xref: /freebsd/sys/netinet/tcp_output.c (revision 1b6c76a2fe091c74f08427e6c870851025a9cf67)
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 <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/domain.h>
44 #include <sys/kernel.h>
45 #include <sys/lock.h>
46 #include <sys/mbuf.h>
47 #include <sys/mutex.h>
48 #include <sys/protosw.h>
49 #include <sys/socket.h>
50 #include <sys/socketvar.h>
51 #include <sys/sysctl.h>
52 
53 #include <net/route.h>
54 
55 #include <netinet/in.h>
56 #include <netinet/in_systm.h>
57 #include <netinet/ip.h>
58 #include <netinet/in_pcb.h>
59 #include <netinet/ip_var.h>
60 #ifdef INET6
61 #include <netinet6/in6_pcb.h>
62 #include <netinet/ip6.h>
63 #include <netinet6/ip6_var.h>
64 #endif
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 #include <machine/in_cksum.h>
81 
82 #ifdef notyet
83 extern struct mbuf *m_copypack();
84 #endif
85 
86 static int path_mtu_discovery = 1;
87 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
88 	&path_mtu_discovery, 1, "Enable Path MTU Discovery");
89 
90 int ss_fltsz = 1;
91 SYSCTL_INT(_net_inet_tcp, OID_AUTO, slowstart_flightsize, CTLFLAG_RW,
92 	&ss_fltsz, 1, "Slow start flight size");
93 
94 int ss_fltsz_local = TCP_MAXWIN;               /* something large */
95 SYSCTL_INT(_net_inet_tcp, OID_AUTO, local_slowstart_flightsize, CTLFLAG_RW,
96 	&ss_fltsz_local, 1, "Slow start flight size for local networks");
97 
98 int     tcp_do_newreno = 1;
99 SYSCTL_INT(_net_inet_tcp, OID_AUTO, newreno, CTLFLAG_RW, &tcp_do_newreno,
100         0, "Enable NewReno Algorithms");
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 	int maxburst = TCP_MAXBURST;
122 	struct rmxp_tao *taop;
123 	struct rmxp_tao tao_noncached;
124 #ifdef INET6
125 	int isipv6;
126 #endif
127 
128 #ifdef INET6
129 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
130 #endif
131 
132 	/*
133 	 * Determine length of data that should be transmitted,
134 	 * and flags that will be used.
135 	 * If there is some data or critical controls (SYN, RST)
136 	 * to send, then transmit; otherwise, investigate further.
137 	 */
138 	idle = (tp->snd_max == tp->snd_una);
139 	if (idle && (ticks - tp->t_rcvtime) >= tp->t_rxtcur) {
140 		/*
141 		 * We have been idle for "a while" and no acks are
142 		 * expected to clock out any data we send --
143 		 * slow start to get ack "clock" running again.
144 		 *
145 		 * Set the slow-start flight size depending on whether
146 		 * this is a local network or not.
147 		 */
148 		if (
149 #ifdef INET6
150 		    (isipv6 && in6_localaddr(&tp->t_inpcb->in6p_faddr)) ||
151 		    (!isipv6 &&
152 #endif
153 		     in_localaddr(tp->t_inpcb->inp_faddr)
154 #ifdef INET6
155 		     )
156 #endif
157 		    )
158 			tp->snd_cwnd = tp->t_maxseg * ss_fltsz_local;
159 		else
160 			tp->snd_cwnd = tp->t_maxseg * ss_fltsz;
161 	}
162 again:
163 	sendalot = 0;
164 	off = tp->snd_nxt - tp->snd_una;
165 	win = min(tp->snd_wnd, tp->snd_cwnd);
166 
167 	flags = tcp_outflags[tp->t_state];
168 	/*
169 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
170 	 * state flags.
171 	 */
172 	if (tp->t_flags & TF_NEEDFIN)
173 		flags |= TH_FIN;
174 	if (tp->t_flags & TF_NEEDSYN)
175 		flags |= TH_SYN;
176 
177 	/*
178 	 * If in persist timeout with window of 0, send 1 byte.
179 	 * Otherwise, if window is small but nonzero
180 	 * and timer expired, we will send what we can
181 	 * and go to transmit state.
182 	 */
183 	if (tp->t_force) {
184 		if (win == 0) {
185 			/*
186 			 * If we still have some data to send, then
187 			 * clear the FIN bit.  Usually this would
188 			 * happen below when it realizes that we
189 			 * aren't sending all the data.  However,
190 			 * if we have exactly 1 byte of unsent data,
191 			 * then it won't clear the FIN bit below,
192 			 * and if we are in persist state, we wind
193 			 * up sending the packet without recording
194 			 * that we sent the FIN bit.
195 			 *
196 			 * We can't just blindly clear the FIN bit,
197 			 * because if we don't have any more data
198 			 * to send then the probe will be the FIN
199 			 * itself.
200 			 */
201 			if (off < so->so_snd.sb_cc)
202 				flags &= ~TH_FIN;
203 			win = 1;
204 		} else {
205 			callout_stop(tp->tt_persist);
206 			tp->t_rxtshift = 0;
207 		}
208 	}
209 
210 	len = (long)ulmin(so->so_snd.sb_cc, win) - off;
211 
212 	if ((taop = tcp_gettaocache(tp->t_inpcb)) == NULL) {
213 		taop = &tao_noncached;
214 		bzero(taop, sizeof(*taop));
215 	}
216 
217 	/*
218 	 * Lop off SYN bit if it has already been sent.  However, if this
219 	 * is SYN-SENT state and if segment contains data and if we don't
220 	 * know that foreign host supports TAO, suppress sending segment.
221 	 */
222 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
223 		flags &= ~TH_SYN;
224 		off--, len++;
225 		if (len > 0 && tp->t_state == TCPS_SYN_SENT &&
226 		    taop->tao_ccsent == 0)
227 			return 0;
228 	}
229 
230 	/*
231 	 * Be careful not to send data and/or FIN on SYN segments
232 	 * in cases when no CC option will be sent.
233 	 * This measure is needed to prevent interoperability problems
234 	 * with not fully conformant TCP implementations.
235 	 */
236 	if ((flags & TH_SYN) &&
237 	    ((tp->t_flags & TF_NOOPT) || !(tp->t_flags & TF_REQ_CC) ||
238 	     ((flags & TH_ACK) && !(tp->t_flags & TF_RCVD_CC)))) {
239 		len = 0;
240 		flags &= ~TH_FIN;
241 	}
242 
243 	if (len < 0) {
244 		/*
245 		 * If FIN has been sent but not acked,
246 		 * but we haven't been called to retransmit,
247 		 * len will be -1.  Otherwise, window shrank
248 		 * after we sent into it.  If window shrank to 0,
249 		 * cancel pending retransmit, pull snd_nxt back
250 		 * to (closed) window, and set the persist timer
251 		 * if it isn't already going.  If the window didn't
252 		 * close completely, just wait for an ACK.
253 		 */
254 		len = 0;
255 		if (win == 0) {
256 			callout_stop(tp->tt_rexmt);
257 			tp->t_rxtshift = 0;
258 			tp->snd_nxt = tp->snd_una;
259 			if (!callout_active(tp->tt_persist))
260 				tcp_setpersist(tp);
261 		}
262 	}
263 	if (len > tp->t_maxseg) {
264 		len = tp->t_maxseg;
265 		sendalot = 1;
266 	}
267 	if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
268 		flags &= ~TH_FIN;
269 
270 	win = sbspace(&so->so_rcv);
271 
272 	/*
273 	 * Sender silly window avoidance.  If connection is idle
274 	 * and can send all data, a maximum segment,
275 	 * at least a maximum default-size segment do it,
276 	 * or are forced, do it; otherwise don't bother.
277 	 * If peer's buffer is tiny, then send
278 	 * when window is at least half open.
279 	 * If retransmitting (possibly after persist timer forced us
280 	 * to send into a small window), then must resend.
281 	 */
282 	if (len) {
283 		if (len == tp->t_maxseg)
284 			goto send;
285 		if (!(tp->t_flags & TF_MORETOCOME) &&
286 		    (idle || tp->t_flags & TF_NODELAY) &&
287 		    (tp->t_flags & TF_NOPUSH) == 0 &&
288 		    len + off >= so->so_snd.sb_cc)
289 			goto send;
290 		if (tp->t_force)
291 			goto send;
292 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
293 			goto send;
294 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
295 			goto send;
296 	}
297 
298 	/*
299 	 * Compare available window to amount of window
300 	 * known to peer (as advertised window less
301 	 * next expected input).  If the difference is at least two
302 	 * max size segments, or at least 50% of the maximum possible
303 	 * window, then want to send a window update to peer.
304 	 */
305 	if (win > 0) {
306 		/*
307 		 * "adv" is the amount we can increase the window,
308 		 * taking into account that we are limited by
309 		 * TCP_MAXWIN << tp->rcv_scale.
310 		 */
311 		long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) -
312 			(tp->rcv_adv - tp->rcv_nxt);
313 
314 		if (adv >= (long) (2 * tp->t_maxseg))
315 			goto send;
316 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
317 			goto send;
318 	}
319 
320 	/*
321 	 * Send if we owe peer an ACK.
322 	 */
323 	if (tp->t_flags & TF_ACKNOW)
324 		goto send;
325 	if ((flags & TH_RST) ||
326 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
327 		goto send;
328 	if (SEQ_GT(tp->snd_up, tp->snd_una))
329 		goto send;
330 	/*
331 	 * If our state indicates that FIN should be sent
332 	 * and we have not yet done so, or we're retransmitting the FIN,
333 	 * then we need to send.
334 	 */
335 	if (flags & TH_FIN &&
336 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
337 		goto send;
338 
339 	/*
340 	 * TCP window updates are not reliable, rather a polling protocol
341 	 * using ``persist'' packets is used to insure receipt of window
342 	 * updates.  The three ``states'' for the output side are:
343 	 *	idle			not doing retransmits or persists
344 	 *	persisting		to move a small or zero window
345 	 *	(re)transmitting	and thereby not persisting
346 	 *
347 	 * callout_active(tp->tt_persist)
348 	 *	is true when we are in persist state.
349 	 * tp->t_force
350 	 *	is set when we are called to send a persist packet.
351 	 * callout_active(tp->tt_rexmt)
352 	 *	is set when we are retransmitting
353 	 * The output side is idle when both timers are zero.
354 	 *
355 	 * If send window is too small, there is data to transmit, and no
356 	 * retransmit or persist is pending, then go to persist state.
357 	 * If nothing happens soon, send when timer expires:
358 	 * if window is nonzero, transmit what we can,
359 	 * otherwise force out a byte.
360 	 */
361 	if (so->so_snd.sb_cc && !callout_active(tp->tt_rexmt) &&
362 	    !callout_active(tp->tt_persist)) {
363 		tp->t_rxtshift = 0;
364 		tcp_setpersist(tp);
365 	}
366 
367 	/*
368 	 * No reason to send a segment, just return.
369 	 */
370 	return (0);
371 
372 send:
373 	/*
374 	 * Before ESTABLISHED, force sending of initial options
375 	 * unless TCP set not to do any options.
376 	 * NOTE: we assume that the IP/TCP header plus TCP options
377 	 * always fit in a single mbuf, leaving room for a maximum
378 	 * link header, i.e.
379 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
380 	 */
381 	optlen = 0;
382 #ifdef INET6
383 	if (isipv6)
384 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
385 	else
386 #endif
387 	hdrlen = sizeof (struct tcpiphdr);
388 	if (flags & TH_SYN) {
389 		tp->snd_nxt = tp->iss;
390 		if ((tp->t_flags & TF_NOOPT) == 0) {
391 			u_short mss;
392 
393 			opt[0] = TCPOPT_MAXSEG;
394 			opt[1] = TCPOLEN_MAXSEG;
395 			mss = htons((u_short) tcp_mssopt(tp));
396 			(void)memcpy(opt + 2, &mss, sizeof(mss));
397 			optlen = TCPOLEN_MAXSEG;
398 
399 			if ((tp->t_flags & TF_REQ_SCALE) &&
400 			    ((flags & TH_ACK) == 0 ||
401 			    (tp->t_flags & TF_RCVD_SCALE))) {
402 				*((u_int32_t *)(opt + optlen)) = htonl(
403 					TCPOPT_NOP << 24 |
404 					TCPOPT_WINDOW << 16 |
405 					TCPOLEN_WINDOW << 8 |
406 					tp->request_r_scale);
407 				optlen += 4;
408 			}
409 		}
410  	}
411 
412  	/*
413 	 * Send a timestamp and echo-reply if this is a SYN and our side
414 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
415 	 * and our peer have sent timestamps in our SYN's.
416  	 */
417  	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
418  	    (flags & TH_RST) == 0 &&
419 	    ((flags & TH_ACK) == 0 ||
420 	     (tp->t_flags & TF_RCVD_TSTMP))) {
421 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
422 
423  		/* Form timestamp option as shown in appendix A of RFC 1323. */
424  		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
425  		*lp++ = htonl(ticks);
426  		*lp   = htonl(tp->ts_recent);
427  		optlen += TCPOLEN_TSTAMP_APPA;
428  	}
429 
430  	/*
431 	 * Send `CC-family' options if our side wants to use them (TF_REQ_CC),
432 	 * options are allowed (!TF_NOOPT) and it's not a RST.
433  	 */
434  	if ((tp->t_flags & (TF_REQ_CC|TF_NOOPT)) == TF_REQ_CC &&
435  	     (flags & TH_RST) == 0) {
436 		switch (flags & (TH_SYN|TH_ACK)) {
437 		/*
438 		 * This is a normal ACK, send CC if we received CC before
439 		 * from our peer.
440 		 */
441 		case TH_ACK:
442 			if (!(tp->t_flags & TF_RCVD_CC))
443 				break;
444 			/*FALLTHROUGH*/
445 
446 		/*
447 		 * We can only get here in T/TCP's SYN_SENT* state, when
448 		 * we're a sending a non-SYN segment without waiting for
449 		 * the ACK of our SYN.  A check above assures that we only
450 		 * do this if our peer understands T/TCP.
451 		 */
452 		case 0:
453 			opt[optlen++] = TCPOPT_NOP;
454 			opt[optlen++] = TCPOPT_NOP;
455 			opt[optlen++] = TCPOPT_CC;
456 			opt[optlen++] = TCPOLEN_CC;
457 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
458 
459 			optlen += 4;
460 			break;
461 
462 		/*
463 		 * This is our initial SYN, check whether we have to use
464 		 * CC or CC.new.
465 		 */
466 		case TH_SYN:
467 			opt[optlen++] = TCPOPT_NOP;
468 			opt[optlen++] = TCPOPT_NOP;
469 			opt[optlen++] = tp->t_flags & TF_SENDCCNEW ?
470 						TCPOPT_CCNEW : TCPOPT_CC;
471 			opt[optlen++] = TCPOLEN_CC;
472 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
473  			optlen += 4;
474 			break;
475 
476 		/*
477 		 * This is a SYN,ACK; send CC and CC.echo if we received
478 		 * CC from our peer.
479 		 */
480 		case (TH_SYN|TH_ACK):
481 			if (tp->t_flags & TF_RCVD_CC) {
482 				opt[optlen++] = TCPOPT_NOP;
483 				opt[optlen++] = TCPOPT_NOP;
484 				opt[optlen++] = TCPOPT_CC;
485 				opt[optlen++] = TCPOLEN_CC;
486 				*(u_int32_t *)&opt[optlen] =
487 					htonl(tp->cc_send);
488 				optlen += 4;
489 				opt[optlen++] = TCPOPT_NOP;
490 				opt[optlen++] = TCPOPT_NOP;
491 				opt[optlen++] = TCPOPT_CCECHO;
492 				opt[optlen++] = TCPOLEN_CC;
493 				*(u_int32_t *)&opt[optlen] =
494 					htonl(tp->cc_recv);
495 				optlen += 4;
496 			}
497 			break;
498 		}
499  	}
500 
501  	hdrlen += optlen;
502 
503 #ifdef INET6
504 	if (isipv6)
505 		ipoptlen = ip6_optlen(tp->t_inpcb);
506 	else
507 #endif
508       {
509 	if (tp->t_inpcb->inp_options) {
510 		ipoptlen = tp->t_inpcb->inp_options->m_len -
511 				offsetof(struct ipoption, ipopt_list);
512 	} else {
513 		ipoptlen = 0;
514 	}
515       }
516 #ifdef IPSEC
517 	ipoptlen += ipsec_hdrsiz_tcp(tp);
518 #endif
519 
520 	/*
521 	 * Adjust data length if insertion of options will
522 	 * bump the packet length beyond the t_maxopd length.
523 	 * Clear the FIN bit because we cut off the tail of
524 	 * the segment.
525 	 */
526 	if (len + optlen + ipoptlen > tp->t_maxopd) {
527 		/*
528 		 * If there is still more to send, don't close the connection.
529 		 */
530 		flags &= ~TH_FIN;
531 		len = tp->t_maxopd - optlen - ipoptlen;
532 		sendalot = 1;
533 	}
534 
535 /*#ifdef DIAGNOSTIC*/
536 #ifdef INET6
537  	if (max_linkhdr + hdrlen > MCLBYTES)
538 		panic("tcphdr too big");
539 #else
540  	if (max_linkhdr + hdrlen > MHLEN)
541 		panic("tcphdr too big");
542 #endif
543 /*#endif*/
544 
545 	/*
546 	 * Grab a header mbuf, attaching a copy of data to
547 	 * be transmitted, and initialize the header from
548 	 * the template for sends on this connection.
549 	 */
550 	if (len) {
551 		if (tp->t_force && len == 1)
552 			tcpstat.tcps_sndprobe++;
553 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
554 			tcpstat.tcps_sndrexmitpack++;
555 			tcpstat.tcps_sndrexmitbyte += len;
556 		} else {
557 			tcpstat.tcps_sndpack++;
558 			tcpstat.tcps_sndbyte += len;
559 		}
560 #ifdef notyet
561 		if ((m = m_copypack(so->so_snd.sb_mb, off,
562 		    (int)len, max_linkhdr + hdrlen)) == 0) {
563 			error = ENOBUFS;
564 			goto out;
565 		}
566 		/*
567 		 * m_copypack left space for our hdr; use it.
568 		 */
569 		m->m_len += hdrlen;
570 		m->m_data -= hdrlen;
571 #else
572 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
573 		if (m == NULL) {
574 			error = ENOBUFS;
575 			goto out;
576 		}
577 #ifdef INET6
578 		if (MHLEN < hdrlen + max_linkhdr) {
579 			MCLGET(m, M_DONTWAIT);
580 			if ((m->m_flags & M_EXT) == 0) {
581 				m_freem(m);
582 				error = ENOBUFS;
583 				goto out;
584 			}
585 		}
586 #endif
587 		m->m_data += max_linkhdr;
588 		m->m_len = hdrlen;
589 		if (len <= MHLEN - hdrlen - max_linkhdr) {
590 			m_copydata(so->so_snd.sb_mb, off, (int) len,
591 			    mtod(m, caddr_t) + hdrlen);
592 			m->m_len += len;
593 		} else {
594 			m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len);
595 			if (m->m_next == 0) {
596 				(void) m_free(m);
597 				error = ENOBUFS;
598 				goto out;
599 			}
600 		}
601 #endif
602 		/*
603 		 * If we're sending everything we've got, set PUSH.
604 		 * (This will keep happy those implementations which only
605 		 * give data to the user when a buffer fills or
606 		 * a PUSH comes in.)
607 		 */
608 		if (off + len == so->so_snd.sb_cc)
609 			flags |= TH_PUSH;
610 	} else {
611 		if (tp->t_flags & TF_ACKNOW)
612 			tcpstat.tcps_sndacks++;
613 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
614 			tcpstat.tcps_sndctrl++;
615 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
616 			tcpstat.tcps_sndurg++;
617 		else
618 			tcpstat.tcps_sndwinup++;
619 
620 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
621 		if (m == NULL) {
622 			error = ENOBUFS;
623 			goto out;
624 		}
625 #ifdef INET6
626 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
627 		    MHLEN >= hdrlen) {
628 			MH_ALIGN(m, hdrlen);
629 		} else
630 #endif
631 		m->m_data += max_linkhdr;
632 		m->m_len = hdrlen;
633 	}
634 	m->m_pkthdr.rcvif = (struct ifnet *)0;
635 #ifdef INET6
636 	if (isipv6) {
637 		ip6 = mtod(m, struct ip6_hdr *);
638 		th = (struct tcphdr *)(ip6 + 1);
639 		tcp_fillheaders(tp, ip6, th);
640 	} else
641 #endif /* INET6 */
642       {
643 	ip = mtod(m, struct ip *);
644 	ipov = (struct ipovly *)ip;
645 	th = (struct tcphdr *)(ip + 1);
646 	/* this picks up the pseudo header (w/o the length) */
647 	tcp_fillheaders(tp, ip, th);
648       }
649 
650 	/*
651 	 * Fill in fields, remembering maximum advertised
652 	 * window for use in delaying messages about window sizes.
653 	 * If resending a FIN, be sure not to use a new sequence number.
654 	 */
655 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
656 	    tp->snd_nxt == tp->snd_max)
657 		tp->snd_nxt--;
658 	/*
659 	 * If we are doing retransmissions, then snd_nxt will
660 	 * not reflect the first unsent octet.  For ACK only
661 	 * packets, we do not want the sequence number of the
662 	 * retransmitted packet, we want the sequence number
663 	 * of the next unsent octet.  So, if there is no data
664 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
665 	 * when filling in ti_seq.  But if we are in persist
666 	 * state, snd_max might reflect one byte beyond the
667 	 * right edge of the window, so use snd_nxt in that
668 	 * case, since we know we aren't doing a retransmission.
669 	 * (retransmit and persist are mutually exclusive...)
670 	 */
671 	if (len || (flags & (TH_SYN|TH_FIN))
672 	    || callout_active(tp->tt_persist))
673 		th->th_seq = htonl(tp->snd_nxt);
674 	else
675 		th->th_seq = htonl(tp->snd_max);
676 	th->th_ack = htonl(tp->rcv_nxt);
677 	if (optlen) {
678 		bcopy(opt, th + 1, optlen);
679 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
680 	}
681 	th->th_flags = flags;
682 	/*
683 	 * Calculate receive window.  Don't shrink window,
684 	 * but avoid silly window syndrome.
685 	 */
686 	if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)tp->t_maxseg)
687 		win = 0;
688 	if (win < (long)(tp->rcv_adv - tp->rcv_nxt))
689 		win = (long)(tp->rcv_adv - tp->rcv_nxt);
690 	if (win > (long)TCP_MAXWIN << tp->rcv_scale)
691 		win = (long)TCP_MAXWIN << tp->rcv_scale;
692 	th->th_win = htons((u_short) (win>>tp->rcv_scale));
693 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
694 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
695 		th->th_flags |= TH_URG;
696 	} else
697 		/*
698 		 * If no urgent pointer to send, then we pull
699 		 * the urgent pointer to the left edge of the send window
700 		 * so that it doesn't drift into the send window on sequence
701 		 * number wraparound.
702 		 */
703 		tp->snd_up = tp->snd_una;		/* drag it along */
704 
705 	/*
706 	 * Put TCP length in extended header, and then
707 	 * checksum extended header and data.
708 	 */
709 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
710 #ifdef INET6
711 	if (isipv6)
712 		/*
713 		 * ip6_plen is not need to be filled now, and will be filled
714 		 * in ip6_output.
715 		 */
716 		th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
717 				       sizeof(struct tcphdr) + optlen + len);
718 	else
719 #endif /* INET6 */
720       {
721 	m->m_pkthdr.csum_flags = CSUM_TCP;
722 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
723 	if (len + optlen)
724 		th->th_sum = in_addword(th->th_sum,
725 		    htons((u_short)(optlen + len)));
726 
727 	/* IP version must be set here for ipv4/ipv6 checking later */
728 	KASSERT(ip->ip_v == IPVERSION,
729 	    ("%s: IP version incorrect: %d", __FUNCTION__, ip->ip_v));
730       }
731 
732 	/*
733 	 * In transmit state, time the transmission and arrange for
734 	 * the retransmit.  In persist state, just set snd_max.
735 	 */
736 	if (tp->t_force == 0 || !callout_active(tp->tt_persist)) {
737 		tcp_seq startseq = tp->snd_nxt;
738 
739 		/*
740 		 * Advance snd_nxt over sequence space of this segment.
741 		 */
742 		if (flags & (TH_SYN|TH_FIN)) {
743 			if (flags & TH_SYN)
744 				tp->snd_nxt++;
745 			if (flags & TH_FIN) {
746 				tp->snd_nxt++;
747 				tp->t_flags |= TF_SENTFIN;
748 			}
749 		}
750 		tp->snd_nxt += len;
751 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
752 			tp->snd_max = tp->snd_nxt;
753 			/*
754 			 * Time this transmission if not a retransmission and
755 			 * not currently timing anything.
756 			 */
757 			if (tp->t_rtttime == 0) {
758 				tp->t_rtttime = ticks;
759 				tp->t_rtseq = startseq;
760 				tcpstat.tcps_segstimed++;
761 			}
762 		}
763 
764 		/*
765 		 * Set retransmit timer if not currently set,
766 		 * and not doing an ack or a keep-alive probe.
767 		 * Initial value for retransmit timer is smoothed
768 		 * round-trip time + 2 * round-trip time variance.
769 		 * Initialize shift counter which is used for backoff
770 		 * of retransmit time.
771 		 */
772 		if (!callout_active(tp->tt_rexmt) &&
773 		    tp->snd_nxt != tp->snd_una) {
774 			if (callout_active(tp->tt_persist)) {
775 				callout_stop(tp->tt_persist);
776 				tp->t_rxtshift = 0;
777 			}
778 			callout_reset(tp->tt_rexmt, tp->t_rxtcur,
779 				      tcp_timer_rexmt, tp);
780 		}
781 	} else
782 		if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
783 			tp->snd_max = tp->snd_nxt + len;
784 
785 #ifdef TCPDEBUG
786 	/*
787 	 * Trace.
788 	 */
789 	if (so->so_options & SO_DEBUG)
790 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
791 #endif
792 
793 	/*
794 	 * Fill in IP length and desired time to live and
795 	 * send to IP level.  There should be a better way
796 	 * to handle ttl and tos; we could keep them in
797 	 * the template, but need a way to checksum without them.
798 	 */
799 	/*
800 	 * m->m_pkthdr.len should have been set before cksum calcuration,
801 	 * because in6_cksum() need it.
802 	 */
803 #ifdef INET6
804 	if (isipv6) {
805 		/*
806 		 * we separately set hoplimit for every segment, since the
807 		 * user might want to change the value via setsockopt.
808 		 * Also, desired default hop limit might be changed via
809 		 * Neighbor Discovery.
810 		 */
811 		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb,
812 					       tp->t_inpcb->in6p_route.ro_rt ?
813 					       tp->t_inpcb->in6p_route.ro_rt->rt_ifp
814 					       : NULL);
815 
816 		/* TODO: IPv6 IP6TOS_ECT bit on */
817 #ifdef IPSEC
818 		if (ipsec_setsocket(m, so) != 0) {
819 			m_freem(m);
820 			error = ENOBUFS;
821 			goto out;
822 		}
823 #endif /*IPSEC*/
824 		error = ip6_output(m,
825 			    tp->t_inpcb->in6p_outputopts,
826 			    &tp->t_inpcb->in6p_route,
827 			    (so->so_options & SO_DONTROUTE), NULL, NULL);
828 	} else
829 #endif /* INET6 */
830     {
831 	struct rtentry *rt;
832 	ip->ip_len = m->m_pkthdr.len;
833 #ifdef INET6
834  	if (INP_CHECK_SOCKAF(so, AF_INET6))
835  		ip->ip_ttl = in6_selecthlim(tp->t_inpcb,
836  					    tp->t_inpcb->in6p_route.ro_rt ?
837  					    tp->t_inpcb->in6p_route.ro_rt->rt_ifp
838  					    : NULL);
839  	else
840 #endif /* INET6 */
841 	ip->ip_ttl = tp->t_inpcb->inp_ip_ttl;	/* XXX */
842 	ip->ip_tos = tp->t_inpcb->inp_ip_tos;	/* XXX */
843 	/*
844 	 * See if we should do MTU discovery.  We do it only if the following
845 	 * are true:
846 	 *	1) we have a valid route to the destination
847 	 *	2) the MTU is not locked (if it is, then discovery has been
848 	 *	   disabled)
849 	 */
850 	if (path_mtu_discovery
851 	    && (rt = tp->t_inpcb->inp_route.ro_rt)
852 	    && rt->rt_flags & RTF_UP
853 	    && !(rt->rt_rmx.rmx_locks & RTV_MTU)) {
854 		ip->ip_off |= IP_DF;
855 	}
856 #ifdef IPSEC
857  	ipsec_setsocket(m, so);
858 #endif /*IPSEC*/
859 	error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
860 	    (so->so_options & SO_DONTROUTE), 0);
861     }
862 	if (error) {
863 
864 		/*
865 		 * We know that the packet was lost, so back out the
866 		 * sequence number advance, if any.
867 		 */
868 		if (tp->t_force == 0 || !callout_active(tp->tt_persist)) {
869 			/*
870 			 * No need to check for TH_FIN here because
871 			 * the TF_SENTFIN flag handles that case.
872 			 */
873 			if ((flags & TH_SYN) == 0)
874 				tp->snd_nxt -= len;
875 		}
876 
877 out:
878 		if (error == ENOBUFS) {
879 	                if (!callout_active(tp->tt_rexmt) &&
880                             !callout_active(tp->tt_persist))
881 	                        callout_reset(tp->tt_rexmt, tp->t_rxtcur,
882                                       tcp_timer_rexmt, tp);
883 			tcp_quench(tp->t_inpcb, 0);
884 			return (0);
885 		}
886 		if (error == EMSGSIZE) {
887 			/*
888 			 * ip_output() will have already fixed the route
889 			 * for us.  tcp_mtudisc() will, as its last action,
890 			 * initiate retransmission, so it is important to
891 			 * not do so here.
892 			 */
893 			tcp_mtudisc(tp->t_inpcb, 0);
894 			return 0;
895 		}
896 		if ((error == EHOSTUNREACH || error == ENETDOWN)
897 		    && TCPS_HAVERCVDSYN(tp->t_state)) {
898 			tp->t_softerror = error;
899 			return (0);
900 		}
901 		return (error);
902 	}
903 	tcpstat.tcps_sndtotal++;
904 
905 	/*
906 	 * Data sent (as far as we can tell).
907 	 * If this advertises a larger window than any other segment,
908 	 * then remember the size of the advertised window.
909 	 * Any pending ACK has now been sent.
910 	 */
911 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
912 		tp->rcv_adv = tp->rcv_nxt + win;
913 	tp->last_ack_sent = tp->rcv_nxt;
914 	tp->t_flags &= ~TF_ACKNOW;
915 	if (tcp_delack_enabled)
916 		callout_stop(tp->tt_delack);
917 	if (sendalot && (!tcp_do_newreno || --maxburst))
918 		goto again;
919 	return (0);
920 }
921 
922 void
923 tcp_setpersist(tp)
924 	register struct tcpcb *tp;
925 {
926 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
927 	int tt;
928 
929 	if (callout_active(tp->tt_rexmt))
930 		panic("tcp_setpersist: retransmit pending");
931 	/*
932 	 * Start/restart persistance timer.
933 	 */
934 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
935 		      TCPTV_PERSMIN, TCPTV_PERSMAX);
936 	callout_reset(tp->tt_persist, tt, tcp_timer_persist, tp);
937 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
938 		tp->t_rxtshift++;
939 }
940