xref: /openbmc/linux/net/ipv4/tcp_timer.c (revision 79f08d9e)
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, int 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 = 1;
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, int 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 	/* Black hole detection */
105 	if (sysctl_tcp_mtu_probing) {
106 		if (!icsk->icsk_mtup.enabled) {
107 			icsk->icsk_mtup.enabled = 1;
108 			tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
109 		} else {
110 			struct tcp_sock *tp = tcp_sk(sk);
111 			int mss;
112 
113 			mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1;
114 			mss = min(sysctl_tcp_base_mss, mss);
115 			mss = max(mss, 68 - tp->tcp_header_len);
116 			icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
117 			tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
118 		}
119 	}
120 }
121 
122 /* This function calculates a "timeout" which is equivalent to the timeout of a
123  * TCP connection after "boundary" unsuccessful, exponentially backed-off
124  * retransmissions with an initial RTO of TCP_RTO_MIN or TCP_TIMEOUT_INIT if
125  * syn_set flag is set.
126  */
127 static bool retransmits_timed_out(struct sock *sk,
128 				  unsigned int boundary,
129 				  unsigned int timeout,
130 				  bool syn_set)
131 {
132 	unsigned int linear_backoff_thresh, start_ts;
133 	unsigned int rto_base = syn_set ? TCP_TIMEOUT_INIT : TCP_RTO_MIN;
134 
135 	if (!inet_csk(sk)->icsk_retransmits)
136 		return false;
137 
138 	if (unlikely(!tcp_sk(sk)->retrans_stamp))
139 		start_ts = TCP_SKB_CB(tcp_write_queue_head(sk))->when;
140 	else
141 		start_ts = tcp_sk(sk)->retrans_stamp;
142 
143 	if (likely(timeout == 0)) {
144 		linear_backoff_thresh = ilog2(TCP_RTO_MAX/rto_base);
145 
146 		if (boundary <= linear_backoff_thresh)
147 			timeout = ((2 << boundary) - 1) * rto_base;
148 		else
149 			timeout = ((2 << linear_backoff_thresh) - 1) * rto_base +
150 				(boundary - linear_backoff_thresh) * TCP_RTO_MAX;
151 	}
152 	return (tcp_time_stamp - start_ts) >= timeout;
153 }
154 
155 /* A write timeout has occurred. Process the after effects. */
156 static int tcp_write_timeout(struct sock *sk)
157 {
158 	struct inet_connection_sock *icsk = inet_csk(sk);
159 	struct tcp_sock *tp = tcp_sk(sk);
160 	int retry_until;
161 	bool do_reset, syn_set = false;
162 
163 	if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
164 		if (icsk->icsk_retransmits) {
165 			dst_negative_advice(sk);
166 			if (tp->syn_fastopen || tp->syn_data)
167 				tcp_fastopen_cache_set(sk, 0, NULL, true);
168 		}
169 		retry_until = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
170 		syn_set = true;
171 	} else {
172 		if (retransmits_timed_out(sk, sysctl_tcp_retries1, 0, 0)) {
173 			/* Black hole detection */
174 			tcp_mtu_probing(icsk, sk);
175 
176 			dst_negative_advice(sk);
177 		}
178 
179 		retry_until = sysctl_tcp_retries2;
180 		if (sock_flag(sk, SOCK_DEAD)) {
181 			const int alive = (icsk->icsk_rto < TCP_RTO_MAX);
182 
183 			retry_until = tcp_orphan_retries(sk, alive);
184 			do_reset = alive ||
185 				!retransmits_timed_out(sk, retry_until, 0, 0);
186 
187 			if (tcp_out_of_resources(sk, do_reset))
188 				return 1;
189 		}
190 	}
191 
192 	if (retransmits_timed_out(sk, retry_until,
193 				  syn_set ? 0 : icsk->icsk_user_timeout, syn_set)) {
194 		/* Has it gone just too far? */
195 		tcp_write_err(sk);
196 		return 1;
197 	}
198 	return 0;
199 }
200 
201 void tcp_delack_timer_handler(struct sock *sk)
202 {
203 	struct tcp_sock *tp = tcp_sk(sk);
204 	struct inet_connection_sock *icsk = inet_csk(sk);
205 
206 	sk_mem_reclaim_partial(sk);
207 
208 	if (sk->sk_state == TCP_CLOSE || !(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
209 		goto out;
210 
211 	if (time_after(icsk->icsk_ack.timeout, jiffies)) {
212 		sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
213 		goto out;
214 	}
215 	icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
216 
217 	if (!skb_queue_empty(&tp->ucopy.prequeue)) {
218 		struct sk_buff *skb;
219 
220 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSCHEDULERFAILED);
221 
222 		while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
223 			sk_backlog_rcv(sk, skb);
224 
225 		tp->ucopy.memory = 0;
226 	}
227 
228 	if (inet_csk_ack_scheduled(sk)) {
229 		if (!icsk->icsk_ack.pingpong) {
230 			/* Delayed ACK missed: inflate ATO. */
231 			icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto);
232 		} else {
233 			/* Delayed ACK missed: leave pingpong mode and
234 			 * deflate ATO.
235 			 */
236 			icsk->icsk_ack.pingpong = 0;
237 			icsk->icsk_ack.ato      = TCP_ATO_MIN;
238 		}
239 		tcp_send_ack(sk);
240 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKS);
241 	}
242 
243 out:
244 	if (sk_under_memory_pressure(sk))
245 		sk_mem_reclaim(sk);
246 }
247 
248 static void tcp_delack_timer(unsigned long data)
249 {
250 	struct sock *sk = (struct sock *)data;
251 
252 	bh_lock_sock(sk);
253 	if (!sock_owned_by_user(sk)) {
254 		tcp_delack_timer_handler(sk);
255 	} else {
256 		inet_csk(sk)->icsk_ack.blocked = 1;
257 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
258 		/* deleguate our work to tcp_release_cb() */
259 		if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &tcp_sk(sk)->tsq_flags))
260 			sock_hold(sk);
261 	}
262 	bh_unlock_sock(sk);
263 	sock_put(sk);
264 }
265 
266 static void tcp_probe_timer(struct sock *sk)
267 {
268 	struct inet_connection_sock *icsk = inet_csk(sk);
269 	struct tcp_sock *tp = tcp_sk(sk);
270 	int max_probes;
271 
272 	if (tp->packets_out || !tcp_send_head(sk)) {
273 		icsk->icsk_probes_out = 0;
274 		return;
275 	}
276 
277 	/* *WARNING* RFC 1122 forbids this
278 	 *
279 	 * It doesn't AFAIK, because we kill the retransmit timer -AK
280 	 *
281 	 * FIXME: We ought not to do it, Solaris 2.5 actually has fixing
282 	 * this behaviour in Solaris down as a bug fix. [AC]
283 	 *
284 	 * Let me to explain. icsk_probes_out is zeroed by incoming ACKs
285 	 * even if they advertise zero window. Hence, connection is killed only
286 	 * if we received no ACKs for normal connection timeout. It is not killed
287 	 * only because window stays zero for some time, window may be zero
288 	 * until armageddon and even later. We are in full accordance
289 	 * with RFCs, only probe timer combines both retransmission timeout
290 	 * and probe timeout in one bottle.				--ANK
291 	 */
292 	max_probes = sysctl_tcp_retries2;
293 
294 	if (sock_flag(sk, SOCK_DEAD)) {
295 		const int alive = ((icsk->icsk_rto << icsk->icsk_backoff) < TCP_RTO_MAX);
296 
297 		max_probes = tcp_orphan_retries(sk, alive);
298 
299 		if (tcp_out_of_resources(sk, alive || icsk->icsk_probes_out <= max_probes))
300 			return;
301 	}
302 
303 	if (icsk->icsk_probes_out > max_probes) {
304 		tcp_write_err(sk);
305 	} else {
306 		/* Only send another probe if we didn't close things up. */
307 		tcp_send_probe0(sk);
308 	}
309 }
310 
311 /*
312  *	Timer for Fast Open socket to retransmit SYNACK. Note that the
313  *	sk here is the child socket, not the parent (listener) socket.
314  */
315 static void tcp_fastopen_synack_timer(struct sock *sk)
316 {
317 	struct inet_connection_sock *icsk = inet_csk(sk);
318 	int max_retries = icsk->icsk_syn_retries ? :
319 	    sysctl_tcp_synack_retries + 1; /* add one more retry for fastopen */
320 	struct request_sock *req;
321 
322 	req = tcp_sk(sk)->fastopen_rsk;
323 	req->rsk_ops->syn_ack_timeout(sk, req);
324 
325 	if (req->num_timeout >= max_retries) {
326 		tcp_write_err(sk);
327 		return;
328 	}
329 	/* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
330 	 * returned from rtx_syn_ack() to make it more persistent like
331 	 * regular retransmit because if the child socket has been accepted
332 	 * it's not good to give up too easily.
333 	 */
334 	inet_rtx_syn_ack(sk, req);
335 	req->num_timeout++;
336 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
337 			  TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
338 }
339 
340 /*
341  *	The TCP retransmit timer.
342  */
343 
344 void tcp_retransmit_timer(struct sock *sk)
345 {
346 	struct tcp_sock *tp = tcp_sk(sk);
347 	struct inet_connection_sock *icsk = inet_csk(sk);
348 
349 	if (tp->fastopen_rsk) {
350 		WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
351 			     sk->sk_state != TCP_FIN_WAIT1);
352 		tcp_fastopen_synack_timer(sk);
353 		/* Before we receive ACK to our SYN-ACK don't retransmit
354 		 * anything else (e.g., data or FIN segments).
355 		 */
356 		return;
357 	}
358 	if (!tp->packets_out)
359 		goto out;
360 
361 	WARN_ON(tcp_write_queue_empty(sk));
362 
363 	tp->tlp_high_seq = 0;
364 
365 	if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
366 	    !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
367 		/* Receiver dastardly shrinks window. Our retransmits
368 		 * become zero probes, but we should not timeout this
369 		 * connection. If the socket is an orphan, time it out,
370 		 * we cannot allow such beasts to hang infinitely.
371 		 */
372 		struct inet_sock *inet = inet_sk(sk);
373 		if (sk->sk_family == AF_INET) {
374 			LIMIT_NETDEBUG(KERN_DEBUG pr_fmt("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n"),
375 				       &inet->inet_daddr,
376 				       ntohs(inet->inet_dport), inet->inet_num,
377 				       tp->snd_una, tp->snd_nxt);
378 		}
379 #if IS_ENABLED(CONFIG_IPV6)
380 		else if (sk->sk_family == AF_INET6) {
381 			LIMIT_NETDEBUG(KERN_DEBUG pr_fmt("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n"),
382 				       &sk->sk_v6_daddr,
383 				       ntohs(inet->inet_dport), inet->inet_num,
384 				       tp->snd_una, tp->snd_nxt);
385 		}
386 #endif
387 		if (tcp_time_stamp - tp->rcv_tstamp > TCP_RTO_MAX) {
388 			tcp_write_err(sk);
389 			goto out;
390 		}
391 		tcp_enter_loss(sk, 0);
392 		tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
393 		__sk_dst_reset(sk);
394 		goto out_reset_timer;
395 	}
396 
397 	if (tcp_write_timeout(sk))
398 		goto out;
399 
400 	if (icsk->icsk_retransmits == 0) {
401 		int mib_idx;
402 
403 		if (icsk->icsk_ca_state == TCP_CA_Recovery) {
404 			if (tcp_is_sack(tp))
405 				mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
406 			else
407 				mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
408 		} else if (icsk->icsk_ca_state == TCP_CA_Loss) {
409 			mib_idx = LINUX_MIB_TCPLOSSFAILURES;
410 		} else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
411 			   tp->sacked_out) {
412 			if (tcp_is_sack(tp))
413 				mib_idx = LINUX_MIB_TCPSACKFAILURES;
414 			else
415 				mib_idx = LINUX_MIB_TCPRENOFAILURES;
416 		} else {
417 			mib_idx = LINUX_MIB_TCPTIMEOUTS;
418 		}
419 		NET_INC_STATS_BH(sock_net(sk), mib_idx);
420 	}
421 
422 	tcp_enter_loss(sk, 0);
423 
424 	if (tcp_retransmit_skb(sk, tcp_write_queue_head(sk)) > 0) {
425 		/* Retransmission failed because of local congestion,
426 		 * do not backoff.
427 		 */
428 		if (!icsk->icsk_retransmits)
429 			icsk->icsk_retransmits = 1;
430 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
431 					  min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL),
432 					  TCP_RTO_MAX);
433 		goto out;
434 	}
435 
436 	/* Increase the timeout each time we retransmit.  Note that
437 	 * we do not increase the rtt estimate.  rto is initialized
438 	 * from rtt, but increases here.  Jacobson (SIGCOMM 88) suggests
439 	 * that doubling rto each time is the least we can get away with.
440 	 * In KA9Q, Karn uses this for the first few times, and then
441 	 * goes to quadratic.  netBSD doubles, but only goes up to *64,
442 	 * and clamps at 1 to 64 sec afterwards.  Note that 120 sec is
443 	 * defined in the protocol as the maximum possible RTT.  I guess
444 	 * we'll have to use something other than TCP to talk to the
445 	 * University of Mars.
446 	 *
447 	 * PAWS allows us longer timeouts and large windows, so once
448 	 * implemented ftp to mars will work nicely. We will have to fix
449 	 * the 120 second clamps though!
450 	 */
451 	icsk->icsk_backoff++;
452 	icsk->icsk_retransmits++;
453 
454 out_reset_timer:
455 	/* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
456 	 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
457 	 * might be increased if the stream oscillates between thin and thick,
458 	 * thus the old value might already be too high compared to the value
459 	 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
460 	 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
461 	 * exponential backoff behaviour to avoid continue hammering
462 	 * linear-timeout retransmissions into a black hole
463 	 */
464 	if (sk->sk_state == TCP_ESTABLISHED &&
465 	    (tp->thin_lto || sysctl_tcp_thin_linear_timeouts) &&
466 	    tcp_stream_is_thin(tp) &&
467 	    icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
468 		icsk->icsk_backoff = 0;
469 		icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX);
470 	} else {
471 		/* Use normal (exponential) backoff */
472 		icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
473 	}
474 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX);
475 	if (retransmits_timed_out(sk, sysctl_tcp_retries1 + 1, 0, 0))
476 		__sk_dst_reset(sk);
477 
478 out:;
479 }
480 
481 void tcp_write_timer_handler(struct sock *sk)
482 {
483 	struct inet_connection_sock *icsk = inet_csk(sk);
484 	int event;
485 
486 	if (sk->sk_state == TCP_CLOSE || !icsk->icsk_pending)
487 		goto out;
488 
489 	if (time_after(icsk->icsk_timeout, jiffies)) {
490 		sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
491 		goto out;
492 	}
493 
494 	event = icsk->icsk_pending;
495 
496 	switch (event) {
497 	case ICSK_TIME_EARLY_RETRANS:
498 		tcp_resume_early_retransmit(sk);
499 		break;
500 	case ICSK_TIME_LOSS_PROBE:
501 		tcp_send_loss_probe(sk);
502 		break;
503 	case ICSK_TIME_RETRANS:
504 		icsk->icsk_pending = 0;
505 		tcp_retransmit_timer(sk);
506 		break;
507 	case ICSK_TIME_PROBE0:
508 		icsk->icsk_pending = 0;
509 		tcp_probe_timer(sk);
510 		break;
511 	}
512 
513 out:
514 	sk_mem_reclaim(sk);
515 }
516 
517 static void tcp_write_timer(unsigned long data)
518 {
519 	struct sock *sk = (struct sock *)data;
520 
521 	bh_lock_sock(sk);
522 	if (!sock_owned_by_user(sk)) {
523 		tcp_write_timer_handler(sk);
524 	} else {
525 		/* deleguate our work to tcp_release_cb() */
526 		if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &tcp_sk(sk)->tsq_flags))
527 			sock_hold(sk);
528 	}
529 	bh_unlock_sock(sk);
530 	sock_put(sk);
531 }
532 
533 /*
534  *	Timer for listening sockets
535  */
536 
537 static void tcp_synack_timer(struct sock *sk)
538 {
539 	inet_csk_reqsk_queue_prune(sk, TCP_SYNQ_INTERVAL,
540 				   TCP_TIMEOUT_INIT, TCP_RTO_MAX);
541 }
542 
543 void tcp_syn_ack_timeout(struct sock *sk, struct request_sock *req)
544 {
545 	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPTIMEOUTS);
546 }
547 EXPORT_SYMBOL(tcp_syn_ack_timeout);
548 
549 void tcp_set_keepalive(struct sock *sk, int val)
550 {
551 	if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
552 		return;
553 
554 	if (val && !sock_flag(sk, SOCK_KEEPOPEN))
555 		inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
556 	else if (!val)
557 		inet_csk_delete_keepalive_timer(sk);
558 }
559 
560 
561 static void tcp_keepalive_timer (unsigned long data)
562 {
563 	struct sock *sk = (struct sock *) data;
564 	struct inet_connection_sock *icsk = inet_csk(sk);
565 	struct tcp_sock *tp = tcp_sk(sk);
566 	u32 elapsed;
567 
568 	/* Only process if socket is not in use. */
569 	bh_lock_sock(sk);
570 	if (sock_owned_by_user(sk)) {
571 		/* Try again later. */
572 		inet_csk_reset_keepalive_timer (sk, HZ/20);
573 		goto out;
574 	}
575 
576 	if (sk->sk_state == TCP_LISTEN) {
577 		tcp_synack_timer(sk);
578 		goto out;
579 	}
580 
581 	if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
582 		if (tp->linger2 >= 0) {
583 			const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
584 
585 			if (tmo > 0) {
586 				tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
587 				goto out;
588 			}
589 		}
590 		tcp_send_active_reset(sk, GFP_ATOMIC);
591 		goto death;
592 	}
593 
594 	if (!sock_flag(sk, SOCK_KEEPOPEN) || sk->sk_state == TCP_CLOSE)
595 		goto out;
596 
597 	elapsed = keepalive_time_when(tp);
598 
599 	/* It is alive without keepalive 8) */
600 	if (tp->packets_out || tcp_send_head(sk))
601 		goto resched;
602 
603 	elapsed = keepalive_time_elapsed(tp);
604 
605 	if (elapsed >= keepalive_time_when(tp)) {
606 		/* If the TCP_USER_TIMEOUT option is enabled, use that
607 		 * to determine when to timeout instead.
608 		 */
609 		if ((icsk->icsk_user_timeout != 0 &&
610 		    elapsed >= icsk->icsk_user_timeout &&
611 		    icsk->icsk_probes_out > 0) ||
612 		    (icsk->icsk_user_timeout == 0 &&
613 		    icsk->icsk_probes_out >= keepalive_probes(tp))) {
614 			tcp_send_active_reset(sk, GFP_ATOMIC);
615 			tcp_write_err(sk);
616 			goto out;
617 		}
618 		if (tcp_write_wakeup(sk) <= 0) {
619 			icsk->icsk_probes_out++;
620 			elapsed = keepalive_intvl_when(tp);
621 		} else {
622 			/* If keepalive was lost due to local congestion,
623 			 * try harder.
624 			 */
625 			elapsed = TCP_RESOURCE_PROBE_INTERVAL;
626 		}
627 	} else {
628 		/* It is tp->rcv_tstamp + keepalive_time_when(tp) */
629 		elapsed = keepalive_time_when(tp) - elapsed;
630 	}
631 
632 	sk_mem_reclaim(sk);
633 
634 resched:
635 	inet_csk_reset_keepalive_timer (sk, elapsed);
636 	goto out;
637 
638 death:
639 	tcp_done(sk);
640 
641 out:
642 	bh_unlock_sock(sk);
643 	sock_put(sk);
644 }
645 
646 void tcp_init_xmit_timers(struct sock *sk)
647 {
648 	inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
649 				  &tcp_keepalive_timer);
650 }
651 EXPORT_SYMBOL(tcp_init_xmit_timers);
652