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
25/**
26 * tcp_write_err() - close socket and save error info
27 * @sk: The socket the error has appeared on.
28 *
29 * Returns: Nothing (void)
30 */
31
32static void tcp_write_err(struct sock *sk)
33{
34 sk->sk_err = sk->sk_err_soft ? : ETIMEDOUT;
35 sk->sk_error_report(sk);
36
37 tcp_write_queue_purge(sk);
38 tcp_done(sk);
39 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT);
40}
41
42/**
43 * tcp_out_of_resources() - Close socket if out of resources
44 * @sk: pointer to current socket
45 * @do_reset: send a last packet with reset flag
46 *
47 * Do not allow orphaned sockets to eat all our resources.
48 * This is direct violation of TCP specs, but it is required
49 * to prevent DoS attacks. It is called when a retransmission timeout
50 * or zero probe timeout occurs on orphaned socket.
51 *
52 * Also close if our net namespace is exiting; in that case there is no
53 * hope of ever communicating again since all netns interfaces are already
54 * down (or about to be down), and we need to release our dst references,
55 * which have been moved to the netns loopback interface, so the namespace
56 * can finish exiting. This condition is only possible if we are a kernel
57 * socket, as those do not hold references to the namespace.
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 * 3. If our net namespace is exiting.
65 */
66static int tcp_out_of_resources(struct sock *sk, bool 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_jiffies32 - 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_jiffies32 - tp->lsndtime) <= TCP_TIMEWAIT_LEN ||
84 /* 2. Window is closed. */
85 (!tp->snd_wnd && !tp->packets_out))
86 do_reset = true;
87 if (do_reset)
88 tcp_send_active_reset(sk, GFP_ATOMIC);
89 tcp_done(sk);
90 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY);
91 return 1;
92 }
93
94 if (!check_net(sock_net(sk))) {
95 /* Not possible to send reset; just close */
96 tcp_done(sk);
97 return 1;
98 }
99
100 return 0;
101}
102
103/**
104 * tcp_orphan_retries() - Returns maximal number of retries on an orphaned socket
105 * @sk: Pointer to the current socket.
106 * @alive: bool, socket alive state
107 */
108static int tcp_orphan_retries(struct sock *sk, bool alive)
109{
110 int retries = sock_net(sk)->ipv4.sysctl_tcp_orphan_retries; /* May be zero. */
111
112 /* We know from an ICMP that something is wrong. */
113 if (sk->sk_err_soft && !alive)
114 retries = 0;
115
116 /* However, if socket sent something recently, select some safe
117 * number of retries. 8 corresponds to >100 seconds with minimal
118 * RTO of 200msec. */
119 if (retries == 0 && alive)
120 retries = 8;
121 return retries;
122}
123
124static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk)
125{
126 const struct net *net = sock_net(sk);
127 int mss;
128
129 /* Black hole detection */
130 if (!net->ipv4.sysctl_tcp_mtu_probing)
131 return;
132
133 if (!icsk->icsk_mtup.enabled) {
134 icsk->icsk_mtup.enabled = 1;
135 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
136 } else {
137 mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1;
138 mss = min(net->ipv4.sysctl_tcp_base_mss, mss);
139 mss = max(mss, 68 - tcp_sk(sk)->tcp_header_len);
140 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
141 }
142 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
143}
144
145
146/**
147 * retransmits_timed_out() - returns true if this connection has timed out
148 * @sk: The current socket
149 * @boundary: max number of retransmissions
150 * @timeout: A custom timeout value.
151 * If set to 0 the default timeout is calculated and used.
152 * Using TCP_RTO_MIN and the number of unsuccessful retransmits.
153 *
154 * The default "timeout" value this function can calculate and use
155 * is equivalent to the timeout of a TCP Connection
156 * after "boundary" unsuccessful, exponentially backed-off
157 * retransmissions with an initial RTO of TCP_RTO_MIN.
158 */
159static bool retransmits_timed_out(struct sock *sk,
160 unsigned int boundary,
161 unsigned int timeout)
162{
163 const unsigned int rto_base = TCP_RTO_MIN;
164 unsigned int linear_backoff_thresh, start_ts;
165
166 if (!inet_csk(sk)->icsk_retransmits)
167 return false;
168
169 start_ts = tcp_sk(sk)->retrans_stamp;
170 if (unlikely(!start_ts)) {
171 struct sk_buff *head = tcp_rtx_queue_head(sk);
172
173 if (!head)
174 return false;
175 start_ts = tcp_skb_timestamp(head);
176 }
177
178 if (likely(timeout == 0)) {
179 linear_backoff_thresh = ilog2(TCP_RTO_MAX/rto_base);
180
181 if (boundary <= linear_backoff_thresh)
182 timeout = ((2 << boundary) - 1) * rto_base;
183 else
184 timeout = ((2 << linear_backoff_thresh) - 1) * rto_base +
185 (boundary - linear_backoff_thresh) * TCP_RTO_MAX;
186 }
187 return (tcp_time_stamp(tcp_sk(sk)) - start_ts) >= jiffies_to_msecs(timeout);
188}
189
190/* A write timeout has occurred. Process the after effects. */
191static int tcp_write_timeout(struct sock *sk)
192{
193 struct inet_connection_sock *icsk = inet_csk(sk);
194 struct tcp_sock *tp = tcp_sk(sk);
195 struct net *net = sock_net(sk);
196 bool expired, do_reset;
197 int retry_until;
198
199 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
200 if (icsk->icsk_retransmits) {
201 dst_negative_advice(sk);
202 } else if (!tp->syn_data && !tp->syn_fastopen) {
203 sk_rethink_txhash(sk);
204 }
205 retry_until = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
206 expired = icsk->icsk_retransmits >= retry_until;
207 } else {
208 if (retransmits_timed_out(sk, net->ipv4.sysctl_tcp_retries1, 0)) {
209 /* Black hole detection */
210 tcp_mtu_probing(icsk, sk);
211
212 dst_negative_advice(sk);
213 } else {
214 sk_rethink_txhash(sk);
215 }
216
217 retry_until = net->ipv4.sysctl_tcp_retries2;
218 if (sock_flag(sk, SOCK_DEAD)) {
219 const bool alive = icsk->icsk_rto < TCP_RTO_MAX;
220
221 retry_until = tcp_orphan_retries(sk, alive);
222 do_reset = alive ||
223 !retransmits_timed_out(sk, retry_until, 0);
224
225 if (tcp_out_of_resources(sk, do_reset))
226 return 1;
227 }
228 expired = retransmits_timed_out(sk, retry_until,
229 icsk->icsk_user_timeout);
230 }
231 tcp_fastopen_active_detect_blackhole(sk, expired);
232
233 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RTO_CB_FLAG))
234 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RTO_CB,
235 icsk->icsk_retransmits,
236 icsk->icsk_rto, (int)expired);
237
238 if (expired) {
239 /* Has it gone just too far? */
240 tcp_write_err(sk);
241 return 1;
242 }
243
244 return 0;
245}
246
247/* Called with BH disabled */
248void tcp_delack_timer_handler(struct sock *sk)
249{
250 struct inet_connection_sock *icsk = inet_csk(sk);
251
252 sk_mem_reclaim_partial(sk);
253
254 if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
255 !(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
256 goto out;
257
258 if (time_after(icsk->icsk_ack.timeout, jiffies)) {
259 sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
260 goto out;
261 }
262 icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
263
264 if (inet_csk_ack_scheduled(sk)) {
265 if (!icsk->icsk_ack.pingpong) {
266 /* Delayed ACK missed: inflate ATO. */
267 icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto);
268 } else {
269 /* Delayed ACK missed: leave pingpong mode and
270 * deflate ATO.
271 */
272 icsk->icsk_ack.pingpong = 0;
273 icsk->icsk_ack.ato = TCP_ATO_MIN;
274 }
275 tcp_mstamp_refresh(tcp_sk(sk));
276 tcp_send_ack(sk);
277 __NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKS);
278 }
279
280out:
281 if (tcp_under_memory_pressure(sk))
282 sk_mem_reclaim(sk);
283}
284
285
286/**
287 * tcp_delack_timer() - The TCP delayed ACK timeout handler
288 * @data: Pointer to the current socket. (gets casted to struct sock *)
289 *
290 * This function gets (indirectly) called when the kernel timer for a TCP packet
291 * of this socket expires. Calls tcp_delack_timer_handler() to do the actual work.
292 *
293 * Returns: Nothing (void)
294 */
295static void tcp_delack_timer(struct timer_list *t)
296{
297 struct inet_connection_sock *icsk =
298 from_timer(icsk, t, icsk_delack_timer);
299 struct sock *sk = &icsk->icsk_inet.sk;
300
301 bh_lock_sock(sk);
302 if (!sock_owned_by_user(sk)) {
303 tcp_delack_timer_handler(sk);
304 } else {
305 icsk->icsk_ack.blocked = 1;
306 __NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
307 /* deleguate our work to tcp_release_cb() */
308 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &sk->sk_tsq_flags))
309 sock_hold(sk);
310 }
311 bh_unlock_sock(sk);
312 sock_put(sk);
313}
314
315static void tcp_probe_timer(struct sock *sk)
316{
317 struct inet_connection_sock *icsk = inet_csk(sk);
318 struct sk_buff *skb = tcp_send_head(sk);
319 struct tcp_sock *tp = tcp_sk(sk);
320 int max_probes;
321 u32 start_ts;
322
323 if (tp->packets_out || !skb) {
324 icsk->icsk_probes_out = 0;
325 return;
326 }
327
328 /* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as
329 * long as the receiver continues to respond probes. We support this by
330 * default and reset icsk_probes_out with incoming ACKs. But if the
331 * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we
332 * kill the socket when the retry count and the time exceeds the
333 * corresponding system limit. We also implement similar policy when
334 * we use RTO to probe window in tcp_retransmit_timer().
335 */
336 start_ts = tcp_skb_timestamp(skb);
337 if (!start_ts)
338 skb->skb_mstamp = tp->tcp_mstamp;
339 else if (icsk->icsk_user_timeout &&
340 (s32)(tcp_time_stamp(tp) - start_ts) >
341 jiffies_to_msecs(icsk->icsk_user_timeout))
342 goto abort;
343
344 max_probes = sock_net(sk)->ipv4.sysctl_tcp_retries2;
345 if (sock_flag(sk, SOCK_DEAD)) {
346 const bool alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX;
347
348 max_probes = tcp_orphan_retries(sk, alive);
349 if (!alive && icsk->icsk_backoff >= max_probes)
350 goto abort;
351 if (tcp_out_of_resources(sk, true))
352 return;
353 }
354
355 if (icsk->icsk_probes_out > max_probes) {
356abort: tcp_write_err(sk);
357 } else {
358 /* Only send another probe if we didn't close things up. */
359 tcp_send_probe0(sk);
360 }
361}
362
363/*
364 * Timer for Fast Open socket to retransmit SYNACK. Note that the
365 * sk here is the child socket, not the parent (listener) socket.
366 */
367static void tcp_fastopen_synack_timer(struct sock *sk)
368{
369 struct inet_connection_sock *icsk = inet_csk(sk);
370 int max_retries = icsk->icsk_syn_retries ? :
371 sock_net(sk)->ipv4.sysctl_tcp_synack_retries + 1; /* add one more retry for fastopen */
372 struct request_sock *req;
373
374 req = tcp_sk(sk)->fastopen_rsk;
375 req->rsk_ops->syn_ack_timeout(req);
376
377 if (req->num_timeout >= max_retries) {
378 tcp_write_err(sk);
379 return;
380 }
381 /* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
382 * returned from rtx_syn_ack() to make it more persistent like
383 * regular retransmit because if the child socket has been accepted
384 * it's not good to give up too easily.
385 */
386 inet_rtx_syn_ack(sk, req);
387 req->num_timeout++;
388 icsk->icsk_retransmits++;
389 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
390 TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
391}
392
393
394/**
395 * tcp_retransmit_timer() - The TCP retransmit timeout handler
396 * @sk: Pointer to the current socket.
397 *
398 * This function gets called when the kernel timer for a TCP packet
399 * of this socket expires.
400 *
401 * It handles retransmission, timer adjustment and other necesarry measures.
402 *
403 * Returns: Nothing (void)
404 */
405void tcp_retransmit_timer(struct sock *sk)
406{
407 struct tcp_sock *tp = tcp_sk(sk);
408 struct net *net = sock_net(sk);
409 struct inet_connection_sock *icsk = inet_csk(sk);
410
411 if (tp->fastopen_rsk) {
412 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
413 sk->sk_state != TCP_FIN_WAIT1);
414 tcp_fastopen_synack_timer(sk);
415 /* Before we receive ACK to our SYN-ACK don't retransmit
416 * anything else (e.g., data or FIN segments).
417 */
418 return;
419 }
420 if (!tp->packets_out)
421 goto out;
422
423 WARN_ON(tcp_rtx_queue_empty(sk));
424
425 tp->tlp_high_seq = 0;
426
427 if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
428 !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
429 /* Receiver dastardly shrinks window. Our retransmits
430 * become zero probes, but we should not timeout this
431 * connection. If the socket is an orphan, time it out,
432 * we cannot allow such beasts to hang infinitely.
433 */
434 struct inet_sock *inet = inet_sk(sk);
435 if (sk->sk_family == AF_INET) {
436 net_dbg_ratelimited("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
437 &inet->inet_daddr,
438 ntohs(inet->inet_dport),
439 inet->inet_num,
440 tp->snd_una, tp->snd_nxt);
441 }
442#if IS_ENABLED(CONFIG_IPV6)
443 else if (sk->sk_family == AF_INET6) {
444 net_dbg_ratelimited("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
445 &sk->sk_v6_daddr,
446 ntohs(inet->inet_dport),
447 inet->inet_num,
448 tp->snd_una, tp->snd_nxt);
449 }
450#endif
451 if (tcp_jiffies32 - tp->rcv_tstamp > TCP_RTO_MAX) {
452 tcp_write_err(sk);
453 goto out;
454 }
455 tcp_enter_loss(sk);
456 tcp_retransmit_skb(sk, tcp_rtx_queue_head(sk), 1);
457 __sk_dst_reset(sk);
458 goto out_reset_timer;
459 }
460
461 if (tcp_write_timeout(sk))
462 goto out;
463
464 if (icsk->icsk_retransmits == 0) {
465 int mib_idx;
466
467 if (icsk->icsk_ca_state == TCP_CA_Recovery) {
468 if (tcp_is_sack(tp))
469 mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
470 else
471 mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
472 } else if (icsk->icsk_ca_state == TCP_CA_Loss) {
473 mib_idx = LINUX_MIB_TCPLOSSFAILURES;
474 } else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
475 tp->sacked_out) {
476 if (tcp_is_sack(tp))
477 mib_idx = LINUX_MIB_TCPSACKFAILURES;
478 else
479 mib_idx = LINUX_MIB_TCPRENOFAILURES;
480 } else {
481 mib_idx = LINUX_MIB_TCPTIMEOUTS;
482 }
483 __NET_INC_STATS(sock_net(sk), mib_idx);
484 }
485
486 tcp_enter_loss(sk);
487
488 if (tcp_retransmit_skb(sk, tcp_rtx_queue_head(sk), 1) > 0) {
489 /* Retransmission failed because of local congestion,
490 * do not backoff.
491 */
492 if (!icsk->icsk_retransmits)
493 icsk->icsk_retransmits = 1;
494 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
495 min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL),
496 TCP_RTO_MAX);
497 goto out;
498 }
499
500 /* Increase the timeout each time we retransmit. Note that
501 * we do not increase the rtt estimate. rto is initialized
502 * from rtt, but increases here. Jacobson (SIGCOMM 88) suggests
503 * that doubling rto each time is the least we can get away with.
504 * In KA9Q, Karn uses this for the first few times, and then
505 * goes to quadratic. netBSD doubles, but only goes up to *64,
506 * and clamps at 1 to 64 sec afterwards. Note that 120 sec is
507 * defined in the protocol as the maximum possible RTT. I guess
508 * we'll have to use something other than TCP to talk to the
509 * University of Mars.
510 *
511 * PAWS allows us longer timeouts and large windows, so once
512 * implemented ftp to mars will work nicely. We will have to fix
513 * the 120 second clamps though!
514 */
515 icsk->icsk_backoff++;
516 icsk->icsk_retransmits++;
517
518out_reset_timer:
519 /* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
520 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
521 * might be increased if the stream oscillates between thin and thick,
522 * thus the old value might already be too high compared to the value
523 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
524 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
525 * exponential backoff behaviour to avoid continue hammering
526 * linear-timeout retransmissions into a black hole
527 */
528 if (sk->sk_state == TCP_ESTABLISHED &&
529 (tp->thin_lto || net->ipv4.sysctl_tcp_thin_linear_timeouts) &&
530 tcp_stream_is_thin(tp) &&
531 icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
532 icsk->icsk_backoff = 0;
533 icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX);
534 } else {
535 /* Use normal (exponential) backoff */
536 icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
537 }
538 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX);
539 if (retransmits_timed_out(sk, net->ipv4.sysctl_tcp_retries1 + 1, 0))
540 __sk_dst_reset(sk);
541
542out:;
543}
544
545/* Called with bottom-half processing disabled.
546 Called by tcp_write_timer() */
547void tcp_write_timer_handler(struct sock *sk)
548{
549 struct inet_connection_sock *icsk = inet_csk(sk);
550 int event;
551
552 if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
553 !icsk->icsk_pending)
554 goto out;
555
556 if (time_after(icsk->icsk_timeout, jiffies)) {
557 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
558 goto out;
559 }
560
561 tcp_mstamp_refresh(tcp_sk(sk));
562 event = icsk->icsk_pending;
563
564 switch (event) {
565 case ICSK_TIME_REO_TIMEOUT:
566 tcp_rack_reo_timeout(sk);
567 break;
568 case ICSK_TIME_LOSS_PROBE:
569 tcp_send_loss_probe(sk);
570 break;
571 case ICSK_TIME_RETRANS:
572 icsk->icsk_pending = 0;
573 tcp_retransmit_timer(sk);
574 break;
575 case ICSK_TIME_PROBE0:
576 icsk->icsk_pending = 0;
577 tcp_probe_timer(sk);
578 break;
579 }
580
581out:
582 sk_mem_reclaim(sk);
583}
584
585static void tcp_write_timer(struct timer_list *t)
586{
587 struct inet_connection_sock *icsk =
588 from_timer(icsk, t, icsk_retransmit_timer);
589 struct sock *sk = &icsk->icsk_inet.sk;
590
591 bh_lock_sock(sk);
592 if (!sock_owned_by_user(sk)) {
593 tcp_write_timer_handler(sk);
594 } else {
595 /* delegate our work to tcp_release_cb() */
596 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &sk->sk_tsq_flags))
597 sock_hold(sk);
598 }
599 bh_unlock_sock(sk);
600 sock_put(sk);
601}
602
603void tcp_syn_ack_timeout(const struct request_sock *req)
604{
605 struct net *net = read_pnet(&inet_rsk(req)->ireq_net);
606
607 __NET_INC_STATS(net, LINUX_MIB_TCPTIMEOUTS);
608}
609EXPORT_SYMBOL(tcp_syn_ack_timeout);
610
611void tcp_set_keepalive(struct sock *sk, int val)
612{
613 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
614 return;
615
616 if (val && !sock_flag(sk, SOCK_KEEPOPEN))
617 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
618 else if (!val)
619 inet_csk_delete_keepalive_timer(sk);
620}
621EXPORT_SYMBOL_GPL(tcp_set_keepalive);
622
623
624static void tcp_keepalive_timer (struct timer_list *t)
625{
626 struct sock *sk = from_timer(sk, t, sk_timer);
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 tcp_mstamp_refresh(tp);
645 if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
646 if (tp->linger2 >= 0) {
647 const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
648
649 if (tmo > 0) {
650 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
651 goto out;
652 }
653 }
654 tcp_send_active_reset(sk, GFP_ATOMIC);
655 goto death;
656 }
657
658 if (!sock_flag(sk, SOCK_KEEPOPEN) ||
659 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_SYN_SENT)))
660 goto out;
661
662 elapsed = keepalive_time_when(tp);
663
664 /* It is alive without keepalive 8) */
665 if (tp->packets_out || !tcp_write_queue_empty(sk))
666 goto resched;
667
668 elapsed = keepalive_time_elapsed(tp);
669
670 if (elapsed >= keepalive_time_when(tp)) {
671 /* If the TCP_USER_TIMEOUT option is enabled, use that
672 * to determine when to timeout instead.
673 */
674 if ((icsk->icsk_user_timeout != 0 &&
675 elapsed >= icsk->icsk_user_timeout &&
676 icsk->icsk_probes_out > 0) ||
677 (icsk->icsk_user_timeout == 0 &&
678 icsk->icsk_probes_out >= keepalive_probes(tp))) {
679 tcp_send_active_reset(sk, GFP_ATOMIC);
680 tcp_write_err(sk);
681 goto out;
682 }
683 if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) {
684 icsk->icsk_probes_out++;
685 elapsed = keepalive_intvl_when(tp);
686 } else {
687 /* If keepalive was lost due to local congestion,
688 * try harder.
689 */
690 elapsed = TCP_RESOURCE_PROBE_INTERVAL;
691 }
692 } else {
693 /* It is tp->rcv_tstamp + keepalive_time_when(tp) */
694 elapsed = keepalive_time_when(tp) - elapsed;
695 }
696
697 sk_mem_reclaim(sk);
698
699resched:
700 inet_csk_reset_keepalive_timer (sk, elapsed);
701 goto out;
702
703death:
704 tcp_done(sk);
705
706out:
707 bh_unlock_sock(sk);
708 sock_put(sk);
709}
710
711void tcp_init_xmit_timers(struct sock *sk)
712{
713 inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
714 &tcp_keepalive_timer);
715 hrtimer_init(&tcp_sk(sk)->pacing_timer, CLOCK_MONOTONIC,
716 HRTIMER_MODE_ABS_PINNED);
717 tcp_sk(sk)->pacing_timer.function = tcp_pace_kick;
718}
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}