xref: /netbsd-src/sys/netinet/tcp_output.c (revision 1ca5c1b28139779176bd5c13ad7c5f25c0bcd5f8)
1 /*	$NetBSD: tcp_output.c,v 1.76 2001/12/03 01:45:43 jmcneill Exp $	*/
2 
3 /*
4 %%% portions-copyright-nrl-95
5 Portions of this software are Copyright 1995-1998 by Randall Atkinson,
6 Ronald Lee, Daniel McDonald, Bao Phan, and Chris Winters. All Rights
7 Reserved. All rights under this copyright have been assigned to the US
8 Naval Research Laboratory (NRL). The NRL Copyright Notice and License
9 Agreement Version 1.1 (January 17, 1995) applies to these portions of the
10 software.
11 You should have received a copy of the license with this software. If you
12 didn't get a copy, you may request one from <license@ipv6.nrl.navy.mil>.
13 
14 */
15 
16 /*
17  * Copyright (C) 1995, 1996, 1997, and 1998 WIDE Project.
18  * All rights reserved.
19  *
20  * Redistribution and use in source and binary forms, with or without
21  * modification, are permitted provided that the following conditions
22  * are met:
23  * 1. Redistributions of source code must retain the above copyright
24  *    notice, this list of conditions and the following disclaimer.
25  * 2. Redistributions in binary form must reproduce the above copyright
26  *    notice, this list of conditions and the following disclaimer in the
27  *    documentation and/or other materials provided with the distribution.
28  * 3. Neither the name of the project nor the names of its contributors
29  *    may be used to endorse or promote products derived from this software
30  *    without specific prior written permission.
31  *
32  * THIS SOFTWARE IS PROVIDED BY THE PROJECT AND CONTRIBUTORS ``AS IS'' AND
33  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
34  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
35  * ARE DISCLAIMED.  IN NO EVENT SHALL THE PROJECT OR CONTRIBUTORS BE LIABLE
36  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
37  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
38  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
39  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
40  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
41  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
42  * SUCH DAMAGE.
43  */
44 
45 /*-
46  * Copyright (c) 1997, 1998, 2001 The NetBSD Foundation, Inc.
47  * All rights reserved.
48  *
49  * This code is derived from software contributed to The NetBSD Foundation
50  * by Jason R. Thorpe and Kevin M. Lahey of the Numerical Aerospace Simulation
51  * Facility, NASA Ames Research Center.
52  *
53  * Redistribution and use in source and binary forms, with or without
54  * modification, are permitted provided that the following conditions
55  * are met:
56  * 1. Redistributions of source code must retain the above copyright
57  *    notice, this list of conditions and the following disclaimer.
58  * 2. Redistributions in binary form must reproduce the above copyright
59  *    notice, this list of conditions and the following disclaimer in the
60  *    documentation and/or other materials provided with the distribution.
61  * 3. All advertising materials mentioning features or use of this software
62  *    must display the following acknowledgement:
63  *	This product includes software developed by the NetBSD
64  *	Foundation, Inc. and its contributors.
65  * 4. Neither the name of The NetBSD Foundation nor the names of its
66  *    contributors may be used to endorse or promote products derived
67  *    from this software without specific prior written permission.
68  *
69  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
70  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
71  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
72  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
73  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
74  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
75  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
76  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
77  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
78  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
79  * POSSIBILITY OF SUCH DAMAGE.
80  */
81 
82 /*
83  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
84  *	The Regents of the University of California.  All rights reserved.
85  *
86  * Redistribution and use in source and binary forms, with or without
87  * modification, are permitted provided that the following conditions
88  * are met:
89  * 1. Redistributions of source code must retain the above copyright
90  *    notice, this list of conditions and the following disclaimer.
91  * 2. Redistributions in binary form must reproduce the above copyright
92  *    notice, this list of conditions and the following disclaimer in the
93  *    documentation and/or other materials provided with the distribution.
94  * 3. All advertising materials mentioning features or use of this software
95  *    must display the following acknowledgement:
96  *	This product includes software developed by the University of
97  *	California, Berkeley and its contributors.
98  * 4. Neither the name of the University nor the names of its contributors
99  *    may be used to endorse or promote products derived from this software
100  *    without specific prior written permission.
101  *
102  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
103  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
104  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
105  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
106  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
107  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
108  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
109  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
110  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
111  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
112  * SUCH DAMAGE.
113  *
114  *	@(#)tcp_output.c	8.4 (Berkeley) 5/24/95
115  */
116 
117 #include <sys/cdefs.h>
118 __KERNEL_RCSID(0, "$NetBSD: tcp_output.c,v 1.76 2001/12/03 01:45:43 jmcneill Exp $");
119 
120 #include "opt_inet.h"
121 #include "opt_ipsec.h"
122 #include "opt_tcp_debug.h"
123 
124 #include <sys/param.h>
125 #include <sys/systm.h>
126 #include <sys/malloc.h>
127 #include <sys/mbuf.h>
128 #include <sys/protosw.h>
129 #include <sys/socket.h>
130 #include <sys/socketvar.h>
131 #include <sys/errno.h>
132 #include <sys/domain.h>
133 #include <sys/kernel.h>
134 
135 #include <net/if.h>
136 #include <net/route.h>
137 
138 #include <netinet/in.h>
139 #include <netinet/in_systm.h>
140 #include <netinet/ip.h>
141 #include <netinet/in_pcb.h>
142 #include <netinet/ip_var.h>
143 
144 #ifdef INET6
145 #ifndef INET
146 #include <netinet/in.h>
147 #endif
148 #include <netinet/ip6.h>
149 #include <netinet6/in6_pcb.h>
150 #include <netinet6/ip6_var.h>
151 #endif
152 
153 #include <netinet/tcp.h>
154 #define	TCPOUTFLAGS
155 #include <netinet/tcp_fsm.h>
156 #include <netinet/tcp_seq.h>
157 #include <netinet/tcp_timer.h>
158 #include <netinet/tcp_var.h>
159 #include <netinet/tcpip.h>
160 #include <netinet/tcp_debug.h>
161 
162 #ifdef notyet
163 extern struct mbuf *m_copypack();
164 #endif
165 
166 #define MAX_TCPOPTLEN	32	/* max # bytes that go in options */
167 
168 /*
169  * Knob to enable Congestion Window Monitoring, and control the
170  * the burst size it allows.  Default burst is 4 packets, per
171  * the Internet draft.
172  */
173 int	tcp_cwm = 1;
174 int	tcp_cwm_burstsize = 4;
175 
176 static
177 #ifndef GPROF
178 __inline
179 #endif
180 void
181 tcp_segsize(struct tcpcb *tp, int *txsegsizep, int *rxsegsizep)
182 {
183 #ifdef INET
184 	struct inpcb *inp = tp->t_inpcb;
185 #endif
186 #ifdef INET6
187 	struct in6pcb *in6p = tp->t_in6pcb;
188 #endif
189 	struct rtentry *rt;
190 	struct ifnet *ifp;
191 	int size;
192 	int iphlen;
193 	int optlen;
194 
195 #ifdef DIAGNOSTIC
196 	if (tp->t_inpcb && tp->t_in6pcb)
197 		panic("tcp_segsize: both t_inpcb and t_in6pcb are set");
198 #endif
199 	switch (tp->t_family) {
200 #ifdef INET
201 	case AF_INET:
202 		iphlen = sizeof(struct ip);
203 		break;
204 #endif
205 #ifdef INET6
206 	case AF_INET6:
207 		iphlen = sizeof(struct ip6_hdr);
208 		break;
209 #endif
210 	default:
211 		size = tcp_mssdflt;
212 		goto out;
213 	}
214 
215 	rt = NULL;
216 #ifdef INET
217 	if (inp)
218 		rt = in_pcbrtentry(inp);
219 #endif
220 #ifdef INET6
221 	if (in6p)
222 		rt = in6_pcbrtentry(in6p);
223 #endif
224 	if (rt == NULL) {
225 		size = tcp_mssdflt;
226 		goto out;
227 	}
228 
229 	ifp = rt->rt_ifp;
230 
231 	size = tcp_mssdflt;
232 	if (rt->rt_rmx.rmx_mtu != 0)
233 		size = rt->rt_rmx.rmx_mtu - iphlen - sizeof(struct tcphdr);
234 	else if (ifp->if_flags & IFF_LOOPBACK)
235 		size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
236 #ifdef INET
237 	else if (inp && ip_mtudisc)
238 		size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
239 	else if (inp && in_localaddr(inp->inp_faddr))
240 		size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
241 #endif
242 #ifdef INET6
243 	else if (in6p) {
244 #ifdef INET
245 		if (IN6_IS_ADDR_V4MAPPED(&in6p->in6p_faddr)) {
246 			/* mapped addr case */
247 			struct in_addr d;
248 			bcopy(&in6p->in6p_faddr.s6_addr32[3], &d, sizeof(d));
249 			if (ip_mtudisc || in_localaddr(d))
250 				size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
251 		} else
252 #endif
253 		{
254 			/*
255 			 * for IPv6, path MTU discovery is always turned on,
256 			 * or the node must use packet size <= 1280.
257 			 */
258 			size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
259 		}
260 	}
261 #endif
262 	/*
263 	 * Now we must make room for whatever extra TCP/IP options are in
264 	 * the packet.
265 	 */
266 	optlen = tcp_optlen(tp);
267 
268 	/*
269 	 * XXX tp->t_ourmss should have the right size, but without this code
270 	 * fragmentation will occur... need more investigation
271 	 */
272 #ifdef INET
273 	if (inp) {
274 #ifdef IPSEC
275 		optlen += ipsec4_hdrsiz_tcp(tp);
276 #endif
277 		optlen += ip_optlen(inp);
278 	}
279 #endif
280 #ifdef INET6
281 #ifdef INET
282 	if (in6p && tp->t_family == AF_INET) {
283 #ifdef IPSEC
284 		optlen += ipsec4_hdrsiz_tcp(tp);
285 #endif
286 		/* XXX size -= ip_optlen(in6p); */
287 	} else
288 #endif
289 	if (in6p && tp->t_family == AF_INET6) {
290 #ifdef IPSEC
291 		optlen += ipsec6_hdrsiz_tcp(tp);
292 #endif
293 		optlen += ip6_optlen(in6p);
294 	}
295 #endif
296 	size -= optlen;
297 
298  out:
299 	/*
300 	 * *rxsegsizep holds *estimated* inbound segment size (estimation
301 	 * assumes that path MTU is the same for both ways).  this is only
302 	 * for silly window avoidance, do not use the value for other purposes.
303 	 *
304 	 * ipseclen is subtracted from both sides, this may not be right.
305 	 * I'm not quite sure about this (could someone comment).
306 	 */
307 	*txsegsizep = min(tp->t_peermss - optlen, size);
308 	*rxsegsizep = min(tp->t_ourmss - optlen, size);
309 
310 	if (*txsegsizep != tp->t_segsz) {
311 		/*
312 		 * If the new segment size is larger, we don't want to
313 		 * mess up the congestion window, but if it is smaller
314 		 * we'll have to reduce the congestion window to ensure
315 		 * that we don't get into trouble with initial windows
316 		 * and the rest.  In any case, if the segment size
317 		 * has changed, chances are the path has, too, and
318 		 * our congestion window will be different.
319 		 */
320 		if (*txsegsizep < tp->t_segsz) {
321 			tp->snd_cwnd = max((tp->snd_cwnd / tp->t_segsz)
322 					   * *txsegsizep, *txsegsizep);
323 			tp->snd_ssthresh = max((tp->snd_ssthresh / tp->t_segsz)
324 						* *txsegsizep, *txsegsizep);
325 		}
326 		tp->t_segsz = *txsegsizep;
327 	}
328 }
329 
330 static
331 #ifndef GPROF
332 __inline
333 #endif
334 int
335 tcp_build_datapkt(struct tcpcb *tp, struct socket *so, int off,
336     long len, int hdrlen, struct mbuf **mp)
337 {
338 	struct mbuf *m;
339 
340 	if (tp->t_force && len == 1)
341 		tcpstat.tcps_sndprobe++;
342 	else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
343 		tcpstat.tcps_sndrexmitpack++;
344 		tcpstat.tcps_sndrexmitbyte += len;
345 	} else {
346 		tcpstat.tcps_sndpack++;
347 		tcpstat.tcps_sndbyte += len;
348 	}
349 #ifdef notyet
350 	if ((m = m_copypack(so->so_snd.sb_mb, off,
351 	    (int)len, max_linkhdr + hdrlen)) == 0)
352 		return (ENOBUFS);
353 	/*
354 	 * m_copypack left space for our hdr; use it.
355 	 */
356 	m->m_len += hdrlen;
357 	m->m_data -= hdrlen;
358 #else
359 	MGETHDR(m, M_DONTWAIT, MT_HEADER);
360 	if (m != NULL &&
361 	    (max_linkhdr + hdrlen > MHLEN ||
362 	     max_linkhdr + hdrlen + len <= MCLBYTES)) {
363 		MCLGET(m, M_DONTWAIT);
364 		if ((m->m_flags & M_EXT) == 0) {
365 			m_freem(m);
366 			m = NULL;
367 		}
368 	}
369 	if (m == NULL)
370 		return (ENOBUFS);
371 	m->m_data += max_linkhdr;
372 	m->m_len = hdrlen;
373 	if (len <= M_TRAILINGSPACE(m)) {
374 		m_copydata(so->so_snd.sb_mb, off, (int) len,
375 		    mtod(m, caddr_t) + hdrlen);
376 		m->m_len += len;
377 	} else {
378 		m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len);
379 		if (m->m_next == NULL) {
380 			m_freem(m);
381 			return (ENOBUFS);
382 		}
383 	}
384 #endif
385 
386 	*mp = m;
387 	return (0);
388 }
389 
390 /*
391  * Tcp output routine: figure out what should be sent and send it.
392  */
393 int
394 tcp_output(tp)
395 	struct tcpcb *tp;
396 {
397 	struct socket *so;
398 	struct route *ro;
399 	long len, win;
400 	int off, flags, error;
401 	struct mbuf *m;
402 	struct ip *ip;
403 #ifdef INET6
404 	struct ip6_hdr *ip6;
405 #endif
406 	struct tcphdr *th;
407 	u_char opt[MAX_TCPOPTLEN];
408 	unsigned optlen, hdrlen;
409 	int idle, sendalot, txsegsize, rxsegsize;
410 	int maxburst = TCP_MAXBURST;
411 	int af;		/* address family on the wire */
412 	int iphdrlen;
413 
414 #ifdef DIAGNOSTIC
415 	if (tp->t_inpcb && tp->t_in6pcb)
416 		panic("tcp_output: both t_inpcb and t_in6pcb are set");
417 #endif
418 	so = NULL;
419 	ro = NULL;
420 	if (tp->t_inpcb) {
421 		so = tp->t_inpcb->inp_socket;
422 		ro = &tp->t_inpcb->inp_route;
423 	}
424 #ifdef INET6
425 	else if (tp->t_in6pcb) {
426 		so = tp->t_in6pcb->in6p_socket;
427 		ro = (struct route *)&tp->t_in6pcb->in6p_route;
428 	}
429 #endif
430 
431 	switch (af = tp->t_family) {
432 #ifdef INET
433 	case AF_INET:
434 		if (tp->t_inpcb)
435 			break;
436 #ifdef INET6
437 		/* mapped addr case */
438 		if (tp->t_in6pcb)
439 			break;
440 #endif
441 		return EINVAL;
442 #endif
443 #ifdef INET6
444 	case AF_INET6:
445 		if (tp->t_in6pcb)
446 			break;
447 		return EINVAL;
448 #endif
449 	default:
450 		return EAFNOSUPPORT;
451 	}
452 
453 	tcp_segsize(tp, &txsegsize, &rxsegsize);
454 
455 	idle = (tp->snd_max == tp->snd_una);
456 
457 	/*
458 	 * Restart Window computation.  From draft-floyd-incr-init-win-03:
459 	 *
460 	 *	Optionally, a TCP MAY set the restart window to the
461 	 *	minimum of the value used for the initial window and
462 	 *	the current value of cwnd (in other words, using a
463 	 *	larger value for the restart window should never increase
464 	 *	the size of cwnd).
465 	 */
466 	if (tcp_cwm) {
467 		/*
468 		 * Hughes/Touch/Heidemann Congestion Window Monitoring.
469 		 * Count the number of packets currently pending
470 		 * acknowledgement, and limit our congestion window
471 		 * to a pre-determined allowed burst size plus that count.
472 		 * This prevents bursting once all pending packets have
473 		 * been acknowledged (i.e. transmission is idle).
474 		 *
475 		 * XXX Link this to Initial Window?
476 		 */
477 		tp->snd_cwnd = min(tp->snd_cwnd,
478 		    (tcp_cwm_burstsize * txsegsize) +
479 		    (tp->snd_nxt - tp->snd_una));
480 	} else {
481 		if (idle && (tcp_now - tp->t_rcvtime) >= tp->t_rxtcur) {
482 			/*
483 			 * We have been idle for "a while" and no acks are
484 			 * expected to clock out any data we send --
485 			 * slow start to get ack "clock" running again.
486 			 */
487 			tp->snd_cwnd = min(tp->snd_cwnd,
488 			    TCP_INITIAL_WINDOW(tcp_init_win, txsegsize));
489 		}
490 	}
491 
492 again:
493 	/*
494 	 * Determine length of data that should be transmitted, and
495 	 * flags that should be used.  If there is some data or critical
496 	 * controls (SYN, RST) to send, then transmit; otherwise,
497 	 * investigate further.
498 	 */
499 	sendalot = 0;
500 	off = tp->snd_nxt - tp->snd_una;
501 	win = min(tp->snd_wnd, tp->snd_cwnd);
502 
503 	flags = tcp_outflags[tp->t_state];
504 	/*
505 	 * If in persist timeout with window of 0, send 1 byte.
506 	 * Otherwise, if window is small but nonzero
507 	 * and timer expired, we will send what we can
508 	 * and go to transmit state.
509 	 */
510 	if (tp->t_force) {
511 		if (win == 0) {
512 			/*
513 			 * If we still have some data to send, then
514 			 * clear the FIN bit.  Usually this would
515 			 * happen below when it realizes that we
516 			 * aren't sending all the data.  However,
517 			 * if we have exactly 1 byte of unset data,
518 			 * then it won't clear the FIN bit below,
519 			 * and if we are in persist state, we wind
520 			 * up sending the packet without recording
521 			 * that we sent the FIN bit.
522 			 *
523 			 * We can't just blindly clear the FIN bit,
524 			 * because if we don't have any more data
525 			 * to send then the probe will be the FIN
526 			 * itself.
527 			 */
528 			if (off < so->so_snd.sb_cc)
529 				flags &= ~TH_FIN;
530 			win = 1;
531 		} else {
532 			TCP_TIMER_DISARM(tp, TCPT_PERSIST);
533 			tp->t_rxtshift = 0;
534 		}
535 	}
536 
537 	if (win < so->so_snd.sb_cc) {
538 		len = win - off;
539 		flags &= ~TH_FIN;
540 	} else
541 		len = so->so_snd.sb_cc - off;
542 
543 	if (len < 0) {
544 		/*
545 		 * If FIN has been sent but not acked,
546 		 * but we haven't been called to retransmit,
547 		 * len will be -1.  Otherwise, window shrank
548 		 * after we sent into it.  If window shrank to 0,
549 		 * cancel pending retransmit, pull snd_nxt back
550 		 * to (closed) window, and set the persist timer
551 		 * if it isn't already going.  If the window didn't
552 		 * close completely, just wait for an ACK.
553 		 *
554 		 * If we have a pending FIN, either it has already been
555 		 * transmitted or it is outside the window, so drop it.
556 		 * If the FIN has been transmitted, but this is not a
557 		 * retransmission, then len must be -1.  Therefore we also
558 		 * prevent here the sending of `gratuitous FINs'.  This
559 		 * eliminates the need to check for that case below (e.g.
560 		 * to back up snd_nxt before the FIN so that the sequence
561 		 * number is correct).
562 		 */
563 		len = 0;
564 		flags &= ~TH_FIN;
565 		if (win == 0) {
566 			TCP_TIMER_DISARM(tp, TCPT_REXMT);
567 			tp->t_rxtshift = 0;
568 			tp->snd_nxt = tp->snd_una;
569 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0)
570 				tcp_setpersist(tp);
571 		}
572 	}
573 	if (len > txsegsize) {
574 		len = txsegsize;
575 		flags &= ~TH_FIN;
576 		sendalot = 1;
577 	}
578 
579 	win = sbspace(&so->so_rcv);
580 
581 	/*
582 	 * Sender silly window avoidance.  If connection is idle
583 	 * and can send all data, a maximum segment,
584 	 * at least a maximum default-size segment do it,
585 	 * or are forced, do it; otherwise don't bother.
586 	 * If peer's buffer is tiny, then send
587 	 * when window is at least half open.
588 	 * If retransmitting (possibly after persist timer forced us
589 	 * to send into a small window), then must resend.
590 	 */
591 	if (len) {
592 		if (len == txsegsize)
593 			goto send;
594 		if ((so->so_state & SS_MORETOCOME) == 0 &&
595 		    ((idle || tp->t_flags & TF_NODELAY) &&
596 		     len + off >= so->so_snd.sb_cc))
597 			goto send;
598 		if (tp->t_force)
599 			goto send;
600 		if (len >= tp->max_sndwnd / 2)
601 			goto send;
602 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
603 			goto send;
604 	}
605 
606 	/*
607 	 * Compare available window to amount of window known to peer
608 	 * (as advertised window less next expected input).  If the
609 	 * difference is at least twice the size of the largest segment
610 	 * we expect to receive (i.e. two segments) or at least 50% of
611 	 * the maximum possible window, then want to send a window update
612 	 * to peer.
613 	 */
614 	if (win > 0) {
615 		/*
616 		 * "adv" is the amount we can increase the window,
617 		 * taking into account that we are limited by
618 		 * TCP_MAXWIN << tp->rcv_scale.
619 		 */
620 		long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) -
621 			(tp->rcv_adv - tp->rcv_nxt);
622 
623 		if (adv >= (long) (2 * rxsegsize))
624 			goto send;
625 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
626 			goto send;
627 	}
628 
629 	/*
630 	 * Send if we owe peer an ACK.
631 	 */
632 	if (tp->t_flags & TF_ACKNOW)
633 		goto send;
634 	if (flags & (TH_SYN|TH_FIN|TH_RST))
635 		goto send;
636 	if (SEQ_GT(tp->snd_up, tp->snd_una))
637 		goto send;
638 
639 	/*
640 	 * TCP window updates are not reliable, rather a polling protocol
641 	 * using ``persist'' packets is used to insure receipt of window
642 	 * updates.  The three ``states'' for the output side are:
643 	 *	idle			not doing retransmits or persists
644 	 *	persisting		to move a small or zero window
645 	 *	(re)transmitting	and thereby not persisting
646 	 *
647 	 * tp->t_timer[TCPT_PERSIST]
648 	 *	is set when we are in persist state.
649 	 * tp->t_force
650 	 *	is set when we are called to send a persist packet.
651 	 * tp->t_timer[TCPT_REXMT]
652 	 *	is set when we are retransmitting
653 	 * The output side is idle when both timers are zero.
654 	 *
655 	 * If send window is too small, there is data to transmit, and no
656 	 * retransmit or persist is pending, then go to persist state.
657 	 * If nothing happens soon, send when timer expires:
658 	 * if window is nonzero, transmit what we can,
659 	 * otherwise force out a byte.
660 	 */
661 	if (so->so_snd.sb_cc && TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
662 	    TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
663 		tp->t_rxtshift = 0;
664 		tcp_setpersist(tp);
665 	}
666 
667 	/*
668 	 * No reason to send a segment, just return.
669 	 */
670 	return (0);
671 
672 send:
673 	/*
674 	 * Before ESTABLISHED, force sending of initial options
675 	 * unless TCP set not to do any options.
676 	 * NOTE: we assume that the IP/TCP header plus TCP options
677 	 * always fit in a single mbuf, leaving room for a maximum
678 	 * link header, i.e.
679 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
680 	 */
681 	optlen = 0;
682 	switch (af) {
683 #ifdef INET
684 	case AF_INET:
685 		iphdrlen = sizeof(struct ip) + sizeof(struct tcphdr);
686 		break;
687 #endif
688 #ifdef INET6
689 	case AF_INET6:
690 		iphdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
691 		break;
692 #endif
693 	default:	/*pacify gcc*/
694 		iphdrlen = 0;
695 		break;
696 	}
697 	hdrlen = iphdrlen;
698 	if (flags & TH_SYN) {
699 		struct rtentry *rt;
700 
701 		rt = NULL;
702 #ifdef INET
703 		if (tp->t_inpcb)
704 			rt = in_pcbrtentry(tp->t_inpcb);
705 #endif
706 #ifdef INET6
707 		if (tp->t_in6pcb)
708 			rt = in6_pcbrtentry(tp->t_in6pcb);
709 #endif
710 
711 		tp->snd_nxt = tp->iss;
712 		tp->t_ourmss = tcp_mss_to_advertise(rt != NULL ?
713 						    rt->rt_ifp : NULL, af);
714 		if ((tp->t_flags & TF_NOOPT) == 0) {
715 			opt[0] = TCPOPT_MAXSEG;
716 			opt[1] = 4;
717 			opt[2] = (tp->t_ourmss >> 8) & 0xff;
718 			opt[3] = tp->t_ourmss & 0xff;
719 			optlen = 4;
720 
721 			if ((tp->t_flags & TF_REQ_SCALE) &&
722 			    ((flags & TH_ACK) == 0 ||
723 			    (tp->t_flags & TF_RCVD_SCALE))) {
724 				*((u_int32_t *) (opt + optlen)) = htonl(
725 					TCPOPT_NOP << 24 |
726 					TCPOPT_WINDOW << 16 |
727 					TCPOLEN_WINDOW << 8 |
728 					tp->request_r_scale);
729 				optlen += 4;
730 			}
731 		}
732  	}
733 
734  	/*
735 	 * Send a timestamp and echo-reply if this is a SYN and our side
736 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
737 	 * and our peer have sent timestamps in our SYN's.
738  	 */
739  	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
740  	     (flags & TH_RST) == 0 &&
741  	    ((flags & (TH_SYN|TH_ACK)) == TH_SYN ||
742 	     (tp->t_flags & TF_RCVD_TSTMP))) {
743 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
744 
745  		/* Form timestamp option as shown in appendix A of RFC 1323. */
746  		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
747  		*lp++ = htonl(TCP_TIMESTAMP(tp));
748  		*lp   = htonl(tp->ts_recent);
749  		optlen += TCPOLEN_TSTAMP_APPA;
750  	}
751 
752  	hdrlen += optlen;
753 
754 #ifdef DIAGNOSTIC
755 	if (len > txsegsize)
756 		panic("tcp data to be sent is larger than segment");
757  	if (max_linkhdr + hdrlen > MCLBYTES)
758 		panic("tcphdr too big");
759 #endif
760 
761 	/*
762 	 * Grab a header mbuf, attaching a copy of data to
763 	 * be transmitted, and initialize the header from
764 	 * the template for sends on this connection.
765 	 */
766 	if (len) {
767 		error = tcp_build_datapkt(tp, so, off, len, hdrlen, &m);
768 		if (error)
769 			goto out;
770 		/*
771 		 * If we're sending everything we've got, set PUSH.
772 		 * (This will keep happy those implementations which only
773 		 * give data to the user when a buffer fills or
774 		 * a PUSH comes in.)
775 		 */
776 		if (off + len == so->so_snd.sb_cc)
777 			flags |= TH_PUSH;
778 	} else {
779 		if (tp->t_flags & TF_ACKNOW)
780 			tcpstat.tcps_sndacks++;
781 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
782 			tcpstat.tcps_sndctrl++;
783 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
784 			tcpstat.tcps_sndurg++;
785 		else
786 			tcpstat.tcps_sndwinup++;
787 
788 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
789 		if (m != NULL && max_linkhdr + hdrlen > MHLEN) {
790 			MCLGET(m, M_DONTWAIT);
791 			if ((m->m_flags & M_EXT) == 0) {
792 				m_freem(m);
793 				m = NULL;
794 			}
795 		}
796 		if (m == NULL) {
797 			error = ENOBUFS;
798 			goto out;
799 		}
800 		m->m_data += max_linkhdr;
801 		m->m_len = hdrlen;
802 	}
803 	m->m_pkthdr.rcvif = (struct ifnet *)0;
804 	switch (af) {
805 #ifdef INET
806 	case AF_INET:
807 		ip = mtod(m, struct ip *);
808 #ifdef INET6
809 		ip6 = NULL;
810 #endif
811 		th = (struct tcphdr *)(ip + 1);
812 		break;
813 #endif
814 #ifdef INET6
815 	case AF_INET6:
816 		ip = NULL;
817 		ip6 = mtod(m, struct ip6_hdr *);
818 		th = (struct tcphdr *)(ip6 + 1);
819 		break;
820 #endif
821 	default:	/*pacify gcc*/
822 		ip = NULL;
823 #ifdef INET6
824 		ip6 = NULL;
825 #endif
826 		th = NULL;
827 		break;
828 	}
829 	if (tp->t_template == 0)
830 		panic("tcp_output");
831 	if (tp->t_template->m_len < iphdrlen)
832 		panic("tcp_output");
833 	bcopy(mtod(tp->t_template, caddr_t), mtod(m, caddr_t), iphdrlen);
834 
835 	/*
836 	 * If we are doing retransmissions, then snd_nxt will
837 	 * not reflect the first unsent octet.  For ACK only
838 	 * packets, we do not want the sequence number of the
839 	 * retransmitted packet, we want the sequence number
840 	 * of the next unsent octet.  So, if there is no data
841 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
842 	 * when filling in ti_seq.  But if we are in persist
843 	 * state, snd_max might reflect one byte beyond the
844 	 * right edge of the window, so use snd_nxt in that
845 	 * case, since we know we aren't doing a retransmission.
846 	 * (retransmit and persist are mutually exclusive...)
847 	 */
848 	if (len || (flags & (TH_SYN|TH_FIN)) ||
849 	    TCP_TIMER_ISARMED(tp, TCPT_PERSIST))
850 		th->th_seq = htonl(tp->snd_nxt);
851 	else
852 		th->th_seq = htonl(tp->snd_max);
853 	th->th_ack = htonl(tp->rcv_nxt);
854 	if (optlen) {
855 		bcopy((caddr_t)opt, (caddr_t)(th + 1), optlen);
856 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
857 	}
858 	th->th_flags = flags;
859 	/*
860 	 * Calculate receive window.  Don't shrink window,
861 	 * but avoid silly window syndrome.
862 	 */
863 	if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)rxsegsize)
864 		win = 0;
865 	if (win > (long)TCP_MAXWIN << tp->rcv_scale)
866 		win = (long)TCP_MAXWIN << tp->rcv_scale;
867 	if (win < (long)(tp->rcv_adv - tp->rcv_nxt))
868 		win = (long)(tp->rcv_adv - tp->rcv_nxt);
869 	th->th_win = htons((u_int16_t) (win>>tp->rcv_scale));
870 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
871 		u_int32_t urp = tp->snd_up - tp->snd_nxt;
872 		if (urp > IP_MAXPACKET)
873 			urp = IP_MAXPACKET;
874 		th->th_urp = htons((u_int16_t)urp);
875 		th->th_flags |= TH_URG;
876 	} else
877 		/*
878 		 * If no urgent pointer to send, then we pull
879 		 * the urgent pointer to the left edge of the send window
880 		 * so that it doesn't drift into the send window on sequence
881 		 * number wraparound.
882 		 */
883 		tp->snd_up = tp->snd_una;		/* drag it along */
884 
885 	/*
886 	 * Set ourselves up to be checksummed just before the packet
887 	 * hits the wire.
888 	 */
889 	switch (af) {
890 #ifdef INET
891 	case AF_INET:
892 		m->m_pkthdr.csum_flags = M_CSUM_TCPv4;
893 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
894 		if (len + optlen) {
895 			/* Fixup the pseudo-header checksum. */
896 			/* XXXJRT Not IP Jumbogram safe. */
897 			th->th_sum = in_cksum_addword(th->th_sum,
898 			    htons((u_int16_t) (len + optlen)));
899 		}
900 		break;
901 #endif
902 #ifdef INET6
903 	case AF_INET6:
904 		/*
905 		 * XXX Actually delaying the checksum is Hard
906 		 * XXX (well, maybe not for Itojun, but it is
907 		 * XXX for me), but we can still take advantage
908 		 * XXX of the cached pseudo-header checksum.
909 		 */
910 		/* equals to hdrlen + len */
911 		m->m_pkthdr.len = sizeof(struct ip6_hdr)
912 			+ sizeof(struct tcphdr) + optlen + len;
913 #ifdef notyet
914 		m->m_pkthdr.csum_flags = M_CSUM_TCPv6;
915 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
916 #endif
917 		if (len + optlen) {
918 			/* Fixup the pseudo-header checksum. */
919 			/* XXXJRT: Not IPv6 Jumbogram safe. */
920 			th->th_sum = in_cksum_addword(th->th_sum,
921 			    htons((u_int16_t) (len + optlen)));
922 		}
923 #ifndef notyet
924 		th->th_sum = in6_cksum(m, 0, sizeof(struct ip6_hdr),
925 		    sizeof(struct tcphdr) + optlen + len);
926 #endif
927 		break;
928 #endif
929 	}
930 
931 	/*
932 	 * In transmit state, time the transmission and arrange for
933 	 * the retransmit.  In persist state, just set snd_max.
934 	 */
935 	if (tp->t_force == 0 || TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
936 		tcp_seq startseq = tp->snd_nxt;
937 
938 		/*
939 		 * Advance snd_nxt over sequence space of this segment.
940 		 * There are no states in which we send both a SYN and a FIN,
941 		 * so we collapse the tests for these flags.
942 		 */
943 		if (flags & (TH_SYN|TH_FIN))
944 			tp->snd_nxt++;
945 		tp->snd_nxt += len;
946 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
947 			tp->snd_max = tp->snd_nxt;
948 			/*
949 			 * Time this transmission if not a retransmission and
950 			 * not currently timing anything.
951 			 */
952 			if (tp->t_rtttime == 0) {
953 				tp->t_rtttime = tcp_now;
954 				tp->t_rtseq = startseq;
955 				tcpstat.tcps_segstimed++;
956 			}
957 		}
958 
959 		/*
960 		 * Set retransmit timer if not currently set,
961 		 * and not doing an ack or a keep-alive probe.
962 		 * Initial value for retransmit timer is smoothed
963 		 * round-trip time + 2 * round-trip time variance.
964 		 * Initialize shift counter which is used for backoff
965 		 * of retransmit time.
966 		 */
967 		if (TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
968 		    tp->snd_nxt != tp->snd_una) {
969 			TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
970 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
971 				TCP_TIMER_DISARM(tp, TCPT_PERSIST);
972 				tp->t_rxtshift = 0;
973 			}
974 		}
975 	} else
976 		if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
977 			tp->snd_max = tp->snd_nxt + len;
978 
979 #ifdef TCP_DEBUG
980 	/*
981 	 * Trace.
982 	 */
983 	if (so->so_options & SO_DEBUG) {
984 		/*
985 		 * need to recover version # field, which was overwritten
986 		 * on ip_cksum computation.
987 		 */
988 		struct ip *sip;
989 		sip = mtod(m, struct ip *);
990 		switch (af) {
991 #ifdef INET
992 		case AF_INET:
993 			sip->ip_v = 4;
994 			break;
995 #endif
996 #ifdef INET6
997 		case AF_INET6:
998 			sip->ip_v = 6;
999 			break;
1000 #endif
1001 		}
1002 		tcp_trace(TA_OUTPUT, tp->t_state, tp, m, 0);
1003 	}
1004 #endif
1005 
1006 	/*
1007 	 * Fill in IP length and desired time to live and
1008 	 * send to IP level.  There should be a better way
1009 	 * to handle ttl and tos; we could keep them in
1010 	 * the template, but need a way to checksum without them.
1011 	 */
1012 	m->m_pkthdr.len = hdrlen + len;
1013 
1014 	switch (af) {
1015 #ifdef INET
1016 	case AF_INET:
1017 		ip->ip_len = m->m_pkthdr.len;
1018 		if (tp->t_inpcb) {
1019 			ip->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl;
1020 			ip->ip_tos = tp->t_inpcb->inp_ip.ip_tos;
1021 		}
1022 #ifdef INET6
1023 		else if (tp->t_in6pcb) {
1024 			ip->ip_ttl = in6_selecthlim(tp->t_in6pcb, NULL); /*XXX*/
1025 			ip->ip_tos = 0;	/*XXX*/
1026 		}
1027 #endif
1028 		break;
1029 #endif
1030 #ifdef INET6
1031 	case AF_INET6:
1032 		ip6->ip6_nxt = IPPROTO_TCP;
1033 		if (tp->t_in6pcb) {
1034 			/*
1035 			 * we separately set hoplimit for every segment, since
1036 			 * the user might want to change the value via
1037 			 * setsockopt. Also, desired default hop limit might
1038 			 * be changed via Neighbor Discovery.
1039 			 */
1040 			ip6->ip6_hlim = in6_selecthlim(tp->t_in6pcb,
1041 				ro->ro_rt ? ro->ro_rt->rt_ifp : NULL);
1042 		}
1043 		/* ip6->ip6_flow = ??? */
1044 		/* ip6_plen will be filled in ip6_output(). */
1045 		break;
1046 #endif
1047 	}
1048 
1049 #ifdef IPSEC
1050 	if (ipsec_setsocket(m, so) != 0) {
1051 		m_freem(m);
1052 		error = ENOBUFS;
1053 		goto out;
1054 	}
1055 #endif /*IPSEC*/
1056 
1057 	switch (af) {
1058 #ifdef INET
1059 	case AF_INET:
1060 	    {
1061 		struct mbuf *opts;
1062 
1063 		if (tp->t_inpcb)
1064 			opts = tp->t_inpcb->inp_options;
1065 		else
1066 			opts = NULL;
1067 		error = ip_output(m, opts, ro,
1068 			(ip_mtudisc ? IP_MTUDISC : 0) |
1069 			(so->so_options & SO_DONTROUTE),
1070 			0);
1071 		break;
1072 	    }
1073 #endif
1074 #ifdef INET6
1075 	case AF_INET6:
1076 	    {
1077 		struct ip6_pktopts *opts;
1078 
1079 		if (tp->t_in6pcb)
1080 			opts = tp->t_in6pcb->in6p_outputopts;
1081 		else
1082 			opts = NULL;
1083 		error = ip6_output(m, opts, (struct route_in6 *)ro,
1084 			so->so_options & SO_DONTROUTE, 0, NULL);
1085 		break;
1086 	    }
1087 #endif
1088 	default:
1089 		error = EAFNOSUPPORT;
1090 		break;
1091 	}
1092 	if (error) {
1093 out:
1094 		if (error == ENOBUFS) {
1095 			tcpstat.tcps_selfquench++;
1096 #ifdef INET
1097 			if (tp->t_inpcb)
1098 				tcp_quench(tp->t_inpcb, 0);
1099 #endif
1100 #ifdef INET6
1101 			if (tp->t_in6pcb)
1102 				tcp6_quench(tp->t_in6pcb, 0);
1103 #endif
1104 			error = 0;
1105 		} else if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1106 		    TCPS_HAVERCVDSYN(tp->t_state)) {
1107 			tp->t_softerror = error;
1108 			error = 0;
1109 		}
1110 
1111 		/* Restart the delayed ACK timer, if necessary. */
1112 		if (tp->t_flags & TF_DELACK)
1113 			TCP_RESTART_DELACK(tp);
1114 
1115 		return (error);
1116 	}
1117 	tcpstat.tcps_sndtotal++;
1118 	if (tp->t_flags & TF_DELACK)
1119 		tcpstat.tcps_delack++;
1120 
1121 	/*
1122 	 * Data sent (as far as we can tell).
1123 	 * If this advertises a larger window than any other segment,
1124 	 * then remember the size of the advertised window.
1125 	 * Any pending ACK has now been sent.
1126 	 */
1127 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
1128 		tp->rcv_adv = tp->rcv_nxt + win;
1129 	tp->last_ack_sent = tp->rcv_nxt;
1130 	tp->t_flags &= ~TF_ACKNOW;
1131 	TCP_CLEAR_DELACK(tp);
1132 #ifdef DIAGNOSTIC
1133 	if (maxburst < 0)
1134 		printf("tcp_output: maxburst exceeded by %d\n", -maxburst);
1135 #endif
1136 	if (sendalot && (!tcp_do_newreno || --maxburst))
1137 		goto again;
1138 	return (0);
1139 }
1140 
1141 void
1142 tcp_setpersist(tp)
1143 	struct tcpcb *tp;
1144 {
1145 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> (1 + 2);
1146 	int nticks;
1147 
1148 	if (TCP_TIMER_ISARMED(tp, TCPT_REXMT))
1149 		panic("tcp_output REXMT");
1150 	/*
1151 	 * Start/restart persistance timer.
1152 	 */
1153 	if (t < tp->t_rttmin)
1154 		t = tp->t_rttmin;
1155 	TCPT_RANGESET(nticks, t * tcp_backoff[tp->t_rxtshift],
1156 	    TCPTV_PERSMIN, TCPTV_PERSMAX);
1157 	TCP_TIMER_ARM(tp, TCPT_PERSIST, nticks);
1158 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1159 		tp->t_rxtshift++;
1160 }
1161