xref: /dflybsd-src/sys/netinet/tcp_output.c (revision ece77bbaa23bf75f3b7bb9d110e2a795e3112878)
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  * $FreeBSD: src/sys/netinet/tcp_output.c,v 1.39.2.20 2003/01/29 22:45:36 hsu Exp $
35  * $DragonFly: src/sys/netinet/tcp_output.c,v 1.9 2004/02/14 21:12:39 dillon Exp $
36  */
37 
38 #include "opt_inet6.h"
39 #include "opt_ipsec.h"
40 #include "opt_tcpdebug.h"
41 
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/kernel.h>
45 #include <sys/sysctl.h>
46 #include <sys/mbuf.h>
47 #include <sys/domain.h>
48 #include <sys/protosw.h>
49 #include <sys/socket.h>
50 #include <sys/socketvar.h>
51 #include <sys/in_cksum.h>
52 
53 #include <net/route.h>
54 
55 #include <netinet/in.h>
56 #include <netinet/in_systm.h>
57 #include <netinet/ip.h>
58 #include <netinet/in_pcb.h>
59 #include <netinet/ip_var.h>
60 #include <netinet6/in6_pcb.h>
61 #include <netinet/ip6.h>
62 #include <netinet6/ip6_var.h>
63 #include <netinet/tcp.h>
64 #define	TCPOUTFLAGS
65 #include <netinet/tcp_fsm.h>
66 #include <netinet/tcp_seq.h>
67 #include <netinet/tcp_timer.h>
68 #include <netinet/tcp_var.h>
69 #include <netinet/tcpip.h>
70 #ifdef TCPDEBUG
71 #include <netinet/tcp_debug.h>
72 #endif
73 
74 #ifdef IPSEC
75 #include <netinet6/ipsec.h>
76 #endif /*IPSEC*/
77 
78 #ifdef FAST_IPSEC
79 #include <netipsec/ipsec.h>
80 #define	IPSEC
81 #endif /*FAST_IPSEC*/
82 
83 #ifdef notyet
84 extern struct mbuf *m_copypack();
85 #endif
86 
87 int path_mtu_discovery = 1;
88 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
89 	&path_mtu_discovery, 1, "Enable Path MTU Discovery");
90 
91 int ss_fltsz = 1;
92 SYSCTL_INT(_net_inet_tcp, OID_AUTO, slowstart_flightsize, CTLFLAG_RW,
93 	&ss_fltsz, 1, "Slow start flight size");
94 
95 int ss_fltsz_local = 4;
96 SYSCTL_INT(_net_inet_tcp, OID_AUTO, local_slowstart_flightsize, CTLFLAG_RW,
97 	&ss_fltsz_local, 1, "Slow start flight size for local networks");
98 
99 int	tcp_do_newreno = 1;
100 SYSCTL_INT(_net_inet_tcp, OID_AUTO, newreno, CTLFLAG_RW, &tcp_do_newreno,
101 	0, "Enable NewReno Algorithms");
102 
103 /*
104  * Tcp output routine: figure out what should be sent and send it.
105  */
106 int
107 tcp_output(tp)
108 	struct tcpcb *tp;
109 {
110 	struct inpcb * const inp = tp->t_inpcb;
111 	struct socket *so = inp->inp_socket;
112 	long len, recvwin, sendwin;
113 	int off, flags, error;
114 	struct mbuf *m;
115 	struct ip *ip = NULL;
116 	struct ipovly *ipov = NULL;
117 	struct tcphdr *th;
118 	u_char opt[TCP_MAXOLEN];
119 	unsigned ipoptlen, optlen, hdrlen;
120 	int idle, sendalot;
121 	struct ip6_hdr *ip6 = NULL;
122 #ifdef INET6
123 	const boolean_t isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
124 #else
125 	const boolean_t isipv6 = FALSE;
126 #endif
127 	struct rmxp_tao *taop;
128 
129 	/*
130 	 * Determine length of data that should be transmitted,
131 	 * and flags that will be used.
132 	 * If there is some data or critical controls (SYN, RST)
133 	 * to send, then transmit; otherwise, investigate further.
134 	 */
135 	if ((tp->snd_max == tp->snd_una) &&
136 	    (ticks - tp->t_rcvtime) >= tp->t_rxtcur) {
137 		/*
138 		 * We have been idle for "a while" and no acks are
139 		 * expected to clock out any data we send --
140 		 * slow start to get ack "clock" running again.
141 		 *
142 		 * Set the slow-start flight size depending on whether
143 		 * this is a local network or not.
144 		 */
145 		if ((isipv6 && in6_localaddr(&inp->in6p_faddr)) ||
146 		    (!isipv6 && in_localaddr(inp->inp_faddr)))
147 			tp->snd_cwnd = tp->t_maxseg * ss_fltsz_local;
148 		else
149 			tp->snd_cwnd = tp->t_maxseg * ss_fltsz;
150 	}
151 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
152 	if (idle && (tp->t_flags & TF_MORETOCOME))
153 		tp->t_flags |= TF_LASTIDLE;
154 	else
155 		tp->t_flags &= ~TF_LASTIDLE;
156 again:
157 	sendalot = 0;
158 	off = tp->snd_nxt - tp->snd_una;
159 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
160 	sendwin = min(sendwin, tp->snd_bwnd);
161 
162 	flags = tcp_outflags[tp->t_state];
163 	/*
164 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
165 	 * state flags.
166 	 */
167 	if (tp->t_flags & TF_NEEDFIN)
168 		flags |= TH_FIN;
169 	if (tp->t_flags & TF_NEEDSYN)
170 		flags |= TH_SYN;
171 
172 	/*
173 	 * If in persist timeout with window of 0, send 1 byte.
174 	 * Otherwise, if window is small but nonzero
175 	 * and timer expired, we will send what we can
176 	 * and go to transmit state.
177 	 */
178 	if (tp->t_force) {
179 		if (sendwin == 0) {
180 			/*
181 			 * If we still have some data to send, then
182 			 * clear the FIN bit.  Usually this would
183 			 * happen below when it realizes that we
184 			 * aren't sending all the data.  However,
185 			 * if we have exactly 1 byte of unsent data,
186 			 * then it won't clear the FIN bit below,
187 			 * and if we are in persist state, we wind
188 			 * up sending the packet without recording
189 			 * that we sent the FIN bit.
190 			 *
191 			 * We can't just blindly clear the FIN bit,
192 			 * because if we don't have any more data
193 			 * to send then the probe will be the FIN
194 			 * itself.
195 			 */
196 			if (off < so->so_snd.sb_cc)
197 				flags &= ~TH_FIN;
198 			sendwin = 1;
199 		} else {
200 			callout_stop(tp->tt_persist);
201 			tp->t_rxtshift = 0;
202 		}
203 	}
204 
205 	/*
206 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
207 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
208 	 * a negative length.  This can also occur when TCP opens up
209 	 * its congestion window while receiving additional duplicate
210 	 * acks after fast-retransmit because TCP will reset snd_nxt
211 	 * to snd_max after the fast-retransmit.
212 	 *
213 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
214 	 * be set to snd_una, the offset will be 0, and the length may
215 	 * wind up 0.
216 	 */
217 	len = (long)ulmin(so->so_snd.sb_cc, sendwin) - off;
218 
219 	/*
220 	 * Lop off SYN bit if it has already been sent.  However, if this
221 	 * is SYN-SENT state and if segment contains data and if we don't
222 	 * know that foreign host supports TAO, suppress sending segment.
223 	 */
224 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
225 		flags &= ~TH_SYN;
226 		off--, len++;
227 		if (len > 0 && tp->t_state == TCPS_SYN_SENT &&
228 		    ((taop = tcp_gettaocache(&inp->inp_inc)) == NULL ||
229 		     taop->tao_ccsent == 0))
230 			return 0;
231 	}
232 
233 	/*
234 	 * Be careful not to send data and/or FIN on SYN segments
235 	 * in cases when no CC option will be sent.
236 	 * This measure is needed to prevent interoperability problems
237 	 * with not fully conformant TCP implementations.
238 	 */
239 	if ((flags & TH_SYN) &&
240 	    ((tp->t_flags & TF_NOOPT) || !(tp->t_flags & TF_REQ_CC) ||
241 	     ((flags & TH_ACK) && !(tp->t_flags & TF_RCVD_CC)))) {
242 		len = 0;
243 		flags &= ~TH_FIN;
244 	}
245 
246 	if (len < 0) {
247 		/*
248 		 * If FIN has been sent but not acked,
249 		 * but we haven't been called to retransmit,
250 		 * len will be < 0.  Otherwise, window shrank
251 		 * after we sent into it.  If window shrank to 0,
252 		 * cancel pending retransmit, pull snd_nxt back
253 		 * to (closed) window, and set the persist timer
254 		 * if it isn't already going.  If the window didn't
255 		 * close completely, just wait for an ACK.
256 		 */
257 		len = 0;
258 		if (sendwin == 0) {
259 			callout_stop(tp->tt_rexmt);
260 			tp->t_rxtshift = 0;
261 			tp->snd_nxt = tp->snd_una;
262 			if (!callout_active(tp->tt_persist))
263 				tcp_setpersist(tp);
264 		}
265 	}
266 
267 	/*
268 	 * len will be >= 0 after this point.  Truncate to the maximum
269 	 * segment length and ensure that FIN is removed if the length
270 	 * no longer contains the last data byte.
271 	 */
272 	if (len > tp->t_maxseg) {
273 		len = tp->t_maxseg;
274 		sendalot = 1;
275 	}
276 	if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
277 		flags &= ~TH_FIN;
278 
279 	recvwin = sbspace(&so->so_rcv);
280 
281 	/*
282 	 * Sender silly window avoidance.   We transmit under the following
283 	 * conditions when len is non-zero:
284 	 *
285 	 *	- We have a full segment
286 	 *	- This is the last buffer in a write()/send() and we are
287 	 *	  either idle or running NODELAY
288 	 *	- we've timed out (e.g. persist timer)
289 	 *	- we have more then 1/2 the maximum send window's worth of
290 	 *	  data (receiver may be limited the window size)
291 	 *	- we need to retransmit
292 	 */
293 	if (len) {
294 		if (len == tp->t_maxseg)
295 			goto send;
296 		/*
297 		 * NOTE! on localhost connections an 'ack' from the remote
298 		 * end may occur synchronously with the output and cause
299 		 * us to flush a buffer queued with moretocome.  XXX
300 		 *
301 		 * note: the len + off check is almost certainly unnecessary.
302 		 */
303 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
304 		    (idle || (tp->t_flags & TF_NODELAY)) &&
305 		    len + off >= so->so_snd.sb_cc &&
306 		    !(tp->t_flags & TF_NOPUSH)) {
307 			goto send;
308 		}
309 		if (tp->t_force)			/* typ. timeout case */
310 			goto send;
311 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
312 			goto send;
313 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
314 			goto send;
315 	}
316 
317 	/*
318 	 * Compare available window to amount of window
319 	 * known to peer (as advertised window less
320 	 * next expected input).  If the difference is at least two
321 	 * max size segments, or at least 50% of the maximum possible
322 	 * window, then want to send a window update to peer.
323 	 */
324 	if (recvwin > 0) {
325 		/*
326 		 * "adv" is the amount we can increase the window,
327 		 * taking into account that we are limited by
328 		 * TCP_MAXWIN << tp->rcv_scale.
329 		 */
330 		long adv = min(recvwin, (long)TCP_MAXWIN << tp->rcv_scale) -
331 			(tp->rcv_adv - tp->rcv_nxt);
332 
333 		if (adv >= (long) (2 * tp->t_maxseg))
334 			goto send;
335 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
336 			goto send;
337 	}
338 
339 	/*
340 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
341 	 * is also a catch-all for the retransmit timer timeout case.
342 	 */
343 	if (tp->t_flags & TF_ACKNOW)
344 		goto send;
345 	if ((flags & TH_RST) ||
346 	    ((flags & TH_SYN) && !(tp->t_flags & TF_NEEDSYN)))
347 		goto send;
348 	if (SEQ_GT(tp->snd_up, tp->snd_una))
349 		goto send;
350 	/*
351 	 * If our state indicates that FIN should be sent
352 	 * and we have not yet done so, then we need to send.
353 	 */
354 	if (flags & TH_FIN &&
355 	    (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una))
356 		goto send;
357 
358 	/*
359 	 * TCP window updates are not reliable, rather a polling protocol
360 	 * using ``persist'' packets is used to insure receipt of window
361 	 * updates.  The three ``states'' for the output side are:
362 	 *	idle			not doing retransmits or persists
363 	 *	persisting		to move a small or zero window
364 	 *	(re)transmitting	and thereby not persisting
365 	 *
366 	 * callout_active(tp->tt_persist)
367 	 *	is true when we are in persist state.
368 	 * tp->t_force
369 	 *	is set when we are called to send a persist packet.
370 	 * callout_active(tp->tt_rexmt)
371 	 *	is set when we are retransmitting
372 	 * The output side is idle when both timers are zero.
373 	 *
374 	 * If send window is too small, there is data to transmit, and no
375 	 * retransmit or persist is pending, then go to persist state.
376 	 * If nothing happens soon, send when timer expires:
377 	 * if window is nonzero, transmit what we can,
378 	 * otherwise force out a byte.
379 	 */
380 	if (so->so_snd.sb_cc && !callout_active(tp->tt_rexmt) &&
381 	    !callout_active(tp->tt_persist)) {
382 		tp->t_rxtshift = 0;
383 		tcp_setpersist(tp);
384 	}
385 
386 	/*
387 	 * No reason to send a segment, just return.
388 	 */
389 	return (0);
390 
391 send:
392 	/*
393 	 * Before ESTABLISHED, force sending of initial options
394 	 * unless TCP set not to do any options.
395 	 * NOTE: we assume that the IP/TCP header plus TCP options
396 	 * always fit in a single mbuf, leaving room for a maximum
397 	 * link header, i.e.
398 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
399 	 */
400 	optlen = 0;
401 	if (isipv6)
402 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
403 	else
404 		hdrlen = sizeof (struct tcpiphdr);
405 	if (flags & TH_SYN) {
406 		tp->snd_nxt = tp->iss;
407 		if (!(tp->t_flags & TF_NOOPT)) {
408 			u_short mss;
409 
410 			opt[0] = TCPOPT_MAXSEG;
411 			opt[1] = TCPOLEN_MAXSEG;
412 			mss = htons((u_short) tcp_mssopt(tp));
413 			(void)memcpy(opt + 2, &mss, sizeof(mss));
414 			optlen = TCPOLEN_MAXSEG;
415 
416 			if ((tp->t_flags & TF_REQ_SCALE) &&
417 			    (!(flags & TH_ACK) ||
418 			     (tp->t_flags & TF_RCVD_SCALE))) {
419 				*((u_int32_t *)(opt + optlen)) = htonl(
420 					TCPOPT_NOP << 24 |
421 					TCPOPT_WINDOW << 16 |
422 					TCPOLEN_WINDOW << 8 |
423 					tp->request_r_scale);
424 				optlen += 4;
425 			}
426 		}
427 	}
428 
429 	/*
430 	 * Send a timestamp and echo-reply if this is a SYN and our side
431 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
432 	 * and our peer have sent timestamps in our SYN's.
433 	 */
434 	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
435 	    !(flags & TH_RST) &&
436 	    (!(flags & TH_ACK) || (tp->t_flags & TF_RCVD_TSTMP))) {
437 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
438 
439 		/* Form timestamp option as shown in appendix A of RFC 1323. */
440 		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
441 		*lp++ = htonl(ticks);
442 		*lp   = htonl(tp->ts_recent);
443 		optlen += TCPOLEN_TSTAMP_APPA;
444 	}
445 
446 	/*
447 	 * Send `CC-family' options if our side wants to use them (TF_REQ_CC),
448 	 * options are allowed (!TF_NOOPT) and it's not a RST.
449 	 */
450 	if ((tp->t_flags & (TF_REQ_CC|TF_NOOPT)) == TF_REQ_CC &&
451 	     !(flags & TH_RST)) {
452 		switch (flags & (TH_SYN|TH_ACK)) {
453 		/*
454 		 * This is a normal ACK, send CC if we received CC before
455 		 * from our peer.
456 		 */
457 		case TH_ACK:
458 			if (!(tp->t_flags & TF_RCVD_CC))
459 				break;
460 			/*FALLTHROUGH*/
461 
462 		/*
463 		 * We can only get here in T/TCP's SYN_SENT* state, when
464 		 * we're a sending a non-SYN segment without waiting for
465 		 * the ACK of our SYN.  A check above assures that we only
466 		 * do this if our peer understands T/TCP.
467 		 */
468 		case 0:
469 			opt[optlen++] = TCPOPT_NOP;
470 			opt[optlen++] = TCPOPT_NOP;
471 			opt[optlen++] = TCPOPT_CC;
472 			opt[optlen++] = TCPOLEN_CC;
473 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
474 
475 			optlen += 4;
476 			break;
477 
478 		/*
479 		 * This is our initial SYN, check whether we have to use
480 		 * CC or CC.new.
481 		 */
482 		case TH_SYN:
483 			opt[optlen++] = TCPOPT_NOP;
484 			opt[optlen++] = TCPOPT_NOP;
485 			opt[optlen++] = tp->t_flags & TF_SENDCCNEW ?
486 						TCPOPT_CCNEW : TCPOPT_CC;
487 			opt[optlen++] = TCPOLEN_CC;
488 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
489 			optlen += 4;
490 			break;
491 
492 		/*
493 		 * This is a SYN,ACK; send CC and CC.echo if we received
494 		 * CC from our peer.
495 		 */
496 		case (TH_SYN|TH_ACK):
497 			if (tp->t_flags & TF_RCVD_CC) {
498 				opt[optlen++] = TCPOPT_NOP;
499 				opt[optlen++] = TCPOPT_NOP;
500 				opt[optlen++] = TCPOPT_CC;
501 				opt[optlen++] = TCPOLEN_CC;
502 				*(u_int32_t *)&opt[optlen] =
503 					htonl(tp->cc_send);
504 				optlen += 4;
505 				opt[optlen++] = TCPOPT_NOP;
506 				opt[optlen++] = TCPOPT_NOP;
507 				opt[optlen++] = TCPOPT_CCECHO;
508 				opt[optlen++] = TCPOLEN_CC;
509 				*(u_int32_t *)&opt[optlen] =
510 					htonl(tp->cc_recv);
511 				optlen += 4;
512 			}
513 			break;
514 		}
515 	}
516 
517 	hdrlen += optlen;
518 
519 	if (isipv6)
520 		ipoptlen = ip6_optlen(inp);
521 	else {
522 		if (inp->inp_options) {
523 			ipoptlen = inp->inp_options->m_len -
524 			    offsetof(struct ipoption, ipopt_list);
525 		} else {
526 			ipoptlen = 0;
527 		}
528 	}
529 #ifdef IPSEC
530 	ipoptlen += ipsec_hdrsiz_tcp(tp);
531 #endif
532 
533 	/*
534 	 * Adjust data length if insertion of options will
535 	 * bump the packet length beyond the t_maxopd length.
536 	 * Clear the FIN bit because we cut off the tail of
537 	 * the segment.
538 	 */
539 	if (len + optlen + ipoptlen > tp->t_maxopd) {
540 		/*
541 		 * If there is still more to send, don't close the connection.
542 		 */
543 		flags &= ~TH_FIN;
544 		len = tp->t_maxopd - optlen - ipoptlen;
545 		sendalot = 1;
546 	}
547 
548 #ifdef INET6
549 	KASSERT(max_linkhdr + hdrlen <= MCLBYTES, ("tcphdr too big"));
550 #else
551 	KASSERT(max_linkhdr + hdrlen <= MHLEN, ("tcphdr too big"));
552 #endif
553 
554 	/*
555 	 * Grab a header mbuf, attaching a copy of data to
556 	 * be transmitted, and initialize the header from
557 	 * the template for sends on this connection.
558 	 */
559 	if (len) {
560 		if (tp->t_force && len == 1)
561 			tcpstat.tcps_sndprobe++;
562 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
563 			tcpstat.tcps_sndrexmitpack++;
564 			tcpstat.tcps_sndrexmitbyte += len;
565 		} else {
566 			tcpstat.tcps_sndpack++;
567 			tcpstat.tcps_sndbyte += len;
568 		}
569 #ifdef notyet
570 		if ((m = m_copypack(so->so_snd.sb_mb, off,
571 		    (int)len, max_linkhdr + hdrlen)) == 0) {
572 			error = ENOBUFS;
573 			goto out;
574 		}
575 		/*
576 		 * m_copypack left space for our hdr; use it.
577 		 */
578 		m->m_len += hdrlen;
579 		m->m_data -= hdrlen;
580 #else
581 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
582 		if (m == NULL) {
583 			error = ENOBUFS;
584 			goto out;
585 		}
586 #ifdef INET6
587 		if (MHLEN < hdrlen + max_linkhdr) {
588 			MCLGET(m, M_DONTWAIT);
589 			if (!(m->m_flags & M_EXT)) {
590 				m_freem(m);
591 				error = ENOBUFS;
592 				goto out;
593 			}
594 		}
595 #endif
596 		m->m_data += max_linkhdr;
597 		m->m_len = hdrlen;
598 		if (len <= MHLEN - hdrlen - max_linkhdr) {
599 			m_copydata(so->so_snd.sb_mb, off, (int) len,
600 			    mtod(m, caddr_t) + hdrlen);
601 			m->m_len += len;
602 		} else {
603 			m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len);
604 			if (m->m_next == 0) {
605 				(void) m_free(m);
606 				error = ENOBUFS;
607 				goto out;
608 			}
609 		}
610 #endif
611 		/*
612 		 * If we're sending everything we've got, set PUSH.
613 		 * (This will keep happy those implementations which only
614 		 * give data to the user when a buffer fills or
615 		 * a PUSH comes in.)
616 		 */
617 		if (off + len == so->so_snd.sb_cc)
618 			flags |= TH_PUSH;
619 	} else {
620 		if (tp->t_flags & TF_ACKNOW)
621 			tcpstat.tcps_sndacks++;
622 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
623 			tcpstat.tcps_sndctrl++;
624 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
625 			tcpstat.tcps_sndurg++;
626 		else
627 			tcpstat.tcps_sndwinup++;
628 
629 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
630 		if (m == NULL) {
631 			error = ENOBUFS;
632 			goto out;
633 		}
634 		if (isipv6 &&
635 		    (hdrlen + max_linkhdr > MHLEN) && hdrlen <= MHLEN)
636 			MH_ALIGN(m, hdrlen);
637 		else
638 			m->m_data += max_linkhdr;
639 		m->m_len = hdrlen;
640 	}
641 	m->m_pkthdr.rcvif = (struct ifnet *)0;
642 	if (isipv6) {
643 		ip6 = mtod(m, struct ip6_hdr *);
644 		th = (struct tcphdr *)(ip6 + 1);
645 		tcp_fillheaders(tp, ip6, th);
646 	} else {
647 		ip = mtod(m, struct ip *);
648 		ipov = (struct ipovly *)ip;
649 		th = (struct tcphdr *)(ip + 1);
650 		/* this picks up the pseudo header (w/o the length) */
651 		tcp_fillheaders(tp, ip, th);
652 	}
653 
654 	/*
655 	 * Fill in fields, remembering maximum advertised
656 	 * window for use in delaying messages about window sizes.
657 	 * If resending a FIN, be sure not to use a new sequence number.
658 	 */
659 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
660 	    tp->snd_nxt == tp->snd_max)
661 		tp->snd_nxt--;
662 	/*
663 	 * If we are doing retransmissions, then snd_nxt will
664 	 * not reflect the first unsent octet.  For ACK only
665 	 * packets, we do not want the sequence number of the
666 	 * retransmitted packet, we want the sequence number
667 	 * of the next unsent octet.  So, if there is no data
668 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
669 	 * when filling in ti_seq.  But if we are in persist
670 	 * state, snd_max might reflect one byte beyond the
671 	 * right edge of the window, so use snd_nxt in that
672 	 * case, since we know we aren't doing a retransmission.
673 	 * (retransmit and persist are mutually exclusive...)
674 	 */
675 	if (len || (flags & (TH_SYN|TH_FIN)) || callout_active(tp->tt_persist))
676 		th->th_seq = htonl(tp->snd_nxt);
677 	else
678 		th->th_seq = htonl(tp->snd_max);
679 	th->th_ack = htonl(tp->rcv_nxt);
680 	if (optlen) {
681 		bcopy(opt, th + 1, optlen);
682 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
683 	}
684 	th->th_flags = flags;
685 	/*
686 	 * Calculate receive window.  Don't shrink window,
687 	 * but avoid silly window syndrome.
688 	 */
689 	if (recvwin < (long)(so->so_rcv.sb_hiwat / 4) &&
690 	    recvwin < (long)tp->t_maxseg)
691 		recvwin = 0;
692 	if (recvwin < (long)(tp->rcv_adv - tp->rcv_nxt))
693 		recvwin = (long)(tp->rcv_adv - tp->rcv_nxt);
694 	if (recvwin > (long)TCP_MAXWIN << tp->rcv_scale)
695 		recvwin = (long)TCP_MAXWIN << tp->rcv_scale;
696 	th->th_win = htons((u_short) (recvwin>>tp->rcv_scale));
697 
698 	/*
699 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
700 	 * a 0 window.  This may cause the remote transmitter to stall.  This
701 	 * flag tells soreceive() to disable delayed acknowledgements when
702 	 * draining the buffer.  This can occur if the receiver is attempting
703 	 * to read more data then can be buffered prior to transmitting on
704 	 * the connection.
705 	 */
706 	if (recvwin == 0)
707 		tp->t_flags |= TF_RXWIN0SENT;
708 	else
709 		tp->t_flags &= ~TF_RXWIN0SENT;
710 
711 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
712 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
713 		th->th_flags |= TH_URG;
714 	} else
715 		/*
716 		 * If no urgent pointer to send, then we pull
717 		 * the urgent pointer to the left edge of the send window
718 		 * so that it doesn't drift into the send window on sequence
719 		 * number wraparound.
720 		 */
721 		tp->snd_up = tp->snd_una;		/* drag it along */
722 
723 	/*
724 	 * Put TCP length in extended header, and then
725 	 * checksum extended header and data.
726 	 */
727 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
728 	if (isipv6) {
729 		/*
730 		 * ip6_plen is not need to be filled now, and will be filled
731 		 * in ip6_output.
732 		 */
733 		th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
734 				       sizeof(struct tcphdr) + optlen + len);
735 	} else {
736 		m->m_pkthdr.csum_flags = CSUM_TCP;
737 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
738 		if (len + optlen)
739 			th->th_sum = in_addword(th->th_sum,
740 						htons((u_short)(optlen + len)));
741 
742 		/* IP version must be set here for ipv4/ipv6 checking later */
743 		KASSERT(ip->ip_v == IPVERSION,
744 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
745 	}
746 
747 	/*
748 	 * In transmit state, time the transmission and arrange for
749 	 * the retransmit.  In persist state, just set snd_max.
750 	 */
751 	if (tp->t_force == 0 || !callout_active(tp->tt_persist)) {
752 		tcp_seq startseq = tp->snd_nxt;
753 
754 		/*
755 		 * Advance snd_nxt over sequence space of this segment.
756 		 */
757 		if (flags & (TH_SYN|TH_FIN)) {
758 			if (flags & TH_SYN)
759 				tp->snd_nxt++;
760 			if (flags & TH_FIN) {
761 				tp->snd_nxt++;
762 				tp->t_flags |= TF_SENTFIN;
763 			}
764 		}
765 		tp->snd_nxt += len;
766 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
767 			tp->snd_max = tp->snd_nxt;
768 			/*
769 			 * Time this transmission if not a retransmission and
770 			 * not currently timing anything.
771 			 */
772 			if (tp->t_rtttime == 0) {
773 				tp->t_rtttime = ticks;
774 				tp->t_rtseq = startseq;
775 				tcpstat.tcps_segstimed++;
776 			}
777 		}
778 
779 		/*
780 		 * Set retransmit timer if not currently set,
781 		 * and not doing a pure ack or a keep-alive probe.
782 		 * Initial value for retransmit timer is smoothed
783 		 * round-trip time + 2 * round-trip time variance.
784 		 * Initialize shift counter which is used for backoff
785 		 * of retransmit time.
786 		 */
787 		if (!callout_active(tp->tt_rexmt) &&
788 		    tp->snd_nxt != tp->snd_una) {
789 			if (callout_active(tp->tt_persist)) {
790 				callout_stop(tp->tt_persist);
791 				tp->t_rxtshift = 0;
792 			}
793 			callout_reset(tp->tt_rexmt, tp->t_rxtcur,
794 				      tcp_timer_rexmt, tp);
795 		}
796 	} else {
797 		/*
798 		 * Persist case, update snd_max but since we are in
799 		 * persist mode (no window) we do not update snd_nxt.
800 		 */
801 		int xlen = len;
802 		if (flags & TH_SYN)
803 			++xlen;
804 		if (flags & TH_FIN) {
805 			++xlen;
806 			tp->t_flags |= TF_SENTFIN;
807 		}
808 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
809 			tp->snd_max = tp->snd_nxt + xlen;
810 	}
811 
812 #ifdef TCPDEBUG
813 	/*
814 	 * Trace.
815 	 */
816 	if (so->so_options & SO_DEBUG)
817 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
818 #endif
819 
820 	/*
821 	 * Fill in IP length and desired time to live and
822 	 * send to IP level.  There should be a better way
823 	 * to handle ttl and tos; we could keep them in
824 	 * the template, but need a way to checksum without them.
825 	 */
826 	/*
827 	 * m->m_pkthdr.len should have been set before cksum calcuration,
828 	 * because in6_cksum() need it.
829 	 */
830 	if (isipv6) {
831 		/*
832 		 * we separately set hoplimit for every segment, since the
833 		 * user might want to change the value via setsockopt.
834 		 * Also, desired default hop limit might be changed via
835 		 * Neighbor Discovery.
836 		 */
837 		ip6->ip6_hlim = in6_selecthlim(inp,
838 		    (inp->in6p_route.ro_rt ?
839 		     inp->in6p_route.ro_rt->rt_ifp : NULL));
840 
841 		/* TODO: IPv6 IP6TOS_ECT bit on */
842 		error = ip6_output(m, inp->in6p_outputopts, &inp->in6p_route,
843 		    (so->so_options & SO_DONTROUTE), NULL, NULL, inp);
844 	} else {
845 		struct rtentry *rt;
846 		ip->ip_len = m->m_pkthdr.len;
847 #ifdef INET6
848 		if (INP_CHECK_SOCKAF(so, AF_INET6))
849 			ip->ip_ttl = in6_selecthlim(inp,
850 			    (inp->in6p_route.ro_rt ?
851 			     inp->in6p_route.ro_rt->rt_ifp : NULL));
852 		else
853 #endif
854 			ip->ip_ttl = inp->inp_ip_ttl;	/* XXX */
855 
856 		ip->ip_tos = inp->inp_ip_tos;	/* XXX */
857 		/*
858 		 * See if we should do MTU discovery.
859 		 * We do it only if the following are true:
860 		 *	1) we have a valid route to the destination
861 		 *	2) the MTU is not locked (if it is,
862 		 *	   then discovery has been disabled)
863 		 */
864 		if (path_mtu_discovery &&
865 		    (rt = inp->inp_route.ro_rt) && (rt->rt_flags & RTF_UP) &&
866 		    !(rt->rt_rmx.rmx_locks & RTV_MTU))
867 			ip->ip_off |= IP_DF;
868 
869 		error = ip_output(m, inp->inp_options, &inp->inp_route,
870 		    (so->so_options & SO_DONTROUTE), NULL, inp);
871 	}
872 	if (error) {
873 
874 		/*
875 		 * We know that the packet was lost, so back out the
876 		 * sequence number advance, if any.
877 		 */
878 		if (tp->t_force == 0 || !callout_active(tp->tt_persist)) {
879 			/*
880 			 * No need to check for TH_FIN here because
881 			 * the TF_SENTFIN flag handles that case.
882 			 */
883 			if (!(flags & TH_SYN))
884 				tp->snd_nxt -= len;
885 		}
886 
887 out:
888 		if (error == ENOBUFS) {
889 			/*
890 			 * If we can't send, make sure there is something
891 			 * to get us going again later.  Persist state
892 			 * is not necessarily right, but it is close enough.
893 			 */
894 			if (!callout_active(tp->tt_rexmt) &&
895 			    !callout_active(tp->tt_persist)) {
896 				tp->t_rxtshift = 0;
897 				tcp_setpersist(tp);
898 			}
899 			tcp_quench(inp, 0);
900 			return (0);
901 		}
902 		if (error == EMSGSIZE) {
903 			/*
904 			 * ip_output() will have already fixed the route
905 			 * for us.  tcp_mtudisc() will, as its last action,
906 			 * initiate retransmission, so it is important to
907 			 * not do so here.
908 			 */
909 			tcp_mtudisc(inp, 0);
910 			return 0;
911 		}
912 		if ((error == EHOSTUNREACH || error == ENETDOWN) &&
913 		    TCPS_HAVERCVDSYN(tp->t_state)) {
914 			tp->t_softerror = error;
915 			return (0);
916 		}
917 		return (error);
918 	}
919 	tcpstat.tcps_sndtotal++;
920 
921 	/*
922 	 * Data sent (as far as we can tell).
923 	 * If this advertises a larger window than any other segment,
924 	 * then remember the size of the advertised window.
925 	 * Any pending ACK has now been sent.
926 	 */
927 	if (recvwin > 0 && SEQ_GT(tp->rcv_nxt + recvwin, tp->rcv_adv))
928 		tp->rcv_adv = tp->rcv_nxt + recvwin;
929 	tp->last_ack_sent = tp->rcv_nxt;
930 	tp->t_flags &= ~TF_ACKNOW;
931 	if (tcp_delack_enabled)
932 		callout_stop(tp->tt_delack);
933 	if (sendalot)
934 		goto again;
935 	return (0);
936 }
937 
938 void
939 tcp_setpersist(tp)
940 	struct tcpcb *tp;
941 {
942 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
943 	int tt;
944 
945 	if (callout_active(tp->tt_rexmt))
946 		panic("tcp_setpersist: retransmit pending");
947 	/*
948 	 * Start/restart persistance timer.
949 	 */
950 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
951 		      TCPTV_PERSMIN, TCPTV_PERSMAX);
952 	callout_reset(tp->tt_persist, tt, tcp_timer_persist, tp);
953 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
954 		tp->t_rxtshift++;
955 }
956