xref: /freebsd/sys/netinet/tcp_output.c (revision 2be1a816b9ff69588e55be0a84cbe2a31efc0f2f)
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 	struct mbuf *m;
132 	struct ip *ip = NULL;
133 	struct ipovly *ipov = NULL;
134 	struct tcphdr *th;
135 	u_char opt[TCP_MAXOLEN];
136 	unsigned ipoptlen, optlen, hdrlen;
137 #ifdef IPSEC
138 	unsigned ipsec_optlen = 0;
139 #endif
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_WLOCK_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 #ifdef IPSEC
460 	/*
461 	 * Pre-calculate here as we save another lookup into the darknesses
462 	 * of IPsec that way and can actually decide if TSO is ok.
463 	 */
464 	ipsec_optlen = ipsec_hdrsiz_tcp(tp);
465 #endif
466 	if (len > tp->t_maxseg) {
467 		if ((tp->t_flags & TF_TSO) && tcp_do_tso &&
468 		    ((tp->t_flags & TF_SIGNATURE) == 0) &&
469 		    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
470 		    tp->t_inpcb->inp_options == NULL &&
471 		    tp->t_inpcb->in6p_options == NULL
472 #ifdef IPSEC
473 		    && ipsec_optlen == 0
474 #endif
475 		    ) {
476 			tso = 1;
477 		} else {
478 			len = tp->t_maxseg;
479 			sendalot = 1;
480 			tso = 0;
481 		}
482 	}
483 	if (sack_rxmit) {
484 		if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
485 			flags &= ~TH_FIN;
486 	} else {
487 		if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
488 			flags &= ~TH_FIN;
489 	}
490 
491 	recwin = sbspace(&so->so_rcv);
492 
493 	/*
494 	 * Sender silly window avoidance.   We transmit under the following
495 	 * conditions when len is non-zero:
496 	 *
497 	 *	- We have a full segment (or more with TSO)
498 	 *	- This is the last buffer in a write()/send() and we are
499 	 *	  either idle or running NODELAY
500 	 *	- we've timed out (e.g. persist timer)
501 	 *	- we have more then 1/2 the maximum send window's worth of
502 	 *	  data (receiver may be limited the window size)
503 	 *	- we need to retransmit
504 	 */
505 	if (len) {
506 		if (len >= tp->t_maxseg)
507 			goto send;
508 		/*
509 		 * NOTE! on localhost connections an 'ack' from the remote
510 		 * end may occur synchronously with the output and cause
511 		 * us to flush a buffer queued with moretocome.  XXX
512 		 *
513 		 * note: the len + off check is almost certainly unnecessary.
514 		 */
515 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
516 		    (idle || (tp->t_flags & TF_NODELAY)) &&
517 		    len + off >= so->so_snd.sb_cc &&
518 		    (tp->t_flags & TF_NOPUSH) == 0) {
519 			goto send;
520 		}
521 		if (tp->t_flags & TF_FORCEDATA)		/* typ. timeout case */
522 			goto send;
523 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
524 			goto send;
525 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
526 			goto send;
527 		if (sack_rxmit)
528 			goto send;
529 	}
530 
531 	/*
532 	 * Compare available window to amount of window
533 	 * known to peer (as advertised window less
534 	 * next expected input).  If the difference is at least two
535 	 * max size segments, or at least 50% of the maximum possible
536 	 * window, then want to send a window update to peer.
537 	 * Skip this if the connection is in T/TCP half-open state.
538 	 * Don't send pure window updates when the peer has closed
539 	 * the connection and won't ever send more data.
540 	 */
541 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
542 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
543 		/*
544 		 * "adv" is the amount we can increase the window,
545 		 * taking into account that we are limited by
546 		 * TCP_MAXWIN << tp->rcv_scale.
547 		 */
548 		long adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale) -
549 			(tp->rcv_adv - tp->rcv_nxt);
550 
551 		if (adv >= (long) (2 * tp->t_maxseg))
552 			goto send;
553 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
554 			goto send;
555 	}
556 
557 	/*
558 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
559 	 * is also a catch-all for the retransmit timer timeout case.
560 	 */
561 	if (tp->t_flags & TF_ACKNOW)
562 		goto send;
563 	if ((flags & TH_RST) ||
564 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
565 		goto send;
566 	if (SEQ_GT(tp->snd_up, tp->snd_una))
567 		goto send;
568 	/*
569 	 * If our state indicates that FIN should be sent
570 	 * and we have not yet done so, then we need to send.
571 	 */
572 	if (flags & TH_FIN &&
573 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
574 		goto send;
575 	/*
576 	 * In SACK, it is possible for tcp_output to fail to send a segment
577 	 * after the retransmission timer has been turned off.  Make sure
578 	 * that the retransmission timer is set.
579 	 */
580 	if ((tp->t_flags & TF_SACK_PERMIT) &&
581 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
582 	    !tcp_timer_active(tp, TT_REXMT) &&
583 	    !tcp_timer_active(tp, TT_PERSIST)) {
584 		tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
585 		goto just_return;
586 	}
587 	/*
588 	 * TCP window updates are not reliable, rather a polling protocol
589 	 * using ``persist'' packets is used to insure receipt of window
590 	 * updates.  The three ``states'' for the output side are:
591 	 *	idle			not doing retransmits or persists
592 	 *	persisting		to move a small or zero window
593 	 *	(re)transmitting	and thereby not persisting
594 	 *
595 	 * tcp_timer_active(tp, TT_PERSIST)
596 	 *	is true when we are in persist state.
597 	 * (tp->t_flags & TF_FORCEDATA)
598 	 *	is set when we are called to send a persist packet.
599 	 * tcp_timer_active(tp, TT_REXMT)
600 	 *	is set when we are retransmitting
601 	 * The output side is idle when both timers are zero.
602 	 *
603 	 * If send window is too small, there is data to transmit, and no
604 	 * retransmit or persist is pending, then go to persist state.
605 	 * If nothing happens soon, send when timer expires:
606 	 * if window is nonzero, transmit what we can,
607 	 * otherwise force out a byte.
608 	 */
609 	if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) &&
610 	    !tcp_timer_active(tp, TT_PERSIST)) {
611 		tp->t_rxtshift = 0;
612 		tcp_setpersist(tp);
613 	}
614 
615 	/*
616 	 * No reason to send a segment, just return.
617 	 */
618 just_return:
619 	SOCKBUF_UNLOCK(&so->so_snd);
620 	return (0);
621 
622 send:
623 	SOCKBUF_LOCK_ASSERT(&so->so_snd);
624 	/*
625 	 * Before ESTABLISHED, force sending of initial options
626 	 * unless TCP set not to do any options.
627 	 * NOTE: we assume that the IP/TCP header plus TCP options
628 	 * always fit in a single mbuf, leaving room for a maximum
629 	 * link header, i.e.
630 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
631 	 */
632 	optlen = 0;
633 #ifdef INET6
634 	if (isipv6)
635 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
636 	else
637 #endif
638 	hdrlen = sizeof (struct tcpiphdr);
639 
640 	/*
641 	 * Compute options for segment.
642 	 * We only have to care about SYN and established connection
643 	 * segments.  Options for SYN-ACK segments are handled in TCP
644 	 * syncache.
645 	 */
646 	if ((tp->t_flags & TF_NOOPT) == 0) {
647 		to.to_flags = 0;
648 		/* Maximum segment size. */
649 		if (flags & TH_SYN) {
650 			tp->snd_nxt = tp->iss;
651 			to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
652 			to.to_flags |= TOF_MSS;
653 		}
654 		/* Window scaling. */
655 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
656 			to.to_wscale = tp->request_r_scale;
657 			to.to_flags |= TOF_SCALE;
658 		}
659 		/* Timestamps. */
660 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
661 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
662 			to.to_tsval = ticks + tp->ts_offset;
663 			to.to_tsecr = tp->ts_recent;
664 			to.to_flags |= TOF_TS;
665 			/* Set receive buffer autosizing timestamp. */
666 			if (tp->rfbuf_ts == 0 &&
667 			    (so->so_rcv.sb_flags & SB_AUTOSIZE))
668 				tp->rfbuf_ts = ticks;
669 		}
670 		/* Selective ACK's. */
671 		if (tp->t_flags & TF_SACK_PERMIT) {
672 			if (flags & TH_SYN)
673 				to.to_flags |= TOF_SACKPERM;
674 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
675 			    (tp->t_flags & TF_SACK_PERMIT) &&
676 			    tp->rcv_numsacks > 0) {
677 				to.to_flags |= TOF_SACK;
678 				to.to_nsacks = tp->rcv_numsacks;
679 				to.to_sacks = (u_char *)tp->sackblks;
680 			}
681 		}
682 #ifdef TCP_SIGNATURE
683 		/* TCP-MD5 (RFC2385). */
684 #ifdef INET6
685 		if (!isipv6 && (tp->t_flags & TF_SIGNATURE))
686 #else
687 		if (tp->t_flags & TF_SIGNATURE)
688 #endif /* INET6 */
689 			to.to_flags |= TOF_SIGNATURE;
690 #endif /* TCP_SIGNATURE */
691 
692 		/* Processing the options. */
693 		hdrlen += optlen = tcp_addoptions(&to, opt);
694 	}
695 
696 #ifdef INET6
697 	if (isipv6)
698 		ipoptlen = ip6_optlen(tp->t_inpcb);
699 	else
700 #endif
701 	if (tp->t_inpcb->inp_options)
702 		ipoptlen = tp->t_inpcb->inp_options->m_len -
703 				offsetof(struct ipoption, ipopt_list);
704 	else
705 		ipoptlen = 0;
706 #ifdef IPSEC
707 	ipoptlen += ipsec_optlen;
708 #endif
709 
710 	/*
711 	 * Adjust data length if insertion of options will
712 	 * bump the packet length beyond the t_maxopd length.
713 	 * Clear the FIN bit because we cut off the tail of
714 	 * the segment.
715 	 *
716 	 * When doing TSO limit a burst to TCP_MAXWIN minus the
717 	 * IP, TCP and Options length to keep ip->ip_len from
718 	 * overflowing.  Prevent the last segment from being
719 	 * fractional thus making them all equal sized and set
720 	 * the flag to continue sending.  TSO is disabled when
721 	 * IP options or IPSEC are present.
722 	 */
723 	if (len + optlen + ipoptlen > tp->t_maxopd) {
724 		flags &= ~TH_FIN;
725 		if (tso) {
726 			if (len > TCP_MAXWIN - hdrlen - optlen) {
727 				len = TCP_MAXWIN - hdrlen - optlen;
728 				len = len - (len % (tp->t_maxopd - optlen));
729 				sendalot = 1;
730 			} else if (tp->t_flags & TF_NEEDFIN)
731 				sendalot = 1;
732 		} else {
733 			len = tp->t_maxopd - optlen - ipoptlen;
734 			sendalot = 1;
735 		}
736 	}
737 
738 /*#ifdef DIAGNOSTIC*/
739 #ifdef INET6
740 	if (max_linkhdr + hdrlen > MCLBYTES)
741 #else
742 	if (max_linkhdr + hdrlen > MHLEN)
743 #endif
744 		panic("tcphdr too big");
745 /*#endif*/
746 
747 	/*
748 	 * Grab a header mbuf, attaching a copy of data to
749 	 * be transmitted, and initialize the header from
750 	 * the template for sends on this connection.
751 	 */
752 	if (len) {
753 		struct mbuf *mb;
754 		u_int moff;
755 
756 		if ((tp->t_flags & TF_FORCEDATA) && len == 1)
757 			tcpstat.tcps_sndprobe++;
758 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
759 			tcpstat.tcps_sndrexmitpack++;
760 			tcpstat.tcps_sndrexmitbyte += len;
761 		} else {
762 			tcpstat.tcps_sndpack++;
763 			tcpstat.tcps_sndbyte += len;
764 		}
765 #ifdef notyet
766 		if ((m = m_copypack(so->so_snd.sb_mb, off,
767 		    (int)len, max_linkhdr + hdrlen)) == 0) {
768 			SOCKBUF_UNLOCK(&so->so_snd);
769 			error = ENOBUFS;
770 			goto out;
771 		}
772 		/*
773 		 * m_copypack left space for our hdr; use it.
774 		 */
775 		m->m_len += hdrlen;
776 		m->m_data -= hdrlen;
777 #else
778 		MGETHDR(m, M_DONTWAIT, MT_DATA);
779 		if (m == NULL) {
780 			SOCKBUF_UNLOCK(&so->so_snd);
781 			error = ENOBUFS;
782 			goto out;
783 		}
784 #ifdef INET6
785 		if (MHLEN < hdrlen + max_linkhdr) {
786 			MCLGET(m, M_DONTWAIT);
787 			if ((m->m_flags & M_EXT) == 0) {
788 				SOCKBUF_UNLOCK(&so->so_snd);
789 				m_freem(m);
790 				error = ENOBUFS;
791 				goto out;
792 			}
793 		}
794 #endif
795 		m->m_data += max_linkhdr;
796 		m->m_len = hdrlen;
797 
798 		/*
799 		 * Start the m_copy functions from the closest mbuf
800 		 * to the offset in the socket buffer chain.
801 		 */
802 		mb = sbsndptr(&so->so_snd, off, len, &moff);
803 
804 		if (len <= MHLEN - hdrlen - max_linkhdr) {
805 			m_copydata(mb, moff, (int)len,
806 			    mtod(m, caddr_t) + hdrlen);
807 			m->m_len += len;
808 		} else {
809 			m->m_next = m_copy(mb, moff, (int)len);
810 			if (m->m_next == NULL) {
811 				SOCKBUF_UNLOCK(&so->so_snd);
812 				(void) m_free(m);
813 				error = ENOBUFS;
814 				goto out;
815 			}
816 		}
817 #endif
818 		/*
819 		 * If we're sending everything we've got, set PUSH.
820 		 * (This will keep happy those implementations which only
821 		 * give data to the user when a buffer fills or
822 		 * a PUSH comes in.)
823 		 */
824 		if (off + len == so->so_snd.sb_cc)
825 			flags |= TH_PUSH;
826 		SOCKBUF_UNLOCK(&so->so_snd);
827 	} else {
828 		SOCKBUF_UNLOCK(&so->so_snd);
829 		if (tp->t_flags & TF_ACKNOW)
830 			tcpstat.tcps_sndacks++;
831 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
832 			tcpstat.tcps_sndctrl++;
833 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
834 			tcpstat.tcps_sndurg++;
835 		else
836 			tcpstat.tcps_sndwinup++;
837 
838 		MGETHDR(m, M_DONTWAIT, MT_DATA);
839 		if (m == NULL) {
840 			error = ENOBUFS;
841 			goto out;
842 		}
843 #ifdef INET6
844 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
845 		    MHLEN >= hdrlen) {
846 			MH_ALIGN(m, hdrlen);
847 		} else
848 #endif
849 		m->m_data += max_linkhdr;
850 		m->m_len = hdrlen;
851 	}
852 	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
853 	m->m_pkthdr.rcvif = (struct ifnet *)0;
854 #ifdef MAC
855 	mac_inpcb_create_mbuf(tp->t_inpcb, m);
856 #endif
857 #ifdef INET6
858 	if (isipv6) {
859 		ip6 = mtod(m, struct ip6_hdr *);
860 		th = (struct tcphdr *)(ip6 + 1);
861 		tcpip_fillheaders(tp->t_inpcb, ip6, th);
862 	} else
863 #endif /* INET6 */
864 	{
865 		ip = mtod(m, struct ip *);
866 		ipov = (struct ipovly *)ip;
867 		th = (struct tcphdr *)(ip + 1);
868 		tcpip_fillheaders(tp->t_inpcb, ip, th);
869 	}
870 
871 	/*
872 	 * Fill in fields, remembering maximum advertised
873 	 * window for use in delaying messages about window sizes.
874 	 * If resending a FIN, be sure not to use a new sequence number.
875 	 */
876 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
877 	    tp->snd_nxt == tp->snd_max)
878 		tp->snd_nxt--;
879 	/*
880 	 * If we are doing retransmissions, then snd_nxt will
881 	 * not reflect the first unsent octet.  For ACK only
882 	 * packets, we do not want the sequence number of the
883 	 * retransmitted packet, we want the sequence number
884 	 * of the next unsent octet.  So, if there is no data
885 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
886 	 * when filling in ti_seq.  But if we are in persist
887 	 * state, snd_max might reflect one byte beyond the
888 	 * right edge of the window, so use snd_nxt in that
889 	 * case, since we know we aren't doing a retransmission.
890 	 * (retransmit and persist are mutually exclusive...)
891 	 */
892 	if (sack_rxmit == 0) {
893 		if (len || (flags & (TH_SYN|TH_FIN)) ||
894 		    tcp_timer_active(tp, TT_PERSIST))
895 			th->th_seq = htonl(tp->snd_nxt);
896 		else
897 			th->th_seq = htonl(tp->snd_max);
898 	} else {
899 		th->th_seq = htonl(p->rxmit);
900 		p->rxmit += len;
901 		tp->sackhint.sack_bytes_rexmit += len;
902 	}
903 	th->th_ack = htonl(tp->rcv_nxt);
904 	if (optlen) {
905 		bcopy(opt, th + 1, optlen);
906 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
907 	}
908 	th->th_flags = flags;
909 	/*
910 	 * Calculate receive window.  Don't shrink window,
911 	 * but avoid silly window syndrome.
912 	 */
913 	if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
914 	    recwin < (long)tp->t_maxseg)
915 		recwin = 0;
916 	if (recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
917 		recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
918 	if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
919 		recwin = (long)TCP_MAXWIN << tp->rcv_scale;
920 
921 	/*
922 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
923 	 * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
924 	 * case is handled in syncache.
925 	 */
926 	if (flags & TH_SYN)
927 		th->th_win = htons((u_short)
928 				(min(sbspace(&so->so_rcv), TCP_MAXWIN)));
929 	else
930 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
931 
932 	/*
933 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
934 	 * a 0 window.  This may cause the remote transmitter to stall.  This
935 	 * flag tells soreceive() to disable delayed acknowledgements when
936 	 * draining the buffer.  This can occur if the receiver is attempting
937 	 * to read more data then can be buffered prior to transmitting on
938 	 * the connection.
939 	 */
940 	if (recwin == 0)
941 		tp->t_flags |= TF_RXWIN0SENT;
942 	else
943 		tp->t_flags &= ~TF_RXWIN0SENT;
944 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
945 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
946 		th->th_flags |= TH_URG;
947 	} else
948 		/*
949 		 * If no urgent pointer to send, then we pull
950 		 * the urgent pointer to the left edge of the send window
951 		 * so that it doesn't drift into the send window on sequence
952 		 * number wraparound.
953 		 */
954 		tp->snd_up = tp->snd_una;		/* drag it along */
955 
956 #ifdef TCP_SIGNATURE
957 #ifdef INET6
958 	if (!isipv6)
959 #endif
960 	if (tp->t_flags & TF_SIGNATURE) {
961 		int sigoff = to.to_signature - opt;
962 		tcp_signature_compute(m, sizeof(struct ip), len, optlen,
963 		    (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
964 	}
965 #endif
966 
967 	/*
968 	 * Put TCP length in extended header, and then
969 	 * checksum extended header and data.
970 	 */
971 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
972 #ifdef INET6
973 	if (isipv6)
974 		/*
975 		 * ip6_plen is not need to be filled now, and will be filled
976 		 * in ip6_output.
977 		 */
978 		th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
979 				       sizeof(struct tcphdr) + optlen + len);
980 	else
981 #endif /* INET6 */
982 	{
983 		m->m_pkthdr.csum_flags = CSUM_TCP;
984 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
985 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
986 		    htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
987 
988 		/* IP version must be set here for ipv4/ipv6 checking later */
989 		KASSERT(ip->ip_v == IPVERSION,
990 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
991 	}
992 
993 	/*
994 	 * Enable TSO and specify the size of the segments.
995 	 * The TCP pseudo header checksum is always provided.
996 	 * XXX: Fixme: This is currently not the case for IPv6.
997 	 */
998 	if (tso) {
999 		m->m_pkthdr.csum_flags = CSUM_TSO;
1000 		m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen;
1001 	}
1002 
1003 	/*
1004 	 * In transmit state, time the transmission and arrange for
1005 	 * the retransmit.  In persist state, just set snd_max.
1006 	 */
1007 	if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1008 	    !tcp_timer_active(tp, TT_PERSIST)) {
1009 		tcp_seq startseq = tp->snd_nxt;
1010 
1011 		/*
1012 		 * Advance snd_nxt over sequence space of this segment.
1013 		 */
1014 		if (flags & (TH_SYN|TH_FIN)) {
1015 			if (flags & TH_SYN)
1016 				tp->snd_nxt++;
1017 			if (flags & TH_FIN) {
1018 				tp->snd_nxt++;
1019 				tp->t_flags |= TF_SENTFIN;
1020 			}
1021 		}
1022 		if (sack_rxmit)
1023 			goto timer;
1024 		tp->snd_nxt += len;
1025 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1026 			tp->snd_max = tp->snd_nxt;
1027 			/*
1028 			 * Time this transmission if not a retransmission and
1029 			 * not currently timing anything.
1030 			 */
1031 			if (tp->t_rtttime == 0) {
1032 				tp->t_rtttime = ticks;
1033 				tp->t_rtseq = startseq;
1034 				tcpstat.tcps_segstimed++;
1035 			}
1036 		}
1037 
1038 		/*
1039 		 * Set retransmit timer if not currently set,
1040 		 * and not doing a pure ack or a keep-alive probe.
1041 		 * Initial value for retransmit timer is smoothed
1042 		 * round-trip time + 2 * round-trip time variance.
1043 		 * Initialize shift counter which is used for backoff
1044 		 * of retransmit time.
1045 		 */
1046 timer:
1047 		if (!tcp_timer_active(tp, TT_REXMT) &&
1048 		    ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1049 		     (tp->snd_nxt != tp->snd_una))) {
1050 			if (tcp_timer_active(tp, TT_PERSIST)) {
1051 				tcp_timer_activate(tp, TT_PERSIST, 0);
1052 				tp->t_rxtshift = 0;
1053 			}
1054 			tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1055 		}
1056 	} else {
1057 		/*
1058 		 * Persist case, update snd_max but since we are in
1059 		 * persist mode (no window) we do not update snd_nxt.
1060 		 */
1061 		int xlen = len;
1062 		if (flags & TH_SYN)
1063 			++xlen;
1064 		if (flags & TH_FIN) {
1065 			++xlen;
1066 			tp->t_flags |= TF_SENTFIN;
1067 		}
1068 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1069 			tp->snd_max = tp->snd_nxt + len;
1070 	}
1071 
1072 #ifdef TCPDEBUG
1073 	/*
1074 	 * Trace.
1075 	 */
1076 	if (so->so_options & SO_DEBUG) {
1077 		u_short save = 0;
1078 #ifdef INET6
1079 		if (!isipv6)
1080 #endif
1081 		{
1082 			save = ipov->ih_len;
1083 			ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1084 		}
1085 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1086 #ifdef INET6
1087 		if (!isipv6)
1088 #endif
1089 		ipov->ih_len = save;
1090 	}
1091 #endif
1092 
1093 	/*
1094 	 * Fill in IP length and desired time to live and
1095 	 * send to IP level.  There should be a better way
1096 	 * to handle ttl and tos; we could keep them in
1097 	 * the template, but need a way to checksum without them.
1098 	 */
1099 	/*
1100 	 * m->m_pkthdr.len should have been set before cksum calcuration,
1101 	 * because in6_cksum() need it.
1102 	 */
1103 #ifdef INET6
1104 	if (isipv6) {
1105 		/*
1106 		 * we separately set hoplimit for every segment, since the
1107 		 * user might want to change the value via setsockopt.
1108 		 * Also, desired default hop limit might be changed via
1109 		 * Neighbor Discovery.
1110 		 */
1111 		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1112 
1113 		/* TODO: IPv6 IP6TOS_ECT bit on */
1114 		error = ip6_output(m,
1115 			    tp->t_inpcb->in6p_outputopts, NULL,
1116 			    ((so->so_options & SO_DONTROUTE) ?
1117 			    IP_ROUTETOIF : 0), NULL, NULL, tp->t_inpcb);
1118 	} else
1119 #endif /* INET6 */
1120     {
1121 	ip->ip_len = m->m_pkthdr.len;
1122 #ifdef INET6
1123 	if (INP_CHECK_SOCKAF(so, AF_INET6))
1124 		ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1125 #endif /* INET6 */
1126 	/*
1127 	 * If we do path MTU discovery, then we set DF on every packet.
1128 	 * This might not be the best thing to do according to RFC3390
1129 	 * Section 2. However the tcp hostcache migitates the problem
1130 	 * so it affects only the first tcp connection with a host.
1131 	 */
1132 	if (path_mtu_discovery)
1133 		ip->ip_off |= IP_DF;
1134 
1135 	error = ip_output(m, tp->t_inpcb->inp_options, NULL,
1136 	    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1137 	    tp->t_inpcb);
1138     }
1139 	if (error) {
1140 
1141 		/*
1142 		 * We know that the packet was lost, so back out the
1143 		 * sequence number advance, if any.
1144 		 *
1145 		 * If the error is EPERM the packet got blocked by the
1146 		 * local firewall.  Normally we should terminate the
1147 		 * connection but the blocking may have been spurious
1148 		 * due to a firewall reconfiguration cycle.  So we treat
1149 		 * it like a packet loss and let the retransmit timer and
1150 		 * timeouts do their work over time.
1151 		 * XXX: It is a POLA question whether calling tcp_drop right
1152 		 * away would be the really correct behavior instead.
1153 		 */
1154 		if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1155 		    !tcp_timer_active(tp, TT_PERSIST)) &&
1156 		    ((flags & TH_SYN) == 0) &&
1157 		    (error != EPERM)) {
1158 			if (sack_rxmit) {
1159 				p->rxmit -= len;
1160 				tp->sackhint.sack_bytes_rexmit -= len;
1161 				KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1162 				    ("sackhint bytes rtx >= 0"));
1163 			} else
1164 				tp->snd_nxt -= len;
1165 		}
1166 out:
1167 		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1168 		switch (error) {
1169 		case EPERM:
1170 			tp->t_softerror = error;
1171 			return (error);
1172 		case ENOBUFS:
1173 	                if (!tcp_timer_active(tp, TT_REXMT) &&
1174 			    !tcp_timer_active(tp, TT_PERSIST))
1175 	                        tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1176 			tp->snd_cwnd = tp->t_maxseg;
1177 			return (0);
1178 		case EMSGSIZE:
1179 			/*
1180 			 * For some reason the interface we used initially
1181 			 * to send segments changed to another or lowered
1182 			 * its MTU.
1183 			 *
1184 			 * tcp_mtudisc() will find out the new MTU and as
1185 			 * its last action, initiate retransmission, so it
1186 			 * is important to not do so here.
1187 			 *
1188 			 * If TSO was active we either got an interface
1189 			 * without TSO capabilits or TSO was turned off.
1190 			 * Disable it for this connection as too and
1191 			 * immediatly retry with MSS sized segments generated
1192 			 * by this function.
1193 			 */
1194 			if (tso)
1195 				tp->t_flags &= ~TF_TSO;
1196 			tcp_mtudisc(tp->t_inpcb, 0);
1197 			return (0);
1198 		case EHOSTDOWN:
1199 		case EHOSTUNREACH:
1200 		case ENETDOWN:
1201 		case ENETUNREACH:
1202 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
1203 				tp->t_softerror = error;
1204 				return (0);
1205 			}
1206 			/* FALLTHROUGH */
1207 		default:
1208 			return (error);
1209 		}
1210 	}
1211 	tcpstat.tcps_sndtotal++;
1212 
1213 	/*
1214 	 * Data sent (as far as we can tell).
1215 	 * If this advertises a larger window than any other segment,
1216 	 * then remember the size of the advertised window.
1217 	 * Any pending ACK has now been sent.
1218 	 */
1219 	if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1220 		tp->rcv_adv = tp->rcv_nxt + recwin;
1221 	tp->last_ack_sent = tp->rcv_nxt;
1222 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1223 	if (tcp_timer_active(tp, TT_DELACK))
1224 		tcp_timer_activate(tp, TT_DELACK, 0);
1225 #if 0
1226 	/*
1227 	 * This completely breaks TCP if newreno is turned on.  What happens
1228 	 * is that if delayed-acks are turned on on the receiver, this code
1229 	 * on the transmitter effectively destroys the TCP window, forcing
1230 	 * it to four packets (1.5Kx4 = 6K window).
1231 	 */
1232 	if (sendalot && (!tcp_do_newreno || --maxburst))
1233 		goto again;
1234 #endif
1235 	if (sendalot)
1236 		goto again;
1237 	return (0);
1238 }
1239 
1240 void
1241 tcp_setpersist(struct tcpcb *tp)
1242 {
1243 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1244 	int tt;
1245 
1246 	if (tcp_timer_active(tp, TT_REXMT))
1247 		panic("tcp_setpersist: retransmit pending");
1248 	/*
1249 	 * Start/restart persistance timer.
1250 	 */
1251 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1252 		      TCPTV_PERSMIN, TCPTV_PERSMAX);
1253 	tcp_timer_activate(tp, TT_PERSIST, tt);
1254 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1255 		tp->t_rxtshift++;
1256 }
1257 
1258 /*
1259  * Insert TCP options according to the supplied parameters to the place
1260  * optp in a consistent way.  Can handle unaligned destinations.
1261  *
1262  * The order of the option processing is crucial for optimal packing and
1263  * alignment for the scarce option space.
1264  *
1265  * The optimal order for a SYN/SYN-ACK segment is:
1266  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1267  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1268  *
1269  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1270  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1271  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1272  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1273  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1274  */
1275 int
1276 tcp_addoptions(struct tcpopt *to, u_char *optp)
1277 {
1278 	u_int mask, optlen = 0;
1279 
1280 	for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1281 		if ((to->to_flags & mask) != mask)
1282 			continue;
1283 		if (optlen == TCP_MAXOLEN)
1284 			break;
1285 		switch (to->to_flags & mask) {
1286 		case TOF_MSS:
1287 			while (optlen % 4) {
1288 				optlen += TCPOLEN_NOP;
1289 				*optp++ = TCPOPT_NOP;
1290 			}
1291 			if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1292 				continue;
1293 			optlen += TCPOLEN_MAXSEG;
1294 			*optp++ = TCPOPT_MAXSEG;
1295 			*optp++ = TCPOLEN_MAXSEG;
1296 			to->to_mss = htons(to->to_mss);
1297 			bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1298 			optp += sizeof(to->to_mss);
1299 			break;
1300 		case TOF_SCALE:
1301 			while (!optlen || optlen % 2 != 1) {
1302 				optlen += TCPOLEN_NOP;
1303 				*optp++ = TCPOPT_NOP;
1304 			}
1305 			if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1306 				continue;
1307 			optlen += TCPOLEN_WINDOW;
1308 			*optp++ = TCPOPT_WINDOW;
1309 			*optp++ = TCPOLEN_WINDOW;
1310 			*optp++ = to->to_wscale;
1311 			break;
1312 		case TOF_SACKPERM:
1313 			while (optlen % 2) {
1314 				optlen += TCPOLEN_NOP;
1315 				*optp++ = TCPOPT_NOP;
1316 			}
1317 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1318 				continue;
1319 			optlen += TCPOLEN_SACK_PERMITTED;
1320 			*optp++ = TCPOPT_SACK_PERMITTED;
1321 			*optp++ = TCPOLEN_SACK_PERMITTED;
1322 			break;
1323 		case TOF_TS:
1324 			while (!optlen || optlen % 4 != 2) {
1325 				optlen += TCPOLEN_NOP;
1326 				*optp++ = TCPOPT_NOP;
1327 			}
1328 			if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1329 				continue;
1330 			optlen += TCPOLEN_TIMESTAMP;
1331 			*optp++ = TCPOPT_TIMESTAMP;
1332 			*optp++ = TCPOLEN_TIMESTAMP;
1333 			to->to_tsval = htonl(to->to_tsval);
1334 			to->to_tsecr = htonl(to->to_tsecr);
1335 			bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1336 			optp += sizeof(to->to_tsval);
1337 			bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1338 			optp += sizeof(to->to_tsecr);
1339 			break;
1340 		case TOF_SIGNATURE:
1341 			{
1342 			int siglen = TCPOLEN_SIGNATURE - 2;
1343 
1344 			while (!optlen || optlen % 4 != 2) {
1345 				optlen += TCPOLEN_NOP;
1346 				*optp++ = TCPOPT_NOP;
1347 			}
1348 			if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE)
1349 				continue;
1350 			optlen += TCPOLEN_SIGNATURE;
1351 			*optp++ = TCPOPT_SIGNATURE;
1352 			*optp++ = TCPOLEN_SIGNATURE;
1353 			to->to_signature = optp;
1354 			while (siglen--)
1355 				 *optp++ = 0;
1356 			break;
1357 			}
1358 		case TOF_SACK:
1359 			{
1360 			int sackblks = 0;
1361 			struct sackblk *sack = (struct sackblk *)to->to_sacks;
1362 			tcp_seq sack_seq;
1363 
1364 			while (!optlen || optlen % 4 != 2) {
1365 				optlen += TCPOLEN_NOP;
1366 				*optp++ = TCPOPT_NOP;
1367 			}
1368 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1369 				continue;
1370 			optlen += TCPOLEN_SACKHDR;
1371 			*optp++ = TCPOPT_SACK;
1372 			sackblks = min(to->to_nsacks,
1373 					(TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1374 			*optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1375 			while (sackblks--) {
1376 				sack_seq = htonl(sack->start);
1377 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1378 				optp += sizeof(sack_seq);
1379 				sack_seq = htonl(sack->end);
1380 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1381 				optp += sizeof(sack_seq);
1382 				optlen += TCPOLEN_SACK;
1383 				sack++;
1384 			}
1385 			tcpstat.tcps_sack_send_blocks++;
1386 			break;
1387 			}
1388 		default:
1389 			panic("%s: unknown TCP option type", __func__);
1390 			break;
1391 		}
1392 	}
1393 
1394 	/* Terminate and pad TCP options to a 4 byte boundary. */
1395 	if (optlen % 4) {
1396 		optlen += TCPOLEN_EOL;
1397 		*optp++ = TCPOPT_EOL;
1398 	}
1399 	/*
1400 	 * According to RFC 793 (STD0007):
1401 	 *   "The content of the header beyond the End-of-Option option
1402 	 *    must be header padding (i.e., zero)."
1403 	 *   and later: "The padding is composed of zeros."
1404 	 */
1405 	while (optlen % 4) {
1406 		optlen += TCPOLEN_PAD;
1407 		*optp++ = TCPOPT_PAD;
1408 	}
1409 
1410 	KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1411 	return (optlen);
1412 }
1413