xref: /openbmc/linux/net/ipv4/tcp_timer.c (revision 4a075bd4)
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 static u32 tcp_clamp_rto_to_user_timeout(const struct sock *sk)
26 {
27 	struct inet_connection_sock *icsk = inet_csk(sk);
28 	u32 elapsed, start_ts;
29 	s32 remaining;
30 
31 	start_ts = tcp_sk(sk)->retrans_stamp;
32 	if (!icsk->icsk_user_timeout)
33 		return icsk->icsk_rto;
34 	elapsed = tcp_time_stamp(tcp_sk(sk)) - start_ts;
35 	remaining = icsk->icsk_user_timeout - elapsed;
36 	if (remaining <= 0)
37 		return 1; /* user timeout has passed; fire ASAP */
38 
39 	return min_t(u32, icsk->icsk_rto, msecs_to_jiffies(remaining));
40 }
41 
42 /**
43  *  tcp_write_err() - close socket and save error info
44  *  @sk:  The socket the error has appeared on.
45  *
46  *  Returns: Nothing (void)
47  */
48 
49 static void tcp_write_err(struct sock *sk)
50 {
51 	sk->sk_err = sk->sk_err_soft ? : ETIMEDOUT;
52 	sk->sk_error_report(sk);
53 
54 	tcp_write_queue_purge(sk);
55 	tcp_done(sk);
56 	__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT);
57 }
58 
59 /**
60  *  tcp_out_of_resources() - Close socket if out of resources
61  *  @sk:        pointer to current socket
62  *  @do_reset:  send a last packet with reset flag
63  *
64  *  Do not allow orphaned sockets to eat all our resources.
65  *  This is direct violation of TCP specs, but it is required
66  *  to prevent DoS attacks. It is called when a retransmission timeout
67  *  or zero probe timeout occurs on orphaned socket.
68  *
69  *  Also close if our net namespace is exiting; in that case there is no
70  *  hope of ever communicating again since all netns interfaces are already
71  *  down (or about to be down), and we need to release our dst references,
72  *  which have been moved to the netns loopback interface, so the namespace
73  *  can finish exiting.  This condition is only possible if we are a kernel
74  *  socket, as those do not hold references to the namespace.
75  *
76  *  Criteria is still not confirmed experimentally and may change.
77  *  We kill the socket, if:
78  *  1. If number of orphaned sockets exceeds an administratively configured
79  *     limit.
80  *  2. If we have strong memory pressure.
81  *  3. If our net namespace is exiting.
82  */
83 static int tcp_out_of_resources(struct sock *sk, bool do_reset)
84 {
85 	struct tcp_sock *tp = tcp_sk(sk);
86 	int shift = 0;
87 
88 	/* If peer does not open window for long time, or did not transmit
89 	 * anything for long time, penalize it. */
90 	if ((s32)(tcp_jiffies32 - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset)
91 		shift++;
92 
93 	/* If some dubious ICMP arrived, penalize even more. */
94 	if (sk->sk_err_soft)
95 		shift++;
96 
97 	if (tcp_check_oom(sk, shift)) {
98 		/* Catch exceptional cases, when connection requires reset.
99 		 *      1. Last segment was sent recently. */
100 		if ((s32)(tcp_jiffies32 - tp->lsndtime) <= TCP_TIMEWAIT_LEN ||
101 		    /*  2. Window is closed. */
102 		    (!tp->snd_wnd && !tp->packets_out))
103 			do_reset = true;
104 		if (do_reset)
105 			tcp_send_active_reset(sk, GFP_ATOMIC);
106 		tcp_done(sk);
107 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY);
108 		return 1;
109 	}
110 
111 	if (!check_net(sock_net(sk))) {
112 		/* Not possible to send reset; just close */
113 		tcp_done(sk);
114 		return 1;
115 	}
116 
117 	return 0;
118 }
119 
120 /**
121  *  tcp_orphan_retries() - Returns maximal number of retries on an orphaned socket
122  *  @sk:    Pointer to the current socket.
123  *  @alive: bool, socket alive state
124  */
125 static int tcp_orphan_retries(struct sock *sk, bool alive)
126 {
127 	int retries = sock_net(sk)->ipv4.sysctl_tcp_orphan_retries; /* May be zero. */
128 
129 	/* We know from an ICMP that something is wrong. */
130 	if (sk->sk_err_soft && !alive)
131 		retries = 0;
132 
133 	/* However, if socket sent something recently, select some safe
134 	 * number of retries. 8 corresponds to >100 seconds with minimal
135 	 * RTO of 200msec. */
136 	if (retries == 0 && alive)
137 		retries = 8;
138 	return retries;
139 }
140 
141 static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk)
142 {
143 	const struct net *net = sock_net(sk);
144 	int mss;
145 
146 	/* Black hole detection */
147 	if (!net->ipv4.sysctl_tcp_mtu_probing)
148 		return;
149 
150 	if (!icsk->icsk_mtup.enabled) {
151 		icsk->icsk_mtup.enabled = 1;
152 		icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
153 	} else {
154 		mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1;
155 		mss = min(net->ipv4.sysctl_tcp_base_mss, mss);
156 		mss = max(mss, 68 - tcp_sk(sk)->tcp_header_len);
157 		icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
158 	}
159 	tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
160 }
161 
162 static unsigned int tcp_model_timeout(struct sock *sk,
163 				      unsigned int boundary,
164 				      unsigned int rto_base)
165 {
166 	unsigned int linear_backoff_thresh, timeout;
167 
168 	linear_backoff_thresh = ilog2(TCP_RTO_MAX / rto_base);
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 	return jiffies_to_msecs(timeout);
175 }
176 /**
177  *  retransmits_timed_out() - returns true if this connection has timed out
178  *  @sk:       The current socket
179  *  @boundary: max number of retransmissions
180  *  @timeout:  A custom timeout value.
181  *             If set to 0 the default timeout is calculated and used.
182  *             Using TCP_RTO_MIN and the number of unsuccessful retransmits.
183  *
184  * The default "timeout" value this function can calculate and use
185  * is equivalent to the timeout of a TCP Connection
186  * after "boundary" unsuccessful, exponentially backed-off
187  * retransmissions with an initial RTO of TCP_RTO_MIN.
188  */
189 static bool retransmits_timed_out(struct sock *sk,
190 				  unsigned int boundary,
191 				  unsigned int timeout)
192 {
193 	unsigned int start_ts;
194 
195 	if (!inet_csk(sk)->icsk_retransmits)
196 		return false;
197 
198 	start_ts = tcp_sk(sk)->retrans_stamp;
199 	if (likely(timeout == 0))
200 		timeout = tcp_model_timeout(sk, boundary, TCP_RTO_MIN);
201 
202 	return (s32)(tcp_time_stamp(tcp_sk(sk)) - start_ts - timeout) >= 0;
203 }
204 
205 /* A write timeout has occurred. Process the after effects. */
206 static int tcp_write_timeout(struct sock *sk)
207 {
208 	struct inet_connection_sock *icsk = inet_csk(sk);
209 	struct tcp_sock *tp = tcp_sk(sk);
210 	struct net *net = sock_net(sk);
211 	bool expired, do_reset;
212 	int retry_until;
213 
214 	if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
215 		if (icsk->icsk_retransmits) {
216 			dst_negative_advice(sk);
217 		} else {
218 			sk_rethink_txhash(sk);
219 		}
220 		retry_until = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
221 		expired = icsk->icsk_retransmits >= retry_until;
222 	} else {
223 		if (retransmits_timed_out(sk, net->ipv4.sysctl_tcp_retries1, 0)) {
224 			/* Black hole detection */
225 			tcp_mtu_probing(icsk, sk);
226 
227 			dst_negative_advice(sk);
228 		} else {
229 			sk_rethink_txhash(sk);
230 		}
231 
232 		retry_until = net->ipv4.sysctl_tcp_retries2;
233 		if (sock_flag(sk, SOCK_DEAD)) {
234 			const bool alive = icsk->icsk_rto < TCP_RTO_MAX;
235 
236 			retry_until = tcp_orphan_retries(sk, alive);
237 			do_reset = alive ||
238 				!retransmits_timed_out(sk, retry_until, 0);
239 
240 			if (tcp_out_of_resources(sk, do_reset))
241 				return 1;
242 		}
243 		expired = retransmits_timed_out(sk, retry_until,
244 						icsk->icsk_user_timeout);
245 	}
246 	tcp_fastopen_active_detect_blackhole(sk, expired);
247 
248 	if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RTO_CB_FLAG))
249 		tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RTO_CB,
250 				  icsk->icsk_retransmits,
251 				  icsk->icsk_rto, (int)expired);
252 
253 	if (expired) {
254 		/* Has it gone just too far? */
255 		tcp_write_err(sk);
256 		return 1;
257 	}
258 
259 	return 0;
260 }
261 
262 /* Called with BH disabled */
263 void tcp_delack_timer_handler(struct sock *sk)
264 {
265 	struct inet_connection_sock *icsk = inet_csk(sk);
266 
267 	sk_mem_reclaim_partial(sk);
268 
269 	if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
270 	    !(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
271 		goto out;
272 
273 	if (time_after(icsk->icsk_ack.timeout, jiffies)) {
274 		sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
275 		goto out;
276 	}
277 	icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
278 
279 	if (inet_csk_ack_scheduled(sk)) {
280 		if (!inet_csk_in_pingpong_mode(sk)) {
281 			/* Delayed ACK missed: inflate ATO. */
282 			icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto);
283 		} else {
284 			/* Delayed ACK missed: leave pingpong mode and
285 			 * deflate ATO.
286 			 */
287 			inet_csk_exit_pingpong_mode(sk);
288 			icsk->icsk_ack.ato      = TCP_ATO_MIN;
289 		}
290 		tcp_mstamp_refresh(tcp_sk(sk));
291 		tcp_send_ack(sk);
292 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKS);
293 	}
294 
295 out:
296 	if (tcp_under_memory_pressure(sk))
297 		sk_mem_reclaim(sk);
298 }
299 
300 
301 /**
302  *  tcp_delack_timer() - The TCP delayed ACK timeout handler
303  *  @data:  Pointer to the current socket. (gets casted to struct sock *)
304  *
305  *  This function gets (indirectly) called when the kernel timer for a TCP packet
306  *  of this socket expires. Calls tcp_delack_timer_handler() to do the actual work.
307  *
308  *  Returns: Nothing (void)
309  */
310 static void tcp_delack_timer(struct timer_list *t)
311 {
312 	struct inet_connection_sock *icsk =
313 			from_timer(icsk, t, icsk_delack_timer);
314 	struct sock *sk = &icsk->icsk_inet.sk;
315 
316 	bh_lock_sock(sk);
317 	if (!sock_owned_by_user(sk)) {
318 		tcp_delack_timer_handler(sk);
319 	} else {
320 		icsk->icsk_ack.blocked = 1;
321 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
322 		/* deleguate our work to tcp_release_cb() */
323 		if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &sk->sk_tsq_flags))
324 			sock_hold(sk);
325 	}
326 	bh_unlock_sock(sk);
327 	sock_put(sk);
328 }
329 
330 static void tcp_probe_timer(struct sock *sk)
331 {
332 	struct inet_connection_sock *icsk = inet_csk(sk);
333 	struct sk_buff *skb = tcp_send_head(sk);
334 	struct tcp_sock *tp = tcp_sk(sk);
335 	int max_probes;
336 
337 	if (tp->packets_out || !skb) {
338 		icsk->icsk_probes_out = 0;
339 		return;
340 	}
341 
342 	/* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as
343 	 * long as the receiver continues to respond probes. We support this by
344 	 * default and reset icsk_probes_out with incoming ACKs. But if the
345 	 * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we
346 	 * kill the socket when the retry count and the time exceeds the
347 	 * corresponding system limit. We also implement similar policy when
348 	 * we use RTO to probe window in tcp_retransmit_timer().
349 	 */
350 	if (icsk->icsk_user_timeout) {
351 		u32 elapsed = tcp_model_timeout(sk, icsk->icsk_probes_out,
352 						tcp_probe0_base(sk));
353 
354 		if (elapsed >= icsk->icsk_user_timeout)
355 			goto abort;
356 	}
357 
358 	max_probes = sock_net(sk)->ipv4.sysctl_tcp_retries2;
359 	if (sock_flag(sk, SOCK_DEAD)) {
360 		const bool alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX;
361 
362 		max_probes = tcp_orphan_retries(sk, alive);
363 		if (!alive && icsk->icsk_backoff >= max_probes)
364 			goto abort;
365 		if (tcp_out_of_resources(sk, true))
366 			return;
367 	}
368 
369 	if (icsk->icsk_probes_out >= max_probes) {
370 abort:		tcp_write_err(sk);
371 	} else {
372 		/* Only send another probe if we didn't close things up. */
373 		tcp_send_probe0(sk);
374 	}
375 }
376 
377 /*
378  *	Timer for Fast Open socket to retransmit SYNACK. Note that the
379  *	sk here is the child socket, not the parent (listener) socket.
380  */
381 static void tcp_fastopen_synack_timer(struct sock *sk)
382 {
383 	struct inet_connection_sock *icsk = inet_csk(sk);
384 	int max_retries = icsk->icsk_syn_retries ? :
385 	    sock_net(sk)->ipv4.sysctl_tcp_synack_retries + 1; /* add one more retry for fastopen */
386 	struct tcp_sock *tp = tcp_sk(sk);
387 	struct request_sock *req;
388 
389 	req = tcp_sk(sk)->fastopen_rsk;
390 	req->rsk_ops->syn_ack_timeout(req);
391 
392 	if (req->num_timeout >= max_retries) {
393 		tcp_write_err(sk);
394 		return;
395 	}
396 	/* Lower cwnd after certain SYNACK timeout like tcp_init_transfer() */
397 	if (icsk->icsk_retransmits == 1)
398 		tcp_enter_loss(sk);
399 	/* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
400 	 * returned from rtx_syn_ack() to make it more persistent like
401 	 * regular retransmit because if the child socket has been accepted
402 	 * it's not good to give up too easily.
403 	 */
404 	inet_rtx_syn_ack(sk, req);
405 	req->num_timeout++;
406 	icsk->icsk_retransmits++;
407 	if (!tp->retrans_stamp)
408 		tp->retrans_stamp = tcp_time_stamp(tp);
409 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
410 			  TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
411 }
412 
413 
414 /**
415  *  tcp_retransmit_timer() - The TCP retransmit timeout handler
416  *  @sk:  Pointer to the current socket.
417  *
418  *  This function gets called when the kernel timer for a TCP packet
419  *  of this socket expires.
420  *
421  *  It handles retransmission, timer adjustment and other necesarry measures.
422  *
423  *  Returns: Nothing (void)
424  */
425 void tcp_retransmit_timer(struct sock *sk)
426 {
427 	struct tcp_sock *tp = tcp_sk(sk);
428 	struct net *net = sock_net(sk);
429 	struct inet_connection_sock *icsk = inet_csk(sk);
430 
431 	if (tp->fastopen_rsk) {
432 		WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
433 			     sk->sk_state != TCP_FIN_WAIT1);
434 		tcp_fastopen_synack_timer(sk);
435 		/* Before we receive ACK to our SYN-ACK don't retransmit
436 		 * anything else (e.g., data or FIN segments).
437 		 */
438 		return;
439 	}
440 	if (!tp->packets_out || WARN_ON_ONCE(tcp_rtx_queue_empty(sk)))
441 		return;
442 
443 	tp->tlp_high_seq = 0;
444 
445 	if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
446 	    !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
447 		/* Receiver dastardly shrinks window. Our retransmits
448 		 * become zero probes, but we should not timeout this
449 		 * connection. If the socket is an orphan, time it out,
450 		 * we cannot allow such beasts to hang infinitely.
451 		 */
452 		struct inet_sock *inet = inet_sk(sk);
453 		if (sk->sk_family == AF_INET) {
454 			net_dbg_ratelimited("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
455 					    &inet->inet_daddr,
456 					    ntohs(inet->inet_dport),
457 					    inet->inet_num,
458 					    tp->snd_una, tp->snd_nxt);
459 		}
460 #if IS_ENABLED(CONFIG_IPV6)
461 		else if (sk->sk_family == AF_INET6) {
462 			net_dbg_ratelimited("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
463 					    &sk->sk_v6_daddr,
464 					    ntohs(inet->inet_dport),
465 					    inet->inet_num,
466 					    tp->snd_una, tp->snd_nxt);
467 		}
468 #endif
469 		if (tcp_jiffies32 - tp->rcv_tstamp > TCP_RTO_MAX) {
470 			tcp_write_err(sk);
471 			goto out;
472 		}
473 		tcp_enter_loss(sk);
474 		tcp_retransmit_skb(sk, tcp_rtx_queue_head(sk), 1);
475 		__sk_dst_reset(sk);
476 		goto out_reset_timer;
477 	}
478 
479 	__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEOUTS);
480 	if (tcp_write_timeout(sk))
481 		goto out;
482 
483 	if (icsk->icsk_retransmits == 0) {
484 		int mib_idx = 0;
485 
486 		if (icsk->icsk_ca_state == TCP_CA_Recovery) {
487 			if (tcp_is_sack(tp))
488 				mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
489 			else
490 				mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
491 		} else if (icsk->icsk_ca_state == TCP_CA_Loss) {
492 			mib_idx = LINUX_MIB_TCPLOSSFAILURES;
493 		} else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
494 			   tp->sacked_out) {
495 			if (tcp_is_sack(tp))
496 				mib_idx = LINUX_MIB_TCPSACKFAILURES;
497 			else
498 				mib_idx = LINUX_MIB_TCPRENOFAILURES;
499 		}
500 		if (mib_idx)
501 			__NET_INC_STATS(sock_net(sk), mib_idx);
502 	}
503 
504 	tcp_enter_loss(sk);
505 
506 	icsk->icsk_retransmits++;
507 	if (tcp_retransmit_skb(sk, tcp_rtx_queue_head(sk), 1) > 0) {
508 		/* Retransmission failed because of local congestion,
509 		 * Let senders fight for local resources conservatively.
510 		 */
511 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
512 					  TCP_RESOURCE_PROBE_INTERVAL,
513 					  TCP_RTO_MAX);
514 		goto out;
515 	}
516 
517 	/* Increase the timeout each time we retransmit.  Note that
518 	 * we do not increase the rtt estimate.  rto is initialized
519 	 * from rtt, but increases here.  Jacobson (SIGCOMM 88) suggests
520 	 * that doubling rto each time is the least we can get away with.
521 	 * In KA9Q, Karn uses this for the first few times, and then
522 	 * goes to quadratic.  netBSD doubles, but only goes up to *64,
523 	 * and clamps at 1 to 64 sec afterwards.  Note that 120 sec is
524 	 * defined in the protocol as the maximum possible RTT.  I guess
525 	 * we'll have to use something other than TCP to talk to the
526 	 * University of Mars.
527 	 *
528 	 * PAWS allows us longer timeouts and large windows, so once
529 	 * implemented ftp to mars will work nicely. We will have to fix
530 	 * the 120 second clamps though!
531 	 */
532 	icsk->icsk_backoff++;
533 
534 out_reset_timer:
535 	/* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
536 	 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
537 	 * might be increased if the stream oscillates between thin and thick,
538 	 * thus the old value might already be too high compared to the value
539 	 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
540 	 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
541 	 * exponential backoff behaviour to avoid continue hammering
542 	 * linear-timeout retransmissions into a black hole
543 	 */
544 	if (sk->sk_state == TCP_ESTABLISHED &&
545 	    (tp->thin_lto || net->ipv4.sysctl_tcp_thin_linear_timeouts) &&
546 	    tcp_stream_is_thin(tp) &&
547 	    icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
548 		icsk->icsk_backoff = 0;
549 		icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX);
550 	} else {
551 		/* Use normal (exponential) backoff */
552 		icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
553 	}
554 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
555 				  tcp_clamp_rto_to_user_timeout(sk), TCP_RTO_MAX);
556 	if (retransmits_timed_out(sk, net->ipv4.sysctl_tcp_retries1 + 1, 0))
557 		__sk_dst_reset(sk);
558 
559 out:;
560 }
561 
562 /* Called with bottom-half processing disabled.
563    Called by tcp_write_timer() */
564 void tcp_write_timer_handler(struct sock *sk)
565 {
566 	struct inet_connection_sock *icsk = inet_csk(sk);
567 	int event;
568 
569 	if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
570 	    !icsk->icsk_pending)
571 		goto out;
572 
573 	if (time_after(icsk->icsk_timeout, jiffies)) {
574 		sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
575 		goto out;
576 	}
577 
578 	tcp_mstamp_refresh(tcp_sk(sk));
579 	event = icsk->icsk_pending;
580 
581 	switch (event) {
582 	case ICSK_TIME_REO_TIMEOUT:
583 		tcp_rack_reo_timeout(sk);
584 		break;
585 	case ICSK_TIME_LOSS_PROBE:
586 		tcp_send_loss_probe(sk);
587 		break;
588 	case ICSK_TIME_RETRANS:
589 		icsk->icsk_pending = 0;
590 		tcp_retransmit_timer(sk);
591 		break;
592 	case ICSK_TIME_PROBE0:
593 		icsk->icsk_pending = 0;
594 		tcp_probe_timer(sk);
595 		break;
596 	}
597 
598 out:
599 	sk_mem_reclaim(sk);
600 }
601 
602 static void tcp_write_timer(struct timer_list *t)
603 {
604 	struct inet_connection_sock *icsk =
605 			from_timer(icsk, t, icsk_retransmit_timer);
606 	struct sock *sk = &icsk->icsk_inet.sk;
607 
608 	bh_lock_sock(sk);
609 	if (!sock_owned_by_user(sk)) {
610 		tcp_write_timer_handler(sk);
611 	} else {
612 		/* delegate our work to tcp_release_cb() */
613 		if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &sk->sk_tsq_flags))
614 			sock_hold(sk);
615 	}
616 	bh_unlock_sock(sk);
617 	sock_put(sk);
618 }
619 
620 void tcp_syn_ack_timeout(const struct request_sock *req)
621 {
622 	struct net *net = read_pnet(&inet_rsk(req)->ireq_net);
623 
624 	__NET_INC_STATS(net, LINUX_MIB_TCPTIMEOUTS);
625 }
626 EXPORT_SYMBOL(tcp_syn_ack_timeout);
627 
628 void tcp_set_keepalive(struct sock *sk, int val)
629 {
630 	if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
631 		return;
632 
633 	if (val && !sock_flag(sk, SOCK_KEEPOPEN))
634 		inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
635 	else if (!val)
636 		inet_csk_delete_keepalive_timer(sk);
637 }
638 EXPORT_SYMBOL_GPL(tcp_set_keepalive);
639 
640 
641 static void tcp_keepalive_timer (struct timer_list *t)
642 {
643 	struct sock *sk = from_timer(sk, t, sk_timer);
644 	struct inet_connection_sock *icsk = inet_csk(sk);
645 	struct tcp_sock *tp = tcp_sk(sk);
646 	u32 elapsed;
647 
648 	/* Only process if socket is not in use. */
649 	bh_lock_sock(sk);
650 	if (sock_owned_by_user(sk)) {
651 		/* Try again later. */
652 		inet_csk_reset_keepalive_timer (sk, HZ/20);
653 		goto out;
654 	}
655 
656 	if (sk->sk_state == TCP_LISTEN) {
657 		pr_err("Hmm... keepalive on a LISTEN ???\n");
658 		goto out;
659 	}
660 
661 	tcp_mstamp_refresh(tp);
662 	if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
663 		if (tp->linger2 >= 0) {
664 			const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
665 
666 			if (tmo > 0) {
667 				tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
668 				goto out;
669 			}
670 		}
671 		tcp_send_active_reset(sk, GFP_ATOMIC);
672 		goto death;
673 	}
674 
675 	if (!sock_flag(sk, SOCK_KEEPOPEN) ||
676 	    ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_SYN_SENT)))
677 		goto out;
678 
679 	elapsed = keepalive_time_when(tp);
680 
681 	/* It is alive without keepalive 8) */
682 	if (tp->packets_out || !tcp_write_queue_empty(sk))
683 		goto resched;
684 
685 	elapsed = keepalive_time_elapsed(tp);
686 
687 	if (elapsed >= keepalive_time_when(tp)) {
688 		/* If the TCP_USER_TIMEOUT option is enabled, use that
689 		 * to determine when to timeout instead.
690 		 */
691 		if ((icsk->icsk_user_timeout != 0 &&
692 		    elapsed >= msecs_to_jiffies(icsk->icsk_user_timeout) &&
693 		    icsk->icsk_probes_out > 0) ||
694 		    (icsk->icsk_user_timeout == 0 &&
695 		    icsk->icsk_probes_out >= keepalive_probes(tp))) {
696 			tcp_send_active_reset(sk, GFP_ATOMIC);
697 			tcp_write_err(sk);
698 			goto out;
699 		}
700 		if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) {
701 			icsk->icsk_probes_out++;
702 			elapsed = keepalive_intvl_when(tp);
703 		} else {
704 			/* If keepalive was lost due to local congestion,
705 			 * try harder.
706 			 */
707 			elapsed = TCP_RESOURCE_PROBE_INTERVAL;
708 		}
709 	} else {
710 		/* It is tp->rcv_tstamp + keepalive_time_when(tp) */
711 		elapsed = keepalive_time_when(tp) - elapsed;
712 	}
713 
714 	sk_mem_reclaim(sk);
715 
716 resched:
717 	inet_csk_reset_keepalive_timer (sk, elapsed);
718 	goto out;
719 
720 death:
721 	tcp_done(sk);
722 
723 out:
724 	bh_unlock_sock(sk);
725 	sock_put(sk);
726 }
727 
728 static enum hrtimer_restart tcp_compressed_ack_kick(struct hrtimer *timer)
729 {
730 	struct tcp_sock *tp = container_of(timer, struct tcp_sock, compressed_ack_timer);
731 	struct sock *sk = (struct sock *)tp;
732 
733 	bh_lock_sock(sk);
734 	if (!sock_owned_by_user(sk)) {
735 		if (tp->compressed_ack > TCP_FASTRETRANS_THRESH)
736 			tcp_send_ack(sk);
737 	} else {
738 		if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED,
739 				      &sk->sk_tsq_flags))
740 			sock_hold(sk);
741 	}
742 	bh_unlock_sock(sk);
743 
744 	sock_put(sk);
745 
746 	return HRTIMER_NORESTART;
747 }
748 
749 void tcp_init_xmit_timers(struct sock *sk)
750 {
751 	inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
752 				  &tcp_keepalive_timer);
753 	hrtimer_init(&tcp_sk(sk)->pacing_timer, CLOCK_MONOTONIC,
754 		     HRTIMER_MODE_ABS_PINNED_SOFT);
755 	tcp_sk(sk)->pacing_timer.function = tcp_pace_kick;
756 
757 	hrtimer_init(&tcp_sk(sk)->compressed_ack_timer, CLOCK_MONOTONIC,
758 		     HRTIMER_MODE_REL_PINNED_SOFT);
759 	tcp_sk(sk)->compressed_ack_timer.function = tcp_compressed_ack_kick;
760 }
761