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
, int 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 /* Black hole detection */
105 if (sysctl_tcp_mtu_probing
) {
106 if (!icsk
->icsk_mtup
.enabled
) {
107 icsk
->icsk_mtup
.enabled
= 1;
108 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
110 struct tcp_sock
*tp
= tcp_sk(sk
);
113 mss
= tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
) >> 1;
114 mss
= min(sysctl_tcp_base_mss
, mss
);
115 mss
= max(mss
, 68 - tp
->tcp_header_len
);
116 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, mss
);
117 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
122 /* This function calculates a "timeout" which is equivalent to the timeout of a
123 * TCP connection after "boundary" unsuccessful, exponentially backed-off
124 * retransmissions with an initial RTO of TCP_RTO_MIN or TCP_TIMEOUT_INIT if
125 * syn_set flag is set.
127 static bool retransmits_timed_out(struct sock
*sk
,
128 unsigned int boundary
,
129 unsigned int timeout
,
132 unsigned int linear_backoff_thresh
, start_ts
;
133 unsigned int rto_base
= syn_set
? TCP_TIMEOUT_INIT
: TCP_RTO_MIN
;
135 if (!inet_csk(sk
)->icsk_retransmits
)
138 if (unlikely(!tcp_sk(sk
)->retrans_stamp
))
139 start_ts
= TCP_SKB_CB(tcp_write_queue_head(sk
))->when
;
141 start_ts
= tcp_sk(sk
)->retrans_stamp
;
143 if (likely(timeout
== 0)) {
144 linear_backoff_thresh
= ilog2(TCP_RTO_MAX
/rto_base
);
146 if (boundary
<= linear_backoff_thresh
)
147 timeout
= ((2 << boundary
) - 1) * rto_base
;
149 timeout
= ((2 << linear_backoff_thresh
) - 1) * rto_base
+
150 (boundary
- linear_backoff_thresh
) * TCP_RTO_MAX
;
152 return (tcp_time_stamp
- start_ts
) >= timeout
;
155 /* A write timeout has occurred. Process the after effects. */
156 static int tcp_write_timeout(struct sock
*sk
)
158 struct inet_connection_sock
*icsk
= inet_csk(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 retry_until
= icsk
->icsk_syn_retries
? : sysctl_tcp_syn_retries
;
168 if (retransmits_timed_out(sk
, sysctl_tcp_retries1
, 0, 0)) {
169 /* Black hole detection */
170 tcp_mtu_probing(icsk
, sk
);
172 dst_negative_advice(sk
);
175 retry_until
= sysctl_tcp_retries2
;
176 if (sock_flag(sk
, SOCK_DEAD
)) {
177 const int alive
= (icsk
->icsk_rto
< TCP_RTO_MAX
);
179 retry_until
= tcp_orphan_retries(sk
, alive
);
181 !retransmits_timed_out(sk
, retry_until
, 0, 0);
183 if (tcp_out_of_resources(sk
, do_reset
))
188 if (retransmits_timed_out(sk
, retry_until
,
189 syn_set
? 0 : icsk
->icsk_user_timeout
, syn_set
)) {
190 /* Has it gone just too far? */
197 void tcp_delack_timer_handler(struct sock
*sk
)
199 struct tcp_sock
*tp
= tcp_sk(sk
);
200 struct inet_connection_sock
*icsk
= inet_csk(sk
);
202 sk_mem_reclaim_partial(sk
);
204 if (sk
->sk_state
== TCP_CLOSE
|| !(icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
))
207 if (time_after(icsk
->icsk_ack
.timeout
, jiffies
)) {
208 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, icsk
->icsk_ack
.timeout
);
211 icsk
->icsk_ack
.pending
&= ~ICSK_ACK_TIMER
;
213 if (!skb_queue_empty(&tp
->ucopy
.prequeue
)) {
216 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSCHEDULERFAILED
);
218 while ((skb
= __skb_dequeue(&tp
->ucopy
.prequeue
)) != NULL
)
219 sk_backlog_rcv(sk
, skb
);
221 tp
->ucopy
.memory
= 0;
224 if (inet_csk_ack_scheduled(sk
)) {
225 if (!icsk
->icsk_ack
.pingpong
) {
226 /* Delayed ACK missed: inflate ATO. */
227 icsk
->icsk_ack
.ato
= min(icsk
->icsk_ack
.ato
<< 1, icsk
->icsk_rto
);
229 /* Delayed ACK missed: leave pingpong mode and
232 icsk
->icsk_ack
.pingpong
= 0;
233 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
236 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKS
);
240 if (sk_under_memory_pressure(sk
))
244 static void tcp_delack_timer(unsigned long data
)
246 struct sock
*sk
= (struct sock
*)data
;
249 if (!sock_owned_by_user(sk
)) {
250 tcp_delack_timer_handler(sk
);
252 inet_csk(sk
)->icsk_ack
.blocked
= 1;
253 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOCKED
);
254 /* deleguate our work to tcp_release_cb() */
255 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED
, &tcp_sk(sk
)->tsq_flags
))
262 static void tcp_probe_timer(struct sock
*sk
)
264 struct inet_connection_sock
*icsk
= inet_csk(sk
);
265 struct tcp_sock
*tp
= tcp_sk(sk
);
268 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
269 icsk
->icsk_probes_out
= 0;
273 /* *WARNING* RFC 1122 forbids this
275 * It doesn't AFAIK, because we kill the retransmit timer -AK
277 * FIXME: We ought not to do it, Solaris 2.5 actually has fixing
278 * this behaviour in Solaris down as a bug fix. [AC]
280 * Let me to explain. icsk_probes_out is zeroed by incoming ACKs
281 * even if they advertise zero window. Hence, connection is killed only
282 * if we received no ACKs for normal connection timeout. It is not killed
283 * only because window stays zero for some time, window may be zero
284 * until armageddon and even later. We are in full accordance
285 * with RFCs, only probe timer combines both retransmission timeout
286 * and probe timeout in one bottle. --ANK
288 max_probes
= sysctl_tcp_retries2
;
290 if (sock_flag(sk
, SOCK_DEAD
)) {
291 const int alive
= ((icsk
->icsk_rto
<< icsk
->icsk_backoff
) < TCP_RTO_MAX
);
293 max_probes
= tcp_orphan_retries(sk
, alive
);
295 if (tcp_out_of_resources(sk
, alive
|| icsk
->icsk_probes_out
<= max_probes
))
299 if (icsk
->icsk_probes_out
> max_probes
) {
302 /* Only send another probe if we didn't close things up. */
308 * Timer for Fast Open socket to retransmit SYNACK. Note that the
309 * sk here is the child socket, not the parent (listener) socket.
311 static void tcp_fastopen_synack_timer(struct sock
*sk
)
313 struct inet_connection_sock
*icsk
= inet_csk(sk
);
314 int max_retries
= icsk
->icsk_syn_retries
? :
315 sysctl_tcp_synack_retries
+ 1; /* add one more retry for fastopen */
316 struct request_sock
*req
;
318 req
= tcp_sk(sk
)->fastopen_rsk
;
319 req
->rsk_ops
->syn_ack_timeout(sk
, req
);
321 if (req
->num_timeout
>= max_retries
) {
325 /* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
326 * returned from rtx_syn_ack() to make it more persistent like
327 * regular retransmit because if the child socket has been accepted
328 * it's not good to give up too easily.
330 inet_rtx_syn_ack(sk
, req
);
332 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
333 TCP_TIMEOUT_INIT
<< req
->num_timeout
, TCP_RTO_MAX
);
337 * The TCP retransmit timer.
340 void tcp_retransmit_timer(struct sock
*sk
)
342 struct tcp_sock
*tp
= tcp_sk(sk
);
343 struct inet_connection_sock
*icsk
= inet_csk(sk
);
345 if (tp
->fastopen_rsk
) {
346 WARN_ON_ONCE(sk
->sk_state
!= TCP_SYN_RECV
&&
347 sk
->sk_state
!= TCP_FIN_WAIT1
);
348 tcp_fastopen_synack_timer(sk
);
349 /* Before we receive ACK to our SYN-ACK don't retransmit
350 * anything else (e.g., data or FIN segments).
354 if (!tp
->packets_out
)
357 WARN_ON(tcp_write_queue_empty(sk
));
359 tp
->tlp_high_seq
= 0;
361 if (!tp
->snd_wnd
&& !sock_flag(sk
, SOCK_DEAD
) &&
362 !((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))) {
363 /* Receiver dastardly shrinks window. Our retransmits
364 * become zero probes, but we should not timeout this
365 * connection. If the socket is an orphan, time it out,
366 * we cannot allow such beasts to hang infinitely.
368 struct inet_sock
*inet
= inet_sk(sk
);
369 if (sk
->sk_family
== AF_INET
) {
370 LIMIT_NETDEBUG(KERN_DEBUG
pr_fmt("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n"),
372 ntohs(inet
->inet_dport
), inet
->inet_num
,
373 tp
->snd_una
, tp
->snd_nxt
);
375 #if IS_ENABLED(CONFIG_IPV6)
376 else if (sk
->sk_family
== AF_INET6
) {
377 struct ipv6_pinfo
*np
= inet6_sk(sk
);
378 LIMIT_NETDEBUG(KERN_DEBUG
pr_fmt("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n"),
380 ntohs(inet
->inet_dport
), inet
->inet_num
,
381 tp
->snd_una
, tp
->snd_nxt
);
384 if (tcp_time_stamp
- tp
->rcv_tstamp
> TCP_RTO_MAX
) {
388 tcp_enter_loss(sk
, 0);
389 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
391 goto out_reset_timer
;
394 if (tcp_write_timeout(sk
))
397 if (icsk
->icsk_retransmits
== 0) {
400 if (icsk
->icsk_ca_state
== TCP_CA_Recovery
) {
402 mib_idx
= LINUX_MIB_TCPSACKRECOVERYFAIL
;
404 mib_idx
= LINUX_MIB_TCPRENORECOVERYFAIL
;
405 } else if (icsk
->icsk_ca_state
== TCP_CA_Loss
) {
406 mib_idx
= LINUX_MIB_TCPLOSSFAILURES
;
407 } else if ((icsk
->icsk_ca_state
== TCP_CA_Disorder
) ||
410 mib_idx
= LINUX_MIB_TCPSACKFAILURES
;
412 mib_idx
= LINUX_MIB_TCPRENOFAILURES
;
414 mib_idx
= LINUX_MIB_TCPTIMEOUTS
;
416 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
419 tcp_enter_loss(sk
, 0);
421 if (tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
)) > 0) {
422 /* Retransmission failed because of local congestion,
425 if (!icsk
->icsk_retransmits
)
426 icsk
->icsk_retransmits
= 1;
427 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
428 min(icsk
->icsk_rto
, TCP_RESOURCE_PROBE_INTERVAL
),
433 /* Increase the timeout each time we retransmit. Note that
434 * we do not increase the rtt estimate. rto is initialized
435 * from rtt, but increases here. Jacobson (SIGCOMM 88) suggests
436 * that doubling rto each time is the least we can get away with.
437 * In KA9Q, Karn uses this for the first few times, and then
438 * goes to quadratic. netBSD doubles, but only goes up to *64,
439 * and clamps at 1 to 64 sec afterwards. Note that 120 sec is
440 * defined in the protocol as the maximum possible RTT. I guess
441 * we'll have to use something other than TCP to talk to the
442 * University of Mars.
444 * PAWS allows us longer timeouts and large windows, so once
445 * implemented ftp to mars will work nicely. We will have to fix
446 * the 120 second clamps though!
448 icsk
->icsk_backoff
++;
449 icsk
->icsk_retransmits
++;
452 /* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
453 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
454 * might be increased if the stream oscillates between thin and thick,
455 * thus the old value might already be too high compared to the value
456 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
457 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
458 * exponential backoff behaviour to avoid continue hammering
459 * linear-timeout retransmissions into a black hole
461 if (sk
->sk_state
== TCP_ESTABLISHED
&&
462 (tp
->thin_lto
|| sysctl_tcp_thin_linear_timeouts
) &&
463 tcp_stream_is_thin(tp
) &&
464 icsk
->icsk_retransmits
<= TCP_THIN_LINEAR_RETRIES
) {
465 icsk
->icsk_backoff
= 0;
466 icsk
->icsk_rto
= min(__tcp_set_rto(tp
), TCP_RTO_MAX
);
468 /* Use normal (exponential) backoff */
469 icsk
->icsk_rto
= min(icsk
->icsk_rto
<< 1, TCP_RTO_MAX
);
471 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, icsk
->icsk_rto
, TCP_RTO_MAX
);
472 if (retransmits_timed_out(sk
, sysctl_tcp_retries1
+ 1, 0, 0))
478 void tcp_write_timer_handler(struct sock
*sk
)
480 struct inet_connection_sock
*icsk
= inet_csk(sk
);
483 if (sk
->sk_state
== TCP_CLOSE
|| !icsk
->icsk_pending
)
486 if (time_after(icsk
->icsk_timeout
, jiffies
)) {
487 sk_reset_timer(sk
, &icsk
->icsk_retransmit_timer
, icsk
->icsk_timeout
);
491 event
= icsk
->icsk_pending
;
494 case ICSK_TIME_EARLY_RETRANS
:
495 tcp_resume_early_retransmit(sk
);
497 case ICSK_TIME_LOSS_PROBE
:
498 tcp_send_loss_probe(sk
);
500 case ICSK_TIME_RETRANS
:
501 icsk
->icsk_pending
= 0;
502 tcp_retransmit_timer(sk
);
504 case ICSK_TIME_PROBE0
:
505 icsk
->icsk_pending
= 0;
514 static void tcp_write_timer(unsigned long data
)
516 struct sock
*sk
= (struct sock
*)data
;
519 if (!sock_owned_by_user(sk
)) {
520 tcp_write_timer_handler(sk
);
522 /* deleguate our work to tcp_release_cb() */
523 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED
, &tcp_sk(sk
)->tsq_flags
))
531 * Timer for listening sockets
534 static void tcp_synack_timer(struct sock
*sk
)
536 inet_csk_reqsk_queue_prune(sk
, TCP_SYNQ_INTERVAL
,
537 TCP_TIMEOUT_INIT
, TCP_RTO_MAX
);
540 void tcp_syn_ack_timeout(struct sock
*sk
, struct request_sock
*req
)
542 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPTIMEOUTS
);
544 EXPORT_SYMBOL(tcp_syn_ack_timeout
);
546 void tcp_set_keepalive(struct sock
*sk
, int val
)
548 if ((1 << sk
->sk_state
) & (TCPF_CLOSE
| TCPF_LISTEN
))
551 if (val
&& !sock_flag(sk
, SOCK_KEEPOPEN
))
552 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tcp_sk(sk
)));
554 inet_csk_delete_keepalive_timer(sk
);
558 static void tcp_keepalive_timer (unsigned long data
)
560 struct sock
*sk
= (struct sock
*) data
;
561 struct inet_connection_sock
*icsk
= inet_csk(sk
);
562 struct tcp_sock
*tp
= tcp_sk(sk
);
565 /* Only process if socket is not in use. */
567 if (sock_owned_by_user(sk
)) {
568 /* Try again later. */
569 inet_csk_reset_keepalive_timer (sk
, HZ
/20);
573 if (sk
->sk_state
== TCP_LISTEN
) {
574 tcp_synack_timer(sk
);
578 if (sk
->sk_state
== TCP_FIN_WAIT2
&& sock_flag(sk
, SOCK_DEAD
)) {
579 if (tp
->linger2
>= 0) {
580 const int tmo
= tcp_fin_time(sk
) - TCP_TIMEWAIT_LEN
;
583 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
587 tcp_send_active_reset(sk
, GFP_ATOMIC
);
591 if (!sock_flag(sk
, SOCK_KEEPOPEN
) || sk
->sk_state
== TCP_CLOSE
)
594 elapsed
= keepalive_time_when(tp
);
596 /* It is alive without keepalive 8) */
597 if (tp
->packets_out
|| tcp_send_head(sk
))
600 elapsed
= keepalive_time_elapsed(tp
);
602 if (elapsed
>= keepalive_time_when(tp
)) {
603 /* If the TCP_USER_TIMEOUT option is enabled, use that
604 * to determine when to timeout instead.
606 if ((icsk
->icsk_user_timeout
!= 0 &&
607 elapsed
>= icsk
->icsk_user_timeout
&&
608 icsk
->icsk_probes_out
> 0) ||
609 (icsk
->icsk_user_timeout
== 0 &&
610 icsk
->icsk_probes_out
>= keepalive_probes(tp
))) {
611 tcp_send_active_reset(sk
, GFP_ATOMIC
);
615 if (tcp_write_wakeup(sk
) <= 0) {
616 icsk
->icsk_probes_out
++;
617 elapsed
= keepalive_intvl_when(tp
);
619 /* If keepalive was lost due to local congestion,
622 elapsed
= TCP_RESOURCE_PROBE_INTERVAL
;
625 /* It is tp->rcv_tstamp + keepalive_time_when(tp) */
626 elapsed
= keepalive_time_when(tp
) - elapsed
;
632 inet_csk_reset_keepalive_timer (sk
, elapsed
);
643 void tcp_init_xmit_timers(struct sock
*sk
)
645 inet_csk_init_xmit_timers(sk
, &tcp_write_timer
, &tcp_delack_timer
,
646 &tcp_keepalive_timer
);
648 EXPORT_SYMBOL(tcp_init_xmit_timers
);