xref: /freebsd/sys/netinet/tcp_output.c (revision 830940567b49bb0c08dfaed40418999e76616909)
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_tcpdebug.h"
39 
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/domain.h>
43 #include <sys/kernel.h>
44 #include <sys/lock.h>
45 #include <sys/mbuf.h>
46 #include <sys/mutex.h>
47 #include <sys/protosw.h>
48 #include <sys/socket.h>
49 #include <sys/socketvar.h>
50 #include <sys/sysctl.h>
51 
52 #include <net/if.h>
53 #include <net/route.h>
54 #include <net/vnet.h>
55 
56 #include <netinet/in.h>
57 #include <netinet/in_systm.h>
58 #include <netinet/ip.h>
59 #include <netinet/in_pcb.h>
60 #include <netinet/ip_var.h>
61 #include <netinet/ip_options.h>
62 #ifdef INET6
63 #include <netinet6/in6_pcb.h>
64 #include <netinet/ip6.h>
65 #include <netinet6/ip6_var.h>
66 #endif
67 #include <netinet/tcp.h>
68 #define	TCPOUTFLAGS
69 #include <netinet/tcp_fsm.h>
70 #include <netinet/tcp_seq.h>
71 #include <netinet/tcp_timer.h>
72 #include <netinet/tcp_var.h>
73 #include <netinet/tcpip.h>
74 #ifdef TCPDEBUG
75 #include <netinet/tcp_debug.h>
76 #endif
77 
78 #ifdef IPSEC
79 #include <netipsec/ipsec.h>
80 #endif /*IPSEC*/
81 
82 #include <machine/in_cksum.h>
83 
84 #include <security/mac/mac_framework.h>
85 
86 #ifdef notyet
87 extern struct mbuf *m_copypack();
88 #endif
89 
90 VNET_DEFINE(int, path_mtu_discovery);
91 VNET_DEFINE(int, ss_fltsz);
92 VNET_DEFINE(int, ss_fltsz_local);
93 VNET_DEFINE(int, tcp_do_newreno);
94 VNET_DEFINE(int, tcp_do_tso);
95 VNET_DEFINE(int, tcp_do_autosndbuf);
96 VNET_DEFINE(int, tcp_autosndbuf_inc);
97 VNET_DEFINE(int, tcp_autosndbuf_max);
98 
99 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
100 	&VNET_NAME(path_mtu_discovery), 1,
101 	"Enable Path MTU Discovery");
102 
103 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, slowstart_flightsize, CTLFLAG_RW,
104 	&VNET_NAME(ss_fltsz), 1,
105 	"Slow start flight size");
106 
107 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, local_slowstart_flightsize,
108 	CTLFLAG_RW, &VNET_NAME(ss_fltsz_local), 1,
109 	"Slow start flight size for local networks");
110 
111 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, newreno, CTLFLAG_RW,
112 	&VNET_NAME(tcp_do_newreno), 0,
113 	"Enable NewReno Algorithms");
114 
115 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW,
116 	&VNET_NAME(tcp_do_tso), 0,
117 	"Enable TCP Segmentation Offload");
118 
119 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW,
120 	&VNET_NAME(tcp_do_autosndbuf), 0,
121 	"Enable automatic send buffer sizing");
122 
123 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW,
124 	&VNET_NAME(tcp_autosndbuf_inc), 0,
125 	"Incrementor step size of automatic send buffer");
126 
127 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW,
128 	&VNET_NAME(tcp_autosndbuf_max), 0,
129 	"Max size of automatic send buffer");
130 
131 
132 /*
133  * Tcp output routine: figure out what should be sent and send it.
134  */
135 int
136 tcp_output(struct tcpcb *tp)
137 {
138 	struct socket *so = tp->t_inpcb->inp_socket;
139 	long len, recwin, sendwin;
140 	int off, flags, error;
141 	struct mbuf *m;
142 	struct ip *ip = NULL;
143 	struct ipovly *ipov = NULL;
144 	struct tcphdr *th;
145 	u_char opt[TCP_MAXOLEN];
146 	unsigned ipoptlen, optlen, hdrlen;
147 #ifdef IPSEC
148 	unsigned ipsec_optlen = 0;
149 #endif
150 	int idle, sendalot;
151 	int sack_rxmit, sack_bytes_rxmt;
152 	struct sackhole *p;
153 	int tso = 0;
154 	struct tcpopt to;
155 #if 0
156 	int maxburst = TCP_MAXBURST;
157 #endif
158 #ifdef INET6
159 	struct ip6_hdr *ip6 = NULL;
160 	int isipv6;
161 
162 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
163 #endif
164 
165 	INP_WLOCK_ASSERT(tp->t_inpcb);
166 
167 	/*
168 	 * Determine length of data that should be transmitted,
169 	 * and flags that will be used.
170 	 * If there is some data or critical controls (SYN, RST)
171 	 * to send, then transmit; otherwise, investigate further.
172 	 */
173 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
174 	if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur) {
175 		/*
176 		 * We have been idle for "a while" and no acks are
177 		 * expected to clock out any data we send --
178 		 * slow start to get ack "clock" running again.
179 		 *
180 		 * Set the slow-start flight size depending on whether
181 		 * this is a local network or not.
182 		 */
183 		int ss = V_ss_fltsz;
184 #ifdef INET6
185 		if (isipv6) {
186 			if (in6_localaddr(&tp->t_inpcb->in6p_faddr))
187 				ss = V_ss_fltsz_local;
188 		} else
189 #endif /* INET6 */
190 		if (in_localaddr(tp->t_inpcb->inp_faddr))
191 			ss = V_ss_fltsz_local;
192 		tp->snd_cwnd = tp->t_maxseg * ss;
193 	}
194 	tp->t_flags &= ~TF_LASTIDLE;
195 	if (idle) {
196 		if (tp->t_flags & TF_MORETOCOME) {
197 			tp->t_flags |= TF_LASTIDLE;
198 			idle = 0;
199 		}
200 	}
201 again:
202 	/*
203 	 * If we've recently taken a timeout, snd_max will be greater than
204 	 * snd_nxt.  There may be SACK information that allows us to avoid
205 	 * resending already delivered data.  Adjust snd_nxt accordingly.
206 	 */
207 	if ((tp->t_flags & TF_SACK_PERMIT) &&
208 	    SEQ_LT(tp->snd_nxt, tp->snd_max))
209 		tcp_sack_adjust(tp);
210 	sendalot = 0;
211 	off = tp->snd_nxt - tp->snd_una;
212 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
213 	sendwin = min(sendwin, tp->snd_bwnd);
214 
215 	flags = tcp_outflags[tp->t_state];
216 	/*
217 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
218 	 * to send out new data (when sendalot is 1), bypass this function.
219 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
220 	 * we're replacing a (future) new transmission with a retransmission
221 	 * now, and we previously incremented snd_cwnd in tcp_input().
222 	 */
223 	/*
224 	 * Still in sack recovery , reset rxmit flag to zero.
225 	 */
226 	sack_rxmit = 0;
227 	sack_bytes_rxmt = 0;
228 	len = 0;
229 	p = NULL;
230 	if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp) &&
231 	    (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
232 		long cwin;
233 
234 		cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
235 		if (cwin < 0)
236 			cwin = 0;
237 		/* Do not retransmit SACK segments beyond snd_recover */
238 		if (SEQ_GT(p->end, tp->snd_recover)) {
239 			/*
240 			 * (At least) part of sack hole extends beyond
241 			 * snd_recover. Check to see if we can rexmit data
242 			 * for this hole.
243 			 */
244 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
245 				/*
246 				 * Can't rexmit any more data for this hole.
247 				 * That data will be rexmitted in the next
248 				 * sack recovery episode, when snd_recover
249 				 * moves past p->rxmit.
250 				 */
251 				p = NULL;
252 				goto after_sack_rexmit;
253 			} else
254 				/* Can rexmit part of the current hole */
255 				len = ((long)ulmin(cwin,
256 						   tp->snd_recover - p->rxmit));
257 		} else
258 			len = ((long)ulmin(cwin, p->end - p->rxmit));
259 		off = p->rxmit - tp->snd_una;
260 		KASSERT(off >= 0,("%s: sack block to the left of una : %d",
261 		    __func__, off));
262 		if (len > 0) {
263 			sack_rxmit = 1;
264 			sendalot = 1;
265 			TCPSTAT_INC(tcps_sack_rexmits);
266 			TCPSTAT_ADD(tcps_sack_rexmit_bytes,
267 			    min(len, tp->t_maxseg));
268 		}
269 	}
270 after_sack_rexmit:
271 	/*
272 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
273 	 * state flags.
274 	 */
275 	if (tp->t_flags & TF_NEEDFIN)
276 		flags |= TH_FIN;
277 	if (tp->t_flags & TF_NEEDSYN)
278 		flags |= TH_SYN;
279 
280 	SOCKBUF_LOCK(&so->so_snd);
281 	/*
282 	 * If in persist timeout with window of 0, send 1 byte.
283 	 * Otherwise, if window is small but nonzero
284 	 * and timer expired, we will send what we can
285 	 * and go to transmit state.
286 	 */
287 	if (tp->t_flags & TF_FORCEDATA) {
288 		if (sendwin == 0) {
289 			/*
290 			 * If we still have some data to send, then
291 			 * clear the FIN bit.  Usually this would
292 			 * happen below when it realizes that we
293 			 * aren't sending all the data.  However,
294 			 * if we have exactly 1 byte of unsent data,
295 			 * then it won't clear the FIN bit below,
296 			 * and if we are in persist state, we wind
297 			 * up sending the packet without recording
298 			 * that we sent the FIN bit.
299 			 *
300 			 * We can't just blindly clear the FIN bit,
301 			 * because if we don't have any more data
302 			 * to send then the probe will be the FIN
303 			 * itself.
304 			 */
305 			if (off < so->so_snd.sb_cc)
306 				flags &= ~TH_FIN;
307 			sendwin = 1;
308 		} else {
309 			tcp_timer_activate(tp, TT_PERSIST, 0);
310 			tp->t_rxtshift = 0;
311 		}
312 	}
313 
314 	/*
315 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
316 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
317 	 * a negative length.  This can also occur when TCP opens up
318 	 * its congestion window while receiving additional duplicate
319 	 * acks after fast-retransmit because TCP will reset snd_nxt
320 	 * to snd_max after the fast-retransmit.
321 	 *
322 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
323 	 * be set to snd_una, the offset will be 0, and the length may
324 	 * wind up 0.
325 	 *
326 	 * If sack_rxmit is true we are retransmitting from the scoreboard
327 	 * in which case len is already set.
328 	 */
329 	if (sack_rxmit == 0) {
330 		if (sack_bytes_rxmt == 0)
331 			len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off);
332 		else {
333 			long cwin;
334 
335                         /*
336 			 * We are inside of a SACK recovery episode and are
337 			 * sending new data, having retransmitted all the
338 			 * data possible in the scoreboard.
339 			 */
340 			len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd)
341 			       - off);
342 			/*
343 			 * Don't remove this (len > 0) check !
344 			 * We explicitly check for len > 0 here (although it
345 			 * isn't really necessary), to work around a gcc
346 			 * optimization issue - to force gcc to compute
347 			 * len above. Without this check, the computation
348 			 * of len is bungled by the optimizer.
349 			 */
350 			if (len > 0) {
351 				cwin = tp->snd_cwnd -
352 					(tp->snd_nxt - tp->sack_newdata) -
353 					sack_bytes_rxmt;
354 				if (cwin < 0)
355 					cwin = 0;
356 				len = lmin(len, cwin);
357 			}
358 		}
359 	}
360 
361 	/*
362 	 * Lop off SYN bit if it has already been sent.  However, if this
363 	 * is SYN-SENT state and if segment contains data and if we don't
364 	 * know that foreign host supports TAO, suppress sending segment.
365 	 */
366 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
367 		if (tp->t_state != TCPS_SYN_RECEIVED)
368 			flags &= ~TH_SYN;
369 		off--, len++;
370 	}
371 
372 	/*
373 	 * Be careful not to send data and/or FIN on SYN segments.
374 	 * This measure is needed to prevent interoperability problems
375 	 * with not fully conformant TCP implementations.
376 	 */
377 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
378 		len = 0;
379 		flags &= ~TH_FIN;
380 	}
381 
382 	if (len < 0) {
383 		/*
384 		 * If FIN has been sent but not acked,
385 		 * but we haven't been called to retransmit,
386 		 * len will be < 0.  Otherwise, window shrank
387 		 * after we sent into it.  If window shrank to 0,
388 		 * cancel pending retransmit, pull snd_nxt back
389 		 * to (closed) window, and set the persist timer
390 		 * if it isn't already going.  If the window didn't
391 		 * close completely, just wait for an ACK.
392 		 */
393 		len = 0;
394 		if (sendwin == 0) {
395 			tcp_timer_activate(tp, TT_REXMT, 0);
396 			tp->t_rxtshift = 0;
397 			tp->snd_nxt = tp->snd_una;
398 			if (!tcp_timer_active(tp, TT_PERSIST))
399 				tcp_setpersist(tp);
400 		}
401 	}
402 
403 	/* len will be >= 0 after this point. */
404 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
405 
406 	/*
407 	 * Automatic sizing of send socket buffer.  Often the send buffer
408 	 * size is not optimally adjusted to the actual network conditions
409 	 * at hand (delay bandwidth product).  Setting the buffer size too
410 	 * small limits throughput on links with high bandwidth and high
411 	 * delay (eg. trans-continental/oceanic links).  Setting the
412 	 * buffer size too big consumes too much real kernel memory,
413 	 * especially with many connections on busy servers.
414 	 *
415 	 * The criteria to step up the send buffer one notch are:
416 	 *  1. receive window of remote host is larger than send buffer
417 	 *     (with a fudge factor of 5/4th);
418 	 *  2. send buffer is filled to 7/8th with data (so we actually
419 	 *     have data to make use of it);
420 	 *  3. send buffer fill has not hit maximal automatic size;
421 	 *  4. our send window (slow start and cogestion controlled) is
422 	 *     larger than sent but unacknowledged data in send buffer.
423 	 *
424 	 * The remote host receive window scaling factor may limit the
425 	 * growing of the send buffer before it reaches its allowed
426 	 * maximum.
427 	 *
428 	 * It scales directly with slow start or congestion window
429 	 * and does at most one step per received ACK.  This fast
430 	 * scaling has the drawback of growing the send buffer beyond
431 	 * what is strictly necessary to make full use of a given
432 	 * delay*bandwith product.  However testing has shown this not
433 	 * to be much of an problem.  At worst we are trading wasting
434 	 * of available bandwith (the non-use of it) for wasting some
435 	 * socket buffer memory.
436 	 *
437 	 * TODO: Shrink send buffer during idle periods together
438 	 * with congestion window.  Requires another timer.  Has to
439 	 * wait for upcoming tcp timer rewrite.
440 	 */
441 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
442 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
443 		    so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
444 		    so->so_snd.sb_cc < V_tcp_autosndbuf_max &&
445 		    sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
446 			if (!sbreserve_locked(&so->so_snd,
447 			    min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
448 			     V_tcp_autosndbuf_max), so, curthread))
449 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
450 		}
451 	}
452 
453 	/*
454 	 * Truncate to the maximum segment length or enable TCP Segmentation
455 	 * Offloading (if supported by hardware) and ensure that FIN is removed
456 	 * if the length no longer contains the last data byte.
457 	 *
458 	 * TSO may only be used if we are in a pure bulk sending state.  The
459 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
460 	 * IP options prevent using TSO.  With TSO the TCP header is the same
461 	 * (except for the sequence number) for all generated packets.  This
462 	 * makes it impossible to transmit any options which vary per generated
463 	 * segment or packet.
464 	 *
465 	 * The length of TSO bursts is limited to TCP_MAXWIN.  That limit and
466 	 * removal of FIN (if not already catched here) are handled later after
467 	 * the exact length of the TCP options are known.
468 	 */
469 #ifdef IPSEC
470 	/*
471 	 * Pre-calculate here as we save another lookup into the darknesses
472 	 * of IPsec that way and can actually decide if TSO is ok.
473 	 */
474 	ipsec_optlen = ipsec_hdrsiz_tcp(tp);
475 #endif
476 	if (len > tp->t_maxseg) {
477 		if ((tp->t_flags & TF_TSO) && V_tcp_do_tso &&
478 		    ((tp->t_flags & TF_SIGNATURE) == 0) &&
479 		    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
480 		    tp->t_inpcb->inp_options == NULL &&
481 		    tp->t_inpcb->in6p_options == NULL
482 #ifdef IPSEC
483 		    && ipsec_optlen == 0
484 #endif
485 		    ) {
486 			tso = 1;
487 		} else {
488 			len = tp->t_maxseg;
489 			sendalot = 1;
490 			tso = 0;
491 		}
492 	}
493 	if (sack_rxmit) {
494 		if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
495 			flags &= ~TH_FIN;
496 	} else {
497 		if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
498 			flags &= ~TH_FIN;
499 	}
500 
501 	recwin = sbspace(&so->so_rcv);
502 
503 	/*
504 	 * Sender silly window avoidance.   We transmit under the following
505 	 * conditions when len is non-zero:
506 	 *
507 	 *	- We have a full segment (or more with TSO)
508 	 *	- This is the last buffer in a write()/send() and we are
509 	 *	  either idle or running NODELAY
510 	 *	- we've timed out (e.g. persist timer)
511 	 *	- we have more then 1/2 the maximum send window's worth of
512 	 *	  data (receiver may be limited the window size)
513 	 *	- we need to retransmit
514 	 */
515 	if (len) {
516 		if (len >= tp->t_maxseg)
517 			goto send;
518 		/*
519 		 * NOTE! on localhost connections an 'ack' from the remote
520 		 * end may occur synchronously with the output and cause
521 		 * us to flush a buffer queued with moretocome.  XXX
522 		 *
523 		 * note: the len + off check is almost certainly unnecessary.
524 		 */
525 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
526 		    (idle || (tp->t_flags & TF_NODELAY)) &&
527 		    len + off >= so->so_snd.sb_cc &&
528 		    (tp->t_flags & TF_NOPUSH) == 0) {
529 			goto send;
530 		}
531 		if (tp->t_flags & TF_FORCEDATA)		/* typ. timeout case */
532 			goto send;
533 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
534 			goto send;
535 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
536 			goto send;
537 		if (sack_rxmit)
538 			goto send;
539 	}
540 
541 	/*
542 	 * Compare available window to amount of window
543 	 * known to peer (as advertised window less
544 	 * next expected input).  If the difference is at least two
545 	 * max size segments, or at least 50% of the maximum possible
546 	 * window, then want to send a window update to peer.
547 	 * Skip this if the connection is in T/TCP half-open state.
548 	 * Don't send pure window updates when the peer has closed
549 	 * the connection and won't ever send more data.
550 	 */
551 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
552 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
553 		/*
554 		 * "adv" is the amount we can increase the window,
555 		 * taking into account that we are limited by
556 		 * TCP_MAXWIN << tp->rcv_scale.
557 		 */
558 		long adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale) -
559 			(tp->rcv_adv - tp->rcv_nxt);
560 
561 		if (adv >= (long) (2 * tp->t_maxseg))
562 			goto send;
563 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
564 			goto send;
565 	}
566 
567 	/*
568 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
569 	 * is also a catch-all for the retransmit timer timeout case.
570 	 */
571 	if (tp->t_flags & TF_ACKNOW)
572 		goto send;
573 	if ((flags & TH_RST) ||
574 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
575 		goto send;
576 	if (SEQ_GT(tp->snd_up, tp->snd_una))
577 		goto send;
578 	/*
579 	 * If our state indicates that FIN should be sent
580 	 * and we have not yet done so, then we need to send.
581 	 */
582 	if (flags & TH_FIN &&
583 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
584 		goto send;
585 	/*
586 	 * In SACK, it is possible for tcp_output to fail to send a segment
587 	 * after the retransmission timer has been turned off.  Make sure
588 	 * that the retransmission timer is set.
589 	 */
590 	if ((tp->t_flags & TF_SACK_PERMIT) &&
591 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
592 	    !tcp_timer_active(tp, TT_REXMT) &&
593 	    !tcp_timer_active(tp, TT_PERSIST)) {
594 		tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
595 		goto just_return;
596 	}
597 	/*
598 	 * TCP window updates are not reliable, rather a polling protocol
599 	 * using ``persist'' packets is used to insure receipt of window
600 	 * updates.  The three ``states'' for the output side are:
601 	 *	idle			not doing retransmits or persists
602 	 *	persisting		to move a small or zero window
603 	 *	(re)transmitting	and thereby not persisting
604 	 *
605 	 * tcp_timer_active(tp, TT_PERSIST)
606 	 *	is true when we are in persist state.
607 	 * (tp->t_flags & TF_FORCEDATA)
608 	 *	is set when we are called to send a persist packet.
609 	 * tcp_timer_active(tp, TT_REXMT)
610 	 *	is set when we are retransmitting
611 	 * The output side is idle when both timers are zero.
612 	 *
613 	 * If send window is too small, there is data to transmit, and no
614 	 * retransmit or persist is pending, then go to persist state.
615 	 * If nothing happens soon, send when timer expires:
616 	 * if window is nonzero, transmit what we can,
617 	 * otherwise force out a byte.
618 	 */
619 	if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) &&
620 	    !tcp_timer_active(tp, TT_PERSIST)) {
621 		tp->t_rxtshift = 0;
622 		tcp_setpersist(tp);
623 	}
624 
625 	/*
626 	 * No reason to send a segment, just return.
627 	 */
628 just_return:
629 	SOCKBUF_UNLOCK(&so->so_snd);
630 	return (0);
631 
632 send:
633 	SOCKBUF_LOCK_ASSERT(&so->so_snd);
634 	/*
635 	 * Before ESTABLISHED, force sending of initial options
636 	 * unless TCP set not to do any options.
637 	 * NOTE: we assume that the IP/TCP header plus TCP options
638 	 * always fit in a single mbuf, leaving room for a maximum
639 	 * link header, i.e.
640 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
641 	 */
642 	optlen = 0;
643 #ifdef INET6
644 	if (isipv6)
645 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
646 	else
647 #endif
648 	hdrlen = sizeof (struct tcpiphdr);
649 
650 	/*
651 	 * Compute options for segment.
652 	 * We only have to care about SYN and established connection
653 	 * segments.  Options for SYN-ACK segments are handled in TCP
654 	 * syncache.
655 	 */
656 	if ((tp->t_flags & TF_NOOPT) == 0) {
657 		to.to_flags = 0;
658 		/* Maximum segment size. */
659 		if (flags & TH_SYN) {
660 			tp->snd_nxt = tp->iss;
661 			to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
662 			to.to_flags |= TOF_MSS;
663 		}
664 		/* Window scaling. */
665 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
666 			to.to_wscale = tp->request_r_scale;
667 			to.to_flags |= TOF_SCALE;
668 		}
669 		/* Timestamps. */
670 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
671 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
672 			to.to_tsval = ticks + tp->ts_offset;
673 			to.to_tsecr = tp->ts_recent;
674 			to.to_flags |= TOF_TS;
675 			/* Set receive buffer autosizing timestamp. */
676 			if (tp->rfbuf_ts == 0 &&
677 			    (so->so_rcv.sb_flags & SB_AUTOSIZE))
678 				tp->rfbuf_ts = ticks;
679 		}
680 		/* Selective ACK's. */
681 		if (tp->t_flags & TF_SACK_PERMIT) {
682 			if (flags & TH_SYN)
683 				to.to_flags |= TOF_SACKPERM;
684 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
685 			    (tp->t_flags & TF_SACK_PERMIT) &&
686 			    tp->rcv_numsacks > 0) {
687 				to.to_flags |= TOF_SACK;
688 				to.to_nsacks = tp->rcv_numsacks;
689 				to.to_sacks = (u_char *)tp->sackblks;
690 			}
691 		}
692 #ifdef TCP_SIGNATURE
693 		/* TCP-MD5 (RFC2385). */
694 		if (tp->t_flags & TF_SIGNATURE)
695 			to.to_flags |= TOF_SIGNATURE;
696 #endif /* TCP_SIGNATURE */
697 
698 		/* Processing the options. */
699 		hdrlen += optlen = tcp_addoptions(&to, opt);
700 	}
701 
702 #ifdef INET6
703 	if (isipv6)
704 		ipoptlen = ip6_optlen(tp->t_inpcb);
705 	else
706 #endif
707 	if (tp->t_inpcb->inp_options)
708 		ipoptlen = tp->t_inpcb->inp_options->m_len -
709 				offsetof(struct ipoption, ipopt_list);
710 	else
711 		ipoptlen = 0;
712 #ifdef IPSEC
713 	ipoptlen += ipsec_optlen;
714 #endif
715 
716 	/*
717 	 * Adjust data length if insertion of options will
718 	 * bump the packet length beyond the t_maxopd length.
719 	 * Clear the FIN bit because we cut off the tail of
720 	 * the segment.
721 	 *
722 	 * When doing TSO limit a burst to TCP_MAXWIN minus the
723 	 * IP, TCP and Options length to keep ip->ip_len from
724 	 * overflowing.  Prevent the last segment from being
725 	 * fractional thus making them all equal sized and set
726 	 * the flag to continue sending.  TSO is disabled when
727 	 * IP options or IPSEC are present.
728 	 */
729 	if (len + optlen + ipoptlen > tp->t_maxopd) {
730 		flags &= ~TH_FIN;
731 		if (tso) {
732 			if (len > TCP_MAXWIN - hdrlen - optlen) {
733 				len = TCP_MAXWIN - hdrlen - optlen;
734 				len = len - (len % (tp->t_maxopd - optlen));
735 				sendalot = 1;
736 			} else if (tp->t_flags & TF_NEEDFIN)
737 				sendalot = 1;
738 		} else {
739 			len = tp->t_maxopd - optlen - ipoptlen;
740 			sendalot = 1;
741 		}
742 	}
743 
744 /*#ifdef DIAGNOSTIC*/
745 #ifdef INET6
746 	if (max_linkhdr + hdrlen > MCLBYTES)
747 #else
748 	if (max_linkhdr + hdrlen > MHLEN)
749 #endif
750 		panic("tcphdr too big");
751 /*#endif*/
752 
753 	/*
754 	 * This KASSERT is here to catch edge cases at a well defined place.
755 	 * Before, those had triggered (random) panic conditions further down.
756 	 */
757 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
758 
759 	/*
760 	 * Grab a header mbuf, attaching a copy of data to
761 	 * be transmitted, and initialize the header from
762 	 * the template for sends on this connection.
763 	 */
764 	if (len) {
765 		struct mbuf *mb;
766 		u_int moff;
767 
768 		if ((tp->t_flags & TF_FORCEDATA) && len == 1)
769 			TCPSTAT_INC(tcps_sndprobe);
770 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
771 			TCPSTAT_INC(tcps_sndrexmitpack);
772 			TCPSTAT_ADD(tcps_sndrexmitbyte, len);
773 		} else {
774 			TCPSTAT_INC(tcps_sndpack);
775 			TCPSTAT_ADD(tcps_sndbyte, len);
776 		}
777 #ifdef notyet
778 		if ((m = m_copypack(so->so_snd.sb_mb, off,
779 		    (int)len, max_linkhdr + hdrlen)) == 0) {
780 			SOCKBUF_UNLOCK(&so->so_snd);
781 			error = ENOBUFS;
782 			goto out;
783 		}
784 		/*
785 		 * m_copypack left space for our hdr; use it.
786 		 */
787 		m->m_len += hdrlen;
788 		m->m_data -= hdrlen;
789 #else
790 		MGETHDR(m, M_DONTWAIT, MT_DATA);
791 		if (m == NULL) {
792 			SOCKBUF_UNLOCK(&so->so_snd);
793 			error = ENOBUFS;
794 			goto out;
795 		}
796 #ifdef INET6
797 		if (MHLEN < hdrlen + max_linkhdr) {
798 			MCLGET(m, M_DONTWAIT);
799 			if ((m->m_flags & M_EXT) == 0) {
800 				SOCKBUF_UNLOCK(&so->so_snd);
801 				m_freem(m);
802 				error = ENOBUFS;
803 				goto out;
804 			}
805 		}
806 #endif
807 		m->m_data += max_linkhdr;
808 		m->m_len = hdrlen;
809 
810 		/*
811 		 * Start the m_copy functions from the closest mbuf
812 		 * to the offset in the socket buffer chain.
813 		 */
814 		mb = sbsndptr(&so->so_snd, off, len, &moff);
815 
816 		if (len <= MHLEN - hdrlen - max_linkhdr) {
817 			m_copydata(mb, moff, (int)len,
818 			    mtod(m, caddr_t) + hdrlen);
819 			m->m_len += len;
820 		} else {
821 			m->m_next = m_copy(mb, moff, (int)len);
822 			if (m->m_next == NULL) {
823 				SOCKBUF_UNLOCK(&so->so_snd);
824 				(void) m_free(m);
825 				error = ENOBUFS;
826 				goto out;
827 			}
828 		}
829 #endif
830 		/*
831 		 * If we're sending everything we've got, set PUSH.
832 		 * (This will keep happy those implementations which only
833 		 * give data to the user when a buffer fills or
834 		 * a PUSH comes in.)
835 		 */
836 		if (off + len == so->so_snd.sb_cc)
837 			flags |= TH_PUSH;
838 		SOCKBUF_UNLOCK(&so->so_snd);
839 	} else {
840 		SOCKBUF_UNLOCK(&so->so_snd);
841 		if (tp->t_flags & TF_ACKNOW)
842 			TCPSTAT_INC(tcps_sndacks);
843 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
844 			TCPSTAT_INC(tcps_sndctrl);
845 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
846 			TCPSTAT_INC(tcps_sndurg);
847 		else
848 			TCPSTAT_INC(tcps_sndwinup);
849 
850 		MGETHDR(m, M_DONTWAIT, MT_DATA);
851 		if (m == NULL) {
852 			error = ENOBUFS;
853 			goto out;
854 		}
855 #ifdef INET6
856 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
857 		    MHLEN >= hdrlen) {
858 			MH_ALIGN(m, hdrlen);
859 		} else
860 #endif
861 		m->m_data += max_linkhdr;
862 		m->m_len = hdrlen;
863 	}
864 	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
865 	m->m_pkthdr.rcvif = (struct ifnet *)0;
866 #ifdef MAC
867 	mac_inpcb_create_mbuf(tp->t_inpcb, m);
868 #endif
869 #ifdef INET6
870 	if (isipv6) {
871 		ip6 = mtod(m, struct ip6_hdr *);
872 		th = (struct tcphdr *)(ip6 + 1);
873 		tcpip_fillheaders(tp->t_inpcb, ip6, th);
874 	} else
875 #endif /* INET6 */
876 	{
877 		ip = mtod(m, struct ip *);
878 		ipov = (struct ipovly *)ip;
879 		th = (struct tcphdr *)(ip + 1);
880 		tcpip_fillheaders(tp->t_inpcb, ip, th);
881 	}
882 
883 	/*
884 	 * Fill in fields, remembering maximum advertised
885 	 * window for use in delaying messages about window sizes.
886 	 * If resending a FIN, be sure not to use a new sequence number.
887 	 */
888 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
889 	    tp->snd_nxt == tp->snd_max)
890 		tp->snd_nxt--;
891 	/*
892 	 * If we are starting a connection, send ECN setup
893 	 * SYN packet. If we are on a retransmit, we may
894 	 * resend those bits a number of times as per
895 	 * RFC 3168.
896 	 */
897 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
898 		if (tp->t_rxtshift >= 1) {
899 			if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
900 				flags |= TH_ECE|TH_CWR;
901 		} else
902 			flags |= TH_ECE|TH_CWR;
903 	}
904 
905 	if (tp->t_state == TCPS_ESTABLISHED &&
906 	    (tp->t_flags & TF_ECN_PERMIT)) {
907 		/*
908 		 * If the peer has ECN, mark data packets with
909 		 * ECN capable transmission (ECT).
910 		 * Ignore pure ack packets, retransmissions and window probes.
911 		 */
912 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
913 		    !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
914 #ifdef INET6
915 			if (isipv6)
916 				ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
917 			else
918 #endif
919 				ip->ip_tos |= IPTOS_ECN_ECT0;
920 			TCPSTAT_INC(tcps_ecn_ect0);
921 		}
922 
923 		/*
924 		 * Reply with proper ECN notifications.
925 		 */
926 		if (tp->t_flags & TF_ECN_SND_CWR) {
927 			flags |= TH_CWR;
928 			tp->t_flags &= ~TF_ECN_SND_CWR;
929 		}
930 		if (tp->t_flags & TF_ECN_SND_ECE)
931 			flags |= TH_ECE;
932 	}
933 
934 	/*
935 	 * If we are doing retransmissions, then snd_nxt will
936 	 * not reflect the first unsent octet.  For ACK only
937 	 * packets, we do not want the sequence number of the
938 	 * retransmitted packet, we want the sequence number
939 	 * of the next unsent octet.  So, if there is no data
940 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
941 	 * when filling in ti_seq.  But if we are in persist
942 	 * state, snd_max might reflect one byte beyond the
943 	 * right edge of the window, so use snd_nxt in that
944 	 * case, since we know we aren't doing a retransmission.
945 	 * (retransmit and persist are mutually exclusive...)
946 	 */
947 	if (sack_rxmit == 0) {
948 		if (len || (flags & (TH_SYN|TH_FIN)) ||
949 		    tcp_timer_active(tp, TT_PERSIST))
950 			th->th_seq = htonl(tp->snd_nxt);
951 		else
952 			th->th_seq = htonl(tp->snd_max);
953 	} else {
954 		th->th_seq = htonl(p->rxmit);
955 		p->rxmit += len;
956 		tp->sackhint.sack_bytes_rexmit += len;
957 	}
958 	th->th_ack = htonl(tp->rcv_nxt);
959 	if (optlen) {
960 		bcopy(opt, th + 1, optlen);
961 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
962 	}
963 	th->th_flags = flags;
964 	/*
965 	 * Calculate receive window.  Don't shrink window,
966 	 * but avoid silly window syndrome.
967 	 */
968 	if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
969 	    recwin < (long)tp->t_maxseg)
970 		recwin = 0;
971 	if (recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
972 		recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
973 	if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
974 		recwin = (long)TCP_MAXWIN << tp->rcv_scale;
975 
976 	/*
977 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
978 	 * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
979 	 * case is handled in syncache.
980 	 */
981 	if (flags & TH_SYN)
982 		th->th_win = htons((u_short)
983 				(min(sbspace(&so->so_rcv), TCP_MAXWIN)));
984 	else
985 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
986 
987 	/*
988 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
989 	 * a 0 window.  This may cause the remote transmitter to stall.  This
990 	 * flag tells soreceive() to disable delayed acknowledgements when
991 	 * draining the buffer.  This can occur if the receiver is attempting
992 	 * to read more data than can be buffered prior to transmitting on
993 	 * the connection.
994 	 */
995 	if (recwin == 0)
996 		tp->t_flags |= TF_RXWIN0SENT;
997 	else
998 		tp->t_flags &= ~TF_RXWIN0SENT;
999 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1000 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1001 		th->th_flags |= TH_URG;
1002 	} else
1003 		/*
1004 		 * If no urgent pointer to send, then we pull
1005 		 * the urgent pointer to the left edge of the send window
1006 		 * so that it doesn't drift into the send window on sequence
1007 		 * number wraparound.
1008 		 */
1009 		tp->snd_up = tp->snd_una;		/* drag it along */
1010 
1011 #ifdef TCP_SIGNATURE
1012 	if (tp->t_flags & TF_SIGNATURE) {
1013 		int sigoff = to.to_signature - opt;
1014 		tcp_signature_compute(m, 0, len, optlen,
1015 		    (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
1016 	}
1017 #endif
1018 
1019 	/*
1020 	 * Put TCP length in extended header, and then
1021 	 * checksum extended header and data.
1022 	 */
1023 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1024 #ifdef INET6
1025 	if (isipv6)
1026 		/*
1027 		 * ip6_plen is not need to be filled now, and will be filled
1028 		 * in ip6_output.
1029 		 */
1030 		th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
1031 				       sizeof(struct tcphdr) + optlen + len);
1032 	else
1033 #endif /* INET6 */
1034 	{
1035 		m->m_pkthdr.csum_flags = CSUM_TCP;
1036 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1037 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1038 		    htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1039 
1040 		/* IP version must be set here for ipv4/ipv6 checking later */
1041 		KASSERT(ip->ip_v == IPVERSION,
1042 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1043 	}
1044 
1045 	/*
1046 	 * Enable TSO and specify the size of the segments.
1047 	 * The TCP pseudo header checksum is always provided.
1048 	 * XXX: Fixme: This is currently not the case for IPv6.
1049 	 */
1050 	if (tso) {
1051 		m->m_pkthdr.csum_flags = CSUM_TSO;
1052 		m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen;
1053 	}
1054 
1055 	/*
1056 	 * In transmit state, time the transmission and arrange for
1057 	 * the retransmit.  In persist state, just set snd_max.
1058 	 */
1059 	if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1060 	    !tcp_timer_active(tp, TT_PERSIST)) {
1061 		tcp_seq startseq = tp->snd_nxt;
1062 
1063 		/*
1064 		 * Advance snd_nxt over sequence space of this segment.
1065 		 */
1066 		if (flags & (TH_SYN|TH_FIN)) {
1067 			if (flags & TH_SYN)
1068 				tp->snd_nxt++;
1069 			if (flags & TH_FIN) {
1070 				tp->snd_nxt++;
1071 				tp->t_flags |= TF_SENTFIN;
1072 			}
1073 		}
1074 		if (sack_rxmit)
1075 			goto timer;
1076 		tp->snd_nxt += len;
1077 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1078 			tp->snd_max = tp->snd_nxt;
1079 			/*
1080 			 * Time this transmission if not a retransmission and
1081 			 * not currently timing anything.
1082 			 */
1083 			if (tp->t_rtttime == 0) {
1084 				tp->t_rtttime = ticks;
1085 				tp->t_rtseq = startseq;
1086 				TCPSTAT_INC(tcps_segstimed);
1087 			}
1088 		}
1089 
1090 		/*
1091 		 * Set retransmit timer if not currently set,
1092 		 * and not doing a pure ack or a keep-alive probe.
1093 		 * Initial value for retransmit timer is smoothed
1094 		 * round-trip time + 2 * round-trip time variance.
1095 		 * Initialize shift counter which is used for backoff
1096 		 * of retransmit time.
1097 		 */
1098 timer:
1099 		if (!tcp_timer_active(tp, TT_REXMT) &&
1100 		    ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1101 		     (tp->snd_nxt != tp->snd_una))) {
1102 			if (tcp_timer_active(tp, TT_PERSIST)) {
1103 				tcp_timer_activate(tp, TT_PERSIST, 0);
1104 				tp->t_rxtshift = 0;
1105 			}
1106 			tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1107 		}
1108 	} else {
1109 		/*
1110 		 * Persist case, update snd_max but since we are in
1111 		 * persist mode (no window) we do not update snd_nxt.
1112 		 */
1113 		int xlen = len;
1114 		if (flags & TH_SYN)
1115 			++xlen;
1116 		if (flags & TH_FIN) {
1117 			++xlen;
1118 			tp->t_flags |= TF_SENTFIN;
1119 		}
1120 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1121 			tp->snd_max = tp->snd_nxt + len;
1122 	}
1123 
1124 #ifdef TCPDEBUG
1125 	/*
1126 	 * Trace.
1127 	 */
1128 	if (so->so_options & SO_DEBUG) {
1129 		u_short save = 0;
1130 #ifdef INET6
1131 		if (!isipv6)
1132 #endif
1133 		{
1134 			save = ipov->ih_len;
1135 			ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1136 		}
1137 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1138 #ifdef INET6
1139 		if (!isipv6)
1140 #endif
1141 		ipov->ih_len = save;
1142 	}
1143 #endif
1144 
1145 	/*
1146 	 * Fill in IP length and desired time to live and
1147 	 * send to IP level.  There should be a better way
1148 	 * to handle ttl and tos; we could keep them in
1149 	 * the template, but need a way to checksum without them.
1150 	 */
1151 	/*
1152 	 * m->m_pkthdr.len should have been set before cksum calcuration,
1153 	 * because in6_cksum() need it.
1154 	 */
1155 #ifdef INET6
1156 	if (isipv6) {
1157 		/*
1158 		 * we separately set hoplimit for every segment, since the
1159 		 * user might want to change the value via setsockopt.
1160 		 * Also, desired default hop limit might be changed via
1161 		 * Neighbor Discovery.
1162 		 */
1163 		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1164 
1165 		/* TODO: IPv6 IP6TOS_ECT bit on */
1166 		error = ip6_output(m,
1167 			    tp->t_inpcb->in6p_outputopts, NULL,
1168 			    ((so->so_options & SO_DONTROUTE) ?
1169 			    IP_ROUTETOIF : 0), NULL, NULL, tp->t_inpcb);
1170 	} else
1171 #endif /* INET6 */
1172     {
1173 	ip->ip_len = m->m_pkthdr.len;
1174 #ifdef INET6
1175 	if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1176 		ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1177 #endif /* INET6 */
1178 	/*
1179 	 * If we do path MTU discovery, then we set DF on every packet.
1180 	 * This might not be the best thing to do according to RFC3390
1181 	 * Section 2. However the tcp hostcache migitates the problem
1182 	 * so it affects only the first tcp connection with a host.
1183 	 */
1184 	if (V_path_mtu_discovery)
1185 		ip->ip_off |= IP_DF;
1186 
1187 	error = ip_output(m, tp->t_inpcb->inp_options, NULL,
1188 	    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1189 	    tp->t_inpcb);
1190     }
1191 	if (error) {
1192 
1193 		/*
1194 		 * We know that the packet was lost, so back out the
1195 		 * sequence number advance, if any.
1196 		 *
1197 		 * If the error is EPERM the packet got blocked by the
1198 		 * local firewall.  Normally we should terminate the
1199 		 * connection but the blocking may have been spurious
1200 		 * due to a firewall reconfiguration cycle.  So we treat
1201 		 * it like a packet loss and let the retransmit timer and
1202 		 * timeouts do their work over time.
1203 		 * XXX: It is a POLA question whether calling tcp_drop right
1204 		 * away would be the really correct behavior instead.
1205 		 */
1206 		if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1207 		    !tcp_timer_active(tp, TT_PERSIST)) &&
1208 		    ((flags & TH_SYN) == 0) &&
1209 		    (error != EPERM)) {
1210 			if (sack_rxmit) {
1211 				p->rxmit -= len;
1212 				tp->sackhint.sack_bytes_rexmit -= len;
1213 				KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1214 				    ("sackhint bytes rtx >= 0"));
1215 			} else
1216 				tp->snd_nxt -= len;
1217 		}
1218 out:
1219 		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1220 		switch (error) {
1221 		case EPERM:
1222 			tp->t_softerror = error;
1223 			return (error);
1224 		case ENOBUFS:
1225 	                if (!tcp_timer_active(tp, TT_REXMT) &&
1226 			    !tcp_timer_active(tp, TT_PERSIST))
1227 	                        tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1228 			tp->snd_cwnd = tp->t_maxseg;
1229 			return (0);
1230 		case EMSGSIZE:
1231 			/*
1232 			 * For some reason the interface we used initially
1233 			 * to send segments changed to another or lowered
1234 			 * its MTU.
1235 			 *
1236 			 * tcp_mtudisc() will find out the new MTU and as
1237 			 * its last action, initiate retransmission, so it
1238 			 * is important to not do so here.
1239 			 *
1240 			 * If TSO was active we either got an interface
1241 			 * without TSO capabilits or TSO was turned off.
1242 			 * Disable it for this connection as too and
1243 			 * immediatly retry with MSS sized segments generated
1244 			 * by this function.
1245 			 */
1246 			if (tso)
1247 				tp->t_flags &= ~TF_TSO;
1248 			tcp_mtudisc(tp->t_inpcb, 0);
1249 			return (0);
1250 		case EHOSTDOWN:
1251 		case EHOSTUNREACH:
1252 		case ENETDOWN:
1253 		case ENETUNREACH:
1254 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
1255 				tp->t_softerror = error;
1256 				return (0);
1257 			}
1258 			/* FALLTHROUGH */
1259 		default:
1260 			return (error);
1261 		}
1262 	}
1263 	TCPSTAT_INC(tcps_sndtotal);
1264 
1265 	/*
1266 	 * Data sent (as far as we can tell).
1267 	 * If this advertises a larger window than any other segment,
1268 	 * then remember the size of the advertised window.
1269 	 * Any pending ACK has now been sent.
1270 	 */
1271 	if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1272 		tp->rcv_adv = tp->rcv_nxt + recwin;
1273 	tp->last_ack_sent = tp->rcv_nxt;
1274 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1275 	if (tcp_timer_active(tp, TT_DELACK))
1276 		tcp_timer_activate(tp, TT_DELACK, 0);
1277 #if 0
1278 	/*
1279 	 * This completely breaks TCP if newreno is turned on.  What happens
1280 	 * is that if delayed-acks are turned on on the receiver, this code
1281 	 * on the transmitter effectively destroys the TCP window, forcing
1282 	 * it to four packets (1.5Kx4 = 6K window).
1283 	 */
1284 	if (sendalot && (!V_tcp_do_newreno || --maxburst))
1285 		goto again;
1286 #endif
1287 	if (sendalot)
1288 		goto again;
1289 	return (0);
1290 }
1291 
1292 void
1293 tcp_setpersist(struct tcpcb *tp)
1294 {
1295 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1296 	int tt;
1297 
1298 	if (tcp_timer_active(tp, TT_REXMT))
1299 		panic("tcp_setpersist: retransmit pending");
1300 	/*
1301 	 * Start/restart persistance timer.
1302 	 */
1303 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1304 		      TCPTV_PERSMIN, TCPTV_PERSMAX);
1305 	tcp_timer_activate(tp, TT_PERSIST, tt);
1306 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1307 		tp->t_rxtshift++;
1308 }
1309 
1310 /*
1311  * Insert TCP options according to the supplied parameters to the place
1312  * optp in a consistent way.  Can handle unaligned destinations.
1313  *
1314  * The order of the option processing is crucial for optimal packing and
1315  * alignment for the scarce option space.
1316  *
1317  * The optimal order for a SYN/SYN-ACK segment is:
1318  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1319  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1320  *
1321  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1322  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1323  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1324  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1325  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1326  */
1327 int
1328 tcp_addoptions(struct tcpopt *to, u_char *optp)
1329 {
1330 	u_int mask, optlen = 0;
1331 
1332 	for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1333 		if ((to->to_flags & mask) != mask)
1334 			continue;
1335 		if (optlen == TCP_MAXOLEN)
1336 			break;
1337 		switch (to->to_flags & mask) {
1338 		case TOF_MSS:
1339 			while (optlen % 4) {
1340 				optlen += TCPOLEN_NOP;
1341 				*optp++ = TCPOPT_NOP;
1342 			}
1343 			if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1344 				continue;
1345 			optlen += TCPOLEN_MAXSEG;
1346 			*optp++ = TCPOPT_MAXSEG;
1347 			*optp++ = TCPOLEN_MAXSEG;
1348 			to->to_mss = htons(to->to_mss);
1349 			bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1350 			optp += sizeof(to->to_mss);
1351 			break;
1352 		case TOF_SCALE:
1353 			while (!optlen || optlen % 2 != 1) {
1354 				optlen += TCPOLEN_NOP;
1355 				*optp++ = TCPOPT_NOP;
1356 			}
1357 			if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1358 				continue;
1359 			optlen += TCPOLEN_WINDOW;
1360 			*optp++ = TCPOPT_WINDOW;
1361 			*optp++ = TCPOLEN_WINDOW;
1362 			*optp++ = to->to_wscale;
1363 			break;
1364 		case TOF_SACKPERM:
1365 			while (optlen % 2) {
1366 				optlen += TCPOLEN_NOP;
1367 				*optp++ = TCPOPT_NOP;
1368 			}
1369 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1370 				continue;
1371 			optlen += TCPOLEN_SACK_PERMITTED;
1372 			*optp++ = TCPOPT_SACK_PERMITTED;
1373 			*optp++ = TCPOLEN_SACK_PERMITTED;
1374 			break;
1375 		case TOF_TS:
1376 			while (!optlen || optlen % 4 != 2) {
1377 				optlen += TCPOLEN_NOP;
1378 				*optp++ = TCPOPT_NOP;
1379 			}
1380 			if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1381 				continue;
1382 			optlen += TCPOLEN_TIMESTAMP;
1383 			*optp++ = TCPOPT_TIMESTAMP;
1384 			*optp++ = TCPOLEN_TIMESTAMP;
1385 			to->to_tsval = htonl(to->to_tsval);
1386 			to->to_tsecr = htonl(to->to_tsecr);
1387 			bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1388 			optp += sizeof(to->to_tsval);
1389 			bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1390 			optp += sizeof(to->to_tsecr);
1391 			break;
1392 		case TOF_SIGNATURE:
1393 			{
1394 			int siglen = TCPOLEN_SIGNATURE - 2;
1395 
1396 			while (!optlen || optlen % 4 != 2) {
1397 				optlen += TCPOLEN_NOP;
1398 				*optp++ = TCPOPT_NOP;
1399 			}
1400 			if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE)
1401 				continue;
1402 			optlen += TCPOLEN_SIGNATURE;
1403 			*optp++ = TCPOPT_SIGNATURE;
1404 			*optp++ = TCPOLEN_SIGNATURE;
1405 			to->to_signature = optp;
1406 			while (siglen--)
1407 				 *optp++ = 0;
1408 			break;
1409 			}
1410 		case TOF_SACK:
1411 			{
1412 			int sackblks = 0;
1413 			struct sackblk *sack = (struct sackblk *)to->to_sacks;
1414 			tcp_seq sack_seq;
1415 
1416 			while (!optlen || optlen % 4 != 2) {
1417 				optlen += TCPOLEN_NOP;
1418 				*optp++ = TCPOPT_NOP;
1419 			}
1420 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1421 				continue;
1422 			optlen += TCPOLEN_SACKHDR;
1423 			*optp++ = TCPOPT_SACK;
1424 			sackblks = min(to->to_nsacks,
1425 					(TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1426 			*optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1427 			while (sackblks--) {
1428 				sack_seq = htonl(sack->start);
1429 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1430 				optp += sizeof(sack_seq);
1431 				sack_seq = htonl(sack->end);
1432 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1433 				optp += sizeof(sack_seq);
1434 				optlen += TCPOLEN_SACK;
1435 				sack++;
1436 			}
1437 			TCPSTAT_INC(tcps_sack_send_blocks);
1438 			break;
1439 			}
1440 		default:
1441 			panic("%s: unknown TCP option type", __func__);
1442 			break;
1443 		}
1444 	}
1445 
1446 	/* Terminate and pad TCP options to a 4 byte boundary. */
1447 	if (optlen % 4) {
1448 		optlen += TCPOLEN_EOL;
1449 		*optp++ = TCPOPT_EOL;
1450 	}
1451 	/*
1452 	 * According to RFC 793 (STD0007):
1453 	 *   "The content of the header beyond the End-of-Option option
1454 	 *    must be header padding (i.e., zero)."
1455 	 *   and later: "The padding is composed of zeros."
1456 	 */
1457 	while (optlen % 4) {
1458 		optlen += TCPOLEN_PAD;
1459 		*optp++ = TCPOPT_PAD;
1460 	}
1461 
1462 	KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1463 	return (optlen);
1464 }
1465