xref: /freebsd/sys/netinet/tcp_output.c (revision 7778ab7e0cc22f0824eb1d1047a7ef8b4785267a)
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, ss_fltsz) = 1;
93 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, slowstart_flightsize, CTLFLAG_RW,
94 	&VNET_NAME(ss_fltsz), 1,
95 	"Slow start flight size");
96 
97 VNET_DEFINE(int, ss_fltsz_local) = 4;
98 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, local_slowstart_flightsize,
99 	CTLFLAG_RW, &VNET_NAME(ss_fltsz_local), 1,
100 	"Slow start flight size for local networks");
101 
102 VNET_DEFINE(int, tcp_do_tso) = 1;
103 #define	V_tcp_do_tso		VNET(tcp_do_tso)
104 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW,
105 	&VNET_NAME(tcp_do_tso), 0,
106 	"Enable TCP Segmentation Offload");
107 
108 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
109 #define	V_tcp_do_autosndbuf	VNET(tcp_do_autosndbuf)
110 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW,
111 	&VNET_NAME(tcp_do_autosndbuf), 0,
112 	"Enable automatic send buffer sizing");
113 
114 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
115 #define	V_tcp_autosndbuf_inc	VNET(tcp_autosndbuf_inc)
116 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW,
117 	&VNET_NAME(tcp_autosndbuf_inc), 0,
118 	"Incrementor step size of automatic send buffer");
119 
120 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
121 #define	V_tcp_autosndbuf_max	VNET(tcp_autosndbuf_max)
122 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW,
123 	&VNET_NAME(tcp_autosndbuf_max), 0,
124 	"Max size of automatic send buffer");
125 
126 static void inline	hhook_run_tcp_est_out(struct tcpcb *tp,
127 			    struct tcphdr *th, struct tcpopt *to,
128 			    long len, int tso);
129 static void inline	cc_after_idle(struct tcpcb *tp);
130 
131 /*
132  * Wrapper for the TCP established ouput helper hook.
133  */
134 static void inline
135 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
136     struct tcpopt *to, long len, int tso)
137 {
138 	struct tcp_hhook_data hhook_data;
139 
140 	if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
141 		hhook_data.tp = tp;
142 		hhook_data.th = th;
143 		hhook_data.to = to;
144 		hhook_data.len = len;
145 		hhook_data.tso = tso;
146 
147 		hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
148 		    tp->osd);
149 	}
150 }
151 
152 /*
153  * CC wrapper hook functions
154  */
155 static void inline
156 cc_after_idle(struct tcpcb *tp)
157 {
158 	INP_WLOCK_ASSERT(tp->t_inpcb);
159 
160 	if (CC_ALGO(tp)->after_idle != NULL)
161 		CC_ALGO(tp)->after_idle(tp->ccv);
162 }
163 
164 /*
165  * Tcp output routine: figure out what should be sent and send it.
166  */
167 int
168 tcp_output(struct tcpcb *tp)
169 {
170 	struct socket *so = tp->t_inpcb->inp_socket;
171 	long len, recwin, sendwin;
172 	int off, flags, error = 0;	/* Keep compiler happy */
173 	struct mbuf *m;
174 	struct ip *ip = NULL;
175 	struct ipovly *ipov = NULL;
176 	struct tcphdr *th;
177 	u_char opt[TCP_MAXOLEN];
178 	unsigned ipoptlen, optlen, hdrlen;
179 #ifdef IPSEC
180 	unsigned ipsec_optlen = 0;
181 #endif
182 	int idle, sendalot;
183 	int sack_rxmit, sack_bytes_rxmt;
184 	struct sackhole *p;
185 	int tso;
186 	struct tcpopt to;
187 #if 0
188 	int maxburst = TCP_MAXBURST;
189 #endif
190 #ifdef INET6
191 	struct ip6_hdr *ip6 = NULL;
192 	int isipv6;
193 
194 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
195 #endif
196 
197 	INP_WLOCK_ASSERT(tp->t_inpcb);
198 
199 	/*
200 	 * Determine length of data that should be transmitted,
201 	 * and flags that will be used.
202 	 * If there is some data or critical controls (SYN, RST)
203 	 * to send, then transmit; otherwise, investigate further.
204 	 */
205 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
206 	if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur)
207 		cc_after_idle(tp);
208 	tp->t_flags &= ~TF_LASTIDLE;
209 	if (idle) {
210 		if (tp->t_flags & TF_MORETOCOME) {
211 			tp->t_flags |= TF_LASTIDLE;
212 			idle = 0;
213 		}
214 	}
215 again:
216 	/*
217 	 * If we've recently taken a timeout, snd_max will be greater than
218 	 * snd_nxt.  There may be SACK information that allows us to avoid
219 	 * resending already delivered data.  Adjust snd_nxt accordingly.
220 	 */
221 	if ((tp->t_flags & TF_SACK_PERMIT) &&
222 	    SEQ_LT(tp->snd_nxt, tp->snd_max))
223 		tcp_sack_adjust(tp);
224 	sendalot = 0;
225 	tso = 0;
226 	off = tp->snd_nxt - tp->snd_una;
227 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
228 
229 	flags = tcp_outflags[tp->t_state];
230 	/*
231 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
232 	 * to send out new data (when sendalot is 1), bypass this function.
233 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
234 	 * we're replacing a (future) new transmission with a retransmission
235 	 * now, and we previously incremented snd_cwnd in tcp_input().
236 	 */
237 	/*
238 	 * Still in sack recovery , reset rxmit flag to zero.
239 	 */
240 	sack_rxmit = 0;
241 	sack_bytes_rxmt = 0;
242 	len = 0;
243 	p = NULL;
244 	if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
245 	    (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
246 		long cwin;
247 
248 		cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
249 		if (cwin < 0)
250 			cwin = 0;
251 		/* Do not retransmit SACK segments beyond snd_recover */
252 		if (SEQ_GT(p->end, tp->snd_recover)) {
253 			/*
254 			 * (At least) part of sack hole extends beyond
255 			 * snd_recover. Check to see if we can rexmit data
256 			 * for this hole.
257 			 */
258 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
259 				/*
260 				 * Can't rexmit any more data for this hole.
261 				 * That data will be rexmitted in the next
262 				 * sack recovery episode, when snd_recover
263 				 * moves past p->rxmit.
264 				 */
265 				p = NULL;
266 				goto after_sack_rexmit;
267 			} else
268 				/* Can rexmit part of the current hole */
269 				len = ((long)ulmin(cwin,
270 						   tp->snd_recover - p->rxmit));
271 		} else
272 			len = ((long)ulmin(cwin, p->end - p->rxmit));
273 		off = p->rxmit - tp->snd_una;
274 		KASSERT(off >= 0,("%s: sack block to the left of una : %d",
275 		    __func__, off));
276 		if (len > 0) {
277 			sack_rxmit = 1;
278 			sendalot = 1;
279 			TCPSTAT_INC(tcps_sack_rexmits);
280 			TCPSTAT_ADD(tcps_sack_rexmit_bytes,
281 			    min(len, tp->t_maxseg));
282 		}
283 	}
284 after_sack_rexmit:
285 	/*
286 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
287 	 * state flags.
288 	 */
289 	if (tp->t_flags & TF_NEEDFIN)
290 		flags |= TH_FIN;
291 	if (tp->t_flags & TF_NEEDSYN)
292 		flags |= TH_SYN;
293 
294 	SOCKBUF_LOCK(&so->so_snd);
295 	/*
296 	 * If in persist timeout with window of 0, send 1 byte.
297 	 * Otherwise, if window is small but nonzero
298 	 * and timer expired, we will send what we can
299 	 * and go to transmit state.
300 	 */
301 	if (tp->t_flags & TF_FORCEDATA) {
302 		if (sendwin == 0) {
303 			/*
304 			 * If we still have some data to send, then
305 			 * clear the FIN bit.  Usually this would
306 			 * happen below when it realizes that we
307 			 * aren't sending all the data.  However,
308 			 * if we have exactly 1 byte of unsent data,
309 			 * then it won't clear the FIN bit below,
310 			 * and if we are in persist state, we wind
311 			 * up sending the packet without recording
312 			 * that we sent the FIN bit.
313 			 *
314 			 * We can't just blindly clear the FIN bit,
315 			 * because if we don't have any more data
316 			 * to send then the probe will be the FIN
317 			 * itself.
318 			 */
319 			if (off < so->so_snd.sb_cc)
320 				flags &= ~TH_FIN;
321 			sendwin = 1;
322 		} else {
323 			tcp_timer_activate(tp, TT_PERSIST, 0);
324 			tp->t_rxtshift = 0;
325 		}
326 	}
327 
328 	/*
329 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
330 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
331 	 * a negative length.  This can also occur when TCP opens up
332 	 * its congestion window while receiving additional duplicate
333 	 * acks after fast-retransmit because TCP will reset snd_nxt
334 	 * to snd_max after the fast-retransmit.
335 	 *
336 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
337 	 * be set to snd_una, the offset will be 0, and the length may
338 	 * wind up 0.
339 	 *
340 	 * If sack_rxmit is true we are retransmitting from the scoreboard
341 	 * in which case len is already set.
342 	 */
343 	if (sack_rxmit == 0) {
344 		if (sack_bytes_rxmt == 0)
345 			len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off);
346 		else {
347 			long cwin;
348 
349                         /*
350 			 * We are inside of a SACK recovery episode and are
351 			 * sending new data, having retransmitted all the
352 			 * data possible in the scoreboard.
353 			 */
354 			len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd)
355 			       - off);
356 			/*
357 			 * Don't remove this (len > 0) check !
358 			 * We explicitly check for len > 0 here (although it
359 			 * isn't really necessary), to work around a gcc
360 			 * optimization issue - to force gcc to compute
361 			 * len above. Without this check, the computation
362 			 * of len is bungled by the optimizer.
363 			 */
364 			if (len > 0) {
365 				cwin = tp->snd_cwnd -
366 					(tp->snd_nxt - tp->sack_newdata) -
367 					sack_bytes_rxmt;
368 				if (cwin < 0)
369 					cwin = 0;
370 				len = lmin(len, cwin);
371 			}
372 		}
373 	}
374 
375 	/*
376 	 * Lop off SYN bit if it has already been sent.  However, if this
377 	 * is SYN-SENT state and if segment contains data and if we don't
378 	 * know that foreign host supports TAO, suppress sending segment.
379 	 */
380 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
381 		if (tp->t_state != TCPS_SYN_RECEIVED)
382 			flags &= ~TH_SYN;
383 		off--, len++;
384 	}
385 
386 	/*
387 	 * Be careful not to send data and/or FIN on SYN segments.
388 	 * This measure is needed to prevent interoperability problems
389 	 * with not fully conformant TCP implementations.
390 	 */
391 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
392 		len = 0;
393 		flags &= ~TH_FIN;
394 	}
395 
396 	if (len < 0) {
397 		/*
398 		 * If FIN has been sent but not acked,
399 		 * but we haven't been called to retransmit,
400 		 * len will be < 0.  Otherwise, window shrank
401 		 * after we sent into it.  If window shrank to 0,
402 		 * cancel pending retransmit, pull snd_nxt back
403 		 * to (closed) window, and set the persist timer
404 		 * if it isn't already going.  If the window didn't
405 		 * close completely, just wait for an ACK.
406 		 */
407 		len = 0;
408 		if (sendwin == 0) {
409 			tcp_timer_activate(tp, TT_REXMT, 0);
410 			tp->t_rxtshift = 0;
411 			tp->snd_nxt = tp->snd_una;
412 			if (!tcp_timer_active(tp, TT_PERSIST))
413 				tcp_setpersist(tp);
414 		}
415 	}
416 
417 	/* len will be >= 0 after this point. */
418 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
419 
420 	/*
421 	 * Automatic sizing of send socket buffer.  Often the send buffer
422 	 * size is not optimally adjusted to the actual network conditions
423 	 * at hand (delay bandwidth product).  Setting the buffer size too
424 	 * small limits throughput on links with high bandwidth and high
425 	 * delay (eg. trans-continental/oceanic links).  Setting the
426 	 * buffer size too big consumes too much real kernel memory,
427 	 * especially with many connections on busy servers.
428 	 *
429 	 * The criteria to step up the send buffer one notch are:
430 	 *  1. receive window of remote host is larger than send buffer
431 	 *     (with a fudge factor of 5/4th);
432 	 *  2. send buffer is filled to 7/8th with data (so we actually
433 	 *     have data to make use of it);
434 	 *  3. send buffer fill has not hit maximal automatic size;
435 	 *  4. our send window (slow start and cogestion controlled) is
436 	 *     larger than sent but unacknowledged data in send buffer.
437 	 *
438 	 * The remote host receive window scaling factor may limit the
439 	 * growing of the send buffer before it reaches its allowed
440 	 * maximum.
441 	 *
442 	 * It scales directly with slow start or congestion window
443 	 * and does at most one step per received ACK.  This fast
444 	 * scaling has the drawback of growing the send buffer beyond
445 	 * what is strictly necessary to make full use of a given
446 	 * delay*bandwith product.  However testing has shown this not
447 	 * to be much of an problem.  At worst we are trading wasting
448 	 * of available bandwith (the non-use of it) for wasting some
449 	 * socket buffer memory.
450 	 *
451 	 * TODO: Shrink send buffer during idle periods together
452 	 * with congestion window.  Requires another timer.  Has to
453 	 * wait for upcoming tcp timer rewrite.
454 	 */
455 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
456 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
457 		    so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
458 		    so->so_snd.sb_cc < V_tcp_autosndbuf_max &&
459 		    sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
460 			if (!sbreserve_locked(&so->so_snd,
461 			    min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
462 			     V_tcp_autosndbuf_max), so, curthread))
463 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
464 		}
465 	}
466 
467 	/*
468 	 * Decide if we can use TCP Segmentation Offloading (if supported by
469 	 * hardware).
470 	 *
471 	 * TSO may only be used if we are in a pure bulk sending state.  The
472 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
473 	 * IP options prevent using TSO.  With TSO the TCP header is the same
474 	 * (except for the sequence number) for all generated packets.  This
475 	 * makes it impossible to transmit any options which vary per generated
476 	 * segment or packet.
477 	 */
478 #ifdef IPSEC
479 	/*
480 	 * Pre-calculate here as we save another lookup into the darknesses
481 	 * of IPsec that way and can actually decide if TSO is ok.
482 	 */
483 	ipsec_optlen = ipsec_hdrsiz_tcp(tp);
484 #endif
485 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
486 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
487 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
488 #ifdef IPSEC
489 	    ipsec_optlen == 0 &&
490 #endif
491 	    tp->t_inpcb->inp_options == NULL &&
492 	    tp->t_inpcb->in6p_options == NULL)
493 		tso = 1;
494 
495 	if (sack_rxmit) {
496 		if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
497 			flags &= ~TH_FIN;
498 	} else {
499 		if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
500 			flags &= ~TH_FIN;
501 	}
502 
503 	recwin = sbspace(&so->so_rcv);
504 
505 	/*
506 	 * Sender silly window avoidance.   We transmit under the following
507 	 * conditions when len is non-zero:
508 	 *
509 	 *	- We have a full segment (or more with TSO)
510 	 *	- This is the last buffer in a write()/send() and we are
511 	 *	  either idle or running NODELAY
512 	 *	- we've timed out (e.g. persist timer)
513 	 *	- we have more then 1/2 the maximum send window's worth of
514 	 *	  data (receiver may be limited the window size)
515 	 *	- we need to retransmit
516 	 */
517 	if (len) {
518 		if (len >= tp->t_maxseg)
519 			goto send;
520 		/*
521 		 * NOTE! on localhost connections an 'ack' from the remote
522 		 * end may occur synchronously with the output and cause
523 		 * us to flush a buffer queued with moretocome.  XXX
524 		 *
525 		 * note: the len + off check is almost certainly unnecessary.
526 		 */
527 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
528 		    (idle || (tp->t_flags & TF_NODELAY)) &&
529 		    len + off >= so->so_snd.sb_cc &&
530 		    (tp->t_flags & TF_NOPUSH) == 0) {
531 			goto send;
532 		}
533 		if (tp->t_flags & TF_FORCEDATA)		/* typ. timeout case */
534 			goto send;
535 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
536 			goto send;
537 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
538 			goto send;
539 		if (sack_rxmit)
540 			goto send;
541 	}
542 
543 	/*
544 	 * Compare available window to amount of window
545 	 * known to peer (as advertised window less
546 	 * next expected input).  If the difference is at least two
547 	 * max size segments, or at least 50% of the maximum possible
548 	 * window, then want to send a window update to peer.
549 	 * Skip this if the connection is in T/TCP half-open state.
550 	 * Don't send pure window updates when the peer has closed
551 	 * the connection and won't ever send more data.
552 	 */
553 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
554 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
555 		/*
556 		 * "adv" is the amount we can increase the window,
557 		 * taking into account that we are limited by
558 		 * TCP_MAXWIN << tp->rcv_scale.
559 		 */
560 		long adv;
561 		int oldwin;
562 
563 		adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale);
564 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
565 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
566 			adv -= oldwin;
567 		} else
568 			oldwin = 0;
569 
570 		/*
571 		 * If the new window size ends up being the same as the old
572 		 * size when it is scaled, then don't force a window update.
573 		 */
574 		if (oldwin >> tp->rcv_scale == (adv + oldwin) >> tp->rcv_scale)
575 			goto dontupdate;
576 		if (adv >= (long) (2 * tp->t_maxseg))
577 			goto send;
578 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
579 			goto send;
580 	}
581 dontupdate:
582 
583 	/*
584 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
585 	 * is also a catch-all for the retransmit timer timeout case.
586 	 */
587 	if (tp->t_flags & TF_ACKNOW)
588 		goto send;
589 	if ((flags & TH_RST) ||
590 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
591 		goto send;
592 	if (SEQ_GT(tp->snd_up, tp->snd_una))
593 		goto send;
594 	/*
595 	 * If our state indicates that FIN should be sent
596 	 * and we have not yet done so, then we need to send.
597 	 */
598 	if (flags & TH_FIN &&
599 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
600 		goto send;
601 	/*
602 	 * In SACK, it is possible for tcp_output to fail to send a segment
603 	 * after the retransmission timer has been turned off.  Make sure
604 	 * that the retransmission timer is set.
605 	 */
606 	if ((tp->t_flags & TF_SACK_PERMIT) &&
607 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
608 	    !tcp_timer_active(tp, TT_REXMT) &&
609 	    !tcp_timer_active(tp, TT_PERSIST)) {
610 		tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
611 		goto just_return;
612 	}
613 	/*
614 	 * TCP window updates are not reliable, rather a polling protocol
615 	 * using ``persist'' packets is used to insure receipt of window
616 	 * updates.  The three ``states'' for the output side are:
617 	 *	idle			not doing retransmits or persists
618 	 *	persisting		to move a small or zero window
619 	 *	(re)transmitting	and thereby not persisting
620 	 *
621 	 * tcp_timer_active(tp, TT_PERSIST)
622 	 *	is true when we are in persist state.
623 	 * (tp->t_flags & TF_FORCEDATA)
624 	 *	is set when we are called to send a persist packet.
625 	 * tcp_timer_active(tp, TT_REXMT)
626 	 *	is set when we are retransmitting
627 	 * The output side is idle when both timers are zero.
628 	 *
629 	 * If send window is too small, there is data to transmit, and no
630 	 * retransmit or persist is pending, then go to persist state.
631 	 * If nothing happens soon, send when timer expires:
632 	 * if window is nonzero, transmit what we can,
633 	 * otherwise force out a byte.
634 	 */
635 	if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) &&
636 	    !tcp_timer_active(tp, TT_PERSIST)) {
637 		tp->t_rxtshift = 0;
638 		tcp_setpersist(tp);
639 	}
640 
641 	/*
642 	 * No reason to send a segment, just return.
643 	 */
644 just_return:
645 	SOCKBUF_UNLOCK(&so->so_snd);
646 	return (0);
647 
648 send:
649 	SOCKBUF_LOCK_ASSERT(&so->so_snd);
650 	/*
651 	 * Before ESTABLISHED, force sending of initial options
652 	 * unless TCP set not to do any options.
653 	 * NOTE: we assume that the IP/TCP header plus TCP options
654 	 * always fit in a single mbuf, leaving room for a maximum
655 	 * link header, i.e.
656 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
657 	 */
658 	optlen = 0;
659 #ifdef INET6
660 	if (isipv6)
661 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
662 	else
663 #endif
664 		hdrlen = sizeof (struct tcpiphdr);
665 
666 	/*
667 	 * Compute options for segment.
668 	 * We only have to care about SYN and established connection
669 	 * segments.  Options for SYN-ACK segments are handled in TCP
670 	 * syncache.
671 	 */
672 	if ((tp->t_flags & TF_NOOPT) == 0) {
673 		to.to_flags = 0;
674 		/* Maximum segment size. */
675 		if (flags & TH_SYN) {
676 			tp->snd_nxt = tp->iss;
677 			to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
678 			to.to_flags |= TOF_MSS;
679 		}
680 		/* Window scaling. */
681 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
682 			to.to_wscale = tp->request_r_scale;
683 			to.to_flags |= TOF_SCALE;
684 		}
685 		/* Timestamps. */
686 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
687 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
688 			to.to_tsval = ticks + tp->ts_offset;
689 			to.to_tsecr = tp->ts_recent;
690 			to.to_flags |= TOF_TS;
691 			/* Set receive buffer autosizing timestamp. */
692 			if (tp->rfbuf_ts == 0 &&
693 			    (so->so_rcv.sb_flags & SB_AUTOSIZE))
694 				tp->rfbuf_ts = ticks;
695 		}
696 		/* Selective ACK's. */
697 		if (tp->t_flags & TF_SACK_PERMIT) {
698 			if (flags & TH_SYN)
699 				to.to_flags |= TOF_SACKPERM;
700 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
701 			    (tp->t_flags & TF_SACK_PERMIT) &&
702 			    tp->rcv_numsacks > 0) {
703 				to.to_flags |= TOF_SACK;
704 				to.to_nsacks = tp->rcv_numsacks;
705 				to.to_sacks = (u_char *)tp->sackblks;
706 			}
707 		}
708 #ifdef TCP_SIGNATURE
709 		/* TCP-MD5 (RFC2385). */
710 		if (tp->t_flags & TF_SIGNATURE)
711 			to.to_flags |= TOF_SIGNATURE;
712 #endif /* TCP_SIGNATURE */
713 
714 		/* Processing the options. */
715 		hdrlen += optlen = tcp_addoptions(&to, opt);
716 	}
717 
718 #ifdef INET6
719 	if (isipv6)
720 		ipoptlen = ip6_optlen(tp->t_inpcb);
721 	else
722 #endif
723 	if (tp->t_inpcb->inp_options)
724 		ipoptlen = tp->t_inpcb->inp_options->m_len -
725 				offsetof(struct ipoption, ipopt_list);
726 	else
727 		ipoptlen = 0;
728 #ifdef IPSEC
729 	ipoptlen += ipsec_optlen;
730 #endif
731 
732 	/*
733 	 * Adjust data length if insertion of options will
734 	 * bump the packet length beyond the t_maxopd length.
735 	 * Clear the FIN bit because we cut off the tail of
736 	 * the segment.
737 	 */
738 	if (len + optlen + ipoptlen > tp->t_maxopd) {
739 		flags &= ~TH_FIN;
740 
741 		if (tso) {
742 			KASSERT(ipoptlen == 0,
743 			    ("%s: TSO can't do IP options", __func__));
744 
745 			/*
746 			 * Limit a burst to IP_MAXPACKET minus IP,
747 			 * TCP and options length to keep ip->ip_len
748 			 * from overflowing.
749 			 */
750 			if (len > IP_MAXPACKET - hdrlen) {
751 				len = IP_MAXPACKET - hdrlen;
752 				sendalot = 1;
753 			}
754 
755 			/*
756 			 * Prevent the last segment from being
757 			 * fractional unless the send sockbuf can
758 			 * be emptied.
759 			 */
760 			if (sendalot && off + len < so->so_snd.sb_cc) {
761 				len -= len % (tp->t_maxopd - optlen);
762 				sendalot = 1;
763 			}
764 
765 			/*
766 			 * Send the FIN in a separate segment
767 			 * after the bulk sending is done.
768 			 * We don't trust the TSO implementations
769 			 * to clear the FIN flag on all but the
770 			 * last segment.
771 			 */
772 			if (tp->t_flags & TF_NEEDFIN)
773 				sendalot = 1;
774 
775 		} else {
776 			len = tp->t_maxopd - optlen - ipoptlen;
777 			sendalot = 1;
778 		}
779 	} else
780 		tso = 0;
781 
782 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
783 	    ("%s: len > IP_MAXPACKET", __func__));
784 
785 /*#ifdef DIAGNOSTIC*/
786 #ifdef INET6
787 	if (max_linkhdr + hdrlen > MCLBYTES)
788 #else
789 	if (max_linkhdr + hdrlen > MHLEN)
790 #endif
791 		panic("tcphdr too big");
792 /*#endif*/
793 
794 	/*
795 	 * This KASSERT is here to catch edge cases at a well defined place.
796 	 * Before, those had triggered (random) panic conditions further down.
797 	 */
798 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
799 
800 	/*
801 	 * Grab a header mbuf, attaching a copy of data to
802 	 * be transmitted, and initialize the header from
803 	 * the template for sends on this connection.
804 	 */
805 	if (len) {
806 		struct mbuf *mb;
807 		u_int moff;
808 
809 		if ((tp->t_flags & TF_FORCEDATA) && len == 1)
810 			TCPSTAT_INC(tcps_sndprobe);
811 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
812 			tp->t_sndrexmitpack++;
813 			TCPSTAT_INC(tcps_sndrexmitpack);
814 			TCPSTAT_ADD(tcps_sndrexmitbyte, len);
815 		} else {
816 			TCPSTAT_INC(tcps_sndpack);
817 			TCPSTAT_ADD(tcps_sndbyte, len);
818 		}
819 		MGETHDR(m, M_DONTWAIT, MT_DATA);
820 		if (m == NULL) {
821 			SOCKBUF_UNLOCK(&so->so_snd);
822 			error = ENOBUFS;
823 			goto out;
824 		}
825 #ifdef INET6
826 		if (MHLEN < hdrlen + max_linkhdr) {
827 			MCLGET(m, M_DONTWAIT);
828 			if ((m->m_flags & M_EXT) == 0) {
829 				SOCKBUF_UNLOCK(&so->so_snd);
830 				m_freem(m);
831 				error = ENOBUFS;
832 				goto out;
833 			}
834 		}
835 #endif
836 		m->m_data += max_linkhdr;
837 		m->m_len = hdrlen;
838 
839 		/*
840 		 * Start the m_copy functions from the closest mbuf
841 		 * to the offset in the socket buffer chain.
842 		 */
843 		mb = sbsndptr(&so->so_snd, off, len, &moff);
844 
845 		if (len <= MHLEN - hdrlen - max_linkhdr) {
846 			m_copydata(mb, moff, (int)len,
847 			    mtod(m, caddr_t) + hdrlen);
848 			m->m_len += len;
849 		} else {
850 			m->m_next = m_copy(mb, moff, (int)len);
851 			if (m->m_next == NULL) {
852 				SOCKBUF_UNLOCK(&so->so_snd);
853 				(void) m_free(m);
854 				error = ENOBUFS;
855 				goto out;
856 			}
857 		}
858 
859 		/*
860 		 * If we're sending everything we've got, set PUSH.
861 		 * (This will keep happy those implementations which only
862 		 * give data to the user when a buffer fills or
863 		 * a PUSH comes in.)
864 		 */
865 		if (off + len == so->so_snd.sb_cc)
866 			flags |= TH_PUSH;
867 		SOCKBUF_UNLOCK(&so->so_snd);
868 	} else {
869 		SOCKBUF_UNLOCK(&so->so_snd);
870 		if (tp->t_flags & TF_ACKNOW)
871 			TCPSTAT_INC(tcps_sndacks);
872 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
873 			TCPSTAT_INC(tcps_sndctrl);
874 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
875 			TCPSTAT_INC(tcps_sndurg);
876 		else
877 			TCPSTAT_INC(tcps_sndwinup);
878 
879 		MGETHDR(m, M_DONTWAIT, MT_DATA);
880 		if (m == NULL) {
881 			error = ENOBUFS;
882 			goto out;
883 		}
884 #ifdef INET6
885 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
886 		    MHLEN >= hdrlen) {
887 			MH_ALIGN(m, hdrlen);
888 		} else
889 #endif
890 		m->m_data += max_linkhdr;
891 		m->m_len = hdrlen;
892 	}
893 	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
894 	m->m_pkthdr.rcvif = (struct ifnet *)0;
895 #ifdef MAC
896 	mac_inpcb_create_mbuf(tp->t_inpcb, m);
897 #endif
898 #ifdef INET6
899 	if (isipv6) {
900 		ip6 = mtod(m, struct ip6_hdr *);
901 		th = (struct tcphdr *)(ip6 + 1);
902 		tcpip_fillheaders(tp->t_inpcb, ip6, th);
903 	} else
904 #endif /* INET6 */
905 	{
906 		ip = mtod(m, struct ip *);
907 		ipov = (struct ipovly *)ip;
908 		th = (struct tcphdr *)(ip + 1);
909 		tcpip_fillheaders(tp->t_inpcb, ip, th);
910 	}
911 
912 	/*
913 	 * Fill in fields, remembering maximum advertised
914 	 * window for use in delaying messages about window sizes.
915 	 * If resending a FIN, be sure not to use a new sequence number.
916 	 */
917 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
918 	    tp->snd_nxt == tp->snd_max)
919 		tp->snd_nxt--;
920 	/*
921 	 * If we are starting a connection, send ECN setup
922 	 * SYN packet. If we are on a retransmit, we may
923 	 * resend those bits a number of times as per
924 	 * RFC 3168.
925 	 */
926 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
927 		if (tp->t_rxtshift >= 1) {
928 			if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
929 				flags |= TH_ECE|TH_CWR;
930 		} else
931 			flags |= TH_ECE|TH_CWR;
932 	}
933 
934 	if (tp->t_state == TCPS_ESTABLISHED &&
935 	    (tp->t_flags & TF_ECN_PERMIT)) {
936 		/*
937 		 * If the peer has ECN, mark data packets with
938 		 * ECN capable transmission (ECT).
939 		 * Ignore pure ack packets, retransmissions and window probes.
940 		 */
941 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
942 		    !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
943 #ifdef INET6
944 			if (isipv6)
945 				ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
946 			else
947 #endif
948 				ip->ip_tos |= IPTOS_ECN_ECT0;
949 			TCPSTAT_INC(tcps_ecn_ect0);
950 		}
951 
952 		/*
953 		 * Reply with proper ECN notifications.
954 		 */
955 		if (tp->t_flags & TF_ECN_SND_CWR) {
956 			flags |= TH_CWR;
957 			tp->t_flags &= ~TF_ECN_SND_CWR;
958 		}
959 		if (tp->t_flags & TF_ECN_SND_ECE)
960 			flags |= TH_ECE;
961 	}
962 
963 	/*
964 	 * If we are doing retransmissions, then snd_nxt will
965 	 * not reflect the first unsent octet.  For ACK only
966 	 * packets, we do not want the sequence number of the
967 	 * retransmitted packet, we want the sequence number
968 	 * of the next unsent octet.  So, if there is no data
969 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
970 	 * when filling in ti_seq.  But if we are in persist
971 	 * state, snd_max might reflect one byte beyond the
972 	 * right edge of the window, so use snd_nxt in that
973 	 * case, since we know we aren't doing a retransmission.
974 	 * (retransmit and persist are mutually exclusive...)
975 	 */
976 	if (sack_rxmit == 0) {
977 		if (len || (flags & (TH_SYN|TH_FIN)) ||
978 		    tcp_timer_active(tp, TT_PERSIST))
979 			th->th_seq = htonl(tp->snd_nxt);
980 		else
981 			th->th_seq = htonl(tp->snd_max);
982 	} else {
983 		th->th_seq = htonl(p->rxmit);
984 		p->rxmit += len;
985 		tp->sackhint.sack_bytes_rexmit += len;
986 	}
987 	th->th_ack = htonl(tp->rcv_nxt);
988 	if (optlen) {
989 		bcopy(opt, th + 1, optlen);
990 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
991 	}
992 	th->th_flags = flags;
993 	/*
994 	 * Calculate receive window.  Don't shrink window,
995 	 * but avoid silly window syndrome.
996 	 */
997 	if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
998 	    recwin < (long)tp->t_maxseg)
999 		recwin = 0;
1000 	if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1001 	    recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
1002 		recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
1003 	if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
1004 		recwin = (long)TCP_MAXWIN << tp->rcv_scale;
1005 
1006 	/*
1007 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1008 	 * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
1009 	 * case is handled in syncache.
1010 	 */
1011 	if (flags & TH_SYN)
1012 		th->th_win = htons((u_short)
1013 				(min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1014 	else
1015 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1016 
1017 	/*
1018 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1019 	 * a 0 window.  This may cause the remote transmitter to stall.  This
1020 	 * flag tells soreceive() to disable delayed acknowledgements when
1021 	 * draining the buffer.  This can occur if the receiver is attempting
1022 	 * to read more data than can be buffered prior to transmitting on
1023 	 * the connection.
1024 	 */
1025 	if (th->th_win == 0) {
1026 		tp->t_sndzerowin++;
1027 		tp->t_flags |= TF_RXWIN0SENT;
1028 	} else
1029 		tp->t_flags &= ~TF_RXWIN0SENT;
1030 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1031 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1032 		th->th_flags |= TH_URG;
1033 	} else
1034 		/*
1035 		 * If no urgent pointer to send, then we pull
1036 		 * the urgent pointer to the left edge of the send window
1037 		 * so that it doesn't drift into the send window on sequence
1038 		 * number wraparound.
1039 		 */
1040 		tp->snd_up = tp->snd_una;		/* drag it along */
1041 
1042 #ifdef TCP_SIGNATURE
1043 	if (tp->t_flags & TF_SIGNATURE) {
1044 		int sigoff = to.to_signature - opt;
1045 		tcp_signature_compute(m, 0, len, optlen,
1046 		    (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
1047 	}
1048 #endif
1049 
1050 	/*
1051 	 * Put TCP length in extended header, and then
1052 	 * checksum extended header and data.
1053 	 */
1054 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1055 #ifdef INET6
1056 	if (isipv6)
1057 		/*
1058 		 * ip6_plen is not need to be filled now, and will be filled
1059 		 * in ip6_output.
1060 		 */
1061 		th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
1062 				       sizeof(struct tcphdr) + optlen + len);
1063 	else
1064 #endif /* INET6 */
1065 	{
1066 		m->m_pkthdr.csum_flags = CSUM_TCP;
1067 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1068 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1069 		    htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1070 
1071 		/* IP version must be set here for ipv4/ipv6 checking later */
1072 		KASSERT(ip->ip_v == IPVERSION,
1073 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1074 	}
1075 
1076 	/*
1077 	 * Enable TSO and specify the size of the segments.
1078 	 * The TCP pseudo header checksum is always provided.
1079 	 * XXX: Fixme: This is currently not the case for IPv6.
1080 	 */
1081 	if (tso) {
1082 		KASSERT(len > tp->t_maxopd - optlen,
1083 		    ("%s: len <= tso_segsz", __func__));
1084 		m->m_pkthdr.csum_flags |= CSUM_TSO;
1085 		m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen;
1086 	}
1087 
1088 #ifdef IPSEC
1089 	KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL),
1090 	    ("%s: mbuf chain shorter than expected: %ld + %u + %u - %u != %u",
1091 	    __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL)));
1092 #else
1093 	KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL),
1094 	    ("%s: mbuf chain shorter than expected: %ld + %u + %u != %u",
1095 	    __func__, len, hdrlen, ipoptlen, m_length(m, NULL)));
1096 #endif
1097 
1098 	/*
1099 	 * In transmit state, time the transmission and arrange for
1100 	 * the retransmit.  In persist state, just set snd_max.
1101 	 */
1102 	if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1103 	    !tcp_timer_active(tp, TT_PERSIST)) {
1104 		tcp_seq startseq = tp->snd_nxt;
1105 
1106 		/*
1107 		 * Advance snd_nxt over sequence space of this segment.
1108 		 */
1109 		if (flags & (TH_SYN|TH_FIN)) {
1110 			if (flags & TH_SYN)
1111 				tp->snd_nxt++;
1112 			if (flags & TH_FIN) {
1113 				tp->snd_nxt++;
1114 				tp->t_flags |= TF_SENTFIN;
1115 			}
1116 		}
1117 		if (sack_rxmit)
1118 			goto timer;
1119 		tp->snd_nxt += len;
1120 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1121 			tp->snd_max = tp->snd_nxt;
1122 			/*
1123 			 * Time this transmission if not a retransmission and
1124 			 * not currently timing anything.
1125 			 */
1126 			if (tp->t_rtttime == 0) {
1127 				tp->t_rtttime = ticks;
1128 				tp->t_rtseq = startseq;
1129 				TCPSTAT_INC(tcps_segstimed);
1130 			}
1131 		}
1132 
1133 		/*
1134 		 * Set retransmit timer if not currently set,
1135 		 * and not doing a pure ack or a keep-alive probe.
1136 		 * Initial value for retransmit timer is smoothed
1137 		 * round-trip time + 2 * round-trip time variance.
1138 		 * Initialize shift counter which is used for backoff
1139 		 * of retransmit time.
1140 		 */
1141 timer:
1142 		if (!tcp_timer_active(tp, TT_REXMT) &&
1143 		    ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1144 		     (tp->snd_nxt != tp->snd_una))) {
1145 			if (tcp_timer_active(tp, TT_PERSIST)) {
1146 				tcp_timer_activate(tp, TT_PERSIST, 0);
1147 				tp->t_rxtshift = 0;
1148 			}
1149 			tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1150 		}
1151 	} else {
1152 		/*
1153 		 * Persist case, update snd_max but since we are in
1154 		 * persist mode (no window) we do not update snd_nxt.
1155 		 */
1156 		int xlen = len;
1157 		if (flags & TH_SYN)
1158 			++xlen;
1159 		if (flags & TH_FIN) {
1160 			++xlen;
1161 			tp->t_flags |= TF_SENTFIN;
1162 		}
1163 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1164 			tp->snd_max = tp->snd_nxt + len;
1165 	}
1166 
1167 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1168 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
1169 
1170 #ifdef TCPDEBUG
1171 	/*
1172 	 * Trace.
1173 	 */
1174 	if (so->so_options & SO_DEBUG) {
1175 		u_short save = 0;
1176 #ifdef INET6
1177 		if (!isipv6)
1178 #endif
1179 		{
1180 			save = ipov->ih_len;
1181 			ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1182 		}
1183 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1184 #ifdef INET6
1185 		if (!isipv6)
1186 #endif
1187 		ipov->ih_len = save;
1188 	}
1189 #endif /* TCPDEBUG */
1190 
1191 	/*
1192 	 * Fill in IP length and desired time to live and
1193 	 * send to IP level.  There should be a better way
1194 	 * to handle ttl and tos; we could keep them in
1195 	 * the template, but need a way to checksum without them.
1196 	 */
1197 	/*
1198 	 * m->m_pkthdr.len should have been set before cksum calcuration,
1199 	 * because in6_cksum() need it.
1200 	 */
1201 #ifdef INET6
1202 	if (isipv6) {
1203 		/*
1204 		 * we separately set hoplimit for every segment, since the
1205 		 * user might want to change the value via setsockopt.
1206 		 * Also, desired default hop limit might be changed via
1207 		 * Neighbor Discovery.
1208 		 */
1209 		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1210 
1211 		/* TODO: IPv6 IP6TOS_ECT bit on */
1212 		error = ip6_output(m,
1213 			    tp->t_inpcb->in6p_outputopts, NULL,
1214 			    ((so->so_options & SO_DONTROUTE) ?
1215 			    IP_ROUTETOIF : 0), NULL, NULL, tp->t_inpcb);
1216 	}
1217 #endif /* INET6 */
1218 #if defined(INET) && defined(INET6)
1219 	else
1220 #endif
1221 #ifdef INET
1222     {
1223 	ip->ip_len = m->m_pkthdr.len;
1224 #ifdef INET6
1225 	if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1226 		ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1227 #endif /* INET6 */
1228 	/*
1229 	 * If we do path MTU discovery, then we set DF on every packet.
1230 	 * This might not be the best thing to do according to RFC3390
1231 	 * Section 2. However the tcp hostcache migitates the problem
1232 	 * so it affects only the first tcp connection with a host.
1233 	 *
1234 	 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1235 	 */
1236 	if (V_path_mtu_discovery && tp->t_maxopd > V_tcp_minmss)
1237 		ip->ip_off |= IP_DF;
1238 
1239 	error = ip_output(m, tp->t_inpcb->inp_options, NULL,
1240 	    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1241 	    tp->t_inpcb);
1242     }
1243 #endif /* INET */
1244 	if (error) {
1245 
1246 		/*
1247 		 * We know that the packet was lost, so back out the
1248 		 * sequence number advance, if any.
1249 		 *
1250 		 * If the error is EPERM the packet got blocked by the
1251 		 * local firewall.  Normally we should terminate the
1252 		 * connection but the blocking may have been spurious
1253 		 * due to a firewall reconfiguration cycle.  So we treat
1254 		 * it like a packet loss and let the retransmit timer and
1255 		 * timeouts do their work over time.
1256 		 * XXX: It is a POLA question whether calling tcp_drop right
1257 		 * away would be the really correct behavior instead.
1258 		 */
1259 		if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1260 		    !tcp_timer_active(tp, TT_PERSIST)) &&
1261 		    ((flags & TH_SYN) == 0) &&
1262 		    (error != EPERM)) {
1263 			if (sack_rxmit) {
1264 				p->rxmit -= len;
1265 				tp->sackhint.sack_bytes_rexmit -= len;
1266 				KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1267 				    ("sackhint bytes rtx >= 0"));
1268 			} else
1269 				tp->snd_nxt -= len;
1270 		}
1271 out:
1272 		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1273 		switch (error) {
1274 		case EPERM:
1275 			tp->t_softerror = error;
1276 			return (error);
1277 		case ENOBUFS:
1278 	                if (!tcp_timer_active(tp, TT_REXMT) &&
1279 			    !tcp_timer_active(tp, TT_PERSIST))
1280 	                        tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1281 			tp->snd_cwnd = tp->t_maxseg;
1282 			return (0);
1283 		case EMSGSIZE:
1284 			/*
1285 			 * For some reason the interface we used initially
1286 			 * to send segments changed to another or lowered
1287 			 * its MTU.
1288 			 *
1289 			 * tcp_mtudisc() will find out the new MTU and as
1290 			 * its last action, initiate retransmission, so it
1291 			 * is important to not do so here.
1292 			 *
1293 			 * If TSO was active we either got an interface
1294 			 * without TSO capabilits or TSO was turned off.
1295 			 * Disable it for this connection as too and
1296 			 * immediatly retry with MSS sized segments generated
1297 			 * by this function.
1298 			 */
1299 			if (tso)
1300 				tp->t_flags &= ~TF_TSO;
1301 			tcp_mtudisc(tp->t_inpcb, 0);
1302 			return (0);
1303 		case EHOSTDOWN:
1304 		case EHOSTUNREACH:
1305 		case ENETDOWN:
1306 		case ENETUNREACH:
1307 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
1308 				tp->t_softerror = error;
1309 				return (0);
1310 			}
1311 			/* FALLTHROUGH */
1312 		default:
1313 			return (error);
1314 		}
1315 	}
1316 	TCPSTAT_INC(tcps_sndtotal);
1317 
1318 	/*
1319 	 * Data sent (as far as we can tell).
1320 	 * If this advertises a larger window than any other segment,
1321 	 * then remember the size of the advertised window.
1322 	 * Any pending ACK has now been sent.
1323 	 */
1324 	if (recwin >= 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1325 		tp->rcv_adv = tp->rcv_nxt + recwin;
1326 	tp->last_ack_sent = tp->rcv_nxt;
1327 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1328 	if (tcp_timer_active(tp, TT_DELACK))
1329 		tcp_timer_activate(tp, TT_DELACK, 0);
1330 #if 0
1331 	/*
1332 	 * This completely breaks TCP if newreno is turned on.  What happens
1333 	 * is that if delayed-acks are turned on on the receiver, this code
1334 	 * on the transmitter effectively destroys the TCP window, forcing
1335 	 * it to four packets (1.5Kx4 = 6K window).
1336 	 */
1337 	if (sendalot && --maxburst)
1338 		goto again;
1339 #endif
1340 	if (sendalot)
1341 		goto again;
1342 	return (0);
1343 }
1344 
1345 void
1346 tcp_setpersist(struct tcpcb *tp)
1347 {
1348 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1349 	int tt;
1350 
1351 	tp->t_flags &= ~TF_PREVVALID;
1352 	if (tcp_timer_active(tp, TT_REXMT))
1353 		panic("tcp_setpersist: retransmit pending");
1354 	/*
1355 	 * Start/restart persistance timer.
1356 	 */
1357 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1358 		      TCPTV_PERSMIN, TCPTV_PERSMAX);
1359 	tcp_timer_activate(tp, TT_PERSIST, tt);
1360 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1361 		tp->t_rxtshift++;
1362 }
1363 
1364 /*
1365  * Insert TCP options according to the supplied parameters to the place
1366  * optp in a consistent way.  Can handle unaligned destinations.
1367  *
1368  * The order of the option processing is crucial for optimal packing and
1369  * alignment for the scarce option space.
1370  *
1371  * The optimal order for a SYN/SYN-ACK segment is:
1372  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1373  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1374  *
1375  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1376  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1377  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1378  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1379  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1380  */
1381 int
1382 tcp_addoptions(struct tcpopt *to, u_char *optp)
1383 {
1384 	u_int mask, optlen = 0;
1385 
1386 	for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1387 		if ((to->to_flags & mask) != mask)
1388 			continue;
1389 		if (optlen == TCP_MAXOLEN)
1390 			break;
1391 		switch (to->to_flags & mask) {
1392 		case TOF_MSS:
1393 			while (optlen % 4) {
1394 				optlen += TCPOLEN_NOP;
1395 				*optp++ = TCPOPT_NOP;
1396 			}
1397 			if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1398 				continue;
1399 			optlen += TCPOLEN_MAXSEG;
1400 			*optp++ = TCPOPT_MAXSEG;
1401 			*optp++ = TCPOLEN_MAXSEG;
1402 			to->to_mss = htons(to->to_mss);
1403 			bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1404 			optp += sizeof(to->to_mss);
1405 			break;
1406 		case TOF_SCALE:
1407 			while (!optlen || optlen % 2 != 1) {
1408 				optlen += TCPOLEN_NOP;
1409 				*optp++ = TCPOPT_NOP;
1410 			}
1411 			if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1412 				continue;
1413 			optlen += TCPOLEN_WINDOW;
1414 			*optp++ = TCPOPT_WINDOW;
1415 			*optp++ = TCPOLEN_WINDOW;
1416 			*optp++ = to->to_wscale;
1417 			break;
1418 		case TOF_SACKPERM:
1419 			while (optlen % 2) {
1420 				optlen += TCPOLEN_NOP;
1421 				*optp++ = TCPOPT_NOP;
1422 			}
1423 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1424 				continue;
1425 			optlen += TCPOLEN_SACK_PERMITTED;
1426 			*optp++ = TCPOPT_SACK_PERMITTED;
1427 			*optp++ = TCPOLEN_SACK_PERMITTED;
1428 			break;
1429 		case TOF_TS:
1430 			while (!optlen || optlen % 4 != 2) {
1431 				optlen += TCPOLEN_NOP;
1432 				*optp++ = TCPOPT_NOP;
1433 			}
1434 			if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1435 				continue;
1436 			optlen += TCPOLEN_TIMESTAMP;
1437 			*optp++ = TCPOPT_TIMESTAMP;
1438 			*optp++ = TCPOLEN_TIMESTAMP;
1439 			to->to_tsval = htonl(to->to_tsval);
1440 			to->to_tsecr = htonl(to->to_tsecr);
1441 			bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1442 			optp += sizeof(to->to_tsval);
1443 			bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1444 			optp += sizeof(to->to_tsecr);
1445 			break;
1446 		case TOF_SIGNATURE:
1447 			{
1448 			int siglen = TCPOLEN_SIGNATURE - 2;
1449 
1450 			while (!optlen || optlen % 4 != 2) {
1451 				optlen += TCPOLEN_NOP;
1452 				*optp++ = TCPOPT_NOP;
1453 			}
1454 			if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE)
1455 				continue;
1456 			optlen += TCPOLEN_SIGNATURE;
1457 			*optp++ = TCPOPT_SIGNATURE;
1458 			*optp++ = TCPOLEN_SIGNATURE;
1459 			to->to_signature = optp;
1460 			while (siglen--)
1461 				 *optp++ = 0;
1462 			break;
1463 			}
1464 		case TOF_SACK:
1465 			{
1466 			int sackblks = 0;
1467 			struct sackblk *sack = (struct sackblk *)to->to_sacks;
1468 			tcp_seq sack_seq;
1469 
1470 			while (!optlen || optlen % 4 != 2) {
1471 				optlen += TCPOLEN_NOP;
1472 				*optp++ = TCPOPT_NOP;
1473 			}
1474 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1475 				continue;
1476 			optlen += TCPOLEN_SACKHDR;
1477 			*optp++ = TCPOPT_SACK;
1478 			sackblks = min(to->to_nsacks,
1479 					(TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1480 			*optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1481 			while (sackblks--) {
1482 				sack_seq = htonl(sack->start);
1483 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1484 				optp += sizeof(sack_seq);
1485 				sack_seq = htonl(sack->end);
1486 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1487 				optp += sizeof(sack_seq);
1488 				optlen += TCPOLEN_SACK;
1489 				sack++;
1490 			}
1491 			TCPSTAT_INC(tcps_sack_send_blocks);
1492 			break;
1493 			}
1494 		default:
1495 			panic("%s: unknown TCP option type", __func__);
1496 			break;
1497 		}
1498 	}
1499 
1500 	/* Terminate and pad TCP options to a 4 byte boundary. */
1501 	if (optlen % 4) {
1502 		optlen += TCPOLEN_EOL;
1503 		*optp++ = TCPOPT_EOL;
1504 	}
1505 	/*
1506 	 * According to RFC 793 (STD0007):
1507 	 *   "The content of the header beyond the End-of-Option option
1508 	 *    must be header padding (i.e., zero)."
1509 	 *   and later: "The padding is composed of zeros."
1510 	 */
1511 	while (optlen % 4) {
1512 		optlen += TCPOLEN_PAD;
1513 		*optp++ = TCPOPT_PAD;
1514 	}
1515 
1516 	KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1517 	return (optlen);
1518 }
1519