xref: /freebsd/sys/netinet/tcp_output.c (revision 23f282aa31e9b6fceacd449020e936e98d6f2298)
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  * $FreeBSD$
35  */
36 
37 #include "opt_inet6.h"
38 #include "opt_ipsec.h"
39 #include "opt_tcpdebug.h"
40 
41 #include <stddef.h>
42 
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/kernel.h>
46 #include <sys/sysctl.h>
47 #include <sys/mbuf.h>
48 #include <sys/protosw.h>
49 #include <sys/socket.h>
50 #include <sys/socketvar.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 #ifdef INET6
58 #include <netinet/ip6.h>
59 #endif
60 #include <netinet/in_pcb.h>
61 #ifdef INET6
62 #include <netinet6/in6_pcb.h>
63 #endif
64 #include <netinet/ip_var.h>
65 #ifdef INET6
66 #include <netinet6/ip6_var.h>
67 #endif
68 #include <netinet/tcp.h>
69 #define	TCPOUTFLAGS
70 #include <netinet/tcp_fsm.h>
71 #include <netinet/tcp_seq.h>
72 #include <netinet/tcp_timer.h>
73 #include <netinet/tcp_var.h>
74 #include <netinet/tcpip.h>
75 #ifdef TCPDEBUG
76 #include <netinet/tcp_debug.h>
77 #endif
78 
79 #ifdef IPSEC
80 #include <netinet6/ipsec.h>
81 #endif /*IPSEC*/
82 
83 #include <machine/in_cksum.h>
84 
85 #ifdef notyet
86 extern struct mbuf *m_copypack();
87 #endif
88 
89 static 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 = TCP_MAXWIN;               /* something large */
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 /*
102  * Tcp output routine: figure out what should be sent and send it.
103  */
104 int
105 tcp_output(tp)
106 	register struct tcpcb *tp;
107 {
108 	register struct socket *so = tp->t_inpcb->inp_socket;
109 	register long len, win;
110 	int off, flags, error;
111 	register struct mbuf *m;
112 	struct ip *ip = NULL;
113 	register struct ipovly *ipov = NULL;
114 #ifdef INET6
115 	struct ip6_hdr *ip6 = NULL;
116 #endif /* INET6 */
117 	register struct tcphdr *th;
118 	u_char opt[TCP_MAXOLEN];
119 	unsigned ipoptlen, optlen, hdrlen;
120 	int idle, sendalot;
121 	struct rmxp_tao *taop;
122 	struct rmxp_tao tao_noncached;
123 #ifdef INET6
124 	int isipv6;
125 #endif
126 
127 #ifdef INET6
128 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
129 #endif
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 	idle = (tp->snd_max == tp->snd_una);
138 	if (idle && (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 (
148 #ifdef INET6
149 		    (isipv6 && in6_localaddr(&tp->t_inpcb->in6p_faddr)) ||
150 		    (!isipv6 &&
151 #endif
152 		     in_localaddr(tp->t_inpcb->inp_faddr)
153 #ifdef INET6
154 		     )
155 #endif
156 		    )
157 			tp->snd_cwnd = tp->t_maxseg * ss_fltsz_local;
158 		else
159 			tp->snd_cwnd = tp->t_maxseg * ss_fltsz;
160 	}
161 again:
162 	sendalot = 0;
163 	off = tp->snd_nxt - tp->snd_una;
164 	win = min(tp->snd_wnd, tp->snd_cwnd);
165 
166 	flags = tcp_outflags[tp->t_state];
167 	/*
168 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
169 	 * state flags.
170 	 */
171 	if (tp->t_flags & TF_NEEDFIN)
172 		flags |= TH_FIN;
173 	if (tp->t_flags & TF_NEEDSYN)
174 		flags |= TH_SYN;
175 
176 	/*
177 	 * If in persist timeout with window of 0, send 1 byte.
178 	 * Otherwise, if window is small but nonzero
179 	 * and timer expired, we will send what we can
180 	 * and go to transmit state.
181 	 */
182 	if (tp->t_force) {
183 		if (win == 0) {
184 			/*
185 			 * If we still have some data to send, then
186 			 * clear the FIN bit.  Usually this would
187 			 * happen below when it realizes that we
188 			 * aren't sending all the data.  However,
189 			 * if we have exactly 1 byte of unsent data,
190 			 * then it won't clear the FIN bit below,
191 			 * and if we are in persist state, we wind
192 			 * up sending the packet without recording
193 			 * that we sent the FIN bit.
194 			 *
195 			 * We can't just blindly clear the FIN bit,
196 			 * because if we don't have any more data
197 			 * to send then the probe will be the FIN
198 			 * itself.
199 			 */
200 			if (off < so->so_snd.sb_cc)
201 				flags &= ~TH_FIN;
202 			win = 1;
203 		} else {
204 			callout_stop(tp->tt_persist);
205 			tp->t_rxtshift = 0;
206 		}
207 	}
208 
209 	len = (long)ulmin(so->so_snd.sb_cc, win) - off;
210 
211 	if ((taop = tcp_gettaocache(tp->t_inpcb)) == NULL) {
212 		taop = &tao_noncached;
213 		bzero(taop, sizeof(*taop));
214 	}
215 
216 	/*
217 	 * Lop off SYN bit if it has already been sent.  However, if this
218 	 * is SYN-SENT state and if segment contains data and if we don't
219 	 * know that foreign host supports TAO, suppress sending segment.
220 	 */
221 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
222 		flags &= ~TH_SYN;
223 		off--, len++;
224 		if (len > 0 && tp->t_state == TCPS_SYN_SENT &&
225 		    taop->tao_ccsent == 0)
226 			return 0;
227 	}
228 
229 	/*
230 	 * Be careful not to send data and/or FIN on SYN segments
231 	 * in cases when no CC option will be sent.
232 	 * This measure is needed to prevent interoperability problems
233 	 * with not fully conformant TCP implementations.
234 	 */
235 	if ((flags & TH_SYN) &&
236 	    ((tp->t_flags & TF_NOOPT) || !(tp->t_flags & TF_REQ_CC) ||
237 	     ((flags & TH_ACK) && !(tp->t_flags & TF_RCVD_CC)))) {
238 		len = 0;
239 		flags &= ~TH_FIN;
240 	}
241 
242 	if (len < 0) {
243 		/*
244 		 * If FIN has been sent but not acked,
245 		 * but we haven't been called to retransmit,
246 		 * len will be -1.  Otherwise, window shrank
247 		 * after we sent into it.  If window shrank to 0,
248 		 * cancel pending retransmit, pull snd_nxt back
249 		 * to (closed) window, and set the persist timer
250 		 * if it isn't already going.  If the window didn't
251 		 * close completely, just wait for an ACK.
252 		 */
253 		len = 0;
254 		if (win == 0) {
255 			callout_stop(tp->tt_rexmt);
256 			tp->t_rxtshift = 0;
257 			tp->snd_nxt = tp->snd_una;
258 			if (!callout_active(tp->tt_persist))
259 				tcp_setpersist(tp);
260 		}
261 	}
262 	if (len > tp->t_maxseg) {
263 		len = tp->t_maxseg;
264 		sendalot = 1;
265 	}
266 	if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
267 		flags &= ~TH_FIN;
268 
269 	win = sbspace(&so->so_rcv);
270 
271 	/*
272 	 * Sender silly window avoidance.  If connection is idle
273 	 * and can send all data, a maximum segment,
274 	 * at least a maximum default-size segment do it,
275 	 * or are forced, do it; otherwise don't bother.
276 	 * If peer's buffer is tiny, then send
277 	 * when window is at least half open.
278 	 * If retransmitting (possibly after persist timer forced us
279 	 * to send into a small window), then must resend.
280 	 */
281 	if (len) {
282 		if (len == tp->t_maxseg)
283 			goto send;
284 		if (!(tp->t_flags & TF_MORETOCOME) &&
285 		    (idle || tp->t_flags & TF_NODELAY) &&
286 		    (tp->t_flags & TF_NOPUSH) == 0 &&
287 		    len + off >= so->so_snd.sb_cc)
288 			goto send;
289 		if (tp->t_force)
290 			goto send;
291 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
292 			goto send;
293 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
294 			goto send;
295 	}
296 
297 	/*
298 	 * Compare available window to amount of window
299 	 * known to peer (as advertised window less
300 	 * next expected input).  If the difference is at least two
301 	 * max size segments, or at least 50% of the maximum possible
302 	 * window, then want to send a window update to peer.
303 	 */
304 	if (win > 0) {
305 		/*
306 		 * "adv" is the amount we can increase the window,
307 		 * taking into account that we are limited by
308 		 * TCP_MAXWIN << tp->rcv_scale.
309 		 */
310 		long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) -
311 			(tp->rcv_adv - tp->rcv_nxt);
312 
313 		if (adv >= (long) (2 * tp->t_maxseg))
314 			goto send;
315 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
316 			goto send;
317 	}
318 
319 	/*
320 	 * Send if we owe peer an ACK.
321 	 */
322 	if (tp->t_flags & TF_ACKNOW)
323 		goto send;
324 	if ((flags & TH_RST) ||
325 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
326 		goto send;
327 	if (SEQ_GT(tp->snd_up, tp->snd_una))
328 		goto send;
329 	/*
330 	 * If our state indicates that FIN should be sent
331 	 * and we have not yet done so, or we're retransmitting the FIN,
332 	 * then we need to send.
333 	 */
334 	if (flags & TH_FIN &&
335 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
336 		goto send;
337 
338 	/*
339 	 * TCP window updates are not reliable, rather a polling protocol
340 	 * using ``persist'' packets is used to insure receipt of window
341 	 * updates.  The three ``states'' for the output side are:
342 	 *	idle			not doing retransmits or persists
343 	 *	persisting		to move a small or zero window
344 	 *	(re)transmitting	and thereby not persisting
345 	 *
346 	 * callout_active(tp->tt_persist)
347 	 *	is true when we are in persist state.
348 	 * tp->t_force
349 	 *	is set when we are called to send a persist packet.
350 	 * callout_active(tp->tt_rexmt)
351 	 *	is set when we are retransmitting
352 	 * The output side is idle when both timers are zero.
353 	 *
354 	 * If send window is too small, there is data to transmit, and no
355 	 * retransmit or persist is pending, then go to persist state.
356 	 * If nothing happens soon, send when timer expires:
357 	 * if window is nonzero, transmit what we can,
358 	 * otherwise force out a byte.
359 	 */
360 	if (so->so_snd.sb_cc && !callout_active(tp->tt_rexmt) &&
361 	    !callout_active(tp->tt_persist)) {
362 		tp->t_rxtshift = 0;
363 		tcp_setpersist(tp);
364 	}
365 
366 	/*
367 	 * No reason to send a segment, just return.
368 	 */
369 	return (0);
370 
371 send:
372 	/*
373 	 * Before ESTABLISHED, force sending of initial options
374 	 * unless TCP set not to do any options.
375 	 * NOTE: we assume that the IP/TCP header plus TCP options
376 	 * always fit in a single mbuf, leaving room for a maximum
377 	 * link header, i.e.
378 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MHLEN
379 	 */
380 	optlen = 0;
381 #ifdef INET6
382 	if (isipv6)
383 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
384 	else
385 #endif
386 	hdrlen = sizeof (struct tcpiphdr);
387 	if (flags & TH_SYN) {
388 		tp->snd_nxt = tp->iss;
389 		if ((tp->t_flags & TF_NOOPT) == 0) {
390 			u_short mss;
391 
392 			opt[0] = TCPOPT_MAXSEG;
393 			opt[1] = TCPOLEN_MAXSEG;
394 			mss = htons((u_short) tcp_mssopt(tp));
395 			(void)memcpy(opt + 2, &mss, sizeof(mss));
396 			optlen = TCPOLEN_MAXSEG;
397 
398 			if ((tp->t_flags & TF_REQ_SCALE) &&
399 			    ((flags & TH_ACK) == 0 ||
400 			    (tp->t_flags & TF_RCVD_SCALE))) {
401 				*((u_int32_t *)(opt + optlen)) = htonl(
402 					TCPOPT_NOP << 24 |
403 					TCPOPT_WINDOW << 16 |
404 					TCPOLEN_WINDOW << 8 |
405 					tp->request_r_scale);
406 				optlen += 4;
407 			}
408 		}
409  	}
410 
411  	/*
412 	 * Send a timestamp and echo-reply if this is a SYN and our side
413 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
414 	 * and our peer have sent timestamps in our SYN's.
415  	 */
416  	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
417  	    (flags & TH_RST) == 0 &&
418 	    ((flags & TH_ACK) == 0 ||
419 	     (tp->t_flags & TF_RCVD_TSTMP))) {
420 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
421 
422  		/* Form timestamp option as shown in appendix A of RFC 1323. */
423  		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
424  		*lp++ = htonl(ticks);
425  		*lp   = htonl(tp->ts_recent);
426  		optlen += TCPOLEN_TSTAMP_APPA;
427  	}
428 
429  	/*
430 	 * Send `CC-family' options if our side wants to use them (TF_REQ_CC),
431 	 * options are allowed (!TF_NOOPT) and it's not a RST.
432  	 */
433  	if ((tp->t_flags & (TF_REQ_CC|TF_NOOPT)) == TF_REQ_CC &&
434  	     (flags & TH_RST) == 0) {
435 		switch (flags & (TH_SYN|TH_ACK)) {
436 		/*
437 		 * This is a normal ACK, send CC if we received CC before
438 		 * from our peer.
439 		 */
440 		case TH_ACK:
441 			if (!(tp->t_flags & TF_RCVD_CC))
442 				break;
443 			/*FALLTHROUGH*/
444 
445 		/*
446 		 * We can only get here in T/TCP's SYN_SENT* state, when
447 		 * we're a sending a non-SYN segment without waiting for
448 		 * the ACK of our SYN.  A check above assures that we only
449 		 * do this if our peer understands T/TCP.
450 		 */
451 		case 0:
452 			opt[optlen++] = TCPOPT_NOP;
453 			opt[optlen++] = TCPOPT_NOP;
454 			opt[optlen++] = TCPOPT_CC;
455 			opt[optlen++] = TCPOLEN_CC;
456 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
457 
458 			optlen += 4;
459 			break;
460 
461 		/*
462 		 * This is our initial SYN, check whether we have to use
463 		 * CC or CC.new.
464 		 */
465 		case TH_SYN:
466 			opt[optlen++] = TCPOPT_NOP;
467 			opt[optlen++] = TCPOPT_NOP;
468 			opt[optlen++] = tp->t_flags & TF_SENDCCNEW ?
469 						TCPOPT_CCNEW : TCPOPT_CC;
470 			opt[optlen++] = TCPOLEN_CC;
471 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
472  			optlen += 4;
473 			break;
474 
475 		/*
476 		 * This is a SYN,ACK; send CC and CC.echo if we received
477 		 * CC from our peer.
478 		 */
479 		case (TH_SYN|TH_ACK):
480 			if (tp->t_flags & TF_RCVD_CC) {
481 				opt[optlen++] = TCPOPT_NOP;
482 				opt[optlen++] = TCPOPT_NOP;
483 				opt[optlen++] = TCPOPT_CC;
484 				opt[optlen++] = TCPOLEN_CC;
485 				*(u_int32_t *)&opt[optlen] =
486 					htonl(tp->cc_send);
487 				optlen += 4;
488 				opt[optlen++] = TCPOPT_NOP;
489 				opt[optlen++] = TCPOPT_NOP;
490 				opt[optlen++] = TCPOPT_CCECHO;
491 				opt[optlen++] = TCPOLEN_CC;
492 				*(u_int32_t *)&opt[optlen] =
493 					htonl(tp->cc_recv);
494 				optlen += 4;
495 			}
496 			break;
497 		}
498  	}
499 
500  	hdrlen += optlen;
501 
502 #ifdef INET6
503 	if (isipv6)
504 		ipoptlen = ip6_optlen(tp->t_inpcb);
505 	else
506 #endif
507       {
508 	if (tp->t_inpcb->inp_options) {
509 		ipoptlen = tp->t_inpcb->inp_options->m_len -
510 				offsetof(struct ipoption, ipopt_list);
511 	} else {
512 		ipoptlen = 0;
513 	}
514       }
515 #ifdef IPSEC
516 	ipoptlen += ipsec_hdrsiz_tcp(tp);
517 #endif
518 
519 	/*
520 	 * Adjust data length if insertion of options will
521 	 * bump the packet length beyond the t_maxopd length.
522 	 * Clear the FIN bit because we cut off the tail of
523 	 * the segment.
524 	 */
525 	if (len + optlen + ipoptlen > tp->t_maxopd) {
526 		/*
527 		 * If there is still more to send, don't close the connection.
528 		 */
529 		flags &= ~TH_FIN;
530 		len = tp->t_maxopd - optlen - ipoptlen;
531 		sendalot = 1;
532 	}
533 
534 /*#ifdef DIAGNOSTIC*/
535 #ifdef INET6
536  	if (max_linkhdr + hdrlen > MCLBYTES)
537 		panic("tcphdr too big");
538 #else
539  	if (max_linkhdr + hdrlen > MHLEN)
540 		panic("tcphdr too big");
541 #endif
542 /*#endif*/
543 
544 	/*
545 	 * Grab a header mbuf, attaching a copy of data to
546 	 * be transmitted, and initialize the header from
547 	 * the template for sends on this connection.
548 	 */
549 	if (len) {
550 		if (tp->t_force && len == 1)
551 			tcpstat.tcps_sndprobe++;
552 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
553 			tcpstat.tcps_sndrexmitpack++;
554 			tcpstat.tcps_sndrexmitbyte += len;
555 		} else {
556 			tcpstat.tcps_sndpack++;
557 			tcpstat.tcps_sndbyte += len;
558 		}
559 #ifdef notyet
560 		if ((m = m_copypack(so->so_snd.sb_mb, off,
561 		    (int)len, max_linkhdr + hdrlen)) == 0) {
562 			error = ENOBUFS;
563 			goto out;
564 		}
565 		/*
566 		 * m_copypack left space for our hdr; use it.
567 		 */
568 		m->m_len += hdrlen;
569 		m->m_data -= hdrlen;
570 #else
571 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
572 		if (m == NULL) {
573 			error = ENOBUFS;
574 			goto out;
575 		}
576 #ifdef INET6
577 		if (MHLEN < hdrlen + max_linkhdr) {
578 			MCLGET(m, M_DONTWAIT);
579 			if ((m->m_flags & M_EXT) == 0) {
580 				m_freem(m);
581 				error = ENOBUFS;
582 				goto out;
583 			}
584 		}
585 #endif
586 		m->m_data += max_linkhdr;
587 		m->m_len = hdrlen;
588 		if (len <= MHLEN - hdrlen - max_linkhdr) {
589 			m_copydata(so->so_snd.sb_mb, off, (int) len,
590 			    mtod(m, caddr_t) + hdrlen);
591 			m->m_len += len;
592 		} else {
593 			m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len);
594 			if (m->m_next == 0) {
595 				(void) m_free(m);
596 				error = ENOBUFS;
597 				goto out;
598 			}
599 		}
600 #endif
601 		/*
602 		 * If we're sending everything we've got, set PUSH.
603 		 * (This will keep happy those implementations which only
604 		 * give data to the user when a buffer fills or
605 		 * a PUSH comes in.)
606 		 */
607 		if (off + len == so->so_snd.sb_cc)
608 			flags |= TH_PUSH;
609 	} else {
610 		if (tp->t_flags & TF_ACKNOW)
611 			tcpstat.tcps_sndacks++;
612 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
613 			tcpstat.tcps_sndctrl++;
614 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
615 			tcpstat.tcps_sndurg++;
616 		else
617 			tcpstat.tcps_sndwinup++;
618 
619 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
620 		if (m == NULL) {
621 			error = ENOBUFS;
622 			goto out;
623 		}
624 #ifdef INET6
625 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
626 		    MHLEN >= hdrlen) {
627 			MH_ALIGN(m, hdrlen);
628 		} else
629 #endif
630 		m->m_data += max_linkhdr;
631 		m->m_len = hdrlen;
632 	}
633 	m->m_pkthdr.rcvif = (struct ifnet *)0;
634 	if (tp->t_template == 0)
635 		panic("tcp_output");
636 #ifdef INET6
637 	if (isipv6) {
638 		ip6 = mtod(m, struct ip6_hdr *);
639 		th = (struct tcphdr *)(ip6 + 1);
640 		bcopy((caddr_t)tp->t_template->tt_ipgen, (caddr_t)ip6,
641 		      sizeof(struct ip6_hdr));
642 		bcopy((caddr_t)&tp->t_template->tt_t, (caddr_t)th,
643 		      sizeof(struct tcphdr));
644 	} else
645 #endif /* INET6 */
646       {
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 	bcopy((caddr_t)tp->t_template->tt_ipgen, (caddr_t)ip,
652 	      sizeof(struct ip));
653 	bcopy((caddr_t)&tp->t_template->tt_t, (caddr_t)th,
654 	      sizeof(struct tcphdr));
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))
679 	    || callout_active(tp->tt_persist))
680 		th->th_seq = htonl(tp->snd_nxt);
681 	else
682 		th->th_seq = htonl(tp->snd_max);
683 	th->th_ack = htonl(tp->rcv_nxt);
684 	if (optlen) {
685 		bcopy(opt, th + 1, optlen);
686 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
687 	}
688 	th->th_flags = flags;
689 	/*
690 	 * Calculate receive window.  Don't shrink window,
691 	 * but avoid silly window syndrome.
692 	 */
693 	if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)tp->t_maxseg)
694 		win = 0;
695 	if (win < (long)(tp->rcv_adv - tp->rcv_nxt))
696 		win = (long)(tp->rcv_adv - tp->rcv_nxt);
697 	if (win > (long)TCP_MAXWIN << tp->rcv_scale)
698 		win = (long)TCP_MAXWIN << tp->rcv_scale;
699 	th->th_win = htons((u_short) (win>>tp->rcv_scale));
700 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
701 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
702 		th->th_flags |= TH_URG;
703 	} else
704 		/*
705 		 * If no urgent pointer to send, then we pull
706 		 * the urgent pointer to the left edge of the send window
707 		 * so that it doesn't drift into the send window on sequence
708 		 * number wraparound.
709 		 */
710 		tp->snd_up = tp->snd_una;		/* drag it along */
711 
712 	/*
713 	 * Put TCP length in extended header, and then
714 	 * checksum extended header and data.
715 	 */
716 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
717 #ifdef INET6
718 	if (isipv6)
719 		/*
720 		 * ip6_plen is not need to be filled now, and will be filled
721 		 * in ip6_output.
722 		 */
723 		th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
724 				       sizeof(struct tcphdr) + optlen + len);
725 	else
726 #endif /* INET6 */
727       {
728 	m->m_pkthdr.csum_flags = CSUM_TCP;
729 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
730 	if (len + optlen)
731 		th->th_sum = in_addword(th->th_sum,
732 		    htons((u_short)(optlen + len)));
733 
734 	/* IP version must be set here for ipv4/ipv6 checking later */
735 	KASSERT(ip->ip_v == IPVERSION,
736 	    ("%s: IP version incorrect: %d", __FUNCTION__, ip->ip_v));
737       }
738 
739 	/*
740 	 * In transmit state, time the transmission and arrange for
741 	 * the retransmit.  In persist state, just set snd_max.
742 	 */
743 	if (tp->t_force == 0 || !callout_active(tp->tt_persist)) {
744 		tcp_seq startseq = tp->snd_nxt;
745 
746 		/*
747 		 * Advance snd_nxt over sequence space of this segment.
748 		 */
749 		if (flags & (TH_SYN|TH_FIN)) {
750 			if (flags & TH_SYN)
751 				tp->snd_nxt++;
752 			if (flags & TH_FIN) {
753 				tp->snd_nxt++;
754 				tp->t_flags |= TF_SENTFIN;
755 			}
756 		}
757 		tp->snd_nxt += len;
758 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
759 			tp->snd_max = tp->snd_nxt;
760 			/*
761 			 * Time this transmission if not a retransmission and
762 			 * not currently timing anything.
763 			 */
764 			if (tp->t_rtttime == 0) {
765 				tp->t_rtttime = ticks;
766 				tp->t_rtseq = startseq;
767 				tcpstat.tcps_segstimed++;
768 			}
769 		}
770 
771 		/*
772 		 * Set retransmit timer if not currently set,
773 		 * and not doing an ack or a keep-alive probe.
774 		 * Initial value for retransmit timer is smoothed
775 		 * round-trip time + 2 * round-trip time variance.
776 		 * Initialize shift counter which is used for backoff
777 		 * of retransmit time.
778 		 */
779 		if (!callout_active(tp->tt_rexmt) &&
780 		    tp->snd_nxt != tp->snd_una) {
781 			callout_reset(tp->tt_rexmt, tp->t_rxtcur,
782 				      tcp_timer_rexmt, tp);
783 			if (callout_active(tp->tt_persist)) {
784 				callout_stop(tp->tt_persist);
785 				tp->t_rxtshift = 0;
786 			}
787 		}
788 	} else
789 		if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
790 			tp->snd_max = tp->snd_nxt + len;
791 
792 #ifdef TCPDEBUG
793 	/*
794 	 * Trace.
795 	 */
796 	if (so->so_options & SO_DEBUG)
797 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
798 #endif
799 
800 	/*
801 	 * Fill in IP length and desired time to live and
802 	 * send to IP level.  There should be a better way
803 	 * to handle ttl and tos; we could keep them in
804 	 * the template, but need a way to checksum without them.
805 	 */
806 	/*
807 	 * m->m_pkthdr.len should have been set before cksum calcuration,
808 	 * because in6_cksum() need it.
809 	 */
810 #ifdef INET6
811 	if (isipv6) {
812 		/*
813 		 * we separately set hoplimit for every segment, since the
814 		 * user might want to change the value via setsockopt.
815 		 * Also, desired default hop limit might be changed via
816 		 * Neighbor Discovery.
817 		 */
818 		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb,
819 					       tp->t_inpcb->in6p_route.ro_rt ?
820 					       tp->t_inpcb->in6p_route.ro_rt->rt_ifp
821 					       : NULL);
822 
823 		/* TODO: IPv6 IP6TOS_ECT bit on */
824 #ifdef IPSEC
825 		m->m_pkthdr.rcvif = (struct ifnet *)so;
826 #endif /*IPSEC*/
827 		error = ip6_output(m,
828 			    tp->t_inpcb->in6p_outputopts,
829 			    &tp->t_inpcb->in6p_route,
830 			    (so->so_options & SO_DONTROUTE)|IPV6_SOCKINMRCVIF,
831 			    NULL, NULL);
832 	} else
833 #endif /* INET6 */
834     {
835 	struct rtentry *rt;
836 	ip->ip_len = m->m_pkthdr.len;
837 	ip->ip_ttl = tp->t_inpcb->inp_ip_ttl;	/* XXX */
838 	ip->ip_tos = tp->t_inpcb->inp_ip_tos;	/* XXX */
839 	/*
840 	 * See if we should do MTU discovery.  We do it only if the following
841 	 * are true:
842 	 *	1) we have a valid route to the destination
843 	 *	2) the MTU is not locked (if it is, then discovery has been
844 	 *	   disabled)
845 	 */
846 	if (path_mtu_discovery
847 	    && (rt = tp->t_inpcb->inp_route.ro_rt)
848 	    && rt->rt_flags & RTF_UP
849 	    && !(rt->rt_rmx.rmx_locks & RTV_MTU)) {
850 		ip->ip_off |= IP_DF;
851 	}
852 	error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
853 	    (so->so_options & SO_DONTROUTE)|IP_SOCKINMRCVIF, 0);
854     }
855 	if (error) {
856 out:
857 		if (error == ENOBUFS) {
858 			tcp_quench(tp->t_inpcb, 0);
859 			return (0);
860 		}
861 		if (error == EMSGSIZE) {
862 			/*
863 			 * ip_output() will have already fixed the route
864 			 * for us.  tcp_mtudisc() will, as its last action,
865 			 * initiate retransmission, so it is important to
866 			 * not do so here.
867 			 */
868 			tcp_mtudisc(tp->t_inpcb, 0);
869 			return 0;
870 		}
871 		if ((error == EHOSTUNREACH || error == ENETDOWN)
872 		    && TCPS_HAVERCVDSYN(tp->t_state)) {
873 			tp->t_softerror = error;
874 			return (0);
875 		}
876 		return (error);
877 	}
878 	tcpstat.tcps_sndtotal++;
879 
880 	/*
881 	 * Data sent (as far as we can tell).
882 	 * If this advertises a larger window than any other segment,
883 	 * then remember the size of the advertised window.
884 	 * Any pending ACK has now been sent.
885 	 */
886 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
887 		tp->rcv_adv = tp->rcv_nxt + win;
888 	tp->last_ack_sent = tp->rcv_nxt;
889 	tp->t_flags &= ~TF_ACKNOW;
890 	if (tcp_delack_enabled)
891 		callout_stop(tp->tt_delack);
892 	if (sendalot)
893 		goto again;
894 	return (0);
895 }
896 
897 void
898 tcp_setpersist(tp)
899 	register struct tcpcb *tp;
900 {
901 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
902 	int tt;
903 
904 	if (callout_active(tp->tt_rexmt))
905 		panic("tcp_setpersist: retransmit pending");
906 	/*
907 	 * Start/restart persistance timer.
908 	 */
909 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
910 		      TCPTV_PERSMIN, TCPTV_PERSMAX);
911 	callout_reset(tp->tt_persist, tt, tcp_timer_persist, tp);
912 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
913 		tp->t_rxtshift++;
914 }
915