xref: /linux/net/ipv4/tcp_timer.c (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
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 	tcp_done_with_error(sk, READ_ONCE(sk->sk_err_soft) ? : ETIMEDOUT);
78 	__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT);
79 }
80 
81 /**
82  *  tcp_out_of_resources() - Close socket if out of resources
83  *  @sk:        pointer to current socket
84  *  @do_reset:  send a last packet with reset flag
85  *
86  *  Do not allow orphaned sockets to eat all our resources.
87  *  This is direct violation of TCP specs, but it is required
88  *  to prevent DoS attacks. It is called when a retransmission timeout
89  *  or zero probe timeout occurs on orphaned socket.
90  *
91  *  Also close if our net namespace is exiting; in that case there is no
92  *  hope of ever communicating again since all netns interfaces are already
93  *  down (or about to be down), and we need to release our dst references,
94  *  which have been moved to the netns loopback interface, so the namespace
95  *  can finish exiting.  This condition is only possible if we are a kernel
96  *  socket, as those do not hold references to the namespace.
97  *
98  *  Criteria is still not confirmed experimentally and may change.
99  *  We kill the socket, if:
100  *  1. If number of orphaned sockets exceeds an administratively configured
101  *     limit.
102  *  2. If we have strong memory pressure.
103  *  3. If our net namespace is exiting.
104  */
105 static int tcp_out_of_resources(struct sock *sk, bool do_reset)
106 {
107 	struct tcp_sock *tp = tcp_sk(sk);
108 	int shift = 0;
109 
110 	/* If peer does not open window for long time, or did not transmit
111 	 * anything for long time, penalize it. */
112 	if ((s32)(tcp_jiffies32 - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset)
113 		shift++;
114 
115 	/* If some dubious ICMP arrived, penalize even more. */
116 	if (READ_ONCE(sk->sk_err_soft))
117 		shift++;
118 
119 	if (tcp_check_oom(sk, shift)) {
120 		/* Catch exceptional cases, when connection requires reset.
121 		 *      1. Last segment was sent recently. */
122 		if ((s32)(tcp_jiffies32 - tp->lsndtime) <= TCP_TIMEWAIT_LEN ||
123 		    /*  2. Window is closed. */
124 		    (!tp->snd_wnd && !tp->packets_out))
125 			do_reset = true;
126 		if (do_reset)
127 			tcp_send_active_reset(sk, GFP_ATOMIC,
128 					      SK_RST_REASON_TCP_ABORT_ON_MEMORY);
129 		tcp_done(sk);
130 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY);
131 		return 1;
132 	}
133 
134 	if (!check_net(sock_net(sk))) {
135 		/* Not possible to send reset; just close */
136 		tcp_done(sk);
137 		return 1;
138 	}
139 
140 	return 0;
141 }
142 
143 /**
144  *  tcp_orphan_retries() - Returns maximal number of retries on an orphaned socket
145  *  @sk:    Pointer to the current socket.
146  *  @alive: bool, socket alive state
147  */
148 static int tcp_orphan_retries(struct sock *sk, bool alive)
149 {
150 	int retries = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_orphan_retries); /* May be zero. */
151 
152 	/* We know from an ICMP that something is wrong. */
153 	if (READ_ONCE(sk->sk_err_soft) && !alive)
154 		retries = 0;
155 
156 	/* However, if socket sent something recently, select some safe
157 	 * number of retries. 8 corresponds to >100 seconds with minimal
158 	 * RTO of 200msec. */
159 	if (retries == 0 && alive)
160 		retries = 8;
161 	return retries;
162 }
163 
164 static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk)
165 {
166 	const struct net *net = sock_net(sk);
167 	int mss;
168 
169 	/* Black hole detection */
170 	if (!READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing))
171 		return;
172 
173 	if (!icsk->icsk_mtup.enabled) {
174 		icsk->icsk_mtup.enabled = 1;
175 		icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
176 	} else {
177 		mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1;
178 		mss = min(READ_ONCE(net->ipv4.sysctl_tcp_base_mss), mss);
179 		mss = max(mss, READ_ONCE(net->ipv4.sysctl_tcp_mtu_probe_floor));
180 		mss = max(mss, READ_ONCE(net->ipv4.sysctl_tcp_min_snd_mss));
181 		icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
182 	}
183 	tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
184 }
185 
186 static unsigned int tcp_model_timeout(struct sock *sk,
187 				      unsigned int boundary,
188 				      unsigned int rto_base)
189 {
190 	unsigned int linear_backoff_thresh, timeout;
191 
192 	linear_backoff_thresh = ilog2(TCP_RTO_MAX / rto_base);
193 	if (boundary <= linear_backoff_thresh)
194 		timeout = ((2 << boundary) - 1) * rto_base;
195 	else
196 		timeout = ((2 << linear_backoff_thresh) - 1) * rto_base +
197 			(boundary - linear_backoff_thresh) * TCP_RTO_MAX;
198 	return jiffies_to_msecs(timeout);
199 }
200 /**
201  *  retransmits_timed_out() - returns true if this connection has timed out
202  *  @sk:       The current socket
203  *  @boundary: max number of retransmissions
204  *  @timeout:  A custom timeout value.
205  *             If set to 0 the default timeout is calculated and used.
206  *             Using TCP_RTO_MIN and the number of unsuccessful retransmits.
207  *
208  * The default "timeout" value this function can calculate and use
209  * is equivalent to the timeout of a TCP Connection
210  * after "boundary" unsuccessful, exponentially backed-off
211  * retransmissions with an initial RTO of TCP_RTO_MIN.
212  */
213 static bool retransmits_timed_out(struct sock *sk,
214 				  unsigned int boundary,
215 				  unsigned int timeout)
216 {
217 	struct tcp_sock *tp = tcp_sk(sk);
218 	unsigned int start_ts, delta;
219 
220 	if (!inet_csk(sk)->icsk_retransmits)
221 		return false;
222 
223 	start_ts = tp->retrans_stamp;
224 	if (likely(timeout == 0)) {
225 		unsigned int rto_base = TCP_RTO_MIN;
226 
227 		if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
228 			rto_base = tcp_timeout_init(sk);
229 		timeout = tcp_model_timeout(sk, boundary, rto_base);
230 	}
231 
232 	if (tp->tcp_usec_ts) {
233 		/* delta maybe off up to a jiffy due to timer granularity. */
234 		delta = tp->tcp_mstamp - start_ts + jiffies_to_usecs(1);
235 		return (s32)(delta - timeout * USEC_PER_MSEC) >= 0;
236 	}
237 	return (s32)(tcp_time_stamp_ts(tp) - start_ts - timeout) >= 0;
238 }
239 
240 /* A write timeout has occurred. Process the after effects. */
241 static int tcp_write_timeout(struct sock *sk)
242 {
243 	struct inet_connection_sock *icsk = inet_csk(sk);
244 	struct tcp_sock *tp = tcp_sk(sk);
245 	struct net *net = sock_net(sk);
246 	bool expired = false, do_reset;
247 	int retry_until, max_retransmits;
248 
249 	if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
250 		if (icsk->icsk_retransmits)
251 			__dst_negative_advice(sk);
252 		/* Paired with WRITE_ONCE() in tcp_sock_set_syncnt() */
253 		retry_until = READ_ONCE(icsk->icsk_syn_retries) ? :
254 			READ_ONCE(net->ipv4.sysctl_tcp_syn_retries);
255 
256 		max_retransmits = retry_until;
257 		if (sk->sk_state == TCP_SYN_SENT)
258 			max_retransmits += READ_ONCE(net->ipv4.sysctl_tcp_syn_linear_timeouts);
259 
260 		expired = icsk->icsk_retransmits >= max_retransmits;
261 	} else {
262 		if (retransmits_timed_out(sk, READ_ONCE(net->ipv4.sysctl_tcp_retries1), 0)) {
263 			/* Black hole detection */
264 			tcp_mtu_probing(icsk, sk);
265 
266 			__dst_negative_advice(sk);
267 		}
268 
269 		retry_until = READ_ONCE(net->ipv4.sysctl_tcp_retries2);
270 		if (sock_flag(sk, SOCK_DEAD)) {
271 			const bool alive = icsk->icsk_rto < TCP_RTO_MAX;
272 
273 			retry_until = tcp_orphan_retries(sk, alive);
274 			do_reset = alive ||
275 				!retransmits_timed_out(sk, retry_until, 0);
276 
277 			if (tcp_out_of_resources(sk, do_reset))
278 				return 1;
279 		}
280 	}
281 	if (!expired)
282 		expired = retransmits_timed_out(sk, retry_until,
283 						READ_ONCE(icsk->icsk_user_timeout));
284 	tcp_fastopen_active_detect_blackhole(sk, expired);
285 	mptcp_active_detect_blackhole(sk, expired);
286 
287 	if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RTO_CB_FLAG))
288 		tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RTO_CB,
289 				  icsk->icsk_retransmits,
290 				  icsk->icsk_rto, (int)expired);
291 
292 	if (expired) {
293 		/* Has it gone just too far? */
294 		tcp_write_err(sk);
295 		return 1;
296 	}
297 
298 	if (sk_rethink_txhash(sk)) {
299 		tp->timeout_rehash++;
300 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEOUTREHASH);
301 	}
302 
303 	return 0;
304 }
305 
306 /* Called with BH disabled */
307 void tcp_delack_timer_handler(struct sock *sk)
308 {
309 	struct inet_connection_sock *icsk = inet_csk(sk);
310 	struct tcp_sock *tp = tcp_sk(sk);
311 
312 	if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
313 		return;
314 
315 	/* Handling the sack compression case */
316 	if (tp->compressed_ack) {
317 		tcp_mstamp_refresh(tp);
318 		tcp_sack_compress_send_ack(sk);
319 		return;
320 	}
321 
322 	if (!(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
323 		return;
324 
325 	if (time_after(icsk->icsk_ack.timeout, jiffies)) {
326 		sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
327 		return;
328 	}
329 	icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
330 
331 	if (inet_csk_ack_scheduled(sk)) {
332 		if (!inet_csk_in_pingpong_mode(sk)) {
333 			/* Delayed ACK missed: inflate ATO. */
334 			icsk->icsk_ack.ato = min_t(u32, icsk->icsk_ack.ato << 1, icsk->icsk_rto);
335 		} else {
336 			/* Delayed ACK missed: leave pingpong mode and
337 			 * deflate ATO.
338 			 */
339 			inet_csk_exit_pingpong_mode(sk);
340 			icsk->icsk_ack.ato      = TCP_ATO_MIN;
341 		}
342 		tcp_mstamp_refresh(tp);
343 		tcp_send_ack(sk);
344 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKS);
345 	}
346 }
347 
348 
349 /**
350  *  tcp_delack_timer() - The TCP delayed ACK timeout handler
351  *  @t:  Pointer to the timer. (gets casted to struct sock *)
352  *
353  *  This function gets (indirectly) called when the kernel timer for a TCP packet
354  *  of this socket expires. Calls tcp_delack_timer_handler() to do the actual work.
355  *
356  *  Returns: Nothing (void)
357  */
358 static void tcp_delack_timer(struct timer_list *t)
359 {
360 	struct inet_connection_sock *icsk =
361 			from_timer(icsk, t, icsk_delack_timer);
362 	struct sock *sk = &icsk->icsk_inet.sk;
363 
364 	bh_lock_sock(sk);
365 	if (!sock_owned_by_user(sk)) {
366 		tcp_delack_timer_handler(sk);
367 	} else {
368 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
369 		/* deleguate our work to tcp_release_cb() */
370 		if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &sk->sk_tsq_flags))
371 			sock_hold(sk);
372 	}
373 	bh_unlock_sock(sk);
374 	sock_put(sk);
375 }
376 
377 static void tcp_probe_timer(struct sock *sk)
378 {
379 	struct inet_connection_sock *icsk = inet_csk(sk);
380 	struct sk_buff *skb = tcp_send_head(sk);
381 	struct tcp_sock *tp = tcp_sk(sk);
382 	int max_probes;
383 
384 	if (tp->packets_out || !skb) {
385 		icsk->icsk_probes_out = 0;
386 		icsk->icsk_probes_tstamp = 0;
387 		return;
388 	}
389 
390 	/* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as
391 	 * long as the receiver continues to respond probes. We support this by
392 	 * default and reset icsk_probes_out with incoming ACKs. But if the
393 	 * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we
394 	 * kill the socket when the retry count and the time exceeds the
395 	 * corresponding system limit. We also implement similar policy when
396 	 * we use RTO to probe window in tcp_retransmit_timer().
397 	 */
398 	if (!icsk->icsk_probes_tstamp) {
399 		icsk->icsk_probes_tstamp = tcp_jiffies32;
400 	} else {
401 		u32 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
402 
403 		if (user_timeout &&
404 		    (s32)(tcp_jiffies32 - icsk->icsk_probes_tstamp) >=
405 		     msecs_to_jiffies(user_timeout))
406 			goto abort;
407 	}
408 	max_probes = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retries2);
409 	if (sock_flag(sk, SOCK_DEAD)) {
410 		const bool alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX;
411 
412 		max_probes = tcp_orphan_retries(sk, alive);
413 		if (!alive && icsk->icsk_backoff >= max_probes)
414 			goto abort;
415 		if (tcp_out_of_resources(sk, true))
416 			return;
417 	}
418 
419 	if (icsk->icsk_probes_out >= max_probes) {
420 abort:		tcp_write_err(sk);
421 	} else {
422 		/* Only send another probe if we didn't close things up. */
423 		tcp_send_probe0(sk);
424 	}
425 }
426 
427 static void tcp_update_rto_stats(struct sock *sk)
428 {
429 	struct inet_connection_sock *icsk = inet_csk(sk);
430 	struct tcp_sock *tp = tcp_sk(sk);
431 
432 	if (!icsk->icsk_retransmits) {
433 		tp->total_rto_recoveries++;
434 		tp->rto_stamp = tcp_time_stamp_ms(tp);
435 	}
436 	icsk->icsk_retransmits++;
437 	tp->total_rto++;
438 }
439 
440 /*
441  *	Timer for Fast Open socket to retransmit SYNACK. Note that the
442  *	sk here is the child socket, not the parent (listener) socket.
443  */
444 static void tcp_fastopen_synack_timer(struct sock *sk, struct request_sock *req)
445 {
446 	struct inet_connection_sock *icsk = inet_csk(sk);
447 	struct tcp_sock *tp = tcp_sk(sk);
448 	int max_retries;
449 
450 	req->rsk_ops->syn_ack_timeout(req);
451 
452 	/* Add one more retry for fastopen.
453 	 * Paired with WRITE_ONCE() in tcp_sock_set_syncnt()
454 	 */
455 	max_retries = READ_ONCE(icsk->icsk_syn_retries) ? :
456 		READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_synack_retries) + 1;
457 
458 	if (req->num_timeout >= max_retries) {
459 		tcp_write_err(sk);
460 		return;
461 	}
462 	/* Lower cwnd after certain SYNACK timeout like tcp_init_transfer() */
463 	if (icsk->icsk_retransmits == 1)
464 		tcp_enter_loss(sk);
465 	/* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
466 	 * returned from rtx_syn_ack() to make it more persistent like
467 	 * regular retransmit because if the child socket has been accepted
468 	 * it's not good to give up too easily.
469 	 */
470 	inet_rtx_syn_ack(sk, req);
471 	req->num_timeout++;
472 	tcp_update_rto_stats(sk);
473 	if (!tp->retrans_stamp)
474 		tp->retrans_stamp = tcp_time_stamp_ts(tp);
475 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
476 			  req->timeout << req->num_timeout, TCP_RTO_MAX);
477 }
478 
479 static bool tcp_rtx_probe0_timed_out(const struct sock *sk,
480 				     const struct sk_buff *skb,
481 				     u32 rtx_delta)
482 {
483 	const struct inet_connection_sock *icsk = inet_csk(sk);
484 	u32 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
485 	const struct tcp_sock *tp = tcp_sk(sk);
486 	int timeout = TCP_RTO_MAX * 2;
487 	s32 rcv_delta;
488 
489 	if (user_timeout) {
490 		/* If user application specified a TCP_USER_TIMEOUT,
491 		 * it does not want win 0 packets to 'reset the timer'
492 		 * while retransmits are not making progress.
493 		 */
494 		if (rtx_delta > user_timeout)
495 			return true;
496 		timeout = min_t(u32, timeout, msecs_to_jiffies(user_timeout));
497 	}
498 	/* Note: timer interrupt might have been delayed by at least one jiffy,
499 	 * and tp->rcv_tstamp might very well have been written recently.
500 	 * rcv_delta can thus be negative.
501 	 */
502 	rcv_delta = icsk->icsk_timeout - tp->rcv_tstamp;
503 	if (rcv_delta <= timeout)
504 		return false;
505 
506 	return msecs_to_jiffies(rtx_delta) > timeout;
507 }
508 
509 /**
510  *  tcp_retransmit_timer() - The TCP retransmit timeout handler
511  *  @sk:  Pointer to the current socket.
512  *
513  *  This function gets called when the kernel timer for a TCP packet
514  *  of this socket expires.
515  *
516  *  It handles retransmission, timer adjustment and other necessary measures.
517  *
518  *  Returns: Nothing (void)
519  */
520 void tcp_retransmit_timer(struct sock *sk)
521 {
522 	struct tcp_sock *tp = tcp_sk(sk);
523 	struct net *net = sock_net(sk);
524 	struct inet_connection_sock *icsk = inet_csk(sk);
525 	struct request_sock *req;
526 	struct sk_buff *skb;
527 
528 	req = rcu_dereference_protected(tp->fastopen_rsk,
529 					lockdep_sock_is_held(sk));
530 	if (req) {
531 		WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
532 			     sk->sk_state != TCP_FIN_WAIT1);
533 		tcp_fastopen_synack_timer(sk, req);
534 		/* Before we receive ACK to our SYN-ACK don't retransmit
535 		 * anything else (e.g., data or FIN segments).
536 		 */
537 		return;
538 	}
539 
540 	if (!tp->packets_out)
541 		return;
542 
543 	skb = tcp_rtx_queue_head(sk);
544 	if (WARN_ON_ONCE(!skb))
545 		return;
546 
547 	if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
548 	    !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
549 		/* Receiver dastardly shrinks window. Our retransmits
550 		 * become zero probes, but we should not timeout this
551 		 * connection. If the socket is an orphan, time it out,
552 		 * we cannot allow such beasts to hang infinitely.
553 		 */
554 		struct inet_sock *inet = inet_sk(sk);
555 		u32 rtx_delta;
556 
557 		rtx_delta = tcp_time_stamp_ts(tp) - (tp->retrans_stamp ?:
558 				tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb));
559 		if (tp->tcp_usec_ts)
560 			rtx_delta /= USEC_PER_MSEC;
561 
562 		if (sk->sk_family == AF_INET) {
563 			net_dbg_ratelimited("Probing zero-window on %pI4:%u/%u, seq=%u:%u, recv %ums ago, lasting %ums\n",
564 				&inet->inet_daddr, ntohs(inet->inet_dport),
565 				inet->inet_num, tp->snd_una, tp->snd_nxt,
566 				jiffies_to_msecs(jiffies - tp->rcv_tstamp),
567 				rtx_delta);
568 		}
569 #if IS_ENABLED(CONFIG_IPV6)
570 		else if (sk->sk_family == AF_INET6) {
571 			net_dbg_ratelimited("Probing zero-window on %pI6:%u/%u, seq=%u:%u, recv %ums ago, lasting %ums\n",
572 				&sk->sk_v6_daddr, ntohs(inet->inet_dport),
573 				inet->inet_num, tp->snd_una, tp->snd_nxt,
574 				jiffies_to_msecs(jiffies - tp->rcv_tstamp),
575 				rtx_delta);
576 		}
577 #endif
578 		if (tcp_rtx_probe0_timed_out(sk, skb, rtx_delta)) {
579 			tcp_write_err(sk);
580 			goto out;
581 		}
582 		tcp_enter_loss(sk);
583 		tcp_retransmit_skb(sk, skb, 1);
584 		__sk_dst_reset(sk);
585 		goto out_reset_timer;
586 	}
587 
588 	__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEOUTS);
589 	if (tcp_write_timeout(sk))
590 		goto out;
591 
592 	if (icsk->icsk_retransmits == 0) {
593 		int mib_idx = 0;
594 
595 		if (icsk->icsk_ca_state == TCP_CA_Recovery) {
596 			if (tcp_is_sack(tp))
597 				mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
598 			else
599 				mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
600 		} else if (icsk->icsk_ca_state == TCP_CA_Loss) {
601 			mib_idx = LINUX_MIB_TCPLOSSFAILURES;
602 		} else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
603 			   tp->sacked_out) {
604 			if (tcp_is_sack(tp))
605 				mib_idx = LINUX_MIB_TCPSACKFAILURES;
606 			else
607 				mib_idx = LINUX_MIB_TCPRENOFAILURES;
608 		}
609 		if (mib_idx)
610 			__NET_INC_STATS(sock_net(sk), mib_idx);
611 	}
612 
613 	tcp_enter_loss(sk);
614 
615 	tcp_update_rto_stats(sk);
616 	if (tcp_retransmit_skb(sk, tcp_rtx_queue_head(sk), 1) > 0) {
617 		/* Retransmission failed because of local congestion,
618 		 * Let senders fight for local resources conservatively.
619 		 */
620 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
621 					  TCP_RESOURCE_PROBE_INTERVAL,
622 					  TCP_RTO_MAX);
623 		goto out;
624 	}
625 
626 	/* Increase the timeout each time we retransmit.  Note that
627 	 * we do not increase the rtt estimate.  rto is initialized
628 	 * from rtt, but increases here.  Jacobson (SIGCOMM 88) suggests
629 	 * that doubling rto each time is the least we can get away with.
630 	 * In KA9Q, Karn uses this for the first few times, and then
631 	 * goes to quadratic.  netBSD doubles, but only goes up to *64,
632 	 * and clamps at 1 to 64 sec afterwards.  Note that 120 sec is
633 	 * defined in the protocol as the maximum possible RTT.  I guess
634 	 * we'll have to use something other than TCP to talk to the
635 	 * University of Mars.
636 	 *
637 	 * PAWS allows us longer timeouts and large windows, so once
638 	 * implemented ftp to mars will work nicely. We will have to fix
639 	 * the 120 second clamps though!
640 	 */
641 
642 out_reset_timer:
643 	/* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
644 	 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
645 	 * might be increased if the stream oscillates between thin and thick,
646 	 * thus the old value might already be too high compared to the value
647 	 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
648 	 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
649 	 * exponential backoff behaviour to avoid continue hammering
650 	 * linear-timeout retransmissions into a black hole
651 	 */
652 	if (sk->sk_state == TCP_ESTABLISHED &&
653 	    (tp->thin_lto || READ_ONCE(net->ipv4.sysctl_tcp_thin_linear_timeouts)) &&
654 	    tcp_stream_is_thin(tp) &&
655 	    icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
656 		icsk->icsk_backoff = 0;
657 		icsk->icsk_rto = clamp(__tcp_set_rto(tp),
658 				       tcp_rto_min(sk),
659 				       TCP_RTO_MAX);
660 	} else if (sk->sk_state != TCP_SYN_SENT ||
661 		   tp->total_rto >
662 		   READ_ONCE(net->ipv4.sysctl_tcp_syn_linear_timeouts)) {
663 		/* Use normal (exponential) backoff unless linear timeouts are
664 		 * activated.
665 		 */
666 		icsk->icsk_backoff++;
667 		icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
668 	}
669 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
670 				  tcp_clamp_rto_to_user_timeout(sk), TCP_RTO_MAX);
671 	if (retransmits_timed_out(sk, READ_ONCE(net->ipv4.sysctl_tcp_retries1) + 1, 0))
672 		__sk_dst_reset(sk);
673 
674 out:;
675 }
676 
677 /* Called with bottom-half processing disabled.
678    Called by tcp_write_timer() */
679 void tcp_write_timer_handler(struct sock *sk)
680 {
681 	struct inet_connection_sock *icsk = inet_csk(sk);
682 	int event;
683 
684 	if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
685 	    !icsk->icsk_pending)
686 		return;
687 
688 	if (time_after(icsk->icsk_timeout, jiffies)) {
689 		sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
690 		return;
691 	}
692 
693 	tcp_mstamp_refresh(tcp_sk(sk));
694 	event = icsk->icsk_pending;
695 
696 	switch (event) {
697 	case ICSK_TIME_REO_TIMEOUT:
698 		tcp_rack_reo_timeout(sk);
699 		break;
700 	case ICSK_TIME_LOSS_PROBE:
701 		tcp_send_loss_probe(sk);
702 		break;
703 	case ICSK_TIME_RETRANS:
704 		icsk->icsk_pending = 0;
705 		tcp_retransmit_timer(sk);
706 		break;
707 	case ICSK_TIME_PROBE0:
708 		icsk->icsk_pending = 0;
709 		tcp_probe_timer(sk);
710 		break;
711 	}
712 }
713 
714 static void tcp_write_timer(struct timer_list *t)
715 {
716 	struct inet_connection_sock *icsk =
717 			from_timer(icsk, t, icsk_retransmit_timer);
718 	struct sock *sk = &icsk->icsk_inet.sk;
719 
720 	bh_lock_sock(sk);
721 	if (!sock_owned_by_user(sk)) {
722 		tcp_write_timer_handler(sk);
723 	} else {
724 		/* delegate our work to tcp_release_cb() */
725 		if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &sk->sk_tsq_flags))
726 			sock_hold(sk);
727 	}
728 	bh_unlock_sock(sk);
729 	sock_put(sk);
730 }
731 
732 void tcp_syn_ack_timeout(const struct request_sock *req)
733 {
734 	struct net *net = read_pnet(&inet_rsk(req)->ireq_net);
735 
736 	__NET_INC_STATS(net, LINUX_MIB_TCPTIMEOUTS);
737 }
738 EXPORT_SYMBOL(tcp_syn_ack_timeout);
739 
740 void tcp_set_keepalive(struct sock *sk, int val)
741 {
742 	if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
743 		return;
744 
745 	if (val && !sock_flag(sk, SOCK_KEEPOPEN))
746 		inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
747 	else if (!val)
748 		inet_csk_delete_keepalive_timer(sk);
749 }
750 EXPORT_SYMBOL_GPL(tcp_set_keepalive);
751 
752 
753 static void tcp_keepalive_timer (struct timer_list *t)
754 {
755 	struct sock *sk = from_timer(sk, t, sk_timer);
756 	struct inet_connection_sock *icsk = inet_csk(sk);
757 	struct tcp_sock *tp = tcp_sk(sk);
758 	u32 elapsed;
759 
760 	/* Only process if socket is not in use. */
761 	bh_lock_sock(sk);
762 	if (sock_owned_by_user(sk)) {
763 		/* Try again later. */
764 		inet_csk_reset_keepalive_timer (sk, HZ/20);
765 		goto out;
766 	}
767 
768 	if (sk->sk_state == TCP_LISTEN) {
769 		pr_err("Hmm... keepalive on a LISTEN ???\n");
770 		goto out;
771 	}
772 
773 	tcp_mstamp_refresh(tp);
774 	if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
775 		if (READ_ONCE(tp->linger2) >= 0) {
776 			const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
777 
778 			if (tmo > 0) {
779 				tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
780 				goto out;
781 			}
782 		}
783 		tcp_send_active_reset(sk, GFP_ATOMIC, SK_RST_REASON_TCP_STATE);
784 		goto death;
785 	}
786 
787 	if (!sock_flag(sk, SOCK_KEEPOPEN) ||
788 	    ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_SYN_SENT)))
789 		goto out;
790 
791 	elapsed = keepalive_time_when(tp);
792 
793 	/* It is alive without keepalive 8) */
794 	if (tp->packets_out || !tcp_write_queue_empty(sk))
795 		goto resched;
796 
797 	elapsed = keepalive_time_elapsed(tp);
798 
799 	if (elapsed >= keepalive_time_when(tp)) {
800 		u32 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
801 
802 		/* If the TCP_USER_TIMEOUT option is enabled, use that
803 		 * to determine when to timeout instead.
804 		 */
805 		if ((user_timeout != 0 &&
806 		    elapsed >= msecs_to_jiffies(user_timeout) &&
807 		    icsk->icsk_probes_out > 0) ||
808 		    (user_timeout == 0 &&
809 		    icsk->icsk_probes_out >= keepalive_probes(tp))) {
810 			tcp_send_active_reset(sk, GFP_ATOMIC,
811 					      SK_RST_REASON_TCP_KEEPALIVE_TIMEOUT);
812 			tcp_write_err(sk);
813 			goto out;
814 		}
815 		if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) {
816 			icsk->icsk_probes_out++;
817 			elapsed = keepalive_intvl_when(tp);
818 		} else {
819 			/* If keepalive was lost due to local congestion,
820 			 * try harder.
821 			 */
822 			elapsed = TCP_RESOURCE_PROBE_INTERVAL;
823 		}
824 	} else {
825 		/* It is tp->rcv_tstamp + keepalive_time_when(tp) */
826 		elapsed = keepalive_time_when(tp) - elapsed;
827 	}
828 
829 resched:
830 	inet_csk_reset_keepalive_timer (sk, elapsed);
831 	goto out;
832 
833 death:
834 	tcp_done(sk);
835 
836 out:
837 	bh_unlock_sock(sk);
838 	sock_put(sk);
839 }
840 
841 static enum hrtimer_restart tcp_compressed_ack_kick(struct hrtimer *timer)
842 {
843 	struct tcp_sock *tp = container_of(timer, struct tcp_sock, compressed_ack_timer);
844 	struct sock *sk = (struct sock *)tp;
845 
846 	bh_lock_sock(sk);
847 	if (!sock_owned_by_user(sk)) {
848 		if (tp->compressed_ack) {
849 			/* Since we have to send one ack finally,
850 			 * subtract one from tp->compressed_ack to keep
851 			 * LINUX_MIB_TCPACKCOMPRESSED accurate.
852 			 */
853 			tp->compressed_ack--;
854 			tcp_send_ack(sk);
855 		}
856 	} else {
857 		if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED,
858 				      &sk->sk_tsq_flags))
859 			sock_hold(sk);
860 	}
861 	bh_unlock_sock(sk);
862 
863 	sock_put(sk);
864 
865 	return HRTIMER_NORESTART;
866 }
867 
868 void tcp_init_xmit_timers(struct sock *sk)
869 {
870 	inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
871 				  &tcp_keepalive_timer);
872 	hrtimer_init(&tcp_sk(sk)->pacing_timer, CLOCK_MONOTONIC,
873 		     HRTIMER_MODE_ABS_PINNED_SOFT);
874 	tcp_sk(sk)->pacing_timer.function = tcp_pace_kick;
875 
876 	hrtimer_init(&tcp_sk(sk)->compressed_ack_timer, CLOCK_MONOTONIC,
877 		     HRTIMER_MODE_REL_PINNED_SOFT);
878 	tcp_sk(sk)->compressed_ack_timer.function = tcp_compressed_ack_kick;
879 }
880