Loading...
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_timer(unsigned long);
36static void tcp_delack_timer(unsigned long);
37static void tcp_keepalive_timer (unsigned long data);
38
39void 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}
44EXPORT_SYMBOL(tcp_init_xmit_timers);
45
46static 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 */
66static 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. */
97static 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
113static 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 */
138static 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. */
167static 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
208static 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
260out:
261 if (sk_under_memory_pressure(sk))
262 sk_mem_reclaim(sk);
263out_unlock:
264 bh_unlock_sock(sk);
265 sock_put(sk);
266}
267
268static 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
317void 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->early_retrans_delayed) {
323 tcp_resume_early_retransmit(sk);
324 return;
325 }
326
327 if (!tp->packets_out)
328 goto out;
329
330 WARN_ON(tcp_write_queue_empty(sk));
331
332 if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
333 !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
334 /* Receiver dastardly shrinks window. Our retransmits
335 * become zero probes, but we should not timeout this
336 * connection. If the socket is an orphan, time it out,
337 * we cannot allow such beasts to hang infinitely.
338 */
339 struct inet_sock *inet = inet_sk(sk);
340 if (sk->sk_family == AF_INET) {
341 LIMIT_NETDEBUG(KERN_DEBUG pr_fmt("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n"),
342 &inet->inet_daddr,
343 ntohs(inet->inet_dport), inet->inet_num,
344 tp->snd_una, tp->snd_nxt);
345 }
346#if IS_ENABLED(CONFIG_IPV6)
347 else if (sk->sk_family == AF_INET6) {
348 struct ipv6_pinfo *np = inet6_sk(sk);
349 LIMIT_NETDEBUG(KERN_DEBUG pr_fmt("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n"),
350 &np->daddr,
351 ntohs(inet->inet_dport), inet->inet_num,
352 tp->snd_una, tp->snd_nxt);
353 }
354#endif
355 if (tcp_time_stamp - tp->rcv_tstamp > TCP_RTO_MAX) {
356 tcp_write_err(sk);
357 goto out;
358 }
359 tcp_enter_loss(sk, 0);
360 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
361 __sk_dst_reset(sk);
362 goto out_reset_timer;
363 }
364
365 if (tcp_write_timeout(sk))
366 goto out;
367
368 if (icsk->icsk_retransmits == 0) {
369 int mib_idx;
370
371 if (icsk->icsk_ca_state == TCP_CA_Recovery) {
372 if (tcp_is_sack(tp))
373 mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
374 else
375 mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
376 } else if (icsk->icsk_ca_state == TCP_CA_Loss) {
377 mib_idx = LINUX_MIB_TCPLOSSFAILURES;
378 } else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
379 tp->sacked_out) {
380 if (tcp_is_sack(tp))
381 mib_idx = LINUX_MIB_TCPSACKFAILURES;
382 else
383 mib_idx = LINUX_MIB_TCPRENOFAILURES;
384 } else {
385 mib_idx = LINUX_MIB_TCPTIMEOUTS;
386 }
387 NET_INC_STATS_BH(sock_net(sk), mib_idx);
388 }
389
390 if (tcp_use_frto(sk)) {
391 tcp_enter_frto(sk);
392 } else {
393 tcp_enter_loss(sk, 0);
394 }
395
396 if (tcp_retransmit_skb(sk, tcp_write_queue_head(sk)) > 0) {
397 /* Retransmission failed because of local congestion,
398 * do not backoff.
399 */
400 if (!icsk->icsk_retransmits)
401 icsk->icsk_retransmits = 1;
402 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
403 min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL),
404 TCP_RTO_MAX);
405 goto out;
406 }
407
408 /* Increase the timeout each time we retransmit. Note that
409 * we do not increase the rtt estimate. rto is initialized
410 * from rtt, but increases here. Jacobson (SIGCOMM 88) suggests
411 * that doubling rto each time is the least we can get away with.
412 * In KA9Q, Karn uses this for the first few times, and then
413 * goes to quadratic. netBSD doubles, but only goes up to *64,
414 * and clamps at 1 to 64 sec afterwards. Note that 120 sec is
415 * defined in the protocol as the maximum possible RTT. I guess
416 * we'll have to use something other than TCP to talk to the
417 * University of Mars.
418 *
419 * PAWS allows us longer timeouts and large windows, so once
420 * implemented ftp to mars will work nicely. We will have to fix
421 * the 120 second clamps though!
422 */
423 icsk->icsk_backoff++;
424 icsk->icsk_retransmits++;
425
426out_reset_timer:
427 /* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
428 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
429 * might be increased if the stream oscillates between thin and thick,
430 * thus the old value might already be too high compared to the value
431 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
432 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
433 * exponential backoff behaviour to avoid continue hammering
434 * linear-timeout retransmissions into a black hole
435 */
436 if (sk->sk_state == TCP_ESTABLISHED &&
437 (tp->thin_lto || sysctl_tcp_thin_linear_timeouts) &&
438 tcp_stream_is_thin(tp) &&
439 icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
440 icsk->icsk_backoff = 0;
441 icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX);
442 } else {
443 /* Use normal (exponential) backoff */
444 icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
445 }
446 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX);
447 if (retransmits_timed_out(sk, sysctl_tcp_retries1 + 1, 0, 0))
448 __sk_dst_reset(sk);
449
450out:;
451}
452
453static void tcp_write_timer(unsigned long data)
454{
455 struct sock *sk = (struct sock *)data;
456 struct inet_connection_sock *icsk = inet_csk(sk);
457 int event;
458
459 bh_lock_sock(sk);
460 if (sock_owned_by_user(sk)) {
461 /* Try again later */
462 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + (HZ / 20));
463 goto out_unlock;
464 }
465
466 if (sk->sk_state == TCP_CLOSE || !icsk->icsk_pending)
467 goto out;
468
469 if (time_after(icsk->icsk_timeout, jiffies)) {
470 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
471 goto out;
472 }
473
474 event = icsk->icsk_pending;
475 icsk->icsk_pending = 0;
476
477 switch (event) {
478 case ICSK_TIME_RETRANS:
479 tcp_retransmit_timer(sk);
480 break;
481 case ICSK_TIME_PROBE0:
482 tcp_probe_timer(sk);
483 break;
484 }
485
486out:
487 sk_mem_reclaim(sk);
488out_unlock:
489 bh_unlock_sock(sk);
490 sock_put(sk);
491}
492
493/*
494 * Timer for listening sockets
495 */
496
497static void tcp_synack_timer(struct sock *sk)
498{
499 inet_csk_reqsk_queue_prune(sk, TCP_SYNQ_INTERVAL,
500 TCP_TIMEOUT_INIT, TCP_RTO_MAX);
501}
502
503void tcp_syn_ack_timeout(struct sock *sk, struct request_sock *req)
504{
505 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPTIMEOUTS);
506}
507EXPORT_SYMBOL(tcp_syn_ack_timeout);
508
509void tcp_set_keepalive(struct sock *sk, int val)
510{
511 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
512 return;
513
514 if (val && !sock_flag(sk, SOCK_KEEPOPEN))
515 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
516 else if (!val)
517 inet_csk_delete_keepalive_timer(sk);
518}
519
520
521static void tcp_keepalive_timer (unsigned long data)
522{
523 struct sock *sk = (struct sock *) data;
524 struct inet_connection_sock *icsk = inet_csk(sk);
525 struct tcp_sock *tp = tcp_sk(sk);
526 u32 elapsed;
527
528 /* Only process if socket is not in use. */
529 bh_lock_sock(sk);
530 if (sock_owned_by_user(sk)) {
531 /* Try again later. */
532 inet_csk_reset_keepalive_timer (sk, HZ/20);
533 goto out;
534 }
535
536 if (sk->sk_state == TCP_LISTEN) {
537 tcp_synack_timer(sk);
538 goto out;
539 }
540
541 if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
542 if (tp->linger2 >= 0) {
543 const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
544
545 if (tmo > 0) {
546 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
547 goto out;
548 }
549 }
550 tcp_send_active_reset(sk, GFP_ATOMIC);
551 goto death;
552 }
553
554 if (!sock_flag(sk, SOCK_KEEPOPEN) || sk->sk_state == TCP_CLOSE)
555 goto out;
556
557 elapsed = keepalive_time_when(tp);
558
559 /* It is alive without keepalive 8) */
560 if (tp->packets_out || tcp_send_head(sk))
561 goto resched;
562
563 elapsed = keepalive_time_elapsed(tp);
564
565 if (elapsed >= keepalive_time_when(tp)) {
566 /* If the TCP_USER_TIMEOUT option is enabled, use that
567 * to determine when to timeout instead.
568 */
569 if ((icsk->icsk_user_timeout != 0 &&
570 elapsed >= icsk->icsk_user_timeout &&
571 icsk->icsk_probes_out > 0) ||
572 (icsk->icsk_user_timeout == 0 &&
573 icsk->icsk_probes_out >= keepalive_probes(tp))) {
574 tcp_send_active_reset(sk, GFP_ATOMIC);
575 tcp_write_err(sk);
576 goto out;
577 }
578 if (tcp_write_wakeup(sk) <= 0) {
579 icsk->icsk_probes_out++;
580 elapsed = keepalive_intvl_when(tp);
581 } else {
582 /* If keepalive was lost due to local congestion,
583 * try harder.
584 */
585 elapsed = TCP_RESOURCE_PROBE_INTERVAL;
586 }
587 } else {
588 /* It is tp->rcv_tstamp + keepalive_time_when(tp) */
589 elapsed = keepalive_time_when(tp) - elapsed;
590 }
591
592 sk_mem_reclaim(sk);
593
594resched:
595 inet_csk_reset_keepalive_timer (sk, elapsed);
596 goto out;
597
598death:
599 tcp_done(sk);
600
601out:
602 bh_unlock_sock(sk);
603 sock_put(sk);
604}
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
27static void tcp_write_err(struct sock *sk)
28{
29 sk->sk_err = sk->sk_err_soft ? : ETIMEDOUT;
30 sk->sk_error_report(sk);
31
32 tcp_done(sk);
33 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT);
34}
35
36/* Do not allow orphaned sockets to eat all our resources.
37 * This is direct violation of TCP specs, but it is required
38 * to prevent DoS attacks. It is called when a retransmission timeout
39 * or zero probe timeout occurs on orphaned socket.
40 *
41 * Criteria is still not confirmed experimentally and may change.
42 * We kill the socket, if:
43 * 1. If number of orphaned sockets exceeds an administratively configured
44 * limit.
45 * 2. If we have strong memory pressure.
46 */
47static int tcp_out_of_resources(struct sock *sk, bool do_reset)
48{
49 struct tcp_sock *tp = tcp_sk(sk);
50 int shift = 0;
51
52 /* If peer does not open window for long time, or did not transmit
53 * anything for long time, penalize it. */
54 if ((s32)(tcp_time_stamp - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset)
55 shift++;
56
57 /* If some dubious ICMP arrived, penalize even more. */
58 if (sk->sk_err_soft)
59 shift++;
60
61 if (tcp_check_oom(sk, shift)) {
62 /* Catch exceptional cases, when connection requires reset.
63 * 1. Last segment was sent recently. */
64 if ((s32)(tcp_time_stamp - tp->lsndtime) <= TCP_TIMEWAIT_LEN ||
65 /* 2. Window is closed. */
66 (!tp->snd_wnd && !tp->packets_out))
67 do_reset = true;
68 if (do_reset)
69 tcp_send_active_reset(sk, GFP_ATOMIC);
70 tcp_done(sk);
71 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY);
72 return 1;
73 }
74 return 0;
75}
76
77/* Calculate maximal number or retries on an orphaned socket. */
78static int tcp_orphan_retries(struct sock *sk, bool alive)
79{
80 int retries = sock_net(sk)->ipv4.sysctl_tcp_orphan_retries; /* May be zero. */
81
82 /* We know from an ICMP that something is wrong. */
83 if (sk->sk_err_soft && !alive)
84 retries = 0;
85
86 /* However, if socket sent something recently, select some safe
87 * number of retries. 8 corresponds to >100 seconds with minimal
88 * RTO of 200msec. */
89 if (retries == 0 && alive)
90 retries = 8;
91 return retries;
92}
93
94static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk)
95{
96 struct net *net = sock_net(sk);
97
98 /* Black hole detection */
99 if (net->ipv4.sysctl_tcp_mtu_probing) {
100 if (!icsk->icsk_mtup.enabled) {
101 icsk->icsk_mtup.enabled = 1;
102 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
103 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
104 } else {
105 struct net *net = sock_net(sk);
106 struct tcp_sock *tp = tcp_sk(sk);
107 int mss;
108
109 mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1;
110 mss = min(net->ipv4.sysctl_tcp_base_mss, mss);
111 mss = max(mss, 68 - tp->tcp_header_len);
112 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
113 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
114 }
115 }
116}
117
118/* This function calculates a "timeout" which is equivalent to the timeout of a
119 * TCP connection after "boundary" unsuccessful, exponentially backed-off
120 * retransmissions with an initial RTO of TCP_RTO_MIN or TCP_TIMEOUT_INIT if
121 * syn_set flag is set.
122 */
123static bool retransmits_timed_out(struct sock *sk,
124 unsigned int boundary,
125 unsigned int timeout,
126 bool syn_set)
127{
128 unsigned int linear_backoff_thresh, start_ts;
129 unsigned int rto_base = syn_set ? TCP_TIMEOUT_INIT : TCP_RTO_MIN;
130
131 if (!inet_csk(sk)->icsk_retransmits)
132 return false;
133
134 start_ts = tcp_sk(sk)->retrans_stamp;
135 if (unlikely(!start_ts))
136 start_ts = tcp_skb_timestamp(tcp_write_queue_head(sk));
137
138 if (likely(timeout == 0)) {
139 linear_backoff_thresh = ilog2(TCP_RTO_MAX/rto_base);
140
141 if (boundary <= linear_backoff_thresh)
142 timeout = ((2 << boundary) - 1) * rto_base;
143 else
144 timeout = ((2 << linear_backoff_thresh) - 1) * rto_base +
145 (boundary - linear_backoff_thresh) * TCP_RTO_MAX;
146 }
147 return (tcp_time_stamp - start_ts) >= timeout;
148}
149
150/* A write timeout has occurred. Process the after effects. */
151static int tcp_write_timeout(struct sock *sk)
152{
153 struct inet_connection_sock *icsk = inet_csk(sk);
154 struct tcp_sock *tp = tcp_sk(sk);
155 struct net *net = sock_net(sk);
156 int retry_until;
157 bool do_reset, syn_set = false;
158
159 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
160 if (icsk->icsk_retransmits) {
161 dst_negative_advice(sk);
162 if (tp->syn_fastopen || tp->syn_data)
163 tcp_fastopen_cache_set(sk, 0, NULL, true, 0);
164 if (tp->syn_data && icsk->icsk_retransmits == 1)
165 NET_INC_STATS_BH(sock_net(sk),
166 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
167 }
168 retry_until = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
169 syn_set = true;
170 } else {
171 if (retransmits_timed_out(sk, net->ipv4.sysctl_tcp_retries1, 0, 0)) {
172 /* Some middle-boxes may black-hole Fast Open _after_
173 * the handshake. Therefore we conservatively disable
174 * Fast Open on this path on recurring timeouts with
175 * few or zero bytes acked after Fast Open.
176 */
177 if (tp->syn_data_acked &&
178 tp->bytes_acked <= tp->rx_opt.mss_clamp) {
179 tcp_fastopen_cache_set(sk, 0, NULL, true, 0);
180 if (icsk->icsk_retransmits == net->ipv4.sysctl_tcp_retries1)
181 NET_INC_STATS_BH(sock_net(sk),
182 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
183 }
184 /* Black hole detection */
185 tcp_mtu_probing(icsk, sk);
186
187 dst_negative_advice(sk);
188 }
189
190 retry_until = net->ipv4.sysctl_tcp_retries2;
191 if (sock_flag(sk, SOCK_DEAD)) {
192 const bool alive = icsk->icsk_rto < TCP_RTO_MAX;
193
194 retry_until = tcp_orphan_retries(sk, alive);
195 do_reset = alive ||
196 !retransmits_timed_out(sk, retry_until, 0, 0);
197
198 if (tcp_out_of_resources(sk, do_reset))
199 return 1;
200 }
201 }
202
203 if (retransmits_timed_out(sk, retry_until,
204 syn_set ? 0 : icsk->icsk_user_timeout, syn_set)) {
205 /* Has it gone just too far? */
206 tcp_write_err(sk);
207 return 1;
208 }
209 return 0;
210}
211
212void tcp_delack_timer_handler(struct sock *sk)
213{
214 struct tcp_sock *tp = tcp_sk(sk);
215 struct inet_connection_sock *icsk = inet_csk(sk);
216
217 sk_mem_reclaim_partial(sk);
218
219 if (sk->sk_state == TCP_CLOSE || !(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
220 goto out;
221
222 if (time_after(icsk->icsk_ack.timeout, jiffies)) {
223 sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
224 goto out;
225 }
226 icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
227
228 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
229 struct sk_buff *skb;
230
231 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSCHEDULERFAILED);
232
233 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
234 sk_backlog_rcv(sk, skb);
235
236 tp->ucopy.memory = 0;
237 }
238
239 if (inet_csk_ack_scheduled(sk)) {
240 if (!icsk->icsk_ack.pingpong) {
241 /* Delayed ACK missed: inflate ATO. */
242 icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto);
243 } else {
244 /* Delayed ACK missed: leave pingpong mode and
245 * deflate ATO.
246 */
247 icsk->icsk_ack.pingpong = 0;
248 icsk->icsk_ack.ato = TCP_ATO_MIN;
249 }
250 tcp_send_ack(sk);
251 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKS);
252 }
253
254out:
255 if (tcp_under_memory_pressure(sk))
256 sk_mem_reclaim(sk);
257}
258
259static void tcp_delack_timer(unsigned long data)
260{
261 struct sock *sk = (struct sock *)data;
262
263 bh_lock_sock(sk);
264 if (!sock_owned_by_user(sk)) {
265 tcp_delack_timer_handler(sk);
266 } else {
267 inet_csk(sk)->icsk_ack.blocked = 1;
268 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
269 /* deleguate our work to tcp_release_cb() */
270 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &tcp_sk(sk)->tsq_flags))
271 sock_hold(sk);
272 }
273 bh_unlock_sock(sk);
274 sock_put(sk);
275}
276
277static void tcp_probe_timer(struct sock *sk)
278{
279 struct inet_connection_sock *icsk = inet_csk(sk);
280 struct tcp_sock *tp = tcp_sk(sk);
281 int max_probes;
282 u32 start_ts;
283
284 if (tp->packets_out || !tcp_send_head(sk)) {
285 icsk->icsk_probes_out = 0;
286 return;
287 }
288
289 /* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as
290 * long as the receiver continues to respond probes. We support this by
291 * default and reset icsk_probes_out with incoming ACKs. But if the
292 * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we
293 * kill the socket when the retry count and the time exceeds the
294 * corresponding system limit. We also implement similar policy when
295 * we use RTO to probe window in tcp_retransmit_timer().
296 */
297 start_ts = tcp_skb_timestamp(tcp_send_head(sk));
298 if (!start_ts)
299 skb_mstamp_get(&tcp_send_head(sk)->skb_mstamp);
300 else if (icsk->icsk_user_timeout &&
301 (s32)(tcp_time_stamp - start_ts) > icsk->icsk_user_timeout)
302 goto abort;
303
304 max_probes = sock_net(sk)->ipv4.sysctl_tcp_retries2;
305 if (sock_flag(sk, SOCK_DEAD)) {
306 const bool alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX;
307
308 max_probes = tcp_orphan_retries(sk, alive);
309 if (!alive && icsk->icsk_backoff >= max_probes)
310 goto abort;
311 if (tcp_out_of_resources(sk, true))
312 return;
313 }
314
315 if (icsk->icsk_probes_out > max_probes) {
316abort: tcp_write_err(sk);
317 } else {
318 /* Only send another probe if we didn't close things up. */
319 tcp_send_probe0(sk);
320 }
321}
322
323/*
324 * Timer for Fast Open socket to retransmit SYNACK. Note that the
325 * sk here is the child socket, not the parent (listener) socket.
326 */
327static void tcp_fastopen_synack_timer(struct sock *sk)
328{
329 struct inet_connection_sock *icsk = inet_csk(sk);
330 int max_retries = icsk->icsk_syn_retries ? :
331 sock_net(sk)->ipv4.sysctl_tcp_synack_retries + 1; /* add one more retry for fastopen */
332 struct request_sock *req;
333
334 req = tcp_sk(sk)->fastopen_rsk;
335 req->rsk_ops->syn_ack_timeout(req);
336
337 if (req->num_timeout >= max_retries) {
338 tcp_write_err(sk);
339 return;
340 }
341 /* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
342 * returned from rtx_syn_ack() to make it more persistent like
343 * regular retransmit because if the child socket has been accepted
344 * it's not good to give up too easily.
345 */
346 inet_rtx_syn_ack(sk, req);
347 req->num_timeout++;
348 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
349 TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
350}
351
352/*
353 * The TCP retransmit timer.
354 */
355
356void tcp_retransmit_timer(struct sock *sk)
357{
358 struct tcp_sock *tp = tcp_sk(sk);
359 struct net *net = sock_net(sk);
360 struct inet_connection_sock *icsk = inet_csk(sk);
361
362 if (tp->fastopen_rsk) {
363 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
364 sk->sk_state != TCP_FIN_WAIT1);
365 tcp_fastopen_synack_timer(sk);
366 /* Before we receive ACK to our SYN-ACK don't retransmit
367 * anything else (e.g., data or FIN segments).
368 */
369 return;
370 }
371 if (!tp->packets_out)
372 goto out;
373
374 WARN_ON(tcp_write_queue_empty(sk));
375
376 tp->tlp_high_seq = 0;
377
378 if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
379 !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
380 /* Receiver dastardly shrinks window. Our retransmits
381 * become zero probes, but we should not timeout this
382 * connection. If the socket is an orphan, time it out,
383 * we cannot allow such beasts to hang infinitely.
384 */
385 struct inet_sock *inet = inet_sk(sk);
386 if (sk->sk_family == AF_INET) {
387 net_dbg_ratelimited("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
388 &inet->inet_daddr,
389 ntohs(inet->inet_dport),
390 inet->inet_num,
391 tp->snd_una, tp->snd_nxt);
392 }
393#if IS_ENABLED(CONFIG_IPV6)
394 else if (sk->sk_family == AF_INET6) {
395 net_dbg_ratelimited("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
396 &sk->sk_v6_daddr,
397 ntohs(inet->inet_dport),
398 inet->inet_num,
399 tp->snd_una, tp->snd_nxt);
400 }
401#endif
402 if (tcp_time_stamp - tp->rcv_tstamp > TCP_RTO_MAX) {
403 tcp_write_err(sk);
404 goto out;
405 }
406 tcp_enter_loss(sk);
407 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
408 __sk_dst_reset(sk);
409 goto out_reset_timer;
410 }
411
412 if (tcp_write_timeout(sk))
413 goto out;
414
415 if (icsk->icsk_retransmits == 0) {
416 int mib_idx;
417
418 if (icsk->icsk_ca_state == TCP_CA_Recovery) {
419 if (tcp_is_sack(tp))
420 mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
421 else
422 mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
423 } else if (icsk->icsk_ca_state == TCP_CA_Loss) {
424 mib_idx = LINUX_MIB_TCPLOSSFAILURES;
425 } else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
426 tp->sacked_out) {
427 if (tcp_is_sack(tp))
428 mib_idx = LINUX_MIB_TCPSACKFAILURES;
429 else
430 mib_idx = LINUX_MIB_TCPRENOFAILURES;
431 } else {
432 mib_idx = LINUX_MIB_TCPTIMEOUTS;
433 }
434 NET_INC_STATS_BH(sock_net(sk), mib_idx);
435 }
436
437 tcp_enter_loss(sk);
438
439 if (tcp_retransmit_skb(sk, tcp_write_queue_head(sk)) > 0) {
440 /* Retransmission failed because of local congestion,
441 * do not backoff.
442 */
443 if (!icsk->icsk_retransmits)
444 icsk->icsk_retransmits = 1;
445 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
446 min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL),
447 TCP_RTO_MAX);
448 goto out;
449 }
450
451 /* Increase the timeout each time we retransmit. Note that
452 * we do not increase the rtt estimate. rto is initialized
453 * from rtt, but increases here. Jacobson (SIGCOMM 88) suggests
454 * that doubling rto each time is the least we can get away with.
455 * In KA9Q, Karn uses this for the first few times, and then
456 * goes to quadratic. netBSD doubles, but only goes up to *64,
457 * and clamps at 1 to 64 sec afterwards. Note that 120 sec is
458 * defined in the protocol as the maximum possible RTT. I guess
459 * we'll have to use something other than TCP to talk to the
460 * University of Mars.
461 *
462 * PAWS allows us longer timeouts and large windows, so once
463 * implemented ftp to mars will work nicely. We will have to fix
464 * the 120 second clamps though!
465 */
466 icsk->icsk_backoff++;
467 icsk->icsk_retransmits++;
468
469out_reset_timer:
470 /* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
471 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
472 * might be increased if the stream oscillates between thin and thick,
473 * thus the old value might already be too high compared to the value
474 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
475 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
476 * exponential backoff behaviour to avoid continue hammering
477 * linear-timeout retransmissions into a black hole
478 */
479 if (sk->sk_state == TCP_ESTABLISHED &&
480 (tp->thin_lto || sysctl_tcp_thin_linear_timeouts) &&
481 tcp_stream_is_thin(tp) &&
482 icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
483 icsk->icsk_backoff = 0;
484 icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX);
485 } else {
486 /* Use normal (exponential) backoff */
487 icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
488 }
489 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX);
490 if (retransmits_timed_out(sk, net->ipv4.sysctl_tcp_retries1 + 1, 0, 0))
491 __sk_dst_reset(sk);
492
493out:;
494}
495
496void tcp_write_timer_handler(struct sock *sk)
497{
498 struct inet_connection_sock *icsk = inet_csk(sk);
499 int event;
500
501 if (sk->sk_state == TCP_CLOSE || !icsk->icsk_pending)
502 goto out;
503
504 if (time_after(icsk->icsk_timeout, jiffies)) {
505 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
506 goto out;
507 }
508
509 event = icsk->icsk_pending;
510
511 switch (event) {
512 case ICSK_TIME_EARLY_RETRANS:
513 tcp_resume_early_retransmit(sk);
514 break;
515 case ICSK_TIME_LOSS_PROBE:
516 tcp_send_loss_probe(sk);
517 break;
518 case ICSK_TIME_RETRANS:
519 icsk->icsk_pending = 0;
520 tcp_retransmit_timer(sk);
521 break;
522 case ICSK_TIME_PROBE0:
523 icsk->icsk_pending = 0;
524 tcp_probe_timer(sk);
525 break;
526 }
527
528out:
529 sk_mem_reclaim(sk);
530}
531
532static void tcp_write_timer(unsigned long data)
533{
534 struct sock *sk = (struct sock *)data;
535
536 bh_lock_sock(sk);
537 if (!sock_owned_by_user(sk)) {
538 tcp_write_timer_handler(sk);
539 } else {
540 /* deleguate our work to tcp_release_cb() */
541 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &tcp_sk(sk)->tsq_flags))
542 sock_hold(sk);
543 }
544 bh_unlock_sock(sk);
545 sock_put(sk);
546}
547
548void tcp_syn_ack_timeout(const struct request_sock *req)
549{
550 struct net *net = read_pnet(&inet_rsk(req)->ireq_net);
551
552 NET_INC_STATS_BH(net, LINUX_MIB_TCPTIMEOUTS);
553}
554EXPORT_SYMBOL(tcp_syn_ack_timeout);
555
556void tcp_set_keepalive(struct sock *sk, int val)
557{
558 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
559 return;
560
561 if (val && !sock_flag(sk, SOCK_KEEPOPEN))
562 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
563 else if (!val)
564 inet_csk_delete_keepalive_timer(sk);
565}
566
567
568static void tcp_keepalive_timer (unsigned long data)
569{
570 struct sock *sk = (struct sock *) data;
571 struct inet_connection_sock *icsk = inet_csk(sk);
572 struct tcp_sock *tp = tcp_sk(sk);
573 u32 elapsed;
574
575 /* Only process if socket is not in use. */
576 bh_lock_sock(sk);
577 if (sock_owned_by_user(sk)) {
578 /* Try again later. */
579 inet_csk_reset_keepalive_timer (sk, HZ/20);
580 goto out;
581 }
582
583 if (sk->sk_state == TCP_LISTEN) {
584 pr_err("Hmm... keepalive on a LISTEN ???\n");
585 goto out;
586 }
587
588 if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
589 if (tp->linger2 >= 0) {
590 const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
591
592 if (tmo > 0) {
593 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
594 goto out;
595 }
596 }
597 tcp_send_active_reset(sk, GFP_ATOMIC);
598 goto death;
599 }
600
601 if (!sock_flag(sk, SOCK_KEEPOPEN) || sk->sk_state == TCP_CLOSE)
602 goto out;
603
604 elapsed = keepalive_time_when(tp);
605
606 /* It is alive without keepalive 8) */
607 if (tp->packets_out || tcp_send_head(sk))
608 goto resched;
609
610 elapsed = keepalive_time_elapsed(tp);
611
612 if (elapsed >= keepalive_time_when(tp)) {
613 /* If the TCP_USER_TIMEOUT option is enabled, use that
614 * to determine when to timeout instead.
615 */
616 if ((icsk->icsk_user_timeout != 0 &&
617 elapsed >= icsk->icsk_user_timeout &&
618 icsk->icsk_probes_out > 0) ||
619 (icsk->icsk_user_timeout == 0 &&
620 icsk->icsk_probes_out >= keepalive_probes(tp))) {
621 tcp_send_active_reset(sk, GFP_ATOMIC);
622 tcp_write_err(sk);
623 goto out;
624 }
625 if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) {
626 icsk->icsk_probes_out++;
627 elapsed = keepalive_intvl_when(tp);
628 } else {
629 /* If keepalive was lost due to local congestion,
630 * try harder.
631 */
632 elapsed = TCP_RESOURCE_PROBE_INTERVAL;
633 }
634 } else {
635 /* It is tp->rcv_tstamp + keepalive_time_when(tp) */
636 elapsed = keepalive_time_when(tp) - elapsed;
637 }
638
639 sk_mem_reclaim(sk);
640
641resched:
642 inet_csk_reset_keepalive_timer (sk, elapsed);
643 goto out;
644
645death:
646 tcp_done(sk);
647
648out:
649 bh_unlock_sock(sk);
650 sock_put(sk);
651}
652
653void tcp_init_xmit_timers(struct sock *sk)
654{
655 inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
656 &tcp_keepalive_timer);
657}