xref: /freebsd/sys/netinet/tcp_output.c (revision 35a04710d7286aa9538917fd7f8e417dbee95b82)
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 
53 #include <net/route.h>
54 
55 #include <netinet/in.h>
56 #include <netinet/in_systm.h>
57 #include <netinet/ip.h>
58 #include <netinet/in_pcb.h>
59 #include <netinet/ip_var.h>
60 #include <netinet/ip_options.h>
61 #ifdef INET6
62 #include <netinet6/in6_pcb.h>
63 #include <netinet/ip6.h>
64 #include <netinet6/ip6_var.h>
65 #endif
66 #include <netinet/tcp.h>
67 #define	TCPOUTFLAGS
68 #include <netinet/tcp_fsm.h>
69 #include <netinet/tcp_seq.h>
70 #include <netinet/tcp_timer.h>
71 #include <netinet/tcp_var.h>
72 #include <netinet/tcpip.h>
73 #ifdef TCPDEBUG
74 #include <netinet/tcp_debug.h>
75 #endif
76 
77 #ifdef IPSEC
78 #include <netipsec/ipsec.h>
79 #endif /*IPSEC*/
80 
81 #include <machine/in_cksum.h>
82 
83 #include <security/mac/mac_framework.h>
84 
85 #ifdef notyet
86 extern struct mbuf *m_copypack();
87 #endif
88 
89 int path_mtu_discovery = 1;
90 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
91 	&path_mtu_discovery, 1, "Enable Path MTU Discovery");
92 
93 int ss_fltsz = 1;
94 SYSCTL_INT(_net_inet_tcp, OID_AUTO, slowstart_flightsize, CTLFLAG_RW,
95 	&ss_fltsz, 1, "Slow start flight size");
96 
97 int ss_fltsz_local = 4;
98 SYSCTL_INT(_net_inet_tcp, OID_AUTO, local_slowstart_flightsize, CTLFLAG_RW,
99 	&ss_fltsz_local, 1, "Slow start flight size for local networks");
100 
101 int     tcp_do_newreno = 1;
102 SYSCTL_INT(_net_inet_tcp, OID_AUTO, newreno, CTLFLAG_RW,
103 	&tcp_do_newreno, 0, "Enable NewReno Algorithms");
104 
105 int	tcp_do_tso = 1;
106 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW,
107 	&tcp_do_tso, 0, "Enable TCP Segmentation Offload");
108 
109 int	tcp_do_autosndbuf = 1;
110 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW,
111 	&tcp_do_autosndbuf, 0, "Enable automatic send buffer sizing");
112 
113 int	tcp_autosndbuf_inc = 8*1024;
114 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW,
115 	&tcp_autosndbuf_inc, 0, "Incrementor step size of automatic send buffer");
116 
117 int	tcp_autosndbuf_max = 256*1024;
118 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW,
119 	&tcp_autosndbuf_max, 0, "Max size of automatic send buffer");
120 
121 
122 /*
123  * Tcp output routine: figure out what should be sent and send it.
124  */
125 int
126 tcp_output(struct tcpcb *tp)
127 {
128 	struct socket *so = tp->t_inpcb->inp_socket;
129 	long len, recwin, sendwin;
130 	int off, flags, error;
131 #ifdef TCP_SIGNATURE
132 	int sigoff = 0;
133 #endif
134 	struct mbuf *m;
135 	struct ip *ip = NULL;
136 	struct ipovly *ipov = NULL;
137 	struct tcphdr *th;
138 	u_char opt[TCP_MAXOLEN];
139 	unsigned ipoptlen, optlen, hdrlen;
140 #ifdef IPSEC
141 	unsigned ipsec_optlen = 0;
142 #endif
143 	int idle, sendalot;
144 	int sack_rxmit, sack_bytes_rxmt;
145 	struct sackhole *p;
146 	int tso = 0;
147 	struct tcpopt to;
148 #if 0
149 	int maxburst = TCP_MAXBURST;
150 #endif
151 #ifdef INET6
152 	struct ip6_hdr *ip6 = NULL;
153 	int isipv6;
154 
155 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
156 #endif
157 
158 	INP_LOCK_ASSERT(tp->t_inpcb);
159 
160 	/*
161 	 * Determine length of data that should be transmitted,
162 	 * and flags that will be used.
163 	 * If there is some data or critical controls (SYN, RST)
164 	 * to send, then transmit; otherwise, investigate further.
165 	 */
166 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
167 	if (idle && (ticks - tp->t_rcvtime) >= tp->t_rxtcur) {
168 		/*
169 		 * We have been idle for "a while" and no acks are
170 		 * expected to clock out any data we send --
171 		 * slow start to get ack "clock" running again.
172 		 *
173 		 * Set the slow-start flight size depending on whether
174 		 * this is a local network or not.
175 		 */
176 		int ss = ss_fltsz;
177 #ifdef INET6
178 		if (isipv6) {
179 			if (in6_localaddr(&tp->t_inpcb->in6p_faddr))
180 				ss = ss_fltsz_local;
181 		} else
182 #endif /* INET6 */
183 		if (in_localaddr(tp->t_inpcb->inp_faddr))
184 			ss = ss_fltsz_local;
185 		tp->snd_cwnd = tp->t_maxseg * ss;
186 	}
187 	tp->t_flags &= ~TF_LASTIDLE;
188 	if (idle) {
189 		if (tp->t_flags & TF_MORETOCOME) {
190 			tp->t_flags |= TF_LASTIDLE;
191 			idle = 0;
192 		}
193 	}
194 again:
195 	/*
196 	 * If we've recently taken a timeout, snd_max will be greater than
197 	 * snd_nxt.  There may be SACK information that allows us to avoid
198 	 * resending already delivered data.  Adjust snd_nxt accordingly.
199 	 */
200 	if ((tp->t_flags & TF_SACK_PERMIT) &&
201 	    SEQ_LT(tp->snd_nxt, tp->snd_max))
202 		tcp_sack_adjust(tp);
203 	sendalot = 0;
204 	off = tp->snd_nxt - tp->snd_una;
205 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
206 	sendwin = min(sendwin, tp->snd_bwnd);
207 
208 	flags = tcp_outflags[tp->t_state];
209 	/*
210 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
211 	 * to send out new data (when sendalot is 1), bypass this function.
212 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
213 	 * we're replacing a (future) new transmission with a retransmission
214 	 * now, and we previously incremented snd_cwnd in tcp_input().
215 	 */
216 	/*
217 	 * Still in sack recovery , reset rxmit flag to zero.
218 	 */
219 	sack_rxmit = 0;
220 	sack_bytes_rxmt = 0;
221 	len = 0;
222 	p = NULL;
223 	if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp) &&
224 	    (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
225 		long cwin;
226 
227 		cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
228 		if (cwin < 0)
229 			cwin = 0;
230 		/* Do not retransmit SACK segments beyond snd_recover */
231 		if (SEQ_GT(p->end, tp->snd_recover)) {
232 			/*
233 			 * (At least) part of sack hole extends beyond
234 			 * snd_recover. Check to see if we can rexmit data
235 			 * for this hole.
236 			 */
237 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
238 				/*
239 				 * Can't rexmit any more data for this hole.
240 				 * That data will be rexmitted in the next
241 				 * sack recovery episode, when snd_recover
242 				 * moves past p->rxmit.
243 				 */
244 				p = NULL;
245 				goto after_sack_rexmit;
246 			} else
247 				/* Can rexmit part of the current hole */
248 				len = ((long)ulmin(cwin,
249 						   tp->snd_recover - p->rxmit));
250 		} else
251 			len = ((long)ulmin(cwin, p->end - p->rxmit));
252 		off = p->rxmit - tp->snd_una;
253 		KASSERT(off >= 0,("%s: sack block to the left of una : %d",
254 		    __func__, off));
255 		if (len > 0) {
256 			sack_rxmit = 1;
257 			sendalot = 1;
258 			tcpstat.tcps_sack_rexmits++;
259 			tcpstat.tcps_sack_rexmit_bytes +=
260 			    min(len, tp->t_maxseg);
261 		}
262 	}
263 after_sack_rexmit:
264 	/*
265 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
266 	 * state flags.
267 	 */
268 	if (tp->t_flags & TF_NEEDFIN)
269 		flags |= TH_FIN;
270 	if (tp->t_flags & TF_NEEDSYN)
271 		flags |= TH_SYN;
272 
273 	SOCKBUF_LOCK(&so->so_snd);
274 	/*
275 	 * If in persist timeout with window of 0, send 1 byte.
276 	 * Otherwise, if window is small but nonzero
277 	 * and timer expired, we will send what we can
278 	 * and go to transmit state.
279 	 */
280 	if (tp->t_flags & TF_FORCEDATA) {
281 		if (sendwin == 0) {
282 			/*
283 			 * If we still have some data to send, then
284 			 * clear the FIN bit.  Usually this would
285 			 * happen below when it realizes that we
286 			 * aren't sending all the data.  However,
287 			 * if we have exactly 1 byte of unsent data,
288 			 * then it won't clear the FIN bit below,
289 			 * and if we are in persist state, we wind
290 			 * up sending the packet without recording
291 			 * that we sent the FIN bit.
292 			 *
293 			 * We can't just blindly clear the FIN bit,
294 			 * because if we don't have any more data
295 			 * to send then the probe will be the FIN
296 			 * itself.
297 			 */
298 			if (off < so->so_snd.sb_cc)
299 				flags &= ~TH_FIN;
300 			sendwin = 1;
301 		} else {
302 			tcp_timer_activate(tp, TT_PERSIST, 0);
303 			tp->t_rxtshift = 0;
304 		}
305 	}
306 
307 	/*
308 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
309 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
310 	 * a negative length.  This can also occur when TCP opens up
311 	 * its congestion window while receiving additional duplicate
312 	 * acks after fast-retransmit because TCP will reset snd_nxt
313 	 * to snd_max after the fast-retransmit.
314 	 *
315 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
316 	 * be set to snd_una, the offset will be 0, and the length may
317 	 * wind up 0.
318 	 *
319 	 * If sack_rxmit is true we are retransmitting from the scoreboard
320 	 * in which case len is already set.
321 	 */
322 	if (sack_rxmit == 0) {
323 		if (sack_bytes_rxmt == 0)
324 			len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off);
325 		else {
326 			long cwin;
327 
328                         /*
329 			 * We are inside of a SACK recovery episode and are
330 			 * sending new data, having retransmitted all the
331 			 * data possible in the scoreboard.
332 			 */
333 			len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd)
334 			       - off);
335 			/*
336 			 * Don't remove this (len > 0) check !
337 			 * We explicitly check for len > 0 here (although it
338 			 * isn't really necessary), to work around a gcc
339 			 * optimization issue - to force gcc to compute
340 			 * len above. Without this check, the computation
341 			 * of len is bungled by the optimizer.
342 			 */
343 			if (len > 0) {
344 				cwin = tp->snd_cwnd -
345 					(tp->snd_nxt - tp->sack_newdata) -
346 					sack_bytes_rxmt;
347 				if (cwin < 0)
348 					cwin = 0;
349 				len = lmin(len, cwin);
350 			}
351 		}
352 	}
353 
354 	/*
355 	 * Lop off SYN bit if it has already been sent.  However, if this
356 	 * is SYN-SENT state and if segment contains data and if we don't
357 	 * know that foreign host supports TAO, suppress sending segment.
358 	 */
359 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
360 		if (tp->t_state != TCPS_SYN_RECEIVED)
361 			flags &= ~TH_SYN;
362 		off--, len++;
363 	}
364 
365 	/*
366 	 * Be careful not to send data and/or FIN on SYN segments.
367 	 * This measure is needed to prevent interoperability problems
368 	 * with not fully conformant TCP implementations.
369 	 */
370 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
371 		len = 0;
372 		flags &= ~TH_FIN;
373 	}
374 
375 	if (len < 0) {
376 		/*
377 		 * If FIN has been sent but not acked,
378 		 * but we haven't been called to retransmit,
379 		 * len will be < 0.  Otherwise, window shrank
380 		 * after we sent into it.  If window shrank to 0,
381 		 * cancel pending retransmit, pull snd_nxt back
382 		 * to (closed) window, and set the persist timer
383 		 * if it isn't already going.  If the window didn't
384 		 * close completely, just wait for an ACK.
385 		 */
386 		len = 0;
387 		if (sendwin == 0) {
388 			tcp_timer_activate(tp, TT_REXMT, 0);
389 			tp->t_rxtshift = 0;
390 			tp->snd_nxt = tp->snd_una;
391 			if (!tcp_timer_active(tp, TT_PERSIST))
392 				tcp_setpersist(tp);
393 		}
394 	}
395 
396 	/* len will be >= 0 after this point. */
397 	KASSERT(len >= 0, ("%s: len < 0", __func__));
398 
399 	/*
400 	 * Automatic sizing of send socket buffer.  Often the send buffer
401 	 * size is not optimally adjusted to the actual network conditions
402 	 * at hand (delay bandwidth product).  Setting the buffer size too
403 	 * small limits throughput on links with high bandwidth and high
404 	 * delay (eg. trans-continental/oceanic links).  Setting the
405 	 * buffer size too big consumes too much real kernel memory,
406 	 * especially with many connections on busy servers.
407 	 *
408 	 * The criteria to step up the send buffer one notch are:
409 	 *  1. receive window of remote host is larger than send buffer
410 	 *     (with a fudge factor of 5/4th);
411 	 *  2. send buffer is filled to 7/8th with data (so we actually
412 	 *     have data to make use of it);
413 	 *  3. send buffer fill has not hit maximal automatic size;
414 	 *  4. our send window (slow start and cogestion controlled) is
415 	 *     larger than sent but unacknowledged data in send buffer.
416 	 *
417 	 * The remote host receive window scaling factor may limit the
418 	 * growing of the send buffer before it reaches its allowed
419 	 * maximum.
420 	 *
421 	 * It scales directly with slow start or congestion window
422 	 * and does at most one step per received ACK.  This fast
423 	 * scaling has the drawback of growing the send buffer beyond
424 	 * what is strictly necessary to make full use of a given
425 	 * delay*bandwith product.  However testing has shown this not
426 	 * to be much of an problem.  At worst we are trading wasting
427 	 * of available bandwith (the non-use of it) for wasting some
428 	 * socket buffer memory.
429 	 *
430 	 * TODO: Shrink send buffer during idle periods together
431 	 * with congestion window.  Requires another timer.  Has to
432 	 * wait for upcoming tcp timer rewrite.
433 	 */
434 	if (tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
435 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
436 		    so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
437 		    so->so_snd.sb_cc < tcp_autosndbuf_max &&
438 		    sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
439 			if (!sbreserve_locked(&so->so_snd,
440 			    min(so->so_snd.sb_hiwat + tcp_autosndbuf_inc,
441 			     tcp_autosndbuf_max), so, curthread))
442 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
443 		}
444 	}
445 
446 	/*
447 	 * Truncate to the maximum segment length or enable TCP Segmentation
448 	 * Offloading (if supported by hardware) and ensure that FIN is removed
449 	 * if the length no longer contains the last data byte.
450 	 *
451 	 * TSO may only be used if we are in a pure bulk sending state.  The
452 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
453 	 * IP options prevent using TSO.  With TSO the TCP header is the same
454 	 * (except for the sequence number) for all generated packets.  This
455 	 * makes it impossible to transmit any options which vary per generated
456 	 * segment or packet.
457 	 *
458 	 * The length of TSO bursts is limited to TCP_MAXWIN.  That limit and
459 	 * removal of FIN (if not already catched here) are handled later after
460 	 * the exact length of the TCP options are known.
461 	 */
462 #ifdef IPSEC
463 	/*
464 	 * Pre-calculate here as we save another lookup into the darknesses
465 	 * of IPsec that way and can actually decide if TSO is ok.
466 	 */
467 	ipsec_optlen = ipsec_hdrsiz_tcp(tp);
468 #endif
469 	if (len > tp->t_maxseg) {
470 		if ((tp->t_flags & TF_TSO) && tcp_do_tso &&
471 		    ((tp->t_flags & TF_SIGNATURE) == 0) &&
472 		    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
473 		    tp->t_inpcb->inp_options == NULL &&
474 		    tp->t_inpcb->in6p_options == NULL
475 #ifdef IPSEC
476 		    && ipsec_optlen == 0
477 #endif
478 		    ) {
479 			tso = 1;
480 		} else {
481 			len = tp->t_maxseg;
482 			sendalot = 1;
483 			tso = 0;
484 		}
485 	}
486 	if (sack_rxmit) {
487 		if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
488 			flags &= ~TH_FIN;
489 	} else {
490 		if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
491 			flags &= ~TH_FIN;
492 	}
493 
494 	recwin = sbspace(&so->so_rcv);
495 
496 	/*
497 	 * Sender silly window avoidance.   We transmit under the following
498 	 * conditions when len is non-zero:
499 	 *
500 	 *	- We have a full segment (or more with TSO)
501 	 *	- This is the last buffer in a write()/send() and we are
502 	 *	  either idle or running NODELAY
503 	 *	- we've timed out (e.g. persist timer)
504 	 *	- we have more then 1/2 the maximum send window's worth of
505 	 *	  data (receiver may be limited the window size)
506 	 *	- we need to retransmit
507 	 */
508 	if (len) {
509 		if (len >= tp->t_maxseg)
510 			goto send;
511 		/*
512 		 * NOTE! on localhost connections an 'ack' from the remote
513 		 * end may occur synchronously with the output and cause
514 		 * us to flush a buffer queued with moretocome.  XXX
515 		 *
516 		 * note: the len + off check is almost certainly unnecessary.
517 		 */
518 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
519 		    (idle || (tp->t_flags & TF_NODELAY)) &&
520 		    len + off >= so->so_snd.sb_cc &&
521 		    (tp->t_flags & TF_NOPUSH) == 0) {
522 			goto send;
523 		}
524 		if (tp->t_flags & TF_FORCEDATA)		/* typ. timeout case */
525 			goto send;
526 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
527 			goto send;
528 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
529 			goto send;
530 		if (sack_rxmit)
531 			goto send;
532 	}
533 
534 	/*
535 	 * Compare available window to amount of window
536 	 * known to peer (as advertised window less
537 	 * next expected input).  If the difference is at least two
538 	 * max size segments, or at least 50% of the maximum possible
539 	 * window, then want to send a window update to peer.
540 	 * Skip this if the connection is in T/TCP half-open state.
541 	 * Don't send pure window updates when the peer has closed
542 	 * the connection and won't ever send more data.
543 	 */
544 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
545 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
546 		/*
547 		 * "adv" is the amount we can increase the window,
548 		 * taking into account that we are limited by
549 		 * TCP_MAXWIN << tp->rcv_scale.
550 		 */
551 		long adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale) -
552 			(tp->rcv_adv - tp->rcv_nxt);
553 
554 		if (adv >= (long) (2 * tp->t_maxseg))
555 			goto send;
556 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
557 			goto send;
558 	}
559 
560 	/*
561 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
562 	 * is also a catch-all for the retransmit timer timeout case.
563 	 */
564 	if (tp->t_flags & TF_ACKNOW)
565 		goto send;
566 	if ((flags & TH_RST) ||
567 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
568 		goto send;
569 	if (SEQ_GT(tp->snd_up, tp->snd_una))
570 		goto send;
571 	/*
572 	 * If our state indicates that FIN should be sent
573 	 * and we have not yet done so, then we need to send.
574 	 */
575 	if (flags & TH_FIN &&
576 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
577 		goto send;
578 	/*
579 	 * In SACK, it is possible for tcp_output to fail to send a segment
580 	 * after the retransmission timer has been turned off.  Make sure
581 	 * that the retransmission timer is set.
582 	 */
583 	if ((tp->t_flags & TF_SACK_PERMIT) &&
584 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
585 	    !tcp_timer_active(tp, TT_REXMT) &&
586 	    !tcp_timer_active(tp, TT_PERSIST)) {
587 		tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
588 		goto just_return;
589 	}
590 	/*
591 	 * TCP window updates are not reliable, rather a polling protocol
592 	 * using ``persist'' packets is used to insure receipt of window
593 	 * updates.  The three ``states'' for the output side are:
594 	 *	idle			not doing retransmits or persists
595 	 *	persisting		to move a small or zero window
596 	 *	(re)transmitting	and thereby not persisting
597 	 *
598 	 * tcp_timer_active(tp, TT_PERSIST)
599 	 *	is true when we are in persist state.
600 	 * (tp->t_flags & TF_FORCEDATA)
601 	 *	is set when we are called to send a persist packet.
602 	 * tcp_timer_active(tp, TT_REXMT)
603 	 *	is set when we are retransmitting
604 	 * The output side is idle when both timers are zero.
605 	 *
606 	 * If send window is too small, there is data to transmit, and no
607 	 * retransmit or persist is pending, then go to persist state.
608 	 * If nothing happens soon, send when timer expires:
609 	 * if window is nonzero, transmit what we can,
610 	 * otherwise force out a byte.
611 	 */
612 	if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) &&
613 	    !tcp_timer_active(tp, TT_PERSIST)) {
614 		tp->t_rxtshift = 0;
615 		tcp_setpersist(tp);
616 	}
617 
618 	/*
619 	 * No reason to send a segment, just return.
620 	 */
621 just_return:
622 	SOCKBUF_UNLOCK(&so->so_snd);
623 	return (0);
624 
625 send:
626 	SOCKBUF_LOCK_ASSERT(&so->so_snd);
627 	/*
628 	 * Before ESTABLISHED, force sending of initial options
629 	 * unless TCP set not to do any options.
630 	 * NOTE: we assume that the IP/TCP header plus TCP options
631 	 * always fit in a single mbuf, leaving room for a maximum
632 	 * link header, i.e.
633 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
634 	 */
635 	optlen = 0;
636 #ifdef INET6
637 	if (isipv6)
638 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
639 	else
640 #endif
641 	hdrlen = sizeof (struct tcpiphdr);
642 
643 	/*
644 	 * Compute options for segment.
645 	 * We only have to care about SYN and established connection
646 	 * segments.  Options for SYN-ACK segments are handled in TCP
647 	 * syncache.
648 	 */
649 	if ((tp->t_flags & TF_NOOPT) == 0) {
650 		to.to_flags = 0;
651 		/* Maximum segment size. */
652 		if (flags & TH_SYN) {
653 			tp->snd_nxt = tp->iss;
654 			to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
655 			to.to_flags |= TOF_MSS;
656 		}
657 		/* Window scaling. */
658 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
659 			to.to_wscale = tp->request_r_scale;
660 			to.to_flags |= TOF_SCALE;
661 		}
662 		/* Timestamps. */
663 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
664 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
665 			to.to_tsval = ticks + tp->ts_offset;
666 			to.to_tsecr = tp->ts_recent;
667 			to.to_flags |= TOF_TS;
668 			/* Set receive buffer autosizing timestamp. */
669 			if (tp->rfbuf_ts == 0 &&
670 			    (so->so_rcv.sb_flags & SB_AUTOSIZE))
671 				tp->rfbuf_ts = ticks;
672 		}
673 		/* Selective ACK's. */
674 		if (tp->t_flags & TF_SACK_PERMIT) {
675 			if (flags & TH_SYN)
676 				to.to_flags |= TOF_SACKPERM;
677 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
678 			    (tp->t_flags & TF_SACK_PERMIT) &&
679 			    tp->rcv_numsacks > 0) {
680 				to.to_flags |= TOF_SACK;
681 				to.to_nsacks = tp->rcv_numsacks;
682 				to.to_sacks = (u_char *)tp->sackblks;
683 			}
684 		}
685 #ifdef TCP_SIGNATURE
686 		/* TCP-MD5 (RFC2385). */
687 #ifdef INET6
688 		if (!isipv6 && (tp->t_flags & TF_SIGNATURE))
689 #else
690 		if (tp->t_flags & TF_SIGNATURE)
691 #endif /* INET6 */
692 			to.to_flags |= TOF_SIGNATURE;
693 #endif /* TCP_SIGNATURE */
694 
695 		/* Processing the options. */
696 		hdrlen += optlen = tcp_addoptions(&to, opt);
697 
698 #ifdef TCP_SIGNATURE
699 		sigoff = to.to_signature - (u_char *)&to;
700 #endif /* TCP_SIGNATURE */
701 	}
702 
703 #ifdef INET6
704 	if (isipv6)
705 		ipoptlen = ip6_optlen(tp->t_inpcb);
706 	else
707 #endif
708 	if (tp->t_inpcb->inp_options)
709 		ipoptlen = tp->t_inpcb->inp_options->m_len -
710 				offsetof(struct ipoption, ipopt_list);
711 	else
712 		ipoptlen = 0;
713 #ifdef IPSEC
714 	ipoptlen += ipsec_optlen;
715 #endif
716 
717 	/*
718 	 * Adjust data length if insertion of options will
719 	 * bump the packet length beyond the t_maxopd length.
720 	 * Clear the FIN bit because we cut off the tail of
721 	 * the segment.
722 	 *
723 	 * When doing TSO limit a burst to TCP_MAXWIN minus the
724 	 * IP, TCP and Options length to keep ip->ip_len from
725 	 * overflowing.  Prevent the last segment from being
726 	 * fractional thus making them all equal sized and set
727 	 * the flag to continue sending.  TSO is disabled when
728 	 * IP options or IPSEC are present.
729 	 */
730 	if (len + optlen + ipoptlen > tp->t_maxopd) {
731 		flags &= ~TH_FIN;
732 		if (tso) {
733 			if (len > TCP_MAXWIN - hdrlen - optlen) {
734 				len = TCP_MAXWIN - hdrlen - optlen;
735 				len = len - (len % (tp->t_maxopd - optlen));
736 				sendalot = 1;
737 			} else if (tp->t_flags & TF_NEEDFIN)
738 				sendalot = 1;
739 		} else {
740 			len = tp->t_maxopd - optlen - ipoptlen;
741 			sendalot = 1;
742 		}
743 	}
744 
745 /*#ifdef DIAGNOSTIC*/
746 #ifdef INET6
747 	if (max_linkhdr + hdrlen > MCLBYTES)
748 #else
749 	if (max_linkhdr + hdrlen > MHLEN)
750 #endif
751 		panic("tcphdr too big");
752 /*#endif*/
753 
754 	/*
755 	 * Grab a header mbuf, attaching a copy of data to
756 	 * be transmitted, and initialize the header from
757 	 * the template for sends on this connection.
758 	 */
759 	if (len) {
760 		struct mbuf *mb;
761 		u_int moff;
762 
763 		if ((tp->t_flags & TF_FORCEDATA) && len == 1)
764 			tcpstat.tcps_sndprobe++;
765 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
766 			tcpstat.tcps_sndrexmitpack++;
767 			tcpstat.tcps_sndrexmitbyte += len;
768 		} else {
769 			tcpstat.tcps_sndpack++;
770 			tcpstat.tcps_sndbyte += len;
771 		}
772 #ifdef notyet
773 		if ((m = m_copypack(so->so_snd.sb_mb, off,
774 		    (int)len, max_linkhdr + hdrlen)) == 0) {
775 			SOCKBUF_UNLOCK(&so->so_snd);
776 			error = ENOBUFS;
777 			goto out;
778 		}
779 		/*
780 		 * m_copypack left space for our hdr; use it.
781 		 */
782 		m->m_len += hdrlen;
783 		m->m_data -= hdrlen;
784 #else
785 		MGETHDR(m, M_DONTWAIT, MT_DATA);
786 		if (m == NULL) {
787 			SOCKBUF_UNLOCK(&so->so_snd);
788 			error = ENOBUFS;
789 			goto out;
790 		}
791 #ifdef INET6
792 		if (MHLEN < hdrlen + max_linkhdr) {
793 			MCLGET(m, M_DONTWAIT);
794 			if ((m->m_flags & M_EXT) == 0) {
795 				SOCKBUF_UNLOCK(&so->so_snd);
796 				m_freem(m);
797 				error = ENOBUFS;
798 				goto out;
799 			}
800 		}
801 #endif
802 		m->m_data += max_linkhdr;
803 		m->m_len = hdrlen;
804 
805 		/*
806 		 * Start the m_copy functions from the closest mbuf
807 		 * to the offset in the socket buffer chain.
808 		 */
809 		mb = sbsndptr(&so->so_snd, off, len, &moff);
810 
811 		if (len <= MHLEN - hdrlen - max_linkhdr) {
812 			m_copydata(mb, moff, (int)len,
813 			    mtod(m, caddr_t) + hdrlen);
814 			m->m_len += len;
815 		} else {
816 			m->m_next = m_copy(mb, moff, (int)len);
817 			if (m->m_next == NULL) {
818 				SOCKBUF_UNLOCK(&so->so_snd);
819 				(void) m_free(m);
820 				error = ENOBUFS;
821 				goto out;
822 			}
823 		}
824 #endif
825 		/*
826 		 * If we're sending everything we've got, set PUSH.
827 		 * (This will keep happy those implementations which only
828 		 * give data to the user when a buffer fills or
829 		 * a PUSH comes in.)
830 		 */
831 		if (off + len == so->so_snd.sb_cc)
832 			flags |= TH_PUSH;
833 		SOCKBUF_UNLOCK(&so->so_snd);
834 	} else {
835 		SOCKBUF_UNLOCK(&so->so_snd);
836 		if (tp->t_flags & TF_ACKNOW)
837 			tcpstat.tcps_sndacks++;
838 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
839 			tcpstat.tcps_sndctrl++;
840 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
841 			tcpstat.tcps_sndurg++;
842 		else
843 			tcpstat.tcps_sndwinup++;
844 
845 		MGETHDR(m, M_DONTWAIT, MT_DATA);
846 		if (m == NULL) {
847 			error = ENOBUFS;
848 			goto out;
849 		}
850 #ifdef INET6
851 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
852 		    MHLEN >= hdrlen) {
853 			MH_ALIGN(m, hdrlen);
854 		} else
855 #endif
856 		m->m_data += max_linkhdr;
857 		m->m_len = hdrlen;
858 	}
859 	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
860 	m->m_pkthdr.rcvif = (struct ifnet *)0;
861 #ifdef MAC
862 	mac_inpcb_create_mbuf(tp->t_inpcb, m);
863 #endif
864 #ifdef INET6
865 	if (isipv6) {
866 		ip6 = mtod(m, struct ip6_hdr *);
867 		th = (struct tcphdr *)(ip6 + 1);
868 		tcpip_fillheaders(tp->t_inpcb, ip6, th);
869 	} else
870 #endif /* INET6 */
871 	{
872 		ip = mtod(m, struct ip *);
873 		ipov = (struct ipovly *)ip;
874 		th = (struct tcphdr *)(ip + 1);
875 		tcpip_fillheaders(tp->t_inpcb, ip, th);
876 	}
877 
878 	/*
879 	 * Fill in fields, remembering maximum advertised
880 	 * window for use in delaying messages about window sizes.
881 	 * If resending a FIN, be sure not to use a new sequence number.
882 	 */
883 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
884 	    tp->snd_nxt == tp->snd_max)
885 		tp->snd_nxt--;
886 	/*
887 	 * If we are doing retransmissions, then snd_nxt will
888 	 * not reflect the first unsent octet.  For ACK only
889 	 * packets, we do not want the sequence number of the
890 	 * retransmitted packet, we want the sequence number
891 	 * of the next unsent octet.  So, if there is no data
892 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
893 	 * when filling in ti_seq.  But if we are in persist
894 	 * state, snd_max might reflect one byte beyond the
895 	 * right edge of the window, so use snd_nxt in that
896 	 * case, since we know we aren't doing a retransmission.
897 	 * (retransmit and persist are mutually exclusive...)
898 	 */
899 	if (sack_rxmit == 0) {
900 		if (len || (flags & (TH_SYN|TH_FIN)) ||
901 		    tcp_timer_active(tp, TT_PERSIST))
902 			th->th_seq = htonl(tp->snd_nxt);
903 		else
904 			th->th_seq = htonl(tp->snd_max);
905 	} else {
906 		th->th_seq = htonl(p->rxmit);
907 		p->rxmit += len;
908 		tp->sackhint.sack_bytes_rexmit += len;
909 	}
910 	th->th_ack = htonl(tp->rcv_nxt);
911 	if (optlen) {
912 		bcopy(opt, th + 1, optlen);
913 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
914 	}
915 	th->th_flags = flags;
916 	/*
917 	 * Calculate receive window.  Don't shrink window,
918 	 * but avoid silly window syndrome.
919 	 */
920 	if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
921 	    recwin < (long)tp->t_maxseg)
922 		recwin = 0;
923 	if (recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
924 		recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
925 	if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
926 		recwin = (long)TCP_MAXWIN << tp->rcv_scale;
927 
928 	/*
929 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
930 	 * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
931 	 * case is handled in syncache.
932 	 */
933 	if (flags & TH_SYN)
934 		th->th_win = htons((u_short)
935 				(min(sbspace(&so->so_rcv), TCP_MAXWIN)));
936 	else
937 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
938 
939 	/*
940 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
941 	 * a 0 window.  This may cause the remote transmitter to stall.  This
942 	 * flag tells soreceive() to disable delayed acknowledgements when
943 	 * draining the buffer.  This can occur if the receiver is attempting
944 	 * to read more data then can be buffered prior to transmitting on
945 	 * the connection.
946 	 */
947 	if (recwin == 0)
948 		tp->t_flags |= TF_RXWIN0SENT;
949 	else
950 		tp->t_flags &= ~TF_RXWIN0SENT;
951 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
952 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
953 		th->th_flags |= TH_URG;
954 	} else
955 		/*
956 		 * If no urgent pointer to send, then we pull
957 		 * the urgent pointer to the left edge of the send window
958 		 * so that it doesn't drift into the send window on sequence
959 		 * number wraparound.
960 		 */
961 		tp->snd_up = tp->snd_una;		/* drag it along */
962 
963 #ifdef TCP_SIGNATURE
964 #ifdef INET6
965 	if (!isipv6)
966 #endif
967 	if (tp->t_flags & TF_SIGNATURE)
968 		tcp_signature_compute(m, sizeof(struct ip), len, optlen,
969 		    (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
970 #endif
971 
972 	/*
973 	 * Put TCP length in extended header, and then
974 	 * checksum extended header and data.
975 	 */
976 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
977 #ifdef INET6
978 	if (isipv6)
979 		/*
980 		 * ip6_plen is not need to be filled now, and will be filled
981 		 * in ip6_output.
982 		 */
983 		th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
984 				       sizeof(struct tcphdr) + optlen + len);
985 	else
986 #endif /* INET6 */
987 	{
988 		m->m_pkthdr.csum_flags = CSUM_TCP;
989 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
990 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
991 		    htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
992 
993 		/* IP version must be set here for ipv4/ipv6 checking later */
994 		KASSERT(ip->ip_v == IPVERSION,
995 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
996 	}
997 
998 	/*
999 	 * Enable TSO and specify the size of the segments.
1000 	 * The TCP pseudo header checksum is always provided.
1001 	 * XXX: Fixme: This is currently not the case for IPv6.
1002 	 */
1003 	if (tso) {
1004 		m->m_pkthdr.csum_flags = CSUM_TSO;
1005 		m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen;
1006 	}
1007 
1008 	/*
1009 	 * In transmit state, time the transmission and arrange for
1010 	 * the retransmit.  In persist state, just set snd_max.
1011 	 */
1012 	if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1013 	    !tcp_timer_active(tp, TT_PERSIST)) {
1014 		tcp_seq startseq = tp->snd_nxt;
1015 
1016 		/*
1017 		 * Advance snd_nxt over sequence space of this segment.
1018 		 */
1019 		if (flags & (TH_SYN|TH_FIN)) {
1020 			if (flags & TH_SYN)
1021 				tp->snd_nxt++;
1022 			if (flags & TH_FIN) {
1023 				tp->snd_nxt++;
1024 				tp->t_flags |= TF_SENTFIN;
1025 			}
1026 		}
1027 		if (sack_rxmit)
1028 			goto timer;
1029 		tp->snd_nxt += len;
1030 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1031 			tp->snd_max = tp->snd_nxt;
1032 			/*
1033 			 * Time this transmission if not a retransmission and
1034 			 * not currently timing anything.
1035 			 */
1036 			if (tp->t_rtttime == 0) {
1037 				tp->t_rtttime = ticks;
1038 				tp->t_rtseq = startseq;
1039 				tcpstat.tcps_segstimed++;
1040 			}
1041 		}
1042 
1043 		/*
1044 		 * Set retransmit timer if not currently set,
1045 		 * and not doing a pure ack or a keep-alive probe.
1046 		 * Initial value for retransmit timer is smoothed
1047 		 * round-trip time + 2 * round-trip time variance.
1048 		 * Initialize shift counter which is used for backoff
1049 		 * of retransmit time.
1050 		 */
1051 timer:
1052 		if (!tcp_timer_active(tp, TT_REXMT) &&
1053 		    ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1054 		     (tp->snd_nxt != tp->snd_una))) {
1055 			if (tcp_timer_active(tp, TT_PERSIST)) {
1056 				tcp_timer_activate(tp, TT_PERSIST, 0);
1057 				tp->t_rxtshift = 0;
1058 			}
1059 			tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1060 		}
1061 	} else {
1062 		/*
1063 		 * Persist case, update snd_max but since we are in
1064 		 * persist mode (no window) we do not update snd_nxt.
1065 		 */
1066 		int xlen = len;
1067 		if (flags & TH_SYN)
1068 			++xlen;
1069 		if (flags & TH_FIN) {
1070 			++xlen;
1071 			tp->t_flags |= TF_SENTFIN;
1072 		}
1073 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1074 			tp->snd_max = tp->snd_nxt + len;
1075 	}
1076 
1077 #ifdef TCPDEBUG
1078 	/*
1079 	 * Trace.
1080 	 */
1081 	if (so->so_options & SO_DEBUG) {
1082 		u_short save = 0;
1083 #ifdef INET6
1084 		if (!isipv6)
1085 #endif
1086 		{
1087 			save = ipov->ih_len;
1088 			ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1089 		}
1090 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1091 #ifdef INET6
1092 		if (!isipv6)
1093 #endif
1094 		ipov->ih_len = save;
1095 	}
1096 #endif
1097 
1098 	/*
1099 	 * Fill in IP length and desired time to live and
1100 	 * send to IP level.  There should be a better way
1101 	 * to handle ttl and tos; we could keep them in
1102 	 * the template, but need a way to checksum without them.
1103 	 */
1104 	/*
1105 	 * m->m_pkthdr.len should have been set before cksum calcuration,
1106 	 * because in6_cksum() need it.
1107 	 */
1108 #ifdef INET6
1109 	if (isipv6) {
1110 		/*
1111 		 * we separately set hoplimit for every segment, since the
1112 		 * user might want to change the value via setsockopt.
1113 		 * Also, desired default hop limit might be changed via
1114 		 * Neighbor Discovery.
1115 		 */
1116 		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1117 
1118 		/* TODO: IPv6 IP6TOS_ECT bit on */
1119 		error = ip6_output(m,
1120 			    tp->t_inpcb->in6p_outputopts, NULL,
1121 			    ((so->so_options & SO_DONTROUTE) ?
1122 			    IP_ROUTETOIF : 0), NULL, NULL, tp->t_inpcb);
1123 	} else
1124 #endif /* INET6 */
1125     {
1126 	ip->ip_len = m->m_pkthdr.len;
1127 #ifdef INET6
1128 	if (INP_CHECK_SOCKAF(so, AF_INET6))
1129 		ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1130 #endif /* INET6 */
1131 	/*
1132 	 * If we do path MTU discovery, then we set DF on every packet.
1133 	 * This might not be the best thing to do according to RFC3390
1134 	 * Section 2. However the tcp hostcache migitates the problem
1135 	 * so it affects only the first tcp connection with a host.
1136 	 */
1137 	if (path_mtu_discovery)
1138 		ip->ip_off |= IP_DF;
1139 
1140 	error = ip_output(m, tp->t_inpcb->inp_options, NULL,
1141 	    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1142 	    tp->t_inpcb);
1143     }
1144 	if (error) {
1145 
1146 		/*
1147 		 * We know that the packet was lost, so back out the
1148 		 * sequence number advance, if any.
1149 		 *
1150 		 * If the error is EPERM the packet got blocked by the
1151 		 * local firewall.  Normally we should terminate the
1152 		 * connection but the blocking may have been spurious
1153 		 * due to a firewall reconfiguration cycle.  So we treat
1154 		 * it like a packet loss and let the retransmit timer and
1155 		 * timeouts do their work over time.
1156 		 * XXX: It is a POLA question whether calling tcp_drop right
1157 		 * away would be the really correct behavior instead.
1158 		 */
1159 		if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1160 		    !tcp_timer_active(tp, TT_PERSIST)) &&
1161 		    ((flags & TH_SYN) == 0) &&
1162 		    (error != EPERM)) {
1163 			if (sack_rxmit) {
1164 				p->rxmit -= len;
1165 				tp->sackhint.sack_bytes_rexmit -= len;
1166 				KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1167 				    ("sackhint bytes rtx >= 0"));
1168 			} else
1169 				tp->snd_nxt -= len;
1170 		}
1171 out:
1172 		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1173 		switch (error) {
1174 		case EPERM:
1175 			tp->t_softerror = error;
1176 			return (error);
1177 		case ENOBUFS:
1178 	                if (!tcp_timer_active(tp, TT_REXMT) &&
1179 			    !tcp_timer_active(tp, TT_PERSIST))
1180 	                        tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1181 			tp->snd_cwnd = tp->t_maxseg;
1182 			return (0);
1183 		case EMSGSIZE:
1184 			/*
1185 			 * For some reason the interface we used initially
1186 			 * to send segments changed to another or lowered
1187 			 * its MTU.
1188 			 *
1189 			 * tcp_mtudisc() will find out the new MTU and as
1190 			 * its last action, initiate retransmission, so it
1191 			 * is important to not do so here.
1192 			 *
1193 			 * If TSO was active we either got an interface
1194 			 * without TSO capabilits or TSO was turned off.
1195 			 * Disable it for this connection as too and
1196 			 * immediatly retry with MSS sized segments generated
1197 			 * by this function.
1198 			 */
1199 			if (tso)
1200 				tp->t_flags &= ~TF_TSO;
1201 			tcp_mtudisc(tp->t_inpcb, 0);
1202 			return (0);
1203 		case EHOSTDOWN:
1204 		case EHOSTUNREACH:
1205 		case ENETDOWN:
1206 		case ENETUNREACH:
1207 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
1208 				tp->t_softerror = error;
1209 				return (0);
1210 			}
1211 			/* FALLTHROUGH */
1212 		default:
1213 			return (error);
1214 		}
1215 	}
1216 	tcpstat.tcps_sndtotal++;
1217 
1218 	/*
1219 	 * Data sent (as far as we can tell).
1220 	 * If this advertises a larger window than any other segment,
1221 	 * then remember the size of the advertised window.
1222 	 * Any pending ACK has now been sent.
1223 	 */
1224 	if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1225 		tp->rcv_adv = tp->rcv_nxt + recwin;
1226 	tp->last_ack_sent = tp->rcv_nxt;
1227 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1228 	if (tcp_timer_active(tp, TT_DELACK))
1229 		tcp_timer_activate(tp, TT_DELACK, 0);
1230 #if 0
1231 	/*
1232 	 * This completely breaks TCP if newreno is turned on.  What happens
1233 	 * is that if delayed-acks are turned on on the receiver, this code
1234 	 * on the transmitter effectively destroys the TCP window, forcing
1235 	 * it to four packets (1.5Kx4 = 6K window).
1236 	 */
1237 	if (sendalot && (!tcp_do_newreno || --maxburst))
1238 		goto again;
1239 #endif
1240 	if (sendalot)
1241 		goto again;
1242 	return (0);
1243 }
1244 
1245 void
1246 tcp_setpersist(struct tcpcb *tp)
1247 {
1248 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1249 	int tt;
1250 
1251 	if (tcp_timer_active(tp, TT_REXMT))
1252 		panic("tcp_setpersist: retransmit pending");
1253 	/*
1254 	 * Start/restart persistance timer.
1255 	 */
1256 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1257 		      TCPTV_PERSMIN, TCPTV_PERSMAX);
1258 	tcp_timer_activate(tp, TT_PERSIST, tt);
1259 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1260 		tp->t_rxtshift++;
1261 }
1262 
1263 /*
1264  * Insert TCP options according to the supplied parameters to the place
1265  * optp in a consistent way.  Can handle unaligned destinations.
1266  *
1267  * The order of the option processing is crucial for optimal packing and
1268  * alignment for the scarce option space.
1269  *
1270  * The optimal order for a SYN/SYN-ACK segment is:
1271  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1272  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1273  *
1274  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1275  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1276  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1277  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1278  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1279  */
1280 int
1281 tcp_addoptions(struct tcpopt *to, u_char *optp)
1282 {
1283 	u_int mask, optlen = 0;
1284 
1285 	for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1286 		if ((to->to_flags & mask) != mask)
1287 			continue;
1288 		switch (to->to_flags & mask) {
1289 		case TOF_MSS:
1290 			while (optlen % 4) {
1291 				optlen += TCPOLEN_NOP;
1292 				*optp++ = TCPOPT_NOP;
1293 			}
1294 			optlen += TCPOLEN_MAXSEG;
1295 			*optp++ = TCPOPT_MAXSEG;
1296 			*optp++ = TCPOLEN_MAXSEG;
1297 			to->to_mss = htons(to->to_mss);
1298 			bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1299 			optp += sizeof(to->to_mss);
1300 			break;
1301 		case TOF_SCALE:
1302 			while (!optlen || optlen % 2 != 1) {
1303 				optlen += TCPOLEN_NOP;
1304 				*optp++ = TCPOPT_NOP;
1305 			}
1306 			optlen += TCPOLEN_WINDOW;
1307 			*optp++ = TCPOPT_WINDOW;
1308 			*optp++ = TCPOLEN_WINDOW;
1309 			*optp++ = to->to_wscale;
1310 			break;
1311 		case TOF_SACKPERM:
1312 			while (optlen % 2) {
1313 				optlen += TCPOLEN_NOP;
1314 				*optp++ = TCPOPT_NOP;
1315 			}
1316 			optlen += TCPOLEN_SACK_PERMITTED;
1317 			*optp++ = TCPOPT_SACK_PERMITTED;
1318 			*optp++ = TCPOLEN_SACK_PERMITTED;
1319 			break;
1320 		case TOF_TS:
1321 			while (!optlen || optlen % 4 != 2) {
1322 				optlen += TCPOLEN_NOP;
1323 				*optp++ = TCPOPT_NOP;
1324 			}
1325 			optlen += TCPOLEN_TIMESTAMP;
1326 			*optp++ = TCPOPT_TIMESTAMP;
1327 			*optp++ = TCPOLEN_TIMESTAMP;
1328 			to->to_tsval = htonl(to->to_tsval);
1329 			to->to_tsecr = htonl(to->to_tsecr);
1330 			bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1331 			optp += sizeof(to->to_tsval);
1332 			bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1333 			optp += sizeof(to->to_tsecr);
1334 			break;
1335 		case TOF_SIGNATURE:
1336 			{
1337 			int siglen = TCPOLEN_SIGNATURE - 2;
1338 
1339 			while (!optlen || optlen % 4 != 2) {
1340 				optlen += TCPOLEN_NOP;
1341 				*optp++ = TCPOPT_NOP;
1342 			}
1343 			if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE)
1344 				continue;
1345 			optlen += TCPOLEN_SIGNATURE;
1346 			*optp++ = TCPOPT_SIGNATURE;
1347 			*optp++ = TCPOLEN_SIGNATURE;
1348 			to->to_signature = optp;
1349 			while (siglen--)
1350 				 *optp++ = 0;
1351 			break;
1352 			}
1353 		case TOF_SACK:
1354 			{
1355 			int sackblks = 0;
1356 			struct sackblk *sack = (struct sackblk *)to->to_sacks;
1357 			tcp_seq sack_seq;
1358 
1359 			while (!optlen || optlen % 4 != 2) {
1360 				optlen += TCPOLEN_NOP;
1361 				*optp++ = TCPOPT_NOP;
1362 			}
1363 			if (TCP_MAXOLEN - optlen < 2 + TCPOLEN_SACK)
1364 				continue;
1365 			optlen += TCPOLEN_SACKHDR;
1366 			*optp++ = TCPOPT_SACK;
1367 			sackblks = min(to->to_nsacks,
1368 					(TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1369 			*optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1370 			while (sackblks--) {
1371 				sack_seq = htonl(sack->start);
1372 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1373 				optp += sizeof(sack_seq);
1374 				sack_seq = htonl(sack->end);
1375 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1376 				optp += sizeof(sack_seq);
1377 				optlen += TCPOLEN_SACK;
1378 				sack++;
1379 			}
1380 			tcpstat.tcps_sack_send_blocks++;
1381 			break;
1382 			}
1383 		default:
1384 			panic("%s: unknown TCP option type", __func__);
1385 			break;
1386 		}
1387 	}
1388 
1389 	/* Terminate and pad TCP options to a 4 byte boundary. */
1390 	if (optlen % 4) {
1391 		optlen += TCPOLEN_EOL;
1392 		*optp++ = TCPOPT_EOL;
1393 	}
1394 	while (optlen % 4) {
1395 		optlen += TCPOLEN_NOP;
1396 		*optp++ = TCPOPT_NOP;
1397 	}
1398 
1399 	KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1400 	return (optlen);
1401 }
1402