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).
21 #include <linux/module.h>
22 #include <linux/gfp.h>
26 * tcp_write_err() - close socket and save error info
27 * @sk: The socket the error has appeared on.
29 * Returns: Nothing (void)
32 static void tcp_write_err(struct sock *sk)
34 sk->sk_err = sk->sk_err_soft ? : ETIMEDOUT;
35 sk->sk_error_report(sk);
38 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT);
42 * tcp_out_of_resources() - Close socket if out of resources
43 * @sk: pointer to current socket
44 * @do_reset: send a last packet with reset flag
46 * Do not allow orphaned sockets to eat all our resources.
47 * This is direct violation of TCP specs, but it is required
48 * to prevent DoS attacks. It is called when a retransmission timeout
49 * or zero probe timeout occurs on orphaned socket.
51 * Criteria is still not confirmed experimentally and may change.
52 * We kill the socket, if:
53 * 1. If number of orphaned sockets exceeds an administratively configured
55 * 2. If we have strong memory pressure.
57 static int tcp_out_of_resources(struct sock *sk, bool do_reset)
59 struct tcp_sock *tp = tcp_sk(sk);
62 /* If peer does not open window for long time, or did not transmit
63 * anything for long time, penalize it. */
64 if ((s32)(tcp_jiffies32 - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset)
67 /* If some dubious ICMP arrived, penalize even more. */
71 if (tcp_check_oom(sk, shift)) {
72 /* Catch exceptional cases, when connection requires reset.
73 * 1. Last segment was sent recently. */
74 if ((s32)(tcp_jiffies32 - tp->lsndtime) <= TCP_TIMEWAIT_LEN ||
75 /* 2. Window is closed. */
76 (!tp->snd_wnd && !tp->packets_out))
79 tcp_send_active_reset(sk, GFP_ATOMIC);
81 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY);
88 * tcp_orphan_retries() - Returns maximal number of retries on an orphaned socket
89 * @sk: Pointer to the current socket.
90 * @alive: bool, socket alive state
92 static int tcp_orphan_retries(struct sock *sk, bool alive)
94 int retries = sock_net(sk)->ipv4.sysctl_tcp_orphan_retries; /* May be zero. */
96 /* We know from an ICMP that something is wrong. */
97 if (sk->sk_err_soft && !alive)
100 /* However, if socket sent something recently, select some safe
101 * number of retries. 8 corresponds to >100 seconds with minimal
103 if (retries == 0 && alive)
108 static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk)
110 const struct net *net = sock_net(sk);
113 /* Black hole detection */
114 if (!net->ipv4.sysctl_tcp_mtu_probing)
117 if (!icsk->icsk_mtup.enabled) {
118 icsk->icsk_mtup.enabled = 1;
119 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
121 mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1;
122 mss = min(net->ipv4.sysctl_tcp_base_mss, mss);
123 mss = max(mss, 68 - tcp_sk(sk)->tcp_header_len);
124 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
126 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
131 * retransmits_timed_out() - returns true if this connection has timed out
132 * @sk: The current socket
133 * @boundary: max number of retransmissions
134 * @timeout: A custom timeout value.
135 * If set to 0 the default timeout is calculated and used.
136 * Using TCP_RTO_MIN and the number of unsuccessful retransmits.
138 * The default "timeout" value this function can calculate and use
139 * is equivalent to the timeout of a TCP Connection
140 * after "boundary" unsuccessful, exponentially backed-off
141 * retransmissions with an initial RTO of TCP_RTO_MIN.
143 static bool retransmits_timed_out(struct sock *sk,
144 unsigned int boundary,
145 unsigned int timeout)
147 const unsigned int rto_base = TCP_RTO_MIN;
148 unsigned int linear_backoff_thresh, start_ts;
150 if (!inet_csk(sk)->icsk_retransmits)
153 start_ts = tcp_sk(sk)->retrans_stamp;
154 if (unlikely(!start_ts)) {
155 struct sk_buff *head = tcp_rtx_queue_head(sk);
159 start_ts = tcp_skb_timestamp(head);
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(tcp_sk(sk)) - start_ts) >= jiffies_to_msecs(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);
179 struct net *net = sock_net(sk);
180 bool expired, do_reset;
183 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
184 if (icsk->icsk_retransmits) {
185 dst_negative_advice(sk);
186 if (tp->syn_fastopen || tp->syn_data)
187 tcp_fastopen_cache_set(sk, 0, NULL, true, 0);
188 if (tp->syn_data && icsk->icsk_retransmits == 1)
189 NET_INC_STATS(sock_net(sk),
190 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
191 } else if (!tp->syn_data && !tp->syn_fastopen) {
192 sk_rethink_txhash(sk);
194 retry_until = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
195 expired = icsk->icsk_retransmits >= retry_until;
197 if (retransmits_timed_out(sk, net->ipv4.sysctl_tcp_retries1, 0)) {
198 /* Some middle-boxes may black-hole Fast Open _after_
199 * the handshake. Therefore we conservatively disable
200 * Fast Open on this path on recurring timeouts after
201 * successful Fast Open.
203 if (tp->syn_data_acked) {
204 tcp_fastopen_cache_set(sk, 0, NULL, true, 0);
205 if (icsk->icsk_retransmits == net->ipv4.sysctl_tcp_retries1)
206 NET_INC_STATS(sock_net(sk),
207 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
209 /* Black hole detection */
210 tcp_mtu_probing(icsk, sk);
212 dst_negative_advice(sk);
214 sk_rethink_txhash(sk);
217 retry_until = net->ipv4.sysctl_tcp_retries2;
218 if (sock_flag(sk, SOCK_DEAD)) {
219 const bool alive = icsk->icsk_rto < TCP_RTO_MAX;
221 retry_until = tcp_orphan_retries(sk, alive);
223 !retransmits_timed_out(sk, retry_until, 0);
225 if (tcp_out_of_resources(sk, do_reset))
228 expired = retransmits_timed_out(sk, retry_until,
229 icsk->icsk_user_timeout);
232 /* Has it gone just too far? */
239 /* Called with BH disabled */
240 void tcp_delack_timer_handler(struct sock *sk)
242 struct inet_connection_sock *icsk = inet_csk(sk);
244 sk_mem_reclaim_partial(sk);
246 if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
247 !(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
250 if (time_after(icsk->icsk_ack.timeout, jiffies)) {
251 sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
254 icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
256 if (inet_csk_ack_scheduled(sk)) {
257 if (!icsk->icsk_ack.pingpong) {
258 /* Delayed ACK missed: inflate ATO. */
259 icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto);
261 /* Delayed ACK missed: leave pingpong mode and
264 icsk->icsk_ack.pingpong = 0;
265 icsk->icsk_ack.ato = TCP_ATO_MIN;
267 tcp_mstamp_refresh(tcp_sk(sk));
269 __NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKS);
273 if (tcp_under_memory_pressure(sk))
279 * tcp_delack_timer() - The TCP delayed ACK timeout handler
280 * @data: Pointer to the current socket. (gets casted to struct sock *)
282 * This function gets (indirectly) called when the kernel timer for a TCP packet
283 * of this socket expires. Calls tcp_delack_timer_handler() to do the actual work.
285 * Returns: Nothing (void)
287 static void tcp_delack_timer(struct timer_list *t)
289 struct inet_connection_sock *icsk =
290 from_timer(icsk, t, icsk_delack_timer);
291 struct sock *sk = &icsk->icsk_inet.sk;
294 if (!sock_owned_by_user(sk)) {
295 tcp_delack_timer_handler(sk);
297 icsk->icsk_ack.blocked = 1;
298 __NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
299 /* deleguate our work to tcp_release_cb() */
300 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &sk->sk_tsq_flags))
307 static void tcp_probe_timer(struct sock *sk)
309 struct inet_connection_sock *icsk = inet_csk(sk);
310 struct sk_buff *skb = tcp_send_head(sk);
311 struct tcp_sock *tp = tcp_sk(sk);
315 if (tp->packets_out || !skb) {
316 icsk->icsk_probes_out = 0;
320 /* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as
321 * long as the receiver continues to respond probes. We support this by
322 * default and reset icsk_probes_out with incoming ACKs. But if the
323 * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we
324 * kill the socket when the retry count and the time exceeds the
325 * corresponding system limit. We also implement similar policy when
326 * we use RTO to probe window in tcp_retransmit_timer().
328 start_ts = tcp_skb_timestamp(skb);
330 skb->skb_mstamp = tp->tcp_mstamp;
331 else if (icsk->icsk_user_timeout &&
332 (s32)(tcp_time_stamp(tp) - start_ts) >
333 jiffies_to_msecs(icsk->icsk_user_timeout))
336 max_probes = sock_net(sk)->ipv4.sysctl_tcp_retries2;
337 if (sock_flag(sk, SOCK_DEAD)) {
338 const bool alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX;
340 max_probes = tcp_orphan_retries(sk, alive);
341 if (!alive && icsk->icsk_backoff >= max_probes)
343 if (tcp_out_of_resources(sk, true))
347 if (icsk->icsk_probes_out > max_probes) {
348 abort: tcp_write_err(sk);
350 /* Only send another probe if we didn't close things up. */
356 * Timer for Fast Open socket to retransmit SYNACK. Note that the
357 * sk here is the child socket, not the parent (listener) socket.
359 static void tcp_fastopen_synack_timer(struct sock *sk)
361 struct inet_connection_sock *icsk = inet_csk(sk);
362 int max_retries = icsk->icsk_syn_retries ? :
363 sock_net(sk)->ipv4.sysctl_tcp_synack_retries + 1; /* add one more retry for fastopen */
364 struct request_sock *req;
366 req = tcp_sk(sk)->fastopen_rsk;
367 req->rsk_ops->syn_ack_timeout(req);
369 if (req->num_timeout >= max_retries) {
373 /* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
374 * returned from rtx_syn_ack() to make it more persistent like
375 * regular retransmit because if the child socket has been accepted
376 * it's not good to give up too easily.
378 inet_rtx_syn_ack(sk, req);
380 icsk->icsk_retransmits++;
381 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
382 TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
387 * tcp_retransmit_timer() - The TCP retransmit timeout handler
388 * @sk: Pointer to the current socket.
390 * This function gets called when the kernel timer for a TCP packet
391 * of this socket expires.
393 * It handles retransmission, timer adjustment and other necesarry measures.
395 * Returns: Nothing (void)
397 void tcp_retransmit_timer(struct sock *sk)
399 struct tcp_sock *tp = tcp_sk(sk);
400 struct net *net = sock_net(sk);
401 struct inet_connection_sock *icsk = inet_csk(sk);
403 if (tp->fastopen_rsk) {
404 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
405 sk->sk_state != TCP_FIN_WAIT1);
406 tcp_fastopen_synack_timer(sk);
407 /* Before we receive ACK to our SYN-ACK don't retransmit
408 * anything else (e.g., data or FIN segments).
412 if (!tp->packets_out)
415 WARN_ON(tcp_rtx_queue_empty(sk));
417 tp->tlp_high_seq = 0;
419 if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
420 !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
421 /* Receiver dastardly shrinks window. Our retransmits
422 * become zero probes, but we should not timeout this
423 * connection. If the socket is an orphan, time it out,
424 * we cannot allow such beasts to hang infinitely.
426 struct inet_sock *inet = inet_sk(sk);
427 if (sk->sk_family == AF_INET) {
428 net_dbg_ratelimited("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
430 ntohs(inet->inet_dport),
432 tp->snd_una, tp->snd_nxt);
434 #if IS_ENABLED(CONFIG_IPV6)
435 else if (sk->sk_family == AF_INET6) {
436 net_dbg_ratelimited("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
438 ntohs(inet->inet_dport),
440 tp->snd_una, tp->snd_nxt);
443 if (tcp_jiffies32 - tp->rcv_tstamp > TCP_RTO_MAX) {
448 tcp_retransmit_skb(sk, tcp_rtx_queue_head(sk), 1);
450 goto out_reset_timer;
453 if (tcp_write_timeout(sk))
456 if (icsk->icsk_retransmits == 0) {
459 if (icsk->icsk_ca_state == TCP_CA_Recovery) {
461 mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
463 mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
464 } else if (icsk->icsk_ca_state == TCP_CA_Loss) {
465 mib_idx = LINUX_MIB_TCPLOSSFAILURES;
466 } else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
469 mib_idx = LINUX_MIB_TCPSACKFAILURES;
471 mib_idx = LINUX_MIB_TCPRENOFAILURES;
473 mib_idx = LINUX_MIB_TCPTIMEOUTS;
475 __NET_INC_STATS(sock_net(sk), mib_idx);
480 if (tcp_retransmit_skb(sk, tcp_rtx_queue_head(sk), 1) > 0) {
481 /* Retransmission failed because of local congestion,
484 if (!icsk->icsk_retransmits)
485 icsk->icsk_retransmits = 1;
486 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
487 min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL),
492 /* Increase the timeout each time we retransmit. Note that
493 * we do not increase the rtt estimate. rto is initialized
494 * from rtt, but increases here. Jacobson (SIGCOMM 88) suggests
495 * that doubling rto each time is the least we can get away with.
496 * In KA9Q, Karn uses this for the first few times, and then
497 * goes to quadratic. netBSD doubles, but only goes up to *64,
498 * and clamps at 1 to 64 sec afterwards. Note that 120 sec is
499 * defined in the protocol as the maximum possible RTT. I guess
500 * we'll have to use something other than TCP to talk to the
501 * University of Mars.
503 * PAWS allows us longer timeouts and large windows, so once
504 * implemented ftp to mars will work nicely. We will have to fix
505 * the 120 second clamps though!
507 icsk->icsk_backoff++;
508 icsk->icsk_retransmits++;
511 /* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
512 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
513 * might be increased if the stream oscillates between thin and thick,
514 * thus the old value might already be too high compared to the value
515 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
516 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
517 * exponential backoff behaviour to avoid continue hammering
518 * linear-timeout retransmissions into a black hole
520 if (sk->sk_state == TCP_ESTABLISHED &&
521 (tp->thin_lto || net->ipv4.sysctl_tcp_thin_linear_timeouts) &&
522 tcp_stream_is_thin(tp) &&
523 icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
524 icsk->icsk_backoff = 0;
525 icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX);
527 /* Use normal (exponential) backoff */
528 icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
530 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX);
531 if (retransmits_timed_out(sk, net->ipv4.sysctl_tcp_retries1 + 1, 0))
537 /* Called with bottom-half processing disabled.
538 Called by tcp_write_timer() */
539 void tcp_write_timer_handler(struct sock *sk)
541 struct inet_connection_sock *icsk = inet_csk(sk);
544 if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
548 if (time_after(icsk->icsk_timeout, jiffies)) {
549 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
553 tcp_mstamp_refresh(tcp_sk(sk));
554 event = icsk->icsk_pending;
557 case ICSK_TIME_REO_TIMEOUT:
558 tcp_rack_reo_timeout(sk);
560 case ICSK_TIME_LOSS_PROBE:
561 tcp_send_loss_probe(sk);
563 case ICSK_TIME_RETRANS:
564 icsk->icsk_pending = 0;
565 tcp_retransmit_timer(sk);
567 case ICSK_TIME_PROBE0:
568 icsk->icsk_pending = 0;
577 static void tcp_write_timer(struct timer_list *t)
579 struct inet_connection_sock *icsk =
580 from_timer(icsk, t, icsk_retransmit_timer);
581 struct sock *sk = &icsk->icsk_inet.sk;
584 if (!sock_owned_by_user(sk)) {
585 tcp_write_timer_handler(sk);
587 /* delegate our work to tcp_release_cb() */
588 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &sk->sk_tsq_flags))
595 void tcp_syn_ack_timeout(const struct request_sock *req)
597 struct net *net = read_pnet(&inet_rsk(req)->ireq_net);
599 __NET_INC_STATS(net, LINUX_MIB_TCPTIMEOUTS);
601 EXPORT_SYMBOL(tcp_syn_ack_timeout);
603 void tcp_set_keepalive(struct sock *sk, int val)
605 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
608 if (val && !sock_flag(sk, SOCK_KEEPOPEN))
609 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
611 inet_csk_delete_keepalive_timer(sk);
613 EXPORT_SYMBOL_GPL(tcp_set_keepalive);
616 static void tcp_keepalive_timer (struct timer_list *t)
618 struct sock *sk = from_timer(sk, t, sk_timer);
619 struct inet_connection_sock *icsk = inet_csk(sk);
620 struct tcp_sock *tp = tcp_sk(sk);
623 /* Only process if socket is not in use. */
625 if (sock_owned_by_user(sk)) {
626 /* Try again later. */
627 inet_csk_reset_keepalive_timer (sk, HZ/20);
631 if (sk->sk_state == TCP_LISTEN) {
632 pr_err("Hmm... keepalive on a LISTEN ???\n");
636 tcp_mstamp_refresh(tp);
637 if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
638 if (tp->linger2 >= 0) {
639 const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
642 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
646 tcp_send_active_reset(sk, GFP_ATOMIC);
650 if (!sock_flag(sk, SOCK_KEEPOPEN) ||
651 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_SYN_SENT)))
654 elapsed = keepalive_time_when(tp);
656 /* It is alive without keepalive 8) */
657 if (tp->packets_out || !tcp_write_queue_empty(sk))
660 elapsed = keepalive_time_elapsed(tp);
662 if (elapsed >= keepalive_time_when(tp)) {
663 /* If the TCP_USER_TIMEOUT option is enabled, use that
664 * to determine when to timeout instead.
666 if ((icsk->icsk_user_timeout != 0 &&
667 elapsed >= icsk->icsk_user_timeout &&
668 icsk->icsk_probes_out > 0) ||
669 (icsk->icsk_user_timeout == 0 &&
670 icsk->icsk_probes_out >= keepalive_probes(tp))) {
671 tcp_send_active_reset(sk, GFP_ATOMIC);
675 if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) {
676 icsk->icsk_probes_out++;
677 elapsed = keepalive_intvl_when(tp);
679 /* If keepalive was lost due to local congestion,
682 elapsed = TCP_RESOURCE_PROBE_INTERVAL;
685 /* It is tp->rcv_tstamp + keepalive_time_when(tp) */
686 elapsed = keepalive_time_when(tp) - elapsed;
692 inet_csk_reset_keepalive_timer (sk, elapsed);
703 void tcp_init_xmit_timers(struct sock *sk)
705 inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
706 &tcp_keepalive_timer);
707 hrtimer_init(&tcp_sk(sk)->pacing_timer, CLOCK_MONOTONIC,
708 HRTIMER_MODE_ABS_PINNED);
709 tcp_sk(sk)->pacing_timer.function = tcp_pace_kick;