xref: /freebsd/sys/netinet/tcp_output.c (revision ef5d438ed4bc17ad7ece3e40fe4d1f9baf3aadf7)
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 1.17 1995/12/05 17:46:35 wollman Exp $
35  */
36 
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/malloc.h>
40 #include <sys/mbuf.h>
41 #include <sys/protosw.h>
42 #include <sys/socket.h>
43 #include <sys/socketvar.h>
44 #include <sys/errno.h>
45 #include <sys/queue.h>
46 
47 #include <net/route.h>
48 
49 #include <netinet/in.h>
50 #include <netinet/in_systm.h>
51 #include <netinet/ip.h>
52 #include <netinet/in_pcb.h>
53 #include <netinet/ip_var.h>
54 #include <netinet/tcp.h>
55 #define	TCPOUTFLAGS
56 #include <netinet/tcp_fsm.h>
57 #include <netinet/tcp_seq.h>
58 #include <netinet/tcp_timer.h>
59 #include <netinet/tcp_var.h>
60 #include <netinet/tcpip.h>
61 #ifdef TCPDEBUG
62 #include <netinet/tcp_debug.h>
63 #endif
64 
65 #ifdef notyet
66 extern struct mbuf *m_copypack();
67 #endif
68 
69 
70 /*
71  * Tcp output routine: figure out what should be sent and send it.
72  */
73 int
74 tcp_output(tp)
75 	register struct tcpcb *tp;
76 {
77 	register struct socket *so = tp->t_inpcb->inp_socket;
78 	register long len, win;
79 	int off, flags, error;
80 	register struct mbuf *m;
81 	register struct tcpiphdr *ti;
82 	u_char opt[TCP_MAXOLEN];
83 	unsigned optlen, hdrlen;
84 	int idle, sendalot;
85 	struct rmxp_tao *taop;
86 	struct rmxp_tao tao_noncached;
87 
88 	/*
89 	 * Determine length of data that should be transmitted,
90 	 * and flags that will be used.
91 	 * If there is some data or critical controls (SYN, RST)
92 	 * to send, then transmit; otherwise, investigate further.
93 	 */
94 	idle = (tp->snd_max == tp->snd_una);
95 	if (idle && tp->t_idle >= tp->t_rxtcur)
96 		/*
97 		 * We have been idle for "a while" and no acks are
98 		 * expected to clock out any data we send --
99 		 * slow start to get ack "clock" running again.
100 		 */
101 		tp->snd_cwnd = tp->t_maxseg;
102 again:
103 	sendalot = 0;
104 	off = tp->snd_nxt - tp->snd_una;
105 	win = min(tp->snd_wnd, tp->snd_cwnd);
106 
107 	flags = tcp_outflags[tp->t_state];
108 	/*
109 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
110 	 * state flags.
111 	 */
112 	if (tp->t_flags & TF_NEEDFIN)
113 		flags |= TH_FIN;
114 	if (tp->t_flags & TF_NEEDSYN)
115 		flags |= TH_SYN;
116 
117 	/*
118 	 * If in persist timeout with window of 0, send 1 byte.
119 	 * Otherwise, if window is small but nonzero
120 	 * and timer expired, we will send what we can
121 	 * and go to transmit state.
122 	 */
123 	if (tp->t_force) {
124 		if (win == 0) {
125 			/*
126 			 * If we still have some data to send, then
127 			 * clear the FIN bit.  Usually this would
128 			 * happen below when it realizes that we
129 			 * aren't sending all the data.  However,
130 			 * if we have exactly 1 byte of unset data,
131 			 * then it won't clear the FIN bit below,
132 			 * and if we are in persist state, we wind
133 			 * up sending the packet without recording
134 			 * that we sent the FIN bit.
135 			 *
136 			 * We can't just blindly clear the FIN bit,
137 			 * because if we don't have any more data
138 			 * to send then the probe will be the FIN
139 			 * itself.
140 			 */
141 			if (off < so->so_snd.sb_cc)
142 				flags &= ~TH_FIN;
143 			win = 1;
144 		} else {
145 			tp->t_timer[TCPT_PERSIST] = 0;
146 			tp->t_rxtshift = 0;
147 		}
148 	}
149 
150 	len = min(so->so_snd.sb_cc, win) - off;
151 
152 	if ((taop = tcp_gettaocache(tp->t_inpcb)) == NULL) {
153 		taop = &tao_noncached;
154 		bzero(taop, sizeof(*taop));
155 	}
156 
157 	/*
158 	 * Lop off SYN bit if it has already been sent.  However, if this
159 	 * is SYN-SENT state and if segment contains data and if we don't
160 	 * know that foreign host supports TAO, suppress sending segment.
161 	 */
162 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
163 		flags &= ~TH_SYN;
164 		off--, len++;
165 		if (len > 0 && tp->t_state == TCPS_SYN_SENT &&
166 		    taop->tao_ccsent == 0)
167 			return 0;
168 	}
169 
170 	/*
171 	 * Be careful not to send data and/or FIN on SYN segments
172 	 * in cases when no CC option will be sent.
173 	 * This measure is needed to prevent interoperability problems
174 	 * with not fully conformant TCP implementations.
175 	 */
176 	if ((flags & TH_SYN) &&
177 	    ((tp->t_flags & TF_NOOPT) || !(tp->t_flags & TF_REQ_CC) ||
178 	     ((flags & TH_ACK) && !(tp->t_flags & TF_RCVD_CC)))) {
179 		len = 0;
180 		flags &= ~TH_FIN;
181 	}
182 
183 	if (len < 0) {
184 		/*
185 		 * If FIN has been sent but not acked,
186 		 * but we haven't been called to retransmit,
187 		 * len will be -1.  Otherwise, window shrank
188 		 * after we sent into it.  If window shrank to 0,
189 		 * cancel pending retransmit and pull snd_nxt
190 		 * back to (closed) window.  We will enter persist
191 		 * state below.  If the window didn't close completely,
192 		 * just wait for an ACK.
193 		 */
194 		len = 0;
195 		if (win == 0) {
196 			tp->t_timer[TCPT_REXMT] = 0;
197 			tp->snd_nxt = tp->snd_una;
198 		}
199 	}
200 	if (len > tp->t_maxseg) {
201 		len = tp->t_maxseg;
202 		sendalot = 1;
203 	}
204 	if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
205 		flags &= ~TH_FIN;
206 
207 	win = sbspace(&so->so_rcv);
208 
209 	/*
210 	 * Sender silly window avoidance.  If connection is idle
211 	 * and can send all data, a maximum segment,
212 	 * at least a maximum default-size segment do it,
213 	 * or are forced, do it; otherwise don't bother.
214 	 * If peer's buffer is tiny, then send
215 	 * when window is at least half open.
216 	 * If retransmitting (possibly after persist timer forced us
217 	 * to send into a small window), then must resend.
218 	 */
219 	if (len) {
220 		if (len == tp->t_maxseg)
221 			goto send;
222 		if ((idle || tp->t_flags & TF_NODELAY) &&
223 		    (tp->t_flags & TF_NOPUSH) == 0 &&
224 		    len + off >= so->so_snd.sb_cc)
225 			goto send;
226 		if (tp->t_force)
227 			goto send;
228 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
229 			goto send;
230 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
231 			goto send;
232 	}
233 
234 	/*
235 	 * Compare available window to amount of window
236 	 * known to peer (as advertised window less
237 	 * next expected input).  If the difference is at least two
238 	 * max size segments, or at least 50% of the maximum possible
239 	 * window, then want to send a window update to peer.
240 	 */
241 	if (win > 0) {
242 		/*
243 		 * "adv" is the amount we can increase the window,
244 		 * taking into account that we are limited by
245 		 * TCP_MAXWIN << tp->rcv_scale.
246 		 */
247 		long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) -
248 			(tp->rcv_adv - tp->rcv_nxt);
249 
250 		if (adv >= (long) (2 * tp->t_maxseg))
251 			goto send;
252 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
253 			goto send;
254 	}
255 
256 	/*
257 	 * Send if we owe peer an ACK.
258 	 */
259 	if (tp->t_flags & TF_ACKNOW)
260 		goto send;
261 	if ((flags & TH_RST) ||
262 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
263 		goto send;
264 	if (SEQ_GT(tp->snd_up, tp->snd_una))
265 		goto send;
266 	/*
267 	 * If our state indicates that FIN should be sent
268 	 * and we have not yet done so, or we're retransmitting the FIN,
269 	 * then we need to send.
270 	 */
271 	if (flags & TH_FIN &&
272 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
273 		goto send;
274 
275 	/*
276 	 * TCP window updates are not reliable, rather a polling protocol
277 	 * using ``persist'' packets is used to insure receipt of window
278 	 * updates.  The three ``states'' for the output side are:
279 	 *	idle			not doing retransmits or persists
280 	 *	persisting		to move a small or zero window
281 	 *	(re)transmitting	and thereby not persisting
282 	 *
283 	 * tp->t_timer[TCPT_PERSIST]
284 	 *	is set when we are in persist state.
285 	 * tp->t_force
286 	 *	is set when we are called to send a persist packet.
287 	 * tp->t_timer[TCPT_REXMT]
288 	 *	is set when we are retransmitting
289 	 * The output side is idle when both timers are zero.
290 	 *
291 	 * If send window is too small, there is data to transmit, and no
292 	 * retransmit or persist is pending, then go to persist state.
293 	 * If nothing happens soon, send when timer expires:
294 	 * if window is nonzero, transmit what we can,
295 	 * otherwise force out a byte.
296 	 */
297 	if (so->so_snd.sb_cc && tp->t_timer[TCPT_REXMT] == 0 &&
298 	    tp->t_timer[TCPT_PERSIST] == 0) {
299 		tp->t_rxtshift = 0;
300 		tcp_setpersist(tp);
301 	}
302 
303 	/*
304 	 * No reason to send a segment, just return.
305 	 */
306 	return (0);
307 
308 send:
309 	/*
310 	 * Before ESTABLISHED, force sending of initial options
311 	 * unless TCP set not to do any options.
312 	 * NOTE: we assume that the IP/TCP header plus TCP options
313 	 * always fit in a single mbuf, leaving room for a maximum
314 	 * link header, i.e.
315 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MHLEN
316 	 */
317 	optlen = 0;
318 	hdrlen = sizeof (struct tcpiphdr);
319 	if (flags & TH_SYN) {
320 		tp->snd_nxt = tp->iss;
321 		if ((tp->t_flags & TF_NOOPT) == 0) {
322 			u_short mss;
323 
324 			opt[0] = TCPOPT_MAXSEG;
325 			opt[1] = TCPOLEN_MAXSEG;
326 			mss = htons((u_short) tcp_mssopt(tp));
327 			(void)memcpy(opt + 2, &mss, sizeof(mss));
328 			optlen = TCPOLEN_MAXSEG;
329 
330 			if ((tp->t_flags & TF_REQ_SCALE) &&
331 			    ((flags & TH_ACK) == 0 ||
332 			    (tp->t_flags & TF_RCVD_SCALE))) {
333 				*((u_long *) (opt + optlen)) = htonl(
334 					TCPOPT_NOP << 24 |
335 					TCPOPT_WINDOW << 16 |
336 					TCPOLEN_WINDOW << 8 |
337 					tp->request_r_scale);
338 				optlen += 4;
339 			}
340 		}
341  	}
342 
343  	/*
344 	 * Send a timestamp and echo-reply if this is a SYN and our side
345 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
346 	 * and our peer have sent timestamps in our SYN's.
347  	 */
348  	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
349  	    (flags & TH_RST) == 0 &&
350 	    ((flags & TH_ACK) == 0 ||
351 	     (tp->t_flags & TF_RCVD_TSTMP))) {
352 		u_long *lp = (u_long *)(opt + optlen);
353 
354  		/* Form timestamp option as shown in appendix A of RFC 1323. */
355  		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
356  		*lp++ = htonl(tcp_now);
357  		*lp   = htonl(tp->ts_recent);
358  		optlen += TCPOLEN_TSTAMP_APPA;
359  	}
360 
361  	/*
362 	 * Send `CC-family' options if our side wants to use them (TF_REQ_CC),
363 	 * options are allowed (!TF_NOOPT) and it's not a RST.
364  	 */
365  	if ((tp->t_flags & (TF_REQ_CC|TF_NOOPT)) == TF_REQ_CC &&
366  	     (flags & TH_RST) == 0) {
367 		switch (flags & (TH_SYN|TH_ACK)) {
368 		/*
369 		 * This is a normal ACK, send CC if we received CC before
370 		 * from our peer.
371 		 */
372 		case TH_ACK:
373 			if (!(tp->t_flags & TF_RCVD_CC))
374 				break;
375 			/*FALLTHROUGH*/
376 
377 		/*
378 		 * We can only get here in T/TCP's SYN_SENT* state, when
379 		 * we're a sending a non-SYN segment without waiting for
380 		 * the ACK of our SYN.  A check above assures that we only
381 		 * do this if our peer understands T/TCP.
382 		 */
383 		case 0:
384 			opt[optlen++] = TCPOPT_NOP;
385 			opt[optlen++] = TCPOPT_NOP;
386 			opt[optlen++] = TCPOPT_CC;
387 			opt[optlen++] = TCPOLEN_CC;
388 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
389 
390 			optlen += 4;
391 			break;
392 
393 		/*
394 		 * This is our initial SYN, check whether we have to use
395 		 * CC or CC.new.
396 		 */
397 		case TH_SYN:
398 			opt[optlen++] = TCPOPT_NOP;
399 			opt[optlen++] = TCPOPT_NOP;
400 			opt[optlen++] = tp->t_flags & TF_SENDCCNEW ?
401 						TCPOPT_CCNEW : TCPOPT_CC;
402 			opt[optlen++] = TCPOLEN_CC;
403 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
404  			optlen += 4;
405 			break;
406 
407 		/*
408 		 * This is a SYN,ACK; send CC and CC.echo if we received
409 		 * CC from our peer.
410 		 */
411 		case (TH_SYN|TH_ACK):
412 			if (tp->t_flags & TF_RCVD_CC) {
413 				opt[optlen++] = TCPOPT_NOP;
414 				opt[optlen++] = TCPOPT_NOP;
415 				opt[optlen++] = TCPOPT_CC;
416 				opt[optlen++] = TCPOLEN_CC;
417 				*(u_int32_t *)&opt[optlen] =
418 					htonl(tp->cc_send);
419 				optlen += 4;
420 				opt[optlen++] = TCPOPT_NOP;
421 				opt[optlen++] = TCPOPT_NOP;
422 				opt[optlen++] = TCPOPT_CCECHO;
423 				opt[optlen++] = TCPOLEN_CC;
424 				*(u_int32_t *)&opt[optlen] =
425 					htonl(tp->cc_recv);
426 				optlen += 4;
427 			}
428 			break;
429 		}
430  	}
431 
432  	hdrlen += optlen;
433 
434 	/*
435 	 * Adjust data length if insertion of options will
436 	 * bump the packet length beyond the t_maxopd length.
437 	 * Clear the FIN bit because we cut off the tail of
438 	 * the segment.
439 	 */
440 	 if (len + optlen > tp->t_maxopd) {
441 		/*
442 		 * If there is still more to send, don't close the connection.
443 		 */
444 		flags &= ~TH_FIN;
445 		len = tp->t_maxopd - optlen;
446 		sendalot = 1;
447 	}
448 
449 /*#ifdef DIAGNOSTIC*/
450  	if (max_linkhdr + hdrlen > MHLEN)
451 		panic("tcphdr too big");
452 /*#endif*/
453 
454 	/*
455 	 * Grab a header mbuf, attaching a copy of data to
456 	 * be transmitted, and initialize the header from
457 	 * the template for sends on this connection.
458 	 */
459 	if (len) {
460 		if (tp->t_force && len == 1)
461 			tcpstat.tcps_sndprobe++;
462 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
463 			tcpstat.tcps_sndrexmitpack++;
464 			tcpstat.tcps_sndrexmitbyte += len;
465 		} else {
466 			tcpstat.tcps_sndpack++;
467 			tcpstat.tcps_sndbyte += len;
468 		}
469 #ifdef notyet
470 		if ((m = m_copypack(so->so_snd.sb_mb, off,
471 		    (int)len, max_linkhdr + hdrlen)) == 0) {
472 			error = ENOBUFS;
473 			goto out;
474 		}
475 		/*
476 		 * m_copypack left space for our hdr; use it.
477 		 */
478 		m->m_len += hdrlen;
479 		m->m_data -= hdrlen;
480 #else
481 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
482 		if (m == NULL) {
483 			error = ENOBUFS;
484 			goto out;
485 		}
486 		m->m_data += max_linkhdr;
487 		m->m_len = hdrlen;
488 		if (len <= MHLEN - hdrlen - max_linkhdr) {
489 			m_copydata(so->so_snd.sb_mb, off, (int) len,
490 			    mtod(m, caddr_t) + hdrlen);
491 			m->m_len += len;
492 		} else {
493 			m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len);
494 			if (m->m_next == 0) {
495 				(void) m_free(m);
496 				error = ENOBUFS;
497 				goto out;
498 			}
499 		}
500 #endif
501 		/*
502 		 * If we're sending everything we've got, set PUSH.
503 		 * (This will keep happy those implementations which only
504 		 * give data to the user when a buffer fills or
505 		 * a PUSH comes in.)
506 		 */
507 		if (off + len == so->so_snd.sb_cc)
508 			flags |= TH_PUSH;
509 	} else {
510 		if (tp->t_flags & TF_ACKNOW)
511 			tcpstat.tcps_sndacks++;
512 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
513 			tcpstat.tcps_sndctrl++;
514 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
515 			tcpstat.tcps_sndurg++;
516 		else
517 			tcpstat.tcps_sndwinup++;
518 
519 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
520 		if (m == NULL) {
521 			error = ENOBUFS;
522 			goto out;
523 		}
524 		m->m_data += max_linkhdr;
525 		m->m_len = hdrlen;
526 	}
527 	m->m_pkthdr.rcvif = (struct ifnet *)0;
528 	ti = mtod(m, struct tcpiphdr *);
529 	if (tp->t_template == 0)
530 		panic("tcp_output");
531 	(void)memcpy(ti, tp->t_template, sizeof (struct tcpiphdr));
532 
533 	/*
534 	 * Fill in fields, remembering maximum advertised
535 	 * window for use in delaying messages about window sizes.
536 	 * If resending a FIN, be sure not to use a new sequence number.
537 	 */
538 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
539 	    tp->snd_nxt == tp->snd_max)
540 		tp->snd_nxt--;
541 	/*
542 	 * If we are doing retransmissions, then snd_nxt will
543 	 * not reflect the first unsent octet.  For ACK only
544 	 * packets, we do not want the sequence number of the
545 	 * retransmitted packet, we want the sequence number
546 	 * of the next unsent octet.  So, if there is no data
547 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
548 	 * when filling in ti_seq.  But if we are in persist
549 	 * state, snd_max might reflect one byte beyond the
550 	 * right edge of the window, so use snd_nxt in that
551 	 * case, since we know we aren't doing a retransmission.
552 	 * (retransmit and persist are mutually exclusive...)
553 	 */
554 	if (len || (flags & (TH_SYN|TH_FIN)) || tp->t_timer[TCPT_PERSIST])
555 		ti->ti_seq = htonl(tp->snd_nxt);
556 	else
557 		ti->ti_seq = htonl(tp->snd_max);
558 	ti->ti_ack = htonl(tp->rcv_nxt);
559 	if (optlen) {
560 		(void)memcpy(ti + 1, opt, optlen);
561 		ti->ti_off = (sizeof (struct tcphdr) + optlen) >> 2;
562 	}
563 	ti->ti_flags = flags;
564 	/*
565 	 * Calculate receive window.  Don't shrink window,
566 	 * but avoid silly window syndrome.
567 	 */
568 	if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)tp->t_maxseg)
569 		win = 0;
570 	if (win > (long)TCP_MAXWIN << tp->rcv_scale)
571 		win = (long)TCP_MAXWIN << tp->rcv_scale;
572 	if (win < (long)(tp->rcv_adv - tp->rcv_nxt))
573 		win = (long)(tp->rcv_adv - tp->rcv_nxt);
574 	ti->ti_win = htons((u_short) (win>>tp->rcv_scale));
575 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
576 		ti->ti_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
577 		ti->ti_flags |= TH_URG;
578 	} else
579 		/*
580 		 * If no urgent pointer to send, then we pull
581 		 * the urgent pointer to the left edge of the send window
582 		 * so that it doesn't drift into the send window on sequence
583 		 * number wraparound.
584 		 */
585 		tp->snd_up = tp->snd_una;		/* drag it along */
586 
587 	/*
588 	 * Put TCP length in extended header, and then
589 	 * checksum extended header and data.
590 	 */
591 	if (len + optlen)
592 		ti->ti_len = htons((u_short)(sizeof (struct tcphdr) +
593 		    optlen + len));
594 	ti->ti_sum = in_cksum(m, (int)(hdrlen + len));
595 
596 	/*
597 	 * In transmit state, time the transmission and arrange for
598 	 * the retransmit.  In persist state, just set snd_max.
599 	 */
600 	if (tp->t_force == 0 || tp->t_timer[TCPT_PERSIST] == 0) {
601 		tcp_seq startseq = tp->snd_nxt;
602 
603 		/*
604 		 * Advance snd_nxt over sequence space of this segment.
605 		 */
606 		if (flags & (TH_SYN|TH_FIN)) {
607 			if (flags & TH_SYN)
608 				tp->snd_nxt++;
609 			if (flags & TH_FIN) {
610 				tp->snd_nxt++;
611 				tp->t_flags |= TF_SENTFIN;
612 			}
613 		}
614 		tp->snd_nxt += len;
615 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
616 			tp->snd_max = tp->snd_nxt;
617 			/*
618 			 * Time this transmission if not a retransmission and
619 			 * not currently timing anything.
620 			 */
621 			if (tp->t_rtt == 0) {
622 				tp->t_rtt = 1;
623 				tp->t_rtseq = startseq;
624 				tcpstat.tcps_segstimed++;
625 			}
626 		}
627 
628 		/*
629 		 * Set retransmit timer if not currently set,
630 		 * and not doing an ack or a keep-alive probe.
631 		 * Initial value for retransmit timer is smoothed
632 		 * round-trip time + 2 * round-trip time variance.
633 		 * Initialize shift counter which is used for backoff
634 		 * of retransmit time.
635 		 */
636 		if (tp->t_timer[TCPT_REXMT] == 0 &&
637 		    tp->snd_nxt != tp->snd_una) {
638 			tp->t_timer[TCPT_REXMT] = tp->t_rxtcur;
639 			if (tp->t_timer[TCPT_PERSIST]) {
640 				tp->t_timer[TCPT_PERSIST] = 0;
641 				tp->t_rxtshift = 0;
642 			}
643 		}
644 	} else
645 		if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
646 			tp->snd_max = tp->snd_nxt + len;
647 
648 #ifdef TCPDEBUG
649 	/*
650 	 * Trace.
651 	 */
652 	if (so->so_options & SO_DEBUG)
653 		tcp_trace(TA_OUTPUT, tp->t_state, tp, ti, 0);
654 #endif
655 
656 	/*
657 	 * Fill in IP length and desired time to live and
658 	 * send to IP level.  There should be a better way
659 	 * to handle ttl and tos; we could keep them in
660 	 * the template, but need a way to checksum without them.
661 	 */
662 	m->m_pkthdr.len = hdrlen + len;
663 #ifdef TUBA
664 	if (tp->t_tuba_pcb)
665 		error = tuba_output(m, tp);
666 	else
667 #endif
668     {
669 #if 1
670 	struct rtentry *rt;
671 #endif
672 	((struct ip *)ti)->ip_len = m->m_pkthdr.len;
673 	((struct ip *)ti)->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl;	/* XXX */
674 	((struct ip *)ti)->ip_tos = tp->t_inpcb->inp_ip.ip_tos;	/* XXX */
675 #if 1
676 	/*
677 	 * See if we should do MTU discovery.  We do it only if the following
678 	 * are true:
679 	 *	1) we have a valid route to the destination
680 	 *	2) the MTU is not locked (if it is, then discovery has been
681 	 *	   disabled)
682 	 */
683 	if ((rt = tp->t_inpcb->inp_route.ro_rt)
684 	    && rt->rt_flags & RTF_UP
685 	    && !(rt->rt_rmx.rmx_locks & RTV_MTU)) {
686 		((struct ip *)ti)->ip_off |= IP_DF;
687 	}
688 #endif
689 	error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
690 	    so->so_options & SO_DONTROUTE, 0);
691     }
692 	if (error) {
693 out:
694 		if (error == ENOBUFS) {
695 			tcp_quench(tp->t_inpcb, 0);
696 			return (0);
697 		}
698 #if 1
699 		if (error == EMSGSIZE) {
700 			/*
701 			 * ip_output() will have already fixed the route
702 			 * for us.  tcp_mtudisc() will, as its last action,
703 			 * initiate retransmission, so it is important to
704 			 * not do so here.
705 			 */
706 			tcp_mtudisc(tp->t_inpcb, 0);
707 			return 0;
708 		}
709 #endif
710 		if ((error == EHOSTUNREACH || error == ENETDOWN)
711 		    && TCPS_HAVERCVDSYN(tp->t_state)) {
712 			tp->t_softerror = error;
713 			return (0);
714 		}
715 		return (error);
716 	}
717 	tcpstat.tcps_sndtotal++;
718 
719 	/*
720 	 * Data sent (as far as we can tell).
721 	 * If this advertises a larger window than any other segment,
722 	 * then remember the size of the advertised window.
723 	 * Any pending ACK has now been sent.
724 	 */
725 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
726 		tp->rcv_adv = tp->rcv_nxt + win;
727 	tp->last_ack_sent = tp->rcv_nxt;
728 	tp->t_flags &= ~(TF_ACKNOW|TF_DELACK);
729 	if (sendalot)
730 		goto again;
731 	return (0);
732 }
733 
734 void
735 tcp_setpersist(tp)
736 	register struct tcpcb *tp;
737 {
738 	register t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
739 
740 	if (tp->t_timer[TCPT_REXMT])
741 		panic("tcp_output REXMT");
742 	/*
743 	 * Start/restart persistance timer.
744 	 */
745 	TCPT_RANGESET(tp->t_timer[TCPT_PERSIST],
746 	    t * tcp_backoff[tp->t_rxtshift],
747 	    TCPTV_PERSMIN, TCPTV_PERSMAX);
748 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
749 		tp->t_rxtshift++;
750 }
751