xref: /openbmc/linux/net/ipv4/tcp_minisocks.c (revision 6c33a6f4)
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 <linux/mm.h>
23 #include <linux/module.h>
24 #include <linux/slab.h>
25 #include <linux/sysctl.h>
26 #include <linux/workqueue.h>
27 #include <linux/static_key.h>
28 #include <net/tcp.h>
29 #include <net/inet_common.h>
30 #include <net/xfrm.h>
31 #include <net/busy_poll.h>
32 
33 static bool tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win)
34 {
35 	if (seq == s_win)
36 		return true;
37 	if (after(end_seq, s_win) && before(seq, e_win))
38 		return true;
39 	return seq == e_win && seq == end_seq;
40 }
41 
42 static enum tcp_tw_status
43 tcp_timewait_check_oow_rate_limit(struct inet_timewait_sock *tw,
44 				  const struct sk_buff *skb, int mib_idx)
45 {
46 	struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
47 
48 	if (!tcp_oow_rate_limited(twsk_net(tw), skb, mib_idx,
49 				  &tcptw->tw_last_oow_ack_time)) {
50 		/* Send ACK. Note, we do not put the bucket,
51 		 * it will be released by caller.
52 		 */
53 		return TCP_TW_ACK;
54 	}
55 
56 	/* We are rate-limiting, so just release the tw sock and drop skb. */
57 	inet_twsk_put(tw);
58 	return TCP_TW_SUCCESS;
59 }
60 
61 /*
62  * * Main purpose of TIME-WAIT state is to close connection gracefully,
63  *   when one of ends sits in LAST-ACK or CLOSING retransmitting FIN
64  *   (and, probably, tail of data) and one or more our ACKs are lost.
65  * * What is TIME-WAIT timeout? It is associated with maximal packet
66  *   lifetime in the internet, which results in wrong conclusion, that
67  *   it is set to catch "old duplicate segments" wandering out of their path.
68  *   It is not quite correct. This timeout is calculated so that it exceeds
69  *   maximal retransmission timeout enough to allow to lose one (or more)
70  *   segments sent by peer and our ACKs. This time may be calculated from RTO.
71  * * When TIME-WAIT socket receives RST, it means that another end
72  *   finally closed and we are allowed to kill TIME-WAIT too.
73  * * Second purpose of TIME-WAIT is catching old duplicate segments.
74  *   Well, certainly it is pure paranoia, but if we load TIME-WAIT
75  *   with this semantics, we MUST NOT kill TIME-WAIT state with RSTs.
76  * * If we invented some more clever way to catch duplicates
77  *   (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs.
78  *
79  * The algorithm below is based on FORMAL INTERPRETATION of RFCs.
80  * When you compare it to RFCs, please, read section SEGMENT ARRIVES
81  * from the very beginning.
82  *
83  * NOTE. With recycling (and later with fin-wait-2) TW bucket
84  * is _not_ stateless. It means, that strictly speaking we must
85  * spinlock it. I do not want! Well, probability of misbehaviour
86  * is ridiculously low and, seems, we could use some mb() tricks
87  * to avoid misread sequence numbers, states etc.  --ANK
88  *
89  * We don't need to initialize tmp_out.sack_ok as we don't use the results
90  */
91 enum tcp_tw_status
92 tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb,
93 			   const struct tcphdr *th)
94 {
95 	struct tcp_options_received tmp_opt;
96 	struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
97 	bool paws_reject = false;
98 
99 	tmp_opt.saw_tstamp = 0;
100 	if (th->doff > (sizeof(*th) >> 2) && tcptw->tw_ts_recent_stamp) {
101 		tcp_parse_options(twsk_net(tw), skb, &tmp_opt, 0, NULL);
102 
103 		if (tmp_opt.saw_tstamp) {
104 			if (tmp_opt.rcv_tsecr)
105 				tmp_opt.rcv_tsecr -= tcptw->tw_ts_offset;
106 			tmp_opt.ts_recent	= tcptw->tw_ts_recent;
107 			tmp_opt.ts_recent_stamp	= tcptw->tw_ts_recent_stamp;
108 			paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
109 		}
110 	}
111 
112 	if (tw->tw_substate == TCP_FIN_WAIT2) {
113 		/* Just repeat all the checks of tcp_rcv_state_process() */
114 
115 		/* Out of window, send ACK */
116 		if (paws_reject ||
117 		    !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
118 				   tcptw->tw_rcv_nxt,
119 				   tcptw->tw_rcv_nxt + tcptw->tw_rcv_wnd))
120 			return tcp_timewait_check_oow_rate_limit(
121 				tw, skb, LINUX_MIB_TCPACKSKIPPEDFINWAIT2);
122 
123 		if (th->rst)
124 			goto kill;
125 
126 		if (th->syn && !before(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt))
127 			return TCP_TW_RST;
128 
129 		/* Dup ACK? */
130 		if (!th->ack ||
131 		    !after(TCP_SKB_CB(skb)->end_seq, tcptw->tw_rcv_nxt) ||
132 		    TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) {
133 			inet_twsk_put(tw);
134 			return TCP_TW_SUCCESS;
135 		}
136 
137 		/* New data or FIN. If new data arrive after half-duplex close,
138 		 * reset.
139 		 */
140 		if (!th->fin ||
141 		    TCP_SKB_CB(skb)->end_seq != tcptw->tw_rcv_nxt + 1)
142 			return TCP_TW_RST;
143 
144 		/* FIN arrived, enter true time-wait state. */
145 		tw->tw_substate	  = TCP_TIME_WAIT;
146 		tcptw->tw_rcv_nxt = TCP_SKB_CB(skb)->end_seq;
147 		if (tmp_opt.saw_tstamp) {
148 			tcptw->tw_ts_recent_stamp = ktime_get_seconds();
149 			tcptw->tw_ts_recent	  = tmp_opt.rcv_tsval;
150 		}
151 
152 		inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
153 		return TCP_TW_ACK;
154 	}
155 
156 	/*
157 	 *	Now real TIME-WAIT state.
158 	 *
159 	 *	RFC 1122:
160 	 *	"When a connection is [...] on TIME-WAIT state [...]
161 	 *	[a TCP] MAY accept a new SYN from the remote TCP to
162 	 *	reopen the connection directly, if it:
163 	 *
164 	 *	(1)  assigns its initial sequence number for the new
165 	 *	connection to be larger than the largest sequence
166 	 *	number it used on the previous connection incarnation,
167 	 *	and
168 	 *
169 	 *	(2)  returns to TIME-WAIT state if the SYN turns out
170 	 *	to be an old duplicate".
171 	 */
172 
173 	if (!paws_reject &&
174 	    (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt &&
175 	     (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) {
176 		/* In window segment, it may be only reset or bare ack. */
177 
178 		if (th->rst) {
179 			/* This is TIME_WAIT assassination, in two flavors.
180 			 * Oh well... nobody has a sufficient solution to this
181 			 * protocol bug yet.
182 			 */
183 			if (twsk_net(tw)->ipv4.sysctl_tcp_rfc1337 == 0) {
184 kill:
185 				inet_twsk_deschedule_put(tw);
186 				return TCP_TW_SUCCESS;
187 			}
188 		} else {
189 			inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
190 		}
191 
192 		if (tmp_opt.saw_tstamp) {
193 			tcptw->tw_ts_recent	  = tmp_opt.rcv_tsval;
194 			tcptw->tw_ts_recent_stamp = ktime_get_seconds();
195 		}
196 
197 		inet_twsk_put(tw);
198 		return TCP_TW_SUCCESS;
199 	}
200 
201 	/* Out of window segment.
202 
203 	   All the segments are ACKed immediately.
204 
205 	   The only exception is new SYN. We accept it, if it is
206 	   not old duplicate and we are not in danger to be killed
207 	   by delayed old duplicates. RFC check is that it has
208 	   newer sequence number works at rates <40Mbit/sec.
209 	   However, if paws works, it is reliable AND even more,
210 	   we even may relax silly seq space cutoff.
211 
212 	   RED-PEN: we violate main RFC requirement, if this SYN will appear
213 	   old duplicate (i.e. we receive RST in reply to SYN-ACK),
214 	   we must return socket to time-wait state. It is not good,
215 	   but not fatal yet.
216 	 */
217 
218 	if (th->syn && !th->rst && !th->ack && !paws_reject &&
219 	    (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) ||
220 	     (tmp_opt.saw_tstamp &&
221 	      (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) {
222 		u32 isn = tcptw->tw_snd_nxt + 65535 + 2;
223 		if (isn == 0)
224 			isn++;
225 		TCP_SKB_CB(skb)->tcp_tw_isn = isn;
226 		return TCP_TW_SYN;
227 	}
228 
229 	if (paws_reject)
230 		__NET_INC_STATS(twsk_net(tw), LINUX_MIB_PAWSESTABREJECTED);
231 
232 	if (!th->rst) {
233 		/* In this case we must reset the TIMEWAIT timer.
234 		 *
235 		 * If it is ACKless SYN it may be both old duplicate
236 		 * and new good SYN with random sequence number <rcv_nxt.
237 		 * Do not reschedule in the last case.
238 		 */
239 		if (paws_reject || th->ack)
240 			inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN);
241 
242 		return tcp_timewait_check_oow_rate_limit(
243 			tw, skb, LINUX_MIB_TCPACKSKIPPEDTIMEWAIT);
244 	}
245 	inet_twsk_put(tw);
246 	return TCP_TW_SUCCESS;
247 }
248 EXPORT_SYMBOL(tcp_timewait_state_process);
249 
250 /*
251  * Move a socket to time-wait or dead fin-wait-2 state.
252  */
253 void tcp_time_wait(struct sock *sk, int state, int timeo)
254 {
255 	const struct inet_connection_sock *icsk = inet_csk(sk);
256 	const struct tcp_sock *tp = tcp_sk(sk);
257 	struct inet_timewait_sock *tw;
258 	struct inet_timewait_death_row *tcp_death_row = &sock_net(sk)->ipv4.tcp_death_row;
259 
260 	tw = inet_twsk_alloc(sk, tcp_death_row, state);
261 
262 	if (tw) {
263 		struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
264 		const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1);
265 		struct inet_sock *inet = inet_sk(sk);
266 
267 		tw->tw_transparent	= inet->transparent;
268 		tw->tw_mark		= sk->sk_mark;
269 		tw->tw_priority		= sk->sk_priority;
270 		tw->tw_rcv_wscale	= tp->rx_opt.rcv_wscale;
271 		tcptw->tw_rcv_nxt	= tp->rcv_nxt;
272 		tcptw->tw_snd_nxt	= tp->snd_nxt;
273 		tcptw->tw_rcv_wnd	= tcp_receive_window(tp);
274 		tcptw->tw_ts_recent	= tp->rx_opt.ts_recent;
275 		tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp;
276 		tcptw->tw_ts_offset	= tp->tsoffset;
277 		tcptw->tw_last_oow_ack_time = 0;
278 		tcptw->tw_tx_delay	= tp->tcp_tx_delay;
279 #if IS_ENABLED(CONFIG_IPV6)
280 		if (tw->tw_family == PF_INET6) {
281 			struct ipv6_pinfo *np = inet6_sk(sk);
282 
283 			tw->tw_v6_daddr = sk->sk_v6_daddr;
284 			tw->tw_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
285 			tw->tw_tclass = np->tclass;
286 			tw->tw_flowlabel = be32_to_cpu(np->flow_label & IPV6_FLOWLABEL_MASK);
287 			tw->tw_txhash = sk->sk_txhash;
288 			tw->tw_ipv6only = sk->sk_ipv6only;
289 		}
290 #endif
291 
292 #ifdef CONFIG_TCP_MD5SIG
293 		/*
294 		 * The timewait bucket does not have the key DB from the
295 		 * sock structure. We just make a quick copy of the
296 		 * md5 key being used (if indeed we are using one)
297 		 * so the timewait ack generating code has the key.
298 		 */
299 		do {
300 			tcptw->tw_md5_key = NULL;
301 			if (static_branch_unlikely(&tcp_md5_needed)) {
302 				struct tcp_md5sig_key *key;
303 
304 				key = tp->af_specific->md5_lookup(sk, sk);
305 				if (key) {
306 					tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC);
307 					BUG_ON(tcptw->tw_md5_key && !tcp_alloc_md5sig_pool());
308 				}
309 			}
310 		} while (0);
311 #endif
312 
313 		/* Get the TIME_WAIT timeout firing. */
314 		if (timeo < rto)
315 			timeo = rto;
316 
317 		if (state == TCP_TIME_WAIT)
318 			timeo = TCP_TIMEWAIT_LEN;
319 
320 		/* tw_timer is pinned, so we need to make sure BH are disabled
321 		 * in following section, otherwise timer handler could run before
322 		 * we complete the initialization.
323 		 */
324 		local_bh_disable();
325 		inet_twsk_schedule(tw, timeo);
326 		/* Linkage updates.
327 		 * Note that access to tw after this point is illegal.
328 		 */
329 		inet_twsk_hashdance(tw, sk, &tcp_hashinfo);
330 		local_bh_enable();
331 	} else {
332 		/* Sorry, if we're out of memory, just CLOSE this
333 		 * socket up.  We've got bigger problems than
334 		 * non-graceful socket closings.
335 		 */
336 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEWAITOVERFLOW);
337 	}
338 
339 	tcp_update_metrics(sk);
340 	tcp_done(sk);
341 }
342 EXPORT_SYMBOL(tcp_time_wait);
343 
344 void tcp_twsk_destructor(struct sock *sk)
345 {
346 #ifdef CONFIG_TCP_MD5SIG
347 	if (static_branch_unlikely(&tcp_md5_needed)) {
348 		struct tcp_timewait_sock *twsk = tcp_twsk(sk);
349 
350 		if (twsk->tw_md5_key)
351 			kfree_rcu(twsk->tw_md5_key, rcu);
352 	}
353 #endif
354 }
355 EXPORT_SYMBOL_GPL(tcp_twsk_destructor);
356 
357 /* Warning : This function is called without sk_listener being locked.
358  * Be sure to read socket fields once, as their value could change under us.
359  */
360 void tcp_openreq_init_rwin(struct request_sock *req,
361 			   const struct sock *sk_listener,
362 			   const struct dst_entry *dst)
363 {
364 	struct inet_request_sock *ireq = inet_rsk(req);
365 	const struct tcp_sock *tp = tcp_sk(sk_listener);
366 	int full_space = tcp_full_space(sk_listener);
367 	u32 window_clamp;
368 	__u8 rcv_wscale;
369 	u32 rcv_wnd;
370 	int mss;
371 
372 	mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
373 	window_clamp = READ_ONCE(tp->window_clamp);
374 	/* Set this up on the first call only */
375 	req->rsk_window_clamp = window_clamp ? : dst_metric(dst, RTAX_WINDOW);
376 
377 	/* limit the window selection if the user enforce a smaller rx buffer */
378 	if (sk_listener->sk_userlocks & SOCK_RCVBUF_LOCK &&
379 	    (req->rsk_window_clamp > full_space || req->rsk_window_clamp == 0))
380 		req->rsk_window_clamp = full_space;
381 
382 	rcv_wnd = tcp_rwnd_init_bpf((struct sock *)req);
383 	if (rcv_wnd == 0)
384 		rcv_wnd = dst_metric(dst, RTAX_INITRWND);
385 	else if (full_space < rcv_wnd * mss)
386 		full_space = rcv_wnd * mss;
387 
388 	/* tcp_full_space because it is guaranteed to be the first packet */
389 	tcp_select_initial_window(sk_listener, full_space,
390 		mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
391 		&req->rsk_rcv_wnd,
392 		&req->rsk_window_clamp,
393 		ireq->wscale_ok,
394 		&rcv_wscale,
395 		rcv_wnd);
396 	ireq->rcv_wscale = rcv_wscale;
397 }
398 EXPORT_SYMBOL(tcp_openreq_init_rwin);
399 
400 static void tcp_ecn_openreq_child(struct tcp_sock *tp,
401 				  const struct request_sock *req)
402 {
403 	tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0;
404 }
405 
406 void tcp_ca_openreq_child(struct sock *sk, const struct dst_entry *dst)
407 {
408 	struct inet_connection_sock *icsk = inet_csk(sk);
409 	u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
410 	bool ca_got_dst = false;
411 
412 	if (ca_key != TCP_CA_UNSPEC) {
413 		const struct tcp_congestion_ops *ca;
414 
415 		rcu_read_lock();
416 		ca = tcp_ca_find_key(ca_key);
417 		if (likely(ca && bpf_try_module_get(ca, ca->owner))) {
418 			icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
419 			icsk->icsk_ca_ops = ca;
420 			ca_got_dst = true;
421 		}
422 		rcu_read_unlock();
423 	}
424 
425 	/* If no valid choice made yet, assign current system default ca. */
426 	if (!ca_got_dst &&
427 	    (!icsk->icsk_ca_setsockopt ||
428 	     !bpf_try_module_get(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner)))
429 		tcp_assign_congestion_control(sk);
430 
431 	tcp_set_ca_state(sk, TCP_CA_Open);
432 }
433 EXPORT_SYMBOL_GPL(tcp_ca_openreq_child);
434 
435 static void smc_check_reset_syn_req(struct tcp_sock *oldtp,
436 				    struct request_sock *req,
437 				    struct tcp_sock *newtp)
438 {
439 #if IS_ENABLED(CONFIG_SMC)
440 	struct inet_request_sock *ireq;
441 
442 	if (static_branch_unlikely(&tcp_have_smc)) {
443 		ireq = inet_rsk(req);
444 		if (oldtp->syn_smc && !ireq->smc_ok)
445 			newtp->syn_smc = 0;
446 	}
447 #endif
448 }
449 
450 /* This is not only more efficient than what we used to do, it eliminates
451  * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM
452  *
453  * Actually, we could lots of memory writes here. tp of listening
454  * socket contains all necessary default parameters.
455  */
456 struct sock *tcp_create_openreq_child(const struct sock *sk,
457 				      struct request_sock *req,
458 				      struct sk_buff *skb)
459 {
460 	struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC);
461 	const struct inet_request_sock *ireq = inet_rsk(req);
462 	struct tcp_request_sock *treq = tcp_rsk(req);
463 	struct inet_connection_sock *newicsk;
464 	struct tcp_sock *oldtp, *newtp;
465 	u32 seq;
466 
467 	if (!newsk)
468 		return NULL;
469 
470 	newicsk = inet_csk(newsk);
471 	newtp = tcp_sk(newsk);
472 	oldtp = tcp_sk(sk);
473 
474 	smc_check_reset_syn_req(oldtp, req, newtp);
475 
476 	/* Now setup tcp_sock */
477 	newtp->pred_flags = 0;
478 
479 	seq = treq->rcv_isn + 1;
480 	newtp->rcv_wup = seq;
481 	WRITE_ONCE(newtp->copied_seq, seq);
482 	WRITE_ONCE(newtp->rcv_nxt, seq);
483 	newtp->segs_in = 1;
484 
485 	seq = treq->snt_isn + 1;
486 	newtp->snd_sml = newtp->snd_una = seq;
487 	WRITE_ONCE(newtp->snd_nxt, seq);
488 	newtp->snd_up = seq;
489 
490 	INIT_LIST_HEAD(&newtp->tsq_node);
491 	INIT_LIST_HEAD(&newtp->tsorted_sent_queue);
492 
493 	tcp_init_wl(newtp, treq->rcv_isn);
494 
495 	minmax_reset(&newtp->rtt_min, tcp_jiffies32, ~0U);
496 	newicsk->icsk_ack.lrcvtime = tcp_jiffies32;
497 
498 	newtp->lsndtime = tcp_jiffies32;
499 	newsk->sk_txhash = treq->txhash;
500 	newtp->total_retrans = req->num_retrans;
501 
502 	tcp_init_xmit_timers(newsk);
503 	WRITE_ONCE(newtp->write_seq, newtp->pushed_seq = treq->snt_isn + 1);
504 
505 	if (sock_flag(newsk, SOCK_KEEPOPEN))
506 		inet_csk_reset_keepalive_timer(newsk,
507 					       keepalive_time_when(newtp));
508 
509 	newtp->rx_opt.tstamp_ok = ireq->tstamp_ok;
510 	newtp->rx_opt.sack_ok = ireq->sack_ok;
511 	newtp->window_clamp = req->rsk_window_clamp;
512 	newtp->rcv_ssthresh = req->rsk_rcv_wnd;
513 	newtp->rcv_wnd = req->rsk_rcv_wnd;
514 	newtp->rx_opt.wscale_ok = ireq->wscale_ok;
515 	if (newtp->rx_opt.wscale_ok) {
516 		newtp->rx_opt.snd_wscale = ireq->snd_wscale;
517 		newtp->rx_opt.rcv_wscale = ireq->rcv_wscale;
518 	} else {
519 		newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0;
520 		newtp->window_clamp = min(newtp->window_clamp, 65535U);
521 	}
522 	newtp->snd_wnd = ntohs(tcp_hdr(skb)->window) << newtp->rx_opt.snd_wscale;
523 	newtp->max_window = newtp->snd_wnd;
524 
525 	if (newtp->rx_opt.tstamp_ok) {
526 		newtp->rx_opt.ts_recent = req->ts_recent;
527 		newtp->rx_opt.ts_recent_stamp = ktime_get_seconds();
528 		newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
529 	} else {
530 		newtp->rx_opt.ts_recent_stamp = 0;
531 		newtp->tcp_header_len = sizeof(struct tcphdr);
532 	}
533 	if (req->num_timeout) {
534 		newtp->undo_marker = treq->snt_isn;
535 		newtp->retrans_stamp = div_u64(treq->snt_synack,
536 					       USEC_PER_SEC / TCP_TS_HZ);
537 	}
538 	newtp->tsoffset = treq->ts_off;
539 #ifdef CONFIG_TCP_MD5SIG
540 	newtp->md5sig_info = NULL;	/*XXX*/
541 	if (newtp->af_specific->md5_lookup(sk, newsk))
542 		newtp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
543 #endif
544 	if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len)
545 		newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len;
546 	newtp->rx_opt.mss_clamp = req->mss;
547 	tcp_ecn_openreq_child(newtp, req);
548 	newtp->fastopen_req = NULL;
549 	RCU_INIT_POINTER(newtp->fastopen_rsk, NULL);
550 
551 	__TCP_INC_STATS(sock_net(sk), TCP_MIB_PASSIVEOPENS);
552 
553 	return newsk;
554 }
555 EXPORT_SYMBOL(tcp_create_openreq_child);
556 
557 /*
558  * Process an incoming packet for SYN_RECV sockets represented as a
559  * request_sock. Normally sk is the listener socket but for TFO it
560  * points to the child socket.
561  *
562  * XXX (TFO) - The current impl contains a special check for ack
563  * validation and inside tcp_v4_reqsk_send_ack(). Can we do better?
564  *
565  * We don't need to initialize tmp_opt.sack_ok as we don't use the results
566  */
567 
568 struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb,
569 			   struct request_sock *req,
570 			   bool fastopen, bool *req_stolen)
571 {
572 	struct tcp_options_received tmp_opt;
573 	struct sock *child;
574 	const struct tcphdr *th = tcp_hdr(skb);
575 	__be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK);
576 	bool paws_reject = false;
577 	bool own_req;
578 
579 	tmp_opt.saw_tstamp = 0;
580 	if (th->doff > (sizeof(struct tcphdr)>>2)) {
581 		tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0, NULL);
582 
583 		if (tmp_opt.saw_tstamp) {
584 			tmp_opt.ts_recent = req->ts_recent;
585 			if (tmp_opt.rcv_tsecr)
586 				tmp_opt.rcv_tsecr -= tcp_rsk(req)->ts_off;
587 			/* We do not store true stamp, but it is not required,
588 			 * it can be estimated (approximately)
589 			 * from another data.
590 			 */
591 			tmp_opt.ts_recent_stamp = ktime_get_seconds() - ((TCP_TIMEOUT_INIT/HZ)<<req->num_timeout);
592 			paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
593 		}
594 	}
595 
596 	/* Check for pure retransmitted SYN. */
597 	if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn &&
598 	    flg == TCP_FLAG_SYN &&
599 	    !paws_reject) {
600 		/*
601 		 * RFC793 draws (Incorrectly! It was fixed in RFC1122)
602 		 * this case on figure 6 and figure 8, but formal
603 		 * protocol description says NOTHING.
604 		 * To be more exact, it says that we should send ACK,
605 		 * because this segment (at least, if it has no data)
606 		 * is out of window.
607 		 *
608 		 *  CONCLUSION: RFC793 (even with RFC1122) DOES NOT
609 		 *  describe SYN-RECV state. All the description
610 		 *  is wrong, we cannot believe to it and should
611 		 *  rely only on common sense and implementation
612 		 *  experience.
613 		 *
614 		 * Enforce "SYN-ACK" according to figure 8, figure 6
615 		 * of RFC793, fixed by RFC1122.
616 		 *
617 		 * Note that even if there is new data in the SYN packet
618 		 * they will be thrown away too.
619 		 *
620 		 * Reset timer after retransmitting SYNACK, similar to
621 		 * the idea of fast retransmit in recovery.
622 		 */
623 		if (!tcp_oow_rate_limited(sock_net(sk), skb,
624 					  LINUX_MIB_TCPACKSKIPPEDSYNRECV,
625 					  &tcp_rsk(req)->last_oow_ack_time) &&
626 
627 		    !inet_rtx_syn_ack(sk, req)) {
628 			unsigned long expires = jiffies;
629 
630 			expires += min(TCP_TIMEOUT_INIT << req->num_timeout,
631 				       TCP_RTO_MAX);
632 			if (!fastopen)
633 				mod_timer_pending(&req->rsk_timer, expires);
634 			else
635 				req->rsk_timer.expires = expires;
636 		}
637 		return NULL;
638 	}
639 
640 	/* Further reproduces section "SEGMENT ARRIVES"
641 	   for state SYN-RECEIVED of RFC793.
642 	   It is broken, however, it does not work only
643 	   when SYNs are crossed.
644 
645 	   You would think that SYN crossing is impossible here, since
646 	   we should have a SYN_SENT socket (from connect()) on our end,
647 	   but this is not true if the crossed SYNs were sent to both
648 	   ends by a malicious third party.  We must defend against this,
649 	   and to do that we first verify the ACK (as per RFC793, page
650 	   36) and reset if it is invalid.  Is this a true full defense?
651 	   To convince ourselves, let us consider a way in which the ACK
652 	   test can still pass in this 'malicious crossed SYNs' case.
653 	   Malicious sender sends identical SYNs (and thus identical sequence
654 	   numbers) to both A and B:
655 
656 		A: gets SYN, seq=7
657 		B: gets SYN, seq=7
658 
659 	   By our good fortune, both A and B select the same initial
660 	   send sequence number of seven :-)
661 
662 		A: sends SYN|ACK, seq=7, ack_seq=8
663 		B: sends SYN|ACK, seq=7, ack_seq=8
664 
665 	   So we are now A eating this SYN|ACK, ACK test passes.  So
666 	   does sequence test, SYN is truncated, and thus we consider
667 	   it a bare ACK.
668 
669 	   If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this
670 	   bare ACK.  Otherwise, we create an established connection.  Both
671 	   ends (listening sockets) accept the new incoming connection and try
672 	   to talk to each other. 8-)
673 
674 	   Note: This case is both harmless, and rare.  Possibility is about the
675 	   same as us discovering intelligent life on another plant tomorrow.
676 
677 	   But generally, we should (RFC lies!) to accept ACK
678 	   from SYNACK both here and in tcp_rcv_state_process().
679 	   tcp_rcv_state_process() does not, hence, we do not too.
680 
681 	   Note that the case is absolutely generic:
682 	   we cannot optimize anything here without
683 	   violating protocol. All the checks must be made
684 	   before attempt to create socket.
685 	 */
686 
687 	/* RFC793 page 36: "If the connection is in any non-synchronized state ...
688 	 *                  and the incoming segment acknowledges something not yet
689 	 *                  sent (the segment carries an unacceptable ACK) ...
690 	 *                  a reset is sent."
691 	 *
692 	 * Invalid ACK: reset will be sent by listening socket.
693 	 * Note that the ACK validity check for a Fast Open socket is done
694 	 * elsewhere and is checked directly against the child socket rather
695 	 * than req because user data may have been sent out.
696 	 */
697 	if ((flg & TCP_FLAG_ACK) && !fastopen &&
698 	    (TCP_SKB_CB(skb)->ack_seq !=
699 	     tcp_rsk(req)->snt_isn + 1))
700 		return sk;
701 
702 	/* Also, it would be not so bad idea to check rcv_tsecr, which
703 	 * is essentially ACK extension and too early or too late values
704 	 * should cause reset in unsynchronized states.
705 	 */
706 
707 	/* RFC793: "first check sequence number". */
708 
709 	if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
710 					  tcp_rsk(req)->rcv_nxt, tcp_rsk(req)->rcv_nxt + req->rsk_rcv_wnd)) {
711 		/* Out of window: send ACK and drop. */
712 		if (!(flg & TCP_FLAG_RST) &&
713 		    !tcp_oow_rate_limited(sock_net(sk), skb,
714 					  LINUX_MIB_TCPACKSKIPPEDSYNRECV,
715 					  &tcp_rsk(req)->last_oow_ack_time))
716 			req->rsk_ops->send_ack(sk, skb, req);
717 		if (paws_reject)
718 			__NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
719 		return NULL;
720 	}
721 
722 	/* In sequence, PAWS is OK. */
723 
724 	if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_nxt))
725 		req->ts_recent = tmp_opt.rcv_tsval;
726 
727 	if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) {
728 		/* Truncate SYN, it is out of window starting
729 		   at tcp_rsk(req)->rcv_isn + 1. */
730 		flg &= ~TCP_FLAG_SYN;
731 	}
732 
733 	/* RFC793: "second check the RST bit" and
734 	 *	   "fourth, check the SYN bit"
735 	 */
736 	if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) {
737 		__TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
738 		goto embryonic_reset;
739 	}
740 
741 	/* ACK sequence verified above, just make sure ACK is
742 	 * set.  If ACK not set, just silently drop the packet.
743 	 *
744 	 * XXX (TFO) - if we ever allow "data after SYN", the
745 	 * following check needs to be removed.
746 	 */
747 	if (!(flg & TCP_FLAG_ACK))
748 		return NULL;
749 
750 	/* For Fast Open no more processing is needed (sk is the
751 	 * child socket).
752 	 */
753 	if (fastopen)
754 		return sk;
755 
756 	/* While TCP_DEFER_ACCEPT is active, drop bare ACK. */
757 	if (req->num_timeout < inet_csk(sk)->icsk_accept_queue.rskq_defer_accept &&
758 	    TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) {
759 		inet_rsk(req)->acked = 1;
760 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP);
761 		return NULL;
762 	}
763 
764 	/* OK, ACK is valid, create big socket and
765 	 * feed this segment to it. It will repeat all
766 	 * the tests. THIS SEGMENT MUST MOVE SOCKET TO
767 	 * ESTABLISHED STATE. If it will be dropped after
768 	 * socket is created, wait for troubles.
769 	 */
770 	child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL,
771 							 req, &own_req);
772 	if (!child)
773 		goto listen_overflow;
774 
775 	sock_rps_save_rxhash(child, skb);
776 	tcp_synack_rtt_meas(child, req);
777 	*req_stolen = !own_req;
778 	return inet_csk_complete_hashdance(sk, child, req, own_req);
779 
780 listen_overflow:
781 	if (!sock_net(sk)->ipv4.sysctl_tcp_abort_on_overflow) {
782 		inet_rsk(req)->acked = 1;
783 		return NULL;
784 	}
785 
786 embryonic_reset:
787 	if (!(flg & TCP_FLAG_RST)) {
788 		/* Received a bad SYN pkt - for TFO We try not to reset
789 		 * the local connection unless it's really necessary to
790 		 * avoid becoming vulnerable to outside attack aiming at
791 		 * resetting legit local connections.
792 		 */
793 		req->rsk_ops->send_reset(sk, skb);
794 	} else if (fastopen) { /* received a valid RST pkt */
795 		reqsk_fastopen_remove(sk, req, true);
796 		tcp_reset(sk);
797 	}
798 	if (!fastopen) {
799 		inet_csk_reqsk_queue_drop(sk, req);
800 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_EMBRYONICRSTS);
801 	}
802 	return NULL;
803 }
804 EXPORT_SYMBOL(tcp_check_req);
805 
806 /*
807  * Queue segment on the new socket if the new socket is active,
808  * otherwise we just shortcircuit this and continue with
809  * the new socket.
810  *
811  * For the vast majority of cases child->sk_state will be TCP_SYN_RECV
812  * when entering. But other states are possible due to a race condition
813  * where after __inet_lookup_established() fails but before the listener
814  * locked is obtained, other packets cause the same connection to
815  * be created.
816  */
817 
818 int tcp_child_process(struct sock *parent, struct sock *child,
819 		      struct sk_buff *skb)
820 {
821 	int ret = 0;
822 	int state = child->sk_state;
823 
824 	/* record NAPI ID of child */
825 	sk_mark_napi_id(child, skb);
826 
827 	tcp_segs_in(tcp_sk(child), skb);
828 	if (!sock_owned_by_user(child)) {
829 		ret = tcp_rcv_state_process(child, skb);
830 		/* Wakeup parent, send SIGIO */
831 		if (state == TCP_SYN_RECV && child->sk_state != state)
832 			parent->sk_data_ready(parent);
833 	} else {
834 		/* Alas, it is possible again, because we do lookup
835 		 * in main socket hash table and lock on listening
836 		 * socket does not protect us more.
837 		 */
838 		__sk_add_backlog(child, skb);
839 	}
840 
841 	bh_unlock_sock(child);
842 	sock_put(child);
843 	return ret;
844 }
845 EXPORT_SYMBOL(tcp_child_process);
846