xref: /freebsd/sys/netinet/tcp_output.c (revision ae77177087c655fc883075af4f425b37e032cd05)
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 VNET_DEFINE(int, path_mtu_discovery) = 1;
88 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
89 	&VNET_NAME(path_mtu_discovery), 1,
90 	"Enable Path MTU Discovery");
91 
92 VNET_DEFINE(int, tcp_do_tso) = 1;
93 #define	V_tcp_do_tso		VNET(tcp_do_tso)
94 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW,
95 	&VNET_NAME(tcp_do_tso), 0,
96 	"Enable TCP Segmentation Offload");
97 
98 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
99 #define	V_tcp_sendspace	VNET(tcp_sendspace)
100 SYSCTL_VNET_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_RW,
101 	&VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
102 
103 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
104 #define	V_tcp_do_autosndbuf	VNET(tcp_do_autosndbuf)
105 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW,
106 	&VNET_NAME(tcp_do_autosndbuf), 0,
107 	"Enable automatic send buffer sizing");
108 
109 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
110 #define	V_tcp_autosndbuf_inc	VNET(tcp_autosndbuf_inc)
111 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW,
112 	&VNET_NAME(tcp_autosndbuf_inc), 0,
113 	"Incrementor step size of automatic send buffer");
114 
115 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
116 #define	V_tcp_autosndbuf_max	VNET(tcp_autosndbuf_max)
117 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW,
118 	&VNET_NAME(tcp_autosndbuf_max), 0,
119 	"Max size of automatic send buffer");
120 
121 static void inline	hhook_run_tcp_est_out(struct tcpcb *tp,
122 			    struct tcphdr *th, struct tcpopt *to,
123 			    long len, int tso);
124 static void inline	cc_after_idle(struct tcpcb *tp);
125 
126 /*
127  * Wrapper for the TCP established ouput helper hook.
128  */
129 static void inline
130 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
131     struct tcpopt *to, long len, int tso)
132 {
133 	struct tcp_hhook_data hhook_data;
134 
135 	if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
136 		hhook_data.tp = tp;
137 		hhook_data.th = th;
138 		hhook_data.to = to;
139 		hhook_data.len = len;
140 		hhook_data.tso = tso;
141 
142 		hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
143 		    tp->osd);
144 	}
145 }
146 
147 /*
148  * CC wrapper hook functions
149  */
150 static void inline
151 cc_after_idle(struct tcpcb *tp)
152 {
153 	INP_WLOCK_ASSERT(tp->t_inpcb);
154 
155 	if (CC_ALGO(tp)->after_idle != NULL)
156 		CC_ALGO(tp)->after_idle(tp->ccv);
157 }
158 
159 /*
160  * Tcp output routine: figure out what should be sent and send it.
161  */
162 int
163 tcp_output(struct tcpcb *tp)
164 {
165 	struct socket *so = tp->t_inpcb->inp_socket;
166 	long len, recwin, sendwin;
167 	int off, flags, error = 0;	/* Keep compiler happy */
168 	struct mbuf *m;
169 	struct ip *ip = NULL;
170 	struct ipovly *ipov = NULL;
171 	struct tcphdr *th;
172 	u_char opt[TCP_MAXOLEN];
173 	unsigned ipoptlen, optlen, hdrlen;
174 #ifdef IPSEC
175 	unsigned ipsec_optlen = 0;
176 #endif
177 	int idle, sendalot;
178 	int sack_rxmit, sack_bytes_rxmt;
179 	struct sackhole *p;
180 	int tso;
181 	struct tcpopt to;
182 #if 0
183 	int maxburst = TCP_MAXBURST;
184 #endif
185 #ifdef INET6
186 	struct ip6_hdr *ip6 = NULL;
187 	int isipv6;
188 
189 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
190 #endif
191 
192 	INP_WLOCK_ASSERT(tp->t_inpcb);
193 
194 	/*
195 	 * Determine length of data that should be transmitted,
196 	 * and flags that will be used.
197 	 * If there is some data or critical controls (SYN, RST)
198 	 * to send, then transmit; otherwise, investigate further.
199 	 */
200 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
201 	if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur)
202 		cc_after_idle(tp);
203 	tp->t_flags &= ~TF_LASTIDLE;
204 	if (idle) {
205 		if (tp->t_flags & TF_MORETOCOME) {
206 			tp->t_flags |= TF_LASTIDLE;
207 			idle = 0;
208 		}
209 	}
210 again:
211 	/*
212 	 * If we've recently taken a timeout, snd_max will be greater than
213 	 * snd_nxt.  There may be SACK information that allows us to avoid
214 	 * resending already delivered data.  Adjust snd_nxt accordingly.
215 	 */
216 	if ((tp->t_flags & TF_SACK_PERMIT) &&
217 	    SEQ_LT(tp->snd_nxt, tp->snd_max))
218 		tcp_sack_adjust(tp);
219 	sendalot = 0;
220 	tso = 0;
221 	off = tp->snd_nxt - tp->snd_una;
222 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
223 
224 	flags = tcp_outflags[tp->t_state];
225 	/*
226 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
227 	 * to send out new data (when sendalot is 1), bypass this function.
228 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
229 	 * we're replacing a (future) new transmission with a retransmission
230 	 * now, and we previously incremented snd_cwnd in tcp_input().
231 	 */
232 	/*
233 	 * Still in sack recovery , reset rxmit flag to zero.
234 	 */
235 	sack_rxmit = 0;
236 	sack_bytes_rxmt = 0;
237 	len = 0;
238 	p = NULL;
239 	if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
240 	    (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
241 		long cwin;
242 
243 		cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
244 		if (cwin < 0)
245 			cwin = 0;
246 		/* Do not retransmit SACK segments beyond snd_recover */
247 		if (SEQ_GT(p->end, tp->snd_recover)) {
248 			/*
249 			 * (At least) part of sack hole extends beyond
250 			 * snd_recover. Check to see if we can rexmit data
251 			 * for this hole.
252 			 */
253 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
254 				/*
255 				 * Can't rexmit any more data for this hole.
256 				 * That data will be rexmitted in the next
257 				 * sack recovery episode, when snd_recover
258 				 * moves past p->rxmit.
259 				 */
260 				p = NULL;
261 				goto after_sack_rexmit;
262 			} else
263 				/* Can rexmit part of the current hole */
264 				len = ((long)ulmin(cwin,
265 						   tp->snd_recover - p->rxmit));
266 		} else
267 			len = ((long)ulmin(cwin, p->end - p->rxmit));
268 		off = p->rxmit - tp->snd_una;
269 		KASSERT(off >= 0,("%s: sack block to the left of una : %d",
270 		    __func__, off));
271 		if (len > 0) {
272 			sack_rxmit = 1;
273 			sendalot = 1;
274 			TCPSTAT_INC(tcps_sack_rexmits);
275 			TCPSTAT_ADD(tcps_sack_rexmit_bytes,
276 			    min(len, tp->t_maxseg));
277 		}
278 	}
279 after_sack_rexmit:
280 	/*
281 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
282 	 * state flags.
283 	 */
284 	if (tp->t_flags & TF_NEEDFIN)
285 		flags |= TH_FIN;
286 	if (tp->t_flags & TF_NEEDSYN)
287 		flags |= TH_SYN;
288 
289 	SOCKBUF_LOCK(&so->so_snd);
290 	/*
291 	 * If in persist timeout with window of 0, send 1 byte.
292 	 * Otherwise, if window is small but nonzero
293 	 * and timer expired, we will send what we can
294 	 * and go to transmit state.
295 	 */
296 	if (tp->t_flags & TF_FORCEDATA) {
297 		if (sendwin == 0) {
298 			/*
299 			 * If we still have some data to send, then
300 			 * clear the FIN bit.  Usually this would
301 			 * happen below when it realizes that we
302 			 * aren't sending all the data.  However,
303 			 * if we have exactly 1 byte of unsent data,
304 			 * then it won't clear the FIN bit below,
305 			 * and if we are in persist state, we wind
306 			 * up sending the packet without recording
307 			 * that we sent the FIN bit.
308 			 *
309 			 * We can't just blindly clear the FIN bit,
310 			 * because if we don't have any more data
311 			 * to send then the probe will be the FIN
312 			 * itself.
313 			 */
314 			if (off < so->so_snd.sb_cc)
315 				flags &= ~TH_FIN;
316 			sendwin = 1;
317 		} else {
318 			tcp_timer_activate(tp, TT_PERSIST, 0);
319 			tp->t_rxtshift = 0;
320 		}
321 	}
322 
323 	/*
324 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
325 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
326 	 * a negative length.  This can also occur when TCP opens up
327 	 * its congestion window while receiving additional duplicate
328 	 * acks after fast-retransmit because TCP will reset snd_nxt
329 	 * to snd_max after the fast-retransmit.
330 	 *
331 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
332 	 * be set to snd_una, the offset will be 0, and the length may
333 	 * wind up 0.
334 	 *
335 	 * If sack_rxmit is true we are retransmitting from the scoreboard
336 	 * in which case len is already set.
337 	 */
338 	if (sack_rxmit == 0) {
339 		if (sack_bytes_rxmt == 0)
340 			len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off);
341 		else {
342 			long cwin;
343 
344                         /*
345 			 * We are inside of a SACK recovery episode and are
346 			 * sending new data, having retransmitted all the
347 			 * data possible in the scoreboard.
348 			 */
349 			len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd)
350 			       - off);
351 			/*
352 			 * Don't remove this (len > 0) check !
353 			 * We explicitly check for len > 0 here (although it
354 			 * isn't really necessary), to work around a gcc
355 			 * optimization issue - to force gcc to compute
356 			 * len above. Without this check, the computation
357 			 * of len is bungled by the optimizer.
358 			 */
359 			if (len > 0) {
360 				cwin = tp->snd_cwnd -
361 					(tp->snd_nxt - tp->sack_newdata) -
362 					sack_bytes_rxmt;
363 				if (cwin < 0)
364 					cwin = 0;
365 				len = lmin(len, cwin);
366 			}
367 		}
368 	}
369 
370 	/*
371 	 * Lop off SYN bit if it has already been sent.  However, if this
372 	 * is SYN-SENT state and if segment contains data and if we don't
373 	 * know that foreign host supports TAO, suppress sending segment.
374 	 */
375 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
376 		if (tp->t_state != TCPS_SYN_RECEIVED)
377 			flags &= ~TH_SYN;
378 		off--, len++;
379 	}
380 
381 	/*
382 	 * Be careful not to send data and/or FIN on SYN segments.
383 	 * This measure is needed to prevent interoperability problems
384 	 * with not fully conformant TCP implementations.
385 	 */
386 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
387 		len = 0;
388 		flags &= ~TH_FIN;
389 	}
390 
391 	if (len < 0) {
392 		/*
393 		 * If FIN has been sent but not acked,
394 		 * but we haven't been called to retransmit,
395 		 * len will be < 0.  Otherwise, window shrank
396 		 * after we sent into it.  If window shrank to 0,
397 		 * cancel pending retransmit, pull snd_nxt back
398 		 * to (closed) window, and set the persist timer
399 		 * if it isn't already going.  If the window didn't
400 		 * close completely, just wait for an ACK.
401 		 */
402 		len = 0;
403 		if (sendwin == 0) {
404 			tcp_timer_activate(tp, TT_REXMT, 0);
405 			tp->t_rxtshift = 0;
406 			tp->snd_nxt = tp->snd_una;
407 			if (!tcp_timer_active(tp, TT_PERSIST))
408 				tcp_setpersist(tp);
409 		}
410 	}
411 
412 	/* len will be >= 0 after this point. */
413 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
414 
415 	/*
416 	 * Automatic sizing of send socket buffer.  Often the send buffer
417 	 * size is not optimally adjusted to the actual network conditions
418 	 * at hand (delay bandwidth product).  Setting the buffer size too
419 	 * small limits throughput on links with high bandwidth and high
420 	 * delay (eg. trans-continental/oceanic links).  Setting the
421 	 * buffer size too big consumes too much real kernel memory,
422 	 * especially with many connections on busy servers.
423 	 *
424 	 * The criteria to step up the send buffer one notch are:
425 	 *  1. receive window of remote host is larger than send buffer
426 	 *     (with a fudge factor of 5/4th);
427 	 *  2. send buffer is filled to 7/8th with data (so we actually
428 	 *     have data to make use of it);
429 	 *  3. send buffer fill has not hit maximal automatic size;
430 	 *  4. our send window (slow start and cogestion controlled) is
431 	 *     larger than sent but unacknowledged data in send buffer.
432 	 *
433 	 * The remote host receive window scaling factor may limit the
434 	 * growing of the send buffer before it reaches its allowed
435 	 * maximum.
436 	 *
437 	 * It scales directly with slow start or congestion window
438 	 * and does at most one step per received ACK.  This fast
439 	 * scaling has the drawback of growing the send buffer beyond
440 	 * what is strictly necessary to make full use of a given
441 	 * delay*bandwith product.  However testing has shown this not
442 	 * to be much of an problem.  At worst we are trading wasting
443 	 * of available bandwith (the non-use of it) for wasting some
444 	 * socket buffer memory.
445 	 *
446 	 * TODO: Shrink send buffer during idle periods together
447 	 * with congestion window.  Requires another timer.  Has to
448 	 * wait for upcoming tcp timer rewrite.
449 	 */
450 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
451 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
452 		    so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
453 		    so->so_snd.sb_cc < V_tcp_autosndbuf_max &&
454 		    sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
455 			if (!sbreserve_locked(&so->so_snd,
456 			    min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
457 			     V_tcp_autosndbuf_max), so, curthread))
458 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
459 		}
460 	}
461 
462 	/*
463 	 * Decide if we can use TCP Segmentation Offloading (if supported by
464 	 * hardware).
465 	 *
466 	 * TSO may only be used if we are in a pure bulk sending state.  The
467 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
468 	 * IP options prevent using TSO.  With TSO the TCP header is the same
469 	 * (except for the sequence number) for all generated packets.  This
470 	 * makes it impossible to transmit any options which vary per generated
471 	 * segment or packet.
472 	 */
473 #ifdef IPSEC
474 	/*
475 	 * Pre-calculate here as we save another lookup into the darknesses
476 	 * of IPsec that way and can actually decide if TSO is ok.
477 	 */
478 	ipsec_optlen = ipsec_hdrsiz_tcp(tp);
479 #endif
480 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
481 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
482 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
483 #ifdef IPSEC
484 	    ipsec_optlen == 0 &&
485 #endif
486 	    tp->t_inpcb->inp_options == NULL &&
487 	    tp->t_inpcb->in6p_options == NULL)
488 		tso = 1;
489 
490 	if (sack_rxmit) {
491 		if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
492 			flags &= ~TH_FIN;
493 	} else {
494 		if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
495 			flags &= ~TH_FIN;
496 	}
497 
498 	recwin = sbspace(&so->so_rcv);
499 
500 	/*
501 	 * Sender silly window avoidance.   We transmit under the following
502 	 * conditions when len is non-zero:
503 	 *
504 	 *	- We have a full segment (or more with TSO)
505 	 *	- This is the last buffer in a write()/send() and we are
506 	 *	  either idle or running NODELAY
507 	 *	- we've timed out (e.g. persist timer)
508 	 *	- we have more then 1/2 the maximum send window's worth of
509 	 *	  data (receiver may be limited the window size)
510 	 *	- we need to retransmit
511 	 */
512 	if (len) {
513 		if (len >= tp->t_maxseg)
514 			goto send;
515 		/*
516 		 * NOTE! on localhost connections an 'ack' from the remote
517 		 * end may occur synchronously with the output and cause
518 		 * us to flush a buffer queued with moretocome.  XXX
519 		 *
520 		 * note: the len + off check is almost certainly unnecessary.
521 		 */
522 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
523 		    (idle || (tp->t_flags & TF_NODELAY)) &&
524 		    len + off >= so->so_snd.sb_cc &&
525 		    (tp->t_flags & TF_NOPUSH) == 0) {
526 			goto send;
527 		}
528 		if (tp->t_flags & TF_FORCEDATA)		/* typ. timeout case */
529 			goto send;
530 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
531 			goto send;
532 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
533 			goto send;
534 		if (sack_rxmit)
535 			goto send;
536 	}
537 
538 	/*
539 	 * Compare available window to amount of window
540 	 * known to peer (as advertised window less
541 	 * next expected input).  If the difference is at least two
542 	 * max size segments, or at least 50% of the maximum possible
543 	 * window, then want to send a window update to peer.
544 	 * Skip this if the connection is in T/TCP half-open state.
545 	 * Don't send pure window updates when the peer has closed
546 	 * the connection and won't ever send more data.
547 	 */
548 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
549 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
550 		/*
551 		 * "adv" is the amount we can increase the window,
552 		 * taking into account that we are limited by
553 		 * TCP_MAXWIN << tp->rcv_scale.
554 		 */
555 		long adv;
556 		int oldwin;
557 
558 		adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale);
559 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
560 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
561 			adv -= oldwin;
562 		} else
563 			oldwin = 0;
564 
565 		/*
566 		 * If the new window size ends up being the same as the old
567 		 * size when it is scaled, then don't force a window update.
568 		 */
569 		if (oldwin >> tp->rcv_scale == (adv + oldwin) >> tp->rcv_scale)
570 			goto dontupdate;
571 		if (adv >= (long) (2 * tp->t_maxseg))
572 			goto send;
573 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
574 			goto send;
575 	}
576 dontupdate:
577 
578 	/*
579 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
580 	 * is also a catch-all for the retransmit timer timeout case.
581 	 */
582 	if (tp->t_flags & TF_ACKNOW)
583 		goto send;
584 	if ((flags & TH_RST) ||
585 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
586 		goto send;
587 	if (SEQ_GT(tp->snd_up, tp->snd_una))
588 		goto send;
589 	/*
590 	 * If our state indicates that FIN should be sent
591 	 * and we have not yet done so, then we need to send.
592 	 */
593 	if (flags & TH_FIN &&
594 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
595 		goto send;
596 	/*
597 	 * In SACK, it is possible for tcp_output to fail to send a segment
598 	 * after the retransmission timer has been turned off.  Make sure
599 	 * that the retransmission timer is set.
600 	 */
601 	if ((tp->t_flags & TF_SACK_PERMIT) &&
602 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
603 	    !tcp_timer_active(tp, TT_REXMT) &&
604 	    !tcp_timer_active(tp, TT_PERSIST)) {
605 		tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
606 		goto just_return;
607 	}
608 	/*
609 	 * TCP window updates are not reliable, rather a polling protocol
610 	 * using ``persist'' packets is used to insure receipt of window
611 	 * updates.  The three ``states'' for the output side are:
612 	 *	idle			not doing retransmits or persists
613 	 *	persisting		to move a small or zero window
614 	 *	(re)transmitting	and thereby not persisting
615 	 *
616 	 * tcp_timer_active(tp, TT_PERSIST)
617 	 *	is true when we are in persist state.
618 	 * (tp->t_flags & TF_FORCEDATA)
619 	 *	is set when we are called to send a persist packet.
620 	 * tcp_timer_active(tp, TT_REXMT)
621 	 *	is set when we are retransmitting
622 	 * The output side is idle when both timers are zero.
623 	 *
624 	 * If send window is too small, there is data to transmit, and no
625 	 * retransmit or persist is pending, then go to persist state.
626 	 * If nothing happens soon, send when timer expires:
627 	 * if window is nonzero, transmit what we can,
628 	 * otherwise force out a byte.
629 	 */
630 	if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) &&
631 	    !tcp_timer_active(tp, TT_PERSIST)) {
632 		tp->t_rxtshift = 0;
633 		tcp_setpersist(tp);
634 	}
635 
636 	/*
637 	 * No reason to send a segment, just return.
638 	 */
639 just_return:
640 	SOCKBUF_UNLOCK(&so->so_snd);
641 	return (0);
642 
643 send:
644 	SOCKBUF_LOCK_ASSERT(&so->so_snd);
645 	/*
646 	 * Before ESTABLISHED, force sending of initial options
647 	 * unless TCP set not to do any options.
648 	 * NOTE: we assume that the IP/TCP header plus TCP options
649 	 * always fit in a single mbuf, leaving room for a maximum
650 	 * link header, i.e.
651 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
652 	 */
653 	optlen = 0;
654 #ifdef INET6
655 	if (isipv6)
656 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
657 	else
658 #endif
659 		hdrlen = sizeof (struct tcpiphdr);
660 
661 	/*
662 	 * Compute options for segment.
663 	 * We only have to care about SYN and established connection
664 	 * segments.  Options for SYN-ACK segments are handled in TCP
665 	 * syncache.
666 	 */
667 	if ((tp->t_flags & TF_NOOPT) == 0) {
668 		to.to_flags = 0;
669 		/* Maximum segment size. */
670 		if (flags & TH_SYN) {
671 			tp->snd_nxt = tp->iss;
672 			to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
673 			to.to_flags |= TOF_MSS;
674 		}
675 		/* Window scaling. */
676 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
677 			to.to_wscale = tp->request_r_scale;
678 			to.to_flags |= TOF_SCALE;
679 		}
680 		/* Timestamps. */
681 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
682 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
683 			to.to_tsval = tcp_ts_getticks() + tp->ts_offset;
684 			to.to_tsecr = tp->ts_recent;
685 			to.to_flags |= TOF_TS;
686 			/* Set receive buffer autosizing timestamp. */
687 			if (tp->rfbuf_ts == 0 &&
688 			    (so->so_rcv.sb_flags & SB_AUTOSIZE))
689 				tp->rfbuf_ts = tcp_ts_getticks();
690 		}
691 		/* Selective ACK's. */
692 		if (tp->t_flags & TF_SACK_PERMIT) {
693 			if (flags & TH_SYN)
694 				to.to_flags |= TOF_SACKPERM;
695 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
696 			    (tp->t_flags & TF_SACK_PERMIT) &&
697 			    tp->rcv_numsacks > 0) {
698 				to.to_flags |= TOF_SACK;
699 				to.to_nsacks = tp->rcv_numsacks;
700 				to.to_sacks = (u_char *)tp->sackblks;
701 			}
702 		}
703 #ifdef TCP_SIGNATURE
704 		/* TCP-MD5 (RFC2385). */
705 		if (tp->t_flags & TF_SIGNATURE)
706 			to.to_flags |= TOF_SIGNATURE;
707 #endif /* TCP_SIGNATURE */
708 
709 		/* Processing the options. */
710 		hdrlen += optlen = tcp_addoptions(&to, opt);
711 	}
712 
713 #ifdef INET6
714 	if (isipv6)
715 		ipoptlen = ip6_optlen(tp->t_inpcb);
716 	else
717 #endif
718 	if (tp->t_inpcb->inp_options)
719 		ipoptlen = tp->t_inpcb->inp_options->m_len -
720 				offsetof(struct ipoption, ipopt_list);
721 	else
722 		ipoptlen = 0;
723 #ifdef IPSEC
724 	ipoptlen += ipsec_optlen;
725 #endif
726 
727 	/*
728 	 * Adjust data length if insertion of options will
729 	 * bump the packet length beyond the t_maxopd length.
730 	 * Clear the FIN bit because we cut off the tail of
731 	 * the segment.
732 	 */
733 	if (len + optlen + ipoptlen > tp->t_maxopd) {
734 		flags &= ~TH_FIN;
735 
736 		if (tso) {
737 			KASSERT(ipoptlen == 0,
738 			    ("%s: TSO can't do IP options", __func__));
739 
740 			/*
741 			 * Limit a burst to IP_MAXPACKET minus IP,
742 			 * TCP and options length to keep ip->ip_len
743 			 * from overflowing.
744 			 */
745 			if (len > IP_MAXPACKET - hdrlen) {
746 				len = IP_MAXPACKET - hdrlen;
747 				sendalot = 1;
748 			}
749 
750 			/*
751 			 * Prevent the last segment from being
752 			 * fractional unless the send sockbuf can
753 			 * be emptied.
754 			 */
755 			if (sendalot && off + len < so->so_snd.sb_cc) {
756 				len -= len % (tp->t_maxopd - optlen);
757 				sendalot = 1;
758 			}
759 
760 			/*
761 			 * Send the FIN in a separate segment
762 			 * after the bulk sending is done.
763 			 * We don't trust the TSO implementations
764 			 * to clear the FIN flag on all but the
765 			 * last segment.
766 			 */
767 			if (tp->t_flags & TF_NEEDFIN)
768 				sendalot = 1;
769 
770 		} else {
771 			len = tp->t_maxopd - optlen - ipoptlen;
772 			sendalot = 1;
773 		}
774 	} else
775 		tso = 0;
776 
777 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
778 	    ("%s: len > IP_MAXPACKET", __func__));
779 
780 /*#ifdef DIAGNOSTIC*/
781 #ifdef INET6
782 	if (max_linkhdr + hdrlen > MCLBYTES)
783 #else
784 	if (max_linkhdr + hdrlen > MHLEN)
785 #endif
786 		panic("tcphdr too big");
787 /*#endif*/
788 
789 	/*
790 	 * This KASSERT is here to catch edge cases at a well defined place.
791 	 * Before, those had triggered (random) panic conditions further down.
792 	 */
793 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
794 
795 	/*
796 	 * Grab a header mbuf, attaching a copy of data to
797 	 * be transmitted, and initialize the header from
798 	 * the template for sends on this connection.
799 	 */
800 	if (len) {
801 		struct mbuf *mb;
802 		u_int moff;
803 
804 		if ((tp->t_flags & TF_FORCEDATA) && len == 1)
805 			TCPSTAT_INC(tcps_sndprobe);
806 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
807 			tp->t_sndrexmitpack++;
808 			TCPSTAT_INC(tcps_sndrexmitpack);
809 			TCPSTAT_ADD(tcps_sndrexmitbyte, len);
810 		} else {
811 			TCPSTAT_INC(tcps_sndpack);
812 			TCPSTAT_ADD(tcps_sndbyte, len);
813 		}
814 		MGETHDR(m, M_DONTWAIT, MT_DATA);
815 		if (m == NULL) {
816 			SOCKBUF_UNLOCK(&so->so_snd);
817 			error = ENOBUFS;
818 			goto out;
819 		}
820 #ifdef INET6
821 		if (MHLEN < hdrlen + max_linkhdr) {
822 			MCLGET(m, M_DONTWAIT);
823 			if ((m->m_flags & M_EXT) == 0) {
824 				SOCKBUF_UNLOCK(&so->so_snd);
825 				m_freem(m);
826 				error = ENOBUFS;
827 				goto out;
828 			}
829 		}
830 #endif
831 		m->m_data += max_linkhdr;
832 		m->m_len = hdrlen;
833 
834 		/*
835 		 * Start the m_copy functions from the closest mbuf
836 		 * to the offset in the socket buffer chain.
837 		 */
838 		mb = sbsndptr(&so->so_snd, off, len, &moff);
839 
840 		if (len <= MHLEN - hdrlen - max_linkhdr) {
841 			m_copydata(mb, moff, (int)len,
842 			    mtod(m, caddr_t) + hdrlen);
843 			m->m_len += len;
844 		} else {
845 			m->m_next = m_copy(mb, moff, (int)len);
846 			if (m->m_next == NULL) {
847 				SOCKBUF_UNLOCK(&so->so_snd);
848 				(void) m_free(m);
849 				error = ENOBUFS;
850 				goto out;
851 			}
852 		}
853 
854 		/*
855 		 * If we're sending everything we've got, set PUSH.
856 		 * (This will keep happy those implementations which only
857 		 * give data to the user when a buffer fills or
858 		 * a PUSH comes in.)
859 		 */
860 		if (off + len == so->so_snd.sb_cc)
861 			flags |= TH_PUSH;
862 		SOCKBUF_UNLOCK(&so->so_snd);
863 	} else {
864 		SOCKBUF_UNLOCK(&so->so_snd);
865 		if (tp->t_flags & TF_ACKNOW)
866 			TCPSTAT_INC(tcps_sndacks);
867 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
868 			TCPSTAT_INC(tcps_sndctrl);
869 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
870 			TCPSTAT_INC(tcps_sndurg);
871 		else
872 			TCPSTAT_INC(tcps_sndwinup);
873 
874 		MGETHDR(m, M_DONTWAIT, MT_DATA);
875 		if (m == NULL) {
876 			error = ENOBUFS;
877 			goto out;
878 		}
879 #ifdef INET6
880 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
881 		    MHLEN >= hdrlen) {
882 			MH_ALIGN(m, hdrlen);
883 		} else
884 #endif
885 		m->m_data += max_linkhdr;
886 		m->m_len = hdrlen;
887 	}
888 	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
889 	m->m_pkthdr.rcvif = (struct ifnet *)0;
890 #ifdef MAC
891 	mac_inpcb_create_mbuf(tp->t_inpcb, m);
892 #endif
893 #ifdef INET6
894 	if (isipv6) {
895 		ip6 = mtod(m, struct ip6_hdr *);
896 		th = (struct tcphdr *)(ip6 + 1);
897 		tcpip_fillheaders(tp->t_inpcb, ip6, th);
898 	} else
899 #endif /* INET6 */
900 	{
901 		ip = mtod(m, struct ip *);
902 		ipov = (struct ipovly *)ip;
903 		th = (struct tcphdr *)(ip + 1);
904 		tcpip_fillheaders(tp->t_inpcb, ip, th);
905 	}
906 
907 	/*
908 	 * Fill in fields, remembering maximum advertised
909 	 * window for use in delaying messages about window sizes.
910 	 * If resending a FIN, be sure not to use a new sequence number.
911 	 */
912 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
913 	    tp->snd_nxt == tp->snd_max)
914 		tp->snd_nxt--;
915 	/*
916 	 * If we are starting a connection, send ECN setup
917 	 * SYN packet. If we are on a retransmit, we may
918 	 * resend those bits a number of times as per
919 	 * RFC 3168.
920 	 */
921 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
922 		if (tp->t_rxtshift >= 1) {
923 			if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
924 				flags |= TH_ECE|TH_CWR;
925 		} else
926 			flags |= TH_ECE|TH_CWR;
927 	}
928 
929 	if (tp->t_state == TCPS_ESTABLISHED &&
930 	    (tp->t_flags & TF_ECN_PERMIT)) {
931 		/*
932 		 * If the peer has ECN, mark data packets with
933 		 * ECN capable transmission (ECT).
934 		 * Ignore pure ack packets, retransmissions and window probes.
935 		 */
936 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
937 		    !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
938 #ifdef INET6
939 			if (isipv6)
940 				ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
941 			else
942 #endif
943 				ip->ip_tos |= IPTOS_ECN_ECT0;
944 			TCPSTAT_INC(tcps_ecn_ect0);
945 		}
946 
947 		/*
948 		 * Reply with proper ECN notifications.
949 		 */
950 		if (tp->t_flags & TF_ECN_SND_CWR) {
951 			flags |= TH_CWR;
952 			tp->t_flags &= ~TF_ECN_SND_CWR;
953 		}
954 		if (tp->t_flags & TF_ECN_SND_ECE)
955 			flags |= TH_ECE;
956 	}
957 
958 	/*
959 	 * If we are doing retransmissions, then snd_nxt will
960 	 * not reflect the first unsent octet.  For ACK only
961 	 * packets, we do not want the sequence number of the
962 	 * retransmitted packet, we want the sequence number
963 	 * of the next unsent octet.  So, if there is no data
964 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
965 	 * when filling in ti_seq.  But if we are in persist
966 	 * state, snd_max might reflect one byte beyond the
967 	 * right edge of the window, so use snd_nxt in that
968 	 * case, since we know we aren't doing a retransmission.
969 	 * (retransmit and persist are mutually exclusive...)
970 	 */
971 	if (sack_rxmit == 0) {
972 		if (len || (flags & (TH_SYN|TH_FIN)) ||
973 		    tcp_timer_active(tp, TT_PERSIST))
974 			th->th_seq = htonl(tp->snd_nxt);
975 		else
976 			th->th_seq = htonl(tp->snd_max);
977 	} else {
978 		th->th_seq = htonl(p->rxmit);
979 		p->rxmit += len;
980 		tp->sackhint.sack_bytes_rexmit += len;
981 	}
982 	th->th_ack = htonl(tp->rcv_nxt);
983 	if (optlen) {
984 		bcopy(opt, th + 1, optlen);
985 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
986 	}
987 	th->th_flags = flags;
988 	/*
989 	 * Calculate receive window.  Don't shrink window,
990 	 * but avoid silly window syndrome.
991 	 */
992 	if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
993 	    recwin < (long)tp->t_maxseg)
994 		recwin = 0;
995 	if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
996 	    recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
997 		recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
998 	if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
999 		recwin = (long)TCP_MAXWIN << tp->rcv_scale;
1000 
1001 	/*
1002 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1003 	 * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
1004 	 * case is handled in syncache.
1005 	 */
1006 	if (flags & TH_SYN)
1007 		th->th_win = htons((u_short)
1008 				(min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1009 	else
1010 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1011 
1012 	/*
1013 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1014 	 * a 0 window.  This may cause the remote transmitter to stall.  This
1015 	 * flag tells soreceive() to disable delayed acknowledgements when
1016 	 * draining the buffer.  This can occur if the receiver is attempting
1017 	 * to read more data than can be buffered prior to transmitting on
1018 	 * the connection.
1019 	 */
1020 	if (th->th_win == 0) {
1021 		tp->t_sndzerowin++;
1022 		tp->t_flags |= TF_RXWIN0SENT;
1023 	} else
1024 		tp->t_flags &= ~TF_RXWIN0SENT;
1025 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1026 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1027 		th->th_flags |= TH_URG;
1028 	} else
1029 		/*
1030 		 * If no urgent pointer to send, then we pull
1031 		 * the urgent pointer to the left edge of the send window
1032 		 * so that it doesn't drift into the send window on sequence
1033 		 * number wraparound.
1034 		 */
1035 		tp->snd_up = tp->snd_una;		/* drag it along */
1036 
1037 #ifdef TCP_SIGNATURE
1038 	if (tp->t_flags & TF_SIGNATURE) {
1039 		int sigoff = to.to_signature - opt;
1040 		tcp_signature_compute(m, 0, len, optlen,
1041 		    (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
1042 	}
1043 #endif
1044 
1045 	/*
1046 	 * Put TCP length in extended header, and then
1047 	 * checksum extended header and data.
1048 	 */
1049 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1050 #ifdef INET6
1051 	if (isipv6)
1052 		/*
1053 		 * ip6_plen is not need to be filled now, and will be filled
1054 		 * in ip6_output.
1055 		 */
1056 		th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
1057 				       sizeof(struct tcphdr) + optlen + len);
1058 	else
1059 #endif /* INET6 */
1060 	{
1061 		m->m_pkthdr.csum_flags = CSUM_TCP;
1062 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1063 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1064 		    htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1065 
1066 		/* IP version must be set here for ipv4/ipv6 checking later */
1067 		KASSERT(ip->ip_v == IPVERSION,
1068 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1069 	}
1070 
1071 	/*
1072 	 * Enable TSO and specify the size of the segments.
1073 	 * The TCP pseudo header checksum is always provided.
1074 	 * XXX: Fixme: This is currently not the case for IPv6.
1075 	 */
1076 	if (tso) {
1077 		KASSERT(len > tp->t_maxopd - optlen,
1078 		    ("%s: len <= tso_segsz", __func__));
1079 		m->m_pkthdr.csum_flags |= CSUM_TSO;
1080 		m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen;
1081 	}
1082 
1083 #ifdef IPSEC
1084 	KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL),
1085 	    ("%s: mbuf chain shorter than expected: %ld + %u + %u - %u != %u",
1086 	    __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL)));
1087 #else
1088 	KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL),
1089 	    ("%s: mbuf chain shorter than expected: %ld + %u + %u != %u",
1090 	    __func__, len, hdrlen, ipoptlen, m_length(m, NULL)));
1091 #endif
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 /* TCPDEBUG */
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 	}
1212 #endif /* INET6 */
1213 #if defined(INET) && defined(INET6)
1214 	else
1215 #endif
1216 #ifdef INET
1217     {
1218 	ip->ip_len = m->m_pkthdr.len;
1219 #ifdef INET6
1220 	if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1221 		ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1222 #endif /* INET6 */
1223 	/*
1224 	 * If we do path MTU discovery, then we set DF on every packet.
1225 	 * This might not be the best thing to do according to RFC3390
1226 	 * Section 2. However the tcp hostcache migitates the problem
1227 	 * so it affects only the first tcp connection with a host.
1228 	 *
1229 	 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1230 	 */
1231 	if (V_path_mtu_discovery && tp->t_maxopd > V_tcp_minmss)
1232 		ip->ip_off |= IP_DF;
1233 
1234 	error = ip_output(m, tp->t_inpcb->inp_options, NULL,
1235 	    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1236 	    tp->t_inpcb);
1237     }
1238 #endif /* INET */
1239 	if (error) {
1240 
1241 		/*
1242 		 * We know that the packet was lost, so back out the
1243 		 * sequence number advance, if any.
1244 		 *
1245 		 * If the error is EPERM the packet got blocked by the
1246 		 * local firewall.  Normally we should terminate the
1247 		 * connection but the blocking may have been spurious
1248 		 * due to a firewall reconfiguration cycle.  So we treat
1249 		 * it like a packet loss and let the retransmit timer and
1250 		 * timeouts do their work over time.
1251 		 * XXX: It is a POLA question whether calling tcp_drop right
1252 		 * away would be the really correct behavior instead.
1253 		 */
1254 		if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1255 		    !tcp_timer_active(tp, TT_PERSIST)) &&
1256 		    ((flags & TH_SYN) == 0) &&
1257 		    (error != EPERM)) {
1258 			if (sack_rxmit) {
1259 				p->rxmit -= len;
1260 				tp->sackhint.sack_bytes_rexmit -= len;
1261 				KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1262 				    ("sackhint bytes rtx >= 0"));
1263 			} else
1264 				tp->snd_nxt -= len;
1265 		}
1266 out:
1267 		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1268 		switch (error) {
1269 		case EPERM:
1270 			tp->t_softerror = error;
1271 			return (error);
1272 		case ENOBUFS:
1273 	                if (!tcp_timer_active(tp, TT_REXMT) &&
1274 			    !tcp_timer_active(tp, TT_PERSIST))
1275 	                        tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1276 			tp->snd_cwnd = tp->t_maxseg;
1277 			return (0);
1278 		case EMSGSIZE:
1279 			/*
1280 			 * For some reason the interface we used initially
1281 			 * to send segments changed to another or lowered
1282 			 * its MTU.
1283 			 *
1284 			 * tcp_mtudisc() will find out the new MTU and as
1285 			 * its last action, initiate retransmission, so it
1286 			 * is important to not do so here.
1287 			 *
1288 			 * If TSO was active we either got an interface
1289 			 * without TSO capabilits or TSO was turned off.
1290 			 * Disable it for this connection as too and
1291 			 * immediatly retry with MSS sized segments generated
1292 			 * by this function.
1293 			 */
1294 			if (tso)
1295 				tp->t_flags &= ~TF_TSO;
1296 			tcp_mtudisc(tp->t_inpcb, 0);
1297 			return (0);
1298 		case EHOSTDOWN:
1299 		case EHOSTUNREACH:
1300 		case ENETDOWN:
1301 		case ENETUNREACH:
1302 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
1303 				tp->t_softerror = error;
1304 				return (0);
1305 			}
1306 			/* FALLTHROUGH */
1307 		default:
1308 			return (error);
1309 		}
1310 	}
1311 	TCPSTAT_INC(tcps_sndtotal);
1312 
1313 	/*
1314 	 * Data sent (as far as we can tell).
1315 	 * If this advertises a larger window than any other segment,
1316 	 * then remember the size of the advertised window.
1317 	 * Any pending ACK has now been sent.
1318 	 */
1319 	if (recwin >= 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1320 		tp->rcv_adv = tp->rcv_nxt + recwin;
1321 	tp->last_ack_sent = tp->rcv_nxt;
1322 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1323 	if (tcp_timer_active(tp, TT_DELACK))
1324 		tcp_timer_activate(tp, TT_DELACK, 0);
1325 #if 0
1326 	/*
1327 	 * This completely breaks TCP if newreno is turned on.  What happens
1328 	 * is that if delayed-acks are turned on on the receiver, this code
1329 	 * on the transmitter effectively destroys the TCP window, forcing
1330 	 * it to four packets (1.5Kx4 = 6K window).
1331 	 */
1332 	if (sendalot && --maxburst)
1333 		goto again;
1334 #endif
1335 	if (sendalot)
1336 		goto again;
1337 	return (0);
1338 }
1339 
1340 void
1341 tcp_setpersist(struct tcpcb *tp)
1342 {
1343 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1344 	int tt;
1345 
1346 	tp->t_flags &= ~TF_PREVVALID;
1347 	if (tcp_timer_active(tp, TT_REXMT))
1348 		panic("tcp_setpersist: retransmit pending");
1349 	/*
1350 	 * Start/restart persistance timer.
1351 	 */
1352 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1353 		      TCPTV_PERSMIN, TCPTV_PERSMAX);
1354 	tcp_timer_activate(tp, TT_PERSIST, tt);
1355 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1356 		tp->t_rxtshift++;
1357 }
1358 
1359 /*
1360  * Insert TCP options according to the supplied parameters to the place
1361  * optp in a consistent way.  Can handle unaligned destinations.
1362  *
1363  * The order of the option processing is crucial for optimal packing and
1364  * alignment for the scarce option space.
1365  *
1366  * The optimal order for a SYN/SYN-ACK segment is:
1367  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1368  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1369  *
1370  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1371  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1372  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1373  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1374  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1375  */
1376 int
1377 tcp_addoptions(struct tcpopt *to, u_char *optp)
1378 {
1379 	u_int mask, optlen = 0;
1380 
1381 	for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1382 		if ((to->to_flags & mask) != mask)
1383 			continue;
1384 		if (optlen == TCP_MAXOLEN)
1385 			break;
1386 		switch (to->to_flags & mask) {
1387 		case TOF_MSS:
1388 			while (optlen % 4) {
1389 				optlen += TCPOLEN_NOP;
1390 				*optp++ = TCPOPT_NOP;
1391 			}
1392 			if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1393 				continue;
1394 			optlen += TCPOLEN_MAXSEG;
1395 			*optp++ = TCPOPT_MAXSEG;
1396 			*optp++ = TCPOLEN_MAXSEG;
1397 			to->to_mss = htons(to->to_mss);
1398 			bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1399 			optp += sizeof(to->to_mss);
1400 			break;
1401 		case TOF_SCALE:
1402 			while (!optlen || optlen % 2 != 1) {
1403 				optlen += TCPOLEN_NOP;
1404 				*optp++ = TCPOPT_NOP;
1405 			}
1406 			if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1407 				continue;
1408 			optlen += TCPOLEN_WINDOW;
1409 			*optp++ = TCPOPT_WINDOW;
1410 			*optp++ = TCPOLEN_WINDOW;
1411 			*optp++ = to->to_wscale;
1412 			break;
1413 		case TOF_SACKPERM:
1414 			while (optlen % 2) {
1415 				optlen += TCPOLEN_NOP;
1416 				*optp++ = TCPOPT_NOP;
1417 			}
1418 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1419 				continue;
1420 			optlen += TCPOLEN_SACK_PERMITTED;
1421 			*optp++ = TCPOPT_SACK_PERMITTED;
1422 			*optp++ = TCPOLEN_SACK_PERMITTED;
1423 			break;
1424 		case TOF_TS:
1425 			while (!optlen || optlen % 4 != 2) {
1426 				optlen += TCPOLEN_NOP;
1427 				*optp++ = TCPOPT_NOP;
1428 			}
1429 			if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1430 				continue;
1431 			optlen += TCPOLEN_TIMESTAMP;
1432 			*optp++ = TCPOPT_TIMESTAMP;
1433 			*optp++ = TCPOLEN_TIMESTAMP;
1434 			to->to_tsval = htonl(to->to_tsval);
1435 			to->to_tsecr = htonl(to->to_tsecr);
1436 			bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1437 			optp += sizeof(to->to_tsval);
1438 			bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1439 			optp += sizeof(to->to_tsecr);
1440 			break;
1441 		case TOF_SIGNATURE:
1442 			{
1443 			int siglen = TCPOLEN_SIGNATURE - 2;
1444 
1445 			while (!optlen || optlen % 4 != 2) {
1446 				optlen += TCPOLEN_NOP;
1447 				*optp++ = TCPOPT_NOP;
1448 			}
1449 			if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE)
1450 				continue;
1451 			optlen += TCPOLEN_SIGNATURE;
1452 			*optp++ = TCPOPT_SIGNATURE;
1453 			*optp++ = TCPOLEN_SIGNATURE;
1454 			to->to_signature = optp;
1455 			while (siglen--)
1456 				 *optp++ = 0;
1457 			break;
1458 			}
1459 		case TOF_SACK:
1460 			{
1461 			int sackblks = 0;
1462 			struct sackblk *sack = (struct sackblk *)to->to_sacks;
1463 			tcp_seq sack_seq;
1464 
1465 			while (!optlen || optlen % 4 != 2) {
1466 				optlen += TCPOLEN_NOP;
1467 				*optp++ = TCPOPT_NOP;
1468 			}
1469 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1470 				continue;
1471 			optlen += TCPOLEN_SACKHDR;
1472 			*optp++ = TCPOPT_SACK;
1473 			sackblks = min(to->to_nsacks,
1474 					(TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1475 			*optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1476 			while (sackblks--) {
1477 				sack_seq = htonl(sack->start);
1478 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1479 				optp += sizeof(sack_seq);
1480 				sack_seq = htonl(sack->end);
1481 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1482 				optp += sizeof(sack_seq);
1483 				optlen += TCPOLEN_SACK;
1484 				sack++;
1485 			}
1486 			TCPSTAT_INC(tcps_sack_send_blocks);
1487 			break;
1488 			}
1489 		default:
1490 			panic("%s: unknown TCP option type", __func__);
1491 			break;
1492 		}
1493 	}
1494 
1495 	/* Terminate and pad TCP options to a 4 byte boundary. */
1496 	if (optlen % 4) {
1497 		optlen += TCPOLEN_EOL;
1498 		*optp++ = TCPOPT_EOL;
1499 	}
1500 	/*
1501 	 * According to RFC 793 (STD0007):
1502 	 *   "The content of the header beyond the End-of-Option option
1503 	 *    must be header padding (i.e., zero)."
1504 	 *   and later: "The padding is composed of zeros."
1505 	 */
1506 	while (optlen % 4) {
1507 		optlen += TCPOLEN_PAD;
1508 		*optp++ = TCPOPT_PAD;
1509 	}
1510 
1511 	KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1512 	return (optlen);
1513 }
1514