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