xref: /linux/net/ipv4/tcp_timer.c (revision 5691276b39daf90294c6a81fb6d62d667f634c92)
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 #include <net/rstreason.h>
26 
27 static u32 tcp_clamp_rto_to_user_timeout(const struct sock *sk)
28 {
29 	const struct inet_connection_sock *icsk = inet_csk(sk);
30 	const struct tcp_sock *tp = tcp_sk(sk);
31 	u32 elapsed, user_timeout;
32 	s32 remaining;
33 
34 	user_timeout = READ_ONCE(icsk->icsk_user_timeout);
35 	if (!user_timeout)
36 		return icsk->icsk_rto;
37 
38 	elapsed = tcp_time_stamp_ts(tp) - tp->retrans_stamp;
39 	if (tp->tcp_usec_ts)
40 		elapsed /= USEC_PER_MSEC;
41 
42 	remaining = user_timeout - elapsed;
43 	if (remaining <= 0)
44 		return 1; /* user timeout has passed; fire ASAP */
45 
46 	return min_t(u32, icsk->icsk_rto, msecs_to_jiffies(remaining));
47 }
48 
49 u32 tcp_clamp_probe0_to_user_timeout(const struct sock *sk, u32 when)
50 {
51 	const struct inet_connection_sock *icsk = inet_csk(sk);
52 	u32 remaining, user_timeout;
53 	s32 elapsed;
54 
55 	user_timeout = READ_ONCE(icsk->icsk_user_timeout);
56 	if (!user_timeout || !icsk->icsk_probes_tstamp)
57 		return when;
58 
59 	elapsed = tcp_jiffies32 - icsk->icsk_probes_tstamp;
60 	if (unlikely(elapsed < 0))
61 		elapsed = 0;
62 	remaining = msecs_to_jiffies(user_timeout) - elapsed;
63 	remaining = max_t(u32, remaining, TCP_TIMEOUT_MIN);
64 
65 	return min_t(u32, remaining, when);
66 }
67 
68 /**
69  *  tcp_write_err() - close socket and save error info
70  *  @sk:  The socket the error has appeared on.
71  *
72  *  Returns: Nothing (void)
73  */
74 
75 static void tcp_write_err(struct sock *sk)
76 {
77 	WRITE_ONCE(sk->sk_err, READ_ONCE(sk->sk_err_soft) ? : ETIMEDOUT);
78 	sk_error_report(sk);
79 
80 	tcp_write_queue_purge(sk);
81 	tcp_done(sk);
82 	__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT);
83 }
84 
85 /**
86  *  tcp_out_of_resources() - Close socket if out of resources
87  *  @sk:        pointer to current socket
88  *  @do_reset:  send a last packet with reset flag
89  *
90  *  Do not allow orphaned sockets to eat all our resources.
91  *  This is direct violation of TCP specs, but it is required
92  *  to prevent DoS attacks. It is called when a retransmission timeout
93  *  or zero probe timeout occurs on orphaned socket.
94  *
95  *  Also close if our net namespace is exiting; in that case there is no
96  *  hope of ever communicating again since all netns interfaces are already
97  *  down (or about to be down), and we need to release our dst references,
98  *  which have been moved to the netns loopback interface, so the namespace
99  *  can finish exiting.  This condition is only possible if we are a kernel
100  *  socket, as those do not hold references to the namespace.
101  *
102  *  Criteria is still not confirmed experimentally and may change.
103  *  We kill the socket, if:
104  *  1. If number of orphaned sockets exceeds an administratively configured
105  *     limit.
106  *  2. If we have strong memory pressure.
107  *  3. If our net namespace is exiting.
108  */
109 static int tcp_out_of_resources(struct sock *sk, bool do_reset)
110 {
111 	struct tcp_sock *tp = tcp_sk(sk);
112 	int shift = 0;
113 
114 	/* If peer does not open window for long time, or did not transmit
115 	 * anything for long time, penalize it. */
116 	if ((s32)(tcp_jiffies32 - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset)
117 		shift++;
118 
119 	/* If some dubious ICMP arrived, penalize even more. */
120 	if (READ_ONCE(sk->sk_err_soft))
121 		shift++;
122 
123 	if (tcp_check_oom(sk, shift)) {
124 		/* Catch exceptional cases, when connection requires reset.
125 		 *      1. Last segment was sent recently. */
126 		if ((s32)(tcp_jiffies32 - tp->lsndtime) <= TCP_TIMEWAIT_LEN ||
127 		    /*  2. Window is closed. */
128 		    (!tp->snd_wnd && !tp->packets_out))
129 			do_reset = true;
130 		if (do_reset)
131 			tcp_send_active_reset(sk, GFP_ATOMIC,
132 					      SK_RST_REASON_NOT_SPECIFIED);
133 		tcp_done(sk);
134 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY);
135 		return 1;
136 	}
137 
138 	if (!check_net(sock_net(sk))) {
139 		/* Not possible to send reset; just close */
140 		tcp_done(sk);
141 		return 1;
142 	}
143 
144 	return 0;
145 }
146 
147 /**
148  *  tcp_orphan_retries() - Returns maximal number of retries on an orphaned socket
149  *  @sk:    Pointer to the current socket.
150  *  @alive: bool, socket alive state
151  */
152 static int tcp_orphan_retries(struct sock *sk, bool alive)
153 {
154 	int retries = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_orphan_retries); /* May be zero. */
155 
156 	/* We know from an ICMP that something is wrong. */
157 	if (READ_ONCE(sk->sk_err_soft) && !alive)
158 		retries = 0;
159 
160 	/* However, if socket sent something recently, select some safe
161 	 * number of retries. 8 corresponds to >100 seconds with minimal
162 	 * RTO of 200msec. */
163 	if (retries == 0 && alive)
164 		retries = 8;
165 	return retries;
166 }
167 
168 static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk)
169 {
170 	const struct net *net = sock_net(sk);
171 	int mss;
172 
173 	/* Black hole detection */
174 	if (!READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing))
175 		return;
176 
177 	if (!icsk->icsk_mtup.enabled) {
178 		icsk->icsk_mtup.enabled = 1;
179 		icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
180 	} else {
181 		mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1;
182 		mss = min(READ_ONCE(net->ipv4.sysctl_tcp_base_mss), mss);
183 		mss = max(mss, READ_ONCE(net->ipv4.sysctl_tcp_mtu_probe_floor));
184 		mss = max(mss, READ_ONCE(net->ipv4.sysctl_tcp_min_snd_mss));
185 		icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
186 	}
187 	tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
188 }
189 
190 static unsigned int tcp_model_timeout(struct sock *sk,
191 				      unsigned int boundary,
192 				      unsigned int rto_base)
193 {
194 	unsigned int linear_backoff_thresh, timeout;
195 
196 	linear_backoff_thresh = ilog2(TCP_RTO_MAX / rto_base);
197 	if (boundary <= linear_backoff_thresh)
198 		timeout = ((2 << boundary) - 1) * rto_base;
199 	else
200 		timeout = ((2 << linear_backoff_thresh) - 1) * rto_base +
201 			(boundary - linear_backoff_thresh) * TCP_RTO_MAX;
202 	return jiffies_to_msecs(timeout);
203 }
204 /**
205  *  retransmits_timed_out() - returns true if this connection has timed out
206  *  @sk:       The current socket
207  *  @boundary: max number of retransmissions
208  *  @timeout:  A custom timeout value.
209  *             If set to 0 the default timeout is calculated and used.
210  *             Using TCP_RTO_MIN and the number of unsuccessful retransmits.
211  *
212  * The default "timeout" value this function can calculate and use
213  * is equivalent to the timeout of a TCP Connection
214  * after "boundary" unsuccessful, exponentially backed-off
215  * retransmissions with an initial RTO of TCP_RTO_MIN.
216  */
217 static bool retransmits_timed_out(struct sock *sk,
218 				  unsigned int boundary,
219 				  unsigned int timeout)
220 {
221 	struct tcp_sock *tp = tcp_sk(sk);
222 	unsigned int start_ts, delta;
223 
224 	if (!inet_csk(sk)->icsk_retransmits)
225 		return false;
226 
227 	start_ts = tp->retrans_stamp;
228 	if (likely(timeout == 0)) {
229 		unsigned int rto_base = TCP_RTO_MIN;
230 
231 		if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
232 			rto_base = tcp_timeout_init(sk);
233 		timeout = tcp_model_timeout(sk, boundary, rto_base);
234 	}
235 
236 	if (tp->tcp_usec_ts) {
237 		/* delta maybe off up to a jiffy due to timer granularity. */
238 		delta = tp->tcp_mstamp - start_ts + jiffies_to_usecs(1);
239 		return (s32)(delta - timeout * USEC_PER_MSEC) >= 0;
240 	}
241 	return (s32)(tcp_time_stamp_ts(tp) - start_ts - timeout) >= 0;
242 }
243 
244 /* A write timeout has occurred. Process the after effects. */
245 static int tcp_write_timeout(struct sock *sk)
246 {
247 	struct inet_connection_sock *icsk = inet_csk(sk);
248 	struct tcp_sock *tp = tcp_sk(sk);
249 	struct net *net = sock_net(sk);
250 	bool expired = false, do_reset;
251 	int retry_until, max_retransmits;
252 
253 	if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
254 		if (icsk->icsk_retransmits)
255 			__dst_negative_advice(sk);
256 		/* Paired with WRITE_ONCE() in tcp_sock_set_syncnt() */
257 		retry_until = READ_ONCE(icsk->icsk_syn_retries) ? :
258 			READ_ONCE(net->ipv4.sysctl_tcp_syn_retries);
259 
260 		max_retransmits = retry_until;
261 		if (sk->sk_state == TCP_SYN_SENT)
262 			max_retransmits += READ_ONCE(net->ipv4.sysctl_tcp_syn_linear_timeouts);
263 
264 		expired = icsk->icsk_retransmits >= max_retransmits;
265 	} else {
266 		if (retransmits_timed_out(sk, READ_ONCE(net->ipv4.sysctl_tcp_retries1), 0)) {
267 			/* Black hole detection */
268 			tcp_mtu_probing(icsk, sk);
269 
270 			__dst_negative_advice(sk);
271 		}
272 
273 		retry_until = READ_ONCE(net->ipv4.sysctl_tcp_retries2);
274 		if (sock_flag(sk, SOCK_DEAD)) {
275 			const bool alive = icsk->icsk_rto < TCP_RTO_MAX;
276 
277 			retry_until = tcp_orphan_retries(sk, alive);
278 			do_reset = alive ||
279 				!retransmits_timed_out(sk, retry_until, 0);
280 
281 			if (tcp_out_of_resources(sk, do_reset))
282 				return 1;
283 		}
284 	}
285 	if (!expired)
286 		expired = retransmits_timed_out(sk, retry_until,
287 						READ_ONCE(icsk->icsk_user_timeout));
288 	tcp_fastopen_active_detect_blackhole(sk, expired);
289 
290 	if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RTO_CB_FLAG))
291 		tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RTO_CB,
292 				  icsk->icsk_retransmits,
293 				  icsk->icsk_rto, (int)expired);
294 
295 	if (expired) {
296 		/* Has it gone just too far? */
297 		tcp_write_err(sk);
298 		return 1;
299 	}
300 
301 	if (sk_rethink_txhash(sk)) {
302 		tp->timeout_rehash++;
303 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEOUTREHASH);
304 	}
305 
306 	return 0;
307 }
308 
309 /* Called with BH disabled */
310 void tcp_delack_timer_handler(struct sock *sk)
311 {
312 	struct inet_connection_sock *icsk = inet_csk(sk);
313 	struct tcp_sock *tp = tcp_sk(sk);
314 
315 	if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
316 		return;
317 
318 	/* Handling the sack compression case */
319 	if (tp->compressed_ack) {
320 		tcp_mstamp_refresh(tp);
321 		tcp_sack_compress_send_ack(sk);
322 		return;
323 	}
324 
325 	if (!(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
326 		return;
327 
328 	if (time_after(icsk->icsk_ack.timeout, jiffies)) {
329 		sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
330 		return;
331 	}
332 	icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
333 
334 	if (inet_csk_ack_scheduled(sk)) {
335 		if (!inet_csk_in_pingpong_mode(sk)) {
336 			/* Delayed ACK missed: inflate ATO. */
337 			icsk->icsk_ack.ato = min_t(u32, icsk->icsk_ack.ato << 1, icsk->icsk_rto);
338 		} else {
339 			/* Delayed ACK missed: leave pingpong mode and
340 			 * deflate ATO.
341 			 */
342 			inet_csk_exit_pingpong_mode(sk);
343 			icsk->icsk_ack.ato      = TCP_ATO_MIN;
344 		}
345 		tcp_mstamp_refresh(tp);
346 		tcp_send_ack(sk);
347 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKS);
348 	}
349 }
350 
351 
352 /**
353  *  tcp_delack_timer() - The TCP delayed ACK timeout handler
354  *  @t:  Pointer to the timer. (gets casted to struct sock *)
355  *
356  *  This function gets (indirectly) called when the kernel timer for a TCP packet
357  *  of this socket expires. Calls tcp_delack_timer_handler() to do the actual work.
358  *
359  *  Returns: Nothing (void)
360  */
361 static void tcp_delack_timer(struct timer_list *t)
362 {
363 	struct inet_connection_sock *icsk =
364 			from_timer(icsk, t, icsk_delack_timer);
365 	struct sock *sk = &icsk->icsk_inet.sk;
366 
367 	bh_lock_sock(sk);
368 	if (!sock_owned_by_user(sk)) {
369 		tcp_delack_timer_handler(sk);
370 	} else {
371 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
372 		/* deleguate our work to tcp_release_cb() */
373 		if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &sk->sk_tsq_flags))
374 			sock_hold(sk);
375 	}
376 	bh_unlock_sock(sk);
377 	sock_put(sk);
378 }
379 
380 static void tcp_probe_timer(struct sock *sk)
381 {
382 	struct inet_connection_sock *icsk = inet_csk(sk);
383 	struct sk_buff *skb = tcp_send_head(sk);
384 	struct tcp_sock *tp = tcp_sk(sk);
385 	int max_probes;
386 
387 	if (tp->packets_out || !skb) {
388 		icsk->icsk_probes_out = 0;
389 		icsk->icsk_probes_tstamp = 0;
390 		return;
391 	}
392 
393 	/* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as
394 	 * long as the receiver continues to respond probes. We support this by
395 	 * default and reset icsk_probes_out with incoming ACKs. But if the
396 	 * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we
397 	 * kill the socket when the retry count and the time exceeds the
398 	 * corresponding system limit. We also implement similar policy when
399 	 * we use RTO to probe window in tcp_retransmit_timer().
400 	 */
401 	if (!icsk->icsk_probes_tstamp) {
402 		icsk->icsk_probes_tstamp = tcp_jiffies32;
403 	} else {
404 		u32 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
405 
406 		if (user_timeout &&
407 		    (s32)(tcp_jiffies32 - icsk->icsk_probes_tstamp) >=
408 		     msecs_to_jiffies(user_timeout))
409 			goto abort;
410 	}
411 	max_probes = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retries2);
412 	if (sock_flag(sk, SOCK_DEAD)) {
413 		const bool alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX;
414 
415 		max_probes = tcp_orphan_retries(sk, alive);
416 		if (!alive && icsk->icsk_backoff >= max_probes)
417 			goto abort;
418 		if (tcp_out_of_resources(sk, true))
419 			return;
420 	}
421 
422 	if (icsk->icsk_probes_out >= max_probes) {
423 abort:		tcp_write_err(sk);
424 	} else {
425 		/* Only send another probe if we didn't close things up. */
426 		tcp_send_probe0(sk);
427 	}
428 }
429 
430 static void tcp_update_rto_stats(struct sock *sk)
431 {
432 	struct inet_connection_sock *icsk = inet_csk(sk);
433 	struct tcp_sock *tp = tcp_sk(sk);
434 
435 	if (!icsk->icsk_retransmits) {
436 		tp->total_rto_recoveries++;
437 		tp->rto_stamp = tcp_time_stamp_ms(tp);
438 	}
439 	icsk->icsk_retransmits++;
440 	tp->total_rto++;
441 }
442 
443 /*
444  *	Timer for Fast Open socket to retransmit SYNACK. Note that the
445  *	sk here is the child socket, not the parent (listener) socket.
446  */
447 static void tcp_fastopen_synack_timer(struct sock *sk, struct request_sock *req)
448 {
449 	struct inet_connection_sock *icsk = inet_csk(sk);
450 	struct tcp_sock *tp = tcp_sk(sk);
451 	int max_retries;
452 
453 	req->rsk_ops->syn_ack_timeout(req);
454 
455 	/* Add one more retry for fastopen.
456 	 * Paired with WRITE_ONCE() in tcp_sock_set_syncnt()
457 	 */
458 	max_retries = READ_ONCE(icsk->icsk_syn_retries) ? :
459 		READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_synack_retries) + 1;
460 
461 	if (req->num_timeout >= max_retries) {
462 		tcp_write_err(sk);
463 		return;
464 	}
465 	/* Lower cwnd after certain SYNACK timeout like tcp_init_transfer() */
466 	if (icsk->icsk_retransmits == 1)
467 		tcp_enter_loss(sk);
468 	/* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
469 	 * returned from rtx_syn_ack() to make it more persistent like
470 	 * regular retransmit because if the child socket has been accepted
471 	 * it's not good to give up too easily.
472 	 */
473 	inet_rtx_syn_ack(sk, req);
474 	req->num_timeout++;
475 	tcp_update_rto_stats(sk);
476 	if (!tp->retrans_stamp)
477 		tp->retrans_stamp = tcp_time_stamp_ts(tp);
478 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
479 			  req->timeout << req->num_timeout, TCP_RTO_MAX);
480 }
481 
482 static bool tcp_rtx_probe0_timed_out(const struct sock *sk,
483 				     const struct sk_buff *skb,
484 				     u32 rtx_delta)
485 {
486 	const struct tcp_sock *tp = tcp_sk(sk);
487 	const int timeout = TCP_RTO_MAX * 2;
488 	u32 rcv_delta;
489 
490 	rcv_delta = inet_csk(sk)->icsk_timeout - tp->rcv_tstamp;
491 	if (rcv_delta <= timeout)
492 		return false;
493 
494 	return msecs_to_jiffies(rtx_delta) > timeout;
495 }
496 
497 /**
498  *  tcp_retransmit_timer() - The TCP retransmit timeout handler
499  *  @sk:  Pointer to the current socket.
500  *
501  *  This function gets called when the kernel timer for a TCP packet
502  *  of this socket expires.
503  *
504  *  It handles retransmission, timer adjustment and other necessary measures.
505  *
506  *  Returns: Nothing (void)
507  */
508 void tcp_retransmit_timer(struct sock *sk)
509 {
510 	struct tcp_sock *tp = tcp_sk(sk);
511 	struct net *net = sock_net(sk);
512 	struct inet_connection_sock *icsk = inet_csk(sk);
513 	struct request_sock *req;
514 	struct sk_buff *skb;
515 
516 	req = rcu_dereference_protected(tp->fastopen_rsk,
517 					lockdep_sock_is_held(sk));
518 	if (req) {
519 		WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
520 			     sk->sk_state != TCP_FIN_WAIT1);
521 		tcp_fastopen_synack_timer(sk, req);
522 		/* Before we receive ACK to our SYN-ACK don't retransmit
523 		 * anything else (e.g., data or FIN segments).
524 		 */
525 		return;
526 	}
527 
528 	if (!tp->packets_out)
529 		return;
530 
531 	skb = tcp_rtx_queue_head(sk);
532 	if (WARN_ON_ONCE(!skb))
533 		return;
534 
535 	tp->tlp_high_seq = 0;
536 
537 	if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
538 	    !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
539 		/* Receiver dastardly shrinks window. Our retransmits
540 		 * become zero probes, but we should not timeout this
541 		 * connection. If the socket is an orphan, time it out,
542 		 * we cannot allow such beasts to hang infinitely.
543 		 */
544 		struct inet_sock *inet = inet_sk(sk);
545 		u32 rtx_delta;
546 
547 		rtx_delta = tcp_time_stamp_ts(tp) - (tp->retrans_stamp ?:
548 				tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb));
549 		if (tp->tcp_usec_ts)
550 			rtx_delta /= USEC_PER_MSEC;
551 
552 		if (sk->sk_family == AF_INET) {
553 			net_dbg_ratelimited("Probing zero-window on %pI4:%u/%u, seq=%u:%u, recv %ums ago, lasting %ums\n",
554 				&inet->inet_daddr, ntohs(inet->inet_dport),
555 				inet->inet_num, tp->snd_una, tp->snd_nxt,
556 				jiffies_to_msecs(jiffies - tp->rcv_tstamp),
557 				rtx_delta);
558 		}
559 #if IS_ENABLED(CONFIG_IPV6)
560 		else if (sk->sk_family == AF_INET6) {
561 			net_dbg_ratelimited("Probing zero-window on %pI6:%u/%u, seq=%u:%u, recv %ums ago, lasting %ums\n",
562 				&sk->sk_v6_daddr, ntohs(inet->inet_dport),
563 				inet->inet_num, tp->snd_una, tp->snd_nxt,
564 				jiffies_to_msecs(jiffies - tp->rcv_tstamp),
565 				rtx_delta);
566 		}
567 #endif
568 		if (tcp_rtx_probe0_timed_out(sk, skb, rtx_delta)) {
569 			tcp_write_err(sk);
570 			goto out;
571 		}
572 		tcp_enter_loss(sk);
573 		tcp_retransmit_skb(sk, skb, 1);
574 		__sk_dst_reset(sk);
575 		goto out_reset_timer;
576 	}
577 
578 	__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEOUTS);
579 	if (tcp_write_timeout(sk))
580 		goto out;
581 
582 	if (icsk->icsk_retransmits == 0) {
583 		int mib_idx = 0;
584 
585 		if (icsk->icsk_ca_state == TCP_CA_Recovery) {
586 			if (tcp_is_sack(tp))
587 				mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
588 			else
589 				mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
590 		} else if (icsk->icsk_ca_state == TCP_CA_Loss) {
591 			mib_idx = LINUX_MIB_TCPLOSSFAILURES;
592 		} else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
593 			   tp->sacked_out) {
594 			if (tcp_is_sack(tp))
595 				mib_idx = LINUX_MIB_TCPSACKFAILURES;
596 			else
597 				mib_idx = LINUX_MIB_TCPRENOFAILURES;
598 		}
599 		if (mib_idx)
600 			__NET_INC_STATS(sock_net(sk), mib_idx);
601 	}
602 
603 	tcp_enter_loss(sk);
604 
605 	tcp_update_rto_stats(sk);
606 	if (tcp_retransmit_skb(sk, tcp_rtx_queue_head(sk), 1) > 0) {
607 		/* Retransmission failed because of local congestion,
608 		 * Let senders fight for local resources conservatively.
609 		 */
610 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
611 					  TCP_RESOURCE_PROBE_INTERVAL,
612 					  TCP_RTO_MAX);
613 		goto out;
614 	}
615 
616 	/* Increase the timeout each time we retransmit.  Note that
617 	 * we do not increase the rtt estimate.  rto is initialized
618 	 * from rtt, but increases here.  Jacobson (SIGCOMM 88) suggests
619 	 * that doubling rto each time is the least we can get away with.
620 	 * In KA9Q, Karn uses this for the first few times, and then
621 	 * goes to quadratic.  netBSD doubles, but only goes up to *64,
622 	 * and clamps at 1 to 64 sec afterwards.  Note that 120 sec is
623 	 * defined in the protocol as the maximum possible RTT.  I guess
624 	 * we'll have to use something other than TCP to talk to the
625 	 * University of Mars.
626 	 *
627 	 * PAWS allows us longer timeouts and large windows, so once
628 	 * implemented ftp to mars will work nicely. We will have to fix
629 	 * the 120 second clamps though!
630 	 */
631 
632 out_reset_timer:
633 	/* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
634 	 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
635 	 * might be increased if the stream oscillates between thin and thick,
636 	 * thus the old value might already be too high compared to the value
637 	 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
638 	 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
639 	 * exponential backoff behaviour to avoid continue hammering
640 	 * linear-timeout retransmissions into a black hole
641 	 */
642 	if (sk->sk_state == TCP_ESTABLISHED &&
643 	    (tp->thin_lto || READ_ONCE(net->ipv4.sysctl_tcp_thin_linear_timeouts)) &&
644 	    tcp_stream_is_thin(tp) &&
645 	    icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
646 		icsk->icsk_backoff = 0;
647 		icsk->icsk_rto = clamp(__tcp_set_rto(tp),
648 				       tcp_rto_min(sk),
649 				       TCP_RTO_MAX);
650 	} else if (sk->sk_state != TCP_SYN_SENT ||
651 		   tp->total_rto >
652 		   READ_ONCE(net->ipv4.sysctl_tcp_syn_linear_timeouts)) {
653 		/* Use normal (exponential) backoff unless linear timeouts are
654 		 * activated.
655 		 */
656 		icsk->icsk_backoff++;
657 		icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
658 	}
659 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
660 				  tcp_clamp_rto_to_user_timeout(sk), TCP_RTO_MAX);
661 	if (retransmits_timed_out(sk, READ_ONCE(net->ipv4.sysctl_tcp_retries1) + 1, 0))
662 		__sk_dst_reset(sk);
663 
664 out:;
665 }
666 
667 /* Called with bottom-half processing disabled.
668    Called by tcp_write_timer() */
669 void tcp_write_timer_handler(struct sock *sk)
670 {
671 	struct inet_connection_sock *icsk = inet_csk(sk);
672 	int event;
673 
674 	if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
675 	    !icsk->icsk_pending)
676 		return;
677 
678 	if (time_after(icsk->icsk_timeout, jiffies)) {
679 		sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
680 		return;
681 	}
682 
683 	tcp_mstamp_refresh(tcp_sk(sk));
684 	event = icsk->icsk_pending;
685 
686 	switch (event) {
687 	case ICSK_TIME_REO_TIMEOUT:
688 		tcp_rack_reo_timeout(sk);
689 		break;
690 	case ICSK_TIME_LOSS_PROBE:
691 		tcp_send_loss_probe(sk);
692 		break;
693 	case ICSK_TIME_RETRANS:
694 		icsk->icsk_pending = 0;
695 		tcp_retransmit_timer(sk);
696 		break;
697 	case ICSK_TIME_PROBE0:
698 		icsk->icsk_pending = 0;
699 		tcp_probe_timer(sk);
700 		break;
701 	}
702 }
703 
704 static void tcp_write_timer(struct timer_list *t)
705 {
706 	struct inet_connection_sock *icsk =
707 			from_timer(icsk, t, icsk_retransmit_timer);
708 	struct sock *sk = &icsk->icsk_inet.sk;
709 
710 	bh_lock_sock(sk);
711 	if (!sock_owned_by_user(sk)) {
712 		tcp_write_timer_handler(sk);
713 	} else {
714 		/* delegate our work to tcp_release_cb() */
715 		if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &sk->sk_tsq_flags))
716 			sock_hold(sk);
717 	}
718 	bh_unlock_sock(sk);
719 	sock_put(sk);
720 }
721 
722 void tcp_syn_ack_timeout(const struct request_sock *req)
723 {
724 	struct net *net = read_pnet(&inet_rsk(req)->ireq_net);
725 
726 	__NET_INC_STATS(net, LINUX_MIB_TCPTIMEOUTS);
727 }
728 EXPORT_SYMBOL(tcp_syn_ack_timeout);
729 
730 void tcp_set_keepalive(struct sock *sk, int val)
731 {
732 	if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
733 		return;
734 
735 	if (val && !sock_flag(sk, SOCK_KEEPOPEN))
736 		inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
737 	else if (!val)
738 		inet_csk_delete_keepalive_timer(sk);
739 }
740 EXPORT_SYMBOL_GPL(tcp_set_keepalive);
741 
742 
743 static void tcp_keepalive_timer (struct timer_list *t)
744 {
745 	struct sock *sk = from_timer(sk, t, sk_timer);
746 	struct inet_connection_sock *icsk = inet_csk(sk);
747 	struct tcp_sock *tp = tcp_sk(sk);
748 	u32 elapsed;
749 
750 	/* Only process if socket is not in use. */
751 	bh_lock_sock(sk);
752 	if (sock_owned_by_user(sk)) {
753 		/* Try again later. */
754 		inet_csk_reset_keepalive_timer (sk, HZ/20);
755 		goto out;
756 	}
757 
758 	if (sk->sk_state == TCP_LISTEN) {
759 		pr_err("Hmm... keepalive on a LISTEN ???\n");
760 		goto out;
761 	}
762 
763 	tcp_mstamp_refresh(tp);
764 	if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
765 		if (READ_ONCE(tp->linger2) >= 0) {
766 			const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
767 
768 			if (tmo > 0) {
769 				tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
770 				goto out;
771 			}
772 		}
773 		tcp_send_active_reset(sk, GFP_ATOMIC, SK_RST_REASON_NOT_SPECIFIED);
774 		goto death;
775 	}
776 
777 	if (!sock_flag(sk, SOCK_KEEPOPEN) ||
778 	    ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_SYN_SENT)))
779 		goto out;
780 
781 	elapsed = keepalive_time_when(tp);
782 
783 	/* It is alive without keepalive 8) */
784 	if (tp->packets_out || !tcp_write_queue_empty(sk))
785 		goto resched;
786 
787 	elapsed = keepalive_time_elapsed(tp);
788 
789 	if (elapsed >= keepalive_time_when(tp)) {
790 		u32 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
791 
792 		/* If the TCP_USER_TIMEOUT option is enabled, use that
793 		 * to determine when to timeout instead.
794 		 */
795 		if ((user_timeout != 0 &&
796 		    elapsed >= msecs_to_jiffies(user_timeout) &&
797 		    icsk->icsk_probes_out > 0) ||
798 		    (user_timeout == 0 &&
799 		    icsk->icsk_probes_out >= keepalive_probes(tp))) {
800 			tcp_send_active_reset(sk, GFP_ATOMIC,
801 					      SK_RST_REASON_NOT_SPECIFIED);
802 			tcp_write_err(sk);
803 			goto out;
804 		}
805 		if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) {
806 			icsk->icsk_probes_out++;
807 			elapsed = keepalive_intvl_when(tp);
808 		} else {
809 			/* If keepalive was lost due to local congestion,
810 			 * try harder.
811 			 */
812 			elapsed = TCP_RESOURCE_PROBE_INTERVAL;
813 		}
814 	} else {
815 		/* It is tp->rcv_tstamp + keepalive_time_when(tp) */
816 		elapsed = keepalive_time_when(tp) - elapsed;
817 	}
818 
819 resched:
820 	inet_csk_reset_keepalive_timer (sk, elapsed);
821 	goto out;
822 
823 death:
824 	tcp_done(sk);
825 
826 out:
827 	bh_unlock_sock(sk);
828 	sock_put(sk);
829 }
830 
831 static enum hrtimer_restart tcp_compressed_ack_kick(struct hrtimer *timer)
832 {
833 	struct tcp_sock *tp = container_of(timer, struct tcp_sock, compressed_ack_timer);
834 	struct sock *sk = (struct sock *)tp;
835 
836 	bh_lock_sock(sk);
837 	if (!sock_owned_by_user(sk)) {
838 		if (tp->compressed_ack) {
839 			/* Since we have to send one ack finally,
840 			 * subtract one from tp->compressed_ack to keep
841 			 * LINUX_MIB_TCPACKCOMPRESSED accurate.
842 			 */
843 			tp->compressed_ack--;
844 			tcp_send_ack(sk);
845 		}
846 	} else {
847 		if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED,
848 				      &sk->sk_tsq_flags))
849 			sock_hold(sk);
850 	}
851 	bh_unlock_sock(sk);
852 
853 	sock_put(sk);
854 
855 	return HRTIMER_NORESTART;
856 }
857 
858 void tcp_init_xmit_timers(struct sock *sk)
859 {
860 	inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
861 				  &tcp_keepalive_timer);
862 	hrtimer_init(&tcp_sk(sk)->pacing_timer, CLOCK_MONOTONIC,
863 		     HRTIMER_MODE_ABS_PINNED_SOFT);
864 	tcp_sk(sk)->pacing_timer.function = tcp_pace_kick;
865 
866 	hrtimer_init(&tcp_sk(sk)->compressed_ack_timer, CLOCK_MONOTONIC,
867 		     HRTIMER_MODE_REL_PINNED_SOFT);
868 	tcp_sk(sk)->compressed_ack_timer.function = tcp_compressed_ack_kick;
869 }
870