| 1 |
199 |
simons |
/*
|
| 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 |
|
|
*
|
| 8 |
|
|
* Version: @(#)tcp.c 1.0.16 05/25/93
|
| 9 |
|
|
*
|
| 10 |
|
|
* Authors: Ross Biro, <bir7@leland.Stanford.Edu>
|
| 11 |
|
|
* Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
|
| 12 |
|
|
* Mark Evans, <evansmp@uhura.aston.ac.uk>
|
| 13 |
|
|
* Corey Minyard <wf-rch!minyard@relay.EU.net>
|
| 14 |
|
|
* Florian La Roche, <flla@stud.uni-sb.de>
|
| 15 |
|
|
* Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
|
| 16 |
|
|
* Linus Torvalds, <torvalds@cs.helsinki.fi>
|
| 17 |
|
|
* Alan Cox, <gw4pts@gw4pts.ampr.org>
|
| 18 |
|
|
* Matthew Dillon, <dillon@apollo.west.oic.com>
|
| 19 |
|
|
* Arnt Gulbrandsen, <agulbra@nvg.unit.no>
|
| 20 |
|
|
* Jorge Cwik, <jorge@laser.satlink.net>
|
| 21 |
|
|
*
|
| 22 |
|
|
* Fixes:
|
| 23 |
|
|
*
|
| 24 |
|
|
* Eric Schenk : Fix retransmission timeout counting.
|
| 25 |
|
|
*/
|
| 26 |
|
|
|
| 27 |
|
|
#include <net/tcp.h>
|
| 28 |
|
|
|
| 29 |
|
|
void tcp_delack_timer(unsigned long data)
|
| 30 |
|
|
{
|
| 31 |
|
|
tcp_send_ack((struct sock *) data);
|
| 32 |
|
|
}
|
| 33 |
|
|
|
| 34 |
|
|
/*
|
| 35 |
|
|
* Reset the retransmission timer
|
| 36 |
|
|
*/
|
| 37 |
|
|
|
| 38 |
|
|
void tcp_reset_xmit_timer(struct sock *sk, int why, unsigned long when)
|
| 39 |
|
|
{
|
| 40 |
|
|
del_timer(&sk->retransmit_timer);
|
| 41 |
|
|
sk->ip_xmit_timeout = why;
|
| 42 |
|
|
if (why == TIME_WRITE) {
|
| 43 |
|
|
/* In this case we want to timeout on the first packet
|
| 44 |
|
|
* in the resend queue. If the resend queue is empty,
|
| 45 |
|
|
* then the packet we are sending hasn't made it there yet,
|
| 46 |
|
|
* so we timeout from the current time.
|
| 47 |
|
|
*/
|
| 48 |
|
|
if (sk->send_head) {
|
| 49 |
|
|
sk->retransmit_timer.expires =
|
| 50 |
|
|
sk->send_head->when + when;
|
| 51 |
|
|
} else {
|
| 52 |
|
|
/* This should never happen!
|
| 53 |
|
|
*/
|
| 54 |
|
|
printk(KERN_ERR "Error: send_head NULL in xmit_timer\n");
|
| 55 |
|
|
sk->ip_xmit_timeout = 0;
|
| 56 |
|
|
return;
|
| 57 |
|
|
}
|
| 58 |
|
|
} else {
|
| 59 |
|
|
sk->retransmit_timer.expires = jiffies+when;
|
| 60 |
|
|
}
|
| 61 |
|
|
|
| 62 |
|
|
if (sk->retransmit_timer.expires < jiffies) {
|
| 63 |
|
|
/* We can get here if we reset the timer on an event
|
| 64 |
|
|
* that could not fire because the interrupts were disabled.
|
| 65 |
|
|
* make sure it happens soon.
|
| 66 |
|
|
*/
|
| 67 |
|
|
sk->retransmit_timer.expires = jiffies+2;
|
| 68 |
|
|
}
|
| 69 |
|
|
add_timer(&sk->retransmit_timer);
|
| 70 |
|
|
}
|
| 71 |
|
|
|
| 72 |
|
|
/*
|
| 73 |
|
|
* POLICY:
|
| 74 |
|
|
*
|
| 75 |
|
|
* This is the normal code called for timeouts. It does the retransmission
|
| 76 |
|
|
* and then does backoff. tcp_do_retransmit is separated out because
|
| 77 |
|
|
* tcp_ack needs to send stuff from the retransmit queue without
|
| 78 |
|
|
* initiating a backoff.
|
| 79 |
|
|
*/
|
| 80 |
|
|
|
| 81 |
|
|
|
| 82 |
|
|
static void tcp_retransmit_time(struct sock *sk, int all)
|
| 83 |
|
|
{
|
| 84 |
|
|
/*
|
| 85 |
|
|
* record how many times we've timed out.
|
| 86 |
|
|
* This determines when we should quite trying.
|
| 87 |
|
|
* This needs to be counted here, because we should not be
|
| 88 |
|
|
* counting one per packet we send, but rather one per round
|
| 89 |
|
|
* trip timeout.
|
| 90 |
|
|
*/
|
| 91 |
|
|
sk->retransmits++;
|
| 92 |
|
|
|
| 93 |
|
|
tcp_do_retransmit(sk, all);
|
| 94 |
|
|
|
| 95 |
|
|
/*
|
| 96 |
|
|
* Increase the timeout each time we retransmit. Note that
|
| 97 |
|
|
* we do not increase the rtt estimate. rto is initialized
|
| 98 |
|
|
* from rtt, but increases here. Jacobson (SIGCOMM 88) suggests
|
| 99 |
|
|
* that doubling rto each time is the least we can get away with.
|
| 100 |
|
|
* In KA9Q, Karn uses this for the first few times, and then
|
| 101 |
|
|
* goes to quadratic. netBSD doubles, but only goes up to *64,
|
| 102 |
|
|
* and clamps at 1 to 64 sec afterwards. Note that 120 sec is
|
| 103 |
|
|
* defined in the protocol as the maximum possible RTT. I guess
|
| 104 |
|
|
* we'll have to use something other than TCP to talk to the
|
| 105 |
|
|
* University of Mars.
|
| 106 |
|
|
*
|
| 107 |
|
|
* PAWS allows us longer timeouts and large windows, so once
|
| 108 |
|
|
* implemented ftp to mars will work nicely. We will have to fix
|
| 109 |
|
|
* the 120 second clamps though!
|
| 110 |
|
|
*/
|
| 111 |
|
|
|
| 112 |
|
|
sk->backoff++;
|
| 113 |
|
|
sk->rto = min(sk->rto << 1, 120*HZ);
|
| 114 |
|
|
|
| 115 |
|
|
/* be paranoid about the data structure... */
|
| 116 |
|
|
if (sk->send_head)
|
| 117 |
|
|
tcp_reset_xmit_timer(sk, TIME_WRITE, sk->rto);
|
| 118 |
|
|
else
|
| 119 |
|
|
/* This should never happen! */
|
| 120 |
|
|
printk(KERN_ERR "send_head NULL in tcp_retransmit_time\n");
|
| 121 |
|
|
}
|
| 122 |
|
|
|
| 123 |
|
|
/*
|
| 124 |
|
|
* POLICY:
|
| 125 |
|
|
* Congestion control.
|
| 126 |
|
|
*
|
| 127 |
|
|
* A timer event has trigger a tcp retransmit timeout. The
|
| 128 |
|
|
* socket xmit queue is ready and set up to send. Because
|
| 129 |
|
|
* the ack receive code keeps the queue straight we do
|
| 130 |
|
|
* nothing clever here.
|
| 131 |
|
|
*/
|
| 132 |
|
|
|
| 133 |
|
|
void tcp_retransmit(struct sock *sk, int all)
|
| 134 |
|
|
{
|
| 135 |
|
|
if (all)
|
| 136 |
|
|
{
|
| 137 |
|
|
tcp_retransmit_time(sk, all);
|
| 138 |
|
|
return;
|
| 139 |
|
|
}
|
| 140 |
|
|
|
| 141 |
|
|
/* remember window where we lost */
|
| 142 |
|
|
sk->ssthresh = min(sk->cong_window,
|
| 143 |
|
|
(sk->window_seq-sk->rcv_ack_seq)/max(sk->mss,1)) >> 1;
|
| 144 |
|
|
/* sk->ssthresh in theory can be zero. I guess that's OK */
|
| 145 |
|
|
sk->cong_count = 0;
|
| 146 |
|
|
sk->cong_window = 1;
|
| 147 |
|
|
|
| 148 |
|
|
/* Do the actual retransmit. */
|
| 149 |
|
|
tcp_retransmit_time(sk, all);
|
| 150 |
|
|
}
|
| 151 |
|
|
|
| 152 |
|
|
/*
|
| 153 |
|
|
* A write timeout has occurred. Process the after effects. BROKEN (badly)
|
| 154 |
|
|
*/
|
| 155 |
|
|
|
| 156 |
|
|
static int tcp_write_timeout(struct sock *sk)
|
| 157 |
|
|
{
|
| 158 |
|
|
/*
|
| 159 |
|
|
* Look for a 'soft' timeout.
|
| 160 |
|
|
*/
|
| 161 |
|
|
if ((sk->state == TCP_ESTABLISHED && sk->retransmits && !(sk->retransmits & 7))
|
| 162 |
|
|
|| (sk->state != TCP_ESTABLISHED && sk->retransmits > TCP_RETR1))
|
| 163 |
|
|
{
|
| 164 |
|
|
/*
|
| 165 |
|
|
* Attempt to recover if arp has changed (unlikely!) or
|
| 166 |
|
|
* a route has shifted (not supported prior to 1.3).
|
| 167 |
|
|
*/
|
| 168 |
|
|
ip_rt_advice(&sk->ip_route_cache, 0);
|
| 169 |
|
|
}
|
| 170 |
|
|
|
| 171 |
|
|
/*
|
| 172 |
|
|
* Have we tried to SYN too many times (repent repent 8))
|
| 173 |
|
|
* NOTE: we must be careful to do this test for both
|
| 174 |
|
|
* the SYN_SENT and SYN_RECV states, otherwise we take
|
| 175 |
|
|
* 23 minutes to timeout on the SYN_RECV state, which
|
| 176 |
|
|
* leaves us (more) open to denial of service attacks
|
| 177 |
|
|
* than we would like.
|
| 178 |
|
|
*/
|
| 179 |
|
|
|
| 180 |
|
|
if (sk->retransmits > TCP_SYN_RETRIES
|
| 181 |
|
|
&& (sk->state==TCP_SYN_SENT || sk->state==TCP_SYN_RECV))
|
| 182 |
|
|
{
|
| 183 |
|
|
if(sk->err_soft)
|
| 184 |
|
|
sk->err=sk->err_soft;
|
| 185 |
|
|
else
|
| 186 |
|
|
sk->err=ETIMEDOUT;
|
| 187 |
|
|
sk->error_report(sk);
|
| 188 |
|
|
del_timer(&sk->retransmit_timer);
|
| 189 |
|
|
tcp_statistics.TcpAttemptFails++; /* Is this right ??? - FIXME - */
|
| 190 |
|
|
tcp_set_state(sk,TCP_CLOSE);
|
| 191 |
|
|
/* Don't FIN, we got nothing back */
|
| 192 |
|
|
return 0;
|
| 193 |
|
|
}
|
| 194 |
|
|
/*
|
| 195 |
|
|
* Has it gone just too far ?
|
| 196 |
|
|
*/
|
| 197 |
|
|
if (sk->retransmits > TCP_RETR2)
|
| 198 |
|
|
{
|
| 199 |
|
|
if(sk->err_soft)
|
| 200 |
|
|
sk->err = sk->err_soft;
|
| 201 |
|
|
else
|
| 202 |
|
|
sk->err = ETIMEDOUT;
|
| 203 |
|
|
sk->error_report(sk);
|
| 204 |
|
|
del_timer(&sk->retransmit_timer);
|
| 205 |
|
|
/*
|
| 206 |
|
|
* Time wait the socket
|
| 207 |
|
|
*/
|
| 208 |
|
|
if (sk->state == TCP_FIN_WAIT1 || sk->state == TCP_FIN_WAIT2 || sk->state == TCP_CLOSING )
|
| 209 |
|
|
{
|
| 210 |
|
|
tcp_set_state(sk,TCP_TIME_WAIT);
|
| 211 |
|
|
tcp_reset_msl_timer (sk, TIME_CLOSE, TCP_TIMEWAIT_LEN);
|
| 212 |
|
|
}
|
| 213 |
|
|
else
|
| 214 |
|
|
{
|
| 215 |
|
|
/*
|
| 216 |
|
|
* Clean up time.
|
| 217 |
|
|
*/
|
| 218 |
|
|
tcp_set_state(sk, TCP_CLOSE);
|
| 219 |
|
|
return 0;
|
| 220 |
|
|
}
|
| 221 |
|
|
}
|
| 222 |
|
|
return 1;
|
| 223 |
|
|
}
|
| 224 |
|
|
|
| 225 |
|
|
/*
|
| 226 |
|
|
* It could be we got here because we needed to send an ack,
|
| 227 |
|
|
* so we need to check for that and not just normal retransmit.
|
| 228 |
|
|
*/
|
| 229 |
|
|
static void tcp_time_write_timeout(struct sock * sk)
|
| 230 |
|
|
{
|
| 231 |
|
|
/*
|
| 232 |
|
|
* Retransmission
|
| 233 |
|
|
*/
|
| 234 |
|
|
sk->prot->retransmit (sk, 0);
|
| 235 |
|
|
tcp_write_timeout(sk);
|
| 236 |
|
|
}
|
| 237 |
|
|
|
| 238 |
|
|
|
| 239 |
|
|
/*
|
| 240 |
|
|
* The TCP retransmit timer. This lacks a few small details.
|
| 241 |
|
|
*
|
| 242 |
|
|
* 1. An initial rtt timeout on the probe0 should cause what we can
|
| 243 |
|
|
* of the first write queue buffer to be split and sent.
|
| 244 |
|
|
* 2. On a 'major timeout' as defined by RFC1122 we shouldn't report
|
| 245 |
|
|
* ETIMEDOUT if we know an additional 'soft' error caused this.
|
| 246 |
|
|
* tcp_err should save a 'soft error' for us.
|
| 247 |
|
|
*/
|
| 248 |
|
|
|
| 249 |
|
|
void tcp_retransmit_timer(unsigned long data)
|
| 250 |
|
|
{
|
| 251 |
|
|
struct sock *sk = (struct sock*)data;
|
| 252 |
|
|
int why = sk->ip_xmit_timeout;
|
| 253 |
|
|
|
| 254 |
|
|
/*
|
| 255 |
|
|
* We are reset. We will send no more retransmits.
|
| 256 |
|
|
*/
|
| 257 |
|
|
|
| 258 |
|
|
if(sk->zapped)
|
| 259 |
|
|
return;
|
| 260 |
|
|
|
| 261 |
|
|
/*
|
| 262 |
|
|
* Only process if socket is not in use
|
| 263 |
|
|
*/
|
| 264 |
|
|
|
| 265 |
|
|
if (sk->users)
|
| 266 |
|
|
{
|
| 267 |
|
|
/* Try again in 1 second */
|
| 268 |
|
|
sk->retransmit_timer.expires = jiffies+HZ;
|
| 269 |
|
|
add_timer(&sk->retransmit_timer);
|
| 270 |
|
|
return;
|
| 271 |
|
|
}
|
| 272 |
|
|
|
| 273 |
|
|
if (sk->ack_backlog && !sk->dead)
|
| 274 |
|
|
sk->data_ready(sk,0);
|
| 275 |
|
|
|
| 276 |
|
|
/* Now we need to figure out why the socket was on the timer. */
|
| 277 |
|
|
|
| 278 |
|
|
switch (why)
|
| 279 |
|
|
{
|
| 280 |
|
|
/* Window probing */
|
| 281 |
|
|
case TIME_PROBE0:
|
| 282 |
|
|
tcp_send_probe0(sk);
|
| 283 |
|
|
tcp_write_timeout(sk);
|
| 284 |
|
|
break;
|
| 285 |
|
|
|
| 286 |
|
|
/* Retransmitting */
|
| 287 |
|
|
case TIME_WRITE:
|
| 288 |
|
|
tcp_time_write_timeout(sk);
|
| 289 |
|
|
break;
|
| 290 |
|
|
|
| 291 |
|
|
/* Sending Keepalives */
|
| 292 |
|
|
case TIME_KEEPOPEN:
|
| 293 |
|
|
/*
|
| 294 |
|
|
* this reset_timer() call is a hack, this is not
|
| 295 |
|
|
* how KEEPOPEN is supposed to work.
|
| 296 |
|
|
*/
|
| 297 |
|
|
tcp_reset_xmit_timer (sk, TIME_KEEPOPEN, TCP_TIMEOUT_LEN);
|
| 298 |
|
|
/* Send something to keep the connection open. */
|
| 299 |
|
|
if (sk->prot->write_wakeup)
|
| 300 |
|
|
sk->prot->write_wakeup (sk);
|
| 301 |
|
|
sk->retransmits++;
|
| 302 |
|
|
sk->prot->retransmits++;
|
| 303 |
|
|
tcp_write_timeout(sk);
|
| 304 |
|
|
break;
|
| 305 |
|
|
|
| 306 |
|
|
default:
|
| 307 |
|
|
printk (KERN_ERR "rexmit_timer: timer expired - reason unknown\n");
|
| 308 |
|
|
break;
|
| 309 |
|
|
}
|
| 310 |
|
|
}
|
