xref: /openbmc/linux/net/ipv4/tcp_timer.c (revision b627b4ed)
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 <net/tcp.h>
23 
24 int sysctl_tcp_syn_retries __read_mostly = TCP_SYN_RETRIES;
25 int sysctl_tcp_synack_retries __read_mostly = TCP_SYNACK_RETRIES;
26 int sysctl_tcp_keepalive_time __read_mostly = TCP_KEEPALIVE_TIME;
27 int sysctl_tcp_keepalive_probes __read_mostly = TCP_KEEPALIVE_PROBES;
28 int sysctl_tcp_keepalive_intvl __read_mostly = TCP_KEEPALIVE_INTVL;
29 int sysctl_tcp_retries1 __read_mostly = TCP_RETR1;
30 int sysctl_tcp_retries2 __read_mostly = TCP_RETR2;
31 int sysctl_tcp_orphan_retries __read_mostly;
32 
33 static void tcp_write_timer(unsigned long);
34 static void tcp_delack_timer(unsigned long);
35 static void tcp_keepalive_timer (unsigned long data);
36 
37 void tcp_init_xmit_timers(struct sock *sk)
38 {
39 	inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
40 				  &tcp_keepalive_timer);
41 }
42 
43 EXPORT_SYMBOL(tcp_init_xmit_timers);
44 
45 static void tcp_write_err(struct sock *sk)
46 {
47 	sk->sk_err = sk->sk_err_soft ? : ETIMEDOUT;
48 	sk->sk_error_report(sk);
49 
50 	tcp_done(sk);
51 	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT);
52 }
53 
54 /* Do not allow orphaned sockets to eat all our resources.
55  * This is direct violation of TCP specs, but it is required
56  * to prevent DoS attacks. It is called when a retransmission timeout
57  * or zero probe timeout occurs on orphaned socket.
58  *
59  * Criteria is still not confirmed experimentally and may change.
60  * We kill the socket, if:
61  * 1. If number of orphaned sockets exceeds an administratively configured
62  *    limit.
63  * 2. If we have strong memory pressure.
64  */
65 static int tcp_out_of_resources(struct sock *sk, int do_reset)
66 {
67 	struct tcp_sock *tp = tcp_sk(sk);
68 	int orphans = percpu_counter_read_positive(&tcp_orphan_count);
69 
70 	/* If peer does not open window for long time, or did not transmit
71 	 * anything for long time, penalize it. */
72 	if ((s32)(tcp_time_stamp - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset)
73 		orphans <<= 1;
74 
75 	/* If some dubious ICMP arrived, penalize even more. */
76 	if (sk->sk_err_soft)
77 		orphans <<= 1;
78 
79 	if (tcp_too_many_orphans(sk, orphans)) {
80 		if (net_ratelimit())
81 			printk(KERN_INFO "Out of socket memory\n");
82 
83 		/* Catch exceptional cases, when connection requires reset.
84 		 *      1. Last segment was sent recently. */
85 		if ((s32)(tcp_time_stamp - tp->lsndtime) <= TCP_TIMEWAIT_LEN ||
86 		    /*  2. Window is closed. */
87 		    (!tp->snd_wnd && !tp->packets_out))
88 			do_reset = 1;
89 		if (do_reset)
90 			tcp_send_active_reset(sk, GFP_ATOMIC);
91 		tcp_done(sk);
92 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY);
93 		return 1;
94 	}
95 	return 0;
96 }
97 
98 /* Calculate maximal number or retries on an orphaned socket. */
99 static int tcp_orphan_retries(struct sock *sk, int alive)
100 {
101 	int retries = sysctl_tcp_orphan_retries; /* May be zero. */
102 
103 	/* We know from an ICMP that something is wrong. */
104 	if (sk->sk_err_soft && !alive)
105 		retries = 0;
106 
107 	/* However, if socket sent something recently, select some safe
108 	 * number of retries. 8 corresponds to >100 seconds with minimal
109 	 * RTO of 200msec. */
110 	if (retries == 0 && alive)
111 		retries = 8;
112 	return retries;
113 }
114 
115 static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk)
116 {
117 	/* Black hole detection */
118 	if (sysctl_tcp_mtu_probing) {
119 		if (!icsk->icsk_mtup.enabled) {
120 			icsk->icsk_mtup.enabled = 1;
121 			tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
122 		} else {
123 			struct tcp_sock *tp = tcp_sk(sk);
124 			int mss;
125 
126 			mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1;
127 			mss = min(sysctl_tcp_base_mss, mss);
128 			mss = max(mss, 68 - tp->tcp_header_len);
129 			icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
130 			tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
131 		}
132 	}
133 }
134 
135 /* A write timeout has occurred. Process the after effects. */
136 static int tcp_write_timeout(struct sock *sk)
137 {
138 	struct inet_connection_sock *icsk = inet_csk(sk);
139 	int retry_until;
140 
141 	if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
142 		if (icsk->icsk_retransmits)
143 			dst_negative_advice(&sk->sk_dst_cache);
144 		retry_until = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
145 	} else {
146 		if (icsk->icsk_retransmits >= sysctl_tcp_retries1) {
147 			/* Black hole detection */
148 			tcp_mtu_probing(icsk, sk);
149 
150 			dst_negative_advice(&sk->sk_dst_cache);
151 		}
152 
153 		retry_until = sysctl_tcp_retries2;
154 		if (sock_flag(sk, SOCK_DEAD)) {
155 			const int alive = (icsk->icsk_rto < TCP_RTO_MAX);
156 
157 			retry_until = tcp_orphan_retries(sk, alive);
158 
159 			if (tcp_out_of_resources(sk, alive || icsk->icsk_retransmits < retry_until))
160 				return 1;
161 		}
162 	}
163 
164 	if (icsk->icsk_retransmits >= retry_until) {
165 		/* Has it gone just too far? */
166 		tcp_write_err(sk);
167 		return 1;
168 	}
169 	return 0;
170 }
171 
172 static void tcp_delack_timer(unsigned long data)
173 {
174 	struct sock *sk = (struct sock *)data;
175 	struct tcp_sock *tp = tcp_sk(sk);
176 	struct inet_connection_sock *icsk = inet_csk(sk);
177 
178 	bh_lock_sock(sk);
179 	if (sock_owned_by_user(sk)) {
180 		/* Try again later. */
181 		icsk->icsk_ack.blocked = 1;
182 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
183 		sk_reset_timer(sk, &icsk->icsk_delack_timer, jiffies + TCP_DELACK_MIN);
184 		goto out_unlock;
185 	}
186 
187 	sk_mem_reclaim_partial(sk);
188 
189 	if (sk->sk_state == TCP_CLOSE || !(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
190 		goto out;
191 
192 	if (time_after(icsk->icsk_ack.timeout, jiffies)) {
193 		sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
194 		goto out;
195 	}
196 	icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
197 
198 	if (!skb_queue_empty(&tp->ucopy.prequeue)) {
199 		struct sk_buff *skb;
200 
201 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSCHEDULERFAILED);
202 
203 		while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
204 			sk_backlog_rcv(sk, skb);
205 
206 		tp->ucopy.memory = 0;
207 	}
208 
209 	if (inet_csk_ack_scheduled(sk)) {
210 		if (!icsk->icsk_ack.pingpong) {
211 			/* Delayed ACK missed: inflate ATO. */
212 			icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto);
213 		} else {
214 			/* Delayed ACK missed: leave pingpong mode and
215 			 * deflate ATO.
216 			 */
217 			icsk->icsk_ack.pingpong = 0;
218 			icsk->icsk_ack.ato      = TCP_ATO_MIN;
219 		}
220 		tcp_send_ack(sk);
221 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKS);
222 	}
223 	TCP_CHECK_TIMER(sk);
224 
225 out:
226 	if (tcp_memory_pressure)
227 		sk_mem_reclaim(sk);
228 out_unlock:
229 	bh_unlock_sock(sk);
230 	sock_put(sk);
231 }
232 
233 static void tcp_probe_timer(struct sock *sk)
234 {
235 	struct inet_connection_sock *icsk = inet_csk(sk);
236 	struct tcp_sock *tp = tcp_sk(sk);
237 	int max_probes;
238 
239 	if (tp->packets_out || !tcp_send_head(sk)) {
240 		icsk->icsk_probes_out = 0;
241 		return;
242 	}
243 
244 	/* *WARNING* RFC 1122 forbids this
245 	 *
246 	 * It doesn't AFAIK, because we kill the retransmit timer -AK
247 	 *
248 	 * FIXME: We ought not to do it, Solaris 2.5 actually has fixing
249 	 * this behaviour in Solaris down as a bug fix. [AC]
250 	 *
251 	 * Let me to explain. icsk_probes_out is zeroed by incoming ACKs
252 	 * even if they advertise zero window. Hence, connection is killed only
253 	 * if we received no ACKs for normal connection timeout. It is not killed
254 	 * only because window stays zero for some time, window may be zero
255 	 * until armageddon and even later. We are in full accordance
256 	 * with RFCs, only probe timer combines both retransmission timeout
257 	 * and probe timeout in one bottle.				--ANK
258 	 */
259 	max_probes = sysctl_tcp_retries2;
260 
261 	if (sock_flag(sk, SOCK_DEAD)) {
262 		const int alive = ((icsk->icsk_rto << icsk->icsk_backoff) < TCP_RTO_MAX);
263 
264 		max_probes = tcp_orphan_retries(sk, alive);
265 
266 		if (tcp_out_of_resources(sk, alive || icsk->icsk_probes_out <= max_probes))
267 			return;
268 	}
269 
270 	if (icsk->icsk_probes_out > max_probes) {
271 		tcp_write_err(sk);
272 	} else {
273 		/* Only send another probe if we didn't close things up. */
274 		tcp_send_probe0(sk);
275 	}
276 }
277 
278 /*
279  *	The TCP retransmit timer.
280  */
281 
282 static void tcp_retransmit_timer(struct sock *sk)
283 {
284 	struct tcp_sock *tp = tcp_sk(sk);
285 	struct inet_connection_sock *icsk = inet_csk(sk);
286 
287 	if (!tp->packets_out)
288 		goto out;
289 
290 	WARN_ON(tcp_write_queue_empty(sk));
291 
292 	if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
293 	    !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
294 		/* Receiver dastardly shrinks window. Our retransmits
295 		 * become zero probes, but we should not timeout this
296 		 * connection. If the socket is an orphan, time it out,
297 		 * we cannot allow such beasts to hang infinitely.
298 		 */
299 #ifdef TCP_DEBUG
300 		struct inet_sock *inet = inet_sk(sk);
301 		if (sk->sk_family == AF_INET) {
302 			LIMIT_NETDEBUG(KERN_DEBUG "TCP: Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
303 			       &inet->daddr, ntohs(inet->dport),
304 			       inet->num, tp->snd_una, tp->snd_nxt);
305 		}
306 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
307 		else if (sk->sk_family == AF_INET6) {
308 			struct ipv6_pinfo *np = inet6_sk(sk);
309 			LIMIT_NETDEBUG(KERN_DEBUG "TCP: Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
310 			       &np->daddr, ntohs(inet->dport),
311 			       inet->num, tp->snd_una, tp->snd_nxt);
312 		}
313 #endif
314 #endif
315 		if (tcp_time_stamp - tp->rcv_tstamp > TCP_RTO_MAX) {
316 			tcp_write_err(sk);
317 			goto out;
318 		}
319 		tcp_enter_loss(sk, 0);
320 		tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
321 		__sk_dst_reset(sk);
322 		goto out_reset_timer;
323 	}
324 
325 	if (tcp_write_timeout(sk))
326 		goto out;
327 
328 	if (icsk->icsk_retransmits == 0) {
329 		int mib_idx;
330 
331 		if (icsk->icsk_ca_state == TCP_CA_Disorder) {
332 			if (tcp_is_sack(tp))
333 				mib_idx = LINUX_MIB_TCPSACKFAILURES;
334 			else
335 				mib_idx = LINUX_MIB_TCPRENOFAILURES;
336 		} else if (icsk->icsk_ca_state == TCP_CA_Recovery) {
337 			if (tcp_is_sack(tp))
338 				mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
339 			else
340 				mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
341 		} else if (icsk->icsk_ca_state == TCP_CA_Loss) {
342 			mib_idx = LINUX_MIB_TCPLOSSFAILURES;
343 		} else {
344 			mib_idx = LINUX_MIB_TCPTIMEOUTS;
345 		}
346 		NET_INC_STATS_BH(sock_net(sk), mib_idx);
347 	}
348 
349 	if (tcp_use_frto(sk)) {
350 		tcp_enter_frto(sk);
351 	} else {
352 		tcp_enter_loss(sk, 0);
353 	}
354 
355 	if (tcp_retransmit_skb(sk, tcp_write_queue_head(sk)) > 0) {
356 		/* Retransmission failed because of local congestion,
357 		 * do not backoff.
358 		 */
359 		if (!icsk->icsk_retransmits)
360 			icsk->icsk_retransmits = 1;
361 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
362 					  min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL),
363 					  TCP_RTO_MAX);
364 		goto out;
365 	}
366 
367 	/* Increase the timeout each time we retransmit.  Note that
368 	 * we do not increase the rtt estimate.  rto is initialized
369 	 * from rtt, but increases here.  Jacobson (SIGCOMM 88) suggests
370 	 * that doubling rto each time is the least we can get away with.
371 	 * In KA9Q, Karn uses this for the first few times, and then
372 	 * goes to quadratic.  netBSD doubles, but only goes up to *64,
373 	 * and clamps at 1 to 64 sec afterwards.  Note that 120 sec is
374 	 * defined in the protocol as the maximum possible RTT.  I guess
375 	 * we'll have to use something other than TCP to talk to the
376 	 * University of Mars.
377 	 *
378 	 * PAWS allows us longer timeouts and large windows, so once
379 	 * implemented ftp to mars will work nicely. We will have to fix
380 	 * the 120 second clamps though!
381 	 */
382 	icsk->icsk_backoff++;
383 	icsk->icsk_retransmits++;
384 
385 out_reset_timer:
386 	icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
387 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX);
388 	if (icsk->icsk_retransmits > sysctl_tcp_retries1)
389 		__sk_dst_reset(sk);
390 
391 out:;
392 }
393 
394 static void tcp_write_timer(unsigned long data)
395 {
396 	struct sock *sk = (struct sock *)data;
397 	struct inet_connection_sock *icsk = inet_csk(sk);
398 	int event;
399 
400 	bh_lock_sock(sk);
401 	if (sock_owned_by_user(sk)) {
402 		/* Try again later */
403 		sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + (HZ / 20));
404 		goto out_unlock;
405 	}
406 
407 	if (sk->sk_state == TCP_CLOSE || !icsk->icsk_pending)
408 		goto out;
409 
410 	if (time_after(icsk->icsk_timeout, jiffies)) {
411 		sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
412 		goto out;
413 	}
414 
415 	event = icsk->icsk_pending;
416 	icsk->icsk_pending = 0;
417 
418 	switch (event) {
419 	case ICSK_TIME_RETRANS:
420 		tcp_retransmit_timer(sk);
421 		break;
422 	case ICSK_TIME_PROBE0:
423 		tcp_probe_timer(sk);
424 		break;
425 	}
426 	TCP_CHECK_TIMER(sk);
427 
428 out:
429 	sk_mem_reclaim(sk);
430 out_unlock:
431 	bh_unlock_sock(sk);
432 	sock_put(sk);
433 }
434 
435 /*
436  *	Timer for listening sockets
437  */
438 
439 static void tcp_synack_timer(struct sock *sk)
440 {
441 	inet_csk_reqsk_queue_prune(sk, TCP_SYNQ_INTERVAL,
442 				   TCP_TIMEOUT_INIT, TCP_RTO_MAX);
443 }
444 
445 void tcp_set_keepalive(struct sock *sk, int val)
446 {
447 	if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
448 		return;
449 
450 	if (val && !sock_flag(sk, SOCK_KEEPOPEN))
451 		inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
452 	else if (!val)
453 		inet_csk_delete_keepalive_timer(sk);
454 }
455 
456 
457 static void tcp_keepalive_timer (unsigned long data)
458 {
459 	struct sock *sk = (struct sock *) data;
460 	struct inet_connection_sock *icsk = inet_csk(sk);
461 	struct tcp_sock *tp = tcp_sk(sk);
462 	__u32 elapsed;
463 
464 	/* Only process if socket is not in use. */
465 	bh_lock_sock(sk);
466 	if (sock_owned_by_user(sk)) {
467 		/* Try again later. */
468 		inet_csk_reset_keepalive_timer (sk, HZ/20);
469 		goto out;
470 	}
471 
472 	if (sk->sk_state == TCP_LISTEN) {
473 		tcp_synack_timer(sk);
474 		goto out;
475 	}
476 
477 	if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
478 		if (tp->linger2 >= 0) {
479 			const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
480 
481 			if (tmo > 0) {
482 				tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
483 				goto out;
484 			}
485 		}
486 		tcp_send_active_reset(sk, GFP_ATOMIC);
487 		goto death;
488 	}
489 
490 	if (!sock_flag(sk, SOCK_KEEPOPEN) || sk->sk_state == TCP_CLOSE)
491 		goto out;
492 
493 	elapsed = keepalive_time_when(tp);
494 
495 	/* It is alive without keepalive 8) */
496 	if (tp->packets_out || tcp_send_head(sk))
497 		goto resched;
498 
499 	elapsed = tcp_time_stamp - tp->rcv_tstamp;
500 
501 	if (elapsed >= keepalive_time_when(tp)) {
502 		if ((!tp->keepalive_probes && icsk->icsk_probes_out >= sysctl_tcp_keepalive_probes) ||
503 		     (tp->keepalive_probes && icsk->icsk_probes_out >= tp->keepalive_probes)) {
504 			tcp_send_active_reset(sk, GFP_ATOMIC);
505 			tcp_write_err(sk);
506 			goto out;
507 		}
508 		if (tcp_write_wakeup(sk) <= 0) {
509 			icsk->icsk_probes_out++;
510 			elapsed = keepalive_intvl_when(tp);
511 		} else {
512 			/* If keepalive was lost due to local congestion,
513 			 * try harder.
514 			 */
515 			elapsed = TCP_RESOURCE_PROBE_INTERVAL;
516 		}
517 	} else {
518 		/* It is tp->rcv_tstamp + keepalive_time_when(tp) */
519 		elapsed = keepalive_time_when(tp) - elapsed;
520 	}
521 
522 	TCP_CHECK_TIMER(sk);
523 	sk_mem_reclaim(sk);
524 
525 resched:
526 	inet_csk_reset_keepalive_timer (sk, elapsed);
527 	goto out;
528 
529 death:
530 	tcp_done(sk);
531 
532 out:
533 	bh_unlock_sock(sk);
534 	sock_put(sk);
535 }
536