1 /* $OpenBSD: tcp_output.c,v 1.80 2007/06/01 00:52:38 henning Exp $ */
2 /* $NetBSD: tcp_output.c,v 1.16 1997/06/03 16:17:09 kml Exp $ */
3
4 /*
5 * Copyright (c) 1982, 1986, 1988, 1990, 1993
6 * The Regents of the University of California. All rights reserved.
7 *
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
10 * are met:
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 3. Neither the name of the University nor the names of its contributors
17 * may be used to endorse or promote products derived from this software
18 * without specific prior written permission.
19 *
20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30 * SUCH DAMAGE.
31 *
32 * @(#)COPYRIGHT 1.1 (NRL) 17 January 1995
33 *
34 * NRL grants permission for redistribution and use in source and binary
35 * forms, with or without modification, of the software and documentation
36 * created at NRL provided that the following conditions are met:
37 *
38 * 1. Redistributions of source code must retain the above copyright
39 * notice, this list of conditions and the following disclaimer.
40 * 2. Redistributions in binary form must reproduce the above copyright
41 * notice, this list of conditions and the following disclaimer in the
42 * documentation and/or other materials provided with the distribution.
43 * 3. All advertising materials mentioning features or use of this software
44 * must display the following acknowledgements:
45 * This product includes software developed by the University of
46 * California, Berkeley and its contributors.
47 * This product includes software developed at the Information
48 * Technology Division, US Naval Research Laboratory.
49 * 4. Neither the name of the NRL nor the names of its contributors
50 * may be used to endorse or promote products derived from this software
51 * without specific prior written permission.
52 *
53 * THE SOFTWARE PROVIDED BY NRL IS PROVIDED BY NRL AND CONTRIBUTORS ``AS
54 * IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
55 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
56 * PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NRL OR
57 * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
58 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
59 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
60 * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
61 * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
62 * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
63 * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
64 *
65 * The views and conclusions contained in the software and documentation
66 * are those of the authors and should not be interpreted as representing
67 * official policies, either expressed or implied, of the US Naval
68 * Research Laboratory (NRL).
69 */
70
71 #include <sys/param.h>
72 #include <sys/systm.h>
73 #include <sys/mbuf.h>
74 #include <sys/protosw.h>
75 #include <sys/socket.h>
76 #include <sys/socketvar.h>
77 #include <sys/kernel.h>
78
79 #include <net/route.h>
80 #include <net/if.h>
81
82 #include <netinet/in.h>
83 #include <netinet/in_systm.h>
84 #include <netinet/ip.h>
85 #include <netinet/in_pcb.h>
86 #include <netinet/ip_var.h>
87 #include <netinet/tcp.h>
88 #define TCPOUTFLAGS
89 #include <netinet/tcp_fsm.h>
90 #include <netinet/tcp_seq.h>
91 #include <netinet/tcp_timer.h>
92 #include <netinet/tcp_var.h>
93 #include <netinet/tcpip.h>
94 #include <netinet/tcp_debug.h>
95
96 #ifdef INET6
97 #include <netinet6/tcpipv6.h>
98 #include <netinet6/in6_var.h>
99 #endif /* INET6 */
100
101 #ifdef notyet
102 extern struct mbuf *m_copypack();
103 #endif
104
105 #ifdef TCP_SACK
106 extern int tcprexmtthresh;
107 #endif
108
109 #ifdef TCP_SACK
110 #ifdef TCP_SACK_DEBUG
111 void tcp_print_holes(struct tcpcb *tp);
112
113 void
114 tcp_print_holes(struct tcpcb *tp)
115 {
116 struct sackhole *p = tp->snd_holes;
117 if (p == 0)
118 return;
119 printf("Hole report: start--end dups rxmit\n");
120 while (p) {
121 printf("%x--%x d %d r %x\n", p->start, p->end, p->dups,
122 p->rxmit);
123 p = p->next;
124 }
125 printf("\n");
126 }
127 #endif /* TCP_SACK_DEBUG */
128
129 /*
130 * Returns pointer to a sackhole if there are any pending retransmissions;
131 * NULL otherwise.
132 */
133 struct sackhole *
134 tcp_sack_output(struct tcpcb *tp)
135 {
136 struct sackhole *p;
137
138 if (!tp->sack_enable)
139 return (NULL);
140 p = tp->snd_holes;
141 while (p) {
142 #ifndef TCP_FACK
143 if (p->dups >= tcprexmtthresh && SEQ_LT(p->rxmit, p->end)) {
144 #else
145 /* In FACK, if p->dups is less than tcprexmtthresh, but
146 * snd_fack advances more than tcprextmtthresh * tp->t_maxseg,
147 * tcp_input() will try fast retransmit. This forces output.
148 */
149 if ((p->dups >= tcprexmtthresh ||
150 tp->t_dupacks == tcprexmtthresh) &&
151 SEQ_LT(p->rxmit, p->end)) {
152 #endif /* TCP_FACK */
153 if (SEQ_LT(p->rxmit, tp->snd_una)) {/* old SACK hole */
154 p = p->next;
155 continue;
156 }
157 #ifdef TCP_SACK_DEBUG
158 if (p)
159 tcp_print_holes(tp);
160 #endif
161 return (p);
162 }
163 p = p->next;
164 }
165 return (NULL);
166 }
167
168 /*
169 * After a timeout, the SACK list may be rebuilt. This SACK information
170 * should be used to avoid retransmitting SACKed data. This function
171 * traverses the SACK list to see if snd_nxt should be moved forward.
172 */
173
174 void
175 tcp_sack_adjust(struct tcpcb *tp)
176 {
177 struct sackhole *cur = tp->snd_holes;
178 if (cur == NULL)
179 return; /* No holes */
180 if (SEQ_GEQ(tp->snd_nxt, tp->rcv_lastsack))
181 return; /* We're already beyond any SACKed blocks */
182 /*
183 * Two cases for which we want to advance snd_nxt:
184 * i) snd_nxt lies between end of one hole and beginning of another
185 * ii) snd_nxt lies between end of last hole and rcv_lastsack
186 */
187 while (cur->next) {
188 if (SEQ_LT(tp->snd_nxt, cur->end))
189 return;
190 if (SEQ_GEQ(tp->snd_nxt, cur->next->start))
191 cur = cur->next;
192 else {
193 tp->snd_nxt = cur->next->start;
194 return;
195 }
196 }
197 if (SEQ_LT(tp->snd_nxt, cur->end))
198 return;
199 tp->snd_nxt = tp->rcv_lastsack;
200 return;
201 }
202 #endif /* TCP_SACK */
203
204 /*
205 * Tcp output routine: figure out what should be sent and send it.
206 */
207 int
208 tcp_output(tp)
209 struct tcpcb *tp;
210 {
211 struct socket *so = tp->t_inpcb->inp_socket;
212 long len, win, txmaxseg;
213 int off, flags, error;
214 struct mbuf *m;
215 struct tcphdr *th;
216 u_char opt[MAX_TCPOPTLEN];
217 unsigned int optlen, hdrlen, packetlen;
218 int idle, sendalot = 0;
219 #ifdef TCP_SACK
220 int i, sack_rxmit = 0;
221 struct sackhole *p;
222 #endif
223 #if defined(TCP_SACK)
224 int maxburst = TCP_MAXBURST;
225 #endif
226 #ifdef TCP_SIGNATURE
227 unsigned int sigoff;
228 #endif /* TCP_SIGNATURE */
229 #ifdef TCP_ECN
230 int needect;
231 #endif
232
233 #if defined(TCP_SACK) && defined(TCP_SIGNATURE) && defined(DIAGNOSTIC)
234 if (tp->sack_enable && (tp->t_flags & TF_SIGNATURE))
235 return (EINVAL);
236 #endif /* defined(TCP_SACK) && defined(TCP_SIGNATURE) && defined(DIAGNOSTIC) */
237
238 /*
239 * Determine length of data that should be transmitted,
240 * and flags that will be used.
241 * If there is some data or critical controls (SYN, RST)
242 * to send, then transmit; otherwise, investigate further.
243 */
244 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
245 if (idle && (tcp_now - tp->t_rcvtime) >= tp->t_rxtcur)
246 /*
247 * We have been idle for "a while" and no acks are
248 * expected to clock out any data we send --
249 * slow start to get ack "clock" running again.
250 */
251 tp->snd_cwnd = 2 * tp->t_maxseg;
252
253 /* remember 'idle' for next invocation of tcp_output */
254 if (idle && soissending(so)) {
255 tp->t_flags |= TF_LASTIDLE;
256 idle = 0;
257 } else
258 tp->t_flags &= ~TF_LASTIDLE;
259
260 again:
261 #ifdef TCP_SACK
262 /*
263 * If we've recently taken a timeout, snd_max will be greater than
264 * snd_nxt. There may be SACK information that allows us to avoid
265 * resending already delivered data. Adjust snd_nxt accordingly.
266 */
267 if (tp->sack_enable && SEQ_LT(tp->snd_nxt, tp->snd_max))
268 tcp_sack_adjust(tp);
269 #endif
270 off = tp->snd_nxt - tp->snd_una;
271 #if defined(TCP_SACK) && defined(TCP_FACK)
272 /* Normally, sendable data is limited by off < tp->snd_cwnd.
273 * But in FACK, sendable data is limited by snd_awnd < snd_cwnd,
274 * regardless of offset.
275 */
276 if (tp->sack_enable && (tp->t_dupacks > tcprexmtthresh))
277 win = tp->snd_wnd;
278 else
279 #endif
280 win = ulmin(tp->snd_wnd, tp->snd_cwnd);
281
282 flags = tcp_outflags[tp->t_state];
283
284 #ifdef TCP_SACK
285 /*
286 * Send any SACK-generated retransmissions. If we're explicitly trying
287 * to send out new data (when sendalot is 1), bypass this function.
288 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
289 * we're replacing a (future) new transmission with a retransmission
290 * now, and we previously incremented snd_cwnd in tcp_input().
291 */
292 if (tp->sack_enable && !sendalot) {
293 if (tp->t_dupacks >= tcprexmtthresh &&
294 (p = tcp_sack_output(tp))) {
295 off = p->rxmit - tp->snd_una;
296 sack_rxmit = 1;
297 #if 0
298 /* Coalesce holes into a single retransmission */
299 #endif
300 len = min(tp->t_maxseg, p->end - p->rxmit);
301 #ifndef TCP_FACK
302 /* in FACK, hold snd_cwnd constant during recovery */
303 if (SEQ_LT(tp->snd_una, tp->snd_last))
304 tp->snd_cwnd -= tp->t_maxseg;
305 #endif
306 }
307 }
308 #endif /* TCP_SACK */
309
310 sendalot = 0;
311 /*
312 * If in persist timeout with window of 0, send 1 byte.
313 * Otherwise, if window is small but nonzero
314 * and timer expired, we will send what we can
315 * and go to transmit state.
316 */
317 if (tp->t_force) {
318 if (win == 0) {
319 /*
320 * If we still have some data to send, then
321 * clear the FIN bit. Usually this would
322 * happen below when it realizes that we
323 * aren't sending all the data. However,
324 * if we have exactly 1 byte of unset data,
325 * then it won't clear the FIN bit below,
326 * and if we are in persist state, we wind
327 * up sending the packet without recording
328 * that we sent the FIN bit.
329 *
330 * We can't just blindly clear the FIN bit,
331 * because if we don't have any more data
332 * to send then the probe will be the FIN
333 * itself.
334 */
335 if (off < so->so_snd.sb_cc)
336 flags &= ~TH_FIN;
337 win = 1;
338 } else {
339 TCP_TIMER_DISARM(tp, TCPT_PERSIST);
340 tp->t_rxtshift = 0;
341 }
342 }
343
344 #ifdef TCP_SACK
345 if (!sack_rxmit) {
346 #endif
347 len = ulmin(so->so_snd.sb_cc, win) - off;
348
349 #if defined(TCP_SACK) && defined(TCP_FACK)
350 /*
351 * If we're in fast recovery (SEQ_GT(tp->snd_last, tp->snd_una)), and
352 * amount of outstanding data (snd_awnd) is >= snd_cwnd, then
353 * do not send data (like zero window conditions)
354 */
355 if (tp->sack_enable && len && SEQ_GT(tp->snd_last, tp->snd_una) &&
356 (tp->snd_awnd >= tp->snd_cwnd))
357 len = 0;
358 #endif /* TCP_FACK */
359 #ifdef TCP_SACK
360 }
361 #endif
362
363 if (len < 0) {
364 /*
365 * If FIN has been sent but not acked,
366 * but we haven't been called to retransmit,
367 * len will be -1. Otherwise, window shrank
368 * after we sent into it. If window shrank to 0,
369 * cancel pending retransmit, pull snd_nxt back
370 * to (closed) window, and set the persist timer
371 * if it isn't already going. If the window didn't
372 * close completely, just wait for an ACK.
373 */
374 len = 0;
375 if (win == 0) {
376 TCP_TIMER_DISARM(tp, TCPT_REXMT);
377 tp->t_rxtshift = 0;
378 tp->snd_nxt = tp->snd_una;
379 if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0)
380 tcp_setpersist(tp);
381 }
382 }
383
384 /*
385 * Never send more than half a buffer full. This insures that we can
386 * always keep 2 packets on the wire, no matter what SO_SNDBUF is, and
387 * therefore acks will never be delayed unless we run out of data to
388 * transmit.
389 */
390 txmaxseg = ulmin(so->so_snd.sb_hiwat / 2, tp->t_maxseg);
391
392 if (len > txmaxseg) {
393 len = txmaxseg;
394 sendalot = 1;
395 }
396 if (off + len < so->so_snd.sb_cc)
397 flags &= ~TH_FIN;
398
399 win = sbspace(&so->so_rcv);
400
401 /*
402 * Sender silly window avoidance. If connection is idle
403 * and can send all data, a maximum segment,
404 * at least a maximum default-size segment do it,
405 * or are forced, do it; otherwise don't bother.
406 * If peer's buffer is tiny, then send
407 * when window is at least half open.
408 * If retransmitting (possibly after persist timer forced us
409 * to send into a small window), then must resend.
410 */
411 if (len) {
412 if (len == txmaxseg)
413 goto send;
414 if ((idle || tp->t_flags & TF_NODELAY) &&
415 len + off >= so->so_snd.sb_cc && !soissending(so))
416 goto send;
417 if (tp->t_force)
418 goto send;
419 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
420 goto send;
421 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
422 goto send;
423 #ifdef TCP_SACK
424 if (sack_rxmit)
425 goto send;
426 #endif
427 }
428
429 /*
430 * Compare available window to amount of window
431 * known to peer (as advertised window less
432 * next expected input). If the difference is at least two
433 * max size segments, or at least 50% of the maximum possible
434 * window, then want to send a window update to peer.
435 */
436 if (win > 0) {
437 /*
438 * "adv" is the amount we can increase the window,
439 * taking into account that we are limited by
440 * TCP_MAXWIN << tp->rcv_scale.
441 */
442 long adv = lmin(win, (long)TCP_MAXWIN << tp->rcv_scale) -
443 (tp->rcv_adv - tp->rcv_nxt);
444
445 if (adv >= (long) (2 * tp->t_maxseg))
446 goto send;
447 if (2 * adv >= (long) so->so_rcv.sb_hiwat)
448 goto send;
449 }
450
451 /*
452 * Send if we owe peer an ACK.
453 */
454 if (tp->t_flags & TF_ACKNOW)
455 goto send;
456 if (flags & (TH_SYN|TH_RST))
457 goto send;
458 if (SEQ_GT(tp->snd_up, tp->snd_una))
459 goto send;
460 /*
461 * If our state indicates that FIN should be sent
462 * and we have not yet done so, or we're retransmitting the FIN,
463 * then we need to send.
464 */
465 if (flags & TH_FIN &&
466 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
467 goto send;
468 #ifdef TCP_SACK
469 /*
470 * In SACK, it is possible for tcp_output to fail to send a segment
471 * after the retransmission timer has been turned off. Make sure
472 * that the retransmission timer is set.
473 */
474 if (SEQ_GT(tp->snd_max, tp->snd_una) &&
475 TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
476 TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
477 TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
478 return (0);
479 }
480 #endif /* TCP_SACK */
481
482 /*
483 * TCP window updates are not reliable, rather a polling protocol
484 * using ``persist'' packets is used to insure receipt of window
485 * updates. The three ``states'' for the output side are:
486 * idle not doing retransmits or persists
487 * persisting to move a small or zero window
488 * (re)transmitting and thereby not persisting
489 *
490 * tp->t_timer[TCPT_PERSIST]
491 * is set when we are in persist state.
492 * tp->t_force
493 * is set when we are called to send a persist packet.
494 * tp->t_timer[TCPT_REXMT]
495 * is set when we are retransmitting
496 * The output side is idle when both timers are zero.
497 *
498 * If send window is too small, there is data to transmit, and no
499 * retransmit or persist is pending, then go to persist state.
500 * If nothing happens soon, send when timer expires:
501 * if window is nonzero, transmit what we can,
502 * otherwise force out a byte.
503 */
504 if (so->so_snd.sb_cc && TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
505 TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
506 tp->t_rxtshift = 0;
507 tcp_setpersist(tp);
508 }
509
510 /*
511 * No reason to send a segment, just return.
512 */
513 return (0);
514
515 send:
516 /*
517 * Before ESTABLISHED, force sending of initial options
518 * unless TCP set not to do any options.
519 * NOTE: we assume that the IP/TCP header plus TCP options
520 * always fit in a single mbuf, leaving room for a maximum
521 * link header, i.e.
522 * max_linkhdr + sizeof(network header) + sizeof(struct tcphdr +
523 * optlen <= MHLEN
524 */
525 optlen = 0;
526
527 switch (tp->pf) {
528 case 0: /*default to PF_INET*/
529 #ifdef INET
530 case PF_INET:
531 hdrlen = sizeof(struct ip) + sizeof(struct tcphdr);
532 break;
533 #endif /* INET */
534 #ifdef INET6
535 case PF_INET6:
536 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
537 break;
538 #endif /* INET6 */
539 default:
540 return (EPFNOSUPPORT);
541 }
542
543 if (flags & TH_SYN) {
544 tp->snd_nxt = tp->iss;
545 if ((tp->t_flags & TF_NOOPT) == 0) {
546 u_int16_t mss;
547
548 opt[0] = TCPOPT_MAXSEG;
549 opt[1] = 4;
550 mss = htons((u_int16_t) tcp_mss(tp, 0));
551 bcopy((caddr_t)&mss, (caddr_t)(opt + 2), sizeof(mss));
552 optlen = 4;
553
554 if (flags & TH_ACK)
555 tcp_mss_update(tp);
556 #ifdef TCP_SACK
557 /*
558 * If this is the first SYN of connection (not a SYN
559 * ACK), include SACK_PERMIT_HDR option. If this is a
560 * SYN ACK, include SACK_PERMIT_HDR option if peer has
561 * already done so.
562 */
563 if (tp->sack_enable && ((flags & TH_ACK) == 0 ||
564 (tp->t_flags & TF_SACK_PERMIT))) {
565 *((u_int32_t *) (opt + optlen)) =
566 htonl(TCPOPT_SACK_PERMIT_HDR);
567 optlen += 4;
568 }
569 #endif
570
571 if ((tp->t_flags & TF_REQ_SCALE) &&
572 ((flags & TH_ACK) == 0 ||
573 (tp->t_flags & TF_RCVD_SCALE))) {
574 *((u_int32_t *) (opt + optlen)) = htonl(
575 TCPOPT_NOP << 24 |
576 TCPOPT_WINDOW << 16 |
577 TCPOLEN_WINDOW << 8 |
578 tp->request_r_scale);
579 optlen += 4;
580 }
581 }
582 }
583
584 /*
585 * Send a timestamp and echo-reply if this is a SYN and our side
586 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
587 * and our peer have sent timestamps in our SYN's.
588 */
589 if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
590 (flags & TH_RST) == 0 &&
591 ((flags & (TH_SYN|TH_ACK)) == TH_SYN ||
592 (tp->t_flags & TF_RCVD_TSTMP))) {
593 u_int32_t *lp = (u_int32_t *)(opt + optlen);
594
595 /* Form timestamp option as shown in appendix A of RFC 1323. */
596 *lp++ = htonl(TCPOPT_TSTAMP_HDR);
597 *lp++ = htonl(tcp_now + tp->ts_modulate);
598 *lp = htonl(tp->ts_recent);
599 optlen += TCPOLEN_TSTAMP_APPA;
600 }
601
602 #ifdef TCP_SIGNATURE
603 if (tp->t_flags & TF_SIGNATURE) {
604 u_int8_t *bp = (u_int8_t *)(opt + optlen);
605
606 /* Send signature option */
607 *(bp++) = TCPOPT_SIGNATURE;
608 *(bp++) = TCPOLEN_SIGNATURE;
609 sigoff = optlen + 2;
610
611 {
612 unsigned int i;
613
614 for (i = 0; i < 16; i++)
615 *(bp++) = 0;
616 }
617
618
619 /* Pad options list to the next 32 bit boundary and
620 * terminate it.
621 */
622 *bp++ = TCPOPT_NOP;
623 *bp++ = TCPOPT_EOL;
624
625 optlen += TCPOLEN_SIGLEN;
626 }
627 #endif /* TCP_SIGNATURE */
628
629 #ifdef TCP_SACK
630 /*
631 * Send SACKs if necessary. This should be the last option processed.
632 * Only as many SACKs are sent as are permitted by the maximum options
633 * size. No more than three SACKs are sent.
634 */
635 if (tp->sack_enable && tp->t_state == TCPS_ESTABLISHED &&
636 (tp->t_flags & (TF_SACK_PERMIT|TF_NOOPT)) == TF_SACK_PERMIT &&
637 tp->rcv_numsacks) {
638 u_int32_t *lp = (u_int32_t *)(opt + optlen);
639 u_int32_t *olp = lp++;
640 int count = 0; /* actual number of SACKs inserted */
641 int maxsack = (MAX_TCPOPTLEN - (optlen + 4))/TCPOLEN_SACK;
642
643 tcpstat.tcps_sack_snd_opts++;
644 maxsack = min(maxsack, TCP_MAX_SACK);
645 for (i = 0; (i < tp->rcv_numsacks && count < maxsack); i++) {
646 struct sackblk sack = tp->sackblks[i];
647 if (sack.start == 0 && sack.end == 0)
648 continue;
649 *lp++ = htonl(sack.start);
650 *lp++ = htonl(sack.end);
651 count++;
652 }
653 *olp = htonl(TCPOPT_SACK_HDR|(TCPOLEN_SACK*count+2));
654 optlen += TCPOLEN_SACK*count + 4; /* including leading NOPs */
655 }
656 #endif /* TCP_SACK */
657
658 #ifdef DIAGNOSTIC
659 if (optlen > MAX_TCPOPTLEN)
660 panic("tcp_output: options too long");
661 #endif /* DIAGNOSTIC */
662
663 hdrlen += optlen;
664
665 /*
666 * Adjust data length if insertion of options will
667 * bump the packet length beyond the t_maxopd length.
668 */
669 if (len > tp->t_maxopd - optlen) {
670 len = tp->t_maxopd - optlen;
671 sendalot = 1;
672 flags &= ~TH_FIN;
673 }
674
675 #ifdef DIAGNOSTIC
676 if (max_linkhdr + hdrlen > MCLBYTES)
677 panic("tcphdr too big");
678 #endif
679
680 /*
681 * Grab a header mbuf, attaching a copy of data to
682 * be transmitted, and initialize the header from
683 * the template for sends on this connection.
684 */
685 if (len) {
686 if (tp->t_force && len == 1)
687 tcpstat.tcps_sndprobe++;
688 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
689 tcpstat.tcps_sndrexmitpack++;
690 tcpstat.tcps_sndrexmitbyte += len;
691 } else {
692 tcpstat.tcps_sndpack++;
693 tcpstat.tcps_sndbyte += len;
694 }
695 #ifdef notyet
696 if ((m = m_copypack(so->so_snd.sb_mb, off,
697 (int)len, max_linkhdr + hdrlen)) == 0) {
698 error = ENOBUFS;
699 goto out;
700 }
701 /*
702 * m_copypack left space for our hdr; use it.
703 */
704 m->m_len += hdrlen;
705 m->m_data -= hdrlen;
706 #else
707 MGETHDR(m, M_DONTWAIT, MT_HEADER);
708 if (m != NULL && max_linkhdr + hdrlen > MHLEN) {
709 MCLGET(m, M_DONTWAIT);
710 if ((m->m_flags & M_EXT) == 0) {
711 m_freem(m);
712 m = NULL;
713 }
714 }
715 if (m == NULL) {
716 error = ENOBUFS;
717 goto out;
718 }
719 m->m_data += max_linkhdr;
720 m->m_len = hdrlen;
721 if (len <= M_TRAILINGSPACE(m)) {
722 m_copydata(so->so_snd.sb_mb, off, (int) len,
723 mtod(m, caddr_t) + hdrlen);
724 m->m_len += len;
725 } else {
726 m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len);
727 if (m->m_next == 0) {
728 (void) m_free(m);
729 error = ENOBUFS;
730 goto out;
731 }
732 }
733 #endif
734 /*
735 * If we're sending everything we've got, set PUSH.
736 * (This will keep happy those implementations which only
737 * give data to the user when a buffer fills or
738 * a PUSH comes in.)
739 */
740 if (off + len == so->so_snd.sb_cc && !soissending(so))
741 flags |= TH_PUSH;
742 } else {
743 if (tp->t_flags & TF_ACKNOW)
744 tcpstat.tcps_sndacks++;
745 else if (flags & (TH_SYN|TH_FIN|TH_RST))
746 tcpstat.tcps_sndctrl++;
747 else if (SEQ_GT(tp->snd_up, tp->snd_una))
748 tcpstat.tcps_sndurg++;
749 else
750 tcpstat.tcps_sndwinup++;
751
752 MGETHDR(m, M_DONTWAIT, MT_HEADER);
753 if (m != NULL && max_linkhdr + hdrlen > MHLEN) {
754 MCLGET(m, M_DONTWAIT);
755 if ((m->m_flags & M_EXT) == 0) {
756 m_freem(m);
757 m = NULL;
758 }
759 }
760 if (m == NULL) {
761 error = ENOBUFS;
762 goto out;
763 }
764 m->m_data += max_linkhdr;
765 m->m_len = hdrlen;
766 }
767 m->m_pkthdr.rcvif = (struct ifnet *)0;
768 m->m_pkthdr.len = hdrlen + len;
769
770 if (!tp->t_template)
771 panic("tcp_output");
772 #ifdef DIAGNOSTIC
773 if (tp->t_template->m_len != hdrlen - optlen)
774 panic("tcp_output: template len != hdrlen - optlen");
775 #endif /* DIAGNOSTIC */
776 bcopy(mtod(tp->t_template, caddr_t), mtod(m, caddr_t),
777 tp->t_template->m_len);
778 th = (struct tcphdr *)(mtod(m, caddr_t) + tp->t_template->m_len -
779 sizeof(struct tcphdr));
780
781 /*
782 * Fill in fields, remembering maximum advertised
783 * window for use in delaying messages about window sizes.
784 * If resending a FIN, be sure not to use a new sequence number.
785 */
786 if ((flags & TH_FIN) && (tp->t_flags & TF_SENTFIN) &&
787 (tp->snd_nxt == tp->snd_max))
788 tp->snd_nxt--;
789 /*
790 * If we are doing retransmissions, then snd_nxt will
791 * not reflect the first unsent octet. For ACK only
792 * packets, we do not want the sequence number of the
793 * retransmitted packet, we want the sequence number
794 * of the next unsent octet. So, if there is no data
795 * (and no SYN or FIN), use snd_max instead of snd_nxt
796 * when filling in ti_seq. But if we are in persist
797 * state, snd_max might reflect one byte beyond the
798 * right edge of the window, so use snd_nxt in that
799 * case, since we know we aren't doing a retransmission.
800 * (retransmit and persist are mutually exclusive...)
801 */
802 if (len || (flags & (TH_SYN|TH_FIN)) || TCP_TIMER_ISARMED(tp, TCPT_PERSIST))
803 th->th_seq = htonl(tp->snd_nxt);
804 else
805 th->th_seq = htonl(tp->snd_max);
806
807 #ifdef TCP_SACK
808 if (sack_rxmit) {
809 /*
810 * If sendalot was turned on (due to option stuffing), turn it
811 * off. Properly set th_seq field. Advance the ret'x pointer
812 * by len.
813 */
814 if (sendalot)
815 sendalot = 0;
816 th->th_seq = htonl(p->rxmit);
817 p->rxmit += len;
818 #if defined(TCP_SACK) && defined(TCP_FACK)
819 tp->retran_data += len;
820 #endif /* TCP_FACK */
821 tcpstat.tcps_sack_rexmits++;
822 tcpstat.tcps_sack_rexmit_bytes += len;
823 }
824 #endif /* TCP_SACK */
825
826 th->th_ack = htonl(tp->rcv_nxt);
827 if (optlen) {
828 bcopy((caddr_t)opt, (caddr_t)(th + 1), optlen);
829 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
830 }
831 #ifdef TCP_ECN
832 if (tcp_do_ecn) {
833 /*
834 * if we have received congestion experienced segs,
835 * set ECE bit.
836 */
837 if (tp->t_flags & TF_RCVD_CE) {
838 flags |= TH_ECE;
839 tcpstat.tcps_ecn_sndece++;
840 }
841 if (!(tp->t_flags & TF_DISABLE_ECN)) {
842 /*
843 * if this is a SYN seg, set ECE and CWR.
844 * set only ECE for SYN-ACK if peer supports ECN.
845 */
846 if ((flags & (TH_SYN|TH_ACK)) == TH_SYN)
847 flags |= (TH_ECE|TH_CWR);
848 else if ((tp->t_flags & TF_ECN_PERMIT) &&
849 (flags & (TH_SYN|TH_ACK)) == (TH_SYN|TH_ACK))
850 flags |= TH_ECE;
851 }
852 /*
853 * if we have reduced the congestion window, notify
854 * the peer by setting CWR bit.
855 */
856 if ((tp->t_flags & TF_ECN_PERMIT) &&
857 (tp->t_flags & TF_SEND_CWR)) {
858 flags |= TH_CWR;
859 tp->t_flags &= ~TF_SEND_CWR;
860 tcpstat.tcps_ecn_sndcwr++;
861 }
862 }
863 #endif
864 th->th_flags = flags;
865
866 /*
867 * Calculate receive window. Don't shrink window,
868 * but avoid silly window syndrome.
869 */
870 if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)tp->t_maxseg)
871 win = 0;
872 if (win > (long)TCP_MAXWIN << tp->rcv_scale)
873 win = (long)TCP_MAXWIN << tp->rcv_scale;
874 if (win < (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt))
875 win = (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt);
876 if (flags & TH_RST)
877 win = 0;
878 th->th_win = htons((u_int16_t) (win>>tp->rcv_scale));
879 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
880 u_int32_t urp = tp->snd_up - tp->snd_nxt;
881 if (urp > IP_MAXPACKET)
882 urp = IP_MAXPACKET;
883 th->th_urp = htons((u_int16_t)urp);
884 th->th_flags |= TH_URG;
885 } else
886 /*
887 * If no urgent pointer to send, then we pull
888 * the urgent pointer to the left edge of the send window
889 * so that it doesn't drift into the send window on sequence
890 * number wraparound.
891 */
892 tp->snd_up = tp->snd_una; /* drag it along */
893
894 #ifdef TCP_SIGNATURE
895 if (tp->t_flags & TF_SIGNATURE) {
896 int iphlen;
897 union sockaddr_union src, dst;
898 struct tdb *tdb;
899
900 bzero(&src, sizeof(union sockaddr_union));
901 bzero(&dst, sizeof(union sockaddr_union));
902
903 switch (tp->pf) {
904 case 0: /*default to PF_INET*/
905 #ifdef INET
906 case AF_INET:
907 iphlen = sizeof(struct ip);
908 src.sa.sa_len = sizeof(struct sockaddr_in);
909 src.sa.sa_family = AF_INET;
910 src.sin.sin_addr = mtod(m, struct ip *)->ip_src;
911 dst.sa.sa_len = sizeof(struct sockaddr_in);
912 dst.sa.sa_family = AF_INET;
913 dst.sin.sin_addr = mtod(m, struct ip *)->ip_dst;
914 break;
915 #endif /* INET */
916 #ifdef INET6
917 case AF_INET6:
918 iphlen = sizeof(struct ip6_hdr);
919 src.sa.sa_len = sizeof(struct sockaddr_in6);
920 src.sa.sa_family = AF_INET6;
921 src.sin6.sin6_addr = mtod(m, struct ip6_hdr *)->ip6_src;
922 dst.sa.sa_len = sizeof(struct sockaddr_in6);
923 dst.sa.sa_family = AF_INET6;
924 dst.sin6.sin6_addr = mtod(m, struct ip6_hdr *)->ip6_dst;
925 break;
926 #endif /* INET6 */
927 }
928
929 /* XXX gettdbbysrcdst() should really be called at spltdb(). */
930 /* XXX this is splsoftnet(), currently they are the same. */
931 tdb = gettdbbysrcdst(0, &src, &dst, IPPROTO_TCP);
932 if (tdb == NULL)
933 return (EPERM);
934
935 if (tcp_signature(tdb, tp->pf, m, th, iphlen, 0,
936 mtod(m, caddr_t) + hdrlen - optlen + sigoff) < 0)
937 return (EINVAL);
938 }
939 #endif /* TCP_SIGNATURE */
940
941 /*
942 * Put TCP length in extended header, and then
943 * checksum extended header and data.
944 */
945 switch (tp->pf) {
946 case 0: /*default to PF_INET*/
947 #ifdef INET
948 case AF_INET:
949 /* Defer checksumming until later (ip_output() or hardware) */
950 m->m_pkthdr.csum_flags |= M_TCPV4_CSUM_OUT;
951 if (len + optlen)
952 th->th_sum = in_cksum_addword(th->th_sum,
953 htons((u_int16_t)(len + optlen)));
954 break;
955 #endif /* INET */
956 #ifdef INET6
957 case AF_INET6:
958 th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
959 hdrlen - sizeof(struct ip6_hdr) + len);
960 break;
961 #endif /* INET6 */
962 }
963
964 /*
965 * In transmit state, time the transmission and arrange for
966 * the retransmit. In persist state, just set snd_max.
967 */
968 if (tp->t_force == 0 || TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
969 tcp_seq startseq = tp->snd_nxt;
970
971 /*
972 * Advance snd_nxt over sequence space of this segment.
973 */
974 if (flags & (TH_SYN|TH_FIN)) {
975 if (flags & TH_SYN)
976 tp->snd_nxt++;
977 if (flags & TH_FIN) {
978 tp->snd_nxt++;
979 tp->t_flags |= TF_SENTFIN;
980 }
981 }
982 #ifdef TCP_SACK
983 if (tp->sack_enable) {
984 if (sack_rxmit && (p->rxmit != tp->snd_nxt)) {
985 goto timer;
986 }
987 }
988 #endif
989 tp->snd_nxt += len;
990 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
991 tp->snd_max = tp->snd_nxt;
992 /*
993 * Time this transmission if not a retransmission and
994 * not currently timing anything.
995 */
996 if (tp->t_rtttime == 0) {
997 tp->t_rtttime = tcp_now;
998 tp->t_rtseq = startseq;
999 tcpstat.tcps_segstimed++;
1000 }
1001 }
1002
1003 /*
1004 * Set retransmit timer if not currently set,
1005 * and not doing an ack or a keep-alive probe.
1006 * Initial value for retransmit timer is smoothed
1007 * round-trip time + 2 * round-trip time variance.
1008 * Initialize shift counter which is used for backoff
1009 * of retransmit time.
1010 */
1011 #ifdef TCP_SACK
1012 timer:
1013 if (tp->sack_enable && sack_rxmit &&
1014 TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
1015 tp->snd_nxt != tp->snd_max) {
1016 TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
1017 if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
1018 TCP_TIMER_DISARM(tp, TCPT_PERSIST);
1019 tp->t_rxtshift = 0;
1020 }
1021 }
1022 #endif
1023
1024 if (TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
1025 tp->snd_nxt != tp->snd_una) {
1026 TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
1027 if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
1028 TCP_TIMER_DISARM(tp, TCPT_PERSIST);
1029 tp->t_rxtshift = 0;
1030 }
1031 }
1032 } else
1033 if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
1034 tp->snd_max = tp->snd_nxt + len;
1035
1036 /*
1037 * Trace.
1038 */
1039 if (so->so_options & SO_DEBUG)
1040 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, caddr_t), 0,
1041 len);
1042
1043 /*
1044 * Fill in IP length and desired time to live and
1045 * send to IP level. There should be a better way
1046 * to handle ttl and tos; we could keep them in
1047 * the template, but need a way to checksum without them.
1048 */
1049
1050 #ifdef TCP_ECN
1051 /*
1052 * if peer is ECN capable, set the ECT bit in the IP header.
1053 * but don't set ECT for a pure ack, a retransmit or a window probe.
1054 */
1055 needect = 0;
1056 if (tcp_do_ecn && (tp->t_flags & TF_ECN_PERMIT)) {
1057 if (len == 0 || SEQ_LT(tp->snd_nxt, tp->snd_max) ||
1058 (tp->t_force && len == 1)) {
1059 /* don't set ECT */
1060 } else {
1061 needect = 1;
1062 tcpstat.tcps_ecn_sndect++;
1063 }
1064 }
1065 #endif
1066
1067 switch (tp->pf) {
1068 case 0: /*default to PF_INET*/
1069 #ifdef INET
1070 case AF_INET:
1071 {
1072 struct ip *ip;
1073
1074 ip = mtod(m, struct ip *);
1075 ip->ip_len = htons(m->m_pkthdr.len);
1076 packetlen = m->m_pkthdr.len;
1077 ip->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl;
1078 ip->ip_tos = tp->t_inpcb->inp_ip.ip_tos;
1079 #ifdef TCP_ECN
1080 if (needect)
1081 ip->ip_tos |= IPTOS_ECN_ECT0;
1082 #endif
1083 }
1084 error = ip_output(m, tp->t_inpcb->inp_options,
1085 &tp->t_inpcb->inp_route,
1086 (ip_mtudisc ? IP_MTUDISC : 0) |
1087 (so->so_options & SO_DONTROUTE),
1088 (void *)NULL, tp->t_inpcb);
1089 break;
1090 #endif /* INET */
1091 #ifdef INET6
1092 case AF_INET6:
1093 {
1094 struct ip6_hdr *ip6;
1095
1096 ip6 = mtod(m, struct ip6_hdr *);
1097 ip6->ip6_plen = m->m_pkthdr.len -
1098 sizeof(struct ip6_hdr);
1099 packetlen = m->m_pkthdr.len;
1100 ip6->ip6_nxt = IPPROTO_TCP;
1101 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1102 #ifdef TCP_ECN
1103 if (needect)
1104 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1105 #endif
1106 }
1107 error = ip6_output(m, tp->t_inpcb->inp_outputopts6,
1108 &tp->t_inpcb->inp_route6,
1109 (so->so_options & SO_DONTROUTE), NULL, NULL,
1110 tp->t_inpcb);
1111 break;
1112 #endif /* INET6 */
1113 }
1114
1115 #if defined(TCP_SACK) && defined(TCP_FACK)
1116 /* Update snd_awnd to reflect the new data that was sent. */
1117 tp->snd_awnd = tcp_seq_subtract(tp->snd_max, tp->snd_fack) +
1118 tp->retran_data;
1119 #endif /* defined(TCP_SACK) && defined(TCP_FACK) */
1120
1121 if (error) {
1122 out:
1123 if (error == ENOBUFS) {
1124 /*
1125 * If the interface queue is full, or IP cannot
1126 * get an mbuf, trigger TCP slow start.
1127 */
1128 tp->snd_cwnd = tp->t_maxseg;
1129 return (0);
1130 }
1131 if (error == EMSGSIZE) {
1132 /*
1133 * ip_output() will have already fixed the route
1134 * for us. tcp_mtudisc() will, as its last action,
1135 * initiate retransmission, so it is important to
1136 * not do so here.
1137 */
1138 tcp_mtudisc(tp->t_inpcb, -1);
1139 return (0);
1140 }
1141 if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1142 TCPS_HAVERCVDSYN(tp->t_state)) {
1143 tp->t_softerror = error;
1144 return (0);
1145 }
1146
1147 /* Restart the delayed ACK timer, if necessary. */
1148 if (tp->t_flags & TF_DELACK)
1149 TCP_RESTART_DELACK(tp);
1150
1151 return (error);
1152 }
1153
1154 if (packetlen > tp->t_pmtud_mtu_sent)
1155 tp->t_pmtud_mtu_sent = packetlen;
1156
1157 tcpstat.tcps_sndtotal++;
1158 if (tp->t_flags & TF_DELACK)
1159 tcpstat.tcps_delack++;
1160
1161 /*
1162 * Data sent (as far as we can tell).
1163 * If this advertises a larger window than any other segment,
1164 * then remember the size of the advertised window.
1165 * Any pending ACK has now been sent.
1166 */
1167 if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
1168 tp->rcv_adv = tp->rcv_nxt + win;
1169 tp->last_ack_sent = tp->rcv_nxt;
1170 tp->t_flags &= ~TF_ACKNOW;
1171 TCP_CLEAR_DELACK(tp);
1172 #if defined(TCP_SACK)
1173 if (sendalot && --maxburst)
1174 #else
1175 if (sendalot)
1176 #endif
1177 goto again;
1178 return (0);
1179 }
1180
1181 void
1182 tcp_setpersist(struct tcpcb *tp)
1183 {
1184 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> (1 + TCP_RTT_BASE_SHIFT);
1185 int nticks;
1186
1187 if (TCP_TIMER_ISARMED(tp, TCPT_REXMT))
1188 panic("tcp_output REXMT");
1189 /*
1190 * Start/restart persistance timer.
1191 */
1192 if (t < tp->t_rttmin)
1193 t = tp->t_rttmin;
1194 TCPT_RANGESET(nticks, t * tcp_backoff[tp->t_rxtshift],
1195 TCPTV_PERSMIN, TCPTV_PERSMAX);
1196 TCP_TIMER_ARM(tp, TCPT_PERSIST, nticks);
1197 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1198 tp->t_rxtshift++;
1199 }