xref: /freebsd/sys/netinet/tcp_output.c (revision 2357939bc239bd5334a169b62313806178dd8f30)
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  * 4. Neither the name of the University nor the names of its contributors
14  *    may be used to endorse or promote products derived from this software
15  *    without specific prior written permission.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  *
29  *	@(#)tcp_output.c	8.4 (Berkeley) 5/24/95
30  * $FreeBSD$
31  */
32 
33 #include "opt_inet.h"
34 #include "opt_inet6.h"
35 #include "opt_ipsec.h"
36 #include "opt_mac.h"
37 #include "opt_tcpdebug.h"
38 
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/domain.h>
42 #include <sys/kernel.h>
43 #include <sys/lock.h>
44 #include <sys/mac.h>
45 #include <sys/mbuf.h>
46 #include <sys/mutex.h>
47 #include <sys/protosw.h>
48 #include <sys/socket.h>
49 #include <sys/socketvar.h>
50 #include <sys/sysctl.h>
51 
52 #include <net/route.h>
53 
54 #include <netinet/in.h>
55 #include <netinet/in_systm.h>
56 #include <netinet/ip.h>
57 #include <netinet/in_pcb.h>
58 #include <netinet/ip_var.h>
59 #ifdef INET6
60 #include <netinet6/in6_pcb.h>
61 #include <netinet/ip6.h>
62 #include <netinet6/ip6_var.h>
63 #endif
64 #include <netinet/tcp.h>
65 #define	TCPOUTFLAGS
66 #include <netinet/tcp_fsm.h>
67 #include <netinet/tcp_seq.h>
68 #include <netinet/tcp_timer.h>
69 #include <netinet/tcp_var.h>
70 #include <netinet/tcpip.h>
71 #ifdef TCPDEBUG
72 #include <netinet/tcp_debug.h>
73 #endif
74 
75 #ifdef IPSEC
76 #include <netinet6/ipsec.h>
77 #endif /*IPSEC*/
78 
79 #ifdef FAST_IPSEC
80 #include <netipsec/ipsec.h>
81 #define	IPSEC
82 #endif /*FAST_IPSEC*/
83 
84 #include <machine/in_cksum.h>
85 
86 #ifdef notyet
87 extern struct mbuf *m_copypack();
88 #endif
89 
90 int path_mtu_discovery = 1;
91 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
92 	&path_mtu_discovery, 1, "Enable Path MTU Discovery");
93 
94 int ss_fltsz = 1;
95 SYSCTL_INT(_net_inet_tcp, OID_AUTO, slowstart_flightsize, CTLFLAG_RW,
96 	&ss_fltsz, 1, "Slow start flight size");
97 
98 int ss_fltsz_local = 4;
99 SYSCTL_INT(_net_inet_tcp, OID_AUTO, local_slowstart_flightsize, CTLFLAG_RW,
100 	&ss_fltsz_local, 1, "Slow start flight size for local networks");
101 
102 int     tcp_do_newreno = 1;
103 SYSCTL_INT(_net_inet_tcp, OID_AUTO, newreno, CTLFLAG_RW, &tcp_do_newreno,
104         0, "Enable NewReno Algorithms");
105 
106 /*
107  * Tcp output routine: figure out what should be sent and send it.
108  */
109 int
110 tcp_output(struct tcpcb *tp)
111 {
112 	struct socket *so = tp->t_inpcb->inp_socket;
113 	long len, recwin, sendwin;
114 	int off, flags, error;
115 #ifdef TCP_SIGNATURE
116 	int sigoff = 0;
117 #endif
118 	struct mbuf *m;
119 	struct ip *ip = NULL;
120 	struct ipovly *ipov = NULL;
121 	struct tcphdr *th;
122 	u_char opt[TCP_MAXOLEN];
123 	unsigned ipoptlen, optlen, hdrlen;
124 	int idle, sendalot;
125 #if 0
126 	int maxburst = TCP_MAXBURST;
127 #endif
128 	struct rmxp_tao tao;
129 #ifdef INET6
130 	struct ip6_hdr *ip6 = NULL;
131 	int isipv6;
132 
133 	bzero(&tao, sizeof(tao));
134 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
135 #endif
136 
137 	INP_LOCK_ASSERT(tp->t_inpcb);
138 
139 	/*
140 	 * Determine length of data that should be transmitted,
141 	 * and flags that will be used.
142 	 * If there is some data or critical controls (SYN, RST)
143 	 * to send, then transmit; otherwise, investigate further.
144 	 */
145 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
146 	if (idle && (ticks - tp->t_rcvtime) >= tp->t_rxtcur) {
147 		/*
148 		 * We have been idle for "a while" and no acks are
149 		 * expected to clock out any data we send --
150 		 * slow start to get ack "clock" running again.
151 		 *
152 		 * Set the slow-start flight size depending on whether
153 		 * this is a local network or not.
154 		 */
155 		int ss = ss_fltsz;
156 #ifdef INET6
157 		if (isipv6) {
158 			if (in6_localaddr(&tp->t_inpcb->in6p_faddr))
159 				ss = ss_fltsz_local;
160 		} else
161 #endif /* INET6 */
162 		if (in_localaddr(tp->t_inpcb->inp_faddr))
163 			ss = ss_fltsz_local;
164 		tp->snd_cwnd = tp->t_maxseg * ss;
165 	}
166 	tp->t_flags &= ~TF_LASTIDLE;
167 	if (idle) {
168 		if (tp->t_flags & TF_MORETOCOME) {
169 			tp->t_flags |= TF_LASTIDLE;
170 			idle = 0;
171 		}
172 	}
173 again:
174 	sendalot = 0;
175 	off = tp->snd_nxt - tp->snd_una;
176 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
177 	sendwin = min(sendwin, tp->snd_bwnd);
178 
179 	flags = tcp_outflags[tp->t_state];
180 	/*
181 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
182 	 * state flags.
183 	 */
184 	if (tp->t_flags & TF_NEEDFIN)
185 		flags |= TH_FIN;
186 	if (tp->t_flags & TF_NEEDSYN)
187 		flags |= TH_SYN;
188 
189 	/*
190 	 * If in persist timeout with window of 0, send 1 byte.
191 	 * Otherwise, if window is small but nonzero
192 	 * and timer expired, we will send what we can
193 	 * and go to transmit state.
194 	 */
195 	if (tp->t_force) {
196 		if (sendwin == 0) {
197 			/*
198 			 * If we still have some data to send, then
199 			 * clear the FIN bit.  Usually this would
200 			 * happen below when it realizes that we
201 			 * aren't sending all the data.  However,
202 			 * if we have exactly 1 byte of unsent data,
203 			 * then it won't clear the FIN bit below,
204 			 * and if we are in persist state, we wind
205 			 * up sending the packet without recording
206 			 * that we sent the FIN bit.
207 			 *
208 			 * We can't just blindly clear the FIN bit,
209 			 * because if we don't have any more data
210 			 * to send then the probe will be the FIN
211 			 * itself.
212 			 */
213 			if (off < so->so_snd.sb_cc)
214 				flags &= ~TH_FIN;
215 			sendwin = 1;
216 		} else {
217 			callout_stop(tp->tt_persist);
218 			tp->t_rxtshift = 0;
219 		}
220 	}
221 
222 	/*
223 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
224 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
225 	 * a negative length.  This can also occur when TCP opens up
226 	 * its congestion window while receiving additional duplicate
227 	 * acks after fast-retransmit because TCP will reset snd_nxt
228 	 * to snd_max after the fast-retransmit.
229 	 *
230 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
231 	 * be set to snd_una, the offset will be 0, and the length may
232 	 * wind up 0.
233 	 */
234 	len = (long)ulmin(so->so_snd.sb_cc, sendwin) - off;
235 
236 
237 	/*
238 	 * Lop off SYN bit if it has already been sent.  However, if this
239 	 * is SYN-SENT state and if segment contains data and if we don't
240 	 * know that foreign host supports TAO, suppress sending segment.
241 	 */
242 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
243 		flags &= ~TH_SYN;
244 		off--, len++;
245 		if (tcp_do_rfc1644)
246 			tcp_hc_gettao(&tp->t_inpcb->inp_inc, &tao);
247 		if (len > 0 && tp->t_state == TCPS_SYN_SENT &&
248 		     tao.tao_ccsent == 0)
249 			return 0;
250 	}
251 
252 	/*
253 	 * Be careful not to send data and/or FIN on SYN segments
254 	 * in cases when no CC option will be sent.
255 	 * This measure is needed to prevent interoperability problems
256 	 * with not fully conformant TCP implementations.
257 	 */
258 	if ((flags & TH_SYN) &&
259 	    ((tp->t_flags & TF_NOOPT) || !(tp->t_flags & TF_REQ_CC) ||
260 	     ((flags & TH_ACK) && !(tp->t_flags & TF_RCVD_CC)))) {
261 		len = 0;
262 		flags &= ~TH_FIN;
263 	}
264 
265 	if (len < 0) {
266 		/*
267 		 * If FIN has been sent but not acked,
268 		 * but we haven't been called to retransmit,
269 		 * len will be < 0.  Otherwise, window shrank
270 		 * after we sent into it.  If window shrank to 0,
271 		 * cancel pending retransmit, pull snd_nxt back
272 		 * to (closed) window, and set the persist timer
273 		 * if it isn't already going.  If the window didn't
274 		 * close completely, just wait for an ACK.
275 		 */
276 		len = 0;
277 		if (sendwin == 0) {
278 			callout_stop(tp->tt_rexmt);
279 			tp->t_rxtshift = 0;
280 			tp->snd_nxt = tp->snd_una;
281 			if (!callout_active(tp->tt_persist))
282 				tcp_setpersist(tp);
283 		}
284 	}
285 
286 	/*
287 	 * len will be >= 0 after this point.  Truncate to the maximum
288 	 * segment length and ensure that FIN is removed if the length
289 	 * no longer contains the last data byte.
290 	 */
291 	if (len > tp->t_maxseg) {
292 		len = tp->t_maxseg;
293 		sendalot = 1;
294 	}
295 	if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
296 		flags &= ~TH_FIN;
297 
298 	recwin = sbspace(&so->so_rcv);
299 
300 	/*
301 	 * Sender silly window avoidance.   We transmit under the following
302 	 * conditions when len is non-zero:
303 	 *
304 	 *	- We have a full segment
305 	 *	- This is the last buffer in a write()/send() and we are
306 	 *	  either idle or running NODELAY
307 	 *	- we've timed out (e.g. persist timer)
308 	 *	- we have more then 1/2 the maximum send window's worth of
309 	 *	  data (receiver may be limited the window size)
310 	 *	- we need to retransmit
311 	 */
312 	if (len) {
313 		if (len == tp->t_maxseg)
314 			goto send;
315 		/*
316 		 * NOTE! on localhost connections an 'ack' from the remote
317 		 * end may occur synchronously with the output and cause
318 		 * us to flush a buffer queued with moretocome.  XXX
319 		 *
320 		 * note: the len + off check is almost certainly unnecessary.
321 		 */
322 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
323 		    (idle || (tp->t_flags & TF_NODELAY)) &&
324 		    len + off >= so->so_snd.sb_cc &&
325 		    (tp->t_flags & TF_NOPUSH) == 0) {
326 			goto send;
327 		}
328 		if (tp->t_force)			/* typ. timeout case */
329 			goto send;
330 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
331 			goto send;
332 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
333 			goto send;
334 	}
335 
336 	/*
337 	 * Compare available window to amount of window
338 	 * known to peer (as advertised window less
339 	 * next expected input).  If the difference is at least two
340 	 * max size segments, or at least 50% of the maximum possible
341 	 * window, then want to send a window update to peer.
342 	 * Skip this if the connection is in T/TCP half-open state.
343 	 */
344 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN)) {
345 		/*
346 		 * "adv" is the amount we can increase the window,
347 		 * taking into account that we are limited by
348 		 * TCP_MAXWIN << tp->rcv_scale.
349 		 */
350 		long adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale) -
351 			(tp->rcv_adv - tp->rcv_nxt);
352 
353 		if (adv >= (long) (2 * tp->t_maxseg))
354 			goto send;
355 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
356 			goto send;
357 	}
358 
359 	/*
360 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
361 	 * is also a catch-all for the retransmit timer timeout case.
362 	 */
363 	if (tp->t_flags & TF_ACKNOW)
364 		goto send;
365 	if ((flags & TH_RST) ||
366 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
367 		goto send;
368 	if (SEQ_GT(tp->snd_up, tp->snd_una))
369 		goto send;
370 	/*
371 	 * If our state indicates that FIN should be sent
372 	 * and we have not yet done so, then we need to send.
373 	 */
374 	if (flags & TH_FIN &&
375 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
376 		goto send;
377 
378 	/*
379 	 * TCP window updates are not reliable, rather a polling protocol
380 	 * using ``persist'' packets is used to insure receipt of window
381 	 * updates.  The three ``states'' for the output side are:
382 	 *	idle			not doing retransmits or persists
383 	 *	persisting		to move a small or zero window
384 	 *	(re)transmitting	and thereby not persisting
385 	 *
386 	 * callout_active(tp->tt_persist)
387 	 *	is true when we are in persist state.
388 	 * tp->t_force
389 	 *	is set when we are called to send a persist packet.
390 	 * callout_active(tp->tt_rexmt)
391 	 *	is set when we are retransmitting
392 	 * The output side is idle when both timers are zero.
393 	 *
394 	 * If send window is too small, there is data to transmit, and no
395 	 * retransmit or persist is pending, then go to persist state.
396 	 * If nothing happens soon, send when timer expires:
397 	 * if window is nonzero, transmit what we can,
398 	 * otherwise force out a byte.
399 	 */
400 	if (so->so_snd.sb_cc && !callout_active(tp->tt_rexmt) &&
401 	    !callout_active(tp->tt_persist)) {
402 		tp->t_rxtshift = 0;
403 		tcp_setpersist(tp);
404 	}
405 
406 	/*
407 	 * No reason to send a segment, just return.
408 	 */
409 	return (0);
410 
411 send:
412 	/*
413 	 * Before ESTABLISHED, force sending of initial options
414 	 * unless TCP set not to do any options.
415 	 * NOTE: we assume that the IP/TCP header plus TCP options
416 	 * always fit in a single mbuf, leaving room for a maximum
417 	 * link header, i.e.
418 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
419 	 */
420 	optlen = 0;
421 #ifdef INET6
422 	if (isipv6)
423 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
424 	else
425 #endif
426 	hdrlen = sizeof (struct tcpiphdr);
427 	if (flags & TH_SYN) {
428 		tp->snd_nxt = tp->iss;
429 		if ((tp->t_flags & TF_NOOPT) == 0) {
430 			u_short mss;
431 
432 			opt[0] = TCPOPT_MAXSEG;
433 			opt[1] = TCPOLEN_MAXSEG;
434 			mss = htons((u_short) tcp_mssopt(&tp->t_inpcb->inp_inc));
435 			(void)memcpy(opt + 2, &mss, sizeof(mss));
436 			optlen = TCPOLEN_MAXSEG;
437 
438 			if ((tp->t_flags & TF_REQ_SCALE) &&
439 			    ((flags & TH_ACK) == 0 ||
440 			    (tp->t_flags & TF_RCVD_SCALE))) {
441 				*((u_int32_t *)(opt + optlen)) = htonl(
442 					TCPOPT_NOP << 24 |
443 					TCPOPT_WINDOW << 16 |
444 					TCPOLEN_WINDOW << 8 |
445 					tp->request_r_scale);
446 				optlen += 4;
447 			}
448 		}
449  	}
450 
451  	/*
452 	 * Send a timestamp and echo-reply if this is a SYN and our side
453 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
454 	 * and our peer have sent timestamps in our SYN's.
455  	 */
456  	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
457  	    (flags & TH_RST) == 0 &&
458 	    ((flags & TH_ACK) == 0 ||
459 	     (tp->t_flags & TF_RCVD_TSTMP))) {
460 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
461 
462  		/* Form timestamp option as shown in appendix A of RFC 1323. */
463  		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
464  		*lp++ = htonl(ticks);
465  		*lp   = htonl(tp->ts_recent);
466  		optlen += TCPOLEN_TSTAMP_APPA;
467  	}
468 
469  	/*
470 	 * Send `CC-family' options if our side wants to use them (TF_REQ_CC),
471 	 * options are allowed (!TF_NOOPT) and it's not a RST.
472  	 */
473  	if ((tp->t_flags & (TF_REQ_CC|TF_NOOPT)) == TF_REQ_CC &&
474  	     (flags & TH_RST) == 0) {
475 		switch (flags & (TH_SYN|TH_ACK)) {
476 		/*
477 		 * This is a normal ACK, send CC if we received CC before
478 		 * from our peer.
479 		 */
480 		case TH_ACK:
481 			if (!(tp->t_flags & TF_RCVD_CC))
482 				break;
483 			/*FALLTHROUGH*/
484 
485 		/*
486 		 * We can only get here in T/TCP's SYN_SENT* state, when
487 		 * we're a sending a non-SYN segment without waiting for
488 		 * the ACK of our SYN.  A check above assures that we only
489 		 * do this if our peer understands T/TCP.
490 		 */
491 		case 0:
492 			opt[optlen++] = TCPOPT_NOP;
493 			opt[optlen++] = TCPOPT_NOP;
494 			opt[optlen++] = TCPOPT_CC;
495 			opt[optlen++] = TCPOLEN_CC;
496 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
497 
498 			optlen += 4;
499 			break;
500 
501 		/*
502 		 * This is our initial SYN, check whether we have to use
503 		 * CC or CC.new.
504 		 */
505 		case TH_SYN:
506 			opt[optlen++] = TCPOPT_NOP;
507 			opt[optlen++] = TCPOPT_NOP;
508 			opt[optlen++] = tp->t_flags & TF_SENDCCNEW ?
509 						TCPOPT_CCNEW : TCPOPT_CC;
510 			opt[optlen++] = TCPOLEN_CC;
511 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
512  			optlen += 4;
513 			break;
514 
515 		/*
516 		 * This is a SYN,ACK; send CC and CC.echo if we received
517 		 * CC from our peer.
518 		 */
519 		case (TH_SYN|TH_ACK):
520 			if (tp->t_flags & TF_RCVD_CC) {
521 				opt[optlen++] = TCPOPT_NOP;
522 				opt[optlen++] = TCPOPT_NOP;
523 				opt[optlen++] = TCPOPT_CC;
524 				opt[optlen++] = TCPOLEN_CC;
525 				*(u_int32_t *)&opt[optlen] =
526 					htonl(tp->cc_send);
527 				optlen += 4;
528 				opt[optlen++] = TCPOPT_NOP;
529 				opt[optlen++] = TCPOPT_NOP;
530 				opt[optlen++] = TCPOPT_CCECHO;
531 				opt[optlen++] = TCPOLEN_CC;
532 				*(u_int32_t *)&opt[optlen] =
533 					htonl(tp->cc_recv);
534 				optlen += 4;
535 			}
536 			break;
537 		}
538  	}
539 
540 #ifdef TCP_SIGNATURE
541 #ifdef INET6
542 	if (!isipv6)
543 #endif
544 	if (tp->t_flags & TF_SIGNATURE) {
545 		int i;
546 		u_char *bp;
547 
548 		/* Initialize TCP-MD5 option (RFC2385) */
549 		bp = (u_char *)opt + optlen;
550 		*bp++ = TCPOPT_SIGNATURE;
551 		*bp++ = TCPOLEN_SIGNATURE;
552 		sigoff = optlen + 2;
553 		for (i = 0; i < TCP_SIGLEN; i++)
554 			*bp++ = 0;
555 		optlen += TCPOLEN_SIGNATURE;
556 
557 		/* Terminate options list and maintain 32-bit alignment. */
558 		*bp++ = TCPOPT_NOP;
559 		*bp++ = TCPOPT_EOL;
560 		optlen += 2;
561 	}
562 #endif /* TCP_SIGNATURE */
563 
564  	hdrlen += optlen;
565 
566 #ifdef INET6
567 	if (isipv6)
568 		ipoptlen = ip6_optlen(tp->t_inpcb);
569 	else
570 #endif
571 	if (tp->t_inpcb->inp_options)
572 		ipoptlen = tp->t_inpcb->inp_options->m_len -
573 				offsetof(struct ipoption, ipopt_list);
574 	else
575 		ipoptlen = 0;
576 #ifdef IPSEC
577 	ipoptlen += ipsec_hdrsiz_tcp(tp);
578 #endif
579 
580 	/*
581 	 * Adjust data length if insertion of options will
582 	 * bump the packet length beyond the t_maxopd length.
583 	 * Clear the FIN bit because we cut off the tail of
584 	 * the segment.
585 	 */
586 	if (len + optlen + ipoptlen > tp->t_maxopd) {
587 		/*
588 		 * If there is still more to send, don't close the connection.
589 		 */
590 		flags &= ~TH_FIN;
591 		len = tp->t_maxopd - optlen - ipoptlen;
592 		sendalot = 1;
593 	}
594 
595 /*#ifdef DIAGNOSTIC*/
596 #ifdef INET6
597  	if (max_linkhdr + hdrlen > MCLBYTES)
598 #else
599  	if (max_linkhdr + hdrlen > MHLEN)
600 #endif
601 		panic("tcphdr too big");
602 /*#endif*/
603 
604 	/*
605 	 * Grab a header mbuf, attaching a copy of data to
606 	 * be transmitted, and initialize the header from
607 	 * the template for sends on this connection.
608 	 */
609 	if (len) {
610 		if (tp->t_force && len == 1)
611 			tcpstat.tcps_sndprobe++;
612 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
613 			tcpstat.tcps_sndrexmitpack++;
614 			tcpstat.tcps_sndrexmitbyte += len;
615 		} else {
616 			tcpstat.tcps_sndpack++;
617 			tcpstat.tcps_sndbyte += len;
618 		}
619 #ifdef notyet
620 		if ((m = m_copypack(so->so_snd.sb_mb, off,
621 		    (int)len, max_linkhdr + hdrlen)) == 0) {
622 			error = ENOBUFS;
623 			goto out;
624 		}
625 		/*
626 		 * m_copypack left space for our hdr; use it.
627 		 */
628 		m->m_len += hdrlen;
629 		m->m_data -= hdrlen;
630 #else
631 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
632 		if (m == NULL) {
633 			error = ENOBUFS;
634 			goto out;
635 		}
636 #ifdef INET6
637 		if (MHLEN < hdrlen + max_linkhdr) {
638 			MCLGET(m, M_DONTWAIT);
639 			if ((m->m_flags & M_EXT) == 0) {
640 				m_freem(m);
641 				error = ENOBUFS;
642 				goto out;
643 			}
644 		}
645 #endif
646 		m->m_data += max_linkhdr;
647 		m->m_len = hdrlen;
648 		if (len <= MHLEN - hdrlen - max_linkhdr) {
649 			m_copydata(so->so_snd.sb_mb, off, (int) len,
650 			    mtod(m, caddr_t) + hdrlen);
651 			m->m_len += len;
652 		} else {
653 			m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len);
654 			if (m->m_next == 0) {
655 				(void) m_free(m);
656 				error = ENOBUFS;
657 				goto out;
658 			}
659 		}
660 #endif
661 		/*
662 		 * If we're sending everything we've got, set PUSH.
663 		 * (This will keep happy those implementations which only
664 		 * give data to the user when a buffer fills or
665 		 * a PUSH comes in.)
666 		 */
667 		if (off + len == so->so_snd.sb_cc)
668 			flags |= TH_PUSH;
669 	} else {
670 		if (tp->t_flags & TF_ACKNOW)
671 			tcpstat.tcps_sndacks++;
672 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
673 			tcpstat.tcps_sndctrl++;
674 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
675 			tcpstat.tcps_sndurg++;
676 		else
677 			tcpstat.tcps_sndwinup++;
678 
679 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
680 		if (m == NULL) {
681 			error = ENOBUFS;
682 			goto out;
683 		}
684 #ifdef INET6
685 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
686 		    MHLEN >= hdrlen) {
687 			MH_ALIGN(m, hdrlen);
688 		} else
689 #endif
690 		m->m_data += max_linkhdr;
691 		m->m_len = hdrlen;
692 	}
693 	m->m_pkthdr.rcvif = (struct ifnet *)0;
694 #ifdef MAC
695 	mac_create_mbuf_from_inpcb(tp->t_inpcb, m);
696 #endif
697 #ifdef INET6
698 	if (isipv6) {
699 		ip6 = mtod(m, struct ip6_hdr *);
700 		th = (struct tcphdr *)(ip6 + 1);
701 		tcpip_fillheaders(tp->t_inpcb, ip6, th);
702 	} else
703 #endif /* INET6 */
704       {
705 	ip = mtod(m, struct ip *);
706 	ipov = (struct ipovly *)ip;
707 	th = (struct tcphdr *)(ip + 1);
708 	tcpip_fillheaders(tp->t_inpcb, ip, th);
709       }
710 
711 	/*
712 	 * Fill in fields, remembering maximum advertised
713 	 * window for use in delaying messages about window sizes.
714 	 * If resending a FIN, be sure not to use a new sequence number.
715 	 */
716 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
717 	    tp->snd_nxt == tp->snd_max)
718 		tp->snd_nxt--;
719 	/*
720 	 * If we are doing retransmissions, then snd_nxt will
721 	 * not reflect the first unsent octet.  For ACK only
722 	 * packets, we do not want the sequence number of the
723 	 * retransmitted packet, we want the sequence number
724 	 * of the next unsent octet.  So, if there is no data
725 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
726 	 * when filling in ti_seq.  But if we are in persist
727 	 * state, snd_max might reflect one byte beyond the
728 	 * right edge of the window, so use snd_nxt in that
729 	 * case, since we know we aren't doing a retransmission.
730 	 * (retransmit and persist are mutually exclusive...)
731 	 */
732 	if (len || (flags & (TH_SYN|TH_FIN))
733 	    || callout_active(tp->tt_persist))
734 		th->th_seq = htonl(tp->snd_nxt);
735 	else
736 		th->th_seq = htonl(tp->snd_max);
737 	th->th_ack = htonl(tp->rcv_nxt);
738 	if (optlen) {
739 		bcopy(opt, th + 1, optlen);
740 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
741 	}
742 	th->th_flags = flags;
743 	/*
744 	 * Calculate receive window.  Don't shrink window,
745 	 * but avoid silly window syndrome.
746 	 */
747 	if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
748 	    recwin < (long)tp->t_maxseg)
749 		recwin = 0;
750 	if (recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
751 		recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
752 	if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
753 		recwin = (long)TCP_MAXWIN << tp->rcv_scale;
754 	th->th_win = htons((u_short) (recwin >> tp->rcv_scale));
755 
756 
757 	/*
758 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
759 	 * a 0 window.  This may cause the remote transmitter to stall.  This
760 	 * flag tells soreceive() to disable delayed acknowledgements when
761 	 * draining the buffer.  This can occur if the receiver is attempting
762 	 * to read more data then can be buffered prior to transmitting on
763 	 * the connection.
764 	 */
765 	if (recwin == 0)
766 		tp->t_flags |= TF_RXWIN0SENT;
767 	else
768 		tp->t_flags &= ~TF_RXWIN0SENT;
769 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
770 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
771 		th->th_flags |= TH_URG;
772 	} else
773 		/*
774 		 * If no urgent pointer to send, then we pull
775 		 * the urgent pointer to the left edge of the send window
776 		 * so that it doesn't drift into the send window on sequence
777 		 * number wraparound.
778 		 */
779 		tp->snd_up = tp->snd_una;		/* drag it along */
780 
781 #ifdef TCP_SIGNATURE
782 #ifdef INET6
783 	if (!isipv6)
784 #endif
785 	if (tp->t_flags & TF_SIGNATURE)
786 		tcp_signature_compute(m, sizeof(struct ip), len, optlen,
787 		    (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
788 #endif
789 
790 	/*
791 	 * Put TCP length in extended header, and then
792 	 * checksum extended header and data.
793 	 */
794 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
795 #ifdef INET6
796 	if (isipv6)
797 		/*
798 		 * ip6_plen is not need to be filled now, and will be filled
799 		 * in ip6_output.
800 		 */
801 		th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
802 				       sizeof(struct tcphdr) + optlen + len);
803 	else
804 #endif /* INET6 */
805       {
806 	m->m_pkthdr.csum_flags = CSUM_TCP;
807 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
808 	th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
809 	    htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
810 
811 	/* IP version must be set here for ipv4/ipv6 checking later */
812 	KASSERT(ip->ip_v == IPVERSION,
813 	    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
814       }
815 
816 	/*
817 	 * In transmit state, time the transmission and arrange for
818 	 * the retransmit.  In persist state, just set snd_max.
819 	 */
820 	if (tp->t_force == 0 || !callout_active(tp->tt_persist)) {
821 		tcp_seq startseq = tp->snd_nxt;
822 
823 		/*
824 		 * Advance snd_nxt over sequence space of this segment.
825 		 */
826 		if (flags & (TH_SYN|TH_FIN)) {
827 			if (flags & TH_SYN)
828 				tp->snd_nxt++;
829 			if (flags & TH_FIN) {
830 				tp->snd_nxt++;
831 				tp->t_flags |= TF_SENTFIN;
832 			}
833 		}
834 		tp->snd_nxt += len;
835 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
836 			tp->snd_max = tp->snd_nxt;
837 			/*
838 			 * Time this transmission if not a retransmission and
839 			 * not currently timing anything.
840 			 */
841 			if (tp->t_rtttime == 0) {
842 				tp->t_rtttime = ticks;
843 				tp->t_rtseq = startseq;
844 				tcpstat.tcps_segstimed++;
845 			}
846 		}
847 
848 		/*
849 		 * Set retransmit timer if not currently set,
850 		 * and not doing a pure ack or a keep-alive probe.
851 		 * Initial value for retransmit timer is smoothed
852 		 * round-trip time + 2 * round-trip time variance.
853 		 * Initialize shift counter which is used for backoff
854 		 * of retransmit time.
855 		 */
856 		if (!callout_active(tp->tt_rexmt) &&
857 		    tp->snd_nxt != tp->snd_una) {
858 			if (callout_active(tp->tt_persist)) {
859 				callout_stop(tp->tt_persist);
860 				tp->t_rxtshift = 0;
861 			}
862 			callout_reset(tp->tt_rexmt, tp->t_rxtcur,
863 				      tcp_timer_rexmt, tp);
864 		}
865 	} else {
866 		/*
867 		 * Persist case, update snd_max but since we are in
868 		 * persist mode (no window) we do not update snd_nxt.
869 		 */
870 		int xlen = len;
871 		if (flags & TH_SYN)
872 			++xlen;
873 		if (flags & TH_FIN) {
874 			++xlen;
875 			tp->t_flags |= TF_SENTFIN;
876 		}
877 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
878 			tp->snd_max = tp->snd_nxt + len;
879 	}
880 
881 #ifdef TCPDEBUG
882 	/*
883 	 * Trace.
884 	 */
885 	if (so->so_options & SO_DEBUG) {
886 		u_short save = ipov->ih_len;
887 		ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
888 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
889 		ipov->ih_len = save;
890 	}
891 #endif
892 
893 	/*
894 	 * Fill in IP length and desired time to live and
895 	 * send to IP level.  There should be a better way
896 	 * to handle ttl and tos; we could keep them in
897 	 * the template, but need a way to checksum without them.
898 	 */
899 	/*
900 	 * m->m_pkthdr.len should have been set before cksum calcuration,
901 	 * because in6_cksum() need it.
902 	 */
903 #ifdef INET6
904 	if (isipv6) {
905 		/*
906 		 * we separately set hoplimit for every segment, since the
907 		 * user might want to change the value via setsockopt.
908 		 * Also, desired default hop limit might be changed via
909 		 * Neighbor Discovery.
910 		 */
911 		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
912 
913 		/* TODO: IPv6 IP6TOS_ECT bit on */
914 		error = ip6_output(m,
915 			    tp->t_inpcb->in6p_outputopts, NULL,
916 			    (so->so_options & SO_DONTROUTE), NULL, NULL,
917 			    tp->t_inpcb);
918 	} else
919 #endif /* INET6 */
920     {
921 	ip->ip_len = m->m_pkthdr.len;
922 #ifdef INET6
923  	if (INP_CHECK_SOCKAF(so, AF_INET6))
924  		ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
925 #endif /* INET6 */
926 	/*
927 	 * If we do path MTU discovery, then we set DF on every packet.
928 	 * This might not be the best thing to do according to RFC3390
929 	 * Section 2. However the tcp hostcache migitates the problem
930 	 * so it affects only the first tcp connection with a host.
931 	 */
932 	if (path_mtu_discovery)
933 		ip->ip_off |= IP_DF;
934 
935 	error = ip_output(m, tp->t_inpcb->inp_options, NULL,
936 	    (so->so_options & SO_DONTROUTE), 0, tp->t_inpcb);
937     }
938 	if (error) {
939 
940 		/*
941 		 * We know that the packet was lost, so back out the
942 		 * sequence number advance, if any.
943 		 */
944 		if (tp->t_force == 0 || !callout_active(tp->tt_persist)) {
945 			/*
946 			 * No need to check for TH_FIN here because
947 			 * the TF_SENTFIN flag handles that case.
948 			 */
949 			if ((flags & TH_SYN) == 0)
950 				tp->snd_nxt -= len;
951 		}
952 
953 out:
954 		if (error == ENOBUFS) {
955 	                if (!callout_active(tp->tt_rexmt) &&
956                             !callout_active(tp->tt_persist))
957 	                        callout_reset(tp->tt_rexmt, tp->t_rxtcur,
958                                       tcp_timer_rexmt, tp);
959 			tcp_quench(tp->t_inpcb, 0);
960 			return (0);
961 		}
962 		if (error == EMSGSIZE) {
963 			/*
964 			 * ip_output() will have already fixed the route
965 			 * for us.  tcp_mtudisc() will, as its last action,
966 			 * initiate retransmission, so it is important to
967 			 * not do so here.
968 			 */
969 			tcp_mtudisc(tp->t_inpcb, 0);
970 			return 0;
971 		}
972 		if ((error == EHOSTUNREACH || error == ENETDOWN)
973 		    && TCPS_HAVERCVDSYN(tp->t_state)) {
974 			tp->t_softerror = error;
975 			return (0);
976 		}
977 		return (error);
978 	}
979 	tcpstat.tcps_sndtotal++;
980 
981 	/*
982 	 * Data sent (as far as we can tell).
983 	 * If this advertises a larger window than any other segment,
984 	 * then remember the size of the advertised window.
985 	 * Any pending ACK has now been sent.
986 	 */
987 	if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
988 		tp->rcv_adv = tp->rcv_nxt + recwin;
989 	tp->last_ack_sent = tp->rcv_nxt;
990 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
991 	if (callout_active(tp->tt_delack))
992 		callout_stop(tp->tt_delack);
993 #if 0
994 	/*
995 	 * This completely breaks TCP if newreno is turned on.  What happens
996 	 * is that if delayed-acks are turned on on the receiver, this code
997 	 * on the transmitter effectively destroys the TCP window, forcing
998 	 * it to four packets (1.5Kx4 = 6K window).
999 	 */
1000 	if (sendalot && (!tcp_do_newreno || --maxburst))
1001 		goto again;
1002 #endif
1003 	if (sendalot)
1004 		goto again;
1005 	return (0);
1006 }
1007 
1008 void
1009 tcp_setpersist(tp)
1010 	register struct tcpcb *tp;
1011 {
1012 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1013 	int tt;
1014 
1015 	if (callout_active(tp->tt_rexmt))
1016 		panic("tcp_setpersist: retransmit pending");
1017 	/*
1018 	 * Start/restart persistance timer.
1019 	 */
1020 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1021 		      TCPTV_PERSMIN, TCPTV_PERSMAX);
1022 	callout_reset(tp->tt_persist, tt, tcp_timer_persist, tp);
1023 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1024 		tp->t_rxtshift++;
1025 }
1026