xref: /linux/net/ipv4/tcp_timer.c (revision 293d5b43948309434568f4dcbb36cce4c3c51bd5)
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_thin_linear_timeouts __read_mostly;
26 
27 /**
28  *  tcp_write_err() - close socket and save error info
29  *  @sk:  The socket the error has appeared on.
30  *
31  *  Returns: Nothing (void)
32  */
33 
34 static void tcp_write_err(struct sock *sk)
35 {
36 	sk->sk_err = sk->sk_err_soft ? : ETIMEDOUT;
37 	sk->sk_error_report(sk);
38 
39 	tcp_done(sk);
40 	__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT);
41 }
42 
43 /**
44  *  tcp_out_of_resources() - Close socket if out of resources
45  *  @sk:        pointer to current socket
46  *  @do_reset:  send a last packet with reset flag
47  *
48  *  Do not allow orphaned sockets to eat all our resources.
49  *  This is direct violation of TCP specs, but it is required
50  *  to prevent DoS attacks. It is called when a retransmission timeout
51  *  or zero probe timeout occurs on orphaned socket.
52  *
53  *  Criteria is still not confirmed experimentally and may change.
54  *  We kill the socket, if:
55  *  1. If number of orphaned sockets exceeds an administratively configured
56  *     limit.
57  *  2. If we have strong memory pressure.
58  */
59 static int tcp_out_of_resources(struct sock *sk, bool do_reset)
60 {
61 	struct tcp_sock *tp = tcp_sk(sk);
62 	int shift = 0;
63 
64 	/* If peer does not open window for long time, or did not transmit
65 	 * anything for long time, penalize it. */
66 	if ((s32)(tcp_time_stamp - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset)
67 		shift++;
68 
69 	/* If some dubious ICMP arrived, penalize even more. */
70 	if (sk->sk_err_soft)
71 		shift++;
72 
73 	if (tcp_check_oom(sk, shift)) {
74 		/* Catch exceptional cases, when connection requires reset.
75 		 *      1. Last segment was sent recently. */
76 		if ((s32)(tcp_time_stamp - tp->lsndtime) <= TCP_TIMEWAIT_LEN ||
77 		    /*  2. Window is closed. */
78 		    (!tp->snd_wnd && !tp->packets_out))
79 			do_reset = true;
80 		if (do_reset)
81 			tcp_send_active_reset(sk, GFP_ATOMIC);
82 		tcp_done(sk);
83 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY);
84 		return 1;
85 	}
86 	return 0;
87 }
88 
89 /**
90  *  tcp_orphan_retries() - Returns maximal number of retries on an orphaned socket
91  *  @sk:    Pointer to the current socket.
92  *  @alive: bool, socket alive state
93  */
94 static int tcp_orphan_retries(struct sock *sk, bool alive)
95 {
96 	int retries = sock_net(sk)->ipv4.sysctl_tcp_orphan_retries; /* May be zero. */
97 
98 	/* We know from an ICMP that something is wrong. */
99 	if (sk->sk_err_soft && !alive)
100 		retries = 0;
101 
102 	/* However, if socket sent something recently, select some safe
103 	 * number of retries. 8 corresponds to >100 seconds with minimal
104 	 * RTO of 200msec. */
105 	if (retries == 0 && alive)
106 		retries = 8;
107 	return retries;
108 }
109 
110 static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk)
111 {
112 	struct net *net = sock_net(sk);
113 
114 	/* Black hole detection */
115 	if (net->ipv4.sysctl_tcp_mtu_probing) {
116 		if (!icsk->icsk_mtup.enabled) {
117 			icsk->icsk_mtup.enabled = 1;
118 			icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
119 			tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
120 		} else {
121 			struct net *net = sock_net(sk);
122 			struct tcp_sock *tp = tcp_sk(sk);
123 			int mss;
124 
125 			mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1;
126 			mss = min(net->ipv4.sysctl_tcp_base_mss, mss);
127 			mss = max(mss, 68 - tp->tcp_header_len);
128 			icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
129 			tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
130 		}
131 	}
132 }
133 
134 
135 /**
136  *  retransmits_timed_out() - returns true if this connection has timed out
137  *  @sk:       The current socket
138  *  @boundary: max number of retransmissions
139  *  @timeout:  A custom timeout value.
140  *             If set to 0 the default timeout is calculated and used.
141  *             Using TCP_RTO_MIN and the number of unsuccessful retransmits.
142  *  @syn_set:  true if the SYN Bit was set.
143  *
144  * The default "timeout" value this function can calculate and use
145  * is equivalent to the timeout of a TCP Connection
146  * after "boundary" unsuccessful, exponentially backed-off
147  * retransmissions with an initial RTO of TCP_RTO_MIN or TCP_TIMEOUT_INIT if
148  * syn_set flag is set.
149  *
150  */
151 static bool retransmits_timed_out(struct sock *sk,
152 				  unsigned int boundary,
153 				  unsigned int timeout,
154 				  bool syn_set)
155 {
156 	unsigned int linear_backoff_thresh, start_ts;
157 	unsigned int rto_base = syn_set ? TCP_TIMEOUT_INIT : TCP_RTO_MIN;
158 
159 	if (!inet_csk(sk)->icsk_retransmits)
160 		return false;
161 
162 	start_ts = tcp_sk(sk)->retrans_stamp;
163 	if (unlikely(!start_ts))
164 		start_ts = tcp_skb_timestamp(tcp_write_queue_head(sk));
165 
166 	if (likely(timeout == 0)) {
167 		linear_backoff_thresh = ilog2(TCP_RTO_MAX/rto_base);
168 
169 		if (boundary <= linear_backoff_thresh)
170 			timeout = ((2 << boundary) - 1) * rto_base;
171 		else
172 			timeout = ((2 << linear_backoff_thresh) - 1) * rto_base +
173 				(boundary - linear_backoff_thresh) * TCP_RTO_MAX;
174 	}
175 	return (tcp_time_stamp - start_ts) >= timeout;
176 }
177 
178 /* A write timeout has occurred. Process the after effects. */
179 static int tcp_write_timeout(struct sock *sk)
180 {
181 	struct inet_connection_sock *icsk = inet_csk(sk);
182 	struct tcp_sock *tp = tcp_sk(sk);
183 	struct net *net = sock_net(sk);
184 	int retry_until;
185 	bool do_reset, syn_set = false;
186 
187 	if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
188 		if (icsk->icsk_retransmits) {
189 			dst_negative_advice(sk);
190 			if (tp->syn_fastopen || tp->syn_data)
191 				tcp_fastopen_cache_set(sk, 0, NULL, true, 0);
192 			if (tp->syn_data && icsk->icsk_retransmits == 1)
193 				NET_INC_STATS(sock_net(sk),
194 					      LINUX_MIB_TCPFASTOPENACTIVEFAIL);
195 		}
196 		retry_until = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
197 		syn_set = true;
198 	} else {
199 		if (retransmits_timed_out(sk, net->ipv4.sysctl_tcp_retries1, 0, 0)) {
200 			/* Some middle-boxes may black-hole Fast Open _after_
201 			 * the handshake. Therefore we conservatively disable
202 			 * Fast Open on this path on recurring timeouts with
203 			 * few or zero bytes acked after Fast Open.
204 			 */
205 			if (tp->syn_data_acked &&
206 			    tp->bytes_acked <= tp->rx_opt.mss_clamp) {
207 				tcp_fastopen_cache_set(sk, 0, NULL, true, 0);
208 				if (icsk->icsk_retransmits == net->ipv4.sysctl_tcp_retries1)
209 					NET_INC_STATS(sock_net(sk),
210 						      LINUX_MIB_TCPFASTOPENACTIVEFAIL);
211 			}
212 			/* Black hole detection */
213 			tcp_mtu_probing(icsk, sk);
214 
215 			dst_negative_advice(sk);
216 		}
217 
218 		retry_until = net->ipv4.sysctl_tcp_retries2;
219 		if (sock_flag(sk, SOCK_DEAD)) {
220 			const bool alive = icsk->icsk_rto < TCP_RTO_MAX;
221 
222 			retry_until = tcp_orphan_retries(sk, alive);
223 			do_reset = alive ||
224 				!retransmits_timed_out(sk, retry_until, 0, 0);
225 
226 			if (tcp_out_of_resources(sk, do_reset))
227 				return 1;
228 		}
229 	}
230 
231 	if (retransmits_timed_out(sk, retry_until,
232 				  syn_set ? 0 : icsk->icsk_user_timeout, syn_set)) {
233 		/* Has it gone just too far? */
234 		tcp_write_err(sk);
235 		return 1;
236 	}
237 	return 0;
238 }
239 
240 /* Called with BH disabled */
241 void tcp_delack_timer_handler(struct sock *sk)
242 {
243 	struct tcp_sock *tp = tcp_sk(sk);
244 	struct inet_connection_sock *icsk = inet_csk(sk);
245 
246 	sk_mem_reclaim_partial(sk);
247 
248 	if (sk->sk_state == TCP_CLOSE || !(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
249 		goto out;
250 
251 	if (time_after(icsk->icsk_ack.timeout, jiffies)) {
252 		sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
253 		goto out;
254 	}
255 	icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
256 
257 	if (!skb_queue_empty(&tp->ucopy.prequeue)) {
258 		struct sk_buff *skb;
259 
260 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSCHEDULERFAILED);
261 
262 		while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
263 			sk_backlog_rcv(sk, skb);
264 
265 		tp->ucopy.memory = 0;
266 	}
267 
268 	if (inet_csk_ack_scheduled(sk)) {
269 		if (!icsk->icsk_ack.pingpong) {
270 			/* Delayed ACK missed: inflate ATO. */
271 			icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto);
272 		} else {
273 			/* Delayed ACK missed: leave pingpong mode and
274 			 * deflate ATO.
275 			 */
276 			icsk->icsk_ack.pingpong = 0;
277 			icsk->icsk_ack.ato      = TCP_ATO_MIN;
278 		}
279 		tcp_send_ack(sk);
280 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKS);
281 	}
282 
283 out:
284 	if (tcp_under_memory_pressure(sk))
285 		sk_mem_reclaim(sk);
286 }
287 
288 
289 /**
290  *  tcp_delack_timer() - The TCP delayed ACK timeout handler
291  *  @data:  Pointer to the current socket. (gets casted to struct sock *)
292  *
293  *  This function gets (indirectly) called when the kernel timer for a TCP packet
294  *  of this socket expires. Calls tcp_delack_timer_handler() to do the actual work.
295  *
296  *  Returns: Nothing (void)
297  */
298 static void tcp_delack_timer(unsigned long data)
299 {
300 	struct sock *sk = (struct sock *)data;
301 
302 	bh_lock_sock(sk);
303 	if (!sock_owned_by_user(sk)) {
304 		tcp_delack_timer_handler(sk);
305 	} else {
306 		inet_csk(sk)->icsk_ack.blocked = 1;
307 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
308 		/* deleguate our work to tcp_release_cb() */
309 		if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &tcp_sk(sk)->tsq_flags))
310 			sock_hold(sk);
311 	}
312 	bh_unlock_sock(sk);
313 	sock_put(sk);
314 }
315 
316 static void tcp_probe_timer(struct sock *sk)
317 {
318 	struct inet_connection_sock *icsk = inet_csk(sk);
319 	struct tcp_sock *tp = tcp_sk(sk);
320 	int max_probes;
321 	u32 start_ts;
322 
323 	if (tp->packets_out || !tcp_send_head(sk)) {
324 		icsk->icsk_probes_out = 0;
325 		return;
326 	}
327 
328 	/* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as
329 	 * long as the receiver continues to respond probes. We support this by
330 	 * default and reset icsk_probes_out with incoming ACKs. But if the
331 	 * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we
332 	 * kill the socket when the retry count and the time exceeds the
333 	 * corresponding system limit. We also implement similar policy when
334 	 * we use RTO to probe window in tcp_retransmit_timer().
335 	 */
336 	start_ts = tcp_skb_timestamp(tcp_send_head(sk));
337 	if (!start_ts)
338 		skb_mstamp_get(&tcp_send_head(sk)->skb_mstamp);
339 	else if (icsk->icsk_user_timeout &&
340 		 (s32)(tcp_time_stamp - start_ts) > icsk->icsk_user_timeout)
341 		goto abort;
342 
343 	max_probes = sock_net(sk)->ipv4.sysctl_tcp_retries2;
344 	if (sock_flag(sk, SOCK_DEAD)) {
345 		const bool alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX;
346 
347 		max_probes = tcp_orphan_retries(sk, alive);
348 		if (!alive && icsk->icsk_backoff >= max_probes)
349 			goto abort;
350 		if (tcp_out_of_resources(sk, true))
351 			return;
352 	}
353 
354 	if (icsk->icsk_probes_out > max_probes) {
355 abort:		tcp_write_err(sk);
356 	} else {
357 		/* Only send another probe if we didn't close things up. */
358 		tcp_send_probe0(sk);
359 	}
360 }
361 
362 /*
363  *	Timer for Fast Open socket to retransmit SYNACK. Note that the
364  *	sk here is the child socket, not the parent (listener) socket.
365  */
366 static void tcp_fastopen_synack_timer(struct sock *sk)
367 {
368 	struct inet_connection_sock *icsk = inet_csk(sk);
369 	int max_retries = icsk->icsk_syn_retries ? :
370 	    sock_net(sk)->ipv4.sysctl_tcp_synack_retries + 1; /* add one more retry for fastopen */
371 	struct request_sock *req;
372 
373 	req = tcp_sk(sk)->fastopen_rsk;
374 	req->rsk_ops->syn_ack_timeout(req);
375 
376 	if (req->num_timeout >= max_retries) {
377 		tcp_write_err(sk);
378 		return;
379 	}
380 	/* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
381 	 * returned from rtx_syn_ack() to make it more persistent like
382 	 * regular retransmit because if the child socket has been accepted
383 	 * it's not good to give up too easily.
384 	 */
385 	inet_rtx_syn_ack(sk, req);
386 	req->num_timeout++;
387 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
388 			  TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
389 }
390 
391 
392 /**
393  *  tcp_retransmit_timer() - The TCP retransmit timeout handler
394  *  @sk:  Pointer to the current socket.
395  *
396  *  This function gets called when the kernel timer for a TCP packet
397  *  of this socket expires.
398  *
399  *  It handles retransmission, timer adjustment and other necesarry measures.
400  *
401  *  Returns: Nothing (void)
402  */
403 void tcp_retransmit_timer(struct sock *sk)
404 {
405 	struct tcp_sock *tp = tcp_sk(sk);
406 	struct net *net = sock_net(sk);
407 	struct inet_connection_sock *icsk = inet_csk(sk);
408 
409 	if (tp->fastopen_rsk) {
410 		WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
411 			     sk->sk_state != TCP_FIN_WAIT1);
412 		tcp_fastopen_synack_timer(sk);
413 		/* Before we receive ACK to our SYN-ACK don't retransmit
414 		 * anything else (e.g., data or FIN segments).
415 		 */
416 		return;
417 	}
418 	if (!tp->packets_out)
419 		goto out;
420 
421 	WARN_ON(tcp_write_queue_empty(sk));
422 
423 	tp->tlp_high_seq = 0;
424 
425 	if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
426 	    !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
427 		/* Receiver dastardly shrinks window. Our retransmits
428 		 * become zero probes, but we should not timeout this
429 		 * connection. If the socket is an orphan, time it out,
430 		 * we cannot allow such beasts to hang infinitely.
431 		 */
432 		struct inet_sock *inet = inet_sk(sk);
433 		if (sk->sk_family == AF_INET) {
434 			net_dbg_ratelimited("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
435 					    &inet->inet_daddr,
436 					    ntohs(inet->inet_dport),
437 					    inet->inet_num,
438 					    tp->snd_una, tp->snd_nxt);
439 		}
440 #if IS_ENABLED(CONFIG_IPV6)
441 		else if (sk->sk_family == AF_INET6) {
442 			net_dbg_ratelimited("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
443 					    &sk->sk_v6_daddr,
444 					    ntohs(inet->inet_dport),
445 					    inet->inet_num,
446 					    tp->snd_una, tp->snd_nxt);
447 		}
448 #endif
449 		if (tcp_time_stamp - tp->rcv_tstamp > TCP_RTO_MAX) {
450 			tcp_write_err(sk);
451 			goto out;
452 		}
453 		tcp_enter_loss(sk);
454 		tcp_retransmit_skb(sk, tcp_write_queue_head(sk), 1);
455 		__sk_dst_reset(sk);
456 		goto out_reset_timer;
457 	}
458 
459 	if (tcp_write_timeout(sk))
460 		goto out;
461 
462 	if (icsk->icsk_retransmits == 0) {
463 		int mib_idx;
464 
465 		if (icsk->icsk_ca_state == TCP_CA_Recovery) {
466 			if (tcp_is_sack(tp))
467 				mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
468 			else
469 				mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
470 		} else if (icsk->icsk_ca_state == TCP_CA_Loss) {
471 			mib_idx = LINUX_MIB_TCPLOSSFAILURES;
472 		} else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
473 			   tp->sacked_out) {
474 			if (tcp_is_sack(tp))
475 				mib_idx = LINUX_MIB_TCPSACKFAILURES;
476 			else
477 				mib_idx = LINUX_MIB_TCPRENOFAILURES;
478 		} else {
479 			mib_idx = LINUX_MIB_TCPTIMEOUTS;
480 		}
481 		__NET_INC_STATS(sock_net(sk), mib_idx);
482 	}
483 
484 	tcp_enter_loss(sk);
485 
486 	if (tcp_retransmit_skb(sk, tcp_write_queue_head(sk), 1) > 0) {
487 		/* Retransmission failed because of local congestion,
488 		 * do not backoff.
489 		 */
490 		if (!icsk->icsk_retransmits)
491 			icsk->icsk_retransmits = 1;
492 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
493 					  min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL),
494 					  TCP_RTO_MAX);
495 		goto out;
496 	}
497 
498 	/* Increase the timeout each time we retransmit.  Note that
499 	 * we do not increase the rtt estimate.  rto is initialized
500 	 * from rtt, but increases here.  Jacobson (SIGCOMM 88) suggests
501 	 * that doubling rto each time is the least we can get away with.
502 	 * In KA9Q, Karn uses this for the first few times, and then
503 	 * goes to quadratic.  netBSD doubles, but only goes up to *64,
504 	 * and clamps at 1 to 64 sec afterwards.  Note that 120 sec is
505 	 * defined in the protocol as the maximum possible RTT.  I guess
506 	 * we'll have to use something other than TCP to talk to the
507 	 * University of Mars.
508 	 *
509 	 * PAWS allows us longer timeouts and large windows, so once
510 	 * implemented ftp to mars will work nicely. We will have to fix
511 	 * the 120 second clamps though!
512 	 */
513 	icsk->icsk_backoff++;
514 	icsk->icsk_retransmits++;
515 
516 out_reset_timer:
517 	/* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
518 	 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
519 	 * might be increased if the stream oscillates between thin and thick,
520 	 * thus the old value might already be too high compared to the value
521 	 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
522 	 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
523 	 * exponential backoff behaviour to avoid continue hammering
524 	 * linear-timeout retransmissions into a black hole
525 	 */
526 	if (sk->sk_state == TCP_ESTABLISHED &&
527 	    (tp->thin_lto || sysctl_tcp_thin_linear_timeouts) &&
528 	    tcp_stream_is_thin(tp) &&
529 	    icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
530 		icsk->icsk_backoff = 0;
531 		icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX);
532 	} else {
533 		/* Use normal (exponential) backoff */
534 		icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
535 	}
536 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX);
537 	if (retransmits_timed_out(sk, net->ipv4.sysctl_tcp_retries1 + 1, 0, 0))
538 		__sk_dst_reset(sk);
539 
540 out:;
541 }
542 
543 /* Called with bottom-half processing disabled.
544    Called by tcp_write_timer() */
545 void tcp_write_timer_handler(struct sock *sk)
546 {
547 	struct inet_connection_sock *icsk = inet_csk(sk);
548 	int event;
549 
550 	if (sk->sk_state == TCP_CLOSE || !icsk->icsk_pending)
551 		goto out;
552 
553 	if (time_after(icsk->icsk_timeout, jiffies)) {
554 		sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
555 		goto out;
556 	}
557 
558 	event = icsk->icsk_pending;
559 
560 	switch (event) {
561 	case ICSK_TIME_EARLY_RETRANS:
562 		tcp_resume_early_retransmit(sk);
563 		break;
564 	case ICSK_TIME_LOSS_PROBE:
565 		tcp_send_loss_probe(sk);
566 		break;
567 	case ICSK_TIME_RETRANS:
568 		icsk->icsk_pending = 0;
569 		tcp_retransmit_timer(sk);
570 		break;
571 	case ICSK_TIME_PROBE0:
572 		icsk->icsk_pending = 0;
573 		tcp_probe_timer(sk);
574 		break;
575 	}
576 
577 out:
578 	sk_mem_reclaim(sk);
579 }
580 
581 static void tcp_write_timer(unsigned long data)
582 {
583 	struct sock *sk = (struct sock *)data;
584 
585 	bh_lock_sock(sk);
586 	if (!sock_owned_by_user(sk)) {
587 		tcp_write_timer_handler(sk);
588 	} else {
589 		/* delegate our work to tcp_release_cb() */
590 		if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &tcp_sk(sk)->tsq_flags))
591 			sock_hold(sk);
592 	}
593 	bh_unlock_sock(sk);
594 	sock_put(sk);
595 }
596 
597 void tcp_syn_ack_timeout(const struct request_sock *req)
598 {
599 	struct net *net = read_pnet(&inet_rsk(req)->ireq_net);
600 
601 	__NET_INC_STATS(net, LINUX_MIB_TCPTIMEOUTS);
602 }
603 EXPORT_SYMBOL(tcp_syn_ack_timeout);
604 
605 void tcp_set_keepalive(struct sock *sk, int val)
606 {
607 	if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
608 		return;
609 
610 	if (val && !sock_flag(sk, SOCK_KEEPOPEN))
611 		inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
612 	else if (!val)
613 		inet_csk_delete_keepalive_timer(sk);
614 }
615 
616 
617 static void tcp_keepalive_timer (unsigned long data)
618 {
619 	struct sock *sk = (struct sock *) data;
620 	struct inet_connection_sock *icsk = inet_csk(sk);
621 	struct tcp_sock *tp = tcp_sk(sk);
622 	u32 elapsed;
623 
624 	/* Only process if socket is not in use. */
625 	bh_lock_sock(sk);
626 	if (sock_owned_by_user(sk)) {
627 		/* Try again later. */
628 		inet_csk_reset_keepalive_timer (sk, HZ/20);
629 		goto out;
630 	}
631 
632 	if (sk->sk_state == TCP_LISTEN) {
633 		pr_err("Hmm... keepalive on a LISTEN ???\n");
634 		goto out;
635 	}
636 
637 	if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
638 		if (tp->linger2 >= 0) {
639 			const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
640 
641 			if (tmo > 0) {
642 				tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
643 				goto out;
644 			}
645 		}
646 		tcp_send_active_reset(sk, GFP_ATOMIC);
647 		goto death;
648 	}
649 
650 	if (!sock_flag(sk, SOCK_KEEPOPEN) || sk->sk_state == TCP_CLOSE)
651 		goto out;
652 
653 	elapsed = keepalive_time_when(tp);
654 
655 	/* It is alive without keepalive 8) */
656 	if (tp->packets_out || tcp_send_head(sk))
657 		goto resched;
658 
659 	elapsed = keepalive_time_elapsed(tp);
660 
661 	if (elapsed >= keepalive_time_when(tp)) {
662 		/* If the TCP_USER_TIMEOUT option is enabled, use that
663 		 * to determine when to timeout instead.
664 		 */
665 		if ((icsk->icsk_user_timeout != 0 &&
666 		    elapsed >= icsk->icsk_user_timeout &&
667 		    icsk->icsk_probes_out > 0) ||
668 		    (icsk->icsk_user_timeout == 0 &&
669 		    icsk->icsk_probes_out >= keepalive_probes(tp))) {
670 			tcp_send_active_reset(sk, GFP_ATOMIC);
671 			tcp_write_err(sk);
672 			goto out;
673 		}
674 		if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) {
675 			icsk->icsk_probes_out++;
676 			elapsed = keepalive_intvl_when(tp);
677 		} else {
678 			/* If keepalive was lost due to local congestion,
679 			 * try harder.
680 			 */
681 			elapsed = TCP_RESOURCE_PROBE_INTERVAL;
682 		}
683 	} else {
684 		/* It is tp->rcv_tstamp + keepalive_time_when(tp) */
685 		elapsed = keepalive_time_when(tp) - elapsed;
686 	}
687 
688 	sk_mem_reclaim(sk);
689 
690 resched:
691 	inet_csk_reset_keepalive_timer (sk, elapsed);
692 	goto out;
693 
694 death:
695 	tcp_done(sk);
696 
697 out:
698 	bh_unlock_sock(sk);
699 	sock_put(sk);
700 }
701 
702 void tcp_init_xmit_timers(struct sock *sk)
703 {
704 	inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
705 				  &tcp_keepalive_timer);
706 }
707