xref: /linux/net/ipv4/tcp_minisocks.c (revision 6dfafbd0299a60bfb5d5e277fdf100037c7ded07)
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/tcp_ecn.h>
24 #include <net/xfrm.h>
25 #include <net/busy_poll.h>
26 #include <net/rstreason.h>
27 #include <net/psp.h>
28 
29 static bool tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win)
30 {
31 	if (seq == s_win)
32 		return true;
33 	if (after(end_seq, s_win) && before(seq, e_win))
34 		return true;
35 	return seq == e_win && seq == end_seq;
36 }
37 
38 static enum tcp_tw_status
39 tcp_timewait_check_oow_rate_limit(struct inet_timewait_sock *tw,
40 				  const struct sk_buff *skb, int mib_idx)
41 {
42 	struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
43 
44 	if (!tcp_oow_rate_limited(twsk_net(tw), skb, mib_idx,
45 				  &tcptw->tw_last_oow_ack_time)) {
46 		/* Send ACK. Note, we do not put the bucket,
47 		 * it will be released by caller.
48 		 */
49 		return TCP_TW_ACK_OOW;
50 	}
51 
52 	/* We are rate-limiting, so just release the tw sock and drop skb. */
53 	inet_twsk_put(tw);
54 	return TCP_TW_SUCCESS;
55 }
56 
57 static void twsk_rcv_nxt_update(struct tcp_timewait_sock *tcptw, u32 seq,
58 				u32 rcv_nxt)
59 {
60 #ifdef CONFIG_TCP_AO
61 	struct tcp_ao_info *ao;
62 
63 	ao = rcu_dereference(tcptw->ao_info);
64 	if (unlikely(ao && seq < rcv_nxt))
65 		WRITE_ONCE(ao->rcv_sne, ao->rcv_sne + 1);
66 #endif
67 	WRITE_ONCE(tcptw->tw_rcv_nxt, seq);
68 }
69 
70 /*
71  * * Main purpose of TIME-WAIT state is to close connection gracefully,
72  *   when one of ends sits in LAST-ACK or CLOSING retransmitting FIN
73  *   (and, probably, tail of data) and one or more our ACKs are lost.
74  * * What is TIME-WAIT timeout? It is associated with maximal packet
75  *   lifetime in the internet, which results in wrong conclusion, that
76  *   it is set to catch "old duplicate segments" wandering out of their path.
77  *   It is not quite correct. This timeout is calculated so that it exceeds
78  *   maximal retransmission timeout enough to allow to lose one (or more)
79  *   segments sent by peer and our ACKs. This time may be calculated from RTO.
80  * * When TIME-WAIT socket receives RST, it means that another end
81  *   finally closed and we are allowed to kill TIME-WAIT too.
82  * * Second purpose of TIME-WAIT is catching old duplicate segments.
83  *   Well, certainly it is pure paranoia, but if we load TIME-WAIT
84  *   with this semantics, we MUST NOT kill TIME-WAIT state with RSTs.
85  * * If we invented some more clever way to catch duplicates
86  *   (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs.
87  *
88  * The algorithm below is based on FORMAL INTERPRETATION of RFCs.
89  * When you compare it to RFCs, please, read section SEGMENT ARRIVES
90  * from the very beginning.
91  *
92  * NOTE. With recycling (and later with fin-wait-2) TW bucket
93  * is _not_ stateless. It means, that strictly speaking we must
94  * spinlock it. I do not want! Well, probability of misbehaviour
95  * is ridiculously low and, seems, we could use some mb() tricks
96  * to avoid misread sequence numbers, states etc.  --ANK
97  *
98  * We don't need to initialize tmp_out.sack_ok as we don't use the results
99  */
100 enum tcp_tw_status
101 tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb,
102 			   const struct tcphdr *th, u32 *tw_isn,
103 			   enum skb_drop_reason *drop_reason)
104 {
105 	struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
106 	u32 rcv_nxt = READ_ONCE(tcptw->tw_rcv_nxt);
107 	struct tcp_options_received tmp_opt;
108 	enum skb_drop_reason psp_drop;
109 	bool paws_reject = false;
110 	int ts_recent_stamp;
111 
112 	/* Instead of dropping immediately, wait to see what value is
113 	 * returned. We will accept a non psp-encapsulated syn in the
114 	 * case where TCP_TW_SYN is returned.
115 	 */
116 	psp_drop = psp_twsk_rx_policy_check(tw, skb);
117 
118 	tmp_opt.saw_tstamp = 0;
119 	ts_recent_stamp = READ_ONCE(tcptw->tw_ts_recent_stamp);
120 	if (th->doff > (sizeof(*th) >> 2) && ts_recent_stamp) {
121 		tcp_parse_options(twsk_net(tw), skb, &tmp_opt, 0, NULL);
122 
123 		if (tmp_opt.saw_tstamp) {
124 			if (tmp_opt.rcv_tsecr)
125 				tmp_opt.rcv_tsecr -= tcptw->tw_ts_offset;
126 			tmp_opt.ts_recent	= READ_ONCE(tcptw->tw_ts_recent);
127 			tmp_opt.ts_recent_stamp	= ts_recent_stamp;
128 			paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
129 		}
130 	}
131 
132 	if (READ_ONCE(tw->tw_substate) == TCP_FIN_WAIT2) {
133 		/* Just repeat all the checks of tcp_rcv_state_process() */
134 
135 		if (psp_drop)
136 			goto out_put;
137 
138 		/* Out of window, send ACK */
139 		if (paws_reject ||
140 		    !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
141 				   rcv_nxt,
142 				   rcv_nxt + tcptw->tw_rcv_wnd))
143 			return tcp_timewait_check_oow_rate_limit(
144 				tw, skb, LINUX_MIB_TCPACKSKIPPEDFINWAIT2);
145 
146 		if (th->rst)
147 			goto kill;
148 
149 		if (th->syn && !before(TCP_SKB_CB(skb)->seq, rcv_nxt))
150 			return TCP_TW_RST;
151 
152 		/* Dup ACK? */
153 		if (!th->ack ||
154 		    !after(TCP_SKB_CB(skb)->end_seq, rcv_nxt) ||
155 		    TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) {
156 			inet_twsk_put(tw);
157 			return TCP_TW_SUCCESS;
158 		}
159 
160 		/* New data or FIN. If new data arrive after half-duplex close,
161 		 * reset.
162 		 */
163 		if (!th->fin ||
164 		    TCP_SKB_CB(skb)->end_seq != rcv_nxt + 1)
165 			return TCP_TW_RST;
166 
167 		/* FIN arrived, enter true time-wait state. */
168 		WRITE_ONCE(tw->tw_substate, TCP_TIME_WAIT);
169 		twsk_rcv_nxt_update(tcptw, TCP_SKB_CB(skb)->end_seq,
170 				    rcv_nxt);
171 
172 		if (tmp_opt.saw_tstamp) {
173 			u64 ts = tcp_clock_ms();
174 
175 			WRITE_ONCE(tw->tw_entry_stamp, ts);
176 			WRITE_ONCE(tcptw->tw_ts_recent_stamp,
177 				   div_u64(ts, MSEC_PER_SEC));
178 			WRITE_ONCE(tcptw->tw_ts_recent,
179 				   tmp_opt.rcv_tsval);
180 		}
181 
182 		inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
183 		return TCP_TW_ACK;
184 	}
185 
186 	/*
187 	 *	Now real TIME-WAIT state.
188 	 *
189 	 *	RFC 1122:
190 	 *	"When a connection is [...] on TIME-WAIT state [...]
191 	 *	[a TCP] MAY accept a new SYN from the remote TCP to
192 	 *	reopen the connection directly, if it:
193 	 *
194 	 *	(1)  assigns its initial sequence number for the new
195 	 *	connection to be larger than the largest sequence
196 	 *	number it used on the previous connection incarnation,
197 	 *	and
198 	 *
199 	 *	(2)  returns to TIME-WAIT state if the SYN turns out
200 	 *	to be an old duplicate".
201 	 */
202 
203 	if (!paws_reject &&
204 	    (TCP_SKB_CB(skb)->seq == rcv_nxt &&
205 	     (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) {
206 		/* In window segment, it may be only reset or bare ack. */
207 
208 		if (psp_drop)
209 			goto out_put;
210 
211 		if (th->rst) {
212 			/* This is TIME_WAIT assassination, in two flavors.
213 			 * Oh well... nobody has a sufficient solution to this
214 			 * protocol bug yet.
215 			 */
216 			if (!READ_ONCE(twsk_net(tw)->ipv4.sysctl_tcp_rfc1337)) {
217 kill:
218 				inet_twsk_deschedule_put(tw);
219 				return TCP_TW_SUCCESS;
220 			}
221 		} else {
222 			inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
223 		}
224 
225 		if (tmp_opt.saw_tstamp) {
226 			WRITE_ONCE(tcptw->tw_ts_recent,
227 				   tmp_opt.rcv_tsval);
228 			WRITE_ONCE(tcptw->tw_ts_recent_stamp,
229 				   ktime_get_seconds());
230 		}
231 
232 		inet_twsk_put(tw);
233 		return TCP_TW_SUCCESS;
234 	}
235 
236 	/* Out of window segment.
237 
238 	   All the segments are ACKed immediately.
239 
240 	   The only exception is new SYN. We accept it, if it is
241 	   not old duplicate and we are not in danger to be killed
242 	   by delayed old duplicates. RFC check is that it has
243 	   newer sequence number works at rates <40Mbit/sec.
244 	   However, if paws works, it is reliable AND even more,
245 	   we even may relax silly seq space cutoff.
246 
247 	   RED-PEN: we violate main RFC requirement, if this SYN will appear
248 	   old duplicate (i.e. we receive RST in reply to SYN-ACK),
249 	   we must return socket to time-wait state. It is not good,
250 	   but not fatal yet.
251 	 */
252 
253 	if (th->syn && !th->rst && !th->ack && !paws_reject &&
254 	    (after(TCP_SKB_CB(skb)->seq, rcv_nxt) ||
255 	     (tmp_opt.saw_tstamp &&
256 	      (s32)(READ_ONCE(tcptw->tw_ts_recent) - tmp_opt.rcv_tsval) < 0))) {
257 		u32 isn = tcptw->tw_snd_nxt + 65535 + 2;
258 		if (isn == 0)
259 			isn++;
260 		*tw_isn = isn;
261 		return TCP_TW_SYN;
262 	}
263 
264 	if (psp_drop)
265 		goto out_put;
266 
267 	if (paws_reject) {
268 		*drop_reason = SKB_DROP_REASON_TCP_RFC7323_TW_PAWS;
269 		__NET_INC_STATS(twsk_net(tw), LINUX_MIB_PAWS_TW_REJECTED);
270 	}
271 
272 	if (!th->rst) {
273 		/* In this case we must reset the TIMEWAIT timer.
274 		 *
275 		 * If it is ACKless SYN it may be both old duplicate
276 		 * and new good SYN with random sequence number <rcv_nxt.
277 		 * Do not reschedule in the last case.
278 		 */
279 		if (paws_reject || th->ack)
280 			inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
281 
282 		return tcp_timewait_check_oow_rate_limit(
283 			tw, skb, LINUX_MIB_TCPACKSKIPPEDTIMEWAIT);
284 	}
285 
286 out_put:
287 	inet_twsk_put(tw);
288 	return TCP_TW_SUCCESS;
289 }
290 EXPORT_IPV6_MOD(tcp_timewait_state_process);
291 
292 static void tcp_time_wait_init(struct sock *sk, struct tcp_timewait_sock *tcptw)
293 {
294 #ifdef CONFIG_TCP_MD5SIG
295 	const struct tcp_sock *tp = tcp_sk(sk);
296 	struct tcp_md5sig_key *key;
297 
298 	/*
299 	 * The timewait bucket does not have the key DB from the
300 	 * sock structure. We just make a quick copy of the
301 	 * md5 key being used (if indeed we are using one)
302 	 * so the timewait ack generating code has the key.
303 	 */
304 	tcptw->tw_md5_key = NULL;
305 	if (!static_branch_unlikely(&tcp_md5_needed.key))
306 		return;
307 
308 	key = tp->af_specific->md5_lookup(sk, sk);
309 	if (key) {
310 		tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC);
311 		if (!tcptw->tw_md5_key)
312 			return;
313 		if (!static_key_fast_inc_not_disabled(&tcp_md5_needed.key.key))
314 			goto out_free;
315 	}
316 	return;
317 out_free:
318 	WARN_ON_ONCE(1);
319 	kfree(tcptw->tw_md5_key);
320 	tcptw->tw_md5_key = NULL;
321 #endif
322 }
323 
324 /*
325  * Move a socket to time-wait or dead fin-wait-2 state.
326  */
327 void tcp_time_wait(struct sock *sk, int state, int timeo)
328 {
329 	const struct inet_connection_sock *icsk = inet_csk(sk);
330 	struct tcp_sock *tp = tcp_sk(sk);
331 	struct net *net = sock_net(sk);
332 	struct inet_timewait_sock *tw;
333 
334 	tw = inet_twsk_alloc(sk, &net->ipv4.tcp_death_row, state);
335 
336 	if (tw) {
337 		struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
338 		const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1);
339 
340 		tw->tw_mark		= sk->sk_mark;
341 		tw->tw_priority		= READ_ONCE(sk->sk_priority);
342 		tw->tw_rcv_wscale	= tp->rx_opt.rcv_wscale;
343 		/* refreshed when we enter true TIME-WAIT state */
344 		tw->tw_entry_stamp	= tcp_time_stamp_ms(tp);
345 		tcptw->tw_rcv_nxt	= tp->rcv_nxt;
346 		tcptw->tw_snd_nxt	= tp->snd_nxt;
347 		tcptw->tw_rcv_wnd	= tcp_receive_window(tp);
348 		tcptw->tw_ts_recent	= tp->rx_opt.ts_recent;
349 		tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp;
350 		tcptw->tw_ts_offset	= tp->tsoffset;
351 		tw->tw_usec_ts		= tp->tcp_usec_ts;
352 		tcptw->tw_last_oow_ack_time = 0;
353 		tcptw->tw_tx_delay	= tp->tcp_tx_delay;
354 		tw->tw_txhash		= sk->sk_txhash;
355 		tw->tw_tx_queue_mapping = sk->sk_tx_queue_mapping;
356 #ifdef CONFIG_SOCK_RX_QUEUE_MAPPING
357 		tw->tw_rx_queue_mapping = sk->sk_rx_queue_mapping;
358 #endif
359 #if IS_ENABLED(CONFIG_IPV6)
360 		if (tw->tw_family == PF_INET6) {
361 			struct ipv6_pinfo *np = inet6_sk(sk);
362 
363 			tw->tw_v6_daddr = sk->sk_v6_daddr;
364 			tw->tw_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
365 			tw->tw_tclass = np->tclass;
366 			tw->tw_flowlabel = be32_to_cpu(np->flow_label & IPV6_FLOWLABEL_MASK);
367 			tw->tw_ipv6only = sk->sk_ipv6only;
368 		}
369 #endif
370 
371 		tcp_time_wait_init(sk, tcptw);
372 		tcp_ao_time_wait(tcptw, tp);
373 
374 		/* Get the TIME_WAIT timeout firing. */
375 		if (timeo < rto)
376 			timeo = rto;
377 
378 		if (state == TCP_TIME_WAIT)
379 			timeo = TCP_TIMEWAIT_LEN;
380 
381 		/* Linkage updates.
382 		 * Note that access to tw after this point is illegal.
383 		 */
384 		inet_twsk_hashdance_schedule(tw, sk, net->ipv4.tcp_death_row.hashinfo, timeo);
385 	} else {
386 		/* Sorry, if we're out of memory, just CLOSE this
387 		 * socket up.  We've got bigger problems than
388 		 * non-graceful socket closings.
389 		 */
390 		NET_INC_STATS(net, LINUX_MIB_TCPTIMEWAITOVERFLOW);
391 	}
392 
393 	tcp_update_metrics(sk);
394 	tcp_done(sk);
395 }
396 EXPORT_SYMBOL(tcp_time_wait);
397 
398 void tcp_twsk_destructor(struct sock *sk)
399 {
400 #ifdef CONFIG_TCP_MD5SIG
401 	if (static_branch_unlikely(&tcp_md5_needed.key)) {
402 		struct tcp_timewait_sock *twsk = tcp_twsk(sk);
403 
404 		if (twsk->tw_md5_key) {
405 			kfree(twsk->tw_md5_key);
406 			static_branch_slow_dec_deferred(&tcp_md5_needed);
407 		}
408 	}
409 #endif
410 	tcp_ao_destroy_sock(sk, true);
411 	psp_twsk_assoc_free(inet_twsk(sk));
412 }
413 
414 void tcp_twsk_purge(struct list_head *net_exit_list)
415 {
416 	bool purged_once = false;
417 	struct net *net;
418 
419 	list_for_each_entry(net, net_exit_list, exit_list) {
420 		if (net->ipv4.tcp_death_row.hashinfo->pernet) {
421 			/* Even if tw_refcount == 1, we must clean up kernel reqsk */
422 			inet_twsk_purge(net->ipv4.tcp_death_row.hashinfo);
423 		} else if (!purged_once) {
424 			inet_twsk_purge(&tcp_hashinfo);
425 			purged_once = true;
426 		}
427 	}
428 }
429 
430 /* Warning : This function is called without sk_listener being locked.
431  * Be sure to read socket fields once, as their value could change under us.
432  */
433 void tcp_openreq_init_rwin(struct request_sock *req,
434 			   const struct sock *sk_listener,
435 			   const struct dst_entry *dst)
436 {
437 	struct inet_request_sock *ireq = inet_rsk(req);
438 	const struct tcp_sock *tp = tcp_sk(sk_listener);
439 	int full_space = tcp_full_space(sk_listener);
440 	u32 window_clamp;
441 	__u8 rcv_wscale;
442 	u32 rcv_wnd;
443 	int mss;
444 
445 	mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
446 	window_clamp = READ_ONCE(tp->window_clamp);
447 	/* Set this up on the first call only */
448 	req->rsk_window_clamp = window_clamp ? : dst_metric(dst, RTAX_WINDOW);
449 
450 	/* limit the window selection if the user enforce a smaller rx buffer */
451 	if (sk_listener->sk_userlocks & SOCK_RCVBUF_LOCK &&
452 	    (req->rsk_window_clamp > full_space || req->rsk_window_clamp == 0))
453 		req->rsk_window_clamp = full_space;
454 
455 	rcv_wnd = tcp_rwnd_init_bpf((struct sock *)req);
456 	if (rcv_wnd == 0)
457 		rcv_wnd = dst_metric(dst, RTAX_INITRWND);
458 	else if (full_space < rcv_wnd * mss)
459 		full_space = rcv_wnd * mss;
460 
461 	/* tcp_full_space because it is guaranteed to be the first packet */
462 	tcp_select_initial_window(sk_listener, full_space,
463 		mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
464 		&req->rsk_rcv_wnd,
465 		&req->rsk_window_clamp,
466 		ireq->wscale_ok,
467 		&rcv_wscale,
468 		rcv_wnd);
469 	ireq->rcv_wscale = rcv_wscale;
470 }
471 
472 static void tcp_ecn_openreq_child(struct sock *sk,
473 				  const struct request_sock *req,
474 				  const struct sk_buff *skb)
475 {
476 	const struct tcp_request_sock *treq = tcp_rsk(req);
477 	struct tcp_sock *tp = tcp_sk(sk);
478 
479 	if (treq->accecn_ok) {
480 		tcp_ecn_mode_set(tp, TCP_ECN_MODE_ACCECN);
481 		tp->syn_ect_snt = treq->syn_ect_snt;
482 		tcp_accecn_third_ack(sk, skb, treq->syn_ect_snt);
483 		tp->saw_accecn_opt = treq->saw_accecn_opt;
484 		tp->prev_ecnfield = treq->syn_ect_rcv;
485 		tp->accecn_opt_demand = 1;
486 		tcp_ecn_received_counters_payload(sk, skb);
487 	} else {
488 		tcp_ecn_mode_set(tp, inet_rsk(req)->ecn_ok ?
489 				     TCP_ECN_MODE_RFC3168 :
490 				     TCP_ECN_DISABLED);
491 	}
492 }
493 
494 void tcp_ca_openreq_child(struct sock *sk, const struct dst_entry *dst)
495 {
496 	struct inet_connection_sock *icsk = inet_csk(sk);
497 	u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
498 	bool ca_got_dst = false;
499 
500 	if (ca_key != TCP_CA_UNSPEC) {
501 		const struct tcp_congestion_ops *ca;
502 
503 		rcu_read_lock();
504 		ca = tcp_ca_find_key(ca_key);
505 		if (likely(ca && bpf_try_module_get(ca, ca->owner))) {
506 			icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
507 			icsk->icsk_ca_ops = ca;
508 			ca_got_dst = true;
509 		}
510 		rcu_read_unlock();
511 	}
512 
513 	/* If no valid choice made yet, assign current system default ca. */
514 	if (!ca_got_dst &&
515 	    (!icsk->icsk_ca_setsockopt ||
516 	     !bpf_try_module_get(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner)))
517 		tcp_assign_congestion_control(sk);
518 
519 	tcp_set_ca_state(sk, TCP_CA_Open);
520 }
521 EXPORT_IPV6_MOD_GPL(tcp_ca_openreq_child);
522 
523 static void smc_check_reset_syn_req(const struct tcp_sock *oldtp,
524 				    struct request_sock *req,
525 				    struct tcp_sock *newtp)
526 {
527 #if IS_ENABLED(CONFIG_SMC)
528 	struct inet_request_sock *ireq;
529 
530 	if (static_branch_unlikely(&tcp_have_smc)) {
531 		ireq = inet_rsk(req);
532 		if (oldtp->syn_smc && !ireq->smc_ok)
533 			newtp->syn_smc = 0;
534 	}
535 #endif
536 }
537 
538 /* This is not only more efficient than what we used to do, it eliminates
539  * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM
540  *
541  * Actually, we could lots of memory writes here. tp of listening
542  * socket contains all necessary default parameters.
543  */
544 struct sock *tcp_create_openreq_child(const struct sock *sk,
545 				      struct request_sock *req,
546 				      struct sk_buff *skb)
547 {
548 	struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC);
549 	const struct inet_request_sock *ireq = inet_rsk(req);
550 	struct tcp_request_sock *treq = tcp_rsk(req);
551 	struct inet_connection_sock *newicsk;
552 	const struct tcp_sock *oldtp;
553 	struct tcp_sock *newtp;
554 	u32 seq;
555 
556 	if (!newsk)
557 		return NULL;
558 
559 	newicsk = inet_csk(newsk);
560 	newtp = tcp_sk(newsk);
561 	oldtp = tcp_sk(sk);
562 
563 	smc_check_reset_syn_req(oldtp, req, newtp);
564 
565 	/* Now setup tcp_sock */
566 	newtp->pred_flags = 0;
567 
568 	seq = treq->rcv_isn + 1;
569 	newtp->rcv_wup = seq;
570 	WRITE_ONCE(newtp->copied_seq, seq);
571 	WRITE_ONCE(newtp->rcv_nxt, seq);
572 	newtp->segs_in = 1;
573 
574 	seq = treq->snt_isn + 1;
575 	newtp->snd_sml = newtp->snd_una = seq;
576 	WRITE_ONCE(newtp->snd_nxt, seq);
577 	newtp->snd_up = seq;
578 
579 	INIT_LIST_HEAD(&newtp->tsq_node);
580 	INIT_LIST_HEAD(&newtp->tsorted_sent_queue);
581 
582 	tcp_init_wl(newtp, treq->rcv_isn);
583 
584 	minmax_reset(&newtp->rtt_min, tcp_jiffies32, ~0U);
585 	newicsk->icsk_ack.lrcvtime = tcp_jiffies32;
586 
587 	newtp->lsndtime = tcp_jiffies32;
588 	newsk->sk_txhash = READ_ONCE(treq->txhash);
589 	newtp->total_retrans = req->num_retrans;
590 
591 	tcp_init_xmit_timers(newsk);
592 	WRITE_ONCE(newtp->write_seq, newtp->pushed_seq = treq->snt_isn + 1);
593 
594 	if (sock_flag(newsk, SOCK_KEEPOPEN))
595 		tcp_reset_keepalive_timer(newsk, keepalive_time_when(newtp));
596 
597 	newtp->rx_opt.tstamp_ok = ireq->tstamp_ok;
598 	newtp->rx_opt.sack_ok = ireq->sack_ok;
599 	newtp->window_clamp = req->rsk_window_clamp;
600 	newtp->rcv_ssthresh = req->rsk_rcv_wnd;
601 	newtp->rcv_wnd = req->rsk_rcv_wnd;
602 	newtp->rx_opt.wscale_ok = ireq->wscale_ok;
603 	if (newtp->rx_opt.wscale_ok) {
604 		newtp->rx_opt.snd_wscale = ireq->snd_wscale;
605 		newtp->rx_opt.rcv_wscale = ireq->rcv_wscale;
606 	} else {
607 		newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0;
608 		newtp->window_clamp = min(newtp->window_clamp, 65535U);
609 	}
610 	newtp->snd_wnd = ntohs(tcp_hdr(skb)->window) << newtp->rx_opt.snd_wscale;
611 	newtp->max_window = newtp->snd_wnd;
612 
613 	if (newtp->rx_opt.tstamp_ok) {
614 		newtp->tcp_usec_ts = treq->req_usec_ts;
615 		newtp->rx_opt.ts_recent = req->ts_recent;
616 		newtp->rx_opt.ts_recent_stamp = ktime_get_seconds();
617 		newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
618 	} else {
619 		newtp->tcp_usec_ts = 0;
620 		newtp->rx_opt.ts_recent_stamp = 0;
621 		newtp->tcp_header_len = sizeof(struct tcphdr);
622 	}
623 	if (req->num_timeout) {
624 		newtp->total_rto = req->num_timeout;
625 		newtp->undo_marker = treq->snt_isn;
626 		if (newtp->tcp_usec_ts) {
627 			newtp->retrans_stamp = treq->snt_synack;
628 			newtp->total_rto_time = (u32)(tcp_clock_us() -
629 						      newtp->retrans_stamp) / USEC_PER_MSEC;
630 		} else {
631 			newtp->retrans_stamp = div_u64(treq->snt_synack,
632 						       USEC_PER_SEC / TCP_TS_HZ);
633 			newtp->total_rto_time = tcp_clock_ms() -
634 						newtp->retrans_stamp;
635 		}
636 		newtp->total_rto_recoveries = 1;
637 	}
638 	newtp->tsoffset = treq->ts_off;
639 #ifdef CONFIG_TCP_MD5SIG
640 	newtp->md5sig_info = NULL;	/*XXX*/
641 #endif
642 #ifdef CONFIG_TCP_AO
643 	newtp->ao_info = NULL;
644 
645 	if (tcp_rsk_used_ao(req)) {
646 		struct tcp_ao_key *ao_key;
647 
648 		ao_key = treq->af_specific->ao_lookup(sk, req, tcp_rsk(req)->ao_keyid, -1);
649 		if (ao_key)
650 			newtp->tcp_header_len += tcp_ao_len_aligned(ao_key);
651 	}
652  #endif
653 	if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len)
654 		newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len;
655 	newtp->rx_opt.mss_clamp = req->mss;
656 	tcp_ecn_openreq_child(newsk, req, skb);
657 	newtp->fastopen_req = NULL;
658 	RCU_INIT_POINTER(newtp->fastopen_rsk, NULL);
659 
660 	newtp->bpf_chg_cc_inprogress = 0;
661 	tcp_bpf_clone(sk, newsk);
662 
663 	__TCP_INC_STATS(sock_net(sk), TCP_MIB_PASSIVEOPENS);
664 
665 	xa_init_flags(&newsk->sk_user_frags, XA_FLAGS_ALLOC1);
666 
667 	return newsk;
668 }
669 EXPORT_SYMBOL(tcp_create_openreq_child);
670 
671 /*
672  * Process an incoming packet for SYN_RECV sockets represented as a
673  * request_sock. Normally sk is the listener socket but for TFO it
674  * points to the child socket.
675  *
676  * XXX (TFO) - The current impl contains a special check for ack
677  * validation and inside tcp_v4_reqsk_send_ack(). Can we do better?
678  *
679  * We don't need to initialize tmp_opt.sack_ok as we don't use the results
680  *
681  * Note: If @fastopen is true, this can be called from process context.
682  *       Otherwise, this is from BH context.
683  */
684 
685 struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb,
686 			   struct request_sock *req,
687 			   bool fastopen, bool *req_stolen,
688 			   enum skb_drop_reason *drop_reason)
689 {
690 	struct tcp_options_received tmp_opt;
691 	struct sock *child;
692 	const struct tcphdr *th = tcp_hdr(skb);
693 	__be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK);
694 	bool tsecr_reject = false;
695 	bool paws_reject = false;
696 	bool own_req;
697 
698 	tmp_opt.saw_tstamp = 0;
699 	tmp_opt.accecn = 0;
700 	if (th->doff > (sizeof(struct tcphdr)>>2)) {
701 		tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0, NULL);
702 
703 		if (tmp_opt.saw_tstamp) {
704 			tmp_opt.ts_recent = req->ts_recent;
705 			if (tmp_opt.rcv_tsecr) {
706 				if (inet_rsk(req)->tstamp_ok && !fastopen)
707 					tsecr_reject = !between(tmp_opt.rcv_tsecr,
708 							tcp_rsk(req)->snt_tsval_first,
709 							READ_ONCE(tcp_rsk(req)->snt_tsval_last));
710 				tmp_opt.rcv_tsecr -= tcp_rsk(req)->ts_off;
711 			}
712 			/* We do not store true stamp, but it is not required,
713 			 * it can be estimated (approximately)
714 			 * from another data.
715 			 */
716 			tmp_opt.ts_recent_stamp = ktime_get_seconds() -
717 				tcp_reqsk_timeout(req) / HZ;
718 			paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
719 		}
720 	}
721 
722 	/* Check for pure retransmitted SYN. */
723 	if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn &&
724 	    flg == TCP_FLAG_SYN &&
725 	    !paws_reject) {
726 		/*
727 		 * RFC793 draws (Incorrectly! It was fixed in RFC1122)
728 		 * this case on figure 6 and figure 8, but formal
729 		 * protocol description says NOTHING.
730 		 * To be more exact, it says that we should send ACK,
731 		 * because this segment (at least, if it has no data)
732 		 * is out of window.
733 		 *
734 		 *  CONCLUSION: RFC793 (even with RFC1122) DOES NOT
735 		 *  describe SYN-RECV state. All the description
736 		 *  is wrong, we cannot believe to it and should
737 		 *  rely only on common sense and implementation
738 		 *  experience.
739 		 *
740 		 * Enforce "SYN-ACK" according to figure 8, figure 6
741 		 * of RFC793, fixed by RFC1122.
742 		 *
743 		 * Note that even if there is new data in the SYN packet
744 		 * they will be thrown away too.
745 		 *
746 		 * Reset timer after retransmitting SYNACK, similar to
747 		 * the idea of fast retransmit in recovery.
748 		 */
749 		if (!tcp_oow_rate_limited(sock_net(sk), skb,
750 					  LINUX_MIB_TCPACKSKIPPEDSYNRECV,
751 					  &tcp_rsk(req)->last_oow_ack_time) &&
752 
753 		    !tcp_rtx_synack(sk, req)) {
754 			unsigned long expires = jiffies;
755 
756 			expires += tcp_reqsk_timeout(req);
757 			if (!fastopen)
758 				mod_timer_pending(&req->rsk_timer, expires);
759 			else
760 				req->rsk_timer.expires = expires;
761 		}
762 		return NULL;
763 	}
764 
765 	/* Further reproduces section "SEGMENT ARRIVES"
766 	   for state SYN-RECEIVED of RFC793.
767 	   It is broken, however, it does not work only
768 	   when SYNs are crossed.
769 
770 	   You would think that SYN crossing is impossible here, since
771 	   we should have a SYN_SENT socket (from connect()) on our end,
772 	   but this is not true if the crossed SYNs were sent to both
773 	   ends by a malicious third party.  We must defend against this,
774 	   and to do that we first verify the ACK (as per RFC793, page
775 	   36) and reset if it is invalid.  Is this a true full defense?
776 	   To convince ourselves, let us consider a way in which the ACK
777 	   test can still pass in this 'malicious crossed SYNs' case.
778 	   Malicious sender sends identical SYNs (and thus identical sequence
779 	   numbers) to both A and B:
780 
781 		A: gets SYN, seq=7
782 		B: gets SYN, seq=7
783 
784 	   By our good fortune, both A and B select the same initial
785 	   send sequence number of seven :-)
786 
787 		A: sends SYN|ACK, seq=7, ack_seq=8
788 		B: sends SYN|ACK, seq=7, ack_seq=8
789 
790 	   So we are now A eating this SYN|ACK, ACK test passes.  So
791 	   does sequence test, SYN is truncated, and thus we consider
792 	   it a bare ACK.
793 
794 	   If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this
795 	   bare ACK.  Otherwise, we create an established connection.  Both
796 	   ends (listening sockets) accept the new incoming connection and try
797 	   to talk to each other. 8-)
798 
799 	   Note: This case is both harmless, and rare.  Possibility is about the
800 	   same as us discovering intelligent life on another plant tomorrow.
801 
802 	   But generally, we should (RFC lies!) to accept ACK
803 	   from SYNACK both here and in tcp_rcv_state_process().
804 	   tcp_rcv_state_process() does not, hence, we do not too.
805 
806 	   Note that the case is absolutely generic:
807 	   we cannot optimize anything here without
808 	   violating protocol. All the checks must be made
809 	   before attempt to create socket.
810 	 */
811 
812 	/* RFC793 page 36: "If the connection is in any non-synchronized state ...
813 	 *                  and the incoming segment acknowledges something not yet
814 	 *                  sent (the segment carries an unacceptable ACK) ...
815 	 *                  a reset is sent."
816 	 *
817 	 * Invalid ACK: reset will be sent by listening socket.
818 	 * Note that the ACK validity check for a Fast Open socket is done
819 	 * elsewhere and is checked directly against the child socket rather
820 	 * than req because user data may have been sent out.
821 	 */
822 	if ((flg & TCP_FLAG_ACK) && !fastopen &&
823 	    (TCP_SKB_CB(skb)->ack_seq !=
824 	     tcp_rsk(req)->snt_isn + 1))
825 		return sk;
826 
827 	/* RFC793: "first check sequence number". */
828 
829 	if (paws_reject || tsecr_reject ||
830 	    !tcp_in_window(TCP_SKB_CB(skb)->seq,
831 			   TCP_SKB_CB(skb)->end_seq,
832 			   tcp_rsk(req)->rcv_nxt,
833 			   tcp_rsk(req)->rcv_nxt +
834 			   tcp_synack_window(req))) {
835 		/* Out of window: send ACK and drop. */
836 		if (!(flg & TCP_FLAG_RST) &&
837 		    !tcp_oow_rate_limited(sock_net(sk), skb,
838 					  LINUX_MIB_TCPACKSKIPPEDSYNRECV,
839 					  &tcp_rsk(req)->last_oow_ack_time))
840 			req->rsk_ops->send_ack(sk, skb, req);
841 		if (paws_reject) {
842 			SKB_DR_SET(*drop_reason, TCP_RFC7323_PAWS);
843 			NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
844 		} else if (tsecr_reject) {
845 			SKB_DR_SET(*drop_reason, TCP_RFC7323_TSECR);
846 			NET_INC_STATS(sock_net(sk), LINUX_MIB_TSECRREJECTED);
847 		} else {
848 			SKB_DR_SET(*drop_reason, TCP_OVERWINDOW);
849 		}
850 		return NULL;
851 	}
852 
853 	/* In sequence, PAWS is OK. */
854 
855 	if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) {
856 		/* Truncate SYN, it is out of window starting
857 		   at tcp_rsk(req)->rcv_isn + 1. */
858 		flg &= ~TCP_FLAG_SYN;
859 	}
860 
861 	/* RFC793: "second check the RST bit" and
862 	 *	   "fourth, check the SYN bit"
863 	 */
864 	if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) {
865 		TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
866 		goto embryonic_reset;
867 	}
868 
869 	/* ACK sequence verified above, just make sure ACK is
870 	 * set.  If ACK not set, just silently drop the packet.
871 	 *
872 	 * XXX (TFO) - if we ever allow "data after SYN", the
873 	 * following check needs to be removed.
874 	 */
875 	if (!(flg & TCP_FLAG_ACK))
876 		return NULL;
877 
878 	if (tcp_rsk(req)->accecn_ok && tmp_opt.accecn &&
879 	    tcp_rsk(req)->saw_accecn_opt < TCP_ACCECN_OPT_COUNTER_SEEN) {
880 		u8 saw_opt = tcp_accecn_option_init(skb, tmp_opt.accecn);
881 
882 		tcp_rsk(req)->saw_accecn_opt = saw_opt;
883 		if (tcp_rsk(req)->saw_accecn_opt == TCP_ACCECN_OPT_FAIL_SEEN) {
884 			u8 fail_mode = TCP_ACCECN_OPT_FAIL_RECV;
885 
886 			tcp_rsk(req)->accecn_fail_mode |= fail_mode;
887 		}
888 	}
889 
890 	/* For Fast Open no more processing is needed (sk is the
891 	 * child socket).
892 	 */
893 	if (fastopen)
894 		return sk;
895 
896 	/* While TCP_DEFER_ACCEPT is active, drop bare ACK. */
897 	if (req->num_timeout < READ_ONCE(inet_csk(sk)->icsk_accept_queue.rskq_defer_accept) &&
898 	    TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) {
899 		inet_rsk(req)->acked = 1;
900 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP);
901 		return NULL;
902 	}
903 
904 	/* OK, ACK is valid, create big socket and
905 	 * feed this segment to it. It will repeat all
906 	 * the tests. THIS SEGMENT MUST MOVE SOCKET TO
907 	 * ESTABLISHED STATE. If it will be dropped after
908 	 * socket is created, wait for troubles.
909 	 */
910 	child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL,
911 							 req, &own_req);
912 	if (!child)
913 		goto listen_overflow;
914 
915 	if (own_req && tmp_opt.saw_tstamp &&
916 	    !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_nxt))
917 		tcp_sk(child)->rx_opt.ts_recent = tmp_opt.rcv_tsval;
918 
919 	if (own_req && rsk_drop_req(req)) {
920 		reqsk_queue_removed(&inet_csk(req->rsk_listener)->icsk_accept_queue, req);
921 		inet_csk_reqsk_queue_drop_and_put(req->rsk_listener, req);
922 		return child;
923 	}
924 
925 	sock_rps_save_rxhash(child, skb);
926 	tcp_synack_rtt_meas(child, req);
927 	*req_stolen = !own_req;
928 	return inet_csk_complete_hashdance(sk, child, req, own_req);
929 
930 listen_overflow:
931 	SKB_DR_SET(*drop_reason, TCP_LISTEN_OVERFLOW);
932 	if (sk != req->rsk_listener)
933 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMIGRATEREQFAILURE);
934 
935 	if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_abort_on_overflow)) {
936 		inet_rsk(req)->acked = 1;
937 		return NULL;
938 	}
939 
940 embryonic_reset:
941 	if (!(flg & TCP_FLAG_RST)) {
942 		/* Received a bad SYN pkt - for TFO We try not to reset
943 		 * the local connection unless it's really necessary to
944 		 * avoid becoming vulnerable to outside attack aiming at
945 		 * resetting legit local connections.
946 		 */
947 		req->rsk_ops->send_reset(sk, skb, SK_RST_REASON_INVALID_SYN);
948 	} else if (fastopen) { /* received a valid RST pkt */
949 		reqsk_fastopen_remove(sk, req, true);
950 		tcp_reset(sk, skb);
951 	}
952 	if (!fastopen) {
953 		bool unlinked = inet_csk_reqsk_queue_drop(sk, req);
954 
955 		if (unlinked)
956 			__NET_INC_STATS(sock_net(sk), LINUX_MIB_EMBRYONICRSTS);
957 		*req_stolen = !unlinked;
958 	}
959 	return NULL;
960 }
961 EXPORT_IPV6_MOD(tcp_check_req);
962 
963 /*
964  * Queue segment on the new socket if the new socket is active,
965  * otherwise we just shortcircuit this and continue with
966  * the new socket.
967  *
968  * For the vast majority of cases child->sk_state will be TCP_SYN_RECV
969  * when entering. But other states are possible due to a race condition
970  * where after __inet_lookup_established() fails but before the listener
971  * locked is obtained, other packets cause the same connection to
972  * be created.
973  */
974 
975 enum skb_drop_reason tcp_child_process(struct sock *parent, struct sock *child,
976 				       struct sk_buff *skb)
977 	__releases(&((child)->sk_lock.slock))
978 {
979 	enum skb_drop_reason reason = SKB_NOT_DROPPED_YET;
980 	int state = child->sk_state;
981 
982 	/* record sk_napi_id and sk_rx_queue_mapping of child. */
983 	sk_mark_napi_id_set(child, skb);
984 
985 	tcp_segs_in(tcp_sk(child), skb);
986 	if (!sock_owned_by_user(child)) {
987 		reason = tcp_rcv_state_process(child, skb);
988 		/* Wakeup parent, send SIGIO */
989 		if (state == TCP_SYN_RECV && child->sk_state != state)
990 			parent->sk_data_ready(parent);
991 	} else {
992 		/* Alas, it is possible again, because we do lookup
993 		 * in main socket hash table and lock on listening
994 		 * socket does not protect us more.
995 		 */
996 		__sk_add_backlog(child, skb);
997 	}
998 
999 	bh_unlock_sock(child);
1000 	sock_put(child);
1001 	return reason;
1002 }
1003 EXPORT_IPV6_MOD(tcp_child_process);
1004