xref: /freebsd/sys/netinet/tcp_output.c (revision a3cf0ef5a295c885c895fabfd56470c0d1db322d)
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) = 1;
91 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
92 	&VNET_NAME(path_mtu_discovery), 1,
93 	"Enable Path MTU Discovery");
94 
95 VNET_DEFINE(int, ss_fltsz) = 1;
96 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, slowstart_flightsize, CTLFLAG_RW,
97 	&VNET_NAME(ss_fltsz), 1,
98 	"Slow start flight size");
99 
100 VNET_DEFINE(int, ss_fltsz_local) = 4;
101 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, local_slowstart_flightsize,
102 	CTLFLAG_RW, &VNET_NAME(ss_fltsz_local), 1,
103 	"Slow start flight size for local networks");
104 
105 VNET_DEFINE(int, tcp_do_newreno) = 1;
106 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, newreno, CTLFLAG_RW,
107 	&VNET_NAME(tcp_do_newreno), 0,
108 	"Enable NewReno Algorithms");
109 
110 VNET_DEFINE(int, tcp_do_tso) = 1;
111 #define	V_tcp_do_tso		VNET(tcp_do_tso)
112 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW,
113 	&VNET_NAME(tcp_do_tso), 0,
114 	"Enable TCP Segmentation Offload");
115 
116 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
117 #define	V_tcp_do_autosndbuf	VNET(tcp_do_autosndbuf)
118 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW,
119 	&VNET_NAME(tcp_do_autosndbuf), 0,
120 	"Enable automatic send buffer sizing");
121 
122 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
123 #define	V_tcp_autosndbuf_inc	VNET(tcp_autosndbuf_inc)
124 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW,
125 	&VNET_NAME(tcp_autosndbuf_inc), 0,
126 	"Incrementor step size of automatic send buffer");
127 
128 VNET_DEFINE(int, tcp_autosndbuf_max) = 256*1024;
129 #define	V_tcp_autosndbuf_max	VNET(tcp_autosndbuf_max)
130 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW,
131 	&VNET_NAME(tcp_autosndbuf_max), 0,
132 	"Max size of automatic send buffer");
133 
134 
135 /*
136  * Tcp output routine: figure out what should be sent and send it.
137  */
138 int
139 tcp_output(struct tcpcb *tp)
140 {
141 	struct socket *so = tp->t_inpcb->inp_socket;
142 	long len, recwin, sendwin;
143 	int off, flags, error, rw;
144 	struct mbuf *m;
145 	struct ip *ip = NULL;
146 	struct ipovly *ipov = NULL;
147 	struct tcphdr *th;
148 	u_char opt[TCP_MAXOLEN];
149 	unsigned ipoptlen, optlen, hdrlen;
150 #ifdef IPSEC
151 	unsigned ipsec_optlen = 0;
152 #endif
153 	int idle, sendalot;
154 	int sack_rxmit, sack_bytes_rxmt;
155 	struct sackhole *p;
156 	int tso;
157 	struct tcpopt to;
158 #if 0
159 	int maxburst = TCP_MAXBURST;
160 #endif
161 #ifdef INET6
162 	struct ip6_hdr *ip6 = NULL;
163 	int isipv6;
164 
165 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
166 #endif
167 
168 	INP_WLOCK_ASSERT(tp->t_inpcb);
169 
170 	/*
171 	 * Determine length of data that should be transmitted,
172 	 * and flags that will be used.
173 	 * If there is some data or critical controls (SYN, RST)
174 	 * to send, then transmit; otherwise, investigate further.
175 	 */
176 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
177 	if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur) {
178 		/*
179 		 * If we've been idle for more than one retransmit
180 		 * timeout the old congestion window is no longer
181 		 * current and we have to reduce it to the restart
182 		 * window before we can transmit again.
183 		 *
184 		 * The restart window is the initial window or the last
185 		 * CWND, whichever is smaller.
186 		 *
187 		 * This is done to prevent us from flooding the path with
188 		 * a full CWND at wirespeed, overloading router and switch
189 		 * buffers along the way.
190 		 *
191 		 * See RFC5681 Section 4.1. "Restarting Idle Connections".
192 		 */
193 		if (V_tcp_do_rfc3390)
194 			rw = min(4 * tp->t_maxseg,
195 				 max(2 * tp->t_maxseg, 4380));
196 #ifdef INET6
197 		else if ((isipv6 ? in6_localaddr(&tp->t_inpcb->in6p_faddr) :
198 			  in_localaddr(tp->t_inpcb->inp_faddr)))
199 #else
200 		else if (in_localaddr(tp->t_inpcb->inp_faddr))
201 #endif
202 			rw = V_ss_fltsz_local * tp->t_maxseg;
203 		else
204 			rw = V_ss_fltsz * tp->t_maxseg;
205 
206 		tp->snd_cwnd = min(rw, tp->snd_cwnd);
207 	}
208 	tp->t_flags &= ~TF_LASTIDLE;
209 	if (idle) {
210 		if (tp->t_flags & TF_MORETOCOME) {
211 			tp->t_flags |= TF_LASTIDLE;
212 			idle = 0;
213 		}
214 	}
215 again:
216 	/*
217 	 * If we've recently taken a timeout, snd_max will be greater than
218 	 * snd_nxt.  There may be SACK information that allows us to avoid
219 	 * resending already delivered data.  Adjust snd_nxt accordingly.
220 	 */
221 	if ((tp->t_flags & TF_SACK_PERMIT) &&
222 	    SEQ_LT(tp->snd_nxt, tp->snd_max))
223 		tcp_sack_adjust(tp);
224 	sendalot = 0;
225 	tso = 0;
226 	off = tp->snd_nxt - tp->snd_una;
227 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
228 
229 	flags = tcp_outflags[tp->t_state];
230 	/*
231 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
232 	 * to send out new data (when sendalot is 1), bypass this function.
233 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
234 	 * we're replacing a (future) new transmission with a retransmission
235 	 * now, and we previously incremented snd_cwnd in tcp_input().
236 	 */
237 	/*
238 	 * Still in sack recovery , reset rxmit flag to zero.
239 	 */
240 	sack_rxmit = 0;
241 	sack_bytes_rxmt = 0;
242 	len = 0;
243 	p = NULL;
244 	if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp) &&
245 	    (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
246 		long cwin;
247 
248 		cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
249 		if (cwin < 0)
250 			cwin = 0;
251 		/* Do not retransmit SACK segments beyond snd_recover */
252 		if (SEQ_GT(p->end, tp->snd_recover)) {
253 			/*
254 			 * (At least) part of sack hole extends beyond
255 			 * snd_recover. Check to see if we can rexmit data
256 			 * for this hole.
257 			 */
258 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
259 				/*
260 				 * Can't rexmit any more data for this hole.
261 				 * That data will be rexmitted in the next
262 				 * sack recovery episode, when snd_recover
263 				 * moves past p->rxmit.
264 				 */
265 				p = NULL;
266 				goto after_sack_rexmit;
267 			} else
268 				/* Can rexmit part of the current hole */
269 				len = ((long)ulmin(cwin,
270 						   tp->snd_recover - p->rxmit));
271 		} else
272 			len = ((long)ulmin(cwin, p->end - p->rxmit));
273 		off = p->rxmit - tp->snd_una;
274 		KASSERT(off >= 0,("%s: sack block to the left of una : %d",
275 		    __func__, off));
276 		if (len > 0) {
277 			sack_rxmit = 1;
278 			sendalot = 1;
279 			TCPSTAT_INC(tcps_sack_rexmits);
280 			TCPSTAT_ADD(tcps_sack_rexmit_bytes,
281 			    min(len, tp->t_maxseg));
282 		}
283 	}
284 after_sack_rexmit:
285 	/*
286 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
287 	 * state flags.
288 	 */
289 	if (tp->t_flags & TF_NEEDFIN)
290 		flags |= TH_FIN;
291 	if (tp->t_flags & TF_NEEDSYN)
292 		flags |= TH_SYN;
293 
294 	SOCKBUF_LOCK(&so->so_snd);
295 	/*
296 	 * If in persist timeout with window of 0, send 1 byte.
297 	 * Otherwise, if window is small but nonzero
298 	 * and timer expired, we will send what we can
299 	 * and go to transmit state.
300 	 */
301 	if (tp->t_flags & TF_FORCEDATA) {
302 		if (sendwin == 0) {
303 			/*
304 			 * If we still have some data to send, then
305 			 * clear the FIN bit.  Usually this would
306 			 * happen below when it realizes that we
307 			 * aren't sending all the data.  However,
308 			 * if we have exactly 1 byte of unsent data,
309 			 * then it won't clear the FIN bit below,
310 			 * and if we are in persist state, we wind
311 			 * up sending the packet without recording
312 			 * that we sent the FIN bit.
313 			 *
314 			 * We can't just blindly clear the FIN bit,
315 			 * because if we don't have any more data
316 			 * to send then the probe will be the FIN
317 			 * itself.
318 			 */
319 			if (off < so->so_snd.sb_cc)
320 				flags &= ~TH_FIN;
321 			sendwin = 1;
322 		} else {
323 			tcp_timer_activate(tp, TT_PERSIST, 0);
324 			tp->t_rxtshift = 0;
325 		}
326 	}
327 
328 	/*
329 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
330 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
331 	 * a negative length.  This can also occur when TCP opens up
332 	 * its congestion window while receiving additional duplicate
333 	 * acks after fast-retransmit because TCP will reset snd_nxt
334 	 * to snd_max after the fast-retransmit.
335 	 *
336 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
337 	 * be set to snd_una, the offset will be 0, and the length may
338 	 * wind up 0.
339 	 *
340 	 * If sack_rxmit is true we are retransmitting from the scoreboard
341 	 * in which case len is already set.
342 	 */
343 	if (sack_rxmit == 0) {
344 		if (sack_bytes_rxmt == 0)
345 			len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off);
346 		else {
347 			long cwin;
348 
349                         /*
350 			 * We are inside of a SACK recovery episode and are
351 			 * sending new data, having retransmitted all the
352 			 * data possible in the scoreboard.
353 			 */
354 			len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd)
355 			       - off);
356 			/*
357 			 * Don't remove this (len > 0) check !
358 			 * We explicitly check for len > 0 here (although it
359 			 * isn't really necessary), to work around a gcc
360 			 * optimization issue - to force gcc to compute
361 			 * len above. Without this check, the computation
362 			 * of len is bungled by the optimizer.
363 			 */
364 			if (len > 0) {
365 				cwin = tp->snd_cwnd -
366 					(tp->snd_nxt - tp->sack_newdata) -
367 					sack_bytes_rxmt;
368 				if (cwin < 0)
369 					cwin = 0;
370 				len = lmin(len, cwin);
371 			}
372 		}
373 	}
374 
375 	/*
376 	 * Lop off SYN bit if it has already been sent.  However, if this
377 	 * is SYN-SENT state and if segment contains data and if we don't
378 	 * know that foreign host supports TAO, suppress sending segment.
379 	 */
380 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
381 		if (tp->t_state != TCPS_SYN_RECEIVED)
382 			flags &= ~TH_SYN;
383 		off--, len++;
384 	}
385 
386 	/*
387 	 * Be careful not to send data and/or FIN on SYN segments.
388 	 * This measure is needed to prevent interoperability problems
389 	 * with not fully conformant TCP implementations.
390 	 */
391 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
392 		len = 0;
393 		flags &= ~TH_FIN;
394 	}
395 
396 	if (len < 0) {
397 		/*
398 		 * If FIN has been sent but not acked,
399 		 * but we haven't been called to retransmit,
400 		 * len will be < 0.  Otherwise, window shrank
401 		 * after we sent into it.  If window shrank to 0,
402 		 * cancel pending retransmit, pull snd_nxt back
403 		 * to (closed) window, and set the persist timer
404 		 * if it isn't already going.  If the window didn't
405 		 * close completely, just wait for an ACK.
406 		 */
407 		len = 0;
408 		if (sendwin == 0) {
409 			tcp_timer_activate(tp, TT_REXMT, 0);
410 			tp->t_rxtshift = 0;
411 			tp->snd_nxt = tp->snd_una;
412 			if (!tcp_timer_active(tp, TT_PERSIST))
413 				tcp_setpersist(tp);
414 		}
415 	}
416 
417 	/* len will be >= 0 after this point. */
418 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
419 
420 	/*
421 	 * Automatic sizing of send socket buffer.  Often the send buffer
422 	 * size is not optimally adjusted to the actual network conditions
423 	 * at hand (delay bandwidth product).  Setting the buffer size too
424 	 * small limits throughput on links with high bandwidth and high
425 	 * delay (eg. trans-continental/oceanic links).  Setting the
426 	 * buffer size too big consumes too much real kernel memory,
427 	 * especially with many connections on busy servers.
428 	 *
429 	 * The criteria to step up the send buffer one notch are:
430 	 *  1. receive window of remote host is larger than send buffer
431 	 *     (with a fudge factor of 5/4th);
432 	 *  2. send buffer is filled to 7/8th with data (so we actually
433 	 *     have data to make use of it);
434 	 *  3. send buffer fill has not hit maximal automatic size;
435 	 *  4. our send window (slow start and cogestion controlled) is
436 	 *     larger than sent but unacknowledged data in send buffer.
437 	 *
438 	 * The remote host receive window scaling factor may limit the
439 	 * growing of the send buffer before it reaches its allowed
440 	 * maximum.
441 	 *
442 	 * It scales directly with slow start or congestion window
443 	 * and does at most one step per received ACK.  This fast
444 	 * scaling has the drawback of growing the send buffer beyond
445 	 * what is strictly necessary to make full use of a given
446 	 * delay*bandwith product.  However testing has shown this not
447 	 * to be much of an problem.  At worst we are trading wasting
448 	 * of available bandwith (the non-use of it) for wasting some
449 	 * socket buffer memory.
450 	 *
451 	 * TODO: Shrink send buffer during idle periods together
452 	 * with congestion window.  Requires another timer.  Has to
453 	 * wait for upcoming tcp timer rewrite.
454 	 */
455 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
456 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
457 		    so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
458 		    so->so_snd.sb_cc < V_tcp_autosndbuf_max &&
459 		    sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
460 			if (!sbreserve_locked(&so->so_snd,
461 			    min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
462 			     V_tcp_autosndbuf_max), so, curthread))
463 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
464 		}
465 	}
466 
467 	/*
468 	 * Decide if we can use TCP Segmentation Offloading (if supported by
469 	 * hardware).
470 	 *
471 	 * TSO may only be used if we are in a pure bulk sending state.  The
472 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
473 	 * IP options prevent using TSO.  With TSO the TCP header is the same
474 	 * (except for the sequence number) for all generated packets.  This
475 	 * makes it impossible to transmit any options which vary per generated
476 	 * segment or packet.
477 	 */
478 #ifdef IPSEC
479 	/*
480 	 * Pre-calculate here as we save another lookup into the darknesses
481 	 * of IPsec that way and can actually decide if TSO is ok.
482 	 */
483 	ipsec_optlen = ipsec_hdrsiz_tcp(tp);
484 #endif
485 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
486 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
487 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
488 #ifdef IPSEC
489 	    ipsec_optlen == 0 &&
490 #endif
491 	    tp->t_inpcb->inp_options == NULL &&
492 	    tp->t_inpcb->in6p_options == NULL)
493 		tso = 1;
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 	if (len + optlen + ipoptlen > tp->t_maxopd) {
725 		flags &= ~TH_FIN;
726 
727 		if (tso) {
728 			KASSERT(ipoptlen == 0,
729 			    ("%s: TSO can't do IP options", __func__));
730 
731 			/*
732 			 * Limit a burst to IP_MAXPACKET minus IP,
733 			 * TCP and options length to keep ip->ip_len
734 			 * from overflowing.
735 			 */
736 			if (len > IP_MAXPACKET - hdrlen) {
737 				len = IP_MAXPACKET - hdrlen;
738 				sendalot = 1;
739 			}
740 
741 			/*
742 			 * Prevent the last segment from being
743 			 * fractional unless the send sockbuf can
744 			 * be emptied.
745 			 */
746 			if (sendalot && off + len < so->so_snd.sb_cc) {
747 				len -= len % (tp->t_maxopd - optlen);
748 				sendalot = 1;
749 			}
750 
751 			/*
752 			 * Send the FIN in a separate segment
753 			 * after the bulk sending is done.
754 			 * We don't trust the TSO implementations
755 			 * to clear the FIN flag on all but the
756 			 * last segment.
757 			 */
758 			if (tp->t_flags & TF_NEEDFIN)
759 				sendalot = 1;
760 
761 		} else {
762 			len = tp->t_maxopd - optlen - ipoptlen;
763 			sendalot = 1;
764 		}
765 	} else
766 		tso = 0;
767 
768 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
769 	    ("%s: len > IP_MAXPACKET", __func__));
770 
771 /*#ifdef DIAGNOSTIC*/
772 #ifdef INET6
773 	if (max_linkhdr + hdrlen > MCLBYTES)
774 #else
775 	if (max_linkhdr + hdrlen > MHLEN)
776 #endif
777 		panic("tcphdr too big");
778 /*#endif*/
779 
780 	/*
781 	 * This KASSERT is here to catch edge cases at a well defined place.
782 	 * Before, those had triggered (random) panic conditions further down.
783 	 */
784 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
785 
786 	/*
787 	 * Grab a header mbuf, attaching a copy of data to
788 	 * be transmitted, and initialize the header from
789 	 * the template for sends on this connection.
790 	 */
791 	if (len) {
792 		struct mbuf *mb;
793 		u_int moff;
794 
795 		if ((tp->t_flags & TF_FORCEDATA) && len == 1)
796 			TCPSTAT_INC(tcps_sndprobe);
797 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
798 			TCPSTAT_INC(tcps_sndrexmitpack);
799 			TCPSTAT_ADD(tcps_sndrexmitbyte, len);
800 		} else {
801 			TCPSTAT_INC(tcps_sndpack);
802 			TCPSTAT_ADD(tcps_sndbyte, len);
803 		}
804 #ifdef notyet
805 		if ((m = m_copypack(so->so_snd.sb_mb, off,
806 		    (int)len, max_linkhdr + hdrlen)) == 0) {
807 			SOCKBUF_UNLOCK(&so->so_snd);
808 			error = ENOBUFS;
809 			goto out;
810 		}
811 		/*
812 		 * m_copypack left space for our hdr; use it.
813 		 */
814 		m->m_len += hdrlen;
815 		m->m_data -= hdrlen;
816 #else
817 		MGETHDR(m, M_DONTWAIT, MT_DATA);
818 		if (m == NULL) {
819 			SOCKBUF_UNLOCK(&so->so_snd);
820 			error = ENOBUFS;
821 			goto out;
822 		}
823 #ifdef INET6
824 		if (MHLEN < hdrlen + max_linkhdr) {
825 			MCLGET(m, M_DONTWAIT);
826 			if ((m->m_flags & M_EXT) == 0) {
827 				SOCKBUF_UNLOCK(&so->so_snd);
828 				m_freem(m);
829 				error = ENOBUFS;
830 				goto out;
831 			}
832 		}
833 #endif
834 		m->m_data += max_linkhdr;
835 		m->m_len = hdrlen;
836 
837 		/*
838 		 * Start the m_copy functions from the closest mbuf
839 		 * to the offset in the socket buffer chain.
840 		 */
841 		mb = sbsndptr(&so->so_snd, off, len, &moff);
842 
843 		if (len <= MHLEN - hdrlen - max_linkhdr) {
844 			m_copydata(mb, moff, (int)len,
845 			    mtod(m, caddr_t) + hdrlen);
846 			m->m_len += len;
847 		} else {
848 			m->m_next = m_copy(mb, moff, (int)len);
849 			if (m->m_next == NULL) {
850 				SOCKBUF_UNLOCK(&so->so_snd);
851 				(void) m_free(m);
852 				error = ENOBUFS;
853 				goto out;
854 			}
855 		}
856 #endif
857 		/*
858 		 * If we're sending everything we've got, set PUSH.
859 		 * (This will keep happy those implementations which only
860 		 * give data to the user when a buffer fills or
861 		 * a PUSH comes in.)
862 		 */
863 		if (off + len == so->so_snd.sb_cc)
864 			flags |= TH_PUSH;
865 		SOCKBUF_UNLOCK(&so->so_snd);
866 	} else {
867 		SOCKBUF_UNLOCK(&so->so_snd);
868 		if (tp->t_flags & TF_ACKNOW)
869 			TCPSTAT_INC(tcps_sndacks);
870 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
871 			TCPSTAT_INC(tcps_sndctrl);
872 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
873 			TCPSTAT_INC(tcps_sndurg);
874 		else
875 			TCPSTAT_INC(tcps_sndwinup);
876 
877 		MGETHDR(m, M_DONTWAIT, MT_DATA);
878 		if (m == NULL) {
879 			error = ENOBUFS;
880 			goto out;
881 		}
882 #ifdef INET6
883 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
884 		    MHLEN >= hdrlen) {
885 			MH_ALIGN(m, hdrlen);
886 		} else
887 #endif
888 		m->m_data += max_linkhdr;
889 		m->m_len = hdrlen;
890 	}
891 	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
892 	m->m_pkthdr.rcvif = (struct ifnet *)0;
893 #ifdef MAC
894 	mac_inpcb_create_mbuf(tp->t_inpcb, m);
895 #endif
896 #ifdef INET6
897 	if (isipv6) {
898 		ip6 = mtod(m, struct ip6_hdr *);
899 		th = (struct tcphdr *)(ip6 + 1);
900 		tcpip_fillheaders(tp->t_inpcb, ip6, th);
901 	} else
902 #endif /* INET6 */
903 	{
904 		ip = mtod(m, struct ip *);
905 		ipov = (struct ipovly *)ip;
906 		th = (struct tcphdr *)(ip + 1);
907 		tcpip_fillheaders(tp->t_inpcb, ip, th);
908 	}
909 
910 	/*
911 	 * Fill in fields, remembering maximum advertised
912 	 * window for use in delaying messages about window sizes.
913 	 * If resending a FIN, be sure not to use a new sequence number.
914 	 */
915 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
916 	    tp->snd_nxt == tp->snd_max)
917 		tp->snd_nxt--;
918 	/*
919 	 * If we are starting a connection, send ECN setup
920 	 * SYN packet. If we are on a retransmit, we may
921 	 * resend those bits a number of times as per
922 	 * RFC 3168.
923 	 */
924 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
925 		if (tp->t_rxtshift >= 1) {
926 			if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
927 				flags |= TH_ECE|TH_CWR;
928 		} else
929 			flags |= TH_ECE|TH_CWR;
930 	}
931 
932 	if (tp->t_state == TCPS_ESTABLISHED &&
933 	    (tp->t_flags & TF_ECN_PERMIT)) {
934 		/*
935 		 * If the peer has ECN, mark data packets with
936 		 * ECN capable transmission (ECT).
937 		 * Ignore pure ack packets, retransmissions and window probes.
938 		 */
939 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
940 		    !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
941 #ifdef INET6
942 			if (isipv6)
943 				ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
944 			else
945 #endif
946 				ip->ip_tos |= IPTOS_ECN_ECT0;
947 			TCPSTAT_INC(tcps_ecn_ect0);
948 		}
949 
950 		/*
951 		 * Reply with proper ECN notifications.
952 		 */
953 		if (tp->t_flags & TF_ECN_SND_CWR) {
954 			flags |= TH_CWR;
955 			tp->t_flags &= ~TF_ECN_SND_CWR;
956 		}
957 		if (tp->t_flags & TF_ECN_SND_ECE)
958 			flags |= TH_ECE;
959 	}
960 
961 	/*
962 	 * If we are doing retransmissions, then snd_nxt will
963 	 * not reflect the first unsent octet.  For ACK only
964 	 * packets, we do not want the sequence number of the
965 	 * retransmitted packet, we want the sequence number
966 	 * of the next unsent octet.  So, if there is no data
967 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
968 	 * when filling in ti_seq.  But if we are in persist
969 	 * state, snd_max might reflect one byte beyond the
970 	 * right edge of the window, so use snd_nxt in that
971 	 * case, since we know we aren't doing a retransmission.
972 	 * (retransmit and persist are mutually exclusive...)
973 	 */
974 	if (sack_rxmit == 0) {
975 		if (len || (flags & (TH_SYN|TH_FIN)) ||
976 		    tcp_timer_active(tp, TT_PERSIST))
977 			th->th_seq = htonl(tp->snd_nxt);
978 		else
979 			th->th_seq = htonl(tp->snd_max);
980 	} else {
981 		th->th_seq = htonl(p->rxmit);
982 		p->rxmit += len;
983 		tp->sackhint.sack_bytes_rexmit += len;
984 	}
985 	th->th_ack = htonl(tp->rcv_nxt);
986 	if (optlen) {
987 		bcopy(opt, th + 1, optlen);
988 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
989 	}
990 	th->th_flags = flags;
991 	/*
992 	 * Calculate receive window.  Don't shrink window,
993 	 * but avoid silly window syndrome.
994 	 */
995 	if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
996 	    recwin < (long)tp->t_maxseg)
997 		recwin = 0;
998 	if (recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
999 		recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
1000 	if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
1001 		recwin = (long)TCP_MAXWIN << tp->rcv_scale;
1002 
1003 	/*
1004 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1005 	 * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
1006 	 * case is handled in syncache.
1007 	 */
1008 	if (flags & TH_SYN)
1009 		th->th_win = htons((u_short)
1010 				(min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1011 	else
1012 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1013 
1014 	/*
1015 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1016 	 * a 0 window.  This may cause the remote transmitter to stall.  This
1017 	 * flag tells soreceive() to disable delayed acknowledgements when
1018 	 * draining the buffer.  This can occur if the receiver is attempting
1019 	 * to read more data than can be buffered prior to transmitting on
1020 	 * the connection.
1021 	 */
1022 	if (th->th_win == 0)
1023 		tp->t_flags |= TF_RXWIN0SENT;
1024 	else
1025 		tp->t_flags &= ~TF_RXWIN0SENT;
1026 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1027 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1028 		th->th_flags |= TH_URG;
1029 	} else
1030 		/*
1031 		 * If no urgent pointer to send, then we pull
1032 		 * the urgent pointer to the left edge of the send window
1033 		 * so that it doesn't drift into the send window on sequence
1034 		 * number wraparound.
1035 		 */
1036 		tp->snd_up = tp->snd_una;		/* drag it along */
1037 
1038 #ifdef TCP_SIGNATURE
1039 	if (tp->t_flags & TF_SIGNATURE) {
1040 		int sigoff = to.to_signature - opt;
1041 		tcp_signature_compute(m, 0, len, optlen,
1042 		    (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
1043 	}
1044 #endif
1045 
1046 	/*
1047 	 * Put TCP length in extended header, and then
1048 	 * checksum extended header and data.
1049 	 */
1050 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1051 #ifdef INET6
1052 	if (isipv6)
1053 		/*
1054 		 * ip6_plen is not need to be filled now, and will be filled
1055 		 * in ip6_output.
1056 		 */
1057 		th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
1058 				       sizeof(struct tcphdr) + optlen + len);
1059 	else
1060 #endif /* INET6 */
1061 	{
1062 		m->m_pkthdr.csum_flags = CSUM_TCP;
1063 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1064 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1065 		    htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1066 
1067 		/* IP version must be set here for ipv4/ipv6 checking later */
1068 		KASSERT(ip->ip_v == IPVERSION,
1069 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1070 	}
1071 
1072 	/*
1073 	 * Enable TSO and specify the size of the segments.
1074 	 * The TCP pseudo header checksum is always provided.
1075 	 * XXX: Fixme: This is currently not the case for IPv6.
1076 	 */
1077 	if (tso) {
1078 		KASSERT(len > tp->t_maxopd - optlen,
1079 		    ("%s: len <= tso_segsz", __func__));
1080 		m->m_pkthdr.csum_flags |= CSUM_TSO;
1081 		m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen;
1082 	}
1083 
1084 	KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL),
1085 	    ("%s: mbuf chain shorter than expected", __func__));
1086 
1087 	/*
1088 	 * In transmit state, time the transmission and arrange for
1089 	 * the retransmit.  In persist state, just set snd_max.
1090 	 */
1091 	if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1092 	    !tcp_timer_active(tp, TT_PERSIST)) {
1093 		tcp_seq startseq = tp->snd_nxt;
1094 
1095 		/*
1096 		 * Advance snd_nxt over sequence space of this segment.
1097 		 */
1098 		if (flags & (TH_SYN|TH_FIN)) {
1099 			if (flags & TH_SYN)
1100 				tp->snd_nxt++;
1101 			if (flags & TH_FIN) {
1102 				tp->snd_nxt++;
1103 				tp->t_flags |= TF_SENTFIN;
1104 			}
1105 		}
1106 		if (sack_rxmit)
1107 			goto timer;
1108 		tp->snd_nxt += len;
1109 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1110 			tp->snd_max = tp->snd_nxt;
1111 			/*
1112 			 * Time this transmission if not a retransmission and
1113 			 * not currently timing anything.
1114 			 */
1115 			if (tp->t_rtttime == 0) {
1116 				tp->t_rtttime = ticks;
1117 				tp->t_rtseq = startseq;
1118 				TCPSTAT_INC(tcps_segstimed);
1119 			}
1120 		}
1121 
1122 		/*
1123 		 * Set retransmit timer if not currently set,
1124 		 * and not doing a pure ack or a keep-alive probe.
1125 		 * Initial value for retransmit timer is smoothed
1126 		 * round-trip time + 2 * round-trip time variance.
1127 		 * Initialize shift counter which is used for backoff
1128 		 * of retransmit time.
1129 		 */
1130 timer:
1131 		if (!tcp_timer_active(tp, TT_REXMT) &&
1132 		    ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1133 		     (tp->snd_nxt != tp->snd_una))) {
1134 			if (tcp_timer_active(tp, TT_PERSIST)) {
1135 				tcp_timer_activate(tp, TT_PERSIST, 0);
1136 				tp->t_rxtshift = 0;
1137 			}
1138 			tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1139 		}
1140 	} else {
1141 		/*
1142 		 * Persist case, update snd_max but since we are in
1143 		 * persist mode (no window) we do not update snd_nxt.
1144 		 */
1145 		int xlen = len;
1146 		if (flags & TH_SYN)
1147 			++xlen;
1148 		if (flags & TH_FIN) {
1149 			++xlen;
1150 			tp->t_flags |= TF_SENTFIN;
1151 		}
1152 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1153 			tp->snd_max = tp->snd_nxt + len;
1154 	}
1155 
1156 #ifdef TCPDEBUG
1157 	/*
1158 	 * Trace.
1159 	 */
1160 	if (so->so_options & SO_DEBUG) {
1161 		u_short save = 0;
1162 #ifdef INET6
1163 		if (!isipv6)
1164 #endif
1165 		{
1166 			save = ipov->ih_len;
1167 			ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1168 		}
1169 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1170 #ifdef INET6
1171 		if (!isipv6)
1172 #endif
1173 		ipov->ih_len = save;
1174 	}
1175 #endif
1176 
1177 	/*
1178 	 * Fill in IP length and desired time to live and
1179 	 * send to IP level.  There should be a better way
1180 	 * to handle ttl and tos; we could keep them in
1181 	 * the template, but need a way to checksum without them.
1182 	 */
1183 	/*
1184 	 * m->m_pkthdr.len should have been set before cksum calcuration,
1185 	 * because in6_cksum() need it.
1186 	 */
1187 #ifdef INET6
1188 	if (isipv6) {
1189 		/*
1190 		 * we separately set hoplimit for every segment, since the
1191 		 * user might want to change the value via setsockopt.
1192 		 * Also, desired default hop limit might be changed via
1193 		 * Neighbor Discovery.
1194 		 */
1195 		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1196 
1197 		/* TODO: IPv6 IP6TOS_ECT bit on */
1198 		error = ip6_output(m,
1199 			    tp->t_inpcb->in6p_outputopts, NULL,
1200 			    ((so->so_options & SO_DONTROUTE) ?
1201 			    IP_ROUTETOIF : 0), NULL, NULL, tp->t_inpcb);
1202 	} else
1203 #endif /* INET6 */
1204     {
1205 	ip->ip_len = m->m_pkthdr.len;
1206 #ifdef INET6
1207 	if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1208 		ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1209 #endif /* INET6 */
1210 	/*
1211 	 * If we do path MTU discovery, then we set DF on every packet.
1212 	 * This might not be the best thing to do according to RFC3390
1213 	 * Section 2. However the tcp hostcache migitates the problem
1214 	 * so it affects only the first tcp connection with a host.
1215 	 *
1216 	 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1217 	 */
1218 	if (V_path_mtu_discovery && tp->t_maxopd > V_tcp_minmss)
1219 		ip->ip_off |= IP_DF;
1220 
1221 	error = ip_output(m, tp->t_inpcb->inp_options, NULL,
1222 	    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1223 	    tp->t_inpcb);
1224     }
1225 	if (error) {
1226 
1227 		/*
1228 		 * We know that the packet was lost, so back out the
1229 		 * sequence number advance, if any.
1230 		 *
1231 		 * If the error is EPERM the packet got blocked by the
1232 		 * local firewall.  Normally we should terminate the
1233 		 * connection but the blocking may have been spurious
1234 		 * due to a firewall reconfiguration cycle.  So we treat
1235 		 * it like a packet loss and let the retransmit timer and
1236 		 * timeouts do their work over time.
1237 		 * XXX: It is a POLA question whether calling tcp_drop right
1238 		 * away would be the really correct behavior instead.
1239 		 */
1240 		if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1241 		    !tcp_timer_active(tp, TT_PERSIST)) &&
1242 		    ((flags & TH_SYN) == 0) &&
1243 		    (error != EPERM)) {
1244 			if (sack_rxmit) {
1245 				p->rxmit -= len;
1246 				tp->sackhint.sack_bytes_rexmit -= len;
1247 				KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1248 				    ("sackhint bytes rtx >= 0"));
1249 			} else
1250 				tp->snd_nxt -= len;
1251 		}
1252 out:
1253 		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1254 		switch (error) {
1255 		case EPERM:
1256 			tp->t_softerror = error;
1257 			return (error);
1258 		case ENOBUFS:
1259 	                if (!tcp_timer_active(tp, TT_REXMT) &&
1260 			    !tcp_timer_active(tp, TT_PERSIST))
1261 	                        tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1262 			tp->snd_cwnd = tp->t_maxseg;
1263 			return (0);
1264 		case EMSGSIZE:
1265 			/*
1266 			 * For some reason the interface we used initially
1267 			 * to send segments changed to another or lowered
1268 			 * its MTU.
1269 			 *
1270 			 * tcp_mtudisc() will find out the new MTU and as
1271 			 * its last action, initiate retransmission, so it
1272 			 * is important to not do so here.
1273 			 *
1274 			 * If TSO was active we either got an interface
1275 			 * without TSO capabilits or TSO was turned off.
1276 			 * Disable it for this connection as too and
1277 			 * immediatly retry with MSS sized segments generated
1278 			 * by this function.
1279 			 */
1280 			if (tso)
1281 				tp->t_flags &= ~TF_TSO;
1282 			tcp_mtudisc(tp->t_inpcb, 0);
1283 			return (0);
1284 		case EHOSTDOWN:
1285 		case EHOSTUNREACH:
1286 		case ENETDOWN:
1287 		case ENETUNREACH:
1288 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
1289 				tp->t_softerror = error;
1290 				return (0);
1291 			}
1292 			/* FALLTHROUGH */
1293 		default:
1294 			return (error);
1295 		}
1296 	}
1297 	TCPSTAT_INC(tcps_sndtotal);
1298 
1299 	/*
1300 	 * Data sent (as far as we can tell).
1301 	 * If this advertises a larger window than any other segment,
1302 	 * then remember the size of the advertised window.
1303 	 * Any pending ACK has now been sent.
1304 	 */
1305 	if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1306 		tp->rcv_adv = tp->rcv_nxt + recwin;
1307 	tp->last_ack_sent = tp->rcv_nxt;
1308 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1309 	if (tcp_timer_active(tp, TT_DELACK))
1310 		tcp_timer_activate(tp, TT_DELACK, 0);
1311 #if 0
1312 	/*
1313 	 * This completely breaks TCP if newreno is turned on.  What happens
1314 	 * is that if delayed-acks are turned on on the receiver, this code
1315 	 * on the transmitter effectively destroys the TCP window, forcing
1316 	 * it to four packets (1.5Kx4 = 6K window).
1317 	 */
1318 	if (sendalot && (!V_tcp_do_newreno || --maxburst))
1319 		goto again;
1320 #endif
1321 	if (sendalot)
1322 		goto again;
1323 	return (0);
1324 }
1325 
1326 void
1327 tcp_setpersist(struct tcpcb *tp)
1328 {
1329 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1330 	int tt;
1331 
1332 	if (tcp_timer_active(tp, TT_REXMT))
1333 		panic("tcp_setpersist: retransmit pending");
1334 	/*
1335 	 * Start/restart persistance timer.
1336 	 */
1337 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1338 		      TCPTV_PERSMIN, TCPTV_PERSMAX);
1339 	tcp_timer_activate(tp, TT_PERSIST, tt);
1340 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1341 		tp->t_rxtshift++;
1342 }
1343 
1344 /*
1345  * Insert TCP options according to the supplied parameters to the place
1346  * optp in a consistent way.  Can handle unaligned destinations.
1347  *
1348  * The order of the option processing is crucial for optimal packing and
1349  * alignment for the scarce option space.
1350  *
1351  * The optimal order for a SYN/SYN-ACK segment is:
1352  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1353  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1354  *
1355  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1356  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1357  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1358  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1359  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1360  */
1361 int
1362 tcp_addoptions(struct tcpopt *to, u_char *optp)
1363 {
1364 	u_int mask, optlen = 0;
1365 
1366 	for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1367 		if ((to->to_flags & mask) != mask)
1368 			continue;
1369 		if (optlen == TCP_MAXOLEN)
1370 			break;
1371 		switch (to->to_flags & mask) {
1372 		case TOF_MSS:
1373 			while (optlen % 4) {
1374 				optlen += TCPOLEN_NOP;
1375 				*optp++ = TCPOPT_NOP;
1376 			}
1377 			if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1378 				continue;
1379 			optlen += TCPOLEN_MAXSEG;
1380 			*optp++ = TCPOPT_MAXSEG;
1381 			*optp++ = TCPOLEN_MAXSEG;
1382 			to->to_mss = htons(to->to_mss);
1383 			bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1384 			optp += sizeof(to->to_mss);
1385 			break;
1386 		case TOF_SCALE:
1387 			while (!optlen || optlen % 2 != 1) {
1388 				optlen += TCPOLEN_NOP;
1389 				*optp++ = TCPOPT_NOP;
1390 			}
1391 			if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1392 				continue;
1393 			optlen += TCPOLEN_WINDOW;
1394 			*optp++ = TCPOPT_WINDOW;
1395 			*optp++ = TCPOLEN_WINDOW;
1396 			*optp++ = to->to_wscale;
1397 			break;
1398 		case TOF_SACKPERM:
1399 			while (optlen % 2) {
1400 				optlen += TCPOLEN_NOP;
1401 				*optp++ = TCPOPT_NOP;
1402 			}
1403 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1404 				continue;
1405 			optlen += TCPOLEN_SACK_PERMITTED;
1406 			*optp++ = TCPOPT_SACK_PERMITTED;
1407 			*optp++ = TCPOLEN_SACK_PERMITTED;
1408 			break;
1409 		case TOF_TS:
1410 			while (!optlen || optlen % 4 != 2) {
1411 				optlen += TCPOLEN_NOP;
1412 				*optp++ = TCPOPT_NOP;
1413 			}
1414 			if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1415 				continue;
1416 			optlen += TCPOLEN_TIMESTAMP;
1417 			*optp++ = TCPOPT_TIMESTAMP;
1418 			*optp++ = TCPOLEN_TIMESTAMP;
1419 			to->to_tsval = htonl(to->to_tsval);
1420 			to->to_tsecr = htonl(to->to_tsecr);
1421 			bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1422 			optp += sizeof(to->to_tsval);
1423 			bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1424 			optp += sizeof(to->to_tsecr);
1425 			break;
1426 		case TOF_SIGNATURE:
1427 			{
1428 			int siglen = TCPOLEN_SIGNATURE - 2;
1429 
1430 			while (!optlen || optlen % 4 != 2) {
1431 				optlen += TCPOLEN_NOP;
1432 				*optp++ = TCPOPT_NOP;
1433 			}
1434 			if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE)
1435 				continue;
1436 			optlen += TCPOLEN_SIGNATURE;
1437 			*optp++ = TCPOPT_SIGNATURE;
1438 			*optp++ = TCPOLEN_SIGNATURE;
1439 			to->to_signature = optp;
1440 			while (siglen--)
1441 				 *optp++ = 0;
1442 			break;
1443 			}
1444 		case TOF_SACK:
1445 			{
1446 			int sackblks = 0;
1447 			struct sackblk *sack = (struct sackblk *)to->to_sacks;
1448 			tcp_seq sack_seq;
1449 
1450 			while (!optlen || optlen % 4 != 2) {
1451 				optlen += TCPOLEN_NOP;
1452 				*optp++ = TCPOPT_NOP;
1453 			}
1454 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1455 				continue;
1456 			optlen += TCPOLEN_SACKHDR;
1457 			*optp++ = TCPOPT_SACK;
1458 			sackblks = min(to->to_nsacks,
1459 					(TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1460 			*optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1461 			while (sackblks--) {
1462 				sack_seq = htonl(sack->start);
1463 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1464 				optp += sizeof(sack_seq);
1465 				sack_seq = htonl(sack->end);
1466 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1467 				optp += sizeof(sack_seq);
1468 				optlen += TCPOLEN_SACK;
1469 				sack++;
1470 			}
1471 			TCPSTAT_INC(tcps_sack_send_blocks);
1472 			break;
1473 			}
1474 		default:
1475 			panic("%s: unknown TCP option type", __func__);
1476 			break;
1477 		}
1478 	}
1479 
1480 	/* Terminate and pad TCP options to a 4 byte boundary. */
1481 	if (optlen % 4) {
1482 		optlen += TCPOLEN_EOL;
1483 		*optp++ = TCPOPT_EOL;
1484 	}
1485 	/*
1486 	 * According to RFC 793 (STD0007):
1487 	 *   "The content of the header beyond the End-of-Option option
1488 	 *    must be header padding (i.e., zero)."
1489 	 *   and later: "The padding is composed of zeros."
1490 	 */
1491 	while (optlen % 4) {
1492 		optlen += TCPOLEN_PAD;
1493 		*optp++ = TCPOPT_PAD;
1494 	}
1495 
1496 	KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1497 	return (optlen);
1498 }
1499