xref: /linux/net/ipv4/tcp_timer.c (revision 04317b129e4eb5c6f4a58bb899b2019c1545320b)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * INET		An implementation of the TCP/IP protocol suite for the LINUX
4  *		operating system.  INET is implemented using the  BSD Socket
5  *		interface as the means of communication with the user level.
6  *
7  *		Implementation of the Transmission Control Protocol(TCP).
8  *
9  * Authors:	Ross Biro
10  *		Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11  *		Mark Evans, <evansmp@uhura.aston.ac.uk>
12  *		Corey Minyard <wf-rch!minyard@relay.EU.net>
13  *		Florian La Roche, <flla@stud.uni-sb.de>
14  *		Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15  *		Linus Torvalds, <torvalds@cs.helsinki.fi>
16  *		Alan Cox, <gw4pts@gw4pts.ampr.org>
17  *		Matthew Dillon, <dillon@apollo.west.oic.com>
18  *		Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19  *		Jorge Cwik, <jorge@laser.satlink.net>
20  */
21 
22 #include <linux/module.h>
23 #include <linux/gfp.h>
24 #include <net/tcp.h>
25 
26 static u32 tcp_clamp_rto_to_user_timeout(const struct sock *sk)
27 {
28 	struct inet_connection_sock *icsk = inet_csk(sk);
29 	u32 elapsed, start_ts, user_timeout;
30 	s32 remaining;
31 
32 	start_ts = tcp_sk(sk)->retrans_stamp;
33 	user_timeout = READ_ONCE(icsk->icsk_user_timeout);
34 	if (!user_timeout)
35 		return icsk->icsk_rto;
36 	elapsed = tcp_time_stamp(tcp_sk(sk)) - start_ts;
37 	remaining = user_timeout - elapsed;
38 	if (remaining <= 0)
39 		return 1; /* user timeout has passed; fire ASAP */
40 
41 	return min_t(u32, icsk->icsk_rto, msecs_to_jiffies(remaining));
42 }
43 
44 u32 tcp_clamp_probe0_to_user_timeout(const struct sock *sk, u32 when)
45 {
46 	struct inet_connection_sock *icsk = inet_csk(sk);
47 	u32 remaining, user_timeout;
48 	s32 elapsed;
49 
50 	user_timeout = READ_ONCE(icsk->icsk_user_timeout);
51 	if (!user_timeout || !icsk->icsk_probes_tstamp)
52 		return when;
53 
54 	elapsed = tcp_jiffies32 - icsk->icsk_probes_tstamp;
55 	if (unlikely(elapsed < 0))
56 		elapsed = 0;
57 	remaining = msecs_to_jiffies(user_timeout) - elapsed;
58 	remaining = max_t(u32, remaining, TCP_TIMEOUT_MIN);
59 
60 	return min_t(u32, remaining, when);
61 }
62 
63 /**
64  *  tcp_write_err() - close socket and save error info
65  *  @sk:  The socket the error has appeared on.
66  *
67  *  Returns: Nothing (void)
68  */
69 
70 static void tcp_write_err(struct sock *sk)
71 {
72 	WRITE_ONCE(sk->sk_err, READ_ONCE(sk->sk_err_soft) ? : ETIMEDOUT);
73 	sk_error_report(sk);
74 
75 	tcp_write_queue_purge(sk);
76 	tcp_done(sk);
77 	__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT);
78 }
79 
80 /**
81  *  tcp_out_of_resources() - Close socket if out of resources
82  *  @sk:        pointer to current socket
83  *  @do_reset:  send a last packet with reset flag
84  *
85  *  Do not allow orphaned sockets to eat all our resources.
86  *  This is direct violation of TCP specs, but it is required
87  *  to prevent DoS attacks. It is called when a retransmission timeout
88  *  or zero probe timeout occurs on orphaned socket.
89  *
90  *  Also close if our net namespace is exiting; in that case there is no
91  *  hope of ever communicating again since all netns interfaces are already
92  *  down (or about to be down), and we need to release our dst references,
93  *  which have been moved to the netns loopback interface, so the namespace
94  *  can finish exiting.  This condition is only possible if we are a kernel
95  *  socket, as those do not hold references to the namespace.
96  *
97  *  Criteria is still not confirmed experimentally and may change.
98  *  We kill the socket, if:
99  *  1. If number of orphaned sockets exceeds an administratively configured
100  *     limit.
101  *  2. If we have strong memory pressure.
102  *  3. If our net namespace is exiting.
103  */
104 static int tcp_out_of_resources(struct sock *sk, bool do_reset)
105 {
106 	struct tcp_sock *tp = tcp_sk(sk);
107 	int shift = 0;
108 
109 	/* If peer does not open window for long time, or did not transmit
110 	 * anything for long time, penalize it. */
111 	if ((s32)(tcp_jiffies32 - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset)
112 		shift++;
113 
114 	/* If some dubious ICMP arrived, penalize even more. */
115 	if (READ_ONCE(sk->sk_err_soft))
116 		shift++;
117 
118 	if (tcp_check_oom(sk, shift)) {
119 		/* Catch exceptional cases, when connection requires reset.
120 		 *      1. Last segment was sent recently. */
121 		if ((s32)(tcp_jiffies32 - tp->lsndtime) <= TCP_TIMEWAIT_LEN ||
122 		    /*  2. Window is closed. */
123 		    (!tp->snd_wnd && !tp->packets_out))
124 			do_reset = true;
125 		if (do_reset)
126 			tcp_send_active_reset(sk, GFP_ATOMIC);
127 		tcp_done(sk);
128 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY);
129 		return 1;
130 	}
131 
132 	if (!check_net(sock_net(sk))) {
133 		/* Not possible to send reset; just close */
134 		tcp_done(sk);
135 		return 1;
136 	}
137 
138 	return 0;
139 }
140 
141 /**
142  *  tcp_orphan_retries() - Returns maximal number of retries on an orphaned socket
143  *  @sk:    Pointer to the current socket.
144  *  @alive: bool, socket alive state
145  */
146 static int tcp_orphan_retries(struct sock *sk, bool alive)
147 {
148 	int retries = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_orphan_retries); /* May be zero. */
149 
150 	/* We know from an ICMP that something is wrong. */
151 	if (READ_ONCE(sk->sk_err_soft) && !alive)
152 		retries = 0;
153 
154 	/* However, if socket sent something recently, select some safe
155 	 * number of retries. 8 corresponds to >100 seconds with minimal
156 	 * RTO of 200msec. */
157 	if (retries == 0 && alive)
158 		retries = 8;
159 	return retries;
160 }
161 
162 static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk)
163 {
164 	const struct net *net = sock_net(sk);
165 	int mss;
166 
167 	/* Black hole detection */
168 	if (!READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing))
169 		return;
170 
171 	if (!icsk->icsk_mtup.enabled) {
172 		icsk->icsk_mtup.enabled = 1;
173 		icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
174 	} else {
175 		mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1;
176 		mss = min(READ_ONCE(net->ipv4.sysctl_tcp_base_mss), mss);
177 		mss = max(mss, READ_ONCE(net->ipv4.sysctl_tcp_mtu_probe_floor));
178 		mss = max(mss, READ_ONCE(net->ipv4.sysctl_tcp_min_snd_mss));
179 		icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
180 	}
181 	tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
182 }
183 
184 static unsigned int tcp_model_timeout(struct sock *sk,
185 				      unsigned int boundary,
186 				      unsigned int rto_base)
187 {
188 	unsigned int linear_backoff_thresh, timeout;
189 
190 	linear_backoff_thresh = ilog2(TCP_RTO_MAX / rto_base);
191 	if (boundary <= linear_backoff_thresh)
192 		timeout = ((2 << boundary) - 1) * rto_base;
193 	else
194 		timeout = ((2 << linear_backoff_thresh) - 1) * rto_base +
195 			(boundary - linear_backoff_thresh) * TCP_RTO_MAX;
196 	return jiffies_to_msecs(timeout);
197 }
198 /**
199  *  retransmits_timed_out() - returns true if this connection has timed out
200  *  @sk:       The current socket
201  *  @boundary: max number of retransmissions
202  *  @timeout:  A custom timeout value.
203  *             If set to 0 the default timeout is calculated and used.
204  *             Using TCP_RTO_MIN and the number of unsuccessful retransmits.
205  *
206  * The default "timeout" value this function can calculate and use
207  * is equivalent to the timeout of a TCP Connection
208  * after "boundary" unsuccessful, exponentially backed-off
209  * retransmissions with an initial RTO of TCP_RTO_MIN.
210  */
211 static bool retransmits_timed_out(struct sock *sk,
212 				  unsigned int boundary,
213 				  unsigned int timeout)
214 {
215 	unsigned int start_ts;
216 
217 	if (!inet_csk(sk)->icsk_retransmits)
218 		return false;
219 
220 	start_ts = tcp_sk(sk)->retrans_stamp;
221 	if (likely(timeout == 0)) {
222 		unsigned int rto_base = TCP_RTO_MIN;
223 
224 		if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
225 			rto_base = tcp_timeout_init(sk);
226 		timeout = tcp_model_timeout(sk, boundary, rto_base);
227 	}
228 
229 	return (s32)(tcp_time_stamp(tcp_sk(sk)) - start_ts - timeout) >= 0;
230 }
231 
232 /* A write timeout has occurred. Process the after effects. */
233 static int tcp_write_timeout(struct sock *sk)
234 {
235 	struct inet_connection_sock *icsk = inet_csk(sk);
236 	struct tcp_sock *tp = tcp_sk(sk);
237 	struct net *net = sock_net(sk);
238 	bool expired = false, do_reset;
239 	int retry_until, max_retransmits;
240 
241 	if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
242 		if (icsk->icsk_retransmits)
243 			__dst_negative_advice(sk);
244 		/* Paired with WRITE_ONCE() in tcp_sock_set_syncnt() */
245 		retry_until = READ_ONCE(icsk->icsk_syn_retries) ? :
246 			READ_ONCE(net->ipv4.sysctl_tcp_syn_retries);
247 
248 		max_retransmits = retry_until;
249 		if (sk->sk_state == TCP_SYN_SENT)
250 			max_retransmits += READ_ONCE(net->ipv4.sysctl_tcp_syn_linear_timeouts);
251 
252 		expired = icsk->icsk_retransmits >= max_retransmits;
253 	} else {
254 		if (retransmits_timed_out(sk, READ_ONCE(net->ipv4.sysctl_tcp_retries1), 0)) {
255 			/* Black hole detection */
256 			tcp_mtu_probing(icsk, sk);
257 
258 			__dst_negative_advice(sk);
259 		}
260 
261 		retry_until = READ_ONCE(net->ipv4.sysctl_tcp_retries2);
262 		if (sock_flag(sk, SOCK_DEAD)) {
263 			const bool alive = icsk->icsk_rto < TCP_RTO_MAX;
264 
265 			retry_until = tcp_orphan_retries(sk, alive);
266 			do_reset = alive ||
267 				!retransmits_timed_out(sk, retry_until, 0);
268 
269 			if (tcp_out_of_resources(sk, do_reset))
270 				return 1;
271 		}
272 	}
273 	if (!expired)
274 		expired = retransmits_timed_out(sk, retry_until,
275 						READ_ONCE(icsk->icsk_user_timeout));
276 	tcp_fastopen_active_detect_blackhole(sk, expired);
277 
278 	if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RTO_CB_FLAG))
279 		tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RTO_CB,
280 				  icsk->icsk_retransmits,
281 				  icsk->icsk_rto, (int)expired);
282 
283 	if (expired) {
284 		/* Has it gone just too far? */
285 		tcp_write_err(sk);
286 		return 1;
287 	}
288 
289 	if (sk_rethink_txhash(sk)) {
290 		tp->timeout_rehash++;
291 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEOUTREHASH);
292 	}
293 
294 	return 0;
295 }
296 
297 /* Called with BH disabled */
298 void tcp_delack_timer_handler(struct sock *sk)
299 {
300 	struct inet_connection_sock *icsk = inet_csk(sk);
301 	struct tcp_sock *tp = tcp_sk(sk);
302 
303 	if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
304 		return;
305 
306 	/* Handling the sack compression case */
307 	if (tp->compressed_ack) {
308 		tcp_mstamp_refresh(tp);
309 		tcp_sack_compress_send_ack(sk);
310 		return;
311 	}
312 
313 	if (!(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
314 		return;
315 
316 	if (time_after(icsk->icsk_ack.timeout, jiffies)) {
317 		sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
318 		return;
319 	}
320 	icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
321 
322 	if (inet_csk_ack_scheduled(sk)) {
323 		if (!inet_csk_in_pingpong_mode(sk)) {
324 			/* Delayed ACK missed: inflate ATO. */
325 			icsk->icsk_ack.ato = min_t(u32, icsk->icsk_ack.ato << 1, icsk->icsk_rto);
326 		} else {
327 			/* Delayed ACK missed: leave pingpong mode and
328 			 * deflate ATO.
329 			 */
330 			inet_csk_exit_pingpong_mode(sk);
331 			icsk->icsk_ack.ato      = TCP_ATO_MIN;
332 		}
333 		tcp_mstamp_refresh(tp);
334 		tcp_send_ack(sk);
335 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKS);
336 	}
337 }
338 
339 
340 /**
341  *  tcp_delack_timer() - The TCP delayed ACK timeout handler
342  *  @t:  Pointer to the timer. (gets casted to struct sock *)
343  *
344  *  This function gets (indirectly) called when the kernel timer for a TCP packet
345  *  of this socket expires. Calls tcp_delack_timer_handler() to do the actual work.
346  *
347  *  Returns: Nothing (void)
348  */
349 static void tcp_delack_timer(struct timer_list *t)
350 {
351 	struct inet_connection_sock *icsk =
352 			from_timer(icsk, t, icsk_delack_timer);
353 	struct sock *sk = &icsk->icsk_inet.sk;
354 
355 	bh_lock_sock(sk);
356 	if (!sock_owned_by_user(sk)) {
357 		tcp_delack_timer_handler(sk);
358 	} else {
359 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
360 		/* deleguate our work to tcp_release_cb() */
361 		if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &sk->sk_tsq_flags))
362 			sock_hold(sk);
363 	}
364 	bh_unlock_sock(sk);
365 	sock_put(sk);
366 }
367 
368 static void tcp_probe_timer(struct sock *sk)
369 {
370 	struct inet_connection_sock *icsk = inet_csk(sk);
371 	struct sk_buff *skb = tcp_send_head(sk);
372 	struct tcp_sock *tp = tcp_sk(sk);
373 	int max_probes;
374 
375 	if (tp->packets_out || !skb) {
376 		icsk->icsk_probes_out = 0;
377 		icsk->icsk_probes_tstamp = 0;
378 		return;
379 	}
380 
381 	/* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as
382 	 * long as the receiver continues to respond probes. We support this by
383 	 * default and reset icsk_probes_out with incoming ACKs. But if the
384 	 * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we
385 	 * kill the socket when the retry count and the time exceeds the
386 	 * corresponding system limit. We also implement similar policy when
387 	 * we use RTO to probe window in tcp_retransmit_timer().
388 	 */
389 	if (!icsk->icsk_probes_tstamp) {
390 		icsk->icsk_probes_tstamp = tcp_jiffies32;
391 	} else {
392 		u32 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
393 
394 		if (user_timeout &&
395 		    (s32)(tcp_jiffies32 - icsk->icsk_probes_tstamp) >=
396 		     msecs_to_jiffies(user_timeout))
397 			goto abort;
398 	}
399 	max_probes = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retries2);
400 	if (sock_flag(sk, SOCK_DEAD)) {
401 		const bool alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX;
402 
403 		max_probes = tcp_orphan_retries(sk, alive);
404 		if (!alive && icsk->icsk_backoff >= max_probes)
405 			goto abort;
406 		if (tcp_out_of_resources(sk, true))
407 			return;
408 	}
409 
410 	if (icsk->icsk_probes_out >= max_probes) {
411 abort:		tcp_write_err(sk);
412 	} else {
413 		/* Only send another probe if we didn't close things up. */
414 		tcp_send_probe0(sk);
415 	}
416 }
417 
418 static void tcp_update_rto_stats(struct sock *sk)
419 {
420 	struct inet_connection_sock *icsk = inet_csk(sk);
421 	struct tcp_sock *tp = tcp_sk(sk);
422 
423 	if (!icsk->icsk_retransmits) {
424 		tp->total_rto_recoveries++;
425 		tp->rto_stamp = tcp_time_stamp(tp);
426 	}
427 	icsk->icsk_retransmits++;
428 	tp->total_rto++;
429 }
430 
431 /*
432  *	Timer for Fast Open socket to retransmit SYNACK. Note that the
433  *	sk here is the child socket, not the parent (listener) socket.
434  */
435 static void tcp_fastopen_synack_timer(struct sock *sk, struct request_sock *req)
436 {
437 	struct inet_connection_sock *icsk = inet_csk(sk);
438 	struct tcp_sock *tp = tcp_sk(sk);
439 	int max_retries;
440 
441 	req->rsk_ops->syn_ack_timeout(req);
442 
443 	/* Add one more retry for fastopen.
444 	 * Paired with WRITE_ONCE() in tcp_sock_set_syncnt()
445 	 */
446 	max_retries = READ_ONCE(icsk->icsk_syn_retries) ? :
447 		READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_synack_retries) + 1;
448 
449 	if (req->num_timeout >= max_retries) {
450 		tcp_write_err(sk);
451 		return;
452 	}
453 	/* Lower cwnd after certain SYNACK timeout like tcp_init_transfer() */
454 	if (icsk->icsk_retransmits == 1)
455 		tcp_enter_loss(sk);
456 	/* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
457 	 * returned from rtx_syn_ack() to make it more persistent like
458 	 * regular retransmit because if the child socket has been accepted
459 	 * it's not good to give up too easily.
460 	 */
461 	inet_rtx_syn_ack(sk, req);
462 	req->num_timeout++;
463 	tcp_update_rto_stats(sk);
464 	if (!tp->retrans_stamp)
465 		tp->retrans_stamp = tcp_time_stamp(tp);
466 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
467 			  req->timeout << req->num_timeout, TCP_RTO_MAX);
468 }
469 
470 static bool tcp_rtx_probe0_timed_out(const struct sock *sk,
471 				     const struct sk_buff *skb)
472 {
473 	const struct tcp_sock *tp = tcp_sk(sk);
474 	const int timeout = TCP_RTO_MAX * 2;
475 	u32 rcv_delta, rtx_delta;
476 
477 	rcv_delta = inet_csk(sk)->icsk_timeout - tp->rcv_tstamp;
478 	if (rcv_delta <= timeout)
479 		return false;
480 
481 	rtx_delta = (u32)msecs_to_jiffies(tcp_time_stamp(tp) -
482 			(tp->retrans_stamp ?: tcp_skb_timestamp(skb)));
483 
484 	return rtx_delta > timeout;
485 }
486 
487 /**
488  *  tcp_retransmit_timer() - The TCP retransmit timeout handler
489  *  @sk:  Pointer to the current socket.
490  *
491  *  This function gets called when the kernel timer for a TCP packet
492  *  of this socket expires.
493  *
494  *  It handles retransmission, timer adjustment and other necessary measures.
495  *
496  *  Returns: Nothing (void)
497  */
498 void tcp_retransmit_timer(struct sock *sk)
499 {
500 	struct tcp_sock *tp = tcp_sk(sk);
501 	struct net *net = sock_net(sk);
502 	struct inet_connection_sock *icsk = inet_csk(sk);
503 	struct request_sock *req;
504 	struct sk_buff *skb;
505 
506 	req = rcu_dereference_protected(tp->fastopen_rsk,
507 					lockdep_sock_is_held(sk));
508 	if (req) {
509 		WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
510 			     sk->sk_state != TCP_FIN_WAIT1);
511 		tcp_fastopen_synack_timer(sk, req);
512 		/* Before we receive ACK to our SYN-ACK don't retransmit
513 		 * anything else (e.g., data or FIN segments).
514 		 */
515 		return;
516 	}
517 
518 	if (!tp->packets_out)
519 		return;
520 
521 	skb = tcp_rtx_queue_head(sk);
522 	if (WARN_ON_ONCE(!skb))
523 		return;
524 
525 	tp->tlp_high_seq = 0;
526 
527 	if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
528 	    !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
529 		/* Receiver dastardly shrinks window. Our retransmits
530 		 * become zero probes, but we should not timeout this
531 		 * connection. If the socket is an orphan, time it out,
532 		 * we cannot allow such beasts to hang infinitely.
533 		 */
534 		struct inet_sock *inet = inet_sk(sk);
535 		u32 rtx_delta;
536 
537 		rtx_delta = tcp_time_stamp(tp) - (tp->retrans_stamp ?: tcp_skb_timestamp(skb));
538 		if (sk->sk_family == AF_INET) {
539 			net_dbg_ratelimited("Probing zero-window on %pI4:%u/%u, seq=%u:%u, recv %ums ago, lasting %ums\n",
540 				&inet->inet_daddr, ntohs(inet->inet_dport),
541 				inet->inet_num, tp->snd_una, tp->snd_nxt,
542 				jiffies_to_msecs(jiffies - tp->rcv_tstamp),
543 				rtx_delta);
544 		}
545 #if IS_ENABLED(CONFIG_IPV6)
546 		else if (sk->sk_family == AF_INET6) {
547 			net_dbg_ratelimited("Probing zero-window on %pI6:%u/%u, seq=%u:%u, recv %ums ago, lasting %ums\n",
548 				&sk->sk_v6_daddr, ntohs(inet->inet_dport),
549 				inet->inet_num, tp->snd_una, tp->snd_nxt,
550 				jiffies_to_msecs(jiffies - tp->rcv_tstamp),
551 				rtx_delta);
552 		}
553 #endif
554 		if (tcp_rtx_probe0_timed_out(sk, skb)) {
555 			tcp_write_err(sk);
556 			goto out;
557 		}
558 		tcp_enter_loss(sk);
559 		tcp_retransmit_skb(sk, skb, 1);
560 		__sk_dst_reset(sk);
561 		goto out_reset_timer;
562 	}
563 
564 	__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEOUTS);
565 	if (tcp_write_timeout(sk))
566 		goto out;
567 
568 	if (icsk->icsk_retransmits == 0) {
569 		int mib_idx = 0;
570 
571 		if (icsk->icsk_ca_state == TCP_CA_Recovery) {
572 			if (tcp_is_sack(tp))
573 				mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
574 			else
575 				mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
576 		} else if (icsk->icsk_ca_state == TCP_CA_Loss) {
577 			mib_idx = LINUX_MIB_TCPLOSSFAILURES;
578 		} else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
579 			   tp->sacked_out) {
580 			if (tcp_is_sack(tp))
581 				mib_idx = LINUX_MIB_TCPSACKFAILURES;
582 			else
583 				mib_idx = LINUX_MIB_TCPRENOFAILURES;
584 		}
585 		if (mib_idx)
586 			__NET_INC_STATS(sock_net(sk), mib_idx);
587 	}
588 
589 	tcp_enter_loss(sk);
590 
591 	tcp_update_rto_stats(sk);
592 	if (tcp_retransmit_skb(sk, tcp_rtx_queue_head(sk), 1) > 0) {
593 		/* Retransmission failed because of local congestion,
594 		 * Let senders fight for local resources conservatively.
595 		 */
596 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
597 					  TCP_RESOURCE_PROBE_INTERVAL,
598 					  TCP_RTO_MAX);
599 		goto out;
600 	}
601 
602 	/* Increase the timeout each time we retransmit.  Note that
603 	 * we do not increase the rtt estimate.  rto is initialized
604 	 * from rtt, but increases here.  Jacobson (SIGCOMM 88) suggests
605 	 * that doubling rto each time is the least we can get away with.
606 	 * In KA9Q, Karn uses this for the first few times, and then
607 	 * goes to quadratic.  netBSD doubles, but only goes up to *64,
608 	 * and clamps at 1 to 64 sec afterwards.  Note that 120 sec is
609 	 * defined in the protocol as the maximum possible RTT.  I guess
610 	 * we'll have to use something other than TCP to talk to the
611 	 * University of Mars.
612 	 *
613 	 * PAWS allows us longer timeouts and large windows, so once
614 	 * implemented ftp to mars will work nicely. We will have to fix
615 	 * the 120 second clamps though!
616 	 */
617 	icsk->icsk_backoff++;
618 
619 out_reset_timer:
620 	/* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
621 	 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
622 	 * might be increased if the stream oscillates between thin and thick,
623 	 * thus the old value might already be too high compared to the value
624 	 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
625 	 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
626 	 * exponential backoff behaviour to avoid continue hammering
627 	 * linear-timeout retransmissions into a black hole
628 	 */
629 	if (sk->sk_state == TCP_ESTABLISHED &&
630 	    (tp->thin_lto || READ_ONCE(net->ipv4.sysctl_tcp_thin_linear_timeouts)) &&
631 	    tcp_stream_is_thin(tp) &&
632 	    icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
633 		icsk->icsk_backoff = 0;
634 		icsk->icsk_rto = clamp(__tcp_set_rto(tp),
635 				       tcp_rto_min(sk),
636 				       TCP_RTO_MAX);
637 	} else if (sk->sk_state != TCP_SYN_SENT ||
638 		   icsk->icsk_backoff >
639 		   READ_ONCE(net->ipv4.sysctl_tcp_syn_linear_timeouts)) {
640 		/* Use normal (exponential) backoff unless linear timeouts are
641 		 * activated.
642 		 */
643 		icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
644 	}
645 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
646 				  tcp_clamp_rto_to_user_timeout(sk), TCP_RTO_MAX);
647 	if (retransmits_timed_out(sk, READ_ONCE(net->ipv4.sysctl_tcp_retries1) + 1, 0))
648 		__sk_dst_reset(sk);
649 
650 out:;
651 }
652 
653 /* Called with bottom-half processing disabled.
654    Called by tcp_write_timer() */
655 void tcp_write_timer_handler(struct sock *sk)
656 {
657 	struct inet_connection_sock *icsk = inet_csk(sk);
658 	int event;
659 
660 	if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
661 	    !icsk->icsk_pending)
662 		return;
663 
664 	if (time_after(icsk->icsk_timeout, jiffies)) {
665 		sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
666 		return;
667 	}
668 
669 	tcp_mstamp_refresh(tcp_sk(sk));
670 	event = icsk->icsk_pending;
671 
672 	switch (event) {
673 	case ICSK_TIME_REO_TIMEOUT:
674 		tcp_rack_reo_timeout(sk);
675 		break;
676 	case ICSK_TIME_LOSS_PROBE:
677 		tcp_send_loss_probe(sk);
678 		break;
679 	case ICSK_TIME_RETRANS:
680 		icsk->icsk_pending = 0;
681 		tcp_retransmit_timer(sk);
682 		break;
683 	case ICSK_TIME_PROBE0:
684 		icsk->icsk_pending = 0;
685 		tcp_probe_timer(sk);
686 		break;
687 	}
688 }
689 
690 static void tcp_write_timer(struct timer_list *t)
691 {
692 	struct inet_connection_sock *icsk =
693 			from_timer(icsk, t, icsk_retransmit_timer);
694 	struct sock *sk = &icsk->icsk_inet.sk;
695 
696 	bh_lock_sock(sk);
697 	if (!sock_owned_by_user(sk)) {
698 		tcp_write_timer_handler(sk);
699 	} else {
700 		/* delegate our work to tcp_release_cb() */
701 		if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &sk->sk_tsq_flags))
702 			sock_hold(sk);
703 	}
704 	bh_unlock_sock(sk);
705 	sock_put(sk);
706 }
707 
708 void tcp_syn_ack_timeout(const struct request_sock *req)
709 {
710 	struct net *net = read_pnet(&inet_rsk(req)->ireq_net);
711 
712 	__NET_INC_STATS(net, LINUX_MIB_TCPTIMEOUTS);
713 }
714 EXPORT_SYMBOL(tcp_syn_ack_timeout);
715 
716 void tcp_set_keepalive(struct sock *sk, int val)
717 {
718 	if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
719 		return;
720 
721 	if (val && !sock_flag(sk, SOCK_KEEPOPEN))
722 		inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
723 	else if (!val)
724 		inet_csk_delete_keepalive_timer(sk);
725 }
726 EXPORT_SYMBOL_GPL(tcp_set_keepalive);
727 
728 
729 static void tcp_keepalive_timer (struct timer_list *t)
730 {
731 	struct sock *sk = from_timer(sk, t, sk_timer);
732 	struct inet_connection_sock *icsk = inet_csk(sk);
733 	struct tcp_sock *tp = tcp_sk(sk);
734 	u32 elapsed;
735 
736 	/* Only process if socket is not in use. */
737 	bh_lock_sock(sk);
738 	if (sock_owned_by_user(sk)) {
739 		/* Try again later. */
740 		inet_csk_reset_keepalive_timer (sk, HZ/20);
741 		goto out;
742 	}
743 
744 	if (sk->sk_state == TCP_LISTEN) {
745 		pr_err("Hmm... keepalive on a LISTEN ???\n");
746 		goto out;
747 	}
748 
749 	tcp_mstamp_refresh(tp);
750 	if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
751 		if (READ_ONCE(tp->linger2) >= 0) {
752 			const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
753 
754 			if (tmo > 0) {
755 				tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
756 				goto out;
757 			}
758 		}
759 		tcp_send_active_reset(sk, GFP_ATOMIC);
760 		goto death;
761 	}
762 
763 	if (!sock_flag(sk, SOCK_KEEPOPEN) ||
764 	    ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_SYN_SENT)))
765 		goto out;
766 
767 	elapsed = keepalive_time_when(tp);
768 
769 	/* It is alive without keepalive 8) */
770 	if (tp->packets_out || !tcp_write_queue_empty(sk))
771 		goto resched;
772 
773 	elapsed = keepalive_time_elapsed(tp);
774 
775 	if (elapsed >= keepalive_time_when(tp)) {
776 		u32 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
777 
778 		/* If the TCP_USER_TIMEOUT option is enabled, use that
779 		 * to determine when to timeout instead.
780 		 */
781 		if ((user_timeout != 0 &&
782 		    elapsed >= msecs_to_jiffies(user_timeout) &&
783 		    icsk->icsk_probes_out > 0) ||
784 		    (user_timeout == 0 &&
785 		    icsk->icsk_probes_out >= keepalive_probes(tp))) {
786 			tcp_send_active_reset(sk, GFP_ATOMIC);
787 			tcp_write_err(sk);
788 			goto out;
789 		}
790 		if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) {
791 			icsk->icsk_probes_out++;
792 			elapsed = keepalive_intvl_when(tp);
793 		} else {
794 			/* If keepalive was lost due to local congestion,
795 			 * try harder.
796 			 */
797 			elapsed = TCP_RESOURCE_PROBE_INTERVAL;
798 		}
799 	} else {
800 		/* It is tp->rcv_tstamp + keepalive_time_when(tp) */
801 		elapsed = keepalive_time_when(tp) - elapsed;
802 	}
803 
804 resched:
805 	inet_csk_reset_keepalive_timer (sk, elapsed);
806 	goto out;
807 
808 death:
809 	tcp_done(sk);
810 
811 out:
812 	bh_unlock_sock(sk);
813 	sock_put(sk);
814 }
815 
816 static enum hrtimer_restart tcp_compressed_ack_kick(struct hrtimer *timer)
817 {
818 	struct tcp_sock *tp = container_of(timer, struct tcp_sock, compressed_ack_timer);
819 	struct sock *sk = (struct sock *)tp;
820 
821 	bh_lock_sock(sk);
822 	if (!sock_owned_by_user(sk)) {
823 		if (tp->compressed_ack) {
824 			/* Since we have to send one ack finally,
825 			 * subtract one from tp->compressed_ack to keep
826 			 * LINUX_MIB_TCPACKCOMPRESSED accurate.
827 			 */
828 			tp->compressed_ack--;
829 			tcp_send_ack(sk);
830 		}
831 	} else {
832 		if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED,
833 				      &sk->sk_tsq_flags))
834 			sock_hold(sk);
835 	}
836 	bh_unlock_sock(sk);
837 
838 	sock_put(sk);
839 
840 	return HRTIMER_NORESTART;
841 }
842 
843 void tcp_init_xmit_timers(struct sock *sk)
844 {
845 	inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
846 				  &tcp_keepalive_timer);
847 	hrtimer_init(&tcp_sk(sk)->pacing_timer, CLOCK_MONOTONIC,
848 		     HRTIMER_MODE_ABS_PINNED_SOFT);
849 	tcp_sk(sk)->pacing_timer.function = tcp_pace_kick;
850 
851 	hrtimer_init(&tcp_sk(sk)->compressed_ack_timer, CLOCK_MONOTONIC,
852 		     HRTIMER_MODE_REL_PINNED_SOFT);
853 	tcp_sk(sk)->compressed_ack_timer.function = tcp_compressed_ack_kick;
854 }
855