xref: /freebsd/sys/netinet/tcp_output.c (revision 79775f8f1be8dc6ce2b705276b7626cb24347f7d)
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  */
31 
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34 
35 #include "opt_inet.h"
36 #include "opt_inet6.h"
37 #include "opt_ipsec.h"
38 #include "opt_mac.h"
39 #include "opt_tcpdebug.h"
40 
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/domain.h>
44 #include <sys/kernel.h>
45 #include <sys/lock.h>
46 #include <sys/mbuf.h>
47 #include <sys/mutex.h>
48 #include <sys/protosw.h>
49 #include <sys/socket.h>
50 #include <sys/socketvar.h>
51 #include <sys/sysctl.h>
52 #include <sys/vimage.h>
53 
54 #include <net/route.h>
55 
56 #include <netinet/in.h>
57 #include <netinet/in_systm.h>
58 #include <netinet/ip.h>
59 #include <netinet/in_pcb.h>
60 #include <netinet/ip_var.h>
61 #include <netinet/ip_options.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 <netipsec/ipsec.h>
80 #endif /*IPSEC*/
81 
82 #include <machine/in_cksum.h>
83 
84 #include <security/mac/mac_framework.h>
85 
86 #ifdef notyet
87 extern struct mbuf *m_copypack();
88 #endif
89 
90 int path_mtu_discovery = 1;
91 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
92 	&path_mtu_discovery, 1, "Enable Path MTU Discovery");
93 
94 int ss_fltsz = 1;
95 SYSCTL_INT(_net_inet_tcp, OID_AUTO, slowstart_flightsize, CTLFLAG_RW,
96 	&ss_fltsz, 1, "Slow start flight size");
97 
98 int ss_fltsz_local = 4;
99 SYSCTL_INT(_net_inet_tcp, OID_AUTO, local_slowstart_flightsize, CTLFLAG_RW,
100 	&ss_fltsz_local, 1, "Slow start flight size for local networks");
101 
102 int     tcp_do_newreno = 1;
103 SYSCTL_INT(_net_inet_tcp, OID_AUTO, newreno, CTLFLAG_RW,
104 	&tcp_do_newreno, 0, "Enable NewReno Algorithms");
105 
106 int	tcp_do_tso = 1;
107 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW,
108 	&tcp_do_tso, 0, "Enable TCP Segmentation Offload");
109 
110 int	tcp_do_autosndbuf = 1;
111 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW,
112 	&tcp_do_autosndbuf, 0, "Enable automatic send buffer sizing");
113 
114 int	tcp_autosndbuf_inc = 8*1024;
115 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW,
116 	&tcp_autosndbuf_inc, 0, "Incrementor step size of automatic send buffer");
117 
118 int	tcp_autosndbuf_max = 256*1024;
119 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW,
120 	&tcp_autosndbuf_max, 0, "Max size of automatic send buffer");
121 
122 
123 /*
124  * Tcp output routine: figure out what should be sent and send it.
125  */
126 int
127 tcp_output(struct tcpcb *tp)
128 {
129 	struct socket *so = tp->t_inpcb->inp_socket;
130 	long len, recwin, sendwin;
131 	int off, flags, error;
132 	struct mbuf *m;
133 	struct ip *ip = NULL;
134 	struct ipovly *ipov = NULL;
135 	struct tcphdr *th;
136 	u_char opt[TCP_MAXOLEN];
137 	unsigned ipoptlen, optlen, hdrlen;
138 #ifdef IPSEC
139 	unsigned ipsec_optlen = 0;
140 #endif
141 	int idle, sendalot;
142 	int sack_rxmit, sack_bytes_rxmt;
143 	struct sackhole *p;
144 	int tso = 0;
145 	struct tcpopt to;
146 #if 0
147 	int maxburst = TCP_MAXBURST;
148 #endif
149 #ifdef INET6
150 	struct ip6_hdr *ip6 = NULL;
151 	int isipv6;
152 
153 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
154 #endif
155 
156 	INP_WLOCK_ASSERT(tp->t_inpcb);
157 
158 	/*
159 	 * Determine length of data that should be transmitted,
160 	 * and flags that will be used.
161 	 * If there is some data or critical controls (SYN, RST)
162 	 * to send, then transmit; otherwise, investigate further.
163 	 */
164 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
165 	if (idle && (ticks - tp->t_rcvtime) >= tp->t_rxtcur) {
166 		/*
167 		 * We have been idle for "a while" and no acks are
168 		 * expected to clock out any data we send --
169 		 * slow start to get ack "clock" running again.
170 		 *
171 		 * Set the slow-start flight size depending on whether
172 		 * this is a local network or not.
173 		 */
174 		int ss = V_ss_fltsz;
175 #ifdef INET6
176 		if (isipv6) {
177 			if (in6_localaddr(&tp->t_inpcb->in6p_faddr))
178 				ss = V_ss_fltsz_local;
179 		} else
180 #endif /* INET6 */
181 		if (in_localaddr(tp->t_inpcb->inp_faddr))
182 			ss = V_ss_fltsz_local;
183 		tp->snd_cwnd = tp->t_maxseg * ss;
184 	}
185 	tp->t_flags &= ~TF_LASTIDLE;
186 	if (idle) {
187 		if (tp->t_flags & TF_MORETOCOME) {
188 			tp->t_flags |= TF_LASTIDLE;
189 			idle = 0;
190 		}
191 	}
192 again:
193 	/*
194 	 * If we've recently taken a timeout, snd_max will be greater than
195 	 * snd_nxt.  There may be SACK information that allows us to avoid
196 	 * resending already delivered data.  Adjust snd_nxt accordingly.
197 	 */
198 	if ((tp->t_flags & TF_SACK_PERMIT) &&
199 	    SEQ_LT(tp->snd_nxt, tp->snd_max))
200 		tcp_sack_adjust(tp);
201 	sendalot = 0;
202 	off = tp->snd_nxt - tp->snd_una;
203 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
204 	sendwin = min(sendwin, tp->snd_bwnd);
205 
206 	flags = tcp_outflags[tp->t_state];
207 	/*
208 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
209 	 * to send out new data (when sendalot is 1), bypass this function.
210 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
211 	 * we're replacing a (future) new transmission with a retransmission
212 	 * now, and we previously incremented snd_cwnd in tcp_input().
213 	 */
214 	/*
215 	 * Still in sack recovery , reset rxmit flag to zero.
216 	 */
217 	sack_rxmit = 0;
218 	sack_bytes_rxmt = 0;
219 	len = 0;
220 	p = NULL;
221 	if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp) &&
222 	    (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
223 		long cwin;
224 
225 		cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
226 		if (cwin < 0)
227 			cwin = 0;
228 		/* Do not retransmit SACK segments beyond snd_recover */
229 		if (SEQ_GT(p->end, tp->snd_recover)) {
230 			/*
231 			 * (At least) part of sack hole extends beyond
232 			 * snd_recover. Check to see if we can rexmit data
233 			 * for this hole.
234 			 */
235 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
236 				/*
237 				 * Can't rexmit any more data for this hole.
238 				 * That data will be rexmitted in the next
239 				 * sack recovery episode, when snd_recover
240 				 * moves past p->rxmit.
241 				 */
242 				p = NULL;
243 				goto after_sack_rexmit;
244 			} else
245 				/* Can rexmit part of the current hole */
246 				len = ((long)ulmin(cwin,
247 						   tp->snd_recover - p->rxmit));
248 		} else
249 			len = ((long)ulmin(cwin, p->end - p->rxmit));
250 		off = p->rxmit - tp->snd_una;
251 		KASSERT(off >= 0,("%s: sack block to the left of una : %d",
252 		    __func__, off));
253 		if (len > 0) {
254 			sack_rxmit = 1;
255 			sendalot = 1;
256 			V_tcpstat.tcps_sack_rexmits++;
257 			V_tcpstat.tcps_sack_rexmit_bytes +=
258 			    min(len, tp->t_maxseg);
259 		}
260 	}
261 after_sack_rexmit:
262 	/*
263 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
264 	 * state flags.
265 	 */
266 	if (tp->t_flags & TF_NEEDFIN)
267 		flags |= TH_FIN;
268 	if (tp->t_flags & TF_NEEDSYN)
269 		flags |= TH_SYN;
270 
271 	SOCKBUF_LOCK(&so->so_snd);
272 	/*
273 	 * If in persist timeout with window of 0, send 1 byte.
274 	 * Otherwise, if window is small but nonzero
275 	 * and timer expired, we will send what we can
276 	 * and go to transmit state.
277 	 */
278 	if (tp->t_flags & TF_FORCEDATA) {
279 		if (sendwin == 0) {
280 			/*
281 			 * If we still have some data to send, then
282 			 * clear the FIN bit.  Usually this would
283 			 * happen below when it realizes that we
284 			 * aren't sending all the data.  However,
285 			 * if we have exactly 1 byte of unsent data,
286 			 * then it won't clear the FIN bit below,
287 			 * and if we are in persist state, we wind
288 			 * up sending the packet without recording
289 			 * that we sent the FIN bit.
290 			 *
291 			 * We can't just blindly clear the FIN bit,
292 			 * because if we don't have any more data
293 			 * to send then the probe will be the FIN
294 			 * itself.
295 			 */
296 			if (off < so->so_snd.sb_cc)
297 				flags &= ~TH_FIN;
298 			sendwin = 1;
299 		} else {
300 			tcp_timer_activate(tp, TT_PERSIST, 0);
301 			tp->t_rxtshift = 0;
302 		}
303 	}
304 
305 	/*
306 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
307 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
308 	 * a negative length.  This can also occur when TCP opens up
309 	 * its congestion window while receiving additional duplicate
310 	 * acks after fast-retransmit because TCP will reset snd_nxt
311 	 * to snd_max after the fast-retransmit.
312 	 *
313 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
314 	 * be set to snd_una, the offset will be 0, and the length may
315 	 * wind up 0.
316 	 *
317 	 * If sack_rxmit is true we are retransmitting from the scoreboard
318 	 * in which case len is already set.
319 	 */
320 	if (sack_rxmit == 0) {
321 		if (sack_bytes_rxmt == 0)
322 			len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off);
323 		else {
324 			long cwin;
325 
326                         /*
327 			 * We are inside of a SACK recovery episode and are
328 			 * sending new data, having retransmitted all the
329 			 * data possible in the scoreboard.
330 			 */
331 			len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd)
332 			       - off);
333 			/*
334 			 * Don't remove this (len > 0) check !
335 			 * We explicitly check for len > 0 here (although it
336 			 * isn't really necessary), to work around a gcc
337 			 * optimization issue - to force gcc to compute
338 			 * len above. Without this check, the computation
339 			 * of len is bungled by the optimizer.
340 			 */
341 			if (len > 0) {
342 				cwin = tp->snd_cwnd -
343 					(tp->snd_nxt - tp->sack_newdata) -
344 					sack_bytes_rxmt;
345 				if (cwin < 0)
346 					cwin = 0;
347 				len = lmin(len, cwin);
348 			}
349 		}
350 	}
351 
352 	/*
353 	 * Lop off SYN bit if it has already been sent.  However, if this
354 	 * is SYN-SENT state and if segment contains data and if we don't
355 	 * know that foreign host supports TAO, suppress sending segment.
356 	 */
357 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
358 		if (tp->t_state != TCPS_SYN_RECEIVED)
359 			flags &= ~TH_SYN;
360 		off--, len++;
361 	}
362 
363 	/*
364 	 * Be careful not to send data and/or FIN on SYN segments.
365 	 * This measure is needed to prevent interoperability problems
366 	 * with not fully conformant TCP implementations.
367 	 */
368 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
369 		len = 0;
370 		flags &= ~TH_FIN;
371 	}
372 
373 	if (len < 0) {
374 		/*
375 		 * If FIN has been sent but not acked,
376 		 * but we haven't been called to retransmit,
377 		 * len will be < 0.  Otherwise, window shrank
378 		 * after we sent into it.  If window shrank to 0,
379 		 * cancel pending retransmit, pull snd_nxt back
380 		 * to (closed) window, and set the persist timer
381 		 * if it isn't already going.  If the window didn't
382 		 * close completely, just wait for an ACK.
383 		 */
384 		len = 0;
385 		if (sendwin == 0) {
386 			tcp_timer_activate(tp, TT_REXMT, 0);
387 			tp->t_rxtshift = 0;
388 			tp->snd_nxt = tp->snd_una;
389 			if (!tcp_timer_active(tp, TT_PERSIST))
390 				tcp_setpersist(tp);
391 		}
392 	}
393 
394 	/* len will be >= 0 after this point. */
395 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
396 
397 	/*
398 	 * Automatic sizing of send socket buffer.  Often the send buffer
399 	 * size is not optimally adjusted to the actual network conditions
400 	 * at hand (delay bandwidth product).  Setting the buffer size too
401 	 * small limits throughput on links with high bandwidth and high
402 	 * delay (eg. trans-continental/oceanic links).  Setting the
403 	 * buffer size too big consumes too much real kernel memory,
404 	 * especially with many connections on busy servers.
405 	 *
406 	 * The criteria to step up the send buffer one notch are:
407 	 *  1. receive window of remote host is larger than send buffer
408 	 *     (with a fudge factor of 5/4th);
409 	 *  2. send buffer is filled to 7/8th with data (so we actually
410 	 *     have data to make use of it);
411 	 *  3. send buffer fill has not hit maximal automatic size;
412 	 *  4. our send window (slow start and cogestion controlled) is
413 	 *     larger than sent but unacknowledged data in send buffer.
414 	 *
415 	 * The remote host receive window scaling factor may limit the
416 	 * growing of the send buffer before it reaches its allowed
417 	 * maximum.
418 	 *
419 	 * It scales directly with slow start or congestion window
420 	 * and does at most one step per received ACK.  This fast
421 	 * scaling has the drawback of growing the send buffer beyond
422 	 * what is strictly necessary to make full use of a given
423 	 * delay*bandwith product.  However testing has shown this not
424 	 * to be much of an problem.  At worst we are trading wasting
425 	 * of available bandwith (the non-use of it) for wasting some
426 	 * socket buffer memory.
427 	 *
428 	 * TODO: Shrink send buffer during idle periods together
429 	 * with congestion window.  Requires another timer.  Has to
430 	 * wait for upcoming tcp timer rewrite.
431 	 */
432 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
433 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
434 		    so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
435 		    so->so_snd.sb_cc < V_tcp_autosndbuf_max &&
436 		    sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
437 			if (!sbreserve_locked(&so->so_snd,
438 			    min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
439 			     V_tcp_autosndbuf_max), so, curthread))
440 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
441 		}
442 	}
443 
444 	/*
445 	 * Truncate to the maximum segment length or enable TCP Segmentation
446 	 * Offloading (if supported by hardware) and ensure that FIN is removed
447 	 * if the length no longer contains the last data byte.
448 	 *
449 	 * TSO may only be used if we are in a pure bulk sending state.  The
450 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
451 	 * IP options prevent using TSO.  With TSO the TCP header is the same
452 	 * (except for the sequence number) for all generated packets.  This
453 	 * makes it impossible to transmit any options which vary per generated
454 	 * segment or packet.
455 	 *
456 	 * The length of TSO bursts is limited to TCP_MAXWIN.  That limit and
457 	 * removal of FIN (if not already catched here) are handled later after
458 	 * the exact length of the TCP options are known.
459 	 */
460 #ifdef IPSEC
461 	/*
462 	 * Pre-calculate here as we save another lookup into the darknesses
463 	 * of IPsec that way and can actually decide if TSO is ok.
464 	 */
465 	ipsec_optlen = ipsec_hdrsiz_tcp(tp);
466 #endif
467 	if (len > tp->t_maxseg) {
468 		if ((tp->t_flags & TF_TSO) && V_tcp_do_tso &&
469 		    ((tp->t_flags & TF_SIGNATURE) == 0) &&
470 		    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
471 		    tp->t_inpcb->inp_options == NULL &&
472 		    tp->t_inpcb->in6p_options == NULL
473 #ifdef IPSEC
474 		    && ipsec_optlen == 0
475 #endif
476 		    ) {
477 			tso = 1;
478 		} else {
479 			len = tp->t_maxseg;
480 			sendalot = 1;
481 			tso = 0;
482 		}
483 	}
484 	if (sack_rxmit) {
485 		if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
486 			flags &= ~TH_FIN;
487 	} else {
488 		if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
489 			flags &= ~TH_FIN;
490 	}
491 
492 	recwin = sbspace(&so->so_rcv);
493 
494 	/*
495 	 * Sender silly window avoidance.   We transmit under the following
496 	 * conditions when len is non-zero:
497 	 *
498 	 *	- We have a full segment (or more with TSO)
499 	 *	- This is the last buffer in a write()/send() and we are
500 	 *	  either idle or running NODELAY
501 	 *	- we've timed out (e.g. persist timer)
502 	 *	- we have more then 1/2 the maximum send window's worth of
503 	 *	  data (receiver may be limited the window size)
504 	 *	- we need to retransmit
505 	 */
506 	if (len) {
507 		if (len >= tp->t_maxseg)
508 			goto send;
509 		/*
510 		 * NOTE! on localhost connections an 'ack' from the remote
511 		 * end may occur synchronously with the output and cause
512 		 * us to flush a buffer queued with moretocome.  XXX
513 		 *
514 		 * note: the len + off check is almost certainly unnecessary.
515 		 */
516 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
517 		    (idle || (tp->t_flags & TF_NODELAY)) &&
518 		    len + off >= so->so_snd.sb_cc &&
519 		    (tp->t_flags & TF_NOPUSH) == 0) {
520 			goto send;
521 		}
522 		if (tp->t_flags & TF_FORCEDATA)		/* typ. timeout case */
523 			goto send;
524 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
525 			goto send;
526 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
527 			goto send;
528 		if (sack_rxmit)
529 			goto send;
530 	}
531 
532 	/*
533 	 * Compare available window to amount of window
534 	 * known to peer (as advertised window less
535 	 * next expected input).  If the difference is at least two
536 	 * max size segments, or at least 50% of the maximum possible
537 	 * window, then want to send a window update to peer.
538 	 * Skip this if the connection is in T/TCP half-open state.
539 	 * Don't send pure window updates when the peer has closed
540 	 * the connection and won't ever send more data.
541 	 */
542 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
543 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
544 		/*
545 		 * "adv" is the amount we can increase the window,
546 		 * taking into account that we are limited by
547 		 * TCP_MAXWIN << tp->rcv_scale.
548 		 */
549 		long adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale) -
550 			(tp->rcv_adv - tp->rcv_nxt);
551 
552 		if (adv >= (long) (2 * tp->t_maxseg))
553 			goto send;
554 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
555 			goto send;
556 	}
557 
558 	/*
559 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
560 	 * is also a catch-all for the retransmit timer timeout case.
561 	 */
562 	if (tp->t_flags & TF_ACKNOW)
563 		goto send;
564 	if ((flags & TH_RST) ||
565 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
566 		goto send;
567 	if (SEQ_GT(tp->snd_up, tp->snd_una))
568 		goto send;
569 	/*
570 	 * If our state indicates that FIN should be sent
571 	 * and we have not yet done so, then we need to send.
572 	 */
573 	if (flags & TH_FIN &&
574 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
575 		goto send;
576 	/*
577 	 * In SACK, it is possible for tcp_output to fail to send a segment
578 	 * after the retransmission timer has been turned off.  Make sure
579 	 * that the retransmission timer is set.
580 	 */
581 	if ((tp->t_flags & TF_SACK_PERMIT) &&
582 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
583 	    !tcp_timer_active(tp, TT_REXMT) &&
584 	    !tcp_timer_active(tp, TT_PERSIST)) {
585 		tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
586 		goto just_return;
587 	}
588 	/*
589 	 * TCP window updates are not reliable, rather a polling protocol
590 	 * using ``persist'' packets is used to insure receipt of window
591 	 * updates.  The three ``states'' for the output side are:
592 	 *	idle			not doing retransmits or persists
593 	 *	persisting		to move a small or zero window
594 	 *	(re)transmitting	and thereby not persisting
595 	 *
596 	 * tcp_timer_active(tp, TT_PERSIST)
597 	 *	is true when we are in persist state.
598 	 * (tp->t_flags & TF_FORCEDATA)
599 	 *	is set when we are called to send a persist packet.
600 	 * tcp_timer_active(tp, TT_REXMT)
601 	 *	is set when we are retransmitting
602 	 * The output side is idle when both timers are zero.
603 	 *
604 	 * If send window is too small, there is data to transmit, and no
605 	 * retransmit or persist is pending, then go to persist state.
606 	 * If nothing happens soon, send when timer expires:
607 	 * if window is nonzero, transmit what we can,
608 	 * otherwise force out a byte.
609 	 */
610 	if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) &&
611 	    !tcp_timer_active(tp, TT_PERSIST)) {
612 		tp->t_rxtshift = 0;
613 		tcp_setpersist(tp);
614 	}
615 
616 	/*
617 	 * No reason to send a segment, just return.
618 	 */
619 just_return:
620 	SOCKBUF_UNLOCK(&so->so_snd);
621 	return (0);
622 
623 send:
624 	SOCKBUF_LOCK_ASSERT(&so->so_snd);
625 	/*
626 	 * Before ESTABLISHED, force sending of initial options
627 	 * unless TCP set not to do any options.
628 	 * NOTE: we assume that the IP/TCP header plus TCP options
629 	 * always fit in a single mbuf, leaving room for a maximum
630 	 * link header, i.e.
631 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
632 	 */
633 	optlen = 0;
634 #ifdef INET6
635 	if (isipv6)
636 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
637 	else
638 #endif
639 	hdrlen = sizeof (struct tcpiphdr);
640 
641 	/*
642 	 * Compute options for segment.
643 	 * We only have to care about SYN and established connection
644 	 * segments.  Options for SYN-ACK segments are handled in TCP
645 	 * syncache.
646 	 */
647 	if ((tp->t_flags & TF_NOOPT) == 0) {
648 		to.to_flags = 0;
649 		/* Maximum segment size. */
650 		if (flags & TH_SYN) {
651 			tp->snd_nxt = tp->iss;
652 			to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
653 			to.to_flags |= TOF_MSS;
654 		}
655 		/* Window scaling. */
656 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
657 			to.to_wscale = tp->request_r_scale;
658 			to.to_flags |= TOF_SCALE;
659 		}
660 		/* Timestamps. */
661 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
662 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
663 			to.to_tsval = ticks + tp->ts_offset;
664 			to.to_tsecr = tp->ts_recent;
665 			to.to_flags |= TOF_TS;
666 			/* Set receive buffer autosizing timestamp. */
667 			if (tp->rfbuf_ts == 0 &&
668 			    (so->so_rcv.sb_flags & SB_AUTOSIZE))
669 				tp->rfbuf_ts = ticks;
670 		}
671 		/* Selective ACK's. */
672 		if (tp->t_flags & TF_SACK_PERMIT) {
673 			if (flags & TH_SYN)
674 				to.to_flags |= TOF_SACKPERM;
675 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
676 			    (tp->t_flags & TF_SACK_PERMIT) &&
677 			    tp->rcv_numsacks > 0) {
678 				to.to_flags |= TOF_SACK;
679 				to.to_nsacks = tp->rcv_numsacks;
680 				to.to_sacks = (u_char *)tp->sackblks;
681 			}
682 		}
683 #ifdef TCP_SIGNATURE
684 		/* TCP-MD5 (RFC2385). */
685 		if (tp->t_flags & TF_SIGNATURE)
686 			to.to_flags |= TOF_SIGNATURE;
687 #endif /* TCP_SIGNATURE */
688 
689 		/* Processing the options. */
690 		hdrlen += optlen = tcp_addoptions(&to, opt);
691 	}
692 
693 #ifdef INET6
694 	if (isipv6)
695 		ipoptlen = ip6_optlen(tp->t_inpcb);
696 	else
697 #endif
698 	if (tp->t_inpcb->inp_options)
699 		ipoptlen = tp->t_inpcb->inp_options->m_len -
700 				offsetof(struct ipoption, ipopt_list);
701 	else
702 		ipoptlen = 0;
703 #ifdef IPSEC
704 	ipoptlen += ipsec_optlen;
705 #endif
706 
707 	/*
708 	 * Adjust data length if insertion of options will
709 	 * bump the packet length beyond the t_maxopd length.
710 	 * Clear the FIN bit because we cut off the tail of
711 	 * the segment.
712 	 *
713 	 * When doing TSO limit a burst to TCP_MAXWIN minus the
714 	 * IP, TCP and Options length to keep ip->ip_len from
715 	 * overflowing.  Prevent the last segment from being
716 	 * fractional thus making them all equal sized and set
717 	 * the flag to continue sending.  TSO is disabled when
718 	 * IP options or IPSEC are present.
719 	 */
720 	if (len + optlen + ipoptlen > tp->t_maxopd) {
721 		flags &= ~TH_FIN;
722 		if (tso) {
723 			if (len > TCP_MAXWIN - hdrlen - optlen) {
724 				len = TCP_MAXWIN - hdrlen - optlen;
725 				len = len - (len % (tp->t_maxopd - optlen));
726 				sendalot = 1;
727 			} else if (tp->t_flags & TF_NEEDFIN)
728 				sendalot = 1;
729 		} else {
730 			len = tp->t_maxopd - optlen - ipoptlen;
731 			sendalot = 1;
732 		}
733 	}
734 
735 /*#ifdef DIAGNOSTIC*/
736 #ifdef INET6
737 	if (max_linkhdr + hdrlen > MCLBYTES)
738 #else
739 	if (max_linkhdr + hdrlen > MHLEN)
740 #endif
741 		panic("tcphdr too big");
742 /*#endif*/
743 
744 	/*
745 	 * This KASSERT is here to catch edge cases at a well defined place.
746 	 * Before, those had triggered (random) panic conditions further down.
747 	 */
748 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
749 
750 	/*
751 	 * Grab a header mbuf, attaching a copy of data to
752 	 * be transmitted, and initialize the header from
753 	 * the template for sends on this connection.
754 	 */
755 	if (len) {
756 		struct mbuf *mb;
757 		u_int moff;
758 
759 		if ((tp->t_flags & TF_FORCEDATA) && len == 1)
760 			V_tcpstat.tcps_sndprobe++;
761 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
762 			V_tcpstat.tcps_sndrexmitpack++;
763 			V_tcpstat.tcps_sndrexmitbyte += len;
764 		} else {
765 			V_tcpstat.tcps_sndpack++;
766 			V_tcpstat.tcps_sndbyte += len;
767 		}
768 #ifdef notyet
769 		if ((m = m_copypack(so->so_snd.sb_mb, off,
770 		    (int)len, max_linkhdr + hdrlen)) == 0) {
771 			SOCKBUF_UNLOCK(&so->so_snd);
772 			error = ENOBUFS;
773 			goto out;
774 		}
775 		/*
776 		 * m_copypack left space for our hdr; use it.
777 		 */
778 		m->m_len += hdrlen;
779 		m->m_data -= hdrlen;
780 #else
781 		MGETHDR(m, M_DONTWAIT, MT_DATA);
782 		if (m == NULL) {
783 			SOCKBUF_UNLOCK(&so->so_snd);
784 			error = ENOBUFS;
785 			goto out;
786 		}
787 #ifdef INET6
788 		if (MHLEN < hdrlen + max_linkhdr) {
789 			MCLGET(m, M_DONTWAIT);
790 			if ((m->m_flags & M_EXT) == 0) {
791 				SOCKBUF_UNLOCK(&so->so_snd);
792 				m_freem(m);
793 				error = ENOBUFS;
794 				goto out;
795 			}
796 		}
797 #endif
798 		m->m_data += max_linkhdr;
799 		m->m_len = hdrlen;
800 
801 		/*
802 		 * Start the m_copy functions from the closest mbuf
803 		 * to the offset in the socket buffer chain.
804 		 */
805 		mb = sbsndptr(&so->so_snd, off, len, &moff);
806 
807 		if (len <= MHLEN - hdrlen - max_linkhdr) {
808 			m_copydata(mb, moff, (int)len,
809 			    mtod(m, caddr_t) + hdrlen);
810 			m->m_len += len;
811 		} else {
812 			m->m_next = m_copy(mb, moff, (int)len);
813 			if (m->m_next == NULL) {
814 				SOCKBUF_UNLOCK(&so->so_snd);
815 				(void) m_free(m);
816 				error = ENOBUFS;
817 				goto out;
818 			}
819 		}
820 #endif
821 		/*
822 		 * If we're sending everything we've got, set PUSH.
823 		 * (This will keep happy those implementations which only
824 		 * give data to the user when a buffer fills or
825 		 * a PUSH comes in.)
826 		 */
827 		if (off + len == so->so_snd.sb_cc)
828 			flags |= TH_PUSH;
829 		SOCKBUF_UNLOCK(&so->so_snd);
830 	} else {
831 		SOCKBUF_UNLOCK(&so->so_snd);
832 		if (tp->t_flags & TF_ACKNOW)
833 			V_tcpstat.tcps_sndacks++;
834 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
835 			V_tcpstat.tcps_sndctrl++;
836 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
837 			V_tcpstat.tcps_sndurg++;
838 		else
839 			V_tcpstat.tcps_sndwinup++;
840 
841 		MGETHDR(m, M_DONTWAIT, MT_DATA);
842 		if (m == NULL) {
843 			error = ENOBUFS;
844 			goto out;
845 		}
846 #ifdef INET6
847 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
848 		    MHLEN >= hdrlen) {
849 			MH_ALIGN(m, hdrlen);
850 		} else
851 #endif
852 		m->m_data += max_linkhdr;
853 		m->m_len = hdrlen;
854 	}
855 	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
856 	m->m_pkthdr.rcvif = (struct ifnet *)0;
857 #ifdef MAC
858 	mac_inpcb_create_mbuf(tp->t_inpcb, m);
859 #endif
860 #ifdef INET6
861 	if (isipv6) {
862 		ip6 = mtod(m, struct ip6_hdr *);
863 		th = (struct tcphdr *)(ip6 + 1);
864 		tcpip_fillheaders(tp->t_inpcb, ip6, th);
865 	} else
866 #endif /* INET6 */
867 	{
868 		ip = mtod(m, struct ip *);
869 		ipov = (struct ipovly *)ip;
870 		th = (struct tcphdr *)(ip + 1);
871 		tcpip_fillheaders(tp->t_inpcb, ip, th);
872 	}
873 
874 	/*
875 	 * Fill in fields, remembering maximum advertised
876 	 * window for use in delaying messages about window sizes.
877 	 * If resending a FIN, be sure not to use a new sequence number.
878 	 */
879 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
880 	    tp->snd_nxt == tp->snd_max)
881 		tp->snd_nxt--;
882 	/*
883 	 * If we are starting a connection, send ECN setup
884 	 * SYN packet. If we are on a retransmit, we may
885 	 * resend those bits a number of times as per
886 	 * RFC 3168.
887 	 */
888 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
889 		if (tp->t_rxtshift >= 1) {
890 			if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
891 				flags |= TH_ECE|TH_CWR;
892 		} else
893 			flags |= TH_ECE|TH_CWR;
894 	}
895 
896 	if (tp->t_state == TCPS_ESTABLISHED &&
897 	    (tp->t_flags & TF_ECN_PERMIT)) {
898 		/*
899 		 * If the peer has ECN, mark data packets with
900 		 * ECN capable transmission (ECT).
901 		 * Ignore pure ack packets, retransmissions and window probes.
902 		 */
903 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
904 		    !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
905 #ifdef INET6
906 			if (isipv6)
907 				ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
908 			else
909 #endif
910 				ip->ip_tos |= IPTOS_ECN_ECT0;
911 			V_tcpstat.tcps_ecn_ect0++;
912 		}
913 
914 		/*
915 		 * Reply with proper ECN notifications.
916 		 */
917 		if (tp->t_flags & TF_ECN_SND_CWR) {
918 			flags |= TH_CWR;
919 			tp->t_flags &= ~TF_ECN_SND_CWR;
920 		}
921 		if (tp->t_flags & TF_ECN_SND_ECE)
922 			flags |= TH_ECE;
923 	}
924 
925 	/*
926 	 * If we are doing retransmissions, then snd_nxt will
927 	 * not reflect the first unsent octet.  For ACK only
928 	 * packets, we do not want the sequence number of the
929 	 * retransmitted packet, we want the sequence number
930 	 * of the next unsent octet.  So, if there is no data
931 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
932 	 * when filling in ti_seq.  But if we are in persist
933 	 * state, snd_max might reflect one byte beyond the
934 	 * right edge of the window, so use snd_nxt in that
935 	 * case, since we know we aren't doing a retransmission.
936 	 * (retransmit and persist are mutually exclusive...)
937 	 */
938 	if (sack_rxmit == 0) {
939 		if (len || (flags & (TH_SYN|TH_FIN)) ||
940 		    tcp_timer_active(tp, TT_PERSIST))
941 			th->th_seq = htonl(tp->snd_nxt);
942 		else
943 			th->th_seq = htonl(tp->snd_max);
944 	} else {
945 		th->th_seq = htonl(p->rxmit);
946 		p->rxmit += len;
947 		tp->sackhint.sack_bytes_rexmit += len;
948 	}
949 	th->th_ack = htonl(tp->rcv_nxt);
950 	if (optlen) {
951 		bcopy(opt, th + 1, optlen);
952 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
953 	}
954 	th->th_flags = flags;
955 	/*
956 	 * Calculate receive window.  Don't shrink window,
957 	 * but avoid silly window syndrome.
958 	 */
959 	if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
960 	    recwin < (long)tp->t_maxseg)
961 		recwin = 0;
962 	if (recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
963 		recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
964 	if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
965 		recwin = (long)TCP_MAXWIN << tp->rcv_scale;
966 
967 	/*
968 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
969 	 * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
970 	 * case is handled in syncache.
971 	 */
972 	if (flags & TH_SYN)
973 		th->th_win = htons((u_short)
974 				(min(sbspace(&so->so_rcv), TCP_MAXWIN)));
975 	else
976 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
977 
978 	/*
979 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
980 	 * a 0 window.  This may cause the remote transmitter to stall.  This
981 	 * flag tells soreceive() to disable delayed acknowledgements when
982 	 * draining the buffer.  This can occur if the receiver is attempting
983 	 * to read more data than can be buffered prior to transmitting on
984 	 * the connection.
985 	 */
986 	if (recwin == 0)
987 		tp->t_flags |= TF_RXWIN0SENT;
988 	else
989 		tp->t_flags &= ~TF_RXWIN0SENT;
990 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
991 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
992 		th->th_flags |= TH_URG;
993 	} else
994 		/*
995 		 * If no urgent pointer to send, then we pull
996 		 * the urgent pointer to the left edge of the send window
997 		 * so that it doesn't drift into the send window on sequence
998 		 * number wraparound.
999 		 */
1000 		tp->snd_up = tp->snd_una;		/* drag it along */
1001 
1002 #ifdef TCP_SIGNATURE
1003 	if (tp->t_flags & TF_SIGNATURE) {
1004 		int sigoff = to.to_signature - opt;
1005 		tcp_signature_compute(m, 0, len, optlen,
1006 		    (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
1007 	}
1008 #endif
1009 
1010 	/*
1011 	 * Put TCP length in extended header, and then
1012 	 * checksum extended header and data.
1013 	 */
1014 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1015 #ifdef INET6
1016 	if (isipv6)
1017 		/*
1018 		 * ip6_plen is not need to be filled now, and will be filled
1019 		 * in ip6_output.
1020 		 */
1021 		th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
1022 				       sizeof(struct tcphdr) + optlen + len);
1023 	else
1024 #endif /* INET6 */
1025 	{
1026 		m->m_pkthdr.csum_flags = CSUM_TCP;
1027 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1028 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1029 		    htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1030 
1031 		/* IP version must be set here for ipv4/ipv6 checking later */
1032 		KASSERT(ip->ip_v == IPVERSION,
1033 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1034 	}
1035 
1036 	/*
1037 	 * Enable TSO and specify the size of the segments.
1038 	 * The TCP pseudo header checksum is always provided.
1039 	 * XXX: Fixme: This is currently not the case for IPv6.
1040 	 */
1041 	if (tso) {
1042 		m->m_pkthdr.csum_flags = CSUM_TSO;
1043 		m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen;
1044 	}
1045 
1046 	/*
1047 	 * In transmit state, time the transmission and arrange for
1048 	 * the retransmit.  In persist state, just set snd_max.
1049 	 */
1050 	if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1051 	    !tcp_timer_active(tp, TT_PERSIST)) {
1052 		tcp_seq startseq = tp->snd_nxt;
1053 
1054 		/*
1055 		 * Advance snd_nxt over sequence space of this segment.
1056 		 */
1057 		if (flags & (TH_SYN|TH_FIN)) {
1058 			if (flags & TH_SYN)
1059 				tp->snd_nxt++;
1060 			if (flags & TH_FIN) {
1061 				tp->snd_nxt++;
1062 				tp->t_flags |= TF_SENTFIN;
1063 			}
1064 		}
1065 		if (sack_rxmit)
1066 			goto timer;
1067 		tp->snd_nxt += len;
1068 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1069 			tp->snd_max = tp->snd_nxt;
1070 			/*
1071 			 * Time this transmission if not a retransmission and
1072 			 * not currently timing anything.
1073 			 */
1074 			if (tp->t_rtttime == 0) {
1075 				tp->t_rtttime = ticks;
1076 				tp->t_rtseq = startseq;
1077 				V_tcpstat.tcps_segstimed++;
1078 			}
1079 		}
1080 
1081 		/*
1082 		 * Set retransmit timer if not currently set,
1083 		 * and not doing a pure ack or a keep-alive probe.
1084 		 * Initial value for retransmit timer is smoothed
1085 		 * round-trip time + 2 * round-trip time variance.
1086 		 * Initialize shift counter which is used for backoff
1087 		 * of retransmit time.
1088 		 */
1089 timer:
1090 		if (!tcp_timer_active(tp, TT_REXMT) &&
1091 		    ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1092 		     (tp->snd_nxt != tp->snd_una))) {
1093 			if (tcp_timer_active(tp, TT_PERSIST)) {
1094 				tcp_timer_activate(tp, TT_PERSIST, 0);
1095 				tp->t_rxtshift = 0;
1096 			}
1097 			tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1098 		}
1099 	} else {
1100 		/*
1101 		 * Persist case, update snd_max but since we are in
1102 		 * persist mode (no window) we do not update snd_nxt.
1103 		 */
1104 		int xlen = len;
1105 		if (flags & TH_SYN)
1106 			++xlen;
1107 		if (flags & TH_FIN) {
1108 			++xlen;
1109 			tp->t_flags |= TF_SENTFIN;
1110 		}
1111 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1112 			tp->snd_max = tp->snd_nxt + len;
1113 	}
1114 
1115 #ifdef TCPDEBUG
1116 	/*
1117 	 * Trace.
1118 	 */
1119 	if (so->so_options & SO_DEBUG) {
1120 		u_short save = 0;
1121 #ifdef INET6
1122 		if (!isipv6)
1123 #endif
1124 		{
1125 			save = ipov->ih_len;
1126 			ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1127 		}
1128 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1129 #ifdef INET6
1130 		if (!isipv6)
1131 #endif
1132 		ipov->ih_len = save;
1133 	}
1134 #endif
1135 
1136 	/*
1137 	 * Fill in IP length and desired time to live and
1138 	 * send to IP level.  There should be a better way
1139 	 * to handle ttl and tos; we could keep them in
1140 	 * the template, but need a way to checksum without them.
1141 	 */
1142 	/*
1143 	 * m->m_pkthdr.len should have been set before cksum calcuration,
1144 	 * because in6_cksum() need it.
1145 	 */
1146 #ifdef INET6
1147 	if (isipv6) {
1148 		/*
1149 		 * we separately set hoplimit for every segment, since the
1150 		 * user might want to change the value via setsockopt.
1151 		 * Also, desired default hop limit might be changed via
1152 		 * Neighbor Discovery.
1153 		 */
1154 		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1155 
1156 		/* TODO: IPv6 IP6TOS_ECT bit on */
1157 		error = ip6_output(m,
1158 			    tp->t_inpcb->in6p_outputopts, NULL,
1159 			    ((so->so_options & SO_DONTROUTE) ?
1160 			    IP_ROUTETOIF : 0), NULL, NULL, tp->t_inpcb);
1161 	} else
1162 #endif /* INET6 */
1163     {
1164 	ip->ip_len = m->m_pkthdr.len;
1165 #ifdef INET6
1166 	if (INP_CHECK_SOCKAF(so, AF_INET6))
1167 		ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1168 #endif /* INET6 */
1169 	/*
1170 	 * If we do path MTU discovery, then we set DF on every packet.
1171 	 * This might not be the best thing to do according to RFC3390
1172 	 * Section 2. However the tcp hostcache migitates the problem
1173 	 * so it affects only the first tcp connection with a host.
1174 	 */
1175 	if (V_path_mtu_discovery)
1176 		ip->ip_off |= IP_DF;
1177 
1178 	error = ip_output(m, tp->t_inpcb->inp_options, NULL,
1179 	    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1180 	    tp->t_inpcb);
1181     }
1182 	if (error) {
1183 
1184 		/*
1185 		 * We know that the packet was lost, so back out the
1186 		 * sequence number advance, if any.
1187 		 *
1188 		 * If the error is EPERM the packet got blocked by the
1189 		 * local firewall.  Normally we should terminate the
1190 		 * connection but the blocking may have been spurious
1191 		 * due to a firewall reconfiguration cycle.  So we treat
1192 		 * it like a packet loss and let the retransmit timer and
1193 		 * timeouts do their work over time.
1194 		 * XXX: It is a POLA question whether calling tcp_drop right
1195 		 * away would be the really correct behavior instead.
1196 		 */
1197 		if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1198 		    !tcp_timer_active(tp, TT_PERSIST)) &&
1199 		    ((flags & TH_SYN) == 0) &&
1200 		    (error != EPERM)) {
1201 			if (sack_rxmit) {
1202 				p->rxmit -= len;
1203 				tp->sackhint.sack_bytes_rexmit -= len;
1204 				KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1205 				    ("sackhint bytes rtx >= 0"));
1206 			} else
1207 				tp->snd_nxt -= len;
1208 		}
1209 out:
1210 		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1211 		switch (error) {
1212 		case EPERM:
1213 			tp->t_softerror = error;
1214 			return (error);
1215 		case ENOBUFS:
1216 	                if (!tcp_timer_active(tp, TT_REXMT) &&
1217 			    !tcp_timer_active(tp, TT_PERSIST))
1218 	                        tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1219 			tp->snd_cwnd = tp->t_maxseg;
1220 			return (0);
1221 		case EMSGSIZE:
1222 			/*
1223 			 * For some reason the interface we used initially
1224 			 * to send segments changed to another or lowered
1225 			 * its MTU.
1226 			 *
1227 			 * tcp_mtudisc() will find out the new MTU and as
1228 			 * its last action, initiate retransmission, so it
1229 			 * is important to not do so here.
1230 			 *
1231 			 * If TSO was active we either got an interface
1232 			 * without TSO capabilits or TSO was turned off.
1233 			 * Disable it for this connection as too and
1234 			 * immediatly retry with MSS sized segments generated
1235 			 * by this function.
1236 			 */
1237 			if (tso)
1238 				tp->t_flags &= ~TF_TSO;
1239 			tcp_mtudisc(tp->t_inpcb, 0);
1240 			return (0);
1241 		case EHOSTDOWN:
1242 		case EHOSTUNREACH:
1243 		case ENETDOWN:
1244 		case ENETUNREACH:
1245 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
1246 				tp->t_softerror = error;
1247 				return (0);
1248 			}
1249 			/* FALLTHROUGH */
1250 		default:
1251 			return (error);
1252 		}
1253 	}
1254 	V_tcpstat.tcps_sndtotal++;
1255 
1256 	/*
1257 	 * Data sent (as far as we can tell).
1258 	 * If this advertises a larger window than any other segment,
1259 	 * then remember the size of the advertised window.
1260 	 * Any pending ACK has now been sent.
1261 	 */
1262 	if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1263 		tp->rcv_adv = tp->rcv_nxt + recwin;
1264 	tp->last_ack_sent = tp->rcv_nxt;
1265 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1266 	if (tcp_timer_active(tp, TT_DELACK))
1267 		tcp_timer_activate(tp, TT_DELACK, 0);
1268 #if 0
1269 	/*
1270 	 * This completely breaks TCP if newreno is turned on.  What happens
1271 	 * is that if delayed-acks are turned on on the receiver, this code
1272 	 * on the transmitter effectively destroys the TCP window, forcing
1273 	 * it to four packets (1.5Kx4 = 6K window).
1274 	 */
1275 	if (sendalot && (!V_tcp_do_newreno || --maxburst))
1276 		goto again;
1277 #endif
1278 	if (sendalot)
1279 		goto again;
1280 	return (0);
1281 }
1282 
1283 void
1284 tcp_setpersist(struct tcpcb *tp)
1285 {
1286 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1287 	int tt;
1288 
1289 	if (tcp_timer_active(tp, TT_REXMT))
1290 		panic("tcp_setpersist: retransmit pending");
1291 	/*
1292 	 * Start/restart persistance timer.
1293 	 */
1294 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1295 		      TCPTV_PERSMIN, TCPTV_PERSMAX);
1296 	tcp_timer_activate(tp, TT_PERSIST, tt);
1297 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1298 		tp->t_rxtshift++;
1299 }
1300 
1301 /*
1302  * Insert TCP options according to the supplied parameters to the place
1303  * optp in a consistent way.  Can handle unaligned destinations.
1304  *
1305  * The order of the option processing is crucial for optimal packing and
1306  * alignment for the scarce option space.
1307  *
1308  * The optimal order for a SYN/SYN-ACK segment is:
1309  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1310  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1311  *
1312  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1313  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1314  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1315  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1316  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1317  */
1318 int
1319 tcp_addoptions(struct tcpopt *to, u_char *optp)
1320 {
1321 	u_int mask, optlen = 0;
1322 
1323 	for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1324 		if ((to->to_flags & mask) != mask)
1325 			continue;
1326 		if (optlen == TCP_MAXOLEN)
1327 			break;
1328 		switch (to->to_flags & mask) {
1329 		case TOF_MSS:
1330 			while (optlen % 4) {
1331 				optlen += TCPOLEN_NOP;
1332 				*optp++ = TCPOPT_NOP;
1333 			}
1334 			if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1335 				continue;
1336 			optlen += TCPOLEN_MAXSEG;
1337 			*optp++ = TCPOPT_MAXSEG;
1338 			*optp++ = TCPOLEN_MAXSEG;
1339 			to->to_mss = htons(to->to_mss);
1340 			bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1341 			optp += sizeof(to->to_mss);
1342 			break;
1343 		case TOF_SCALE:
1344 			while (!optlen || optlen % 2 != 1) {
1345 				optlen += TCPOLEN_NOP;
1346 				*optp++ = TCPOPT_NOP;
1347 			}
1348 			if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1349 				continue;
1350 			optlen += TCPOLEN_WINDOW;
1351 			*optp++ = TCPOPT_WINDOW;
1352 			*optp++ = TCPOLEN_WINDOW;
1353 			*optp++ = to->to_wscale;
1354 			break;
1355 		case TOF_SACKPERM:
1356 			while (optlen % 2) {
1357 				optlen += TCPOLEN_NOP;
1358 				*optp++ = TCPOPT_NOP;
1359 			}
1360 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1361 				continue;
1362 			optlen += TCPOLEN_SACK_PERMITTED;
1363 			*optp++ = TCPOPT_SACK_PERMITTED;
1364 			*optp++ = TCPOLEN_SACK_PERMITTED;
1365 			break;
1366 		case TOF_TS:
1367 			while (!optlen || optlen % 4 != 2) {
1368 				optlen += TCPOLEN_NOP;
1369 				*optp++ = TCPOPT_NOP;
1370 			}
1371 			if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1372 				continue;
1373 			optlen += TCPOLEN_TIMESTAMP;
1374 			*optp++ = TCPOPT_TIMESTAMP;
1375 			*optp++ = TCPOLEN_TIMESTAMP;
1376 			to->to_tsval = htonl(to->to_tsval);
1377 			to->to_tsecr = htonl(to->to_tsecr);
1378 			bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1379 			optp += sizeof(to->to_tsval);
1380 			bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1381 			optp += sizeof(to->to_tsecr);
1382 			break;
1383 		case TOF_SIGNATURE:
1384 			{
1385 			int siglen = TCPOLEN_SIGNATURE - 2;
1386 
1387 			while (!optlen || optlen % 4 != 2) {
1388 				optlen += TCPOLEN_NOP;
1389 				*optp++ = TCPOPT_NOP;
1390 			}
1391 			if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE)
1392 				continue;
1393 			optlen += TCPOLEN_SIGNATURE;
1394 			*optp++ = TCPOPT_SIGNATURE;
1395 			*optp++ = TCPOLEN_SIGNATURE;
1396 			to->to_signature = optp;
1397 			while (siglen--)
1398 				 *optp++ = 0;
1399 			break;
1400 			}
1401 		case TOF_SACK:
1402 			{
1403 			int sackblks = 0;
1404 			struct sackblk *sack = (struct sackblk *)to->to_sacks;
1405 			tcp_seq sack_seq;
1406 
1407 			while (!optlen || optlen % 4 != 2) {
1408 				optlen += TCPOLEN_NOP;
1409 				*optp++ = TCPOPT_NOP;
1410 			}
1411 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1412 				continue;
1413 			optlen += TCPOLEN_SACKHDR;
1414 			*optp++ = TCPOPT_SACK;
1415 			sackblks = min(to->to_nsacks,
1416 					(TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1417 			*optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1418 			while (sackblks--) {
1419 				sack_seq = htonl(sack->start);
1420 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1421 				optp += sizeof(sack_seq);
1422 				sack_seq = htonl(sack->end);
1423 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1424 				optp += sizeof(sack_seq);
1425 				optlen += TCPOLEN_SACK;
1426 				sack++;
1427 			}
1428 			V_tcpstat.tcps_sack_send_blocks++;
1429 			break;
1430 			}
1431 		default:
1432 			panic("%s: unknown TCP option type", __func__);
1433 			break;
1434 		}
1435 	}
1436 
1437 	/* Terminate and pad TCP options to a 4 byte boundary. */
1438 	if (optlen % 4) {
1439 		optlen += TCPOLEN_EOL;
1440 		*optp++ = TCPOPT_EOL;
1441 	}
1442 	/*
1443 	 * According to RFC 793 (STD0007):
1444 	 *   "The content of the header beyond the End-of-Option option
1445 	 *    must be header padding (i.e., zero)."
1446 	 *   and later: "The padding is composed of zeros."
1447 	 */
1448 	while (optlen % 4) {
1449 		optlen += TCPOLEN_PAD;
1450 		*optp++ = TCPOPT_PAD;
1451 	}
1452 
1453 	KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1454 	return (optlen);
1455 }
1456