc_check_reset_syn_req(oldtp, req, newtp);
534
535 /* Now setup tcp_sock */
536 newtp->pred_flags = 0;
537
538 seq = treq->rcv_isn + 1;
539 newtp->rcv_wup = seq;
540 WRITE_ONCE(newtp->copied_seq, seq);
541 WRITE_ONCE(newtp->rcv_nxt, seq);
542 newtp->segs_in = 1;
543
544 seq = treq->snt_isn + 1;
545 newtp->snd_sml = newtp->snd_una = seq;
546 WRITE_ONCE(newtp->snd_nxt, seq);
547 newtp->snd_up = seq;
548
549 INIT_LIST_HEAD(&newtp->tsq_node);
550 INIT_LIST_HEAD(&newtp->tsorted_sent_queue);
551
552 tcp_init_wl(newtp, treq->rcv_isn);
553
554 minmax_reset(&newtp->rtt_min, tcp_jiffies32, ~0U);
555 newicsk->icsk_ack.lrcvtime = tcp_jiffies32;
556
557 newtp->lsndtime = tcp_jiffies32;
558 newsk->sk_txhash = READ_ONCE(treq->txhash);
559 newtp->total_retrans = req->num_retrans;
560
561 tcp_init_xmit_timers(newsk);
562 WRITE_ONCE(newtp->write_seq, newtp->pushed_seq = treq->snt_isn + 1);
563
564 if (sock_flag(newsk, SOCK_KEEPOPEN))
565 inet_csk_reset_keepalive_timer(newsk,
566 keepalive_time_when(newtp));
567
568 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok;
569 newtp->rx_opt.sack_ok = ireq->sack_ok;
570 newtp->window_clamp = req->rsk_window_clamp;
571 newtp->rcv_ssthresh = req->rsk_rcv_wnd;
572 newtp->rcv_wnd = req->rsk_rcv_wnd;
573 newtp->rx_opt.wscale_ok = ireq->wscale_ok;
574 if (newtp->rx_opt.wscale_ok) {
575 newtp->rx_opt.snd_wscale = ireq->snd_wscale;
576 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale;
577 } else {
578 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0;
579 newtp->window_clamp = min(newtp->window_clamp, 65535U);
580 }
581 newtp->snd_wnd = ntohs(tcp_hdr(skb)->window) << newtp->rx_opt.snd_wscale;
582 newtp->max_window = newtp->snd_wnd;
583
584 if (newtp->rx_opt.tstamp_ok) {
585 newtp->tcp_usec_ts = treq->req_usec_ts;
586 newtp->rx_opt.ts_recent = READ_ONCE(req->ts_recent);
587 newtp->rx_opt.ts_recent_stamp = ktime_get_seconds();
588 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
589 } else {
590 newtp->tcp_usec_ts = 0;
591 newtp->rx_opt.ts_recent_stamp = 0;
592 newtp->tcp_header_len = sizeof(struct tcphdr);
593 }
594 if (req->num_timeout) {
595 newtp->total_rto = req->num_timeout;
596 newtp->undo_marker = treq->snt_isn;
597 if (newtp->tcp_usec_ts) {
598 newtp->retrans_stamp = treq->snt_synack;
599 newtp->total_rto_time = (u32)(tcp_clock_us() -
600 newtp->retrans_stamp) / USEC_PER_MSEC;
601 } else {
602 newtp->retrans_stamp = div_u64(treq->snt_synack,
603 USEC_PER_SEC / TCP_TS_HZ);
604 newtp->total_rto_time = tcp_clock_ms() -
605 newtp->retrans_stamp;
606 }
607 newtp->total_rto_recoveries = 1;
608 }
609 newtp->tsoffset = treq->ts_off;
610#ifdef CONFIG_TCP_MD5SIG
611 newtp->md5sig_info = NULL; /*XXX*/
612#endif
613#ifdef CONFIG_TCP_AO
614 newtp->ao_info = NULL;
615 ao_key = treq->af_specific->ao_lookup(sk, req,
616 tcp_rsk(req)->ao_keyid, -1);
617 if (ao_key)
618 newtp->tcp_header_len += tcp_ao_len_aligned(ao_key);
619 #endif
620 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len)
621 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len;
622 newtp->rx_opt.mss_clamp = req->mss;
623 tcp_ecn_openreq_child(newtp, req);
624 newtp->fastopen_req = NULL;
625 RCU_INIT_POINTER(newtp->fastopen_rsk, NULL);
626
627 newtp->bpf_chg_cc_inprogress = 0;
628 tcp_bpf_clone(sk, newsk);
629
630 __TCP_INC_STATS(sock_net(sk), TCP_MIB_PASSIVEOPENS);
631
632 return newsk;
633}
634EXPORT_SYMBOL(tcp_create_openreq_child);
635
636/*
637 * Process an incoming packet for SYN_RECV sockets represented as a
638 * request_sock. Normally sk is the listener socket but for TFO it
639 * points to the child socket.
640 *
641 * XXX (TFO) - The current impl contains a special check for ack
642 * validation and inside tcp_v4_reqsk_send_ack(). Can we do better?
643 *
644 * We don't need to initialize tmp_opt.sack_ok as we don't use the results
645 *
646 * Note: If @fastopen is true, this can be called from process context.
647 * Otherwise, this is from BH context.
648 */
649
650struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb,
651 struct request_sock *req,
652 bool fastopen, bool *req_stolen)
653{
654 struct tcp_options_received tmp_opt;
655 struct sock *child;
656 const struct tcphdr *th = tcp_hdr(skb);
657 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK);
658 bool paws_reject = false;
659 bool own_req;
660
661 tmp_opt.saw_tstamp = 0;
662 if (th->doff > (sizeof(struct tcphdr)>>2)) {
663 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0, NULL);
664
665 if (tmp_opt.saw_tstamp) {
666 tmp_opt.ts_recent = READ_ONCE(req->ts_recent);
667 if (tmp_opt.rcv_tsecr)
668 tmp_opt.rcv_tsecr -= tcp_rsk(req)->ts_off;
669 /* We do not store true stamp, but it is not required,
670 * it can be estimated (approximately)
671 * from another data.
672 */
673 tmp_opt.ts_recent_stamp = ktime_get_seconds() - reqsk_timeout(req, TCP_RTO_MAX) / HZ;
674 paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
675 }
676 }
677
678 /* Check for pure retransmitted SYN. */
679 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn &&
680 flg == TCP_FLAG_SYN &&
681 !paws_reject) {
682 /*
683 * RFC793 draws (Incorrectly! It was fixed in RFC1122)
684 * this case on figure 6 and figure 8, but formal
685 * protocol description says NOTHING.
686 * To be more exact, it says that we should send ACK,
687 * because this segment (at least, if it has no data)
688 * is out of window.
689 *
690 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT
691 * describe SYN-RECV state. All the description
692 * is wrong, we cannot believe to it and should
693 * rely only on common sense and implementation
694 * experience.
695 *
696 * Enforce "SYN-ACK" according to figure 8, figure 6
697 * of RFC793, fixed by RFC1122.
698 *
699 * Note that even if there is new data in the SYN packet
700 * they will be thrown away too.
701 *
702 * Reset timer after retransmitting SYNACK, similar to
703 * the idea of fast retransmit in recovery.
704 */
705 if (!tcp_oow_rate_limited(sock_net(sk), skb,
706 LINUX_MIB_TCPACKSKIPPEDSYNRECV,
707 &tcp_rsk(req)->last_oow_ack_time) &&
708
709 !inet_rtx_syn_ack(sk, req)) {
710 unsigned long expires = jiffies;
711
712 expires += reqsk_timeout(req, TCP_RTO_MAX);
713 if (!fastopen)
714 mod_timer_pending(&req->rsk_timer, expires);
715 else
716 req->rsk_timer.expires = expires;
717 }
718 return NULL;
719 }
720
721 /* Further reproduces section "SEGMENT ARRIVES"
722 for state SYN-RECEIVED of RFC793.
723 It is broken, however, it does not work only
724 when SYNs are crossed.
725
726 You would think that SYN crossing is impossible here, since
727 we should have a SYN_SENT socket (from connect()) on our end,
728 but this is not true if the crossed SYNs were sent to both
729 ends by a malicious third party. We must defend against this,
730 and to do that we first verify the ACK (as per RFC793, page
731 36) and reset if it is invalid. Is this a true full defense?
732 To convince ourselves, let us consider a way in which the ACK
733 test can still pass in this 'malicious crossed SYNs' case.
734 Malicious sender sends identical SYNs (and thus identical sequence
735 numbers) to both A and B:
736
737 A: gets SYN, seq=7
738 B: gets SYN, seq=7
739
740 By our good fortune, both A and B select the same initial
741 send sequence number of seven :-)
742
743 A: sends SYN|ACK, seq=7, ack_seq=8
744 B: sends SYN|ACK, seq=7, ack_seq=8
745
746 So we are now A eating this SYN|ACK, ACK test passes. So
747 does sequence test, SYN is truncated, and thus we consider
748 it a bare ACK.
749
750 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this
751 bare ACK. Otherwise, we create an established connection. Both
752 ends (listening sockets) accept the new incoming connection and try
753 to talk to each other. 8-)
754
755 Note: This case is both harmless, and rare. Possibility is about the
756 same as us discovering intelligent life on another plant tomorrow.
757
758 But generally, we should (RFC lies!) to accept ACK
759 from SYNACK both here and in tcp_rcv_state_process().
760 tcp_rcv_state_process() does not, hence, we do not too.
761
762 Note that the case is absolutely generic:
763 we cannot optimize anything here without
764 violating protocol. All the checks must be made
765 before attempt to create socket.
766 */
767
768 /* RFC793 page 36: "If the connection is in any non-synchronized state ...
769 * and the incoming segment acknowledges something not yet
770 * sent (the segment carries an unacceptable ACK) ...
771 * a reset is sent."
772 *
773 * Invalid ACK: reset will be sent by listening socket.
774 * Note that the ACK validity check for a Fast Open socket is done
775 * elsewhere and is checked directly against the child socket rather
776 * than req because user data may have been sent out.
777 */
778 if ((flg & TCP_FLAG_ACK) && !fastopen &&
779 (TCP_SKB_CB(skb)->ack_seq !=
780 tcp_rsk(req)->snt_isn + 1))
781 return sk;
782
783 /* Also, it would be not so bad idea to check rcv_tsecr, which
784 * is essentially ACK extension and too early or too late values
785 * should cause reset in unsynchronized states.
786 */
787
788 /* RFC793: "first check sequence number". */
789
790 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
791 tcp_rsk(req)->rcv_nxt, tcp_rsk(req)->rcv_nxt + req->rsk_rcv_wnd)) {
792 /* Out of window: send ACK and drop. */
793 if (!(flg & TCP_FLAG_RST) &&
794 !tcp_oow_rate_limited(sock_net(sk), skb,
795 LINUX_MIB_TCPACKSKIPPEDSYNRECV,
796 &tcp_rsk(req)->last_oow_ack_time))
797 req->rsk_ops->send_ack(sk, skb, req);
798 if (paws_reject)
799 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
800 return NULL;
801 }
802
803 /* In sequence, PAWS is OK. */
804
805 /* TODO: We probably should defer ts_recent change once
806 * we take ownership of @req.
807 */
808 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_nxt))
809 WRITE_ONCE(req->ts_recent, tmp_opt.rcv_tsval);
810
811 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) {
812 /* Truncate SYN, it is out of window starting
813 at tcp_rsk(req)->rcv_isn + 1. */
814 flg &= ~TCP_FLAG_SYN;
815 }
816
817 /* RFC793: "second check the RST bit" and
818 * "fourth, check the SYN bit"
819 */
820 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) {
821 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
822 goto embryonic_reset;
823 }
824
825 /* ACK sequence verified above, just make sure ACK is
826 * set. If ACK not set, just silently drop the packet.
827 *
828 * XXX (TFO) - if we ever allow "data after SYN", the
829 * following check needs to be removed.
830 */
831 if (!(flg & TCP_FLAG_ACK))
832 return NULL;
833
834 /* For Fast Open no more processing is needed (sk is the
835 * child socket).
836 */
837 if (fastopen)
838 return sk;
839
840 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */
841 if (req->num_timeout < READ_ONCE(inet_csk(sk)->icsk_accept_queue.rskq_defer_accept) &&
842 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) {
843 inet_rsk(req)->acked = 1;
844 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP);
845 return NULL;
846 }
847
848 /* OK, ACK is valid, create big socket and
849 * feed this segment to it. It will repeat all
850 * the tests. THIS SEGMENT MUST MOVE SOCKET TO
851 * ESTABLISHED STATE. If it will be dropped after
852 * socket is created, wait for troubles.
853 */
854 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL,
855 req, &own_req);
856 if (!child)
857 goto listen_overflow;
858
859 if (own_req && rsk_drop_req(req)) {
860 reqsk_queue_removed(&inet_csk(req->rsk_listener)->icsk_accept_queue, req);
861 inet_csk_reqsk_queue_drop_and_put(req->rsk_listener, req);
862 return child;
863 }
864
865 sock_rps_save_rxhash(child, skb);
866 tcp_synack_rtt_meas(child, req);
867 *req_stolen = !own_req;
868 return inet_csk_complete_hashdance(sk, child, req, own_req);
869
870listen_overflow:
871 if (sk != req->rsk_listener)
872 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMIGRATEREQFAILURE);
873
874 if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_abort_on_overflow)) {
875 inet_rsk(req)->acked = 1;
876 return NULL;
877 }
878
879embryonic_reset:
880 if (!(flg & TCP_FLAG_RST)) {
881 /* Received a bad SYN pkt - for TFO We try not to reset
882 * the local connection unless it's really necessary to
883 * avoid becoming vulnerable to outside attack aiming at
884 * resetting legit local connections.
885 */
886 req->rsk_ops->send_reset(sk, skb);
887 } else if (fastopen) { /* received a valid RST pkt */
888 reqsk_fastopen_remove(sk, req, true);
889 tcp_reset(sk, skb);
890 }
891 if (!fastopen) {
892 bool unlinked = inet_csk_reqsk_queue_drop(sk, req);
893
894 if (unlinked)
895 __NET_INC_STATS(sock_net(sk), LINUX_MIB_EMBRYONICRSTS);
896 *req_stolen = !unlinked;
897 }
898 return NULL;
899}
900EXPORT_SYMBOL(tcp_check_req);
901
902/*
903 * Queue segment on the new socket if the new socket is active,
904 * otherwise we just shortcircuit this and continue with
905 * the new socket.
906 *
907 * For the vast majority of cases child->sk_state will be TCP_SYN_RECV
908 * when entering. But other states are possible due to a race condition
909 * where after __inet_lookup_established() fails but before the listener
910 * locked is obtained, other packets cause the same connection to
911 * be created.
912 */
913
914int tcp_child_process(struct sock *parent, struct sock *child,
915 struct sk_buff *skb)
916 __releases(&((child)->sk_lock.slock))
917{
918 int ret = 0;
919 int state = child->sk_state;
920
921 /* record sk_napi_id and sk_rx_queue_mapping of child. */
922 sk_mark_napi_id_set(child, skb);
923
924 tcp_segs_in(tcp_sk(child), skb);
925 if (!sock_owned_by_user(child)) {
926 ret = tcp_rcv_state_process(child, skb);
927 /* Wakeup parent, send SIGIO */
928 if (state == TCP_SYN_RECV && child->sk_state != state)
929 parent->sk_data_ready(parent);
930 } else {
931 /* Alas, it is possible again, because we do lookup
932 * in main socket hash table and lock on listening
933 * socket does not protect us more.
934 */
935 __sk_add_backlog(child, skb);
936 }
937
938 bh_unlock_sock(child);
939 sock_put(child);
940 return ret;
941}
942EXPORT_SYMBOL(tcp_child_process);