projects / linux-block.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Linux 6.16-rc6
[linux-block.git] / net / ipv4 / tcp_minisocks.c
... / ...
CommitLineData
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
6 *
7 * Implementation of the Transmission Control Protocol(TCP).
8 *
9 * Authors: Ross Biro
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
20 */
21
22#include <net/tcp.h>
23#include <net/xfrm.h>
24#include <net/busy_poll.h>
25#include <net/rstreason.h>
26
27static bool tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win)
28{
29 if (seq == s_win)
30 return true;
31 if (after(end_seq, s_win) && before(seq, e_win))
32 return true;
33 return seq == e_win && seq == end_seq;
34}
35
36static enum tcp_tw_status
37tcp_timewait_check_oow_rate_limit(struct inet_timewait_sock *tw,
38 const struct sk_buff *skb, int mib_idx)
39{
40 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
41
42 if (!tcp_oow_rate_limited(twsk_net(tw), skb, mib_idx,
43 &tcptw->tw_last_oow_ack_time)) {
44 /* Send ACK. Note, we do not put the bucket,
45 * it will be released by caller.
46 */
47 return TCP_TW_ACK_OOW;
48 }
49
50 /* We are rate-limiting, so just release the tw sock and drop skb. */
51 inet_twsk_put(tw);
52 return TCP_TW_SUCCESS;
53}
54
55static void twsk_rcv_nxt_update(struct tcp_timewait_sock *tcptw, u32 seq,
56 u32 rcv_nxt)
57{
58#ifdef CONFIG_TCP_AO
59 struct tcp_ao_info *ao;
60
61 ao = rcu_dereference(tcptw->ao_info);
62 if (unlikely(ao && seq < rcv_nxt))
63 WRITE_ONCE(ao->rcv_sne, ao->rcv_sne + 1);
64#endif
65 WRITE_ONCE(tcptw->tw_rcv_nxt, seq);
66}
67
68/*
69 * * Main purpose of TIME-WAIT state is to close connection gracefully,
70 * when one of ends sits in LAST-ACK or CLOSING retransmitting FIN
71 * (and, probably, tail of data) and one or more our ACKs are lost.
72 * * What is TIME-WAIT timeout? It is associated with maximal packet
73 * lifetime in the internet, which results in wrong conclusion, that
74 * it is set to catch "old duplicate segments" wandering out of their path.
75 * It is not quite correct. This timeout is calculated so that it exceeds
76 * maximal retransmission timeout enough to allow to lose one (or more)
77 * segments sent by peer and our ACKs. This time may be calculated from RTO.
78 * * When TIME-WAIT socket receives RST, it means that another end
79 * finally closed and we are allowed to kill TIME-WAIT too.
80 * * Second purpose of TIME-WAIT is catching old duplicate segments.
81 * Well, certainly it is pure paranoia, but if we load TIME-WAIT
82 * with this semantics, we MUST NOT kill TIME-WAIT state with RSTs.
83 * * If we invented some more clever way to catch duplicates
84 * (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs.
85 *
86 * The algorithm below is based on FORMAL INTERPRETATION of RFCs.
87 * When you compare it to RFCs, please, read section SEGMENT ARRIVES
88 * from the very beginning.
89 *
90 * NOTE. With recycling (and later with fin-wait-2) TW bucket
91 * is _not_ stateless. It means, that strictly speaking we must
92 * spinlock it. I do not want! Well, probability of misbehaviour
93 * is ridiculously low and, seems, we could use some mb() tricks
94 * to avoid misread sequence numbers, states etc. --ANK
95 *
96 * We don't need to initialize tmp_out.sack_ok as we don't use the results
97 */
98enum tcp_tw_status
99tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb,
100 const struct tcphdr *th, u32 *tw_isn,
101 enum skb_drop_reason *drop_reason)
102{
103 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
104 u32 rcv_nxt = READ_ONCE(tcptw->tw_rcv_nxt);
105 struct tcp_options_received tmp_opt;
106 bool paws_reject = false;
107 int ts_recent_stamp;
108
109 tmp_opt.saw_tstamp = 0;
110 ts_recent_stamp = READ_ONCE(tcptw->tw_ts_recent_stamp);
111 if (th->doff > (sizeof(*th) >> 2) && ts_recent_stamp) {
112 tcp_parse_options(twsk_net(tw), skb, &tmp_opt, 0, NULL);
113
114 if (tmp_opt.saw_tstamp) {
115 if (tmp_opt.rcv_tsecr)
116 tmp_opt.rcv_tsecr -= tcptw->tw_ts_offset;
117 tmp_opt.ts_recent = READ_ONCE(tcptw->tw_ts_recent);
118 tmp_opt.ts_recent_stamp = ts_recent_stamp;
119 paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
120 }
121 }
122
123 if (READ_ONCE(tw->tw_substate) == TCP_FIN_WAIT2) {
124 /* Just repeat all the checks of tcp_rcv_state_process() */
125
126 /* Out of window, send ACK */
127 if (paws_reject ||
128 !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
129 rcv_nxt,
130 rcv_nxt + tcptw->tw_rcv_wnd))
131 return tcp_timewait_check_oow_rate_limit(
132 tw, skb, LINUX_MIB_TCPACKSKIPPEDFINWAIT2);
133
134 if (th->rst)
135 goto kill;
136
137 if (th->syn && !before(TCP_SKB_CB(skb)->seq, rcv_nxt))
138 return TCP_TW_RST;
139
140 /* Dup ACK? */
141 if (!th->ack ||
142 !after(TCP_SKB_CB(skb)->end_seq, rcv_nxt) ||
143 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) {
144 inet_twsk_put(tw);
145 return TCP_TW_SUCCESS;
146 }
147
148 /* New data or FIN. If new data arrive after half-duplex close,
149 * reset.
150 */
151 if (!th->fin ||
152 TCP_SKB_CB(skb)->end_seq != rcv_nxt + 1)
153 return TCP_TW_RST;
154
155 /* FIN arrived, enter true time-wait state. */
156 WRITE_ONCE(tw->tw_substate, TCP_TIME_WAIT);
157 twsk_rcv_nxt_update(tcptw, TCP_SKB_CB(skb)->end_seq,
158 rcv_nxt);
159
160 if (tmp_opt.saw_tstamp) {
161 u64 ts = tcp_clock_ms();
162
163 WRITE_ONCE(tw->tw_entry_stamp, ts);
164 WRITE_ONCE(tcptw->tw_ts_recent_stamp,
165 div_u64(ts, MSEC_PER_SEC));
166 WRITE_ONCE(tcptw->tw_ts_recent,
167 tmp_opt.rcv_tsval);
168 }
169
170 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
171 return TCP_TW_ACK;
172 }
173
174 /*
175 * Now real TIME-WAIT state.
176 *
177 * RFC 1122:
178 * "When a connection is [...] on TIME-WAIT state [...]
179 * [a TCP] MAY accept a new SYN from the remote TCP to
180 * reopen the connection directly, if it:
181 *
182 * (1) assigns its initial sequence number for the new
183 * connection to be larger than the largest sequence
184 * number it used on the previous connection incarnation,
185 * and
186 *
187 * (2) returns to TIME-WAIT state if the SYN turns out
188 * to be an old duplicate".
189 */
190
191 if (!paws_reject &&
192 (TCP_SKB_CB(skb)->seq == rcv_nxt &&
193 (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) {
194 /* In window segment, it may be only reset or bare ack. */
195
196 if (th->rst) {
197 /* This is TIME_WAIT assassination, in two flavors.
198 * Oh well... nobody has a sufficient solution to this
199 * protocol bug yet.
200 */
201 if (!READ_ONCE(twsk_net(tw)->ipv4.sysctl_tcp_rfc1337)) {
202kill:
203 inet_twsk_deschedule_put(tw);
204 return TCP_TW_SUCCESS;
205 }
206 } else {
207 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
208 }
209
210 if (tmp_opt.saw_tstamp) {
211 WRITE_ONCE(tcptw->tw_ts_recent,
212 tmp_opt.rcv_tsval);
213 WRITE_ONCE(tcptw->tw_ts_recent_stamp,
214 ktime_get_seconds());
215 }
216
217 inet_twsk_put(tw);
218 return TCP_TW_SUCCESS;
219 }
220
221 /* Out of window segment.
222
223 All the segments are ACKed immediately.
224
225 The only exception is new SYN. We accept it, if it is
226 not old duplicate and we are not in danger to be killed
227 by delayed old duplicates. RFC check is that it has
228 newer sequence number works at rates <40Mbit/sec.
229 However, if paws works, it is reliable AND even more,
230 we even may relax silly seq space cutoff.
231
232 RED-PEN: we violate main RFC requirement, if this SYN will appear
233 old duplicate (i.e. we receive RST in reply to SYN-ACK),
234 we must return socket to time-wait state. It is not good,
235 but not fatal yet.
236 */
237
238 if (th->syn && !th->rst && !th->ack && !paws_reject &&
239 (after(TCP_SKB_CB(skb)->seq, rcv_nxt) ||
240 (tmp_opt.saw_tstamp &&
241 (s32)(READ_ONCE(tcptw->tw_ts_recent) - tmp_opt.rcv_tsval) < 0))) {
242 u32 isn = tcptw->tw_snd_nxt + 65535 + 2;
243 if (isn == 0)
244 isn++;
245 *tw_isn = isn;
246 return TCP_TW_SYN;
247 }
248
249 if (paws_reject) {
250 *drop_reason = SKB_DROP_REASON_TCP_RFC7323_TW_PAWS;
251 __NET_INC_STATS(twsk_net(tw), LINUX_MIB_PAWS_TW_REJECTED);
252 }
253
254 if (!th->rst) {
255 /* In this case we must reset the TIMEWAIT timer.
256 *
257 * If it is ACKless SYN it may be both old duplicate
258 * and new good SYN with random sequence number <rcv_nxt.
259 * Do not reschedule in the last case.
260 */
261 if (paws_reject || th->ack)
262 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
263
264 return tcp_timewait_check_oow_rate_limit(
265 tw, skb, LINUX_MIB_TCPACKSKIPPEDTIMEWAIT);
266 }
267 inet_twsk_put(tw);
268 return TCP_TW_SUCCESS;
269}
270EXPORT_IPV6_MOD(tcp_timewait_state_process);
271
272static void tcp_time_wait_init(struct sock *sk, struct tcp_timewait_sock *tcptw)
273{
274#ifdef CONFIG_TCP_MD5SIG
275 const struct tcp_sock *tp = tcp_sk(sk);
276 struct tcp_md5sig_key *key;
277
278 /*
279 * The timewait bucket does not have the key DB from the
280 * sock structure. We just make a quick copy of the
281 * md5 key being used (if indeed we are using one)
282 * so the timewait ack generating code has the key.
283 */
284 tcptw->tw_md5_key = NULL;
285 if (!static_branch_unlikely(&tcp_md5_needed.key))
286 return;
287
288 key = tp->af_specific->md5_lookup(sk, sk);
289 if (key) {
290 tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC);
291 if (!tcptw->tw_md5_key)
292 return;
293 if (!static_key_fast_inc_not_disabled(&tcp_md5_needed.key.key))
294 goto out_free;
295 tcp_md5_add_sigpool();
296 }
297 return;
298out_free:
299 WARN_ON_ONCE(1);
300 kfree(tcptw->tw_md5_key);
301 tcptw->tw_md5_key = NULL;
302#endif
303}
304
305/*
306 * Move a socket to time-wait or dead fin-wait-2 state.
307 */
308void tcp_time_wait(struct sock *sk, int state, int timeo)
309{
310 const struct inet_connection_sock *icsk = inet_csk(sk);
311 struct tcp_sock *tp = tcp_sk(sk);
312 struct net *net = sock_net(sk);
313 struct inet_timewait_sock *tw;
314
315 tw = inet_twsk_alloc(sk, &net->ipv4.tcp_death_row, state);
316
317 if (tw) {
318 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
319 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1);
320
321 tw->tw_transparent = inet_test_bit(TRANSPARENT, sk);
322 tw->tw_mark = sk->sk_mark;
323 tw->tw_priority = READ_ONCE(sk->sk_priority);
324 tw->tw_rcv_wscale = tp->rx_opt.rcv_wscale;
325 /* refreshed when we enter true TIME-WAIT state */
326 tw->tw_entry_stamp = tcp_time_stamp_ms(tp);
327 tcptw->tw_rcv_nxt = tp->rcv_nxt;
328 tcptw->tw_snd_nxt = tp->snd_nxt;
329 tcptw->tw_rcv_wnd = tcp_receive_window(tp);
330 tcptw->tw_ts_recent = tp->rx_opt.ts_recent;
331 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp;
332 tcptw->tw_ts_offset = tp->tsoffset;
333 tw->tw_usec_ts = tp->tcp_usec_ts;
334 tcptw->tw_last_oow_ack_time = 0;
335 tcptw->tw_tx_delay = tp->tcp_tx_delay;
336 tw->tw_txhash = sk->sk_txhash;
337 tw->tw_tx_queue_mapping = sk->sk_tx_queue_mapping;
338#ifdef CONFIG_SOCK_RX_QUEUE_MAPPING
339 tw->tw_rx_queue_mapping = sk->sk_rx_queue_mapping;
340#endif
341#if IS_ENABLED(CONFIG_IPV6)
342 if (tw->tw_family == PF_INET6) {
343 struct ipv6_pinfo *np = inet6_sk(sk);
344
345 tw->tw_v6_daddr = sk->sk_v6_daddr;
346 tw->tw_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
347 tw->tw_tclass = np->tclass;
348 tw->tw_flowlabel = be32_to_cpu(np->flow_label & IPV6_FLOWLABEL_MASK);
349 tw->tw_ipv6only = sk->sk_ipv6only;
350 }
351#endif
352
353 tcp_time_wait_init(sk, tcptw);
354 tcp_ao_time_wait(tcptw, tp);
355
356 /* Get the TIME_WAIT timeout firing. */
357 if (timeo < rto)
358 timeo = rto;
359
360 if (state == TCP_TIME_WAIT)
361 timeo = TCP_TIMEWAIT_LEN;
362
363 /* Linkage updates.
364 * Note that access to tw after this point is illegal.
365 */
366 inet_twsk_hashdance_schedule(tw, sk, net->ipv4.tcp_death_row.hashinfo, timeo);
367 } else {
368 /* Sorry, if we're out of memory, just CLOSE this
369 * socket up. We've got bigger problems than
370 * non-graceful socket closings.
371 */
372 NET_INC_STATS(net, LINUX_MIB_TCPTIMEWAITOVERFLOW);
373 }
374
375 tcp_update_metrics(sk);
376 tcp_done(sk);
377}
378EXPORT_SYMBOL(tcp_time_wait);
379
380#ifdef CONFIG_TCP_MD5SIG
381static void tcp_md5_twsk_free_rcu(struct rcu_head *head)
382{
383 struct tcp_md5sig_key *key;
384
385 key = container_of(head, struct tcp_md5sig_key, rcu);
386 kfree(key);
387 static_branch_slow_dec_deferred(&tcp_md5_needed);
388 tcp_md5_release_sigpool();
389}
390#endif
391
392void tcp_twsk_destructor(struct sock *sk)
393{
394#ifdef CONFIG_TCP_MD5SIG
395 if (static_branch_unlikely(&tcp_md5_needed.key)) {
396 struct tcp_timewait_sock *twsk = tcp_twsk(sk);
397
398 if (twsk->tw_md5_key)
399 call_rcu(&twsk->tw_md5_key->rcu, tcp_md5_twsk_free_rcu);
400 }
401#endif
402 tcp_ao_destroy_sock(sk, true);
403}
404EXPORT_IPV6_MOD_GPL(tcp_twsk_destructor);
405
406void tcp_twsk_purge(struct list_head *net_exit_list)
407{
408 bool purged_once = false;
409 struct net *net;
410
411 list_for_each_entry(net, net_exit_list, exit_list) {
412 if (net->ipv4.tcp_death_row.hashinfo->pernet) {
413 /* Even if tw_refcount == 1, we must clean up kernel reqsk */
414 inet_twsk_purge(net->ipv4.tcp_death_row.hashinfo);
415 } else if (!purged_once) {
416 inet_twsk_purge(&tcp_hashinfo);
417 purged_once = true;
418 }
419 }
420}
421
422/* Warning : This function is called without sk_listener being locked.
423 * Be sure to read socket fields once, as their value could change under us.
424 */
425void tcp_openreq_init_rwin(struct request_sock *req,
426 const struct sock *sk_listener,
427 const struct dst_entry *dst)
428{
429 struct inet_request_sock *ireq = inet_rsk(req);
430 const struct tcp_sock *tp = tcp_sk(sk_listener);
431 int full_space = tcp_full_space(sk_listener);
432 u32 window_clamp;
433 __u8 rcv_wscale;
434 u32 rcv_wnd;
435 int mss;
436
437 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
438 window_clamp = READ_ONCE(tp->window_clamp);
439 /* Set this up on the first call only */
440 req->rsk_window_clamp = window_clamp ? : dst_metric(dst, RTAX_WINDOW);
441
442 /* limit the window selection if the user enforce a smaller rx buffer */
443 if (sk_listener->sk_userlocks & SOCK_RCVBUF_LOCK &&
444 (req->rsk_window_clamp > full_space || req->rsk_window_clamp == 0))
445 req->rsk_window_clamp = full_space;
446
447 rcv_wnd = tcp_rwnd_init_bpf((struct sock *)req);
448 if (rcv_wnd == 0)
449 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
450 else if (full_space < rcv_wnd * mss)
451 full_space = rcv_wnd * mss;
452
453 /* tcp_full_space because it is guaranteed to be the first packet */
454 tcp_select_initial_window(sk_listener, full_space,
455 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
456 &req->rsk_rcv_wnd,
457 &req->rsk_window_clamp,
458 ireq->wscale_ok,
459 &rcv_wscale,
460 rcv_wnd);
461 ireq->rcv_wscale = rcv_wscale;
462}
463
464static void tcp_ecn_openreq_child(struct tcp_sock *tp,
465 const struct request_sock *req)
466{
467 tcp_ecn_mode_set(tp, inet_rsk(req)->ecn_ok ?
468 TCP_ECN_MODE_RFC3168 :
469 TCP_ECN_DISABLED);
470}
471
472void tcp_ca_openreq_child(struct sock *sk, const struct dst_entry *dst)
473{
474 struct inet_connection_sock *icsk = inet_csk(sk);
475 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
476 bool ca_got_dst = false;
477
478 if (ca_key != TCP_CA_UNSPEC) {
479 const struct tcp_congestion_ops *ca;
480
481 rcu_read_lock();
482 ca = tcp_ca_find_key(ca_key);
483 if (likely(ca && bpf_try_module_get(ca, ca->owner))) {
484 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
485 icsk->icsk_ca_ops = ca;
486 ca_got_dst = true;
487 }
488 rcu_read_unlock();
489 }
490
491 /* If no valid choice made yet, assign current system default ca. */
492 if (!ca_got_dst &&
493 (!icsk->icsk_ca_setsockopt ||
494 !bpf_try_module_get(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner)))
495 tcp_assign_congestion_control(sk);
496
497 tcp_set_ca_state(sk, TCP_CA_Open);
498}
499EXPORT_IPV6_MOD_GPL(tcp_ca_openreq_child);
500
501static void smc_check_reset_syn_req(const struct tcp_sock *oldtp,
502 struct request_sock *req,
503 struct tcp_sock *newtp)
504{
505#if IS_ENABLED(CONFIG_SMC)
506 struct inet_request_sock *ireq;
507
508 if (static_branch_unlikely(&tcp_have_smc)) {
509 ireq = inet_rsk(req);
510 if (oldtp->syn_smc && !ireq->smc_ok)
511 newtp->syn_smc = 0;
512 }
513#endif
514}
515
516/* This is not only more efficient than what we used to do, it eliminates
517 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM
518 *
519 * Actually, we could lots of memory writes here. tp of listening
520 * socket contains all necessary default parameters.
521 */
522struct sock *tcp_create_openreq_child(const struct sock *sk,
523 struct request_sock *req,
524 struct sk_buff *skb)
525{
526 struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC);
527 const struct inet_request_sock *ireq = inet_rsk(req);
528 struct tcp_request_sock *treq = tcp_rsk(req);
529 struct inet_connection_sock *newicsk;
530 const struct tcp_sock *oldtp;
531 struct tcp_sock *newtp;
532 u32 seq;
533
534 if (!newsk)
535 return NULL;
536
537 newicsk = inet_csk(newsk);
538 newtp = tcp_sk(newsk);
539 oldtp = tcp_sk(sk);
540
541 smc_check_reset_syn_req(oldtp, req, newtp);
542
543 /* Now setup tcp_sock */
544 newtp->pred_flags = 0;
545
546 seq = treq->rcv_isn + 1;
547 newtp->rcv_wup = seq;
548 WRITE_ONCE(newtp->copied_seq, seq);
549 WRITE_ONCE(newtp->rcv_nxt, seq);
550 newtp->segs_in = 1;
551
552 seq = treq->snt_isn + 1;
553 newtp->snd_sml = newtp->snd_una = seq;
554 WRITE_ONCE(newtp->snd_nxt, seq);
555 newtp->snd_up = seq;
556
557 INIT_LIST_HEAD(&newtp->tsq_node);
558 INIT_LIST_HEAD(&newtp->tsorted_sent_queue);
559
560 tcp_init_wl(newtp, treq->rcv_isn);
561
562 minmax_reset(&newtp->rtt_min, tcp_jiffies32, ~0U);
563 newicsk->icsk_ack.lrcvtime = tcp_jiffies32;
564
565 newtp->lsndtime = tcp_jiffies32;
566 newsk->sk_txhash = READ_ONCE(treq->txhash);
567 newtp->total_retrans = req->num_retrans;
568
569 tcp_init_xmit_timers(newsk);
570 WRITE_ONCE(newtp->write_seq, newtp->pushed_seq = treq->snt_isn + 1);
571
572 if (sock_flag(newsk, SOCK_KEEPOPEN))
573 tcp_reset_keepalive_timer(newsk, keepalive_time_when(newtp));
574
575 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok;
576 newtp->rx_opt.sack_ok = ireq->sack_ok;
577 newtp->window_clamp = req->rsk_window_clamp;
578 newtp->rcv_ssthresh = req->rsk_rcv_wnd;
579 newtp->rcv_wnd = req->rsk_rcv_wnd;
580 newtp->rx_opt.wscale_ok = ireq->wscale_ok;
581 if (newtp->rx_opt.wscale_ok) {
582 newtp->rx_opt.snd_wscale = ireq->snd_wscale;
583 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale;
584 } else {
585 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0;
586 newtp->window_clamp = min(newtp->window_clamp, 65535U);
587 }
588 newtp->snd_wnd = ntohs(tcp_hdr(skb)->window) << newtp->rx_opt.snd_wscale;
589 newtp->max_window = newtp->snd_wnd;
590
591 if (newtp->rx_opt.tstamp_ok) {
592 newtp->tcp_usec_ts = treq->req_usec_ts;
593 newtp->rx_opt.ts_recent = req->ts_recent;
594 newtp->rx_opt.ts_recent_stamp = ktime_get_seconds();
595 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
596 } else {
597 newtp->tcp_usec_ts = 0;
598 newtp->rx_opt.ts_recent_stamp = 0;
599 newtp->tcp_header_len = sizeof(struct tcphdr);
600 }
601 if (req->num_timeout) {
602 newtp->total_rto = req->num_timeout;
603 newtp->undo_marker = treq->snt_isn;
604 if (newtp->tcp_usec_ts) {
605 newtp->retrans_stamp = treq->snt_synack;
606 newtp->total_rto_time = (u32)(tcp_clock_us() -
607 newtp->retrans_stamp) / USEC_PER_MSEC;
608 } else {
609 newtp->retrans_stamp = div_u64(treq->snt_synack,
610 USEC_PER_SEC / TCP_TS_HZ);
611 newtp->total_rto_time = tcp_clock_ms() -
612 newtp->retrans_stamp;
613 }
614 newtp->total_rto_recoveries = 1;
615 }
616 newtp->tsoffset = treq->ts_off;
617#ifdef CONFIG_TCP_MD5SIG
618 newtp->md5sig_info = NULL; /*XXX*/
619#endif
620#ifdef CONFIG_TCP_AO
621 newtp->ao_info = NULL;
622
623 if (tcp_rsk_used_ao(req)) {
624 struct tcp_ao_key *ao_key;
625
626 ao_key = treq->af_specific->ao_lookup(sk, req, tcp_rsk(req)->ao_keyid, -1);
627 if (ao_key)
628 newtp->tcp_header_len += tcp_ao_len_aligned(ao_key);
629 }
630 #endif
631 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len)
632 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len;
633 newtp->rx_opt.mss_clamp = req->mss;
634 tcp_ecn_openreq_child(newtp, req);
635 newtp->fastopen_req = NULL;
636 RCU_INIT_POINTER(newtp->fastopen_rsk, NULL);
637
638 newtp->bpf_chg_cc_inprogress = 0;
639 tcp_bpf_clone(sk, newsk);
640
641 __TCP_INC_STATS(sock_net(sk), TCP_MIB_PASSIVEOPENS);
642
643 xa_init_flags(&newsk->sk_user_frags, XA_FLAGS_ALLOC1);
644
645 return newsk;
646}
647EXPORT_SYMBOL(tcp_create_openreq_child);
648
649/*
650 * Process an incoming packet for SYN_RECV sockets represented as a
651 * request_sock. Normally sk is the listener socket but for TFO it
652 * points to the child socket.
653 *
654 * XXX (TFO) - The current impl contains a special check for ack
655 * validation and inside tcp_v4_reqsk_send_ack(). Can we do better?
656 *
657 * We don't need to initialize tmp_opt.sack_ok as we don't use the results
658 *
659 * Note: If @fastopen is true, this can be called from process context.
660 * Otherwise, this is from BH context.
661 */
662
663struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb,
664 struct request_sock *req,
665 bool fastopen, bool *req_stolen,
666 enum skb_drop_reason *drop_reason)
667{
668 struct tcp_options_received tmp_opt;
669 struct sock *child;
670 const struct tcphdr *th = tcp_hdr(skb);
671 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK);
672 bool tsecr_reject = false;
673 bool paws_reject = false;
674 bool own_req;
675
676 tmp_opt.saw_tstamp = 0;
677 if (th->doff > (sizeof(struct tcphdr)>>2)) {
678 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0, NULL);
679
680 if (tmp_opt.saw_tstamp) {
681 tmp_opt.ts_recent = req->ts_recent;
682 if (tmp_opt.rcv_tsecr) {
683 if (inet_rsk(req)->tstamp_ok && !fastopen)
684 tsecr_reject = !between(tmp_opt.rcv_tsecr,
685 tcp_rsk(req)->snt_tsval_first,
686 READ_ONCE(tcp_rsk(req)->snt_tsval_last));
687 tmp_opt.rcv_tsecr -= tcp_rsk(req)->ts_off;
688 }
689 /* We do not store true stamp, but it is not required,
690 * it can be estimated (approximately)
691 * from another data.
692 */
693 tmp_opt.ts_recent_stamp = ktime_get_seconds() - reqsk_timeout(req, TCP_RTO_MAX) / HZ;
694 paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
695 }
696 }
697
698 /* Check for pure retransmitted SYN. */
699 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn &&
700 flg == TCP_FLAG_SYN &&
701 !paws_reject) {
702 /*
703 * RFC793 draws (Incorrectly! It was fixed in RFC1122)
704 * this case on figure 6 and figure 8, but formal
705 * protocol description says NOTHING.
706 * To be more exact, it says that we should send ACK,
707 * because this segment (at least, if it has no data)
708 * is out of window.
709 *
710 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT
711 * describe SYN-RECV state. All the description
712 * is wrong, we cannot believe to it and should
713 * rely only on common sense and implementation
714 * experience.
715 *
716 * Enforce "SYN-ACK" according to figure 8, figure 6
717 * of RFC793, fixed by RFC1122.
718 *
719 * Note that even if there is new data in the SYN packet
720 * they will be thrown away too.
721 *
722 * Reset timer after retransmitting SYNACK, similar to
723 * the idea of fast retransmit in recovery.
724 */
725 if (!tcp_oow_rate_limited(sock_net(sk), skb,
726 LINUX_MIB_TCPACKSKIPPEDSYNRECV,
727 &tcp_rsk(req)->last_oow_ack_time) &&
728
729 !inet_rtx_syn_ack(sk, req)) {
730 unsigned long expires = jiffies;
731
732 expires += reqsk_timeout(req, TCP_RTO_MAX);
733 if (!fastopen)
734 mod_timer_pending(&req->rsk_timer, expires);
735 else
736 req->rsk_timer.expires = expires;
737 }
738 return NULL;
739 }
740
741 /* Further reproduces section "SEGMENT ARRIVES"
742 for state SYN-RECEIVED of RFC793.
743 It is broken, however, it does not work only
744 when SYNs are crossed.
745
746 You would think that SYN crossing is impossible here, since
747 we should have a SYN_SENT socket (from connect()) on our end,
748 but this is not true if the crossed SYNs were sent to both
749 ends by a malicious third party. We must defend against this,
750 and to do that we first verify the ACK (as per RFC793, page
751 36) and reset if it is invalid. Is this a true full defense?
752 To convince ourselves, let us consider a way in which the ACK
753 test can still pass in this 'malicious crossed SYNs' case.
754 Malicious sender sends identical SYNs (and thus identical sequence
755 numbers) to both A and B:
756
757 A: gets SYN, seq=7
758 B: gets SYN, seq=7
759
760 By our good fortune, both A and B select the same initial
761 send sequence number of seven :-)
762
763 A: sends SYN|ACK, seq=7, ack_seq=8
764 B: sends SYN|ACK, seq=7, ack_seq=8
765
766 So we are now A eating this SYN|ACK, ACK test passes. So
767 does sequence test, SYN is truncated, and thus we consider
768 it a bare ACK.
769
770 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this
771 bare ACK. Otherwise, we create an established connection. Both
772 ends (listening sockets) accept the new incoming connection and try
773 to talk to each other. 8-)
774
775 Note: This case is both harmless, and rare. Possibility is about the
776 same as us discovering intelligent life on another plant tomorrow.
777
778 But generally, we should (RFC lies!) to accept ACK
779 from SYNACK both here and in tcp_rcv_state_process().
780 tcp_rcv_state_process() does not, hence, we do not too.
781
782 Note that the case is absolutely generic:
783 we cannot optimize anything here without
784 violating protocol. All the checks must be made
785 before attempt to create socket.
786 */
787
788 /* RFC793 page 36: "If the connection is in any non-synchronized state ...
789 * and the incoming segment acknowledges something not yet
790 * sent (the segment carries an unacceptable ACK) ...
791 * a reset is sent."
792 *
793 * Invalid ACK: reset will be sent by listening socket.
794 * Note that the ACK validity check for a Fast Open socket is done
795 * elsewhere and is checked directly against the child socket rather
796 * than req because user data may have been sent out.
797 */
798 if ((flg & TCP_FLAG_ACK) && !fastopen &&
799 (TCP_SKB_CB(skb)->ack_seq !=
800 tcp_rsk(req)->snt_isn + 1))
801 return sk;
802
803 /* RFC793: "first check sequence number". */
804
805 if (paws_reject || tsecr_reject ||
806 !tcp_in_window(TCP_SKB_CB(skb)->seq,
807 TCP_SKB_CB(skb)->end_seq,
808 tcp_rsk(req)->rcv_nxt,
809 tcp_rsk(req)->rcv_nxt +
810 tcp_synack_window(req))) {
811 /* Out of window: send ACK and drop. */
812 if (!(flg & TCP_FLAG_RST) &&
813 !tcp_oow_rate_limited(sock_net(sk), skb,
814 LINUX_MIB_TCPACKSKIPPEDSYNRECV,
815 &tcp_rsk(req)->last_oow_ack_time))
816 req->rsk_ops->send_ack(sk, skb, req);
817 if (paws_reject) {
818 SKB_DR_SET(*drop_reason, TCP_RFC7323_PAWS);
819 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
820 } else if (tsecr_reject) {
821 SKB_DR_SET(*drop_reason, TCP_RFC7323_TSECR);
822 NET_INC_STATS(sock_net(sk), LINUX_MIB_TSECRREJECTED);
823 } else {
824 SKB_DR_SET(*drop_reason, TCP_OVERWINDOW);
825 }
826 return NULL;
827 }
828
829 /* In sequence, PAWS is OK. */
830
831 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) {
832 /* Truncate SYN, it is out of window starting
833 at tcp_rsk(req)->rcv_isn + 1. */
834 flg &= ~TCP_FLAG_SYN;
835 }
836
837 /* RFC793: "second check the RST bit" and
838 * "fourth, check the SYN bit"
839 */
840 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) {
841 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
842 goto embryonic_reset;
843 }
844
845 /* ACK sequence verified above, just make sure ACK is
846 * set. If ACK not set, just silently drop the packet.
847 *
848 * XXX (TFO) - if we ever allow "data after SYN", the
849 * following check needs to be removed.
850 */
851 if (!(flg & TCP_FLAG_ACK))
852 return NULL;
853
854 /* For Fast Open no more processing is needed (sk is the
855 * child socket).
856 */
857 if (fastopen)
858 return sk;
859
860 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */
861 if (req->num_timeout < READ_ONCE(inet_csk(sk)->icsk_accept_queue.rskq_defer_accept) &&
862 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) {
863 inet_rsk(req)->acked = 1;
864 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP);
865 return NULL;
866 }
867
868 /* OK, ACK is valid, create big socket and
869 * feed this segment to it. It will repeat all
870 * the tests. THIS SEGMENT MUST MOVE SOCKET TO
871 * ESTABLISHED STATE. If it will be dropped after
872 * socket is created, wait for troubles.
873 */
874 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL,
875 req, &own_req);
876 if (!child)
877 goto listen_overflow;
878
879 if (own_req && tmp_opt.saw_tstamp &&
880 !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_nxt))
881 tcp_sk(child)->rx_opt.ts_recent = tmp_opt.rcv_tsval;
882
883 if (own_req && rsk_drop_req(req)) {
884 reqsk_queue_removed(&inet_csk(req->rsk_listener)->icsk_accept_queue, req);
885 inet_csk_reqsk_queue_drop_and_put(req->rsk_listener, req);
886 return child;
887 }
888
889 sock_rps_save_rxhash(child, skb);
890 tcp_synack_rtt_meas(child, req);
891 *req_stolen = !own_req;
892 return inet_csk_complete_hashdance(sk, child, req, own_req);
893
894listen_overflow:
895 SKB_DR_SET(*drop_reason, TCP_LISTEN_OVERFLOW);
896 if (sk != req->rsk_listener)
897 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMIGRATEREQFAILURE);
898
899 if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_abort_on_overflow)) {
900 inet_rsk(req)->acked = 1;
901 return NULL;
902 }
903
904embryonic_reset:
905 if (!(flg & TCP_FLAG_RST)) {
906 /* Received a bad SYN pkt - for TFO We try not to reset
907 * the local connection unless it's really necessary to
908 * avoid becoming vulnerable to outside attack aiming at
909 * resetting legit local connections.
910 */
911 req->rsk_ops->send_reset(sk, skb, SK_RST_REASON_INVALID_SYN);
912 } else if (fastopen) { /* received a valid RST pkt */
913 reqsk_fastopen_remove(sk, req, true);
914 tcp_reset(sk, skb);
915 }
916 if (!fastopen) {
917 bool unlinked = inet_csk_reqsk_queue_drop(sk, req);
918
919 if (unlinked)
920 __NET_INC_STATS(sock_net(sk), LINUX_MIB_EMBRYONICRSTS);
921 *req_stolen = !unlinked;
922 }
923 return NULL;
924}
925EXPORT_IPV6_MOD(tcp_check_req);
926
927/*
928 * Queue segment on the new socket if the new socket is active,
929 * otherwise we just shortcircuit this and continue with
930 * the new socket.
931 *
932 * For the vast majority of cases child->sk_state will be TCP_SYN_RECV
933 * when entering. But other states are possible due to a race condition
934 * where after __inet_lookup_established() fails but before the listener
935 * locked is obtained, other packets cause the same connection to
936 * be created.
937 */
938
939enum skb_drop_reason tcp_child_process(struct sock *parent, struct sock *child,
940 struct sk_buff *skb)
941 __releases(&((child)->sk_lock.slock))
942{
943 enum skb_drop_reason reason = SKB_NOT_DROPPED_YET;
944 int state = child->sk_state;
945
946 /* record sk_napi_id and sk_rx_queue_mapping of child. */
947 sk_mark_napi_id_set(child, skb);
948
949 tcp_segs_in(tcp_sk(child), skb);
950 if (!sock_owned_by_user(child)) {
951 reason = tcp_rcv_state_process(child, skb);
952 /* Wakeup parent, send SIGIO */
953 if (state == TCP_SYN_RECV && child->sk_state != state)
954 parent->sk_data_ready(parent);
955 } else {
956 /* Alas, it is possible again, because we do lookup
957 * in main socket hash table and lock on listening
958 * socket does not protect us more.
959 */
960 __sk_add_backlog(child, skb);
961 }
962
963 bh_unlock_sock(child);
964 sock_put(child);
965 return reason;
966}
967EXPORT_IPV6_MOD(tcp_child_process);
Linux 6.x block layer and io_uring tree(s)