xref: /freebsd/sys/netinet/tcp_output.c (revision b3aaa0cc21c63d388230c7ef2a80abd631ff20d5)
1 /*-
2  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  * 4. Neither the name of the University nor the names of its contributors
14  *    may be used to endorse or promote products derived from this software
15  *    without specific prior written permission.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  *
29  *	@(#)tcp_output.c	8.4 (Berkeley) 5/24/95
30  */
31 
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34 
35 #include "opt_inet.h"
36 #include "opt_inet6.h"
37 #include "opt_ipsec.h"
38 #include "opt_mac.h"
39 #include "opt_tcpdebug.h"
40 
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/domain.h>
44 #include <sys/kernel.h>
45 #include <sys/lock.h>
46 #include <sys/mbuf.h>
47 #include <sys/mutex.h>
48 #include <sys/protosw.h>
49 #include <sys/socket.h>
50 #include <sys/socketvar.h>
51 #include <sys/sysctl.h>
52 #include <sys/vimage.h>
53 
54 #include <net/if.h>
55 #include <net/route.h>
56 
57 #include <netinet/in.h>
58 #include <netinet/in_systm.h>
59 #include <netinet/ip.h>
60 #include <netinet/in_pcb.h>
61 #include <netinet/ip_var.h>
62 #include <netinet/ip_options.h>
63 #ifdef INET6
64 #include <netinet6/in6_pcb.h>
65 #include <netinet/ip6.h>
66 #include <netinet6/ip6_var.h>
67 #endif
68 #include <netinet/tcp.h>
69 #define	TCPOUTFLAGS
70 #include <netinet/tcp_fsm.h>
71 #include <netinet/tcp_seq.h>
72 #include <netinet/tcp_timer.h>
73 #include <netinet/tcp_var.h>
74 #include <netinet/tcpip.h>
75 #ifdef TCPDEBUG
76 #include <netinet/tcp_debug.h>
77 #endif
78 #include <netinet/vinet.h>
79 
80 #ifdef IPSEC
81 #include <netipsec/ipsec.h>
82 #endif /*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 #ifdef VIMAGE_GLOBALS
93 int path_mtu_discovery;
94 int ss_fltsz;
95 int ss_fltsz_local;
96 int tcp_do_newreno;
97 int tcp_do_tso;
98 int tcp_do_autosndbuf;
99 int tcp_autosndbuf_inc;
100 int tcp_autosndbuf_max;
101 #endif
102 
103 SYSCTL_V_INT(V_NET, vnet_inet, _net_inet_tcp, OID_AUTO, path_mtu_discovery,
104 	CTLFLAG_RW, path_mtu_discovery, 1, "Enable Path MTU Discovery");
105 
106 SYSCTL_V_INT(V_NET, vnet_inet, _net_inet_tcp, OID_AUTO,
107 	slowstart_flightsize, CTLFLAG_RW,
108 	ss_fltsz, 1, "Slow start flight size");
109 
110 SYSCTL_V_INT(V_NET, vnet_inet, _net_inet_tcp, OID_AUTO,
111 	local_slowstart_flightsize, CTLFLAG_RW,
112 	ss_fltsz_local, 1, "Slow start flight size for local networks");
113 
114 SYSCTL_V_INT(V_NET, vnet_inet, _net_inet_tcp, OID_AUTO, newreno, CTLFLAG_RW,
115 	tcp_do_newreno, 0, "Enable NewReno Algorithms");
116 
117 SYSCTL_V_INT(V_NET, vnet_inet, _net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW,
118 	tcp_do_tso, 0, "Enable TCP Segmentation Offload");
119 
120 SYSCTL_V_INT(V_NET, vnet_inet, _net_inet_tcp, OID_AUTO, sendbuf_auto,
121 	CTLFLAG_RW,
122 	tcp_do_autosndbuf, 0, "Enable automatic send buffer sizing");
123 
124 SYSCTL_V_INT(V_NET, vnet_inet, _net_inet_tcp, OID_AUTO, sendbuf_inc,
125 	CTLFLAG_RW, tcp_autosndbuf_inc, 0,
126 	"Incrementor step size of automatic send buffer");
127 
128 SYSCTL_V_INT(V_NET, vnet_inet, _net_inet_tcp, OID_AUTO, sendbuf_max,
129 	CTLFLAG_RW, tcp_autosndbuf_max, 0,
130 	"Max size of automatic send buffer");
131 
132 
133 /*
134  * Tcp output routine: figure out what should be sent and send it.
135  */
136 int
137 tcp_output(struct tcpcb *tp)
138 {
139 	INIT_VNET_INET(tp->t_inpcb->inp_vnet);
140 	struct socket *so = tp->t_inpcb->inp_socket;
141 	long len, recwin, sendwin;
142 	int off, flags, error;
143 	struct mbuf *m;
144 	struct ip *ip = NULL;
145 	struct ipovly *ipov = NULL;
146 	struct tcphdr *th;
147 	u_char opt[TCP_MAXOLEN];
148 	unsigned ipoptlen, optlen, hdrlen;
149 #ifdef IPSEC
150 	unsigned ipsec_optlen = 0;
151 #endif
152 	int idle, sendalot;
153 	int sack_rxmit, sack_bytes_rxmt;
154 	struct sackhole *p;
155 	int tso = 0;
156 	struct tcpopt to;
157 #if 0
158 	int maxburst = TCP_MAXBURST;
159 #endif
160 #ifdef INET6
161 	struct ip6_hdr *ip6 = NULL;
162 	int isipv6;
163 
164 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
165 #endif
166 
167 	INP_WLOCK_ASSERT(tp->t_inpcb);
168 
169 	/*
170 	 * Determine length of data that should be transmitted,
171 	 * and flags that will be used.
172 	 * If there is some data or critical controls (SYN, RST)
173 	 * to send, then transmit; otherwise, investigate further.
174 	 */
175 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
176 	if (idle && (ticks - tp->t_rcvtime) >= tp->t_rxtcur) {
177 		/*
178 		 * We have been idle for "a while" and no acks are
179 		 * expected to clock out any data we send --
180 		 * slow start to get ack "clock" running again.
181 		 *
182 		 * Set the slow-start flight size depending on whether
183 		 * this is a local network or not.
184 		 */
185 		int ss = V_ss_fltsz;
186 #ifdef INET6
187 		if (isipv6) {
188 			if (in6_localaddr(&tp->t_inpcb->in6p_faddr))
189 				ss = V_ss_fltsz_local;
190 		} else
191 #endif /* INET6 */
192 		if (in_localaddr(tp->t_inpcb->inp_faddr))
193 			ss = V_ss_fltsz_local;
194 		tp->snd_cwnd = tp->t_maxseg * ss;
195 	}
196 	tp->t_flags &= ~TF_LASTIDLE;
197 	if (idle) {
198 		if (tp->t_flags & TF_MORETOCOME) {
199 			tp->t_flags |= TF_LASTIDLE;
200 			idle = 0;
201 		}
202 	}
203 again:
204 	/*
205 	 * If we've recently taken a timeout, snd_max will be greater than
206 	 * snd_nxt.  There may be SACK information that allows us to avoid
207 	 * resending already delivered data.  Adjust snd_nxt accordingly.
208 	 */
209 	if ((tp->t_flags & TF_SACK_PERMIT) &&
210 	    SEQ_LT(tp->snd_nxt, tp->snd_max))
211 		tcp_sack_adjust(tp);
212 	sendalot = 0;
213 	off = tp->snd_nxt - tp->snd_una;
214 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
215 	sendwin = min(sendwin, tp->snd_bwnd);
216 
217 	flags = tcp_outflags[tp->t_state];
218 	/*
219 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
220 	 * to send out new data (when sendalot is 1), bypass this function.
221 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
222 	 * we're replacing a (future) new transmission with a retransmission
223 	 * now, and we previously incremented snd_cwnd in tcp_input().
224 	 */
225 	/*
226 	 * Still in sack recovery , reset rxmit flag to zero.
227 	 */
228 	sack_rxmit = 0;
229 	sack_bytes_rxmt = 0;
230 	len = 0;
231 	p = NULL;
232 	if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp) &&
233 	    (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
234 		long cwin;
235 
236 		cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
237 		if (cwin < 0)
238 			cwin = 0;
239 		/* Do not retransmit SACK segments beyond snd_recover */
240 		if (SEQ_GT(p->end, tp->snd_recover)) {
241 			/*
242 			 * (At least) part of sack hole extends beyond
243 			 * snd_recover. Check to see if we can rexmit data
244 			 * for this hole.
245 			 */
246 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
247 				/*
248 				 * Can't rexmit any more data for this hole.
249 				 * That data will be rexmitted in the next
250 				 * sack recovery episode, when snd_recover
251 				 * moves past p->rxmit.
252 				 */
253 				p = NULL;
254 				goto after_sack_rexmit;
255 			} else
256 				/* Can rexmit part of the current hole */
257 				len = ((long)ulmin(cwin,
258 						   tp->snd_recover - p->rxmit));
259 		} else
260 			len = ((long)ulmin(cwin, p->end - p->rxmit));
261 		off = p->rxmit - tp->snd_una;
262 		KASSERT(off >= 0,("%s: sack block to the left of una : %d",
263 		    __func__, off));
264 		if (len > 0) {
265 			sack_rxmit = 1;
266 			sendalot = 1;
267 			V_tcpstat.tcps_sack_rexmits++;
268 			V_tcpstat.tcps_sack_rexmit_bytes +=
269 			    min(len, tp->t_maxseg);
270 		}
271 	}
272 after_sack_rexmit:
273 	/*
274 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
275 	 * state flags.
276 	 */
277 	if (tp->t_flags & TF_NEEDFIN)
278 		flags |= TH_FIN;
279 	if (tp->t_flags & TF_NEEDSYN)
280 		flags |= TH_SYN;
281 
282 	SOCKBUF_LOCK(&so->so_snd);
283 	/*
284 	 * If in persist timeout with window of 0, send 1 byte.
285 	 * Otherwise, if window is small but nonzero
286 	 * and timer expired, we will send what we can
287 	 * and go to transmit state.
288 	 */
289 	if (tp->t_flags & TF_FORCEDATA) {
290 		if (sendwin == 0) {
291 			/*
292 			 * If we still have some data to send, then
293 			 * clear the FIN bit.  Usually this would
294 			 * happen below when it realizes that we
295 			 * aren't sending all the data.  However,
296 			 * if we have exactly 1 byte of unsent data,
297 			 * then it won't clear the FIN bit below,
298 			 * and if we are in persist state, we wind
299 			 * up sending the packet without recording
300 			 * that we sent the FIN bit.
301 			 *
302 			 * We can't just blindly clear the FIN bit,
303 			 * because if we don't have any more data
304 			 * to send then the probe will be the FIN
305 			 * itself.
306 			 */
307 			if (off < so->so_snd.sb_cc)
308 				flags &= ~TH_FIN;
309 			sendwin = 1;
310 		} else {
311 			tcp_timer_activate(tp, TT_PERSIST, 0);
312 			tp->t_rxtshift = 0;
313 		}
314 	}
315 
316 	/*
317 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
318 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
319 	 * a negative length.  This can also occur when TCP opens up
320 	 * its congestion window while receiving additional duplicate
321 	 * acks after fast-retransmit because TCP will reset snd_nxt
322 	 * to snd_max after the fast-retransmit.
323 	 *
324 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
325 	 * be set to snd_una, the offset will be 0, and the length may
326 	 * wind up 0.
327 	 *
328 	 * If sack_rxmit is true we are retransmitting from the scoreboard
329 	 * in which case len is already set.
330 	 */
331 	if (sack_rxmit == 0) {
332 		if (sack_bytes_rxmt == 0)
333 			len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off);
334 		else {
335 			long cwin;
336 
337                         /*
338 			 * We are inside of a SACK recovery episode and are
339 			 * sending new data, having retransmitted all the
340 			 * data possible in the scoreboard.
341 			 */
342 			len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd)
343 			       - off);
344 			/*
345 			 * Don't remove this (len > 0) check !
346 			 * We explicitly check for len > 0 here (although it
347 			 * isn't really necessary), to work around a gcc
348 			 * optimization issue - to force gcc to compute
349 			 * len above. Without this check, the computation
350 			 * of len is bungled by the optimizer.
351 			 */
352 			if (len > 0) {
353 				cwin = tp->snd_cwnd -
354 					(tp->snd_nxt - tp->sack_newdata) -
355 					sack_bytes_rxmt;
356 				if (cwin < 0)
357 					cwin = 0;
358 				len = lmin(len, cwin);
359 			}
360 		}
361 	}
362 
363 	/*
364 	 * Lop off SYN bit if it has already been sent.  However, if this
365 	 * is SYN-SENT state and if segment contains data and if we don't
366 	 * know that foreign host supports TAO, suppress sending segment.
367 	 */
368 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
369 		if (tp->t_state != TCPS_SYN_RECEIVED)
370 			flags &= ~TH_SYN;
371 		off--, len++;
372 	}
373 
374 	/*
375 	 * Be careful not to send data and/or FIN on SYN segments.
376 	 * This measure is needed to prevent interoperability problems
377 	 * with not fully conformant TCP implementations.
378 	 */
379 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
380 		len = 0;
381 		flags &= ~TH_FIN;
382 	}
383 
384 	if (len < 0) {
385 		/*
386 		 * If FIN has been sent but not acked,
387 		 * but we haven't been called to retransmit,
388 		 * len will be < 0.  Otherwise, window shrank
389 		 * after we sent into it.  If window shrank to 0,
390 		 * cancel pending retransmit, pull snd_nxt back
391 		 * to (closed) window, and set the persist timer
392 		 * if it isn't already going.  If the window didn't
393 		 * close completely, just wait for an ACK.
394 		 */
395 		len = 0;
396 		if (sendwin == 0) {
397 			tcp_timer_activate(tp, TT_REXMT, 0);
398 			tp->t_rxtshift = 0;
399 			tp->snd_nxt = tp->snd_una;
400 			if (!tcp_timer_active(tp, TT_PERSIST))
401 				tcp_setpersist(tp);
402 		}
403 	}
404 
405 	/* len will be >= 0 after this point. */
406 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
407 
408 	/*
409 	 * Automatic sizing of send socket buffer.  Often the send buffer
410 	 * size is not optimally adjusted to the actual network conditions
411 	 * at hand (delay bandwidth product).  Setting the buffer size too
412 	 * small limits throughput on links with high bandwidth and high
413 	 * delay (eg. trans-continental/oceanic links).  Setting the
414 	 * buffer size too big consumes too much real kernel memory,
415 	 * especially with many connections on busy servers.
416 	 *
417 	 * The criteria to step up the send buffer one notch are:
418 	 *  1. receive window of remote host is larger than send buffer
419 	 *     (with a fudge factor of 5/4th);
420 	 *  2. send buffer is filled to 7/8th with data (so we actually
421 	 *     have data to make use of it);
422 	 *  3. send buffer fill has not hit maximal automatic size;
423 	 *  4. our send window (slow start and cogestion controlled) is
424 	 *     larger than sent but unacknowledged data in send buffer.
425 	 *
426 	 * The remote host receive window scaling factor may limit the
427 	 * growing of the send buffer before it reaches its allowed
428 	 * maximum.
429 	 *
430 	 * It scales directly with slow start or congestion window
431 	 * and does at most one step per received ACK.  This fast
432 	 * scaling has the drawback of growing the send buffer beyond
433 	 * what is strictly necessary to make full use of a given
434 	 * delay*bandwith product.  However testing has shown this not
435 	 * to be much of an problem.  At worst we are trading wasting
436 	 * of available bandwith (the non-use of it) for wasting some
437 	 * socket buffer memory.
438 	 *
439 	 * TODO: Shrink send buffer during idle periods together
440 	 * with congestion window.  Requires another timer.  Has to
441 	 * wait for upcoming tcp timer rewrite.
442 	 */
443 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
444 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
445 		    so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
446 		    so->so_snd.sb_cc < V_tcp_autosndbuf_max &&
447 		    sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
448 			if (!sbreserve_locked(&so->so_snd,
449 			    min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
450 			     V_tcp_autosndbuf_max), so, curthread))
451 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
452 		}
453 	}
454 
455 	/*
456 	 * Truncate to the maximum segment length or enable TCP Segmentation
457 	 * Offloading (if supported by hardware) and ensure that FIN is removed
458 	 * if the length no longer contains the last data byte.
459 	 *
460 	 * TSO may only be used if we are in a pure bulk sending state.  The
461 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
462 	 * IP options prevent using TSO.  With TSO the TCP header is the same
463 	 * (except for the sequence number) for all generated packets.  This
464 	 * makes it impossible to transmit any options which vary per generated
465 	 * segment or packet.
466 	 *
467 	 * The length of TSO bursts is limited to TCP_MAXWIN.  That limit and
468 	 * removal of FIN (if not already catched here) are handled later after
469 	 * the exact length of the TCP options are known.
470 	 */
471 #ifdef IPSEC
472 	/*
473 	 * Pre-calculate here as we save another lookup into the darknesses
474 	 * of IPsec that way and can actually decide if TSO is ok.
475 	 */
476 	ipsec_optlen = ipsec_hdrsiz_tcp(tp);
477 #endif
478 	if (len > tp->t_maxseg) {
479 		if ((tp->t_flags & TF_TSO) && V_tcp_do_tso &&
480 		    ((tp->t_flags & TF_SIGNATURE) == 0) &&
481 		    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
482 		    tp->t_inpcb->inp_options == NULL &&
483 		    tp->t_inpcb->in6p_options == NULL
484 #ifdef IPSEC
485 		    && ipsec_optlen == 0
486 #endif
487 		    ) {
488 			tso = 1;
489 		} else {
490 			len = tp->t_maxseg;
491 			sendalot = 1;
492 			tso = 0;
493 		}
494 	}
495 	if (sack_rxmit) {
496 		if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
497 			flags &= ~TH_FIN;
498 	} else {
499 		if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
500 			flags &= ~TH_FIN;
501 	}
502 
503 	recwin = sbspace(&so->so_rcv);
504 
505 	/*
506 	 * Sender silly window avoidance.   We transmit under the following
507 	 * conditions when len is non-zero:
508 	 *
509 	 *	- We have a full segment (or more with TSO)
510 	 *	- This is the last buffer in a write()/send() and we are
511 	 *	  either idle or running NODELAY
512 	 *	- we've timed out (e.g. persist timer)
513 	 *	- we have more then 1/2 the maximum send window's worth of
514 	 *	  data (receiver may be limited the window size)
515 	 *	- we need to retransmit
516 	 */
517 	if (len) {
518 		if (len >= tp->t_maxseg)
519 			goto send;
520 		/*
521 		 * NOTE! on localhost connections an 'ack' from the remote
522 		 * end may occur synchronously with the output and cause
523 		 * us to flush a buffer queued with moretocome.  XXX
524 		 *
525 		 * note: the len + off check is almost certainly unnecessary.
526 		 */
527 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
528 		    (idle || (tp->t_flags & TF_NODELAY)) &&
529 		    len + off >= so->so_snd.sb_cc &&
530 		    (tp->t_flags & TF_NOPUSH) == 0) {
531 			goto send;
532 		}
533 		if (tp->t_flags & TF_FORCEDATA)		/* typ. timeout case */
534 			goto send;
535 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
536 			goto send;
537 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
538 			goto send;
539 		if (sack_rxmit)
540 			goto send;
541 	}
542 
543 	/*
544 	 * Compare available window to amount of window
545 	 * known to peer (as advertised window less
546 	 * next expected input).  If the difference is at least two
547 	 * max size segments, or at least 50% of the maximum possible
548 	 * window, then want to send a window update to peer.
549 	 * Skip this if the connection is in T/TCP half-open state.
550 	 * Don't send pure window updates when the peer has closed
551 	 * the connection and won't ever send more data.
552 	 */
553 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
554 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
555 		/*
556 		 * "adv" is the amount we can increase the window,
557 		 * taking into account that we are limited by
558 		 * TCP_MAXWIN << tp->rcv_scale.
559 		 */
560 		long adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale) -
561 			(tp->rcv_adv - tp->rcv_nxt);
562 
563 		if (adv >= (long) (2 * tp->t_maxseg))
564 			goto send;
565 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
566 			goto send;
567 	}
568 
569 	/*
570 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
571 	 * is also a catch-all for the retransmit timer timeout case.
572 	 */
573 	if (tp->t_flags & TF_ACKNOW)
574 		goto send;
575 	if ((flags & TH_RST) ||
576 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
577 		goto send;
578 	if (SEQ_GT(tp->snd_up, tp->snd_una))
579 		goto send;
580 	/*
581 	 * If our state indicates that FIN should be sent
582 	 * and we have not yet done so, then we need to send.
583 	 */
584 	if (flags & TH_FIN &&
585 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
586 		goto send;
587 	/*
588 	 * In SACK, it is possible for tcp_output to fail to send a segment
589 	 * after the retransmission timer has been turned off.  Make sure
590 	 * that the retransmission timer is set.
591 	 */
592 	if ((tp->t_flags & TF_SACK_PERMIT) &&
593 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
594 	    !tcp_timer_active(tp, TT_REXMT) &&
595 	    !tcp_timer_active(tp, TT_PERSIST)) {
596 		tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
597 		goto just_return;
598 	}
599 	/*
600 	 * TCP window updates are not reliable, rather a polling protocol
601 	 * using ``persist'' packets is used to insure receipt of window
602 	 * updates.  The three ``states'' for the output side are:
603 	 *	idle			not doing retransmits or persists
604 	 *	persisting		to move a small or zero window
605 	 *	(re)transmitting	and thereby not persisting
606 	 *
607 	 * tcp_timer_active(tp, TT_PERSIST)
608 	 *	is true when we are in persist state.
609 	 * (tp->t_flags & TF_FORCEDATA)
610 	 *	is set when we are called to send a persist packet.
611 	 * tcp_timer_active(tp, TT_REXMT)
612 	 *	is set when we are retransmitting
613 	 * The output side is idle when both timers are zero.
614 	 *
615 	 * If send window is too small, there is data to transmit, and no
616 	 * retransmit or persist is pending, then go to persist state.
617 	 * If nothing happens soon, send when timer expires:
618 	 * if window is nonzero, transmit what we can,
619 	 * otherwise force out a byte.
620 	 */
621 	if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) &&
622 	    !tcp_timer_active(tp, TT_PERSIST)) {
623 		tp->t_rxtshift = 0;
624 		tcp_setpersist(tp);
625 	}
626 
627 	/*
628 	 * No reason to send a segment, just return.
629 	 */
630 just_return:
631 	SOCKBUF_UNLOCK(&so->so_snd);
632 	return (0);
633 
634 send:
635 	SOCKBUF_LOCK_ASSERT(&so->so_snd);
636 	/*
637 	 * Before ESTABLISHED, force sending of initial options
638 	 * unless TCP set not to do any options.
639 	 * NOTE: we assume that the IP/TCP header plus TCP options
640 	 * always fit in a single mbuf, leaving room for a maximum
641 	 * link header, i.e.
642 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
643 	 */
644 	optlen = 0;
645 #ifdef INET6
646 	if (isipv6)
647 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
648 	else
649 #endif
650 	hdrlen = sizeof (struct tcpiphdr);
651 
652 	/*
653 	 * Compute options for segment.
654 	 * We only have to care about SYN and established connection
655 	 * segments.  Options for SYN-ACK segments are handled in TCP
656 	 * syncache.
657 	 */
658 	if ((tp->t_flags & TF_NOOPT) == 0) {
659 		to.to_flags = 0;
660 		/* Maximum segment size. */
661 		if (flags & TH_SYN) {
662 			tp->snd_nxt = tp->iss;
663 			to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
664 			to.to_flags |= TOF_MSS;
665 		}
666 		/* Window scaling. */
667 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
668 			to.to_wscale = tp->request_r_scale;
669 			to.to_flags |= TOF_SCALE;
670 		}
671 		/* Timestamps. */
672 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
673 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
674 			to.to_tsval = ticks + tp->ts_offset;
675 			to.to_tsecr = tp->ts_recent;
676 			to.to_flags |= TOF_TS;
677 			/* Set receive buffer autosizing timestamp. */
678 			if (tp->rfbuf_ts == 0 &&
679 			    (so->so_rcv.sb_flags & SB_AUTOSIZE))
680 				tp->rfbuf_ts = ticks;
681 		}
682 		/* Selective ACK's. */
683 		if (tp->t_flags & TF_SACK_PERMIT) {
684 			if (flags & TH_SYN)
685 				to.to_flags |= TOF_SACKPERM;
686 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
687 			    (tp->t_flags & TF_SACK_PERMIT) &&
688 			    tp->rcv_numsacks > 0) {
689 				to.to_flags |= TOF_SACK;
690 				to.to_nsacks = tp->rcv_numsacks;
691 				to.to_sacks = (u_char *)tp->sackblks;
692 			}
693 		}
694 #ifdef TCP_SIGNATURE
695 		/* TCP-MD5 (RFC2385). */
696 		if (tp->t_flags & TF_SIGNATURE)
697 			to.to_flags |= TOF_SIGNATURE;
698 #endif /* TCP_SIGNATURE */
699 
700 		/* Processing the options. */
701 		hdrlen += optlen = tcp_addoptions(&to, opt);
702 	}
703 
704 #ifdef INET6
705 	if (isipv6)
706 		ipoptlen = ip6_optlen(tp->t_inpcb);
707 	else
708 #endif
709 	if (tp->t_inpcb->inp_options)
710 		ipoptlen = tp->t_inpcb->inp_options->m_len -
711 				offsetof(struct ipoption, ipopt_list);
712 	else
713 		ipoptlen = 0;
714 #ifdef IPSEC
715 	ipoptlen += ipsec_optlen;
716 #endif
717 
718 	/*
719 	 * Adjust data length if insertion of options will
720 	 * bump the packet length beyond the t_maxopd length.
721 	 * Clear the FIN bit because we cut off the tail of
722 	 * the segment.
723 	 *
724 	 * When doing TSO limit a burst to TCP_MAXWIN minus the
725 	 * IP, TCP and Options length to keep ip->ip_len from
726 	 * overflowing.  Prevent the last segment from being
727 	 * fractional thus making them all equal sized and set
728 	 * the flag to continue sending.  TSO is disabled when
729 	 * IP options or IPSEC are present.
730 	 */
731 	if (len + optlen + ipoptlen > tp->t_maxopd) {
732 		flags &= ~TH_FIN;
733 		if (tso) {
734 			if (len > TCP_MAXWIN - hdrlen - optlen) {
735 				len = TCP_MAXWIN - hdrlen - optlen;
736 				len = len - (len % (tp->t_maxopd - optlen));
737 				sendalot = 1;
738 			} else if (tp->t_flags & TF_NEEDFIN)
739 				sendalot = 1;
740 		} else {
741 			len = tp->t_maxopd - optlen - ipoptlen;
742 			sendalot = 1;
743 		}
744 	}
745 
746 /*#ifdef DIAGNOSTIC*/
747 #ifdef INET6
748 	if (max_linkhdr + hdrlen > MCLBYTES)
749 #else
750 	if (max_linkhdr + hdrlen > MHLEN)
751 #endif
752 		panic("tcphdr too big");
753 /*#endif*/
754 
755 	/*
756 	 * This KASSERT is here to catch edge cases at a well defined place.
757 	 * Before, those had triggered (random) panic conditions further down.
758 	 */
759 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
760 
761 	/*
762 	 * Grab a header mbuf, attaching a copy of data to
763 	 * be transmitted, and initialize the header from
764 	 * the template for sends on this connection.
765 	 */
766 	if (len) {
767 		struct mbuf *mb;
768 		u_int moff;
769 
770 		if ((tp->t_flags & TF_FORCEDATA) && len == 1)
771 			V_tcpstat.tcps_sndprobe++;
772 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
773 			V_tcpstat.tcps_sndrexmitpack++;
774 			V_tcpstat.tcps_sndrexmitbyte += len;
775 		} else {
776 			V_tcpstat.tcps_sndpack++;
777 			V_tcpstat.tcps_sndbyte += len;
778 		}
779 #ifdef notyet
780 		if ((m = m_copypack(so->so_snd.sb_mb, off,
781 		    (int)len, max_linkhdr + hdrlen)) == 0) {
782 			SOCKBUF_UNLOCK(&so->so_snd);
783 			error = ENOBUFS;
784 			goto out;
785 		}
786 		/*
787 		 * m_copypack left space for our hdr; use it.
788 		 */
789 		m->m_len += hdrlen;
790 		m->m_data -= hdrlen;
791 #else
792 		MGETHDR(m, M_DONTWAIT, MT_DATA);
793 		if (m == NULL) {
794 			SOCKBUF_UNLOCK(&so->so_snd);
795 			error = ENOBUFS;
796 			goto out;
797 		}
798 #ifdef INET6
799 		if (MHLEN < hdrlen + max_linkhdr) {
800 			MCLGET(m, M_DONTWAIT);
801 			if ((m->m_flags & M_EXT) == 0) {
802 				SOCKBUF_UNLOCK(&so->so_snd);
803 				m_freem(m);
804 				error = ENOBUFS;
805 				goto out;
806 			}
807 		}
808 #endif
809 		m->m_data += max_linkhdr;
810 		m->m_len = hdrlen;
811 
812 		/*
813 		 * Start the m_copy functions from the closest mbuf
814 		 * to the offset in the socket buffer chain.
815 		 */
816 		mb = sbsndptr(&so->so_snd, off, len, &moff);
817 
818 		if (len <= MHLEN - hdrlen - max_linkhdr) {
819 			m_copydata(mb, moff, (int)len,
820 			    mtod(m, caddr_t) + hdrlen);
821 			m->m_len += len;
822 		} else {
823 			m->m_next = m_copy(mb, moff, (int)len);
824 			if (m->m_next == NULL) {
825 				SOCKBUF_UNLOCK(&so->so_snd);
826 				(void) m_free(m);
827 				error = ENOBUFS;
828 				goto out;
829 			}
830 		}
831 #endif
832 		/*
833 		 * If we're sending everything we've got, set PUSH.
834 		 * (This will keep happy those implementations which only
835 		 * give data to the user when a buffer fills or
836 		 * a PUSH comes in.)
837 		 */
838 		if (off + len == so->so_snd.sb_cc)
839 			flags |= TH_PUSH;
840 		SOCKBUF_UNLOCK(&so->so_snd);
841 	} else {
842 		SOCKBUF_UNLOCK(&so->so_snd);
843 		if (tp->t_flags & TF_ACKNOW)
844 			V_tcpstat.tcps_sndacks++;
845 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
846 			V_tcpstat.tcps_sndctrl++;
847 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
848 			V_tcpstat.tcps_sndurg++;
849 		else
850 			V_tcpstat.tcps_sndwinup++;
851 
852 		MGETHDR(m, M_DONTWAIT, MT_DATA);
853 		if (m == NULL) {
854 			error = ENOBUFS;
855 			goto out;
856 		}
857 #ifdef INET6
858 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
859 		    MHLEN >= hdrlen) {
860 			MH_ALIGN(m, hdrlen);
861 		} else
862 #endif
863 		m->m_data += max_linkhdr;
864 		m->m_len = hdrlen;
865 	}
866 	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
867 	m->m_pkthdr.rcvif = (struct ifnet *)0;
868 #ifdef MAC
869 	mac_inpcb_create_mbuf(tp->t_inpcb, m);
870 #endif
871 #ifdef INET6
872 	if (isipv6) {
873 		ip6 = mtod(m, struct ip6_hdr *);
874 		th = (struct tcphdr *)(ip6 + 1);
875 		tcpip_fillheaders(tp->t_inpcb, ip6, th);
876 	} else
877 #endif /* INET6 */
878 	{
879 		ip = mtod(m, struct ip *);
880 		ipov = (struct ipovly *)ip;
881 		th = (struct tcphdr *)(ip + 1);
882 		tcpip_fillheaders(tp->t_inpcb, ip, th);
883 	}
884 
885 	/*
886 	 * Fill in fields, remembering maximum advertised
887 	 * window for use in delaying messages about window sizes.
888 	 * If resending a FIN, be sure not to use a new sequence number.
889 	 */
890 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
891 	    tp->snd_nxt == tp->snd_max)
892 		tp->snd_nxt--;
893 	/*
894 	 * If we are starting a connection, send ECN setup
895 	 * SYN packet. If we are on a retransmit, we may
896 	 * resend those bits a number of times as per
897 	 * RFC 3168.
898 	 */
899 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
900 		if (tp->t_rxtshift >= 1) {
901 			if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
902 				flags |= TH_ECE|TH_CWR;
903 		} else
904 			flags |= TH_ECE|TH_CWR;
905 	}
906 
907 	if (tp->t_state == TCPS_ESTABLISHED &&
908 	    (tp->t_flags & TF_ECN_PERMIT)) {
909 		/*
910 		 * If the peer has ECN, mark data packets with
911 		 * ECN capable transmission (ECT).
912 		 * Ignore pure ack packets, retransmissions and window probes.
913 		 */
914 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
915 		    !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
916 #ifdef INET6
917 			if (isipv6)
918 				ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
919 			else
920 #endif
921 				ip->ip_tos |= IPTOS_ECN_ECT0;
922 			V_tcpstat.tcps_ecn_ect0++;
923 		}
924 
925 		/*
926 		 * Reply with proper ECN notifications.
927 		 */
928 		if (tp->t_flags & TF_ECN_SND_CWR) {
929 			flags |= TH_CWR;
930 			tp->t_flags &= ~TF_ECN_SND_CWR;
931 		}
932 		if (tp->t_flags & TF_ECN_SND_ECE)
933 			flags |= TH_ECE;
934 	}
935 
936 	/*
937 	 * If we are doing retransmissions, then snd_nxt will
938 	 * not reflect the first unsent octet.  For ACK only
939 	 * packets, we do not want the sequence number of the
940 	 * retransmitted packet, we want the sequence number
941 	 * of the next unsent octet.  So, if there is no data
942 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
943 	 * when filling in ti_seq.  But if we are in persist
944 	 * state, snd_max might reflect one byte beyond the
945 	 * right edge of the window, so use snd_nxt in that
946 	 * case, since we know we aren't doing a retransmission.
947 	 * (retransmit and persist are mutually exclusive...)
948 	 */
949 	if (sack_rxmit == 0) {
950 		if (len || (flags & (TH_SYN|TH_FIN)) ||
951 		    tcp_timer_active(tp, TT_PERSIST))
952 			th->th_seq = htonl(tp->snd_nxt);
953 		else
954 			th->th_seq = htonl(tp->snd_max);
955 	} else {
956 		th->th_seq = htonl(p->rxmit);
957 		p->rxmit += len;
958 		tp->sackhint.sack_bytes_rexmit += len;
959 	}
960 	th->th_ack = htonl(tp->rcv_nxt);
961 	if (optlen) {
962 		bcopy(opt, th + 1, optlen);
963 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
964 	}
965 	th->th_flags = flags;
966 	/*
967 	 * Calculate receive window.  Don't shrink window,
968 	 * but avoid silly window syndrome.
969 	 */
970 	if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
971 	    recwin < (long)tp->t_maxseg)
972 		recwin = 0;
973 	if (recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
974 		recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
975 	if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
976 		recwin = (long)TCP_MAXWIN << tp->rcv_scale;
977 
978 	/*
979 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
980 	 * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
981 	 * case is handled in syncache.
982 	 */
983 	if (flags & TH_SYN)
984 		th->th_win = htons((u_short)
985 				(min(sbspace(&so->so_rcv), TCP_MAXWIN)));
986 	else
987 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
988 
989 	/*
990 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
991 	 * a 0 window.  This may cause the remote transmitter to stall.  This
992 	 * flag tells soreceive() to disable delayed acknowledgements when
993 	 * draining the buffer.  This can occur if the receiver is attempting
994 	 * to read more data than can be buffered prior to transmitting on
995 	 * the connection.
996 	 */
997 	if (recwin == 0)
998 		tp->t_flags |= TF_RXWIN0SENT;
999 	else
1000 		tp->t_flags &= ~TF_RXWIN0SENT;
1001 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1002 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1003 		th->th_flags |= TH_URG;
1004 	} else
1005 		/*
1006 		 * If no urgent pointer to send, then we pull
1007 		 * the urgent pointer to the left edge of the send window
1008 		 * so that it doesn't drift into the send window on sequence
1009 		 * number wraparound.
1010 		 */
1011 		tp->snd_up = tp->snd_una;		/* drag it along */
1012 
1013 #ifdef TCP_SIGNATURE
1014 	if (tp->t_flags & TF_SIGNATURE) {
1015 		int sigoff = to.to_signature - opt;
1016 		tcp_signature_compute(m, 0, len, optlen,
1017 		    (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
1018 	}
1019 #endif
1020 
1021 	/*
1022 	 * Put TCP length in extended header, and then
1023 	 * checksum extended header and data.
1024 	 */
1025 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1026 #ifdef INET6
1027 	if (isipv6)
1028 		/*
1029 		 * ip6_plen is not need to be filled now, and will be filled
1030 		 * in ip6_output.
1031 		 */
1032 		th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
1033 				       sizeof(struct tcphdr) + optlen + len);
1034 	else
1035 #endif /* INET6 */
1036 	{
1037 		m->m_pkthdr.csum_flags = CSUM_TCP;
1038 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1039 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1040 		    htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1041 
1042 		/* IP version must be set here for ipv4/ipv6 checking later */
1043 		KASSERT(ip->ip_v == IPVERSION,
1044 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1045 	}
1046 
1047 	/*
1048 	 * Enable TSO and specify the size of the segments.
1049 	 * The TCP pseudo header checksum is always provided.
1050 	 * XXX: Fixme: This is currently not the case for IPv6.
1051 	 */
1052 	if (tso) {
1053 		m->m_pkthdr.csum_flags = CSUM_TSO;
1054 		m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen;
1055 	}
1056 
1057 	/*
1058 	 * In transmit state, time the transmission and arrange for
1059 	 * the retransmit.  In persist state, just set snd_max.
1060 	 */
1061 	if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1062 	    !tcp_timer_active(tp, TT_PERSIST)) {
1063 		tcp_seq startseq = tp->snd_nxt;
1064 
1065 		/*
1066 		 * Advance snd_nxt over sequence space of this segment.
1067 		 */
1068 		if (flags & (TH_SYN|TH_FIN)) {
1069 			if (flags & TH_SYN)
1070 				tp->snd_nxt++;
1071 			if (flags & TH_FIN) {
1072 				tp->snd_nxt++;
1073 				tp->t_flags |= TF_SENTFIN;
1074 			}
1075 		}
1076 		if (sack_rxmit)
1077 			goto timer;
1078 		tp->snd_nxt += len;
1079 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1080 			tp->snd_max = tp->snd_nxt;
1081 			/*
1082 			 * Time this transmission if not a retransmission and
1083 			 * not currently timing anything.
1084 			 */
1085 			if (tp->t_rtttime == 0) {
1086 				tp->t_rtttime = ticks;
1087 				tp->t_rtseq = startseq;
1088 				V_tcpstat.tcps_segstimed++;
1089 			}
1090 		}
1091 
1092 		/*
1093 		 * Set retransmit timer if not currently set,
1094 		 * and not doing a pure ack or a keep-alive probe.
1095 		 * Initial value for retransmit timer is smoothed
1096 		 * round-trip time + 2 * round-trip time variance.
1097 		 * Initialize shift counter which is used for backoff
1098 		 * of retransmit time.
1099 		 */
1100 timer:
1101 		if (!tcp_timer_active(tp, TT_REXMT) &&
1102 		    ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1103 		     (tp->snd_nxt != tp->snd_una))) {
1104 			if (tcp_timer_active(tp, TT_PERSIST)) {
1105 				tcp_timer_activate(tp, TT_PERSIST, 0);
1106 				tp->t_rxtshift = 0;
1107 			}
1108 			tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1109 		}
1110 	} else {
1111 		/*
1112 		 * Persist case, update snd_max but since we are in
1113 		 * persist mode (no window) we do not update snd_nxt.
1114 		 */
1115 		int xlen = len;
1116 		if (flags & TH_SYN)
1117 			++xlen;
1118 		if (flags & TH_FIN) {
1119 			++xlen;
1120 			tp->t_flags |= TF_SENTFIN;
1121 		}
1122 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1123 			tp->snd_max = tp->snd_nxt + len;
1124 	}
1125 
1126 #ifdef TCPDEBUG
1127 	/*
1128 	 * Trace.
1129 	 */
1130 	if (so->so_options & SO_DEBUG) {
1131 		u_short save = 0;
1132 #ifdef INET6
1133 		if (!isipv6)
1134 #endif
1135 		{
1136 			save = ipov->ih_len;
1137 			ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1138 		}
1139 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1140 #ifdef INET6
1141 		if (!isipv6)
1142 #endif
1143 		ipov->ih_len = save;
1144 	}
1145 #endif
1146 
1147 	/*
1148 	 * Fill in IP length and desired time to live and
1149 	 * send to IP level.  There should be a better way
1150 	 * to handle ttl and tos; we could keep them in
1151 	 * the template, but need a way to checksum without them.
1152 	 */
1153 	/*
1154 	 * m->m_pkthdr.len should have been set before cksum calcuration,
1155 	 * because in6_cksum() need it.
1156 	 */
1157 #ifdef INET6
1158 	if (isipv6) {
1159 		/*
1160 		 * we separately set hoplimit for every segment, since the
1161 		 * user might want to change the value via setsockopt.
1162 		 * Also, desired default hop limit might be changed via
1163 		 * Neighbor Discovery.
1164 		 */
1165 		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1166 
1167 		/* TODO: IPv6 IP6TOS_ECT bit on */
1168 		error = ip6_output(m,
1169 			    tp->t_inpcb->in6p_outputopts, NULL,
1170 			    ((so->so_options & SO_DONTROUTE) ?
1171 			    IP_ROUTETOIF : 0), NULL, NULL, tp->t_inpcb);
1172 	} else
1173 #endif /* INET6 */
1174     {
1175 	ip->ip_len = m->m_pkthdr.len;
1176 #ifdef INET6
1177 	if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1178 		ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1179 #endif /* INET6 */
1180 	/*
1181 	 * If we do path MTU discovery, then we set DF on every packet.
1182 	 * This might not be the best thing to do according to RFC3390
1183 	 * Section 2. However the tcp hostcache migitates the problem
1184 	 * so it affects only the first tcp connection with a host.
1185 	 */
1186 	if (V_path_mtu_discovery)
1187 		ip->ip_off |= IP_DF;
1188 
1189 	error = ip_output(m, tp->t_inpcb->inp_options, NULL,
1190 	    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1191 	    tp->t_inpcb);
1192     }
1193 	if (error) {
1194 
1195 		/*
1196 		 * We know that the packet was lost, so back out the
1197 		 * sequence number advance, if any.
1198 		 *
1199 		 * If the error is EPERM the packet got blocked by the
1200 		 * local firewall.  Normally we should terminate the
1201 		 * connection but the blocking may have been spurious
1202 		 * due to a firewall reconfiguration cycle.  So we treat
1203 		 * it like a packet loss and let the retransmit timer and
1204 		 * timeouts do their work over time.
1205 		 * XXX: It is a POLA question whether calling tcp_drop right
1206 		 * away would be the really correct behavior instead.
1207 		 */
1208 		if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1209 		    !tcp_timer_active(tp, TT_PERSIST)) &&
1210 		    ((flags & TH_SYN) == 0) &&
1211 		    (error != EPERM)) {
1212 			if (sack_rxmit) {
1213 				p->rxmit -= len;
1214 				tp->sackhint.sack_bytes_rexmit -= len;
1215 				KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1216 				    ("sackhint bytes rtx >= 0"));
1217 			} else
1218 				tp->snd_nxt -= len;
1219 		}
1220 out:
1221 		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1222 		switch (error) {
1223 		case EPERM:
1224 			tp->t_softerror = error;
1225 			return (error);
1226 		case ENOBUFS:
1227 	                if (!tcp_timer_active(tp, TT_REXMT) &&
1228 			    !tcp_timer_active(tp, TT_PERSIST))
1229 	                        tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1230 			tp->snd_cwnd = tp->t_maxseg;
1231 			return (0);
1232 		case EMSGSIZE:
1233 			/*
1234 			 * For some reason the interface we used initially
1235 			 * to send segments changed to another or lowered
1236 			 * its MTU.
1237 			 *
1238 			 * tcp_mtudisc() will find out the new MTU and as
1239 			 * its last action, initiate retransmission, so it
1240 			 * is important to not do so here.
1241 			 *
1242 			 * If TSO was active we either got an interface
1243 			 * without TSO capabilits or TSO was turned off.
1244 			 * Disable it for this connection as too and
1245 			 * immediatly retry with MSS sized segments generated
1246 			 * by this function.
1247 			 */
1248 			if (tso)
1249 				tp->t_flags &= ~TF_TSO;
1250 			tcp_mtudisc(tp->t_inpcb, 0);
1251 			return (0);
1252 		case EHOSTDOWN:
1253 		case EHOSTUNREACH:
1254 		case ENETDOWN:
1255 		case ENETUNREACH:
1256 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
1257 				tp->t_softerror = error;
1258 				return (0);
1259 			}
1260 			/* FALLTHROUGH */
1261 		default:
1262 			return (error);
1263 		}
1264 	}
1265 	V_tcpstat.tcps_sndtotal++;
1266 
1267 	/*
1268 	 * Data sent (as far as we can tell).
1269 	 * If this advertises a larger window than any other segment,
1270 	 * then remember the size of the advertised window.
1271 	 * Any pending ACK has now been sent.
1272 	 */
1273 	if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1274 		tp->rcv_adv = tp->rcv_nxt + recwin;
1275 	tp->last_ack_sent = tp->rcv_nxt;
1276 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1277 	if (tcp_timer_active(tp, TT_DELACK))
1278 		tcp_timer_activate(tp, TT_DELACK, 0);
1279 #if 0
1280 	/*
1281 	 * This completely breaks TCP if newreno is turned on.  What happens
1282 	 * is that if delayed-acks are turned on on the receiver, this code
1283 	 * on the transmitter effectively destroys the TCP window, forcing
1284 	 * it to four packets (1.5Kx4 = 6K window).
1285 	 */
1286 	if (sendalot && (!V_tcp_do_newreno || --maxburst))
1287 		goto again;
1288 #endif
1289 	if (sendalot)
1290 		goto again;
1291 	return (0);
1292 }
1293 
1294 void
1295 tcp_setpersist(struct tcpcb *tp)
1296 {
1297 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1298 	int tt;
1299 
1300 	if (tcp_timer_active(tp, TT_REXMT))
1301 		panic("tcp_setpersist: retransmit pending");
1302 	/*
1303 	 * Start/restart persistance timer.
1304 	 */
1305 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1306 		      TCPTV_PERSMIN, TCPTV_PERSMAX);
1307 	tcp_timer_activate(tp, TT_PERSIST, tt);
1308 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1309 		tp->t_rxtshift++;
1310 }
1311 
1312 /*
1313  * Insert TCP options according to the supplied parameters to the place
1314  * optp in a consistent way.  Can handle unaligned destinations.
1315  *
1316  * The order of the option processing is crucial for optimal packing and
1317  * alignment for the scarce option space.
1318  *
1319  * The optimal order for a SYN/SYN-ACK segment is:
1320  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1321  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1322  *
1323  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1324  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1325  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1326  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1327  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1328  */
1329 int
1330 tcp_addoptions(struct tcpopt *to, u_char *optp)
1331 {
1332 	INIT_VNET_INET(curvnet);
1333 	u_int mask, optlen = 0;
1334 
1335 	for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1336 		if ((to->to_flags & mask) != mask)
1337 			continue;
1338 		if (optlen == TCP_MAXOLEN)
1339 			break;
1340 		switch (to->to_flags & mask) {
1341 		case TOF_MSS:
1342 			while (optlen % 4) {
1343 				optlen += TCPOLEN_NOP;
1344 				*optp++ = TCPOPT_NOP;
1345 			}
1346 			if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1347 				continue;
1348 			optlen += TCPOLEN_MAXSEG;
1349 			*optp++ = TCPOPT_MAXSEG;
1350 			*optp++ = TCPOLEN_MAXSEG;
1351 			to->to_mss = htons(to->to_mss);
1352 			bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1353 			optp += sizeof(to->to_mss);
1354 			break;
1355 		case TOF_SCALE:
1356 			while (!optlen || optlen % 2 != 1) {
1357 				optlen += TCPOLEN_NOP;
1358 				*optp++ = TCPOPT_NOP;
1359 			}
1360 			if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1361 				continue;
1362 			optlen += TCPOLEN_WINDOW;
1363 			*optp++ = TCPOPT_WINDOW;
1364 			*optp++ = TCPOLEN_WINDOW;
1365 			*optp++ = to->to_wscale;
1366 			break;
1367 		case TOF_SACKPERM:
1368 			while (optlen % 2) {
1369 				optlen += TCPOLEN_NOP;
1370 				*optp++ = TCPOPT_NOP;
1371 			}
1372 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1373 				continue;
1374 			optlen += TCPOLEN_SACK_PERMITTED;
1375 			*optp++ = TCPOPT_SACK_PERMITTED;
1376 			*optp++ = TCPOLEN_SACK_PERMITTED;
1377 			break;
1378 		case TOF_TS:
1379 			while (!optlen || optlen % 4 != 2) {
1380 				optlen += TCPOLEN_NOP;
1381 				*optp++ = TCPOPT_NOP;
1382 			}
1383 			if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1384 				continue;
1385 			optlen += TCPOLEN_TIMESTAMP;
1386 			*optp++ = TCPOPT_TIMESTAMP;
1387 			*optp++ = TCPOLEN_TIMESTAMP;
1388 			to->to_tsval = htonl(to->to_tsval);
1389 			to->to_tsecr = htonl(to->to_tsecr);
1390 			bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1391 			optp += sizeof(to->to_tsval);
1392 			bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1393 			optp += sizeof(to->to_tsecr);
1394 			break;
1395 		case TOF_SIGNATURE:
1396 			{
1397 			int siglen = TCPOLEN_SIGNATURE - 2;
1398 
1399 			while (!optlen || optlen % 4 != 2) {
1400 				optlen += TCPOLEN_NOP;
1401 				*optp++ = TCPOPT_NOP;
1402 			}
1403 			if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE)
1404 				continue;
1405 			optlen += TCPOLEN_SIGNATURE;
1406 			*optp++ = TCPOPT_SIGNATURE;
1407 			*optp++ = TCPOLEN_SIGNATURE;
1408 			to->to_signature = optp;
1409 			while (siglen--)
1410 				 *optp++ = 0;
1411 			break;
1412 			}
1413 		case TOF_SACK:
1414 			{
1415 			int sackblks = 0;
1416 			struct sackblk *sack = (struct sackblk *)to->to_sacks;
1417 			tcp_seq sack_seq;
1418 
1419 			while (!optlen || optlen % 4 != 2) {
1420 				optlen += TCPOLEN_NOP;
1421 				*optp++ = TCPOPT_NOP;
1422 			}
1423 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1424 				continue;
1425 			optlen += TCPOLEN_SACKHDR;
1426 			*optp++ = TCPOPT_SACK;
1427 			sackblks = min(to->to_nsacks,
1428 					(TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1429 			*optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1430 			while (sackblks--) {
1431 				sack_seq = htonl(sack->start);
1432 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1433 				optp += sizeof(sack_seq);
1434 				sack_seq = htonl(sack->end);
1435 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1436 				optp += sizeof(sack_seq);
1437 				optlen += TCPOLEN_SACK;
1438 				sack++;
1439 			}
1440 			V_tcpstat.tcps_sack_send_blocks++;
1441 			break;
1442 			}
1443 		default:
1444 			panic("%s: unknown TCP option type", __func__);
1445 			break;
1446 		}
1447 	}
1448 
1449 	/* Terminate and pad TCP options to a 4 byte boundary. */
1450 	if (optlen % 4) {
1451 		optlen += TCPOLEN_EOL;
1452 		*optp++ = TCPOPT_EOL;
1453 	}
1454 	/*
1455 	 * According to RFC 793 (STD0007):
1456 	 *   "The content of the header beyond the End-of-Option option
1457 	 *    must be header padding (i.e., zero)."
1458 	 *   and later: "The padding is composed of zeros."
1459 	 */
1460 	while (optlen % 4) {
1461 		optlen += TCPOLEN_PAD;
1462 		*optp++ = TCPOPT_PAD;
1463 	}
1464 
1465 	KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1466 	return (optlen);
1467 }
1468