Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
6 *
7 * Implementation of the Transmission Control Protocol(TCP).
8 *
9 * Authors: Ross Biro
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
20 */
21
22#include <linux/module.h>
23#include <linux/gfp.h>
24#include <net/tcp.h>
25#include <net/rstreason.h>
26
27static u32 tcp_clamp_rto_to_user_timeout(const struct sock *sk)
28{
29 const struct inet_connection_sock *icsk = inet_csk(sk);
30 const struct tcp_sock *tp = tcp_sk(sk);
31 u32 elapsed, user_timeout;
32 s32 remaining;
33
34 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
35 if (!user_timeout)
36 return icsk->icsk_rto;
37
38 elapsed = tcp_time_stamp_ts(tp) - tp->retrans_stamp;
39 if (tp->tcp_usec_ts)
40 elapsed /= USEC_PER_MSEC;
41
42 remaining = user_timeout - elapsed;
43 if (remaining <= 0)
44 return 1; /* user timeout has passed; fire ASAP */
45
46 return min_t(u32, icsk->icsk_rto, msecs_to_jiffies(remaining));
47}
48
49u32 tcp_clamp_probe0_to_user_timeout(const struct sock *sk, u32 when)
50{
51 const struct inet_connection_sock *icsk = inet_csk(sk);
52 u32 remaining, user_timeout;
53 s32 elapsed;
54
55 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
56 if (!user_timeout || !icsk->icsk_probes_tstamp)
57 return when;
58
59 elapsed = tcp_jiffies32 - icsk->icsk_probes_tstamp;
60 if (unlikely(elapsed < 0))
61 elapsed = 0;
62 remaining = msecs_to_jiffies(user_timeout) - elapsed;
63 remaining = max_t(u32, remaining, TCP_TIMEOUT_MIN);
64
65 return min_t(u32, remaining, when);
66}
67
68/**
69 * tcp_write_err() - close socket and save error info
70 * @sk: The socket the error has appeared on.
71 *
72 * Returns: Nothing (void)
73 */
74
75static void tcp_write_err(struct sock *sk)
76{
77 tcp_done_with_error(sk, READ_ONCE(sk->sk_err_soft) ? : ETIMEDOUT);
78 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT);
79}
80
81/**
82 * tcp_out_of_resources() - Close socket if out of resources
83 * @sk: pointer to current socket
84 * @do_reset: send a last packet with reset flag
85 *
86 * Do not allow orphaned sockets to eat all our resources.
87 * This is direct violation of TCP specs, but it is required
88 * to prevent DoS attacks. It is called when a retransmission timeout
89 * or zero probe timeout occurs on orphaned socket.
90 *
91 * Also close if our net namespace is exiting; in that case there is no
92 * hope of ever communicating again since all netns interfaces are already
93 * down (or about to be down), and we need to release our dst references,
94 * which have been moved to the netns loopback interface, so the namespace
95 * can finish exiting. This condition is only possible if we are a kernel
96 * socket, as those do not hold references to the namespace.
97 *
98 * Criteria is still not confirmed experimentally and may change.
99 * We kill the socket, if:
100 * 1. If number of orphaned sockets exceeds an administratively configured
101 * limit.
102 * 2. If we have strong memory pressure.
103 * 3. If our net namespace is exiting.
104 */
105static int tcp_out_of_resources(struct sock *sk, bool do_reset)
106{
107 struct tcp_sock *tp = tcp_sk(sk);
108 int shift = 0;
109
110 /* If peer does not open window for long time, or did not transmit
111 * anything for long time, penalize it. */
112 if ((s32)(tcp_jiffies32 - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset)
113 shift++;
114
115 /* If some dubious ICMP arrived, penalize even more. */
116 if (READ_ONCE(sk->sk_err_soft))
117 shift++;
118
119 if (tcp_check_oom(sk, shift)) {
120 /* Catch exceptional cases, when connection requires reset.
121 * 1. Last segment was sent recently. */
122 if ((s32)(tcp_jiffies32 - tp->lsndtime) <= TCP_TIMEWAIT_LEN ||
123 /* 2. Window is closed. */
124 (!tp->snd_wnd && !tp->packets_out))
125 do_reset = true;
126 if (do_reset)
127 tcp_send_active_reset(sk, GFP_ATOMIC,
128 SK_RST_REASON_TCP_ABORT_ON_MEMORY);
129 tcp_done(sk);
130 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY);
131 return 1;
132 }
133
134 if (!check_net(sock_net(sk))) {
135 /* Not possible to send reset; just close */
136 tcp_done(sk);
137 return 1;
138 }
139
140 return 0;
141}
142
143/**
144 * tcp_orphan_retries() - Returns maximal number of retries on an orphaned socket
145 * @sk: Pointer to the current socket.
146 * @alive: bool, socket alive state
147 */
148static int tcp_orphan_retries(struct sock *sk, bool alive)
149{
150 int retries = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_orphan_retries); /* May be zero. */
151
152 /* We know from an ICMP that something is wrong. */
153 if (READ_ONCE(sk->sk_err_soft) && !alive)
154 retries = 0;
155
156 /* However, if socket sent something recently, select some safe
157 * number of retries. 8 corresponds to >100 seconds with minimal
158 * RTO of 200msec. */
159 if (retries == 0 && alive)
160 retries = 8;
161 return retries;
162}
163
164static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk)
165{
166 const struct net *net = sock_net(sk);
167 int mss;
168
169 /* Black hole detection */
170 if (!READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing))
171 return;
172
173 if (!icsk->icsk_mtup.enabled) {
174 icsk->icsk_mtup.enabled = 1;
175 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
176 } else {
177 mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1;
178 mss = min(READ_ONCE(net->ipv4.sysctl_tcp_base_mss), mss);
179 mss = max(mss, READ_ONCE(net->ipv4.sysctl_tcp_mtu_probe_floor));
180 mss = max(mss, READ_ONCE(net->ipv4.sysctl_tcp_min_snd_mss));
181 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
182 }
183 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
184}
185
186static unsigned int tcp_model_timeout(struct sock *sk,
187 unsigned int boundary,
188 unsigned int rto_base)
189{
190 unsigned int linear_backoff_thresh, timeout;
191
192 linear_backoff_thresh = ilog2(TCP_RTO_MAX / rto_base);
193 if (boundary <= linear_backoff_thresh)
194 timeout = ((2 << boundary) - 1) * rto_base;
195 else
196 timeout = ((2 << linear_backoff_thresh) - 1) * rto_base +
197 (boundary - linear_backoff_thresh) * TCP_RTO_MAX;
198 return jiffies_to_msecs(timeout);
199}
200/**
201 * retransmits_timed_out() - returns true if this connection has timed out
202 * @sk: The current socket
203 * @boundary: max number of retransmissions
204 * @timeout: A custom timeout value.
205 * If set to 0 the default timeout is calculated and used.
206 * Using TCP_RTO_MIN and the number of unsuccessful retransmits.
207 *
208 * The default "timeout" value this function can calculate and use
209 * is equivalent to the timeout of a TCP Connection
210 * after "boundary" unsuccessful, exponentially backed-off
211 * retransmissions with an initial RTO of TCP_RTO_MIN.
212 */
213static bool retransmits_timed_out(struct sock *sk,
214 unsigned int boundary,
215 unsigned int timeout)
216{
217 struct tcp_sock *tp = tcp_sk(sk);
218 unsigned int start_ts, delta;
219
220 if (!inet_csk(sk)->icsk_retransmits)
221 return false;
222
223 start_ts = tp->retrans_stamp;
224 if (likely(timeout == 0)) {
225 unsigned int rto_base = TCP_RTO_MIN;
226
227 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
228 rto_base = tcp_timeout_init(sk);
229 timeout = tcp_model_timeout(sk, boundary, rto_base);
230 }
231
232 if (tp->tcp_usec_ts) {
233 /* delta maybe off up to a jiffy due to timer granularity. */
234 delta = tp->tcp_mstamp - start_ts + jiffies_to_usecs(1);
235 return (s32)(delta - timeout * USEC_PER_MSEC) >= 0;
236 }
237 return (s32)(tcp_time_stamp_ts(tp) - start_ts - timeout) >= 0;
238}
239
240/* A write timeout has occurred. Process the after effects. */
241static int tcp_write_timeout(struct sock *sk)
242{
243 struct inet_connection_sock *icsk = inet_csk(sk);
244 struct tcp_sock *tp = tcp_sk(sk);
245 struct net *net = sock_net(sk);
246 bool expired = false, do_reset;
247 int retry_until, max_retransmits;
248
249 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
250 if (icsk->icsk_retransmits)
251 __dst_negative_advice(sk);
252 /* Paired with WRITE_ONCE() in tcp_sock_set_syncnt() */
253 retry_until = READ_ONCE(icsk->icsk_syn_retries) ? :
254 READ_ONCE(net->ipv4.sysctl_tcp_syn_retries);
255
256 max_retransmits = retry_until;
257 if (sk->sk_state == TCP_SYN_SENT)
258 max_retransmits += READ_ONCE(net->ipv4.sysctl_tcp_syn_linear_timeouts);
259
260 expired = icsk->icsk_retransmits >= max_retransmits;
261 } else {
262 if (retransmits_timed_out(sk, READ_ONCE(net->ipv4.sysctl_tcp_retries1), 0)) {
263 /* Black hole detection */
264 tcp_mtu_probing(icsk, sk);
265
266 __dst_negative_advice(sk);
267 }
268
269 retry_until = READ_ONCE(net->ipv4.sysctl_tcp_retries2);
270 if (sock_flag(sk, SOCK_DEAD)) {
271 const bool alive = icsk->icsk_rto < TCP_RTO_MAX;
272
273 retry_until = tcp_orphan_retries(sk, alive);
274 do_reset = alive ||
275 !retransmits_timed_out(sk, retry_until, 0);
276
277 if (tcp_out_of_resources(sk, do_reset))
278 return 1;
279 }
280 }
281 if (!expired)
282 expired = retransmits_timed_out(sk, retry_until,
283 READ_ONCE(icsk->icsk_user_timeout));
284 tcp_fastopen_active_detect_blackhole(sk, expired);
285 mptcp_active_detect_blackhole(sk, expired);
286
287 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RTO_CB_FLAG))
288 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RTO_CB,
289 icsk->icsk_retransmits,
290 icsk->icsk_rto, (int)expired);
291
292 if (expired) {
293 /* Has it gone just too far? */
294 tcp_write_err(sk);
295 return 1;
296 }
297
298 if (sk_rethink_txhash(sk)) {
299 tp->timeout_rehash++;
300 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEOUTREHASH);
301 }
302
303 return 0;
304}
305
306/* Called with BH disabled */
307void tcp_delack_timer_handler(struct sock *sk)
308{
309 struct inet_connection_sock *icsk = inet_csk(sk);
310 struct tcp_sock *tp = tcp_sk(sk);
311
312 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
313 return;
314
315 /* Handling the sack compression case */
316 if (tp->compressed_ack) {
317 tcp_mstamp_refresh(tp);
318 tcp_sack_compress_send_ack(sk);
319 return;
320 }
321
322 if (!(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
323 return;
324
325 if (time_after(icsk->icsk_ack.timeout, jiffies)) {
326 sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
327 return;
328 }
329 icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
330
331 if (inet_csk_ack_scheduled(sk)) {
332 if (!inet_csk_in_pingpong_mode(sk)) {
333 /* Delayed ACK missed: inflate ATO. */
334 icsk->icsk_ack.ato = min_t(u32, icsk->icsk_ack.ato << 1, icsk->icsk_rto);
335 } else {
336 /* Delayed ACK missed: leave pingpong mode and
337 * deflate ATO.
338 */
339 inet_csk_exit_pingpong_mode(sk);
340 icsk->icsk_ack.ato = TCP_ATO_MIN;
341 }
342 tcp_mstamp_refresh(tp);
343 tcp_send_ack(sk);
344 __NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKS);
345 }
346}
347
348
349/**
350 * tcp_delack_timer() - The TCP delayed ACK timeout handler
351 * @t: Pointer to the timer. (gets casted to struct sock *)
352 *
353 * This function gets (indirectly) called when the kernel timer for a TCP packet
354 * of this socket expires. Calls tcp_delack_timer_handler() to do the actual work.
355 *
356 * Returns: Nothing (void)
357 */
358static void tcp_delack_timer(struct timer_list *t)
359{
360 struct inet_connection_sock *icsk =
361 from_timer(icsk, t, icsk_delack_timer);
362 struct sock *sk = &icsk->icsk_inet.sk;
363
364 /* Avoid taking socket spinlock if there is no ACK to send.
365 * The compressed_ack check is racy, but a separate hrtimer
366 * will take care of it eventually.
367 */
368 if (!(smp_load_acquire(&icsk->icsk_ack.pending) & ICSK_ACK_TIMER) &&
369 !READ_ONCE(tcp_sk(sk)->compressed_ack))
370 goto out;
371
372 bh_lock_sock(sk);
373 if (!sock_owned_by_user(sk)) {
374 tcp_delack_timer_handler(sk);
375 } else {
376 __NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
377 /* deleguate our work to tcp_release_cb() */
378 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &sk->sk_tsq_flags))
379 sock_hold(sk);
380 }
381 bh_unlock_sock(sk);
382out:
383 sock_put(sk);
384}
385
386static void tcp_probe_timer(struct sock *sk)
387{
388 struct inet_connection_sock *icsk = inet_csk(sk);
389 struct sk_buff *skb = tcp_send_head(sk);
390 struct tcp_sock *tp = tcp_sk(sk);
391 int max_probes;
392
393 if (tp->packets_out || !skb) {
394 icsk->icsk_probes_out = 0;
395 icsk->icsk_probes_tstamp = 0;
396 return;
397 }
398
399 /* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as
400 * long as the receiver continues to respond probes. We support this by
401 * default and reset icsk_probes_out with incoming ACKs. But if the
402 * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we
403 * kill the socket when the retry count and the time exceeds the
404 * corresponding system limit. We also implement similar policy when
405 * we use RTO to probe window in tcp_retransmit_timer().
406 */
407 if (!icsk->icsk_probes_tstamp) {
408 icsk->icsk_probes_tstamp = tcp_jiffies32;
409 } else {
410 u32 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
411
412 if (user_timeout &&
413 (s32)(tcp_jiffies32 - icsk->icsk_probes_tstamp) >=
414 msecs_to_jiffies(user_timeout))
415 goto abort;
416 }
417 max_probes = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retries2);
418 if (sock_flag(sk, SOCK_DEAD)) {
419 const bool alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX;
420
421 max_probes = tcp_orphan_retries(sk, alive);
422 if (!alive && icsk->icsk_backoff >= max_probes)
423 goto abort;
424 if (tcp_out_of_resources(sk, true))
425 return;
426 }
427
428 if (icsk->icsk_probes_out >= max_probes) {
429abort: tcp_write_err(sk);
430 } else {
431 /* Only send another probe if we didn't close things up. */
432 tcp_send_probe0(sk);
433 }
434}
435
436static void tcp_update_rto_stats(struct sock *sk)
437{
438 struct inet_connection_sock *icsk = inet_csk(sk);
439 struct tcp_sock *tp = tcp_sk(sk);
440
441 if (!icsk->icsk_retransmits) {
442 tp->total_rto_recoveries++;
443 tp->rto_stamp = tcp_time_stamp_ms(tp);
444 }
445 icsk->icsk_retransmits++;
446 tp->total_rto++;
447}
448
449/*
450 * Timer for Fast Open socket to retransmit SYNACK. Note that the
451 * sk here is the child socket, not the parent (listener) socket.
452 */
453static void tcp_fastopen_synack_timer(struct sock *sk, struct request_sock *req)
454{
455 struct inet_connection_sock *icsk = inet_csk(sk);
456 struct tcp_sock *tp = tcp_sk(sk);
457 int max_retries;
458
459 req->rsk_ops->syn_ack_timeout(req);
460
461 /* Add one more retry for fastopen.
462 * Paired with WRITE_ONCE() in tcp_sock_set_syncnt()
463 */
464 max_retries = READ_ONCE(icsk->icsk_syn_retries) ? :
465 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_synack_retries) + 1;
466
467 if (req->num_timeout >= max_retries) {
468 tcp_write_err(sk);
469 return;
470 }
471 /* Lower cwnd after certain SYNACK timeout like tcp_init_transfer() */
472 if (icsk->icsk_retransmits == 1)
473 tcp_enter_loss(sk);
474 /* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
475 * returned from rtx_syn_ack() to make it more persistent like
476 * regular retransmit because if the child socket has been accepted
477 * it's not good to give up too easily.
478 */
479 inet_rtx_syn_ack(sk, req);
480 req->num_timeout++;
481 tcp_update_rto_stats(sk);
482 if (!tp->retrans_stamp)
483 tp->retrans_stamp = tcp_time_stamp_ts(tp);
484 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
485 req->timeout << req->num_timeout, TCP_RTO_MAX);
486}
487
488static bool tcp_rtx_probe0_timed_out(const struct sock *sk,
489 const struct sk_buff *skb,
490 u32 rtx_delta)
491{
492 const struct inet_connection_sock *icsk = inet_csk(sk);
493 u32 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
494 const struct tcp_sock *tp = tcp_sk(sk);
495 int timeout = TCP_RTO_MAX * 2;
496 s32 rcv_delta;
497
498 if (user_timeout) {
499 /* If user application specified a TCP_USER_TIMEOUT,
500 * it does not want win 0 packets to 'reset the timer'
501 * while retransmits are not making progress.
502 */
503 if (rtx_delta > user_timeout)
504 return true;
505 timeout = min_t(u32, timeout, msecs_to_jiffies(user_timeout));
506 }
507 /* Note: timer interrupt might have been delayed by at least one jiffy,
508 * and tp->rcv_tstamp might very well have been written recently.
509 * rcv_delta can thus be negative.
510 */
511 rcv_delta = icsk->icsk_timeout - tp->rcv_tstamp;
512 if (rcv_delta <= timeout)
513 return false;
514
515 return msecs_to_jiffies(rtx_delta) > timeout;
516}
517
518/**
519 * tcp_retransmit_timer() - The TCP retransmit timeout handler
520 * @sk: Pointer to the current socket.
521 *
522 * This function gets called when the kernel timer for a TCP packet
523 * of this socket expires.
524 *
525 * It handles retransmission, timer adjustment and other necessary measures.
526 *
527 * Returns: Nothing (void)
528 */
529void tcp_retransmit_timer(struct sock *sk)
530{
531 struct tcp_sock *tp = tcp_sk(sk);
532 struct net *net = sock_net(sk);
533 struct inet_connection_sock *icsk = inet_csk(sk);
534 struct request_sock *req;
535 struct sk_buff *skb;
536
537 req = rcu_dereference_protected(tp->fastopen_rsk,
538 lockdep_sock_is_held(sk));
539 if (req) {
540 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
541 sk->sk_state != TCP_FIN_WAIT1);
542 tcp_fastopen_synack_timer(sk, req);
543 /* Before we receive ACK to our SYN-ACK don't retransmit
544 * anything else (e.g., data or FIN segments).
545 */
546 return;
547 }
548
549 if (!tp->packets_out)
550 return;
551
552 skb = tcp_rtx_queue_head(sk);
553 if (WARN_ON_ONCE(!skb))
554 return;
555
556 if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
557 !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
558 /* Receiver dastardly shrinks window. Our retransmits
559 * become zero probes, but we should not timeout this
560 * connection. If the socket is an orphan, time it out,
561 * we cannot allow such beasts to hang infinitely.
562 */
563 struct inet_sock *inet = inet_sk(sk);
564 u32 rtx_delta;
565
566 rtx_delta = tcp_time_stamp_ts(tp) - (tp->retrans_stamp ?:
567 tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb));
568 if (tp->tcp_usec_ts)
569 rtx_delta /= USEC_PER_MSEC;
570
571 if (sk->sk_family == AF_INET) {
572 net_dbg_ratelimited("Probing zero-window on %pI4:%u/%u, seq=%u:%u, recv %ums ago, lasting %ums\n",
573 &inet->inet_daddr, ntohs(inet->inet_dport),
574 inet->inet_num, tp->snd_una, tp->snd_nxt,
575 jiffies_to_msecs(jiffies - tp->rcv_tstamp),
576 rtx_delta);
577 }
578#if IS_ENABLED(CONFIG_IPV6)
579 else if (sk->sk_family == AF_INET6) {
580 net_dbg_ratelimited("Probing zero-window on %pI6:%u/%u, seq=%u:%u, recv %ums ago, lasting %ums\n",
581 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
582 inet->inet_num, tp->snd_una, tp->snd_nxt,
583 jiffies_to_msecs(jiffies - tp->rcv_tstamp),
584 rtx_delta);
585 }
586#endif
587 if (tcp_rtx_probe0_timed_out(sk, skb, rtx_delta)) {
588 tcp_write_err(sk);
589 goto out;
590 }
591 tcp_enter_loss(sk);
592 tcp_retransmit_skb(sk, skb, 1);
593 __sk_dst_reset(sk);
594 goto out_reset_timer;
595 }
596
597 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEOUTS);
598 if (tcp_write_timeout(sk))
599 goto out;
600
601 if (icsk->icsk_retransmits == 0) {
602 int mib_idx = 0;
603
604 if (icsk->icsk_ca_state == TCP_CA_Recovery) {
605 if (tcp_is_sack(tp))
606 mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
607 else
608 mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
609 } else if (icsk->icsk_ca_state == TCP_CA_Loss) {
610 mib_idx = LINUX_MIB_TCPLOSSFAILURES;
611 } else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
612 tp->sacked_out) {
613 if (tcp_is_sack(tp))
614 mib_idx = LINUX_MIB_TCPSACKFAILURES;
615 else
616 mib_idx = LINUX_MIB_TCPRENOFAILURES;
617 }
618 if (mib_idx)
619 __NET_INC_STATS(sock_net(sk), mib_idx);
620 }
621
622 tcp_enter_loss(sk);
623
624 tcp_update_rto_stats(sk);
625 if (tcp_retransmit_skb(sk, tcp_rtx_queue_head(sk), 1) > 0) {
626 /* Retransmission failed because of local congestion,
627 * Let senders fight for local resources conservatively.
628 */
629 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
630 TCP_RESOURCE_PROBE_INTERVAL,
631 TCP_RTO_MAX);
632 goto out;
633 }
634
635 /* Increase the timeout each time we retransmit. Note that
636 * we do not increase the rtt estimate. rto is initialized
637 * from rtt, but increases here. Jacobson (SIGCOMM 88) suggests
638 * that doubling rto each time is the least we can get away with.
639 * In KA9Q, Karn uses this for the first few times, and then
640 * goes to quadratic. netBSD doubles, but only goes up to *64,
641 * and clamps at 1 to 64 sec afterwards. Note that 120 sec is
642 * defined in the protocol as the maximum possible RTT. I guess
643 * we'll have to use something other than TCP to talk to the
644 * University of Mars.
645 *
646 * PAWS allows us longer timeouts and large windows, so once
647 * implemented ftp to mars will work nicely. We will have to fix
648 * the 120 second clamps though!
649 */
650
651out_reset_timer:
652 /* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
653 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
654 * might be increased if the stream oscillates between thin and thick,
655 * thus the old value might already be too high compared to the value
656 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
657 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
658 * exponential backoff behaviour to avoid continue hammering
659 * linear-timeout retransmissions into a black hole
660 */
661 if (sk->sk_state == TCP_ESTABLISHED &&
662 (tp->thin_lto || READ_ONCE(net->ipv4.sysctl_tcp_thin_linear_timeouts)) &&
663 tcp_stream_is_thin(tp) &&
664 icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
665 icsk->icsk_backoff = 0;
666 icsk->icsk_rto = clamp(__tcp_set_rto(tp),
667 tcp_rto_min(sk),
668 TCP_RTO_MAX);
669 } else if (sk->sk_state != TCP_SYN_SENT ||
670 tp->total_rto >
671 READ_ONCE(net->ipv4.sysctl_tcp_syn_linear_timeouts)) {
672 /* Use normal (exponential) backoff unless linear timeouts are
673 * activated.
674 */
675 icsk->icsk_backoff++;
676 icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
677 }
678 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
679 tcp_clamp_rto_to_user_timeout(sk), TCP_RTO_MAX);
680 if (retransmits_timed_out(sk, READ_ONCE(net->ipv4.sysctl_tcp_retries1) + 1, 0))
681 __sk_dst_reset(sk);
682
683out:;
684}
685
686/* Called with bottom-half processing disabled.
687 Called by tcp_write_timer() */
688void tcp_write_timer_handler(struct sock *sk)
689{
690 struct inet_connection_sock *icsk = inet_csk(sk);
691 int event;
692
693 if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
694 !icsk->icsk_pending)
695 return;
696
697 if (time_after(icsk->icsk_timeout, jiffies)) {
698 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
699 return;
700 }
701
702 tcp_mstamp_refresh(tcp_sk(sk));
703 event = icsk->icsk_pending;
704
705 switch (event) {
706 case ICSK_TIME_REO_TIMEOUT:
707 tcp_rack_reo_timeout(sk);
708 break;
709 case ICSK_TIME_LOSS_PROBE:
710 tcp_send_loss_probe(sk);
711 break;
712 case ICSK_TIME_RETRANS:
713 smp_store_release(&icsk->icsk_pending, 0);
714 tcp_retransmit_timer(sk);
715 break;
716 case ICSK_TIME_PROBE0:
717 smp_store_release(&icsk->icsk_pending, 0);
718 tcp_probe_timer(sk);
719 break;
720 }
721}
722
723static void tcp_write_timer(struct timer_list *t)
724{
725 struct inet_connection_sock *icsk =
726 from_timer(icsk, t, icsk_retransmit_timer);
727 struct sock *sk = &icsk->icsk_inet.sk;
728
729 /* Avoid locking the socket when there is no pending event. */
730 if (!smp_load_acquire(&icsk->icsk_pending))
731 goto out;
732
733 bh_lock_sock(sk);
734 if (!sock_owned_by_user(sk)) {
735 tcp_write_timer_handler(sk);
736 } else {
737 /* delegate our work to tcp_release_cb() */
738 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &sk->sk_tsq_flags))
739 sock_hold(sk);
740 }
741 bh_unlock_sock(sk);
742out:
743 sock_put(sk);
744}
745
746void tcp_syn_ack_timeout(const struct request_sock *req)
747{
748 struct net *net = read_pnet(&inet_rsk(req)->ireq_net);
749
750 __NET_INC_STATS(net, LINUX_MIB_TCPTIMEOUTS);
751}
752EXPORT_SYMBOL(tcp_syn_ack_timeout);
753
754void tcp_set_keepalive(struct sock *sk, int val)
755{
756 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
757 return;
758
759 if (val && !sock_flag(sk, SOCK_KEEPOPEN))
760 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
761 else if (!val)
762 inet_csk_delete_keepalive_timer(sk);
763}
764EXPORT_SYMBOL_GPL(tcp_set_keepalive);
765
766
767static void tcp_keepalive_timer (struct timer_list *t)
768{
769 struct sock *sk = from_timer(sk, t, sk_timer);
770 struct inet_connection_sock *icsk = inet_csk(sk);
771 struct tcp_sock *tp = tcp_sk(sk);
772 u32 elapsed;
773
774 /* Only process if socket is not in use. */
775 bh_lock_sock(sk);
776 if (sock_owned_by_user(sk)) {
777 /* Try again later. */
778 inet_csk_reset_keepalive_timer (sk, HZ/20);
779 goto out;
780 }
781
782 if (sk->sk_state == TCP_LISTEN) {
783 pr_err("Hmm... keepalive on a LISTEN ???\n");
784 goto out;
785 }
786
787 tcp_mstamp_refresh(tp);
788 if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
789 if (READ_ONCE(tp->linger2) >= 0) {
790 const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
791
792 if (tmo > 0) {
793 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
794 goto out;
795 }
796 }
797 tcp_send_active_reset(sk, GFP_ATOMIC, SK_RST_REASON_TCP_STATE);
798 goto death;
799 }
800
801 if (!sock_flag(sk, SOCK_KEEPOPEN) ||
802 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_SYN_SENT)))
803 goto out;
804
805 elapsed = keepalive_time_when(tp);
806
807 /* It is alive without keepalive 8) */
808 if (tp->packets_out || !tcp_write_queue_empty(sk))
809 goto resched;
810
811 elapsed = keepalive_time_elapsed(tp);
812
813 if (elapsed >= keepalive_time_when(tp)) {
814 u32 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
815
816 /* If the TCP_USER_TIMEOUT option is enabled, use that
817 * to determine when to timeout instead.
818 */
819 if ((user_timeout != 0 &&
820 elapsed >= msecs_to_jiffies(user_timeout) &&
821 icsk->icsk_probes_out > 0) ||
822 (user_timeout == 0 &&
823 icsk->icsk_probes_out >= keepalive_probes(tp))) {
824 tcp_send_active_reset(sk, GFP_ATOMIC,
825 SK_RST_REASON_TCP_KEEPALIVE_TIMEOUT);
826 tcp_write_err(sk);
827 goto out;
828 }
829 if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) {
830 icsk->icsk_probes_out++;
831 elapsed = keepalive_intvl_when(tp);
832 } else {
833 /* If keepalive was lost due to local congestion,
834 * try harder.
835 */
836 elapsed = TCP_RESOURCE_PROBE_INTERVAL;
837 }
838 } else {
839 /* It is tp->rcv_tstamp + keepalive_time_when(tp) */
840 elapsed = keepalive_time_when(tp) - elapsed;
841 }
842
843resched:
844 inet_csk_reset_keepalive_timer (sk, elapsed);
845 goto out;
846
847death:
848 tcp_done(sk);
849
850out:
851 bh_unlock_sock(sk);
852 sock_put(sk);
853}
854
855static enum hrtimer_restart tcp_compressed_ack_kick(struct hrtimer *timer)
856{
857 struct tcp_sock *tp = container_of(timer, struct tcp_sock, compressed_ack_timer);
858 struct sock *sk = (struct sock *)tp;
859
860 bh_lock_sock(sk);
861 if (!sock_owned_by_user(sk)) {
862 if (tp->compressed_ack) {
863 /* Since we have to send one ack finally,
864 * subtract one from tp->compressed_ack to keep
865 * LINUX_MIB_TCPACKCOMPRESSED accurate.
866 */
867 tp->compressed_ack--;
868 tcp_mstamp_refresh(tp);
869 tcp_send_ack(sk);
870 }
871 } else {
872 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED,
873 &sk->sk_tsq_flags))
874 sock_hold(sk);
875 }
876 bh_unlock_sock(sk);
877
878 sock_put(sk);
879
880 return HRTIMER_NORESTART;
881}
882
883void tcp_init_xmit_timers(struct sock *sk)
884{
885 inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
886 &tcp_keepalive_timer);
887 hrtimer_init(&tcp_sk(sk)->pacing_timer, CLOCK_MONOTONIC,
888 HRTIMER_MODE_ABS_PINNED_SOFT);
889 tcp_sk(sk)->pacing_timer.function = tcp_pace_kick;
890
891 hrtimer_init(&tcp_sk(sk)->compressed_ack_timer, CLOCK_MONOTONIC,
892 HRTIMER_MODE_REL_PINNED_SOFT);
893 tcp_sk(sk)->compressed_ack_timer.function = tcp_compressed_ack_kick;
894}
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}