xref: /openbsd/sys/netinet/tcp_output.c (revision 666d181c)
1 /*	$OpenBSD: tcp_output.c,v 1.150 2025/01/01 13:44:22 bluhm Exp $	*/
2 /*	$NetBSD: tcp_output.c,v 1.16 1997/06/03 16:17:09 kml Exp $	*/
3 
4 /*
5  * Copyright (c) 1982, 1986, 1988, 1990, 1993
6  *	The Regents of the University of California.  All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. Neither the name of the University nor the names of its contributors
17  *    may be used to endorse or promote products derived from this software
18  *    without specific prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  *
32  *	@(#)COPYRIGHT	1.1 (NRL) 17 January 1995
33  *
34  * NRL grants permission for redistribution and use in source and binary
35  * forms, with or without modification, of the software and documentation
36  * created at NRL provided that the following conditions are met:
37  *
38  * 1. Redistributions of source code must retain the above copyright
39  *    notice, this list of conditions and the following disclaimer.
40  * 2. Redistributions in binary form must reproduce the above copyright
41  *    notice, this list of conditions and the following disclaimer in the
42  *    documentation and/or other materials provided with the distribution.
43  * 3. All advertising materials mentioning features or use of this software
44  *    must display the following acknowledgements:
45  *	This product includes software developed by the University of
46  *	California, Berkeley and its contributors.
47  *	This product includes software developed at the Information
48  *	Technology Division, US Naval Research Laboratory.
49  * 4. Neither the name of the NRL nor the names of its contributors
50  *    may be used to endorse or promote products derived from this software
51  *    without specific prior written permission.
52  *
53  * THE SOFTWARE PROVIDED BY NRL IS PROVIDED BY NRL AND CONTRIBUTORS ``AS
54  * IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
55  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
56  * PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL NRL OR
57  * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
58  * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
59  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
60  * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
61  * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
62  * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
63  * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
64  *
65  * The views and conclusions contained in the software and documentation
66  * are those of the authors and should not be interpreted as representing
67  * official policies, either expressed or implied, of the US Naval
68  * Research Laboratory (NRL).
69  */
70 
71 #include "pf.h"
72 #include "stoeplitz.h"
73 
74 #include <sys/param.h>
75 #include <sys/systm.h>
76 #include <sys/mbuf.h>
77 #include <sys/protosw.h>
78 #include <sys/socket.h>
79 #include <sys/socketvar.h>
80 #include <sys/kernel.h>
81 
82 #include <net/if.h>
83 #include <net/if_var.h>
84 #include <net/route.h>
85 #if NPF > 0
86 #include <net/pfvar.h>
87 #endif
88 
89 #include <netinet/in.h>
90 #include <netinet/ip.h>
91 #include <netinet/in_pcb.h>
92 #include <netinet/ip_var.h>
93 #include <netinet6/ip6_var.h>
94 #include <netinet/tcp.h>
95 #define	TCPOUTFLAGS
96 #include <netinet/tcp_fsm.h>
97 #include <netinet/tcp_seq.h>
98 #include <netinet/tcp_timer.h>
99 #include <netinet/tcp_var.h>
100 #include <netinet/tcp_debug.h>
101 
102 #ifdef notyet
103 extern struct mbuf *m_copypack();
104 #endif
105 
106 #ifdef TCP_SACK_DEBUG
107 void tcp_print_holes(struct tcpcb *tp);
108 
109 void
110 tcp_print_holes(struct tcpcb *tp)
111 {
112 	struct sackhole *p = tp->snd_holes;
113 	if (p == NULL)
114 		return;
115 	printf("Hole report: start--end dups rxmit\n");
116 	while (p) {
117 		printf("%x--%x d %d r %x\n", p->start, p->end, p->dups,
118 		    p->rxmit);
119 		p = p->next;
120 	}
121 	printf("\n");
122 }
123 #endif /* TCP_SACK_DEBUG */
124 
125 /*
126  * Returns pointer to a sackhole if there are any pending retransmissions;
127  * NULL otherwise.
128  */
129 struct sackhole *
130 tcp_sack_output(struct tcpcb *tp)
131 {
132 	struct sackhole *p;
133 
134 	if (!tp->sack_enable)
135 		return (NULL);
136 	p = tp->snd_holes;
137 	while (p) {
138 		if (p->dups >= tcprexmtthresh && SEQ_LT(p->rxmit, p->end)) {
139 			if (SEQ_LT(p->rxmit, tp->snd_una)) {/* old SACK hole */
140 				p = p->next;
141 				continue;
142 			}
143 #ifdef TCP_SACK_DEBUG
144 			if (p)
145 				tcp_print_holes(tp);
146 #endif
147 			return (p);
148 		}
149 		p = p->next;
150 	}
151 	return (NULL);
152 }
153 
154 /*
155  * After a timeout, the SACK list may be rebuilt.  This SACK information
156  * should be used to avoid retransmitting SACKed data.  This function
157  * traverses the SACK list to see if snd_nxt should be moved forward.
158  */
159 
160 void
161 tcp_sack_adjust(struct tcpcb *tp)
162 {
163 	struct sackhole *cur = tp->snd_holes;
164 	if (cur == NULL)
165 		return; /* No holes */
166 	if (SEQ_GEQ(tp->snd_nxt, tp->rcv_lastsack))
167 		return; /* We're already beyond any SACKed blocks */
168 	/*
169 	 * Two cases for which we want to advance snd_nxt:
170 	 * i) snd_nxt lies between end of one hole and beginning of another
171 	 * ii) snd_nxt lies between end of last hole and rcv_lastsack
172 	 */
173 	while (cur->next) {
174 		if (SEQ_LT(tp->snd_nxt, cur->end))
175 			return;
176 		if (SEQ_GEQ(tp->snd_nxt, cur->next->start))
177 			cur = cur->next;
178 		else {
179 			tp->snd_nxt = cur->next->start;
180 			return;
181 		}
182 	}
183 	if (SEQ_LT(tp->snd_nxt, cur->end))
184 		return;
185 	tp->snd_nxt = tp->rcv_lastsack;
186 	return;
187 }
188 
189 /*
190  * Tcp output routine: figure out what should be sent and send it.
191  */
192 int
193 tcp_output(struct tcpcb *tp)
194 {
195 	struct socket *so = tp->t_inpcb->inp_socket;
196 	long len, win, rcv_hiwat, txmaxseg;
197 	int off, flags, error;
198 	struct mbuf *m;
199 	struct tcphdr *th;
200 	u_int32_t optbuf[howmany(MAX_TCPOPTLEN, sizeof(u_int32_t))];
201 	u_char *opt = (u_char *)optbuf;
202 	unsigned int optlen, hdrlen, packetlen;
203 	int doing_sosend, idle, sendalot = 0;
204 	int i, sack_rxmit = 0;
205 	struct sackhole *p;
206 	uint64_t now;
207 #ifdef TCP_SIGNATURE
208 	unsigned int sigoff;
209 #endif /* TCP_SIGNATURE */
210 #ifdef TCP_ECN
211 	int needect;
212 #endif
213 	int tso;
214 
215 	if (tp->t_flags & TF_BLOCKOUTPUT) {
216 		tp->t_flags |= TF_NEEDOUTPUT;
217 		return (0);
218 	} else
219 		tp->t_flags &= ~TF_NEEDOUTPUT;
220 
221 #if defined(TCP_SIGNATURE) && defined(DIAGNOSTIC)
222 	if (tp->sack_enable && (tp->t_flags & TF_SIGNATURE))
223 		return (EINVAL);
224 #endif /* defined(TCP_SIGNATURE) && defined(DIAGNOSTIC) */
225 
226 	now = tcp_now();
227 
228 	mtx_enter(&so->so_snd.sb_mtx);
229 	doing_sosend = soissending(so);
230 	mtx_leave(&so->so_snd.sb_mtx);
231 
232 	/*
233 	 * Determine length of data that should be transmitted,
234 	 * and flags that will be used.
235 	 * If there is some data or critical controls (SYN, RST)
236 	 * to send, then transmit; otherwise, investigate further.
237 	 */
238 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
239 	if (idle && (now - tp->t_rcvtime) >= tp->t_rxtcur)
240 		/*
241 		 * We have been idle for "a while" and no acks are
242 		 * expected to clock out any data we send --
243 		 * slow start to get ack "clock" running again.
244 		 */
245 		tp->snd_cwnd = 2 * tp->t_maxseg;
246 
247 	/* remember 'idle' for next invocation of tcp_output */
248 	if (idle && doing_sosend) {
249 		tp->t_flags |= TF_LASTIDLE;
250 		idle = 0;
251 	} else
252 		tp->t_flags &= ~TF_LASTIDLE;
253 
254 again:
255 	/*
256 	 * If we've recently taken a timeout, snd_max will be greater than
257 	 * snd_nxt.  There may be SACK information that allows us to avoid
258 	 * resending already delivered data.  Adjust snd_nxt accordingly.
259 	 */
260 	if (tp->sack_enable && SEQ_LT(tp->snd_nxt, tp->snd_max))
261 		tcp_sack_adjust(tp);
262 	off = tp->snd_nxt - tp->snd_una;
263 	win = ulmin(tp->snd_wnd, tp->snd_cwnd);
264 
265 	flags = tcp_outflags[tp->t_state];
266 
267 	/*
268 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
269 	 * to send out new data (when sendalot is 1), bypass this function.
270 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
271 	 * we're replacing a (future) new transmission with a retransmission
272 	 * now, and we previously incremented snd_cwnd in tcp_input().
273 	 */
274 	if (tp->sack_enable && !sendalot) {
275 		if (tp->t_dupacks >= tcprexmtthresh &&
276 		    (p = tcp_sack_output(tp))) {
277 			off = p->rxmit - tp->snd_una;
278 			sack_rxmit = 1;
279 			/* Coalesce holes into a single retransmission */
280 			len = min(tp->t_maxseg, p->end - p->rxmit);
281 			if (SEQ_LT(tp->snd_una, tp->snd_last))
282 				tp->snd_cwnd -= tp->t_maxseg;
283 		}
284 	}
285 
286 	sendalot = 0;
287 	tso = 0;
288 	/*
289 	 * If in persist timeout with window of 0, send 1 byte.
290 	 * Otherwise, if window is small but nonzero
291 	 * and timer expired, we will send what we can
292 	 * and go to transmit state.
293 	 */
294 	if (tp->t_force) {
295 		if (win == 0) {
296 			/*
297 			 * If we still have some data to send, then
298 			 * clear the FIN bit.  Usually this would
299 			 * happen below when it realizes that we
300 			 * aren't sending all the data.  However,
301 			 * if we have exactly 1 byte of unset data,
302 			 * then it won't clear the FIN bit below,
303 			 * and if we are in persist state, we wind
304 			 * up sending the packet without recording
305 			 * that we sent the FIN bit.
306 			 *
307 			 * We can't just blindly clear the FIN bit,
308 			 * because if we don't have any more data
309 			 * to send then the probe will be the FIN
310 			 * itself.
311 			 */
312 			if (off < so->so_snd.sb_cc)
313 				flags &= ~TH_FIN;
314 			win = 1;
315 		} else {
316 			TCP_TIMER_DISARM(tp, TCPT_PERSIST);
317 			tp->t_rxtshift = 0;
318 		}
319 	}
320 
321 	if (!sack_rxmit) {
322 		len = ulmin(so->so_snd.sb_cc, win) - off;
323 	}
324 
325 	if (len < 0) {
326 		/*
327 		 * If FIN has been sent but not acked,
328 		 * but we haven't been called to retransmit,
329 		 * len will be -1.  Otherwise, window shrank
330 		 * after we sent into it.  If window shrank to 0,
331 		 * cancel pending retransmit, pull snd_nxt back
332 		 * to (closed) window, and set the persist timer
333 		 * if it isn't already going.  If the window didn't
334 		 * close completely, just wait for an ACK.
335 		 */
336 		len = 0;
337 		if (win == 0) {
338 			TCP_TIMER_DISARM(tp, TCPT_REXMT);
339 			tp->t_rxtshift = 0;
340 			tp->snd_nxt = tp->snd_una;
341 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0)
342 				tcp_setpersist(tp);
343 		}
344 	}
345 
346 	/*
347 	 * Never send more than half a buffer full.  This insures that we can
348 	 * always keep 2 packets on the wire, no matter what SO_SNDBUF is, and
349 	 * therefore acks will never be delayed unless we run out of data to
350 	 * transmit.
351 	 */
352 	txmaxseg = ulmin(so->so_snd.sb_hiwat / 2, tp->t_maxseg);
353 
354 	if (len > txmaxseg) {
355 		if (atomic_load_int(&tcp_do_tso) &&
356 		    tp->t_inpcb->inp_options == NULL &&
357 		    tp->t_inpcb->inp_outputopts6 == NULL &&
358 #ifdef TCP_SIGNATURE
359 		    ((tp->t_flags & TF_SIGNATURE) == 0) &&
360 #endif
361 		    len >= 2 * tp->t_maxseg &&
362 		    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
363 		    !(flags & (TH_SYN|TH_RST|TH_FIN))) {
364 			tso = 1;
365 			/* avoid small chopped packets */
366 			if (len > (len / tp->t_maxseg) * tp->t_maxseg) {
367 				len = (len / tp->t_maxseg) * tp->t_maxseg;
368 				sendalot = 1;
369 			}
370 		} else {
371 			len = txmaxseg;
372 			sendalot = 1;
373 		}
374 	}
375 	if (off + len < so->so_snd.sb_cc)
376 		flags &= ~TH_FIN;
377 
378 	mtx_enter(&so->so_rcv.sb_mtx);
379 	win = sbspace_locked(so, &so->so_rcv);
380 	rcv_hiwat = (long) so->so_rcv.sb_hiwat;
381 	mtx_leave(&so->so_rcv.sb_mtx);
382 
383 	/*
384 	 * Sender silly window avoidance.  If connection is idle
385 	 * and can send all data, a maximum segment,
386 	 * at least a maximum default-size segment do it,
387 	 * or are forced, do it; otherwise don't bother.
388 	 * If peer's buffer is tiny, then send
389 	 * when window is at least half open.
390 	 * If retransmitting (possibly after persist timer forced us
391 	 * to send into a small window), then must resend.
392 	 */
393 	if (len) {
394 		if (len >= txmaxseg)
395 			goto send;
396 		if ((idle || (tp->t_flags & TF_NODELAY)) &&
397 		    len + off >= so->so_snd.sb_cc && !doing_sosend &&
398 		    (tp->t_flags & TF_NOPUSH) == 0)
399 			goto send;
400 		if (tp->t_force)
401 			goto send;
402 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
403 			goto send;
404 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
405 			goto send;
406 		if (sack_rxmit)
407 			goto send;
408 	}
409 
410 	/*
411 	 * Compare available window to amount of window
412 	 * known to peer (as advertised window less
413 	 * next expected input).  If the difference is at least two
414 	 * max size segments, or at least 50% of the maximum possible
415 	 * window, then want to send a window update to peer.
416 	 */
417 	if (win > 0) {
418 		/*
419 		 * "adv" is the amount we can increase the window,
420 		 * taking into account that we are limited by
421 		 * TCP_MAXWIN << tp->rcv_scale.
422 		 */
423 		long adv = lmin(win, (long)TCP_MAXWIN << tp->rcv_scale) -
424 			(tp->rcv_adv - tp->rcv_nxt);
425 
426 		if (adv >= (long) (2 * tp->t_maxseg))
427 			goto send;
428 		if (2 * adv >= rcv_hiwat)
429 			goto send;
430 	}
431 
432 	/*
433 	 * Send if we owe peer an ACK.
434 	 */
435 	if (tp->t_flags & TF_ACKNOW)
436 		goto send;
437 	if (flags & (TH_SYN|TH_RST))
438 		goto send;
439 	if (SEQ_GT(tp->snd_up, tp->snd_una))
440 		goto send;
441 	/*
442 	 * If our state indicates that FIN should be sent
443 	 * and we have not yet done so, or we're retransmitting the FIN,
444 	 * then we need to send.
445 	 */
446 	if (flags & TH_FIN &&
447 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
448 		goto send;
449 	/*
450 	 * In SACK, it is possible for tcp_output to fail to send a segment
451 	 * after the retransmission timer has been turned off.  Make sure
452 	 * that the retransmission timer is set.
453 	 */
454 	if (SEQ_GT(tp->snd_max, tp->snd_una) &&
455 	    TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
456 	    TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
457 		TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
458 		return (0);
459 	}
460 
461 	/*
462 	 * TCP window updates are not reliable, rather a polling protocol
463 	 * using ``persist'' packets is used to insure receipt of window
464 	 * updates.  The three ``states'' for the output side are:
465 	 *	idle			not doing retransmits or persists
466 	 *	persisting		to move a small or zero window
467 	 *	(re)transmitting	and thereby not persisting
468 	 *
469 	 * tp->t_timer[TCPT_PERSIST]
470 	 *	is set when we are in persist state.
471 	 * tp->t_force
472 	 *	is set when we are called to send a persist packet.
473 	 * tp->t_timer[TCPT_REXMT]
474 	 *	is set when we are retransmitting
475 	 * The output side is idle when both timers are zero.
476 	 *
477 	 * If send window is too small, there is data to transmit, and no
478 	 * retransmit or persist is pending, then go to persist state.
479 	 * If nothing happens soon, send when timer expires:
480 	 * if window is nonzero, transmit what we can,
481 	 * otherwise force out a byte.
482 	 */
483 	if (so->so_snd.sb_cc && TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
484 	    TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
485 		tp->t_rxtshift = 0;
486 		tcp_setpersist(tp);
487 	}
488 
489 	/*
490 	 * No reason to send a segment, just return.
491 	 */
492 	return (0);
493 
494 send:
495 	/*
496 	 * Before ESTABLISHED, force sending of initial options
497 	 * unless TCP set not to do any options.
498 	 * NOTE: we assume that the IP/TCP header plus TCP options
499 	 * always fit in a single mbuf, leaving room for a maximum
500 	 * link header, i.e.
501 	 *	max_linkhdr + sizeof(network header) + sizeof(struct tcphdr +
502 	 *		optlen <= MHLEN
503 	 */
504 	optlen = 0;
505 
506 	switch (tp->pf) {
507 	case 0:	/*default to PF_INET*/
508 	case PF_INET:
509 		hdrlen = sizeof(struct ip) + sizeof(struct tcphdr);
510 		break;
511 #ifdef INET6
512 	case PF_INET6:
513 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
514 		break;
515 #endif /* INET6 */
516 	default:
517 		return (EPFNOSUPPORT);
518 	}
519 
520 	if (flags & TH_SYN) {
521 		tp->snd_nxt = tp->iss;
522 		if ((tp->t_flags & TF_NOOPT) == 0) {
523 			u_int16_t mss;
524 
525 			opt[0] = TCPOPT_MAXSEG;
526 			opt[1] = 4;
527 			mss = htons((u_int16_t) tcp_mss(tp, 0));
528 			memcpy(opt + 2, &mss, sizeof(mss));
529 			optlen = 4;
530 
531 			if (flags & TH_ACK)
532 				tcp_mss_update(tp);
533 			/*
534 			 * If this is the first SYN of connection (not a SYN
535 			 * ACK), include SACK_PERMIT_HDR option.  If this is a
536 			 * SYN ACK, include SACK_PERMIT_HDR option if peer has
537 			 * already done so.
538 			 */
539 			if (tp->sack_enable && ((flags & TH_ACK) == 0 ||
540 			    (tp->t_flags & TF_SACK_PERMIT))) {
541 				*((u_int32_t *) (opt + optlen)) =
542 				    htonl(TCPOPT_SACK_PERMIT_HDR);
543 				optlen += 4;
544 			}
545 			if ((tp->t_flags & TF_REQ_SCALE) &&
546 			    ((flags & TH_ACK) == 0 ||
547 			    (tp->t_flags & TF_RCVD_SCALE))) {
548 				*((u_int32_t *) (opt + optlen)) = htonl(
549 					TCPOPT_NOP << 24 |
550 					TCPOPT_WINDOW << 16 |
551 					TCPOLEN_WINDOW << 8 |
552 					tp->request_r_scale);
553 				optlen += 4;
554 			}
555 		}
556 	}
557 
558 	/*
559 	 * Send a timestamp and echo-reply if this is a SYN and our side
560 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
561 	 * and our peer have sent timestamps in our SYN's.
562 	 */
563 	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
564 	     (flags & TH_RST) == 0 &&
565 	    ((flags & (TH_SYN|TH_ACK)) == TH_SYN ||
566 	     (tp->t_flags & TF_RCVD_TSTMP))) {
567 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
568 
569 		/* Form timestamp option as shown in appendix A of RFC 1323. */
570 		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
571 		*lp++ = htonl(now + tp->ts_modulate);
572 		*lp   = htonl(tp->ts_recent);
573 		optlen += TCPOLEN_TSTAMP_APPA;
574 	}
575 	/* Set receive buffer autosizing timestamp. */
576 	if (tp->rfbuf_ts == 0) {
577 		tp->rfbuf_ts = now;
578 		tp->rfbuf_cnt = 0;
579 	}
580 
581 #ifdef TCP_SIGNATURE
582 	if (tp->t_flags & TF_SIGNATURE) {
583 		u_int8_t *bp = (u_int8_t *)(opt + optlen);
584 
585 		/* Send signature option */
586 		*(bp++) = TCPOPT_SIGNATURE;
587 		*(bp++) = TCPOLEN_SIGNATURE;
588 		sigoff = optlen + 2;
589 
590 		{
591 			unsigned int i;
592 
593 			for (i = 0; i < 16; i++)
594 				*(bp++) = 0;
595 		}
596 
597 
598 		/* Pad options list to the next 32 bit boundary and
599 		 * terminate it.
600 		 */
601 		*bp++ = TCPOPT_NOP;
602 		*bp++ = TCPOPT_NOP;
603 
604 		optlen += TCPOLEN_SIGLEN;
605 	}
606 #endif /* TCP_SIGNATURE */
607 
608 	/*
609 	 * Send SACKs if necessary.  This should be the last option processed.
610 	 * Only as many SACKs are sent as are permitted by the maximum options
611 	 * size.  No more than three SACKs are sent.
612 	 */
613 	if (tp->sack_enable && tp->t_state == TCPS_ESTABLISHED &&
614 	    (tp->t_flags & (TF_SACK_PERMIT|TF_NOOPT)) == TF_SACK_PERMIT &&
615 	    tp->rcv_numsacks) {
616 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
617 		u_int32_t *olp = lp++;
618 		int count = 0;  /* actual number of SACKs inserted */
619 		int maxsack = (MAX_TCPOPTLEN - (optlen + 4))/TCPOLEN_SACK;
620 
621 		tcpstat_inc(tcps_sack_snd_opts);
622 		maxsack = min(maxsack, TCP_MAX_SACK);
623 		for (i = 0; (i < tp->rcv_numsacks && count < maxsack); i++) {
624 			struct sackblk sack = tp->sackblks[i];
625 			if (sack.start == 0 && sack.end == 0)
626 				continue;
627 			*lp++ = htonl(sack.start);
628 			*lp++ = htonl(sack.end);
629 			count++;
630 		}
631 		*olp = htonl(TCPOPT_SACK_HDR|(TCPOLEN_SACK*count+2));
632 		optlen += TCPOLEN_SACK*count + 4; /* including leading NOPs */
633 	}
634 
635 #ifdef DIAGNOSTIC
636 	if (optlen > MAX_TCPOPTLEN)
637 		panic("tcp_output: options too long");
638 #endif /* DIAGNOSTIC */
639 
640 	hdrlen += optlen;
641 
642 	/*
643 	 * Adjust data length if insertion of options will
644 	 * bump the packet length beyond the t_maxopd length.
645 	 * Clear the FIN bit because we cut off the tail of
646 	 * the segment.
647 	 */
648 	if (len > tp->t_maxopd - optlen) {
649 		if (tso) {
650 			if (len + hdrlen + max_linkhdr > MAXMCLBYTES) {
651 				len = MAXMCLBYTES - hdrlen - max_linkhdr;
652 				sendalot = 1;
653 			}
654 		} else {
655 			len = tp->t_maxopd - optlen;
656 			sendalot = 1;
657 		}
658 		flags &= ~TH_FIN;
659 	}
660 
661 #ifdef DIAGNOSTIC
662 	if (max_linkhdr + hdrlen > MCLBYTES)
663 		panic("tcphdr too big");
664 #endif
665 
666 	/*
667 	 * Grab a header mbuf, attaching a copy of data to
668 	 * be transmitted, and initialize the header from
669 	 * the template for sends on this connection.
670 	 */
671 	if (len) {
672 		if (tp->t_force && len == 1)
673 			tcpstat_inc(tcps_sndprobe);
674 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
675 			tcpstat_pkt(tcps_sndrexmitpack, tcps_sndrexmitbyte,
676 			    len);
677 			tp->t_sndrexmitpack++;
678 		} else {
679 			tcpstat_pkt(tcps_sndpack, tcps_sndbyte, len);
680 		}
681 #ifdef notyet
682 		if ((m = m_copypack(so->so_snd.sb_mb, off,
683 		    (int)len, max_linkhdr + hdrlen)) == 0) {
684 			error = ENOBUFS;
685 			goto out;
686 		}
687 		/*
688 		 * m_copypack left space for our hdr; use it.
689 		 */
690 		m->m_len += hdrlen;
691 		m->m_data -= hdrlen;
692 #else
693 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
694 		if (m != NULL && max_linkhdr + hdrlen > MHLEN) {
695 			MCLGET(m, M_DONTWAIT);
696 			if ((m->m_flags & M_EXT) == 0) {
697 				m_freem(m);
698 				m = NULL;
699 			}
700 		}
701 		if (m == NULL) {
702 			error = ENOBUFS;
703 			goto out;
704 		}
705 		m->m_data += max_linkhdr;
706 		m->m_len = hdrlen;
707 		if (len <= m_trailingspace(m)) {
708 			m_copydata(so->so_snd.sb_mb, off, (int) len,
709 			    mtod(m, caddr_t) + hdrlen);
710 			m->m_len += len;
711 		} else {
712 			m->m_next = m_copym(so->so_snd.sb_mb, off, (int) len,
713 			    M_NOWAIT);
714 			if (m->m_next == 0) {
715 				(void) m_free(m);
716 				error = ENOBUFS;
717 				goto out;
718 			}
719 		}
720 		if (so->so_snd.sb_mb->m_flags & M_PKTHDR)
721 			m->m_pkthdr.ph_loopcnt =
722 			    so->so_snd.sb_mb->m_pkthdr.ph_loopcnt;
723 #endif
724 		/*
725 		 * If we're sending everything we've got, set PUSH.
726 		 * (This will keep happy those implementations which only
727 		 * give data to the user when a buffer fills or
728 		 * a PUSH comes in.)
729 		 */
730 		if (off + len == so->so_snd.sb_cc && !doing_sosend)
731 			flags |= TH_PUSH;
732 		tp->t_sndtime = now;
733 	} else {
734 		if (tp->t_flags & TF_ACKNOW)
735 			tcpstat_inc(tcps_sndacks);
736 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
737 			tcpstat_inc(tcps_sndctrl);
738 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
739 			tcpstat_inc(tcps_sndurg);
740 		else
741 			tcpstat_inc(tcps_sndwinup);
742 
743 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
744 		if (m != NULL && max_linkhdr + hdrlen > MHLEN) {
745 			MCLGET(m, M_DONTWAIT);
746 			if ((m->m_flags & M_EXT) == 0) {
747 				m_freem(m);
748 				m = NULL;
749 			}
750 		}
751 		if (m == NULL) {
752 			error = ENOBUFS;
753 			goto out;
754 		}
755 		m->m_data += max_linkhdr;
756 		m->m_len = hdrlen;
757 	}
758 	m->m_pkthdr.ph_ifidx = 0;
759 	m->m_pkthdr.len = hdrlen + len;
760 
761 	/* Enable TSO and specify the size of the resulting segments. */
762 	if (tso) {
763 		SET(m->m_pkthdr.csum_flags, M_TCP_TSO);
764 		m->m_pkthdr.ph_mss = tp->t_maxseg;
765 	}
766 
767 	if (!tp->t_template)
768 		panic("tcp_output");
769 #ifdef DIAGNOSTIC
770 	if (tp->t_template->m_len != hdrlen - optlen)
771 		panic("tcp_output: template len != hdrlen - optlen");
772 #endif /* DIAGNOSTIC */
773 	memcpy(mtod(m, caddr_t), mtod(tp->t_template, caddr_t),
774 	    tp->t_template->m_len);
775 	th = (struct tcphdr *)(mtod(m, caddr_t) + tp->t_template->m_len -
776 		sizeof(struct tcphdr));
777 
778 	/*
779 	 * Fill in fields, remembering maximum advertised
780 	 * window for use in delaying messages about window sizes.
781 	 * If resending a FIN, be sure not to use a new sequence number.
782 	 */
783 	if ((flags & TH_FIN) && (tp->t_flags & TF_SENTFIN) &&
784 	    (tp->snd_nxt == tp->snd_max))
785 		tp->snd_nxt--;
786 	/*
787 	 * If we are doing retransmissions, then snd_nxt will
788 	 * not reflect the first unsent octet.  For ACK only
789 	 * packets, we do not want the sequence number of the
790 	 * retransmitted packet, we want the sequence number
791 	 * of the next unsent octet.  So, if there is no data
792 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
793 	 * when filling in ti_seq.  But if we are in persist
794 	 * state, snd_max might reflect one byte beyond the
795 	 * right edge of the window, so use snd_nxt in that
796 	 * case, since we know we aren't doing a retransmission.
797 	 * (retransmit and persist are mutually exclusive...)
798 	 */
799 	if (len || (flags & (TH_SYN|TH_FIN)) ||
800 	    TCP_TIMER_ISARMED(tp, TCPT_PERSIST))
801 		th->th_seq = htonl(tp->snd_nxt);
802 	else
803 		th->th_seq = htonl(tp->snd_max);
804 
805 	if (sack_rxmit) {
806 		/*
807 		 * If sendalot was turned on (due to option stuffing), turn it
808 		 * off. Properly set th_seq field.  Advance the ret'x pointer
809 		 * by len.
810 		 */
811 		if (sendalot)
812 			sendalot = 0;
813 		th->th_seq = htonl(p->rxmit);
814 		p->rxmit += len;
815 		tcpstat_pkt(tcps_sack_rexmits, tcps_sack_rexmit_bytes, len);
816 	}
817 
818 	th->th_ack = htonl(tp->rcv_nxt);
819 	if (optlen) {
820 		memcpy(th + 1, opt, optlen);
821 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
822 	}
823 #ifdef TCP_ECN
824 	if (tcp_do_ecn) {
825 		/*
826 		 * if we have received congestion experienced segs,
827 		 * set ECE bit.
828 		 */
829 		if (tp->t_flags & TF_RCVD_CE) {
830 			flags |= TH_ECE;
831 			tcpstat_inc(tcps_ecn_sndece);
832 		}
833 		if (!(tp->t_flags & TF_DISABLE_ECN)) {
834 			/*
835 			 * if this is a SYN seg, set ECE and CWR.
836 			 * set only ECE for SYN-ACK if peer supports ECN.
837 			 */
838 			if ((flags & (TH_SYN|TH_ACK)) == TH_SYN)
839 				flags |= (TH_ECE|TH_CWR);
840 			else if ((tp->t_flags & TF_ECN_PERMIT) &&
841 			    (flags & (TH_SYN|TH_ACK)) == (TH_SYN|TH_ACK))
842 				flags |= TH_ECE;
843 		}
844 		/*
845 		 * if we have reduced the congestion window, notify
846 		 * the peer by setting CWR bit.
847 		 */
848 		if ((tp->t_flags & TF_ECN_PERMIT) &&
849 		    (tp->t_flags & TF_SEND_CWR)) {
850 			flags |= TH_CWR;
851 			tp->t_flags &= ~TF_SEND_CWR;
852 			tcpstat_inc(tcps_ecn_sndcwr);
853 		}
854 	}
855 #endif
856 	th->th_flags = flags;
857 
858 	/*
859 	 * Calculate receive window.  Don't shrink window,
860 	 * but avoid silly window syndrome.
861 	 */
862 	if (win < (rcv_hiwat / 4) && win < (long)tp->t_maxseg)
863 		win = 0;
864 	if (win > (long)TCP_MAXWIN << tp->rcv_scale)
865 		win = (long)TCP_MAXWIN << tp->rcv_scale;
866 	if (win < (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt))
867 		win = (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt);
868 	if (flags & TH_RST)
869 		win = 0;
870 	th->th_win = htons((u_int16_t) (win>>tp->rcv_scale));
871 	if (th->th_win == 0)
872 		tp->t_sndzerowin++;
873 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
874 		u_int32_t urp = tp->snd_up - tp->snd_nxt;
875 		if (urp > IP_MAXPACKET)
876 			urp = IP_MAXPACKET;
877 		th->th_urp = htons((u_int16_t)urp);
878 		th->th_flags |= TH_URG;
879 	} else
880 		/*
881 		 * If no urgent pointer to send, then we pull
882 		 * the urgent pointer to the left edge of the send window
883 		 * so that it doesn't drift into the send window on sequence
884 		 * number wraparound.
885 		 */
886 		tp->snd_up = tp->snd_una;		/* drag it along */
887 
888 #ifdef TCP_SIGNATURE
889 	if (tp->t_flags & TF_SIGNATURE) {
890 		int iphlen;
891 		union sockaddr_union src, dst;
892 		struct tdb *tdb;
893 
894 		bzero(&src, sizeof(union sockaddr_union));
895 		bzero(&dst, sizeof(union sockaddr_union));
896 
897 		switch (tp->pf) {
898 		case 0:	/*default to PF_INET*/
899 		case AF_INET:
900 			iphlen = sizeof(struct ip);
901 			src.sa.sa_len = sizeof(struct sockaddr_in);
902 			src.sa.sa_family = AF_INET;
903 			src.sin.sin_addr = mtod(m, struct ip *)->ip_src;
904 			dst.sa.sa_len = sizeof(struct sockaddr_in);
905 			dst.sa.sa_family = AF_INET;
906 			dst.sin.sin_addr = mtod(m, struct ip *)->ip_dst;
907 			break;
908 #ifdef INET6
909 		case AF_INET6:
910 			iphlen = sizeof(struct ip6_hdr);
911 			src.sa.sa_len = sizeof(struct sockaddr_in6);
912 			src.sa.sa_family = AF_INET6;
913 			src.sin6.sin6_addr = mtod(m, struct ip6_hdr *)->ip6_src;
914 			dst.sa.sa_len = sizeof(struct sockaddr_in6);
915 			dst.sa.sa_family = AF_INET6;
916 			dst.sin6.sin6_addr = mtod(m, struct ip6_hdr *)->ip6_dst;
917 			break;
918 #endif /* INET6 */
919 		}
920 
921 		tdb = gettdbbysrcdst(rtable_l2(tp->t_inpcb->inp_rtableid),
922 		    0, &src, &dst, IPPROTO_TCP);
923 		if (tdb == NULL) {
924 			m_freem(m);
925 			return (EPERM);
926 		}
927 
928 		if (tcp_signature(tdb, tp->pf, m, th, iphlen, 0,
929 		    mtod(m, caddr_t) + hdrlen - optlen + sigoff) < 0) {
930 			m_freem(m);
931 			tdb_unref(tdb);
932 			return (EINVAL);
933 		}
934 		tdb_unref(tdb);
935 	}
936 #endif /* TCP_SIGNATURE */
937 
938 	/* Defer checksumming until later (ip_output() or hardware) */
939 	m->m_pkthdr.csum_flags |= M_TCP_CSUM_OUT;
940 
941 	/*
942 	 * In transmit state, time the transmission and arrange for
943 	 * the retransmit.  In persist state, just set snd_max.
944 	 */
945 	if (tp->t_force == 0 || TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
946 		tcp_seq startseq = tp->snd_nxt;
947 
948 		/*
949 		 * Advance snd_nxt over sequence space of this segment.
950 		 */
951 		if (flags & (TH_SYN|TH_FIN)) {
952 			if (flags & TH_SYN)
953 				tp->snd_nxt++;
954 			if (flags & TH_FIN) {
955 				tp->snd_nxt++;
956 				tp->t_flags |= TF_SENTFIN;
957 			}
958 		}
959 		if (tp->sack_enable) {
960 			if (sack_rxmit && (p->rxmit != tp->snd_nxt)) {
961 				goto timer;
962 			}
963 		}
964 		tp->snd_nxt += len;
965 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
966 			tp->snd_max = tp->snd_nxt;
967 			/*
968 			 * Time this transmission if not a retransmission and
969 			 * not currently timing anything.
970 			 */
971 			if (tp->t_rtttime == 0) {
972 				tp->t_rtttime = now;
973 				tp->t_rtseq = startseq;
974 				tcpstat_inc(tcps_segstimed);
975 			}
976 		}
977 
978 		/*
979 		 * Set retransmit timer if not currently set,
980 		 * and not doing an ack or a keep-alive probe.
981 		 * Initial value for retransmit timer is smoothed
982 		 * round-trip time + 2 * round-trip time variance.
983 		 * Initialize shift counter which is used for backoff
984 		 * of retransmit time.
985 		 */
986  timer:
987 		if (tp->sack_enable && sack_rxmit &&
988 		    TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
989 		    tp->snd_nxt != tp->snd_max) {
990 			TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
991 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
992 				TCP_TIMER_DISARM(tp, TCPT_PERSIST);
993 				tp->t_rxtshift = 0;
994 			}
995 		}
996 
997 		if (TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
998 		    tp->snd_nxt != tp->snd_una) {
999 			TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
1000 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
1001 				TCP_TIMER_DISARM(tp, TCPT_PERSIST);
1002 				tp->t_rxtshift = 0;
1003 			}
1004 		}
1005 
1006 		if (len == 0 && so->so_snd.sb_cc &&
1007 		    TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
1008 		    TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
1009 			/*
1010 			 * Avoid a situation where we do not set persist timer
1011 			 * after a zero window condition. For example:
1012 			 * 1) A -> B: packet with enough data to fill the window
1013 			 * 2) B -> A: ACK for #1 + new data (0 window
1014 			 *    advertisement)
1015 			 * 3) A -> B: ACK for #2, 0 len packet
1016 			 *
1017 			 * In this case, A will not activate the persist timer,
1018 			 * because it chose to send a packet. Unless tcp_output
1019 			 * is called for some other reason (delayed ack timer,
1020 			 * another input packet from B, socket syscall), A will
1021 			 * not send zero window probes.
1022 			 *
1023 			 * So, if you send a 0-length packet, but there is data
1024 			 * in the socket buffer, and neither the rexmt or
1025 			 * persist timer is already set, then activate the
1026 			 * persist timer.
1027 			 */
1028 			tp->t_rxtshift = 0;
1029 			tcp_setpersist(tp);
1030 		}
1031 	} else
1032 		if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
1033 			tp->snd_max = tp->snd_nxt + len;
1034 
1035 	tcp_update_sndspace(tp);
1036 
1037 	/*
1038 	 * Trace.
1039 	 */
1040 	if (so->so_options & SO_DEBUG)
1041 		tcp_trace(TA_OUTPUT, tp->t_state, tp, tp, mtod(m, caddr_t), 0,
1042 			len);
1043 
1044 	/*
1045 	 * Fill in IP length and desired time to live and
1046 	 * send to IP level.  There should be a better way
1047 	 * to handle ttl and tos; we could keep them in
1048 	 * the template, but need a way to checksum without them.
1049 	 */
1050 
1051 #ifdef TCP_ECN
1052 	/*
1053 	 * if peer is ECN capable, set the ECT bit in the IP header.
1054 	 * but don't set ECT for a pure ack, a retransmit or a window probe.
1055 	 */
1056 	needect = 0;
1057 	if (tcp_do_ecn && (tp->t_flags & TF_ECN_PERMIT)) {
1058 		if (len == 0 || SEQ_LT(tp->snd_nxt, tp->snd_max) ||
1059 		    (tp->t_force && len == 1)) {
1060 			/* don't set ECT */
1061 		} else {
1062 			needect = 1;
1063 			tcpstat_inc(tcps_ecn_sndect);
1064 		}
1065 	}
1066 #endif
1067 
1068 	/* force routing table */
1069 	m->m_pkthdr.ph_rtableid = tp->t_inpcb->inp_rtableid;
1070 
1071 #if NPF > 0
1072 	pf_mbuf_link_inpcb(m, tp->t_inpcb);
1073 #endif
1074 
1075 	switch (tp->pf) {
1076 	case 0:	/*default to PF_INET*/
1077 	case AF_INET:
1078 		{
1079 			struct ip *ip;
1080 
1081 			ip = mtod(m, struct ip *);
1082 			ip->ip_len = htons(m->m_pkthdr.len);
1083 			packetlen = m->m_pkthdr.len;
1084 			ip->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl;
1085 			ip->ip_tos = tp->t_inpcb->inp_ip.ip_tos;
1086 #ifdef TCP_ECN
1087 			if (needect)
1088 				ip->ip_tos |= IPTOS_ECN_ECT0;
1089 #endif
1090 		}
1091 #if NSTOEPLITZ > 0
1092 		m->m_pkthdr.ph_flowid = tp->t_inpcb->inp_flowid;
1093 		SET(m->m_pkthdr.csum_flags, M_FLOWID);
1094 #endif
1095 		error = ip_output(m, tp->t_inpcb->inp_options,
1096 		    &tp->t_inpcb->inp_route,
1097 		    (ip_mtudisc ? IP_MTUDISC : 0), NULL,
1098 		    &tp->t_inpcb->inp_seclevel, 0);
1099 		break;
1100 #ifdef INET6
1101 	case AF_INET6:
1102 		{
1103 			struct ip6_hdr *ip6;
1104 
1105 			ip6 = mtod(m, struct ip6_hdr *);
1106 			ip6->ip6_plen = m->m_pkthdr.len -
1107 				sizeof(struct ip6_hdr);
1108 			packetlen = m->m_pkthdr.len;
1109 			ip6->ip6_nxt = IPPROTO_TCP;
1110 			ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb);
1111 #ifdef TCP_ECN
1112 			if (needect)
1113 				ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1114 #endif
1115 		}
1116 		error = ip6_output(m, tp->t_inpcb->inp_outputopts6,
1117 		    &tp->t_inpcb->inp_route, 0, NULL,
1118 		    &tp->t_inpcb->inp_seclevel);
1119 		break;
1120 #endif /* INET6 */
1121 	}
1122 
1123 	if (error) {
1124 out:
1125 		if (error == ENOBUFS) {
1126 			/*
1127 			 * If the interface queue is full, or IP cannot
1128 			 * get an mbuf, trigger TCP slow start.
1129 			 */
1130 			tp->snd_cwnd = tp->t_maxseg;
1131 			return (0);
1132 		}
1133 		if (error == EMSGSIZE) {
1134 			/*
1135 			 * ip_output() will have already fixed the route
1136 			 * for us.  tcp_mtudisc() will, as its last action,
1137 			 * initiate retransmission, so it is important to
1138 			 * not do so here.
1139 			 */
1140 			tcp_mtudisc(tp->t_inpcb, -1);
1141 			return (0);
1142 		}
1143 		if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1144 		    TCPS_HAVERCVDSYN(tp->t_state)) {
1145 			tp->t_softerror = error;
1146 			return (0);
1147 		}
1148 
1149 		/* Restart the delayed ACK timer, if necessary. */
1150 		if (TCP_TIMER_ISARMED(tp, TCPT_DELACK))
1151 			TCP_TIMER_ARM(tp, TCPT_DELACK, tcp_delack_msecs);
1152 
1153 		return (error);
1154 	}
1155 
1156 	if (packetlen > tp->t_pmtud_mtu_sent)
1157 		tp->t_pmtud_mtu_sent = packetlen;
1158 
1159 	tcpstat_inc(tcps_sndtotal);
1160 	if (TCP_TIMER_ISARMED(tp, TCPT_DELACK))
1161 		tcpstat_inc(tcps_delack);
1162 
1163 	/*
1164 	 * Data sent (as far as we can tell).
1165 	 * If this advertises a larger window than any other segment,
1166 	 * then remember the size of the advertised window.
1167 	 * Any pending ACK has now been sent.
1168 	 */
1169 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
1170 		tp->rcv_adv = tp->rcv_nxt + win;
1171 	tp->last_ack_sent = tp->rcv_nxt;
1172 	tp->t_sndacktime = now;
1173 	tp->t_flags &= ~TF_ACKNOW;
1174 	TCP_TIMER_DISARM(tp, TCPT_DELACK);
1175 	if (sendalot)
1176 		goto again;
1177 	return (0);
1178 }
1179 
1180 void
1181 tcp_setpersist(struct tcpcb *tp)
1182 {
1183 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> (1 + TCP_RTT_BASE_SHIFT);
1184 	int msec;
1185 
1186 	if (TCP_TIMER_ISARMED(tp, TCPT_REXMT))
1187 		panic("tcp_output REXMT");
1188 	/*
1189 	 * Start/restart persistence timer.
1190 	 */
1191 	if (t < tp->t_rttmin)
1192 		t = tp->t_rttmin;
1193 	TCPT_RANGESET(msec, t * tcp_backoff[tp->t_rxtshift],
1194 	    TCPTV_PERSMIN, TCPTV_PERSMAX);
1195 	TCP_TIMER_ARM(tp, TCPT_PERSIST, msec);
1196 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1197 		tp->t_rxtshift++;
1198 }
1199 
1200 int
1201 tcp_chopper(struct mbuf *m0, struct mbuf_list *ml, struct ifnet *ifp,
1202     u_int mss)
1203 {
1204 	struct ip *ip = NULL;
1205 #ifdef INET6
1206 	struct ip6_hdr *ip6 = NULL;
1207 #endif
1208 	struct tcphdr *th;
1209 	int firstlen, iphlen, hlen, tlen, off;
1210 	int error;
1211 
1212 	ml_init(ml);
1213 	ml_enqueue(ml, m0);
1214 
1215 	if (mss == 0) {
1216 		error = EINVAL;
1217 		goto bad;
1218 	}
1219 
1220 	ip = mtod(m0, struct ip *);
1221 	switch (ip->ip_v) {
1222 	case 4:
1223 		iphlen = ip->ip_hl << 2;
1224 		if (ISSET(ip->ip_off, htons(IP_OFFMASK | IP_MF)) ||
1225 		    iphlen != sizeof(struct ip) || ip->ip_p != IPPROTO_TCP) {
1226 			/* only TCP without fragment or IP option supported */
1227 			error = EPROTOTYPE;
1228 			goto bad;
1229 		}
1230 		break;
1231 #ifdef INET6
1232 	case 6:
1233 		ip = NULL;
1234 		ip6 = mtod(m0, struct ip6_hdr *);
1235 		iphlen = sizeof(struct ip6_hdr);
1236 		if (ip6->ip6_nxt != IPPROTO_TCP) {
1237 			/* only TCP without IPv6 header chain supported */
1238 			error = EPROTOTYPE;
1239 			goto bad;
1240 		}
1241 		break;
1242 #endif
1243 	default:
1244 		panic("%s: unknown ip version %d", __func__, ip->ip_v);
1245 	}
1246 
1247 	tlen = m0->m_pkthdr.len;
1248 	if (tlen < iphlen + sizeof(struct tcphdr)) {
1249 		error = ENOPROTOOPT;
1250 		goto bad;
1251 	}
1252 	/* IP and TCP header should be contiguous, this check is paranoia */
1253 	if (m0->m_len < iphlen + sizeof(*th)) {
1254 		ml_dequeue(ml);
1255 		if ((m0 = m_pullup(m0, iphlen + sizeof(*th))) == NULL) {
1256 			error = ENOBUFS;
1257 			goto bad;
1258 		}
1259 		ml_enqueue(ml, m0);
1260 	}
1261 	th = (struct tcphdr *)(mtod(m0, caddr_t) + iphlen);
1262 	hlen = iphlen + (th->th_off << 2);
1263 	if (tlen < hlen) {
1264 		error = ENOPROTOOPT;
1265 		goto bad;
1266 	}
1267 	firstlen = MIN(tlen - hlen, mss);
1268 
1269 	CLR(m0->m_pkthdr.csum_flags, M_TCP_TSO);
1270 
1271 	/*
1272 	 * Loop through length of payload after first segment,
1273 	 * make new header and copy data of each part and link onto chain.
1274 	 */
1275 	for (off = hlen + firstlen; off < tlen; off += mss) {
1276 		struct mbuf *m;
1277 		struct tcphdr *mhth;
1278 		int len;
1279 
1280 		len = MIN(tlen - off, mss);
1281 
1282 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
1283 		if (m == NULL) {
1284 			error = ENOBUFS;
1285 			goto bad;
1286 		}
1287 		ml_enqueue(ml, m);
1288 		if ((error = m_dup_pkthdr(m, m0, M_DONTWAIT)) != 0)
1289 			goto bad;
1290 
1291 		/* IP and TCP header to the end, space for link layer header */
1292 		m->m_len = hlen;
1293 		m_align(m, hlen);
1294 
1295 		/* copy and adjust TCP header */
1296 		mhth = (struct tcphdr *)(mtod(m, caddr_t) + iphlen);
1297 		memcpy(mhth, th, hlen - iphlen);
1298 		mhth->th_seq = htonl(ntohl(th->th_seq) + (off - hlen));
1299 		if (off + len < tlen)
1300 			CLR(mhth->th_flags, TH_PUSH|TH_FIN);
1301 
1302 		/* add mbuf chain with payload */
1303 		m->m_pkthdr.len = hlen + len;
1304 		if ((m->m_next = m_copym(m0, off, len, M_DONTWAIT)) == NULL) {
1305 			error = ENOBUFS;
1306 			goto bad;
1307 		}
1308 
1309 		/* copy and adjust IP header, calculate checksum */
1310 		SET(m->m_pkthdr.csum_flags, M_TCP_CSUM_OUT);
1311 		if (ip) {
1312 			struct ip *mhip;
1313 
1314 			mhip = mtod(m, struct ip *);
1315 			*mhip = *ip;
1316 			mhip->ip_len = htons(hlen + len);
1317 			mhip->ip_id = htons(ip_randomid());
1318 			in_hdr_cksum_out(m, ifp);
1319 			in_proto_cksum_out(m, ifp);
1320 		}
1321 #ifdef INET6
1322 		if (ip6) {
1323 			struct ip6_hdr *mhip6;
1324 
1325 			mhip6 = mtod(m, struct ip6_hdr *);
1326 			*mhip6 = *ip6;
1327 			mhip6->ip6_plen = htons(hlen - iphlen + len);
1328 			in6_proto_cksum_out(m, ifp);
1329 		}
1330 #endif
1331 	}
1332 
1333 	/*
1334 	 * Update first segment by trimming what's been copied out
1335 	 * and updating header, then send each segment (in order).
1336 	 */
1337 	if (hlen + firstlen < tlen) {
1338 		m_adj(m0, hlen + firstlen - tlen);
1339 		CLR(th->th_flags, TH_PUSH|TH_FIN);
1340 	}
1341 	/* adjust IP header, calculate checksum */
1342 	SET(m0->m_pkthdr.csum_flags, M_TCP_CSUM_OUT);
1343 	if (ip) {
1344 		ip->ip_len = htons(m0->m_pkthdr.len);
1345 		in_hdr_cksum_out(m0, ifp);
1346 		in_proto_cksum_out(m0, ifp);
1347 	}
1348 #ifdef INET6
1349 	if (ip6) {
1350 		ip6->ip6_plen = htons(m0->m_pkthdr.len - iphlen);
1351 		in6_proto_cksum_out(m0, ifp);
1352 	}
1353 #endif
1354 
1355 	tcpstat_add(tcps_outpkttso, ml_len(ml));
1356 	return 0;
1357 
1358  bad:
1359 	tcpstat_inc(tcps_outbadtso);
1360 	ml_purge(ml);
1361 	return error;
1362 }
1363 
1364 int
1365 tcp_if_output_tso(struct ifnet *ifp, struct mbuf **mp, struct sockaddr *dst,
1366     struct rtentry *rt, uint32_t ifcap, u_int mtu)
1367 {
1368 	struct mbuf_list ml;
1369 	int error;
1370 
1371 	/* caller must fail later or fragment */
1372 	if (!ISSET((*mp)->m_pkthdr.csum_flags, M_TCP_TSO))
1373 		return 0;
1374 	if ((*mp)->m_pkthdr.ph_mss > mtu) {
1375 		CLR((*mp)->m_pkthdr.csum_flags, M_TCP_TSO);
1376 		return 0;
1377 	}
1378 
1379 	/* network interface hardware will do TSO */
1380 	if (in_ifcap_cksum(*mp, ifp, ifcap)) {
1381 		if (ISSET(ifcap, IFCAP_TSOv4)) {
1382 			in_hdr_cksum_out(*mp, ifp);
1383 			in_proto_cksum_out(*mp, ifp);
1384 		}
1385 #ifdef INET6
1386 		if (ISSET(ifcap, IFCAP_TSOv6))
1387 			in6_proto_cksum_out(*mp, ifp);
1388 #endif
1389 		error = ifp->if_output(ifp, *mp, dst, rt);
1390 		if (!error)
1391 			tcpstat_inc(tcps_outhwtso);
1392 		goto done;
1393 	}
1394 
1395 	/* as fallback do TSO in software */
1396 	if ((error = tcp_chopper(*mp, &ml, ifp, (*mp)->m_pkthdr.ph_mss)) ||
1397 	    (error = if_output_ml(ifp, &ml, dst, rt)))
1398 		goto done;
1399 	tcpstat_inc(tcps_outswtso);
1400 
1401  done:
1402 	*mp = NULL;
1403 	return error;
1404 }
1405