xref: /freebsd/sys/netinet/tcp_output.c (revision fa030de01267cc324a0b57479715e4a53356b665)
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 	int idle, sendalot;
141 	int sack_rxmit, sack_bytes_rxmt;
142 	struct sackhole *p;
143 	int tso = 0;
144 	struct tcpopt to;
145 #if 0
146 	int maxburst = TCP_MAXBURST;
147 #endif
148 #ifdef INET6
149 	struct ip6_hdr *ip6 = NULL;
150 	int isipv6;
151 
152 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
153 #endif
154 
155 	INP_LOCK_ASSERT(tp->t_inpcb);
156 
157 	/*
158 	 * Determine length of data that should be transmitted,
159 	 * and flags that will be used.
160 	 * If there is some data or critical controls (SYN, RST)
161 	 * to send, then transmit; otherwise, investigate further.
162 	 */
163 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
164 	if (idle && (ticks - tp->t_rcvtime) >= tp->t_rxtcur) {
165 		/*
166 		 * We have been idle for "a while" and no acks are
167 		 * expected to clock out any data we send --
168 		 * slow start to get ack "clock" running again.
169 		 *
170 		 * Set the slow-start flight size depending on whether
171 		 * this is a local network or not.
172 		 */
173 		int ss = ss_fltsz;
174 #ifdef INET6
175 		if (isipv6) {
176 			if (in6_localaddr(&tp->t_inpcb->in6p_faddr))
177 				ss = ss_fltsz_local;
178 		} else
179 #endif /* INET6 */
180 		if (in_localaddr(tp->t_inpcb->inp_faddr))
181 			ss = ss_fltsz_local;
182 		tp->snd_cwnd = tp->t_maxseg * ss;
183 	}
184 	tp->t_flags &= ~TF_LASTIDLE;
185 	if (idle) {
186 		if (tp->t_flags & TF_MORETOCOME) {
187 			tp->t_flags |= TF_LASTIDLE;
188 			idle = 0;
189 		}
190 	}
191 again:
192 	/*
193 	 * If we've recently taken a timeout, snd_max will be greater than
194 	 * snd_nxt.  There may be SACK information that allows us to avoid
195 	 * resending already delivered data.  Adjust snd_nxt accordingly.
196 	 */
197 	if ((tp->t_flags & TF_SACK_PERMIT) &&
198 	    SEQ_LT(tp->snd_nxt, tp->snd_max))
199 		tcp_sack_adjust(tp);
200 	sendalot = 0;
201 	off = tp->snd_nxt - tp->snd_una;
202 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
203 	sendwin = min(sendwin, tp->snd_bwnd);
204 
205 	flags = tcp_outflags[tp->t_state];
206 	/*
207 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
208 	 * to send out new data (when sendalot is 1), bypass this function.
209 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
210 	 * we're replacing a (future) new transmission with a retransmission
211 	 * now, and we previously incremented snd_cwnd in tcp_input().
212 	 */
213 	/*
214 	 * Still in sack recovery , reset rxmit flag to zero.
215 	 */
216 	sack_rxmit = 0;
217 	sack_bytes_rxmt = 0;
218 	len = 0;
219 	p = NULL;
220 	if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp) &&
221 	    (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
222 		long cwin;
223 
224 		cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
225 		if (cwin < 0)
226 			cwin = 0;
227 		/* Do not retransmit SACK segments beyond snd_recover */
228 		if (SEQ_GT(p->end, tp->snd_recover)) {
229 			/*
230 			 * (At least) part of sack hole extends beyond
231 			 * snd_recover. Check to see if we can rexmit data
232 			 * for this hole.
233 			 */
234 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
235 				/*
236 				 * Can't rexmit any more data for this hole.
237 				 * That data will be rexmitted in the next
238 				 * sack recovery episode, when snd_recover
239 				 * moves past p->rxmit.
240 				 */
241 				p = NULL;
242 				goto after_sack_rexmit;
243 			} else
244 				/* Can rexmit part of the current hole */
245 				len = ((long)ulmin(cwin,
246 						   tp->snd_recover - p->rxmit));
247 		} else
248 			len = ((long)ulmin(cwin, p->end - p->rxmit));
249 		off = p->rxmit - tp->snd_una;
250 		KASSERT(off >= 0,("%s: sack block to the left of una : %d",
251 		    __func__, off));
252 		if (len > 0) {
253 			sack_rxmit = 1;
254 			sendalot = 1;
255 			tcpstat.tcps_sack_rexmits++;
256 			tcpstat.tcps_sack_rexmit_bytes +=
257 			    min(len, tp->t_maxseg);
258 		}
259 	}
260 after_sack_rexmit:
261 	/*
262 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
263 	 * state flags.
264 	 */
265 	if (tp->t_flags & TF_NEEDFIN)
266 		flags |= TH_FIN;
267 	if (tp->t_flags & TF_NEEDSYN)
268 		flags |= TH_SYN;
269 
270 	SOCKBUF_LOCK(&so->so_snd);
271 	/*
272 	 * If in persist timeout with window of 0, send 1 byte.
273 	 * Otherwise, if window is small but nonzero
274 	 * and timer expired, we will send what we can
275 	 * and go to transmit state.
276 	 */
277 	if (tp->t_flags & TF_FORCEDATA) {
278 		if (sendwin == 0) {
279 			/*
280 			 * If we still have some data to send, then
281 			 * clear the FIN bit.  Usually this would
282 			 * happen below when it realizes that we
283 			 * aren't sending all the data.  However,
284 			 * if we have exactly 1 byte of unsent data,
285 			 * then it won't clear the FIN bit below,
286 			 * and if we are in persist state, we wind
287 			 * up sending the packet without recording
288 			 * that we sent the FIN bit.
289 			 *
290 			 * We can't just blindly clear the FIN bit,
291 			 * because if we don't have any more data
292 			 * to send then the probe will be the FIN
293 			 * itself.
294 			 */
295 			if (off < so->so_snd.sb_cc)
296 				flags &= ~TH_FIN;
297 			sendwin = 1;
298 		} else {
299 			tcp_timer_activate(tp, TT_PERSIST, 0);
300 			tp->t_rxtshift = 0;
301 		}
302 	}
303 
304 	/*
305 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
306 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
307 	 * a negative length.  This can also occur when TCP opens up
308 	 * its congestion window while receiving additional duplicate
309 	 * acks after fast-retransmit because TCP will reset snd_nxt
310 	 * to snd_max after the fast-retransmit.
311 	 *
312 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
313 	 * be set to snd_una, the offset will be 0, and the length may
314 	 * wind up 0.
315 	 *
316 	 * If sack_rxmit is true we are retransmitting from the scoreboard
317 	 * in which case len is already set.
318 	 */
319 	if (sack_rxmit == 0) {
320 		if (sack_bytes_rxmt == 0)
321 			len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off);
322 		else {
323 			long cwin;
324 
325                         /*
326 			 * We are inside of a SACK recovery episode and are
327 			 * sending new data, having retransmitted all the
328 			 * data possible in the scoreboard.
329 			 */
330 			len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd)
331 			       - off);
332 			/*
333 			 * Don't remove this (len > 0) check !
334 			 * We explicitly check for len > 0 here (although it
335 			 * isn't really necessary), to work around a gcc
336 			 * optimization issue - to force gcc to compute
337 			 * len above. Without this check, the computation
338 			 * of len is bungled by the optimizer.
339 			 */
340 			if (len > 0) {
341 				cwin = tp->snd_cwnd -
342 					(tp->snd_nxt - tp->sack_newdata) -
343 					sack_bytes_rxmt;
344 				if (cwin < 0)
345 					cwin = 0;
346 				len = lmin(len, cwin);
347 			}
348 		}
349 	}
350 
351 	/*
352 	 * Lop off SYN bit if it has already been sent.  However, if this
353 	 * is SYN-SENT state and if segment contains data and if we don't
354 	 * know that foreign host supports TAO, suppress sending segment.
355 	 */
356 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
357 		if (tp->t_state != TCPS_SYN_RECEIVED)
358 			flags &= ~TH_SYN;
359 		off--, len++;
360 	}
361 
362 	/*
363 	 * Be careful not to send data and/or FIN on SYN segments.
364 	 * This measure is needed to prevent interoperability problems
365 	 * with not fully conformant TCP implementations.
366 	 */
367 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
368 		len = 0;
369 		flags &= ~TH_FIN;
370 	}
371 
372 	if (len < 0) {
373 		/*
374 		 * If FIN has been sent but not acked,
375 		 * but we haven't been called to retransmit,
376 		 * len will be < 0.  Otherwise, window shrank
377 		 * after we sent into it.  If window shrank to 0,
378 		 * cancel pending retransmit, pull snd_nxt back
379 		 * to (closed) window, and set the persist timer
380 		 * if it isn't already going.  If the window didn't
381 		 * close completely, just wait for an ACK.
382 		 */
383 		len = 0;
384 		if (sendwin == 0) {
385 			tcp_timer_activate(tp, TT_REXMT, 0);
386 			tp->t_rxtshift = 0;
387 			tp->snd_nxt = tp->snd_una;
388 			if (!tcp_timer_active(tp, TT_PERSIST))
389 				tcp_setpersist(tp);
390 		}
391 	}
392 
393 	/* len will be >= 0 after this point. */
394 	KASSERT(len >= 0, ("%s: len < 0", __func__));
395 
396 	/*
397 	 * Automatic sizing of send socket buffer.  Often the send buffer
398 	 * size is not optimally adjusted to the actual network conditions
399 	 * at hand (delay bandwidth product).  Setting the buffer size too
400 	 * small limits throughput on links with high bandwidth and high
401 	 * delay (eg. trans-continental/oceanic links).  Setting the
402 	 * buffer size too big consumes too much real kernel memory,
403 	 * especially with many connections on busy servers.
404 	 *
405 	 * The criteria to step up the send buffer one notch are:
406 	 *  1. receive window of remote host is larger than send buffer
407 	 *     (with a fudge factor of 5/4th);
408 	 *  2. send buffer is filled to 7/8th with data (so we actually
409 	 *     have data to make use of it);
410 	 *  3. send buffer fill has not hit maximal automatic size;
411 	 *  4. our send window (slow start and cogestion controlled) is
412 	 *     larger than sent but unacknowledged data in send buffer.
413 	 *
414 	 * The remote host receive window scaling factor may limit the
415 	 * growing of the send buffer before it reaches its allowed
416 	 * maximum.
417 	 *
418 	 * It scales directly with slow start or congestion window
419 	 * and does at most one step per received ACK.  This fast
420 	 * scaling has the drawback of growing the send buffer beyond
421 	 * what is strictly necessary to make full use of a given
422 	 * delay*bandwith product.  However testing has shown this not
423 	 * to be much of an problem.  At worst we are trading wasting
424 	 * of available bandwith (the non-use of it) for wasting some
425 	 * socket buffer memory.
426 	 *
427 	 * TODO: Shrink send buffer during idle periods together
428 	 * with congestion window.  Requires another timer.  Has to
429 	 * wait for upcoming tcp timer rewrite.
430 	 */
431 	if (tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
432 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
433 		    so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
434 		    so->so_snd.sb_cc < tcp_autosndbuf_max &&
435 		    sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
436 			if (!sbreserve_locked(&so->so_snd,
437 			    min(so->so_snd.sb_hiwat + tcp_autosndbuf_inc,
438 			     tcp_autosndbuf_max), so, curthread))
439 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
440 		}
441 	}
442 
443 	/*
444 	 * Truncate to the maximum segment length or enable TCP Segmentation
445 	 * Offloading (if supported by hardware) and ensure that FIN is removed
446 	 * if the length no longer contains the last data byte.
447 	 *
448 	 * TSO may only be used if we are in a pure bulk sending state.  The
449 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
450 	 * IP options prevent using TSO.  With TSO the TCP header is the same
451 	 * (except for the sequence number) for all generated packets.  This
452 	 * makes it impossible to transmit any options which vary per generated
453 	 * segment or packet.
454 	 *
455 	 * The length of TSO bursts is limited to TCP_MAXWIN.  That limit and
456 	 * removal of FIN (if not already catched here) are handled later after
457 	 * the exact length of the TCP options are known.
458 	 */
459 	if (len > tp->t_maxseg) {
460 		if ((tp->t_flags & TF_TSO) && tcp_do_tso &&
461 		    ((tp->t_flags & TF_SIGNATURE) == 0) &&
462 		    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
463 		    tp->t_inpcb->inp_options == NULL &&
464 		    tp->t_inpcb->in6p_options == NULL &&
465 		    tp->t_inpcb->inp_sp == NULL) {
466 			tso = 1;
467 		} else {
468 			len = tp->t_maxseg;
469 			sendalot = 1;
470 			tso = 0;
471 		}
472 	}
473 	if (sack_rxmit) {
474 		if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
475 			flags &= ~TH_FIN;
476 	} else {
477 		if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
478 			flags &= ~TH_FIN;
479 	}
480 
481 	recwin = sbspace(&so->so_rcv);
482 
483 	/*
484 	 * Sender silly window avoidance.   We transmit under the following
485 	 * conditions when len is non-zero:
486 	 *
487 	 *	- We have a full segment (or more with TSO)
488 	 *	- This is the last buffer in a write()/send() and we are
489 	 *	  either idle or running NODELAY
490 	 *	- we've timed out (e.g. persist timer)
491 	 *	- we have more then 1/2 the maximum send window's worth of
492 	 *	  data (receiver may be limited the window size)
493 	 *	- we need to retransmit
494 	 */
495 	if (len) {
496 		if (len >= tp->t_maxseg)
497 			goto send;
498 		/*
499 		 * NOTE! on localhost connections an 'ack' from the remote
500 		 * end may occur synchronously with the output and cause
501 		 * us to flush a buffer queued with moretocome.  XXX
502 		 *
503 		 * note: the len + off check is almost certainly unnecessary.
504 		 */
505 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
506 		    (idle || (tp->t_flags & TF_NODELAY)) &&
507 		    len + off >= so->so_snd.sb_cc &&
508 		    (tp->t_flags & TF_NOPUSH) == 0) {
509 			goto send;
510 		}
511 		if (tp->t_flags & TF_FORCEDATA)		/* typ. timeout case */
512 			goto send;
513 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
514 			goto send;
515 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
516 			goto send;
517 		if (sack_rxmit)
518 			goto send;
519 	}
520 
521 	/*
522 	 * Compare available window to amount of window
523 	 * known to peer (as advertised window less
524 	 * next expected input).  If the difference is at least two
525 	 * max size segments, or at least 50% of the maximum possible
526 	 * window, then want to send a window update to peer.
527 	 * Skip this if the connection is in T/TCP half-open state.
528 	 * Don't send pure window updates when the peer has closed
529 	 * the connection and won't ever send more data.
530 	 */
531 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
532 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
533 		/*
534 		 * "adv" is the amount we can increase the window,
535 		 * taking into account that we are limited by
536 		 * TCP_MAXWIN << tp->rcv_scale.
537 		 */
538 		long adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale) -
539 			(tp->rcv_adv - tp->rcv_nxt);
540 
541 		if (adv >= (long) (2 * tp->t_maxseg))
542 			goto send;
543 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
544 			goto send;
545 	}
546 
547 	/*
548 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
549 	 * is also a catch-all for the retransmit timer timeout case.
550 	 */
551 	if (tp->t_flags & TF_ACKNOW)
552 		goto send;
553 	if ((flags & TH_RST) ||
554 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
555 		goto send;
556 	if (SEQ_GT(tp->snd_up, tp->snd_una))
557 		goto send;
558 	/*
559 	 * If our state indicates that FIN should be sent
560 	 * and we have not yet done so, then we need to send.
561 	 */
562 	if (flags & TH_FIN &&
563 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
564 		goto send;
565 	/*
566 	 * In SACK, it is possible for tcp_output to fail to send a segment
567 	 * after the retransmission timer has been turned off.  Make sure
568 	 * that the retransmission timer is set.
569 	 */
570 	if ((tp->t_flags & TF_SACK_PERMIT) &&
571 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
572 	    !tcp_timer_active(tp, TT_REXMT) &&
573 	    !tcp_timer_active(tp, TT_PERSIST)) {
574 		tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
575 		goto just_return;
576 	}
577 	/*
578 	 * TCP window updates are not reliable, rather a polling protocol
579 	 * using ``persist'' packets is used to insure receipt of window
580 	 * updates.  The three ``states'' for the output side are:
581 	 *	idle			not doing retransmits or persists
582 	 *	persisting		to move a small or zero window
583 	 *	(re)transmitting	and thereby not persisting
584 	 *
585 	 * tcp_timer_active(tp, TT_PERSIST)
586 	 *	is true when we are in persist state.
587 	 * (tp->t_flags & TF_FORCEDATA)
588 	 *	is set when we are called to send a persist packet.
589 	 * tcp_timer_active(tp, TT_REXMT)
590 	 *	is set when we are retransmitting
591 	 * The output side is idle when both timers are zero.
592 	 *
593 	 * If send window is too small, there is data to transmit, and no
594 	 * retransmit or persist is pending, then go to persist state.
595 	 * If nothing happens soon, send when timer expires:
596 	 * if window is nonzero, transmit what we can,
597 	 * otherwise force out a byte.
598 	 */
599 	if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) &&
600 	    !tcp_timer_active(tp, TT_PERSIST)) {
601 		tp->t_rxtshift = 0;
602 		tcp_setpersist(tp);
603 	}
604 
605 	/*
606 	 * No reason to send a segment, just return.
607 	 */
608 just_return:
609 	SOCKBUF_UNLOCK(&so->so_snd);
610 	return (0);
611 
612 send:
613 	SOCKBUF_LOCK_ASSERT(&so->so_snd);
614 	/*
615 	 * Before ESTABLISHED, force sending of initial options
616 	 * unless TCP set not to do any options.
617 	 * NOTE: we assume that the IP/TCP header plus TCP options
618 	 * always fit in a single mbuf, leaving room for a maximum
619 	 * link header, i.e.
620 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
621 	 */
622 	optlen = 0;
623 #ifdef INET6
624 	if (isipv6)
625 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
626 	else
627 #endif
628 	hdrlen = sizeof (struct tcpiphdr);
629 
630 	/*
631 	 * Compute options for segment.
632 	 * We only have to care about SYN and established connection
633 	 * segments.  Options for SYN-ACK segments are handled in TCP
634 	 * syncache.
635 	 */
636 	if ((tp->t_flags & TF_NOOPT) == 0) {
637 		to.to_flags = 0;
638 		/* Maximum segment size. */
639 		if (flags & TH_SYN) {
640 			tp->snd_nxt = tp->iss;
641 			to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
642 			to.to_flags |= TOF_MSS;
643 		}
644 		/* Window scaling. */
645 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
646 			to.to_wscale = tp->request_r_scale;
647 			to.to_flags |= TOF_SCALE;
648 		}
649 		/* Timestamps. */
650 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
651 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
652 			to.to_tsval = ticks + tp->ts_offset;
653 			to.to_tsecr = tp->ts_recent;
654 			to.to_flags |= TOF_TS;
655 			/* Set receive buffer autosizing timestamp. */
656 			if (tp->rfbuf_ts == 0 &&
657 			    (so->so_rcv.sb_flags & SB_AUTOSIZE))
658 				tp->rfbuf_ts = ticks;
659 		}
660 		/* Selective ACK's. */
661 		if (tp->t_flags & TF_SACK_PERMIT) {
662 			if (flags & TH_SYN)
663 				to.to_flags |= TOF_SACKPERM;
664 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
665 			    (tp->t_flags & TF_SACK_PERMIT) &&
666 			    tp->rcv_numsacks > 0) {
667 				to.to_flags |= TOF_SACK;
668 				to.to_nsacks = tp->rcv_numsacks;
669 				to.to_sacks = (u_char *)tp->sackblks;
670 			}
671 		}
672 #ifdef TCP_SIGNATURE
673 		/* TCP-MD5 (RFC2385). */
674 #ifdef INET6
675 		if (!isipv6 && (tp->t_flags & TF_SIGNATURE))
676 #else
677 		if (tp->t_flags & TF_SIGNATURE)
678 #endif /* INET6 */
679 			to.to_flags |= TOF_SIGNATURE;
680 #endif /* TCP_SIGNATURE */
681 
682 		/* Processing the options. */
683 		hdrlen += optlen = tcp_addoptions(&to, (u_char *)&opt);
684 
685 #ifdef TCP_SIGNATURE
686 		sigoff = to.to_signature - (u_char *)&to;
687 #endif /* TCP_SIGNATURE */
688 	}
689 
690 #ifdef INET6
691 	if (isipv6)
692 		ipoptlen = ip6_optlen(tp->t_inpcb);
693 	else
694 #endif
695 	if (tp->t_inpcb->inp_options)
696 		ipoptlen = tp->t_inpcb->inp_options->m_len -
697 				offsetof(struct ipoption, ipopt_list);
698 	else
699 		ipoptlen = 0;
700 #ifdef IPSEC
701 	ipoptlen += ipsec_hdrsiz_tcp(tp);
702 #endif
703 
704 	/*
705 	 * Adjust data length if insertion of options will
706 	 * bump the packet length beyond the t_maxopd length.
707 	 * Clear the FIN bit because we cut off the tail of
708 	 * the segment.
709 	 *
710 	 * When doing TSO limit a burst to TCP_MAXWIN minus the
711 	 * IP, TCP and Options length to keep ip->ip_len from
712 	 * overflowing.  Prevent the last segment from being
713 	 * fractional thus making them all equal sized and set
714 	 * the flag to continue sending.  TSO is disabled when
715 	 * IP options or IPSEC are present.
716 	 */
717 	if (len + optlen + ipoptlen > tp->t_maxopd) {
718 		flags &= ~TH_FIN;
719 		if (tso) {
720 			if (len > TCP_MAXWIN - hdrlen - optlen) {
721 				len = TCP_MAXWIN - hdrlen - optlen;
722 				len = len - (len % (tp->t_maxopd - optlen));
723 				sendalot = 1;
724 			} else if (tp->t_flags & TF_NEEDFIN)
725 				sendalot = 1;
726 		} else {
727 			len = tp->t_maxopd - optlen - ipoptlen;
728 			sendalot = 1;
729 		}
730 	}
731 
732 /*#ifdef DIAGNOSTIC*/
733 #ifdef INET6
734 	if (max_linkhdr + hdrlen > MCLBYTES)
735 #else
736 	if (max_linkhdr + hdrlen > MHLEN)
737 #endif
738 		panic("tcphdr too big");
739 /*#endif*/
740 
741 	/*
742 	 * Grab a header mbuf, attaching a copy of data to
743 	 * be transmitted, and initialize the header from
744 	 * the template for sends on this connection.
745 	 */
746 	if (len) {
747 		struct mbuf *mb;
748 		u_int moff;
749 
750 		if ((tp->t_flags & TF_FORCEDATA) && len == 1)
751 			tcpstat.tcps_sndprobe++;
752 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
753 			tcpstat.tcps_sndrexmitpack++;
754 			tcpstat.tcps_sndrexmitbyte += len;
755 		} else {
756 			tcpstat.tcps_sndpack++;
757 			tcpstat.tcps_sndbyte += len;
758 		}
759 #ifdef notyet
760 		if ((m = m_copypack(so->so_snd.sb_mb, off,
761 		    (int)len, max_linkhdr + hdrlen)) == 0) {
762 			SOCKBUF_UNLOCK(&so->so_snd);
763 			error = ENOBUFS;
764 			goto out;
765 		}
766 		/*
767 		 * m_copypack left space for our hdr; use it.
768 		 */
769 		m->m_len += hdrlen;
770 		m->m_data -= hdrlen;
771 #else
772 		MGETHDR(m, M_DONTWAIT, MT_DATA);
773 		if (m == NULL) {
774 			SOCKBUF_UNLOCK(&so->so_snd);
775 			error = ENOBUFS;
776 			goto out;
777 		}
778 #ifdef INET6
779 		if (MHLEN < hdrlen + max_linkhdr) {
780 			MCLGET(m, M_DONTWAIT);
781 			if ((m->m_flags & M_EXT) == 0) {
782 				SOCKBUF_UNLOCK(&so->so_snd);
783 				m_freem(m);
784 				error = ENOBUFS;
785 				goto out;
786 			}
787 		}
788 #endif
789 		m->m_data += max_linkhdr;
790 		m->m_len = hdrlen;
791 
792 		/*
793 		 * Start the m_copy functions from the closest mbuf
794 		 * to the offset in the socket buffer chain.
795 		 */
796 		mb = sbsndptr(&so->so_snd, off, len, &moff);
797 
798 		if (len <= MHLEN - hdrlen - max_linkhdr) {
799 			m_copydata(mb, moff, (int)len,
800 			    mtod(m, caddr_t) + hdrlen);
801 			m->m_len += len;
802 		} else {
803 			m->m_next = m_copy(mb, moff, (int)len);
804 			if (m->m_next == NULL) {
805 				SOCKBUF_UNLOCK(&so->so_snd);
806 				(void) m_free(m);
807 				error = ENOBUFS;
808 				goto out;
809 			}
810 		}
811 #endif
812 		/*
813 		 * If we're sending everything we've got, set PUSH.
814 		 * (This will keep happy those implementations which only
815 		 * give data to the user when a buffer fills or
816 		 * a PUSH comes in.)
817 		 */
818 		if (off + len == so->so_snd.sb_cc)
819 			flags |= TH_PUSH;
820 		SOCKBUF_UNLOCK(&so->so_snd);
821 	} else {
822 		SOCKBUF_UNLOCK(&so->so_snd);
823 		if (tp->t_flags & TF_ACKNOW)
824 			tcpstat.tcps_sndacks++;
825 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
826 			tcpstat.tcps_sndctrl++;
827 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
828 			tcpstat.tcps_sndurg++;
829 		else
830 			tcpstat.tcps_sndwinup++;
831 
832 		MGETHDR(m, M_DONTWAIT, MT_DATA);
833 		if (m == NULL) {
834 			error = ENOBUFS;
835 			goto out;
836 		}
837 #ifdef INET6
838 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
839 		    MHLEN >= hdrlen) {
840 			MH_ALIGN(m, hdrlen);
841 		} else
842 #endif
843 		m->m_data += max_linkhdr;
844 		m->m_len = hdrlen;
845 	}
846 	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
847 	m->m_pkthdr.rcvif = (struct ifnet *)0;
848 #ifdef MAC
849 	mac_create_mbuf_from_inpcb(tp->t_inpcb, m);
850 #endif
851 #ifdef INET6
852 	if (isipv6) {
853 		ip6 = mtod(m, struct ip6_hdr *);
854 		th = (struct tcphdr *)(ip6 + 1);
855 		tcpip_fillheaders(tp->t_inpcb, ip6, th);
856 	} else
857 #endif /* INET6 */
858 	{
859 		ip = mtod(m, struct ip *);
860 		ipov = (struct ipovly *)ip;
861 		th = (struct tcphdr *)(ip + 1);
862 		tcpip_fillheaders(tp->t_inpcb, ip, th);
863 	}
864 
865 	/*
866 	 * Fill in fields, remembering maximum advertised
867 	 * window for use in delaying messages about window sizes.
868 	 * If resending a FIN, be sure not to use a new sequence number.
869 	 */
870 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
871 	    tp->snd_nxt == tp->snd_max)
872 		tp->snd_nxt--;
873 	/*
874 	 * If we are doing retransmissions, then snd_nxt will
875 	 * not reflect the first unsent octet.  For ACK only
876 	 * packets, we do not want the sequence number of the
877 	 * retransmitted packet, we want the sequence number
878 	 * of the next unsent octet.  So, if there is no data
879 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
880 	 * when filling in ti_seq.  But if we are in persist
881 	 * state, snd_max might reflect one byte beyond the
882 	 * right edge of the window, so use snd_nxt in that
883 	 * case, since we know we aren't doing a retransmission.
884 	 * (retransmit and persist are mutually exclusive...)
885 	 */
886 	if (sack_rxmit == 0) {
887 		if (len || (flags & (TH_SYN|TH_FIN)) ||
888 		    tcp_timer_active(tp, TT_PERSIST))
889 			th->th_seq = htonl(tp->snd_nxt);
890 		else
891 			th->th_seq = htonl(tp->snd_max);
892 	} else {
893 		th->th_seq = htonl(p->rxmit);
894 		p->rxmit += len;
895 		tp->sackhint.sack_bytes_rexmit += len;
896 	}
897 	th->th_ack = htonl(tp->rcv_nxt);
898 	if (optlen) {
899 		bcopy(opt, th + 1, optlen);
900 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
901 	}
902 	th->th_flags = flags;
903 	/*
904 	 * Calculate receive window.  Don't shrink window,
905 	 * but avoid silly window syndrome.
906 	 */
907 	if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
908 	    recwin < (long)tp->t_maxseg)
909 		recwin = 0;
910 	if (recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
911 		recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
912 	if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
913 		recwin = (long)TCP_MAXWIN << tp->rcv_scale;
914 
915 	/*
916 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
917 	 * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
918 	 * case is handled in syncache.
919 	 */
920 	if (flags & TH_SYN)
921 		th->th_win = htons((u_short)
922 				(min(sbspace(&so->so_rcv), TCP_MAXWIN)));
923 	else
924 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
925 
926 	/*
927 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
928 	 * a 0 window.  This may cause the remote transmitter to stall.  This
929 	 * flag tells soreceive() to disable delayed acknowledgements when
930 	 * draining the buffer.  This can occur if the receiver is attempting
931 	 * to read more data then can be buffered prior to transmitting on
932 	 * the connection.
933 	 */
934 	if (recwin == 0)
935 		tp->t_flags |= TF_RXWIN0SENT;
936 	else
937 		tp->t_flags &= ~TF_RXWIN0SENT;
938 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
939 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
940 		th->th_flags |= TH_URG;
941 	} else
942 		/*
943 		 * If no urgent pointer to send, then we pull
944 		 * the urgent pointer to the left edge of the send window
945 		 * so that it doesn't drift into the send window on sequence
946 		 * number wraparound.
947 		 */
948 		tp->snd_up = tp->snd_una;		/* drag it along */
949 
950 #ifdef TCP_SIGNATURE
951 #ifdef INET6
952 	if (!isipv6)
953 #endif
954 	if (tp->t_flags & TF_SIGNATURE)
955 		tcp_signature_compute(m, sizeof(struct ip), len, optlen,
956 		    (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
957 #endif
958 
959 	/*
960 	 * Put TCP length in extended header, and then
961 	 * checksum extended header and data.
962 	 */
963 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
964 #ifdef INET6
965 	if (isipv6)
966 		/*
967 		 * ip6_plen is not need to be filled now, and will be filled
968 		 * in ip6_output.
969 		 */
970 		th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
971 				       sizeof(struct tcphdr) + optlen + len);
972 	else
973 #endif /* INET6 */
974 	{
975 		m->m_pkthdr.csum_flags = CSUM_TCP;
976 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
977 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
978 		    htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
979 
980 		/* IP version must be set here for ipv4/ipv6 checking later */
981 		KASSERT(ip->ip_v == IPVERSION,
982 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
983 	}
984 
985 	/*
986 	 * Enable TSO and specify the size of the segments.
987 	 * The TCP pseudo header checksum is always provided.
988 	 * XXX: Fixme: This is currently not the case for IPv6.
989 	 */
990 	if (tso) {
991 		m->m_pkthdr.csum_flags = CSUM_TSO;
992 		m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen;
993 	}
994 
995 	/*
996 	 * In transmit state, time the transmission and arrange for
997 	 * the retransmit.  In persist state, just set snd_max.
998 	 */
999 	if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1000 	    !tcp_timer_active(tp, TT_PERSIST)) {
1001 		tcp_seq startseq = tp->snd_nxt;
1002 
1003 		/*
1004 		 * Advance snd_nxt over sequence space of this segment.
1005 		 */
1006 		if (flags & (TH_SYN|TH_FIN)) {
1007 			if (flags & TH_SYN)
1008 				tp->snd_nxt++;
1009 			if (flags & TH_FIN) {
1010 				tp->snd_nxt++;
1011 				tp->t_flags |= TF_SENTFIN;
1012 			}
1013 		}
1014 		if (sack_rxmit)
1015 			goto timer;
1016 		tp->snd_nxt += len;
1017 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1018 			tp->snd_max = tp->snd_nxt;
1019 			/*
1020 			 * Time this transmission if not a retransmission and
1021 			 * not currently timing anything.
1022 			 */
1023 			if (tp->t_rtttime == 0) {
1024 				tp->t_rtttime = ticks;
1025 				tp->t_rtseq = startseq;
1026 				tcpstat.tcps_segstimed++;
1027 			}
1028 		}
1029 
1030 		/*
1031 		 * Set retransmit timer if not currently set,
1032 		 * and not doing a pure ack or a keep-alive probe.
1033 		 * Initial value for retransmit timer is smoothed
1034 		 * round-trip time + 2 * round-trip time variance.
1035 		 * Initialize shift counter which is used for backoff
1036 		 * of retransmit time.
1037 		 */
1038 timer:
1039 		if (!tcp_timer_active(tp, TT_REXMT) &&
1040 		    ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1041 		     (tp->snd_nxt != tp->snd_una))) {
1042 			if (tcp_timer_active(tp, TT_PERSIST)) {
1043 				tcp_timer_activate(tp, TT_PERSIST, 0);
1044 				tp->t_rxtshift = 0;
1045 			}
1046 			tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1047 		}
1048 	} else {
1049 		/*
1050 		 * Persist case, update snd_max but since we are in
1051 		 * persist mode (no window) we do not update snd_nxt.
1052 		 */
1053 		int xlen = len;
1054 		if (flags & TH_SYN)
1055 			++xlen;
1056 		if (flags & TH_FIN) {
1057 			++xlen;
1058 			tp->t_flags |= TF_SENTFIN;
1059 		}
1060 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1061 			tp->snd_max = tp->snd_nxt + len;
1062 	}
1063 
1064 #ifdef TCPDEBUG
1065 	/*
1066 	 * Trace.
1067 	 */
1068 	if (so->so_options & SO_DEBUG) {
1069 		u_short save = 0;
1070 #ifdef INET6
1071 		if (!isipv6)
1072 #endif
1073 		{
1074 			save = ipov->ih_len;
1075 			ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1076 		}
1077 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1078 #ifdef INET6
1079 		if (!isipv6)
1080 #endif
1081 		ipov->ih_len = save;
1082 	}
1083 #endif
1084 
1085 	/*
1086 	 * Fill in IP length and desired time to live and
1087 	 * send to IP level.  There should be a better way
1088 	 * to handle ttl and tos; we could keep them in
1089 	 * the template, but need a way to checksum without them.
1090 	 */
1091 	/*
1092 	 * m->m_pkthdr.len should have been set before cksum calcuration,
1093 	 * because in6_cksum() need it.
1094 	 */
1095 #ifdef INET6
1096 	if (isipv6) {
1097 		/*
1098 		 * we separately set hoplimit for every segment, since the
1099 		 * user might want to change the value via setsockopt.
1100 		 * Also, desired default hop limit might be changed via
1101 		 * Neighbor Discovery.
1102 		 */
1103 		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1104 
1105 		/* TODO: IPv6 IP6TOS_ECT bit on */
1106 		error = ip6_output(m,
1107 			    tp->t_inpcb->in6p_outputopts, NULL,
1108 			    ((so->so_options & SO_DONTROUTE) ?
1109 			    IP_ROUTETOIF : 0), NULL, NULL, tp->t_inpcb);
1110 	} else
1111 #endif /* INET6 */
1112     {
1113 	ip->ip_len = m->m_pkthdr.len;
1114 #ifdef INET6
1115 	if (INP_CHECK_SOCKAF(so, AF_INET6))
1116 		ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1117 #endif /* INET6 */
1118 	/*
1119 	 * If we do path MTU discovery, then we set DF on every packet.
1120 	 * This might not be the best thing to do according to RFC3390
1121 	 * Section 2. However the tcp hostcache migitates the problem
1122 	 * so it affects only the first tcp connection with a host.
1123 	 */
1124 	if (path_mtu_discovery)
1125 		ip->ip_off |= IP_DF;
1126 
1127 	error = ip_output(m, tp->t_inpcb->inp_options, NULL,
1128 	    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1129 	    tp->t_inpcb);
1130     }
1131 	if (error) {
1132 
1133 		/*
1134 		 * We know that the packet was lost, so back out the
1135 		 * sequence number advance, if any.
1136 		 *
1137 		 * If the error is EPERM the packet got blocked by the
1138 		 * local firewall.  Normally we should terminate the
1139 		 * connection but the blocking may have been spurious
1140 		 * due to a firewall reconfiguration cycle.  So we treat
1141 		 * it like a packet loss and let the retransmit timer and
1142 		 * timeouts do their work over time.
1143 		 * XXX: It is a POLA question whether calling tcp_drop right
1144 		 * away would be the really correct behavior instead.
1145 		 */
1146 		if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1147 		    !tcp_timer_active(tp, TT_PERSIST)) &&
1148 		    ((flags & TH_SYN) == 0) &&
1149 		    (error != EPERM)) {
1150 			if (sack_rxmit) {
1151 				p->rxmit -= len;
1152 				tp->sackhint.sack_bytes_rexmit -= len;
1153 				KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1154 				    ("sackhint bytes rtx >= 0"));
1155 			} else
1156 				tp->snd_nxt -= len;
1157 		}
1158 out:
1159 		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1160 		switch (error) {
1161 		case EPERM:
1162 			tp->t_softerror = error;
1163 			return (error);
1164 		case ENOBUFS:
1165 	                if (!tcp_timer_active(tp, TT_REXMT) &&
1166 			    !tcp_timer_active(tp, TT_PERSIST))
1167 	                        tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1168 			tp->snd_cwnd = tp->t_maxseg;
1169 			return (0);
1170 		case EMSGSIZE:
1171 			/*
1172 			 * For some reason the interface we used initially
1173 			 * to send segments changed to another or lowered
1174 			 * its MTU.
1175 			 *
1176 			 * tcp_mtudisc() will find out the new MTU and as
1177 			 * its last action, initiate retransmission, so it
1178 			 * is important to not do so here.
1179 			 *
1180 			 * If TSO was active we either got an interface
1181 			 * without TSO capabilits or TSO was turned off.
1182 			 * Disable it for this connection as too and
1183 			 * immediatly retry with MSS sized segments generated
1184 			 * by this function.
1185 			 */
1186 			if (tso)
1187 				tp->t_flags &= ~TF_TSO;
1188 			tcp_mtudisc(tp->t_inpcb, 0);
1189 			return (0);
1190 		case EHOSTDOWN:
1191 		case EHOSTUNREACH:
1192 		case ENETDOWN:
1193 		case ENETUNREACH:
1194 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
1195 				tp->t_softerror = error;
1196 				return (0);
1197 			}
1198 			/* FALLTHROUGH */
1199 		default:
1200 			return (error);
1201 		}
1202 	}
1203 	tcpstat.tcps_sndtotal++;
1204 
1205 	/*
1206 	 * Data sent (as far as we can tell).
1207 	 * If this advertises a larger window than any other segment,
1208 	 * then remember the size of the advertised window.
1209 	 * Any pending ACK has now been sent.
1210 	 */
1211 	if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1212 		tp->rcv_adv = tp->rcv_nxt + recwin;
1213 	tp->last_ack_sent = tp->rcv_nxt;
1214 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1215 	if (tcp_timer_active(tp, TT_DELACK))
1216 		tcp_timer_activate(tp, TT_DELACK, 0);
1217 #if 0
1218 	/*
1219 	 * This completely breaks TCP if newreno is turned on.  What happens
1220 	 * is that if delayed-acks are turned on on the receiver, this code
1221 	 * on the transmitter effectively destroys the TCP window, forcing
1222 	 * it to four packets (1.5Kx4 = 6K window).
1223 	 */
1224 	if (sendalot && (!tcp_do_newreno || --maxburst))
1225 		goto again;
1226 #endif
1227 	if (sendalot)
1228 		goto again;
1229 	return (0);
1230 }
1231 
1232 void
1233 tcp_setpersist(struct tcpcb *tp)
1234 {
1235 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1236 	int tt;
1237 
1238 	if (tcp_timer_active(tp, TT_REXMT))
1239 		panic("tcp_setpersist: retransmit pending");
1240 	/*
1241 	 * Start/restart persistance timer.
1242 	 */
1243 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1244 		      TCPTV_PERSMIN, TCPTV_PERSMAX);
1245 	tcp_timer_activate(tp, TT_PERSIST, tt);
1246 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1247 		tp->t_rxtshift++;
1248 }
1249 
1250 /*
1251  * Insert TCP options according to the supplied parameters to the place
1252  * optp in a consistent way.  Can handle unaligned destinations.
1253  *
1254  * The order of the option processing is crucial for optimal packing and
1255  * alignment for the scarce option space.
1256  *
1257  * The optimal order for a SYN/SYN-ACK segment is:
1258  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1259  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1260  *
1261  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1262  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1263  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1264  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1265  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1266  */
1267 int
1268 tcp_addoptions(struct tcpopt *to, u_char *optp)
1269 {
1270 	u_int mask, optlen = 0;
1271 
1272 	for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1273 		if ((to->to_flags & mask) != mask)
1274 			continue;
1275 		switch (to->to_flags & mask) {
1276 		case TOF_MSS:
1277 			while (optlen % 4) {
1278 				optlen += TCPOLEN_NOP;
1279 				*optp++ = TCPOPT_NOP;
1280 			}
1281 			optlen += TCPOLEN_MAXSEG;
1282 			*optp++ = TCPOPT_MAXSEG;
1283 			*optp++ = TCPOLEN_MAXSEG;
1284 			to->to_mss = htons(to->to_mss);
1285 			bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1286 			optp += sizeof(to->to_mss);
1287 			break;
1288 		case TOF_SCALE:
1289 			while (!optlen || optlen % 2 != 1) {
1290 				optlen += TCPOLEN_NOP;
1291 				*optp++ = TCPOPT_NOP;
1292 			}
1293 			optlen += TCPOLEN_WINDOW;
1294 			*optp++ = TCPOPT_WINDOW;
1295 			*optp++ = TCPOLEN_WINDOW;
1296 			*optp++ = to->to_wscale;
1297 			break;
1298 		case TOF_SACKPERM:
1299 			while (optlen % 2) {
1300 				optlen += TCPOLEN_NOP;
1301 				*optp++ = TCPOPT_NOP;
1302 			}
1303 			optlen += TCPOLEN_SACK_PERMITTED;
1304 			*optp++ = TCPOPT_SACK_PERMITTED;
1305 			*optp++ = TCPOLEN_SACK_PERMITTED;
1306 			break;
1307 		case TOF_TS:
1308 			while (!optlen || optlen % 4 != 2) {
1309 				optlen += TCPOLEN_NOP;
1310 				*optp++ = TCPOPT_NOP;
1311 			}
1312 			optlen += TCPOLEN_TIMESTAMP;
1313 			*optp++ = TCPOPT_TIMESTAMP;
1314 			*optp++ = TCPOLEN_TIMESTAMP;
1315 			to->to_tsval = htonl(to->to_tsval);
1316 			to->to_tsecr = htonl(to->to_tsecr);
1317 			bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1318 			optp += sizeof(to->to_tsval);
1319 			bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1320 			optp += sizeof(to->to_tsecr);
1321 			break;
1322 		case TOF_SIGNATURE:
1323 			{
1324 			int siglen = TCPOLEN_SIGNATURE - 2;
1325 
1326 			while (!optlen || optlen % 4 != 2) {
1327 				optlen += TCPOLEN_NOP;
1328 				*optp++ = TCPOPT_NOP;
1329 			}
1330 			if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE)
1331 				continue;
1332 			optlen += TCPOLEN_SIGNATURE;
1333 			*optp++ = TCPOPT_SIGNATURE;
1334 			*optp++ = TCPOLEN_SIGNATURE;
1335 			to->to_signature = optp;
1336 			while (siglen--)
1337 				 *optp++ = 0;
1338 			break;
1339 			}
1340 		case TOF_SACK:
1341 			{
1342 			int sackblks = 0;
1343 			struct sackblk *sack = (struct sackblk *)to->to_sacks;
1344 			tcp_seq sack_seq;
1345 
1346 			while (!optlen || optlen % 4 != 2) {
1347 				optlen += TCPOLEN_NOP;
1348 				*optp++ = TCPOPT_NOP;
1349 			}
1350 			if (TCP_MAXOLEN - optlen < 2 + TCPOLEN_SACK)
1351 				continue;
1352 			optlen += TCPOLEN_SACKHDR;
1353 			*optp++ = TCPOPT_SACK;
1354 			sackblks = min(to->to_nsacks,
1355 					(TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1356 			*optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1357 			while (sackblks--) {
1358 				sack_seq = htonl(sack->start);
1359 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1360 				optp += sizeof(sack_seq);
1361 				sack_seq = htonl(sack->end);
1362 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1363 				optp += sizeof(sack_seq);
1364 				optlen += TCPOLEN_SACK;
1365 				sack++;
1366 			}
1367 			tcpstat.tcps_sack_send_blocks++;
1368 			break;
1369 			}
1370 		default:
1371 			panic("%s: unknown TCP option type", __func__);
1372 			break;
1373 		}
1374 	}
1375 
1376 	/* Terminate and pad TCP options to a 4 byte boundary. */
1377 	if (optlen % 4) {
1378 		optlen += TCPOLEN_EOL;
1379 		*optp++ = TCPOPT_EOL;
1380 	}
1381 	while (optlen % 4) {
1382 		optlen += TCPOLEN_NOP;
1383 		*optp++ = TCPOPT_NOP;
1384 	}
1385 
1386 	KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1387 	return (optlen);
1388 }
1389