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.
6 * Implementation of the Transmission Control Protocol(TCP).
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>
21 #include <linux/module.h>
22 #include <linux/gfp.h>
25 int sysctl_tcp_syn_retries __read_mostly = TCP_SYN_RETRIES;
26 int sysctl_tcp_synack_retries __read_mostly = TCP_SYNACK_RETRIES;
27 int sysctl_tcp_keepalive_time __read_mostly = TCP_KEEPALIVE_TIME;
28 int sysctl_tcp_keepalive_probes __read_mostly = TCP_KEEPALIVE_PROBES;
29 int sysctl_tcp_keepalive_intvl __read_mostly = TCP_KEEPALIVE_INTVL;
30 int sysctl_tcp_retries1 __read_mostly = TCP_RETR1;
31 int sysctl_tcp_retries2 __read_mostly = TCP_RETR2;
32 int sysctl_tcp_orphan_retries __read_mostly;
33 int sysctl_tcp_thin_linear_timeouts __read_mostly;
35 static void tcp_write_err(struct sock *sk)
37 sk->sk_err = sk->sk_err_soft ? : ETIMEDOUT;
38 sk->sk_error_report(sk);
41 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT);
44 /* Do not allow orphaned sockets to eat all our resources.
45 * This is direct violation of TCP specs, but it is required
46 * to prevent DoS attacks. It is called when a retransmission timeout
47 * or zero probe timeout occurs on orphaned socket.
49 * Criteria is still not confirmed experimentally and may change.
50 * We kill the socket, if:
51 * 1. If number of orphaned sockets exceeds an administratively configured
53 * 2. If we have strong memory pressure.
55 static int tcp_out_of_resources(struct sock *sk, bool do_reset)
57 struct tcp_sock *tp = tcp_sk(sk);
60 /* If peer does not open window for long time, or did not transmit
61 * anything for long time, penalize it. */
62 if ((s32)(tcp_time_stamp - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset)
65 /* If some dubious ICMP arrived, penalize even more. */
69 if (tcp_check_oom(sk, shift)) {
70 /* Catch exceptional cases, when connection requires reset.
71 * 1. Last segment was sent recently. */
72 if ((s32)(tcp_time_stamp - tp->lsndtime) <= TCP_TIMEWAIT_LEN ||
73 /* 2. Window is closed. */
74 (!tp->snd_wnd && !tp->packets_out))
77 tcp_send_active_reset(sk, GFP_ATOMIC);
79 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY);
85 /* Calculate maximal number or retries on an orphaned socket. */
86 static int tcp_orphan_retries(struct sock *sk, int alive)
88 int retries = sysctl_tcp_orphan_retries; /* May be zero. */
90 /* We know from an ICMP that something is wrong. */
91 if (sk->sk_err_soft && !alive)
94 /* However, if socket sent something recently, select some safe
95 * number of retries. 8 corresponds to >100 seconds with minimal
97 if (retries == 0 && alive)
102 static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk)
104 struct net *net = sock_net(sk);
106 /* Black hole detection */
107 if (net->ipv4.sysctl_tcp_mtu_probing) {
108 if (!icsk->icsk_mtup.enabled) {
109 icsk->icsk_mtup.enabled = 1;
110 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
111 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
113 struct net *net = sock_net(sk);
114 struct tcp_sock *tp = tcp_sk(sk);
117 mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1;
118 mss = min(net->ipv4.sysctl_tcp_base_mss, mss);
119 mss = max(mss, 68 - tp->tcp_header_len);
120 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
121 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
126 /* This function calculates a "timeout" which is equivalent to the timeout of a
127 * TCP connection after "boundary" unsuccessful, exponentially backed-off
128 * retransmissions with an initial RTO of TCP_RTO_MIN or TCP_TIMEOUT_INIT if
129 * syn_set flag is set.
131 static bool retransmits_timed_out(struct sock *sk,
132 unsigned int boundary,
133 unsigned int timeout,
136 unsigned int linear_backoff_thresh, start_ts;
137 unsigned int rto_base = syn_set ? TCP_TIMEOUT_INIT : TCP_RTO_MIN;
139 if (!inet_csk(sk)->icsk_retransmits)
142 start_ts = tcp_sk(sk)->retrans_stamp;
143 if (unlikely(!start_ts))
144 start_ts = tcp_skb_timestamp(tcp_write_queue_head(sk));
146 if (likely(timeout == 0)) {
147 linear_backoff_thresh = ilog2(TCP_RTO_MAX/rto_base);
149 if (boundary <= linear_backoff_thresh)
150 timeout = ((2 << boundary) - 1) * rto_base;
152 timeout = ((2 << linear_backoff_thresh) - 1) * rto_base +
153 (boundary - linear_backoff_thresh) * TCP_RTO_MAX;
155 return (tcp_time_stamp - start_ts) >= timeout;
158 /* A write timeout has occurred. Process the after effects. */
159 static int tcp_write_timeout(struct sock *sk)
161 struct inet_connection_sock *icsk = inet_csk(sk);
162 struct tcp_sock *tp = tcp_sk(sk);
164 bool do_reset, syn_set = false;
166 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
167 if (icsk->icsk_retransmits) {
168 dst_negative_advice(sk);
169 if (tp->syn_fastopen || tp->syn_data)
170 tcp_fastopen_cache_set(sk, 0, NULL, true);
172 NET_INC_STATS_BH(sock_net(sk),
173 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
175 retry_until = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
178 if (retransmits_timed_out(sk, sysctl_tcp_retries1, 0, 0)) {
179 /* Black hole detection */
180 tcp_mtu_probing(icsk, sk);
182 dst_negative_advice(sk);
185 retry_until = sysctl_tcp_retries2;
186 if (sock_flag(sk, SOCK_DEAD)) {
187 const int alive = icsk->icsk_rto < TCP_RTO_MAX;
189 retry_until = tcp_orphan_retries(sk, alive);
191 !retransmits_timed_out(sk, retry_until, 0, 0);
193 if (tcp_out_of_resources(sk, do_reset))
198 if (retransmits_timed_out(sk, retry_until,
199 syn_set ? 0 : icsk->icsk_user_timeout, syn_set)) {
200 /* Has it gone just too far? */
207 void tcp_delack_timer_handler(struct sock *sk)
209 struct tcp_sock *tp = tcp_sk(sk);
210 struct inet_connection_sock *icsk = inet_csk(sk);
212 sk_mem_reclaim_partial(sk);
214 if (sk->sk_state == TCP_CLOSE || !(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
217 if (time_after(icsk->icsk_ack.timeout, jiffies)) {
218 sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
221 icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
223 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
226 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSCHEDULERFAILED);
228 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
229 sk_backlog_rcv(sk, skb);
231 tp->ucopy.memory = 0;
234 if (inet_csk_ack_scheduled(sk)) {
235 if (!icsk->icsk_ack.pingpong) {
236 /* Delayed ACK missed: inflate ATO. */
237 icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto);
239 /* Delayed ACK missed: leave pingpong mode and
242 icsk->icsk_ack.pingpong = 0;
243 icsk->icsk_ack.ato = TCP_ATO_MIN;
246 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKS);
250 if (sk_under_memory_pressure(sk))
254 static void tcp_delack_timer(unsigned long data)
256 struct sock *sk = (struct sock *)data;
259 if (!sock_owned_by_user(sk)) {
260 tcp_delack_timer_handler(sk);
262 inet_csk(sk)->icsk_ack.blocked = 1;
263 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
264 /* deleguate our work to tcp_release_cb() */
265 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &tcp_sk(sk)->tsq_flags))
272 static void tcp_probe_timer(struct sock *sk)
274 struct inet_connection_sock *icsk = inet_csk(sk);
275 struct tcp_sock *tp = tcp_sk(sk);
279 if (tp->packets_out || !tcp_send_head(sk)) {
280 icsk->icsk_probes_out = 0;
284 /* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as
285 * long as the receiver continues to respond probes. We support this by
286 * default and reset icsk_probes_out with incoming ACKs. But if the
287 * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we
288 * kill the socket when the retry count and the time exceeds the
289 * corresponding system limit. We also implement similar policy when
290 * we use RTO to probe window in tcp_retransmit_timer().
292 start_ts = tcp_skb_timestamp(tcp_send_head(sk));
294 skb_mstamp_get(&tcp_send_head(sk)->skb_mstamp);
295 else if (icsk->icsk_user_timeout &&
296 (s32)(tcp_time_stamp - start_ts) > icsk->icsk_user_timeout)
299 max_probes = sysctl_tcp_retries2;
300 if (sock_flag(sk, SOCK_DEAD)) {
301 const int alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX;
303 max_probes = tcp_orphan_retries(sk, alive);
304 if (!alive && icsk->icsk_backoff >= max_probes)
306 if (tcp_out_of_resources(sk, true))
310 if (icsk->icsk_probes_out > max_probes) {
311 abort: tcp_write_err(sk);
313 /* Only send another probe if we didn't close things up. */
319 * Timer for Fast Open socket to retransmit SYNACK. Note that the
320 * sk here is the child socket, not the parent (listener) socket.
322 static void tcp_fastopen_synack_timer(struct sock *sk)
324 struct inet_connection_sock *icsk = inet_csk(sk);
325 int max_retries = icsk->icsk_syn_retries ? :
326 sysctl_tcp_synack_retries + 1; /* add one more retry for fastopen */
327 struct request_sock *req;
329 req = tcp_sk(sk)->fastopen_rsk;
330 req->rsk_ops->syn_ack_timeout(sk, req);
332 if (req->num_timeout >= max_retries) {
336 /* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
337 * returned from rtx_syn_ack() to make it more persistent like
338 * regular retransmit because if the child socket has been accepted
339 * it's not good to give up too easily.
341 inet_rtx_syn_ack(sk, req);
343 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
344 TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
348 * The TCP retransmit timer.
351 void tcp_retransmit_timer(struct sock *sk)
353 struct tcp_sock *tp = tcp_sk(sk);
354 struct inet_connection_sock *icsk = inet_csk(sk);
356 if (tp->fastopen_rsk) {
357 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
358 sk->sk_state != TCP_FIN_WAIT1);
359 tcp_fastopen_synack_timer(sk);
360 /* Before we receive ACK to our SYN-ACK don't retransmit
361 * anything else (e.g., data or FIN segments).
365 if (!tp->packets_out)
368 WARN_ON(tcp_write_queue_empty(sk));
370 tp->tlp_high_seq = 0;
372 if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
373 !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
374 /* Receiver dastardly shrinks window. Our retransmits
375 * become zero probes, but we should not timeout this
376 * connection. If the socket is an orphan, time it out,
377 * we cannot allow such beasts to hang infinitely.
379 struct inet_sock *inet = inet_sk(sk);
380 if (sk->sk_family == AF_INET) {
381 net_dbg_ratelimited("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
383 ntohs(inet->inet_dport),
385 tp->snd_una, tp->snd_nxt);
387 #if IS_ENABLED(CONFIG_IPV6)
388 else if (sk->sk_family == AF_INET6) {
389 net_dbg_ratelimited("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
391 ntohs(inet->inet_dport),
393 tp->snd_una, tp->snd_nxt);
396 if (tcp_time_stamp - tp->rcv_tstamp > TCP_RTO_MAX) {
401 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
403 goto out_reset_timer;
406 if (tcp_write_timeout(sk))
409 if (icsk->icsk_retransmits == 0) {
412 if (icsk->icsk_ca_state == TCP_CA_Recovery) {
414 mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
416 mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
417 } else if (icsk->icsk_ca_state == TCP_CA_Loss) {
418 mib_idx = LINUX_MIB_TCPLOSSFAILURES;
419 } else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
422 mib_idx = LINUX_MIB_TCPSACKFAILURES;
424 mib_idx = LINUX_MIB_TCPRENOFAILURES;
426 mib_idx = LINUX_MIB_TCPTIMEOUTS;
428 NET_INC_STATS_BH(sock_net(sk), mib_idx);
433 if (tcp_retransmit_skb(sk, tcp_write_queue_head(sk)) > 0) {
434 /* Retransmission failed because of local congestion,
437 if (!icsk->icsk_retransmits)
438 icsk->icsk_retransmits = 1;
439 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
440 min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL),
445 /* Increase the timeout each time we retransmit. Note that
446 * we do not increase the rtt estimate. rto is initialized
447 * from rtt, but increases here. Jacobson (SIGCOMM 88) suggests
448 * that doubling rto each time is the least we can get away with.
449 * In KA9Q, Karn uses this for the first few times, and then
450 * goes to quadratic. netBSD doubles, but only goes up to *64,
451 * and clamps at 1 to 64 sec afterwards. Note that 120 sec is
452 * defined in the protocol as the maximum possible RTT. I guess
453 * we'll have to use something other than TCP to talk to the
454 * University of Mars.
456 * PAWS allows us longer timeouts and large windows, so once
457 * implemented ftp to mars will work nicely. We will have to fix
458 * the 120 second clamps though!
460 icsk->icsk_backoff++;
461 icsk->icsk_retransmits++;
464 /* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
465 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
466 * might be increased if the stream oscillates between thin and thick,
467 * thus the old value might already be too high compared to the value
468 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
469 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
470 * exponential backoff behaviour to avoid continue hammering
471 * linear-timeout retransmissions into a black hole
473 if (sk->sk_state == TCP_ESTABLISHED &&
474 (tp->thin_lto || sysctl_tcp_thin_linear_timeouts) &&
475 tcp_stream_is_thin(tp) &&
476 icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
477 icsk->icsk_backoff = 0;
478 icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX);
480 /* Use normal (exponential) backoff */
481 icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
483 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX);
484 if (retransmits_timed_out(sk, sysctl_tcp_retries1 + 1, 0, 0))
490 void tcp_write_timer_handler(struct sock *sk)
492 struct inet_connection_sock *icsk = inet_csk(sk);
495 if (sk->sk_state == TCP_CLOSE || !icsk->icsk_pending)
498 if (time_after(icsk->icsk_timeout, jiffies)) {
499 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
503 event = icsk->icsk_pending;
506 case ICSK_TIME_EARLY_RETRANS:
507 tcp_resume_early_retransmit(sk);
509 case ICSK_TIME_LOSS_PROBE:
510 tcp_send_loss_probe(sk);
512 case ICSK_TIME_RETRANS:
513 icsk->icsk_pending = 0;
514 tcp_retransmit_timer(sk);
516 case ICSK_TIME_PROBE0:
517 icsk->icsk_pending = 0;
526 static void tcp_write_timer(unsigned long data)
528 struct sock *sk = (struct sock *)data;
531 if (!sock_owned_by_user(sk)) {
532 tcp_write_timer_handler(sk);
534 /* deleguate our work to tcp_release_cb() */
535 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &tcp_sk(sk)->tsq_flags))
542 void tcp_syn_ack_timeout(struct sock *sk, struct request_sock *req)
544 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPTIMEOUTS);
546 EXPORT_SYMBOL(tcp_syn_ack_timeout);
548 void tcp_set_keepalive(struct sock *sk, int val)
550 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
553 if (val && !sock_flag(sk, SOCK_KEEPOPEN))
554 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
556 inet_csk_delete_keepalive_timer(sk);
560 static void tcp_keepalive_timer (unsigned long data)
562 struct sock *sk = (struct sock *) data;
563 struct inet_connection_sock *icsk = inet_csk(sk);
564 struct tcp_sock *tp = tcp_sk(sk);
567 /* Only process if socket is not in use. */
569 if (sock_owned_by_user(sk)) {
570 /* Try again later. */
571 inet_csk_reset_keepalive_timer (sk, HZ/20);
575 if (sk->sk_state == TCP_LISTEN) {
576 pr_err("Hmm... keepalive on a LISTEN ???\n");
580 if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
581 if (tp->linger2 >= 0) {
582 const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
585 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
589 tcp_send_active_reset(sk, GFP_ATOMIC);
593 if (!sock_flag(sk, SOCK_KEEPOPEN) || sk->sk_state == TCP_CLOSE)
596 elapsed = keepalive_time_when(tp);
598 /* It is alive without keepalive 8) */
599 if (tp->packets_out || tcp_send_head(sk))
602 elapsed = keepalive_time_elapsed(tp);
604 if (elapsed >= keepalive_time_when(tp)) {
605 /* If the TCP_USER_TIMEOUT option is enabled, use that
606 * to determine when to timeout instead.
608 if ((icsk->icsk_user_timeout != 0 &&
609 elapsed >= icsk->icsk_user_timeout &&
610 icsk->icsk_probes_out > 0) ||
611 (icsk->icsk_user_timeout == 0 &&
612 icsk->icsk_probes_out >= keepalive_probes(tp))) {
613 tcp_send_active_reset(sk, GFP_ATOMIC);
617 if (tcp_write_wakeup(sk) <= 0) {
618 icsk->icsk_probes_out++;
619 elapsed = keepalive_intvl_when(tp);
621 /* If keepalive was lost due to local congestion,
624 elapsed = TCP_RESOURCE_PROBE_INTERVAL;
627 /* It is tp->rcv_tstamp + keepalive_time_when(tp) */
628 elapsed = keepalive_time_when(tp) - elapsed;
634 inet_csk_reset_keepalive_timer (sk, elapsed);
645 void tcp_init_xmit_timers(struct sock *sk)
647 inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
648 &tcp_keepalive_timer);
650 EXPORT_SYMBOL(tcp_init_xmit_timers);