xref: /freebsd/sys/netinet/tcp_output.c (revision a8445737e740901f5f2c8d24c12ef7fc8b00134e)
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 1.30 1998/05/24 18:41:04 fenner Exp $
35  */
36 
37 #include "opt_tcpdebug.h"
38 
39 #include <stddef.h>
40 
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/mbuf.h>
44 #include <sys/protosw.h>
45 #include <sys/socket.h>
46 #include <sys/socketvar.h>
47 
48 #include <net/route.h>
49 
50 #include <netinet/in.h>
51 #include <netinet/in_systm.h>
52 #include <netinet/ip.h>
53 #include <netinet/in_pcb.h>
54 #include <netinet/ip_var.h>
55 #include <netinet/tcp.h>
56 #define	TCPOUTFLAGS
57 #include <netinet/tcp_fsm.h>
58 #include <netinet/tcp_seq.h>
59 #include <netinet/tcp_timer.h>
60 #include <netinet/tcp_var.h>
61 #include <netinet/tcpip.h>
62 #ifdef TCPDEBUG
63 #include <netinet/tcp_debug.h>
64 #endif
65 
66 #ifdef notyet
67 extern struct mbuf *m_copypack();
68 #endif
69 
70 
71 /*
72  * Tcp output routine: figure out what should be sent and send it.
73  */
74 int
75 tcp_output(tp)
76 	register struct tcpcb *tp;
77 {
78 	register struct socket *so = tp->t_inpcb->inp_socket;
79 	register long len, win;
80 	int off, flags, error;
81 	register struct mbuf *m;
82 	register struct tcpiphdr *ti;
83 	u_char opt[TCP_MAXOLEN];
84 	unsigned ipoptlen, optlen, hdrlen;
85 	int idle, sendalot;
86 	struct rmxp_tao *taop;
87 	struct rmxp_tao tao_noncached;
88 
89 	/*
90 	 * Determine length of data that should be transmitted,
91 	 * and flags that will be used.
92 	 * If there is some data or critical controls (SYN, RST)
93 	 * to send, then transmit; otherwise, investigate further.
94 	 */
95 	idle = (tp->snd_max == tp->snd_una);
96 	if (idle && tp->t_idle >= tp->t_rxtcur)
97 		/*
98 		 * We have been idle for "a while" and no acks are
99 		 * expected to clock out any data we send --
100 		 * slow start to get ack "clock" running again.
101 		 */
102 		tp->snd_cwnd = tp->t_maxseg;
103 again:
104 	sendalot = 0;
105 	off = tp->snd_nxt - tp->snd_una;
106 	win = min(tp->snd_wnd, tp->snd_cwnd);
107 
108 	flags = tcp_outflags[tp->t_state];
109 	/*
110 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
111 	 * state flags.
112 	 */
113 	if (tp->t_flags & TF_NEEDFIN)
114 		flags |= TH_FIN;
115 	if (tp->t_flags & TF_NEEDSYN)
116 		flags |= TH_SYN;
117 
118 	/*
119 	 * If in persist timeout with window of 0, send 1 byte.
120 	 * Otherwise, if window is small but nonzero
121 	 * and timer expired, we will send what we can
122 	 * and go to transmit state.
123 	 */
124 	if (tp->t_force) {
125 		if (win == 0) {
126 			/*
127 			 * If we still have some data to send, then
128 			 * clear the FIN bit.  Usually this would
129 			 * happen below when it realizes that we
130 			 * aren't sending all the data.  However,
131 			 * if we have exactly 1 byte of unset data,
132 			 * then it won't clear the FIN bit below,
133 			 * and if we are in persist state, we wind
134 			 * up sending the packet without recording
135 			 * that we sent the FIN bit.
136 			 *
137 			 * We can't just blindly clear the FIN bit,
138 			 * because if we don't have any more data
139 			 * to send then the probe will be the FIN
140 			 * itself.
141 			 */
142 			if (off < so->so_snd.sb_cc)
143 				flags &= ~TH_FIN;
144 			win = 1;
145 		} else {
146 			tp->t_timer[TCPT_PERSIST] = 0;
147 			tp->t_rxtshift = 0;
148 		}
149 	}
150 
151 	len = (long)ulmin(so->so_snd.sb_cc, win) - off;
152 
153 	if ((taop = tcp_gettaocache(tp->t_inpcb)) == NULL) {
154 		taop = &tao_noncached;
155 		bzero(taop, sizeof(*taop));
156 	}
157 
158 	/*
159 	 * Lop off SYN bit if it has already been sent.  However, if this
160 	 * is SYN-SENT state and if segment contains data and if we don't
161 	 * know that foreign host supports TAO, suppress sending segment.
162 	 */
163 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
164 		flags &= ~TH_SYN;
165 		off--, len++;
166 		if (len > 0 && tp->t_state == TCPS_SYN_SENT &&
167 		    taop->tao_ccsent == 0)
168 			return 0;
169 	}
170 
171 	/*
172 	 * Be careful not to send data and/or FIN on SYN segments
173 	 * in cases when no CC option will be sent.
174 	 * This measure is needed to prevent interoperability problems
175 	 * with not fully conformant TCP implementations.
176 	 */
177 	if ((flags & TH_SYN) &&
178 	    ((tp->t_flags & TF_NOOPT) || !(tp->t_flags & TF_REQ_CC) ||
179 	     ((flags & TH_ACK) && !(tp->t_flags & TF_RCVD_CC)))) {
180 		len = 0;
181 		flags &= ~TH_FIN;
182 	}
183 
184 	if (len < 0) {
185 		/*
186 		 * If FIN has been sent but not acked,
187 		 * but we haven't been called to retransmit,
188 		 * len will be -1.  Otherwise, window shrank
189 		 * after we sent into it.  If window shrank to 0,
190 		 * cancel pending retransmit, pull snd_nxt back
191 		 * to (closed) window, and set the persist timer
192 		 * if it isn't already going.  If the window didn't
193 		 * close completely, just wait for an ACK.
194 		 */
195 		len = 0;
196 		if (win == 0) {
197 			tp->t_timer[TCPT_REXMT] = 0;
198 			tp->t_rxtshift = 0;
199 			tp->snd_nxt = tp->snd_una;
200 			if (tp->t_timer[TCPT_PERSIST] == 0)
201 				tcp_setpersist(tp);
202 		}
203 	}
204 	if (len > tp->t_maxseg) {
205 		len = tp->t_maxseg;
206 		sendalot = 1;
207 	}
208 	if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
209 		flags &= ~TH_FIN;
210 
211 	win = sbspace(&so->so_rcv);
212 
213 	/*
214 	 * Sender silly window avoidance.  If connection is idle
215 	 * and can send all data, a maximum segment,
216 	 * at least a maximum default-size segment do it,
217 	 * or are forced, do it; otherwise don't bother.
218 	 * If peer's buffer is tiny, then send
219 	 * when window is at least half open.
220 	 * If retransmitting (possibly after persist timer forced us
221 	 * to send into a small window), then must resend.
222 	 */
223 	if (len) {
224 		if (len == tp->t_maxseg)
225 			goto send;
226 		if ((idle || tp->t_flags & TF_NODELAY) &&
227 		    (tp->t_flags & TF_NOPUSH) == 0 &&
228 		    len + off >= so->so_snd.sb_cc)
229 			goto send;
230 		if (tp->t_force)
231 			goto send;
232 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
233 			goto send;
234 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
235 			goto send;
236 	}
237 
238 	/*
239 	 * Compare available window to amount of window
240 	 * known to peer (as advertised window less
241 	 * next expected input).  If the difference is at least two
242 	 * max size segments, or at least 50% of the maximum possible
243 	 * window, then want to send a window update to peer.
244 	 */
245 	if (win > 0) {
246 		/*
247 		 * "adv" is the amount we can increase the window,
248 		 * taking into account that we are limited by
249 		 * TCP_MAXWIN << tp->rcv_scale.
250 		 */
251 		long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) -
252 			(tp->rcv_adv - tp->rcv_nxt);
253 
254 		if (adv >= (long) (2 * tp->t_maxseg))
255 			goto send;
256 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
257 			goto send;
258 	}
259 
260 	/*
261 	 * Send if we owe peer an ACK.
262 	 */
263 	if (tp->t_flags & TF_ACKNOW)
264 		goto send;
265 	if ((flags & TH_RST) ||
266 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
267 		goto send;
268 	if (SEQ_GT(tp->snd_up, tp->snd_una))
269 		goto send;
270 	/*
271 	 * If our state indicates that FIN should be sent
272 	 * and we have not yet done so, or we're retransmitting the FIN,
273 	 * then we need to send.
274 	 */
275 	if (flags & TH_FIN &&
276 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
277 		goto send;
278 
279 	/*
280 	 * TCP window updates are not reliable, rather a polling protocol
281 	 * using ``persist'' packets is used to insure receipt of window
282 	 * updates.  The three ``states'' for the output side are:
283 	 *	idle			not doing retransmits or persists
284 	 *	persisting		to move a small or zero window
285 	 *	(re)transmitting	and thereby not persisting
286 	 *
287 	 * tp->t_timer[TCPT_PERSIST]
288 	 *	is set when we are in persist state.
289 	 * tp->t_force
290 	 *	is set when we are called to send a persist packet.
291 	 * tp->t_timer[TCPT_REXMT]
292 	 *	is set when we are retransmitting
293 	 * The output side is idle when both timers are zero.
294 	 *
295 	 * If send window is too small, there is data to transmit, and no
296 	 * retransmit or persist is pending, then go to persist state.
297 	 * If nothing happens soon, send when timer expires:
298 	 * if window is nonzero, transmit what we can,
299 	 * otherwise force out a byte.
300 	 */
301 	if (so->so_snd.sb_cc && tp->t_timer[TCPT_REXMT] == 0 &&
302 	    tp->t_timer[TCPT_PERSIST] == 0) {
303 		tp->t_rxtshift = 0;
304 		tcp_setpersist(tp);
305 	}
306 
307 	/*
308 	 * No reason to send a segment, just return.
309 	 */
310 	return (0);
311 
312 send:
313 	/*
314 	 * Before ESTABLISHED, force sending of initial options
315 	 * unless TCP set not to do any options.
316 	 * NOTE: we assume that the IP/TCP header plus TCP options
317 	 * always fit in a single mbuf, leaving room for a maximum
318 	 * link header, i.e.
319 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MHLEN
320 	 */
321 	optlen = 0;
322 	hdrlen = sizeof (struct tcpiphdr);
323 	if (flags & TH_SYN) {
324 		tp->snd_nxt = tp->iss;
325 		if ((tp->t_flags & TF_NOOPT) == 0) {
326 			u_short mss;
327 
328 			opt[0] = TCPOPT_MAXSEG;
329 			opt[1] = TCPOLEN_MAXSEG;
330 			mss = htons((u_short) tcp_mssopt(tp));
331 			(void)memcpy(opt + 2, &mss, sizeof(mss));
332 			optlen = TCPOLEN_MAXSEG;
333 
334 			if ((tp->t_flags & TF_REQ_SCALE) &&
335 			    ((flags & TH_ACK) == 0 ||
336 			    (tp->t_flags & TF_RCVD_SCALE))) {
337 				*((u_int32_t *)(opt + optlen)) = htonl(
338 					TCPOPT_NOP << 24 |
339 					TCPOPT_WINDOW << 16 |
340 					TCPOLEN_WINDOW << 8 |
341 					tp->request_r_scale);
342 				optlen += 4;
343 			}
344 		}
345  	}
346 
347  	/*
348 	 * Send a timestamp and echo-reply if this is a SYN and our side
349 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
350 	 * and our peer have sent timestamps in our SYN's.
351  	 */
352  	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
353  	    (flags & TH_RST) == 0 &&
354 	    ((flags & TH_ACK) == 0 ||
355 	     (tp->t_flags & TF_RCVD_TSTMP))) {
356 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
357 
358  		/* Form timestamp option as shown in appendix A of RFC 1323. */
359  		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
360  		*lp++ = htonl(tcp_now);
361  		*lp   = htonl(tp->ts_recent);
362  		optlen += TCPOLEN_TSTAMP_APPA;
363  	}
364 
365  	/*
366 	 * Send `CC-family' options if our side wants to use them (TF_REQ_CC),
367 	 * options are allowed (!TF_NOOPT) and it's not a RST.
368  	 */
369  	if ((tp->t_flags & (TF_REQ_CC|TF_NOOPT)) == TF_REQ_CC &&
370  	     (flags & TH_RST) == 0) {
371 		switch (flags & (TH_SYN|TH_ACK)) {
372 		/*
373 		 * This is a normal ACK, send CC if we received CC before
374 		 * from our peer.
375 		 */
376 		case TH_ACK:
377 			if (!(tp->t_flags & TF_RCVD_CC))
378 				break;
379 			/*FALLTHROUGH*/
380 
381 		/*
382 		 * We can only get here in T/TCP's SYN_SENT* state, when
383 		 * we're a sending a non-SYN segment without waiting for
384 		 * the ACK of our SYN.  A check above assures that we only
385 		 * do this if our peer understands T/TCP.
386 		 */
387 		case 0:
388 			opt[optlen++] = TCPOPT_NOP;
389 			opt[optlen++] = TCPOPT_NOP;
390 			opt[optlen++] = TCPOPT_CC;
391 			opt[optlen++] = TCPOLEN_CC;
392 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
393 
394 			optlen += 4;
395 			break;
396 
397 		/*
398 		 * This is our initial SYN, check whether we have to use
399 		 * CC or CC.new.
400 		 */
401 		case TH_SYN:
402 			opt[optlen++] = TCPOPT_NOP;
403 			opt[optlen++] = TCPOPT_NOP;
404 			opt[optlen++] = tp->t_flags & TF_SENDCCNEW ?
405 						TCPOPT_CCNEW : TCPOPT_CC;
406 			opt[optlen++] = TCPOLEN_CC;
407 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
408  			optlen += 4;
409 			break;
410 
411 		/*
412 		 * This is a SYN,ACK; send CC and CC.echo if we received
413 		 * CC from our peer.
414 		 */
415 		case (TH_SYN|TH_ACK):
416 			if (tp->t_flags & TF_RCVD_CC) {
417 				opt[optlen++] = TCPOPT_NOP;
418 				opt[optlen++] = TCPOPT_NOP;
419 				opt[optlen++] = TCPOPT_CC;
420 				opt[optlen++] = TCPOLEN_CC;
421 				*(u_int32_t *)&opt[optlen] =
422 					htonl(tp->cc_send);
423 				optlen += 4;
424 				opt[optlen++] = TCPOPT_NOP;
425 				opt[optlen++] = TCPOPT_NOP;
426 				opt[optlen++] = TCPOPT_CCECHO;
427 				opt[optlen++] = TCPOLEN_CC;
428 				*(u_int32_t *)&opt[optlen] =
429 					htonl(tp->cc_recv);
430 				optlen += 4;
431 			}
432 			break;
433 		}
434  	}
435 
436  	hdrlen += optlen;
437 
438 	if (tp->t_inpcb->inp_options) {
439 		ipoptlen = tp->t_inpcb->inp_options->m_len -
440 				offsetof(struct ipoption, ipopt_list);
441 	} else {
442 		ipoptlen = 0;
443 	}
444 
445 	/*
446 	 * Adjust data length if insertion of options will
447 	 * bump the packet length beyond the t_maxopd length.
448 	 * Clear the FIN bit because we cut off the tail of
449 	 * the segment.
450 	 */
451 	if (len + optlen + ipoptlen > tp->t_maxopd) {
452 		/*
453 		 * If there is still more to send, don't close the connection.
454 		 */
455 		flags &= ~TH_FIN;
456 		len = tp->t_maxopd - optlen - ipoptlen;
457 		sendalot = 1;
458 	}
459 
460 /*#ifdef DIAGNOSTIC*/
461  	if (max_linkhdr + hdrlen > MHLEN)
462 		panic("tcphdr too big");
463 /*#endif*/
464 
465 	/*
466 	 * Grab a header mbuf, attaching a copy of data to
467 	 * be transmitted, and initialize the header from
468 	 * the template for sends on this connection.
469 	 */
470 	if (len) {
471 		if (tp->t_force && len == 1)
472 			tcpstat.tcps_sndprobe++;
473 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
474 			tcpstat.tcps_sndrexmitpack++;
475 			tcpstat.tcps_sndrexmitbyte += len;
476 		} else {
477 			tcpstat.tcps_sndpack++;
478 			tcpstat.tcps_sndbyte += len;
479 		}
480 #ifdef notyet
481 		if ((m = m_copypack(so->so_snd.sb_mb, off,
482 		    (int)len, max_linkhdr + hdrlen)) == 0) {
483 			error = ENOBUFS;
484 			goto out;
485 		}
486 		/*
487 		 * m_copypack left space for our hdr; use it.
488 		 */
489 		m->m_len += hdrlen;
490 		m->m_data -= hdrlen;
491 #else
492 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
493 		if (m == NULL) {
494 			error = ENOBUFS;
495 			goto out;
496 		}
497 		m->m_data += max_linkhdr;
498 		m->m_len = hdrlen;
499 		if (len <= MHLEN - hdrlen - max_linkhdr) {
500 			m_copydata(so->so_snd.sb_mb, off, (int) len,
501 			    mtod(m, caddr_t) + hdrlen);
502 			m->m_len += len;
503 		} else {
504 			m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len);
505 			if (m->m_next == 0) {
506 				(void) m_free(m);
507 				error = ENOBUFS;
508 				goto out;
509 			}
510 		}
511 #endif
512 		/*
513 		 * If we're sending everything we've got, set PUSH.
514 		 * (This will keep happy those implementations which only
515 		 * give data to the user when a buffer fills or
516 		 * a PUSH comes in.)
517 		 */
518 		if (off + len == so->so_snd.sb_cc)
519 			flags |= TH_PUSH;
520 	} else {
521 		if (tp->t_flags & TF_ACKNOW)
522 			tcpstat.tcps_sndacks++;
523 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
524 			tcpstat.tcps_sndctrl++;
525 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
526 			tcpstat.tcps_sndurg++;
527 		else
528 			tcpstat.tcps_sndwinup++;
529 
530 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
531 		if (m == NULL) {
532 			error = ENOBUFS;
533 			goto out;
534 		}
535 		m->m_data += max_linkhdr;
536 		m->m_len = hdrlen;
537 	}
538 	m->m_pkthdr.rcvif = (struct ifnet *)0;
539 	ti = mtod(m, struct tcpiphdr *);
540 	if (tp->t_template == 0)
541 		panic("tcp_output");
542 	(void)memcpy(ti, tp->t_template, sizeof (struct tcpiphdr));
543 
544 	/*
545 	 * Fill in fields, remembering maximum advertised
546 	 * window for use in delaying messages about window sizes.
547 	 * If resending a FIN, be sure not to use a new sequence number.
548 	 */
549 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
550 	    tp->snd_nxt == tp->snd_max)
551 		tp->snd_nxt--;
552 	/*
553 	 * If we are doing retransmissions, then snd_nxt will
554 	 * not reflect the first unsent octet.  For ACK only
555 	 * packets, we do not want the sequence number of the
556 	 * retransmitted packet, we want the sequence number
557 	 * of the next unsent octet.  So, if there is no data
558 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
559 	 * when filling in ti_seq.  But if we are in persist
560 	 * state, snd_max might reflect one byte beyond the
561 	 * right edge of the window, so use snd_nxt in that
562 	 * case, since we know we aren't doing a retransmission.
563 	 * (retransmit and persist are mutually exclusive...)
564 	 */
565 	if (len || (flags & (TH_SYN|TH_FIN)) || tp->t_timer[TCPT_PERSIST])
566 		ti->ti_seq = htonl(tp->snd_nxt);
567 	else
568 		ti->ti_seq = htonl(tp->snd_max);
569 	ti->ti_ack = htonl(tp->rcv_nxt);
570 	if (optlen) {
571 		bcopy(opt, ti + 1, optlen);
572 		ti->ti_off = (sizeof (struct tcphdr) + optlen) >> 2;
573 	}
574 	ti->ti_flags = flags;
575 	/*
576 	 * Calculate receive window.  Don't shrink window,
577 	 * but avoid silly window syndrome.
578 	 */
579 	if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)tp->t_maxseg)
580 		win = 0;
581 	if (win < (long)(tp->rcv_adv - tp->rcv_nxt))
582 		win = (long)(tp->rcv_adv - tp->rcv_nxt);
583 	if (win > (long)TCP_MAXWIN << tp->rcv_scale)
584 		win = (long)TCP_MAXWIN << tp->rcv_scale;
585 	ti->ti_win = htons((u_short) (win>>tp->rcv_scale));
586 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
587 		ti->ti_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
588 		ti->ti_flags |= TH_URG;
589 	} else
590 		/*
591 		 * If no urgent pointer to send, then we pull
592 		 * the urgent pointer to the left edge of the send window
593 		 * so that it doesn't drift into the send window on sequence
594 		 * number wraparound.
595 		 */
596 		tp->snd_up = tp->snd_una;		/* drag it along */
597 
598 	/*
599 	 * Put TCP length in extended header, and then
600 	 * checksum extended header and data.
601 	 */
602 	if (len + optlen)
603 		ti->ti_len = htons((u_short)(sizeof (struct tcphdr) +
604 		    optlen + len));
605 	ti->ti_sum = in_cksum(m, (int)(hdrlen + len));
606 
607 	/*
608 	 * In transmit state, time the transmission and arrange for
609 	 * the retransmit.  In persist state, just set snd_max.
610 	 */
611 	if (tp->t_force == 0 || tp->t_timer[TCPT_PERSIST] == 0) {
612 		tcp_seq startseq = tp->snd_nxt;
613 
614 		/*
615 		 * Advance snd_nxt over sequence space of this segment.
616 		 */
617 		if (flags & (TH_SYN|TH_FIN)) {
618 			if (flags & TH_SYN)
619 				tp->snd_nxt++;
620 			if (flags & TH_FIN) {
621 				tp->snd_nxt++;
622 				tp->t_flags |= TF_SENTFIN;
623 			}
624 		}
625 		tp->snd_nxt += len;
626 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
627 			tp->snd_max = tp->snd_nxt;
628 			/*
629 			 * Time this transmission if not a retransmission and
630 			 * not currently timing anything.
631 			 */
632 			if (tp->t_rtt == 0) {
633 				tp->t_rtt = 1;
634 				tp->t_rtseq = startseq;
635 				tcpstat.tcps_segstimed++;
636 			}
637 		}
638 
639 		/*
640 		 * Set retransmit timer if not currently set,
641 		 * and not doing an ack or a keep-alive probe.
642 		 * Initial value for retransmit timer is smoothed
643 		 * round-trip time + 2 * round-trip time variance.
644 		 * Initialize shift counter which is used for backoff
645 		 * of retransmit time.
646 		 */
647 		if (tp->t_timer[TCPT_REXMT] == 0 &&
648 		    tp->snd_nxt != tp->snd_una) {
649 			tp->t_timer[TCPT_REXMT] = tp->t_rxtcur;
650 			if (tp->t_timer[TCPT_PERSIST]) {
651 				tp->t_timer[TCPT_PERSIST] = 0;
652 				tp->t_rxtshift = 0;
653 			}
654 		}
655 	} else
656 		if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
657 			tp->snd_max = tp->snd_nxt + len;
658 
659 #ifdef TCPDEBUG
660 	/*
661 	 * Trace.
662 	 */
663 	if (so->so_options & SO_DEBUG)
664 		tcp_trace(TA_OUTPUT, tp->t_state, tp, ti, 0);
665 #endif
666 
667 	/*
668 	 * Fill in IP length and desired time to live and
669 	 * send to IP level.  There should be a better way
670 	 * to handle ttl and tos; we could keep them in
671 	 * the template, but need a way to checksum without them.
672 	 */
673 	m->m_pkthdr.len = hdrlen + len;
674     {
675 #if 1
676 	struct rtentry *rt;
677 #endif
678 	((struct ip *)ti)->ip_len = m->m_pkthdr.len;
679 	((struct ip *)ti)->ip_ttl = tp->t_inpcb->inp_ip_ttl;	/* XXX */
680 	((struct ip *)ti)->ip_tos = tp->t_inpcb->inp_ip_tos;	/* XXX */
681 #if 1
682 	/*
683 	 * See if we should do MTU discovery.  We do it only if the following
684 	 * are true:
685 	 *	1) we have a valid route to the destination
686 	 *	2) the MTU is not locked (if it is, then discovery has been
687 	 *	   disabled)
688 	 */
689 	if ((rt = tp->t_inpcb->inp_route.ro_rt)
690 	    && rt->rt_flags & RTF_UP
691 	    && !(rt->rt_rmx.rmx_locks & RTV_MTU)) {
692 		((struct ip *)ti)->ip_off |= IP_DF;
693 	}
694 #endif
695 	error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
696 	    so->so_options & SO_DONTROUTE, 0);
697     }
698 	if (error) {
699 out:
700 		if (error == ENOBUFS) {
701 			tcp_quench(tp->t_inpcb, 0);
702 			return (0);
703 		}
704 #if 1
705 		if (error == EMSGSIZE) {
706 			/*
707 			 * ip_output() will have already fixed the route
708 			 * for us.  tcp_mtudisc() will, as its last action,
709 			 * initiate retransmission, so it is important to
710 			 * not do so here.
711 			 */
712 			tcp_mtudisc(tp->t_inpcb, 0);
713 			return 0;
714 		}
715 #endif
716 		if ((error == EHOSTUNREACH || error == ENETDOWN)
717 		    && TCPS_HAVERCVDSYN(tp->t_state)) {
718 			tp->t_softerror = error;
719 			return (0);
720 		}
721 		return (error);
722 	}
723 	tcpstat.tcps_sndtotal++;
724 
725 	/*
726 	 * Data sent (as far as we can tell).
727 	 * If this advertises a larger window than any other segment,
728 	 * then remember the size of the advertised window.
729 	 * Any pending ACK has now been sent.
730 	 */
731 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
732 		tp->rcv_adv = tp->rcv_nxt + win;
733 	tp->last_ack_sent = tp->rcv_nxt;
734 	tp->t_flags &= ~(TF_ACKNOW|TF_DELACK);
735 	if (sendalot)
736 		goto again;
737 	return (0);
738 }
739 
740 void
741 tcp_setpersist(tp)
742 	register struct tcpcb *tp;
743 {
744 	register int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
745 
746 	if (tp->t_timer[TCPT_REXMT])
747 		panic("tcp_output REXMT");
748 	/*
749 	 * Start/restart persistance timer.
750 	 */
751 	TCPT_RANGESET(tp->t_timer[TCPT_PERSIST],
752 	    t * tcp_backoff[tp->t_rxtshift],
753 	    TCPTV_PERSMIN, TCPTV_PERSMAX);
754 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
755 		tp->t_rxtshift++;
756 }
757