blob: 8c65dc147d8bcfb58e14c20b774711ffbcc30d5a [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 */
Yuchung Chengb2482302014-09-29 13:20:38 -070055static int tcp_out_of_resources(struct sock *sk, bool do_reset)
Linus Torvalds1da177e2005-04-16 15:20:36 -070056{
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))
Yuchung Chengb2482302014-09-29 13:20:38 -070075 do_reset = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -070076 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{
Fan Dub0f9ca52015-02-10 09:53:16 +0800104 struct net *net = sock_net(sk);
105
Eric Dumazetce55dd32007-12-21 01:50:43 -0800106 /* Black hole detection */
Fan Dub0f9ca52015-02-10 09:53:16 +0800107 if (net->ipv4.sysctl_tcp_mtu_probing) {
Eric Dumazetce55dd32007-12-21 01:50:43 -0800108 if (!icsk->icsk_mtup.enabled) {
109 icsk->icsk_mtup.enabled = 1;
Fan Du05cbc0d2015-03-06 11:18:24 +0800110 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
Eric Dumazetce55dd32007-12-21 01:50:43 -0800111 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
112 } else {
Fan Dub0f9ca52015-02-10 09:53:16 +0800113 struct net *net = sock_net(sk);
Eric Dumazetce55dd32007-12-21 01:50:43 -0800114 struct tcp_sock *tp = tcp_sk(sk);
David S. Miller829942c2007-12-21 04:29:16 -0800115 int mss;
116
Eric Dumazet8beb5c52007-12-21 05:58:29 -0800117 mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1;
Fan Dub0f9ca52015-02-10 09:53:16 +0800118 mss = min(net->ipv4.sysctl_tcp_base_mss, mss);
Eric Dumazetce55dd32007-12-21 01:50:43 -0800119 mss = max(mss, 68 - tp->tcp_header_len);
120 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
121 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
122 }
123 }
124}
125
Damian Lukowski2f7de572009-12-07 06:06:16 +0000126/* This function calculates a "timeout" which is equivalent to the timeout of a
Daniel Mack3ad2f3f2010-02-03 08:01:28 +0800127 * TCP connection after "boundary" unsuccessful, exponentially backed-off
Damian Lukowski4d22f7d2010-09-28 13:08:32 -0700128 * retransmissions with an initial RTO of TCP_RTO_MIN or TCP_TIMEOUT_INIT if
129 * syn_set flag is set.
Damian Lukowski2f7de572009-12-07 06:06:16 +0000130 */
131static bool retransmits_timed_out(struct sock *sk,
Jerry Chudca43c72010-08-27 19:13:28 +0000132 unsigned int boundary,
David S. Miller21a180c2010-10-04 11:56:38 -0700133 unsigned int timeout,
Damian Lukowski4d22f7d2010-09-28 13:08:32 -0700134 bool syn_set)
Damian Lukowski2f7de572009-12-07 06:06:16 +0000135{
Jerry Chudca43c72010-08-27 19:13:28 +0000136 unsigned int linear_backoff_thresh, start_ts;
Damian Lukowski4d22f7d2010-09-28 13:08:32 -0700137 unsigned int rto_base = syn_set ? TCP_TIMEOUT_INIT : TCP_RTO_MIN;
Damian Lukowski2f7de572009-12-07 06:06:16 +0000138
139 if (!inet_csk(sk)->icsk_retransmits)
140 return false;
141
Eric Dumazet7faee5c2014-09-05 15:33:33 -0700142 start_ts = tcp_sk(sk)->retrans_stamp;
143 if (unlikely(!start_ts))
144 start_ts = tcp_skb_timestamp(tcp_write_queue_head(sk));
Damian Lukowski2f7de572009-12-07 06:06:16 +0000145
Jerry Chudca43c72010-08-27 19:13:28 +0000146 if (likely(timeout == 0)) {
David S. Miller21a180c2010-10-04 11:56:38 -0700147 linear_backoff_thresh = ilog2(TCP_RTO_MAX/rto_base);
Damian Lukowski2f7de572009-12-07 06:06:16 +0000148
Jerry Chudca43c72010-08-27 19:13:28 +0000149 if (boundary <= linear_backoff_thresh)
David S. Miller21a180c2010-10-04 11:56:38 -0700150 timeout = ((2 << boundary) - 1) * rto_base;
Jerry Chudca43c72010-08-27 19:13:28 +0000151 else
David S. Miller21a180c2010-10-04 11:56:38 -0700152 timeout = ((2 << linear_backoff_thresh) - 1) * rto_base +
Jerry Chudca43c72010-08-27 19:13:28 +0000153 (boundary - linear_backoff_thresh) * TCP_RTO_MAX;
154 }
Damian Lukowski2f7de572009-12-07 06:06:16 +0000155 return (tcp_time_stamp - start_ts) >= timeout;
156}
157
Linus Torvalds1da177e2005-04-16 15:20:36 -0700158/* A write timeout has occurred. Process the after effects. */
159static int tcp_write_timeout(struct sock *sk)
160{
John Heffner5d424d52006-03-20 17:53:41 -0800161 struct inet_connection_sock *icsk = inet_csk(sk);
Yuchung Chengc9686012013-10-29 10:09:05 -0700162 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700163 int retry_until;
Rusty Russell3db1cd52011-12-19 13:56:45 +0000164 bool do_reset, syn_set = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700165
166 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
Yuchung Chengc9686012013-10-29 10:09:05 -0700167 if (icsk->icsk_retransmits) {
Eric Dumazetb6c67122010-04-08 23:03:29 +0000168 dst_negative_advice(sk);
Yuchung Chengc9686012013-10-29 10:09:05 -0700169 if (tp->syn_fastopen || tp->syn_data)
Daniel Lee2646c832015-04-06 14:37:27 -0700170 tcp_fastopen_cache_set(sk, 0, NULL, true, 0);
Yuchung Chengf19c29e2014-03-03 12:31:36 -0800171 if (tp->syn_data)
172 NET_INC_STATS_BH(sock_net(sk),
173 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
Yuchung Chengc9686012013-10-29 10:09:05 -0700174 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700175 retry_until = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
Rusty Russell3db1cd52011-12-19 13:56:45 +0000176 syn_set = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177 } else {
David S. Miller21a180c2010-10-04 11:56:38 -0700178 if (retransmits_timed_out(sk, sysctl_tcp_retries1, 0, 0)) {
John Heffner5d424d52006-03-20 17:53:41 -0800179 /* Black hole detection */
Eric Dumazetce55dd32007-12-21 01:50:43 -0800180 tcp_mtu_probing(icsk, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700181
Eric Dumazetb6c67122010-04-08 23:03:29 +0000182 dst_negative_advice(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700183 }
184
185 retry_until = sysctl_tcp_retries2;
186 if (sock_flag(sk, SOCK_DEAD)) {
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -0700187 const int alive = icsk->icsk_rto < TCP_RTO_MAX;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900188
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189 retry_until = tcp_orphan_retries(sk, alive);
Damian Lukowski6fa12c82009-08-26 00:16:34 +0000190 do_reset = alive ||
David S. Miller21a180c2010-10-04 11:56:38 -0700191 !retransmits_timed_out(sk, retry_until, 0, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192
Damian Lukowski6fa12c82009-08-26 00:16:34 +0000193 if (tcp_out_of_resources(sk, do_reset))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700194 return 1;
195 }
196 }
197
Jerry Chudca43c72010-08-27 19:13:28 +0000198 if (retransmits_timed_out(sk, retry_until,
David S. Miller21a180c2010-10-04 11:56:38 -0700199 syn_set ? 0 : icsk->icsk_user_timeout, syn_set)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700200 /* Has it gone just too far? */
201 tcp_write_err(sk);
202 return 1;
203 }
204 return 0;
205}
206
Eric Dumazet6f458df2012-07-20 05:45:50 +0000207void tcp_delack_timer_handler(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700208{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700209 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700210 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700211
David S. Miller9993e7d2008-01-10 21:56:38 -0800212 sk_mem_reclaim_partial(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700214 if (sk->sk_state == TCP_CLOSE || !(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700215 goto out;
216
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700217 if (time_after(icsk->icsk_ack.timeout, jiffies)) {
218 sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700219 goto out;
220 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700221 icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700222
David S. Millerb03efcf2005-07-08 14:57:23 -0700223 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700224 struct sk_buff *skb;
225
Pavel Emelyanovde0744a2008-07-16 20:31:16 -0700226 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSCHEDULERFAILED);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700227
228 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
Peter Zijlstrac57943a2008-10-07 14:18:42 -0700229 sk_backlog_rcv(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700230
231 tp->ucopy.memory = 0;
232 }
233
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700234 if (inet_csk_ack_scheduled(sk)) {
235 if (!icsk->icsk_ack.pingpong) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700236 /* Delayed ACK missed: inflate ATO. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700237 icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700238 } else {
239 /* Delayed ACK missed: leave pingpong mode and
240 * deflate ATO.
241 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700242 icsk->icsk_ack.pingpong = 0;
243 icsk->icsk_ack.ato = TCP_ATO_MIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700244 }
245 tcp_send_ack(sk);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -0700246 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700248
249out:
Glauber Costa180d8cd2011-12-11 21:47:02 +0000250 if (sk_under_memory_pressure(sk))
Hideo Aoki3ab224b2007-12-31 00:11:19 -0800251 sk_mem_reclaim(sk);
Eric Dumazet6f458df2012-07-20 05:45:50 +0000252}
253
254static void tcp_delack_timer(unsigned long data)
255{
256 struct sock *sk = (struct sock *)data;
257
258 bh_lock_sock(sk);
259 if (!sock_owned_by_user(sk)) {
260 tcp_delack_timer_handler(sk);
261 } else {
262 inet_csk(sk)->icsk_ack.blocked = 1;
263 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
264 /* deleguate our work to tcp_release_cb() */
Eric Dumazet144d56e2012-08-20 00:22:46 +0000265 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &tcp_sk(sk)->tsq_flags))
266 sock_hold(sk);
Eric Dumazet6f458df2012-07-20 05:45:50 +0000267 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700268 bh_unlock_sock(sk);
269 sock_put(sk);
270}
271
272static void tcp_probe_timer(struct sock *sk)
273{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300274 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275 struct tcp_sock *tp = tcp_sk(sk);
276 int max_probes;
Yuchung Chengb2482302014-09-29 13:20:38 -0700277 u32 start_ts;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700278
David S. Millerfe067e82007-03-07 12:12:44 -0800279 if (tp->packets_out || !tcp_send_head(sk)) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300280 icsk->icsk_probes_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281 return;
282 }
283
Yuchung Chengb2482302014-09-29 13:20:38 -0700284 /* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as
285 * long as the receiver continues to respond probes. We support this by
286 * default and reset icsk_probes_out with incoming ACKs. But if the
287 * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we
288 * kill the socket when the retry count and the time exceeds the
289 * corresponding system limit. We also implement similar policy when
290 * we use RTO to probe window in tcp_retransmit_timer().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291 */
Yuchung Chengb2482302014-09-29 13:20:38 -0700292 start_ts = tcp_skb_timestamp(tcp_send_head(sk));
293 if (!start_ts)
294 skb_mstamp_get(&tcp_send_head(sk)->skb_mstamp);
295 else if (icsk->icsk_user_timeout &&
296 (s32)(tcp_time_stamp - start_ts) > icsk->icsk_user_timeout)
297 goto abort;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700298
Yuchung Chengb2482302014-09-29 13:20:38 -0700299 max_probes = sysctl_tcp_retries2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300 if (sock_flag(sk, SOCK_DEAD)) {
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -0700301 const int alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900302
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303 max_probes = tcp_orphan_retries(sk, alive);
Yuchung Chengb2482302014-09-29 13:20:38 -0700304 if (!alive && icsk->icsk_backoff >= max_probes)
305 goto abort;
306 if (tcp_out_of_resources(sk, true))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700307 return;
308 }
309
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300310 if (icsk->icsk_probes_out > max_probes) {
Yuchung Chengb2482302014-09-29 13:20:38 -0700311abort: tcp_write_err(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700312 } else {
313 /* Only send another probe if we didn't close things up. */
314 tcp_send_probe0(sk);
315 }
316}
317
318/*
Jerry Chu83368862012-08-31 12:29:12 +0000319 * Timer for Fast Open socket to retransmit SYNACK. Note that the
320 * sk here is the child socket, not the parent (listener) socket.
321 */
322static void tcp_fastopen_synack_timer(struct sock *sk)
323{
324 struct inet_connection_sock *icsk = inet_csk(sk);
325 int max_retries = icsk->icsk_syn_retries ? :
326 sysctl_tcp_synack_retries + 1; /* add one more retry for fastopen */
327 struct request_sock *req;
328
329 req = tcp_sk(sk)->fastopen_rsk;
Eric Dumazet42cb80a2015-03-22 10:22:19 -0700330 req->rsk_ops->syn_ack_timeout(req);
Jerry Chu83368862012-08-31 12:29:12 +0000331
Eric Dumazete6c022a2012-10-27 23:16:46 +0000332 if (req->num_timeout >= max_retries) {
Jerry Chu83368862012-08-31 12:29:12 +0000333 tcp_write_err(sk);
334 return;
335 }
336 /* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
337 * returned from rtx_syn_ack() to make it more persistent like
338 * regular retransmit because if the child socket has been accepted
339 * it's not good to give up too easily.
340 */
Eric Dumazete6c022a2012-10-27 23:16:46 +0000341 inet_rtx_syn_ack(sk, req);
342 req->num_timeout++;
Jerry Chu83368862012-08-31 12:29:12 +0000343 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
Eric Dumazete6c022a2012-10-27 23:16:46 +0000344 TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
Jerry Chu83368862012-08-31 12:29:12 +0000345}
346
347/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700348 * The TCP retransmit timer.
349 */
350
Damian Lukowskif1ecd5d2009-08-26 00:16:31 +0000351void tcp_retransmit_timer(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700352{
353 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700354 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700355
Jerry Chu83368862012-08-31 12:29:12 +0000356 if (tp->fastopen_rsk) {
Jerry Chu37561f62012-10-22 11:26:36 +0000357 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
358 sk->sk_state != TCP_FIN_WAIT1);
Jerry Chu83368862012-08-31 12:29:12 +0000359 tcp_fastopen_synack_timer(sk);
360 /* Before we receive ACK to our SYN-ACK don't retransmit
361 * anything else (e.g., data or FIN segments).
362 */
363 return;
364 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700365 if (!tp->packets_out)
366 goto out;
367
Ilpo Järvinen547b7922008-07-25 21:43:18 -0700368 WARN_ON(tcp_write_queue_empty(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369
Nandita Dukkipati9b717a82013-03-11 10:00:44 +0000370 tp->tlp_high_seq = 0;
371
Linus Torvalds1da177e2005-04-16 15:20:36 -0700372 if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
373 !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
374 /* Receiver dastardly shrinks window. Our retransmits
375 * become zero probes, but we should not timeout this
376 * connection. If the socket is an orphan, time it out,
377 * we cannot allow such beasts to hang infinitely.
378 */
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -0700379 struct inet_sock *inet = inet_sk(sk);
380 if (sk->sk_family == AF_INET) {
Joe Perchesba7a46f2014-11-11 10:59:17 -0800381 net_dbg_ratelimited("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
382 &inet->inet_daddr,
383 ntohs(inet->inet_dport),
384 inet->inet_num,
385 tp->snd_una, tp->snd_nxt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386 }
Eric Dumazetdfd56b82011-12-10 09:48:31 +0000387#if IS_ENABLED(CONFIG_IPV6)
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -0700388 else if (sk->sk_family == AF_INET6) {
Joe Perchesba7a46f2014-11-11 10:59:17 -0800389 net_dbg_ratelimited("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
390 &sk->sk_v6_daddr,
391 ntohs(inet->inet_dport),
392 inet->inet_num,
393 tp->snd_una, tp->snd_nxt);
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -0700394 }
395#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700396 if (tcp_time_stamp - tp->rcv_tstamp > TCP_RTO_MAX) {
397 tcp_write_err(sk);
398 goto out;
399 }
Neal Cardwell5ae344c2014-08-04 19:12:29 -0400400 tcp_enter_loss(sk);
David S. Millerfe067e82007-03-07 12:12:44 -0800401 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700402 __sk_dst_reset(sk);
403 goto out_reset_timer;
404 }
405
406 if (tcp_write_timeout(sk))
407 goto out;
408
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700409 if (icsk->icsk_retransmits == 0) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700410 int mib_idx;
411
Ilpo Järvinenc60ce4e2010-10-14 01:52:09 +0000412 if (icsk->icsk_ca_state == TCP_CA_Recovery) {
Ilpo Järvinenbc079e92009-02-28 04:44:34 +0000413 if (tcp_is_sack(tp))
414 mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
415 else
416 mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300417 } else if (icsk->icsk_ca_state == TCP_CA_Loss) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700418 mib_idx = LINUX_MIB_TCPLOSSFAILURES;
Ilpo Järvinenc60ce4e2010-10-14 01:52:09 +0000419 } else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
420 tp->sacked_out) {
421 if (tcp_is_sack(tp))
422 mib_idx = LINUX_MIB_TCPSACKFAILURES;
423 else
424 mib_idx = LINUX_MIB_TCPRENOFAILURES;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425 } else {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700426 mib_idx = LINUX_MIB_TCPTIMEOUTS;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700427 }
Pavel Emelyanovde0744a2008-07-16 20:31:16 -0700428 NET_INC_STATS_BH(sock_net(sk), mib_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700429 }
430
Neal Cardwell5ae344c2014-08-04 19:12:29 -0400431 tcp_enter_loss(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432
David S. Millerfe067e82007-03-07 12:12:44 -0800433 if (tcp_retransmit_skb(sk, tcp_write_queue_head(sk)) > 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434 /* Retransmission failed because of local congestion,
435 * do not backoff.
436 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700437 if (!icsk->icsk_retransmits)
438 icsk->icsk_retransmits = 1;
439 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -0700440 min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL),
441 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442 goto out;
443 }
444
445 /* Increase the timeout each time we retransmit. Note that
446 * we do not increase the rtt estimate. rto is initialized
447 * from rtt, but increases here. Jacobson (SIGCOMM 88) suggests
448 * that doubling rto each time is the least we can get away with.
449 * In KA9Q, Karn uses this for the first few times, and then
450 * goes to quadratic. netBSD doubles, but only goes up to *64,
451 * and clamps at 1 to 64 sec afterwards. Note that 120 sec is
452 * defined in the protocol as the maximum possible RTT. I guess
453 * we'll have to use something other than TCP to talk to the
454 * University of Mars.
455 *
456 * PAWS allows us longer timeouts and large windows, so once
457 * implemented ftp to mars will work nicely. We will have to fix
458 * the 120 second clamps though!
459 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700460 icsk->icsk_backoff++;
461 icsk->icsk_retransmits++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462
463out_reset_timer:
Andreas Petlund36e31b0a2010-02-18 02:47:01 +0000464 /* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
465 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
466 * might be increased if the stream oscillates between thin and thick,
467 * thus the old value might already be too high compared to the value
468 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
469 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
470 * exponential backoff behaviour to avoid continue hammering
471 * linear-timeout retransmissions into a black hole
472 */
473 if (sk->sk_state == TCP_ESTABLISHED &&
474 (tp->thin_lto || sysctl_tcp_thin_linear_timeouts) &&
475 tcp_stream_is_thin(tp) &&
476 icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
477 icsk->icsk_backoff = 0;
478 icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX);
479 } else {
480 /* Use normal (exponential) backoff */
481 icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
482 }
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -0700483 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX);
David S. Miller21a180c2010-10-04 11:56:38 -0700484 if (retransmits_timed_out(sk, sysctl_tcp_retries1 + 1, 0, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485 __sk_dst_reset(sk);
486
487out:;
488}
489
Eric Dumazet6f458df2012-07-20 05:45:50 +0000490void tcp_write_timer_handler(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700491{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700492 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700493 int event;
494
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700495 if (sk->sk_state == TCP_CLOSE || !icsk->icsk_pending)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700496 goto out;
497
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700498 if (time_after(icsk->icsk_timeout, jiffies)) {
499 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700500 goto out;
501 }
502
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700503 event = icsk->icsk_pending;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700504
505 switch (event) {
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +0000506 case ICSK_TIME_EARLY_RETRANS:
507 tcp_resume_early_retransmit(sk);
508 break;
509 case ICSK_TIME_LOSS_PROBE:
510 tcp_send_loss_probe(sk);
511 break;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700512 case ICSK_TIME_RETRANS:
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +0000513 icsk->icsk_pending = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700514 tcp_retransmit_timer(sk);
515 break;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700516 case ICSK_TIME_PROBE0:
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +0000517 icsk->icsk_pending = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700518 tcp_probe_timer(sk);
519 break;
520 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521
522out:
Hideo Aoki3ab224b2007-12-31 00:11:19 -0800523 sk_mem_reclaim(sk);
Eric Dumazet6f458df2012-07-20 05:45:50 +0000524}
525
526static void tcp_write_timer(unsigned long data)
527{
528 struct sock *sk = (struct sock *)data;
529
530 bh_lock_sock(sk);
531 if (!sock_owned_by_user(sk)) {
532 tcp_write_timer_handler(sk);
533 } else {
534 /* deleguate our work to tcp_release_cb() */
Eric Dumazet144d56e2012-08-20 00:22:46 +0000535 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &tcp_sk(sk)->tsq_flags))
536 sock_hold(sk);
Eric Dumazet6f458df2012-07-20 05:45:50 +0000537 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700538 bh_unlock_sock(sk);
539 sock_put(sk);
540}
541
Eric Dumazet42cb80a2015-03-22 10:22:19 -0700542void tcp_syn_ack_timeout(const struct request_sock *req)
Octavian Purdila72659ec2010-01-17 19:09:39 -0800543{
Eric Dumazet42cb80a2015-03-22 10:22:19 -0700544 struct net *net = read_pnet(&inet_rsk(req)->ireq_net);
545
546 NET_INC_STATS_BH(net, LINUX_MIB_TCPTIMEOUTS);
Octavian Purdila72659ec2010-01-17 19:09:39 -0800547}
548EXPORT_SYMBOL(tcp_syn_ack_timeout);
549
Linus Torvalds1da177e2005-04-16 15:20:36 -0700550void tcp_set_keepalive(struct sock *sk, int val)
551{
552 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
553 return;
554
555 if (val && !sock_flag(sk, SOCK_KEEPOPEN))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700556 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700557 else if (!val)
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700558 inet_csk_delete_keepalive_timer(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559}
560
561
562static void tcp_keepalive_timer (unsigned long data)
563{
564 struct sock *sk = (struct sock *) data;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300565 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700566 struct tcp_sock *tp = tcp_sk(sk);
Flavio Leitner6c37e5d2010-04-26 18:33:27 +0000567 u32 elapsed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568
569 /* Only process if socket is not in use. */
570 bh_lock_sock(sk);
571 if (sock_owned_by_user(sk)) {
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900572 /* Try again later. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700573 inet_csk_reset_keepalive_timer (sk, HZ/20);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574 goto out;
575 }
576
577 if (sk->sk_state == TCP_LISTEN) {
Eric Dumazetfa76ce732015-03-19 19:04:20 -0700578 pr_err("Hmm... keepalive on a LISTEN ???\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579 goto out;
580 }
581
582 if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
583 if (tp->linger2 >= 0) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700584 const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585
586 if (tmo > 0) {
587 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
588 goto out;
589 }
590 }
591 tcp_send_active_reset(sk, GFP_ATOMIC);
592 goto death;
593 }
594
595 if (!sock_flag(sk, SOCK_KEEPOPEN) || sk->sk_state == TCP_CLOSE)
596 goto out;
597
598 elapsed = keepalive_time_when(tp);
599
600 /* It is alive without keepalive 8) */
David S. Millerfe067e82007-03-07 12:12:44 -0800601 if (tp->packets_out || tcp_send_head(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700602 goto resched;
603
Flavio Leitner6c37e5d2010-04-26 18:33:27 +0000604 elapsed = keepalive_time_elapsed(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700605
606 if (elapsed >= keepalive_time_when(tp)) {
Jerry Chudca43c72010-08-27 19:13:28 +0000607 /* If the TCP_USER_TIMEOUT option is enabled, use that
608 * to determine when to timeout instead.
609 */
610 if ((icsk->icsk_user_timeout != 0 &&
611 elapsed >= icsk->icsk_user_timeout &&
612 icsk->icsk_probes_out > 0) ||
613 (icsk->icsk_user_timeout == 0 &&
614 icsk->icsk_probes_out >= keepalive_probes(tp))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700615 tcp_send_active_reset(sk, GFP_ATOMIC);
616 tcp_write_err(sk);
617 goto out;
618 }
619 if (tcp_write_wakeup(sk) <= 0) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300620 icsk->icsk_probes_out++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700621 elapsed = keepalive_intvl_when(tp);
622 } else {
623 /* If keepalive was lost due to local congestion,
624 * try harder.
625 */
626 elapsed = TCP_RESOURCE_PROBE_INTERVAL;
627 }
628 } else {
629 /* It is tp->rcv_tstamp + keepalive_time_when(tp) */
630 elapsed = keepalive_time_when(tp) - elapsed;
631 }
632
Hideo Aoki3ab224b2007-12-31 00:11:19 -0800633 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700634
635resched:
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700636 inet_csk_reset_keepalive_timer (sk, elapsed);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700637 goto out;
638
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900639death:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700640 tcp_done(sk);
641
642out:
643 bh_unlock_sock(sk);
644 sock_put(sk);
645}
Eric Dumazet6f458df2012-07-20 05:45:50 +0000646
647void tcp_init_xmit_timers(struct sock *sk)
648{
649 inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
650 &tcp_keepalive_timer);
651}
652EXPORT_SYMBOL(tcp_init_xmit_timers);