xref: /freebsd/sys/netinet/tcp_output.c (revision 5dcd9c10612684d1c823670cbb5b4715028784e7)
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/hhook.h>
44 #include <sys/kernel.h>
45 #include <sys/lock.h>
46 #include <sys/mbuf.h>
47 #include <sys/mutex.h>
48 #include <sys/protosw.h>
49 #include <sys/socket.h>
50 #include <sys/socketvar.h>
51 #include <sys/sysctl.h>
52 
53 #include <net/if.h>
54 #include <net/route.h>
55 #include <net/vnet.h>
56 
57 #include <netinet/cc.h>
58 #include <netinet/in.h>
59 #include <netinet/in_systm.h>
60 #include <netinet/ip.h>
61 #include <netinet/in_pcb.h>
62 #include <netinet/ip_var.h>
63 #include <netinet/ip_options.h>
64 #ifdef INET6
65 #include <netinet6/in6_pcb.h>
66 #include <netinet/ip6.h>
67 #include <netinet6/ip6_var.h>
68 #endif
69 #define	TCPOUTFLAGS
70 #include <netinet/tcp_fsm.h>
71 #include <netinet/tcp_seq.h>
72 #include <netinet/tcp_timer.h>
73 #include <netinet/tcp_var.h>
74 #include <netinet/tcpip.h>
75 #ifdef TCPDEBUG
76 #include <netinet/tcp_debug.h>
77 #endif
78 
79 #ifdef IPSEC
80 #include <netipsec/ipsec.h>
81 #endif /*IPSEC*/
82 
83 #include <machine/in_cksum.h>
84 
85 #include <security/mac/mac_framework.h>
86 
87 #ifdef notyet
88 extern struct mbuf *m_copypack();
89 #endif
90 
91 VNET_DEFINE(int, path_mtu_discovery) = 1;
92 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
93 	&VNET_NAME(path_mtu_discovery), 1,
94 	"Enable Path MTU Discovery");
95 
96 VNET_DEFINE(int, ss_fltsz) = 1;
97 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, slowstart_flightsize, CTLFLAG_RW,
98 	&VNET_NAME(ss_fltsz), 1,
99 	"Slow start flight size");
100 
101 VNET_DEFINE(int, ss_fltsz_local) = 4;
102 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, local_slowstart_flightsize,
103 	CTLFLAG_RW, &VNET_NAME(ss_fltsz_local), 1,
104 	"Slow start flight size for local networks");
105 
106 VNET_DEFINE(int, tcp_do_tso) = 1;
107 #define	V_tcp_do_tso		VNET(tcp_do_tso)
108 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW,
109 	&VNET_NAME(tcp_do_tso), 0,
110 	"Enable TCP Segmentation Offload");
111 
112 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
113 #define	V_tcp_do_autosndbuf	VNET(tcp_do_autosndbuf)
114 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW,
115 	&VNET_NAME(tcp_do_autosndbuf), 0,
116 	"Enable automatic send buffer sizing");
117 
118 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
119 #define	V_tcp_autosndbuf_inc	VNET(tcp_autosndbuf_inc)
120 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW,
121 	&VNET_NAME(tcp_autosndbuf_inc), 0,
122 	"Incrementor step size of automatic send buffer");
123 
124 VNET_DEFINE(int, tcp_autosndbuf_max) = 256*1024;
125 #define	V_tcp_autosndbuf_max	VNET(tcp_autosndbuf_max)
126 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW,
127 	&VNET_NAME(tcp_autosndbuf_max), 0,
128 	"Max size of automatic send buffer");
129 
130 static void inline	hhook_run_tcp_est_out(struct tcpcb *tp,
131 			    struct tcphdr *th, struct tcpopt *to,
132 			    long len, int tso);
133 static void inline	cc_after_idle(struct tcpcb *tp);
134 
135 /*
136  * Wrapper for the TCP established ouput helper hook.
137  */
138 static void inline
139 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
140     struct tcpopt *to, long len, int tso)
141 {
142 	struct tcp_hhook_data hhook_data;
143 
144 	if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
145 		hhook_data.tp = tp;
146 		hhook_data.th = th;
147 		hhook_data.to = to;
148 		hhook_data.len = len;
149 		hhook_data.tso = tso;
150 
151 		hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
152 		    tp->osd);
153 	}
154 }
155 
156 /*
157  * CC wrapper hook functions
158  */
159 static void inline
160 cc_after_idle(struct tcpcb *tp)
161 {
162 	INP_WLOCK_ASSERT(tp->t_inpcb);
163 
164 	if (CC_ALGO(tp)->after_idle != NULL)
165 		CC_ALGO(tp)->after_idle(tp->ccv);
166 }
167 
168 /*
169  * Tcp output routine: figure out what should be sent and send it.
170  */
171 int
172 tcp_output(struct tcpcb *tp)
173 {
174 	struct socket *so = tp->t_inpcb->inp_socket;
175 	long len, recwin, sendwin;
176 	int off, flags, error;
177 	struct mbuf *m;
178 	struct ip *ip = NULL;
179 	struct ipovly *ipov = NULL;
180 	struct tcphdr *th;
181 	u_char opt[TCP_MAXOLEN];
182 	unsigned ipoptlen, optlen, hdrlen;
183 #ifdef IPSEC
184 	unsigned ipsec_optlen = 0;
185 #endif
186 	int idle, sendalot;
187 	int sack_rxmit, sack_bytes_rxmt;
188 	struct sackhole *p;
189 	int tso;
190 	struct tcpopt to;
191 #if 0
192 	int maxburst = TCP_MAXBURST;
193 #endif
194 #ifdef INET6
195 	struct ip6_hdr *ip6 = NULL;
196 	int isipv6;
197 
198 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
199 #endif
200 
201 	INP_WLOCK_ASSERT(tp->t_inpcb);
202 
203 	/*
204 	 * Determine length of data that should be transmitted,
205 	 * and flags that will be used.
206 	 * If there is some data or critical controls (SYN, RST)
207 	 * to send, then transmit; otherwise, investigate further.
208 	 */
209 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
210 	if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur)
211 		cc_after_idle(tp);
212 	tp->t_flags &= ~TF_LASTIDLE;
213 	if (idle) {
214 		if (tp->t_flags & TF_MORETOCOME) {
215 			tp->t_flags |= TF_LASTIDLE;
216 			idle = 0;
217 		}
218 	}
219 again:
220 	/*
221 	 * If we've recently taken a timeout, snd_max will be greater than
222 	 * snd_nxt.  There may be SACK information that allows us to avoid
223 	 * resending already delivered data.  Adjust snd_nxt accordingly.
224 	 */
225 	if ((tp->t_flags & TF_SACK_PERMIT) &&
226 	    SEQ_LT(tp->snd_nxt, tp->snd_max))
227 		tcp_sack_adjust(tp);
228 	sendalot = 0;
229 	tso = 0;
230 	off = tp->snd_nxt - tp->snd_una;
231 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
232 
233 	flags = tcp_outflags[tp->t_state];
234 	/*
235 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
236 	 * to send out new data (when sendalot is 1), bypass this function.
237 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
238 	 * we're replacing a (future) new transmission with a retransmission
239 	 * now, and we previously incremented snd_cwnd in tcp_input().
240 	 */
241 	/*
242 	 * Still in sack recovery , reset rxmit flag to zero.
243 	 */
244 	sack_rxmit = 0;
245 	sack_bytes_rxmt = 0;
246 	len = 0;
247 	p = NULL;
248 	if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
249 	    (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
250 		long cwin;
251 
252 		cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
253 		if (cwin < 0)
254 			cwin = 0;
255 		/* Do not retransmit SACK segments beyond snd_recover */
256 		if (SEQ_GT(p->end, tp->snd_recover)) {
257 			/*
258 			 * (At least) part of sack hole extends beyond
259 			 * snd_recover. Check to see if we can rexmit data
260 			 * for this hole.
261 			 */
262 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
263 				/*
264 				 * Can't rexmit any more data for this hole.
265 				 * That data will be rexmitted in the next
266 				 * sack recovery episode, when snd_recover
267 				 * moves past p->rxmit.
268 				 */
269 				p = NULL;
270 				goto after_sack_rexmit;
271 			} else
272 				/* Can rexmit part of the current hole */
273 				len = ((long)ulmin(cwin,
274 						   tp->snd_recover - p->rxmit));
275 		} else
276 			len = ((long)ulmin(cwin, p->end - p->rxmit));
277 		off = p->rxmit - tp->snd_una;
278 		KASSERT(off >= 0,("%s: sack block to the left of una : %d",
279 		    __func__, off));
280 		if (len > 0) {
281 			sack_rxmit = 1;
282 			sendalot = 1;
283 			TCPSTAT_INC(tcps_sack_rexmits);
284 			TCPSTAT_ADD(tcps_sack_rexmit_bytes,
285 			    min(len, tp->t_maxseg));
286 		}
287 	}
288 after_sack_rexmit:
289 	/*
290 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
291 	 * state flags.
292 	 */
293 	if (tp->t_flags & TF_NEEDFIN)
294 		flags |= TH_FIN;
295 	if (tp->t_flags & TF_NEEDSYN)
296 		flags |= TH_SYN;
297 
298 	SOCKBUF_LOCK(&so->so_snd);
299 	/*
300 	 * If in persist timeout with window of 0, send 1 byte.
301 	 * Otherwise, if window is small but nonzero
302 	 * and timer expired, we will send what we can
303 	 * and go to transmit state.
304 	 */
305 	if (tp->t_flags & TF_FORCEDATA) {
306 		if (sendwin == 0) {
307 			/*
308 			 * If we still have some data to send, then
309 			 * clear the FIN bit.  Usually this would
310 			 * happen below when it realizes that we
311 			 * aren't sending all the data.  However,
312 			 * if we have exactly 1 byte of unsent data,
313 			 * then it won't clear the FIN bit below,
314 			 * and if we are in persist state, we wind
315 			 * up sending the packet without recording
316 			 * that we sent the FIN bit.
317 			 *
318 			 * We can't just blindly clear the FIN bit,
319 			 * because if we don't have any more data
320 			 * to send then the probe will be the FIN
321 			 * itself.
322 			 */
323 			if (off < so->so_snd.sb_cc)
324 				flags &= ~TH_FIN;
325 			sendwin = 1;
326 		} else {
327 			tcp_timer_activate(tp, TT_PERSIST, 0);
328 			tp->t_rxtshift = 0;
329 		}
330 	}
331 
332 	/*
333 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
334 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
335 	 * a negative length.  This can also occur when TCP opens up
336 	 * its congestion window while receiving additional duplicate
337 	 * acks after fast-retransmit because TCP will reset snd_nxt
338 	 * to snd_max after the fast-retransmit.
339 	 *
340 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
341 	 * be set to snd_una, the offset will be 0, and the length may
342 	 * wind up 0.
343 	 *
344 	 * If sack_rxmit is true we are retransmitting from the scoreboard
345 	 * in which case len is already set.
346 	 */
347 	if (sack_rxmit == 0) {
348 		if (sack_bytes_rxmt == 0)
349 			len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off);
350 		else {
351 			long cwin;
352 
353                         /*
354 			 * We are inside of a SACK recovery episode and are
355 			 * sending new data, having retransmitted all the
356 			 * data possible in the scoreboard.
357 			 */
358 			len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd)
359 			       - off);
360 			/*
361 			 * Don't remove this (len > 0) check !
362 			 * We explicitly check for len > 0 here (although it
363 			 * isn't really necessary), to work around a gcc
364 			 * optimization issue - to force gcc to compute
365 			 * len above. Without this check, the computation
366 			 * of len is bungled by the optimizer.
367 			 */
368 			if (len > 0) {
369 				cwin = tp->snd_cwnd -
370 					(tp->snd_nxt - tp->sack_newdata) -
371 					sack_bytes_rxmt;
372 				if (cwin < 0)
373 					cwin = 0;
374 				len = lmin(len, cwin);
375 			}
376 		}
377 	}
378 
379 	/*
380 	 * Lop off SYN bit if it has already been sent.  However, if this
381 	 * is SYN-SENT state and if segment contains data and if we don't
382 	 * know that foreign host supports TAO, suppress sending segment.
383 	 */
384 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
385 		if (tp->t_state != TCPS_SYN_RECEIVED)
386 			flags &= ~TH_SYN;
387 		off--, len++;
388 	}
389 
390 	/*
391 	 * Be careful not to send data and/or FIN on SYN segments.
392 	 * This measure is needed to prevent interoperability problems
393 	 * with not fully conformant TCP implementations.
394 	 */
395 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
396 		len = 0;
397 		flags &= ~TH_FIN;
398 	}
399 
400 	if (len < 0) {
401 		/*
402 		 * If FIN has been sent but not acked,
403 		 * but we haven't been called to retransmit,
404 		 * len will be < 0.  Otherwise, window shrank
405 		 * after we sent into it.  If window shrank to 0,
406 		 * cancel pending retransmit, pull snd_nxt back
407 		 * to (closed) window, and set the persist timer
408 		 * if it isn't already going.  If the window didn't
409 		 * close completely, just wait for an ACK.
410 		 */
411 		len = 0;
412 		if (sendwin == 0) {
413 			tcp_timer_activate(tp, TT_REXMT, 0);
414 			tp->t_rxtshift = 0;
415 			tp->snd_nxt = tp->snd_una;
416 			if (!tcp_timer_active(tp, TT_PERSIST))
417 				tcp_setpersist(tp);
418 		}
419 	}
420 
421 	/* len will be >= 0 after this point. */
422 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
423 
424 	/*
425 	 * Automatic sizing of send socket buffer.  Often the send buffer
426 	 * size is not optimally adjusted to the actual network conditions
427 	 * at hand (delay bandwidth product).  Setting the buffer size too
428 	 * small limits throughput on links with high bandwidth and high
429 	 * delay (eg. trans-continental/oceanic links).  Setting the
430 	 * buffer size too big consumes too much real kernel memory,
431 	 * especially with many connections on busy servers.
432 	 *
433 	 * The criteria to step up the send buffer one notch are:
434 	 *  1. receive window of remote host is larger than send buffer
435 	 *     (with a fudge factor of 5/4th);
436 	 *  2. send buffer is filled to 7/8th with data (so we actually
437 	 *     have data to make use of it);
438 	 *  3. send buffer fill has not hit maximal automatic size;
439 	 *  4. our send window (slow start and cogestion controlled) is
440 	 *     larger than sent but unacknowledged data in send buffer.
441 	 *
442 	 * The remote host receive window scaling factor may limit the
443 	 * growing of the send buffer before it reaches its allowed
444 	 * maximum.
445 	 *
446 	 * It scales directly with slow start or congestion window
447 	 * and does at most one step per received ACK.  This fast
448 	 * scaling has the drawback of growing the send buffer beyond
449 	 * what is strictly necessary to make full use of a given
450 	 * delay*bandwith product.  However testing has shown this not
451 	 * to be much of an problem.  At worst we are trading wasting
452 	 * of available bandwith (the non-use of it) for wasting some
453 	 * socket buffer memory.
454 	 *
455 	 * TODO: Shrink send buffer during idle periods together
456 	 * with congestion window.  Requires another timer.  Has to
457 	 * wait for upcoming tcp timer rewrite.
458 	 */
459 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
460 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
461 		    so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
462 		    so->so_snd.sb_cc < V_tcp_autosndbuf_max &&
463 		    sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
464 			if (!sbreserve_locked(&so->so_snd,
465 			    min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
466 			     V_tcp_autosndbuf_max), so, curthread))
467 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
468 		}
469 	}
470 
471 	/*
472 	 * Decide if we can use TCP Segmentation Offloading (if supported by
473 	 * hardware).
474 	 *
475 	 * TSO may only be used if we are in a pure bulk sending state.  The
476 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
477 	 * IP options prevent using TSO.  With TSO the TCP header is the same
478 	 * (except for the sequence number) for all generated packets.  This
479 	 * makes it impossible to transmit any options which vary per generated
480 	 * segment or packet.
481 	 */
482 #ifdef IPSEC
483 	/*
484 	 * Pre-calculate here as we save another lookup into the darknesses
485 	 * of IPsec that way and can actually decide if TSO is ok.
486 	 */
487 	ipsec_optlen = ipsec_hdrsiz_tcp(tp);
488 #endif
489 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
490 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
491 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
492 #ifdef IPSEC
493 	    ipsec_optlen == 0 &&
494 #endif
495 	    tp->t_inpcb->inp_options == NULL &&
496 	    tp->t_inpcb->in6p_options == NULL)
497 		tso = 1;
498 
499 	if (sack_rxmit) {
500 		if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
501 			flags &= ~TH_FIN;
502 	} else {
503 		if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
504 			flags &= ~TH_FIN;
505 	}
506 
507 	recwin = sbspace(&so->so_rcv);
508 
509 	/*
510 	 * Sender silly window avoidance.   We transmit under the following
511 	 * conditions when len is non-zero:
512 	 *
513 	 *	- We have a full segment (or more with TSO)
514 	 *	- This is the last buffer in a write()/send() and we are
515 	 *	  either idle or running NODELAY
516 	 *	- we've timed out (e.g. persist timer)
517 	 *	- we have more then 1/2 the maximum send window's worth of
518 	 *	  data (receiver may be limited the window size)
519 	 *	- we need to retransmit
520 	 */
521 	if (len) {
522 		if (len >= tp->t_maxseg)
523 			goto send;
524 		/*
525 		 * NOTE! on localhost connections an 'ack' from the remote
526 		 * end may occur synchronously with the output and cause
527 		 * us to flush a buffer queued with moretocome.  XXX
528 		 *
529 		 * note: the len + off check is almost certainly unnecessary.
530 		 */
531 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
532 		    (idle || (tp->t_flags & TF_NODELAY)) &&
533 		    len + off >= so->so_snd.sb_cc &&
534 		    (tp->t_flags & TF_NOPUSH) == 0) {
535 			goto send;
536 		}
537 		if (tp->t_flags & TF_FORCEDATA)		/* typ. timeout case */
538 			goto send;
539 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
540 			goto send;
541 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
542 			goto send;
543 		if (sack_rxmit)
544 			goto send;
545 	}
546 
547 	/*
548 	 * Compare available window to amount of window
549 	 * known to peer (as advertised window less
550 	 * next expected input).  If the difference is at least two
551 	 * max size segments, or at least 50% of the maximum possible
552 	 * window, then want to send a window update to peer.
553 	 * Skip this if the connection is in T/TCP half-open state.
554 	 * Don't send pure window updates when the peer has closed
555 	 * the connection and won't ever send more data.
556 	 */
557 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
558 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
559 		/*
560 		 * "adv" is the amount we can increase the window,
561 		 * taking into account that we are limited by
562 		 * TCP_MAXWIN << tp->rcv_scale.
563 		 */
564 		long adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale) -
565 			(tp->rcv_adv - tp->rcv_nxt);
566 
567 		if (adv >= (long) (2 * tp->t_maxseg))
568 			goto send;
569 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
570 			goto send;
571 	}
572 
573 	/*
574 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
575 	 * is also a catch-all for the retransmit timer timeout case.
576 	 */
577 	if (tp->t_flags & TF_ACKNOW)
578 		goto send;
579 	if ((flags & TH_RST) ||
580 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
581 		goto send;
582 	if (SEQ_GT(tp->snd_up, tp->snd_una))
583 		goto send;
584 	/*
585 	 * If our state indicates that FIN should be sent
586 	 * and we have not yet done so, then we need to send.
587 	 */
588 	if (flags & TH_FIN &&
589 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
590 		goto send;
591 	/*
592 	 * In SACK, it is possible for tcp_output to fail to send a segment
593 	 * after the retransmission timer has been turned off.  Make sure
594 	 * that the retransmission timer is set.
595 	 */
596 	if ((tp->t_flags & TF_SACK_PERMIT) &&
597 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
598 	    !tcp_timer_active(tp, TT_REXMT) &&
599 	    !tcp_timer_active(tp, TT_PERSIST)) {
600 		tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
601 		goto just_return;
602 	}
603 	/*
604 	 * TCP window updates are not reliable, rather a polling protocol
605 	 * using ``persist'' packets is used to insure receipt of window
606 	 * updates.  The three ``states'' for the output side are:
607 	 *	idle			not doing retransmits or persists
608 	 *	persisting		to move a small or zero window
609 	 *	(re)transmitting	and thereby not persisting
610 	 *
611 	 * tcp_timer_active(tp, TT_PERSIST)
612 	 *	is true when we are in persist state.
613 	 * (tp->t_flags & TF_FORCEDATA)
614 	 *	is set when we are called to send a persist packet.
615 	 * tcp_timer_active(tp, TT_REXMT)
616 	 *	is set when we are retransmitting
617 	 * The output side is idle when both timers are zero.
618 	 *
619 	 * If send window is too small, there is data to transmit, and no
620 	 * retransmit or persist is pending, then go to persist state.
621 	 * If nothing happens soon, send when timer expires:
622 	 * if window is nonzero, transmit what we can,
623 	 * otherwise force out a byte.
624 	 */
625 	if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) &&
626 	    !tcp_timer_active(tp, TT_PERSIST)) {
627 		tp->t_rxtshift = 0;
628 		tcp_setpersist(tp);
629 	}
630 
631 	/*
632 	 * No reason to send a segment, just return.
633 	 */
634 just_return:
635 	SOCKBUF_UNLOCK(&so->so_snd);
636 	return (0);
637 
638 send:
639 	SOCKBUF_LOCK_ASSERT(&so->so_snd);
640 	/*
641 	 * Before ESTABLISHED, force sending of initial options
642 	 * unless TCP set not to do any options.
643 	 * NOTE: we assume that the IP/TCP header plus TCP options
644 	 * always fit in a single mbuf, leaving room for a maximum
645 	 * link header, i.e.
646 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
647 	 */
648 	optlen = 0;
649 #ifdef INET6
650 	if (isipv6)
651 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
652 	else
653 #endif
654 	hdrlen = sizeof (struct tcpiphdr);
655 
656 	/*
657 	 * Compute options for segment.
658 	 * We only have to care about SYN and established connection
659 	 * segments.  Options for SYN-ACK segments are handled in TCP
660 	 * syncache.
661 	 */
662 	if ((tp->t_flags & TF_NOOPT) == 0) {
663 		to.to_flags = 0;
664 		/* Maximum segment size. */
665 		if (flags & TH_SYN) {
666 			tp->snd_nxt = tp->iss;
667 			to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
668 			to.to_flags |= TOF_MSS;
669 		}
670 		/* Window scaling. */
671 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
672 			to.to_wscale = tp->request_r_scale;
673 			to.to_flags |= TOF_SCALE;
674 		}
675 		/* Timestamps. */
676 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
677 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
678 			to.to_tsval = ticks + tp->ts_offset;
679 			to.to_tsecr = tp->ts_recent;
680 			to.to_flags |= TOF_TS;
681 			/* Set receive buffer autosizing timestamp. */
682 			if (tp->rfbuf_ts == 0 &&
683 			    (so->so_rcv.sb_flags & SB_AUTOSIZE))
684 				tp->rfbuf_ts = ticks;
685 		}
686 		/* Selective ACK's. */
687 		if (tp->t_flags & TF_SACK_PERMIT) {
688 			if (flags & TH_SYN)
689 				to.to_flags |= TOF_SACKPERM;
690 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
691 			    (tp->t_flags & TF_SACK_PERMIT) &&
692 			    tp->rcv_numsacks > 0) {
693 				to.to_flags |= TOF_SACK;
694 				to.to_nsacks = tp->rcv_numsacks;
695 				to.to_sacks = (u_char *)tp->sackblks;
696 			}
697 		}
698 #ifdef TCP_SIGNATURE
699 		/* TCP-MD5 (RFC2385). */
700 		if (tp->t_flags & TF_SIGNATURE)
701 			to.to_flags |= TOF_SIGNATURE;
702 #endif /* TCP_SIGNATURE */
703 
704 		/* Processing the options. */
705 		hdrlen += optlen = tcp_addoptions(&to, opt);
706 	}
707 
708 #ifdef INET6
709 	if (isipv6)
710 		ipoptlen = ip6_optlen(tp->t_inpcb);
711 	else
712 #endif
713 	if (tp->t_inpcb->inp_options)
714 		ipoptlen = tp->t_inpcb->inp_options->m_len -
715 				offsetof(struct ipoption, ipopt_list);
716 	else
717 		ipoptlen = 0;
718 #ifdef IPSEC
719 	ipoptlen += ipsec_optlen;
720 #endif
721 
722 	/*
723 	 * Adjust data length if insertion of options will
724 	 * bump the packet length beyond the t_maxopd length.
725 	 * Clear the FIN bit because we cut off the tail of
726 	 * the segment.
727 	 */
728 	if (len + optlen + ipoptlen > tp->t_maxopd) {
729 		flags &= ~TH_FIN;
730 
731 		if (tso) {
732 			KASSERT(ipoptlen == 0,
733 			    ("%s: TSO can't do IP options", __func__));
734 
735 			/*
736 			 * Limit a burst to IP_MAXPACKET minus IP,
737 			 * TCP and options length to keep ip->ip_len
738 			 * from overflowing.
739 			 */
740 			if (len > IP_MAXPACKET - hdrlen) {
741 				len = IP_MAXPACKET - hdrlen;
742 				sendalot = 1;
743 			}
744 
745 			/*
746 			 * Prevent the last segment from being
747 			 * fractional unless the send sockbuf can
748 			 * be emptied.
749 			 */
750 			if (sendalot && off + len < so->so_snd.sb_cc) {
751 				len -= len % (tp->t_maxopd - optlen);
752 				sendalot = 1;
753 			}
754 
755 			/*
756 			 * Send the FIN in a separate segment
757 			 * after the bulk sending is done.
758 			 * We don't trust the TSO implementations
759 			 * to clear the FIN flag on all but the
760 			 * last segment.
761 			 */
762 			if (tp->t_flags & TF_NEEDFIN)
763 				sendalot = 1;
764 
765 		} else {
766 			len = tp->t_maxopd - optlen - ipoptlen;
767 			sendalot = 1;
768 		}
769 	} else
770 		tso = 0;
771 
772 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
773 	    ("%s: len > IP_MAXPACKET", __func__));
774 
775 /*#ifdef DIAGNOSTIC*/
776 #ifdef INET6
777 	if (max_linkhdr + hdrlen > MCLBYTES)
778 #else
779 	if (max_linkhdr + hdrlen > MHLEN)
780 #endif
781 		panic("tcphdr too big");
782 /*#endif*/
783 
784 	/*
785 	 * This KASSERT is here to catch edge cases at a well defined place.
786 	 * Before, those had triggered (random) panic conditions further down.
787 	 */
788 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
789 
790 	/*
791 	 * Grab a header mbuf, attaching a copy of data to
792 	 * be transmitted, and initialize the header from
793 	 * the template for sends on this connection.
794 	 */
795 	if (len) {
796 		struct mbuf *mb;
797 		u_int moff;
798 
799 		if ((tp->t_flags & TF_FORCEDATA) && len == 1)
800 			TCPSTAT_INC(tcps_sndprobe);
801 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
802 			tp->t_sndrexmitpack++;
803 			TCPSTAT_INC(tcps_sndrexmitpack);
804 			TCPSTAT_ADD(tcps_sndrexmitbyte, len);
805 		} else {
806 			TCPSTAT_INC(tcps_sndpack);
807 			TCPSTAT_ADD(tcps_sndbyte, len);
808 		}
809 #ifdef notyet
810 		if ((m = m_copypack(so->so_snd.sb_mb, off,
811 		    (int)len, max_linkhdr + hdrlen)) == 0) {
812 			SOCKBUF_UNLOCK(&so->so_snd);
813 			error = ENOBUFS;
814 			goto out;
815 		}
816 		/*
817 		 * m_copypack left space for our hdr; use it.
818 		 */
819 		m->m_len += hdrlen;
820 		m->m_data -= hdrlen;
821 #else
822 		MGETHDR(m, M_DONTWAIT, MT_DATA);
823 		if (m == NULL) {
824 			SOCKBUF_UNLOCK(&so->so_snd);
825 			error = ENOBUFS;
826 			goto out;
827 		}
828 #ifdef INET6
829 		if (MHLEN < hdrlen + max_linkhdr) {
830 			MCLGET(m, M_DONTWAIT);
831 			if ((m->m_flags & M_EXT) == 0) {
832 				SOCKBUF_UNLOCK(&so->so_snd);
833 				m_freem(m);
834 				error = ENOBUFS;
835 				goto out;
836 			}
837 		}
838 #endif
839 		m->m_data += max_linkhdr;
840 		m->m_len = hdrlen;
841 
842 		/*
843 		 * Start the m_copy functions from the closest mbuf
844 		 * to the offset in the socket buffer chain.
845 		 */
846 		mb = sbsndptr(&so->so_snd, off, len, &moff);
847 
848 		if (len <= MHLEN - hdrlen - max_linkhdr) {
849 			m_copydata(mb, moff, (int)len,
850 			    mtod(m, caddr_t) + hdrlen);
851 			m->m_len += len;
852 		} else {
853 			m->m_next = m_copy(mb, moff, (int)len);
854 			if (m->m_next == NULL) {
855 				SOCKBUF_UNLOCK(&so->so_snd);
856 				(void) m_free(m);
857 				error = ENOBUFS;
858 				goto out;
859 			}
860 		}
861 #endif
862 		/*
863 		 * If we're sending everything we've got, set PUSH.
864 		 * (This will keep happy those implementations which only
865 		 * give data to the user when a buffer fills or
866 		 * a PUSH comes in.)
867 		 */
868 		if (off + len == so->so_snd.sb_cc)
869 			flags |= TH_PUSH;
870 		SOCKBUF_UNLOCK(&so->so_snd);
871 	} else {
872 		SOCKBUF_UNLOCK(&so->so_snd);
873 		if (tp->t_flags & TF_ACKNOW)
874 			TCPSTAT_INC(tcps_sndacks);
875 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
876 			TCPSTAT_INC(tcps_sndctrl);
877 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
878 			TCPSTAT_INC(tcps_sndurg);
879 		else
880 			TCPSTAT_INC(tcps_sndwinup);
881 
882 		MGETHDR(m, M_DONTWAIT, MT_DATA);
883 		if (m == NULL) {
884 			error = ENOBUFS;
885 			goto out;
886 		}
887 #ifdef INET6
888 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
889 		    MHLEN >= hdrlen) {
890 			MH_ALIGN(m, hdrlen);
891 		} else
892 #endif
893 		m->m_data += max_linkhdr;
894 		m->m_len = hdrlen;
895 	}
896 	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
897 	m->m_pkthdr.rcvif = (struct ifnet *)0;
898 #ifdef MAC
899 	mac_inpcb_create_mbuf(tp->t_inpcb, m);
900 #endif
901 #ifdef INET6
902 	if (isipv6) {
903 		ip6 = mtod(m, struct ip6_hdr *);
904 		th = (struct tcphdr *)(ip6 + 1);
905 		tcpip_fillheaders(tp->t_inpcb, ip6, th);
906 	} else
907 #endif /* INET6 */
908 	{
909 		ip = mtod(m, struct ip *);
910 		ipov = (struct ipovly *)ip;
911 		th = (struct tcphdr *)(ip + 1);
912 		tcpip_fillheaders(tp->t_inpcb, ip, th);
913 	}
914 
915 	/*
916 	 * Fill in fields, remembering maximum advertised
917 	 * window for use in delaying messages about window sizes.
918 	 * If resending a FIN, be sure not to use a new sequence number.
919 	 */
920 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
921 	    tp->snd_nxt == tp->snd_max)
922 		tp->snd_nxt--;
923 	/*
924 	 * If we are starting a connection, send ECN setup
925 	 * SYN packet. If we are on a retransmit, we may
926 	 * resend those bits a number of times as per
927 	 * RFC 3168.
928 	 */
929 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
930 		if (tp->t_rxtshift >= 1) {
931 			if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
932 				flags |= TH_ECE|TH_CWR;
933 		} else
934 			flags |= TH_ECE|TH_CWR;
935 	}
936 
937 	if (tp->t_state == TCPS_ESTABLISHED &&
938 	    (tp->t_flags & TF_ECN_PERMIT)) {
939 		/*
940 		 * If the peer has ECN, mark data packets with
941 		 * ECN capable transmission (ECT).
942 		 * Ignore pure ack packets, retransmissions and window probes.
943 		 */
944 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
945 		    !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
946 #ifdef INET6
947 			if (isipv6)
948 				ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
949 			else
950 #endif
951 				ip->ip_tos |= IPTOS_ECN_ECT0;
952 			TCPSTAT_INC(tcps_ecn_ect0);
953 		}
954 
955 		/*
956 		 * Reply with proper ECN notifications.
957 		 */
958 		if (tp->t_flags & TF_ECN_SND_CWR) {
959 			flags |= TH_CWR;
960 			tp->t_flags &= ~TF_ECN_SND_CWR;
961 		}
962 		if (tp->t_flags & TF_ECN_SND_ECE)
963 			flags |= TH_ECE;
964 	}
965 
966 	/*
967 	 * If we are doing retransmissions, then snd_nxt will
968 	 * not reflect the first unsent octet.  For ACK only
969 	 * packets, we do not want the sequence number of the
970 	 * retransmitted packet, we want the sequence number
971 	 * of the next unsent octet.  So, if there is no data
972 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
973 	 * when filling in ti_seq.  But if we are in persist
974 	 * state, snd_max might reflect one byte beyond the
975 	 * right edge of the window, so use snd_nxt in that
976 	 * case, since we know we aren't doing a retransmission.
977 	 * (retransmit and persist are mutually exclusive...)
978 	 */
979 	if (sack_rxmit == 0) {
980 		if (len || (flags & (TH_SYN|TH_FIN)) ||
981 		    tcp_timer_active(tp, TT_PERSIST))
982 			th->th_seq = htonl(tp->snd_nxt);
983 		else
984 			th->th_seq = htonl(tp->snd_max);
985 	} else {
986 		th->th_seq = htonl(p->rxmit);
987 		p->rxmit += len;
988 		tp->sackhint.sack_bytes_rexmit += len;
989 	}
990 	th->th_ack = htonl(tp->rcv_nxt);
991 	if (optlen) {
992 		bcopy(opt, th + 1, optlen);
993 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
994 	}
995 	th->th_flags = flags;
996 	/*
997 	 * Calculate receive window.  Don't shrink window,
998 	 * but avoid silly window syndrome.
999 	 */
1000 	if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
1001 	    recwin < (long)tp->t_maxseg)
1002 		recwin = 0;
1003 	if (recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
1004 		recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
1005 	if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
1006 		recwin = (long)TCP_MAXWIN << tp->rcv_scale;
1007 
1008 	/*
1009 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1010 	 * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
1011 	 * case is handled in syncache.
1012 	 */
1013 	if (flags & TH_SYN)
1014 		th->th_win = htons((u_short)
1015 				(min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1016 	else
1017 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1018 
1019 	/*
1020 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1021 	 * a 0 window.  This may cause the remote transmitter to stall.  This
1022 	 * flag tells soreceive() to disable delayed acknowledgements when
1023 	 * draining the buffer.  This can occur if the receiver is attempting
1024 	 * to read more data than can be buffered prior to transmitting on
1025 	 * the connection.
1026 	 */
1027 	if (th->th_win == 0) {
1028 		tp->t_sndzerowin++;
1029 		tp->t_flags |= TF_RXWIN0SENT;
1030 	} else
1031 		tp->t_flags &= ~TF_RXWIN0SENT;
1032 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1033 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1034 		th->th_flags |= TH_URG;
1035 	} else
1036 		/*
1037 		 * If no urgent pointer to send, then we pull
1038 		 * the urgent pointer to the left edge of the send window
1039 		 * so that it doesn't drift into the send window on sequence
1040 		 * number wraparound.
1041 		 */
1042 		tp->snd_up = tp->snd_una;		/* drag it along */
1043 
1044 #ifdef TCP_SIGNATURE
1045 	if (tp->t_flags & TF_SIGNATURE) {
1046 		int sigoff = to.to_signature - opt;
1047 		tcp_signature_compute(m, 0, len, optlen,
1048 		    (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
1049 	}
1050 #endif
1051 
1052 	/*
1053 	 * Put TCP length in extended header, and then
1054 	 * checksum extended header and data.
1055 	 */
1056 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1057 #ifdef INET6
1058 	if (isipv6)
1059 		/*
1060 		 * ip6_plen is not need to be filled now, and will be filled
1061 		 * in ip6_output.
1062 		 */
1063 		th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
1064 				       sizeof(struct tcphdr) + optlen + len);
1065 	else
1066 #endif /* INET6 */
1067 	{
1068 		m->m_pkthdr.csum_flags = CSUM_TCP;
1069 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1070 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1071 		    htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1072 
1073 		/* IP version must be set here for ipv4/ipv6 checking later */
1074 		KASSERT(ip->ip_v == IPVERSION,
1075 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1076 	}
1077 
1078 	/*
1079 	 * Enable TSO and specify the size of the segments.
1080 	 * The TCP pseudo header checksum is always provided.
1081 	 * XXX: Fixme: This is currently not the case for IPv6.
1082 	 */
1083 	if (tso) {
1084 		KASSERT(len > tp->t_maxopd - optlen,
1085 		    ("%s: len <= tso_segsz", __func__));
1086 		m->m_pkthdr.csum_flags |= CSUM_TSO;
1087 		m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen;
1088 	}
1089 
1090 	KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL),
1091 	    ("%s: mbuf chain shorter than expected", __func__));
1092 
1093 	/*
1094 	 * In transmit state, time the transmission and arrange for
1095 	 * the retransmit.  In persist state, just set snd_max.
1096 	 */
1097 	if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1098 	    !tcp_timer_active(tp, TT_PERSIST)) {
1099 		tcp_seq startseq = tp->snd_nxt;
1100 
1101 		/*
1102 		 * Advance snd_nxt over sequence space of this segment.
1103 		 */
1104 		if (flags & (TH_SYN|TH_FIN)) {
1105 			if (flags & TH_SYN)
1106 				tp->snd_nxt++;
1107 			if (flags & TH_FIN) {
1108 				tp->snd_nxt++;
1109 				tp->t_flags |= TF_SENTFIN;
1110 			}
1111 		}
1112 		if (sack_rxmit)
1113 			goto timer;
1114 		tp->snd_nxt += len;
1115 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1116 			tp->snd_max = tp->snd_nxt;
1117 			/*
1118 			 * Time this transmission if not a retransmission and
1119 			 * not currently timing anything.
1120 			 */
1121 			if (tp->t_rtttime == 0) {
1122 				tp->t_rtttime = ticks;
1123 				tp->t_rtseq = startseq;
1124 				TCPSTAT_INC(tcps_segstimed);
1125 			}
1126 		}
1127 
1128 		/*
1129 		 * Set retransmit timer if not currently set,
1130 		 * and not doing a pure ack or a keep-alive probe.
1131 		 * Initial value for retransmit timer is smoothed
1132 		 * round-trip time + 2 * round-trip time variance.
1133 		 * Initialize shift counter which is used for backoff
1134 		 * of retransmit time.
1135 		 */
1136 timer:
1137 		if (!tcp_timer_active(tp, TT_REXMT) &&
1138 		    ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1139 		     (tp->snd_nxt != tp->snd_una))) {
1140 			if (tcp_timer_active(tp, TT_PERSIST)) {
1141 				tcp_timer_activate(tp, TT_PERSIST, 0);
1142 				tp->t_rxtshift = 0;
1143 			}
1144 			tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1145 		}
1146 	} else {
1147 		/*
1148 		 * Persist case, update snd_max but since we are in
1149 		 * persist mode (no window) we do not update snd_nxt.
1150 		 */
1151 		int xlen = len;
1152 		if (flags & TH_SYN)
1153 			++xlen;
1154 		if (flags & TH_FIN) {
1155 			++xlen;
1156 			tp->t_flags |= TF_SENTFIN;
1157 		}
1158 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1159 			tp->snd_max = tp->snd_nxt + len;
1160 	}
1161 
1162 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1163 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
1164 
1165 #ifdef TCPDEBUG
1166 	/*
1167 	 * Trace.
1168 	 */
1169 	if (so->so_options & SO_DEBUG) {
1170 		u_short save = 0;
1171 #ifdef INET6
1172 		if (!isipv6)
1173 #endif
1174 		{
1175 			save = ipov->ih_len;
1176 			ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1177 		}
1178 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1179 #ifdef INET6
1180 		if (!isipv6)
1181 #endif
1182 		ipov->ih_len = save;
1183 	}
1184 #endif
1185 
1186 	/*
1187 	 * Fill in IP length and desired time to live and
1188 	 * send to IP level.  There should be a better way
1189 	 * to handle ttl and tos; we could keep them in
1190 	 * the template, but need a way to checksum without them.
1191 	 */
1192 	/*
1193 	 * m->m_pkthdr.len should have been set before cksum calcuration,
1194 	 * because in6_cksum() need it.
1195 	 */
1196 #ifdef INET6
1197 	if (isipv6) {
1198 		/*
1199 		 * we separately set hoplimit for every segment, since the
1200 		 * user might want to change the value via setsockopt.
1201 		 * Also, desired default hop limit might be changed via
1202 		 * Neighbor Discovery.
1203 		 */
1204 		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1205 
1206 		/* TODO: IPv6 IP6TOS_ECT bit on */
1207 		error = ip6_output(m,
1208 			    tp->t_inpcb->in6p_outputopts, NULL,
1209 			    ((so->so_options & SO_DONTROUTE) ?
1210 			    IP_ROUTETOIF : 0), NULL, NULL, tp->t_inpcb);
1211 	} else
1212 #endif /* INET6 */
1213     {
1214 	ip->ip_len = m->m_pkthdr.len;
1215 #ifdef INET6
1216 	if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1217 		ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1218 #endif /* INET6 */
1219 	/*
1220 	 * If we do path MTU discovery, then we set DF on every packet.
1221 	 * This might not be the best thing to do according to RFC3390
1222 	 * Section 2. However the tcp hostcache migitates the problem
1223 	 * so it affects only the first tcp connection with a host.
1224 	 *
1225 	 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1226 	 */
1227 	if (V_path_mtu_discovery && tp->t_maxopd > V_tcp_minmss)
1228 		ip->ip_off |= IP_DF;
1229 
1230 	error = ip_output(m, tp->t_inpcb->inp_options, NULL,
1231 	    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1232 	    tp->t_inpcb);
1233     }
1234 	if (error) {
1235 
1236 		/*
1237 		 * We know that the packet was lost, so back out the
1238 		 * sequence number advance, if any.
1239 		 *
1240 		 * If the error is EPERM the packet got blocked by the
1241 		 * local firewall.  Normally we should terminate the
1242 		 * connection but the blocking may have been spurious
1243 		 * due to a firewall reconfiguration cycle.  So we treat
1244 		 * it like a packet loss and let the retransmit timer and
1245 		 * timeouts do their work over time.
1246 		 * XXX: It is a POLA question whether calling tcp_drop right
1247 		 * away would be the really correct behavior instead.
1248 		 */
1249 		if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1250 		    !tcp_timer_active(tp, TT_PERSIST)) &&
1251 		    ((flags & TH_SYN) == 0) &&
1252 		    (error != EPERM)) {
1253 			if (sack_rxmit) {
1254 				p->rxmit -= len;
1255 				tp->sackhint.sack_bytes_rexmit -= len;
1256 				KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1257 				    ("sackhint bytes rtx >= 0"));
1258 			} else
1259 				tp->snd_nxt -= len;
1260 		}
1261 out:
1262 		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1263 		switch (error) {
1264 		case EPERM:
1265 			tp->t_softerror = error;
1266 			return (error);
1267 		case ENOBUFS:
1268 	                if (!tcp_timer_active(tp, TT_REXMT) &&
1269 			    !tcp_timer_active(tp, TT_PERSIST))
1270 	                        tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1271 			tp->snd_cwnd = tp->t_maxseg;
1272 			return (0);
1273 		case EMSGSIZE:
1274 			/*
1275 			 * For some reason the interface we used initially
1276 			 * to send segments changed to another or lowered
1277 			 * its MTU.
1278 			 *
1279 			 * tcp_mtudisc() will find out the new MTU and as
1280 			 * its last action, initiate retransmission, so it
1281 			 * is important to not do so here.
1282 			 *
1283 			 * If TSO was active we either got an interface
1284 			 * without TSO capabilits or TSO was turned off.
1285 			 * Disable it for this connection as too and
1286 			 * immediatly retry with MSS sized segments generated
1287 			 * by this function.
1288 			 */
1289 			if (tso)
1290 				tp->t_flags &= ~TF_TSO;
1291 			tcp_mtudisc(tp->t_inpcb, 0);
1292 			return (0);
1293 		case EHOSTDOWN:
1294 		case EHOSTUNREACH:
1295 		case ENETDOWN:
1296 		case ENETUNREACH:
1297 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
1298 				tp->t_softerror = error;
1299 				return (0);
1300 			}
1301 			/* FALLTHROUGH */
1302 		default:
1303 			return (error);
1304 		}
1305 	}
1306 	TCPSTAT_INC(tcps_sndtotal);
1307 
1308 	/*
1309 	 * Data sent (as far as we can tell).
1310 	 * If this advertises a larger window than any other segment,
1311 	 * then remember the size of the advertised window.
1312 	 * Any pending ACK has now been sent.
1313 	 */
1314 	if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1315 		tp->rcv_adv = tp->rcv_nxt + recwin;
1316 	tp->last_ack_sent = tp->rcv_nxt;
1317 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1318 	if (tcp_timer_active(tp, TT_DELACK))
1319 		tcp_timer_activate(tp, TT_DELACK, 0);
1320 #if 0
1321 	/*
1322 	 * This completely breaks TCP if newreno is turned on.  What happens
1323 	 * is that if delayed-acks are turned on on the receiver, this code
1324 	 * on the transmitter effectively destroys the TCP window, forcing
1325 	 * it to four packets (1.5Kx4 = 6K window).
1326 	 */
1327 	if (sendalot && --maxburst)
1328 		goto again;
1329 #endif
1330 	if (sendalot)
1331 		goto again;
1332 	return (0);
1333 }
1334 
1335 void
1336 tcp_setpersist(struct tcpcb *tp)
1337 {
1338 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1339 	int tt;
1340 
1341 	if (tcp_timer_active(tp, TT_REXMT))
1342 		panic("tcp_setpersist: retransmit pending");
1343 	/*
1344 	 * Start/restart persistance timer.
1345 	 */
1346 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1347 		      TCPTV_PERSMIN, TCPTV_PERSMAX);
1348 	tcp_timer_activate(tp, TT_PERSIST, tt);
1349 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1350 		tp->t_rxtshift++;
1351 }
1352 
1353 /*
1354  * Insert TCP options according to the supplied parameters to the place
1355  * optp in a consistent way.  Can handle unaligned destinations.
1356  *
1357  * The order of the option processing is crucial for optimal packing and
1358  * alignment for the scarce option space.
1359  *
1360  * The optimal order for a SYN/SYN-ACK segment is:
1361  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1362  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1363  *
1364  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1365  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1366  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1367  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1368  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1369  */
1370 int
1371 tcp_addoptions(struct tcpopt *to, u_char *optp)
1372 {
1373 	u_int mask, optlen = 0;
1374 
1375 	for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1376 		if ((to->to_flags & mask) != mask)
1377 			continue;
1378 		if (optlen == TCP_MAXOLEN)
1379 			break;
1380 		switch (to->to_flags & mask) {
1381 		case TOF_MSS:
1382 			while (optlen % 4) {
1383 				optlen += TCPOLEN_NOP;
1384 				*optp++ = TCPOPT_NOP;
1385 			}
1386 			if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1387 				continue;
1388 			optlen += TCPOLEN_MAXSEG;
1389 			*optp++ = TCPOPT_MAXSEG;
1390 			*optp++ = TCPOLEN_MAXSEG;
1391 			to->to_mss = htons(to->to_mss);
1392 			bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1393 			optp += sizeof(to->to_mss);
1394 			break;
1395 		case TOF_SCALE:
1396 			while (!optlen || optlen % 2 != 1) {
1397 				optlen += TCPOLEN_NOP;
1398 				*optp++ = TCPOPT_NOP;
1399 			}
1400 			if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1401 				continue;
1402 			optlen += TCPOLEN_WINDOW;
1403 			*optp++ = TCPOPT_WINDOW;
1404 			*optp++ = TCPOLEN_WINDOW;
1405 			*optp++ = to->to_wscale;
1406 			break;
1407 		case TOF_SACKPERM:
1408 			while (optlen % 2) {
1409 				optlen += TCPOLEN_NOP;
1410 				*optp++ = TCPOPT_NOP;
1411 			}
1412 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1413 				continue;
1414 			optlen += TCPOLEN_SACK_PERMITTED;
1415 			*optp++ = TCPOPT_SACK_PERMITTED;
1416 			*optp++ = TCPOLEN_SACK_PERMITTED;
1417 			break;
1418 		case TOF_TS:
1419 			while (!optlen || optlen % 4 != 2) {
1420 				optlen += TCPOLEN_NOP;
1421 				*optp++ = TCPOPT_NOP;
1422 			}
1423 			if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1424 				continue;
1425 			optlen += TCPOLEN_TIMESTAMP;
1426 			*optp++ = TCPOPT_TIMESTAMP;
1427 			*optp++ = TCPOLEN_TIMESTAMP;
1428 			to->to_tsval = htonl(to->to_tsval);
1429 			to->to_tsecr = htonl(to->to_tsecr);
1430 			bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1431 			optp += sizeof(to->to_tsval);
1432 			bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1433 			optp += sizeof(to->to_tsecr);
1434 			break;
1435 		case TOF_SIGNATURE:
1436 			{
1437 			int siglen = TCPOLEN_SIGNATURE - 2;
1438 
1439 			while (!optlen || optlen % 4 != 2) {
1440 				optlen += TCPOLEN_NOP;
1441 				*optp++ = TCPOPT_NOP;
1442 			}
1443 			if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE)
1444 				continue;
1445 			optlen += TCPOLEN_SIGNATURE;
1446 			*optp++ = TCPOPT_SIGNATURE;
1447 			*optp++ = TCPOLEN_SIGNATURE;
1448 			to->to_signature = optp;
1449 			while (siglen--)
1450 				 *optp++ = 0;
1451 			break;
1452 			}
1453 		case TOF_SACK:
1454 			{
1455 			int sackblks = 0;
1456 			struct sackblk *sack = (struct sackblk *)to->to_sacks;
1457 			tcp_seq sack_seq;
1458 
1459 			while (!optlen || optlen % 4 != 2) {
1460 				optlen += TCPOLEN_NOP;
1461 				*optp++ = TCPOPT_NOP;
1462 			}
1463 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1464 				continue;
1465 			optlen += TCPOLEN_SACKHDR;
1466 			*optp++ = TCPOPT_SACK;
1467 			sackblks = min(to->to_nsacks,
1468 					(TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1469 			*optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1470 			while (sackblks--) {
1471 				sack_seq = htonl(sack->start);
1472 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1473 				optp += sizeof(sack_seq);
1474 				sack_seq = htonl(sack->end);
1475 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1476 				optp += sizeof(sack_seq);
1477 				optlen += TCPOLEN_SACK;
1478 				sack++;
1479 			}
1480 			TCPSTAT_INC(tcps_sack_send_blocks);
1481 			break;
1482 			}
1483 		default:
1484 			panic("%s: unknown TCP option type", __func__);
1485 			break;
1486 		}
1487 	}
1488 
1489 	/* Terminate and pad TCP options to a 4 byte boundary. */
1490 	if (optlen % 4) {
1491 		optlen += TCPOLEN_EOL;
1492 		*optp++ = TCPOPT_EOL;
1493 	}
1494 	/*
1495 	 * According to RFC 793 (STD0007):
1496 	 *   "The content of the header beyond the End-of-Option option
1497 	 *    must be header padding (i.e., zero)."
1498 	 *   and later: "The padding is composed of zeros."
1499 	 */
1500 	while (optlen % 4) {
1501 		optlen += TCPOLEN_PAD;
1502 		*optp++ = TCPOPT_PAD;
1503 	}
1504 
1505 	KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1506 	return (optlen);
1507 }
1508