xref: /openbmc/linux/net/ipv4/tcp_timer.c (revision 2572f00d)
1 /*
2  * INET		An implementation of the TCP/IP protocol suite for the LINUX
3  *		operating system.  INET is implemented using the  BSD Socket
4  *		interface as the means of communication with the user level.
5  *
6  *		Implementation of the Transmission Control Protocol(TCP).
7  *
8  * Authors:	Ross Biro
9  *		Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10  *		Mark Evans, <evansmp@uhura.aston.ac.uk>
11  *		Corey Minyard <wf-rch!minyard@relay.EU.net>
12  *		Florian La Roche, <flla@stud.uni-sb.de>
13  *		Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14  *		Linus Torvalds, <torvalds@cs.helsinki.fi>
15  *		Alan Cox, <gw4pts@gw4pts.ampr.org>
16  *		Matthew Dillon, <dillon@apollo.west.oic.com>
17  *		Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18  *		Jorge Cwik, <jorge@laser.satlink.net>
19  */
20 
21 #include <linux/module.h>
22 #include <linux/gfp.h>
23 #include <net/tcp.h>
24 
25 int sysctl_tcp_syn_retries __read_mostly = TCP_SYN_RETRIES;
26 int sysctl_tcp_synack_retries __read_mostly = TCP_SYNACK_RETRIES;
27 int sysctl_tcp_keepalive_time __read_mostly = TCP_KEEPALIVE_TIME;
28 int sysctl_tcp_keepalive_probes __read_mostly = TCP_KEEPALIVE_PROBES;
29 int sysctl_tcp_keepalive_intvl __read_mostly = TCP_KEEPALIVE_INTVL;
30 int sysctl_tcp_retries1 __read_mostly = TCP_RETR1;
31 int sysctl_tcp_retries2 __read_mostly = TCP_RETR2;
32 int sysctl_tcp_orphan_retries __read_mostly;
33 int sysctl_tcp_thin_linear_timeouts __read_mostly;
34 
35 static void tcp_write_err(struct sock *sk)
36 {
37 	sk->sk_err = sk->sk_err_soft ? : ETIMEDOUT;
38 	sk->sk_error_report(sk);
39 
40 	tcp_done(sk);
41 	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT);
42 }
43 
44 /* Do not allow orphaned sockets to eat all our resources.
45  * This is direct violation of TCP specs, but it is required
46  * to prevent DoS attacks. It is called when a retransmission timeout
47  * or zero probe timeout occurs on orphaned socket.
48  *
49  * Criteria is still not confirmed experimentally and may change.
50  * We kill the socket, if:
51  * 1. If number of orphaned sockets exceeds an administratively configured
52  *    limit.
53  * 2. If we have strong memory pressure.
54  */
55 static int tcp_out_of_resources(struct sock *sk, bool do_reset)
56 {
57 	struct tcp_sock *tp = tcp_sk(sk);
58 	int shift = 0;
59 
60 	/* If peer does not open window for long time, or did not transmit
61 	 * anything for long time, penalize it. */
62 	if ((s32)(tcp_time_stamp - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset)
63 		shift++;
64 
65 	/* If some dubious ICMP arrived, penalize even more. */
66 	if (sk->sk_err_soft)
67 		shift++;
68 
69 	if (tcp_check_oom(sk, shift)) {
70 		/* Catch exceptional cases, when connection requires reset.
71 		 *      1. Last segment was sent recently. */
72 		if ((s32)(tcp_time_stamp - tp->lsndtime) <= TCP_TIMEWAIT_LEN ||
73 		    /*  2. Window is closed. */
74 		    (!tp->snd_wnd && !tp->packets_out))
75 			do_reset = true;
76 		if (do_reset)
77 			tcp_send_active_reset(sk, GFP_ATOMIC);
78 		tcp_done(sk);
79 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY);
80 		return 1;
81 	}
82 	return 0;
83 }
84 
85 /* Calculate maximal number or retries on an orphaned socket. */
86 static int tcp_orphan_retries(struct sock *sk, bool alive)
87 {
88 	int retries = sysctl_tcp_orphan_retries; /* May be zero. */
89 
90 	/* We know from an ICMP that something is wrong. */
91 	if (sk->sk_err_soft && !alive)
92 		retries = 0;
93 
94 	/* However, if socket sent something recently, select some safe
95 	 * number of retries. 8 corresponds to >100 seconds with minimal
96 	 * RTO of 200msec. */
97 	if (retries == 0 && alive)
98 		retries = 8;
99 	return retries;
100 }
101 
102 static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk)
103 {
104 	struct net *net = sock_net(sk);
105 
106 	/* Black hole detection */
107 	if (net->ipv4.sysctl_tcp_mtu_probing) {
108 		if (!icsk->icsk_mtup.enabled) {
109 			icsk->icsk_mtup.enabled = 1;
110 			icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
111 			tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
112 		} else {
113 			struct net *net = sock_net(sk);
114 			struct tcp_sock *tp = tcp_sk(sk);
115 			int mss;
116 
117 			mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1;
118 			mss = min(net->ipv4.sysctl_tcp_base_mss, mss);
119 			mss = max(mss, 68 - tp->tcp_header_len);
120 			icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
121 			tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
122 		}
123 	}
124 }
125 
126 /* This function calculates a "timeout" which is equivalent to the timeout of a
127  * TCP connection after "boundary" unsuccessful, exponentially backed-off
128  * retransmissions with an initial RTO of TCP_RTO_MIN or TCP_TIMEOUT_INIT if
129  * syn_set flag is set.
130  */
131 static bool retransmits_timed_out(struct sock *sk,
132 				  unsigned int boundary,
133 				  unsigned int timeout,
134 				  bool syn_set)
135 {
136 	unsigned int linear_backoff_thresh, start_ts;
137 	unsigned int rto_base = syn_set ? TCP_TIMEOUT_INIT : TCP_RTO_MIN;
138 
139 	if (!inet_csk(sk)->icsk_retransmits)
140 		return false;
141 
142 	start_ts = tcp_sk(sk)->retrans_stamp;
143 	if (unlikely(!start_ts))
144 		start_ts = tcp_skb_timestamp(tcp_write_queue_head(sk));
145 
146 	if (likely(timeout == 0)) {
147 		linear_backoff_thresh = ilog2(TCP_RTO_MAX/rto_base);
148 
149 		if (boundary <= linear_backoff_thresh)
150 			timeout = ((2 << boundary) - 1) * rto_base;
151 		else
152 			timeout = ((2 << linear_backoff_thresh) - 1) * rto_base +
153 				(boundary - linear_backoff_thresh) * TCP_RTO_MAX;
154 	}
155 	return (tcp_time_stamp - start_ts) >= timeout;
156 }
157 
158 /* A write timeout has occurred. Process the after effects. */
159 static int tcp_write_timeout(struct sock *sk)
160 {
161 	struct inet_connection_sock *icsk = inet_csk(sk);
162 	struct tcp_sock *tp = tcp_sk(sk);
163 	int retry_until;
164 	bool do_reset, syn_set = false;
165 
166 	if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
167 		if (icsk->icsk_retransmits) {
168 			dst_negative_advice(sk);
169 			if (tp->syn_fastopen || tp->syn_data)
170 				tcp_fastopen_cache_set(sk, 0, NULL, true, 0);
171 			if (tp->syn_data && icsk->icsk_retransmits == 1)
172 				NET_INC_STATS_BH(sock_net(sk),
173 						 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
174 		}
175 		retry_until = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
176 		syn_set = true;
177 	} else {
178 		if (retransmits_timed_out(sk, sysctl_tcp_retries1, 0, 0)) {
179 			/* Some middle-boxes may black-hole Fast Open _after_
180 			 * the handshake. Therefore we conservatively disable
181 			 * Fast Open on this path on recurring timeouts with
182 			 * few or zero bytes acked after Fast Open.
183 			 */
184 			if (tp->syn_data_acked &&
185 			    tp->bytes_acked <= tp->rx_opt.mss_clamp) {
186 				tcp_fastopen_cache_set(sk, 0, NULL, true, 0);
187 				if (icsk->icsk_retransmits == sysctl_tcp_retries1)
188 					NET_INC_STATS_BH(sock_net(sk),
189 							 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
190 			}
191 			/* Black hole detection */
192 			tcp_mtu_probing(icsk, sk);
193 
194 			dst_negative_advice(sk);
195 		}
196 
197 		retry_until = sysctl_tcp_retries2;
198 		if (sock_flag(sk, SOCK_DEAD)) {
199 			const bool alive = icsk->icsk_rto < TCP_RTO_MAX;
200 
201 			retry_until = tcp_orphan_retries(sk, alive);
202 			do_reset = alive ||
203 				!retransmits_timed_out(sk, retry_until, 0, 0);
204 
205 			if (tcp_out_of_resources(sk, do_reset))
206 				return 1;
207 		}
208 	}
209 
210 	if (retransmits_timed_out(sk, retry_until,
211 				  syn_set ? 0 : icsk->icsk_user_timeout, syn_set)) {
212 		/* Has it gone just too far? */
213 		tcp_write_err(sk);
214 		return 1;
215 	}
216 	return 0;
217 }
218 
219 void tcp_delack_timer_handler(struct sock *sk)
220 {
221 	struct tcp_sock *tp = tcp_sk(sk);
222 	struct inet_connection_sock *icsk = inet_csk(sk);
223 
224 	sk_mem_reclaim_partial(sk);
225 
226 	if (sk->sk_state == TCP_CLOSE || !(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
227 		goto out;
228 
229 	if (time_after(icsk->icsk_ack.timeout, jiffies)) {
230 		sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
231 		goto out;
232 	}
233 	icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
234 
235 	if (!skb_queue_empty(&tp->ucopy.prequeue)) {
236 		struct sk_buff *skb;
237 
238 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSCHEDULERFAILED);
239 
240 		while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
241 			sk_backlog_rcv(sk, skb);
242 
243 		tp->ucopy.memory = 0;
244 	}
245 
246 	if (inet_csk_ack_scheduled(sk)) {
247 		if (!icsk->icsk_ack.pingpong) {
248 			/* Delayed ACK missed: inflate ATO. */
249 			icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto);
250 		} else {
251 			/* Delayed ACK missed: leave pingpong mode and
252 			 * deflate ATO.
253 			 */
254 			icsk->icsk_ack.pingpong = 0;
255 			icsk->icsk_ack.ato      = TCP_ATO_MIN;
256 		}
257 		tcp_send_ack(sk);
258 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKS);
259 	}
260 
261 out:
262 	if (tcp_under_memory_pressure(sk))
263 		sk_mem_reclaim(sk);
264 }
265 
266 static void tcp_delack_timer(unsigned long data)
267 {
268 	struct sock *sk = (struct sock *)data;
269 
270 	bh_lock_sock(sk);
271 	if (!sock_owned_by_user(sk)) {
272 		tcp_delack_timer_handler(sk);
273 	} else {
274 		inet_csk(sk)->icsk_ack.blocked = 1;
275 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
276 		/* deleguate our work to tcp_release_cb() */
277 		if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &tcp_sk(sk)->tsq_flags))
278 			sock_hold(sk);
279 	}
280 	bh_unlock_sock(sk);
281 	sock_put(sk);
282 }
283 
284 static void tcp_probe_timer(struct sock *sk)
285 {
286 	struct inet_connection_sock *icsk = inet_csk(sk);
287 	struct tcp_sock *tp = tcp_sk(sk);
288 	int max_probes;
289 	u32 start_ts;
290 
291 	if (tp->packets_out || !tcp_send_head(sk)) {
292 		icsk->icsk_probes_out = 0;
293 		return;
294 	}
295 
296 	/* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as
297 	 * long as the receiver continues to respond probes. We support this by
298 	 * default and reset icsk_probes_out with incoming ACKs. But if the
299 	 * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we
300 	 * kill the socket when the retry count and the time exceeds the
301 	 * corresponding system limit. We also implement similar policy when
302 	 * we use RTO to probe window in tcp_retransmit_timer().
303 	 */
304 	start_ts = tcp_skb_timestamp(tcp_send_head(sk));
305 	if (!start_ts)
306 		skb_mstamp_get(&tcp_send_head(sk)->skb_mstamp);
307 	else if (icsk->icsk_user_timeout &&
308 		 (s32)(tcp_time_stamp - start_ts) > icsk->icsk_user_timeout)
309 		goto abort;
310 
311 	max_probes = sysctl_tcp_retries2;
312 	if (sock_flag(sk, SOCK_DEAD)) {
313 		const bool alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX;
314 
315 		max_probes = tcp_orphan_retries(sk, alive);
316 		if (!alive && icsk->icsk_backoff >= max_probes)
317 			goto abort;
318 		if (tcp_out_of_resources(sk, true))
319 			return;
320 	}
321 
322 	if (icsk->icsk_probes_out > max_probes) {
323 abort:		tcp_write_err(sk);
324 	} else {
325 		/* Only send another probe if we didn't close things up. */
326 		tcp_send_probe0(sk);
327 	}
328 }
329 
330 /*
331  *	Timer for Fast Open socket to retransmit SYNACK. Note that the
332  *	sk here is the child socket, not the parent (listener) socket.
333  */
334 static void tcp_fastopen_synack_timer(struct sock *sk)
335 {
336 	struct inet_connection_sock *icsk = inet_csk(sk);
337 	int max_retries = icsk->icsk_syn_retries ? :
338 	    sysctl_tcp_synack_retries + 1; /* add one more retry for fastopen */
339 	struct request_sock *req;
340 
341 	req = tcp_sk(sk)->fastopen_rsk;
342 	req->rsk_ops->syn_ack_timeout(req);
343 
344 	if (req->num_timeout >= max_retries) {
345 		tcp_write_err(sk);
346 		return;
347 	}
348 	/* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
349 	 * returned from rtx_syn_ack() to make it more persistent like
350 	 * regular retransmit because if the child socket has been accepted
351 	 * it's not good to give up too easily.
352 	 */
353 	inet_rtx_syn_ack(sk, req);
354 	req->num_timeout++;
355 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
356 			  TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
357 }
358 
359 /*
360  *	The TCP retransmit timer.
361  */
362 
363 void tcp_retransmit_timer(struct sock *sk)
364 {
365 	struct tcp_sock *tp = tcp_sk(sk);
366 	struct inet_connection_sock *icsk = inet_csk(sk);
367 
368 	if (tp->fastopen_rsk) {
369 		WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
370 			     sk->sk_state != TCP_FIN_WAIT1);
371 		tcp_fastopen_synack_timer(sk);
372 		/* Before we receive ACK to our SYN-ACK don't retransmit
373 		 * anything else (e.g., data or FIN segments).
374 		 */
375 		return;
376 	}
377 	if (!tp->packets_out)
378 		goto out;
379 
380 	WARN_ON(tcp_write_queue_empty(sk));
381 
382 	tp->tlp_high_seq = 0;
383 
384 	if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
385 	    !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
386 		/* Receiver dastardly shrinks window. Our retransmits
387 		 * become zero probes, but we should not timeout this
388 		 * connection. If the socket is an orphan, time it out,
389 		 * we cannot allow such beasts to hang infinitely.
390 		 */
391 		struct inet_sock *inet = inet_sk(sk);
392 		if (sk->sk_family == AF_INET) {
393 			net_dbg_ratelimited("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
394 					    &inet->inet_daddr,
395 					    ntohs(inet->inet_dport),
396 					    inet->inet_num,
397 					    tp->snd_una, tp->snd_nxt);
398 		}
399 #if IS_ENABLED(CONFIG_IPV6)
400 		else if (sk->sk_family == AF_INET6) {
401 			net_dbg_ratelimited("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
402 					    &sk->sk_v6_daddr,
403 					    ntohs(inet->inet_dport),
404 					    inet->inet_num,
405 					    tp->snd_una, tp->snd_nxt);
406 		}
407 #endif
408 		if (tcp_time_stamp - tp->rcv_tstamp > TCP_RTO_MAX) {
409 			tcp_write_err(sk);
410 			goto out;
411 		}
412 		tcp_enter_loss(sk);
413 		tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
414 		__sk_dst_reset(sk);
415 		goto out_reset_timer;
416 	}
417 
418 	if (tcp_write_timeout(sk))
419 		goto out;
420 
421 	if (icsk->icsk_retransmits == 0) {
422 		int mib_idx;
423 
424 		if (icsk->icsk_ca_state == TCP_CA_Recovery) {
425 			if (tcp_is_sack(tp))
426 				mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
427 			else
428 				mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
429 		} else if (icsk->icsk_ca_state == TCP_CA_Loss) {
430 			mib_idx = LINUX_MIB_TCPLOSSFAILURES;
431 		} else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
432 			   tp->sacked_out) {
433 			if (tcp_is_sack(tp))
434 				mib_idx = LINUX_MIB_TCPSACKFAILURES;
435 			else
436 				mib_idx = LINUX_MIB_TCPRENOFAILURES;
437 		} else {
438 			mib_idx = LINUX_MIB_TCPTIMEOUTS;
439 		}
440 		NET_INC_STATS_BH(sock_net(sk), mib_idx);
441 	}
442 
443 	tcp_enter_loss(sk);
444 
445 	if (tcp_retransmit_skb(sk, tcp_write_queue_head(sk)) > 0) {
446 		/* Retransmission failed because of local congestion,
447 		 * do not backoff.
448 		 */
449 		if (!icsk->icsk_retransmits)
450 			icsk->icsk_retransmits = 1;
451 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
452 					  min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL),
453 					  TCP_RTO_MAX);
454 		goto out;
455 	}
456 
457 	/* Increase the timeout each time we retransmit.  Note that
458 	 * we do not increase the rtt estimate.  rto is initialized
459 	 * from rtt, but increases here.  Jacobson (SIGCOMM 88) suggests
460 	 * that doubling rto each time is the least we can get away with.
461 	 * In KA9Q, Karn uses this for the first few times, and then
462 	 * goes to quadratic.  netBSD doubles, but only goes up to *64,
463 	 * and clamps at 1 to 64 sec afterwards.  Note that 120 sec is
464 	 * defined in the protocol as the maximum possible RTT.  I guess
465 	 * we'll have to use something other than TCP to talk to the
466 	 * University of Mars.
467 	 *
468 	 * PAWS allows us longer timeouts and large windows, so once
469 	 * implemented ftp to mars will work nicely. We will have to fix
470 	 * the 120 second clamps though!
471 	 */
472 	icsk->icsk_backoff++;
473 	icsk->icsk_retransmits++;
474 
475 out_reset_timer:
476 	/* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
477 	 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
478 	 * might be increased if the stream oscillates between thin and thick,
479 	 * thus the old value might already be too high compared to the value
480 	 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
481 	 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
482 	 * exponential backoff behaviour to avoid continue hammering
483 	 * linear-timeout retransmissions into a black hole
484 	 */
485 	if (sk->sk_state == TCP_ESTABLISHED &&
486 	    (tp->thin_lto || sysctl_tcp_thin_linear_timeouts) &&
487 	    tcp_stream_is_thin(tp) &&
488 	    icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
489 		icsk->icsk_backoff = 0;
490 		icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX);
491 	} else {
492 		/* Use normal (exponential) backoff */
493 		icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
494 	}
495 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX);
496 	if (retransmits_timed_out(sk, sysctl_tcp_retries1 + 1, 0, 0))
497 		__sk_dst_reset(sk);
498 
499 out:;
500 }
501 
502 void tcp_write_timer_handler(struct sock *sk)
503 {
504 	struct inet_connection_sock *icsk = inet_csk(sk);
505 	int event;
506 
507 	if (sk->sk_state == TCP_CLOSE || !icsk->icsk_pending)
508 		goto out;
509 
510 	if (time_after(icsk->icsk_timeout, jiffies)) {
511 		sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
512 		goto out;
513 	}
514 
515 	event = icsk->icsk_pending;
516 
517 	switch (event) {
518 	case ICSK_TIME_EARLY_RETRANS:
519 		tcp_resume_early_retransmit(sk);
520 		break;
521 	case ICSK_TIME_LOSS_PROBE:
522 		tcp_send_loss_probe(sk);
523 		break;
524 	case ICSK_TIME_RETRANS:
525 		icsk->icsk_pending = 0;
526 		tcp_retransmit_timer(sk);
527 		break;
528 	case ICSK_TIME_PROBE0:
529 		icsk->icsk_pending = 0;
530 		tcp_probe_timer(sk);
531 		break;
532 	}
533 
534 out:
535 	sk_mem_reclaim(sk);
536 }
537 
538 static void tcp_write_timer(unsigned long data)
539 {
540 	struct sock *sk = (struct sock *)data;
541 
542 	bh_lock_sock(sk);
543 	if (!sock_owned_by_user(sk)) {
544 		tcp_write_timer_handler(sk);
545 	} else {
546 		/* deleguate our work to tcp_release_cb() */
547 		if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &tcp_sk(sk)->tsq_flags))
548 			sock_hold(sk);
549 	}
550 	bh_unlock_sock(sk);
551 	sock_put(sk);
552 }
553 
554 void tcp_syn_ack_timeout(const struct request_sock *req)
555 {
556 	struct net *net = read_pnet(&inet_rsk(req)->ireq_net);
557 
558 	NET_INC_STATS_BH(net, LINUX_MIB_TCPTIMEOUTS);
559 }
560 EXPORT_SYMBOL(tcp_syn_ack_timeout);
561 
562 void tcp_set_keepalive(struct sock *sk, int val)
563 {
564 	if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
565 		return;
566 
567 	if (val && !sock_flag(sk, SOCK_KEEPOPEN))
568 		inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
569 	else if (!val)
570 		inet_csk_delete_keepalive_timer(sk);
571 }
572 
573 
574 static void tcp_keepalive_timer (unsigned long data)
575 {
576 	struct sock *sk = (struct sock *) data;
577 	struct inet_connection_sock *icsk = inet_csk(sk);
578 	struct tcp_sock *tp = tcp_sk(sk);
579 	u32 elapsed;
580 
581 	/* Only process if socket is not in use. */
582 	bh_lock_sock(sk);
583 	if (sock_owned_by_user(sk)) {
584 		/* Try again later. */
585 		inet_csk_reset_keepalive_timer (sk, HZ/20);
586 		goto out;
587 	}
588 
589 	if (sk->sk_state == TCP_LISTEN) {
590 		pr_err("Hmm... keepalive on a LISTEN ???\n");
591 		goto out;
592 	}
593 
594 	if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
595 		if (tp->linger2 >= 0) {
596 			const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
597 
598 			if (tmo > 0) {
599 				tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
600 				goto out;
601 			}
602 		}
603 		tcp_send_active_reset(sk, GFP_ATOMIC);
604 		goto death;
605 	}
606 
607 	if (!sock_flag(sk, SOCK_KEEPOPEN) || sk->sk_state == TCP_CLOSE)
608 		goto out;
609 
610 	elapsed = keepalive_time_when(tp);
611 
612 	/* It is alive without keepalive 8) */
613 	if (tp->packets_out || tcp_send_head(sk))
614 		goto resched;
615 
616 	elapsed = keepalive_time_elapsed(tp);
617 
618 	if (elapsed >= keepalive_time_when(tp)) {
619 		/* If the TCP_USER_TIMEOUT option is enabled, use that
620 		 * to determine when to timeout instead.
621 		 */
622 		if ((icsk->icsk_user_timeout != 0 &&
623 		    elapsed >= icsk->icsk_user_timeout &&
624 		    icsk->icsk_probes_out > 0) ||
625 		    (icsk->icsk_user_timeout == 0 &&
626 		    icsk->icsk_probes_out >= keepalive_probes(tp))) {
627 			tcp_send_active_reset(sk, GFP_ATOMIC);
628 			tcp_write_err(sk);
629 			goto out;
630 		}
631 		if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) {
632 			icsk->icsk_probes_out++;
633 			elapsed = keepalive_intvl_when(tp);
634 		} else {
635 			/* If keepalive was lost due to local congestion,
636 			 * try harder.
637 			 */
638 			elapsed = TCP_RESOURCE_PROBE_INTERVAL;
639 		}
640 	} else {
641 		/* It is tp->rcv_tstamp + keepalive_time_when(tp) */
642 		elapsed = keepalive_time_when(tp) - elapsed;
643 	}
644 
645 	sk_mem_reclaim(sk);
646 
647 resched:
648 	inet_csk_reset_keepalive_timer (sk, elapsed);
649 	goto out;
650 
651 death:
652 	tcp_done(sk);
653 
654 out:
655 	bh_unlock_sock(sk);
656 	sock_put(sk);
657 }
658 
659 void tcp_init_xmit_timers(struct sock *sk)
660 {
661 	inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
662 				  &tcp_keepalive_timer);
663 }
664