xref: /freebsd/sys/netinet/tcp_output.c (revision 7d99ab9fd0cc2c1ce2ecef0ed6d0672c2a50b0cb)
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 #ifdef TCP_OFFLOAD
79 #include <netinet/tcp_offload.h>
80 #endif
81 
82 #ifdef IPSEC
83 #include <netipsec/ipsec.h>
84 #endif /*IPSEC*/
85 
86 #include <machine/in_cksum.h>
87 
88 #include <security/mac/mac_framework.h>
89 
90 VNET_DEFINE(int, path_mtu_discovery) = 1;
91 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
92 	&VNET_NAME(path_mtu_discovery), 1,
93 	"Enable Path MTU Discovery");
94 
95 VNET_DEFINE(int, tcp_do_tso) = 1;
96 #define	V_tcp_do_tso		VNET(tcp_do_tso)
97 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW,
98 	&VNET_NAME(tcp_do_tso), 0,
99 	"Enable TCP Segmentation Offload");
100 
101 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
102 #define	V_tcp_sendspace	VNET(tcp_sendspace)
103 SYSCTL_VNET_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_RW,
104 	&VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
105 
106 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
107 #define	V_tcp_do_autosndbuf	VNET(tcp_do_autosndbuf)
108 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW,
109 	&VNET_NAME(tcp_do_autosndbuf), 0,
110 	"Enable automatic send buffer sizing");
111 
112 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
113 #define	V_tcp_autosndbuf_inc	VNET(tcp_autosndbuf_inc)
114 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW,
115 	&VNET_NAME(tcp_autosndbuf_inc), 0,
116 	"Incrementor step size of automatic send buffer");
117 
118 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
119 #define	V_tcp_autosndbuf_max	VNET(tcp_autosndbuf_max)
120 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW,
121 	&VNET_NAME(tcp_autosndbuf_max), 0,
122 	"Max size of automatic send buffer");
123 
124 static void inline	hhook_run_tcp_est_out(struct tcpcb *tp,
125 			    struct tcphdr *th, struct tcpopt *to,
126 			    long len, int tso);
127 static void inline	cc_after_idle(struct tcpcb *tp);
128 
129 /*
130  * Wrapper for the TCP established output helper hook.
131  */
132 static void inline
133 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
134     struct tcpopt *to, long len, int tso)
135 {
136 	struct tcp_hhook_data hhook_data;
137 
138 	if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
139 		hhook_data.tp = tp;
140 		hhook_data.th = th;
141 		hhook_data.to = to;
142 		hhook_data.len = len;
143 		hhook_data.tso = tso;
144 
145 		hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
146 		    tp->osd);
147 	}
148 }
149 
150 /*
151  * CC wrapper hook functions
152  */
153 static void inline
154 cc_after_idle(struct tcpcb *tp)
155 {
156 	INP_WLOCK_ASSERT(tp->t_inpcb);
157 
158 	if (CC_ALGO(tp)->after_idle != NULL)
159 		CC_ALGO(tp)->after_idle(tp->ccv);
160 }
161 
162 /*
163  * Tcp output routine: figure out what should be sent and send it.
164  */
165 int
166 tcp_output(struct tcpcb *tp)
167 {
168 	struct socket *so = tp->t_inpcb->inp_socket;
169 	long len, recwin, sendwin;
170 	int off, flags, error = 0;	/* Keep compiler happy */
171 	struct mbuf *m;
172 	struct ip *ip = NULL;
173 	struct ipovly *ipov = NULL;
174 	struct tcphdr *th;
175 	u_char opt[TCP_MAXOLEN];
176 	unsigned ipoptlen, optlen, hdrlen;
177 #ifdef IPSEC
178 	unsigned ipsec_optlen = 0;
179 #endif
180 	int idle, sendalot;
181 	int sack_rxmit, sack_bytes_rxmt;
182 	struct sackhole *p;
183 	int tso, mtu;
184 	struct tcpopt to;
185 #if 0
186 	int maxburst = TCP_MAXBURST;
187 #endif
188 #ifdef INET6
189 	struct ip6_hdr *ip6 = NULL;
190 	int isipv6;
191 
192 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
193 #endif
194 
195 	INP_WLOCK_ASSERT(tp->t_inpcb);
196 
197 #ifdef TCP_OFFLOAD
198 	if (tp->t_flags & TF_TOE)
199 		return (tcp_offload_output(tp));
200 #endif
201 
202 	/*
203 	 * Determine length of data that should be transmitted,
204 	 * and flags that will be used.
205 	 * If there is some data or critical controls (SYN, RST)
206 	 * to send, then transmit; otherwise, investigate further.
207 	 */
208 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
209 	if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur)
210 		cc_after_idle(tp);
211 	tp->t_flags &= ~TF_LASTIDLE;
212 	if (idle) {
213 		if (tp->t_flags & TF_MORETOCOME) {
214 			tp->t_flags |= TF_LASTIDLE;
215 			idle = 0;
216 		}
217 	}
218 again:
219 	/*
220 	 * If we've recently taken a timeout, snd_max will be greater than
221 	 * snd_nxt.  There may be SACK information that allows us to avoid
222 	 * resending already delivered data.  Adjust snd_nxt accordingly.
223 	 */
224 	if ((tp->t_flags & TF_SACK_PERMIT) &&
225 	    SEQ_LT(tp->snd_nxt, tp->snd_max))
226 		tcp_sack_adjust(tp);
227 	sendalot = 0;
228 	tso = 0;
229 	mtu = 0;
230 	off = tp->snd_nxt - tp->snd_una;
231 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
232 
233 	flags = tcp_outflags[tp->t_state];
234 	/*
235 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
236 	 * to send out new data (when sendalot is 1), bypass this function.
237 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
238 	 * we're replacing a (future) new transmission with a retransmission
239 	 * now, and we previously incremented snd_cwnd in tcp_input().
240 	 */
241 	/*
242 	 * Still in sack recovery , reset rxmit flag to zero.
243 	 */
244 	sack_rxmit = 0;
245 	sack_bytes_rxmt = 0;
246 	len = 0;
247 	p = NULL;
248 	if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
249 	    (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
250 		long cwin;
251 
252 		cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
253 		if (cwin < 0)
254 			cwin = 0;
255 		/* Do not retransmit SACK segments beyond snd_recover */
256 		if (SEQ_GT(p->end, tp->snd_recover)) {
257 			/*
258 			 * (At least) part of sack hole extends beyond
259 			 * snd_recover. Check to see if we can rexmit data
260 			 * for this hole.
261 			 */
262 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
263 				/*
264 				 * Can't rexmit any more data for this hole.
265 				 * That data will be rexmitted in the next
266 				 * sack recovery episode, when snd_recover
267 				 * moves past p->rxmit.
268 				 */
269 				p = NULL;
270 				goto after_sack_rexmit;
271 			} else
272 				/* Can rexmit part of the current hole */
273 				len = ((long)ulmin(cwin,
274 						   tp->snd_recover - p->rxmit));
275 		} else
276 			len = ((long)ulmin(cwin, p->end - p->rxmit));
277 		off = p->rxmit - tp->snd_una;
278 		KASSERT(off >= 0,("%s: sack block to the left of una : %d",
279 		    __func__, off));
280 		if (len > 0) {
281 			sack_rxmit = 1;
282 			sendalot = 1;
283 			TCPSTAT_INC(tcps_sack_rexmits);
284 			TCPSTAT_ADD(tcps_sack_rexmit_bytes,
285 			    min(len, tp->t_maxseg));
286 		}
287 	}
288 after_sack_rexmit:
289 	/*
290 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
291 	 * state flags.
292 	 */
293 	if (tp->t_flags & TF_NEEDFIN)
294 		flags |= TH_FIN;
295 	if (tp->t_flags & TF_NEEDSYN)
296 		flags |= TH_SYN;
297 
298 	SOCKBUF_LOCK(&so->so_snd);
299 	/*
300 	 * If in persist timeout with window of 0, send 1 byte.
301 	 * Otherwise, if window is small but nonzero
302 	 * and timer expired, we will send what we can
303 	 * and go to transmit state.
304 	 */
305 	if (tp->t_flags & TF_FORCEDATA) {
306 		if (sendwin == 0) {
307 			/*
308 			 * If we still have some data to send, then
309 			 * clear the FIN bit.  Usually this would
310 			 * happen below when it realizes that we
311 			 * aren't sending all the data.  However,
312 			 * if we have exactly 1 byte of unsent data,
313 			 * then it won't clear the FIN bit below,
314 			 * and if we are in persist state, we wind
315 			 * up sending the packet without recording
316 			 * that we sent the FIN bit.
317 			 *
318 			 * We can't just blindly clear the FIN bit,
319 			 * because if we don't have any more data
320 			 * to send then the probe will be the FIN
321 			 * itself.
322 			 */
323 			if (off < so->so_snd.sb_cc)
324 				flags &= ~TH_FIN;
325 			sendwin = 1;
326 		} else {
327 			tcp_timer_activate(tp, TT_PERSIST, 0);
328 			tp->t_rxtshift = 0;
329 		}
330 	}
331 
332 	/*
333 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
334 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
335 	 * a negative length.  This can also occur when TCP opens up
336 	 * its congestion window while receiving additional duplicate
337 	 * acks after fast-retransmit because TCP will reset snd_nxt
338 	 * to snd_max after the fast-retransmit.
339 	 *
340 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
341 	 * be set to snd_una, the offset will be 0, and the length may
342 	 * wind up 0.
343 	 *
344 	 * If sack_rxmit is true we are retransmitting from the scoreboard
345 	 * in which case len is already set.
346 	 */
347 	if (sack_rxmit == 0) {
348 		if (sack_bytes_rxmt == 0)
349 			len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off);
350 		else {
351 			long cwin;
352 
353                         /*
354 			 * We are inside of a SACK recovery episode and are
355 			 * sending new data, having retransmitted all the
356 			 * data possible in the scoreboard.
357 			 */
358 			len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd)
359 			       - off);
360 			/*
361 			 * Don't remove this (len > 0) check !
362 			 * We explicitly check for len > 0 here (although it
363 			 * isn't really necessary), to work around a gcc
364 			 * optimization issue - to force gcc to compute
365 			 * len above. Without this check, the computation
366 			 * of len is bungled by the optimizer.
367 			 */
368 			if (len > 0) {
369 				cwin = tp->snd_cwnd -
370 					(tp->snd_nxt - tp->sack_newdata) -
371 					sack_bytes_rxmt;
372 				if (cwin < 0)
373 					cwin = 0;
374 				len = lmin(len, cwin);
375 			}
376 		}
377 	}
378 
379 	/*
380 	 * Lop off SYN bit if it has already been sent.  However, if this
381 	 * is SYN-SENT state and if segment contains data and if we don't
382 	 * know that foreign host supports TAO, suppress sending segment.
383 	 */
384 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
385 		if (tp->t_state != TCPS_SYN_RECEIVED)
386 			flags &= ~TH_SYN;
387 		off--, len++;
388 	}
389 
390 	/*
391 	 * Be careful not to send data and/or FIN on SYN segments.
392 	 * This measure is needed to prevent interoperability problems
393 	 * with not fully conformant TCP implementations.
394 	 */
395 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
396 		len = 0;
397 		flags &= ~TH_FIN;
398 	}
399 
400 	if (len < 0) {
401 		/*
402 		 * If FIN has been sent but not acked,
403 		 * but we haven't been called to retransmit,
404 		 * len will be < 0.  Otherwise, window shrank
405 		 * after we sent into it.  If window shrank to 0,
406 		 * cancel pending retransmit, pull snd_nxt back
407 		 * to (closed) window, and set the persist timer
408 		 * if it isn't already going.  If the window didn't
409 		 * close completely, just wait for an ACK.
410 		 */
411 		len = 0;
412 		if (sendwin == 0) {
413 			tcp_timer_activate(tp, TT_REXMT, 0);
414 			tp->t_rxtshift = 0;
415 			tp->snd_nxt = tp->snd_una;
416 			if (!tcp_timer_active(tp, TT_PERSIST))
417 				tcp_setpersist(tp);
418 		}
419 	}
420 
421 	/* len will be >= 0 after this point. */
422 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
423 
424 	/*
425 	 * Automatic sizing of send socket buffer.  Often the send buffer
426 	 * size is not optimally adjusted to the actual network conditions
427 	 * at hand (delay bandwidth product).  Setting the buffer size too
428 	 * small limits throughput on links with high bandwidth and high
429 	 * delay (eg. trans-continental/oceanic links).  Setting the
430 	 * buffer size too big consumes too much real kernel memory,
431 	 * especially with many connections on busy servers.
432 	 *
433 	 * The criteria to step up the send buffer one notch are:
434 	 *  1. receive window of remote host is larger than send buffer
435 	 *     (with a fudge factor of 5/4th);
436 	 *  2. send buffer is filled to 7/8th with data (so we actually
437 	 *     have data to make use of it);
438 	 *  3. send buffer fill has not hit maximal automatic size;
439 	 *  4. our send window (slow start and cogestion controlled) is
440 	 *     larger than sent but unacknowledged data in send buffer.
441 	 *
442 	 * The remote host receive window scaling factor may limit the
443 	 * growing of the send buffer before it reaches its allowed
444 	 * maximum.
445 	 *
446 	 * It scales directly with slow start or congestion window
447 	 * and does at most one step per received ACK.  This fast
448 	 * scaling has the drawback of growing the send buffer beyond
449 	 * what is strictly necessary to make full use of a given
450 	 * delay*bandwith product.  However testing has shown this not
451 	 * to be much of an problem.  At worst we are trading wasting
452 	 * of available bandwith (the non-use of it) for wasting some
453 	 * socket buffer memory.
454 	 *
455 	 * TODO: Shrink send buffer during idle periods together
456 	 * with congestion window.  Requires another timer.  Has to
457 	 * wait for upcoming tcp timer rewrite.
458 	 */
459 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
460 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
461 		    so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
462 		    so->so_snd.sb_cc < V_tcp_autosndbuf_max &&
463 		    sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
464 			if (!sbreserve_locked(&so->so_snd,
465 			    min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
466 			     V_tcp_autosndbuf_max), so, curthread))
467 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
468 		}
469 	}
470 
471 	/*
472 	 * Decide if we can use TCP Segmentation Offloading (if supported by
473 	 * hardware).
474 	 *
475 	 * TSO may only be used if we are in a pure bulk sending state.  The
476 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
477 	 * IP options prevent using TSO.  With TSO the TCP header is the same
478 	 * (except for the sequence number) for all generated packets.  This
479 	 * makes it impossible to transmit any options which vary per generated
480 	 * segment or packet.
481 	 */
482 #ifdef IPSEC
483 	/*
484 	 * Pre-calculate here as we save another lookup into the darknesses
485 	 * of IPsec that way and can actually decide if TSO is ok.
486 	 */
487 	ipsec_optlen = ipsec_hdrsiz_tcp(tp);
488 #endif
489 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
490 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
491 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
492 #ifdef IPSEC
493 	    ipsec_optlen == 0 &&
494 #endif
495 	    tp->t_inpcb->inp_options == NULL &&
496 	    tp->t_inpcb->in6p_options == NULL)
497 		tso = 1;
498 
499 	if (sack_rxmit) {
500 		if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
501 			flags &= ~TH_FIN;
502 	} else {
503 		if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
504 			flags &= ~TH_FIN;
505 	}
506 
507 	recwin = sbspace(&so->so_rcv);
508 
509 	/*
510 	 * Sender silly window avoidance.   We transmit under the following
511 	 * conditions when len is non-zero:
512 	 *
513 	 *	- We have a full segment (or more with TSO)
514 	 *	- This is the last buffer in a write()/send() and we are
515 	 *	  either idle or running NODELAY
516 	 *	- we've timed out (e.g. persist timer)
517 	 *	- we have more then 1/2 the maximum send window's worth of
518 	 *	  data (receiver may be limited the window size)
519 	 *	- we need to retransmit
520 	 */
521 	if (len) {
522 		if (len >= tp->t_maxseg)
523 			goto send;
524 		/*
525 		 * NOTE! on localhost connections an 'ack' from the remote
526 		 * end may occur synchronously with the output and cause
527 		 * us to flush a buffer queued with moretocome.  XXX
528 		 *
529 		 * note: the len + off check is almost certainly unnecessary.
530 		 */
531 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
532 		    (idle || (tp->t_flags & TF_NODELAY)) &&
533 		    len + off >= so->so_snd.sb_cc &&
534 		    (tp->t_flags & TF_NOPUSH) == 0) {
535 			goto send;
536 		}
537 		if (tp->t_flags & TF_FORCEDATA)		/* typ. timeout case */
538 			goto send;
539 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
540 			goto send;
541 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
542 			goto send;
543 		if (sack_rxmit)
544 			goto send;
545 	}
546 
547 	/*
548 	 * Sending of standalone window updates.
549 	 *
550 	 * Window updates are important when we close our window due to a
551 	 * full socket buffer and are opening it again after the application
552 	 * reads data from it.  Once the window has opened again and the
553 	 * remote end starts to send again the ACK clock takes over and
554 	 * provides the most current window information.
555 	 *
556 	 * We must avoid the silly window syndrome whereas every read
557 	 * from the receive buffer, no matter how small, causes a window
558 	 * update to be sent.  We also should avoid sending a flurry of
559 	 * window updates when the socket buffer had queued a lot of data
560 	 * and the application is doing small reads.
561 	 *
562 	 * Prevent a flurry of pointless window updates by only sending
563 	 * an update when we can increase the advertized window by more
564 	 * than 1/4th of the socket buffer capacity.  When the buffer is
565 	 * getting full or is very small be more aggressive and send an
566 	 * update whenever we can increase by two mss sized segments.
567 	 * In all other situations the ACK's to new incoming data will
568 	 * carry further window increases.
569 	 *
570 	 * Don't send an independent window update if a delayed
571 	 * ACK is pending (it will get piggy-backed on it) or the
572 	 * remote side already has done a half-close and won't send
573 	 * more data.  Skip this if the connection is in T/TCP
574 	 * half-open state.
575 	 */
576 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
577 	    !(tp->t_flags & TF_DELACK) &&
578 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
579 		/*
580 		 * "adv" is the amount we could increase the window,
581 		 * taking into account that we are limited by
582 		 * TCP_MAXWIN << tp->rcv_scale.
583 		 */
584 		long adv;
585 		int oldwin;
586 
587 		adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale);
588 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
589 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
590 			adv -= oldwin;
591 		} else
592 			oldwin = 0;
593 
594 		/*
595 		 * If the new window size ends up being the same as the old
596 		 * size when it is scaled, then don't force a window update.
597 		 */
598 		if (oldwin >> tp->rcv_scale == (adv + oldwin) >> tp->rcv_scale)
599 			goto dontupdate;
600 
601 		if (adv >= (long)(2 * tp->t_maxseg) &&
602 		    (adv >= (long)(so->so_rcv.sb_hiwat / 4) ||
603 		     recwin <= (long)(so->so_rcv.sb_hiwat / 8) ||
604 		     so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg))
605 			goto send;
606 	}
607 dontupdate:
608 
609 	/*
610 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
611 	 * is also a catch-all for the retransmit timer timeout case.
612 	 */
613 	if (tp->t_flags & TF_ACKNOW)
614 		goto send;
615 	if ((flags & TH_RST) ||
616 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
617 		goto send;
618 	if (SEQ_GT(tp->snd_up, tp->snd_una))
619 		goto send;
620 	/*
621 	 * If our state indicates that FIN should be sent
622 	 * and we have not yet done so, then we need to send.
623 	 */
624 	if (flags & TH_FIN &&
625 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
626 		goto send;
627 	/*
628 	 * In SACK, it is possible for tcp_output to fail to send a segment
629 	 * after the retransmission timer has been turned off.  Make sure
630 	 * that the retransmission timer is set.
631 	 */
632 	if ((tp->t_flags & TF_SACK_PERMIT) &&
633 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
634 	    !tcp_timer_active(tp, TT_REXMT) &&
635 	    !tcp_timer_active(tp, TT_PERSIST)) {
636 		tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
637 		goto just_return;
638 	}
639 	/*
640 	 * TCP window updates are not reliable, rather a polling protocol
641 	 * using ``persist'' packets is used to insure receipt of window
642 	 * updates.  The three ``states'' for the output side are:
643 	 *	idle			not doing retransmits or persists
644 	 *	persisting		to move a small or zero window
645 	 *	(re)transmitting	and thereby not persisting
646 	 *
647 	 * tcp_timer_active(tp, TT_PERSIST)
648 	 *	is true when we are in persist state.
649 	 * (tp->t_flags & TF_FORCEDATA)
650 	 *	is set when we are called to send a persist packet.
651 	 * tcp_timer_active(tp, TT_REXMT)
652 	 *	is set when we are retransmitting
653 	 * The output side is idle when both timers are zero.
654 	 *
655 	 * If send window is too small, there is data to transmit, and no
656 	 * retransmit or persist is pending, then go to persist state.
657 	 * If nothing happens soon, send when timer expires:
658 	 * if window is nonzero, transmit what we can,
659 	 * otherwise force out a byte.
660 	 */
661 	if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) &&
662 	    !tcp_timer_active(tp, TT_PERSIST)) {
663 		tp->t_rxtshift = 0;
664 		tcp_setpersist(tp);
665 	}
666 
667 	/*
668 	 * No reason to send a segment, just return.
669 	 */
670 just_return:
671 	SOCKBUF_UNLOCK(&so->so_snd);
672 	return (0);
673 
674 send:
675 	SOCKBUF_LOCK_ASSERT(&so->so_snd);
676 	/*
677 	 * Before ESTABLISHED, force sending of initial options
678 	 * unless TCP set not to do any options.
679 	 * NOTE: we assume that the IP/TCP header plus TCP options
680 	 * always fit in a single mbuf, leaving room for a maximum
681 	 * link header, i.e.
682 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
683 	 */
684 	optlen = 0;
685 #ifdef INET6
686 	if (isipv6)
687 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
688 	else
689 #endif
690 		hdrlen = sizeof (struct tcpiphdr);
691 
692 	/*
693 	 * Compute options for segment.
694 	 * We only have to care about SYN and established connection
695 	 * segments.  Options for SYN-ACK segments are handled in TCP
696 	 * syncache.
697 	 */
698 	if ((tp->t_flags & TF_NOOPT) == 0) {
699 		to.to_flags = 0;
700 		/* Maximum segment size. */
701 		if (flags & TH_SYN) {
702 			tp->snd_nxt = tp->iss;
703 			to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
704 			to.to_flags |= TOF_MSS;
705 		}
706 		/* Window scaling. */
707 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
708 			to.to_wscale = tp->request_r_scale;
709 			to.to_flags |= TOF_SCALE;
710 		}
711 		/* Timestamps. */
712 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
713 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
714 			to.to_tsval = tcp_ts_getticks() + tp->ts_offset;
715 			to.to_tsecr = tp->ts_recent;
716 			to.to_flags |= TOF_TS;
717 			/* Set receive buffer autosizing timestamp. */
718 			if (tp->rfbuf_ts == 0 &&
719 			    (so->so_rcv.sb_flags & SB_AUTOSIZE))
720 				tp->rfbuf_ts = tcp_ts_getticks();
721 		}
722 		/* Selective ACK's. */
723 		if (tp->t_flags & TF_SACK_PERMIT) {
724 			if (flags & TH_SYN)
725 				to.to_flags |= TOF_SACKPERM;
726 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
727 			    (tp->t_flags & TF_SACK_PERMIT) &&
728 			    tp->rcv_numsacks > 0) {
729 				to.to_flags |= TOF_SACK;
730 				to.to_nsacks = tp->rcv_numsacks;
731 				to.to_sacks = (u_char *)tp->sackblks;
732 			}
733 		}
734 #ifdef TCP_SIGNATURE
735 		/* TCP-MD5 (RFC2385). */
736 		if (tp->t_flags & TF_SIGNATURE)
737 			to.to_flags |= TOF_SIGNATURE;
738 #endif /* TCP_SIGNATURE */
739 
740 		/* Processing the options. */
741 		hdrlen += optlen = tcp_addoptions(&to, opt);
742 	}
743 
744 #ifdef INET6
745 	if (isipv6)
746 		ipoptlen = ip6_optlen(tp->t_inpcb);
747 	else
748 #endif
749 	if (tp->t_inpcb->inp_options)
750 		ipoptlen = tp->t_inpcb->inp_options->m_len -
751 				offsetof(struct ipoption, ipopt_list);
752 	else
753 		ipoptlen = 0;
754 #ifdef IPSEC
755 	ipoptlen += ipsec_optlen;
756 #endif
757 
758 	/*
759 	 * Adjust data length if insertion of options will
760 	 * bump the packet length beyond the t_maxopd length.
761 	 * Clear the FIN bit because we cut off the tail of
762 	 * the segment.
763 	 */
764 	if (len + optlen + ipoptlen > tp->t_maxopd) {
765 		flags &= ~TH_FIN;
766 
767 		if (tso) {
768 			KASSERT(ipoptlen == 0,
769 			    ("%s: TSO can't do IP options", __func__));
770 
771 			/*
772 			 * Limit a burst to t_tsomax minus IP,
773 			 * TCP and options length to keep ip->ip_len
774 			 * from overflowing or exceeding the maximum
775 			 * length allowed by the network interface.
776 			 */
777 			if (len > tp->t_tsomax - hdrlen) {
778 				len = tp->t_tsomax - hdrlen;
779 				sendalot = 1;
780 			}
781 
782 			/*
783 			 * Prevent the last segment from being
784 			 * fractional unless the send sockbuf can
785 			 * be emptied.
786 			 */
787 			if (sendalot && off + len < so->so_snd.sb_cc) {
788 				len -= len % (tp->t_maxopd - optlen);
789 				sendalot = 1;
790 			}
791 
792 			/*
793 			 * Send the FIN in a separate segment
794 			 * after the bulk sending is done.
795 			 * We don't trust the TSO implementations
796 			 * to clear the FIN flag on all but the
797 			 * last segment.
798 			 */
799 			if (tp->t_flags & TF_NEEDFIN)
800 				sendalot = 1;
801 
802 		} else {
803 			len = tp->t_maxopd - optlen - ipoptlen;
804 			sendalot = 1;
805 		}
806 	} else
807 		tso = 0;
808 
809 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
810 	    ("%s: len > IP_MAXPACKET", __func__));
811 
812 /*#ifdef DIAGNOSTIC*/
813 #ifdef INET6
814 	if (max_linkhdr + hdrlen > MCLBYTES)
815 #else
816 	if (max_linkhdr + hdrlen > MHLEN)
817 #endif
818 		panic("tcphdr too big");
819 /*#endif*/
820 
821 	/*
822 	 * This KASSERT is here to catch edge cases at a well defined place.
823 	 * Before, those had triggered (random) panic conditions further down.
824 	 */
825 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
826 
827 	/*
828 	 * Grab a header mbuf, attaching a copy of data to
829 	 * be transmitted, and initialize the header from
830 	 * the template for sends on this connection.
831 	 */
832 	if (len) {
833 		struct mbuf *mb;
834 		u_int moff;
835 
836 		if ((tp->t_flags & TF_FORCEDATA) && len == 1)
837 			TCPSTAT_INC(tcps_sndprobe);
838 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
839 			tp->t_sndrexmitpack++;
840 			TCPSTAT_INC(tcps_sndrexmitpack);
841 			TCPSTAT_ADD(tcps_sndrexmitbyte, len);
842 		} else {
843 			TCPSTAT_INC(tcps_sndpack);
844 			TCPSTAT_ADD(tcps_sndbyte, len);
845 		}
846 #ifdef INET6
847 		if (MHLEN < hdrlen + max_linkhdr)
848 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
849 		else
850 #endif
851 			m = m_gethdr(M_NOWAIT, MT_DATA);
852 
853 		if (m == NULL) {
854 			SOCKBUF_UNLOCK(&so->so_snd);
855 			error = ENOBUFS;
856 			sack_rxmit = 0;
857 			goto out;
858 		}
859 
860 		m->m_data += max_linkhdr;
861 		m->m_len = hdrlen;
862 
863 		/*
864 		 * Start the m_copy functions from the closest mbuf
865 		 * to the offset in the socket buffer chain.
866 		 */
867 		mb = sbsndptr(&so->so_snd, off, len, &moff);
868 
869 		if (len <= MHLEN - hdrlen - max_linkhdr) {
870 			m_copydata(mb, moff, (int)len,
871 			    mtod(m, caddr_t) + hdrlen);
872 			m->m_len += len;
873 		} else {
874 			m->m_next = m_copy(mb, moff, (int)len);
875 			if (m->m_next == NULL) {
876 				SOCKBUF_UNLOCK(&so->so_snd);
877 				(void) m_free(m);
878 				error = ENOBUFS;
879 				sack_rxmit = 0;
880 				goto out;
881 			}
882 		}
883 
884 		/*
885 		 * If we're sending everything we've got, set PUSH.
886 		 * (This will keep happy those implementations which only
887 		 * give data to the user when a buffer fills or
888 		 * a PUSH comes in.)
889 		 */
890 		if (off + len == so->so_snd.sb_cc)
891 			flags |= TH_PUSH;
892 		SOCKBUF_UNLOCK(&so->so_snd);
893 	} else {
894 		SOCKBUF_UNLOCK(&so->so_snd);
895 		if (tp->t_flags & TF_ACKNOW)
896 			TCPSTAT_INC(tcps_sndacks);
897 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
898 			TCPSTAT_INC(tcps_sndctrl);
899 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
900 			TCPSTAT_INC(tcps_sndurg);
901 		else
902 			TCPSTAT_INC(tcps_sndwinup);
903 
904 		m = m_gethdr(M_NOWAIT, MT_DATA);
905 		if (m == NULL) {
906 			error = ENOBUFS;
907 			sack_rxmit = 0;
908 			goto out;
909 		}
910 #ifdef INET6
911 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
912 		    MHLEN >= hdrlen) {
913 			MH_ALIGN(m, hdrlen);
914 		} else
915 #endif
916 		m->m_data += max_linkhdr;
917 		m->m_len = hdrlen;
918 	}
919 	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
920 	m->m_pkthdr.rcvif = (struct ifnet *)0;
921 #ifdef MAC
922 	mac_inpcb_create_mbuf(tp->t_inpcb, m);
923 #endif
924 #ifdef INET6
925 	if (isipv6) {
926 		ip6 = mtod(m, struct ip6_hdr *);
927 		th = (struct tcphdr *)(ip6 + 1);
928 		tcpip_fillheaders(tp->t_inpcb, ip6, th);
929 	} else
930 #endif /* INET6 */
931 	{
932 		ip = mtod(m, struct ip *);
933 		ipov = (struct ipovly *)ip;
934 		th = (struct tcphdr *)(ip + 1);
935 		tcpip_fillheaders(tp->t_inpcb, ip, th);
936 	}
937 
938 	/*
939 	 * Fill in fields, remembering maximum advertised
940 	 * window for use in delaying messages about window sizes.
941 	 * If resending a FIN, be sure not to use a new sequence number.
942 	 */
943 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
944 	    tp->snd_nxt == tp->snd_max)
945 		tp->snd_nxt--;
946 	/*
947 	 * If we are starting a connection, send ECN setup
948 	 * SYN packet. If we are on a retransmit, we may
949 	 * resend those bits a number of times as per
950 	 * RFC 3168.
951 	 */
952 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
953 		if (tp->t_rxtshift >= 1) {
954 			if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
955 				flags |= TH_ECE|TH_CWR;
956 		} else
957 			flags |= TH_ECE|TH_CWR;
958 	}
959 
960 	if (tp->t_state == TCPS_ESTABLISHED &&
961 	    (tp->t_flags & TF_ECN_PERMIT)) {
962 		/*
963 		 * If the peer has ECN, mark data packets with
964 		 * ECN capable transmission (ECT).
965 		 * Ignore pure ack packets, retransmissions and window probes.
966 		 */
967 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
968 		    !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
969 #ifdef INET6
970 			if (isipv6)
971 				ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
972 			else
973 #endif
974 				ip->ip_tos |= IPTOS_ECN_ECT0;
975 			TCPSTAT_INC(tcps_ecn_ect0);
976 		}
977 
978 		/*
979 		 * Reply with proper ECN notifications.
980 		 */
981 		if (tp->t_flags & TF_ECN_SND_CWR) {
982 			flags |= TH_CWR;
983 			tp->t_flags &= ~TF_ECN_SND_CWR;
984 		}
985 		if (tp->t_flags & TF_ECN_SND_ECE)
986 			flags |= TH_ECE;
987 	}
988 
989 	/*
990 	 * If we are doing retransmissions, then snd_nxt will
991 	 * not reflect the first unsent octet.  For ACK only
992 	 * packets, we do not want the sequence number of the
993 	 * retransmitted packet, we want the sequence number
994 	 * of the next unsent octet.  So, if there is no data
995 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
996 	 * when filling in ti_seq.  But if we are in persist
997 	 * state, snd_max might reflect one byte beyond the
998 	 * right edge of the window, so use snd_nxt in that
999 	 * case, since we know we aren't doing a retransmission.
1000 	 * (retransmit and persist are mutually exclusive...)
1001 	 */
1002 	if (sack_rxmit == 0) {
1003 		if (len || (flags & (TH_SYN|TH_FIN)) ||
1004 		    tcp_timer_active(tp, TT_PERSIST))
1005 			th->th_seq = htonl(tp->snd_nxt);
1006 		else
1007 			th->th_seq = htonl(tp->snd_max);
1008 	} else {
1009 		th->th_seq = htonl(p->rxmit);
1010 		p->rxmit += len;
1011 		tp->sackhint.sack_bytes_rexmit += len;
1012 	}
1013 	th->th_ack = htonl(tp->rcv_nxt);
1014 	if (optlen) {
1015 		bcopy(opt, th + 1, optlen);
1016 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1017 	}
1018 	th->th_flags = flags;
1019 	/*
1020 	 * Calculate receive window.  Don't shrink window,
1021 	 * but avoid silly window syndrome.
1022 	 */
1023 	if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
1024 	    recwin < (long)tp->t_maxseg)
1025 		recwin = 0;
1026 	if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1027 	    recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
1028 		recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
1029 	if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
1030 		recwin = (long)TCP_MAXWIN << tp->rcv_scale;
1031 
1032 	/*
1033 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1034 	 * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
1035 	 * case is handled in syncache.
1036 	 */
1037 	if (flags & TH_SYN)
1038 		th->th_win = htons((u_short)
1039 				(min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1040 	else
1041 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1042 
1043 	/*
1044 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1045 	 * a 0 window.  This may cause the remote transmitter to stall.  This
1046 	 * flag tells soreceive() to disable delayed acknowledgements when
1047 	 * draining the buffer.  This can occur if the receiver is attempting
1048 	 * to read more data than can be buffered prior to transmitting on
1049 	 * the connection.
1050 	 */
1051 	if (th->th_win == 0) {
1052 		tp->t_sndzerowin++;
1053 		tp->t_flags |= TF_RXWIN0SENT;
1054 	} else
1055 		tp->t_flags &= ~TF_RXWIN0SENT;
1056 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1057 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1058 		th->th_flags |= TH_URG;
1059 	} else
1060 		/*
1061 		 * If no urgent pointer to send, then we pull
1062 		 * the urgent pointer to the left edge of the send window
1063 		 * so that it doesn't drift into the send window on sequence
1064 		 * number wraparound.
1065 		 */
1066 		tp->snd_up = tp->snd_una;		/* drag it along */
1067 
1068 #ifdef TCP_SIGNATURE
1069 	if (tp->t_flags & TF_SIGNATURE) {
1070 		int sigoff = to.to_signature - opt;
1071 		tcp_signature_compute(m, 0, len, optlen,
1072 		    (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
1073 	}
1074 #endif
1075 
1076 	/*
1077 	 * Put TCP length in extended header, and then
1078 	 * checksum extended header and data.
1079 	 */
1080 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1081 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1082 #ifdef INET6
1083 	if (isipv6) {
1084 		/*
1085 		 * ip6_plen is not need to be filled now, and will be filled
1086 		 * in ip6_output.
1087 		 */
1088 		m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1089 		th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
1090 		    optlen + len, IPPROTO_TCP, 0);
1091 	}
1092 #endif
1093 #if defined(INET6) && defined(INET)
1094 	else
1095 #endif
1096 #ifdef INET
1097 	{
1098 		m->m_pkthdr.csum_flags = CSUM_TCP;
1099 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1100 		    htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1101 
1102 		/* IP version must be set here for ipv4/ipv6 checking later */
1103 		KASSERT(ip->ip_v == IPVERSION,
1104 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1105 	}
1106 #endif
1107 
1108 	/*
1109 	 * Enable TSO and specify the size of the segments.
1110 	 * The TCP pseudo header checksum is always provided.
1111 	 * XXX: Fixme: This is currently not the case for IPv6.
1112 	 */
1113 	if (tso) {
1114 		KASSERT(len > tp->t_maxopd - optlen,
1115 		    ("%s: len <= tso_segsz", __func__));
1116 		m->m_pkthdr.csum_flags |= CSUM_TSO;
1117 		m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen;
1118 	}
1119 
1120 #ifdef IPSEC
1121 	KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL),
1122 	    ("%s: mbuf chain shorter than expected: %ld + %u + %u - %u != %u",
1123 	    __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL)));
1124 #else
1125 	KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL),
1126 	    ("%s: mbuf chain shorter than expected: %ld + %u + %u != %u",
1127 	    __func__, len, hdrlen, ipoptlen, m_length(m, NULL)));
1128 #endif
1129 
1130 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1131 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
1132 
1133 #ifdef TCPDEBUG
1134 	/*
1135 	 * Trace.
1136 	 */
1137 	if (so->so_options & SO_DEBUG) {
1138 		u_short save = 0;
1139 #ifdef INET6
1140 		if (!isipv6)
1141 #endif
1142 		{
1143 			save = ipov->ih_len;
1144 			ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1145 		}
1146 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1147 #ifdef INET6
1148 		if (!isipv6)
1149 #endif
1150 		ipov->ih_len = save;
1151 	}
1152 #endif /* TCPDEBUG */
1153 
1154 	/*
1155 	 * Fill in IP length and desired time to live and
1156 	 * send to IP level.  There should be a better way
1157 	 * to handle ttl and tos; we could keep them in
1158 	 * the template, but need a way to checksum without them.
1159 	 */
1160 	/*
1161 	 * m->m_pkthdr.len should have been set before cksum calcuration,
1162 	 * because in6_cksum() need it.
1163 	 */
1164 #ifdef INET6
1165 	if (isipv6) {
1166 		struct route_in6 ro;
1167 
1168 		bzero(&ro, sizeof(ro));
1169 		/*
1170 		 * we separately set hoplimit for every segment, since the
1171 		 * user might want to change the value via setsockopt.
1172 		 * Also, desired default hop limit might be changed via
1173 		 * Neighbor Discovery.
1174 		 */
1175 		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1176 
1177 		/* TODO: IPv6 IP6TOS_ECT bit on */
1178 		error = ip6_output(m, tp->t_inpcb->in6p_outputopts, &ro,
1179 		    ((so->so_options & SO_DONTROUTE) ?  IP_ROUTETOIF : 0),
1180 		    NULL, NULL, tp->t_inpcb);
1181 
1182 		if (error == EMSGSIZE && ro.ro_rt != NULL)
1183 			mtu = ro.ro_rt->rt_rmx.rmx_mtu;
1184 		RO_RTFREE(&ro);
1185 	}
1186 #endif /* INET6 */
1187 #if defined(INET) && defined(INET6)
1188 	else
1189 #endif
1190 #ifdef INET
1191     {
1192 	struct route ro;
1193 
1194 	bzero(&ro, sizeof(ro));
1195 	ip->ip_len = htons(m->m_pkthdr.len);
1196 #ifdef INET6
1197 	if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1198 		ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1199 #endif /* INET6 */
1200 	/*
1201 	 * If we do path MTU discovery, then we set DF on every packet.
1202 	 * This might not be the best thing to do according to RFC3390
1203 	 * Section 2. However the tcp hostcache migitates the problem
1204 	 * so it affects only the first tcp connection with a host.
1205 	 *
1206 	 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1207 	 */
1208 	if (V_path_mtu_discovery && tp->t_maxopd > V_tcp_minmss)
1209 		ip->ip_off |= htons(IP_DF);
1210 
1211 	error = ip_output(m, tp->t_inpcb->inp_options, &ro,
1212 	    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1213 	    tp->t_inpcb);
1214 
1215 	if (error == EMSGSIZE && ro.ro_rt != NULL)
1216 		mtu = ro.ro_rt->rt_rmx.rmx_mtu;
1217 	RO_RTFREE(&ro);
1218     }
1219 #endif /* INET */
1220 
1221 out:
1222 	/*
1223 	 * In transmit state, time the transmission and arrange for
1224 	 * the retransmit.  In persist state, just set snd_max.
1225 	 */
1226 	if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1227 	    !tcp_timer_active(tp, TT_PERSIST)) {
1228 		tcp_seq startseq = tp->snd_nxt;
1229 
1230 		/*
1231 		 * Advance snd_nxt over sequence space of this segment.
1232 		 */
1233 		if (flags & (TH_SYN|TH_FIN)) {
1234 			if (flags & TH_SYN)
1235 				tp->snd_nxt++;
1236 			if (flags & TH_FIN) {
1237 				tp->snd_nxt++;
1238 				tp->t_flags |= TF_SENTFIN;
1239 			}
1240 		}
1241 		if (sack_rxmit)
1242 			goto timer;
1243 		tp->snd_nxt += len;
1244 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1245 			tp->snd_max = tp->snd_nxt;
1246 			/*
1247 			 * Time this transmission if not a retransmission and
1248 			 * not currently timing anything.
1249 			 */
1250 			if (tp->t_rtttime == 0) {
1251 				tp->t_rtttime = ticks;
1252 				tp->t_rtseq = startseq;
1253 				TCPSTAT_INC(tcps_segstimed);
1254 			}
1255 		}
1256 
1257 		/*
1258 		 * Set retransmit timer if not currently set,
1259 		 * and not doing a pure ack or a keep-alive probe.
1260 		 * Initial value for retransmit timer is smoothed
1261 		 * round-trip time + 2 * round-trip time variance.
1262 		 * Initialize shift counter which is used for backoff
1263 		 * of retransmit time.
1264 		 */
1265 timer:
1266 		if (!tcp_timer_active(tp, TT_REXMT) &&
1267 		    ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1268 		     (tp->snd_nxt != tp->snd_una))) {
1269 			if (tcp_timer_active(tp, TT_PERSIST)) {
1270 				tcp_timer_activate(tp, TT_PERSIST, 0);
1271 				tp->t_rxtshift = 0;
1272 			}
1273 			tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1274 		}
1275 	} else {
1276 		/*
1277 		 * Persist case, update snd_max but since we are in
1278 		 * persist mode (no window) we do not update snd_nxt.
1279 		 */
1280 		int xlen = len;
1281 		if (flags & TH_SYN)
1282 			++xlen;
1283 		if (flags & TH_FIN) {
1284 			++xlen;
1285 			tp->t_flags |= TF_SENTFIN;
1286 		}
1287 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1288 			tp->snd_max = tp->snd_nxt + len;
1289 	}
1290 
1291 	if (error) {
1292 
1293 		/*
1294 		 * We know that the packet was lost, so back out the
1295 		 * sequence number advance, if any.
1296 		 *
1297 		 * If the error is EPERM the packet got blocked by the
1298 		 * local firewall.  Normally we should terminate the
1299 		 * connection but the blocking may have been spurious
1300 		 * due to a firewall reconfiguration cycle.  So we treat
1301 		 * it like a packet loss and let the retransmit timer and
1302 		 * timeouts do their work over time.
1303 		 * XXX: It is a POLA question whether calling tcp_drop right
1304 		 * away would be the really correct behavior instead.
1305 		 */
1306 		if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1307 		    !tcp_timer_active(tp, TT_PERSIST)) &&
1308 		    ((flags & TH_SYN) == 0) &&
1309 		    (error != EPERM)) {
1310 			if (sack_rxmit) {
1311 				p->rxmit -= len;
1312 				tp->sackhint.sack_bytes_rexmit -= len;
1313 				KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1314 				    ("sackhint bytes rtx >= 0"));
1315 			} else
1316 				tp->snd_nxt -= len;
1317 		}
1318 		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1319 		switch (error) {
1320 		case EPERM:
1321 			tp->t_softerror = error;
1322 			return (error);
1323 		case ENOBUFS:
1324 	                if (!tcp_timer_active(tp, TT_REXMT) &&
1325 			    !tcp_timer_active(tp, TT_PERSIST))
1326 	                        tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1327 			tp->snd_cwnd = tp->t_maxseg;
1328 			return (0);
1329 		case EMSGSIZE:
1330 			/*
1331 			 * For some reason the interface we used initially
1332 			 * to send segments changed to another or lowered
1333 			 * its MTU.
1334 			 * If TSO was active we either got an interface
1335 			 * without TSO capabilits or TSO was turned off.
1336 			 * If we obtained mtu from ip_output() then update
1337 			 * it and try again.
1338 			 */
1339 			if (tso)
1340 				tp->t_flags &= ~TF_TSO;
1341 			if (mtu != 0) {
1342 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
1343 				goto again;
1344 			}
1345 			return (error);
1346 		case EHOSTDOWN:
1347 		case EHOSTUNREACH:
1348 		case ENETDOWN:
1349 		case ENETUNREACH:
1350 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
1351 				tp->t_softerror = error;
1352 				return (0);
1353 			}
1354 			/* FALLTHROUGH */
1355 		default:
1356 			return (error);
1357 		}
1358 	}
1359 	TCPSTAT_INC(tcps_sndtotal);
1360 
1361 	/*
1362 	 * Data sent (as far as we can tell).
1363 	 * If this advertises a larger window than any other segment,
1364 	 * then remember the size of the advertised window.
1365 	 * Any pending ACK has now been sent.
1366 	 */
1367 	if (recwin >= 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1368 		tp->rcv_adv = tp->rcv_nxt + recwin;
1369 	tp->last_ack_sent = tp->rcv_nxt;
1370 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1371 	if (tcp_timer_active(tp, TT_DELACK))
1372 		tcp_timer_activate(tp, TT_DELACK, 0);
1373 #if 0
1374 	/*
1375 	 * This completely breaks TCP if newreno is turned on.  What happens
1376 	 * is that if delayed-acks are turned on on the receiver, this code
1377 	 * on the transmitter effectively destroys the TCP window, forcing
1378 	 * it to four packets (1.5Kx4 = 6K window).
1379 	 */
1380 	if (sendalot && --maxburst)
1381 		goto again;
1382 #endif
1383 	if (sendalot)
1384 		goto again;
1385 	return (0);
1386 }
1387 
1388 void
1389 tcp_setpersist(struct tcpcb *tp)
1390 {
1391 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1392 	int tt;
1393 
1394 	tp->t_flags &= ~TF_PREVVALID;
1395 	if (tcp_timer_active(tp, TT_REXMT))
1396 		panic("tcp_setpersist: retransmit pending");
1397 	/*
1398 	 * Start/restart persistance timer.
1399 	 */
1400 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1401 		      TCPTV_PERSMIN, TCPTV_PERSMAX);
1402 	tcp_timer_activate(tp, TT_PERSIST, tt);
1403 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1404 		tp->t_rxtshift++;
1405 }
1406 
1407 /*
1408  * Insert TCP options according to the supplied parameters to the place
1409  * optp in a consistent way.  Can handle unaligned destinations.
1410  *
1411  * The order of the option processing is crucial for optimal packing and
1412  * alignment for the scarce option space.
1413  *
1414  * The optimal order for a SYN/SYN-ACK segment is:
1415  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1416  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1417  *
1418  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1419  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1420  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1421  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1422  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1423  */
1424 int
1425 tcp_addoptions(struct tcpopt *to, u_char *optp)
1426 {
1427 	u_int mask, optlen = 0;
1428 
1429 	for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1430 		if ((to->to_flags & mask) != mask)
1431 			continue;
1432 		if (optlen == TCP_MAXOLEN)
1433 			break;
1434 		switch (to->to_flags & mask) {
1435 		case TOF_MSS:
1436 			while (optlen % 4) {
1437 				optlen += TCPOLEN_NOP;
1438 				*optp++ = TCPOPT_NOP;
1439 			}
1440 			if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1441 				continue;
1442 			optlen += TCPOLEN_MAXSEG;
1443 			*optp++ = TCPOPT_MAXSEG;
1444 			*optp++ = TCPOLEN_MAXSEG;
1445 			to->to_mss = htons(to->to_mss);
1446 			bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1447 			optp += sizeof(to->to_mss);
1448 			break;
1449 		case TOF_SCALE:
1450 			while (!optlen || optlen % 2 != 1) {
1451 				optlen += TCPOLEN_NOP;
1452 				*optp++ = TCPOPT_NOP;
1453 			}
1454 			if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1455 				continue;
1456 			optlen += TCPOLEN_WINDOW;
1457 			*optp++ = TCPOPT_WINDOW;
1458 			*optp++ = TCPOLEN_WINDOW;
1459 			*optp++ = to->to_wscale;
1460 			break;
1461 		case TOF_SACKPERM:
1462 			while (optlen % 2) {
1463 				optlen += TCPOLEN_NOP;
1464 				*optp++ = TCPOPT_NOP;
1465 			}
1466 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1467 				continue;
1468 			optlen += TCPOLEN_SACK_PERMITTED;
1469 			*optp++ = TCPOPT_SACK_PERMITTED;
1470 			*optp++ = TCPOLEN_SACK_PERMITTED;
1471 			break;
1472 		case TOF_TS:
1473 			while (!optlen || optlen % 4 != 2) {
1474 				optlen += TCPOLEN_NOP;
1475 				*optp++ = TCPOPT_NOP;
1476 			}
1477 			if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1478 				continue;
1479 			optlen += TCPOLEN_TIMESTAMP;
1480 			*optp++ = TCPOPT_TIMESTAMP;
1481 			*optp++ = TCPOLEN_TIMESTAMP;
1482 			to->to_tsval = htonl(to->to_tsval);
1483 			to->to_tsecr = htonl(to->to_tsecr);
1484 			bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1485 			optp += sizeof(to->to_tsval);
1486 			bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1487 			optp += sizeof(to->to_tsecr);
1488 			break;
1489 		case TOF_SIGNATURE:
1490 			{
1491 			int siglen = TCPOLEN_SIGNATURE - 2;
1492 
1493 			while (!optlen || optlen % 4 != 2) {
1494 				optlen += TCPOLEN_NOP;
1495 				*optp++ = TCPOPT_NOP;
1496 			}
1497 			if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE)
1498 				continue;
1499 			optlen += TCPOLEN_SIGNATURE;
1500 			*optp++ = TCPOPT_SIGNATURE;
1501 			*optp++ = TCPOLEN_SIGNATURE;
1502 			to->to_signature = optp;
1503 			while (siglen--)
1504 				 *optp++ = 0;
1505 			break;
1506 			}
1507 		case TOF_SACK:
1508 			{
1509 			int sackblks = 0;
1510 			struct sackblk *sack = (struct sackblk *)to->to_sacks;
1511 			tcp_seq sack_seq;
1512 
1513 			while (!optlen || optlen % 4 != 2) {
1514 				optlen += TCPOLEN_NOP;
1515 				*optp++ = TCPOPT_NOP;
1516 			}
1517 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1518 				continue;
1519 			optlen += TCPOLEN_SACKHDR;
1520 			*optp++ = TCPOPT_SACK;
1521 			sackblks = min(to->to_nsacks,
1522 					(TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1523 			*optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1524 			while (sackblks--) {
1525 				sack_seq = htonl(sack->start);
1526 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1527 				optp += sizeof(sack_seq);
1528 				sack_seq = htonl(sack->end);
1529 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1530 				optp += sizeof(sack_seq);
1531 				optlen += TCPOLEN_SACK;
1532 				sack++;
1533 			}
1534 			TCPSTAT_INC(tcps_sack_send_blocks);
1535 			break;
1536 			}
1537 		default:
1538 			panic("%s: unknown TCP option type", __func__);
1539 			break;
1540 		}
1541 	}
1542 
1543 	/* Terminate and pad TCP options to a 4 byte boundary. */
1544 	if (optlen % 4) {
1545 		optlen += TCPOLEN_EOL;
1546 		*optp++ = TCPOPT_EOL;
1547 	}
1548 	/*
1549 	 * According to RFC 793 (STD0007):
1550 	 *   "The content of the header beyond the End-of-Option option
1551 	 *    must be header padding (i.e., zero)."
1552 	 *   and later: "The padding is composed of zeros."
1553 	 */
1554 	while (optlen % 4) {
1555 		optlen += TCPOLEN_PAD;
1556 		*optp++ = TCPOPT_PAD;
1557 	}
1558 
1559 	KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1560 	return (optlen);
1561 }
1562