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// 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
26static u32 tcp_clamp_rto_to_user_timeout(const struct sock *sk)
27{
28 struct inet_connection_sock *icsk = inet_csk(sk);
29 const struct tcp_sock *tp = tcp_sk(sk);
30 u32 elapsed, user_timeout;
31 s32 remaining;
32
33 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
34 if (!user_timeout)
35 return icsk->icsk_rto;
36
37 elapsed = tcp_time_stamp_ts(tp) - tp->retrans_stamp;
38 if (tp->tcp_usec_ts)
39 elapsed /= USEC_PER_MSEC;
40
41 remaining = user_timeout - elapsed;
42 if (remaining <= 0)
43 return 1; /* user timeout has passed; fire ASAP */
44
45 return min_t(u32, icsk->icsk_rto, msecs_to_jiffies(remaining));
46}
47
48u32 tcp_clamp_probe0_to_user_timeout(const struct sock *sk, u32 when)
49{
50 struct inet_connection_sock *icsk = inet_csk(sk);
51 u32 remaining, user_timeout;
52 s32 elapsed;
53
54 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
55 if (!user_timeout || !icsk->icsk_probes_tstamp)
56 return when;
57
58 elapsed = tcp_jiffies32 - icsk->icsk_probes_tstamp;
59 if (unlikely(elapsed < 0))
60 elapsed = 0;
61 remaining = msecs_to_jiffies(user_timeout) - elapsed;
62 remaining = max_t(u32, remaining, TCP_TIMEOUT_MIN);
63
64 return min_t(u32, remaining, when);
65}
66
67/**
68 * tcp_write_err() - close socket and save error info
69 * @sk: The socket the error has appeared on.
70 *
71 * Returns: Nothing (void)
72 */
73
74static void tcp_write_err(struct sock *sk)
75{
76 WRITE_ONCE(sk->sk_err, READ_ONCE(sk->sk_err_soft) ? : ETIMEDOUT);
77 sk_error_report(sk);
78
79 tcp_write_queue_purge(sk);
80 tcp_done(sk);
81 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT);
82}
83
84/**
85 * tcp_out_of_resources() - Close socket if out of resources
86 * @sk: pointer to current socket
87 * @do_reset: send a last packet with reset flag
88 *
89 * Do not allow orphaned sockets to eat all our resources.
90 * This is direct violation of TCP specs, but it is required
91 * to prevent DoS attacks. It is called when a retransmission timeout
92 * or zero probe timeout occurs on orphaned socket.
93 *
94 * Also close if our net namespace is exiting; in that case there is no
95 * hope of ever communicating again since all netns interfaces are already
96 * down (or about to be down), and we need to release our dst references,
97 * which have been moved to the netns loopback interface, so the namespace
98 * can finish exiting. This condition is only possible if we are a kernel
99 * socket, as those do not hold references to the namespace.
100 *
101 * Criteria is still not confirmed experimentally and may change.
102 * We kill the socket, if:
103 * 1. If number of orphaned sockets exceeds an administratively configured
104 * limit.
105 * 2. If we have strong memory pressure.
106 * 3. If our net namespace is exiting.
107 */
108static int tcp_out_of_resources(struct sock *sk, bool do_reset)
109{
110 struct tcp_sock *tp = tcp_sk(sk);
111 int shift = 0;
112
113 /* If peer does not open window for long time, or did not transmit
114 * anything for long time, penalize it. */
115 if ((s32)(tcp_jiffies32 - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset)
116 shift++;
117
118 /* If some dubious ICMP arrived, penalize even more. */
119 if (READ_ONCE(sk->sk_err_soft))
120 shift++;
121
122 if (tcp_check_oom(sk, shift)) {
123 /* Catch exceptional cases, when connection requires reset.
124 * 1. Last segment was sent recently. */
125 if ((s32)(tcp_jiffies32 - tp->lsndtime) <= TCP_TIMEWAIT_LEN ||
126 /* 2. Window is closed. */
127 (!tp->snd_wnd && !tp->packets_out))
128 do_reset = true;
129 if (do_reset)
130 tcp_send_active_reset(sk, GFP_ATOMIC);
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 / 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;
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;
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
288 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RTO_CB_FLAG))
289 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RTO_CB,
290 icsk->icsk_retransmits,
291 icsk->icsk_rto, (int)expired);
292
293 if (expired) {
294 /* Has it gone just too far? */
295 tcp_write_err(sk);
296 return 1;
297 }
298
299 if (sk_rethink_txhash(sk)) {
300 tp->timeout_rehash++;
301 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEOUTREHASH);
302 }
303
304 return 0;
305}
306
307/* Called with BH disabled */
308void tcp_delack_timer_handler(struct sock *sk)
309{
310 struct inet_connection_sock *icsk = inet_csk(sk);
311 struct tcp_sock *tp = tcp_sk(sk);
312
313 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
314 return;
315
316 /* Handling the sack compression case */
317 if (tp->compressed_ack) {
318 tcp_mstamp_refresh(tp);
319 tcp_sack_compress_send_ack(sk);
320 return;
321 }
322
323 if (!(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
324 return;
325
326 if (time_after(icsk->icsk_ack.timeout, jiffies)) {
327 sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
328 return;
329 }
330 icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
331
332 if (inet_csk_ack_scheduled(sk)) {
333 if (!inet_csk_in_pingpong_mode(sk)) {
334 /* Delayed ACK missed: inflate ATO. */
335 icsk->icsk_ack.ato = min_t(u32, icsk->icsk_ack.ato << 1, icsk->icsk_rto);
336 } else {
337 /* Delayed ACK missed: leave pingpong mode and
338 * deflate ATO.
339 */
340 inet_csk_exit_pingpong_mode(sk);
341 icsk->icsk_ack.ato = TCP_ATO_MIN;
342 }
343 tcp_mstamp_refresh(tp);
344 tcp_send_ack(sk);
345 __NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKS);
346 }
347}
348
349
350/**
351 * tcp_delack_timer() - The TCP delayed ACK timeout handler
352 * @t: Pointer to the timer. (gets casted to struct sock *)
353 *
354 * This function gets (indirectly) called when the kernel timer for a TCP packet
355 * of this socket expires. Calls tcp_delack_timer_handler() to do the actual work.
356 *
357 * Returns: Nothing (void)
358 */
359static void tcp_delack_timer(struct timer_list *t)
360{
361 struct inet_connection_sock *icsk =
362 from_timer(icsk, t, icsk_delack_timer);
363 struct sock *sk = &icsk->icsk_inet.sk;
364
365 bh_lock_sock(sk);
366 if (!sock_owned_by_user(sk)) {
367 tcp_delack_timer_handler(sk);
368 } else {
369 __NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
370 /* deleguate our work to tcp_release_cb() */
371 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &sk->sk_tsq_flags))
372 sock_hold(sk);
373 }
374 bh_unlock_sock(sk);
375 sock_put(sk);
376}
377
378static void tcp_probe_timer(struct sock *sk)
379{
380 struct inet_connection_sock *icsk = inet_csk(sk);
381 struct sk_buff *skb = tcp_send_head(sk);
382 struct tcp_sock *tp = tcp_sk(sk);
383 int max_probes;
384
385 if (tp->packets_out || !skb) {
386 icsk->icsk_probes_out = 0;
387 icsk->icsk_probes_tstamp = 0;
388 return;
389 }
390
391 /* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as
392 * long as the receiver continues to respond probes. We support this by
393 * default and reset icsk_probes_out with incoming ACKs. But if the
394 * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we
395 * kill the socket when the retry count and the time exceeds the
396 * corresponding system limit. We also implement similar policy when
397 * we use RTO to probe window in tcp_retransmit_timer().
398 */
399 if (!icsk->icsk_probes_tstamp) {
400 icsk->icsk_probes_tstamp = tcp_jiffies32;
401 } else {
402 u32 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
403
404 if (user_timeout &&
405 (s32)(tcp_jiffies32 - icsk->icsk_probes_tstamp) >=
406 msecs_to_jiffies(user_timeout))
407 goto abort;
408 }
409 max_probes = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retries2);
410 if (sock_flag(sk, SOCK_DEAD)) {
411 const bool alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX;
412
413 max_probes = tcp_orphan_retries(sk, alive);
414 if (!alive && icsk->icsk_backoff >= max_probes)
415 goto abort;
416 if (tcp_out_of_resources(sk, true))
417 return;
418 }
419
420 if (icsk->icsk_probes_out >= max_probes) {
421abort: tcp_write_err(sk);
422 } else {
423 /* Only send another probe if we didn't close things up. */
424 tcp_send_probe0(sk);
425 }
426}
427
428static void tcp_update_rto_stats(struct sock *sk)
429{
430 struct inet_connection_sock *icsk = inet_csk(sk);
431 struct tcp_sock *tp = tcp_sk(sk);
432
433 if (!icsk->icsk_retransmits) {
434 tp->total_rto_recoveries++;
435 tp->rto_stamp = tcp_time_stamp_ms(tp);
436 }
437 icsk->icsk_retransmits++;
438 tp->total_rto++;
439}
440
441/*
442 * Timer for Fast Open socket to retransmit SYNACK. Note that the
443 * sk here is the child socket, not the parent (listener) socket.
444 */
445static void tcp_fastopen_synack_timer(struct sock *sk, struct request_sock *req)
446{
447 struct inet_connection_sock *icsk = inet_csk(sk);
448 struct tcp_sock *tp = tcp_sk(sk);
449 int max_retries;
450
451 req->rsk_ops->syn_ack_timeout(req);
452
453 /* Add one more retry for fastopen.
454 * Paired with WRITE_ONCE() in tcp_sock_set_syncnt()
455 */
456 max_retries = READ_ONCE(icsk->icsk_syn_retries) ? :
457 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_synack_retries) + 1;
458
459 if (req->num_timeout >= max_retries) {
460 tcp_write_err(sk);
461 return;
462 }
463 /* Lower cwnd after certain SYNACK timeout like tcp_init_transfer() */
464 if (icsk->icsk_retransmits == 1)
465 tcp_enter_loss(sk);
466 /* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
467 * returned from rtx_syn_ack() to make it more persistent like
468 * regular retransmit because if the child socket has been accepted
469 * it's not good to give up too easily.
470 */
471 inet_rtx_syn_ack(sk, req);
472 req->num_timeout++;
473 tcp_update_rto_stats(sk);
474 if (!tp->retrans_stamp)
475 tp->retrans_stamp = tcp_time_stamp_ts(tp);
476 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
477 req->timeout << req->num_timeout, TCP_RTO_MAX);
478}
479
480static bool tcp_rtx_probe0_timed_out(const struct sock *sk,
481 const struct sk_buff *skb,
482 u32 rtx_delta)
483{
484 const struct tcp_sock *tp = tcp_sk(sk);
485 const int timeout = TCP_RTO_MAX * 2;
486 u32 rcv_delta;
487
488 rcv_delta = inet_csk(sk)->icsk_timeout - tp->rcv_tstamp;
489 if (rcv_delta <= timeout)
490 return false;
491
492 return msecs_to_jiffies(rtx_delta) > timeout;
493}
494
495/**
496 * tcp_retransmit_timer() - The TCP retransmit timeout handler
497 * @sk: Pointer to the current socket.
498 *
499 * This function gets called when the kernel timer for a TCP packet
500 * of this socket expires.
501 *
502 * It handles retransmission, timer adjustment and other necessary measures.
503 *
504 * Returns: Nothing (void)
505 */
506void tcp_retransmit_timer(struct sock *sk)
507{
508 struct tcp_sock *tp = tcp_sk(sk);
509 struct net *net = sock_net(sk);
510 struct inet_connection_sock *icsk = inet_csk(sk);
511 struct request_sock *req;
512 struct sk_buff *skb;
513
514 req = rcu_dereference_protected(tp->fastopen_rsk,
515 lockdep_sock_is_held(sk));
516 if (req) {
517 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
518 sk->sk_state != TCP_FIN_WAIT1);
519 tcp_fastopen_synack_timer(sk, req);
520 /* Before we receive ACK to our SYN-ACK don't retransmit
521 * anything else (e.g., data or FIN segments).
522 */
523 return;
524 }
525
526 if (!tp->packets_out)
527 return;
528
529 skb = tcp_rtx_queue_head(sk);
530 if (WARN_ON_ONCE(!skb))
531 return;
532
533 tp->tlp_high_seq = 0;
534
535 if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
536 !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
537 /* Receiver dastardly shrinks window. Our retransmits
538 * become zero probes, but we should not timeout this
539 * connection. If the socket is an orphan, time it out,
540 * we cannot allow such beasts to hang infinitely.
541 */
542 struct inet_sock *inet = inet_sk(sk);
543 u32 rtx_delta;
544
545 rtx_delta = tcp_time_stamp_ts(tp) - (tp->retrans_stamp ?:
546 tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb));
547 if (tp->tcp_usec_ts)
548 rtx_delta /= USEC_PER_MSEC;
549
550 if (sk->sk_family == AF_INET) {
551 net_dbg_ratelimited("Probing zero-window on %pI4:%u/%u, seq=%u:%u, recv %ums ago, lasting %ums\n",
552 &inet->inet_daddr, ntohs(inet->inet_dport),
553 inet->inet_num, tp->snd_una, tp->snd_nxt,
554 jiffies_to_msecs(jiffies - tp->rcv_tstamp),
555 rtx_delta);
556 }
557#if IS_ENABLED(CONFIG_IPV6)
558 else if (sk->sk_family == AF_INET6) {
559 net_dbg_ratelimited("Probing zero-window on %pI6:%u/%u, seq=%u:%u, recv %ums ago, lasting %ums\n",
560 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
561 inet->inet_num, tp->snd_una, tp->snd_nxt,
562 jiffies_to_msecs(jiffies - tp->rcv_tstamp),
563 rtx_delta);
564 }
565#endif
566 if (tcp_rtx_probe0_timed_out(sk, skb, rtx_delta)) {
567 tcp_write_err(sk);
568 goto out;
569 }
570 tcp_enter_loss(sk);
571 tcp_retransmit_skb(sk, skb, 1);
572 __sk_dst_reset(sk);
573 goto out_reset_timer;
574 }
575
576 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEOUTS);
577 if (tcp_write_timeout(sk))
578 goto out;
579
580 if (icsk->icsk_retransmits == 0) {
581 int mib_idx = 0;
582
583 if (icsk->icsk_ca_state == TCP_CA_Recovery) {
584 if (tcp_is_sack(tp))
585 mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
586 else
587 mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
588 } else if (icsk->icsk_ca_state == TCP_CA_Loss) {
589 mib_idx = LINUX_MIB_TCPLOSSFAILURES;
590 } else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
591 tp->sacked_out) {
592 if (tcp_is_sack(tp))
593 mib_idx = LINUX_MIB_TCPSACKFAILURES;
594 else
595 mib_idx = LINUX_MIB_TCPRENOFAILURES;
596 }
597 if (mib_idx)
598 __NET_INC_STATS(sock_net(sk), mib_idx);
599 }
600
601 tcp_enter_loss(sk);
602
603 tcp_update_rto_stats(sk);
604 if (tcp_retransmit_skb(sk, tcp_rtx_queue_head(sk), 1) > 0) {
605 /* Retransmission failed because of local congestion,
606 * Let senders fight for local resources conservatively.
607 */
608 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
609 TCP_RESOURCE_PROBE_INTERVAL,
610 TCP_RTO_MAX);
611 goto out;
612 }
613
614 /* Increase the timeout each time we retransmit. Note that
615 * we do not increase the rtt estimate. rto is initialized
616 * from rtt, but increases here. Jacobson (SIGCOMM 88) suggests
617 * that doubling rto each time is the least we can get away with.
618 * In KA9Q, Karn uses this for the first few times, and then
619 * goes to quadratic. netBSD doubles, but only goes up to *64,
620 * and clamps at 1 to 64 sec afterwards. Note that 120 sec is
621 * defined in the protocol as the maximum possible RTT. I guess
622 * we'll have to use something other than TCP to talk to the
623 * University of Mars.
624 *
625 * PAWS allows us longer timeouts and large windows, so once
626 * implemented ftp to mars will work nicely. We will have to fix
627 * the 120 second clamps though!
628 */
629
630out_reset_timer:
631 /* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
632 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
633 * might be increased if the stream oscillates between thin and thick,
634 * thus the old value might already be too high compared to the value
635 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
636 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
637 * exponential backoff behaviour to avoid continue hammering
638 * linear-timeout retransmissions into a black hole
639 */
640 if (sk->sk_state == TCP_ESTABLISHED &&
641 (tp->thin_lto || READ_ONCE(net->ipv4.sysctl_tcp_thin_linear_timeouts)) &&
642 tcp_stream_is_thin(tp) &&
643 icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
644 icsk->icsk_backoff = 0;
645 icsk->icsk_rto = clamp(__tcp_set_rto(tp),
646 tcp_rto_min(sk),
647 TCP_RTO_MAX);
648 } else if (sk->sk_state != TCP_SYN_SENT ||
649 tp->total_rto >
650 READ_ONCE(net->ipv4.sysctl_tcp_syn_linear_timeouts)) {
651 /* Use normal (exponential) backoff unless linear timeouts are
652 * activated.
653 */
654 icsk->icsk_backoff++;
655 icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
656 }
657 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
658 tcp_clamp_rto_to_user_timeout(sk), TCP_RTO_MAX);
659 if (retransmits_timed_out(sk, READ_ONCE(net->ipv4.sysctl_tcp_retries1) + 1, 0))
660 __sk_dst_reset(sk);
661
662out:;
663}
664
665/* Called with bottom-half processing disabled.
666 Called by tcp_write_timer() */
667void tcp_write_timer_handler(struct sock *sk)
668{
669 struct inet_connection_sock *icsk = inet_csk(sk);
670 int event;
671
672 if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
673 !icsk->icsk_pending)
674 return;
675
676 if (time_after(icsk->icsk_timeout, jiffies)) {
677 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
678 return;
679 }
680
681 tcp_mstamp_refresh(tcp_sk(sk));
682 event = icsk->icsk_pending;
683
684 switch (event) {
685 case ICSK_TIME_REO_TIMEOUT:
686 tcp_rack_reo_timeout(sk);
687 break;
688 case ICSK_TIME_LOSS_PROBE:
689 tcp_send_loss_probe(sk);
690 break;
691 case ICSK_TIME_RETRANS:
692 icsk->icsk_pending = 0;
693 tcp_retransmit_timer(sk);
694 break;
695 case ICSK_TIME_PROBE0:
696 icsk->icsk_pending = 0;
697 tcp_probe_timer(sk);
698 break;
699 }
700}
701
702static void tcp_write_timer(struct timer_list *t)
703{
704 struct inet_connection_sock *icsk =
705 from_timer(icsk, t, icsk_retransmit_timer);
706 struct sock *sk = &icsk->icsk_inet.sk;
707
708 bh_lock_sock(sk);
709 if (!sock_owned_by_user(sk)) {
710 tcp_write_timer_handler(sk);
711 } else {
712 /* delegate our work to tcp_release_cb() */
713 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &sk->sk_tsq_flags))
714 sock_hold(sk);
715 }
716 bh_unlock_sock(sk);
717 sock_put(sk);
718}
719
720void tcp_syn_ack_timeout(const struct request_sock *req)
721{
722 struct net *net = read_pnet(&inet_rsk(req)->ireq_net);
723
724 __NET_INC_STATS(net, LINUX_MIB_TCPTIMEOUTS);
725}
726EXPORT_SYMBOL(tcp_syn_ack_timeout);
727
728void tcp_set_keepalive(struct sock *sk, int val)
729{
730 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
731 return;
732
733 if (val && !sock_flag(sk, SOCK_KEEPOPEN))
734 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
735 else if (!val)
736 inet_csk_delete_keepalive_timer(sk);
737}
738EXPORT_SYMBOL_GPL(tcp_set_keepalive);
739
740
741static void tcp_keepalive_timer (struct timer_list *t)
742{
743 struct sock *sk = from_timer(sk, t, sk_timer);
744 struct inet_connection_sock *icsk = inet_csk(sk);
745 struct tcp_sock *tp = tcp_sk(sk);
746 u32 elapsed;
747
748 /* Only process if socket is not in use. */
749 bh_lock_sock(sk);
750 if (sock_owned_by_user(sk)) {
751 /* Try again later. */
752 inet_csk_reset_keepalive_timer (sk, HZ/20);
753 goto out;
754 }
755
756 if (sk->sk_state == TCP_LISTEN) {
757 pr_err("Hmm... keepalive on a LISTEN ???\n");
758 goto out;
759 }
760
761 tcp_mstamp_refresh(tp);
762 if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
763 if (READ_ONCE(tp->linger2) >= 0) {
764 const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
765
766 if (tmo > 0) {
767 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
768 goto out;
769 }
770 }
771 tcp_send_active_reset(sk, GFP_ATOMIC);
772 goto death;
773 }
774
775 if (!sock_flag(sk, SOCK_KEEPOPEN) ||
776 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_SYN_SENT)))
777 goto out;
778
779 elapsed = keepalive_time_when(tp);
780
781 /* It is alive without keepalive 8) */
782 if (tp->packets_out || !tcp_write_queue_empty(sk))
783 goto resched;
784
785 elapsed = keepalive_time_elapsed(tp);
786
787 if (elapsed >= keepalive_time_when(tp)) {
788 u32 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
789
790 /* If the TCP_USER_TIMEOUT option is enabled, use that
791 * to determine when to timeout instead.
792 */
793 if ((user_timeout != 0 &&
794 elapsed >= msecs_to_jiffies(user_timeout) &&
795 icsk->icsk_probes_out > 0) ||
796 (user_timeout == 0 &&
797 icsk->icsk_probes_out >= keepalive_probes(tp))) {
798 tcp_send_active_reset(sk, GFP_ATOMIC);
799 tcp_write_err(sk);
800 goto out;
801 }
802 if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) {
803 icsk->icsk_probes_out++;
804 elapsed = keepalive_intvl_when(tp);
805 } else {
806 /* If keepalive was lost due to local congestion,
807 * try harder.
808 */
809 elapsed = TCP_RESOURCE_PROBE_INTERVAL;
810 }
811 } else {
812 /* It is tp->rcv_tstamp + keepalive_time_when(tp) */
813 elapsed = keepalive_time_when(tp) - elapsed;
814 }
815
816resched:
817 inet_csk_reset_keepalive_timer (sk, elapsed);
818 goto out;
819
820death:
821 tcp_done(sk);
822
823out:
824 bh_unlock_sock(sk);
825 sock_put(sk);
826}
827
828static enum hrtimer_restart tcp_compressed_ack_kick(struct hrtimer *timer)
829{
830 struct tcp_sock *tp = container_of(timer, struct tcp_sock, compressed_ack_timer);
831 struct sock *sk = (struct sock *)tp;
832
833 bh_lock_sock(sk);
834 if (!sock_owned_by_user(sk)) {
835 if (tp->compressed_ack) {
836 /* Since we have to send one ack finally,
837 * subtract one from tp->compressed_ack to keep
838 * LINUX_MIB_TCPACKCOMPRESSED accurate.
839 */
840 tp->compressed_ack--;
841 tcp_send_ack(sk);
842 }
843 } else {
844 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED,
845 &sk->sk_tsq_flags))
846 sock_hold(sk);
847 }
848 bh_unlock_sock(sk);
849
850 sock_put(sk);
851
852 return HRTIMER_NORESTART;
853}
854
855void tcp_init_xmit_timers(struct sock *sk)
856{
857 inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
858 &tcp_keepalive_timer);
859 hrtimer_init(&tcp_sk(sk)->pacing_timer, CLOCK_MONOTONIC,
860 HRTIMER_MODE_ABS_PINNED_SOFT);
861 tcp_sk(sk)->pacing_timer.function = tcp_pace_kick;
862
863 hrtimer_init(&tcp_sk(sk)->compressed_ack_timer, CLOCK_MONOTONIC,
864 HRTIMER_MODE_REL_PINNED_SOFT);
865 tcp_sk(sk)->compressed_ack_timer.function = tcp_compressed_ack_kick;
866}