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 * Also close if our net namespace is exiting; in that case there is no
50 * hope of ever communicating again since all netns interfaces are already
51 * down (or about to be down), and we need to release our dst references,
52 * which have been moved to the netns loopback interface, so the namespace
53 * can finish exiting. This condition is only possible if we are a kernel
54 * socket, as those do not hold references to the namespace.
56 * Criteria is still not confirmed experimentally and may change.
57 * We kill the socket, if:
58 * 1. If number of orphaned sockets exceeds an administratively configured
60 * 2. If we have strong memory pressure.
61 * 3. If our net namespace is exiting.
63 static int tcp_out_of_resources(struct sock *sk, bool do_reset)
65 struct tcp_sock *tp = tcp_sk(sk);
68 /* If peer does not open window for long time, or did not transmit
69 * anything for long time, penalize it. */
70 if ((s32)(tcp_time_stamp - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset)
73 /* If some dubious ICMP arrived, penalize even more. */
77 if (tcp_check_oom(sk, shift)) {
78 /* Catch exceptional cases, when connection requires reset.
79 * 1. Last segment was sent recently. */
80 if ((s32)(tcp_time_stamp - tp->lsndtime) <= TCP_TIMEWAIT_LEN ||
81 /* 2. Window is closed. */
82 (!tp->snd_wnd && !tp->packets_out))
85 tcp_send_active_reset(sk, GFP_ATOMIC);
87 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY);
91 if (!check_net(sock_net(sk))) {
92 /* Not possible to send reset; just close */
100 /* Calculate maximal number or retries on an orphaned socket. */
101 static int tcp_orphan_retries(struct sock *sk, bool alive)
103 int retries = sysctl_tcp_orphan_retries; /* May be zero. */
105 /* We know from an ICMP that something is wrong. */
106 if (sk->sk_err_soft && !alive)
109 /* However, if socket sent something recently, select some safe
110 * number of retries. 8 corresponds to >100 seconds with minimal
112 if (retries == 0 && alive)
117 static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk)
119 struct net *net = sock_net(sk);
121 /* Black hole detection */
122 if (net->ipv4.sysctl_tcp_mtu_probing) {
123 if (!icsk->icsk_mtup.enabled) {
124 icsk->icsk_mtup.enabled = 1;
125 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
126 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
128 struct net *net = sock_net(sk);
129 struct tcp_sock *tp = tcp_sk(sk);
132 mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1;
133 mss = min(net->ipv4.sysctl_tcp_base_mss, mss);
134 mss = max(mss, 68 - tp->tcp_header_len);
135 mss = max(mss, net->ipv4.sysctl_tcp_min_snd_mss);
136 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
137 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
142 /* This function calculates a "timeout" which is equivalent to the timeout of a
143 * TCP connection after "boundary" unsuccessful, exponentially backed-off
144 * retransmissions with an initial RTO of TCP_RTO_MIN or TCP_TIMEOUT_INIT if
145 * syn_set flag is set.
147 static bool retransmits_timed_out(struct sock *sk,
148 unsigned int boundary,
149 unsigned int timeout,
152 unsigned int linear_backoff_thresh, start_ts;
153 unsigned int rto_base = syn_set ? TCP_TIMEOUT_INIT : TCP_RTO_MIN;
155 if (!inet_csk(sk)->icsk_retransmits)
158 start_ts = tcp_sk(sk)->retrans_stamp;
159 if (unlikely(!start_ts))
160 start_ts = tcp_skb_timestamp(tcp_write_queue_head(sk));
162 if (likely(timeout == 0)) {
163 linear_backoff_thresh = ilog2(TCP_RTO_MAX/rto_base);
165 if (boundary <= linear_backoff_thresh)
166 timeout = ((2 << boundary) - 1) * rto_base;
168 timeout = ((2 << linear_backoff_thresh) - 1) * rto_base +
169 (boundary - linear_backoff_thresh) * TCP_RTO_MAX;
171 return (tcp_time_stamp - start_ts) >= timeout;
174 /* A write timeout has occurred. Process the after effects. */
175 static int tcp_write_timeout(struct sock *sk)
177 struct inet_connection_sock *icsk = inet_csk(sk);
178 struct tcp_sock *tp = tcp_sk(sk);
180 bool do_reset, syn_set = false;
182 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
183 if (icsk->icsk_retransmits) {
184 dst_negative_advice(sk);
185 if (tp->syn_fastopen || tp->syn_data)
186 tcp_fastopen_cache_set(sk, 0, NULL, true, 0);
187 if (tp->syn_data && icsk->icsk_retransmits == 1)
188 NET_INC_STATS_BH(sock_net(sk),
189 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
191 retry_until = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
194 if (retransmits_timed_out(sk, sysctl_tcp_retries1, 0, 0)) {
195 /* Some middle-boxes may black-hole Fast Open _after_
196 * the handshake. Therefore we conservatively disable
197 * Fast Open on this path on recurring timeouts with
198 * few or zero bytes acked after Fast Open.
200 if (tp->syn_data_acked &&
201 tp->bytes_acked <= tp->rx_opt.mss_clamp) {
202 tcp_fastopen_cache_set(sk, 0, NULL, true, 0);
203 if (icsk->icsk_retransmits == sysctl_tcp_retries1)
204 NET_INC_STATS_BH(sock_net(sk),
205 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
207 /* Black hole detection */
208 tcp_mtu_probing(icsk, sk);
210 dst_negative_advice(sk);
213 retry_until = sysctl_tcp_retries2;
214 if (sock_flag(sk, SOCK_DEAD)) {
215 const bool alive = icsk->icsk_rto < TCP_RTO_MAX;
217 retry_until = tcp_orphan_retries(sk, alive);
219 !retransmits_timed_out(sk, retry_until, 0, 0);
221 if (tcp_out_of_resources(sk, do_reset))
226 if (retransmits_timed_out(sk, retry_until,
227 syn_set ? 0 : icsk->icsk_user_timeout, syn_set)) {
228 /* Has it gone just too far? */
235 void tcp_delack_timer_handler(struct sock *sk)
237 struct tcp_sock *tp = tcp_sk(sk);
238 struct inet_connection_sock *icsk = inet_csk(sk);
240 sk_mem_reclaim_partial(sk);
242 if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
243 !(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
246 if (time_after(icsk->icsk_ack.timeout, jiffies)) {
247 sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
250 icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
252 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
255 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSCHEDULERFAILED);
257 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
258 sk_backlog_rcv(sk, skb);
260 tp->ucopy.memory = 0;
263 if (inet_csk_ack_scheduled(sk)) {
264 if (!icsk->icsk_ack.pingpong) {
265 /* Delayed ACK missed: inflate ATO. */
266 icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto);
268 /* Delayed ACK missed: leave pingpong mode and
271 icsk->icsk_ack.pingpong = 0;
272 icsk->icsk_ack.ato = TCP_ATO_MIN;
275 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKS);
279 if (tcp_under_memory_pressure(sk))
283 static void tcp_delack_timer(unsigned long data)
285 struct sock *sk = (struct sock *)data;
288 if (!sock_owned_by_user(sk)) {
289 tcp_delack_timer_handler(sk);
291 inet_csk(sk)->icsk_ack.blocked = 1;
292 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
293 /* deleguate our work to tcp_release_cb() */
294 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &tcp_sk(sk)->tsq_flags))
301 static void tcp_probe_timer(struct sock *sk)
303 struct inet_connection_sock *icsk = inet_csk(sk);
304 struct tcp_sock *tp = tcp_sk(sk);
308 if (tp->packets_out || !tcp_send_head(sk)) {
309 icsk->icsk_probes_out = 0;
313 /* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as
314 * long as the receiver continues to respond probes. We support this by
315 * default and reset icsk_probes_out with incoming ACKs. But if the
316 * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we
317 * kill the socket when the retry count and the time exceeds the
318 * corresponding system limit. We also implement similar policy when
319 * we use RTO to probe window in tcp_retransmit_timer().
321 start_ts = tcp_skb_timestamp(tcp_send_head(sk));
323 skb_mstamp_get(&tcp_send_head(sk)->skb_mstamp);
324 else if (icsk->icsk_user_timeout &&
325 (s32)(tcp_time_stamp - start_ts) > icsk->icsk_user_timeout)
328 max_probes = sysctl_tcp_retries2;
329 if (sock_flag(sk, SOCK_DEAD)) {
330 const bool alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX;
332 max_probes = tcp_orphan_retries(sk, alive);
333 if (!alive && icsk->icsk_backoff >= max_probes)
335 if (tcp_out_of_resources(sk, true))
339 if (icsk->icsk_probes_out >= max_probes) {
340 abort: tcp_write_err(sk);
342 /* Only send another probe if we didn't close things up. */
348 * Timer for Fast Open socket to retransmit SYNACK. Note that the
349 * sk here is the child socket, not the parent (listener) socket.
351 static void tcp_fastopen_synack_timer(struct sock *sk)
353 struct inet_connection_sock *icsk = inet_csk(sk);
354 int max_retries = icsk->icsk_syn_retries ? :
355 sysctl_tcp_synack_retries + 1; /* add one more retry for fastopen */
356 struct request_sock *req;
358 req = tcp_sk(sk)->fastopen_rsk;
359 req->rsk_ops->syn_ack_timeout(req);
361 if (req->num_timeout >= max_retries) {
365 /* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
366 * returned from rtx_syn_ack() to make it more persistent like
367 * regular retransmit because if the child socket has been accepted
368 * it's not good to give up too easily.
370 inet_rtx_syn_ack(sk, req);
372 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
373 TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
377 * The TCP retransmit timer.
380 void tcp_retransmit_timer(struct sock *sk)
382 struct tcp_sock *tp = tcp_sk(sk);
383 struct inet_connection_sock *icsk = inet_csk(sk);
385 if (tp->fastopen_rsk) {
386 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
387 sk->sk_state != TCP_FIN_WAIT1);
388 tcp_fastopen_synack_timer(sk);
389 /* Before we receive ACK to our SYN-ACK don't retransmit
390 * anything else (e.g., data or FIN segments).
394 if (!tp->packets_out)
397 WARN_ON(tcp_write_queue_empty(sk));
399 tp->tlp_high_seq = 0;
401 if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
402 !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
403 /* Receiver dastardly shrinks window. Our retransmits
404 * become zero probes, but we should not timeout this
405 * connection. If the socket is an orphan, time it out,
406 * we cannot allow such beasts to hang infinitely.
408 struct inet_sock *inet = inet_sk(sk);
409 if (sk->sk_family == AF_INET) {
410 net_dbg_ratelimited("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
412 ntohs(inet->inet_dport),
414 tp->snd_una, tp->snd_nxt);
416 #if IS_ENABLED(CONFIG_IPV6)
417 else if (sk->sk_family == AF_INET6) {
418 net_dbg_ratelimited("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
420 ntohs(inet->inet_dport),
422 tp->snd_una, tp->snd_nxt);
425 if (tcp_time_stamp - tp->rcv_tstamp > TCP_RTO_MAX) {
430 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
432 goto out_reset_timer;
435 if (tcp_write_timeout(sk))
438 if (icsk->icsk_retransmits == 0) {
441 if (icsk->icsk_ca_state == TCP_CA_Recovery) {
443 mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
445 mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
446 } else if (icsk->icsk_ca_state == TCP_CA_Loss) {
447 mib_idx = LINUX_MIB_TCPLOSSFAILURES;
448 } else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
451 mib_idx = LINUX_MIB_TCPSACKFAILURES;
453 mib_idx = LINUX_MIB_TCPRENOFAILURES;
455 mib_idx = LINUX_MIB_TCPTIMEOUTS;
457 NET_INC_STATS_BH(sock_net(sk), mib_idx);
462 if (tcp_retransmit_skb(sk, tcp_write_queue_head(sk)) > 0) {
463 /* Retransmission failed because of local congestion,
466 if (!icsk->icsk_retransmits)
467 icsk->icsk_retransmits = 1;
468 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
469 min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL),
474 /* Increase the timeout each time we retransmit. Note that
475 * we do not increase the rtt estimate. rto is initialized
476 * from rtt, but increases here. Jacobson (SIGCOMM 88) suggests
477 * that doubling rto each time is the least we can get away with.
478 * In KA9Q, Karn uses this for the first few times, and then
479 * goes to quadratic. netBSD doubles, but only goes up to *64,
480 * and clamps at 1 to 64 sec afterwards. Note that 120 sec is
481 * defined in the protocol as the maximum possible RTT. I guess
482 * we'll have to use something other than TCP to talk to the
483 * University of Mars.
485 * PAWS allows us longer timeouts and large windows, so once
486 * implemented ftp to mars will work nicely. We will have to fix
487 * the 120 second clamps though!
489 icsk->icsk_backoff++;
490 icsk->icsk_retransmits++;
493 /* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
494 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
495 * might be increased if the stream oscillates between thin and thick,
496 * thus the old value might already be too high compared to the value
497 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
498 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
499 * exponential backoff behaviour to avoid continue hammering
500 * linear-timeout retransmissions into a black hole
502 if (sk->sk_state == TCP_ESTABLISHED &&
503 (tp->thin_lto || sysctl_tcp_thin_linear_timeouts) &&
504 tcp_stream_is_thin(tp) &&
505 icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
506 icsk->icsk_backoff = 0;
507 icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX);
509 /* Use normal (exponential) backoff */
510 icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
512 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX);
513 if (retransmits_timed_out(sk, sysctl_tcp_retries1 + 1, 0, 0))
519 void tcp_write_timer_handler(struct sock *sk)
521 struct inet_connection_sock *icsk = inet_csk(sk);
524 if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
528 if (time_after(icsk->icsk_timeout, jiffies)) {
529 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
533 event = icsk->icsk_pending;
536 case ICSK_TIME_EARLY_RETRANS:
537 tcp_resume_early_retransmit(sk);
539 case ICSK_TIME_LOSS_PROBE:
540 tcp_send_loss_probe(sk);
542 case ICSK_TIME_RETRANS:
543 icsk->icsk_pending = 0;
544 tcp_retransmit_timer(sk);
546 case ICSK_TIME_PROBE0:
547 icsk->icsk_pending = 0;
556 static void tcp_write_timer(unsigned long data)
558 struct sock *sk = (struct sock *)data;
561 if (!sock_owned_by_user(sk)) {
562 tcp_write_timer_handler(sk);
564 /* deleguate our work to tcp_release_cb() */
565 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &tcp_sk(sk)->tsq_flags))
572 void tcp_syn_ack_timeout(const struct request_sock *req)
574 struct net *net = read_pnet(&inet_rsk(req)->ireq_net);
576 NET_INC_STATS_BH(net, LINUX_MIB_TCPTIMEOUTS);
578 EXPORT_SYMBOL(tcp_syn_ack_timeout);
580 void tcp_set_keepalive(struct sock *sk, int val)
582 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
585 if (val && !sock_flag(sk, SOCK_KEEPOPEN))
586 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
588 inet_csk_delete_keepalive_timer(sk);
592 static void tcp_keepalive_timer (unsigned long data)
594 struct sock *sk = (struct sock *) data;
595 struct inet_connection_sock *icsk = inet_csk(sk);
596 struct tcp_sock *tp = tcp_sk(sk);
599 /* Only process if socket is not in use. */
601 if (sock_owned_by_user(sk)) {
602 /* Try again later. */
603 inet_csk_reset_keepalive_timer (sk, HZ/20);
607 if (sk->sk_state == TCP_LISTEN) {
608 pr_err("Hmm... keepalive on a LISTEN ???\n");
612 if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
613 if (tp->linger2 >= 0) {
614 const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
617 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
621 tcp_send_active_reset(sk, GFP_ATOMIC);
625 if (!sock_flag(sk, SOCK_KEEPOPEN) ||
626 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_SYN_SENT)))
629 elapsed = keepalive_time_when(tp);
631 /* It is alive without keepalive 8) */
632 if (tp->packets_out || tcp_send_head(sk))
635 elapsed = keepalive_time_elapsed(tp);
637 if (elapsed >= keepalive_time_when(tp)) {
638 /* If the TCP_USER_TIMEOUT option is enabled, use that
639 * to determine when to timeout instead.
641 if ((icsk->icsk_user_timeout != 0 &&
642 elapsed >= icsk->icsk_user_timeout &&
643 icsk->icsk_probes_out > 0) ||
644 (icsk->icsk_user_timeout == 0 &&
645 icsk->icsk_probes_out >= keepalive_probes(tp))) {
646 tcp_send_active_reset(sk, GFP_ATOMIC);
650 if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) {
651 icsk->icsk_probes_out++;
652 elapsed = keepalive_intvl_when(tp);
654 /* If keepalive was lost due to local congestion,
657 elapsed = TCP_RESOURCE_PROBE_INTERVAL;
660 /* It is tp->rcv_tstamp + keepalive_time_when(tp) */
661 elapsed = keepalive_time_when(tp) - elapsed;
667 inet_csk_reset_keepalive_timer (sk, elapsed);
678 void tcp_init_xmit_timers(struct sock *sk)
680 inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
681 &tcp_keepalive_timer);