xref: /openbmc/linux/net/ipv4/tcp_timer.c (revision 4f205687)
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 static void tcp_write_err(struct sock *sk)
28 {
29 	sk->sk_err = sk->sk_err_soft ? : ETIMEDOUT;
30 	sk->sk_error_report(sk);
31 
32 	tcp_done(sk);
33 	__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT);
34 }
35 
36 /* Do not allow orphaned sockets to eat all our resources.
37  * This is direct violation of TCP specs, but it is required
38  * to prevent DoS attacks. It is called when a retransmission timeout
39  * or zero probe timeout occurs on orphaned socket.
40  *
41  * Criteria is still not confirmed experimentally and may change.
42  * We kill the socket, if:
43  * 1. If number of orphaned sockets exceeds an administratively configured
44  *    limit.
45  * 2. If we have strong memory pressure.
46  */
47 static int tcp_out_of_resources(struct sock *sk, bool do_reset)
48 {
49 	struct tcp_sock *tp = tcp_sk(sk);
50 	int shift = 0;
51 
52 	/* If peer does not open window for long time, or did not transmit
53 	 * anything for long time, penalize it. */
54 	if ((s32)(tcp_time_stamp - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset)
55 		shift++;
56 
57 	/* If some dubious ICMP arrived, penalize even more. */
58 	if (sk->sk_err_soft)
59 		shift++;
60 
61 	if (tcp_check_oom(sk, shift)) {
62 		/* Catch exceptional cases, when connection requires reset.
63 		 *      1. Last segment was sent recently. */
64 		if ((s32)(tcp_time_stamp - tp->lsndtime) <= TCP_TIMEWAIT_LEN ||
65 		    /*  2. Window is closed. */
66 		    (!tp->snd_wnd && !tp->packets_out))
67 			do_reset = true;
68 		if (do_reset)
69 			tcp_send_active_reset(sk, GFP_ATOMIC);
70 		tcp_done(sk);
71 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY);
72 		return 1;
73 	}
74 	return 0;
75 }
76 
77 /* Calculate maximal number or retries on an orphaned socket. */
78 static int tcp_orphan_retries(struct sock *sk, bool alive)
79 {
80 	int retries = sock_net(sk)->ipv4.sysctl_tcp_orphan_retries; /* May be zero. */
81 
82 	/* We know from an ICMP that something is wrong. */
83 	if (sk->sk_err_soft && !alive)
84 		retries = 0;
85 
86 	/* However, if socket sent something recently, select some safe
87 	 * number of retries. 8 corresponds to >100 seconds with minimal
88 	 * RTO of 200msec. */
89 	if (retries == 0 && alive)
90 		retries = 8;
91 	return retries;
92 }
93 
94 static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk)
95 {
96 	struct net *net = sock_net(sk);
97 
98 	/* Black hole detection */
99 	if (net->ipv4.sysctl_tcp_mtu_probing) {
100 		if (!icsk->icsk_mtup.enabled) {
101 			icsk->icsk_mtup.enabled = 1;
102 			icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
103 			tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
104 		} else {
105 			struct net *net = sock_net(sk);
106 			struct tcp_sock *tp = tcp_sk(sk);
107 			int mss;
108 
109 			mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1;
110 			mss = min(net->ipv4.sysctl_tcp_base_mss, mss);
111 			mss = max(mss, 68 - tp->tcp_header_len);
112 			icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
113 			tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
114 		}
115 	}
116 }
117 
118 /* This function calculates a "timeout" which is equivalent to the timeout of a
119  * TCP connection after "boundary" unsuccessful, exponentially backed-off
120  * retransmissions with an initial RTO of TCP_RTO_MIN or TCP_TIMEOUT_INIT if
121  * syn_set flag is set.
122  */
123 static bool retransmits_timed_out(struct sock *sk,
124 				  unsigned int boundary,
125 				  unsigned int timeout,
126 				  bool syn_set)
127 {
128 	unsigned int linear_backoff_thresh, start_ts;
129 	unsigned int rto_base = syn_set ? TCP_TIMEOUT_INIT : TCP_RTO_MIN;
130 
131 	if (!inet_csk(sk)->icsk_retransmits)
132 		return false;
133 
134 	start_ts = tcp_sk(sk)->retrans_stamp;
135 	if (unlikely(!start_ts))
136 		start_ts = tcp_skb_timestamp(tcp_write_queue_head(sk));
137 
138 	if (likely(timeout == 0)) {
139 		linear_backoff_thresh = ilog2(TCP_RTO_MAX/rto_base);
140 
141 		if (boundary <= linear_backoff_thresh)
142 			timeout = ((2 << boundary) - 1) * rto_base;
143 		else
144 			timeout = ((2 << linear_backoff_thresh) - 1) * rto_base +
145 				(boundary - linear_backoff_thresh) * TCP_RTO_MAX;
146 	}
147 	return (tcp_time_stamp - start_ts) >= timeout;
148 }
149 
150 /* A write timeout has occurred. Process the after effects. */
151 static int tcp_write_timeout(struct sock *sk)
152 {
153 	struct inet_connection_sock *icsk = inet_csk(sk);
154 	struct tcp_sock *tp = tcp_sk(sk);
155 	struct net *net = sock_net(sk);
156 	int retry_until;
157 	bool do_reset, syn_set = false;
158 
159 	if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
160 		if (icsk->icsk_retransmits) {
161 			dst_negative_advice(sk);
162 			if (tp->syn_fastopen || tp->syn_data)
163 				tcp_fastopen_cache_set(sk, 0, NULL, true, 0);
164 			if (tp->syn_data && icsk->icsk_retransmits == 1)
165 				NET_INC_STATS(sock_net(sk),
166 					      LINUX_MIB_TCPFASTOPENACTIVEFAIL);
167 		}
168 		retry_until = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
169 		syn_set = true;
170 	} else {
171 		if (retransmits_timed_out(sk, net->ipv4.sysctl_tcp_retries1, 0, 0)) {
172 			/* Some middle-boxes may black-hole Fast Open _after_
173 			 * the handshake. Therefore we conservatively disable
174 			 * Fast Open on this path on recurring timeouts with
175 			 * few or zero bytes acked after Fast Open.
176 			 */
177 			if (tp->syn_data_acked &&
178 			    tp->bytes_acked <= tp->rx_opt.mss_clamp) {
179 				tcp_fastopen_cache_set(sk, 0, NULL, true, 0);
180 				if (icsk->icsk_retransmits == net->ipv4.sysctl_tcp_retries1)
181 					NET_INC_STATS(sock_net(sk),
182 						      LINUX_MIB_TCPFASTOPENACTIVEFAIL);
183 			}
184 			/* Black hole detection */
185 			tcp_mtu_probing(icsk, sk);
186 
187 			dst_negative_advice(sk);
188 		}
189 
190 		retry_until = net->ipv4.sysctl_tcp_retries2;
191 		if (sock_flag(sk, SOCK_DEAD)) {
192 			const bool alive = icsk->icsk_rto < TCP_RTO_MAX;
193 
194 			retry_until = tcp_orphan_retries(sk, alive);
195 			do_reset = alive ||
196 				!retransmits_timed_out(sk, retry_until, 0, 0);
197 
198 			if (tcp_out_of_resources(sk, do_reset))
199 				return 1;
200 		}
201 	}
202 
203 	if (retransmits_timed_out(sk, retry_until,
204 				  syn_set ? 0 : icsk->icsk_user_timeout, syn_set)) {
205 		/* Has it gone just too far? */
206 		tcp_write_err(sk);
207 		return 1;
208 	}
209 	return 0;
210 }
211 
212 /* Called with BH disabled */
213 void tcp_delack_timer_handler(struct sock *sk)
214 {
215 	struct tcp_sock *tp = tcp_sk(sk);
216 	struct inet_connection_sock *icsk = inet_csk(sk);
217 
218 	sk_mem_reclaim_partial(sk);
219 
220 	if (sk->sk_state == TCP_CLOSE || !(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
221 		goto out;
222 
223 	if (time_after(icsk->icsk_ack.timeout, jiffies)) {
224 		sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
225 		goto out;
226 	}
227 	icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
228 
229 	if (!skb_queue_empty(&tp->ucopy.prequeue)) {
230 		struct sk_buff *skb;
231 
232 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSCHEDULERFAILED);
233 
234 		while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
235 			sk_backlog_rcv(sk, skb);
236 
237 		tp->ucopy.memory = 0;
238 	}
239 
240 	if (inet_csk_ack_scheduled(sk)) {
241 		if (!icsk->icsk_ack.pingpong) {
242 			/* Delayed ACK missed: inflate ATO. */
243 			icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto);
244 		} else {
245 			/* Delayed ACK missed: leave pingpong mode and
246 			 * deflate ATO.
247 			 */
248 			icsk->icsk_ack.pingpong = 0;
249 			icsk->icsk_ack.ato      = TCP_ATO_MIN;
250 		}
251 		tcp_send_ack(sk);
252 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKS);
253 	}
254 
255 out:
256 	if (tcp_under_memory_pressure(sk))
257 		sk_mem_reclaim(sk);
258 }
259 
260 static void tcp_delack_timer(unsigned long data)
261 {
262 	struct sock *sk = (struct sock *)data;
263 
264 	bh_lock_sock(sk);
265 	if (!sock_owned_by_user(sk)) {
266 		tcp_delack_timer_handler(sk);
267 	} else {
268 		inet_csk(sk)->icsk_ack.blocked = 1;
269 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
270 		/* deleguate our work to tcp_release_cb() */
271 		if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &tcp_sk(sk)->tsq_flags))
272 			sock_hold(sk);
273 	}
274 	bh_unlock_sock(sk);
275 	sock_put(sk);
276 }
277 
278 static void tcp_probe_timer(struct sock *sk)
279 {
280 	struct inet_connection_sock *icsk = inet_csk(sk);
281 	struct tcp_sock *tp = tcp_sk(sk);
282 	int max_probes;
283 	u32 start_ts;
284 
285 	if (tp->packets_out || !tcp_send_head(sk)) {
286 		icsk->icsk_probes_out = 0;
287 		return;
288 	}
289 
290 	/* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as
291 	 * long as the receiver continues to respond probes. We support this by
292 	 * default and reset icsk_probes_out with incoming ACKs. But if the
293 	 * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we
294 	 * kill the socket when the retry count and the time exceeds the
295 	 * corresponding system limit. We also implement similar policy when
296 	 * we use RTO to probe window in tcp_retransmit_timer().
297 	 */
298 	start_ts = tcp_skb_timestamp(tcp_send_head(sk));
299 	if (!start_ts)
300 		skb_mstamp_get(&tcp_send_head(sk)->skb_mstamp);
301 	else if (icsk->icsk_user_timeout &&
302 		 (s32)(tcp_time_stamp - start_ts) > icsk->icsk_user_timeout)
303 		goto abort;
304 
305 	max_probes = sock_net(sk)->ipv4.sysctl_tcp_retries2;
306 	if (sock_flag(sk, SOCK_DEAD)) {
307 		const bool alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX;
308 
309 		max_probes = tcp_orphan_retries(sk, alive);
310 		if (!alive && icsk->icsk_backoff >= max_probes)
311 			goto abort;
312 		if (tcp_out_of_resources(sk, true))
313 			return;
314 	}
315 
316 	if (icsk->icsk_probes_out > max_probes) {
317 abort:		tcp_write_err(sk);
318 	} else {
319 		/* Only send another probe if we didn't close things up. */
320 		tcp_send_probe0(sk);
321 	}
322 }
323 
324 /*
325  *	Timer for Fast Open socket to retransmit SYNACK. Note that the
326  *	sk here is the child socket, not the parent (listener) socket.
327  */
328 static void tcp_fastopen_synack_timer(struct sock *sk)
329 {
330 	struct inet_connection_sock *icsk = inet_csk(sk);
331 	int max_retries = icsk->icsk_syn_retries ? :
332 	    sock_net(sk)->ipv4.sysctl_tcp_synack_retries + 1; /* add one more retry for fastopen */
333 	struct request_sock *req;
334 
335 	req = tcp_sk(sk)->fastopen_rsk;
336 	req->rsk_ops->syn_ack_timeout(req);
337 
338 	if (req->num_timeout >= max_retries) {
339 		tcp_write_err(sk);
340 		return;
341 	}
342 	/* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
343 	 * returned from rtx_syn_ack() to make it more persistent like
344 	 * regular retransmit because if the child socket has been accepted
345 	 * it's not good to give up too easily.
346 	 */
347 	inet_rtx_syn_ack(sk, req);
348 	req->num_timeout++;
349 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
350 			  TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
351 }
352 
353 /*
354  *	The TCP retransmit timer.
355  */
356 
357 void tcp_retransmit_timer(struct sock *sk)
358 {
359 	struct tcp_sock *tp = tcp_sk(sk);
360 	struct net *net = sock_net(sk);
361 	struct inet_connection_sock *icsk = inet_csk(sk);
362 
363 	if (tp->fastopen_rsk) {
364 		WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
365 			     sk->sk_state != TCP_FIN_WAIT1);
366 		tcp_fastopen_synack_timer(sk);
367 		/* Before we receive ACK to our SYN-ACK don't retransmit
368 		 * anything else (e.g., data or FIN segments).
369 		 */
370 		return;
371 	}
372 	if (!tp->packets_out)
373 		goto out;
374 
375 	WARN_ON(tcp_write_queue_empty(sk));
376 
377 	tp->tlp_high_seq = 0;
378 
379 	if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
380 	    !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
381 		/* Receiver dastardly shrinks window. Our retransmits
382 		 * become zero probes, but we should not timeout this
383 		 * connection. If the socket is an orphan, time it out,
384 		 * we cannot allow such beasts to hang infinitely.
385 		 */
386 		struct inet_sock *inet = inet_sk(sk);
387 		if (sk->sk_family == AF_INET) {
388 			net_dbg_ratelimited("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
389 					    &inet->inet_daddr,
390 					    ntohs(inet->inet_dport),
391 					    inet->inet_num,
392 					    tp->snd_una, tp->snd_nxt);
393 		}
394 #if IS_ENABLED(CONFIG_IPV6)
395 		else if (sk->sk_family == AF_INET6) {
396 			net_dbg_ratelimited("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
397 					    &sk->sk_v6_daddr,
398 					    ntohs(inet->inet_dport),
399 					    inet->inet_num,
400 					    tp->snd_una, tp->snd_nxt);
401 		}
402 #endif
403 		if (tcp_time_stamp - tp->rcv_tstamp > TCP_RTO_MAX) {
404 			tcp_write_err(sk);
405 			goto out;
406 		}
407 		tcp_enter_loss(sk);
408 		tcp_retransmit_skb(sk, tcp_write_queue_head(sk), 1);
409 		__sk_dst_reset(sk);
410 		goto out_reset_timer;
411 	}
412 
413 	if (tcp_write_timeout(sk))
414 		goto out;
415 
416 	if (icsk->icsk_retransmits == 0) {
417 		int mib_idx;
418 
419 		if (icsk->icsk_ca_state == TCP_CA_Recovery) {
420 			if (tcp_is_sack(tp))
421 				mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
422 			else
423 				mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
424 		} else if (icsk->icsk_ca_state == TCP_CA_Loss) {
425 			mib_idx = LINUX_MIB_TCPLOSSFAILURES;
426 		} else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
427 			   tp->sacked_out) {
428 			if (tcp_is_sack(tp))
429 				mib_idx = LINUX_MIB_TCPSACKFAILURES;
430 			else
431 				mib_idx = LINUX_MIB_TCPRENOFAILURES;
432 		} else {
433 			mib_idx = LINUX_MIB_TCPTIMEOUTS;
434 		}
435 		__NET_INC_STATS(sock_net(sk), mib_idx);
436 	}
437 
438 	tcp_enter_loss(sk);
439 
440 	if (tcp_retransmit_skb(sk, tcp_write_queue_head(sk), 1) > 0) {
441 		/* Retransmission failed because of local congestion,
442 		 * do not backoff.
443 		 */
444 		if (!icsk->icsk_retransmits)
445 			icsk->icsk_retransmits = 1;
446 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
447 					  min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL),
448 					  TCP_RTO_MAX);
449 		goto out;
450 	}
451 
452 	/* Increase the timeout each time we retransmit.  Note that
453 	 * we do not increase the rtt estimate.  rto is initialized
454 	 * from rtt, but increases here.  Jacobson (SIGCOMM 88) suggests
455 	 * that doubling rto each time is the least we can get away with.
456 	 * In KA9Q, Karn uses this for the first few times, and then
457 	 * goes to quadratic.  netBSD doubles, but only goes up to *64,
458 	 * and clamps at 1 to 64 sec afterwards.  Note that 120 sec is
459 	 * defined in the protocol as the maximum possible RTT.  I guess
460 	 * we'll have to use something other than TCP to talk to the
461 	 * University of Mars.
462 	 *
463 	 * PAWS allows us longer timeouts and large windows, so once
464 	 * implemented ftp to mars will work nicely. We will have to fix
465 	 * the 120 second clamps though!
466 	 */
467 	icsk->icsk_backoff++;
468 	icsk->icsk_retransmits++;
469 
470 out_reset_timer:
471 	/* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
472 	 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
473 	 * might be increased if the stream oscillates between thin and thick,
474 	 * thus the old value might already be too high compared to the value
475 	 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
476 	 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
477 	 * exponential backoff behaviour to avoid continue hammering
478 	 * linear-timeout retransmissions into a black hole
479 	 */
480 	if (sk->sk_state == TCP_ESTABLISHED &&
481 	    (tp->thin_lto || sysctl_tcp_thin_linear_timeouts) &&
482 	    tcp_stream_is_thin(tp) &&
483 	    icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
484 		icsk->icsk_backoff = 0;
485 		icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX);
486 	} else {
487 		/* Use normal (exponential) backoff */
488 		icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
489 	}
490 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX);
491 	if (retransmits_timed_out(sk, net->ipv4.sysctl_tcp_retries1 + 1, 0, 0))
492 		__sk_dst_reset(sk);
493 
494 out:;
495 }
496 
497 /* Called with BH disabled */
498 void tcp_write_timer_handler(struct sock *sk)
499 {
500 	struct inet_connection_sock *icsk = inet_csk(sk);
501 	int event;
502 
503 	if (sk->sk_state == TCP_CLOSE || !icsk->icsk_pending)
504 		goto out;
505 
506 	if (time_after(icsk->icsk_timeout, jiffies)) {
507 		sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
508 		goto out;
509 	}
510 
511 	event = icsk->icsk_pending;
512 
513 	switch (event) {
514 	case ICSK_TIME_EARLY_RETRANS:
515 		tcp_resume_early_retransmit(sk);
516 		break;
517 	case ICSK_TIME_LOSS_PROBE:
518 		tcp_send_loss_probe(sk);
519 		break;
520 	case ICSK_TIME_RETRANS:
521 		icsk->icsk_pending = 0;
522 		tcp_retransmit_timer(sk);
523 		break;
524 	case ICSK_TIME_PROBE0:
525 		icsk->icsk_pending = 0;
526 		tcp_probe_timer(sk);
527 		break;
528 	}
529 
530 out:
531 	sk_mem_reclaim(sk);
532 }
533 
534 static void tcp_write_timer(unsigned long data)
535 {
536 	struct sock *sk = (struct sock *)data;
537 
538 	bh_lock_sock(sk);
539 	if (!sock_owned_by_user(sk)) {
540 		tcp_write_timer_handler(sk);
541 	} else {
542 		/* deleguate our work to tcp_release_cb() */
543 		if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &tcp_sk(sk)->tsq_flags))
544 			sock_hold(sk);
545 	}
546 	bh_unlock_sock(sk);
547 	sock_put(sk);
548 }
549 
550 void tcp_syn_ack_timeout(const struct request_sock *req)
551 {
552 	struct net *net = read_pnet(&inet_rsk(req)->ireq_net);
553 
554 	__NET_INC_STATS(net, LINUX_MIB_TCPTIMEOUTS);
555 }
556 EXPORT_SYMBOL(tcp_syn_ack_timeout);
557 
558 void tcp_set_keepalive(struct sock *sk, int val)
559 {
560 	if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
561 		return;
562 
563 	if (val && !sock_flag(sk, SOCK_KEEPOPEN))
564 		inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
565 	else if (!val)
566 		inet_csk_delete_keepalive_timer(sk);
567 }
568 
569 
570 static void tcp_keepalive_timer (unsigned long data)
571 {
572 	struct sock *sk = (struct sock *) data;
573 	struct inet_connection_sock *icsk = inet_csk(sk);
574 	struct tcp_sock *tp = tcp_sk(sk);
575 	u32 elapsed;
576 
577 	/* Only process if socket is not in use. */
578 	bh_lock_sock(sk);
579 	if (sock_owned_by_user(sk)) {
580 		/* Try again later. */
581 		inet_csk_reset_keepalive_timer (sk, HZ/20);
582 		goto out;
583 	}
584 
585 	if (sk->sk_state == TCP_LISTEN) {
586 		pr_err("Hmm... keepalive on a LISTEN ???\n");
587 		goto out;
588 	}
589 
590 	if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
591 		if (tp->linger2 >= 0) {
592 			const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
593 
594 			if (tmo > 0) {
595 				tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
596 				goto out;
597 			}
598 		}
599 		tcp_send_active_reset(sk, GFP_ATOMIC);
600 		goto death;
601 	}
602 
603 	if (!sock_flag(sk, SOCK_KEEPOPEN) || sk->sk_state == TCP_CLOSE)
604 		goto out;
605 
606 	elapsed = keepalive_time_when(tp);
607 
608 	/* It is alive without keepalive 8) */
609 	if (tp->packets_out || tcp_send_head(sk))
610 		goto resched;
611 
612 	elapsed = keepalive_time_elapsed(tp);
613 
614 	if (elapsed >= keepalive_time_when(tp)) {
615 		/* If the TCP_USER_TIMEOUT option is enabled, use that
616 		 * to determine when to timeout instead.
617 		 */
618 		if ((icsk->icsk_user_timeout != 0 &&
619 		    elapsed >= icsk->icsk_user_timeout &&
620 		    icsk->icsk_probes_out > 0) ||
621 		    (icsk->icsk_user_timeout == 0 &&
622 		    icsk->icsk_probes_out >= keepalive_probes(tp))) {
623 			tcp_send_active_reset(sk, GFP_ATOMIC);
624 			tcp_write_err(sk);
625 			goto out;
626 		}
627 		if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) {
628 			icsk->icsk_probes_out++;
629 			elapsed = keepalive_intvl_when(tp);
630 		} else {
631 			/* If keepalive was lost due to local congestion,
632 			 * try harder.
633 			 */
634 			elapsed = TCP_RESOURCE_PROBE_INTERVAL;
635 		}
636 	} else {
637 		/* It is tp->rcv_tstamp + keepalive_time_when(tp) */
638 		elapsed = keepalive_time_when(tp) - elapsed;
639 	}
640 
641 	sk_mem_reclaim(sk);
642 
643 resched:
644 	inet_csk_reset_keepalive_timer (sk, elapsed);
645 	goto out;
646 
647 death:
648 	tcp_done(sk);
649 
650 out:
651 	bh_unlock_sock(sk);
652 	sock_put(sk);
653 }
654 
655 void tcp_init_xmit_timers(struct sock *sk)
656 {
657 	inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
658 				  &tcp_keepalive_timer);
659 }
660