xref: /freebsd/sys/netinet/tcp_output.c (revision c4f6a2a9e1b1879b618c436ab4f56ff75c73a0f5)
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_mac.h"
40 #include "opt_tcpdebug.h"
41 
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/domain.h>
45 #include <sys/kernel.h>
46 #include <sys/lock.h>
47 #include <sys/mac.h>
48 #include <sys/mbuf.h>
49 #include <sys/mutex.h>
50 #include <sys/protosw.h>
51 #include <sys/socket.h>
52 #include <sys/socketvar.h>
53 #include <sys/sysctl.h>
54 
55 #include <net/route.h>
56 
57 #include <netinet/in.h>
58 #include <netinet/in_systm.h>
59 #include <netinet/ip.h>
60 #include <netinet/in_pcb.h>
61 #include <netinet/ip_var.h>
62 #ifdef INET6
63 #include <netinet6/in6_pcb.h>
64 #include <netinet/ip6.h>
65 #include <netinet6/ip6_var.h>
66 #endif
67 #include <netinet/tcp.h>
68 #define	TCPOUTFLAGS
69 #include <netinet/tcp_fsm.h>
70 #include <netinet/tcp_seq.h>
71 #include <netinet/tcp_timer.h>
72 #include <netinet/tcp_var.h>
73 #include <netinet/tcpip.h>
74 #ifdef TCPDEBUG
75 #include <netinet/tcp_debug.h>
76 #endif
77 
78 #ifdef IPSEC
79 #include <netinet6/ipsec.h>
80 #endif /*IPSEC*/
81 
82 #include <machine/in_cksum.h>
83 
84 #ifdef notyet
85 extern struct mbuf *m_copypack();
86 #endif
87 
88 int path_mtu_discovery = 1;
89 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
90 	&path_mtu_discovery, 1, "Enable Path MTU Discovery");
91 
92 int ss_fltsz = 1;
93 SYSCTL_INT(_net_inet_tcp, OID_AUTO, slowstart_flightsize, CTLFLAG_RW,
94 	&ss_fltsz, 1, "Slow start flight size");
95 
96 int ss_fltsz_local = 4;
97 SYSCTL_INT(_net_inet_tcp, OID_AUTO, local_slowstart_flightsize, CTLFLAG_RW,
98 	&ss_fltsz_local, 1, "Slow start flight size for local networks");
99 
100 int     tcp_do_newreno = 1;
101 SYSCTL_INT(_net_inet_tcp, OID_AUTO, newreno, CTLFLAG_RW, &tcp_do_newreno,
102         0, "Enable NewReno Algorithms");
103 /*
104  * Tcp output routine: figure out what should be sent and send it.
105  */
106 int
107 tcp_output(struct tcpcb *tp)
108 {
109 	struct socket *so = tp->t_inpcb->inp_socket;
110 	long len, win;
111 	int off, flags, error;
112 	struct mbuf *m;
113 	struct ip *ip = NULL;
114 	struct ipovly *ipov = NULL;
115 	struct tcphdr *th;
116 	u_char opt[TCP_MAXOLEN];
117 	unsigned ipoptlen, optlen, hdrlen;
118 	int idle, sendalot;
119 #if 0
120 	int maxburst = TCP_MAXBURST;
121 #endif
122 	struct rmxp_tao *taop;
123 	struct rmxp_tao tao_noncached;
124 #ifdef INET6
125 	struct ip6_hdr *ip6 = NULL;
126 	int isipv6;
127 
128 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
129 #endif
130 
131 	mtx_assert(&tp->t_inpcb->inp_mtx, MA_OWNED);
132 
133 	/*
134 	 * Determine length of data that should be transmitted,
135 	 * and flags that will be used.
136 	 * If there is some data or critical controls (SYN, RST)
137 	 * to send, then transmit; otherwise, investigate further.
138 	 */
139 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
140 	if (idle && (ticks - tp->t_rcvtime) >= tp->t_rxtcur) {
141 		/*
142 		 * We have been idle for "a while" and no acks are
143 		 * expected to clock out any data we send --
144 		 * slow start to get ack "clock" running again.
145 		 *
146 		 * Set the slow-start flight size depending on whether
147 		 * this is a local network or not.
148 		 */
149 		int ss = ss_fltsz;
150 #ifdef INET6
151 		if (isipv6) {
152 			if (in6_localaddr(&tp->t_inpcb->in6p_faddr))
153 				ss = ss_fltsz_local;
154 		} else
155 #endif /* INET6 */
156 		if (in_localaddr(tp->t_inpcb->inp_faddr))
157 			ss = ss_fltsz_local;
158 		tp->snd_cwnd = tp->t_maxseg * ss;
159 	}
160 	tp->t_flags &= ~TF_LASTIDLE;
161 	if (idle) {
162 		if (tp->t_flags & TF_MORETOCOME) {
163 			tp->t_flags |= TF_LASTIDLE;
164 			idle = 0;
165 		}
166 	}
167 again:
168 	sendalot = 0;
169 	off = tp->snd_nxt - tp->snd_una;
170 	win = min(tp->snd_wnd, tp->snd_cwnd);
171 	win = min(win, tp->snd_bwnd);
172 
173 	flags = tcp_outflags[tp->t_state];
174 	/*
175 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
176 	 * state flags.
177 	 */
178 	if (tp->t_flags & TF_NEEDFIN)
179 		flags |= TH_FIN;
180 	if (tp->t_flags & TF_NEEDSYN)
181 		flags |= TH_SYN;
182 
183 	/*
184 	 * If in persist timeout with window of 0, send 1 byte.
185 	 * Otherwise, if window is small but nonzero
186 	 * and timer expired, we will send what we can
187 	 * and go to transmit state.
188 	 */
189 	if (tp->t_force) {
190 		if (win == 0) {
191 			/*
192 			 * If we still have some data to send, then
193 			 * clear the FIN bit.  Usually this would
194 			 * happen below when it realizes that we
195 			 * aren't sending all the data.  However,
196 			 * if we have exactly 1 byte of unsent data,
197 			 * then it won't clear the FIN bit below,
198 			 * and if we are in persist state, we wind
199 			 * up sending the packet without recording
200 			 * that we sent the FIN bit.
201 			 *
202 			 * We can't just blindly clear the FIN bit,
203 			 * because if we don't have any more data
204 			 * to send then the probe will be the FIN
205 			 * itself.
206 			 */
207 			if (off < so->so_snd.sb_cc)
208 				flags &= ~TH_FIN;
209 			win = 1;
210 		} else {
211 			callout_stop(tp->tt_persist);
212 			tp->t_rxtshift = 0;
213 		}
214 	}
215 
216 	len = (long)ulmin(so->so_snd.sb_cc, win) - off;
217 
218 	if ((taop = tcp_gettaocache(&tp->t_inpcb->inp_inc)) == NULL) {
219 		taop = &tao_noncached;
220 		bzero(taop, sizeof(*taop));
221 	}
222 
223 	/*
224 	 * Lop off SYN bit if it has already been sent.  However, if this
225 	 * is SYN-SENT state and if segment contains data and if we don't
226 	 * know that foreign host supports TAO, suppress sending segment.
227 	 */
228 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
229 		flags &= ~TH_SYN;
230 		off--, len++;
231 		if (len > 0 && tp->t_state == TCPS_SYN_SENT &&
232 		    taop->tao_ccsent == 0)
233 			return 0;
234 	}
235 
236 	/*
237 	 * Be careful not to send data and/or FIN on SYN segments
238 	 * in cases when no CC option will be sent.
239 	 * This measure is needed to prevent interoperability problems
240 	 * with not fully conformant TCP implementations.
241 	 */
242 	if ((flags & TH_SYN) &&
243 	    ((tp->t_flags & TF_NOOPT) || !(tp->t_flags & TF_REQ_CC) ||
244 	     ((flags & TH_ACK) && !(tp->t_flags & TF_RCVD_CC)))) {
245 		len = 0;
246 		flags &= ~TH_FIN;
247 	}
248 
249 	if (len < 0) {
250 		/*
251 		 * If FIN has been sent but not acked,
252 		 * but we haven't been called to retransmit,
253 		 * len will be -1.  Otherwise, window shrank
254 		 * after we sent into it.  If window shrank to 0,
255 		 * cancel pending retransmit, pull snd_nxt back
256 		 * to (closed) window, and set the persist timer
257 		 * if it isn't already going.  If the window didn't
258 		 * close completely, just wait for an ACK.
259 		 */
260 		len = 0;
261 		if (win == 0) {
262 			callout_stop(tp->tt_rexmt);
263 			tp->t_rxtshift = 0;
264 			tp->snd_nxt = tp->snd_una;
265 			if (!callout_active(tp->tt_persist))
266 				tcp_setpersist(tp);
267 		}
268 	}
269 	if (len > tp->t_maxseg) {
270 		len = tp->t_maxseg;
271 		sendalot = 1;
272 	}
273 	if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
274 		flags &= ~TH_FIN;
275 
276 	win = sbspace(&so->so_rcv);
277 
278 	/*
279 	 * Sender silly window avoidance.   We transmit under the following
280 	 * conditions when len is non-zero:
281 	 *
282 	 *	- We have a full segment
283 	 *	- This is the last buffer in a write()/send() and we are
284 	 *	  either idle or running NODELAY
285 	 *	- we've timed out (e.g. persist timer)
286 	 *	- we have more then 1/2 the maximum send window's worth of
287 	 *	  data (receiver may be limited the window size)
288 	 *	- we need to retransmit
289 	 */
290 	if (len) {
291 		if (len == tp->t_maxseg)
292 			goto send;
293 		/*
294 		 * NOTE! on localhost connections an 'ack' from the remote
295 		 * end may occur synchronously with the output and cause
296 		 * us to flush a buffer queued with moretocome.  XXX
297 		 *
298 		 * note: the len + off check is almost certainly unnecessary.
299 		 */
300 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
301 		    (idle || (tp->t_flags & TF_NODELAY)) &&
302 		    len + off >= so->so_snd.sb_cc &&
303 		    (tp->t_flags & TF_NOPUSH) == 0) {
304 			goto send;
305 		}
306 		if (tp->t_force)			/* typ. timeout case */
307 			goto send;
308 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
309 			goto send;
310 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
311 			goto send;
312 	}
313 
314 	/*
315 	 * Compare available window to amount of window
316 	 * known to peer (as advertised window less
317 	 * next expected input).  If the difference is at least two
318 	 * max size segments, or at least 50% of the maximum possible
319 	 * window, then want to send a window update to peer.
320 	 */
321 	if (win > 0) {
322 		/*
323 		 * "adv" is the amount we can increase the window,
324 		 * taking into account that we are limited by
325 		 * TCP_MAXWIN << tp->rcv_scale.
326 		 */
327 		long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) -
328 			(tp->rcv_adv - tp->rcv_nxt);
329 
330 		if (adv >= (long) (2 * tp->t_maxseg))
331 			goto send;
332 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
333 			goto send;
334 	}
335 
336 	/*
337 	 * Send if we owe peer an ACK.
338 	 */
339 	if (tp->t_flags & TF_ACKNOW)
340 		goto send;
341 	if ((flags & TH_RST) ||
342 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
343 		goto send;
344 	if (SEQ_GT(tp->snd_up, tp->snd_una))
345 		goto send;
346 	/*
347 	 * If our state indicates that FIN should be sent
348 	 * and we have not yet done so, or we're retransmitting the FIN,
349 	 * then we need to send.
350 	 */
351 	if (flags & TH_FIN &&
352 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
353 		goto send;
354 
355 	/*
356 	 * TCP window updates are not reliable, rather a polling protocol
357 	 * using ``persist'' packets is used to insure receipt of window
358 	 * updates.  The three ``states'' for the output side are:
359 	 *	idle			not doing retransmits or persists
360 	 *	persisting		to move a small or zero window
361 	 *	(re)transmitting	and thereby not persisting
362 	 *
363 	 * callout_active(tp->tt_persist)
364 	 *	is true when we are in persist state.
365 	 * tp->t_force
366 	 *	is set when we are called to send a persist packet.
367 	 * callout_active(tp->tt_rexmt)
368 	 *	is set when we are retransmitting
369 	 * The output side is idle when both timers are zero.
370 	 *
371 	 * If send window is too small, there is data to transmit, and no
372 	 * retransmit or persist is pending, then go to persist state.
373 	 * If nothing happens soon, send when timer expires:
374 	 * if window is nonzero, transmit what we can,
375 	 * otherwise force out a byte.
376 	 */
377 	if (so->so_snd.sb_cc && !callout_active(tp->tt_rexmt) &&
378 	    !callout_active(tp->tt_persist)) {
379 		tp->t_rxtshift = 0;
380 		tcp_setpersist(tp);
381 	}
382 
383 	/*
384 	 * No reason to send a segment, just return.
385 	 */
386 	return (0);
387 
388 send:
389 	/*
390 	 * Before ESTABLISHED, force sending of initial options
391 	 * unless TCP set not to do any options.
392 	 * NOTE: we assume that the IP/TCP header plus TCP options
393 	 * always fit in a single mbuf, leaving room for a maximum
394 	 * link header, i.e.
395 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
396 	 */
397 	optlen = 0;
398 #ifdef INET6
399 	if (isipv6)
400 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
401 	else
402 #endif
403 	hdrlen = sizeof (struct tcpiphdr);
404 	if (flags & TH_SYN) {
405 		tp->snd_nxt = tp->iss;
406 		if ((tp->t_flags & TF_NOOPT) == 0) {
407 			u_short mss;
408 
409 			opt[0] = TCPOPT_MAXSEG;
410 			opt[1] = TCPOLEN_MAXSEG;
411 			mss = htons((u_short) tcp_mssopt(tp));
412 			(void)memcpy(opt + 2, &mss, sizeof(mss));
413 			optlen = TCPOLEN_MAXSEG;
414 
415 			if ((tp->t_flags & TF_REQ_SCALE) &&
416 			    ((flags & TH_ACK) == 0 ||
417 			    (tp->t_flags & TF_RCVD_SCALE))) {
418 				*((u_int32_t *)(opt + optlen)) = htonl(
419 					TCPOPT_NOP << 24 |
420 					TCPOPT_WINDOW << 16 |
421 					TCPOLEN_WINDOW << 8 |
422 					tp->request_r_scale);
423 				optlen += 4;
424 			}
425 		}
426  	}
427 
428  	/*
429 	 * Send a timestamp and echo-reply if this is a SYN and our side
430 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
431 	 * and our peer have sent timestamps in our SYN's.
432  	 */
433  	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
434  	    (flags & TH_RST) == 0 &&
435 	    ((flags & TH_ACK) == 0 ||
436 	     (tp->t_flags & TF_RCVD_TSTMP))) {
437 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
438 
439  		/* Form timestamp option as shown in appendix A of RFC 1323. */
440  		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
441  		*lp++ = htonl(ticks);
442  		*lp   = htonl(tp->ts_recent);
443  		optlen += TCPOLEN_TSTAMP_APPA;
444  	}
445 
446  	/*
447 	 * Send `CC-family' options if our side wants to use them (TF_REQ_CC),
448 	 * options are allowed (!TF_NOOPT) and it's not a RST.
449  	 */
450  	if ((tp->t_flags & (TF_REQ_CC|TF_NOOPT)) == TF_REQ_CC &&
451  	     (flags & TH_RST) == 0) {
452 		switch (flags & (TH_SYN|TH_ACK)) {
453 		/*
454 		 * This is a normal ACK, send CC if we received CC before
455 		 * from our peer.
456 		 */
457 		case TH_ACK:
458 			if (!(tp->t_flags & TF_RCVD_CC))
459 				break;
460 			/*FALLTHROUGH*/
461 
462 		/*
463 		 * We can only get here in T/TCP's SYN_SENT* state, when
464 		 * we're a sending a non-SYN segment without waiting for
465 		 * the ACK of our SYN.  A check above assures that we only
466 		 * do this if our peer understands T/TCP.
467 		 */
468 		case 0:
469 			opt[optlen++] = TCPOPT_NOP;
470 			opt[optlen++] = TCPOPT_NOP;
471 			opt[optlen++] = TCPOPT_CC;
472 			opt[optlen++] = TCPOLEN_CC;
473 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
474 
475 			optlen += 4;
476 			break;
477 
478 		/*
479 		 * This is our initial SYN, check whether we have to use
480 		 * CC or CC.new.
481 		 */
482 		case TH_SYN:
483 			opt[optlen++] = TCPOPT_NOP;
484 			opt[optlen++] = TCPOPT_NOP;
485 			opt[optlen++] = tp->t_flags & TF_SENDCCNEW ?
486 						TCPOPT_CCNEW : TCPOPT_CC;
487 			opt[optlen++] = TCPOLEN_CC;
488 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
489  			optlen += 4;
490 			break;
491 
492 		/*
493 		 * This is a SYN,ACK; send CC and CC.echo if we received
494 		 * CC from our peer.
495 		 */
496 		case (TH_SYN|TH_ACK):
497 			if (tp->t_flags & TF_RCVD_CC) {
498 				opt[optlen++] = TCPOPT_NOP;
499 				opt[optlen++] = TCPOPT_NOP;
500 				opt[optlen++] = TCPOPT_CC;
501 				opt[optlen++] = TCPOLEN_CC;
502 				*(u_int32_t *)&opt[optlen] =
503 					htonl(tp->cc_send);
504 				optlen += 4;
505 				opt[optlen++] = TCPOPT_NOP;
506 				opt[optlen++] = TCPOPT_NOP;
507 				opt[optlen++] = TCPOPT_CCECHO;
508 				opt[optlen++] = TCPOLEN_CC;
509 				*(u_int32_t *)&opt[optlen] =
510 					htonl(tp->cc_recv);
511 				optlen += 4;
512 			}
513 			break;
514 		}
515  	}
516 
517  	hdrlen += optlen;
518 
519 #ifdef INET6
520 	if (isipv6)
521 		ipoptlen = ip6_optlen(tp->t_inpcb);
522 	else
523 #endif
524 	if (tp->t_inpcb->inp_options)
525 		ipoptlen = tp->t_inpcb->inp_options->m_len -
526 				offsetof(struct ipoption, ipopt_list);
527 	else
528 		ipoptlen = 0;
529 #ifdef IPSEC
530 	ipoptlen += ipsec_hdrsiz_tcp(tp);
531 #endif
532 
533 	/*
534 	 * Adjust data length if insertion of options will
535 	 * bump the packet length beyond the t_maxopd length.
536 	 * Clear the FIN bit because we cut off the tail of
537 	 * the segment.
538 	 */
539 	if (len + optlen + ipoptlen > tp->t_maxopd) {
540 		/*
541 		 * If there is still more to send, don't close the connection.
542 		 */
543 		flags &= ~TH_FIN;
544 		len = tp->t_maxopd - optlen - ipoptlen;
545 		sendalot = 1;
546 	}
547 
548 /*#ifdef DIAGNOSTIC*/
549 #ifdef INET6
550  	if (max_linkhdr + hdrlen > MCLBYTES)
551 #else
552  	if (max_linkhdr + hdrlen > MHLEN)
553 #endif
554 		panic("tcphdr too big");
555 /*#endif*/
556 
557 	/*
558 	 * Grab a header mbuf, attaching a copy of data to
559 	 * be transmitted, and initialize the header from
560 	 * the template for sends on this connection.
561 	 */
562 	if (len) {
563 		if (tp->t_force && len == 1)
564 			tcpstat.tcps_sndprobe++;
565 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
566 			tcpstat.tcps_sndrexmitpack++;
567 			tcpstat.tcps_sndrexmitbyte += len;
568 		} else {
569 			tcpstat.tcps_sndpack++;
570 			tcpstat.tcps_sndbyte += len;
571 		}
572 #ifdef notyet
573 		if ((m = m_copypack(so->so_snd.sb_mb, off,
574 		    (int)len, max_linkhdr + hdrlen)) == 0) {
575 			error = ENOBUFS;
576 			goto out;
577 		}
578 		/*
579 		 * m_copypack left space for our hdr; use it.
580 		 */
581 		m->m_len += hdrlen;
582 		m->m_data -= hdrlen;
583 #else
584 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
585 		if (m == NULL) {
586 			error = ENOBUFS;
587 			goto out;
588 		}
589 #ifdef INET6
590 		if (MHLEN < hdrlen + max_linkhdr) {
591 			MCLGET(m, M_DONTWAIT);
592 			if ((m->m_flags & M_EXT) == 0) {
593 				m_freem(m);
594 				error = ENOBUFS;
595 				goto out;
596 			}
597 		}
598 #endif
599 		m->m_data += max_linkhdr;
600 		m->m_len = hdrlen;
601 		if (len <= MHLEN - hdrlen - max_linkhdr) {
602 			m_copydata(so->so_snd.sb_mb, off, (int) len,
603 			    mtod(m, caddr_t) + hdrlen);
604 			m->m_len += len;
605 		} else {
606 			m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len);
607 			if (m->m_next == 0) {
608 				(void) m_free(m);
609 				error = ENOBUFS;
610 				goto out;
611 			}
612 		}
613 #endif
614 		/*
615 		 * If we're sending everything we've got, set PUSH.
616 		 * (This will keep happy those implementations which only
617 		 * give data to the user when a buffer fills or
618 		 * a PUSH comes in.)
619 		 */
620 		if (off + len == so->so_snd.sb_cc)
621 			flags |= TH_PUSH;
622 	} else {
623 		if (tp->t_flags & TF_ACKNOW)
624 			tcpstat.tcps_sndacks++;
625 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
626 			tcpstat.tcps_sndctrl++;
627 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
628 			tcpstat.tcps_sndurg++;
629 		else
630 			tcpstat.tcps_sndwinup++;
631 
632 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
633 		if (m == NULL) {
634 			error = ENOBUFS;
635 			goto out;
636 		}
637 #ifdef INET6
638 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
639 		    MHLEN >= hdrlen) {
640 			MH_ALIGN(m, hdrlen);
641 		} else
642 #endif
643 		m->m_data += max_linkhdr;
644 		m->m_len = hdrlen;
645 	}
646 	m->m_pkthdr.rcvif = (struct ifnet *)0;
647 #ifdef MAC
648 	mac_create_mbuf_from_socket(so, m);
649 #endif
650 #ifdef INET6
651 	if (isipv6) {
652 		ip6 = mtod(m, struct ip6_hdr *);
653 		th = (struct tcphdr *)(ip6 + 1);
654 		tcp_fillheaders(tp, ip6, th);
655 	} else
656 #endif /* INET6 */
657       {
658 	ip = mtod(m, struct ip *);
659 	ipov = (struct ipovly *)ip;
660 	th = (struct tcphdr *)(ip + 1);
661 	/* this picks up the pseudo header (w/o the length) */
662 	tcp_fillheaders(tp, ip, th);
663       }
664 
665 	/*
666 	 * Fill in fields, remembering maximum advertised
667 	 * window for use in delaying messages about window sizes.
668 	 * If resending a FIN, be sure not to use a new sequence number.
669 	 */
670 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
671 	    tp->snd_nxt == tp->snd_max)
672 		tp->snd_nxt--;
673 	/*
674 	 * If we are doing retransmissions, then snd_nxt will
675 	 * not reflect the first unsent octet.  For ACK only
676 	 * packets, we do not want the sequence number of the
677 	 * retransmitted packet, we want the sequence number
678 	 * of the next unsent octet.  So, if there is no data
679 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
680 	 * when filling in ti_seq.  But if we are in persist
681 	 * state, snd_max might reflect one byte beyond the
682 	 * right edge of the window, so use snd_nxt in that
683 	 * case, since we know we aren't doing a retransmission.
684 	 * (retransmit and persist are mutually exclusive...)
685 	 */
686 	if (len || (flags & (TH_SYN|TH_FIN))
687 	    || callout_active(tp->tt_persist))
688 		th->th_seq = htonl(tp->snd_nxt);
689 	else
690 		th->th_seq = htonl(tp->snd_max);
691 	th->th_ack = htonl(tp->rcv_nxt);
692 	if (optlen) {
693 		bcopy(opt, th + 1, optlen);
694 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
695 	}
696 	th->th_flags = flags;
697 	/*
698 	 * Calculate receive window.  Don't shrink window,
699 	 * but avoid silly window syndrome.
700 	 */
701 	if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)tp->t_maxseg)
702 		win = 0;
703 	if (win < (long)(tp->rcv_adv - tp->rcv_nxt))
704 		win = (long)(tp->rcv_adv - tp->rcv_nxt);
705 	if (win > (long)TCP_MAXWIN << tp->rcv_scale)
706 		win = (long)TCP_MAXWIN << tp->rcv_scale;
707 	th->th_win = htons((u_short) (win>>tp->rcv_scale));
708 
709 
710 	/*
711 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
712 	 * a 0 window.  This may cause the remote transmitter to stall.  This
713 	 * flag tells soreceive() to disable delayed acknowledgements when
714 	 * draining the buffer.  This can occur if the receiver is attempting
715 	 * to read more data then can be buffered prior to transmitting on
716 	 * the connection.
717 	 */
718 	if (win == 0)
719 		tp->t_flags |= TF_RXWIN0SENT;
720 	else
721 		tp->t_flags &= ~TF_RXWIN0SENT;
722 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
723 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
724 		th->th_flags |= TH_URG;
725 	} else
726 		/*
727 		 * If no urgent pointer to send, then we pull
728 		 * the urgent pointer to the left edge of the send window
729 		 * so that it doesn't drift into the send window on sequence
730 		 * number wraparound.
731 		 */
732 		tp->snd_up = tp->snd_una;		/* drag it along */
733 
734 	/*
735 	 * Put TCP length in extended header, and then
736 	 * checksum extended header and data.
737 	 */
738 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
739 #ifdef INET6
740 	if (isipv6)
741 		/*
742 		 * ip6_plen is not need to be filled now, and will be filled
743 		 * in ip6_output.
744 		 */
745 		th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
746 				       sizeof(struct tcphdr) + optlen + len);
747 	else
748 #endif /* INET6 */
749       {
750 	m->m_pkthdr.csum_flags = CSUM_TCP;
751 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
752 	if (len + optlen)
753 		th->th_sum = in_addword(th->th_sum,
754 		    htons((u_short)(optlen + len)));
755 
756 	/* IP version must be set here for ipv4/ipv6 checking later */
757 	KASSERT(ip->ip_v == IPVERSION,
758 	    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
759       }
760 
761 	/*
762 	 * In transmit state, time the transmission and arrange for
763 	 * the retransmit.  In persist state, just set snd_max.
764 	 */
765 	if (tp->t_force == 0 || !callout_active(tp->tt_persist)) {
766 		tcp_seq startseq = tp->snd_nxt;
767 
768 		/*
769 		 * Advance snd_nxt over sequence space of this segment.
770 		 */
771 		if (flags & (TH_SYN|TH_FIN)) {
772 			if (flags & TH_SYN)
773 				tp->snd_nxt++;
774 			if (flags & TH_FIN) {
775 				tp->snd_nxt++;
776 				tp->t_flags |= TF_SENTFIN;
777 			}
778 		}
779 		tp->snd_nxt += len;
780 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
781 			tp->snd_max = tp->snd_nxt;
782 			/*
783 			 * Time this transmission if not a retransmission and
784 			 * not currently timing anything.
785 			 */
786 			if (tp->t_rtttime == 0) {
787 				tp->t_rtttime = ticks;
788 				tp->t_rtseq = startseq;
789 				tcpstat.tcps_segstimed++;
790 			}
791 		}
792 
793 		/*
794 		 * Set retransmit timer if not currently set,
795 		 * and not doing an ack or a keep-alive probe.
796 		 * Initial value for retransmit timer is smoothed
797 		 * round-trip time + 2 * round-trip time variance.
798 		 * Initialize shift counter which is used for backoff
799 		 * of retransmit time.
800 		 */
801 		if (!callout_active(tp->tt_rexmt) &&
802 		    tp->snd_nxt != tp->snd_una) {
803 			if (callout_active(tp->tt_persist)) {
804 				callout_stop(tp->tt_persist);
805 				tp->t_rxtshift = 0;
806 			}
807 			callout_reset(tp->tt_rexmt, tp->t_rxtcur,
808 				      tcp_timer_rexmt, tp);
809 		}
810 	} else
811 		if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
812 			tp->snd_max = tp->snd_nxt + len;
813 
814 #ifdef TCPDEBUG
815 	/*
816 	 * Trace.
817 	 */
818 	if (so->so_options & SO_DEBUG)
819 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
820 #endif
821 
822 	/*
823 	 * Fill in IP length and desired time to live and
824 	 * send to IP level.  There should be a better way
825 	 * to handle ttl and tos; we could keep them in
826 	 * the template, but need a way to checksum without them.
827 	 */
828 	/*
829 	 * m->m_pkthdr.len should have been set before cksum calcuration,
830 	 * because in6_cksum() need it.
831 	 */
832 #ifdef INET6
833 	if (isipv6) {
834 		/*
835 		 * we separately set hoplimit for every segment, since the
836 		 * user might want to change the value via setsockopt.
837 		 * Also, desired default hop limit might be changed via
838 		 * Neighbor Discovery.
839 		 */
840 		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb,
841 					       tp->t_inpcb->in6p_route.ro_rt ?
842 					       tp->t_inpcb->in6p_route.ro_rt->rt_ifp
843 					       : NULL);
844 
845 		/* TODO: IPv6 IP6TOS_ECT bit on */
846 #ifdef IPSEC
847 		if (ipsec_setsocket(m, so) != 0) {
848 			m_freem(m);
849 			error = ENOBUFS;
850 			goto out;
851 		}
852 #endif /*IPSEC*/
853 		error = ip6_output(m,
854 			    tp->t_inpcb->in6p_outputopts,
855 			    &tp->t_inpcb->in6p_route,
856 			    (so->so_options & SO_DONTROUTE), NULL, NULL);
857 	} else
858 #endif /* INET6 */
859     {
860 	struct rtentry *rt;
861 	ip->ip_len = m->m_pkthdr.len;
862 #ifdef INET6
863  	if (INP_CHECK_SOCKAF(so, AF_INET6))
864  		ip->ip_ttl = in6_selecthlim(tp->t_inpcb,
865  					    tp->t_inpcb->in6p_route.ro_rt ?
866  					    tp->t_inpcb->in6p_route.ro_rt->rt_ifp
867  					    : NULL);
868  	else
869 #endif /* INET6 */
870 	ip->ip_ttl = tp->t_inpcb->inp_ip_ttl;	/* XXX */
871 	ip->ip_tos = tp->t_inpcb->inp_ip_tos;	/* XXX */
872 	/*
873 	 * See if we should do MTU discovery.  We do it only if the following
874 	 * are true:
875 	 *	1) we have a valid route to the destination
876 	 *	2) the MTU is not locked (if it is, then discovery has been
877 	 *	   disabled)
878 	 */
879 	if (path_mtu_discovery
880 	    && (rt = tp->t_inpcb->inp_route.ro_rt)
881 	    && rt->rt_flags & RTF_UP
882 	    && !(rt->rt_rmx.rmx_locks & RTV_MTU)) {
883 		ip->ip_off |= IP_DF;
884 	}
885 #ifdef IPSEC
886  	ipsec_setsocket(m, so);
887 #endif /*IPSEC*/
888 	error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
889 	    (so->so_options & SO_DONTROUTE), 0);
890     }
891 	if (error) {
892 
893 		/*
894 		 * We know that the packet was lost, so back out the
895 		 * sequence number advance, if any.
896 		 */
897 		if (tp->t_force == 0 || !callout_active(tp->tt_persist)) {
898 			/*
899 			 * No need to check for TH_FIN here because
900 			 * the TF_SENTFIN flag handles that case.
901 			 */
902 			if ((flags & TH_SYN) == 0)
903 				tp->snd_nxt -= len;
904 		}
905 
906 out:
907 		if (error == ENOBUFS) {
908 	                if (!callout_active(tp->tt_rexmt) &&
909                             !callout_active(tp->tt_persist))
910 	                        callout_reset(tp->tt_rexmt, tp->t_rxtcur,
911                                       tcp_timer_rexmt, tp);
912 			tcp_quench(tp->t_inpcb, 0);
913 			return (0);
914 		}
915 		if (error == EMSGSIZE) {
916 			/*
917 			 * ip_output() will have already fixed the route
918 			 * for us.  tcp_mtudisc() will, as its last action,
919 			 * initiate retransmission, so it is important to
920 			 * not do so here.
921 			 */
922 			tcp_mtudisc(tp->t_inpcb, 0);
923 			return 0;
924 		}
925 		if ((error == EHOSTUNREACH || error == ENETDOWN)
926 		    && TCPS_HAVERCVDSYN(tp->t_state)) {
927 			tp->t_softerror = error;
928 			return (0);
929 		}
930 		return (error);
931 	}
932 	tcpstat.tcps_sndtotal++;
933 
934 	/*
935 	 * Data sent (as far as we can tell).
936 	 * If this advertises a larger window than any other segment,
937 	 * then remember the size of the advertised window.
938 	 * Any pending ACK has now been sent.
939 	 */
940 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
941 		tp->rcv_adv = tp->rcv_nxt + win;
942 	tp->last_ack_sent = tp->rcv_nxt;
943 	tp->t_flags &= ~TF_ACKNOW;
944 	if (tcp_delack_enabled)
945 		callout_stop(tp->tt_delack);
946 #if 0
947 	/*
948 	 * This completely breaks TCP if newreno is turned on.  What happens
949 	 * is that if delayed-acks are turned on on the receiver, this code
950 	 * on the transmitter effectively destroys the TCP window, forcing
951 	 * it to four packets (1.5Kx4 = 6K window).
952 	 */
953 	if (sendalot && (!tcp_do_newreno || --maxburst))
954 		goto again;
955 #endif
956 	if (sendalot)
957 		goto again;
958 	return (0);
959 }
960 
961 void
962 tcp_setpersist(tp)
963 	register struct tcpcb *tp;
964 {
965 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
966 	int tt;
967 
968 	if (callout_active(tp->tt_rexmt))
969 		panic("tcp_setpersist: retransmit pending");
970 	/*
971 	 * Start/restart persistance timer.
972 	 */
973 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
974 		      TCPTV_PERSMIN, TCPTV_PERSMAX);
975 	callout_reset(tp->tt_persist, tt, tcp_timer_persist, tp);
976 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
977 		tp->t_rxtshift++;
978 }
979