xref: /openbmc/linux/net/ipv4/tcp_timer.c (revision f4356947)
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;
30 	s32 remaining;
31 
32 	start_ts = tcp_sk(sk)->retrans_stamp;
33 	if (!icsk->icsk_user_timeout)
34 		return icsk->icsk_rto;
35 	elapsed = tcp_time_stamp(tcp_sk(sk)) - start_ts;
36 	remaining = icsk->icsk_user_timeout - elapsed;
37 	if (remaining <= 0)
38 		return 1; /* user timeout has passed; fire ASAP */
39 
40 	return min_t(u32, icsk->icsk_rto, msecs_to_jiffies(remaining));
41 }
42 
43 u32 tcp_clamp_probe0_to_user_timeout(const struct sock *sk, u32 when)
44 {
45 	struct inet_connection_sock *icsk = inet_csk(sk);
46 	u32 remaining;
47 	s32 elapsed;
48 
49 	if (!icsk->icsk_user_timeout || !icsk->icsk_probes_tstamp)
50 		return when;
51 
52 	elapsed = tcp_jiffies32 - icsk->icsk_probes_tstamp;
53 	if (unlikely(elapsed < 0))
54 		elapsed = 0;
55 	remaining = msecs_to_jiffies(icsk->icsk_user_timeout) - elapsed;
56 	remaining = max_t(u32, remaining, TCP_TIMEOUT_MIN);
57 
58 	return min_t(u32, remaining, when);
59 }
60 
61 /**
62  *  tcp_write_err() - close socket and save error info
63  *  @sk:  The socket the error has appeared on.
64  *
65  *  Returns: Nothing (void)
66  */
67 
68 static void tcp_write_err(struct sock *sk)
69 {
70 	WRITE_ONCE(sk->sk_err, READ_ONCE(sk->sk_err_soft) ? : ETIMEDOUT);
71 	sk_error_report(sk);
72 
73 	tcp_write_queue_purge(sk);
74 	tcp_done(sk);
75 	__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT);
76 }
77 
78 /**
79  *  tcp_out_of_resources() - Close socket if out of resources
80  *  @sk:        pointer to current socket
81  *  @do_reset:  send a last packet with reset flag
82  *
83  *  Do not allow orphaned sockets to eat all our resources.
84  *  This is direct violation of TCP specs, but it is required
85  *  to prevent DoS attacks. It is called when a retransmission timeout
86  *  or zero probe timeout occurs on orphaned socket.
87  *
88  *  Also close if our net namespace is exiting; in that case there is no
89  *  hope of ever communicating again since all netns interfaces are already
90  *  down (or about to be down), and we need to release our dst references,
91  *  which have been moved to the netns loopback interface, so the namespace
92  *  can finish exiting.  This condition is only possible if we are a kernel
93  *  socket, as those do not hold references to the namespace.
94  *
95  *  Criteria is still not confirmed experimentally and may change.
96  *  We kill the socket, if:
97  *  1. If number of orphaned sockets exceeds an administratively configured
98  *     limit.
99  *  2. If we have strong memory pressure.
100  *  3. If our net namespace is exiting.
101  */
102 static int tcp_out_of_resources(struct sock *sk, bool do_reset)
103 {
104 	struct tcp_sock *tp = tcp_sk(sk);
105 	int shift = 0;
106 
107 	/* If peer does not open window for long time, or did not transmit
108 	 * anything for long time, penalize it. */
109 	if ((s32)(tcp_jiffies32 - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset)
110 		shift++;
111 
112 	/* If some dubious ICMP arrived, penalize even more. */
113 	if (READ_ONCE(sk->sk_err_soft))
114 		shift++;
115 
116 	if (tcp_check_oom(sk, shift)) {
117 		/* Catch exceptional cases, when connection requires reset.
118 		 *      1. Last segment was sent recently. */
119 		if ((s32)(tcp_jiffies32 - tp->lsndtime) <= TCP_TIMEWAIT_LEN ||
120 		    /*  2. Window is closed. */
121 		    (!tp->snd_wnd && !tp->packets_out))
122 			do_reset = true;
123 		if (do_reset)
124 			tcp_send_active_reset(sk, GFP_ATOMIC);
125 		tcp_done(sk);
126 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY);
127 		return 1;
128 	}
129 
130 	if (!check_net(sock_net(sk))) {
131 		/* Not possible to send reset; just close */
132 		tcp_done(sk);
133 		return 1;
134 	}
135 
136 	return 0;
137 }
138 
139 /**
140  *  tcp_orphan_retries() - Returns maximal number of retries on an orphaned socket
141  *  @sk:    Pointer to the current socket.
142  *  @alive: bool, socket alive state
143  */
144 static int tcp_orphan_retries(struct sock *sk, bool alive)
145 {
146 	int retries = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_orphan_retries); /* May be zero. */
147 
148 	/* We know from an ICMP that something is wrong. */
149 	if (READ_ONCE(sk->sk_err_soft) && !alive)
150 		retries = 0;
151 
152 	/* However, if socket sent something recently, select some safe
153 	 * number of retries. 8 corresponds to >100 seconds with minimal
154 	 * RTO of 200msec. */
155 	if (retries == 0 && alive)
156 		retries = 8;
157 	return retries;
158 }
159 
160 static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk)
161 {
162 	const struct net *net = sock_net(sk);
163 	int mss;
164 
165 	/* Black hole detection */
166 	if (!READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing))
167 		return;
168 
169 	if (!icsk->icsk_mtup.enabled) {
170 		icsk->icsk_mtup.enabled = 1;
171 		icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
172 	} else {
173 		mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1;
174 		mss = min(READ_ONCE(net->ipv4.sysctl_tcp_base_mss), mss);
175 		mss = max(mss, READ_ONCE(net->ipv4.sysctl_tcp_mtu_probe_floor));
176 		mss = max(mss, READ_ONCE(net->ipv4.sysctl_tcp_min_snd_mss));
177 		icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
178 	}
179 	tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
180 }
181 
182 static unsigned int tcp_model_timeout(struct sock *sk,
183 				      unsigned int boundary,
184 				      unsigned int rto_base)
185 {
186 	unsigned int linear_backoff_thresh, timeout;
187 
188 	linear_backoff_thresh = ilog2(TCP_RTO_MAX / rto_base);
189 	if (boundary <= linear_backoff_thresh)
190 		timeout = ((2 << boundary) - 1) * rto_base;
191 	else
192 		timeout = ((2 << linear_backoff_thresh) - 1) * rto_base +
193 			(boundary - linear_backoff_thresh) * TCP_RTO_MAX;
194 	return jiffies_to_msecs(timeout);
195 }
196 /**
197  *  retransmits_timed_out() - returns true if this connection has timed out
198  *  @sk:       The current socket
199  *  @boundary: max number of retransmissions
200  *  @timeout:  A custom timeout value.
201  *             If set to 0 the default timeout is calculated and used.
202  *             Using TCP_RTO_MIN and the number of unsuccessful retransmits.
203  *
204  * The default "timeout" value this function can calculate and use
205  * is equivalent to the timeout of a TCP Connection
206  * after "boundary" unsuccessful, exponentially backed-off
207  * retransmissions with an initial RTO of TCP_RTO_MIN.
208  */
209 static bool retransmits_timed_out(struct sock *sk,
210 				  unsigned int boundary,
211 				  unsigned int timeout)
212 {
213 	unsigned int start_ts;
214 
215 	if (!inet_csk(sk)->icsk_retransmits)
216 		return false;
217 
218 	start_ts = tcp_sk(sk)->retrans_stamp;
219 	if (likely(timeout == 0)) {
220 		unsigned int rto_base = TCP_RTO_MIN;
221 
222 		if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
223 			rto_base = tcp_timeout_init(sk);
224 		timeout = tcp_model_timeout(sk, boundary, rto_base);
225 	}
226 
227 	return (s32)(tcp_time_stamp(tcp_sk(sk)) - start_ts - timeout) >= 0;
228 }
229 
230 /* A write timeout has occurred. Process the after effects. */
231 static int tcp_write_timeout(struct sock *sk)
232 {
233 	struct inet_connection_sock *icsk = inet_csk(sk);
234 	struct tcp_sock *tp = tcp_sk(sk);
235 	struct net *net = sock_net(sk);
236 	bool expired = false, do_reset;
237 	int retry_until, max_retransmits;
238 
239 	if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
240 		if (icsk->icsk_retransmits)
241 			__dst_negative_advice(sk);
242 		retry_until = icsk->icsk_syn_retries ? :
243 			READ_ONCE(net->ipv4.sysctl_tcp_syn_retries);
244 
245 		max_retransmits = retry_until;
246 		if (sk->sk_state == TCP_SYN_SENT)
247 			max_retransmits += READ_ONCE(net->ipv4.sysctl_tcp_syn_linear_timeouts);
248 
249 		expired = icsk->icsk_retransmits >= max_retransmits;
250 	} else {
251 		if (retransmits_timed_out(sk, READ_ONCE(net->ipv4.sysctl_tcp_retries1), 0)) {
252 			/* Black hole detection */
253 			tcp_mtu_probing(icsk, sk);
254 
255 			__dst_negative_advice(sk);
256 		}
257 
258 		retry_until = READ_ONCE(net->ipv4.sysctl_tcp_retries2);
259 		if (sock_flag(sk, SOCK_DEAD)) {
260 			const bool alive = icsk->icsk_rto < TCP_RTO_MAX;
261 
262 			retry_until = tcp_orphan_retries(sk, alive);
263 			do_reset = alive ||
264 				!retransmits_timed_out(sk, retry_until, 0);
265 
266 			if (tcp_out_of_resources(sk, do_reset))
267 				return 1;
268 		}
269 	}
270 	if (!expired)
271 		expired = retransmits_timed_out(sk, retry_until,
272 						icsk->icsk_user_timeout);
273 	tcp_fastopen_active_detect_blackhole(sk, expired);
274 
275 	if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RTO_CB_FLAG))
276 		tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RTO_CB,
277 				  icsk->icsk_retransmits,
278 				  icsk->icsk_rto, (int)expired);
279 
280 	if (expired) {
281 		/* Has it gone just too far? */
282 		tcp_write_err(sk);
283 		return 1;
284 	}
285 
286 	if (sk_rethink_txhash(sk)) {
287 		tp->timeout_rehash++;
288 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEOUTREHASH);
289 	}
290 
291 	return 0;
292 }
293 
294 /* Called with BH disabled */
295 void tcp_delack_timer_handler(struct sock *sk)
296 {
297 	struct inet_connection_sock *icsk = inet_csk(sk);
298 
299 	if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
300 	    !(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
301 		return;
302 
303 	if (time_after(icsk->icsk_ack.timeout, jiffies)) {
304 		sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
305 		return;
306 	}
307 	icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
308 
309 	if (inet_csk_ack_scheduled(sk)) {
310 		if (!inet_csk_in_pingpong_mode(sk)) {
311 			/* Delayed ACK missed: inflate ATO. */
312 			icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto);
313 		} else {
314 			/* Delayed ACK missed: leave pingpong mode and
315 			 * deflate ATO.
316 			 */
317 			inet_csk_exit_pingpong_mode(sk);
318 			icsk->icsk_ack.ato      = TCP_ATO_MIN;
319 		}
320 		tcp_mstamp_refresh(tcp_sk(sk));
321 		tcp_send_ack(sk);
322 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKS);
323 	}
324 }
325 
326 
327 /**
328  *  tcp_delack_timer() - The TCP delayed ACK timeout handler
329  *  @t:  Pointer to the timer. (gets casted to struct sock *)
330  *
331  *  This function gets (indirectly) called when the kernel timer for a TCP packet
332  *  of this socket expires. Calls tcp_delack_timer_handler() to do the actual work.
333  *
334  *  Returns: Nothing (void)
335  */
336 static void tcp_delack_timer(struct timer_list *t)
337 {
338 	struct inet_connection_sock *icsk =
339 			from_timer(icsk, t, icsk_delack_timer);
340 	struct sock *sk = &icsk->icsk_inet.sk;
341 
342 	bh_lock_sock(sk);
343 	if (!sock_owned_by_user(sk)) {
344 		tcp_delack_timer_handler(sk);
345 	} else {
346 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
347 		/* deleguate our work to tcp_release_cb() */
348 		if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &sk->sk_tsq_flags))
349 			sock_hold(sk);
350 	}
351 	bh_unlock_sock(sk);
352 	sock_put(sk);
353 }
354 
355 static void tcp_probe_timer(struct sock *sk)
356 {
357 	struct inet_connection_sock *icsk = inet_csk(sk);
358 	struct sk_buff *skb = tcp_send_head(sk);
359 	struct tcp_sock *tp = tcp_sk(sk);
360 	int max_probes;
361 
362 	if (tp->packets_out || !skb) {
363 		icsk->icsk_probes_out = 0;
364 		icsk->icsk_probes_tstamp = 0;
365 		return;
366 	}
367 
368 	/* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as
369 	 * long as the receiver continues to respond probes. We support this by
370 	 * default and reset icsk_probes_out with incoming ACKs. But if the
371 	 * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we
372 	 * kill the socket when the retry count and the time exceeds the
373 	 * corresponding system limit. We also implement similar policy when
374 	 * we use RTO to probe window in tcp_retransmit_timer().
375 	 */
376 	if (!icsk->icsk_probes_tstamp)
377 		icsk->icsk_probes_tstamp = tcp_jiffies32;
378 	else if (icsk->icsk_user_timeout &&
379 		 (s32)(tcp_jiffies32 - icsk->icsk_probes_tstamp) >=
380 		 msecs_to_jiffies(icsk->icsk_user_timeout))
381 		goto abort;
382 
383 	max_probes = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retries2);
384 	if (sock_flag(sk, SOCK_DEAD)) {
385 		const bool alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX;
386 
387 		max_probes = tcp_orphan_retries(sk, alive);
388 		if (!alive && icsk->icsk_backoff >= max_probes)
389 			goto abort;
390 		if (tcp_out_of_resources(sk, true))
391 			return;
392 	}
393 
394 	if (icsk->icsk_probes_out >= max_probes) {
395 abort:		tcp_write_err(sk);
396 	} else {
397 		/* Only send another probe if we didn't close things up. */
398 		tcp_send_probe0(sk);
399 	}
400 }
401 
402 /*
403  *	Timer for Fast Open socket to retransmit SYNACK. Note that the
404  *	sk here is the child socket, not the parent (listener) socket.
405  */
406 static void tcp_fastopen_synack_timer(struct sock *sk, struct request_sock *req)
407 {
408 	struct inet_connection_sock *icsk = inet_csk(sk);
409 	struct tcp_sock *tp = tcp_sk(sk);
410 	int max_retries;
411 
412 	req->rsk_ops->syn_ack_timeout(req);
413 
414 	/* add one more retry for fastopen */
415 	max_retries = icsk->icsk_syn_retries ? :
416 		READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_synack_retries) + 1;
417 
418 	if (req->num_timeout >= max_retries) {
419 		tcp_write_err(sk);
420 		return;
421 	}
422 	/* Lower cwnd after certain SYNACK timeout like tcp_init_transfer() */
423 	if (icsk->icsk_retransmits == 1)
424 		tcp_enter_loss(sk);
425 	/* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
426 	 * returned from rtx_syn_ack() to make it more persistent like
427 	 * regular retransmit because if the child socket has been accepted
428 	 * it's not good to give up too easily.
429 	 */
430 	inet_rtx_syn_ack(sk, req);
431 	req->num_timeout++;
432 	icsk->icsk_retransmits++;
433 	if (!tp->retrans_stamp)
434 		tp->retrans_stamp = tcp_time_stamp(tp);
435 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
436 			  req->timeout << req->num_timeout, TCP_RTO_MAX);
437 }
438 
439 
440 /**
441  *  tcp_retransmit_timer() - The TCP retransmit timeout handler
442  *  @sk:  Pointer to the current socket.
443  *
444  *  This function gets called when the kernel timer for a TCP packet
445  *  of this socket expires.
446  *
447  *  It handles retransmission, timer adjustment and other necessary measures.
448  *
449  *  Returns: Nothing (void)
450  */
451 void tcp_retransmit_timer(struct sock *sk)
452 {
453 	struct tcp_sock *tp = tcp_sk(sk);
454 	struct net *net = sock_net(sk);
455 	struct inet_connection_sock *icsk = inet_csk(sk);
456 	struct request_sock *req;
457 	struct sk_buff *skb;
458 
459 	req = rcu_dereference_protected(tp->fastopen_rsk,
460 					lockdep_sock_is_held(sk));
461 	if (req) {
462 		WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
463 			     sk->sk_state != TCP_FIN_WAIT1);
464 		tcp_fastopen_synack_timer(sk, req);
465 		/* Before we receive ACK to our SYN-ACK don't retransmit
466 		 * anything else (e.g., data or FIN segments).
467 		 */
468 		return;
469 	}
470 
471 	if (!tp->packets_out)
472 		return;
473 
474 	skb = tcp_rtx_queue_head(sk);
475 	if (WARN_ON_ONCE(!skb))
476 		return;
477 
478 	tp->tlp_high_seq = 0;
479 
480 	if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
481 	    !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
482 		/* Receiver dastardly shrinks window. Our retransmits
483 		 * become zero probes, but we should not timeout this
484 		 * connection. If the socket is an orphan, time it out,
485 		 * we cannot allow such beasts to hang infinitely.
486 		 */
487 		struct inet_sock *inet = inet_sk(sk);
488 		if (sk->sk_family == AF_INET) {
489 			net_dbg_ratelimited("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
490 					    &inet->inet_daddr,
491 					    ntohs(inet->inet_dport),
492 					    inet->inet_num,
493 					    tp->snd_una, tp->snd_nxt);
494 		}
495 #if IS_ENABLED(CONFIG_IPV6)
496 		else if (sk->sk_family == AF_INET6) {
497 			net_dbg_ratelimited("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
498 					    &sk->sk_v6_daddr,
499 					    ntohs(inet->inet_dport),
500 					    inet->inet_num,
501 					    tp->snd_una, tp->snd_nxt);
502 		}
503 #endif
504 		if (tcp_jiffies32 - tp->rcv_tstamp > TCP_RTO_MAX) {
505 			tcp_write_err(sk);
506 			goto out;
507 		}
508 		tcp_enter_loss(sk);
509 		tcp_retransmit_skb(sk, skb, 1);
510 		__sk_dst_reset(sk);
511 		goto out_reset_timer;
512 	}
513 
514 	__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEOUTS);
515 	if (tcp_write_timeout(sk))
516 		goto out;
517 
518 	if (icsk->icsk_retransmits == 0) {
519 		int mib_idx = 0;
520 
521 		if (icsk->icsk_ca_state == TCP_CA_Recovery) {
522 			if (tcp_is_sack(tp))
523 				mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
524 			else
525 				mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
526 		} else if (icsk->icsk_ca_state == TCP_CA_Loss) {
527 			mib_idx = LINUX_MIB_TCPLOSSFAILURES;
528 		} else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
529 			   tp->sacked_out) {
530 			if (tcp_is_sack(tp))
531 				mib_idx = LINUX_MIB_TCPSACKFAILURES;
532 			else
533 				mib_idx = LINUX_MIB_TCPRENOFAILURES;
534 		}
535 		if (mib_idx)
536 			__NET_INC_STATS(sock_net(sk), mib_idx);
537 	}
538 
539 	tcp_enter_loss(sk);
540 
541 	icsk->icsk_retransmits++;
542 	if (tcp_retransmit_skb(sk, tcp_rtx_queue_head(sk), 1) > 0) {
543 		/* Retransmission failed because of local congestion,
544 		 * Let senders fight for local resources conservatively.
545 		 */
546 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
547 					  TCP_RESOURCE_PROBE_INTERVAL,
548 					  TCP_RTO_MAX);
549 		goto out;
550 	}
551 
552 	/* Increase the timeout each time we retransmit.  Note that
553 	 * we do not increase the rtt estimate.  rto is initialized
554 	 * from rtt, but increases here.  Jacobson (SIGCOMM 88) suggests
555 	 * that doubling rto each time is the least we can get away with.
556 	 * In KA9Q, Karn uses this for the first few times, and then
557 	 * goes to quadratic.  netBSD doubles, but only goes up to *64,
558 	 * and clamps at 1 to 64 sec afterwards.  Note that 120 sec is
559 	 * defined in the protocol as the maximum possible RTT.  I guess
560 	 * we'll have to use something other than TCP to talk to the
561 	 * University of Mars.
562 	 *
563 	 * PAWS allows us longer timeouts and large windows, so once
564 	 * implemented ftp to mars will work nicely. We will have to fix
565 	 * the 120 second clamps though!
566 	 */
567 	icsk->icsk_backoff++;
568 
569 out_reset_timer:
570 	/* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
571 	 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
572 	 * might be increased if the stream oscillates between thin and thick,
573 	 * thus the old value might already be too high compared to the value
574 	 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
575 	 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
576 	 * exponential backoff behaviour to avoid continue hammering
577 	 * linear-timeout retransmissions into a black hole
578 	 */
579 	if (sk->sk_state == TCP_ESTABLISHED &&
580 	    (tp->thin_lto || READ_ONCE(net->ipv4.sysctl_tcp_thin_linear_timeouts)) &&
581 	    tcp_stream_is_thin(tp) &&
582 	    icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
583 		icsk->icsk_backoff = 0;
584 		icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX);
585 	} else if (sk->sk_state != TCP_SYN_SENT ||
586 		   icsk->icsk_backoff >
587 		   READ_ONCE(net->ipv4.sysctl_tcp_syn_linear_timeouts)) {
588 		/* Use normal (exponential) backoff unless linear timeouts are
589 		 * activated.
590 		 */
591 		icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
592 	}
593 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
594 				  tcp_clamp_rto_to_user_timeout(sk), TCP_RTO_MAX);
595 	if (retransmits_timed_out(sk, READ_ONCE(net->ipv4.sysctl_tcp_retries1) + 1, 0))
596 		__sk_dst_reset(sk);
597 
598 out:;
599 }
600 
601 /* Called with bottom-half processing disabled.
602    Called by tcp_write_timer() */
603 void tcp_write_timer_handler(struct sock *sk)
604 {
605 	struct inet_connection_sock *icsk = inet_csk(sk);
606 	int event;
607 
608 	if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
609 	    !icsk->icsk_pending)
610 		return;
611 
612 	if (time_after(icsk->icsk_timeout, jiffies)) {
613 		sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
614 		return;
615 	}
616 
617 	tcp_mstamp_refresh(tcp_sk(sk));
618 	event = icsk->icsk_pending;
619 
620 	switch (event) {
621 	case ICSK_TIME_REO_TIMEOUT:
622 		tcp_rack_reo_timeout(sk);
623 		break;
624 	case ICSK_TIME_LOSS_PROBE:
625 		tcp_send_loss_probe(sk);
626 		break;
627 	case ICSK_TIME_RETRANS:
628 		icsk->icsk_pending = 0;
629 		tcp_retransmit_timer(sk);
630 		break;
631 	case ICSK_TIME_PROBE0:
632 		icsk->icsk_pending = 0;
633 		tcp_probe_timer(sk);
634 		break;
635 	}
636 }
637 
638 static void tcp_write_timer(struct timer_list *t)
639 {
640 	struct inet_connection_sock *icsk =
641 			from_timer(icsk, t, icsk_retransmit_timer);
642 	struct sock *sk = &icsk->icsk_inet.sk;
643 
644 	bh_lock_sock(sk);
645 	if (!sock_owned_by_user(sk)) {
646 		tcp_write_timer_handler(sk);
647 	} else {
648 		/* delegate our work to tcp_release_cb() */
649 		if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &sk->sk_tsq_flags))
650 			sock_hold(sk);
651 	}
652 	bh_unlock_sock(sk);
653 	sock_put(sk);
654 }
655 
656 void tcp_syn_ack_timeout(const struct request_sock *req)
657 {
658 	struct net *net = read_pnet(&inet_rsk(req)->ireq_net);
659 
660 	__NET_INC_STATS(net, LINUX_MIB_TCPTIMEOUTS);
661 }
662 EXPORT_SYMBOL(tcp_syn_ack_timeout);
663 
664 void tcp_set_keepalive(struct sock *sk, int val)
665 {
666 	if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
667 		return;
668 
669 	if (val && !sock_flag(sk, SOCK_KEEPOPEN))
670 		inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
671 	else if (!val)
672 		inet_csk_delete_keepalive_timer(sk);
673 }
674 EXPORT_SYMBOL_GPL(tcp_set_keepalive);
675 
676 
677 static void tcp_keepalive_timer (struct timer_list *t)
678 {
679 	struct sock *sk = from_timer(sk, t, sk_timer);
680 	struct inet_connection_sock *icsk = inet_csk(sk);
681 	struct tcp_sock *tp = tcp_sk(sk);
682 	u32 elapsed;
683 
684 	/* Only process if socket is not in use. */
685 	bh_lock_sock(sk);
686 	if (sock_owned_by_user(sk)) {
687 		/* Try again later. */
688 		inet_csk_reset_keepalive_timer (sk, HZ/20);
689 		goto out;
690 	}
691 
692 	if (sk->sk_state == TCP_LISTEN) {
693 		pr_err("Hmm... keepalive on a LISTEN ???\n");
694 		goto out;
695 	}
696 
697 	tcp_mstamp_refresh(tp);
698 	if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
699 		if (tp->linger2 >= 0) {
700 			const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
701 
702 			if (tmo > 0) {
703 				tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
704 				goto out;
705 			}
706 		}
707 		tcp_send_active_reset(sk, GFP_ATOMIC);
708 		goto death;
709 	}
710 
711 	if (!sock_flag(sk, SOCK_KEEPOPEN) ||
712 	    ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_SYN_SENT)))
713 		goto out;
714 
715 	elapsed = keepalive_time_when(tp);
716 
717 	/* It is alive without keepalive 8) */
718 	if (tp->packets_out || !tcp_write_queue_empty(sk))
719 		goto resched;
720 
721 	elapsed = keepalive_time_elapsed(tp);
722 
723 	if (elapsed >= keepalive_time_when(tp)) {
724 		/* If the TCP_USER_TIMEOUT option is enabled, use that
725 		 * to determine when to timeout instead.
726 		 */
727 		if ((icsk->icsk_user_timeout != 0 &&
728 		    elapsed >= msecs_to_jiffies(icsk->icsk_user_timeout) &&
729 		    icsk->icsk_probes_out > 0) ||
730 		    (icsk->icsk_user_timeout == 0 &&
731 		    icsk->icsk_probes_out >= keepalive_probes(tp))) {
732 			tcp_send_active_reset(sk, GFP_ATOMIC);
733 			tcp_write_err(sk);
734 			goto out;
735 		}
736 		if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) {
737 			icsk->icsk_probes_out++;
738 			elapsed = keepalive_intvl_when(tp);
739 		} else {
740 			/* If keepalive was lost due to local congestion,
741 			 * try harder.
742 			 */
743 			elapsed = TCP_RESOURCE_PROBE_INTERVAL;
744 		}
745 	} else {
746 		/* It is tp->rcv_tstamp + keepalive_time_when(tp) */
747 		elapsed = keepalive_time_when(tp) - elapsed;
748 	}
749 
750 resched:
751 	inet_csk_reset_keepalive_timer (sk, elapsed);
752 	goto out;
753 
754 death:
755 	tcp_done(sk);
756 
757 out:
758 	bh_unlock_sock(sk);
759 	sock_put(sk);
760 }
761 
762 static enum hrtimer_restart tcp_compressed_ack_kick(struct hrtimer *timer)
763 {
764 	struct tcp_sock *tp = container_of(timer, struct tcp_sock, compressed_ack_timer);
765 	struct sock *sk = (struct sock *)tp;
766 
767 	bh_lock_sock(sk);
768 	if (!sock_owned_by_user(sk)) {
769 		if (tp->compressed_ack) {
770 			/* Since we have to send one ack finally,
771 			 * subtract one from tp->compressed_ack to keep
772 			 * LINUX_MIB_TCPACKCOMPRESSED accurate.
773 			 */
774 			tp->compressed_ack--;
775 			tcp_send_ack(sk);
776 		}
777 	} else {
778 		if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED,
779 				      &sk->sk_tsq_flags))
780 			sock_hold(sk);
781 	}
782 	bh_unlock_sock(sk);
783 
784 	sock_put(sk);
785 
786 	return HRTIMER_NORESTART;
787 }
788 
789 void tcp_init_xmit_timers(struct sock *sk)
790 {
791 	inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
792 				  &tcp_keepalive_timer);
793 	hrtimer_init(&tcp_sk(sk)->pacing_timer, CLOCK_MONOTONIC,
794 		     HRTIMER_MODE_ABS_PINNED_SOFT);
795 	tcp_sk(sk)->pacing_timer.function = tcp_pace_kick;
796 
797 	hrtimer_init(&tcp_sk(sk)->compressed_ack_timer, CLOCK_MONOTONIC,
798 		     HRTIMER_MODE_REL_PINNED_SOFT);
799 	tcp_sk(sk)->compressed_ack_timer.function = tcp_compressed_ack_kick;
800 }
801