blob: b78aac30c4986dc78fcd8dae03a527326baa32a3 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
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.
5 *
6 * Implementation of the Transmission Control Protocol(TCP).
7 *
Jesper Juhl02c30a82005-05-05 16:16:16 -07008 * Authors: Ross Biro
Linus Torvalds1da177e2005-04-16 15:20:36 -07009 * 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>
19 */
20
21#include <linux/module.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090022#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070023#include <net/tcp.h>
24
Brian Haleyab32ea52006-09-22 14:15:41 -070025int sysctl_tcp_syn_retries __read_mostly = TCP_SYN_RETRIES;
26int sysctl_tcp_synack_retries __read_mostly = TCP_SYNACK_RETRIES;
27int sysctl_tcp_keepalive_time __read_mostly = TCP_KEEPALIVE_TIME;
28int sysctl_tcp_keepalive_probes __read_mostly = TCP_KEEPALIVE_PROBES;
29int sysctl_tcp_keepalive_intvl __read_mostly = TCP_KEEPALIVE_INTVL;
30int sysctl_tcp_retries1 __read_mostly = TCP_RETR1;
31int sysctl_tcp_retries2 __read_mostly = TCP_RETR2;
32int sysctl_tcp_orphan_retries __read_mostly;
Andreas Petlund36e31b0a2010-02-18 02:47:01 +000033int sysctl_tcp_thin_linear_timeouts __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -070034
Linus Torvalds1da177e2005-04-16 15:20:36 -070035static void tcp_write_err(struct sock *sk)
36{
37 sk->sk_err = sk->sk_err_soft ? : ETIMEDOUT;
38 sk->sk_error_report(sk);
39
40 tcp_done(sk);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -070041 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT);
Linus Torvalds1da177e2005-04-16 15:20:36 -070042}
43
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.
48 *
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -080049 * Criteria is still not confirmed experimentally and may change.
Linus Torvalds1da177e2005-04-16 15:20:36 -070050 * We kill the socket, if:
51 * 1. If number of orphaned sockets exceeds an administratively configured
52 * limit.
53 * 2. If we have strong memory pressure.
54 */
55static int tcp_out_of_resources(struct sock *sk, int do_reset)
56{
57 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerad1af0f2010-08-25 02:27:49 -070058 int shift = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -070059
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +090060 /* If peer does not open window for long time, or did not transmit
Linus Torvalds1da177e2005-04-16 15:20:36 -070061 * anything for long time, penalize it. */
62 if ((s32)(tcp_time_stamp - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset)
David S. Millerad1af0f2010-08-25 02:27:49 -070063 shift++;
Linus Torvalds1da177e2005-04-16 15:20:36 -070064
65 /* If some dubious ICMP arrived, penalize even more. */
66 if (sk->sk_err_soft)
David S. Millerad1af0f2010-08-25 02:27:49 -070067 shift++;
Linus Torvalds1da177e2005-04-16 15:20:36 -070068
Arun Sharmaefcdbf22012-01-30 14:16:06 -080069 if (tcp_check_oom(sk, shift)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -070070 /* 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))
75 do_reset = 1;
76 if (do_reset)
77 tcp_send_active_reset(sk, GFP_ATOMIC);
78 tcp_done(sk);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -070079 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY);
Linus Torvalds1da177e2005-04-16 15:20:36 -070080 return 1;
81 }
82 return 0;
83}
84
85/* Calculate maximal number or retries on an orphaned socket. */
86static int tcp_orphan_retries(struct sock *sk, int alive)
87{
88 int retries = sysctl_tcp_orphan_retries; /* May be zero. */
89
90 /* We know from an ICMP that something is wrong. */
91 if (sk->sk_err_soft && !alive)
92 retries = 0;
93
94 /* However, if socket sent something recently, select some safe
95 * number of retries. 8 corresponds to >100 seconds with minimal
96 * RTO of 200msec. */
97 if (retries == 0 && alive)
98 retries = 8;
99 return retries;
100}
101
Eric Dumazetce55dd32007-12-21 01:50:43 -0800102static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk)
103{
Eric Dumazetce55dd32007-12-21 01:50:43 -0800104 /* 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);
109 } else {
110 struct tcp_sock *tp = tcp_sk(sk);
David S. Miller829942c2007-12-21 04:29:16 -0800111 int mss;
112
Eric Dumazet8beb5c52007-12-21 05:58:29 -0800113 mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1;
Eric Dumazetce55dd32007-12-21 01:50:43 -0800114 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);
118 }
119 }
120}
121
Damian Lukowski2f7de572009-12-07 06:06:16 +0000122/* This function calculates a "timeout" which is equivalent to the timeout of a
Daniel Mack3ad2f3f2010-02-03 08:01:28 +0800123 * TCP connection after "boundary" unsuccessful, exponentially backed-off
Damian Lukowski4d22f7d2010-09-28 13:08:32 -0700124 * retransmissions with an initial RTO of TCP_RTO_MIN or TCP_TIMEOUT_INIT if
125 * syn_set flag is set.
Damian Lukowski2f7de572009-12-07 06:06:16 +0000126 */
127static bool retransmits_timed_out(struct sock *sk,
Jerry Chudca43c72010-08-27 19:13:28 +0000128 unsigned int boundary,
David S. Miller21a180c2010-10-04 11:56:38 -0700129 unsigned int timeout,
Damian Lukowski4d22f7d2010-09-28 13:08:32 -0700130 bool syn_set)
Damian Lukowski2f7de572009-12-07 06:06:16 +0000131{
Jerry Chudca43c72010-08-27 19:13:28 +0000132 unsigned int linear_backoff_thresh, start_ts;
Damian Lukowski4d22f7d2010-09-28 13:08:32 -0700133 unsigned int rto_base = syn_set ? TCP_TIMEOUT_INIT : TCP_RTO_MIN;
Damian Lukowski2f7de572009-12-07 06:06:16 +0000134
135 if (!inet_csk(sk)->icsk_retransmits)
136 return false;
137
138 if (unlikely(!tcp_sk(sk)->retrans_stamp))
139 start_ts = TCP_SKB_CB(tcp_write_queue_head(sk))->when;
140 else
141 start_ts = tcp_sk(sk)->retrans_stamp;
142
Jerry Chudca43c72010-08-27 19:13:28 +0000143 if (likely(timeout == 0)) {
David S. Miller21a180c2010-10-04 11:56:38 -0700144 linear_backoff_thresh = ilog2(TCP_RTO_MAX/rto_base);
Damian Lukowski2f7de572009-12-07 06:06:16 +0000145
Jerry Chudca43c72010-08-27 19:13:28 +0000146 if (boundary <= linear_backoff_thresh)
David S. Miller21a180c2010-10-04 11:56:38 -0700147 timeout = ((2 << boundary) - 1) * rto_base;
Jerry Chudca43c72010-08-27 19:13:28 +0000148 else
David S. Miller21a180c2010-10-04 11:56:38 -0700149 timeout = ((2 << linear_backoff_thresh) - 1) * rto_base +
Jerry Chudca43c72010-08-27 19:13:28 +0000150 (boundary - linear_backoff_thresh) * TCP_RTO_MAX;
151 }
Damian Lukowski2f7de572009-12-07 06:06:16 +0000152 return (tcp_time_stamp - start_ts) >= timeout;
153}
154
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155/* A write timeout has occurred. Process the after effects. */
156static int tcp_write_timeout(struct sock *sk)
157{
John Heffner5d424d52006-03-20 17:53:41 -0800158 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159 int retry_until;
Rusty Russell3db1cd52011-12-19 13:56:45 +0000160 bool do_reset, syn_set = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700161
162 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700163 if (icsk->icsk_retransmits)
Eric Dumazetb6c67122010-04-08 23:03:29 +0000164 dst_negative_advice(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700165 retry_until = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
Rusty Russell3db1cd52011-12-19 13:56:45 +0000166 syn_set = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700167 } else {
David S. Miller21a180c2010-10-04 11:56:38 -0700168 if (retransmits_timed_out(sk, sysctl_tcp_retries1, 0, 0)) {
John Heffner5d424d52006-03-20 17:53:41 -0800169 /* Black hole detection */
Eric Dumazetce55dd32007-12-21 01:50:43 -0800170 tcp_mtu_probing(icsk, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700171
Eric Dumazetb6c67122010-04-08 23:03:29 +0000172 dst_negative_advice(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700173 }
174
175 retry_until = sysctl_tcp_retries2;
176 if (sock_flag(sk, SOCK_DEAD)) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700177 const int alive = (icsk->icsk_rto < TCP_RTO_MAX);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900178
Linus Torvalds1da177e2005-04-16 15:20:36 -0700179 retry_until = tcp_orphan_retries(sk, alive);
Damian Lukowski6fa12c82009-08-26 00:16:34 +0000180 do_reset = alive ||
David S. Miller21a180c2010-10-04 11:56:38 -0700181 !retransmits_timed_out(sk, retry_until, 0, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182
Damian Lukowski6fa12c82009-08-26 00:16:34 +0000183 if (tcp_out_of_resources(sk, do_reset))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184 return 1;
185 }
186 }
187
Jerry Chudca43c72010-08-27 19:13:28 +0000188 if (retransmits_timed_out(sk, retry_until,
David S. Miller21a180c2010-10-04 11:56:38 -0700189 syn_set ? 0 : icsk->icsk_user_timeout, syn_set)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700190 /* Has it gone just too far? */
191 tcp_write_err(sk);
192 return 1;
193 }
194 return 0;
195}
196
Eric Dumazet6f458df2012-07-20 05:45:50 +0000197void tcp_delack_timer_handler(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700198{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700199 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700200 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700201
David S. Miller9993e7d2008-01-10 21:56:38 -0800202 sk_mem_reclaim_partial(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700204 if (sk->sk_state == TCP_CLOSE || !(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700205 goto out;
206
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700207 if (time_after(icsk->icsk_ack.timeout, jiffies)) {
208 sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700209 goto out;
210 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700211 icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700212
David S. Millerb03efcf2005-07-08 14:57:23 -0700213 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214 struct sk_buff *skb;
215
Pavel Emelyanovde0744a2008-07-16 20:31:16 -0700216 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSCHEDULERFAILED);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700217
218 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
Peter Zijlstrac57943a2008-10-07 14:18:42 -0700219 sk_backlog_rcv(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700220
221 tp->ucopy.memory = 0;
222 }
223
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700224 if (inet_csk_ack_scheduled(sk)) {
225 if (!icsk->icsk_ack.pingpong) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700226 /* Delayed ACK missed: inflate ATO. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700227 icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228 } else {
229 /* Delayed ACK missed: leave pingpong mode and
230 * deflate ATO.
231 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700232 icsk->icsk_ack.pingpong = 0;
233 icsk->icsk_ack.ato = TCP_ATO_MIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700234 }
235 tcp_send_ack(sk);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -0700236 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700237 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700238
239out:
Glauber Costa180d8cd2011-12-11 21:47:02 +0000240 if (sk_under_memory_pressure(sk))
Hideo Aoki3ab224b2007-12-31 00:11:19 -0800241 sk_mem_reclaim(sk);
Eric Dumazet6f458df2012-07-20 05:45:50 +0000242}
243
244static void tcp_delack_timer(unsigned long data)
245{
246 struct sock *sk = (struct sock *)data;
247
248 bh_lock_sock(sk);
249 if (!sock_owned_by_user(sk)) {
250 tcp_delack_timer_handler(sk);
251 } else {
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() */
Eric Dumazet144d56e2012-08-20 00:22:46 +0000255 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &tcp_sk(sk)->tsq_flags))
256 sock_hold(sk);
Eric Dumazet6f458df2012-07-20 05:45:50 +0000257 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700258 bh_unlock_sock(sk);
259 sock_put(sk);
260}
261
262static void tcp_probe_timer(struct sock *sk)
263{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300264 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700265 struct tcp_sock *tp = tcp_sk(sk);
266 int max_probes;
267
David S. Millerfe067e82007-03-07 12:12:44 -0800268 if (tp->packets_out || !tcp_send_head(sk)) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300269 icsk->icsk_probes_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700270 return;
271 }
272
273 /* *WARNING* RFC 1122 forbids this
274 *
275 * It doesn't AFAIK, because we kill the retransmit timer -AK
276 *
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]
279 *
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300280 * Let me to explain. icsk_probes_out is zeroed by incoming ACKs
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281 * 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
287 */
288 max_probes = sysctl_tcp_retries2;
289
290 if (sock_flag(sk, SOCK_DEAD)) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700291 const int alive = ((icsk->icsk_rto << icsk->icsk_backoff) < TCP_RTO_MAX);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900292
Linus Torvalds1da177e2005-04-16 15:20:36 -0700293 max_probes = tcp_orphan_retries(sk, alive);
294
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300295 if (tcp_out_of_resources(sk, alive || icsk->icsk_probes_out <= max_probes))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296 return;
297 }
298
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300299 if (icsk->icsk_probes_out > max_probes) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300 tcp_write_err(sk);
301 } else {
302 /* Only send another probe if we didn't close things up. */
303 tcp_send_probe0(sk);
304 }
305}
306
307/*
Jerry Chu83368862012-08-31 12:29:12 +0000308 * Timer for Fast Open socket to retransmit SYNACK. Note that the
309 * sk here is the child socket, not the parent (listener) socket.
310 */
311static void tcp_fastopen_synack_timer(struct sock *sk)
312{
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;
317
318 req = tcp_sk(sk)->fastopen_rsk;
319 req->rsk_ops->syn_ack_timeout(sk, req);
320
Eric Dumazete6c022a2012-10-27 23:16:46 +0000321 if (req->num_timeout >= max_retries) {
Jerry Chu83368862012-08-31 12:29:12 +0000322 tcp_write_err(sk);
323 return;
324 }
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.
329 */
Eric Dumazete6c022a2012-10-27 23:16:46 +0000330 inet_rtx_syn_ack(sk, req);
331 req->num_timeout++;
Jerry Chu83368862012-08-31 12:29:12 +0000332 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
Eric Dumazete6c022a2012-10-27 23:16:46 +0000333 TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
Jerry Chu83368862012-08-31 12:29:12 +0000334}
335
336/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700337 * The TCP retransmit timer.
338 */
339
Damian Lukowskif1ecd5d2009-08-26 00:16:31 +0000340void tcp_retransmit_timer(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700341{
342 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700343 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700344
Yuchung Cheng750ea2b2012-05-02 13:30:04 +0000345 if (tp->early_retrans_delayed) {
346 tcp_resume_early_retransmit(sk);
347 return;
348 }
Jerry Chu83368862012-08-31 12:29:12 +0000349 if (tp->fastopen_rsk) {
Jerry Chu37561f62012-10-22 11:26:36 +0000350 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
351 sk->sk_state != TCP_FIN_WAIT1);
Jerry Chu83368862012-08-31 12:29:12 +0000352 tcp_fastopen_synack_timer(sk);
353 /* Before we receive ACK to our SYN-ACK don't retransmit
354 * anything else (e.g., data or FIN segments).
355 */
356 return;
357 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700358 if (!tp->packets_out)
359 goto out;
360
Ilpo Järvinen547b7922008-07-25 21:43:18 -0700361 WARN_ON(tcp_write_queue_empty(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700362
363 if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
364 !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
365 /* Receiver dastardly shrinks window. Our retransmits
366 * become zero probes, but we should not timeout this
367 * connection. If the socket is an orphan, time it out,
368 * we cannot allow such beasts to hang infinitely.
369 */
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -0700370 struct inet_sock *inet = inet_sk(sk);
371 if (sk->sk_family == AF_INET) {
Joe Perchesafd465032012-03-12 07:03:32 +0000372 LIMIT_NETDEBUG(KERN_DEBUG pr_fmt("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n"),
373 &inet->inet_daddr,
374 ntohs(inet->inet_dport), inet->inet_num,
375 tp->snd_una, tp->snd_nxt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376 }
Eric Dumazetdfd56b82011-12-10 09:48:31 +0000377#if IS_ENABLED(CONFIG_IPV6)
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -0700378 else if (sk->sk_family == AF_INET6) {
379 struct ipv6_pinfo *np = inet6_sk(sk);
Joe Perchesafd465032012-03-12 07:03:32 +0000380 LIMIT_NETDEBUG(KERN_DEBUG pr_fmt("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n"),
381 &np->daddr,
382 ntohs(inet->inet_dport), inet->inet_num,
383 tp->snd_una, tp->snd_nxt);
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -0700384 }
385#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386 if (tcp_time_stamp - tp->rcv_tstamp > TCP_RTO_MAX) {
387 tcp_write_err(sk);
388 goto out;
389 }
390 tcp_enter_loss(sk, 0);
David S. Millerfe067e82007-03-07 12:12:44 -0800391 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700392 __sk_dst_reset(sk);
393 goto out_reset_timer;
394 }
395
396 if (tcp_write_timeout(sk))
397 goto out;
398
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700399 if (icsk->icsk_retransmits == 0) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700400 int mib_idx;
401
Ilpo Järvinenc60ce4e2010-10-14 01:52:09 +0000402 if (icsk->icsk_ca_state == TCP_CA_Recovery) {
Ilpo Järvinenbc079e92009-02-28 04:44:34 +0000403 if (tcp_is_sack(tp))
404 mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
405 else
406 mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300407 } else if (icsk->icsk_ca_state == TCP_CA_Loss) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700408 mib_idx = LINUX_MIB_TCPLOSSFAILURES;
Ilpo Järvinenc60ce4e2010-10-14 01:52:09 +0000409 } else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
410 tp->sacked_out) {
411 if (tcp_is_sack(tp))
412 mib_idx = LINUX_MIB_TCPSACKFAILURES;
413 else
414 mib_idx = LINUX_MIB_TCPRENOFAILURES;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415 } else {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700416 mib_idx = LINUX_MIB_TCPTIMEOUTS;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417 }
Pavel Emelyanovde0744a2008-07-16 20:31:16 -0700418 NET_INC_STATS_BH(sock_net(sk), mib_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419 }
420
421 if (tcp_use_frto(sk)) {
422 tcp_enter_frto(sk);
423 } else {
424 tcp_enter_loss(sk, 0);
425 }
426
David S. Millerfe067e82007-03-07 12:12:44 -0800427 if (tcp_retransmit_skb(sk, tcp_write_queue_head(sk)) > 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428 /* Retransmission failed because of local congestion,
429 * do not backoff.
430 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700431 if (!icsk->icsk_retransmits)
432 icsk->icsk_retransmits = 1;
433 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -0700434 min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL),
435 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700436 goto out;
437 }
438
439 /* Increase the timeout each time we retransmit. Note that
440 * we do not increase the rtt estimate. rto is initialized
441 * from rtt, but increases here. Jacobson (SIGCOMM 88) suggests
442 * that doubling rto each time is the least we can get away with.
443 * In KA9Q, Karn uses this for the first few times, and then
444 * goes to quadratic. netBSD doubles, but only goes up to *64,
445 * and clamps at 1 to 64 sec afterwards. Note that 120 sec is
446 * defined in the protocol as the maximum possible RTT. I guess
447 * we'll have to use something other than TCP to talk to the
448 * University of Mars.
449 *
450 * PAWS allows us longer timeouts and large windows, so once
451 * implemented ftp to mars will work nicely. We will have to fix
452 * the 120 second clamps though!
453 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700454 icsk->icsk_backoff++;
455 icsk->icsk_retransmits++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456
457out_reset_timer:
Andreas Petlund36e31b0a2010-02-18 02:47:01 +0000458 /* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
459 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
460 * might be increased if the stream oscillates between thin and thick,
461 * thus the old value might already be too high compared to the value
462 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
463 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
464 * exponential backoff behaviour to avoid continue hammering
465 * linear-timeout retransmissions into a black hole
466 */
467 if (sk->sk_state == TCP_ESTABLISHED &&
468 (tp->thin_lto || sysctl_tcp_thin_linear_timeouts) &&
469 tcp_stream_is_thin(tp) &&
470 icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
471 icsk->icsk_backoff = 0;
472 icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX);
473 } else {
474 /* Use normal (exponential) backoff */
475 icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
476 }
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -0700477 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX);
David S. Miller21a180c2010-10-04 11:56:38 -0700478 if (retransmits_timed_out(sk, sysctl_tcp_retries1 + 1, 0, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700479 __sk_dst_reset(sk);
480
481out:;
482}
483
Eric Dumazet6f458df2012-07-20 05:45:50 +0000484void tcp_write_timer_handler(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700486 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700487 int event;
488
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700489 if (sk->sk_state == TCP_CLOSE || !icsk->icsk_pending)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490 goto out;
491
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700492 if (time_after(icsk->icsk_timeout, jiffies)) {
493 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700494 goto out;
495 }
496
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700497 event = icsk->icsk_pending;
498 icsk->icsk_pending = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700499
500 switch (event) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700501 case ICSK_TIME_RETRANS:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700502 tcp_retransmit_timer(sk);
503 break;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700504 case ICSK_TIME_PROBE0:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700505 tcp_probe_timer(sk);
506 break;
507 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700508
509out:
Hideo Aoki3ab224b2007-12-31 00:11:19 -0800510 sk_mem_reclaim(sk);
Eric Dumazet6f458df2012-07-20 05:45:50 +0000511}
512
513static void tcp_write_timer(unsigned long data)
514{
515 struct sock *sk = (struct sock *)data;
516
517 bh_lock_sock(sk);
518 if (!sock_owned_by_user(sk)) {
519 tcp_write_timer_handler(sk);
520 } else {
521 /* deleguate our work to tcp_release_cb() */
Eric Dumazet144d56e2012-08-20 00:22:46 +0000522 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &tcp_sk(sk)->tsq_flags))
523 sock_hold(sk);
Eric Dumazet6f458df2012-07-20 05:45:50 +0000524 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700525 bh_unlock_sock(sk);
526 sock_put(sk);
527}
528
Arnaldo Carvalho de Melo295f7322005-08-09 20:11:56 -0700529/*
530 * Timer for listening sockets
531 */
532
533static void tcp_synack_timer(struct sock *sk)
534{
Arnaldo Carvalho de Meloa019d6f2005-08-09 20:15:09 -0700535 inet_csk_reqsk_queue_prune(sk, TCP_SYNQ_INTERVAL,
536 TCP_TIMEOUT_INIT, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700537}
538
Octavian Purdila72659ec2010-01-17 19:09:39 -0800539void tcp_syn_ack_timeout(struct sock *sk, struct request_sock *req)
540{
541 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPTIMEOUTS);
542}
543EXPORT_SYMBOL(tcp_syn_ack_timeout);
544
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545void tcp_set_keepalive(struct sock *sk, int val)
546{
547 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
548 return;
549
550 if (val && !sock_flag(sk, SOCK_KEEPOPEN))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700551 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700552 else if (!val)
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700553 inet_csk_delete_keepalive_timer(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700554}
555
556
557static void tcp_keepalive_timer (unsigned long data)
558{
559 struct sock *sk = (struct sock *) data;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300560 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700561 struct tcp_sock *tp = tcp_sk(sk);
Flavio Leitner6c37e5d2010-04-26 18:33:27 +0000562 u32 elapsed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700563
564 /* Only process if socket is not in use. */
565 bh_lock_sock(sk);
566 if (sock_owned_by_user(sk)) {
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900567 /* Try again later. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700568 inet_csk_reset_keepalive_timer (sk, HZ/20);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700569 goto out;
570 }
571
572 if (sk->sk_state == TCP_LISTEN) {
573 tcp_synack_timer(sk);
574 goto out;
575 }
576
577 if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
578 if (tp->linger2 >= 0) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700579 const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700580
581 if (tmo > 0) {
582 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
583 goto out;
584 }
585 }
586 tcp_send_active_reset(sk, GFP_ATOMIC);
587 goto death;
588 }
589
590 if (!sock_flag(sk, SOCK_KEEPOPEN) || sk->sk_state == TCP_CLOSE)
591 goto out;
592
593 elapsed = keepalive_time_when(tp);
594
595 /* It is alive without keepalive 8) */
David S. Millerfe067e82007-03-07 12:12:44 -0800596 if (tp->packets_out || tcp_send_head(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700597 goto resched;
598
Flavio Leitner6c37e5d2010-04-26 18:33:27 +0000599 elapsed = keepalive_time_elapsed(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700600
601 if (elapsed >= keepalive_time_when(tp)) {
Jerry Chudca43c72010-08-27 19:13:28 +0000602 /* If the TCP_USER_TIMEOUT option is enabled, use that
603 * to determine when to timeout instead.
604 */
605 if ((icsk->icsk_user_timeout != 0 &&
606 elapsed >= icsk->icsk_user_timeout &&
607 icsk->icsk_probes_out > 0) ||
608 (icsk->icsk_user_timeout == 0 &&
609 icsk->icsk_probes_out >= keepalive_probes(tp))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700610 tcp_send_active_reset(sk, GFP_ATOMIC);
611 tcp_write_err(sk);
612 goto out;
613 }
614 if (tcp_write_wakeup(sk) <= 0) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300615 icsk->icsk_probes_out++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700616 elapsed = keepalive_intvl_when(tp);
617 } else {
618 /* If keepalive was lost due to local congestion,
619 * try harder.
620 */
621 elapsed = TCP_RESOURCE_PROBE_INTERVAL;
622 }
623 } else {
624 /* It is tp->rcv_tstamp + keepalive_time_when(tp) */
625 elapsed = keepalive_time_when(tp) - elapsed;
626 }
627
Hideo Aoki3ab224b2007-12-31 00:11:19 -0800628 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700629
630resched:
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700631 inet_csk_reset_keepalive_timer (sk, elapsed);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700632 goto out;
633
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900634death:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700635 tcp_done(sk);
636
637out:
638 bh_unlock_sock(sk);
639 sock_put(sk);
640}
Eric Dumazet6f458df2012-07-20 05:45:50 +0000641
642void tcp_init_xmit_timers(struct sock *sk)
643{
644 inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
645 &tcp_keepalive_timer);
646}
647EXPORT_SYMBOL(tcp_init_xmit_timers);