xref: /freebsd/sys/netinet/tcp_output.c (revision b2db760808f74bb53c232900091c9da801ebbfcc)
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 	sendwin = min(sendwin, tp->snd_bwnd);
229 
230 	flags = tcp_outflags[tp->t_state];
231 	/*
232 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
233 	 * to send out new data (when sendalot is 1), bypass this function.
234 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
235 	 * we're replacing a (future) new transmission with a retransmission
236 	 * now, and we previously incremented snd_cwnd in tcp_input().
237 	 */
238 	/*
239 	 * Still in sack recovery , reset rxmit flag to zero.
240 	 */
241 	sack_rxmit = 0;
242 	sack_bytes_rxmt = 0;
243 	len = 0;
244 	p = NULL;
245 	if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp) &&
246 	    (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
247 		long cwin;
248 
249 		cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
250 		if (cwin < 0)
251 			cwin = 0;
252 		/* Do not retransmit SACK segments beyond snd_recover */
253 		if (SEQ_GT(p->end, tp->snd_recover)) {
254 			/*
255 			 * (At least) part of sack hole extends beyond
256 			 * snd_recover. Check to see if we can rexmit data
257 			 * for this hole.
258 			 */
259 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
260 				/*
261 				 * Can't rexmit any more data for this hole.
262 				 * That data will be rexmitted in the next
263 				 * sack recovery episode, when snd_recover
264 				 * moves past p->rxmit.
265 				 */
266 				p = NULL;
267 				goto after_sack_rexmit;
268 			} else
269 				/* Can rexmit part of the current hole */
270 				len = ((long)ulmin(cwin,
271 						   tp->snd_recover - p->rxmit));
272 		} else
273 			len = ((long)ulmin(cwin, p->end - p->rxmit));
274 		off = p->rxmit - tp->snd_una;
275 		KASSERT(off >= 0,("%s: sack block to the left of una : %d",
276 		    __func__, off));
277 		if (len > 0) {
278 			sack_rxmit = 1;
279 			sendalot = 1;
280 			TCPSTAT_INC(tcps_sack_rexmits);
281 			TCPSTAT_ADD(tcps_sack_rexmit_bytes,
282 			    min(len, tp->t_maxseg));
283 		}
284 	}
285 after_sack_rexmit:
286 	/*
287 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
288 	 * state flags.
289 	 */
290 	if (tp->t_flags & TF_NEEDFIN)
291 		flags |= TH_FIN;
292 	if (tp->t_flags & TF_NEEDSYN)
293 		flags |= TH_SYN;
294 
295 	SOCKBUF_LOCK(&so->so_snd);
296 	/*
297 	 * If in persist timeout with window of 0, send 1 byte.
298 	 * Otherwise, if window is small but nonzero
299 	 * and timer expired, we will send what we can
300 	 * and go to transmit state.
301 	 */
302 	if (tp->t_flags & TF_FORCEDATA) {
303 		if (sendwin == 0) {
304 			/*
305 			 * If we still have some data to send, then
306 			 * clear the FIN bit.  Usually this would
307 			 * happen below when it realizes that we
308 			 * aren't sending all the data.  However,
309 			 * if we have exactly 1 byte of unsent data,
310 			 * then it won't clear the FIN bit below,
311 			 * and if we are in persist state, we wind
312 			 * up sending the packet without recording
313 			 * that we sent the FIN bit.
314 			 *
315 			 * We can't just blindly clear the FIN bit,
316 			 * because if we don't have any more data
317 			 * to send then the probe will be the FIN
318 			 * itself.
319 			 */
320 			if (off < so->so_snd.sb_cc)
321 				flags &= ~TH_FIN;
322 			sendwin = 1;
323 		} else {
324 			tcp_timer_activate(tp, TT_PERSIST, 0);
325 			tp->t_rxtshift = 0;
326 		}
327 	}
328 
329 	/*
330 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
331 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
332 	 * a negative length.  This can also occur when TCP opens up
333 	 * its congestion window while receiving additional duplicate
334 	 * acks after fast-retransmit because TCP will reset snd_nxt
335 	 * to snd_max after the fast-retransmit.
336 	 *
337 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
338 	 * be set to snd_una, the offset will be 0, and the length may
339 	 * wind up 0.
340 	 *
341 	 * If sack_rxmit is true we are retransmitting from the scoreboard
342 	 * in which case len is already set.
343 	 */
344 	if (sack_rxmit == 0) {
345 		if (sack_bytes_rxmt == 0)
346 			len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off);
347 		else {
348 			long cwin;
349 
350                         /*
351 			 * We are inside of a SACK recovery episode and are
352 			 * sending new data, having retransmitted all the
353 			 * data possible in the scoreboard.
354 			 */
355 			len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd)
356 			       - off);
357 			/*
358 			 * Don't remove this (len > 0) check !
359 			 * We explicitly check for len > 0 here (although it
360 			 * isn't really necessary), to work around a gcc
361 			 * optimization issue - to force gcc to compute
362 			 * len above. Without this check, the computation
363 			 * of len is bungled by the optimizer.
364 			 */
365 			if (len > 0) {
366 				cwin = tp->snd_cwnd -
367 					(tp->snd_nxt - tp->sack_newdata) -
368 					sack_bytes_rxmt;
369 				if (cwin < 0)
370 					cwin = 0;
371 				len = lmin(len, cwin);
372 			}
373 		}
374 	}
375 
376 	/*
377 	 * Lop off SYN bit if it has already been sent.  However, if this
378 	 * is SYN-SENT state and if segment contains data and if we don't
379 	 * know that foreign host supports TAO, suppress sending segment.
380 	 */
381 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
382 		if (tp->t_state != TCPS_SYN_RECEIVED)
383 			flags &= ~TH_SYN;
384 		off--, len++;
385 	}
386 
387 	/*
388 	 * Be careful not to send data and/or FIN on SYN segments.
389 	 * This measure is needed to prevent interoperability problems
390 	 * with not fully conformant TCP implementations.
391 	 */
392 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
393 		len = 0;
394 		flags &= ~TH_FIN;
395 	}
396 
397 	if (len < 0) {
398 		/*
399 		 * If FIN has been sent but not acked,
400 		 * but we haven't been called to retransmit,
401 		 * len will be < 0.  Otherwise, window shrank
402 		 * after we sent into it.  If window shrank to 0,
403 		 * cancel pending retransmit, pull snd_nxt back
404 		 * to (closed) window, and set the persist timer
405 		 * if it isn't already going.  If the window didn't
406 		 * close completely, just wait for an ACK.
407 		 */
408 		len = 0;
409 		if (sendwin == 0) {
410 			tcp_timer_activate(tp, TT_REXMT, 0);
411 			tp->t_rxtshift = 0;
412 			tp->snd_nxt = tp->snd_una;
413 			if (!tcp_timer_active(tp, TT_PERSIST))
414 				tcp_setpersist(tp);
415 		}
416 	}
417 
418 	/* len will be >= 0 after this point. */
419 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
420 
421 	/*
422 	 * Automatic sizing of send socket buffer.  Often the send buffer
423 	 * size is not optimally adjusted to the actual network conditions
424 	 * at hand (delay bandwidth product).  Setting the buffer size too
425 	 * small limits throughput on links with high bandwidth and high
426 	 * delay (eg. trans-continental/oceanic links).  Setting the
427 	 * buffer size too big consumes too much real kernel memory,
428 	 * especially with many connections on busy servers.
429 	 *
430 	 * The criteria to step up the send buffer one notch are:
431 	 *  1. receive window of remote host is larger than send buffer
432 	 *     (with a fudge factor of 5/4th);
433 	 *  2. send buffer is filled to 7/8th with data (so we actually
434 	 *     have data to make use of it);
435 	 *  3. send buffer fill has not hit maximal automatic size;
436 	 *  4. our send window (slow start and cogestion controlled) is
437 	 *     larger than sent but unacknowledged data in send buffer.
438 	 *
439 	 * The remote host receive window scaling factor may limit the
440 	 * growing of the send buffer before it reaches its allowed
441 	 * maximum.
442 	 *
443 	 * It scales directly with slow start or congestion window
444 	 * and does at most one step per received ACK.  This fast
445 	 * scaling has the drawback of growing the send buffer beyond
446 	 * what is strictly necessary to make full use of a given
447 	 * delay*bandwith product.  However testing has shown this not
448 	 * to be much of an problem.  At worst we are trading wasting
449 	 * of available bandwith (the non-use of it) for wasting some
450 	 * socket buffer memory.
451 	 *
452 	 * TODO: Shrink send buffer during idle periods together
453 	 * with congestion window.  Requires another timer.  Has to
454 	 * wait for upcoming tcp timer rewrite.
455 	 */
456 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
457 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
458 		    so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
459 		    so->so_snd.sb_cc < V_tcp_autosndbuf_max &&
460 		    sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
461 			if (!sbreserve_locked(&so->so_snd,
462 			    min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
463 			     V_tcp_autosndbuf_max), so, curthread))
464 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
465 		}
466 	}
467 
468 	/*
469 	 * Truncate to the maximum segment length or enable TCP Segmentation
470 	 * Offloading (if supported by hardware) and ensure that FIN is removed
471 	 * if the length no longer contains the last data byte.
472 	 *
473 	 * TSO may only be used if we are in a pure bulk sending state.  The
474 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
475 	 * IP options prevent using TSO.  With TSO the TCP header is the same
476 	 * (except for the sequence number) for all generated packets.  This
477 	 * makes it impossible to transmit any options which vary per generated
478 	 * segment or packet.
479 	 *
480 	 * The length of TSO bursts is limited to TCP_MAXWIN.  That limit and
481 	 * removal of FIN (if not already catched here) are handled later after
482 	 * the exact length of the TCP options are known.
483 	 */
484 #ifdef IPSEC
485 	/*
486 	 * Pre-calculate here as we save another lookup into the darknesses
487 	 * of IPsec that way and can actually decide if TSO is ok.
488 	 */
489 	ipsec_optlen = ipsec_hdrsiz_tcp(tp);
490 #endif
491 	if (len > tp->t_maxseg) {
492 		if ((tp->t_flags & TF_TSO) && V_tcp_do_tso &&
493 		    ((tp->t_flags & TF_SIGNATURE) == 0) &&
494 		    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
495 		    tp->t_inpcb->inp_options == NULL &&
496 		    tp->t_inpcb->in6p_options == NULL
497 #ifdef IPSEC
498 		    && ipsec_optlen == 0
499 #endif
500 		    ) {
501 			tso = 1;
502 		} else {
503 			len = tp->t_maxseg;
504 			sendalot = 1;
505 		}
506 	}
507 
508 	if (sack_rxmit) {
509 		if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
510 			flags &= ~TH_FIN;
511 	} else {
512 		if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
513 			flags &= ~TH_FIN;
514 	}
515 
516 	recwin = sbspace(&so->so_rcv);
517 
518 	/*
519 	 * Sender silly window avoidance.   We transmit under the following
520 	 * conditions when len is non-zero:
521 	 *
522 	 *	- We have a full segment (or more with TSO)
523 	 *	- This is the last buffer in a write()/send() and we are
524 	 *	  either idle or running NODELAY
525 	 *	- we've timed out (e.g. persist timer)
526 	 *	- we have more then 1/2 the maximum send window's worth of
527 	 *	  data (receiver may be limited the window size)
528 	 *	- we need to retransmit
529 	 */
530 	if (len) {
531 		if (len >= tp->t_maxseg)
532 			goto send;
533 		/*
534 		 * NOTE! on localhost connections an 'ack' from the remote
535 		 * end may occur synchronously with the output and cause
536 		 * us to flush a buffer queued with moretocome.  XXX
537 		 *
538 		 * note: the len + off check is almost certainly unnecessary.
539 		 */
540 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
541 		    (idle || (tp->t_flags & TF_NODELAY)) &&
542 		    len + off >= so->so_snd.sb_cc &&
543 		    (tp->t_flags & TF_NOPUSH) == 0) {
544 			goto send;
545 		}
546 		if (tp->t_flags & TF_FORCEDATA)		/* typ. timeout case */
547 			goto send;
548 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
549 			goto send;
550 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
551 			goto send;
552 		if (sack_rxmit)
553 			goto send;
554 	}
555 
556 	/*
557 	 * Compare available window to amount of window
558 	 * known to peer (as advertised window less
559 	 * next expected input).  If the difference is at least two
560 	 * max size segments, or at least 50% of the maximum possible
561 	 * window, then want to send a window update to peer.
562 	 * Skip this if the connection is in T/TCP half-open state.
563 	 * Don't send pure window updates when the peer has closed
564 	 * the connection and won't ever send more data.
565 	 */
566 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
567 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
568 		/*
569 		 * "adv" is the amount we can increase the window,
570 		 * taking into account that we are limited by
571 		 * TCP_MAXWIN << tp->rcv_scale.
572 		 */
573 		long adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale) -
574 			(tp->rcv_adv - tp->rcv_nxt);
575 
576 		if (adv >= (long) (2 * tp->t_maxseg))
577 			goto send;
578 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
579 			goto send;
580 	}
581 
582 	/*
583 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
584 	 * is also a catch-all for the retransmit timer timeout case.
585 	 */
586 	if (tp->t_flags & TF_ACKNOW)
587 		goto send;
588 	if ((flags & TH_RST) ||
589 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
590 		goto send;
591 	if (SEQ_GT(tp->snd_up, tp->snd_una))
592 		goto send;
593 	/*
594 	 * If our state indicates that FIN should be sent
595 	 * and we have not yet done so, then we need to send.
596 	 */
597 	if (flags & TH_FIN &&
598 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
599 		goto send;
600 	/*
601 	 * In SACK, it is possible for tcp_output to fail to send a segment
602 	 * after the retransmission timer has been turned off.  Make sure
603 	 * that the retransmission timer is set.
604 	 */
605 	if ((tp->t_flags & TF_SACK_PERMIT) &&
606 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
607 	    !tcp_timer_active(tp, TT_REXMT) &&
608 	    !tcp_timer_active(tp, TT_PERSIST)) {
609 		tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
610 		goto just_return;
611 	}
612 	/*
613 	 * TCP window updates are not reliable, rather a polling protocol
614 	 * using ``persist'' packets is used to insure receipt of window
615 	 * updates.  The three ``states'' for the output side are:
616 	 *	idle			not doing retransmits or persists
617 	 *	persisting		to move a small or zero window
618 	 *	(re)transmitting	and thereby not persisting
619 	 *
620 	 * tcp_timer_active(tp, TT_PERSIST)
621 	 *	is true when we are in persist state.
622 	 * (tp->t_flags & TF_FORCEDATA)
623 	 *	is set when we are called to send a persist packet.
624 	 * tcp_timer_active(tp, TT_REXMT)
625 	 *	is set when we are retransmitting
626 	 * The output side is idle when both timers are zero.
627 	 *
628 	 * If send window is too small, there is data to transmit, and no
629 	 * retransmit or persist is pending, then go to persist state.
630 	 * If nothing happens soon, send when timer expires:
631 	 * if window is nonzero, transmit what we can,
632 	 * otherwise force out a byte.
633 	 */
634 	if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) &&
635 	    !tcp_timer_active(tp, TT_PERSIST)) {
636 		tp->t_rxtshift = 0;
637 		tcp_setpersist(tp);
638 	}
639 
640 	/*
641 	 * No reason to send a segment, just return.
642 	 */
643 just_return:
644 	SOCKBUF_UNLOCK(&so->so_snd);
645 	return (0);
646 
647 send:
648 	SOCKBUF_LOCK_ASSERT(&so->so_snd);
649 	/*
650 	 * Before ESTABLISHED, force sending of initial options
651 	 * unless TCP set not to do any options.
652 	 * NOTE: we assume that the IP/TCP header plus TCP options
653 	 * always fit in a single mbuf, leaving room for a maximum
654 	 * link header, i.e.
655 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
656 	 */
657 	optlen = 0;
658 #ifdef INET6
659 	if (isipv6)
660 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
661 	else
662 #endif
663 	hdrlen = sizeof (struct tcpiphdr);
664 
665 	/*
666 	 * Compute options for segment.
667 	 * We only have to care about SYN and established connection
668 	 * segments.  Options for SYN-ACK segments are handled in TCP
669 	 * syncache.
670 	 */
671 	if ((tp->t_flags & TF_NOOPT) == 0) {
672 		to.to_flags = 0;
673 		/* Maximum segment size. */
674 		if (flags & TH_SYN) {
675 			tp->snd_nxt = tp->iss;
676 			to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
677 			to.to_flags |= TOF_MSS;
678 		}
679 		/* Window scaling. */
680 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
681 			to.to_wscale = tp->request_r_scale;
682 			to.to_flags |= TOF_SCALE;
683 		}
684 		/* Timestamps. */
685 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
686 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
687 			to.to_tsval = ticks + tp->ts_offset;
688 			to.to_tsecr = tp->ts_recent;
689 			to.to_flags |= TOF_TS;
690 			/* Set receive buffer autosizing timestamp. */
691 			if (tp->rfbuf_ts == 0 &&
692 			    (so->so_rcv.sb_flags & SB_AUTOSIZE))
693 				tp->rfbuf_ts = ticks;
694 		}
695 		/* Selective ACK's. */
696 		if (tp->t_flags & TF_SACK_PERMIT) {
697 			if (flags & TH_SYN)
698 				to.to_flags |= TOF_SACKPERM;
699 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
700 			    (tp->t_flags & TF_SACK_PERMIT) &&
701 			    tp->rcv_numsacks > 0) {
702 				to.to_flags |= TOF_SACK;
703 				to.to_nsacks = tp->rcv_numsacks;
704 				to.to_sacks = (u_char *)tp->sackblks;
705 			}
706 		}
707 #ifdef TCP_SIGNATURE
708 		/* TCP-MD5 (RFC2385). */
709 		if (tp->t_flags & TF_SIGNATURE)
710 			to.to_flags |= TOF_SIGNATURE;
711 #endif /* TCP_SIGNATURE */
712 
713 		/* Processing the options. */
714 		hdrlen += optlen = tcp_addoptions(&to, opt);
715 	}
716 
717 #ifdef INET6
718 	if (isipv6)
719 		ipoptlen = ip6_optlen(tp->t_inpcb);
720 	else
721 #endif
722 	if (tp->t_inpcb->inp_options)
723 		ipoptlen = tp->t_inpcb->inp_options->m_len -
724 				offsetof(struct ipoption, ipopt_list);
725 	else
726 		ipoptlen = 0;
727 #ifdef IPSEC
728 	ipoptlen += ipsec_optlen;
729 #endif
730 
731 	/*
732 	 * Adjust data length if insertion of options will
733 	 * bump the packet length beyond the t_maxopd length.
734 	 * Clear the FIN bit because we cut off the tail of
735 	 * the segment.
736 	 *
737 	 * When doing TSO limit a burst to TCP_MAXWIN minus the
738 	 * IP, TCP and Options length to keep ip->ip_len from
739 	 * overflowing.  Prevent the last segment from being
740 	 * fractional thus making them all equal sized and set
741 	 * the flag to continue sending.  TSO is disabled when
742 	 * IP options or IPSEC are present.
743 	 */
744 	if (len + optlen + ipoptlen > tp->t_maxopd) {
745 		flags &= ~TH_FIN;
746 		if (tso) {
747 			if (len > TCP_MAXWIN - hdrlen - optlen) {
748 				len = TCP_MAXWIN - hdrlen - optlen;
749 				len = len - (len % (tp->t_maxopd - optlen));
750 				sendalot = 1;
751 			} else if (tp->t_flags & TF_NEEDFIN)
752 				sendalot = 1;
753 		} else {
754 			len = tp->t_maxopd - optlen - ipoptlen;
755 			sendalot = 1;
756 		}
757 	}
758 
759 /*#ifdef DIAGNOSTIC*/
760 #ifdef INET6
761 	if (max_linkhdr + hdrlen > MCLBYTES)
762 #else
763 	if (max_linkhdr + hdrlen > MHLEN)
764 #endif
765 		panic("tcphdr too big");
766 /*#endif*/
767 
768 	/*
769 	 * This KASSERT is here to catch edge cases at a well defined place.
770 	 * Before, those had triggered (random) panic conditions further down.
771 	 */
772 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
773 
774 	/*
775 	 * Grab a header mbuf, attaching a copy of data to
776 	 * be transmitted, and initialize the header from
777 	 * the template for sends on this connection.
778 	 */
779 	if (len) {
780 		struct mbuf *mb;
781 		u_int moff;
782 
783 		if ((tp->t_flags & TF_FORCEDATA) && len == 1)
784 			TCPSTAT_INC(tcps_sndprobe);
785 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
786 			TCPSTAT_INC(tcps_sndrexmitpack);
787 			TCPSTAT_ADD(tcps_sndrexmitbyte, len);
788 		} else {
789 			TCPSTAT_INC(tcps_sndpack);
790 			TCPSTAT_ADD(tcps_sndbyte, len);
791 		}
792 #ifdef notyet
793 		if ((m = m_copypack(so->so_snd.sb_mb, off,
794 		    (int)len, max_linkhdr + hdrlen)) == 0) {
795 			SOCKBUF_UNLOCK(&so->so_snd);
796 			error = ENOBUFS;
797 			goto out;
798 		}
799 		/*
800 		 * m_copypack left space for our hdr; use it.
801 		 */
802 		m->m_len += hdrlen;
803 		m->m_data -= hdrlen;
804 #else
805 		MGETHDR(m, M_DONTWAIT, MT_DATA);
806 		if (m == NULL) {
807 			SOCKBUF_UNLOCK(&so->so_snd);
808 			error = ENOBUFS;
809 			goto out;
810 		}
811 #ifdef INET6
812 		if (MHLEN < hdrlen + max_linkhdr) {
813 			MCLGET(m, M_DONTWAIT);
814 			if ((m->m_flags & M_EXT) == 0) {
815 				SOCKBUF_UNLOCK(&so->so_snd);
816 				m_freem(m);
817 				error = ENOBUFS;
818 				goto out;
819 			}
820 		}
821 #endif
822 		m->m_data += max_linkhdr;
823 		m->m_len = hdrlen;
824 
825 		/*
826 		 * Start the m_copy functions from the closest mbuf
827 		 * to the offset in the socket buffer chain.
828 		 */
829 		mb = sbsndptr(&so->so_snd, off, len, &moff);
830 
831 		if (len <= MHLEN - hdrlen - max_linkhdr) {
832 			m_copydata(mb, moff, (int)len,
833 			    mtod(m, caddr_t) + hdrlen);
834 			m->m_len += len;
835 		} else {
836 			m->m_next = m_copy(mb, moff, (int)len);
837 			if (m->m_next == NULL) {
838 				SOCKBUF_UNLOCK(&so->so_snd);
839 				(void) m_free(m);
840 				error = ENOBUFS;
841 				goto out;
842 			}
843 		}
844 #endif
845 		/*
846 		 * If we're sending everything we've got, set PUSH.
847 		 * (This will keep happy those implementations which only
848 		 * give data to the user when a buffer fills or
849 		 * a PUSH comes in.)
850 		 */
851 		if (off + len == so->so_snd.sb_cc)
852 			flags |= TH_PUSH;
853 		SOCKBUF_UNLOCK(&so->so_snd);
854 	} else {
855 		SOCKBUF_UNLOCK(&so->so_snd);
856 		if (tp->t_flags & TF_ACKNOW)
857 			TCPSTAT_INC(tcps_sndacks);
858 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
859 			TCPSTAT_INC(tcps_sndctrl);
860 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
861 			TCPSTAT_INC(tcps_sndurg);
862 		else
863 			TCPSTAT_INC(tcps_sndwinup);
864 
865 		MGETHDR(m, M_DONTWAIT, MT_DATA);
866 		if (m == NULL) {
867 			error = ENOBUFS;
868 			goto out;
869 		}
870 #ifdef INET6
871 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
872 		    MHLEN >= hdrlen) {
873 			MH_ALIGN(m, hdrlen);
874 		} else
875 #endif
876 		m->m_data += max_linkhdr;
877 		m->m_len = hdrlen;
878 	}
879 	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
880 	m->m_pkthdr.rcvif = (struct ifnet *)0;
881 #ifdef MAC
882 	mac_inpcb_create_mbuf(tp->t_inpcb, m);
883 #endif
884 #ifdef INET6
885 	if (isipv6) {
886 		ip6 = mtod(m, struct ip6_hdr *);
887 		th = (struct tcphdr *)(ip6 + 1);
888 		tcpip_fillheaders(tp->t_inpcb, ip6, th);
889 	} else
890 #endif /* INET6 */
891 	{
892 		ip = mtod(m, struct ip *);
893 		ipov = (struct ipovly *)ip;
894 		th = (struct tcphdr *)(ip + 1);
895 		tcpip_fillheaders(tp->t_inpcb, ip, th);
896 	}
897 
898 	/*
899 	 * Fill in fields, remembering maximum advertised
900 	 * window for use in delaying messages about window sizes.
901 	 * If resending a FIN, be sure not to use a new sequence number.
902 	 */
903 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
904 	    tp->snd_nxt == tp->snd_max)
905 		tp->snd_nxt--;
906 	/*
907 	 * If we are starting a connection, send ECN setup
908 	 * SYN packet. If we are on a retransmit, we may
909 	 * resend those bits a number of times as per
910 	 * RFC 3168.
911 	 */
912 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
913 		if (tp->t_rxtshift >= 1) {
914 			if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
915 				flags |= TH_ECE|TH_CWR;
916 		} else
917 			flags |= TH_ECE|TH_CWR;
918 	}
919 
920 	if (tp->t_state == TCPS_ESTABLISHED &&
921 	    (tp->t_flags & TF_ECN_PERMIT)) {
922 		/*
923 		 * If the peer has ECN, mark data packets with
924 		 * ECN capable transmission (ECT).
925 		 * Ignore pure ack packets, retransmissions and window probes.
926 		 */
927 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
928 		    !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
929 #ifdef INET6
930 			if (isipv6)
931 				ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
932 			else
933 #endif
934 				ip->ip_tos |= IPTOS_ECN_ECT0;
935 			TCPSTAT_INC(tcps_ecn_ect0);
936 		}
937 
938 		/*
939 		 * Reply with proper ECN notifications.
940 		 */
941 		if (tp->t_flags & TF_ECN_SND_CWR) {
942 			flags |= TH_CWR;
943 			tp->t_flags &= ~TF_ECN_SND_CWR;
944 		}
945 		if (tp->t_flags & TF_ECN_SND_ECE)
946 			flags |= TH_ECE;
947 	}
948 
949 	/*
950 	 * If we are doing retransmissions, then snd_nxt will
951 	 * not reflect the first unsent octet.  For ACK only
952 	 * packets, we do not want the sequence number of the
953 	 * retransmitted packet, we want the sequence number
954 	 * of the next unsent octet.  So, if there is no data
955 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
956 	 * when filling in ti_seq.  But if we are in persist
957 	 * state, snd_max might reflect one byte beyond the
958 	 * right edge of the window, so use snd_nxt in that
959 	 * case, since we know we aren't doing a retransmission.
960 	 * (retransmit and persist are mutually exclusive...)
961 	 */
962 	if (sack_rxmit == 0) {
963 		if (len || (flags & (TH_SYN|TH_FIN)) ||
964 		    tcp_timer_active(tp, TT_PERSIST))
965 			th->th_seq = htonl(tp->snd_nxt);
966 		else
967 			th->th_seq = htonl(tp->snd_max);
968 	} else {
969 		th->th_seq = htonl(p->rxmit);
970 		p->rxmit += len;
971 		tp->sackhint.sack_bytes_rexmit += len;
972 	}
973 	th->th_ack = htonl(tp->rcv_nxt);
974 	if (optlen) {
975 		bcopy(opt, th + 1, optlen);
976 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
977 	}
978 	th->th_flags = flags;
979 	/*
980 	 * Calculate receive window.  Don't shrink window,
981 	 * but avoid silly window syndrome.
982 	 */
983 	if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
984 	    recwin < (long)tp->t_maxseg)
985 		recwin = 0;
986 	if (recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
987 		recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
988 	if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
989 		recwin = (long)TCP_MAXWIN << tp->rcv_scale;
990 
991 	/*
992 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
993 	 * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
994 	 * case is handled in syncache.
995 	 */
996 	if (flags & TH_SYN)
997 		th->th_win = htons((u_short)
998 				(min(sbspace(&so->so_rcv), TCP_MAXWIN)));
999 	else
1000 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1001 
1002 	/*
1003 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1004 	 * a 0 window.  This may cause the remote transmitter to stall.  This
1005 	 * flag tells soreceive() to disable delayed acknowledgements when
1006 	 * draining the buffer.  This can occur if the receiver is attempting
1007 	 * to read more data than can be buffered prior to transmitting on
1008 	 * the connection.
1009 	 */
1010 	if (th->th_win == 0)
1011 		tp->t_flags |= TF_RXWIN0SENT;
1012 	else
1013 		tp->t_flags &= ~TF_RXWIN0SENT;
1014 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1015 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1016 		th->th_flags |= TH_URG;
1017 	} else
1018 		/*
1019 		 * If no urgent pointer to send, then we pull
1020 		 * the urgent pointer to the left edge of the send window
1021 		 * so that it doesn't drift into the send window on sequence
1022 		 * number wraparound.
1023 		 */
1024 		tp->snd_up = tp->snd_una;		/* drag it along */
1025 
1026 #ifdef TCP_SIGNATURE
1027 	if (tp->t_flags & TF_SIGNATURE) {
1028 		int sigoff = to.to_signature - opt;
1029 		tcp_signature_compute(m, 0, len, optlen,
1030 		    (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
1031 	}
1032 #endif
1033 
1034 	/*
1035 	 * Put TCP length in extended header, and then
1036 	 * checksum extended header and data.
1037 	 */
1038 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1039 #ifdef INET6
1040 	if (isipv6)
1041 		/*
1042 		 * ip6_plen is not need to be filled now, and will be filled
1043 		 * in ip6_output.
1044 		 */
1045 		th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
1046 				       sizeof(struct tcphdr) + optlen + len);
1047 	else
1048 #endif /* INET6 */
1049 	{
1050 		m->m_pkthdr.csum_flags = CSUM_TCP;
1051 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1052 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1053 		    htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1054 
1055 		/* IP version must be set here for ipv4/ipv6 checking later */
1056 		KASSERT(ip->ip_v == IPVERSION,
1057 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1058 	}
1059 
1060 	/*
1061 	 * Enable TSO and specify the size of the segments.
1062 	 * The TCP pseudo header checksum is always provided.
1063 	 * XXX: Fixme: This is currently not the case for IPv6.
1064 	 */
1065 	if (tso) {
1066 		KASSERT(len > tp->t_maxopd - optlen,
1067 		    ("%s: len <= tso_segsz", __func__));
1068 		m->m_pkthdr.csum_flags |= CSUM_TSO;
1069 		m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen;
1070 	}
1071 
1072 	/*
1073 	 * In transmit state, time the transmission and arrange for
1074 	 * the retransmit.  In persist state, just set snd_max.
1075 	 */
1076 	if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1077 	    !tcp_timer_active(tp, TT_PERSIST)) {
1078 		tcp_seq startseq = tp->snd_nxt;
1079 
1080 		/*
1081 		 * Advance snd_nxt over sequence space of this segment.
1082 		 */
1083 		if (flags & (TH_SYN|TH_FIN)) {
1084 			if (flags & TH_SYN)
1085 				tp->snd_nxt++;
1086 			if (flags & TH_FIN) {
1087 				tp->snd_nxt++;
1088 				tp->t_flags |= TF_SENTFIN;
1089 			}
1090 		}
1091 		if (sack_rxmit)
1092 			goto timer;
1093 		tp->snd_nxt += len;
1094 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1095 			tp->snd_max = tp->snd_nxt;
1096 			/*
1097 			 * Time this transmission if not a retransmission and
1098 			 * not currently timing anything.
1099 			 */
1100 			if (tp->t_rtttime == 0) {
1101 				tp->t_rtttime = ticks;
1102 				tp->t_rtseq = startseq;
1103 				TCPSTAT_INC(tcps_segstimed);
1104 			}
1105 		}
1106 
1107 		/*
1108 		 * Set retransmit timer if not currently set,
1109 		 * and not doing a pure ack or a keep-alive probe.
1110 		 * Initial value for retransmit timer is smoothed
1111 		 * round-trip time + 2 * round-trip time variance.
1112 		 * Initialize shift counter which is used for backoff
1113 		 * of retransmit time.
1114 		 */
1115 timer:
1116 		if (!tcp_timer_active(tp, TT_REXMT) &&
1117 		    ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1118 		     (tp->snd_nxt != tp->snd_una))) {
1119 			if (tcp_timer_active(tp, TT_PERSIST)) {
1120 				tcp_timer_activate(tp, TT_PERSIST, 0);
1121 				tp->t_rxtshift = 0;
1122 			}
1123 			tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1124 		}
1125 	} else {
1126 		/*
1127 		 * Persist case, update snd_max but since we are in
1128 		 * persist mode (no window) we do not update snd_nxt.
1129 		 */
1130 		int xlen = len;
1131 		if (flags & TH_SYN)
1132 			++xlen;
1133 		if (flags & TH_FIN) {
1134 			++xlen;
1135 			tp->t_flags |= TF_SENTFIN;
1136 		}
1137 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1138 			tp->snd_max = tp->snd_nxt + len;
1139 	}
1140 
1141 #ifdef TCPDEBUG
1142 	/*
1143 	 * Trace.
1144 	 */
1145 	if (so->so_options & SO_DEBUG) {
1146 		u_short save = 0;
1147 #ifdef INET6
1148 		if (!isipv6)
1149 #endif
1150 		{
1151 			save = ipov->ih_len;
1152 			ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1153 		}
1154 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1155 #ifdef INET6
1156 		if (!isipv6)
1157 #endif
1158 		ipov->ih_len = save;
1159 	}
1160 #endif
1161 
1162 	/*
1163 	 * Fill in IP length and desired time to live and
1164 	 * send to IP level.  There should be a better way
1165 	 * to handle ttl and tos; we could keep them in
1166 	 * the template, but need a way to checksum without them.
1167 	 */
1168 	/*
1169 	 * m->m_pkthdr.len should have been set before cksum calcuration,
1170 	 * because in6_cksum() need it.
1171 	 */
1172 #ifdef INET6
1173 	if (isipv6) {
1174 		/*
1175 		 * we separately set hoplimit for every segment, since the
1176 		 * user might want to change the value via setsockopt.
1177 		 * Also, desired default hop limit might be changed via
1178 		 * Neighbor Discovery.
1179 		 */
1180 		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1181 
1182 		/* TODO: IPv6 IP6TOS_ECT bit on */
1183 		error = ip6_output(m,
1184 			    tp->t_inpcb->in6p_outputopts, NULL,
1185 			    ((so->so_options & SO_DONTROUTE) ?
1186 			    IP_ROUTETOIF : 0), NULL, NULL, tp->t_inpcb);
1187 	} else
1188 #endif /* INET6 */
1189     {
1190 	ip->ip_len = m->m_pkthdr.len;
1191 #ifdef INET6
1192 	if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1193 		ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1194 #endif /* INET6 */
1195 	/*
1196 	 * If we do path MTU discovery, then we set DF on every packet.
1197 	 * This might not be the best thing to do according to RFC3390
1198 	 * Section 2. However the tcp hostcache migitates the problem
1199 	 * so it affects only the first tcp connection with a host.
1200 	 *
1201 	 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1202 	 */
1203 	if (V_path_mtu_discovery && tp->t_maxopd > V_tcp_minmss)
1204 		ip->ip_off |= IP_DF;
1205 
1206 	error = ip_output(m, tp->t_inpcb->inp_options, NULL,
1207 	    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1208 	    tp->t_inpcb);
1209     }
1210 	if (error) {
1211 
1212 		/*
1213 		 * We know that the packet was lost, so back out the
1214 		 * sequence number advance, if any.
1215 		 *
1216 		 * If the error is EPERM the packet got blocked by the
1217 		 * local firewall.  Normally we should terminate the
1218 		 * connection but the blocking may have been spurious
1219 		 * due to a firewall reconfiguration cycle.  So we treat
1220 		 * it like a packet loss and let the retransmit timer and
1221 		 * timeouts do their work over time.
1222 		 * XXX: It is a POLA question whether calling tcp_drop right
1223 		 * away would be the really correct behavior instead.
1224 		 */
1225 		if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1226 		    !tcp_timer_active(tp, TT_PERSIST)) &&
1227 		    ((flags & TH_SYN) == 0) &&
1228 		    (error != EPERM)) {
1229 			if (sack_rxmit) {
1230 				p->rxmit -= len;
1231 				tp->sackhint.sack_bytes_rexmit -= len;
1232 				KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1233 				    ("sackhint bytes rtx >= 0"));
1234 			} else
1235 				tp->snd_nxt -= len;
1236 		}
1237 out:
1238 		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1239 		switch (error) {
1240 		case EPERM:
1241 			tp->t_softerror = error;
1242 			return (error);
1243 		case ENOBUFS:
1244 	                if (!tcp_timer_active(tp, TT_REXMT) &&
1245 			    !tcp_timer_active(tp, TT_PERSIST))
1246 	                        tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1247 			tp->snd_cwnd = tp->t_maxseg;
1248 			return (0);
1249 		case EMSGSIZE:
1250 			/*
1251 			 * For some reason the interface we used initially
1252 			 * to send segments changed to another or lowered
1253 			 * its MTU.
1254 			 *
1255 			 * tcp_mtudisc() will find out the new MTU and as
1256 			 * its last action, initiate retransmission, so it
1257 			 * is important to not do so here.
1258 			 *
1259 			 * If TSO was active we either got an interface
1260 			 * without TSO capabilits or TSO was turned off.
1261 			 * Disable it for this connection as too and
1262 			 * immediatly retry with MSS sized segments generated
1263 			 * by this function.
1264 			 */
1265 			if (tso)
1266 				tp->t_flags &= ~TF_TSO;
1267 			tcp_mtudisc(tp->t_inpcb, 0);
1268 			return (0);
1269 		case EHOSTDOWN:
1270 		case EHOSTUNREACH:
1271 		case ENETDOWN:
1272 		case ENETUNREACH:
1273 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
1274 				tp->t_softerror = error;
1275 				return (0);
1276 			}
1277 			/* FALLTHROUGH */
1278 		default:
1279 			return (error);
1280 		}
1281 	}
1282 	TCPSTAT_INC(tcps_sndtotal);
1283 
1284 	/*
1285 	 * Data sent (as far as we can tell).
1286 	 * If this advertises a larger window than any other segment,
1287 	 * then remember the size of the advertised window.
1288 	 * Any pending ACK has now been sent.
1289 	 */
1290 	if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1291 		tp->rcv_adv = tp->rcv_nxt + recwin;
1292 	tp->last_ack_sent = tp->rcv_nxt;
1293 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1294 	if (tcp_timer_active(tp, TT_DELACK))
1295 		tcp_timer_activate(tp, TT_DELACK, 0);
1296 #if 0
1297 	/*
1298 	 * This completely breaks TCP if newreno is turned on.  What happens
1299 	 * is that if delayed-acks are turned on on the receiver, this code
1300 	 * on the transmitter effectively destroys the TCP window, forcing
1301 	 * it to four packets (1.5Kx4 = 6K window).
1302 	 */
1303 	if (sendalot && (!V_tcp_do_newreno || --maxburst))
1304 		goto again;
1305 #endif
1306 	if (sendalot)
1307 		goto again;
1308 	return (0);
1309 }
1310 
1311 void
1312 tcp_setpersist(struct tcpcb *tp)
1313 {
1314 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1315 	int tt;
1316 
1317 	if (tcp_timer_active(tp, TT_REXMT))
1318 		panic("tcp_setpersist: retransmit pending");
1319 	/*
1320 	 * Start/restart persistance timer.
1321 	 */
1322 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1323 		      TCPTV_PERSMIN, TCPTV_PERSMAX);
1324 	tcp_timer_activate(tp, TT_PERSIST, tt);
1325 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1326 		tp->t_rxtshift++;
1327 }
1328 
1329 /*
1330  * Insert TCP options according to the supplied parameters to the place
1331  * optp in a consistent way.  Can handle unaligned destinations.
1332  *
1333  * The order of the option processing is crucial for optimal packing and
1334  * alignment for the scarce option space.
1335  *
1336  * The optimal order for a SYN/SYN-ACK segment is:
1337  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1338  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1339  *
1340  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1341  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1342  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1343  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1344  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1345  */
1346 int
1347 tcp_addoptions(struct tcpopt *to, u_char *optp)
1348 {
1349 	u_int mask, optlen = 0;
1350 
1351 	for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1352 		if ((to->to_flags & mask) != mask)
1353 			continue;
1354 		if (optlen == TCP_MAXOLEN)
1355 			break;
1356 		switch (to->to_flags & mask) {
1357 		case TOF_MSS:
1358 			while (optlen % 4) {
1359 				optlen += TCPOLEN_NOP;
1360 				*optp++ = TCPOPT_NOP;
1361 			}
1362 			if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1363 				continue;
1364 			optlen += TCPOLEN_MAXSEG;
1365 			*optp++ = TCPOPT_MAXSEG;
1366 			*optp++ = TCPOLEN_MAXSEG;
1367 			to->to_mss = htons(to->to_mss);
1368 			bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1369 			optp += sizeof(to->to_mss);
1370 			break;
1371 		case TOF_SCALE:
1372 			while (!optlen || optlen % 2 != 1) {
1373 				optlen += TCPOLEN_NOP;
1374 				*optp++ = TCPOPT_NOP;
1375 			}
1376 			if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1377 				continue;
1378 			optlen += TCPOLEN_WINDOW;
1379 			*optp++ = TCPOPT_WINDOW;
1380 			*optp++ = TCPOLEN_WINDOW;
1381 			*optp++ = to->to_wscale;
1382 			break;
1383 		case TOF_SACKPERM:
1384 			while (optlen % 2) {
1385 				optlen += TCPOLEN_NOP;
1386 				*optp++ = TCPOPT_NOP;
1387 			}
1388 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1389 				continue;
1390 			optlen += TCPOLEN_SACK_PERMITTED;
1391 			*optp++ = TCPOPT_SACK_PERMITTED;
1392 			*optp++ = TCPOLEN_SACK_PERMITTED;
1393 			break;
1394 		case TOF_TS:
1395 			while (!optlen || optlen % 4 != 2) {
1396 				optlen += TCPOLEN_NOP;
1397 				*optp++ = TCPOPT_NOP;
1398 			}
1399 			if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1400 				continue;
1401 			optlen += TCPOLEN_TIMESTAMP;
1402 			*optp++ = TCPOPT_TIMESTAMP;
1403 			*optp++ = TCPOLEN_TIMESTAMP;
1404 			to->to_tsval = htonl(to->to_tsval);
1405 			to->to_tsecr = htonl(to->to_tsecr);
1406 			bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1407 			optp += sizeof(to->to_tsval);
1408 			bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1409 			optp += sizeof(to->to_tsecr);
1410 			break;
1411 		case TOF_SIGNATURE:
1412 			{
1413 			int siglen = TCPOLEN_SIGNATURE - 2;
1414 
1415 			while (!optlen || optlen % 4 != 2) {
1416 				optlen += TCPOLEN_NOP;
1417 				*optp++ = TCPOPT_NOP;
1418 			}
1419 			if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE)
1420 				continue;
1421 			optlen += TCPOLEN_SIGNATURE;
1422 			*optp++ = TCPOPT_SIGNATURE;
1423 			*optp++ = TCPOLEN_SIGNATURE;
1424 			to->to_signature = optp;
1425 			while (siglen--)
1426 				 *optp++ = 0;
1427 			break;
1428 			}
1429 		case TOF_SACK:
1430 			{
1431 			int sackblks = 0;
1432 			struct sackblk *sack = (struct sackblk *)to->to_sacks;
1433 			tcp_seq sack_seq;
1434 
1435 			while (!optlen || optlen % 4 != 2) {
1436 				optlen += TCPOLEN_NOP;
1437 				*optp++ = TCPOPT_NOP;
1438 			}
1439 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1440 				continue;
1441 			optlen += TCPOLEN_SACKHDR;
1442 			*optp++ = TCPOPT_SACK;
1443 			sackblks = min(to->to_nsacks,
1444 					(TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1445 			*optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1446 			while (sackblks--) {
1447 				sack_seq = htonl(sack->start);
1448 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1449 				optp += sizeof(sack_seq);
1450 				sack_seq = htonl(sack->end);
1451 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1452 				optp += sizeof(sack_seq);
1453 				optlen += TCPOLEN_SACK;
1454 				sack++;
1455 			}
1456 			TCPSTAT_INC(tcps_sack_send_blocks);
1457 			break;
1458 			}
1459 		default:
1460 			panic("%s: unknown TCP option type", __func__);
1461 			break;
1462 		}
1463 	}
1464 
1465 	/* Terminate and pad TCP options to a 4 byte boundary. */
1466 	if (optlen % 4) {
1467 		optlen += TCPOLEN_EOL;
1468 		*optp++ = TCPOPT_EOL;
1469 	}
1470 	/*
1471 	 * According to RFC 793 (STD0007):
1472 	 *   "The content of the header beyond the End-of-Option option
1473 	 *    must be header padding (i.e., zero)."
1474 	 *   and later: "The padding is composed of zeros."
1475 	 */
1476 	while (optlen % 4) {
1477 		optlen += TCPOLEN_PAD;
1478 		*optp++ = TCPOPT_PAD;
1479 	}
1480 
1481 	KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1482 	return (optlen);
1483 }
1484