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_retries1 __read_mostly
= TCP_RETR1
;
26 int sysctl_tcp_retries2 __read_mostly
= TCP_RETR2
;
27 int sysctl_tcp_orphan_retries __read_mostly
;
28 int sysctl_tcp_thin_linear_timeouts __read_mostly
;
30 static void tcp_write_err(struct sock
*sk
)
32 sk
->sk_err
= sk
->sk_err_soft
? : ETIMEDOUT
;
33 sk
->sk_error_report(sk
);
36 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONTIMEOUT
);
39 /* Do not allow orphaned sockets to eat all our resources.
40 * This is direct violation of TCP specs, but it is required
41 * to prevent DoS attacks. It is called when a retransmission timeout
42 * or zero probe timeout occurs on orphaned socket.
44 * Criteria is still not confirmed experimentally and may change.
45 * We kill the socket, if:
46 * 1. If number of orphaned sockets exceeds an administratively configured
48 * 2. If we have strong memory pressure.
50 static int tcp_out_of_resources(struct sock
*sk
, bool do_reset
)
52 struct tcp_sock
*tp
= tcp_sk(sk
);
55 /* If peer does not open window for long time, or did not transmit
56 * anything for long time, penalize it. */
57 if ((s32
)(tcp_time_stamp
- tp
->lsndtime
) > 2*TCP_RTO_MAX
|| !do_reset
)
60 /* If some dubious ICMP arrived, penalize even more. */
64 if (tcp_check_oom(sk
, shift
)) {
65 /* Catch exceptional cases, when connection requires reset.
66 * 1. Last segment was sent recently. */
67 if ((s32
)(tcp_time_stamp
- tp
->lsndtime
) <= TCP_TIMEWAIT_LEN
||
68 /* 2. Window is closed. */
69 (!tp
->snd_wnd
&& !tp
->packets_out
))
72 tcp_send_active_reset(sk
, GFP_ATOMIC
);
74 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONMEMORY
);
80 /* Calculate maximal number or retries on an orphaned socket. */
81 static int tcp_orphan_retries(struct sock
*sk
, bool alive
)
83 int retries
= sysctl_tcp_orphan_retries
; /* May be zero. */
85 /* We know from an ICMP that something is wrong. */
86 if (sk
->sk_err_soft
&& !alive
)
89 /* However, if socket sent something recently, select some safe
90 * number of retries. 8 corresponds to >100 seconds with minimal
92 if (retries
== 0 && alive
)
97 static void tcp_mtu_probing(struct inet_connection_sock
*icsk
, struct sock
*sk
)
99 struct net
*net
= sock_net(sk
);
101 /* Black hole detection */
102 if (net
->ipv4
.sysctl_tcp_mtu_probing
) {
103 if (!icsk
->icsk_mtup
.enabled
) {
104 icsk
->icsk_mtup
.enabled
= 1;
105 icsk
->icsk_mtup
.probe_timestamp
= tcp_time_stamp
;
106 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
108 struct net
*net
= sock_net(sk
);
109 struct tcp_sock
*tp
= tcp_sk(sk
);
112 mss
= tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
) >> 1;
113 mss
= min(net
->ipv4
.sysctl_tcp_base_mss
, mss
);
114 mss
= max(mss
, 68 - tp
->tcp_header_len
);
115 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, mss
);
116 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
121 /* This function calculates a "timeout" which is equivalent to the timeout of a
122 * TCP connection after "boundary" unsuccessful, exponentially backed-off
123 * retransmissions with an initial RTO of TCP_RTO_MIN or TCP_TIMEOUT_INIT if
124 * syn_set flag is set.
126 static bool retransmits_timed_out(struct sock
*sk
,
127 unsigned int boundary
,
128 unsigned int timeout
,
131 unsigned int linear_backoff_thresh
, start_ts
;
132 unsigned int rto_base
= syn_set
? TCP_TIMEOUT_INIT
: TCP_RTO_MIN
;
134 if (!inet_csk(sk
)->icsk_retransmits
)
137 start_ts
= tcp_sk(sk
)->retrans_stamp
;
138 if (unlikely(!start_ts
))
139 start_ts
= tcp_skb_timestamp(tcp_write_queue_head(sk
));
141 if (likely(timeout
== 0)) {
142 linear_backoff_thresh
= ilog2(TCP_RTO_MAX
/rto_base
);
144 if (boundary
<= linear_backoff_thresh
)
145 timeout
= ((2 << boundary
) - 1) * rto_base
;
147 timeout
= ((2 << linear_backoff_thresh
) - 1) * rto_base
+
148 (boundary
- linear_backoff_thresh
) * TCP_RTO_MAX
;
150 return (tcp_time_stamp
- start_ts
) >= timeout
;
153 /* A write timeout has occurred. Process the after effects. */
154 static int tcp_write_timeout(struct sock
*sk
)
156 struct inet_connection_sock
*icsk
= inet_csk(sk
);
157 struct tcp_sock
*tp
= tcp_sk(sk
);
158 struct net
*net
= sock_net(sk
);
160 bool do_reset
, syn_set
= false;
162 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
)) {
163 if (icsk
->icsk_retransmits
) {
164 dst_negative_advice(sk
);
165 if (tp
->syn_fastopen
|| tp
->syn_data
)
166 tcp_fastopen_cache_set(sk
, 0, NULL
, true, 0);
167 if (tp
->syn_data
&& icsk
->icsk_retransmits
== 1)
168 NET_INC_STATS_BH(sock_net(sk
),
169 LINUX_MIB_TCPFASTOPENACTIVEFAIL
);
171 retry_until
= icsk
->icsk_syn_retries
? : net
->ipv4
.sysctl_tcp_syn_retries
;
174 if (retransmits_timed_out(sk
, sysctl_tcp_retries1
, 0, 0)) {
175 /* Some middle-boxes may black-hole Fast Open _after_
176 * the handshake. Therefore we conservatively disable
177 * Fast Open on this path on recurring timeouts with
178 * few or zero bytes acked after Fast Open.
180 if (tp
->syn_data_acked
&&
181 tp
->bytes_acked
<= tp
->rx_opt
.mss_clamp
) {
182 tcp_fastopen_cache_set(sk
, 0, NULL
, true, 0);
183 if (icsk
->icsk_retransmits
== sysctl_tcp_retries1
)
184 NET_INC_STATS_BH(sock_net(sk
),
185 LINUX_MIB_TCPFASTOPENACTIVEFAIL
);
187 /* Black hole detection */
188 tcp_mtu_probing(icsk
, sk
);
190 dst_negative_advice(sk
);
193 retry_until
= sysctl_tcp_retries2
;
194 if (sock_flag(sk
, SOCK_DEAD
)) {
195 const bool alive
= icsk
->icsk_rto
< TCP_RTO_MAX
;
197 retry_until
= tcp_orphan_retries(sk
, alive
);
199 !retransmits_timed_out(sk
, retry_until
, 0, 0);
201 if (tcp_out_of_resources(sk
, do_reset
))
206 if (retransmits_timed_out(sk
, retry_until
,
207 syn_set
? 0 : icsk
->icsk_user_timeout
, syn_set
)) {
208 /* Has it gone just too far? */
215 void tcp_delack_timer_handler(struct sock
*sk
)
217 struct tcp_sock
*tp
= tcp_sk(sk
);
218 struct inet_connection_sock
*icsk
= inet_csk(sk
);
220 sk_mem_reclaim_partial(sk
);
222 if (sk
->sk_state
== TCP_CLOSE
|| !(icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
))
225 if (time_after(icsk
->icsk_ack
.timeout
, jiffies
)) {
226 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, icsk
->icsk_ack
.timeout
);
229 icsk
->icsk_ack
.pending
&= ~ICSK_ACK_TIMER
;
231 if (!skb_queue_empty(&tp
->ucopy
.prequeue
)) {
234 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSCHEDULERFAILED
);
236 while ((skb
= __skb_dequeue(&tp
->ucopy
.prequeue
)) != NULL
)
237 sk_backlog_rcv(sk
, skb
);
239 tp
->ucopy
.memory
= 0;
242 if (inet_csk_ack_scheduled(sk
)) {
243 if (!icsk
->icsk_ack
.pingpong
) {
244 /* Delayed ACK missed: inflate ATO. */
245 icsk
->icsk_ack
.ato
= min(icsk
->icsk_ack
.ato
<< 1, icsk
->icsk_rto
);
247 /* Delayed ACK missed: leave pingpong mode and
250 icsk
->icsk_ack
.pingpong
= 0;
251 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
254 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKS
);
258 if (tcp_under_memory_pressure(sk
))
262 static void tcp_delack_timer(unsigned long data
)
264 struct sock
*sk
= (struct sock
*)data
;
267 if (!sock_owned_by_user(sk
)) {
268 tcp_delack_timer_handler(sk
);
270 inet_csk(sk
)->icsk_ack
.blocked
= 1;
271 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOCKED
);
272 /* deleguate our work to tcp_release_cb() */
273 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED
, &tcp_sk(sk
)->tsq_flags
))
280 static void tcp_probe_timer(struct sock
*sk
)
282 struct inet_connection_sock
*icsk
= inet_csk(sk
);
283 struct tcp_sock
*tp
= tcp_sk(sk
);
287 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
288 icsk
->icsk_probes_out
= 0;
292 /* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as
293 * long as the receiver continues to respond probes. We support this by
294 * default and reset icsk_probes_out with incoming ACKs. But if the
295 * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we
296 * kill the socket when the retry count and the time exceeds the
297 * corresponding system limit. We also implement similar policy when
298 * we use RTO to probe window in tcp_retransmit_timer().
300 start_ts
= tcp_skb_timestamp(tcp_send_head(sk
));
302 skb_mstamp_get(&tcp_send_head(sk
)->skb_mstamp
);
303 else if (icsk
->icsk_user_timeout
&&
304 (s32
)(tcp_time_stamp
- start_ts
) > icsk
->icsk_user_timeout
)
307 max_probes
= sysctl_tcp_retries2
;
308 if (sock_flag(sk
, SOCK_DEAD
)) {
309 const bool alive
= inet_csk_rto_backoff(icsk
, TCP_RTO_MAX
) < TCP_RTO_MAX
;
311 max_probes
= tcp_orphan_retries(sk
, alive
);
312 if (!alive
&& icsk
->icsk_backoff
>= max_probes
)
314 if (tcp_out_of_resources(sk
, true))
318 if (icsk
->icsk_probes_out
> max_probes
) {
319 abort
: tcp_write_err(sk
);
321 /* Only send another probe if we didn't close things up. */
327 * Timer for Fast Open socket to retransmit SYNACK. Note that the
328 * sk here is the child socket, not the parent (listener) socket.
330 static void tcp_fastopen_synack_timer(struct sock
*sk
)
332 struct inet_connection_sock
*icsk
= inet_csk(sk
);
333 int max_retries
= icsk
->icsk_syn_retries
? :
334 sock_net(sk
)->ipv4
.sysctl_tcp_synack_retries
+ 1; /* add one more retry for fastopen */
335 struct request_sock
*req
;
337 req
= tcp_sk(sk
)->fastopen_rsk
;
338 req
->rsk_ops
->syn_ack_timeout(req
);
340 if (req
->num_timeout
>= max_retries
) {
344 /* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
345 * returned from rtx_syn_ack() to make it more persistent like
346 * regular retransmit because if the child socket has been accepted
347 * it's not good to give up too easily.
349 inet_rtx_syn_ack(sk
, req
);
351 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
352 TCP_TIMEOUT_INIT
<< req
->num_timeout
, TCP_RTO_MAX
);
356 * The TCP retransmit timer.
359 void tcp_retransmit_timer(struct sock
*sk
)
361 struct tcp_sock
*tp
= tcp_sk(sk
);
362 struct inet_connection_sock
*icsk
= inet_csk(sk
);
364 if (tp
->fastopen_rsk
) {
365 WARN_ON_ONCE(sk
->sk_state
!= TCP_SYN_RECV
&&
366 sk
->sk_state
!= TCP_FIN_WAIT1
);
367 tcp_fastopen_synack_timer(sk
);
368 /* Before we receive ACK to our SYN-ACK don't retransmit
369 * anything else (e.g., data or FIN segments).
373 if (!tp
->packets_out
)
376 WARN_ON(tcp_write_queue_empty(sk
));
378 tp
->tlp_high_seq
= 0;
380 if (!tp
->snd_wnd
&& !sock_flag(sk
, SOCK_DEAD
) &&
381 !((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))) {
382 /* Receiver dastardly shrinks window. Our retransmits
383 * become zero probes, but we should not timeout this
384 * connection. If the socket is an orphan, time it out,
385 * we cannot allow such beasts to hang infinitely.
387 struct inet_sock
*inet
= inet_sk(sk
);
388 if (sk
->sk_family
== AF_INET
) {
389 net_dbg_ratelimited("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
391 ntohs(inet
->inet_dport
),
393 tp
->snd_una
, tp
->snd_nxt
);
395 #if IS_ENABLED(CONFIG_IPV6)
396 else if (sk
->sk_family
== AF_INET6
) {
397 net_dbg_ratelimited("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
399 ntohs(inet
->inet_dport
),
401 tp
->snd_una
, tp
->snd_nxt
);
404 if (tcp_time_stamp
- tp
->rcv_tstamp
> TCP_RTO_MAX
) {
409 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
411 goto out_reset_timer
;
414 if (tcp_write_timeout(sk
))
417 if (icsk
->icsk_retransmits
== 0) {
420 if (icsk
->icsk_ca_state
== TCP_CA_Recovery
) {
422 mib_idx
= LINUX_MIB_TCPSACKRECOVERYFAIL
;
424 mib_idx
= LINUX_MIB_TCPRENORECOVERYFAIL
;
425 } else if (icsk
->icsk_ca_state
== TCP_CA_Loss
) {
426 mib_idx
= LINUX_MIB_TCPLOSSFAILURES
;
427 } else if ((icsk
->icsk_ca_state
== TCP_CA_Disorder
) ||
430 mib_idx
= LINUX_MIB_TCPSACKFAILURES
;
432 mib_idx
= LINUX_MIB_TCPRENOFAILURES
;
434 mib_idx
= LINUX_MIB_TCPTIMEOUTS
;
436 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
441 if (tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
)) > 0) {
442 /* Retransmission failed because of local congestion,
445 if (!icsk
->icsk_retransmits
)
446 icsk
->icsk_retransmits
= 1;
447 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
448 min(icsk
->icsk_rto
, TCP_RESOURCE_PROBE_INTERVAL
),
453 /* Increase the timeout each time we retransmit. Note that
454 * we do not increase the rtt estimate. rto is initialized
455 * from rtt, but increases here. Jacobson (SIGCOMM 88) suggests
456 * that doubling rto each time is the least we can get away with.
457 * In KA9Q, Karn uses this for the first few times, and then
458 * goes to quadratic. netBSD doubles, but only goes up to *64,
459 * and clamps at 1 to 64 sec afterwards. Note that 120 sec is
460 * defined in the protocol as the maximum possible RTT. I guess
461 * we'll have to use something other than TCP to talk to the
462 * University of Mars.
464 * PAWS allows us longer timeouts and large windows, so once
465 * implemented ftp to mars will work nicely. We will have to fix
466 * the 120 second clamps though!
468 icsk
->icsk_backoff
++;
469 icsk
->icsk_retransmits
++;
472 /* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
473 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
474 * might be increased if the stream oscillates between thin and thick,
475 * thus the old value might already be too high compared to the value
476 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
477 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
478 * exponential backoff behaviour to avoid continue hammering
479 * linear-timeout retransmissions into a black hole
481 if (sk
->sk_state
== TCP_ESTABLISHED
&&
482 (tp
->thin_lto
|| sysctl_tcp_thin_linear_timeouts
) &&
483 tcp_stream_is_thin(tp
) &&
484 icsk
->icsk_retransmits
<= TCP_THIN_LINEAR_RETRIES
) {
485 icsk
->icsk_backoff
= 0;
486 icsk
->icsk_rto
= min(__tcp_set_rto(tp
), TCP_RTO_MAX
);
488 /* Use normal (exponential) backoff */
489 icsk
->icsk_rto
= min(icsk
->icsk_rto
<< 1, TCP_RTO_MAX
);
491 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, icsk
->icsk_rto
, TCP_RTO_MAX
);
492 if (retransmits_timed_out(sk
, sysctl_tcp_retries1
+ 1, 0, 0))
498 void tcp_write_timer_handler(struct sock
*sk
)
500 struct inet_connection_sock
*icsk
= inet_csk(sk
);
503 if (sk
->sk_state
== TCP_CLOSE
|| !icsk
->icsk_pending
)
506 if (time_after(icsk
->icsk_timeout
, jiffies
)) {
507 sk_reset_timer(sk
, &icsk
->icsk_retransmit_timer
, icsk
->icsk_timeout
);
511 event
= icsk
->icsk_pending
;
514 case ICSK_TIME_EARLY_RETRANS
:
515 tcp_resume_early_retransmit(sk
);
517 case ICSK_TIME_LOSS_PROBE
:
518 tcp_send_loss_probe(sk
);
520 case ICSK_TIME_RETRANS
:
521 icsk
->icsk_pending
= 0;
522 tcp_retransmit_timer(sk
);
524 case ICSK_TIME_PROBE0
:
525 icsk
->icsk_pending
= 0;
534 static void tcp_write_timer(unsigned long data
)
536 struct sock
*sk
= (struct sock
*)data
;
539 if (!sock_owned_by_user(sk
)) {
540 tcp_write_timer_handler(sk
);
542 /* deleguate our work to tcp_release_cb() */
543 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED
, &tcp_sk(sk
)->tsq_flags
))
550 void tcp_syn_ack_timeout(const struct request_sock
*req
)
552 struct net
*net
= read_pnet(&inet_rsk(req
)->ireq_net
);
554 NET_INC_STATS_BH(net
, LINUX_MIB_TCPTIMEOUTS
);
556 EXPORT_SYMBOL(tcp_syn_ack_timeout
);
558 void tcp_set_keepalive(struct sock
*sk
, int val
)
560 if ((1 << sk
->sk_state
) & (TCPF_CLOSE
| TCPF_LISTEN
))
563 if (val
&& !sock_flag(sk
, SOCK_KEEPOPEN
))
564 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tcp_sk(sk
)));
566 inet_csk_delete_keepalive_timer(sk
);
570 static void tcp_keepalive_timer (unsigned long data
)
572 struct sock
*sk
= (struct sock
*) data
;
573 struct inet_connection_sock
*icsk
= inet_csk(sk
);
574 struct tcp_sock
*tp
= tcp_sk(sk
);
577 /* Only process if socket is not in use. */
579 if (sock_owned_by_user(sk
)) {
580 /* Try again later. */
581 inet_csk_reset_keepalive_timer (sk
, HZ
/20);
585 if (sk
->sk_state
== TCP_LISTEN
) {
586 pr_err("Hmm... keepalive on a LISTEN ???\n");
590 if (sk
->sk_state
== TCP_FIN_WAIT2
&& sock_flag(sk
, SOCK_DEAD
)) {
591 if (tp
->linger2
>= 0) {
592 const int tmo
= tcp_fin_time(sk
) - TCP_TIMEWAIT_LEN
;
595 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
599 tcp_send_active_reset(sk
, GFP_ATOMIC
);
603 if (!sock_flag(sk
, SOCK_KEEPOPEN
) || sk
->sk_state
== TCP_CLOSE
)
606 elapsed
= keepalive_time_when(tp
);
608 /* It is alive without keepalive 8) */
609 if (tp
->packets_out
|| tcp_send_head(sk
))
612 elapsed
= keepalive_time_elapsed(tp
);
614 if (elapsed
>= keepalive_time_when(tp
)) {
615 /* If the TCP_USER_TIMEOUT option is enabled, use that
616 * to determine when to timeout instead.
618 if ((icsk
->icsk_user_timeout
!= 0 &&
619 elapsed
>= icsk
->icsk_user_timeout
&&
620 icsk
->icsk_probes_out
> 0) ||
621 (icsk
->icsk_user_timeout
== 0 &&
622 icsk
->icsk_probes_out
>= keepalive_probes(tp
))) {
623 tcp_send_active_reset(sk
, GFP_ATOMIC
);
627 if (tcp_write_wakeup(sk
, LINUX_MIB_TCPKEEPALIVE
) <= 0) {
628 icsk
->icsk_probes_out
++;
629 elapsed
= keepalive_intvl_when(tp
);
631 /* If keepalive was lost due to local congestion,
634 elapsed
= TCP_RESOURCE_PROBE_INTERVAL
;
637 /* It is tp->rcv_tstamp + keepalive_time_when(tp) */
638 elapsed
= keepalive_time_when(tp
) - elapsed
;
644 inet_csk_reset_keepalive_timer (sk
, elapsed
);
655 void tcp_init_xmit_timers(struct sock
*sk
)
657 inet_csk_init_xmit_timers(sk
, &tcp_write_timer
, &tcp_delack_timer
,
658 &tcp_keepalive_timer
);