xref: /dflybsd-src/sys/netinet/tcp_output.c (revision bf22d4c1f95f57623b2b3030738e116d3a547284)
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 1.11 2004/04/07 17:01:25 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 #include <sys/thread.h>
53 #include <sys/globaldata.h>
54 
55 #include <net/route.h>
56 
57 #include <netinet/in.h>
58 #include <netinet/in_systm.h>
59 #include <netinet/ip.h>
60 #include <netinet/in_pcb.h>
61 #include <netinet/ip_var.h>
62 #include <netinet6/in6_pcb.h>
63 #include <netinet/ip6.h>
64 #include <netinet6/ip6_var.h>
65 #include <netinet/tcp.h>
66 #define	TCPOUTFLAGS
67 #include <netinet/tcp_fsm.h>
68 #include <netinet/tcp_seq.h>
69 #include <netinet/tcp_timer.h>
70 #include <netinet/tcp_var.h>
71 #include <netinet/tcpip.h>
72 #ifdef TCPDEBUG
73 #include <netinet/tcp_debug.h>
74 #endif
75 
76 #ifdef IPSEC
77 #include <netinet6/ipsec.h>
78 #endif /*IPSEC*/
79 
80 #ifdef FAST_IPSEC
81 #include <netipsec/ipsec.h>
82 #define	IPSEC
83 #endif /*FAST_IPSEC*/
84 
85 #ifdef notyet
86 extern struct mbuf *m_copypack();
87 #endif
88 
89 int path_mtu_discovery = 1;
90 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
91 	&path_mtu_discovery, 1, "Enable Path MTU Discovery");
92 
93 int ss_fltsz = 1;
94 SYSCTL_INT(_net_inet_tcp, OID_AUTO, slowstart_flightsize, CTLFLAG_RW,
95 	&ss_fltsz, 1, "Slow start flight size");
96 
97 int ss_fltsz_local = 4;
98 SYSCTL_INT(_net_inet_tcp, OID_AUTO, local_slowstart_flightsize, CTLFLAG_RW,
99 	&ss_fltsz_local, 1, "Slow start flight size for local networks");
100 
101 int	tcp_do_newreno = 1;
102 SYSCTL_INT(_net_inet_tcp, OID_AUTO, newreno, CTLFLAG_RW, &tcp_do_newreno,
103 	0, "Enable NewReno Algorithms");
104 
105 /*
106  * Tcp output routine: figure out what should be sent and send it.
107  */
108 int
109 tcp_output(tp)
110 	struct tcpcb *tp;
111 {
112 	struct inpcb * const inp = tp->t_inpcb;
113 	struct socket *so = inp->inp_socket;
114 	long len, recvwin, sendwin;
115 	int off, flags, error;
116 	struct mbuf *m;
117 	struct ip *ip = NULL;
118 	struct ipovly *ipov = NULL;
119 	struct tcphdr *th;
120 	u_char opt[TCP_MAXOLEN];
121 	unsigned ipoptlen, optlen, hdrlen;
122 	int idle, sendalot;
123 	struct ip6_hdr *ip6 = NULL;
124 #ifdef INET6
125 	const boolean_t isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
126 #else
127 	const boolean_t isipv6 = FALSE;
128 #endif
129 	struct rmxp_tao *taop;
130 
131 	/*
132 	 * Determine length of data that should be transmitted,
133 	 * and flags that will be used.
134 	 * If there is some data or critical controls (SYN, RST)
135 	 * to send, then transmit; otherwise, investigate further.
136 	 */
137 	if ((tp->snd_max == tp->snd_una) &&
138 	    (ticks - tp->t_rcvtime) >= tp->t_rxtcur) {
139 		/*
140 		 * We have been idle for "a while" and no acks are
141 		 * expected to clock out any data we send --
142 		 * slow start to get ack "clock" running again.
143 		 *
144 		 * Set the slow-start flight size depending on whether
145 		 * this is a local network or not.
146 		 */
147 		if ((isipv6 && in6_localaddr(&inp->in6p_faddr)) ||
148 		    (!isipv6 && in_localaddr(inp->inp_faddr)))
149 			tp->snd_cwnd = tp->t_maxseg * ss_fltsz_local;
150 		else
151 			tp->snd_cwnd = tp->t_maxseg * ss_fltsz;
152 	}
153 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
154 	if (idle && (tp->t_flags & TF_MORETOCOME))
155 		tp->t_flags |= TF_LASTIDLE;
156 	else
157 		tp->t_flags &= ~TF_LASTIDLE;
158 
159 again:
160 	sendalot = 0;
161 	off = tp->snd_nxt - tp->snd_una;
162 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
163 	sendwin = min(sendwin, tp->snd_bwnd);
164 
165 	flags = tcp_outflags[tp->t_state];
166 	/*
167 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
168 	 * state flags.
169 	 */
170 	if (tp->t_flags & TF_NEEDFIN)
171 		flags |= TH_FIN;
172 	if (tp->t_flags & TF_NEEDSYN)
173 		flags |= TH_SYN;
174 
175 	/*
176 	 * If in persist timeout with window of 0, send 1 byte.
177 	 * Otherwise, if window is small but nonzero
178 	 * and timer expired, we will send what we can
179 	 * and go to transmit state.
180 	 */
181 	if (tp->t_force) {
182 		if (sendwin == 0) {
183 			/*
184 			 * If we still have some data to send, then
185 			 * clear the FIN bit.  Usually this would
186 			 * happen below when it realizes that we
187 			 * aren't sending all the data.  However,
188 			 * if we have exactly 1 byte of unsent data,
189 			 * then it won't clear the FIN bit below,
190 			 * and if we are in persist state, we wind
191 			 * up sending the packet without recording
192 			 * that we sent the FIN bit.
193 			 *
194 			 * We can't just blindly clear the FIN bit,
195 			 * because if we don't have any more data
196 			 * to send then the probe will be the FIN
197 			 * itself.
198 			 */
199 			if (off < so->so_snd.sb_cc)
200 				flags &= ~TH_FIN;
201 			sendwin = 1;
202 		} else {
203 			callout_stop(tp->tt_persist);
204 			tp->t_rxtshift = 0;
205 		}
206 	}
207 
208 	/*
209 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
210 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
211 	 * a negative length.  This can also occur when TCP opens up
212 	 * its congestion window while receiving additional duplicate
213 	 * acks after fast-retransmit because TCP will reset snd_nxt
214 	 * to snd_max after the fast-retransmit.
215 	 *
216 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
217 	 * be set to snd_una, the offset will be 0, and the length may
218 	 * wind up 0.
219 	 */
220 	len = (long)ulmin(so->so_snd.sb_cc, sendwin) - off;
221 
222 	/*
223 	 * Lop off SYN bit if it has already been sent.  However, if this
224 	 * is SYN-SENT state and if segment contains data and if we don't
225 	 * know that foreign host supports TAO, suppress sending segment.
226 	 */
227 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
228 		flags &= ~TH_SYN;
229 		off--, len++;
230 		if (len > 0 && tp->t_state == TCPS_SYN_SENT &&
231 		    ((taop = tcp_gettaocache(&inp->inp_inc)) == NULL ||
232 		     taop->tao_ccsent == 0))
233 			return 0;
234 	}
235 
236 	/*
237 	 * Be careful not to send data and/or FIN on SYN segments
238 	 * in cases when no CC option will be sent.
239 	 * This measure is needed to prevent interoperability problems
240 	 * with not fully conformant TCP implementations.
241 	 */
242 	if ((flags & TH_SYN) &&
243 	    ((tp->t_flags & TF_NOOPT) || !(tp->t_flags & TF_REQ_CC) ||
244 	     ((flags & TH_ACK) && !(tp->t_flags & TF_RCVD_CC)))) {
245 		len = 0;
246 		flags &= ~TH_FIN;
247 	}
248 
249 	if (len < 0) {
250 		/*
251 		 * If FIN has been sent but not acked,
252 		 * but we haven't been called to retransmit,
253 		 * len will be < 0.  Otherwise, window shrank
254 		 * after we sent into it.  If window shrank to 0,
255 		 * cancel pending retransmit, pull snd_nxt back
256 		 * to (closed) window, and set the persist timer
257 		 * if it isn't already going.  If the window didn't
258 		 * close completely, just wait for an ACK.
259 		 */
260 		len = 0;
261 		if (sendwin == 0) {
262 			callout_stop(tp->tt_rexmt);
263 			tp->t_rxtshift = 0;
264 			tp->snd_nxt = tp->snd_una;
265 			if (!callout_active(tp->tt_persist))
266 				tcp_setpersist(tp);
267 		}
268 	}
269 
270 	/*
271 	 * len will be >= 0 after this point.  Truncate to the maximum
272 	 * segment length and ensure that FIN is removed if the length
273 	 * no longer contains the last data byte.
274 	 */
275 	if (len > tp->t_maxseg) {
276 		len = tp->t_maxseg;
277 		sendalot = 1;
278 	}
279 	if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
280 		flags &= ~TH_FIN;
281 
282 	recvwin = sbspace(&so->so_rcv);
283 
284 	/*
285 	 * Sender silly window avoidance.   We transmit under the following
286 	 * conditions when len is non-zero:
287 	 *
288 	 *	- We have a full segment
289 	 *	- This is the last buffer in a write()/send() and we are
290 	 *	  either idle or running NODELAY
291 	 *	- we've timed out (e.g. persist timer)
292 	 *	- we have more then 1/2 the maximum send window's worth of
293 	 *	  data (receiver may be limited the window size)
294 	 *	- we need to retransmit
295 	 */
296 	if (len) {
297 		if (len == tp->t_maxseg)
298 			goto send;
299 		/*
300 		 * NOTE! on localhost connections an 'ack' from the remote
301 		 * end may occur synchronously with the output and cause
302 		 * us to flush a buffer queued with moretocome.  XXX
303 		 *
304 		 * note: the len + off check is almost certainly unnecessary.
305 		 */
306 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
307 		    (idle || (tp->t_flags & TF_NODELAY)) &&
308 		    len + off >= so->so_snd.sb_cc &&
309 		    !(tp->t_flags & TF_NOPUSH)) {
310 			goto send;
311 		}
312 		if (tp->t_force)			/* typ. timeout case */
313 			goto send;
314 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
315 			goto send;
316 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
317 			goto send;
318 	}
319 
320 	/*
321 	 * Compare available window to amount of window
322 	 * known to peer (as advertised window less
323 	 * next expected input).  If the difference is at least two
324 	 * max size segments, or at least 50% of the maximum possible
325 	 * window, then want to send a window update to peer.
326 	 */
327 	if (recvwin > 0) {
328 		/*
329 		 * "adv" is the amount we can increase the window,
330 		 * taking into account that we are limited by
331 		 * TCP_MAXWIN << tp->rcv_scale.
332 		 */
333 		long adv = min(recvwin, (long)TCP_MAXWIN << tp->rcv_scale) -
334 			(tp->rcv_adv - tp->rcv_nxt);
335 
336 		if (adv >= (long) (2 * tp->t_maxseg))
337 			goto send;
338 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
339 			goto send;
340 	}
341 
342 	/*
343 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
344 	 * is also a catch-all for the retransmit timer timeout case.
345 	 */
346 	if (tp->t_flags & TF_ACKNOW)
347 		goto send;
348 	if ((flags & TH_RST) ||
349 	    ((flags & TH_SYN) && !(tp->t_flags & TF_NEEDSYN)))
350 		goto send;
351 	if (SEQ_GT(tp->snd_up, tp->snd_una))
352 		goto send;
353 	/*
354 	 * If our state indicates that FIN should be sent
355 	 * and we have not yet done so, then we need to send.
356 	 */
357 	if (flags & TH_FIN &&
358 	    (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una))
359 		goto send;
360 
361 	/*
362 	 * TCP window updates are not reliable, rather a polling protocol
363 	 * using ``persist'' packets is used to insure receipt of window
364 	 * updates.  The three ``states'' for the output side are:
365 	 *	idle			not doing retransmits or persists
366 	 *	persisting		to move a small or zero window
367 	 *	(re)transmitting	and thereby not persisting
368 	 *
369 	 * callout_active(tp->tt_persist)
370 	 *	is true when we are in persist state.
371 	 * tp->t_force
372 	 *	is set when we are called to send a persist packet.
373 	 * callout_active(tp->tt_rexmt)
374 	 *	is set when we are retransmitting
375 	 * The output side is idle when both timers are zero.
376 	 *
377 	 * If send window is too small, there is data to transmit, and no
378 	 * retransmit or persist is pending, then go to persist state.
379 	 * If nothing happens soon, send when timer expires:
380 	 * if window is nonzero, transmit what we can,
381 	 * otherwise force out a byte.
382 	 */
383 	if (so->so_snd.sb_cc && !callout_active(tp->tt_rexmt) &&
384 	    !callout_active(tp->tt_persist)) {
385 		tp->t_rxtshift = 0;
386 		tcp_setpersist(tp);
387 	}
388 
389 	/*
390 	 * No reason to send a segment, just return.
391 	 */
392 	return (0);
393 
394 send:
395 	/*
396 	 * Before ESTABLISHED, force sending of initial options
397 	 * unless TCP set not to do any options.
398 	 * NOTE: we assume that the IP/TCP header plus TCP options
399 	 * always fit in a single mbuf, leaving room for a maximum
400 	 * link header, i.e.
401 	 *	max_linkhdr + sizeof(struct tcpiphdr) + optlen <= MCLBYTES
402 	 */
403 	optlen = 0;
404 	if (isipv6)
405 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
406 	else
407 		hdrlen = sizeof(struct tcpiphdr);
408 	if (flags & TH_SYN) {
409 		tp->snd_nxt = tp->iss;
410 		if (!(tp->t_flags & TF_NOOPT)) {
411 			u_short mss;
412 
413 			opt[0] = TCPOPT_MAXSEG;
414 			opt[1] = TCPOLEN_MAXSEG;
415 			mss = htons((u_short) tcp_mssopt(tp));
416 			(void)memcpy(opt + 2, &mss, sizeof(mss));
417 			optlen = TCPOLEN_MAXSEG;
418 
419 			if ((tp->t_flags & TF_REQ_SCALE) &&
420 			    (!(flags & TH_ACK) ||
421 			     (tp->t_flags & TF_RCVD_SCALE))) {
422 				*((u_int32_t *)(opt + optlen)) = htonl(
423 					TCPOPT_NOP << 24 |
424 					TCPOPT_WINDOW << 16 |
425 					TCPOLEN_WINDOW << 8 |
426 					tp->request_r_scale);
427 				optlen += 4;
428 			}
429 		}
430 	}
431 
432 	/*
433 	 * Send a timestamp and echo-reply if this is a SYN and our side
434 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
435 	 * and our peer have sent timestamps in our SYN's.
436 	 */
437 	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
438 	    !(flags & TH_RST) &&
439 	    (!(flags & TH_ACK) || (tp->t_flags & TF_RCVD_TSTMP))) {
440 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
441 
442 		/* Form timestamp option as shown in appendix A of RFC 1323. */
443 		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
444 		*lp++ = htonl(ticks);
445 		*lp   = htonl(tp->ts_recent);
446 		optlen += TCPOLEN_TSTAMP_APPA;
447 	}
448 
449 	/*
450 	 * Send `CC-family' options if our side wants to use them (TF_REQ_CC),
451 	 * options are allowed (!TF_NOOPT) and it's not a RST.
452 	 */
453 	if ((tp->t_flags & (TF_REQ_CC|TF_NOOPT)) == TF_REQ_CC &&
454 	     !(flags & TH_RST)) {
455 		switch (flags & (TH_SYN|TH_ACK)) {
456 		/*
457 		 * This is a normal ACK, send CC if we received CC before
458 		 * from our peer.
459 		 */
460 		case TH_ACK:
461 			if (!(tp->t_flags & TF_RCVD_CC))
462 				break;
463 			/*FALLTHROUGH*/
464 
465 		/*
466 		 * We can only get here in T/TCP's SYN_SENT* state, when
467 		 * we're a sending a non-SYN segment without waiting for
468 		 * the ACK of our SYN.  A check above assures that we only
469 		 * do this if our peer understands T/TCP.
470 		 */
471 		case 0:
472 			opt[optlen++] = TCPOPT_NOP;
473 			opt[optlen++] = TCPOPT_NOP;
474 			opt[optlen++] = TCPOPT_CC;
475 			opt[optlen++] = TCPOLEN_CC;
476 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
477 
478 			optlen += 4;
479 			break;
480 
481 		/*
482 		 * This is our initial SYN, check whether we have to use
483 		 * CC or CC.new.
484 		 */
485 		case TH_SYN:
486 			opt[optlen++] = TCPOPT_NOP;
487 			opt[optlen++] = TCPOPT_NOP;
488 			opt[optlen++] = tp->t_flags & TF_SENDCCNEW ?
489 						TCPOPT_CCNEW : TCPOPT_CC;
490 			opt[optlen++] = TCPOLEN_CC;
491 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
492 			optlen += 4;
493 			break;
494 
495 		/*
496 		 * This is a SYN,ACK; send CC and CC.echo if we received
497 		 * CC from our peer.
498 		 */
499 		case (TH_SYN|TH_ACK):
500 			if (tp->t_flags & TF_RCVD_CC) {
501 				opt[optlen++] = TCPOPT_NOP;
502 				opt[optlen++] = TCPOPT_NOP;
503 				opt[optlen++] = TCPOPT_CC;
504 				opt[optlen++] = TCPOLEN_CC;
505 				*(u_int32_t *)&opt[optlen] =
506 					htonl(tp->cc_send);
507 				optlen += 4;
508 				opt[optlen++] = TCPOPT_NOP;
509 				opt[optlen++] = TCPOPT_NOP;
510 				opt[optlen++] = TCPOPT_CCECHO;
511 				opt[optlen++] = TCPOLEN_CC;
512 				*(u_int32_t *)&opt[optlen] =
513 					htonl(tp->cc_recv);
514 				optlen += 4;
515 			}
516 			break;
517 		}
518 	}
519 
520 	hdrlen += optlen;
521 
522 	if (isipv6)
523 		ipoptlen = ip6_optlen(inp);
524 	else {
525 		if (inp->inp_options) {
526 			ipoptlen = inp->inp_options->m_len -
527 			    offsetof(struct ipoption, ipopt_list);
528 		} else {
529 			ipoptlen = 0;
530 		}
531 	}
532 #ifdef IPSEC
533 	ipoptlen += ipsec_hdrsiz_tcp(tp);
534 #endif
535 
536 	/*
537 	 * Adjust data length if insertion of options will
538 	 * bump the packet length beyond the t_maxopd length.
539 	 * Clear the FIN bit because we cut off the tail of
540 	 * the segment.
541 	 */
542 	if (len + optlen + ipoptlen > tp->t_maxopd) {
543 		/*
544 		 * If there is still more to send, don't close the connection.
545 		 */
546 		flags &= ~TH_FIN;
547 		len = tp->t_maxopd - optlen - ipoptlen;
548 		sendalot = 1;
549 	}
550 
551 #ifdef INET6
552 	KASSERT(max_linkhdr + hdrlen <= MCLBYTES, ("tcphdr too big"));
553 #else
554 	KASSERT(max_linkhdr + hdrlen <= MHLEN, ("tcphdr too big"));
555 #endif
556 
557 	/*
558 	 * Grab a header mbuf, attaching a copy of data to
559 	 * be transmitted, and initialize the header from
560 	 * the template for sends on this connection.
561 	 */
562 	if (len) {
563 		if (tp->t_force && len == 1)
564 			tcpstat.tcps_sndprobe++;
565 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
566 			tcpstat.tcps_sndrexmitpack++;
567 			tcpstat.tcps_sndrexmitbyte += len;
568 		} else {
569 			tcpstat.tcps_sndpack++;
570 			tcpstat.tcps_sndbyte += len;
571 		}
572 #ifdef notyet
573 		if ((m = m_copypack(so->so_snd.sb_mb, off,
574 		    (int)len, max_linkhdr + hdrlen)) == 0) {
575 			error = ENOBUFS;
576 			goto out;
577 		}
578 		/*
579 		 * m_copypack left space for our hdr; use it.
580 		 */
581 		m->m_len += hdrlen;
582 		m->m_data -= hdrlen;
583 #else
584 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
585 		if (m == NULL) {
586 			error = ENOBUFS;
587 			goto out;
588 		}
589 #ifdef INET6
590 		if (MHLEN < hdrlen + max_linkhdr) {
591 			MCLGET(m, M_DONTWAIT);
592 			if (!(m->m_flags & M_EXT)) {
593 				m_freem(m);
594 				error = ENOBUFS;
595 				goto out;
596 			}
597 		}
598 #endif
599 		m->m_data += max_linkhdr;
600 		m->m_len = hdrlen;
601 		if (len <= MHLEN - hdrlen - max_linkhdr) {
602 			m_copydata(so->so_snd.sb_mb, off, (int) len,
603 			    mtod(m, caddr_t) + hdrlen);
604 			m->m_len += len;
605 		} else {
606 			m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len);
607 			if (m->m_next == 0) {
608 				(void) m_free(m);
609 				error = ENOBUFS;
610 				goto out;
611 			}
612 		}
613 #endif
614 		/*
615 		 * If we're sending everything we've got, set PUSH.
616 		 * (This will keep happy those implementations which only
617 		 * give data to the user when a buffer fills or
618 		 * a PUSH comes in.)
619 		 */
620 		if (off + len == so->so_snd.sb_cc)
621 			flags |= TH_PUSH;
622 	} else {
623 		if (tp->t_flags & TF_ACKNOW)
624 			tcpstat.tcps_sndacks++;
625 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
626 			tcpstat.tcps_sndctrl++;
627 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
628 			tcpstat.tcps_sndurg++;
629 		else
630 			tcpstat.tcps_sndwinup++;
631 
632 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
633 		if (m == NULL) {
634 			error = ENOBUFS;
635 			goto out;
636 		}
637 		if (isipv6 &&
638 		    (hdrlen + max_linkhdr > MHLEN) && hdrlen <= MHLEN)
639 			MH_ALIGN(m, hdrlen);
640 		else
641 			m->m_data += max_linkhdr;
642 		m->m_len = hdrlen;
643 	}
644 	m->m_pkthdr.rcvif = (struct ifnet *)0;
645 	if (isipv6) {
646 		ip6 = mtod(m, struct ip6_hdr *);
647 		th = (struct tcphdr *)(ip6 + 1);
648 		tcp_fillheaders(tp, ip6, th);
649 	} else {
650 		ip = mtod(m, struct ip *);
651 		ipov = (struct ipovly *)ip;
652 		th = (struct tcphdr *)(ip + 1);
653 		/* this picks up the pseudo header (w/o the length) */
654 		tcp_fillheaders(tp, ip, th);
655 	}
656 
657 	/*
658 	 * Fill in fields, remembering maximum advertised
659 	 * window for use in delaying messages about window sizes.
660 	 * If resending a FIN, be sure not to use a new sequence number.
661 	 */
662 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
663 	    tp->snd_nxt == tp->snd_max)
664 		tp->snd_nxt--;
665 	/*
666 	 * If we are doing retransmissions, then snd_nxt will
667 	 * not reflect the first unsent octet.  For ACK only
668 	 * packets, we do not want the sequence number of the
669 	 * retransmitted packet, we want the sequence number
670 	 * of the next unsent octet.  So, if there is no data
671 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
672 	 * when filling in ti_seq.  But if we are in persist
673 	 * state, snd_max might reflect one byte beyond the
674 	 * right edge of the window, so use snd_nxt in that
675 	 * case, since we know we aren't doing a retransmission.
676 	 * (retransmit and persist are mutually exclusive...)
677 	 */
678 	if (len || (flags & (TH_SYN|TH_FIN)) || callout_active(tp->tt_persist))
679 		th->th_seq = htonl(tp->snd_nxt);
680 	else
681 		th->th_seq = htonl(tp->snd_max);
682 	th->th_ack = htonl(tp->rcv_nxt);
683 	if (optlen) {
684 		bcopy(opt, th + 1, optlen);
685 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
686 	}
687 	th->th_flags = flags;
688 	/*
689 	 * Calculate receive window.  Don't shrink window,
690 	 * but avoid silly window syndrome.
691 	 */
692 	if (recvwin < (long)(so->so_rcv.sb_hiwat / 4) &&
693 	    recvwin < (long)tp->t_maxseg)
694 		recvwin = 0;
695 	if (recvwin < (long)(tp->rcv_adv - tp->rcv_nxt))
696 		recvwin = (long)(tp->rcv_adv - tp->rcv_nxt);
697 	if (recvwin > (long)TCP_MAXWIN << tp->rcv_scale)
698 		recvwin = (long)TCP_MAXWIN << tp->rcv_scale;
699 	th->th_win = htons((u_short) (recvwin>>tp->rcv_scale));
700 
701 	/*
702 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
703 	 * a 0 window.  This may cause the remote transmitter to stall.  This
704 	 * flag tells soreceive() to disable delayed acknowledgements when
705 	 * draining the buffer.  This can occur if the receiver is attempting
706 	 * to read more data then can be buffered prior to transmitting on
707 	 * the connection.
708 	 */
709 	if (recvwin == 0)
710 		tp->t_flags |= TF_RXWIN0SENT;
711 	else
712 		tp->t_flags &= ~TF_RXWIN0SENT;
713 
714 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
715 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
716 		th->th_flags |= TH_URG;
717 	} else {
718 		/*
719 		 * If no urgent pointer to send, then we pull
720 		 * the urgent pointer to the left edge of the send window
721 		 * so that it doesn't drift into the send window on sequence
722 		 * number wraparound.
723 		 */
724 		tp->snd_up = tp->snd_una;		/* drag it along */
725 	}
726 
727 	/*
728 	 * Put TCP length in extended header, and then
729 	 * checksum extended header and data.
730 	 */
731 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
732 	if (isipv6) {
733 		/*
734 		 * ip6_plen is not need to be filled now, and will be filled
735 		 * in ip6_output.
736 		 */
737 		th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
738 				       sizeof(struct tcphdr) + optlen + len);
739 	} else {
740 		m->m_pkthdr.csum_flags = CSUM_TCP;
741 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
742 		if (len + optlen)
743 			th->th_sum = in_addword(th->th_sum,
744 						htons((u_short)(optlen + len)));
745 
746 		/* IP version must be set here for ipv4/ipv6 checking later */
747 		KASSERT(ip->ip_v == IPVERSION,
748 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
749 	}
750 
751 	/*
752 	 * In transmit state, time the transmission and arrange for
753 	 * the retransmit.  In persist state, just set snd_max.
754 	 */
755 	if (tp->t_force == 0 || !callout_active(tp->tt_persist)) {
756 		tcp_seq startseq = tp->snd_nxt;
757 
758 		/*
759 		 * Advance snd_nxt over sequence space of this segment.
760 		 */
761 		if (flags & (TH_SYN|TH_FIN)) {
762 			if (flags & TH_SYN)
763 				tp->snd_nxt++;
764 			if (flags & TH_FIN) {
765 				tp->snd_nxt++;
766 				tp->t_flags |= TF_SENTFIN;
767 			}
768 		}
769 		tp->snd_nxt += len;
770 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
771 			tp->snd_max = tp->snd_nxt;
772 			/*
773 			 * Time this transmission if not a retransmission and
774 			 * not currently timing anything.
775 			 */
776 			if (tp->t_rtttime == 0) {
777 				tp->t_rtttime = ticks;
778 				tp->t_rtseq = startseq;
779 				tcpstat.tcps_segstimed++;
780 			}
781 		}
782 
783 		/*
784 		 * Set retransmit timer if not currently set,
785 		 * and not doing a pure ack or a keep-alive probe.
786 		 * Initial value for retransmit timer is smoothed
787 		 * round-trip time + 2 * round-trip time variance.
788 		 * Initialize shift counter which is used for backoff
789 		 * of retransmit time.
790 		 */
791 		if (!callout_active(tp->tt_rexmt) &&
792 		    tp->snd_nxt != tp->snd_una) {
793 			if (callout_active(tp->tt_persist)) {
794 				callout_stop(tp->tt_persist);
795 				tp->t_rxtshift = 0;
796 			}
797 			callout_reset(tp->tt_rexmt, tp->t_rxtcur,
798 				      tcp_timer_rexmt, tp);
799 		}
800 	} else {
801 		/*
802 		 * Persist case, update snd_max but since we are in
803 		 * persist mode (no window) we do not update snd_nxt.
804 		 */
805 		int xlen = len;
806 		if (flags & TH_SYN)
807 			++xlen;
808 		if (flags & TH_FIN) {
809 			++xlen;
810 			tp->t_flags |= TF_SENTFIN;
811 		}
812 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
813 			tp->snd_max = tp->snd_nxt + xlen;
814 	}
815 
816 #ifdef TCPDEBUG
817 	/*
818 	 * Trace.
819 	 */
820 	if (so->so_options & SO_DEBUG)
821 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
822 #endif
823 
824 	/*
825 	 * Fill in IP length and desired time to live and
826 	 * send to IP level.  There should be a better way
827 	 * to handle ttl and tos; we could keep them in
828 	 * the template, but need a way to checksum without them.
829 	 */
830 	/*
831 	 * m->m_pkthdr.len should have been set before cksum calcuration,
832 	 * because in6_cksum() need it.
833 	 */
834 	if (isipv6) {
835 		/*
836 		 * we separately set hoplimit for every segment, since the
837 		 * user might want to change the value via setsockopt.
838 		 * Also, desired default hop limit might be changed via
839 		 * Neighbor Discovery.
840 		 */
841 		ip6->ip6_hlim = in6_selecthlim(inp,
842 		    (inp->in6p_route.ro_rt ?
843 		     inp->in6p_route.ro_rt->rt_ifp : NULL));
844 
845 		/* TODO: IPv6 IP6TOS_ECT bit on */
846 		error = ip6_output(m, inp->in6p_outputopts, &inp->in6p_route,
847 		    (so->so_options & SO_DONTROUTE), NULL, NULL, inp);
848 	} else {
849 		struct rtentry *rt;
850 		ip->ip_len = m->m_pkthdr.len;
851 #ifdef INET6
852 		if (INP_CHECK_SOCKAF(so, AF_INET6))
853 			ip->ip_ttl = in6_selecthlim(inp,
854 			    (inp->in6p_route.ro_rt ?
855 			     inp->in6p_route.ro_rt->rt_ifp : NULL));
856 		else
857 #endif
858 			ip->ip_ttl = inp->inp_ip_ttl;	/* XXX */
859 
860 		ip->ip_tos = inp->inp_ip_tos;	/* XXX */
861 		/*
862 		 * See if we should do MTU discovery.
863 		 * We do it only if the following are true:
864 		 *	1) we have a valid route to the destination
865 		 *	2) the MTU is not locked (if it is,
866 		 *	   then discovery has been disabled)
867 		 */
868 		if (path_mtu_discovery &&
869 		    (rt = inp->inp_route.ro_rt) && (rt->rt_flags & RTF_UP) &&
870 		    !(rt->rt_rmx.rmx_locks & RTV_MTU))
871 			ip->ip_off |= IP_DF;
872 
873 		error = ip_output(m, inp->inp_options, &inp->inp_route,
874 		    (so->so_options & SO_DONTROUTE), NULL, inp);
875 	}
876 	if (error) {
877 
878 		/*
879 		 * We know that the packet was lost, so back out the
880 		 * sequence number advance, if any.
881 		 */
882 		if (tp->t_force == 0 || !callout_active(tp->tt_persist)) {
883 			/*
884 			 * No need to check for TH_FIN here because
885 			 * the TF_SENTFIN flag handles that case.
886 			 */
887 			if (!(flags & TH_SYN))
888 				tp->snd_nxt -= len;
889 		}
890 
891 out:
892 		if (error == ENOBUFS) {
893 			/*
894 			 * If we can't send, make sure there is something
895 			 * to get us going again later.  Persist state
896 			 * is not necessarily right, but it is close enough.
897 			 */
898 			if (!callout_active(tp->tt_rexmt) &&
899 			    !callout_active(tp->tt_persist)) {
900 				tp->t_rxtshift = 0;
901 				tcp_setpersist(tp);
902 			}
903 			tcp_quench(inp, 0);
904 			return (0);
905 		}
906 		if (error == EMSGSIZE) {
907 			/*
908 			 * ip_output() will have already fixed the route
909 			 * for us.  tcp_mtudisc() will, as its last action,
910 			 * initiate retransmission, so it is important to
911 			 * not do so here.
912 			 */
913 			tcp_mtudisc(inp, 0);
914 			return 0;
915 		}
916 		if ((error == EHOSTUNREACH || error == ENETDOWN) &&
917 		    TCPS_HAVERCVDSYN(tp->t_state)) {
918 			tp->t_softerror = error;
919 			return (0);
920 		}
921 		return (error);
922 	}
923 	tcpstat.tcps_sndtotal++;
924 
925 	/*
926 	 * Data sent (as far as we can tell).
927 	 * If this advertises a larger window than any other segment,
928 	 * then remember the size of the advertised window.
929 	 * Any pending ACK has now been sent.
930 	 */
931 	if (recvwin > 0 && SEQ_GT(tp->rcv_nxt + recvwin, tp->rcv_adv))
932 		tp->rcv_adv = tp->rcv_nxt + recvwin;
933 	tp->last_ack_sent = tp->rcv_nxt;
934 	tp->t_flags &= ~TF_ACKNOW;
935 	if (tcp_delack_enabled)
936 		callout_stop(tp->tt_delack);
937 	if (sendalot)
938 		goto again;
939 	return (0);
940 }
941 
942 void
943 tcp_setpersist(tp)
944 	struct tcpcb *tp;
945 {
946 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
947 	int tt;
948 
949 	if (callout_active(tp->tt_rexmt))
950 		panic("tcp_setpersist: retransmit pending");
951 	/*
952 	 * Start/restart persistance timer.
953 	 */
954 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
955 		      TCPTV_PERSMIN, TCPTV_PERSMAX);
956 	callout_reset(tp->tt_persist, tt, tcp_timer_persist, tp);
957 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
958 		tp->t_rxtshift++;
959 }
960