xref: /kernel/linux/linux-5.10/net/ipv4/tcp_timer.c (revision 8c2ecf20)
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
26static 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
43u32 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
68static void tcp_write_err(struct sock *sk)
69{
70	sk->sk_err = sk->sk_err_soft ? : ETIMEDOUT;
71	sk->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 */
102static 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 (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 */
144static 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 (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
160static 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
182static 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 */
209static 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. */
231static 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;
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		expired = icsk->icsk_retransmits >= retry_until;
245	} else {
246		if (retransmits_timed_out(sk, READ_ONCE(net->ipv4.sysctl_tcp_retries1), 0)) {
247			/* Black hole detection */
248			tcp_mtu_probing(icsk, sk);
249
250			__dst_negative_advice(sk);
251		}
252
253		retry_until = READ_ONCE(net->ipv4.sysctl_tcp_retries2);
254		if (sock_flag(sk, SOCK_DEAD)) {
255			const bool alive = icsk->icsk_rto < TCP_RTO_MAX;
256
257			retry_until = tcp_orphan_retries(sk, alive);
258			do_reset = alive ||
259				!retransmits_timed_out(sk, retry_until, 0);
260
261			if (tcp_out_of_resources(sk, do_reset))
262				return 1;
263		}
264	}
265	if (!expired)
266		expired = retransmits_timed_out(sk, retry_until,
267						icsk->icsk_user_timeout);
268	tcp_fastopen_active_detect_blackhole(sk, expired);
269
270	if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RTO_CB_FLAG))
271		tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RTO_CB,
272				  icsk->icsk_retransmits,
273				  icsk->icsk_rto, (int)expired);
274
275	if (expired) {
276		/* Has it gone just too far? */
277		tcp_write_err(sk);
278		return 1;
279	}
280
281	if (sk_rethink_txhash(sk)) {
282		tp->timeout_rehash++;
283		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEOUTREHASH);
284	}
285
286	return 0;
287}
288
289/* Called with BH disabled */
290void tcp_delack_timer_handler(struct sock *sk)
291{
292	struct inet_connection_sock *icsk = inet_csk(sk);
293
294	sk_mem_reclaim_partial(sk);
295
296	if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
297	    !(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
298		goto out;
299
300	if (time_after(icsk->icsk_ack.timeout, jiffies)) {
301		sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
302		goto out;
303	}
304	icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
305
306	if (inet_csk_ack_scheduled(sk)) {
307		if (!inet_csk_in_pingpong_mode(sk)) {
308			/* Delayed ACK missed: inflate ATO. */
309			icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto);
310		} else {
311			/* Delayed ACK missed: leave pingpong mode and
312			 * deflate ATO.
313			 */
314			inet_csk_exit_pingpong_mode(sk);
315			icsk->icsk_ack.ato      = TCP_ATO_MIN;
316		}
317		tcp_mstamp_refresh(tcp_sk(sk));
318		tcp_send_ack(sk);
319		__NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKS);
320	}
321
322out:
323	if (tcp_under_memory_pressure(sk))
324		sk_mem_reclaim(sk);
325}
326
327
328/**
329 *  tcp_delack_timer() - The TCP delayed ACK timeout handler
330 *  @t:  Pointer to the timer. (gets casted to struct sock *)
331 *
332 *  This function gets (indirectly) called when the kernel timer for a TCP packet
333 *  of this socket expires. Calls tcp_delack_timer_handler() to do the actual work.
334 *
335 *  Returns: Nothing (void)
336 */
337static void tcp_delack_timer(struct timer_list *t)
338{
339	struct inet_connection_sock *icsk =
340			from_timer(icsk, t, icsk_delack_timer);
341	struct sock *sk = &icsk->icsk_inet.sk;
342
343	bh_lock_sock(sk);
344	if (!sock_owned_by_user(sk)) {
345		tcp_delack_timer_handler(sk);
346	} else {
347		__NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
348		/* deleguate our work to tcp_release_cb() */
349		if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &sk->sk_tsq_flags))
350			sock_hold(sk);
351	}
352	bh_unlock_sock(sk);
353	sock_put(sk);
354}
355
356static void tcp_probe_timer(struct sock *sk)
357{
358	struct inet_connection_sock *icsk = inet_csk(sk);
359	struct sk_buff *skb = tcp_send_head(sk);
360	struct tcp_sock *tp = tcp_sk(sk);
361	int max_probes;
362
363	if (tp->packets_out || !skb) {
364		icsk->icsk_probes_out = 0;
365		icsk->icsk_probes_tstamp = 0;
366		return;
367	}
368
369	/* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as
370	 * long as the receiver continues to respond probes. We support this by
371	 * default and reset icsk_probes_out with incoming ACKs. But if the
372	 * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we
373	 * kill the socket when the retry count and the time exceeds the
374	 * corresponding system limit. We also implement similar policy when
375	 * we use RTO to probe window in tcp_retransmit_timer().
376	 */
377	if (!icsk->icsk_probes_tstamp)
378		icsk->icsk_probes_tstamp = tcp_jiffies32;
379	else if (icsk->icsk_user_timeout &&
380		 (s32)(tcp_jiffies32 - icsk->icsk_probes_tstamp) >=
381		 msecs_to_jiffies(icsk->icsk_user_timeout))
382		goto abort;
383
384	max_probes = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retries2);
385	if (sock_flag(sk, SOCK_DEAD)) {
386		const bool alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX;
387
388		max_probes = tcp_orphan_retries(sk, alive);
389		if (!alive && icsk->icsk_backoff >= max_probes)
390			goto abort;
391		if (tcp_out_of_resources(sk, true))
392			return;
393	}
394
395	if (icsk->icsk_probes_out >= max_probes) {
396abort:		tcp_write_err(sk);
397	} else {
398		/* Only send another probe if we didn't close things up. */
399		tcp_send_probe0(sk);
400	}
401}
402
403/*
404 *	Timer for Fast Open socket to retransmit SYNACK. Note that the
405 *	sk here is the child socket, not the parent (listener) socket.
406 */
407static void tcp_fastopen_synack_timer(struct sock *sk, struct request_sock *req)
408{
409	struct inet_connection_sock *icsk = inet_csk(sk);
410	struct tcp_sock *tp = tcp_sk(sk);
411	int max_retries;
412
413	req->rsk_ops->syn_ack_timeout(req);
414
415	/* add one more retry for fastopen */
416	max_retries = icsk->icsk_syn_retries ? :
417		READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_synack_retries) + 1;
418
419	if (req->num_timeout >= max_retries) {
420		tcp_write_err(sk);
421		return;
422	}
423	/* Lower cwnd after certain SYNACK timeout like tcp_init_transfer() */
424	if (icsk->icsk_retransmits == 1)
425		tcp_enter_loss(sk);
426	/* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
427	 * returned from rtx_syn_ack() to make it more persistent like
428	 * regular retransmit because if the child socket has been accepted
429	 * it's not good to give up too easily.
430	 */
431	inet_rtx_syn_ack(sk, req);
432	req->num_timeout++;
433	icsk->icsk_retransmits++;
434	if (!tp->retrans_stamp)
435		tp->retrans_stamp = tcp_time_stamp(tp);
436	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
437			  TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
438}
439
440static bool tcp_rtx_probe0_timed_out(const struct sock *sk,
441				     const struct sk_buff *skb)
442{
443	const struct inet_connection_sock *icsk = inet_csk(sk);
444	u32 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
445	const struct tcp_sock *tp = tcp_sk(sk);
446	int timeout = TCP_RTO_MAX * 2;
447	u32 rtx_delta;
448	s32 rcv_delta;
449
450	rtx_delta = (u32)msecs_to_jiffies(tcp_time_stamp(tp) -
451			(tp->retrans_stamp ?: tcp_skb_timestamp(skb)));
452
453	if (user_timeout) {
454		/* If user application specified a TCP_USER_TIMEOUT,
455		 * it does not want win 0 packets to 'reset the timer'
456		 * while retransmits are not making progress.
457		 */
458		if (rtx_delta > user_timeout)
459			return true;
460		timeout = min_t(u32, timeout, msecs_to_jiffies(user_timeout));
461	}
462
463	/* Note: timer interrupt might have been delayed by at least one jiffy,
464	 * and tp->rcv_tstamp might very well have been written recently.
465	 * rcv_delta can thus be negative.
466	 */
467	rcv_delta = icsk->icsk_timeout - tp->rcv_tstamp;
468	if (rcv_delta <= timeout)
469		return false;
470
471	return rtx_delta > timeout;
472}
473
474/**
475 *  tcp_retransmit_timer() - The TCP retransmit timeout handler
476 *  @sk:  Pointer to the current socket.
477 *
478 *  This function gets called when the kernel timer for a TCP packet
479 *  of this socket expires.
480 *
481 *  It handles retransmission, timer adjustment and other necesarry measures.
482 *
483 *  Returns: Nothing (void)
484 */
485void tcp_retransmit_timer(struct sock *sk)
486{
487	struct tcp_sock *tp = tcp_sk(sk);
488	struct net *net = sock_net(sk);
489	struct inet_connection_sock *icsk = inet_csk(sk);
490	struct request_sock *req;
491	struct sk_buff *skb;
492
493	req = rcu_dereference_protected(tp->fastopen_rsk,
494					lockdep_sock_is_held(sk));
495	if (req) {
496		WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
497			     sk->sk_state != TCP_FIN_WAIT1);
498		tcp_fastopen_synack_timer(sk, req);
499		/* Before we receive ACK to our SYN-ACK don't retransmit
500		 * anything else (e.g., data or FIN segments).
501		 */
502		return;
503	}
504
505	if (!tp->packets_out)
506		return;
507
508	skb = tcp_rtx_queue_head(sk);
509	if (WARN_ON_ONCE(!skb))
510		return;
511
512	tp->tlp_high_seq = 0;
513
514	if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
515	    !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
516		/* Receiver dastardly shrinks window. Our retransmits
517		 * become zero probes, but we should not timeout this
518		 * connection. If the socket is an orphan, time it out,
519		 * we cannot allow such beasts to hang infinitely.
520		 */
521		struct inet_sock *inet = inet_sk(sk);
522		if (sk->sk_family == AF_INET) {
523			net_dbg_ratelimited("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
524					    &inet->inet_daddr,
525					    ntohs(inet->inet_dport),
526					    inet->inet_num,
527					    tp->snd_una, tp->snd_nxt);
528		}
529#if IS_ENABLED(CONFIG_IPV6)
530		else if (sk->sk_family == AF_INET6) {
531			net_dbg_ratelimited("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
532					    &sk->sk_v6_daddr,
533					    ntohs(inet->inet_dport),
534					    inet->inet_num,
535					    tp->snd_una, tp->snd_nxt);
536		}
537#endif
538		if (tcp_rtx_probe0_timed_out(sk, skb)) {
539			tcp_write_err(sk);
540			goto out;
541		}
542		tcp_enter_loss(sk);
543		tcp_retransmit_skb(sk, skb, 1);
544		__sk_dst_reset(sk);
545		goto out_reset_timer;
546	}
547
548	__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEOUTS);
549	if (tcp_write_timeout(sk))
550		goto out;
551
552	if (icsk->icsk_retransmits == 0) {
553		int mib_idx = 0;
554
555		if (icsk->icsk_ca_state == TCP_CA_Recovery) {
556			if (tcp_is_sack(tp))
557				mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
558			else
559				mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
560		} else if (icsk->icsk_ca_state == TCP_CA_Loss) {
561			mib_idx = LINUX_MIB_TCPLOSSFAILURES;
562		} else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
563			   tp->sacked_out) {
564			if (tcp_is_sack(tp))
565				mib_idx = LINUX_MIB_TCPSACKFAILURES;
566			else
567				mib_idx = LINUX_MIB_TCPRENOFAILURES;
568		}
569		if (mib_idx)
570			__NET_INC_STATS(sock_net(sk), mib_idx);
571	}
572
573	tcp_enter_loss(sk);
574
575	icsk->icsk_retransmits++;
576	if (tcp_retransmit_skb(sk, tcp_rtx_queue_head(sk), 1) > 0) {
577		/* Retransmission failed because of local congestion,
578		 * Let senders fight for local resources conservatively.
579		 */
580		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
581					  TCP_RESOURCE_PROBE_INTERVAL,
582					  TCP_RTO_MAX);
583		goto out;
584	}
585
586	/* Increase the timeout each time we retransmit.  Note that
587	 * we do not increase the rtt estimate.  rto is initialized
588	 * from rtt, but increases here.  Jacobson (SIGCOMM 88) suggests
589	 * that doubling rto each time is the least we can get away with.
590	 * In KA9Q, Karn uses this for the first few times, and then
591	 * goes to quadratic.  netBSD doubles, but only goes up to *64,
592	 * and clamps at 1 to 64 sec afterwards.  Note that 120 sec is
593	 * defined in the protocol as the maximum possible RTT.  I guess
594	 * we'll have to use something other than TCP to talk to the
595	 * University of Mars.
596	 *
597	 * PAWS allows us longer timeouts and large windows, so once
598	 * implemented ftp to mars will work nicely. We will have to fix
599	 * the 120 second clamps though!
600	 */
601	icsk->icsk_backoff++;
602
603out_reset_timer:
604	/* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
605	 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
606	 * might be increased if the stream oscillates between thin and thick,
607	 * thus the old value might already be too high compared to the value
608	 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
609	 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
610	 * exponential backoff behaviour to avoid continue hammering
611	 * linear-timeout retransmissions into a black hole
612	 */
613	if (sk->sk_state == TCP_ESTABLISHED &&
614	    (tp->thin_lto || READ_ONCE(net->ipv4.sysctl_tcp_thin_linear_timeouts)) &&
615	    tcp_stream_is_thin(tp) &&
616	    icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
617		icsk->icsk_backoff = 0;
618		icsk->icsk_rto = clamp(__tcp_set_rto(tp),
619				       tcp_rto_min(sk),
620				       TCP_RTO_MAX);
621	} else {
622		/* Use normal (exponential) backoff */
623		icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
624	}
625	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
626				  tcp_clamp_rto_to_user_timeout(sk), TCP_RTO_MAX);
627	if (retransmits_timed_out(sk, READ_ONCE(net->ipv4.sysctl_tcp_retries1) + 1, 0))
628		__sk_dst_reset(sk);
629
630out:;
631}
632
633/* Called with bottom-half processing disabled.
634   Called by tcp_write_timer() */
635void tcp_write_timer_handler(struct sock *sk)
636{
637	struct inet_connection_sock *icsk = inet_csk(sk);
638	int event;
639
640	if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
641	    !icsk->icsk_pending)
642		goto out;
643
644	if (time_after(icsk->icsk_timeout, jiffies)) {
645		sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
646		goto out;
647	}
648
649	tcp_mstamp_refresh(tcp_sk(sk));
650	event = icsk->icsk_pending;
651
652	switch (event) {
653	case ICSK_TIME_REO_TIMEOUT:
654		tcp_rack_reo_timeout(sk);
655		break;
656	case ICSK_TIME_LOSS_PROBE:
657		tcp_send_loss_probe(sk);
658		break;
659	case ICSK_TIME_RETRANS:
660		icsk->icsk_pending = 0;
661		tcp_retransmit_timer(sk);
662		break;
663	case ICSK_TIME_PROBE0:
664		icsk->icsk_pending = 0;
665		tcp_probe_timer(sk);
666		break;
667	}
668
669out:
670	sk_mem_reclaim(sk);
671}
672
673static void tcp_write_timer(struct timer_list *t)
674{
675	struct inet_connection_sock *icsk =
676			from_timer(icsk, t, icsk_retransmit_timer);
677	struct sock *sk = &icsk->icsk_inet.sk;
678
679	bh_lock_sock(sk);
680	if (!sock_owned_by_user(sk)) {
681		tcp_write_timer_handler(sk);
682	} else {
683		/* delegate our work to tcp_release_cb() */
684		if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &sk->sk_tsq_flags))
685			sock_hold(sk);
686	}
687	bh_unlock_sock(sk);
688	sock_put(sk);
689}
690
691void tcp_syn_ack_timeout(const struct request_sock *req)
692{
693	struct net *net = read_pnet(&inet_rsk(req)->ireq_net);
694
695	__NET_INC_STATS(net, LINUX_MIB_TCPTIMEOUTS);
696}
697EXPORT_SYMBOL(tcp_syn_ack_timeout);
698
699void tcp_set_keepalive(struct sock *sk, int val)
700{
701	if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
702		return;
703
704	if (val && !sock_flag(sk, SOCK_KEEPOPEN))
705		inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
706	else if (!val)
707		inet_csk_delete_keepalive_timer(sk);
708}
709EXPORT_SYMBOL_GPL(tcp_set_keepalive);
710
711
712static void tcp_keepalive_timer (struct timer_list *t)
713{
714	struct sock *sk = from_timer(sk, t, sk_timer);
715	struct inet_connection_sock *icsk = inet_csk(sk);
716	struct tcp_sock *tp = tcp_sk(sk);
717	u32 elapsed;
718
719	/* Only process if socket is not in use. */
720	bh_lock_sock(sk);
721	if (sock_owned_by_user(sk)) {
722		/* Try again later. */
723		inet_csk_reset_keepalive_timer (sk, HZ/20);
724		goto out;
725	}
726
727	if (sk->sk_state == TCP_LISTEN) {
728		pr_err("Hmm... keepalive on a LISTEN ???\n");
729		goto out;
730	}
731
732	tcp_mstamp_refresh(tp);
733	if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
734		if (tp->linger2 >= 0) {
735			const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
736
737			if (tmo > 0) {
738				tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
739				goto out;
740			}
741		}
742		tcp_send_active_reset(sk, GFP_ATOMIC);
743		goto death;
744	}
745
746	if (!sock_flag(sk, SOCK_KEEPOPEN) ||
747	    ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_SYN_SENT)))
748		goto out;
749
750	elapsed = keepalive_time_when(tp);
751
752	/* It is alive without keepalive 8) */
753	if (tp->packets_out || !tcp_write_queue_empty(sk))
754		goto resched;
755
756	elapsed = keepalive_time_elapsed(tp);
757
758	if (elapsed >= keepalive_time_when(tp)) {
759		/* If the TCP_USER_TIMEOUT option is enabled, use that
760		 * to determine when to timeout instead.
761		 */
762		if ((icsk->icsk_user_timeout != 0 &&
763		    elapsed >= msecs_to_jiffies(icsk->icsk_user_timeout) &&
764		    icsk->icsk_probes_out > 0) ||
765		    (icsk->icsk_user_timeout == 0 &&
766		    icsk->icsk_probes_out >= keepalive_probes(tp))) {
767			tcp_send_active_reset(sk, GFP_ATOMIC);
768			tcp_write_err(sk);
769			goto out;
770		}
771		if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) {
772			icsk->icsk_probes_out++;
773			elapsed = keepalive_intvl_when(tp);
774		} else {
775			/* If keepalive was lost due to local congestion,
776			 * try harder.
777			 */
778			elapsed = TCP_RESOURCE_PROBE_INTERVAL;
779		}
780	} else {
781		/* It is tp->rcv_tstamp + keepalive_time_when(tp) */
782		elapsed = keepalive_time_when(tp) - elapsed;
783	}
784
785	sk_mem_reclaim(sk);
786
787resched:
788	inet_csk_reset_keepalive_timer (sk, elapsed);
789	goto out;
790
791death:
792	tcp_done(sk);
793
794out:
795	bh_unlock_sock(sk);
796	sock_put(sk);
797}
798
799static enum hrtimer_restart tcp_compressed_ack_kick(struct hrtimer *timer)
800{
801	struct tcp_sock *tp = container_of(timer, struct tcp_sock, compressed_ack_timer);
802	struct sock *sk = (struct sock *)tp;
803
804	bh_lock_sock(sk);
805	if (!sock_owned_by_user(sk)) {
806		if (tp->compressed_ack) {
807			/* Since we have to send one ack finally,
808			 * substract one from tp->compressed_ack to keep
809			 * LINUX_MIB_TCPACKCOMPRESSED accurate.
810			 */
811			tp->compressed_ack--;
812			tcp_send_ack(sk);
813		}
814	} else {
815		if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED,
816				      &sk->sk_tsq_flags))
817			sock_hold(sk);
818	}
819	bh_unlock_sock(sk);
820
821	sock_put(sk);
822
823	return HRTIMER_NORESTART;
824}
825
826void tcp_init_xmit_timers(struct sock *sk)
827{
828	inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
829				  &tcp_keepalive_timer);
830	hrtimer_init(&tcp_sk(sk)->pacing_timer, CLOCK_MONOTONIC,
831		     HRTIMER_MODE_ABS_PINNED_SOFT);
832	tcp_sk(sk)->pacing_timer.function = tcp_pace_kick;
833
834	hrtimer_init(&tcp_sk(sk)->compressed_ack_timer, CLOCK_MONOTONIC,
835		     HRTIMER_MODE_REL_PINNED_SOFT);
836	tcp_sk(sk)->compressed_ack_timer.function = tcp_compressed_ack_kick;
837}
838