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 
25 int sysctl_tcp_syn_retries __read_mostly = TCP_SYN_RETRIES;
26 int sysctl_tcp_synack_retries __read_mostly = TCP_SYNACK_RETRIES;
27 int sysctl_tcp_keepalive_time __read_mostly = TCP_KEEPALIVE_TIME;
28 int sysctl_tcp_keepalive_probes __read_mostly = TCP_KEEPALIVE_PROBES;
29 int sysctl_tcp_keepalive_intvl __read_mostly = TCP_KEEPALIVE_INTVL;
30 int sysctl_tcp_retries1 __read_mostly = TCP_RETR1;
31 int sysctl_tcp_retries2 __read_mostly = TCP_RETR2;
32 int sysctl_tcp_orphan_retries __read_mostly;
33 int sysctl_tcp_thin_linear_timeouts __read_mostly;
34 
35 static void tcp_write_timer(unsigned long);
36 static void tcp_delack_timer(unsigned long);
37 static void tcp_keepalive_timer (unsigned long data);
38 
tcp_init_xmit_timers(struct sock * sk)39 void tcp_init_xmit_timers(struct sock *sk)
40 {
41 	inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
42 				  &tcp_keepalive_timer);
43 }
44 EXPORT_SYMBOL(tcp_init_xmit_timers);
45 
tcp_write_err(struct sock * sk)46 static void tcp_write_err(struct sock *sk)
47 {
48 	sk->sk_err = sk->sk_err_soft ? : ETIMEDOUT;
49 	sk->sk_error_report(sk);
50 
51 	tcp_done(sk);
52 	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT);
53 }
54 
55 /* Do not allow orphaned sockets to eat all our resources.
56  * This is direct violation of TCP specs, but it is required
57  * to prevent DoS attacks. It is called when a retransmission timeout
58  * or zero probe timeout occurs on orphaned socket.
59  *
60  * Criteria is still not confirmed experimentally and may change.
61  * We kill the socket, if:
62  * 1. If number of orphaned sockets exceeds an administratively configured
63  *    limit.
64  * 2. If we have strong memory pressure.
65  */
tcp_out_of_resources(struct sock * sk,int do_reset)66 static int tcp_out_of_resources(struct sock *sk, int do_reset)
67 {
68 	struct tcp_sock *tp = tcp_sk(sk);
69 	int shift = 0;
70 
71 	/* If peer does not open window for long time, or did not transmit
72 	 * anything for long time, penalize it. */
73 	if ((s32)(tcp_time_stamp - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset)
74 		shift++;
75 
76 	/* If some dubious ICMP arrived, penalize even more. */
77 	if (sk->sk_err_soft)
78 		shift++;
79 
80 	if (tcp_check_oom(sk, shift)) {
81 		/* Catch exceptional cases, when connection requires reset.
82 		 *      1. Last segment was sent recently. */
83 		if ((s32)(tcp_time_stamp - tp->lsndtime) <= TCP_TIMEWAIT_LEN ||
84 		    /*  2. Window is closed. */
85 		    (!tp->snd_wnd && !tp->packets_out))
86 			do_reset = 1;
87 		if (do_reset)
88 			tcp_send_active_reset(sk, GFP_ATOMIC);
89 		tcp_done(sk);
90 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY);
91 		return 1;
92 	}
93 	return 0;
94 }
95 
96 /* Calculate maximal number or retries on an orphaned socket. */
tcp_orphan_retries(struct sock * sk,int alive)97 static int tcp_orphan_retries(struct sock *sk, int alive)
98 {
99 	int retries = sysctl_tcp_orphan_retries; /* May be zero. */
100 
101 	/* We know from an ICMP that something is wrong. */
102 	if (sk->sk_err_soft && !alive)
103 		retries = 0;
104 
105 	/* However, if socket sent something recently, select some safe
106 	 * number of retries. 8 corresponds to >100 seconds with minimal
107 	 * RTO of 200msec. */
108 	if (retries == 0 && alive)
109 		retries = 8;
110 	return retries;
111 }
112 
tcp_mtu_probing(struct inet_connection_sock * icsk,struct sock * sk)113 static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk)
114 {
115 	/* Black hole detection */
116 	if (sysctl_tcp_mtu_probing) {
117 		if (!icsk->icsk_mtup.enabled) {
118 			icsk->icsk_mtup.enabled = 1;
119 			tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
120 		} else {
121 			struct tcp_sock *tp = tcp_sk(sk);
122 			int mss;
123 
124 			mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1;
125 			mss = min(sysctl_tcp_base_mss, mss);
126 			mss = max(mss, 68 - tp->tcp_header_len);
127 			icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
128 			tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
129 		}
130 	}
131 }
132 
133 /* This function calculates a "timeout" which is equivalent to the timeout of a
134  * TCP connection after "boundary" unsuccessful, exponentially backed-off
135  * retransmissions with an initial RTO of TCP_RTO_MIN or TCP_TIMEOUT_INIT if
136  * syn_set flag is set.
137  */
retransmits_timed_out(struct sock * sk,unsigned int boundary,unsigned int timeout,bool syn_set)138 static bool retransmits_timed_out(struct sock *sk,
139 				  unsigned int boundary,
140 				  unsigned int timeout,
141 				  bool syn_set)
142 {
143 	unsigned int linear_backoff_thresh, start_ts;
144 	unsigned int rto_base = syn_set ? TCP_TIMEOUT_INIT : TCP_RTO_MIN;
145 
146 	if (!inet_csk(sk)->icsk_retransmits)
147 		return false;
148 
149 	if (unlikely(!tcp_sk(sk)->retrans_stamp))
150 		start_ts = TCP_SKB_CB(tcp_write_queue_head(sk))->when;
151 	else
152 		start_ts = tcp_sk(sk)->retrans_stamp;
153 
154 	if (likely(timeout == 0)) {
155 		linear_backoff_thresh = ilog2(TCP_RTO_MAX/rto_base);
156 
157 		if (boundary <= linear_backoff_thresh)
158 			timeout = ((2 << boundary) - 1) * rto_base;
159 		else
160 			timeout = ((2 << linear_backoff_thresh) - 1) * rto_base +
161 				(boundary - linear_backoff_thresh) * TCP_RTO_MAX;
162 	}
163 	return (tcp_time_stamp - start_ts) >= timeout;
164 }
165 
166 /* A write timeout has occurred. Process the after effects. */
tcp_write_timeout(struct sock * sk)167 static int tcp_write_timeout(struct sock *sk)
168 {
169 	struct inet_connection_sock *icsk = inet_csk(sk);
170 	int retry_until;
171 	bool do_reset, syn_set = false;
172 
173 	if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
174 		if (icsk->icsk_retransmits)
175 			dst_negative_advice(sk);
176 		retry_until = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
177 		syn_set = true;
178 	} else {
179 		if (retransmits_timed_out(sk, sysctl_tcp_retries1, 0, 0)) {
180 			/* Black hole detection */
181 			tcp_mtu_probing(icsk, sk);
182 
183 			dst_negative_advice(sk);
184 		}
185 
186 		retry_until = sysctl_tcp_retries2;
187 		if (sock_flag(sk, SOCK_DEAD)) {
188 			const int alive = (icsk->icsk_rto < TCP_RTO_MAX);
189 
190 			retry_until = tcp_orphan_retries(sk, alive);
191 			do_reset = alive ||
192 				!retransmits_timed_out(sk, retry_until, 0, 0);
193 
194 			if (tcp_out_of_resources(sk, do_reset))
195 				return 1;
196 		}
197 	}
198 
199 	if (retransmits_timed_out(sk, retry_until,
200 				  syn_set ? 0 : icsk->icsk_user_timeout, syn_set)) {
201 		/* Has it gone just too far? */
202 		tcp_write_err(sk);
203 		return 1;
204 	}
205 	return 0;
206 }
207 
tcp_delack_timer(unsigned long data)208 static void tcp_delack_timer(unsigned long data)
209 {
210 	struct sock *sk = (struct sock *)data;
211 	struct tcp_sock *tp = tcp_sk(sk);
212 	struct inet_connection_sock *icsk = inet_csk(sk);
213 
214 	bh_lock_sock(sk);
215 	if (sock_owned_by_user(sk)) {
216 		/* Try again later. */
217 		icsk->icsk_ack.blocked = 1;
218 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
219 		sk_reset_timer(sk, &icsk->icsk_delack_timer, jiffies + TCP_DELACK_MIN);
220 		goto out_unlock;
221 	}
222 
223 	sk_mem_reclaim_partial(sk);
224 
225 	if (sk->sk_state == TCP_CLOSE || !(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
226 		goto out;
227 
228 	if (time_after(icsk->icsk_ack.timeout, jiffies)) {
229 		sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
230 		goto out;
231 	}
232 	icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
233 
234 	if (!skb_queue_empty(&tp->ucopy.prequeue)) {
235 		struct sk_buff *skb;
236 
237 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSCHEDULERFAILED);
238 
239 		while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
240 			sk_backlog_rcv(sk, skb);
241 
242 		tp->ucopy.memory = 0;
243 	}
244 
245 	if (inet_csk_ack_scheduled(sk)) {
246 		if (!icsk->icsk_ack.pingpong) {
247 			/* Delayed ACK missed: inflate ATO. */
248 			icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto);
249 		} else {
250 			/* Delayed ACK missed: leave pingpong mode and
251 			 * deflate ATO.
252 			 */
253 			icsk->icsk_ack.pingpong = 0;
254 			icsk->icsk_ack.ato      = TCP_ATO_MIN;
255 		}
256 		tcp_send_ack(sk);
257 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKS);
258 	}
259 
260 out:
261 	if (sk_under_memory_pressure(sk))
262 		sk_mem_reclaim(sk);
263 out_unlock:
264 	bh_unlock_sock(sk);
265 	sock_put(sk);
266 }
267 
tcp_probe_timer(struct sock * sk)268 static void tcp_probe_timer(struct sock *sk)
269 {
270 	struct inet_connection_sock *icsk = inet_csk(sk);
271 	struct tcp_sock *tp = tcp_sk(sk);
272 	int max_probes;
273 
274 	if (tp->packets_out || !tcp_send_head(sk)) {
275 		icsk->icsk_probes_out = 0;
276 		return;
277 	}
278 
279 	/* *WARNING* RFC 1122 forbids this
280 	 *
281 	 * It doesn't AFAIK, because we kill the retransmit timer -AK
282 	 *
283 	 * FIXME: We ought not to do it, Solaris 2.5 actually has fixing
284 	 * this behaviour in Solaris down as a bug fix. [AC]
285 	 *
286 	 * Let me to explain. icsk_probes_out is zeroed by incoming ACKs
287 	 * even if they advertise zero window. Hence, connection is killed only
288 	 * if we received no ACKs for normal connection timeout. It is not killed
289 	 * only because window stays zero for some time, window may be zero
290 	 * until armageddon and even later. We are in full accordance
291 	 * with RFCs, only probe timer combines both retransmission timeout
292 	 * and probe timeout in one bottle.				--ANK
293 	 */
294 	max_probes = sysctl_tcp_retries2;
295 
296 	if (sock_flag(sk, SOCK_DEAD)) {
297 		const int alive = ((icsk->icsk_rto << icsk->icsk_backoff) < TCP_RTO_MAX);
298 
299 		max_probes = tcp_orphan_retries(sk, alive);
300 
301 		if (tcp_out_of_resources(sk, alive || icsk->icsk_probes_out <= max_probes))
302 			return;
303 	}
304 
305 	if (icsk->icsk_probes_out > max_probes) {
306 		tcp_write_err(sk);
307 	} else {
308 		/* Only send another probe if we didn't close things up. */
309 		tcp_send_probe0(sk);
310 	}
311 }
312 
313 /*
314  *	The TCP retransmit timer.
315  */
316 
tcp_retransmit_timer(struct sock * sk)317 void tcp_retransmit_timer(struct sock *sk)
318 {
319 	struct tcp_sock *tp = tcp_sk(sk);
320 	struct inet_connection_sock *icsk = inet_csk(sk);
321 
322 	if (!tp->packets_out)
323 		goto out;
324 
325 	WARN_ON(tcp_write_queue_empty(sk));
326 
327 	if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
328 	    !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
329 		/* Receiver dastardly shrinks window. Our retransmits
330 		 * become zero probes, but we should not timeout this
331 		 * connection. If the socket is an orphan, time it out,
332 		 * we cannot allow such beasts to hang infinitely.
333 		 */
334 		struct inet_sock *inet = inet_sk(sk);
335 		if (sk->sk_family == AF_INET) {
336 			LIMIT_NETDEBUG(KERN_DEBUG "TCP: Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
337 			       &inet->inet_daddr, ntohs(inet->inet_dport),
338 			       inet->inet_num, tp->snd_una, tp->snd_nxt);
339 		}
340 #if IS_ENABLED(CONFIG_IPV6)
341 		else if (sk->sk_family == AF_INET6) {
342 			struct ipv6_pinfo *np = inet6_sk(sk);
343 			LIMIT_NETDEBUG(KERN_DEBUG "TCP: Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
344 			       &np->daddr, ntohs(inet->inet_dport),
345 			       inet->inet_num, tp->snd_una, tp->snd_nxt);
346 		}
347 #endif
348 		if (tcp_time_stamp - tp->rcv_tstamp > TCP_RTO_MAX) {
349 			tcp_write_err(sk);
350 			goto out;
351 		}
352 		tcp_enter_loss(sk, 0);
353 		tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
354 		__sk_dst_reset(sk);
355 		goto out_reset_timer;
356 	}
357 
358 	if (tcp_write_timeout(sk))
359 		goto out;
360 
361 	if (icsk->icsk_retransmits == 0) {
362 		int mib_idx;
363 
364 		if (icsk->icsk_ca_state == TCP_CA_Recovery) {
365 			if (tcp_is_sack(tp))
366 				mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
367 			else
368 				mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
369 		} else if (icsk->icsk_ca_state == TCP_CA_Loss) {
370 			mib_idx = LINUX_MIB_TCPLOSSFAILURES;
371 		} else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
372 			   tp->sacked_out) {
373 			if (tcp_is_sack(tp))
374 				mib_idx = LINUX_MIB_TCPSACKFAILURES;
375 			else
376 				mib_idx = LINUX_MIB_TCPRENOFAILURES;
377 		} else {
378 			mib_idx = LINUX_MIB_TCPTIMEOUTS;
379 		}
380 		NET_INC_STATS_BH(sock_net(sk), mib_idx);
381 	}
382 
383 	if (tcp_use_frto(sk)) {
384 		tcp_enter_frto(sk);
385 	} else {
386 		tcp_enter_loss(sk, 0);
387 	}
388 
389 	if (tcp_retransmit_skb(sk, tcp_write_queue_head(sk)) > 0) {
390 		/* Retransmission failed because of local congestion,
391 		 * do not backoff.
392 		 */
393 		if (!icsk->icsk_retransmits)
394 			icsk->icsk_retransmits = 1;
395 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
396 					  min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL),
397 					  TCP_RTO_MAX);
398 		goto out;
399 	}
400 
401 	/* Increase the timeout each time we retransmit.  Note that
402 	 * we do not increase the rtt estimate.  rto is initialized
403 	 * from rtt, but increases here.  Jacobson (SIGCOMM 88) suggests
404 	 * that doubling rto each time is the least we can get away with.
405 	 * In KA9Q, Karn uses this for the first few times, and then
406 	 * goes to quadratic.  netBSD doubles, but only goes up to *64,
407 	 * and clamps at 1 to 64 sec afterwards.  Note that 120 sec is
408 	 * defined in the protocol as the maximum possible RTT.  I guess
409 	 * we'll have to use something other than TCP to talk to the
410 	 * University of Mars.
411 	 *
412 	 * PAWS allows us longer timeouts and large windows, so once
413 	 * implemented ftp to mars will work nicely. We will have to fix
414 	 * the 120 second clamps though!
415 	 */
416 	icsk->icsk_backoff++;
417 	icsk->icsk_retransmits++;
418 
419 out_reset_timer:
420 	/* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
421 	 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
422 	 * might be increased if the stream oscillates between thin and thick,
423 	 * thus the old value might already be too high compared to the value
424 	 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
425 	 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
426 	 * exponential backoff behaviour to avoid continue hammering
427 	 * linear-timeout retransmissions into a black hole
428 	 */
429 	if (sk->sk_state == TCP_ESTABLISHED &&
430 	    (tp->thin_lto || sysctl_tcp_thin_linear_timeouts) &&
431 	    tcp_stream_is_thin(tp) &&
432 	    icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
433 		icsk->icsk_backoff = 0;
434 		icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX);
435 	} else {
436 		/* Use normal (exponential) backoff */
437 		icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
438 	}
439 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX);
440 	if (retransmits_timed_out(sk, sysctl_tcp_retries1 + 1, 0, 0))
441 		__sk_dst_reset(sk);
442 
443 out:;
444 }
445 
tcp_write_timer(unsigned long data)446 static void tcp_write_timer(unsigned long data)
447 {
448 	struct sock *sk = (struct sock *)data;
449 	struct inet_connection_sock *icsk = inet_csk(sk);
450 	int event;
451 
452 	bh_lock_sock(sk);
453 	if (sock_owned_by_user(sk)) {
454 		/* Try again later */
455 		sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + (HZ / 20));
456 		goto out_unlock;
457 	}
458 
459 	if (sk->sk_state == TCP_CLOSE || !icsk->icsk_pending)
460 		goto out;
461 
462 	if (time_after(icsk->icsk_timeout, jiffies)) {
463 		sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
464 		goto out;
465 	}
466 
467 	event = icsk->icsk_pending;
468 	icsk->icsk_pending = 0;
469 
470 	switch (event) {
471 	case ICSK_TIME_RETRANS:
472 		tcp_retransmit_timer(sk);
473 		break;
474 	case ICSK_TIME_PROBE0:
475 		tcp_probe_timer(sk);
476 		break;
477 	}
478 
479 out:
480 	sk_mem_reclaim(sk);
481 out_unlock:
482 	bh_unlock_sock(sk);
483 	sock_put(sk);
484 }
485 
486 /*
487  *	Timer for listening sockets
488  */
489 
tcp_synack_timer(struct sock * sk)490 static void tcp_synack_timer(struct sock *sk)
491 {
492 	inet_csk_reqsk_queue_prune(sk, TCP_SYNQ_INTERVAL,
493 				   TCP_TIMEOUT_INIT, TCP_RTO_MAX);
494 }
495 
tcp_syn_ack_timeout(struct sock * sk,struct request_sock * req)496 void tcp_syn_ack_timeout(struct sock *sk, struct request_sock *req)
497 {
498 	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPTIMEOUTS);
499 }
500 EXPORT_SYMBOL(tcp_syn_ack_timeout);
501 
tcp_set_keepalive(struct sock * sk,int val)502 void tcp_set_keepalive(struct sock *sk, int val)
503 {
504 	if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
505 		return;
506 
507 	if (val && !sock_flag(sk, SOCK_KEEPOPEN))
508 		inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
509 	else if (!val)
510 		inet_csk_delete_keepalive_timer(sk);
511 }
512 
513 
tcp_keepalive_timer(unsigned long data)514 static void tcp_keepalive_timer (unsigned long data)
515 {
516 	struct sock *sk = (struct sock *) data;
517 	struct inet_connection_sock *icsk = inet_csk(sk);
518 	struct tcp_sock *tp = tcp_sk(sk);
519 	u32 elapsed;
520 
521 	/* Only process if socket is not in use. */
522 	bh_lock_sock(sk);
523 	if (sock_owned_by_user(sk)) {
524 		/* Try again later. */
525 		inet_csk_reset_keepalive_timer (sk, HZ/20);
526 		goto out;
527 	}
528 
529 	if (sk->sk_state == TCP_LISTEN) {
530 		tcp_synack_timer(sk);
531 		goto out;
532 	}
533 
534 	if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
535 		if (tp->linger2 >= 0) {
536 			const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
537 
538 			if (tmo > 0) {
539 				tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
540 				goto out;
541 			}
542 		}
543 		tcp_send_active_reset(sk, GFP_ATOMIC);
544 		goto death;
545 	}
546 
547 	if (!sock_flag(sk, SOCK_KEEPOPEN) || sk->sk_state == TCP_CLOSE)
548 		goto out;
549 
550 	elapsed = keepalive_time_when(tp);
551 
552 	/* It is alive without keepalive 8) */
553 	if (tp->packets_out || tcp_send_head(sk))
554 		goto resched;
555 
556 	elapsed = keepalive_time_elapsed(tp);
557 
558 	if (elapsed >= keepalive_time_when(tp)) {
559 		/* If the TCP_USER_TIMEOUT option is enabled, use that
560 		 * to determine when to timeout instead.
561 		 */
562 		if ((icsk->icsk_user_timeout != 0 &&
563 		    elapsed >= icsk->icsk_user_timeout &&
564 		    icsk->icsk_probes_out > 0) ||
565 		    (icsk->icsk_user_timeout == 0 &&
566 		    icsk->icsk_probes_out >= keepalive_probes(tp))) {
567 			tcp_send_active_reset(sk, GFP_ATOMIC);
568 			tcp_write_err(sk);
569 			goto out;
570 		}
571 		if (tcp_write_wakeup(sk) <= 0) {
572 			icsk->icsk_probes_out++;
573 			elapsed = keepalive_intvl_when(tp);
574 		} else {
575 			/* If keepalive was lost due to local congestion,
576 			 * try harder.
577 			 */
578 			elapsed = TCP_RESOURCE_PROBE_INTERVAL;
579 		}
580 	} else {
581 		/* It is tp->rcv_tstamp + keepalive_time_when(tp) */
582 		elapsed = keepalive_time_when(tp) - elapsed;
583 	}
584 
585 	sk_mem_reclaim(sk);
586 
587 resched:
588 	inet_csk_reset_keepalive_timer (sk, elapsed);
589 	goto out;
590 
591 death:
592 	tcp_done(sk);
593 
594 out:
595 	bh_unlock_sock(sk);
596 	sock_put(sk);
597 }
598