xref: /freebsd/sys/netinet/tcp_output.c (revision 0de89efe5c443f213c7ea28773ef2dc6cf3af2ed)
1 /*
2  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  * 3. All advertising materials mentioning features or use of this software
14  *    must display the following acknowledgement:
15  *	This product includes software developed by the University of
16  *	California, Berkeley and its contributors.
17  * 4. Neither the name of the University nor the names of its contributors
18  *    may be used to endorse or promote products derived from this software
19  *    without specific prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  *
33  *	@(#)tcp_output.c	8.4 (Berkeley) 5/24/95
34  *	$Id: tcp_output.c,v 1.25 1997/08/02 14:32:56 bde Exp $
35  */
36 
37 #include "opt_tcpdebug.h"
38 
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/mbuf.h>
42 #include <sys/protosw.h>
43 #include <sys/socket.h>
44 #include <sys/socketvar.h>
45 
46 #include <net/route.h>
47 
48 #include <netinet/in.h>
49 #include <netinet/in_systm.h>
50 #include <netinet/ip.h>
51 #include <netinet/in_pcb.h>
52 #include <netinet/ip_var.h>
53 #include <netinet/tcp.h>
54 #define	TCPOUTFLAGS
55 #include <netinet/tcp_fsm.h>
56 #include <netinet/tcp_seq.h>
57 #include <netinet/tcp_timer.h>
58 #include <netinet/tcp_var.h>
59 #include <netinet/tcpip.h>
60 #ifdef TCPDEBUG
61 #include <netinet/tcp_debug.h>
62 #endif
63 
64 #ifdef notyet
65 extern struct mbuf *m_copypack();
66 #endif
67 
68 
69 /*
70  * Tcp output routine: figure out what should be sent and send it.
71  */
72 int
73 tcp_output(tp)
74 	register struct tcpcb *tp;
75 {
76 	register struct socket *so = tp->t_inpcb->inp_socket;
77 	register long len, win;
78 	int off, flags, error;
79 	register struct mbuf *m;
80 	register struct tcpiphdr *ti;
81 	u_char opt[TCP_MAXOLEN];
82 	unsigned optlen, hdrlen;
83 	int idle, sendalot;
84 	struct rmxp_tao *taop;
85 	struct rmxp_tao tao_noncached;
86 
87 	/*
88 	 * Determine length of data that should be transmitted,
89 	 * and flags that will be used.
90 	 * If there is some data or critical controls (SYN, RST)
91 	 * to send, then transmit; otherwise, investigate further.
92 	 */
93 	idle = (tp->snd_max == tp->snd_una);
94 	if (idle && tp->t_idle >= tp->t_rxtcur)
95 		/*
96 		 * We have been idle for "a while" and no acks are
97 		 * expected to clock out any data we send --
98 		 * slow start to get ack "clock" running again.
99 		 */
100 		tp->snd_cwnd = tp->t_maxseg;
101 again:
102 	sendalot = 0;
103 	off = tp->snd_nxt - tp->snd_una;
104 	win = min(tp->snd_wnd, tp->snd_cwnd);
105 
106 	flags = tcp_outflags[tp->t_state];
107 	/*
108 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
109 	 * state flags.
110 	 */
111 	if (tp->t_flags & TF_NEEDFIN)
112 		flags |= TH_FIN;
113 	if (tp->t_flags & TF_NEEDSYN)
114 		flags |= TH_SYN;
115 
116 	/*
117 	 * If in persist timeout with window of 0, send 1 byte.
118 	 * Otherwise, if window is small but nonzero
119 	 * and timer expired, we will send what we can
120 	 * and go to transmit state.
121 	 */
122 	if (tp->t_force) {
123 		if (win == 0) {
124 			/*
125 			 * If we still have some data to send, then
126 			 * clear the FIN bit.  Usually this would
127 			 * happen below when it realizes that we
128 			 * aren't sending all the data.  However,
129 			 * if we have exactly 1 byte of unset data,
130 			 * then it won't clear the FIN bit below,
131 			 * and if we are in persist state, we wind
132 			 * up sending the packet without recording
133 			 * that we sent the FIN bit.
134 			 *
135 			 * We can't just blindly clear the FIN bit,
136 			 * because if we don't have any more data
137 			 * to send then the probe will be the FIN
138 			 * itself.
139 			 */
140 			if (off < so->so_snd.sb_cc)
141 				flags &= ~TH_FIN;
142 			win = 1;
143 		} else {
144 			tp->t_timer[TCPT_PERSIST] = 0;
145 			tp->t_rxtshift = 0;
146 		}
147 	}
148 
149 	len = min(so->so_snd.sb_cc, win) - off;
150 
151 	if ((taop = tcp_gettaocache(tp->t_inpcb)) == NULL) {
152 		taop = &tao_noncached;
153 		bzero(taop, sizeof(*taop));
154 	}
155 
156 	/*
157 	 * Lop off SYN bit if it has already been sent.  However, if this
158 	 * is SYN-SENT state and if segment contains data and if we don't
159 	 * know that foreign host supports TAO, suppress sending segment.
160 	 */
161 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
162 		flags &= ~TH_SYN;
163 		off--, len++;
164 		if (len > 0 && tp->t_state == TCPS_SYN_SENT &&
165 		    taop->tao_ccsent == 0)
166 			return 0;
167 	}
168 
169 	/*
170 	 * Be careful not to send data and/or FIN on SYN segments
171 	 * in cases when no CC option will be sent.
172 	 * This measure is needed to prevent interoperability problems
173 	 * with not fully conformant TCP implementations.
174 	 */
175 	if ((flags & TH_SYN) &&
176 	    ((tp->t_flags & TF_NOOPT) || !(tp->t_flags & TF_REQ_CC) ||
177 	     ((flags & TH_ACK) && !(tp->t_flags & TF_RCVD_CC)))) {
178 		len = 0;
179 		flags &= ~TH_FIN;
180 	}
181 
182 	if (len < 0) {
183 		/*
184 		 * If FIN has been sent but not acked,
185 		 * but we haven't been called to retransmit,
186 		 * len will be -1.  Otherwise, window shrank
187 		 * after we sent into it.  If window shrank to 0,
188 		 * cancel pending retransmit, pull snd_nxt back
189 		 * to (closed) window, and set the persist timer
190 		 * if it isn't already going.  If the window didn't
191 		 * close completely, just wait for an ACK.
192 		 */
193 		len = 0;
194 		if (win == 0) {
195 			tp->t_timer[TCPT_REXMT] = 0;
196 			tp->t_rxtshift = 0;
197 			tp->snd_nxt = tp->snd_una;
198 			if (tp->t_timer[TCPT_PERSIST] == 0)
199 				tcp_setpersist(tp);
200 		}
201 	}
202 	if (len > tp->t_maxseg) {
203 		len = tp->t_maxseg;
204 		sendalot = 1;
205 	}
206 	if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
207 		flags &= ~TH_FIN;
208 
209 	win = sbspace(&so->so_rcv);
210 
211 	/*
212 	 * Sender silly window avoidance.  If connection is idle
213 	 * and can send all data, a maximum segment,
214 	 * at least a maximum default-size segment do it,
215 	 * or are forced, do it; otherwise don't bother.
216 	 * If peer's buffer is tiny, then send
217 	 * when window is at least half open.
218 	 * If retransmitting (possibly after persist timer forced us
219 	 * to send into a small window), then must resend.
220 	 */
221 	if (len) {
222 		if (len == tp->t_maxseg)
223 			goto send;
224 		if ((idle || tp->t_flags & TF_NODELAY) &&
225 		    (tp->t_flags & TF_NOPUSH) == 0 &&
226 		    len + off >= so->so_snd.sb_cc)
227 			goto send;
228 		if (tp->t_force)
229 			goto send;
230 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
231 			goto send;
232 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
233 			goto send;
234 	}
235 
236 	/*
237 	 * Compare available window to amount of window
238 	 * known to peer (as advertised window less
239 	 * next expected input).  If the difference is at least two
240 	 * max size segments, or at least 50% of the maximum possible
241 	 * window, then want to send a window update to peer.
242 	 */
243 	if (win > 0) {
244 		/*
245 		 * "adv" is the amount we can increase the window,
246 		 * taking into account that we are limited by
247 		 * TCP_MAXWIN << tp->rcv_scale.
248 		 */
249 		long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) -
250 			(tp->rcv_adv - tp->rcv_nxt);
251 
252 		if (adv >= (long) (2 * tp->t_maxseg))
253 			goto send;
254 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
255 			goto send;
256 	}
257 
258 	/*
259 	 * Send if we owe peer an ACK.
260 	 */
261 	if (tp->t_flags & TF_ACKNOW)
262 		goto send;
263 	if ((flags & TH_RST) ||
264 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
265 		goto send;
266 	if (SEQ_GT(tp->snd_up, tp->snd_una))
267 		goto send;
268 	/*
269 	 * If our state indicates that FIN should be sent
270 	 * and we have not yet done so, or we're retransmitting the FIN,
271 	 * then we need to send.
272 	 */
273 	if (flags & TH_FIN &&
274 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
275 		goto send;
276 
277 	/*
278 	 * TCP window updates are not reliable, rather a polling protocol
279 	 * using ``persist'' packets is used to insure receipt of window
280 	 * updates.  The three ``states'' for the output side are:
281 	 *	idle			not doing retransmits or persists
282 	 *	persisting		to move a small or zero window
283 	 *	(re)transmitting	and thereby not persisting
284 	 *
285 	 * tp->t_timer[TCPT_PERSIST]
286 	 *	is set when we are in persist state.
287 	 * tp->t_force
288 	 *	is set when we are called to send a persist packet.
289 	 * tp->t_timer[TCPT_REXMT]
290 	 *	is set when we are retransmitting
291 	 * The output side is idle when both timers are zero.
292 	 *
293 	 * If send window is too small, there is data to transmit, and no
294 	 * retransmit or persist is pending, then go to persist state.
295 	 * If nothing happens soon, send when timer expires:
296 	 * if window is nonzero, transmit what we can,
297 	 * otherwise force out a byte.
298 	 */
299 	if (so->so_snd.sb_cc && tp->t_timer[TCPT_REXMT] == 0 &&
300 	    tp->t_timer[TCPT_PERSIST] == 0) {
301 		tp->t_rxtshift = 0;
302 		tcp_setpersist(tp);
303 	}
304 
305 	/*
306 	 * No reason to send a segment, just return.
307 	 */
308 	return (0);
309 
310 send:
311 	/*
312 	 * Before ESTABLISHED, force sending of initial options
313 	 * unless TCP set not to do any options.
314 	 * NOTE: we assume that the IP/TCP header plus TCP options
315 	 * always fit in a single mbuf, leaving room for a maximum
316 	 * link header, i.e.
317 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MHLEN
318 	 */
319 	optlen = 0;
320 	hdrlen = sizeof (struct tcpiphdr);
321 	if (flags & TH_SYN) {
322 		tp->snd_nxt = tp->iss;
323 		if ((tp->t_flags & TF_NOOPT) == 0) {
324 			u_short mss;
325 
326 			opt[0] = TCPOPT_MAXSEG;
327 			opt[1] = TCPOLEN_MAXSEG;
328 			mss = htons((u_short) tcp_mssopt(tp));
329 			(void)memcpy(opt + 2, &mss, sizeof(mss));
330 			optlen = TCPOLEN_MAXSEG;
331 
332 			if ((tp->t_flags & TF_REQ_SCALE) &&
333 			    ((flags & TH_ACK) == 0 ||
334 			    (tp->t_flags & TF_RCVD_SCALE))) {
335 				*((u_long *) (opt + optlen)) = htonl(
336 					TCPOPT_NOP << 24 |
337 					TCPOPT_WINDOW << 16 |
338 					TCPOLEN_WINDOW << 8 |
339 					tp->request_r_scale);
340 				optlen += 4;
341 			}
342 		}
343  	}
344 
345  	/*
346 	 * Send a timestamp and echo-reply if this is a SYN and our side
347 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
348 	 * and our peer have sent timestamps in our SYN's.
349  	 */
350  	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
351  	    (flags & TH_RST) == 0 &&
352 	    ((flags & TH_ACK) == 0 ||
353 	     (tp->t_flags & TF_RCVD_TSTMP))) {
354 		u_long *lp = (u_long *)(opt + optlen);
355 
356  		/* Form timestamp option as shown in appendix A of RFC 1323. */
357  		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
358  		*lp++ = htonl(tcp_now);
359  		*lp   = htonl(tp->ts_recent);
360  		optlen += TCPOLEN_TSTAMP_APPA;
361  	}
362 
363  	/*
364 	 * Send `CC-family' options if our side wants to use them (TF_REQ_CC),
365 	 * options are allowed (!TF_NOOPT) and it's not a RST.
366  	 */
367  	if ((tp->t_flags & (TF_REQ_CC|TF_NOOPT)) == TF_REQ_CC &&
368  	     (flags & TH_RST) == 0) {
369 		switch (flags & (TH_SYN|TH_ACK)) {
370 		/*
371 		 * This is a normal ACK, send CC if we received CC before
372 		 * from our peer.
373 		 */
374 		case TH_ACK:
375 			if (!(tp->t_flags & TF_RCVD_CC))
376 				break;
377 			/*FALLTHROUGH*/
378 
379 		/*
380 		 * We can only get here in T/TCP's SYN_SENT* state, when
381 		 * we're a sending a non-SYN segment without waiting for
382 		 * the ACK of our SYN.  A check above assures that we only
383 		 * do this if our peer understands T/TCP.
384 		 */
385 		case 0:
386 			opt[optlen++] = TCPOPT_NOP;
387 			opt[optlen++] = TCPOPT_NOP;
388 			opt[optlen++] = TCPOPT_CC;
389 			opt[optlen++] = TCPOLEN_CC;
390 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
391 
392 			optlen += 4;
393 			break;
394 
395 		/*
396 		 * This is our initial SYN, check whether we have to use
397 		 * CC or CC.new.
398 		 */
399 		case TH_SYN:
400 			opt[optlen++] = TCPOPT_NOP;
401 			opt[optlen++] = TCPOPT_NOP;
402 			opt[optlen++] = tp->t_flags & TF_SENDCCNEW ?
403 						TCPOPT_CCNEW : TCPOPT_CC;
404 			opt[optlen++] = TCPOLEN_CC;
405 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
406  			optlen += 4;
407 			break;
408 
409 		/*
410 		 * This is a SYN,ACK; send CC and CC.echo if we received
411 		 * CC from our peer.
412 		 */
413 		case (TH_SYN|TH_ACK):
414 			if (tp->t_flags & TF_RCVD_CC) {
415 				opt[optlen++] = TCPOPT_NOP;
416 				opt[optlen++] = TCPOPT_NOP;
417 				opt[optlen++] = TCPOPT_CC;
418 				opt[optlen++] = TCPOLEN_CC;
419 				*(u_int32_t *)&opt[optlen] =
420 					htonl(tp->cc_send);
421 				optlen += 4;
422 				opt[optlen++] = TCPOPT_NOP;
423 				opt[optlen++] = TCPOPT_NOP;
424 				opt[optlen++] = TCPOPT_CCECHO;
425 				opt[optlen++] = TCPOLEN_CC;
426 				*(u_int32_t *)&opt[optlen] =
427 					htonl(tp->cc_recv);
428 				optlen += 4;
429 			}
430 			break;
431 		}
432  	}
433 
434  	hdrlen += optlen;
435 
436 	/*
437 	 * Adjust data length if insertion of options will
438 	 * bump the packet length beyond the t_maxopd length.
439 	 * Clear the FIN bit because we cut off the tail of
440 	 * the segment.
441 	 */
442 	 if (len + optlen > tp->t_maxopd) {
443 		/*
444 		 * If there is still more to send, don't close the connection.
445 		 */
446 		flags &= ~TH_FIN;
447 		len = tp->t_maxopd - optlen;
448 		sendalot = 1;
449 	}
450 
451 /*#ifdef DIAGNOSTIC*/
452  	if (max_linkhdr + hdrlen > MHLEN)
453 		panic("tcphdr too big");
454 /*#endif*/
455 
456 	/*
457 	 * Grab a header mbuf, attaching a copy of data to
458 	 * be transmitted, and initialize the header from
459 	 * the template for sends on this connection.
460 	 */
461 	if (len) {
462 		if (tp->t_force && len == 1)
463 			tcpstat.tcps_sndprobe++;
464 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
465 			tcpstat.tcps_sndrexmitpack++;
466 			tcpstat.tcps_sndrexmitbyte += len;
467 		} else {
468 			tcpstat.tcps_sndpack++;
469 			tcpstat.tcps_sndbyte += len;
470 		}
471 #ifdef notyet
472 		if ((m = m_copypack(so->so_snd.sb_mb, off,
473 		    (int)len, max_linkhdr + hdrlen)) == 0) {
474 			error = ENOBUFS;
475 			goto out;
476 		}
477 		/*
478 		 * m_copypack left space for our hdr; use it.
479 		 */
480 		m->m_len += hdrlen;
481 		m->m_data -= hdrlen;
482 #else
483 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
484 		if (m == NULL) {
485 			error = ENOBUFS;
486 			goto out;
487 		}
488 		m->m_data += max_linkhdr;
489 		m->m_len = hdrlen;
490 		if (len <= MHLEN - hdrlen - max_linkhdr) {
491 			m_copydata(so->so_snd.sb_mb, off, (int) len,
492 			    mtod(m, caddr_t) + hdrlen);
493 			m->m_len += len;
494 		} else {
495 			m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len);
496 			if (m->m_next == 0) {
497 				(void) m_free(m);
498 				error = ENOBUFS;
499 				goto out;
500 			}
501 		}
502 #endif
503 		/*
504 		 * If we're sending everything we've got, set PUSH.
505 		 * (This will keep happy those implementations which only
506 		 * give data to the user when a buffer fills or
507 		 * a PUSH comes in.)
508 		 */
509 		if (off + len == so->so_snd.sb_cc)
510 			flags |= TH_PUSH;
511 	} else {
512 		if (tp->t_flags & TF_ACKNOW)
513 			tcpstat.tcps_sndacks++;
514 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
515 			tcpstat.tcps_sndctrl++;
516 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
517 			tcpstat.tcps_sndurg++;
518 		else
519 			tcpstat.tcps_sndwinup++;
520 
521 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
522 		if (m == NULL) {
523 			error = ENOBUFS;
524 			goto out;
525 		}
526 		m->m_data += max_linkhdr;
527 		m->m_len = hdrlen;
528 	}
529 	m->m_pkthdr.rcvif = (struct ifnet *)0;
530 	ti = mtod(m, struct tcpiphdr *);
531 	if (tp->t_template == 0)
532 		panic("tcp_output");
533 	(void)memcpy(ti, tp->t_template, sizeof (struct tcpiphdr));
534 
535 	/*
536 	 * Fill in fields, remembering maximum advertised
537 	 * window for use in delaying messages about window sizes.
538 	 * If resending a FIN, be sure not to use a new sequence number.
539 	 */
540 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
541 	    tp->snd_nxt == tp->snd_max)
542 		tp->snd_nxt--;
543 	/*
544 	 * If we are doing retransmissions, then snd_nxt will
545 	 * not reflect the first unsent octet.  For ACK only
546 	 * packets, we do not want the sequence number of the
547 	 * retransmitted packet, we want the sequence number
548 	 * of the next unsent octet.  So, if there is no data
549 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
550 	 * when filling in ti_seq.  But if we are in persist
551 	 * state, snd_max might reflect one byte beyond the
552 	 * right edge of the window, so use snd_nxt in that
553 	 * case, since we know we aren't doing a retransmission.
554 	 * (retransmit and persist are mutually exclusive...)
555 	 */
556 	if (len || (flags & (TH_SYN|TH_FIN)) || tp->t_timer[TCPT_PERSIST])
557 		ti->ti_seq = htonl(tp->snd_nxt);
558 	else
559 		ti->ti_seq = htonl(tp->snd_max);
560 	ti->ti_ack = htonl(tp->rcv_nxt);
561 	if (optlen) {
562 		bcopy(opt, ti + 1, optlen);
563 		ti->ti_off = (sizeof (struct tcphdr) + optlen) >> 2;
564 	}
565 	ti->ti_flags = flags;
566 	/*
567 	 * Calculate receive window.  Don't shrink window,
568 	 * but avoid silly window syndrome.
569 	 */
570 	if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)tp->t_maxseg)
571 		win = 0;
572 	if (win > (long)TCP_MAXWIN << tp->rcv_scale)
573 		win = (long)TCP_MAXWIN << tp->rcv_scale;
574 	if (win < (long)(tp->rcv_adv - tp->rcv_nxt))
575 		win = (long)(tp->rcv_adv - tp->rcv_nxt);
576 	ti->ti_win = htons((u_short) (win>>tp->rcv_scale));
577 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
578 		ti->ti_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
579 		ti->ti_flags |= TH_URG;
580 	} else
581 		/*
582 		 * If no urgent pointer to send, then we pull
583 		 * the urgent pointer to the left edge of the send window
584 		 * so that it doesn't drift into the send window on sequence
585 		 * number wraparound.
586 		 */
587 		tp->snd_up = tp->snd_una;		/* drag it along */
588 
589 	/*
590 	 * Put TCP length in extended header, and then
591 	 * checksum extended header and data.
592 	 */
593 	if (len + optlen)
594 		ti->ti_len = htons((u_short)(sizeof (struct tcphdr) +
595 		    optlen + len));
596 	ti->ti_sum = in_cksum(m, (int)(hdrlen + len));
597 
598 	/*
599 	 * In transmit state, time the transmission and arrange for
600 	 * the retransmit.  In persist state, just set snd_max.
601 	 */
602 	if (tp->t_force == 0 || tp->t_timer[TCPT_PERSIST] == 0) {
603 		tcp_seq startseq = tp->snd_nxt;
604 
605 		/*
606 		 * Advance snd_nxt over sequence space of this segment.
607 		 */
608 		if (flags & (TH_SYN|TH_FIN)) {
609 			if (flags & TH_SYN)
610 				tp->snd_nxt++;
611 			if (flags & TH_FIN) {
612 				tp->snd_nxt++;
613 				tp->t_flags |= TF_SENTFIN;
614 			}
615 		}
616 		tp->snd_nxt += len;
617 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
618 			tp->snd_max = tp->snd_nxt;
619 			/*
620 			 * Time this transmission if not a retransmission and
621 			 * not currently timing anything.
622 			 */
623 			if (tp->t_rtt == 0) {
624 				tp->t_rtt = 1;
625 				tp->t_rtseq = startseq;
626 				tcpstat.tcps_segstimed++;
627 			}
628 		}
629 
630 		/*
631 		 * Set retransmit timer if not currently set,
632 		 * and not doing an ack or a keep-alive probe.
633 		 * Initial value for retransmit timer is smoothed
634 		 * round-trip time + 2 * round-trip time variance.
635 		 * Initialize shift counter which is used for backoff
636 		 * of retransmit time.
637 		 */
638 		if (tp->t_timer[TCPT_REXMT] == 0 &&
639 		    tp->snd_nxt != tp->snd_una) {
640 			tp->t_timer[TCPT_REXMT] = tp->t_rxtcur;
641 			if (tp->t_timer[TCPT_PERSIST]) {
642 				tp->t_timer[TCPT_PERSIST] = 0;
643 				tp->t_rxtshift = 0;
644 			}
645 		}
646 	} else
647 		if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
648 			tp->snd_max = tp->snd_nxt + len;
649 
650 #ifdef TCPDEBUG
651 	/*
652 	 * Trace.
653 	 */
654 	if (so->so_options & SO_DEBUG)
655 		tcp_trace(TA_OUTPUT, tp->t_state, tp, ti, 0);
656 #endif
657 
658 	/*
659 	 * Fill in IP length and desired time to live and
660 	 * send to IP level.  There should be a better way
661 	 * to handle ttl and tos; we could keep them in
662 	 * the template, but need a way to checksum without them.
663 	 */
664 	m->m_pkthdr.len = hdrlen + len;
665 #ifdef TUBA
666 	if (tp->t_tuba_pcb)
667 		error = tuba_output(m, tp);
668 	else
669 #endif
670     {
671 #if 1
672 	struct rtentry *rt;
673 #endif
674 	((struct ip *)ti)->ip_len = m->m_pkthdr.len;
675 	((struct ip *)ti)->ip_ttl = tp->t_inpcb->inp_ip_ttl;	/* XXX */
676 	((struct ip *)ti)->ip_tos = tp->t_inpcb->inp_ip_tos;	/* XXX */
677 #if 1
678 	/*
679 	 * See if we should do MTU discovery.  We do it only if the following
680 	 * are true:
681 	 *	1) we have a valid route to the destination
682 	 *	2) the MTU is not locked (if it is, then discovery has been
683 	 *	   disabled)
684 	 */
685 	if ((rt = tp->t_inpcb->inp_route.ro_rt)
686 	    && rt->rt_flags & RTF_UP
687 	    && !(rt->rt_rmx.rmx_locks & RTV_MTU)) {
688 		((struct ip *)ti)->ip_off |= IP_DF;
689 	}
690 #endif
691 	error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
692 	    so->so_options & SO_DONTROUTE, 0);
693     }
694 	if (error) {
695 out:
696 		if (error == ENOBUFS) {
697 			tcp_quench(tp->t_inpcb, 0);
698 			return (0);
699 		}
700 #if 1
701 		if (error == EMSGSIZE) {
702 			/*
703 			 * ip_output() will have already fixed the route
704 			 * for us.  tcp_mtudisc() will, as its last action,
705 			 * initiate retransmission, so it is important to
706 			 * not do so here.
707 			 */
708 			tcp_mtudisc(tp->t_inpcb, 0);
709 			return 0;
710 		}
711 #endif
712 		if ((error == EHOSTUNREACH || error == ENETDOWN)
713 		    && TCPS_HAVERCVDSYN(tp->t_state)) {
714 			tp->t_softerror = error;
715 			return (0);
716 		}
717 		return (error);
718 	}
719 	tcpstat.tcps_sndtotal++;
720 
721 	/*
722 	 * Data sent (as far as we can tell).
723 	 * If this advertises a larger window than any other segment,
724 	 * then remember the size of the advertised window.
725 	 * Any pending ACK has now been sent.
726 	 */
727 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
728 		tp->rcv_adv = tp->rcv_nxt + win;
729 	tp->last_ack_sent = tp->rcv_nxt;
730 	tp->t_flags &= ~(TF_ACKNOW|TF_DELACK);
731 	if (sendalot)
732 		goto again;
733 	return (0);
734 }
735 
736 void
737 tcp_setpersist(tp)
738 	register struct tcpcb *tp;
739 {
740 	register t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
741 
742 	if (tp->t_timer[TCPT_REXMT])
743 		panic("tcp_output REXMT");
744 	/*
745 	 * Start/restart persistance timer.
746 	 */
747 	TCPT_RANGESET(tp->t_timer[TCPT_PERSIST],
748 	    t * tcp_backoff[tp->t_rxtshift],
749 	    TCPTV_PERSMIN, TCPTV_PERSMAX);
750 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
751 		tp->t_rxtshift++;
752 }
753