tcp_timer.c

  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
 27static 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
 49u32 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
 75static 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 */
105static 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(sk) || !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 */
148static 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
164static 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
186static 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(sk) / 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(sk);
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 */
213static 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. */
241static 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(sk);
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 */
307void 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_delack_timeout(icsk), jiffies)) {
326		sk_reset_timer(sk, &icsk->icsk_delack_timer,
327			       icsk_delack_timeout(icsk));
328		return;
329	}
330	icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
331
332	if (inet_csk_ack_scheduled(sk)) {
333		if (!inet_csk_in_pingpong_mode(sk)) {
334			/* Delayed ACK missed: inflate ATO. */
335			icsk->icsk_ack.ato = min_t(u32, icsk->icsk_ack.ato << 1, icsk->icsk_rto);
336		} else {
337			/* Delayed ACK missed: leave pingpong mode and
338			 * deflate ATO.
339			 */
340			inet_csk_exit_pingpong_mode(sk);
341			icsk->icsk_ack.ato      = TCP_ATO_MIN;
342		}
343		tcp_mstamp_refresh(tp);
344		tcp_send_ack(sk);
345		__NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKS);
346	}
347}
348
349
350/**
351 *  tcp_delack_timer() - The TCP delayed ACK timeout handler
352 *  @t:  Pointer to the timer. (gets casted to struct sock *)
353 *
354 *  This function gets (indirectly) called when the kernel timer for a TCP packet
355 *  of this socket expires. Calls tcp_delack_timer_handler() to do the actual work.
356 *
357 *  Returns: Nothing (void)
358 */
359static void tcp_delack_timer(struct timer_list *t)
360{
361	struct inet_connection_sock *icsk =
362			timer_container_of(icsk, t, icsk_delack_timer);
363	struct sock *sk = &icsk->icsk_inet.sk;
364
365	/* Avoid taking socket spinlock if there is no ACK to send.
366	 * The compressed_ack check is racy, but a separate hrtimer
367	 * will take care of it eventually.
368	 */
369	if (!(smp_load_acquire(&icsk->icsk_ack.pending) & ICSK_ACK_TIMER) &&
370	    !READ_ONCE(tcp_sk(sk)->compressed_ack))
371		goto out;
372
373	bh_lock_sock(sk);
374	if (!sock_owned_by_user(sk)) {
375		tcp_delack_timer_handler(sk);
376	} else {
377		__NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
378		/* deleguate our work to tcp_release_cb() */
379		if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &sk->sk_tsq_flags))
380			sock_hold(sk);
381	}
382	bh_unlock_sock(sk);
383out:
384	sock_put(sk);
385}
386
387static void tcp_probe_timer(struct sock *sk)
388{
389	struct inet_connection_sock *icsk = inet_csk(sk);
390	struct sk_buff *skb = tcp_send_head(sk);
391	struct tcp_sock *tp = tcp_sk(sk);
392	int max_probes;
393
394	if (tp->packets_out || !skb) {
395		WRITE_ONCE(icsk->icsk_probes_out, 0);
396		icsk->icsk_probes_tstamp = 0;
397		return;
398	}
399
400	/* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as
401	 * long as the receiver continues to respond probes. We support this by
402	 * default and reset icsk_probes_out with incoming ACKs. But if the
403	 * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we
404	 * kill the socket when the retry count and the time exceeds the
405	 * corresponding system limit. We also implement similar policy when
406	 * we use RTO to probe window in tcp_retransmit_timer().
407	 */
408	if (!icsk->icsk_probes_tstamp) {
409		icsk->icsk_probes_tstamp = tcp_jiffies32;
410	} else {
411		u32 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
412
413		if (user_timeout &&
414		    (s32)(tcp_jiffies32 - icsk->icsk_probes_tstamp) >=
415		     msecs_to_jiffies(user_timeout))
416			goto abort;
417	}
418	max_probes = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retries2);
419	if (sock_flag(sk, SOCK_DEAD)) {
420		unsigned int rto_max = tcp_rto_max(sk);
421		const bool alive = inet_csk_rto_backoff(icsk, rto_max) < rto_max;
422
423		max_probes = tcp_orphan_retries(sk, alive);
424		if (!alive && icsk->icsk_backoff >= max_probes)
425			goto abort;
426		if (tcp_out_of_resources(sk, true))
427			return;
428	}
429
430	if (icsk->icsk_probes_out >= max_probes) {
431abort:		tcp_write_err(sk);
432	} else {
433		/* Only send another probe if we didn't close things up. */
434		tcp_send_probe0(sk);
435	}
436}
437
438static void tcp_update_rto_stats(struct sock *sk)
439{
440	struct inet_connection_sock *icsk = inet_csk(sk);
441	struct tcp_sock *tp = tcp_sk(sk);
442
443	if (!icsk->icsk_retransmits) {
444		tp->total_rto_recoveries++;
445		tp->rto_stamp = tcp_time_stamp_ms(tp);
446	}
447	WRITE_ONCE(icsk->icsk_retransmits, icsk->icsk_retransmits + 1);
448	tp->total_rto++;
449}
450
451/*
452 *	Timer for Fast Open socket to retransmit SYNACK. Note that the
453 *	sk here is the child socket, not the parent (listener) socket.
454 */
455static void tcp_fastopen_synack_timer(struct sock *sk, struct request_sock *req)
456{
457	struct inet_connection_sock *icsk = inet_csk(sk);
458	struct tcp_sock *tp = tcp_sk(sk);
459	int max_retries;
460
461	tcp_syn_ack_timeout(req);
462
463	/* Add one more retry for fastopen.
464	 * Paired with WRITE_ONCE() in tcp_sock_set_syncnt()
465	 */
466	max_retries = READ_ONCE(icsk->icsk_syn_retries) ? :
467		READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_synack_retries) + 1;
468
469	if (req->num_timeout >= max_retries) {
470		tcp_write_err(sk);
471		return;
472	}
473	/* Lower cwnd after certain SYNACK timeout like tcp_init_transfer() */
474	if (icsk->icsk_retransmits == 1)
475		tcp_enter_loss(sk);
476	/* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
477	 * returned from rtx_syn_ack() to make it more persistent like
478	 * regular retransmit because if the child socket has been accepted
479	 * it's not good to give up too easily.
480	 */
481	tcp_rtx_synack(sk, req);
482	req->num_timeout++;
483	tcp_update_rto_stats(sk);
484	if (!tp->retrans_stamp)
485		tp->retrans_stamp = tcp_time_stamp_ts(tp);
486	tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
487			  req->timeout << req->num_timeout, false);
488}
489
490static bool tcp_rtx_probe0_timed_out(const struct sock *sk,
491				     const struct sk_buff *skb,
492				     u32 rtx_delta)
493{
494	const struct inet_connection_sock *icsk = inet_csk(sk);
495	u32 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
496	const struct tcp_sock *tp = tcp_sk(sk);
497	int timeout = tcp_rto_max(sk) * 2;
498	s32 rcv_delta;
499
500	if (user_timeout) {
501		/* If user application specified a TCP_USER_TIMEOUT,
502		 * it does not want win 0 packets to 'reset the timer'
503		 * while retransmits are not making progress.
504		 */
505		if (rtx_delta > user_timeout)
506			return true;
507		timeout = min_t(u32, timeout, msecs_to_jiffies(user_timeout));
508	}
509	/* Note: timer interrupt might have been delayed by at least one jiffy,
510	 * and tp->rcv_tstamp might very well have been written recently.
511	 * rcv_delta can thus be negative.
512	 */
513	rcv_delta = tcp_timeout_expires(sk) - tp->rcv_tstamp;
514	if (rcv_delta <= timeout)
515		return false;
516
517	return msecs_to_jiffies(rtx_delta) > timeout;
518}
519
520/**
521 *  tcp_retransmit_timer() - The TCP retransmit timeout handler
522 *  @sk:  Pointer to the current socket.
523 *
524 *  This function gets called when the kernel timer for a TCP packet
525 *  of this socket expires.
526 *
527 *  It handles retransmission, timer adjustment and other necessary measures.
528 *
529 *  Returns: Nothing (void)
530 */
531void tcp_retransmit_timer(struct sock *sk)
532{
533	struct tcp_sock *tp = tcp_sk(sk);
534	struct net *net = sock_net(sk);
535	struct inet_connection_sock *icsk = inet_csk(sk);
536	struct request_sock *req;
537	struct sk_buff *skb;
538
539	req = rcu_dereference_protected(tp->fastopen_rsk,
540					lockdep_sock_is_held(sk));
541	if (req) {
542		WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
543			     sk->sk_state != TCP_FIN_WAIT1);
544		tcp_fastopen_synack_timer(sk, req);
545		/* Before we receive ACK to our SYN-ACK don't retransmit
546		 * anything else (e.g., data or FIN segments).
547		 */
548		return;
549	}
550
551	if (!tp->packets_out)
552		return;
553
554	skb = tcp_rtx_queue_head(sk);
555	if (WARN_ON_ONCE(!skb))
556		return;
557
558	if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
559	    !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
560		/* Receiver dastardly shrinks window. Our retransmits
561		 * become zero probes, but we should not timeout this
562		 * connection. If the socket is an orphan, time it out,
563		 * we cannot allow such beasts to hang infinitely.
564		 */
565		struct inet_sock *inet = inet_sk(sk);
566		u32 rtx_delta;
567
568		rtx_delta = tcp_time_stamp_ts(tp) - (tp->retrans_stamp ?: 
569				tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb));
570		if (tp->tcp_usec_ts)
571			rtx_delta /= USEC_PER_MSEC;
572
573		if (sk->sk_family == AF_INET) {
574			net_dbg_ratelimited("Probing zero-window on %pI4:%u/%u, seq=%u:%u, recv %ums ago, lasting %ums\n",
575				&inet->inet_daddr, ntohs(inet->inet_dport),
576				inet->inet_num, tp->snd_una, tp->snd_nxt,
577				jiffies_to_msecs(jiffies - tp->rcv_tstamp),
578				rtx_delta);
579		}
580#if IS_ENABLED(CONFIG_IPV6)
581		else if (sk->sk_family == AF_INET6) {
582			net_dbg_ratelimited("Probing zero-window on %pI6:%u/%u, seq=%u:%u, recv %ums ago, lasting %ums\n",
583				&sk->sk_v6_daddr, ntohs(inet->inet_dport),
584				inet->inet_num, tp->snd_una, tp->snd_nxt,
585				jiffies_to_msecs(jiffies - tp->rcv_tstamp),
586				rtx_delta);
587		}
588#endif
589		if (tcp_rtx_probe0_timed_out(sk, skb, rtx_delta)) {
590			tcp_write_err(sk);
591			goto out;
592		}
593		tcp_enter_loss(sk);
594		tcp_retransmit_skb(sk, skb, 1);
595		__sk_dst_reset(sk);
596		goto out_reset_timer;
597	}
598
599	__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEOUTS);
600	if (tcp_write_timeout(sk))
601		goto out;
602
603	if (icsk->icsk_retransmits == 0) {
604		int mib_idx = 0;
605
606		if (icsk->icsk_ca_state == TCP_CA_Recovery) {
607			if (tcp_is_sack(tp))
608				mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
609			else
610				mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
611		} else if (icsk->icsk_ca_state == TCP_CA_Loss) {
612			mib_idx = LINUX_MIB_TCPLOSSFAILURES;
613		} else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
614			   tp->sacked_out) {
615			if (tcp_is_sack(tp))
616				mib_idx = LINUX_MIB_TCPSACKFAILURES;
617			else
618				mib_idx = LINUX_MIB_TCPRENOFAILURES;
619		}
620		if (mib_idx)
621			__NET_INC_STATS(sock_net(sk), mib_idx);
622	}
623
624	tcp_enter_loss(sk);
625
626	tcp_update_rto_stats(sk);
627	if (tcp_retransmit_skb(sk, tcp_rtx_queue_head(sk), 1) > 0) {
628		/* Retransmission failed because of local congestion,
629		 * Let senders fight for local resources conservatively.
630		 */
631		tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
632				     TCP_RESOURCE_PROBE_INTERVAL,
633				     false);
634		goto out;
635	}
636
637	/* Increase the timeout each time we retransmit.  Note that
638	 * we do not increase the rtt estimate.  rto is initialized
639	 * from rtt, but increases here.  Jacobson (SIGCOMM 88) suggests
640	 * that doubling rto each time is the least we can get away with.
641	 * In KA9Q, Karn uses this for the first few times, and then
642	 * goes to quadratic.  netBSD doubles, but only goes up to *64,
643	 * and clamps at 1 to 64 sec afterwards.  Note that 120 sec is
644	 * defined in the protocol as the maximum possible RTT.  I guess
645	 * we'll have to use something other than TCP to talk to the
646	 * University of Mars.
647	 *
648	 * PAWS allows us longer timeouts and large windows, so once
649	 * implemented ftp to mars will work nicely. We will have to fix
650	 * the 120 second clamps though!
651	 */
652
653out_reset_timer:
654	/* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
655	 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
656	 * might be increased if the stream oscillates between thin and thick,
657	 * thus the old value might already be too high compared to the value
658	 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
659	 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
660	 * exponential backoff behaviour to avoid continue hammering
661	 * linear-timeout retransmissions into a black hole
662	 */
663	if (sk->sk_state == TCP_ESTABLISHED &&
664	    (tp->thin_lto || READ_ONCE(net->ipv4.sysctl_tcp_thin_linear_timeouts)) &&
665	    tcp_stream_is_thin(tp) &&
666	    icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
667		icsk->icsk_backoff = 0;
668		icsk->icsk_rto = clamp(__tcp_set_rto(tp),
669				       tcp_rto_min(sk),
670				       tcp_rto_max(sk));
671	} else if (sk->sk_state != TCP_SYN_SENT ||
672		   tp->total_rto >
673		   READ_ONCE(net->ipv4.sysctl_tcp_syn_linear_timeouts)) {
674		/* Use normal (exponential) backoff unless linear timeouts are
675		 * activated.
676		 */
677		icsk->icsk_backoff++;
678		icsk->icsk_rto = min(icsk->icsk_rto << 1, tcp_rto_max(sk));
679	}
680	tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
681			     tcp_clamp_rto_to_user_timeout(sk), false);
682	if (retransmits_timed_out(sk, READ_ONCE(net->ipv4.sysctl_tcp_retries1) + 1, 0))
683		__sk_dst_reset(sk);
684
685out:;
686}
687
688/* Called with bottom-half processing disabled.
689 * Called by tcp_write_timer() and tcp_release_cb().
690 */
691void tcp_write_timer_handler(struct sock *sk)
692{
693	struct inet_connection_sock *icsk = inet_csk(sk);
694	int event;
695
696	if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
697	    !icsk->icsk_pending)
698		return;
699
700	if (time_after(tcp_timeout_expires(sk), jiffies)) {
701		sk_reset_timer(sk, &sk->tcp_retransmit_timer,
702			       tcp_timeout_expires(sk));
703		return;
704	}
705	tcp_mstamp_refresh(tcp_sk(sk));
706	event = icsk->icsk_pending;
707
708	switch (event) {
709	case ICSK_TIME_REO_TIMEOUT:
710		tcp_rack_reo_timeout(sk);
711		break;
712	case ICSK_TIME_LOSS_PROBE:
713		tcp_send_loss_probe(sk);
714		break;
715	case ICSK_TIME_RETRANS:
716		smp_store_release(&icsk->icsk_pending, 0);
717		tcp_retransmit_timer(sk);
718		break;
719	case ICSK_TIME_PROBE0:
720		smp_store_release(&icsk->icsk_pending, 0);
721		tcp_probe_timer(sk);
722		break;
723	}
724}
725
726static void tcp_write_timer(struct timer_list *t)
727{
728	struct sock *sk = timer_container_of(sk, t, tcp_retransmit_timer);
729
730	/* Avoid locking the socket when there is no pending event. */
731	if (!smp_load_acquire(&inet_csk(sk)->icsk_pending))
732		goto out;
733
734	bh_lock_sock(sk);
735	if (!sock_owned_by_user(sk)) {
736		tcp_write_timer_handler(sk);
737	} else {
738		/* delegate our work to tcp_release_cb() */
739		if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &sk->sk_tsq_flags))
740			sock_hold(sk);
741	}
742	bh_unlock_sock(sk);
743out:
744	sock_put(sk);
745}
746
747void tcp_syn_ack_timeout(const struct request_sock *req)
748{
749	struct net *net = read_pnet(&inet_rsk(req)->ireq_net);
750
751	__NET_INC_STATS(net, LINUX_MIB_TCPTIMEOUTS);
752}
753
754void tcp_reset_keepalive_timer(struct sock *sk, unsigned long len)
755{
756	sk_reset_timer(sk, &inet_csk(sk)->icsk_keepalive_timer, jiffies + len);
757}
758
759static void tcp_delete_keepalive_timer(struct sock *sk)
760{
761	sk_stop_timer(sk, &inet_csk(sk)->icsk_keepalive_timer);
762}
763
764void tcp_set_keepalive(struct sock *sk, int val)
765{
766	if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
767		return;
768
769	if (val && !sock_flag(sk, SOCK_KEEPOPEN))
770		tcp_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
771	else if (!val)
772		tcp_delete_keepalive_timer(sk);
773}
774EXPORT_IPV6_MOD_GPL(tcp_set_keepalive);
775
776static void tcp_keepalive_timer(struct timer_list *t)
777{
778	struct inet_connection_sock *icsk =
779		timer_container_of(icsk, t, icsk_keepalive_timer);
780	struct sock *sk = &icsk->icsk_inet.sk;
781	struct tcp_sock *tp = tcp_sk(sk);
782	u32 elapsed;
783
784	/* Only process if socket is not in use. */
785	bh_lock_sock(sk);
786	if (sock_owned_by_user(sk)) {
787		/* Try again later. */
788		tcp_reset_keepalive_timer(sk, HZ/20);
789		goto out;
790	}
791
792	if (sk->sk_state == TCP_LISTEN) {
793		pr_err("Hmm... keepalive on a LISTEN ???\n");
794		goto out;
795	}
796
797	tcp_mstamp_refresh(tp);
798	if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
799		if (READ_ONCE(tp->linger2) >= 0) {
800			const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
801
802			if (tmo > 0) {
803				tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
804				goto out;
805			}
806		}
807		tcp_send_active_reset(sk, GFP_ATOMIC, SK_RST_REASON_TCP_STATE);
808		goto death;
809	}
810
811	if (!sock_flag(sk, SOCK_KEEPOPEN) ||
812	    ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_SYN_SENT)))
813		goto out;
814
815	elapsed = keepalive_time_when(tp);
816
817	/* It is alive without keepalive 8) */
818	if (tp->packets_out || !tcp_write_queue_empty(sk))
819		goto resched;
820
821	elapsed = keepalive_time_elapsed(tp);
822
823	if (elapsed >= keepalive_time_when(tp)) {
824		u32 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
825
826		/* If the TCP_USER_TIMEOUT option is enabled, use that
827		 * to determine when to timeout instead.
828		 */
829		if ((user_timeout != 0 &&
830		    elapsed >= msecs_to_jiffies(user_timeout) &&
831		    icsk->icsk_probes_out > 0) ||
832		    (user_timeout == 0 &&
833		    icsk->icsk_probes_out >= keepalive_probes(tp))) {
834			tcp_send_active_reset(sk, GFP_ATOMIC,
835					      SK_RST_REASON_TCP_KEEPALIVE_TIMEOUT);
836			tcp_write_err(sk);
837			goto out;
838		}
839		if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) {
840			WRITE_ONCE(icsk->icsk_probes_out, icsk->icsk_probes_out + 1);
841			elapsed = keepalive_intvl_when(tp);
842		} else {
843			/* If keepalive was lost due to local congestion,
844			 * try harder.
845			 */
846			elapsed = TCP_RESOURCE_PROBE_INTERVAL;
847		}
848	} else {
849		/* It is tp->rcv_tstamp + keepalive_time_when(tp) */
850		elapsed = keepalive_time_when(tp) - elapsed;
851	}
852
853resched:
854	tcp_reset_keepalive_timer(sk, elapsed);
855	goto out;
856
857death:
858	tcp_done(sk);
859
860out:
861	bh_unlock_sock(sk);
862	sock_put(sk);
863}
864
865static enum hrtimer_restart tcp_compressed_ack_kick(struct hrtimer *timer)
866{
867	struct tcp_sock *tp = container_of(timer, struct tcp_sock, compressed_ack_timer);
868	struct sock *sk = (struct sock *)tp;
869
870	bh_lock_sock(sk);
871	if (!sock_owned_by_user(sk)) {
872		if (tp->compressed_ack) {
873			/* Since we have to send one ack finally,
874			 * subtract one from tp->compressed_ack to keep
875			 * LINUX_MIB_TCPACKCOMPRESSED accurate.
876			 */
877			tp->compressed_ack--;
878			tcp_mstamp_refresh(tp);
879			tcp_send_ack(sk);
880		}
881	} else {
882		if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED,
883				      &sk->sk_tsq_flags))
884			sock_hold(sk);
885	}
886	bh_unlock_sock(sk);
887
888	sock_put(sk);
889
890	return HRTIMER_NORESTART;
891}
892
893void tcp_init_xmit_timers(struct sock *sk)
894{
895	inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
896				  &tcp_keepalive_timer);
897	hrtimer_setup(&tcp_sk(sk)->pacing_timer, tcp_pace_kick, CLOCK_MONOTONIC,
898		      HRTIMER_MODE_ABS_PINNED_SOFT);
899
900	hrtimer_setup(&tcp_sk(sk)->compressed_ack_timer, tcp_compressed_ack_kick, CLOCK_MONOTONIC,
901		      HRTIMER_MODE_REL_PINNED_SOFT);
902}