Linux Audio

Check our new training course

Loading...
v3.15
  1/*
  2 * INET		An implementation of the TCP/IP protocol suite for the LINUX
  3 *		operating system.  INET is implemented using the  BSD Socket
  4 *		interface as the means of communication with the user level.
  5 *
  6 *		Implementation of the Transmission Control Protocol(TCP).
  7 *
  8 * Authors:	Ross Biro
  9 *		Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
 10 *		Mark Evans, <evansmp@uhura.aston.ac.uk>
 11 *		Corey Minyard <wf-rch!minyard@relay.EU.net>
 12 *		Florian La Roche, <flla@stud.uni-sb.de>
 13 *		Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
 14 *		Linus Torvalds, <torvalds@cs.helsinki.fi>
 15 *		Alan Cox, <gw4pts@gw4pts.ampr.org>
 16 *		Matthew Dillon, <dillon@apollo.west.oic.com>
 17 *		Arnt Gulbrandsen, <agulbra@nvg.unit.no>
 18 *		Jorge Cwik, <jorge@laser.satlink.net>
 19 */
 20
 21#include <linux/module.h>
 22#include <linux/gfp.h>
 23#include <net/tcp.h>
 24
 25int sysctl_tcp_syn_retries __read_mostly = TCP_SYN_RETRIES;
 26int sysctl_tcp_synack_retries __read_mostly = TCP_SYNACK_RETRIES;
 27int sysctl_tcp_keepalive_time __read_mostly = TCP_KEEPALIVE_TIME;
 28int sysctl_tcp_keepalive_probes __read_mostly = TCP_KEEPALIVE_PROBES;
 29int sysctl_tcp_keepalive_intvl __read_mostly = TCP_KEEPALIVE_INTVL;
 30int sysctl_tcp_retries1 __read_mostly = TCP_RETR1;
 31int sysctl_tcp_retries2 __read_mostly = TCP_RETR2;
 32int sysctl_tcp_orphan_retries __read_mostly;
 33int sysctl_tcp_thin_linear_timeouts __read_mostly;
 34
 
 
 
 
 
 
 
 35static void tcp_write_err(struct sock *sk)
 36{
 37	sk->sk_err = sk->sk_err_soft ? : ETIMEDOUT;
 38	sk->sk_error_report(sk);
 39
 40	tcp_done(sk);
 41	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT);
 42}
 43
 44/* Do not allow orphaned sockets to eat all our resources.
 45 * This is direct violation of TCP specs, but it is required
 46 * to prevent DoS attacks. It is called when a retransmission timeout
 47 * or zero probe timeout occurs on orphaned socket.
 48 *
 49 * Criteria is still not confirmed experimentally and may change.
 50 * We kill the socket, if:
 51 * 1. If number of orphaned sockets exceeds an administratively configured
 52 *    limit.
 53 * 2. If we have strong memory pressure.
 
 
 
 
 
 54 */
 55static int tcp_out_of_resources(struct sock *sk, int do_reset)
 56{
 57	struct tcp_sock *tp = tcp_sk(sk);
 58	int shift = 0;
 59
 60	/* If peer does not open window for long time, or did not transmit
 61	 * anything for long time, penalize it. */
 62	if ((s32)(tcp_time_stamp - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset)
 63		shift++;
 64
 65	/* If some dubious ICMP arrived, penalize even more. */
 66	if (sk->sk_err_soft)
 67		shift++;
 68
 69	if (tcp_check_oom(sk, shift)) {
 70		/* Catch exceptional cases, when connection requires reset.
 71		 *      1. Last segment was sent recently. */
 72		if ((s32)(tcp_time_stamp - tp->lsndtime) <= TCP_TIMEWAIT_LEN ||
 73		    /*  2. Window is closed. */
 74		    (!tp->snd_wnd && !tp->packets_out))
 75			do_reset = 1;
 76		if (do_reset)
 77			tcp_send_active_reset(sk, GFP_ATOMIC);
 78		tcp_done(sk);
 79		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY);
 80		return 1;
 81	}
 82	return 0;
 83}
 84
 85/* Calculate maximal number or retries on an orphaned socket. */
 86static int tcp_orphan_retries(struct sock *sk, int alive)
 
 
 
 
 87{
 88	int retries = sysctl_tcp_orphan_retries; /* May be zero. */
 89
 90	/* We know from an ICMP that something is wrong. */
 91	if (sk->sk_err_soft && !alive)
 92		retries = 0;
 93
 94	/* However, if socket sent something recently, select some safe
 95	 * number of retries. 8 corresponds to >100 seconds with minimal
 96	 * RTO of 200msec. */
 97	if (retries == 0 && alive)
 98		retries = 8;
 99	return retries;
100}
101
102static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk)
103{
 
 
104	/* Black hole detection */
105	if (sysctl_tcp_mtu_probing) {
106		if (!icsk->icsk_mtup.enabled) {
107			icsk->icsk_mtup.enabled = 1;
 
108			tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
109		} else {
 
110			struct tcp_sock *tp = tcp_sk(sk);
111			int mss;
112
113			mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1;
114			mss = min(sysctl_tcp_base_mss, mss);
115			mss = max(mss, 68 - tp->tcp_header_len);
116			icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
117			tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
118		}
119	}
120}
121
122/* This function calculates a "timeout" which is equivalent to the timeout of a
123 * TCP connection after "boundary" unsuccessful, exponentially backed-off
 
 
 
 
 
 
 
 
 
 
 
124 * retransmissions with an initial RTO of TCP_RTO_MIN or TCP_TIMEOUT_INIT if
125 * syn_set flag is set.
 
126 */
127static bool retransmits_timed_out(struct sock *sk,
128				  unsigned int boundary,
129				  unsigned int timeout,
130				  bool syn_set)
131{
132	unsigned int linear_backoff_thresh, start_ts;
133	unsigned int rto_base = syn_set ? TCP_TIMEOUT_INIT : TCP_RTO_MIN;
134
135	if (!inet_csk(sk)->icsk_retransmits)
136		return false;
137
138	if (unlikely(!tcp_sk(sk)->retrans_stamp))
139		start_ts = TCP_SKB_CB(tcp_write_queue_head(sk))->when;
140	else
141		start_ts = tcp_sk(sk)->retrans_stamp;
142
143	if (likely(timeout == 0)) {
144		linear_backoff_thresh = ilog2(TCP_RTO_MAX/rto_base);
145
146		if (boundary <= linear_backoff_thresh)
147			timeout = ((2 << boundary) - 1) * rto_base;
148		else
149			timeout = ((2 << linear_backoff_thresh) - 1) * rto_base +
150				(boundary - linear_backoff_thresh) * TCP_RTO_MAX;
151	}
152	return (tcp_time_stamp - start_ts) >= timeout;
153}
154
155/* A write timeout has occurred. Process the after effects. */
156static int tcp_write_timeout(struct sock *sk)
157{
158	struct inet_connection_sock *icsk = inet_csk(sk);
159	struct tcp_sock *tp = tcp_sk(sk);
 
160	int retry_until;
161	bool do_reset, syn_set = false;
162
163	if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
164		if (icsk->icsk_retransmits) {
165			dst_negative_advice(sk);
166			if (tp->syn_fastopen || tp->syn_data)
167				tcp_fastopen_cache_set(sk, 0, NULL, true);
168			if (tp->syn_data)
169				NET_INC_STATS_BH(sock_net(sk),
170						 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
 
 
171		}
172		retry_until = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
173		syn_set = true;
174	} else {
175		if (retransmits_timed_out(sk, sysctl_tcp_retries1, 0, 0)) {
 
 
 
 
 
 
 
 
 
 
 
 
176			/* Black hole detection */
177			tcp_mtu_probing(icsk, sk);
178
179			dst_negative_advice(sk);
 
 
180		}
181
182		retry_until = sysctl_tcp_retries2;
183		if (sock_flag(sk, SOCK_DEAD)) {
184			const int alive = (icsk->icsk_rto < TCP_RTO_MAX);
185
186			retry_until = tcp_orphan_retries(sk, alive);
187			do_reset = alive ||
188				!retransmits_timed_out(sk, retry_until, 0, 0);
189
190			if (tcp_out_of_resources(sk, do_reset))
191				return 1;
192		}
193	}
194
195	if (retransmits_timed_out(sk, retry_until,
196				  syn_set ? 0 : icsk->icsk_user_timeout, syn_set)) {
197		/* Has it gone just too far? */
198		tcp_write_err(sk);
199		return 1;
200	}
201	return 0;
202}
203
 
204void tcp_delack_timer_handler(struct sock *sk)
205{
206	struct tcp_sock *tp = tcp_sk(sk);
207	struct inet_connection_sock *icsk = inet_csk(sk);
208
209	sk_mem_reclaim_partial(sk);
210
211	if (sk->sk_state == TCP_CLOSE || !(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
 
212		goto out;
213
214	if (time_after(icsk->icsk_ack.timeout, jiffies)) {
215		sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
216		goto out;
217	}
218	icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
219
220	if (!skb_queue_empty(&tp->ucopy.prequeue)) {
221		struct sk_buff *skb;
222
223		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSCHEDULERFAILED);
224
225		while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
226			sk_backlog_rcv(sk, skb);
227
228		tp->ucopy.memory = 0;
229	}
230
231	if (inet_csk_ack_scheduled(sk)) {
232		if (!icsk->icsk_ack.pingpong) {
233			/* Delayed ACK missed: inflate ATO. */
234			icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto);
235		} else {
236			/* Delayed ACK missed: leave pingpong mode and
237			 * deflate ATO.
238			 */
239			icsk->icsk_ack.pingpong = 0;
240			icsk->icsk_ack.ato      = TCP_ATO_MIN;
241		}
242		tcp_send_ack(sk);
243		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKS);
244	}
245
246out:
247	if (sk_under_memory_pressure(sk))
248		sk_mem_reclaim(sk);
249}
250
 
 
 
 
 
 
 
 
 
 
251static void tcp_delack_timer(unsigned long data)
252{
253	struct sock *sk = (struct sock *)data;
254
255	bh_lock_sock(sk);
256	if (!sock_owned_by_user(sk)) {
257		tcp_delack_timer_handler(sk);
258	} else {
259		inet_csk(sk)->icsk_ack.blocked = 1;
260		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
261		/* deleguate our work to tcp_release_cb() */
262		if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &tcp_sk(sk)->tsq_flags))
263			sock_hold(sk);
264	}
265	bh_unlock_sock(sk);
266	sock_put(sk);
267}
268
269static void tcp_probe_timer(struct sock *sk)
270{
271	struct inet_connection_sock *icsk = inet_csk(sk);
272	struct tcp_sock *tp = tcp_sk(sk);
273	int max_probes;
 
274
275	if (tp->packets_out || !tcp_send_head(sk)) {
276		icsk->icsk_probes_out = 0;
277		return;
278	}
279
280	/* *WARNING* RFC 1122 forbids this
281	 *
282	 * It doesn't AFAIK, because we kill the retransmit timer -AK
283	 *
284	 * FIXME: We ought not to do it, Solaris 2.5 actually has fixing
285	 * this behaviour in Solaris down as a bug fix. [AC]
286	 *
287	 * Let me to explain. icsk_probes_out is zeroed by incoming ACKs
288	 * even if they advertise zero window. Hence, connection is killed only
289	 * if we received no ACKs for normal connection timeout. It is not killed
290	 * only because window stays zero for some time, window may be zero
291	 * until armageddon and even later. We are in full accordance
292	 * with RFCs, only probe timer combines both retransmission timeout
293	 * and probe timeout in one bottle.				--ANK
294	 */
295	max_probes = sysctl_tcp_retries2;
 
 
 
 
 
296
 
297	if (sock_flag(sk, SOCK_DEAD)) {
298		const int alive = ((icsk->icsk_rto << icsk->icsk_backoff) < TCP_RTO_MAX);
299
300		max_probes = tcp_orphan_retries(sk, alive);
301
302		if (tcp_out_of_resources(sk, alive || icsk->icsk_probes_out <= max_probes))
 
303			return;
304	}
305
306	if (icsk->icsk_probes_out > max_probes) {
307		tcp_write_err(sk);
308	} else {
309		/* Only send another probe if we didn't close things up. */
310		tcp_send_probe0(sk);
311	}
312}
313
314/*
315 *	Timer for Fast Open socket to retransmit SYNACK. Note that the
316 *	sk here is the child socket, not the parent (listener) socket.
317 */
318static void tcp_fastopen_synack_timer(struct sock *sk)
319{
320	struct inet_connection_sock *icsk = inet_csk(sk);
321	int max_retries = icsk->icsk_syn_retries ? :
322	    sysctl_tcp_synack_retries + 1; /* add one more retry for fastopen */
323	struct request_sock *req;
324
325	req = tcp_sk(sk)->fastopen_rsk;
326	req->rsk_ops->syn_ack_timeout(sk, req);
327
328	if (req->num_timeout >= max_retries) {
329		tcp_write_err(sk);
330		return;
331	}
332	/* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
333	 * returned from rtx_syn_ack() to make it more persistent like
334	 * regular retransmit because if the child socket has been accepted
335	 * it's not good to give up too easily.
336	 */
337	inet_rtx_syn_ack(sk, req);
338	req->num_timeout++;
 
339	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
340			  TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
341}
342
343/*
344 *	The TCP retransmit timer.
345 */
346
 
 
 
 
 
 
 
 
 
 
 
347void tcp_retransmit_timer(struct sock *sk)
348{
349	struct tcp_sock *tp = tcp_sk(sk);
 
350	struct inet_connection_sock *icsk = inet_csk(sk);
351
352	if (tp->fastopen_rsk) {
353		WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
354			     sk->sk_state != TCP_FIN_WAIT1);
355		tcp_fastopen_synack_timer(sk);
356		/* Before we receive ACK to our SYN-ACK don't retransmit
357		 * anything else (e.g., data or FIN segments).
358		 */
359		return;
360	}
361	if (!tp->packets_out)
362		goto out;
363
364	WARN_ON(tcp_write_queue_empty(sk));
365
366	tp->tlp_high_seq = 0;
367
368	if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
369	    !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
370		/* Receiver dastardly shrinks window. Our retransmits
371		 * become zero probes, but we should not timeout this
372		 * connection. If the socket is an orphan, time it out,
373		 * we cannot allow such beasts to hang infinitely.
374		 */
375		struct inet_sock *inet = inet_sk(sk);
376		if (sk->sk_family == AF_INET) {
377			LIMIT_NETDEBUG(KERN_DEBUG pr_fmt("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n"),
378				       &inet->inet_daddr,
379				       ntohs(inet->inet_dport), inet->inet_num,
380				       tp->snd_una, tp->snd_nxt);
 
381		}
382#if IS_ENABLED(CONFIG_IPV6)
383		else if (sk->sk_family == AF_INET6) {
384			LIMIT_NETDEBUG(KERN_DEBUG pr_fmt("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n"),
385				       &sk->sk_v6_daddr,
386				       ntohs(inet->inet_dport), inet->inet_num,
387				       tp->snd_una, tp->snd_nxt);
 
388		}
389#endif
390		if (tcp_time_stamp - tp->rcv_tstamp > TCP_RTO_MAX) {
391			tcp_write_err(sk);
392			goto out;
393		}
394		tcp_enter_loss(sk, 0);
395		tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
396		__sk_dst_reset(sk);
397		goto out_reset_timer;
398	}
399
400	if (tcp_write_timeout(sk))
401		goto out;
402
403	if (icsk->icsk_retransmits == 0) {
404		int mib_idx;
405
406		if (icsk->icsk_ca_state == TCP_CA_Recovery) {
407			if (tcp_is_sack(tp))
408				mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
409			else
410				mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
411		} else if (icsk->icsk_ca_state == TCP_CA_Loss) {
412			mib_idx = LINUX_MIB_TCPLOSSFAILURES;
413		} else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
414			   tp->sacked_out) {
415			if (tcp_is_sack(tp))
416				mib_idx = LINUX_MIB_TCPSACKFAILURES;
417			else
418				mib_idx = LINUX_MIB_TCPRENOFAILURES;
419		} else {
420			mib_idx = LINUX_MIB_TCPTIMEOUTS;
421		}
422		NET_INC_STATS_BH(sock_net(sk), mib_idx);
423	}
424
425	tcp_enter_loss(sk, 0);
426
427	if (tcp_retransmit_skb(sk, tcp_write_queue_head(sk)) > 0) {
428		/* Retransmission failed because of local congestion,
429		 * do not backoff.
430		 */
431		if (!icsk->icsk_retransmits)
432			icsk->icsk_retransmits = 1;
433		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
434					  min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL),
435					  TCP_RTO_MAX);
436		goto out;
437	}
438
439	/* Increase the timeout each time we retransmit.  Note that
440	 * we do not increase the rtt estimate.  rto is initialized
441	 * from rtt, but increases here.  Jacobson (SIGCOMM 88) suggests
442	 * that doubling rto each time is the least we can get away with.
443	 * In KA9Q, Karn uses this for the first few times, and then
444	 * goes to quadratic.  netBSD doubles, but only goes up to *64,
445	 * and clamps at 1 to 64 sec afterwards.  Note that 120 sec is
446	 * defined in the protocol as the maximum possible RTT.  I guess
447	 * we'll have to use something other than TCP to talk to the
448	 * University of Mars.
449	 *
450	 * PAWS allows us longer timeouts and large windows, so once
451	 * implemented ftp to mars will work nicely. We will have to fix
452	 * the 120 second clamps though!
453	 */
454	icsk->icsk_backoff++;
455	icsk->icsk_retransmits++;
456
457out_reset_timer:
458	/* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
459	 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
460	 * might be increased if the stream oscillates between thin and thick,
461	 * thus the old value might already be too high compared to the value
462	 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
463	 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
464	 * exponential backoff behaviour to avoid continue hammering
465	 * linear-timeout retransmissions into a black hole
466	 */
467	if (sk->sk_state == TCP_ESTABLISHED &&
468	    (tp->thin_lto || sysctl_tcp_thin_linear_timeouts) &&
469	    tcp_stream_is_thin(tp) &&
470	    icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
471		icsk->icsk_backoff = 0;
472		icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX);
473	} else {
474		/* Use normal (exponential) backoff */
475		icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
476	}
477	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX);
478	if (retransmits_timed_out(sk, sysctl_tcp_retries1 + 1, 0, 0))
479		__sk_dst_reset(sk);
480
481out:;
482}
483
 
 
484void tcp_write_timer_handler(struct sock *sk)
485{
486	struct inet_connection_sock *icsk = inet_csk(sk);
487	int event;
488
489	if (sk->sk_state == TCP_CLOSE || !icsk->icsk_pending)
 
490		goto out;
491
492	if (time_after(icsk->icsk_timeout, jiffies)) {
493		sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
494		goto out;
495	}
496
497	event = icsk->icsk_pending;
498
499	switch (event) {
500	case ICSK_TIME_EARLY_RETRANS:
501		tcp_resume_early_retransmit(sk);
502		break;
503	case ICSK_TIME_LOSS_PROBE:
504		tcp_send_loss_probe(sk);
505		break;
506	case ICSK_TIME_RETRANS:
507		icsk->icsk_pending = 0;
508		tcp_retransmit_timer(sk);
509		break;
510	case ICSK_TIME_PROBE0:
511		icsk->icsk_pending = 0;
512		tcp_probe_timer(sk);
513		break;
514	}
515
516out:
517	sk_mem_reclaim(sk);
518}
519
520static void tcp_write_timer(unsigned long data)
521{
522	struct sock *sk = (struct sock *)data;
523
524	bh_lock_sock(sk);
525	if (!sock_owned_by_user(sk)) {
526		tcp_write_timer_handler(sk);
527	} else {
528		/* deleguate our work to tcp_release_cb() */
529		if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &tcp_sk(sk)->tsq_flags))
530			sock_hold(sk);
531	}
532	bh_unlock_sock(sk);
533	sock_put(sk);
534}
535
536/*
537 *	Timer for listening sockets
538 */
539
540static void tcp_synack_timer(struct sock *sk)
541{
542	inet_csk_reqsk_queue_prune(sk, TCP_SYNQ_INTERVAL,
543				   TCP_TIMEOUT_INIT, TCP_RTO_MAX);
544}
545
546void tcp_syn_ack_timeout(struct sock *sk, struct request_sock *req)
547{
548	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPTIMEOUTS);
549}
550EXPORT_SYMBOL(tcp_syn_ack_timeout);
551
552void tcp_set_keepalive(struct sock *sk, int val)
553{
554	if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
555		return;
556
557	if (val && !sock_flag(sk, SOCK_KEEPOPEN))
558		inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
559	else if (!val)
560		inet_csk_delete_keepalive_timer(sk);
561}
562
563
564static void tcp_keepalive_timer (unsigned long data)
565{
566	struct sock *sk = (struct sock *) data;
567	struct inet_connection_sock *icsk = inet_csk(sk);
568	struct tcp_sock *tp = tcp_sk(sk);
569	u32 elapsed;
570
571	/* Only process if socket is not in use. */
572	bh_lock_sock(sk);
573	if (sock_owned_by_user(sk)) {
574		/* Try again later. */
575		inet_csk_reset_keepalive_timer (sk, HZ/20);
576		goto out;
577	}
578
579	if (sk->sk_state == TCP_LISTEN) {
580		tcp_synack_timer(sk);
581		goto out;
582	}
583
584	if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
585		if (tp->linger2 >= 0) {
586			const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
587
588			if (tmo > 0) {
589				tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
590				goto out;
591			}
592		}
593		tcp_send_active_reset(sk, GFP_ATOMIC);
594		goto death;
595	}
596
597	if (!sock_flag(sk, SOCK_KEEPOPEN) || sk->sk_state == TCP_CLOSE)
598		goto out;
599
600	elapsed = keepalive_time_when(tp);
601
602	/* It is alive without keepalive 8) */
603	if (tp->packets_out || tcp_send_head(sk))
604		goto resched;
605
606	elapsed = keepalive_time_elapsed(tp);
607
608	if (elapsed >= keepalive_time_when(tp)) {
609		/* If the TCP_USER_TIMEOUT option is enabled, use that
610		 * to determine when to timeout instead.
611		 */
612		if ((icsk->icsk_user_timeout != 0 &&
613		    elapsed >= icsk->icsk_user_timeout &&
614		    icsk->icsk_probes_out > 0) ||
615		    (icsk->icsk_user_timeout == 0 &&
616		    icsk->icsk_probes_out >= keepalive_probes(tp))) {
617			tcp_send_active_reset(sk, GFP_ATOMIC);
618			tcp_write_err(sk);
619			goto out;
620		}
621		if (tcp_write_wakeup(sk) <= 0) {
622			icsk->icsk_probes_out++;
623			elapsed = keepalive_intvl_when(tp);
624		} else {
625			/* If keepalive was lost due to local congestion,
626			 * try harder.
627			 */
628			elapsed = TCP_RESOURCE_PROBE_INTERVAL;
629		}
630	} else {
631		/* It is tp->rcv_tstamp + keepalive_time_when(tp) */
632		elapsed = keepalive_time_when(tp) - elapsed;
633	}
634
635	sk_mem_reclaim(sk);
636
637resched:
638	inet_csk_reset_keepalive_timer (sk, elapsed);
639	goto out;
640
641death:
642	tcp_done(sk);
643
644out:
645	bh_unlock_sock(sk);
646	sock_put(sk);
647}
648
649void tcp_init_xmit_timers(struct sock *sk)
650{
651	inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
652				  &tcp_keepalive_timer);
653}
654EXPORT_SYMBOL(tcp_init_xmit_timers);
v4.10.11
  1/*
  2 * INET		An implementation of the TCP/IP protocol suite for the LINUX
  3 *		operating system.  INET is implemented using the  BSD Socket
  4 *		interface as the means of communication with the user level.
  5 *
  6 *		Implementation of the Transmission Control Protocol(TCP).
  7 *
  8 * Authors:	Ross Biro
  9 *		Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
 10 *		Mark Evans, <evansmp@uhura.aston.ac.uk>
 11 *		Corey Minyard <wf-rch!minyard@relay.EU.net>
 12 *		Florian La Roche, <flla@stud.uni-sb.de>
 13 *		Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
 14 *		Linus Torvalds, <torvalds@cs.helsinki.fi>
 15 *		Alan Cox, <gw4pts@gw4pts.ampr.org>
 16 *		Matthew Dillon, <dillon@apollo.west.oic.com>
 17 *		Arnt Gulbrandsen, <agulbra@nvg.unit.no>
 18 *		Jorge Cwik, <jorge@laser.satlink.net>
 19 */
 20
 21#include <linux/module.h>
 22#include <linux/gfp.h>
 23#include <net/tcp.h>
 24
 
 
 
 
 
 
 
 
 25int sysctl_tcp_thin_linear_timeouts __read_mostly;
 26
 27/**
 28 *  tcp_write_err() - close socket and save error info
 29 *  @sk:  The socket the error has appeared on.
 30 *
 31 *  Returns: Nothing (void)
 32 */
 33
 34static void tcp_write_err(struct sock *sk)
 35{
 36	sk->sk_err = sk->sk_err_soft ? : ETIMEDOUT;
 37	sk->sk_error_report(sk);
 38
 39	tcp_done(sk);
 40	__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT);
 41}
 42
 43/**
 44 *  tcp_out_of_resources() - Close socket if out of resources
 45 *  @sk:        pointer to current socket
 46 *  @do_reset:  send a last packet with reset flag
 47 *
 48 *  Do not allow orphaned sockets to eat all our resources.
 49 *  This is direct violation of TCP specs, but it is required
 50 *  to prevent DoS attacks. It is called when a retransmission timeout
 51 *  or zero probe timeout occurs on orphaned socket.
 52 *
 53 *  Criteria is still not confirmed experimentally and may change.
 54 *  We kill the socket, if:
 55 *  1. If number of orphaned sockets exceeds an administratively configured
 56 *     limit.
 57 *  2. If we have strong memory pressure.
 58 */
 59static int tcp_out_of_resources(struct sock *sk, bool do_reset)
 60{
 61	struct tcp_sock *tp = tcp_sk(sk);
 62	int shift = 0;
 63
 64	/* If peer does not open window for long time, or did not transmit
 65	 * anything for long time, penalize it. */
 66	if ((s32)(tcp_time_stamp - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset)
 67		shift++;
 68
 69	/* If some dubious ICMP arrived, penalize even more. */
 70	if (sk->sk_err_soft)
 71		shift++;
 72
 73	if (tcp_check_oom(sk, shift)) {
 74		/* Catch exceptional cases, when connection requires reset.
 75		 *      1. Last segment was sent recently. */
 76		if ((s32)(tcp_time_stamp - tp->lsndtime) <= TCP_TIMEWAIT_LEN ||
 77		    /*  2. Window is closed. */
 78		    (!tp->snd_wnd && !tp->packets_out))
 79			do_reset = true;
 80		if (do_reset)
 81			tcp_send_active_reset(sk, GFP_ATOMIC);
 82		tcp_done(sk);
 83		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY);
 84		return 1;
 85	}
 86	return 0;
 87}
 88
 89/**
 90 *  tcp_orphan_retries() - Returns maximal number of retries on an orphaned socket
 91 *  @sk:    Pointer to the current socket.
 92 *  @alive: bool, socket alive state
 93 */
 94static int tcp_orphan_retries(struct sock *sk, bool alive)
 95{
 96	int retries = sock_net(sk)->ipv4.sysctl_tcp_orphan_retries; /* May be zero. */
 97
 98	/* We know from an ICMP that something is wrong. */
 99	if (sk->sk_err_soft && !alive)
100		retries = 0;
101
102	/* However, if socket sent something recently, select some safe
103	 * number of retries. 8 corresponds to >100 seconds with minimal
104	 * RTO of 200msec. */
105	if (retries == 0 && alive)
106		retries = 8;
107	return retries;
108}
109
110static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk)
111{
112	struct net *net = sock_net(sk);
113
114	/* Black hole detection */
115	if (net->ipv4.sysctl_tcp_mtu_probing) {
116		if (!icsk->icsk_mtup.enabled) {
117			icsk->icsk_mtup.enabled = 1;
118			icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
119			tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
120		} else {
121			struct net *net = sock_net(sk);
122			struct tcp_sock *tp = tcp_sk(sk);
123			int mss;
124
125			mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1;
126			mss = min(net->ipv4.sysctl_tcp_base_mss, mss);
127			mss = max(mss, 68 - tp->tcp_header_len);
128			icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
129			tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
130		}
131	}
132}
133
134
135/**
136 *  retransmits_timed_out() - returns true if this connection has timed out
137 *  @sk:       The current socket
138 *  @boundary: max number of retransmissions
139 *  @timeout:  A custom timeout value.
140 *             If set to 0 the default timeout is calculated and used.
141 *             Using TCP_RTO_MIN and the number of unsuccessful retransmits.
142 *  @syn_set:  true if the SYN Bit was set.
143 *
144 * The default "timeout" value this function can calculate and use
145 * is equivalent to the timeout of a TCP Connection
146 * after "boundary" unsuccessful, exponentially backed-off
147 * retransmissions with an initial RTO of TCP_RTO_MIN or TCP_TIMEOUT_INIT if
148 * syn_set flag is set.
149 *
150 */
151static bool retransmits_timed_out(struct sock *sk,
152				  unsigned int boundary,
153				  unsigned int timeout,
154				  bool syn_set)
155{
156	unsigned int linear_backoff_thresh, start_ts;
157	unsigned int rto_base = syn_set ? TCP_TIMEOUT_INIT : TCP_RTO_MIN;
158
159	if (!inet_csk(sk)->icsk_retransmits)
160		return false;
161
162	start_ts = tcp_sk(sk)->retrans_stamp;
163	if (unlikely(!start_ts))
164		start_ts = tcp_skb_timestamp(tcp_write_queue_head(sk));
 
165
166	if (likely(timeout == 0)) {
167		linear_backoff_thresh = ilog2(TCP_RTO_MAX/rto_base);
168
169		if (boundary <= linear_backoff_thresh)
170			timeout = ((2 << boundary) - 1) * rto_base;
171		else
172			timeout = ((2 << linear_backoff_thresh) - 1) * rto_base +
173				(boundary - linear_backoff_thresh) * TCP_RTO_MAX;
174	}
175	return (tcp_time_stamp - start_ts) >= timeout;
176}
177
178/* A write timeout has occurred. Process the after effects. */
179static int tcp_write_timeout(struct sock *sk)
180{
181	struct inet_connection_sock *icsk = inet_csk(sk);
182	struct tcp_sock *tp = tcp_sk(sk);
183	struct net *net = sock_net(sk);
184	int retry_until;
185	bool do_reset, syn_set = false;
186
187	if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
188		if (icsk->icsk_retransmits) {
189			dst_negative_advice(sk);
190			if (tp->syn_fastopen || tp->syn_data)
191				tcp_fastopen_cache_set(sk, 0, NULL, true, 0);
192			if (tp->syn_data && icsk->icsk_retransmits == 1)
193				NET_INC_STATS(sock_net(sk),
194					      LINUX_MIB_TCPFASTOPENACTIVEFAIL);
195		} else if (!tp->syn_data && !tp->syn_fastopen) {
196			sk_rethink_txhash(sk);
197		}
198		retry_until = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
199		syn_set = true;
200	} else {
201		if (retransmits_timed_out(sk, net->ipv4.sysctl_tcp_retries1, 0, 0)) {
202			/* Some middle-boxes may black-hole Fast Open _after_
203			 * the handshake. Therefore we conservatively disable
204			 * Fast Open on this path on recurring timeouts with
205			 * few or zero bytes acked after Fast Open.
206			 */
207			if (tp->syn_data_acked &&
208			    tp->bytes_acked <= tp->rx_opt.mss_clamp) {
209				tcp_fastopen_cache_set(sk, 0, NULL, true, 0);
210				if (icsk->icsk_retransmits == net->ipv4.sysctl_tcp_retries1)
211					NET_INC_STATS(sock_net(sk),
212						      LINUX_MIB_TCPFASTOPENACTIVEFAIL);
213			}
214			/* Black hole detection */
215			tcp_mtu_probing(icsk, sk);
216
217			dst_negative_advice(sk);
218		} else {
219			sk_rethink_txhash(sk);
220		}
221
222		retry_until = net->ipv4.sysctl_tcp_retries2;
223		if (sock_flag(sk, SOCK_DEAD)) {
224			const bool alive = icsk->icsk_rto < TCP_RTO_MAX;
225
226			retry_until = tcp_orphan_retries(sk, alive);
227			do_reset = alive ||
228				!retransmits_timed_out(sk, retry_until, 0, 0);
229
230			if (tcp_out_of_resources(sk, do_reset))
231				return 1;
232		}
233	}
234
235	if (retransmits_timed_out(sk, retry_until,
236				  syn_set ? 0 : icsk->icsk_user_timeout, syn_set)) {
237		/* Has it gone just too far? */
238		tcp_write_err(sk);
239		return 1;
240	}
241	return 0;
242}
243
244/* Called with BH disabled */
245void tcp_delack_timer_handler(struct sock *sk)
246{
247	struct tcp_sock *tp = tcp_sk(sk);
248	struct inet_connection_sock *icsk = inet_csk(sk);
249
250	sk_mem_reclaim_partial(sk);
251
252	if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
253	    !(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
254		goto out;
255
256	if (time_after(icsk->icsk_ack.timeout, jiffies)) {
257		sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
258		goto out;
259	}
260	icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
261
262	if (!skb_queue_empty(&tp->ucopy.prequeue)) {
263		struct sk_buff *skb;
264
265		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSCHEDULERFAILED);
266
267		while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
268			sk_backlog_rcv(sk, skb);
269
270		tp->ucopy.memory = 0;
271	}
272
273	if (inet_csk_ack_scheduled(sk)) {
274		if (!icsk->icsk_ack.pingpong) {
275			/* Delayed ACK missed: inflate ATO. */
276			icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto);
277		} else {
278			/* Delayed ACK missed: leave pingpong mode and
279			 * deflate ATO.
280			 */
281			icsk->icsk_ack.pingpong = 0;
282			icsk->icsk_ack.ato      = TCP_ATO_MIN;
283		}
284		tcp_send_ack(sk);
285		__NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKS);
286	}
287
288out:
289	if (tcp_under_memory_pressure(sk))
290		sk_mem_reclaim(sk);
291}
292
293
294/**
295 *  tcp_delack_timer() - The TCP delayed ACK timeout handler
296 *  @data:  Pointer to the current socket. (gets casted to struct sock *)
297 *
298 *  This function gets (indirectly) called when the kernel timer for a TCP packet
299 *  of this socket expires. Calls tcp_delack_timer_handler() to do the actual work.
300 *
301 *  Returns: Nothing (void)
302 */
303static void tcp_delack_timer(unsigned long data)
304{
305	struct sock *sk = (struct sock *)data;
306
307	bh_lock_sock(sk);
308	if (!sock_owned_by_user(sk)) {
309		tcp_delack_timer_handler(sk);
310	} else {
311		inet_csk(sk)->icsk_ack.blocked = 1;
312		__NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
313		/* deleguate our work to tcp_release_cb() */
314		if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &sk->sk_tsq_flags))
315			sock_hold(sk);
316	}
317	bh_unlock_sock(sk);
318	sock_put(sk);
319}
320
321static void tcp_probe_timer(struct sock *sk)
322{
323	struct inet_connection_sock *icsk = inet_csk(sk);
324	struct tcp_sock *tp = tcp_sk(sk);
325	int max_probes;
326	u32 start_ts;
327
328	if (tp->packets_out || !tcp_send_head(sk)) {
329		icsk->icsk_probes_out = 0;
330		return;
331	}
332
333	/* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as
334	 * long as the receiver continues to respond probes. We support this by
335	 * default and reset icsk_probes_out with incoming ACKs. But if the
336	 * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we
337	 * kill the socket when the retry count and the time exceeds the
338	 * corresponding system limit. We also implement similar policy when
339	 * we use RTO to probe window in tcp_retransmit_timer().
 
 
 
 
 
 
 
340	 */
341	start_ts = tcp_skb_timestamp(tcp_send_head(sk));
342	if (!start_ts)
343		skb_mstamp_get(&tcp_send_head(sk)->skb_mstamp);
344	else if (icsk->icsk_user_timeout &&
345		 (s32)(tcp_time_stamp - start_ts) > icsk->icsk_user_timeout)
346		goto abort;
347
348	max_probes = sock_net(sk)->ipv4.sysctl_tcp_retries2;
349	if (sock_flag(sk, SOCK_DEAD)) {
350		const bool alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX;
351
352		max_probes = tcp_orphan_retries(sk, alive);
353		if (!alive && icsk->icsk_backoff >= max_probes)
354			goto abort;
355		if (tcp_out_of_resources(sk, true))
356			return;
357	}
358
359	if (icsk->icsk_probes_out > max_probes) {
360abort:		tcp_write_err(sk);
361	} else {
362		/* Only send another probe if we didn't close things up. */
363		tcp_send_probe0(sk);
364	}
365}
366
367/*
368 *	Timer for Fast Open socket to retransmit SYNACK. Note that the
369 *	sk here is the child socket, not the parent (listener) socket.
370 */
371static void tcp_fastopen_synack_timer(struct sock *sk)
372{
373	struct inet_connection_sock *icsk = inet_csk(sk);
374	int max_retries = icsk->icsk_syn_retries ? :
375	    sock_net(sk)->ipv4.sysctl_tcp_synack_retries + 1; /* add one more retry for fastopen */
376	struct request_sock *req;
377
378	req = tcp_sk(sk)->fastopen_rsk;
379	req->rsk_ops->syn_ack_timeout(req);
380
381	if (req->num_timeout >= max_retries) {
382		tcp_write_err(sk);
383		return;
384	}
385	/* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
386	 * returned from rtx_syn_ack() to make it more persistent like
387	 * regular retransmit because if the child socket has been accepted
388	 * it's not good to give up too easily.
389	 */
390	inet_rtx_syn_ack(sk, req);
391	req->num_timeout++;
392	icsk->icsk_retransmits++;
393	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
394			  TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
395}
396
 
 
 
397
398/**
399 *  tcp_retransmit_timer() - The TCP retransmit timeout handler
400 *  @sk:  Pointer to the current socket.
401 *
402 *  This function gets called when the kernel timer for a TCP packet
403 *  of this socket expires.
404 *
405 *  It handles retransmission, timer adjustment and other necesarry measures.
406 *
407 *  Returns: Nothing (void)
408 */
409void tcp_retransmit_timer(struct sock *sk)
410{
411	struct tcp_sock *tp = tcp_sk(sk);
412	struct net *net = sock_net(sk);
413	struct inet_connection_sock *icsk = inet_csk(sk);
414
415	if (tp->fastopen_rsk) {
416		WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
417			     sk->sk_state != TCP_FIN_WAIT1);
418		tcp_fastopen_synack_timer(sk);
419		/* Before we receive ACK to our SYN-ACK don't retransmit
420		 * anything else (e.g., data or FIN segments).
421		 */
422		return;
423	}
424	if (!tp->packets_out)
425		goto out;
426
427	WARN_ON(tcp_write_queue_empty(sk));
428
429	tp->tlp_high_seq = 0;
430
431	if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
432	    !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
433		/* Receiver dastardly shrinks window. Our retransmits
434		 * become zero probes, but we should not timeout this
435		 * connection. If the socket is an orphan, time it out,
436		 * we cannot allow such beasts to hang infinitely.
437		 */
438		struct inet_sock *inet = inet_sk(sk);
439		if (sk->sk_family == AF_INET) {
440			net_dbg_ratelimited("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
441					    &inet->inet_daddr,
442					    ntohs(inet->inet_dport),
443					    inet->inet_num,
444					    tp->snd_una, tp->snd_nxt);
445		}
446#if IS_ENABLED(CONFIG_IPV6)
447		else if (sk->sk_family == AF_INET6) {
448			net_dbg_ratelimited("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
449					    &sk->sk_v6_daddr,
450					    ntohs(inet->inet_dport),
451					    inet->inet_num,
452					    tp->snd_una, tp->snd_nxt);
453		}
454#endif
455		if (tcp_time_stamp - tp->rcv_tstamp > TCP_RTO_MAX) {
456			tcp_write_err(sk);
457			goto out;
458		}
459		tcp_enter_loss(sk);
460		tcp_retransmit_skb(sk, tcp_write_queue_head(sk), 1);
461		__sk_dst_reset(sk);
462		goto out_reset_timer;
463	}
464
465	if (tcp_write_timeout(sk))
466		goto out;
467
468	if (icsk->icsk_retransmits == 0) {
469		int mib_idx;
470
471		if (icsk->icsk_ca_state == TCP_CA_Recovery) {
472			if (tcp_is_sack(tp))
473				mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
474			else
475				mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
476		} else if (icsk->icsk_ca_state == TCP_CA_Loss) {
477			mib_idx = LINUX_MIB_TCPLOSSFAILURES;
478		} else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
479			   tp->sacked_out) {
480			if (tcp_is_sack(tp))
481				mib_idx = LINUX_MIB_TCPSACKFAILURES;
482			else
483				mib_idx = LINUX_MIB_TCPRENOFAILURES;
484		} else {
485			mib_idx = LINUX_MIB_TCPTIMEOUTS;
486		}
487		__NET_INC_STATS(sock_net(sk), mib_idx);
488	}
489
490	tcp_enter_loss(sk);
491
492	if (tcp_retransmit_skb(sk, tcp_write_queue_head(sk), 1) > 0) {
493		/* Retransmission failed because of local congestion,
494		 * do not backoff.
495		 */
496		if (!icsk->icsk_retransmits)
497			icsk->icsk_retransmits = 1;
498		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
499					  min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL),
500					  TCP_RTO_MAX);
501		goto out;
502	}
503
504	/* Increase the timeout each time we retransmit.  Note that
505	 * we do not increase the rtt estimate.  rto is initialized
506	 * from rtt, but increases here.  Jacobson (SIGCOMM 88) suggests
507	 * that doubling rto each time is the least we can get away with.
508	 * In KA9Q, Karn uses this for the first few times, and then
509	 * goes to quadratic.  netBSD doubles, but only goes up to *64,
510	 * and clamps at 1 to 64 sec afterwards.  Note that 120 sec is
511	 * defined in the protocol as the maximum possible RTT.  I guess
512	 * we'll have to use something other than TCP to talk to the
513	 * University of Mars.
514	 *
515	 * PAWS allows us longer timeouts and large windows, so once
516	 * implemented ftp to mars will work nicely. We will have to fix
517	 * the 120 second clamps though!
518	 */
519	icsk->icsk_backoff++;
520	icsk->icsk_retransmits++;
521
522out_reset_timer:
523	/* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
524	 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
525	 * might be increased if the stream oscillates between thin and thick,
526	 * thus the old value might already be too high compared to the value
527	 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
528	 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
529	 * exponential backoff behaviour to avoid continue hammering
530	 * linear-timeout retransmissions into a black hole
531	 */
532	if (sk->sk_state == TCP_ESTABLISHED &&
533	    (tp->thin_lto || sysctl_tcp_thin_linear_timeouts) &&
534	    tcp_stream_is_thin(tp) &&
535	    icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
536		icsk->icsk_backoff = 0;
537		icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX);
538	} else {
539		/* Use normal (exponential) backoff */
540		icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
541	}
542	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX);
543	if (retransmits_timed_out(sk, net->ipv4.sysctl_tcp_retries1 + 1, 0, 0))
544		__sk_dst_reset(sk);
545
546out:;
547}
548
549/* Called with bottom-half processing disabled.
550   Called by tcp_write_timer() */
551void tcp_write_timer_handler(struct sock *sk)
552{
553	struct inet_connection_sock *icsk = inet_csk(sk);
554	int event;
555
556	if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
557	    !icsk->icsk_pending)
558		goto out;
559
560	if (time_after(icsk->icsk_timeout, jiffies)) {
561		sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
562		goto out;
563	}
564
565	event = icsk->icsk_pending;
566
567	switch (event) {
568	case ICSK_TIME_EARLY_RETRANS:
569		tcp_resume_early_retransmit(sk);
570		break;
571	case ICSK_TIME_LOSS_PROBE:
572		tcp_send_loss_probe(sk);
573		break;
574	case ICSK_TIME_RETRANS:
575		icsk->icsk_pending = 0;
576		tcp_retransmit_timer(sk);
577		break;
578	case ICSK_TIME_PROBE0:
579		icsk->icsk_pending = 0;
580		tcp_probe_timer(sk);
581		break;
582	}
583
584out:
585	sk_mem_reclaim(sk);
586}
587
588static void tcp_write_timer(unsigned long data)
589{
590	struct sock *sk = (struct sock *)data;
591
592	bh_lock_sock(sk);
593	if (!sock_owned_by_user(sk)) {
594		tcp_write_timer_handler(sk);
595	} else {
596		/* delegate our work to tcp_release_cb() */
597		if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &sk->sk_tsq_flags))
598			sock_hold(sk);
599	}
600	bh_unlock_sock(sk);
601	sock_put(sk);
602}
603
604void tcp_syn_ack_timeout(const struct request_sock *req)
 
 
 
 
605{
606	struct net *net = read_pnet(&inet_rsk(req)->ireq_net);
 
 
607
608	__NET_INC_STATS(net, LINUX_MIB_TCPTIMEOUTS);
 
 
609}
610EXPORT_SYMBOL(tcp_syn_ack_timeout);
611
612void tcp_set_keepalive(struct sock *sk, int val)
613{
614	if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
615		return;
616
617	if (val && !sock_flag(sk, SOCK_KEEPOPEN))
618		inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
619	else if (!val)
620		inet_csk_delete_keepalive_timer(sk);
621}
622
623
624static void tcp_keepalive_timer (unsigned long data)
625{
626	struct sock *sk = (struct sock *) data;
627	struct inet_connection_sock *icsk = inet_csk(sk);
628	struct tcp_sock *tp = tcp_sk(sk);
629	u32 elapsed;
630
631	/* Only process if socket is not in use. */
632	bh_lock_sock(sk);
633	if (sock_owned_by_user(sk)) {
634		/* Try again later. */
635		inet_csk_reset_keepalive_timer (sk, HZ/20);
636		goto out;
637	}
638
639	if (sk->sk_state == TCP_LISTEN) {
640		pr_err("Hmm... keepalive on a LISTEN ???\n");
641		goto out;
642	}
643
644	if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
645		if (tp->linger2 >= 0) {
646			const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
647
648			if (tmo > 0) {
649				tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
650				goto out;
651			}
652		}
653		tcp_send_active_reset(sk, GFP_ATOMIC);
654		goto death;
655	}
656
657	if (!sock_flag(sk, SOCK_KEEPOPEN) || sk->sk_state == TCP_CLOSE)
658		goto out;
659
660	elapsed = keepalive_time_when(tp);
661
662	/* It is alive without keepalive 8) */
663	if (tp->packets_out || tcp_send_head(sk))
664		goto resched;
665
666	elapsed = keepalive_time_elapsed(tp);
667
668	if (elapsed >= keepalive_time_when(tp)) {
669		/* If the TCP_USER_TIMEOUT option is enabled, use that
670		 * to determine when to timeout instead.
671		 */
672		if ((icsk->icsk_user_timeout != 0 &&
673		    elapsed >= icsk->icsk_user_timeout &&
674		    icsk->icsk_probes_out > 0) ||
675		    (icsk->icsk_user_timeout == 0 &&
676		    icsk->icsk_probes_out >= keepalive_probes(tp))) {
677			tcp_send_active_reset(sk, GFP_ATOMIC);
678			tcp_write_err(sk);
679			goto out;
680		}
681		if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) {
682			icsk->icsk_probes_out++;
683			elapsed = keepalive_intvl_when(tp);
684		} else {
685			/* If keepalive was lost due to local congestion,
686			 * try harder.
687			 */
688			elapsed = TCP_RESOURCE_PROBE_INTERVAL;
689		}
690	} else {
691		/* It is tp->rcv_tstamp + keepalive_time_when(tp) */
692		elapsed = keepalive_time_when(tp) - elapsed;
693	}
694
695	sk_mem_reclaim(sk);
696
697resched:
698	inet_csk_reset_keepalive_timer (sk, elapsed);
699	goto out;
700
701death:
702	tcp_done(sk);
703
704out:
705	bh_unlock_sock(sk);
706	sock_put(sk);
707}
708
709void tcp_init_xmit_timers(struct sock *sk)
710{
711	inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
712				  &tcp_keepalive_timer);
713}