xref: /freebsd/sys/netinet/tcp_output.c (revision 78007886c995898a9494648343e5236bca1cbba3)
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  * $FreeBSD$
31  */
32 
33 #include "opt_inet.h"
34 #include "opt_inet6.h"
35 #include "opt_ipsec.h"
36 #include "opt_mac.h"
37 #include "opt_tcpdebug.h"
38 
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/domain.h>
42 #include <sys/kernel.h>
43 #include <sys/lock.h>
44 #include <sys/mbuf.h>
45 #include <sys/mutex.h>
46 #include <sys/protosw.h>
47 #include <sys/socket.h>
48 #include <sys/socketvar.h>
49 #include <sys/sysctl.h>
50 
51 #include <net/route.h>
52 
53 #include <netinet/in.h>
54 #include <netinet/in_systm.h>
55 #include <netinet/ip.h>
56 #include <netinet/in_pcb.h>
57 #include <netinet/ip_var.h>
58 #include <netinet/ip_options.h>
59 #ifdef INET6
60 #include <netinet6/in6_pcb.h>
61 #include <netinet/ip6.h>
62 #include <netinet6/ip6_var.h>
63 #endif
64 #include <netinet/tcp.h>
65 #define	TCPOUTFLAGS
66 #include <netinet/tcp_fsm.h>
67 #include <netinet/tcp_seq.h>
68 #include <netinet/tcp_timer.h>
69 #include <netinet/tcp_var.h>
70 #include <netinet/tcpip.h>
71 #ifdef TCPDEBUG
72 #include <netinet/tcp_debug.h>
73 #endif
74 
75 #ifdef IPSEC
76 #include <netinet6/ipsec.h>
77 #endif /*IPSEC*/
78 
79 #ifdef FAST_IPSEC
80 #include <netipsec/ipsec.h>
81 #define	IPSEC
82 #endif /*FAST_IPSEC*/
83 
84 #include <machine/in_cksum.h>
85 
86 #include <security/mac/mac_framework.h>
87 
88 #ifdef notyet
89 extern struct mbuf *m_copypack();
90 #endif
91 
92 int path_mtu_discovery = 1;
93 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
94 	&path_mtu_discovery, 1, "Enable Path MTU Discovery");
95 
96 int ss_fltsz = 1;
97 SYSCTL_INT(_net_inet_tcp, OID_AUTO, slowstart_flightsize, CTLFLAG_RW,
98 	&ss_fltsz, 1, "Slow start flight size");
99 
100 int ss_fltsz_local = 4;
101 SYSCTL_INT(_net_inet_tcp, OID_AUTO, local_slowstart_flightsize, CTLFLAG_RW,
102 	&ss_fltsz_local, 1, "Slow start flight size for local networks");
103 
104 int     tcp_do_newreno = 1;
105 SYSCTL_INT(_net_inet_tcp, OID_AUTO, newreno, CTLFLAG_RW,
106 	&tcp_do_newreno, 0, "Enable NewReno Algorithms");
107 
108 int	tcp_do_tso = 1;
109 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW,
110 	&tcp_do_tso, 0, "Enable TCP Segmentation Offload");
111 
112 int	tcp_do_autosndbuf = 1;
113 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW,
114 	&tcp_do_autosndbuf, 0, "Enable automatic send buffer sizing");
115 
116 int	tcp_autosndbuf_inc = 8*1024;
117 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW,
118 	&tcp_autosndbuf_inc, 0, "Incrementor step size of automatic send buffer");
119 
120 int	tcp_autosndbuf_max = 256*1024;
121 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW,
122 	&tcp_autosndbuf_max, 0, "Max size of automatic send buffer");
123 
124 
125 /*
126  * Tcp output routine: figure out what should be sent and send it.
127  */
128 int
129 tcp_output(struct tcpcb *tp)
130 {
131 	struct socket *so = tp->t_inpcb->inp_socket;
132 	long len, recwin, sendwin;
133 	int off, flags, error;
134 #ifdef TCP_SIGNATURE
135 	int sigoff = 0;
136 #endif
137 	struct mbuf *m;
138 	struct ip *ip = NULL;
139 	struct ipovly *ipov = NULL;
140 	struct tcphdr *th;
141 	u_char opt[TCP_MAXOLEN];
142 	unsigned ipoptlen, optlen, hdrlen;
143 	int idle, sendalot;
144 	int sack_rxmit, sack_bytes_rxmt;
145 	struct sackhole *p;
146 	int tso = 0;
147 	struct tcpopt to;
148 #if 0
149 	int maxburst = TCP_MAXBURST;
150 #endif
151 #ifdef INET6
152 	struct ip6_hdr *ip6 = NULL;
153 	int isipv6;
154 
155 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
156 #endif
157 
158 	INP_LOCK_ASSERT(tp->t_inpcb);
159 
160 	/*
161 	 * Determine length of data that should be transmitted,
162 	 * and flags that will be used.
163 	 * If there is some data or critical controls (SYN, RST)
164 	 * to send, then transmit; otherwise, investigate further.
165 	 */
166 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
167 	if (idle && (ticks - tp->t_rcvtime) >= tp->t_rxtcur) {
168 		/*
169 		 * We have been idle for "a while" and no acks are
170 		 * expected to clock out any data we send --
171 		 * slow start to get ack "clock" running again.
172 		 *
173 		 * Set the slow-start flight size depending on whether
174 		 * this is a local network or not.
175 		 */
176 		int ss = ss_fltsz;
177 #ifdef INET6
178 		if (isipv6) {
179 			if (in6_localaddr(&tp->t_inpcb->in6p_faddr))
180 				ss = ss_fltsz_local;
181 		} else
182 #endif /* INET6 */
183 		if (in_localaddr(tp->t_inpcb->inp_faddr))
184 			ss = ss_fltsz_local;
185 		tp->snd_cwnd = tp->t_maxseg * ss;
186 	}
187 	tp->t_flags &= ~TF_LASTIDLE;
188 	if (idle) {
189 		if (tp->t_flags & TF_MORETOCOME) {
190 			tp->t_flags |= TF_LASTIDLE;
191 			idle = 0;
192 		}
193 	}
194 again:
195 	/*
196 	 * If we've recently taken a timeout, snd_max will be greater than
197 	 * snd_nxt.  There may be SACK information that allows us to avoid
198 	 * resending already delivered data.  Adjust snd_nxt accordingly.
199 	 */
200 	if (tp->sack_enable && SEQ_LT(tp->snd_nxt, tp->snd_max))
201 		tcp_sack_adjust(tp);
202 	sendalot = 0;
203 	off = tp->snd_nxt - tp->snd_una;
204 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
205 	sendwin = min(sendwin, tp->snd_bwnd);
206 
207 	flags = tcp_outflags[tp->t_state];
208 	/*
209 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
210 	 * to send out new data (when sendalot is 1), bypass this function.
211 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
212 	 * we're replacing a (future) new transmission with a retransmission
213 	 * now, and we previously incremented snd_cwnd in tcp_input().
214 	 */
215 	/*
216 	 * Still in sack recovery , reset rxmit flag to zero.
217 	 */
218 	sack_rxmit = 0;
219 	sack_bytes_rxmt = 0;
220 	len = 0;
221 	p = NULL;
222 	if (tp->sack_enable && IN_FASTRECOVERY(tp) &&
223 	    (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
224 		long cwin;
225 
226 		cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
227 		if (cwin < 0)
228 			cwin = 0;
229 		/* Do not retransmit SACK segments beyond snd_recover */
230 		if (SEQ_GT(p->end, tp->snd_recover)) {
231 			/*
232 			 * (At least) part of sack hole extends beyond
233 			 * snd_recover. Check to see if we can rexmit data
234 			 * for this hole.
235 			 */
236 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
237 				/*
238 				 * Can't rexmit any more data for this hole.
239 				 * That data will be rexmitted in the next
240 				 * sack recovery episode, when snd_recover
241 				 * moves past p->rxmit.
242 				 */
243 				p = NULL;
244 				goto after_sack_rexmit;
245 			} else
246 				/* Can rexmit part of the current hole */
247 				len = ((long)ulmin(cwin,
248 						   tp->snd_recover - p->rxmit));
249 		} else
250 			len = ((long)ulmin(cwin, p->end - p->rxmit));
251 		off = p->rxmit - tp->snd_una;
252 		KASSERT(off >= 0,("%s: sack block to the left of una : %d",
253 		    __func__, off));
254 		if (len > 0) {
255 			sack_rxmit = 1;
256 			sendalot = 1;
257 			tcpstat.tcps_sack_rexmits++;
258 			tcpstat.tcps_sack_rexmit_bytes +=
259 			    min(len, tp->t_maxseg);
260 		}
261 	}
262 after_sack_rexmit:
263 	/*
264 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
265 	 * state flags.
266 	 */
267 	if (tp->t_flags & TF_NEEDFIN)
268 		flags |= TH_FIN;
269 	if (tp->t_flags & TF_NEEDSYN)
270 		flags |= TH_SYN;
271 
272 	SOCKBUF_LOCK(&so->so_snd);
273 	/*
274 	 * If in persist timeout with window of 0, send 1 byte.
275 	 * Otherwise, if window is small but nonzero
276 	 * and timer expired, we will send what we can
277 	 * and go to transmit state.
278 	 */
279 	if (tp->t_flags & TF_FORCEDATA) {
280 		if (sendwin == 0) {
281 			/*
282 			 * If we still have some data to send, then
283 			 * clear the FIN bit.  Usually this would
284 			 * happen below when it realizes that we
285 			 * aren't sending all the data.  However,
286 			 * if we have exactly 1 byte of unsent data,
287 			 * then it won't clear the FIN bit below,
288 			 * and if we are in persist state, we wind
289 			 * up sending the packet without recording
290 			 * that we sent the FIN bit.
291 			 *
292 			 * We can't just blindly clear the FIN bit,
293 			 * because if we don't have any more data
294 			 * to send then the probe will be the FIN
295 			 * itself.
296 			 */
297 			if (off < so->so_snd.sb_cc)
298 				flags &= ~TH_FIN;
299 			sendwin = 1;
300 		} else {
301 			tcp_timer_activate(tp, TT_PERSIST, 0);
302 			tp->t_rxtshift = 0;
303 		}
304 	}
305 
306 	/*
307 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
308 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
309 	 * a negative length.  This can also occur when TCP opens up
310 	 * its congestion window while receiving additional duplicate
311 	 * acks after fast-retransmit because TCP will reset snd_nxt
312 	 * to snd_max after the fast-retransmit.
313 	 *
314 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
315 	 * be set to snd_una, the offset will be 0, and the length may
316 	 * wind up 0.
317 	 *
318 	 * If sack_rxmit is true we are retransmitting from the scoreboard
319 	 * in which case len is already set.
320 	 */
321 	if (sack_rxmit == 0) {
322 		if (sack_bytes_rxmt == 0)
323 			len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off);
324 		else {
325 			long cwin;
326 
327                         /*
328 			 * We are inside of a SACK recovery episode and are
329 			 * sending new data, having retransmitted all the
330 			 * data possible in the scoreboard.
331 			 */
332 			len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd)
333 			       - off);
334 			/*
335 			 * Don't remove this (len > 0) check !
336 			 * We explicitly check for len > 0 here (although it
337 			 * isn't really necessary), to work around a gcc
338 			 * optimization issue - to force gcc to compute
339 			 * len above. Without this check, the computation
340 			 * of len is bungled by the optimizer.
341 			 */
342 			if (len > 0) {
343 				cwin = tp->snd_cwnd -
344 					(tp->snd_nxt - tp->sack_newdata) -
345 					sack_bytes_rxmt;
346 				if (cwin < 0)
347 					cwin = 0;
348 				len = lmin(len, cwin);
349 			}
350 		}
351 	}
352 
353 	/*
354 	 * Lop off SYN bit if it has already been sent.  However, if this
355 	 * is SYN-SENT state and if segment contains data and if we don't
356 	 * know that foreign host supports TAO, suppress sending segment.
357 	 */
358 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
359 		if (tp->t_state != TCPS_SYN_RECEIVED)
360 			flags &= ~TH_SYN;
361 		off--, len++;
362 	}
363 
364 	/*
365 	 * Be careful not to send data and/or FIN on SYN segments.
366 	 * This measure is needed to prevent interoperability problems
367 	 * with not fully conformant TCP implementations.
368 	 */
369 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
370 		len = 0;
371 		flags &= ~TH_FIN;
372 	}
373 
374 	if (len < 0) {
375 		/*
376 		 * If FIN has been sent but not acked,
377 		 * but we haven't been called to retransmit,
378 		 * len will be < 0.  Otherwise, window shrank
379 		 * after we sent into it.  If window shrank to 0,
380 		 * cancel pending retransmit, pull snd_nxt back
381 		 * to (closed) window, and set the persist timer
382 		 * if it isn't already going.  If the window didn't
383 		 * close completely, just wait for an ACK.
384 		 */
385 		len = 0;
386 		if (sendwin == 0) {
387 			tcp_timer_activate(tp, TT_REXMT, 0);
388 			tp->t_rxtshift = 0;
389 			tp->snd_nxt = tp->snd_una;
390 			if (!tcp_timer_active(tp, TT_PERSIST))
391 				tcp_setpersist(tp);
392 		}
393 	}
394 
395 	/* len will be >= 0 after this point. */
396 	KASSERT(len >= 0, ("%s: len < 0", __func__));
397 
398 	/*
399 	 * Automatic sizing of send socket buffer.  Often the send buffer
400 	 * size is not optimally adjusted to the actual network conditions
401 	 * at hand (delay bandwidth product).  Setting the buffer size too
402 	 * small limits throughput on links with high bandwidth and high
403 	 * delay (eg. trans-continental/oceanic links).  Setting the
404 	 * buffer size too big consumes too much real kernel memory,
405 	 * especially with many connections on busy servers.
406 	 *
407 	 * The criteria to step up the send buffer one notch are:
408 	 *  1. receive window of remote host is larger than send buffer
409 	 *     (with a fudge factor of 5/4th);
410 	 *  2. send buffer is filled to 7/8th with data (so we actually
411 	 *     have data to make use of it);
412 	 *  3. send buffer fill has not hit maximal automatic size;
413 	 *  4. our send window (slow start and cogestion controlled) is
414 	 *     larger than sent but unacknowledged data in send buffer.
415 	 *
416 	 * The remote host receive window scaling factor may limit the
417 	 * growing of the send buffer before it reaches its allowed
418 	 * maximum.
419 	 *
420 	 * It scales directly with slow start or congestion window
421 	 * and does at most one step per received ACK.  This fast
422 	 * scaling has the drawback of growing the send buffer beyond
423 	 * what is strictly necessary to make full use of a given
424 	 * delay*bandwith product.  However testing has shown this not
425 	 * to be much of an problem.  At worst we are trading wasting
426 	 * of available bandwith (the non-use of it) for wasting some
427 	 * socket buffer memory.
428 	 *
429 	 * TODO: Shrink send buffer during idle periods together
430 	 * with congestion window.  Requires another timer.  Has to
431 	 * wait for upcoming tcp timer rewrite.
432 	 */
433 	if (tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
434 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
435 		    so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
436 		    so->so_snd.sb_cc < tcp_autosndbuf_max &&
437 		    sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
438 			if (!sbreserve_locked(&so->so_snd,
439 			    min(so->so_snd.sb_hiwat + tcp_autosndbuf_inc,
440 			     tcp_autosndbuf_max), so, curthread))
441 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
442 		}
443 	}
444 
445 	/*
446 	 * Truncate to the maximum segment length or enable TCP Segmentation
447 	 * Offloading (if supported by hardware) and ensure that FIN is removed
448 	 * if the length no longer contains the last data byte.
449 	 *
450 	 * TSO may only be used if we are in a pure bulk sending state.  The
451 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
452 	 * IP options prevent using TSO.  With TSO the TCP header is the same
453 	 * (except for the sequence number) for all generated packets.  This
454 	 * makes it impossible to transmit any options which vary per generated
455 	 * segment or packet.
456 	 *
457 	 * The length of TSO bursts is limited to TCP_MAXWIN.  That limit and
458 	 * removal of FIN (if not already catched here) are handled later after
459 	 * the exact length of the TCP options are known.
460 	 */
461 	if (len > tp->t_maxseg) {
462 		if ((tp->t_flags & TF_TSO) && tcp_do_tso &&
463 		    ((tp->t_flags & TF_SIGNATURE) == 0) &&
464 		    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
465 		    tp->t_inpcb->inp_options == NULL &&
466 		    tp->t_inpcb->in6p_options == NULL &&
467 		    tp->t_inpcb->inp_sp == NULL) {
468 			tso = 1;
469 		} else {
470 			len = tp->t_maxseg;
471 			sendalot = 1;
472 			tso = 0;
473 		}
474 	}
475 	if (sack_rxmit) {
476 		if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
477 			flags &= ~TH_FIN;
478 	} else {
479 		if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
480 			flags &= ~TH_FIN;
481 	}
482 
483 	recwin = sbspace(&so->so_rcv);
484 
485 	/*
486 	 * Sender silly window avoidance.   We transmit under the following
487 	 * conditions when len is non-zero:
488 	 *
489 	 *	- We have a full segment (or more with TSO)
490 	 *	- This is the last buffer in a write()/send() and we are
491 	 *	  either idle or running NODELAY
492 	 *	- we've timed out (e.g. persist timer)
493 	 *	- we have more then 1/2 the maximum send window's worth of
494 	 *	  data (receiver may be limited the window size)
495 	 *	- we need to retransmit
496 	 */
497 	if (len) {
498 		if (len >= tp->t_maxseg)
499 			goto send;
500 		/*
501 		 * NOTE! on localhost connections an 'ack' from the remote
502 		 * end may occur synchronously with the output and cause
503 		 * us to flush a buffer queued with moretocome.  XXX
504 		 *
505 		 * note: the len + off check is almost certainly unnecessary.
506 		 */
507 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
508 		    (idle || (tp->t_flags & TF_NODELAY)) &&
509 		    len + off >= so->so_snd.sb_cc &&
510 		    (tp->t_flags & TF_NOPUSH) == 0) {
511 			goto send;
512 		}
513 		if (tp->t_flags & TF_FORCEDATA)		/* typ. timeout case */
514 			goto send;
515 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
516 			goto send;
517 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
518 			goto send;
519 		if (sack_rxmit)
520 			goto send;
521 	}
522 
523 	/*
524 	 * Compare available window to amount of window
525 	 * known to peer (as advertised window less
526 	 * next expected input).  If the difference is at least two
527 	 * max size segments, or at least 50% of the maximum possible
528 	 * window, then want to send a window update to peer.
529 	 * Skip this if the connection is in T/TCP half-open state.
530 	 */
531 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN)) {
532 		/*
533 		 * "adv" is the amount we can increase the window,
534 		 * taking into account that we are limited by
535 		 * TCP_MAXWIN << tp->rcv_scale.
536 		 */
537 		long adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale) -
538 			(tp->rcv_adv - tp->rcv_nxt);
539 
540 		if (adv >= (long) (2 * tp->t_maxseg))
541 			goto send;
542 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
543 			goto send;
544 	}
545 
546 	/*
547 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
548 	 * is also a catch-all for the retransmit timer timeout case.
549 	 */
550 	if (tp->t_flags & TF_ACKNOW)
551 		goto send;
552 	if ((flags & TH_RST) ||
553 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
554 		goto send;
555 	if (SEQ_GT(tp->snd_up, tp->snd_una))
556 		goto send;
557 	/*
558 	 * If our state indicates that FIN should be sent
559 	 * and we have not yet done so, then we need to send.
560 	 */
561 	if (flags & TH_FIN &&
562 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
563 		goto send;
564 	/*
565 	 * In SACK, it is possible for tcp_output to fail to send a segment
566 	 * after the retransmission timer has been turned off.  Make sure
567 	 * that the retransmission timer is set.
568 	 */
569 	if (tp->sack_enable && SEQ_GT(tp->snd_max, tp->snd_una) &&
570 	    !tcp_timer_active(tp, TT_REXMT) &&
571 	    !tcp_timer_active(tp, TT_PERSIST)) {
572 		tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
573 		goto just_return;
574 	}
575 	/*
576 	 * TCP window updates are not reliable, rather a polling protocol
577 	 * using ``persist'' packets is used to insure receipt of window
578 	 * updates.  The three ``states'' for the output side are:
579 	 *	idle			not doing retransmits or persists
580 	 *	persisting		to move a small or zero window
581 	 *	(re)transmitting	and thereby not persisting
582 	 *
583 	 * tcp_timer_active(tp, TT_PERSIST)
584 	 *	is true when we are in persist state.
585 	 * (tp->t_flags & TF_FORCEDATA)
586 	 *	is set when we are called to send a persist packet.
587 	 * tcp_timer_active(tp, TT_REXMT)
588 	 *	is set when we are retransmitting
589 	 * The output side is idle when both timers are zero.
590 	 *
591 	 * If send window is too small, there is data to transmit, and no
592 	 * retransmit or persist is pending, then go to persist state.
593 	 * If nothing happens soon, send when timer expires:
594 	 * if window is nonzero, transmit what we can,
595 	 * otherwise force out a byte.
596 	 */
597 	if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) &&
598 	    !tcp_timer_active(tp, TT_PERSIST)) {
599 		tp->t_rxtshift = 0;
600 		tcp_setpersist(tp);
601 	}
602 
603 	/*
604 	 * No reason to send a segment, just return.
605 	 */
606 just_return:
607 	SOCKBUF_UNLOCK(&so->so_snd);
608 	return (0);
609 
610 send:
611 	SOCKBUF_LOCK_ASSERT(&so->so_snd);
612 	/*
613 	 * Before ESTABLISHED, force sending of initial options
614 	 * unless TCP set not to do any options.
615 	 * NOTE: we assume that the IP/TCP header plus TCP options
616 	 * always fit in a single mbuf, leaving room for a maximum
617 	 * link header, i.e.
618 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
619 	 */
620 	optlen = 0;
621 #ifdef INET6
622 	if (isipv6)
623 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
624 	else
625 #endif
626 	hdrlen = sizeof (struct tcpiphdr);
627 
628 	/*
629 	 * Compute options for segment.
630 	 * We only have to care about SYN and established connection
631 	 * segments.  Options for SYN-ACK segments are handled in TCP
632 	 * syncache.
633 	 */
634 	if ((tp->t_flags & TF_NOOPT) == 0) {
635 		to.to_flags = 0;
636 		/* Maximum segment size. */
637 		if (flags & TH_SYN) {
638 			tp->snd_nxt = tp->iss;
639 			to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
640 			to.to_flags |= TOF_MSS;
641 		}
642 		/* Window scaling. */
643 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
644 			to.to_wscale = tp->request_r_scale;
645 			to.to_flags |= TOF_SCALE;
646 		}
647 		/* Timestamps. */
648 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
649 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
650 			to.to_tsval = ticks + tp->ts_offset;
651 			to.to_tsecr = tp->ts_recent;
652 			to.to_flags |= TOF_TS;
653 			/* Set receive buffer autosizing timestamp. */
654 			if (tp->rfbuf_ts == 0 &&
655 			    (so->so_rcv.sb_flags & SB_AUTOSIZE))
656 				tp->rfbuf_ts = ticks;
657 		}
658 		/* Selective ACK's. */
659 		if (tp->sack_enable) {
660 			if (flags & TH_SYN)
661 				to.to_flags |= TOF_SACKPERM;
662 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
663 			    (tp->t_flags & TF_SACK_PERMIT) &&
664 			    tp->rcv_numsacks > 0) {
665 				to.to_flags |= TOF_SACK;
666 				to.to_nsacks = tp->rcv_numsacks;
667 				to.to_sacks = (u_char *)tp->sackblks;
668 			}
669 		}
670 #ifdef TCP_SIGNATURE
671 		/* TCP-MD5 (RFC2385). */
672 #ifdef INET6
673 		if (!isipv6 && (tp->t_flags & TF_SIGNATURE))
674 #else
675 		if (tp->t_flags & TF_SIGNATURE)
676 #endif /* INET6 */
677 			to.to_flags |= TOF_SIGNATURE;
678 #endif /* TCP_SIGNATURE */
679 
680 		/* Processing the options. */
681 		hdrlen += optlen = tcp_addoptions(&to, (u_char *)&opt);
682 
683 #ifdef TCP_SIGNATURE
684 		sigoff = to.to_signature - (u_char *)&to;
685 #endif /* TCP_SIGNATURE */
686 	}
687 
688 #ifdef INET6
689 	if (isipv6)
690 		ipoptlen = ip6_optlen(tp->t_inpcb);
691 	else
692 #endif
693 	if (tp->t_inpcb->inp_options)
694 		ipoptlen = tp->t_inpcb->inp_options->m_len -
695 				offsetof(struct ipoption, ipopt_list);
696 	else
697 		ipoptlen = 0;
698 #ifdef IPSEC
699 	ipoptlen += ipsec_hdrsiz_tcp(tp);
700 #endif
701 
702 	/*
703 	 * Adjust data length if insertion of options will
704 	 * bump the packet length beyond the t_maxopd length.
705 	 * Clear the FIN bit because we cut off the tail of
706 	 * the segment.
707 	 *
708 	 * When doing TSO limit a burst to TCP_MAXWIN minus the
709 	 * IP, TCP and Options length to keep ip->ip_len from
710 	 * overflowing.  Prevent the last segment from being
711 	 * fractional thus making them all equal sized and set
712 	 * the flag to continue sending.  TSO is disabled when
713 	 * IP options or IPSEC are present.
714 	 */
715 	if (len + optlen + ipoptlen > tp->t_maxopd) {
716 		flags &= ~TH_FIN;
717 		if (tso) {
718 			if (len > TCP_MAXWIN - hdrlen - optlen) {
719 				len = TCP_MAXWIN - hdrlen - optlen;
720 				len = len - (len % (tp->t_maxopd - optlen));
721 				sendalot = 1;
722 			} else if (tp->t_flags & TF_NEEDFIN)
723 				sendalot = 1;
724 		} else {
725 			len = tp->t_maxopd - optlen - ipoptlen;
726 			sendalot = 1;
727 		}
728 	}
729 
730 /*#ifdef DIAGNOSTIC*/
731 #ifdef INET6
732 	if (max_linkhdr + hdrlen > MCLBYTES)
733 #else
734 	if (max_linkhdr + hdrlen > MHLEN)
735 #endif
736 		panic("tcphdr too big");
737 /*#endif*/
738 
739 	/*
740 	 * Grab a header mbuf, attaching a copy of data to
741 	 * be transmitted, and initialize the header from
742 	 * the template for sends on this connection.
743 	 */
744 	if (len) {
745 		struct mbuf *mb;
746 		u_int moff;
747 
748 		if ((tp->t_flags & TF_FORCEDATA) && len == 1)
749 			tcpstat.tcps_sndprobe++;
750 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
751 			tcpstat.tcps_sndrexmitpack++;
752 			tcpstat.tcps_sndrexmitbyte += len;
753 		} else {
754 			tcpstat.tcps_sndpack++;
755 			tcpstat.tcps_sndbyte += len;
756 		}
757 #ifdef notyet
758 		if ((m = m_copypack(so->so_snd.sb_mb, off,
759 		    (int)len, max_linkhdr + hdrlen)) == 0) {
760 			SOCKBUF_UNLOCK(&so->so_snd);
761 			error = ENOBUFS;
762 			goto out;
763 		}
764 		/*
765 		 * m_copypack left space for our hdr; use it.
766 		 */
767 		m->m_len += hdrlen;
768 		m->m_data -= hdrlen;
769 #else
770 		MGETHDR(m, M_DONTWAIT, MT_DATA);
771 		if (m == NULL) {
772 			SOCKBUF_UNLOCK(&so->so_snd);
773 			error = ENOBUFS;
774 			goto out;
775 		}
776 #ifdef INET6
777 		if (MHLEN < hdrlen + max_linkhdr) {
778 			MCLGET(m, M_DONTWAIT);
779 			if ((m->m_flags & M_EXT) == 0) {
780 				SOCKBUF_UNLOCK(&so->so_snd);
781 				m_freem(m);
782 				error = ENOBUFS;
783 				goto out;
784 			}
785 		}
786 #endif
787 		m->m_data += max_linkhdr;
788 		m->m_len = hdrlen;
789 
790 		/*
791 		 * Start the m_copy functions from the closest mbuf
792 		 * to the offset in the socket buffer chain.
793 		 */
794 		mb = sbsndptr(&so->so_snd, off, len, &moff);
795 
796 		if (len <= MHLEN - hdrlen - max_linkhdr) {
797 			m_copydata(mb, moff, (int)len,
798 			    mtod(m, caddr_t) + hdrlen);
799 			m->m_len += len;
800 		} else {
801 			m->m_next = m_copy(mb, moff, (int)len);
802 			if (m->m_next == NULL) {
803 				SOCKBUF_UNLOCK(&so->so_snd);
804 				(void) m_free(m);
805 				error = ENOBUFS;
806 				goto out;
807 			}
808 		}
809 #endif
810 		/*
811 		 * If we're sending everything we've got, set PUSH.
812 		 * (This will keep happy those implementations which only
813 		 * give data to the user when a buffer fills or
814 		 * a PUSH comes in.)
815 		 */
816 		if (off + len == so->so_snd.sb_cc)
817 			flags |= TH_PUSH;
818 		SOCKBUF_UNLOCK(&so->so_snd);
819 	} else {
820 		SOCKBUF_UNLOCK(&so->so_snd);
821 		if (tp->t_flags & TF_ACKNOW)
822 			tcpstat.tcps_sndacks++;
823 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
824 			tcpstat.tcps_sndctrl++;
825 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
826 			tcpstat.tcps_sndurg++;
827 		else
828 			tcpstat.tcps_sndwinup++;
829 
830 		MGETHDR(m, M_DONTWAIT, MT_DATA);
831 		if (m == NULL) {
832 			error = ENOBUFS;
833 			goto out;
834 		}
835 #ifdef INET6
836 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
837 		    MHLEN >= hdrlen) {
838 			MH_ALIGN(m, hdrlen);
839 		} else
840 #endif
841 		m->m_data += max_linkhdr;
842 		m->m_len = hdrlen;
843 	}
844 	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
845 	m->m_pkthdr.rcvif = (struct ifnet *)0;
846 #ifdef MAC
847 	mac_create_mbuf_from_inpcb(tp->t_inpcb, m);
848 #endif
849 #ifdef INET6
850 	if (isipv6) {
851 		ip6 = mtod(m, struct ip6_hdr *);
852 		th = (struct tcphdr *)(ip6 + 1);
853 		tcpip_fillheaders(tp->t_inpcb, ip6, th);
854 	} else
855 #endif /* INET6 */
856 	{
857 		ip = mtod(m, struct ip *);
858 		ipov = (struct ipovly *)ip;
859 		th = (struct tcphdr *)(ip + 1);
860 		tcpip_fillheaders(tp->t_inpcb, ip, th);
861 	}
862 
863 	/*
864 	 * Fill in fields, remembering maximum advertised
865 	 * window for use in delaying messages about window sizes.
866 	 * If resending a FIN, be sure not to use a new sequence number.
867 	 */
868 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
869 	    tp->snd_nxt == tp->snd_max)
870 		tp->snd_nxt--;
871 	/*
872 	 * If we are doing retransmissions, then snd_nxt will
873 	 * not reflect the first unsent octet.  For ACK only
874 	 * packets, we do not want the sequence number of the
875 	 * retransmitted packet, we want the sequence number
876 	 * of the next unsent octet.  So, if there is no data
877 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
878 	 * when filling in ti_seq.  But if we are in persist
879 	 * state, snd_max might reflect one byte beyond the
880 	 * right edge of the window, so use snd_nxt in that
881 	 * case, since we know we aren't doing a retransmission.
882 	 * (retransmit and persist are mutually exclusive...)
883 	 */
884 	if (sack_rxmit == 0) {
885 		if (len || (flags & (TH_SYN|TH_FIN)) ||
886 		    tcp_timer_active(tp, TT_PERSIST))
887 			th->th_seq = htonl(tp->snd_nxt);
888 		else
889 			th->th_seq = htonl(tp->snd_max);
890 	} else {
891 		th->th_seq = htonl(p->rxmit);
892 		p->rxmit += len;
893 		tp->sackhint.sack_bytes_rexmit += len;
894 	}
895 	th->th_ack = htonl(tp->rcv_nxt);
896 	if (optlen) {
897 		bcopy(opt, th + 1, optlen);
898 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
899 	}
900 	th->th_flags = flags;
901 	/*
902 	 * Calculate receive window.  Don't shrink window,
903 	 * but avoid silly window syndrome.
904 	 *
905 	 * XXX: RFC1323:  The Window field in a SYN (i.e., a <SYN> or
906 	 * <SYN,ACK>) segment itself is never scaled.
907 	 */
908 	if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
909 	    recwin < (long)tp->t_maxseg)
910 		recwin = 0;
911 	if (recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
912 		recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
913 	if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
914 		recwin = (long)TCP_MAXWIN << tp->rcv_scale;
915 	th->th_win = htons((u_short) (recwin >> tp->rcv_scale));
916 
917 
918 	/*
919 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
920 	 * a 0 window.  This may cause the remote transmitter to stall.  This
921 	 * flag tells soreceive() to disable delayed acknowledgements when
922 	 * draining the buffer.  This can occur if the receiver is attempting
923 	 * to read more data then can be buffered prior to transmitting on
924 	 * the connection.
925 	 */
926 	if (recwin == 0)
927 		tp->t_flags |= TF_RXWIN0SENT;
928 	else
929 		tp->t_flags &= ~TF_RXWIN0SENT;
930 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
931 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
932 		th->th_flags |= TH_URG;
933 	} else
934 		/*
935 		 * If no urgent pointer to send, then we pull
936 		 * the urgent pointer to the left edge of the send window
937 		 * so that it doesn't drift into the send window on sequence
938 		 * number wraparound.
939 		 */
940 		tp->snd_up = tp->snd_una;		/* drag it along */
941 
942 #ifdef TCP_SIGNATURE
943 #ifdef INET6
944 	if (!isipv6)
945 #endif
946 	if (tp->t_flags & TF_SIGNATURE)
947 		tcp_signature_compute(m, sizeof(struct ip), len, optlen,
948 		    (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
949 #endif
950 
951 	/*
952 	 * Put TCP length in extended header, and then
953 	 * checksum extended header and data.
954 	 */
955 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
956 #ifdef INET6
957 	if (isipv6)
958 		/*
959 		 * ip6_plen is not need to be filled now, and will be filled
960 		 * in ip6_output.
961 		 */
962 		th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
963 				       sizeof(struct tcphdr) + optlen + len);
964 	else
965 #endif /* INET6 */
966 	{
967 		m->m_pkthdr.csum_flags = CSUM_TCP;
968 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
969 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
970 		    htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
971 
972 		/* IP version must be set here for ipv4/ipv6 checking later */
973 		KASSERT(ip->ip_v == IPVERSION,
974 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
975 	}
976 
977 	/*
978 	 * Enable TSO and specify the size of the segments.
979 	 * The TCP pseudo header checksum is always provided.
980 	 * XXX: Fixme: This is currently not the case for IPv6.
981 	 */
982 	if (tso) {
983 		m->m_pkthdr.csum_flags = CSUM_TSO;
984 		m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen;
985 	}
986 
987 	/*
988 	 * In transmit state, time the transmission and arrange for
989 	 * the retransmit.  In persist state, just set snd_max.
990 	 */
991 	if ((tp->t_flags & TF_FORCEDATA) == 0 ||
992 	    !tcp_timer_active(tp, TT_PERSIST)) {
993 		tcp_seq startseq = tp->snd_nxt;
994 
995 		/*
996 		 * Advance snd_nxt over sequence space of this segment.
997 		 */
998 		if (flags & (TH_SYN|TH_FIN)) {
999 			if (flags & TH_SYN)
1000 				tp->snd_nxt++;
1001 			if (flags & TH_FIN) {
1002 				tp->snd_nxt++;
1003 				tp->t_flags |= TF_SENTFIN;
1004 			}
1005 		}
1006 		if (sack_rxmit)
1007 			goto timer;
1008 		tp->snd_nxt += len;
1009 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1010 			tp->snd_max = tp->snd_nxt;
1011 			/*
1012 			 * Time this transmission if not a retransmission and
1013 			 * not currently timing anything.
1014 			 */
1015 			if (tp->t_rtttime == 0) {
1016 				tp->t_rtttime = ticks;
1017 				tp->t_rtseq = startseq;
1018 				tcpstat.tcps_segstimed++;
1019 			}
1020 		}
1021 
1022 		/*
1023 		 * Set retransmit timer if not currently set,
1024 		 * and not doing a pure ack or a keep-alive probe.
1025 		 * Initial value for retransmit timer is smoothed
1026 		 * round-trip time + 2 * round-trip time variance.
1027 		 * Initialize shift counter which is used for backoff
1028 		 * of retransmit time.
1029 		 */
1030 timer:
1031 		if (!tcp_timer_active(tp, TT_PERSIST) &&
1032 		    ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1033 		     (tp->snd_nxt != tp->snd_una))) {
1034 			if (tcp_timer_active(tp, TT_PERSIST)) {
1035 				tcp_timer_activate(tp, TT_PERSIST, 0);
1036 				tp->t_rxtshift = 0;
1037 			}
1038 			tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1039 		}
1040 	} else {
1041 		/*
1042 		 * Persist case, update snd_max but since we are in
1043 		 * persist mode (no window) we do not update snd_nxt.
1044 		 */
1045 		int xlen = len;
1046 		if (flags & TH_SYN)
1047 			++xlen;
1048 		if (flags & TH_FIN) {
1049 			++xlen;
1050 			tp->t_flags |= TF_SENTFIN;
1051 		}
1052 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1053 			tp->snd_max = tp->snd_nxt + len;
1054 	}
1055 
1056 #ifdef TCPDEBUG
1057 	/*
1058 	 * Trace.
1059 	 */
1060 	if (so->so_options & SO_DEBUG) {
1061 		u_short save = 0;
1062 #ifdef INET6
1063 		if (!isipv6)
1064 #endif
1065 		{
1066 			save = ipov->ih_len;
1067 			ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1068 		}
1069 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1070 #ifdef INET6
1071 		if (!isipv6)
1072 #endif
1073 		ipov->ih_len = save;
1074 	}
1075 #endif
1076 
1077 	/*
1078 	 * Fill in IP length and desired time to live and
1079 	 * send to IP level.  There should be a better way
1080 	 * to handle ttl and tos; we could keep them in
1081 	 * the template, but need a way to checksum without them.
1082 	 */
1083 	/*
1084 	 * m->m_pkthdr.len should have been set before cksum calcuration,
1085 	 * because in6_cksum() need it.
1086 	 */
1087 #ifdef INET6
1088 	if (isipv6) {
1089 		/*
1090 		 * we separately set hoplimit for every segment, since the
1091 		 * user might want to change the value via setsockopt.
1092 		 * Also, desired default hop limit might be changed via
1093 		 * Neighbor Discovery.
1094 		 */
1095 		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1096 
1097 		/* TODO: IPv6 IP6TOS_ECT bit on */
1098 		error = ip6_output(m,
1099 			    tp->t_inpcb->in6p_outputopts, NULL,
1100 			    ((so->so_options & SO_DONTROUTE) ?
1101 			    IP_ROUTETOIF : 0), NULL, NULL, tp->t_inpcb);
1102 	} else
1103 #endif /* INET6 */
1104     {
1105 	ip->ip_len = m->m_pkthdr.len;
1106 #ifdef INET6
1107 	if (INP_CHECK_SOCKAF(so, AF_INET6))
1108 		ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1109 #endif /* INET6 */
1110 	/*
1111 	 * If we do path MTU discovery, then we set DF on every packet.
1112 	 * This might not be the best thing to do according to RFC3390
1113 	 * Section 2. However the tcp hostcache migitates the problem
1114 	 * so it affects only the first tcp connection with a host.
1115 	 */
1116 	if (path_mtu_discovery)
1117 		ip->ip_off |= IP_DF;
1118 
1119 	error = ip_output(m, tp->t_inpcb->inp_options, NULL,
1120 	    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1121 	    tp->t_inpcb);
1122     }
1123 	if (error) {
1124 
1125 		/*
1126 		 * We know that the packet was lost, so back out the
1127 		 * sequence number advance, if any.
1128 		 *
1129 		 * If the error is EPERM the packet got blocked by the
1130 		 * local firewall.  Normally we should terminate the
1131 		 * connection but the blocking may have been spurious
1132 		 * due to a firewall reconfiguration cycle.  So we treat
1133 		 * it like a packet loss and let the retransmit timer and
1134 		 * timeouts do their work over time.
1135 		 * XXX: It is a POLA question whether calling tcp_drop right
1136 		 * away would be the really correct behavior instead.
1137 		 */
1138 		if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1139 		    !tcp_timer_active(tp, TT_PERSIST)) &&
1140 		    ((flags & TH_SYN) == 0) &&
1141 		    (error != EPERM)) {
1142 			if (sack_rxmit) {
1143 				p->rxmit -= len;
1144 				tp->sackhint.sack_bytes_rexmit -= len;
1145 				KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1146 				    ("sackhint bytes rtx >= 0"));
1147 			} else
1148 				tp->snd_nxt -= len;
1149 		}
1150 out:
1151 		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1152 		switch (error) {
1153 		case EPERM:
1154 			tp->t_softerror = error;
1155 			return (error);
1156 		case ENOBUFS:
1157 	                if (!tcp_timer_active(tp, TT_REXMT) &&
1158 			    !tcp_timer_active(tp, TT_PERSIST))
1159 	                        tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1160 			tp->snd_cwnd = tp->t_maxseg;
1161 			return (0);
1162 		case EMSGSIZE:
1163 			/*
1164 			 * For some reason the interface we used initially
1165 			 * to send segments changed to another or lowered
1166 			 * its MTU.
1167 			 *
1168 			 * tcp_mtudisc() will find out the new MTU and as
1169 			 * its last action, initiate retransmission, so it
1170 			 * is important to not do so here.
1171 			 *
1172 			 * If TSO was active we either got an interface
1173 			 * without TSO capabilits or TSO was turned off.
1174 			 * Disable it for this connection as too and
1175 			 * immediatly retry with MSS sized segments generated
1176 			 * by this function.
1177 			 */
1178 			if (tso)
1179 				tp->t_flags &= ~TF_TSO;
1180 			tcp_mtudisc(tp->t_inpcb, 0);
1181 			return (0);
1182 		case EHOSTDOWN:
1183 		case EHOSTUNREACH:
1184 		case ENETDOWN:
1185 		case ENETUNREACH:
1186 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
1187 				tp->t_softerror = error;
1188 				return (0);
1189 			}
1190 			/* FALLTHROUGH */
1191 		default:
1192 			return (error);
1193 		}
1194 	}
1195 	tcpstat.tcps_sndtotal++;
1196 
1197 	/*
1198 	 * Data sent (as far as we can tell).
1199 	 * If this advertises a larger window than any other segment,
1200 	 * then remember the size of the advertised window.
1201 	 * Any pending ACK has now been sent.
1202 	 */
1203 	if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1204 		tp->rcv_adv = tp->rcv_nxt + recwin;
1205 	tp->last_ack_sent = tp->rcv_nxt;
1206 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1207 	if (tcp_timer_active(tp, TT_DELACK))
1208 		tcp_timer_activate(tp, TT_DELACK, 0);
1209 #if 0
1210 	/*
1211 	 * This completely breaks TCP if newreno is turned on.  What happens
1212 	 * is that if delayed-acks are turned on on the receiver, this code
1213 	 * on the transmitter effectively destroys the TCP window, forcing
1214 	 * it to four packets (1.5Kx4 = 6K window).
1215 	 */
1216 	if (sendalot && (!tcp_do_newreno || --maxburst))
1217 		goto again;
1218 #endif
1219 	if (sendalot)
1220 		goto again;
1221 	return (0);
1222 }
1223 
1224 void
1225 tcp_setpersist(struct tcpcb *tp)
1226 {
1227 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1228 	int tt;
1229 
1230 	if (tcp_timer_active(tp, TT_REXMT))
1231 		panic("tcp_setpersist: retransmit pending");
1232 	/*
1233 	 * Start/restart persistance timer.
1234 	 */
1235 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1236 		      TCPTV_PERSMIN, TCPTV_PERSMAX);
1237 	tcp_timer_activate(tp, TT_PERSIST, tt);
1238 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1239 		tp->t_rxtshift++;
1240 }
1241 
1242 /*
1243  * Insert TCP options according to the supplied parameters to the place
1244  * optp in a consistent way.  Can handle unaligned destinations.
1245  *
1246  * The order of the option processing is crucial for optimal packing and
1247  * alignment for the scarce option space.
1248  *
1249  * The optimal order for a SYN/SYN-ACK segment is:
1250  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1251  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1252  *
1253  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1254  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1255  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1256  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1257  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1258  */
1259 int
1260 tcp_addoptions(struct tcpopt *to, u_char *optp)
1261 {
1262 	u_int mask, optlen = 0;
1263 
1264 	for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1265 		if ((to->to_flags & mask) != mask)
1266 			continue;
1267 		switch (to->to_flags & mask) {
1268 		case TOF_MSS:
1269 			while (optlen % 4) {
1270 				optlen += TCPOLEN_NOP;
1271 				*optp++ = TCPOPT_NOP;
1272 			}
1273 			optlen += TCPOLEN_MAXSEG;
1274 			*optp++ = TCPOPT_MAXSEG;
1275 			*optp++ = TCPOLEN_MAXSEG;
1276 			to->to_mss = htons(to->to_mss);
1277 			bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1278 			optp += sizeof(to->to_mss);
1279 			break;
1280 		case TOF_SCALE:
1281 			while (!optlen || optlen % 2 != 1) {
1282 				optlen += TCPOLEN_NOP;
1283 				*optp++ = TCPOPT_NOP;
1284 			}
1285 			optlen += TCPOLEN_WINDOW;
1286 			*optp++ = TCPOPT_WINDOW;
1287 			*optp++ = TCPOLEN_WINDOW;
1288 			*optp++ = to->to_wscale;
1289 			break;
1290 		case TOF_SACKPERM:
1291 			while (optlen % 2) {
1292 				optlen += TCPOLEN_NOP;
1293 				*optp++ = TCPOPT_NOP;
1294 			}
1295 			optlen += TCPOLEN_SACK_PERMITTED;
1296 			*optp++ = TCPOPT_SACK_PERMITTED;
1297 			*optp++ = TCPOLEN_SACK_PERMITTED;
1298 			break;
1299 		case TOF_TS:
1300 			while (!optlen || optlen % 4 != 2) {
1301 				optlen += TCPOLEN_NOP;
1302 				*optp++ = TCPOPT_NOP;
1303 			}
1304 			optlen += TCPOLEN_TIMESTAMP;
1305 			*optp++ = TCPOPT_TIMESTAMP;
1306 			*optp++ = TCPOLEN_TIMESTAMP;
1307 			to->to_tsval = htonl(to->to_tsval);
1308 			to->to_tsecr = htonl(to->to_tsecr);
1309 			bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1310 			optp += sizeof(to->to_tsval);
1311 			bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1312 			optp += sizeof(to->to_tsecr);
1313 			break;
1314 		case TOF_SIGNATURE:
1315 			{
1316 			int siglen = TCPOLEN_SIGNATURE - 2;
1317 
1318 			while (!optlen || optlen % 4 != 2) {
1319 				optlen += TCPOLEN_NOP;
1320 				*optp++ = TCPOPT_NOP;
1321 			}
1322 			if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE)
1323 				continue;
1324 			optlen += TCPOLEN_SIGNATURE;
1325 			*optp++ = TCPOPT_SIGNATURE;
1326 			*optp++ = TCPOLEN_SIGNATURE;
1327 			to->to_signature = optp;
1328 			while (siglen--)
1329 				 *optp++ = 0;
1330 			break;
1331 			}
1332 		case TOF_SACK:
1333 			{
1334 			int sackblks = 0;
1335 			struct sackblk *sack = (struct sackblk *)to->to_sacks;
1336 			tcp_seq sack_seq;
1337 
1338 			while (!optlen || optlen % 4 != 2) {
1339 				optlen += TCPOLEN_NOP;
1340 				*optp++ = TCPOPT_NOP;
1341 			}
1342 			if (TCP_MAXOLEN - optlen < 2 + TCPOLEN_SACK)
1343 				continue;
1344 			optlen += TCPOLEN_SACKHDR;
1345 			*optp++ = TCPOPT_SACK;
1346 			sackblks = min(to->to_nsacks,
1347 					(TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1348 			*optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1349 			while (sackblks--) {
1350 				sack_seq = htonl(sack->start);
1351 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1352 				optp += sizeof(sack_seq);
1353 				sack_seq = htonl(sack->end);
1354 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1355 				optp += sizeof(sack_seq);
1356 				optlen += TCPOLEN_SACK;
1357 				sack++;
1358 			}
1359 			tcpstat.tcps_sack_send_blocks++;
1360 			break;
1361 			}
1362 		default:
1363 			panic("%s: unknown TCP option type", __func__);
1364 			break;
1365 		}
1366 	}
1367 
1368 	/* Terminate and pad TCP options to a 4 byte boundary. */
1369 	if (optlen % 4) {
1370 		optlen += TCPOLEN_EOL;
1371 		*optp++ = TCPOPT_EOL;
1372 	}
1373 	while (optlen % 4) {
1374 		optlen += TCPOLEN_NOP;
1375 		*optp++ = TCPOPT_NOP;
1376 	}
1377 
1378 	KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1379 	return (optlen);
1380 }
1381