xref: /freebsd/sys/netinet/tcp_output.c (revision 6af83ee0d2941d18880b6aaa2b4facd1d30c6106)
1 /*-
2  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  * 4. Neither the name of the University nor the names of its contributors
14  *    may be used to endorse or promote products derived from this software
15  *    without specific prior written permission.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  *
29  *	@(#)tcp_output.c	8.4 (Berkeley) 5/24/95
30  * $FreeBSD$
31  */
32 
33 #include "opt_inet.h"
34 #include "opt_inet6.h"
35 #include "opt_ipsec.h"
36 #include "opt_mac.h"
37 #include "opt_tcpdebug.h"
38 #include "opt_tcp_sack.h"
39 
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/domain.h>
43 #include <sys/kernel.h>
44 #include <sys/lock.h>
45 #include <sys/mac.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 #ifdef FAST_IPSEC
81 #include <netipsec/ipsec.h>
82 #define	IPSEC
83 #endif /*FAST_IPSEC*/
84 
85 #include <machine/in_cksum.h>
86 
87 #ifdef notyet
88 extern struct mbuf *m_copypack();
89 #endif
90 
91 int path_mtu_discovery = 1;
92 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
93 	&path_mtu_discovery, 1, "Enable Path MTU Discovery");
94 
95 int ss_fltsz = 1;
96 SYSCTL_INT(_net_inet_tcp, OID_AUTO, slowstart_flightsize, CTLFLAG_RW,
97 	&ss_fltsz, 1, "Slow start flight size");
98 
99 int ss_fltsz_local = 4;
100 SYSCTL_INT(_net_inet_tcp, OID_AUTO, local_slowstart_flightsize, CTLFLAG_RW,
101 	&ss_fltsz_local, 1, "Slow start flight size for local networks");
102 
103 int     tcp_do_newreno = 1;
104 SYSCTL_INT(_net_inet_tcp, OID_AUTO, newreno, CTLFLAG_RW, &tcp_do_newreno,
105 	0, "Enable NewReno Algorithms");
106 
107 /*
108  * Tcp output routine: figure out what should be sent and send it.
109  */
110 int
111 tcp_output(struct tcpcb *tp)
112 {
113 	struct socket *so = tp->t_inpcb->inp_socket;
114 	long len, recwin, sendwin;
115 	int off, flags, error;
116 #ifdef TCP_SIGNATURE
117 	int sigoff = 0;
118 #endif
119 	struct mbuf *m;
120 	struct ip *ip = NULL;
121 	struct ipovly *ipov = NULL;
122 	struct tcphdr *th;
123 	u_char opt[TCP_MAXOLEN];
124 	unsigned ipoptlen, optlen, hdrlen;
125 	int idle, sendalot;
126 	int i, sack_rxmit;
127 	int sack_bytes_rxmt;
128 	struct sackhole *p;
129 #if 0
130 	int maxburst = TCP_MAXBURST;
131 #endif
132 #ifdef INET6
133 	struct ip6_hdr *ip6 = NULL;
134 	int isipv6;
135 
136 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
137 #endif
138 
139 	INP_LOCK_ASSERT(tp->t_inpcb);
140 
141 	/*
142 	 * Determine length of data that should be transmitted,
143 	 * and flags that will be used.
144 	 * If there is some data or critical controls (SYN, RST)
145 	 * to send, then transmit; otherwise, investigate further.
146 	 */
147 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
148 	if (idle && (ticks - tp->t_rcvtime) >= tp->t_rxtcur) {
149 		/*
150 		 * We have been idle for "a while" and no acks are
151 		 * expected to clock out any data we send --
152 		 * slow start to get ack "clock" running again.
153 		 *
154 		 * Set the slow-start flight size depending on whether
155 		 * this is a local network or not.
156 		 */
157 		int ss = ss_fltsz;
158 #ifdef INET6
159 		if (isipv6) {
160 			if (in6_localaddr(&tp->t_inpcb->in6p_faddr))
161 				ss = ss_fltsz_local;
162 		} else
163 #endif /* INET6 */
164 		if (in_localaddr(tp->t_inpcb->inp_faddr))
165 			ss = ss_fltsz_local;
166 		tp->snd_cwnd = tp->t_maxseg * ss;
167 	}
168 	tp->t_flags &= ~TF_LASTIDLE;
169 	if (idle) {
170 		if (tp->t_flags & TF_MORETOCOME) {
171 			tp->t_flags |= TF_LASTIDLE;
172 			idle = 0;
173 		}
174 	}
175 again:
176 	/*
177 	 * If we've recently taken a timeout, snd_max will be greater than
178 	 * snd_nxt.  There may be SACK information that allows us to avoid
179 	 * resending already delivered data.  Adjust snd_nxt accordingly.
180 	 */
181 	if (tp->sack_enable && SEQ_LT(tp->snd_nxt, tp->snd_max))
182 		tcp_sack_adjust(tp);
183 	sendalot = 0;
184 	off = tp->snd_nxt - tp->snd_una;
185 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
186 	sendwin = min(sendwin, tp->snd_bwnd);
187 
188 	flags = tcp_outflags[tp->t_state];
189 	/*
190 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
191 	 * to send out new data (when sendalot is 1), bypass this function.
192 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
193 	 * we're replacing a (future) new transmission with a retransmission
194 	 * now, and we previously incremented snd_cwnd in tcp_input().
195 	 */
196 	/*
197 	 * Still in sack recovery , reset rxmit flag to zero.
198 	 */
199 	sack_rxmit = 0;
200 	sack_bytes_rxmt = 0;
201 	len = 0;
202 	p = NULL;
203 	if (tp->sack_enable && IN_FASTRECOVERY(tp) &&
204 	    (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
205 		long cwin;
206 
207 		cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
208 		if (cwin < 0)
209 			cwin = 0;
210 		/* Do not retransmit SACK segments beyond snd_recover */
211 		if (SEQ_GT(p->end, tp->snd_recover)) {
212 			/*
213 			 * (At least) part of sack hole extends beyond
214 			 * snd_recover. Check to see if we can rexmit data
215 			 * for this hole.
216 			 */
217 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
218 				/*
219 				 * Can't rexmit any more data for this hole.
220 				 * That data will be rexmitted in the next
221 				 * sack recovery episode, when snd_recover
222 				 * moves past p->rxmit.
223 				 */
224 				p = NULL;
225 				goto after_sack_rexmit;
226 			} else
227 				/* Can rexmit part of the current hole */
228 				len = ((long)ulmin(cwin,
229 						   tp->snd_recover - p->rxmit));
230 		} else
231 			len = ((long)ulmin(cwin, p->end - p->rxmit));
232 		off = p->rxmit - tp->snd_una;
233 		KASSERT(off >= 0,("%s: sack block to the left of una : %d",
234 		    __func__, off));
235 		if (len > 0) {
236 			sack_rxmit = 1;
237 			sendalot = 1;
238 			tcpstat.tcps_sack_rexmits++;
239 			tcpstat.tcps_sack_rexmit_bytes +=
240 			    min(len, tp->t_maxseg);
241 		}
242 	}
243 after_sack_rexmit:
244 	/*
245 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
246 	 * state flags.
247 	 */
248 	if (tp->t_flags & TF_NEEDFIN)
249 		flags |= TH_FIN;
250 	if (tp->t_flags & TF_NEEDSYN)
251 		flags |= TH_SYN;
252 
253 	SOCKBUF_LOCK(&so->so_snd);
254 	/*
255 	 * If in persist timeout with window of 0, send 1 byte.
256 	 * Otherwise, if window is small but nonzero
257 	 * and timer expired, we will send what we can
258 	 * and go to transmit state.
259 	 */
260 	if (tp->t_force) {
261 		if (sendwin == 0) {
262 			/*
263 			 * If we still have some data to send, then
264 			 * clear the FIN bit.  Usually this would
265 			 * happen below when it realizes that we
266 			 * aren't sending all the data.  However,
267 			 * if we have exactly 1 byte of unsent data,
268 			 * then it won't clear the FIN bit below,
269 			 * and if we are in persist state, we wind
270 			 * up sending the packet without recording
271 			 * that we sent the FIN bit.
272 			 *
273 			 * We can't just blindly clear the FIN bit,
274 			 * because if we don't have any more data
275 			 * to send then the probe will be the FIN
276 			 * itself.
277 			 */
278 			if (off < so->so_snd.sb_cc)
279 				flags &= ~TH_FIN;
280 			sendwin = 1;
281 		} else {
282 			callout_stop(tp->tt_persist);
283 			tp->t_rxtshift = 0;
284 		}
285 	}
286 
287 	/*
288 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
289 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
290 	 * a negative length.  This can also occur when TCP opens up
291 	 * its congestion window while receiving additional duplicate
292 	 * acks after fast-retransmit because TCP will reset snd_nxt
293 	 * to snd_max after the fast-retransmit.
294 	 *
295 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
296 	 * be set to snd_una, the offset will be 0, and the length may
297 	 * wind up 0.
298 	 *
299 	 * If sack_rxmit is true we are retransmitting from the scoreboard
300 	 * in which case len is already set.
301 	 */
302 	if (sack_rxmit == 0) {
303 		if (sack_bytes_rxmt == 0)
304 			len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off);
305 		else {
306 			long cwin;
307 
308                         /*
309 			 * We are inside of a SACK recovery episode and are
310 			 * sending new data, having retransmitted all the
311 			 * data possible in the scoreboard.
312 			 */
313 			len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd)
314 			       - off);
315 			/*
316 			 * Don't remove this (len > 0) check !
317 			 * We explicitly check for len > 0 here (although it
318 			 * isn't really necessary), to work around a gcc
319 			 * optimization issue - to force gcc to compute
320 			 * len above. Without this check, the computation
321 			 * of len is bungled by the optimizer.
322 			 */
323 			if (len > 0) {
324 				cwin = tp->snd_cwnd -
325 					(tp->snd_nxt - tp->sack_newdata) -
326 					sack_bytes_rxmt;
327 				if (cwin < 0)
328 					cwin = 0;
329 				len = lmin(len, cwin);
330 			}
331 		}
332 	}
333 
334 	/*
335 	 * Lop off SYN bit if it has already been sent.  However, if this
336 	 * is SYN-SENT state and if segment contains data and if we don't
337 	 * know that foreign host supports TAO, suppress sending segment.
338 	 */
339 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
340 		flags &= ~TH_SYN;
341 		off--, len++;
342 	}
343 
344 	/*
345 	 * Be careful not to send data and/or FIN on SYN segments.
346 	 * This measure is needed to prevent interoperability problems
347 	 * with not fully conformant TCP implementations.
348 	 */
349 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
350 		len = 0;
351 		flags &= ~TH_FIN;
352 	}
353 
354 	if (len < 0) {
355 		/*
356 		 * If FIN has been sent but not acked,
357 		 * but we haven't been called to retransmit,
358 		 * len will be < 0.  Otherwise, window shrank
359 		 * after we sent into it.  If window shrank to 0,
360 		 * cancel pending retransmit, pull snd_nxt back
361 		 * to (closed) window, and set the persist timer
362 		 * if it isn't already going.  If the window didn't
363 		 * close completely, just wait for an ACK.
364 		 */
365 		len = 0;
366 		if (sendwin == 0) {
367 			callout_stop(tp->tt_rexmt);
368 			tp->t_rxtshift = 0;
369 			tp->snd_nxt = tp->snd_una;
370 			if (!callout_active(tp->tt_persist))
371 				tcp_setpersist(tp);
372 		}
373 	}
374 
375 	/*
376 	 * len will be >= 0 after this point.  Truncate to the maximum
377 	 * segment length and ensure that FIN is removed if the length
378 	 * no longer contains the last data byte.
379 	 */
380 	if (len > tp->t_maxseg) {
381 		len = tp->t_maxseg;
382 		sendalot = 1;
383 	}
384 	if (sack_rxmit) {
385 		if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
386 			flags &= ~TH_FIN;
387 	} else {
388 		if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
389 			flags &= ~TH_FIN;
390 	}
391 
392 	recwin = sbspace(&so->so_rcv);
393 
394 	/*
395 	 * Sender silly window avoidance.   We transmit under the following
396 	 * conditions when len is non-zero:
397 	 *
398 	 *	- We have a full segment
399 	 *	- This is the last buffer in a write()/send() and we are
400 	 *	  either idle or running NODELAY
401 	 *	- we've timed out (e.g. persist timer)
402 	 *	- we have more then 1/2 the maximum send window's worth of
403 	 *	  data (receiver may be limited the window size)
404 	 *	- we need to retransmit
405 	 */
406 	if (len) {
407 		if (len == tp->t_maxseg)
408 			goto send;
409 		/*
410 		 * NOTE! on localhost connections an 'ack' from the remote
411 		 * end may occur synchronously with the output and cause
412 		 * us to flush a buffer queued with moretocome.  XXX
413 		 *
414 		 * note: the len + off check is almost certainly unnecessary.
415 		 */
416 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
417 		    (idle || (tp->t_flags & TF_NODELAY)) &&
418 		    len + off >= so->so_snd.sb_cc &&
419 		    (tp->t_flags & TF_NOPUSH) == 0) {
420 			goto send;
421 		}
422 		if (tp->t_force)			/* typ. timeout case */
423 			goto send;
424 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
425 			goto send;
426 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
427 			goto send;
428 		if (sack_rxmit)
429 			goto send;
430 	}
431 
432 	/*
433 	 * Compare available window to amount of window
434 	 * known to peer (as advertised window less
435 	 * next expected input).  If the difference is at least two
436 	 * max size segments, or at least 50% of the maximum possible
437 	 * window, then want to send a window update to peer.
438 	 * Skip this if the connection is in T/TCP half-open state.
439 	 */
440 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN)) {
441 		/*
442 		 * "adv" is the amount we can increase the window,
443 		 * taking into account that we are limited by
444 		 * TCP_MAXWIN << tp->rcv_scale.
445 		 */
446 		long adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale) -
447 			(tp->rcv_adv - tp->rcv_nxt);
448 
449 		if (adv >= (long) (2 * tp->t_maxseg))
450 			goto send;
451 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
452 			goto send;
453 	}
454 
455 	/*
456 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
457 	 * is also a catch-all for the retransmit timer timeout case.
458 	 */
459 	if (tp->t_flags & TF_ACKNOW)
460 		goto send;
461 	if ((flags & TH_RST) ||
462 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
463 		goto send;
464 	if (SEQ_GT(tp->snd_up, tp->snd_una))
465 		goto send;
466 	/*
467 	 * If our state indicates that FIN should be sent
468 	 * and we have not yet done so, then we need to send.
469 	 */
470 	if (flags & TH_FIN &&
471 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
472 		goto send;
473 	/*
474 	 * In SACK, it is possible for tcp_output to fail to send a segment
475 	 * after the retransmission timer has been turned off.  Make sure
476 	 * that the retransmission timer is set.
477 	 */
478 	if (tp->sack_enable && SEQ_GT(tp->snd_max, tp->snd_una) &&
479 	    !callout_active(tp->tt_rexmt) &&
480 	    !callout_active(tp->tt_persist)) {
481 		callout_reset(tp->tt_rexmt, tp->t_rxtcur,
482 			      tcp_timer_rexmt, tp);
483 		goto just_return;
484 	}
485 	/*
486 	 * TCP window updates are not reliable, rather a polling protocol
487 	 * using ``persist'' packets is used to insure receipt of window
488 	 * updates.  The three ``states'' for the output side are:
489 	 *	idle			not doing retransmits or persists
490 	 *	persisting		to move a small or zero window
491 	 *	(re)transmitting	and thereby not persisting
492 	 *
493 	 * callout_active(tp->tt_persist)
494 	 *	is true when we are in persist state.
495 	 * tp->t_force
496 	 *	is set when we are called to send a persist packet.
497 	 * callout_active(tp->tt_rexmt)
498 	 *	is set when we are retransmitting
499 	 * The output side is idle when both timers are zero.
500 	 *
501 	 * If send window is too small, there is data to transmit, and no
502 	 * retransmit or persist is pending, then go to persist state.
503 	 * If nothing happens soon, send when timer expires:
504 	 * if window is nonzero, transmit what we can,
505 	 * otherwise force out a byte.
506 	 */
507 	if (so->so_snd.sb_cc && !callout_active(tp->tt_rexmt) &&
508 	    !callout_active(tp->tt_persist)) {
509 		tp->t_rxtshift = 0;
510 		tcp_setpersist(tp);
511 	}
512 
513 	/*
514 	 * No reason to send a segment, just return.
515 	 */
516 just_return:
517 	SOCKBUF_UNLOCK(&so->so_snd);
518 	return (0);
519 
520 send:
521 	SOCKBUF_LOCK_ASSERT(&so->so_snd);
522 	/*
523 	 * Before ESTABLISHED, force sending of initial options
524 	 * unless TCP set not to do any options.
525 	 * NOTE: we assume that the IP/TCP header plus TCP options
526 	 * always fit in a single mbuf, leaving room for a maximum
527 	 * link header, i.e.
528 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
529 	 */
530 	optlen = 0;
531 #ifdef INET6
532 	if (isipv6)
533 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
534 	else
535 #endif
536 	hdrlen = sizeof (struct tcpiphdr);
537 	if (flags & TH_SYN) {
538 		tp->snd_nxt = tp->iss;
539 		if ((tp->t_flags & TF_NOOPT) == 0) {
540 			u_short mss;
541 
542 			opt[0] = TCPOPT_MAXSEG;
543 			opt[1] = TCPOLEN_MAXSEG;
544 			mss = htons((u_short) tcp_mssopt(&tp->t_inpcb->inp_inc));
545 			(void)memcpy(opt + 2, &mss, sizeof(mss));
546 			optlen = TCPOLEN_MAXSEG;
547 
548 			/*
549 			 * If this is the first SYN of connection (not a SYN
550 			 * ACK), include SACK_PERMIT_HDR option.  If this is a
551 			 * SYN ACK, include SACK_PERMIT_HDR option if peer has
552 			 * already done so. This is only for active connect,
553 			 * since the syncache takes care of the passive connect.
554 			 */
555 			if (tp->sack_enable && ((flags & TH_ACK) == 0 ||
556 			    (tp->t_flags & TF_SACK_PERMIT))) {
557 				*((u_int32_t *) (opt + optlen)) =
558 					htonl(TCPOPT_SACK_PERMIT_HDR);
559 				optlen += 4;
560 			}
561 			if ((tp->t_flags & TF_REQ_SCALE) &&
562 			    ((flags & TH_ACK) == 0 ||
563 			    (tp->t_flags & TF_RCVD_SCALE))) {
564 				*((u_int32_t *)(opt + optlen)) = htonl(
565 					TCPOPT_NOP << 24 |
566 					TCPOPT_WINDOW << 16 |
567 					TCPOLEN_WINDOW << 8 |
568 					tp->request_r_scale);
569 				optlen += 4;
570 			}
571 		}
572 	}
573 
574 	/*
575 	 * Send a timestamp and echo-reply if this is a SYN and our side
576 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
577 	 * and our peer have sent timestamps in our SYN's.
578 	 */
579 	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
580 	    (flags & TH_RST) == 0 &&
581 	    ((flags & TH_ACK) == 0 ||
582 	     (tp->t_flags & TF_RCVD_TSTMP))) {
583 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
584 
585 		/* Form timestamp option as shown in appendix A of RFC 1323. */
586 		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
587 		*lp++ = htonl(ticks);
588 		*lp   = htonl(tp->ts_recent);
589 		optlen += TCPOLEN_TSTAMP_APPA;
590 	}
591 
592 	/*
593 	 * Send SACKs if necessary.  This should be the last option processed.
594 	 * Only as many SACKs are sent as are permitted by the maximum options
595 	 * size.  No more than three SACKs are sent.
596 	 */
597 	if (tp->sack_enable && tp->t_state == TCPS_ESTABLISHED &&
598 	    (tp->t_flags & (TF_SACK_PERMIT|TF_NOOPT)) == TF_SACK_PERMIT &&
599 	    tp->rcv_numsacks) {
600 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
601 		u_int32_t *olp = lp++;
602 		int count = 0;  /* actual number of SACKs inserted */
603 		int maxsack = (MAX_TCPOPTLEN - (optlen + 4))/TCPOLEN_SACK;
604 
605 		tcpstat.tcps_sack_send_blocks++;
606 		maxsack = min(maxsack, TCP_MAX_SACK);
607 		for (i = 0; (i < tp->rcv_numsacks && count < maxsack); i++) {
608 			struct sackblk sack = tp->sackblks[i];
609 			if (sack.start == 0 && sack.end == 0)
610 				continue;
611 			*lp++ = htonl(sack.start);
612 			*lp++ = htonl(sack.end);
613 			count++;
614 		}
615 		*olp = htonl(TCPOPT_SACK_HDR|(TCPOLEN_SACK*count+2));
616 		optlen += TCPOLEN_SACK*count + 4; /* including leading NOPs */
617 	}
618 
619 #ifdef TCP_SIGNATURE
620 #ifdef INET6
621 	if (!isipv6)
622 #endif
623 	if (tp->t_flags & TF_SIGNATURE) {
624 		int i;
625 		u_char *bp;
626 
627 		/* Initialize TCP-MD5 option (RFC2385) */
628 		bp = (u_char *)opt + optlen;
629 		*bp++ = TCPOPT_SIGNATURE;
630 		*bp++ = TCPOLEN_SIGNATURE;
631 		sigoff = optlen + 2;
632 		for (i = 0; i < TCP_SIGLEN; i++)
633 			*bp++ = 0;
634 		optlen += TCPOLEN_SIGNATURE;
635 
636 		/* Terminate options list and maintain 32-bit alignment. */
637 		*bp++ = TCPOPT_NOP;
638 		*bp++ = TCPOPT_EOL;
639 		optlen += 2;
640 	}
641 #endif /* TCP_SIGNATURE */
642 
643 	hdrlen += optlen;
644 
645 #ifdef INET6
646 	if (isipv6)
647 		ipoptlen = ip6_optlen(tp->t_inpcb);
648 	else
649 #endif
650 	if (tp->t_inpcb->inp_options)
651 		ipoptlen = tp->t_inpcb->inp_options->m_len -
652 				offsetof(struct ipoption, ipopt_list);
653 	else
654 		ipoptlen = 0;
655 #ifdef IPSEC
656 	ipoptlen += ipsec_hdrsiz_tcp(tp);
657 #endif
658 
659 	/*
660 	 * Adjust data length if insertion of options will
661 	 * bump the packet length beyond the t_maxopd length.
662 	 * Clear the FIN bit because we cut off the tail of
663 	 * the segment.
664 	 */
665 	if (len + optlen + ipoptlen > tp->t_maxopd) {
666 		/*
667 		 * If there is still more to send, don't close the connection.
668 		 */
669 		flags &= ~TH_FIN;
670 		len = tp->t_maxopd - optlen - ipoptlen;
671 		sendalot = 1;
672 	}
673 
674 /*#ifdef DIAGNOSTIC*/
675 #ifdef INET6
676 	if (max_linkhdr + hdrlen > MCLBYTES)
677 #else
678 	if (max_linkhdr + hdrlen > MHLEN)
679 #endif
680 		panic("tcphdr too big");
681 /*#endif*/
682 
683 	/*
684 	 * Grab a header mbuf, attaching a copy of data to
685 	 * be transmitted, and initialize the header from
686 	 * the template for sends on this connection.
687 	 */
688 	if (len) {
689 		if (tp->t_force && len == 1)
690 			tcpstat.tcps_sndprobe++;
691 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
692 			tcpstat.tcps_sndrexmitpack++;
693 			tcpstat.tcps_sndrexmitbyte += len;
694 		} else {
695 			tcpstat.tcps_sndpack++;
696 			tcpstat.tcps_sndbyte += len;
697 		}
698 #ifdef notyet
699 		if ((m = m_copypack(so->so_snd.sb_mb, off,
700 		    (int)len, max_linkhdr + hdrlen)) == 0) {
701 			SOCKBUF_UNLOCK(&so->so_snd);
702 			error = ENOBUFS;
703 			goto out;
704 		}
705 		/*
706 		 * m_copypack left space for our hdr; use it.
707 		 */
708 		m->m_len += hdrlen;
709 		m->m_data -= hdrlen;
710 #else
711 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
712 		if (m == NULL) {
713 			SOCKBUF_UNLOCK(&so->so_snd);
714 			error = ENOBUFS;
715 			goto out;
716 		}
717 #ifdef INET6
718 		if (MHLEN < hdrlen + max_linkhdr) {
719 			MCLGET(m, M_DONTWAIT);
720 			if ((m->m_flags & M_EXT) == 0) {
721 				SOCKBUF_UNLOCK(&so->so_snd);
722 				m_freem(m);
723 				error = ENOBUFS;
724 				goto out;
725 			}
726 		}
727 #endif
728 		m->m_data += max_linkhdr;
729 		m->m_len = hdrlen;
730 		if (len <= MHLEN - hdrlen - max_linkhdr) {
731 			m_copydata(so->so_snd.sb_mb, off, (int) len,
732 			    mtod(m, caddr_t) + hdrlen);
733 			m->m_len += len;
734 		} else {
735 			m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len);
736 			if (m->m_next == 0) {
737 				SOCKBUF_UNLOCK(&so->so_snd);
738 				(void) m_free(m);
739 				error = ENOBUFS;
740 				goto out;
741 			}
742 		}
743 #endif
744 		/*
745 		 * If we're sending everything we've got, set PUSH.
746 		 * (This will keep happy those implementations which only
747 		 * give data to the user when a buffer fills or
748 		 * a PUSH comes in.)
749 		 */
750 		if (off + len == so->so_snd.sb_cc)
751 			flags |= TH_PUSH;
752 		SOCKBUF_UNLOCK(&so->so_snd);
753 	} else {
754 		SOCKBUF_UNLOCK(&so->so_snd);
755 		if (tp->t_flags & TF_ACKNOW)
756 			tcpstat.tcps_sndacks++;
757 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
758 			tcpstat.tcps_sndctrl++;
759 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
760 			tcpstat.tcps_sndurg++;
761 		else
762 			tcpstat.tcps_sndwinup++;
763 
764 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
765 		if (m == NULL) {
766 			error = ENOBUFS;
767 			goto out;
768 		}
769 #ifdef INET6
770 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
771 		    MHLEN >= hdrlen) {
772 			MH_ALIGN(m, hdrlen);
773 		} else
774 #endif
775 		m->m_data += max_linkhdr;
776 		m->m_len = hdrlen;
777 	}
778 	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
779 	m->m_pkthdr.rcvif = (struct ifnet *)0;
780 #ifdef MAC
781 	mac_create_mbuf_from_inpcb(tp->t_inpcb, m);
782 #endif
783 #ifdef INET6
784 	if (isipv6) {
785 		ip6 = mtod(m, struct ip6_hdr *);
786 		th = (struct tcphdr *)(ip6 + 1);
787 		tcpip_fillheaders(tp->t_inpcb, ip6, th);
788 	} else
789 #endif /* INET6 */
790 	{
791 		ip = mtod(m, struct ip *);
792 		ipov = (struct ipovly *)ip;
793 		th = (struct tcphdr *)(ip + 1);
794 		tcpip_fillheaders(tp->t_inpcb, ip, th);
795 	}
796 
797 	/*
798 	 * Fill in fields, remembering maximum advertised
799 	 * window for use in delaying messages about window sizes.
800 	 * If resending a FIN, be sure not to use a new sequence number.
801 	 */
802 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
803 	    tp->snd_nxt == tp->snd_max)
804 		tp->snd_nxt--;
805 	/*
806 	 * If we are doing retransmissions, then snd_nxt will
807 	 * not reflect the first unsent octet.  For ACK only
808 	 * packets, we do not want the sequence number of the
809 	 * retransmitted packet, we want the sequence number
810 	 * of the next unsent octet.  So, if there is no data
811 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
812 	 * when filling in ti_seq.  But if we are in persist
813 	 * state, snd_max might reflect one byte beyond the
814 	 * right edge of the window, so use snd_nxt in that
815 	 * case, since we know we aren't doing a retransmission.
816 	 * (retransmit and persist are mutually exclusive...)
817 	 */
818 	if (sack_rxmit == 0) {
819 		if (len || (flags & (TH_SYN|TH_FIN))
820 		    || callout_active(tp->tt_persist))
821 			th->th_seq = htonl(tp->snd_nxt);
822 		else
823 			th->th_seq = htonl(tp->snd_max);
824 	} else {
825 		th->th_seq = htonl(p->rxmit);
826 		p->rxmit += len;
827 	}
828 	th->th_ack = htonl(tp->rcv_nxt);
829 	if (optlen) {
830 		bcopy(opt, th + 1, optlen);
831 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
832 	}
833 	th->th_flags = flags;
834 	/*
835 	 * Calculate receive window.  Don't shrink window,
836 	 * but avoid silly window syndrome.
837 	 */
838 	if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
839 	    recwin < (long)tp->t_maxseg)
840 		recwin = 0;
841 	if (recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
842 		recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
843 	if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
844 		recwin = (long)TCP_MAXWIN << tp->rcv_scale;
845 	th->th_win = htons((u_short) (recwin >> tp->rcv_scale));
846 
847 
848 	/*
849 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
850 	 * a 0 window.  This may cause the remote transmitter to stall.  This
851 	 * flag tells soreceive() to disable delayed acknowledgements when
852 	 * draining the buffer.  This can occur if the receiver is attempting
853 	 * to read more data then can be buffered prior to transmitting on
854 	 * the connection.
855 	 */
856 	if (recwin == 0)
857 		tp->t_flags |= TF_RXWIN0SENT;
858 	else
859 		tp->t_flags &= ~TF_RXWIN0SENT;
860 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
861 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
862 		th->th_flags |= TH_URG;
863 	} else
864 		/*
865 		 * If no urgent pointer to send, then we pull
866 		 * the urgent pointer to the left edge of the send window
867 		 * so that it doesn't drift into the send window on sequence
868 		 * number wraparound.
869 		 */
870 		tp->snd_up = tp->snd_una;		/* drag it along */
871 
872 #ifdef TCP_SIGNATURE
873 #ifdef INET6
874 	if (!isipv6)
875 #endif
876 	if (tp->t_flags & TF_SIGNATURE)
877 		tcp_signature_compute(m, sizeof(struct ip), len, optlen,
878 		    (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
879 #endif
880 
881 	/*
882 	 * Put TCP length in extended header, and then
883 	 * checksum extended header and data.
884 	 */
885 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
886 #ifdef INET6
887 	if (isipv6)
888 		/*
889 		 * ip6_plen is not need to be filled now, and will be filled
890 		 * in ip6_output.
891 		 */
892 		th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
893 				       sizeof(struct tcphdr) + optlen + len);
894 	else
895 #endif /* INET6 */
896 	{
897 		m->m_pkthdr.csum_flags = CSUM_TCP;
898 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
899 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
900 		    htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
901 
902 		/* IP version must be set here for ipv4/ipv6 checking later */
903 		KASSERT(ip->ip_v == IPVERSION,
904 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
905 	}
906 
907 	/*
908 	 * In transmit state, time the transmission and arrange for
909 	 * the retransmit.  In persist state, just set snd_max.
910 	 */
911 	if (tp->t_force == 0 || !callout_active(tp->tt_persist)) {
912 		tcp_seq startseq = tp->snd_nxt;
913 
914 		/*
915 		 * Advance snd_nxt over sequence space of this segment.
916 		 */
917 		if (flags & (TH_SYN|TH_FIN)) {
918 			if (flags & TH_SYN)
919 				tp->snd_nxt++;
920 			if (flags & TH_FIN) {
921 				tp->snd_nxt++;
922 				tp->t_flags |= TF_SENTFIN;
923 			}
924 		}
925 		if (sack_rxmit)
926 			goto timer;
927 		tp->snd_nxt += len;
928 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
929 			tp->snd_max = tp->snd_nxt;
930 			/*
931 			 * Time this transmission if not a retransmission and
932 			 * not currently timing anything.
933 			 */
934 			if (tp->t_rtttime == 0) {
935 				tp->t_rtttime = ticks;
936 				tp->t_rtseq = startseq;
937 				tcpstat.tcps_segstimed++;
938 			}
939 		}
940 
941 		/*
942 		 * Set retransmit timer if not currently set,
943 		 * and not doing a pure ack or a keep-alive probe.
944 		 * Initial value for retransmit timer is smoothed
945 		 * round-trip time + 2 * round-trip time variance.
946 		 * Initialize shift counter which is used for backoff
947 		 * of retransmit time.
948 		 */
949 timer:
950 		if (!callout_active(tp->tt_rexmt) &&
951 		    ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
952 		     (tp->snd_nxt != tp->snd_una))) {
953 			if (callout_active(tp->tt_persist)) {
954 				callout_stop(tp->tt_persist);
955 				tp->t_rxtshift = 0;
956 			}
957 			callout_reset(tp->tt_rexmt, tp->t_rxtcur,
958 				      tcp_timer_rexmt, tp);
959 		}
960 	} else {
961 		/*
962 		 * Persist case, update snd_max but since we are in
963 		 * persist mode (no window) we do not update snd_nxt.
964 		 */
965 		int xlen = len;
966 		if (flags & TH_SYN)
967 			++xlen;
968 		if (flags & TH_FIN) {
969 			++xlen;
970 			tp->t_flags |= TF_SENTFIN;
971 		}
972 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
973 			tp->snd_max = tp->snd_nxt + len;
974 	}
975 
976 #ifdef TCPDEBUG
977 	/*
978 	 * Trace.
979 	 */
980 	if (so->so_options & SO_DEBUG) {
981 		u_short save = 0;
982 #ifdef INET6
983 		if (!isipv6)
984 #endif
985 		{
986 			save = ipov->ih_len;
987 			ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
988 		}
989 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
990 #ifdef INET6
991 		if (!isipv6)
992 #endif
993 		ipov->ih_len = save;
994 	}
995 #endif
996 
997 	/*
998 	 * Fill in IP length and desired time to live and
999 	 * send to IP level.  There should be a better way
1000 	 * to handle ttl and tos; we could keep them in
1001 	 * the template, but need a way to checksum without them.
1002 	 */
1003 	/*
1004 	 * m->m_pkthdr.len should have been set before cksum calcuration,
1005 	 * because in6_cksum() need it.
1006 	 */
1007 #ifdef INET6
1008 	if (isipv6) {
1009 		/*
1010 		 * we separately set hoplimit for every segment, since the
1011 		 * user might want to change the value via setsockopt.
1012 		 * Also, desired default hop limit might be changed via
1013 		 * Neighbor Discovery.
1014 		 */
1015 		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1016 
1017 		/* TODO: IPv6 IP6TOS_ECT bit on */
1018 		error = ip6_output(m,
1019 			    tp->t_inpcb->in6p_outputopts, NULL,
1020 			    ((so->so_options & SO_DONTROUTE) ?
1021 			    IP_ROUTETOIF : 0), NULL, NULL, tp->t_inpcb);
1022 	} else
1023 #endif /* INET6 */
1024     {
1025 	ip->ip_len = m->m_pkthdr.len;
1026 #ifdef INET6
1027 	if (INP_CHECK_SOCKAF(so, AF_INET6))
1028 		ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1029 #endif /* INET6 */
1030 	/*
1031 	 * If we do path MTU discovery, then we set DF on every packet.
1032 	 * This might not be the best thing to do according to RFC3390
1033 	 * Section 2. However the tcp hostcache migitates the problem
1034 	 * so it affects only the first tcp connection with a host.
1035 	 */
1036 	if (path_mtu_discovery)
1037 		ip->ip_off |= IP_DF;
1038 
1039 	error = ip_output(m, tp->t_inpcb->inp_options, NULL,
1040 	    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1041 	    tp->t_inpcb);
1042     }
1043 	if (error) {
1044 
1045 		/*
1046 		 * We know that the packet was lost, so back out the
1047 		 * sequence number advance, if any.
1048 		 */
1049 		if (tp->t_force == 0 || !callout_active(tp->tt_persist)) {
1050 			/*
1051 			 * No need to check for TH_FIN here because
1052 			 * the TF_SENTFIN flag handles that case.
1053 			 */
1054 			if ((flags & TH_SYN) == 0) {
1055 				if (sack_rxmit)
1056 					p->rxmit -= len;
1057 				else
1058 					tp->snd_nxt -= len;
1059 			}
1060 		}
1061 
1062 out:
1063 		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1064 		if (error == ENOBUFS) {
1065 	                if (!callout_active(tp->tt_rexmt) &&
1066 			    !callout_active(tp->tt_persist))
1067 	                        callout_reset(tp->tt_rexmt, tp->t_rxtcur,
1068 				    tcp_timer_rexmt, tp);
1069 			tcp_quench(tp->t_inpcb, 0);
1070 			return (0);
1071 		}
1072 		if (error == EMSGSIZE) {
1073 			/*
1074 			 * ip_output() will have already fixed the route
1075 			 * for us.  tcp_mtudisc() will, as its last action,
1076 			 * initiate retransmission, so it is important to
1077 			 * not do so here.
1078 			 */
1079 			tcp_mtudisc(tp->t_inpcb, 0);
1080 			return 0;
1081 		}
1082 		if ((error == EHOSTUNREACH || error == ENETDOWN)
1083 		    && TCPS_HAVERCVDSYN(tp->t_state)) {
1084 			tp->t_softerror = error;
1085 			return (0);
1086 		}
1087 		return (error);
1088 	}
1089 	tcpstat.tcps_sndtotal++;
1090 
1091 	/*
1092 	 * Data sent (as far as we can tell).
1093 	 * If this advertises a larger window than any other segment,
1094 	 * then remember the size of the advertised window.
1095 	 * Any pending ACK has now been sent.
1096 	 */
1097 	if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1098 		tp->rcv_adv = tp->rcv_nxt + recwin;
1099 	tp->last_ack_sent = tp->rcv_nxt;
1100 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1101 	if (callout_active(tp->tt_delack))
1102 		callout_stop(tp->tt_delack);
1103 #if 0
1104 	/*
1105 	 * This completely breaks TCP if newreno is turned on.  What happens
1106 	 * is that if delayed-acks are turned on on the receiver, this code
1107 	 * on the transmitter effectively destroys the TCP window, forcing
1108 	 * it to four packets (1.5Kx4 = 6K window).
1109 	 */
1110 	if (sendalot && (!tcp_do_newreno || --maxburst))
1111 		goto again;
1112 #endif
1113 	if (sendalot)
1114 		goto again;
1115 	return (0);
1116 }
1117 
1118 void
1119 tcp_setpersist(tp)
1120 	register struct tcpcb *tp;
1121 {
1122 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1123 	int tt;
1124 
1125 	if (callout_active(tp->tt_rexmt))
1126 		panic("tcp_setpersist: retransmit pending");
1127 	/*
1128 	 * Start/restart persistance timer.
1129 	 */
1130 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1131 		      TCPTV_PERSMIN, TCPTV_PERSMAX);
1132 	callout_reset(tp->tt_persist, tt, tcp_timer_persist, tp);
1133 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1134 		tp->t_rxtshift++;
1135 }
1136