projects/linux·blob master · refs

907 lines · 26.6 KiB

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, <[email protected]>
11 * Mark Evans, <[email protected]>
12 * Corey Minyard <[email protected]>
13 * Florian La Roche, <[email protected]>
14 * Charles Hedrick, <[email protected]>
15 * Linus Torvalds, <[email protected]>
16 * Alan Cox, <[email protected]>
17 * Matthew Dillon, <[email protected]>
18 * Arnt Gulbrandsen, <[email protected]>
19 * Jorge Cwik, <[email protected]>
20 */
21
22#include <linux/module.h>
23#include <linux/gfp.h>
24#include <net/tcp.h>
25#include <net/tcp_ecn.h>
26#include <net/rstreason.h>
27
28static u32 tcp_clamp_rto_to_user_timeout(const struct sock *sk)
29{
30 const struct inet_connection_sock *icsk = inet_csk(sk);
31 const struct tcp_sock *tp = tcp_sk(sk);
32 u32 elapsed, user_timeout;
33 s32 remaining;
34
35 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
36 if (!user_timeout)
37 return icsk->icsk_rto;
38
39 elapsed = tcp_time_stamp_ts(tp) - tp->retrans_stamp;
40 if (tp->tcp_usec_ts)
41 elapsed /= USEC_PER_MSEC;
42
43 remaining = user_timeout - elapsed;
44 if (remaining <= 0)
45 return 1; /* user timeout has passed; fire ASAP */
46
47 return min_t(u32, icsk->icsk_rto, msecs_to_jiffies(remaining));
48}
49
50u32 tcp_clamp_probe0_to_user_timeout(const struct sock *sk, u32 when)
51{
52 const struct inet_connection_sock *icsk = inet_csk(sk);
53 u32 user_timeout;
54 s32 remaining;
55 s32 elapsed;
56
57 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
58 if (!user_timeout || !icsk->icsk_probes_tstamp)
59 return when;
60
61 elapsed = tcp_jiffies32 - icsk->icsk_probes_tstamp;
62 if (unlikely(elapsed < 0))
63 elapsed = 0;
64 remaining = msecs_to_jiffies(user_timeout) - elapsed;
65 remaining = max_t(int, remaining, TCP_TIMEOUT_MIN);
66
67 return min_t(u32, remaining, when);
68}
69
70/**
71 * tcp_write_err() - close socket and save error info
72 * @sk: The socket the error has appeared on.
73 *
74 * Returns: Nothing (void)
75 */
76
77static void tcp_write_err(struct sock *sk)
78{
79 tcp_done_with_error(sk, READ_ONCE(sk->sk_err_soft) ? : ETIMEDOUT);
80 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT);
81}
82
83/**
84 * tcp_out_of_resources() - Close socket if out of resources
85 * @sk: pointer to current socket
86 * @do_reset: send a last packet with reset flag
87 *
88 * Do not allow orphaned sockets to eat all our resources.
89 * This is direct violation of TCP specs, but it is required
90 * to prevent DoS attacks. It is called when a retransmission timeout
91 * or zero probe timeout occurs on orphaned socket.
92 *
93 * Also close if our net namespace is exiting; in that case there is no
94 * hope of ever communicating again since all netns interfaces are already
95 * down (or about to be down), and we need to release our dst references,
96 * which have been moved to the netns loopback interface, so the namespace
97 * can finish exiting. This condition is only possible if we are a kernel
98 * socket, as those do not hold references to the namespace.
99 *
100 * Criteria is still not confirmed experimentally and may change.
101 * We kill the socket, if:
102 * 1. If number of orphaned sockets exceeds an administratively configured
103 * limit.
104 * 2. If we have strong memory pressure.
105 * 3. If our net namespace is exiting.
106 */
107static int tcp_out_of_resources(struct sock *sk, bool do_reset)
108{
109 struct tcp_sock *tp = tcp_sk(sk);
110 int shift = 0;
111
112 /* If peer does not open window for long time, or did not transmit
113 * anything for long time, penalize it. */
114 if ((s32)(tcp_jiffies32 - tp->lsndtime) > 2*tcp_rto_max(sk) || !do_reset)
115 shift++;
116
117 /* If some dubious ICMP arrived, penalize even more. */
118 if (READ_ONCE(sk->sk_err_soft))
119 shift++;
120
121 if (tcp_check_oom(sk, shift)) {
122 /* Catch exceptional cases, when connection requires reset.
123 * 1. Last segment was sent recently. */
124 if ((s32)(tcp_jiffies32 - tp->lsndtime) <= TCP_TIMEWAIT_LEN ||
125 /* 2. Window is closed. */
126 (!tp->snd_wnd && !tp->packets_out))
127 do_reset = true;
128 if (do_reset)
129 tcp_send_active_reset(sk,
130 SK_RST_REASON_TCP_ABORT_ON_MEMORY);
131 tcp_done(sk);
132 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY);
133 return 1;
134 }
135
136 if (!check_net(sock_net(sk))) {
137 /* Not possible to send reset; just close */
138 tcp_done(sk);
139 return 1;
140 }
141
142 return 0;
143}
144
145/**
146 * tcp_orphan_retries() - Returns maximal number of retries on an orphaned socket
147 * @sk: Pointer to the current socket.
148 * @alive: bool, socket alive state
149 */
150static int tcp_orphan_retries(struct sock *sk, bool alive)
151{
152 int retries = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_orphan_retries); /* May be zero. */
153
154 /* We know from an ICMP that something is wrong. */
155 if (READ_ONCE(sk->sk_err_soft) && !alive)
156 retries = 0;
157
158 /* However, if socket sent something recently, select some safe
159 * number of retries. 8 corresponds to >100 seconds with minimal
160 * RTO of 200msec. */
161 if (retries == 0 && alive)
162 retries = 8;
163 return retries;
164}
165
166static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk)
167{
168 const struct net *net = sock_net(sk);
169 int mss;
170
171 /* Black hole detection */
172 if (!READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing))
173 return;
174
175 if (!icsk->icsk_mtup.enabled) {
176 icsk->icsk_mtup.enabled = 1;
177 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
178 } else {
179 mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1;
180 mss = min(READ_ONCE(net->ipv4.sysctl_tcp_base_mss), mss);
181 mss = max(mss, READ_ONCE(net->ipv4.sysctl_tcp_mtu_probe_floor));
182 mss = max(mss, READ_ONCE(net->ipv4.sysctl_tcp_min_snd_mss));
183 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
184 }
185 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
186}
187
188static unsigned int tcp_model_timeout(struct sock *sk,
189 unsigned int boundary,
190 unsigned int rto_base)
191{
192 unsigned int linear_backoff_thresh, timeout;
193
194 linear_backoff_thresh = ilog2(tcp_rto_max(sk) / rto_base);
195 if (boundary <= linear_backoff_thresh)
196 timeout = ((2 << boundary) - 1) * rto_base;
197 else
198 timeout = ((2 << linear_backoff_thresh) - 1) * rto_base +
199 (boundary - linear_backoff_thresh) * tcp_rto_max(sk);
200 return jiffies_to_msecs(timeout);
201}
202/**
203 * retransmits_timed_out() - returns true if this connection has timed out
204 * @sk: The current socket
205 * @boundary: max number of retransmissions
206 * @timeout: A custom timeout value.
207 * If set to 0 the default timeout is calculated and used.
208 * Using TCP_RTO_MIN and the number of unsuccessful retransmits.
209 *
210 * The default "timeout" value this function can calculate and use
211 * is equivalent to the timeout of a TCP Connection
212 * after "boundary" unsuccessful, exponentially backed-off
213 * retransmissions with an initial RTO of TCP_RTO_MIN.
214 */
215static bool retransmits_timed_out(struct sock *sk,
216 unsigned int boundary,
217 unsigned int timeout)
218{
219 struct tcp_sock *tp = tcp_sk(sk);
220 unsigned int start_ts, delta;
221
222 if (!inet_csk(sk)->icsk_retransmits)
223 return false;
224
225 start_ts = tp->retrans_stamp;
226 if (likely(timeout == 0)) {
227 unsigned int rto_base = TCP_RTO_MIN;
228
229 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
230 rto_base = tcp_timeout_init(sk);
231 timeout = tcp_model_timeout(sk, boundary, rto_base);
232 }
233
234 if (tp->tcp_usec_ts) {
235 /* delta maybe off up to a jiffy due to timer granularity. */
236 delta = tp->tcp_mstamp - start_ts + jiffies_to_usecs(1);
237 return (s32)(delta - timeout * USEC_PER_MSEC) >= 0;
238 }
239 return (s32)(tcp_time_stamp_ts(tp) - start_ts - timeout) >= 0;
240}
241
242/* A write timeout has occurred. Process the after effects. */
243static int tcp_write_timeout(struct sock *sk)
244{
245 struct inet_connection_sock *icsk = inet_csk(sk);
246 struct tcp_sock *tp = tcp_sk(sk);
247 struct net *net = sock_net(sk);
248 bool expired = false, do_reset;
249 int retry_until, max_retransmits;
250
251 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
252 if (icsk->icsk_retransmits)
253 __dst_negative_advice(sk);
254 /* Paired with WRITE_ONCE() in tcp_sock_set_syncnt() */
255 retry_until = READ_ONCE(icsk->icsk_syn_retries) ? :
256 READ_ONCE(net->ipv4.sysctl_tcp_syn_retries);
257
258 max_retransmits = retry_until;
259 if (sk->sk_state == TCP_SYN_SENT)
260 max_retransmits += READ_ONCE(net->ipv4.sysctl_tcp_syn_linear_timeouts);
261
262 expired = icsk->icsk_retransmits >= max_retransmits;
263 } else {
264 if (retransmits_timed_out(sk, READ_ONCE(net->ipv4.sysctl_tcp_retries1), 0)) {
265 /* Black hole detection */
266 tcp_mtu_probing(icsk, sk);
267
268 __dst_negative_advice(sk);
269 }
270
271 retry_until = READ_ONCE(net->ipv4.sysctl_tcp_retries2);
272 if (sock_flag(sk, SOCK_DEAD)) {
273 const bool alive = icsk->icsk_rto < tcp_rto_max(sk);
274
275 retry_until = tcp_orphan_retries(sk, alive);
276 do_reset = alive ||
277 !retransmits_timed_out(sk, retry_until, 0);
278
279 if (tcp_out_of_resources(sk, do_reset))
280 return 1;
281 }
282 }
283 if (!expired)
284 expired = retransmits_timed_out(sk, retry_until,
285 READ_ONCE(icsk->icsk_user_timeout));
286 tcp_fastopen_active_detect_blackhole(sk, expired);
287 mptcp_active_detect_blackhole(sk, expired);
288
289 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RTO_CB_FLAG))
290 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RTO_CB,
291 icsk->icsk_retransmits,
292 icsk->icsk_rto, (int)expired);
293
294 if (expired) {
295 /* Has it gone just too far? */
296 tcp_write_err(sk);
297 return 1;
298 }
299
300 if (__sk_rethink_txhash_reset_dst(sk)) {
301 WRITE_ONCE(tp->timeout_rehash, tp->timeout_rehash + 1);
302 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEOUTREHASH);
303 }
304
305 return 0;
306}
307
308/* Called with BH disabled */
309void tcp_delack_timer_handler(struct sock *sk)
310{
311 struct inet_connection_sock *icsk = inet_csk(sk);
312 struct tcp_sock *tp = tcp_sk(sk);
313
314 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
315 return;
316
317 /* Handling the sack compression case */
318 if (tp->compressed_ack) {
319 tcp_mstamp_refresh(tp);
320 tcp_sack_compress_send_ack(sk);
321 return;
322 }
323
324 if (!(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
325 return;
326
327 if (time_after(icsk_delack_timeout(icsk), jiffies)) {
328 sk_reset_timer(sk, &icsk->icsk_delack_timer,
329 icsk_delack_timeout(icsk));
330 return;
331 }
332 icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
333
334 if (inet_csk_ack_scheduled(sk)) {
335 if (!inet_csk_in_pingpong_mode(sk)) {
336 /* Delayed ACK missed: inflate ATO. */
337 icsk->icsk_ack.ato = min3((u32)icsk->icsk_ack.ato << 1,
338 icsk->icsk_rto,
339 (u32)TCP_DELACK_MAX);
340 } else {
341 /* Delayed ACK missed: leave pingpong mode and
342 * deflate ATO.
343 */
344 inet_csk_exit_pingpong_mode(sk);
345 icsk->icsk_ack.ato = TCP_ATO_MIN;
346 }
347 tcp_mstamp_refresh(tp);
348 tcp_send_ack(sk);
349 __NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKS);
350 }
351}
352
353
354/**
355 * tcp_delack_timer() - The TCP delayed ACK timeout handler
356 * @t: Pointer to the timer. (gets casted to struct sock *)
357 *
358 * This function gets (indirectly) called when the kernel timer for a TCP packet
359 * of this socket expires. Calls tcp_delack_timer_handler() to do the actual work.
360 *
361 * Returns: Nothing (void)
362 */
363static void tcp_delack_timer(struct timer_list *t)
364{
365 struct inet_connection_sock *icsk =
366 timer_container_of(icsk, t, icsk_delack_timer);
367 struct sock *sk = &icsk->icsk_inet.sk;
368
369 /* Avoid taking socket spinlock if there is no ACK to send.
370 * The compressed_ack check is racy, but a separate hrtimer
371 * will take care of it eventually.
372 */
373 if (!(smp_load_acquire(&icsk->icsk_ack.pending) & ICSK_ACK_TIMER) &&
374 !READ_ONCE(tcp_sk(sk)->compressed_ack))
375 goto out;
376
377 bh_lock_sock(sk);
378 if (!sock_owned_by_user(sk)) {
379 tcp_delack_timer_handler(sk);
380 } else {
381 __NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
382 /* deleguate our work to tcp_release_cb() */
383 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &sk->sk_tsq_flags))
384 sock_hold(sk);
385 }
386 bh_unlock_sock(sk);
387out:
388 sock_put(sk);
389}
390
391static void tcp_probe_timer(struct sock *sk)
392{
393 struct inet_connection_sock *icsk = inet_csk(sk);
394 struct sk_buff *skb = tcp_send_head(sk);
395 struct tcp_sock *tp = tcp_sk(sk);
396 int max_probes;
397
398 if (tp->packets_out || !skb) {
399 WRITE_ONCE(icsk->icsk_probes_out, 0);
400 icsk->icsk_probes_tstamp = 0;
401 return;
402 }
403
404 /* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as
405 * long as the receiver continues to respond probes. We support this by
406 * default and reset icsk_probes_out with incoming ACKs. But if the
407 * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we
408 * kill the socket when the retry count and the time exceeds the
409 * corresponding system limit. We also implement similar policy when
410 * we use RTO to probe window in tcp_retransmit_timer().
411 */
412 if (!icsk->icsk_probes_tstamp) {
413 icsk->icsk_probes_tstamp = tcp_jiffies32;
414 } else {
415 u32 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
416
417 if (user_timeout &&
418 (s32)(tcp_jiffies32 - icsk->icsk_probes_tstamp) >=
419 msecs_to_jiffies(user_timeout))
420 goto abort;
421 }
422 max_probes = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retries2);
423 if (sock_flag(sk, SOCK_DEAD)) {
424 unsigned int rto_max = tcp_rto_max(sk);
425 const bool alive = inet_csk_rto_backoff(icsk, rto_max) < rto_max;
426
427 max_probes = tcp_orphan_retries(sk, alive);
428 if (!alive && icsk->icsk_backoff >= max_probes)
429 goto abort;
430 if (tcp_out_of_resources(sk, true))
431 return;
432 }
433
434 if (icsk->icsk_probes_out >= max_probes) {
435abort: tcp_write_err(sk);
436 } else {
437 /* Only send another probe if we didn't close things up. */
438 tcp_send_probe0(sk);
439 }
440}
441
442static void tcp_update_rto_stats(struct sock *sk)
443{
444 struct inet_connection_sock *icsk = inet_csk(sk);
445 struct tcp_sock *tp = tcp_sk(sk);
446
447 if (!icsk->icsk_retransmits) {
448 tp->total_rto_recoveries++;
449 tp->rto_stamp = tcp_time_stamp_ms(tp);
450 }
451 WRITE_ONCE(icsk->icsk_retransmits, icsk->icsk_retransmits + 1);
452 tp->total_rto++;
453}
454
455/*
456 * Timer for Fast Open socket to retransmit SYNACK. Note that the
457 * sk here is the child socket, not the parent (listener) socket.
458 */
459static void tcp_fastopen_synack_timer(struct sock *sk, struct request_sock *req)
460{
461 struct inet_connection_sock *icsk = inet_csk(sk);
462 struct tcp_sock *tp = tcp_sk(sk);
463 int max_retries;
464
465 tcp_syn_ack_timeout(req);
466
467 /* Add one more retry for fastopen.
468 * Paired with WRITE_ONCE() in tcp_sock_set_syncnt()
469 */
470 max_retries = READ_ONCE(icsk->icsk_syn_retries) ? :
471 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_synack_retries) + 1;
472
473 if (req->num_timeout >= max_retries) {
474 tcp_write_err(sk);
475 return;
476 }
477 /* Lower cwnd after certain SYNACK timeout like tcp_init_transfer() */
478 if (icsk->icsk_retransmits == 1)
479 tcp_enter_loss(sk);
480 /* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
481 * returned from rtx_syn_ack() to make it more persistent like
482 * regular retransmit because if the child socket has been accepted
483 * it's not good to give up too easily.
484 */
485 tcp_rtx_synack(sk, req);
486 if (req->num_retrans > 1 && tcp_rsk(req)->accecn_ok)
487 tcp_rsk(req)->accecn_fail_mode |= TCP_ACCECN_ACE_FAIL_SEND;
488 req->num_timeout++;
489 tcp_update_rto_stats(sk);
490 if (!tp->retrans_stamp)
491 tp->retrans_stamp = tcp_time_stamp_ts(tp);
492 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
493 req->timeout << req->num_timeout, false);
494}
495
496static bool tcp_rtx_probe0_timed_out(const struct sock *sk,
497 const struct sk_buff *skb,
498 u32 rtx_delta)
499{
500 const struct inet_connection_sock *icsk = inet_csk(sk);
501 u32 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
502 const struct tcp_sock *tp = tcp_sk(sk);
503 int timeout = tcp_rto_max(sk) * 2;
504 s32 rcv_delta;
505
506 if (user_timeout) {
507 /* If user application specified a TCP_USER_TIMEOUT,
508 * it does not want win 0 packets to 'reset the timer'
509 * while retransmits are not making progress.
510 */
511 if (rtx_delta > user_timeout)
512 return true;
513 timeout = min_t(u32, timeout, msecs_to_jiffies(user_timeout));
514 }
515 /* Note: timer interrupt might have been delayed by at least one jiffy,
516 * and tp->rcv_tstamp might very well have been written recently.
517 * rcv_delta can thus be negative.
518 */
519 rcv_delta = tcp_timeout_expires(sk) - tp->rcv_tstamp;
520 if (rcv_delta <= timeout)
521 return false;
522
523 return msecs_to_jiffies(rtx_delta) > timeout;
524}
525
526/**
527 * tcp_retransmit_timer() - The TCP retransmit timeout handler
528 * @sk: Pointer to the current socket.
529 *
530 * This function gets called when the kernel timer for a TCP packet
531 * of this socket expires.
532 *
533 * It handles retransmission, timer adjustment and other necessary measures.
534 *
535 * Returns: Nothing (void)
536 */
537void tcp_retransmit_timer(struct sock *sk)
538{
539 struct tcp_sock *tp = tcp_sk(sk);
540 struct net *net = sock_net(sk);
541 struct inet_connection_sock *icsk = inet_csk(sk);
542 struct request_sock *req;
543 struct sk_buff *skb;
544
545 req = rcu_dereference_protected(tp->fastopen_rsk,
546 lockdep_sock_is_held(sk));
547 if (req) {
548 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
549 sk->sk_state != TCP_FIN_WAIT1);
550 tcp_fastopen_synack_timer(sk, req);
551 /* Before we receive ACK to our SYN-ACK don't retransmit
552 * anything else (e.g., data or FIN segments).
553 */
554 return;
555 }
556
557 if (!tp->packets_out)
558 return;
559
560 skb = tcp_rtx_queue_head(sk);
561 if (WARN_ON_ONCE(!skb))
562 return;
563
564 if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
565 !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
566 /* Receiver dastardly shrinks window. Our retransmits
567 * become zero probes, but we should not timeout this
568 * connection. If the socket is an orphan, time it out,
569 * we cannot allow such beasts to hang infinitely.
570 */
571 struct inet_sock *inet = inet_sk(sk);
572 u32 rtx_delta;
573
574 rtx_delta = tcp_time_stamp_ts(tp) - (tp->retrans_stamp ?:
575 tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb));
576 if (tp->tcp_usec_ts)
577 rtx_delta /= USEC_PER_MSEC;
578
579 if (sk->sk_family == AF_INET) {
580 net_dbg_ratelimited("Probing zero-window on %pI4:%u/%u, seq=%u:%u, recv %ums ago, lasting %ums\n",
581 &inet->inet_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#if IS_ENABLED(CONFIG_IPV6)
587 else if (sk->sk_family == AF_INET6) {
588 net_dbg_ratelimited("Probing zero-window on %pI6:%u/%u, seq=%u:%u, recv %ums ago, lasting %ums\n",
589 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
590 inet->inet_num, tp->snd_una, tp->snd_nxt,
591 jiffies_to_msecs(jiffies - tp->rcv_tstamp),
592 rtx_delta);
593 }
594#endif
595 if (tcp_rtx_probe0_timed_out(sk, skb, rtx_delta)) {
596 tcp_write_err(sk);
597 goto out;
598 }
599 tcp_enter_loss(sk);
600 tcp_retransmit_skb(sk, skb, 1);
601 __sk_dst_reset(sk);
602 goto out_reset_timer;
603 }
604
605 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEOUTS);
606 if (tcp_write_timeout(sk))
607 goto out;
608
609 if (icsk->icsk_retransmits == 0) {
610 int mib_idx = 0;
611
612 if (icsk->icsk_ca_state == TCP_CA_Recovery) {
613 if (tcp_is_sack(tp))
614 mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
615 else
616 mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
617 } else if (icsk->icsk_ca_state == TCP_CA_Loss) {
618 mib_idx = LINUX_MIB_TCPLOSSFAILURES;
619 } else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
620 tp->sacked_out) {
621 if (tcp_is_sack(tp))
622 mib_idx = LINUX_MIB_TCPSACKFAILURES;
623 else
624 mib_idx = LINUX_MIB_TCPRENOFAILURES;
625 }
626 if (mib_idx)
627 __NET_INC_STATS(sock_net(sk), mib_idx);
628 }
629
630 tcp_enter_loss(sk);
631
632 tcp_update_rto_stats(sk);
633 if (tcp_retransmit_skb(sk, tcp_rtx_queue_head(sk), 1) > 0) {
634 /* Retransmission failed because of local congestion,
635 * Let senders fight for local resources conservatively.
636 */
637 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
638 TCP_RESOURCE_PROBE_INTERVAL,
639 false);
640 goto out;
641 }
642
643 /* Increase the timeout each time we retransmit. Note that
644 * we do not increase the rtt estimate. rto is initialized
645 * from rtt, but increases here. Jacobson (SIGCOMM 88) suggests
646 * that doubling rto each time is the least we can get away with.
647 * In KA9Q, Karn uses this for the first few times, and then
648 * goes to quadratic. netBSD doubles, but only goes up to *64,
649 * and clamps at 1 to 64 sec afterwards. Note that 120 sec is
650 * defined in the protocol as the maximum possible RTT. I guess
651 * we'll have to use something other than TCP to talk to the
652 * University of Mars.
653 *
654 * PAWS allows us longer timeouts and large windows, so once
655 * implemented ftp to mars will work nicely. We will have to fix
656 * the 120 second clamps though!
657 */
658
659out_reset_timer:
660 /* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
661 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
662 * might be increased if the stream oscillates between thin and thick,
663 * thus the old value might already be too high compared to the value
664 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
665 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
666 * exponential backoff behaviour to avoid continue hammering
667 * linear-timeout retransmissions into a black hole
668 */
669 if (sk->sk_state == TCP_ESTABLISHED &&
670 (tp->thin_lto || READ_ONCE(net->ipv4.sysctl_tcp_thin_linear_timeouts)) &&
671 tcp_stream_is_thin(tp) &&
672 icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
673 icsk->icsk_backoff = 0;
674 icsk->icsk_rto = clamp(__tcp_set_rto(tp),
675 tcp_rto_min(sk),
676 tcp_rto_max(sk));
677 } else if (sk->sk_state != TCP_SYN_SENT ||
678 tp->total_rto >
679 READ_ONCE(net->ipv4.sysctl_tcp_syn_linear_timeouts)) {
680 /* Use normal (exponential) backoff unless linear timeouts are
681 * activated.
682 */
683 icsk->icsk_backoff++;
684 icsk->icsk_rto = min(icsk->icsk_rto << 1, tcp_rto_max(sk));
685 }
686 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
687 tcp_clamp_rto_to_user_timeout(sk), false);
688 if (retransmits_timed_out(sk, READ_ONCE(net->ipv4.sysctl_tcp_retries1) + 1, 0))
689 __sk_dst_reset(sk);
690
691out:;
692}
693
694/* Called with bottom-half processing disabled.
695 * Called by tcp_write_timer() and tcp_release_cb().
696 */
697void tcp_write_timer_handler(struct sock *sk)
698{
699 struct inet_connection_sock *icsk = inet_csk(sk);
700 int event;
701
702 if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
703 !icsk->icsk_pending)
704 return;
705
706 if (time_after(tcp_timeout_expires(sk), jiffies)) {
707 sk_reset_timer(sk, &sk->tcp_retransmit_timer,
708 tcp_timeout_expires(sk));
709 return;
710 }
711 tcp_mstamp_refresh(tcp_sk(sk));
712 event = icsk->icsk_pending;
713
714 switch (event) {
715 case ICSK_TIME_REO_TIMEOUT:
716 tcp_rack_reo_timeout(sk);
717 break;
718 case ICSK_TIME_LOSS_PROBE:
719 tcp_send_loss_probe(sk);
720 break;
721 case ICSK_TIME_RETRANS:
722 smp_store_release(&icsk->icsk_pending, 0);
723 tcp_retransmit_timer(sk);
724 break;
725 case ICSK_TIME_PROBE0:
726 smp_store_release(&icsk->icsk_pending, 0);
727 tcp_probe_timer(sk);
728 break;
729 }
730}
731
732static void tcp_write_timer(struct timer_list *t)
733{
734 struct sock *sk = timer_container_of(sk, t, tcp_retransmit_timer);
735
736 /* Avoid locking the socket when there is no pending event. */
737 if (!smp_load_acquire(&inet_csk(sk)->icsk_pending))
738 goto out;
739
740 bh_lock_sock(sk);
741 if (!sock_owned_by_user(sk)) {
742 tcp_write_timer_handler(sk);
743 } else {
744 /* delegate our work to tcp_release_cb() */
745 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &sk->sk_tsq_flags))
746 sock_hold(sk);
747 }
748 bh_unlock_sock(sk);
749out:
750 sock_put(sk);
751}
752
753noinline_for_tracing void tcp_syn_ack_timeout(const struct request_sock *req)
754{
755 struct net *net = read_pnet(&inet_rsk(req)->ireq_net);
756
757 __NET_INC_STATS(net, LINUX_MIB_TCPTIMEOUTS);
758}
759
760void tcp_reset_keepalive_timer(struct sock *sk, unsigned long len)
761{
762 sk_reset_timer(sk, &inet_csk(sk)->icsk_keepalive_timer, jiffies + len);
763}
764
765static void tcp_delete_keepalive_timer(struct sock *sk)
766{
767 sk_stop_timer(sk, &inet_csk(sk)->icsk_keepalive_timer);
768}
769
770void tcp_set_keepalive(struct sock *sk, int val)
771{
772 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
773 return;
774
775 if (val && !sock_flag(sk, SOCK_KEEPOPEN))
776 tcp_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
777 else if (!val)
778 tcp_delete_keepalive_timer(sk);
779}
780
781static void tcp_keepalive_timer(struct timer_list *t)
782{
783 struct inet_connection_sock *icsk =
784 timer_container_of(icsk, t, icsk_keepalive_timer);
785 struct sock *sk = &icsk->icsk_inet.sk;
786 struct tcp_sock *tp = tcp_sk(sk);
787 u32 elapsed;
788
789 /* Only process if socket is not in use. */
790 bh_lock_sock(sk);
791 if (sock_owned_by_user(sk)) {
792 /* Try again later. */
793 tcp_reset_keepalive_timer(sk, HZ/20);
794 goto out;
795 }
796
797 if (sk->sk_state == TCP_LISTEN) {
798 pr_err("Hmm... keepalive on a LISTEN ???\n");
799 goto out;
800 }
801
802 tcp_mstamp_refresh(tp);
803 if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
804 if (READ_ONCE(tp->linger2) >= 0) {
805 const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
806
807 if (tmo > 0) {
808 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
809 goto out;
810 }
811 }
812 tcp_send_active_reset(sk, SK_RST_REASON_TCP_STATE);
813 goto death;
814 }
815
816 if (!sock_flag(sk, SOCK_KEEPOPEN) ||
817 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_SYN_SENT)))
818 goto out;
819
820 elapsed = keepalive_time_when(tp);
821
822 /* It is alive without keepalive 8) */
823 if (tp->packets_out || !tcp_write_queue_empty(sk))
824 goto resched;
825
826 elapsed = keepalive_time_elapsed(tp);
827
828 if (elapsed >= keepalive_time_when(tp)) {
829 u32 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
830
831 /* If the TCP_USER_TIMEOUT option is enabled, use that
832 * to determine when to timeout instead.
833 */
834 if ((user_timeout != 0 &&
835 elapsed >= msecs_to_jiffies(user_timeout) &&
836 icsk->icsk_probes_out > 0) ||
837 (user_timeout == 0 &&
838 icsk->icsk_probes_out >= keepalive_probes(tp))) {
839 tcp_send_active_reset(sk,
840 SK_RST_REASON_TCP_KEEPALIVE_TIMEOUT);
841 tcp_write_err(sk);
842 goto out;
843 }
844 if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) {
845 WRITE_ONCE(icsk->icsk_probes_out, icsk->icsk_probes_out + 1);
846 elapsed = keepalive_intvl_when(tp);
847 } else {
848 /* If keepalive was lost due to local congestion,
849 * try harder.
850 */
851 elapsed = TCP_RESOURCE_PROBE_INTERVAL;
852 }
853 } else {
854 /* It is tp->rcv_tstamp + keepalive_time_when(tp) */
855 elapsed = keepalive_time_when(tp) - elapsed;
856 }
857
858resched:
859 tcp_reset_keepalive_timer(sk, elapsed);
860 goto out;
861
862death:
863 tcp_done(sk);
864
865out:
866 bh_unlock_sock(sk);
867 sock_put(sk);
868}
869
870static enum hrtimer_restart tcp_compressed_ack_kick(struct hrtimer *timer)
871{
872 struct tcp_sock *tp = container_of(timer, struct tcp_sock, compressed_ack_timer);
873 struct sock *sk = (struct sock *)tp;
874
875 bh_lock_sock(sk);
876 if (!sock_owned_by_user(sk)) {
877 if (tp->compressed_ack) {
878 /* Since we have to send one ack finally,
879 * subtract one from tp->compressed_ack to keep
880 * LINUX_MIB_TCPACKCOMPRESSED accurate.
881 */
882 tp->compressed_ack--;
883 tcp_mstamp_refresh(tp);
884 tcp_send_ack(sk);
885 }
886 } else {
887 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED,
888 &sk->sk_tsq_flags))
889 sock_hold(sk);
890 }
891 bh_unlock_sock(sk);
892
893 sock_put(sk);
894
895 return HRTIMER_NORESTART;
896}
897
898void tcp_init_xmit_timers(struct sock *sk)
899{
900 inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
901 &tcp_keepalive_timer);
902 hrtimer_setup(&tcp_sk(sk)->pacing_timer, tcp_pace_kick, CLOCK_MONOTONIC,
903 HRTIMER_MODE_ABS_PINNED_SOFT);
904
905 hrtimer_setup(&tcp_sk(sk)->compressed_ack_timer, tcp_compressed_ack_kick, CLOCK_MONOTONIC,
906 HRTIMER_MODE_REL_PINNED_SOFT);
907}