xref: /freebsd/sys/netinet/tcp_output.c (revision ff0ba87247820afbdfdc1b307c803f7923d0e4d3)
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/sdt.h>
50 #include <sys/socket.h>
51 #include <sys/socketvar.h>
52 #include <sys/sysctl.h>
53 
54 #include <net/if.h>
55 #include <net/route.h>
56 #include <net/vnet.h>
57 
58 #include <netinet/cc.h>
59 #include <netinet/in.h>
60 #include <netinet/in_kdtrace.h>
61 #include <netinet/in_systm.h>
62 #include <netinet/ip.h>
63 #include <netinet/in_pcb.h>
64 #include <netinet/ip_var.h>
65 #include <netinet/ip_options.h>
66 #ifdef INET6
67 #include <netinet6/in6_pcb.h>
68 #include <netinet/ip6.h>
69 #include <netinet6/ip6_var.h>
70 #endif
71 #define	TCPOUTFLAGS
72 #include <netinet/tcp_fsm.h>
73 #include <netinet/tcp_seq.h>
74 #include <netinet/tcp_timer.h>
75 #include <netinet/tcp_var.h>
76 #include <netinet/tcpip.h>
77 #ifdef TCPDEBUG
78 #include <netinet/tcp_debug.h>
79 #endif
80 #ifdef TCP_OFFLOAD
81 #include <netinet/tcp_offload.h>
82 #endif
83 
84 #ifdef IPSEC
85 #include <netipsec/ipsec.h>
86 #endif /*IPSEC*/
87 
88 #include <machine/in_cksum.h>
89 
90 #include <security/mac/mac_framework.h>
91 
92 VNET_DEFINE(int, path_mtu_discovery) = 1;
93 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW,
94 	&VNET_NAME(path_mtu_discovery), 1,
95 	"Enable Path MTU Discovery");
96 
97 VNET_DEFINE(int, tcp_do_tso) = 1;
98 #define	V_tcp_do_tso		VNET(tcp_do_tso)
99 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW,
100 	&VNET_NAME(tcp_do_tso), 0,
101 	"Enable TCP Segmentation Offload");
102 
103 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
104 #define	V_tcp_sendspace	VNET(tcp_sendspace)
105 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW,
106 	&VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
107 
108 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
109 #define	V_tcp_do_autosndbuf	VNET(tcp_do_autosndbuf)
110 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | 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_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | 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_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | 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 output 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, mtu;
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 #ifdef TCP_OFFLOAD
200 	if (tp->t_flags & TF_TOE)
201 		return (tcp_offload_output(tp));
202 #endif
203 
204 	/*
205 	 * Determine length of data that should be transmitted,
206 	 * and flags that will be used.
207 	 * If there is some data or critical controls (SYN, RST)
208 	 * to send, then transmit; otherwise, investigate further.
209 	 */
210 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
211 	if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur)
212 		cc_after_idle(tp);
213 	tp->t_flags &= ~TF_LASTIDLE;
214 	if (idle) {
215 		if (tp->t_flags & TF_MORETOCOME) {
216 			tp->t_flags |= TF_LASTIDLE;
217 			idle = 0;
218 		}
219 	}
220 again:
221 	/*
222 	 * If we've recently taken a timeout, snd_max will be greater than
223 	 * snd_nxt.  There may be SACK information that allows us to avoid
224 	 * resending already delivered data.  Adjust snd_nxt accordingly.
225 	 */
226 	if ((tp->t_flags & TF_SACK_PERMIT) &&
227 	    SEQ_LT(tp->snd_nxt, tp->snd_max))
228 		tcp_sack_adjust(tp);
229 	sendalot = 0;
230 	tso = 0;
231 	mtu = 0;
232 	off = tp->snd_nxt - tp->snd_una;
233 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
234 
235 	flags = tcp_outflags[tp->t_state];
236 	/*
237 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
238 	 * to send out new data (when sendalot is 1), bypass this function.
239 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
240 	 * we're replacing a (future) new transmission with a retransmission
241 	 * now, and we previously incremented snd_cwnd in tcp_input().
242 	 */
243 	/*
244 	 * Still in sack recovery , reset rxmit flag to zero.
245 	 */
246 	sack_rxmit = 0;
247 	sack_bytes_rxmt = 0;
248 	len = 0;
249 	p = NULL;
250 	if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
251 	    (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
252 		long cwin;
253 
254 		cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
255 		if (cwin < 0)
256 			cwin = 0;
257 		/* Do not retransmit SACK segments beyond snd_recover */
258 		if (SEQ_GT(p->end, tp->snd_recover)) {
259 			/*
260 			 * (At least) part of sack hole extends beyond
261 			 * snd_recover. Check to see if we can rexmit data
262 			 * for this hole.
263 			 */
264 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
265 				/*
266 				 * Can't rexmit any more data for this hole.
267 				 * That data will be rexmitted in the next
268 				 * sack recovery episode, when snd_recover
269 				 * moves past p->rxmit.
270 				 */
271 				p = NULL;
272 				goto after_sack_rexmit;
273 			} else
274 				/* Can rexmit part of the current hole */
275 				len = ((long)ulmin(cwin,
276 						   tp->snd_recover - p->rxmit));
277 		} else
278 			len = ((long)ulmin(cwin, p->end - p->rxmit));
279 		off = p->rxmit - tp->snd_una;
280 		KASSERT(off >= 0,("%s: sack block to the left of una : %d",
281 		    __func__, off));
282 		if (len > 0) {
283 			sack_rxmit = 1;
284 			sendalot = 1;
285 			TCPSTAT_INC(tcps_sack_rexmits);
286 			TCPSTAT_ADD(tcps_sack_rexmit_bytes,
287 			    min(len, tp->t_maxseg));
288 		}
289 	}
290 after_sack_rexmit:
291 	/*
292 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
293 	 * state flags.
294 	 */
295 	if (tp->t_flags & TF_NEEDFIN)
296 		flags |= TH_FIN;
297 	if (tp->t_flags & TF_NEEDSYN)
298 		flags |= TH_SYN;
299 
300 	SOCKBUF_LOCK(&so->so_snd);
301 	/*
302 	 * If in persist timeout with window of 0, send 1 byte.
303 	 * Otherwise, if window is small but nonzero
304 	 * and timer expired, we will send what we can
305 	 * and go to transmit state.
306 	 */
307 	if (tp->t_flags & TF_FORCEDATA) {
308 		if (sendwin == 0) {
309 			/*
310 			 * If we still have some data to send, then
311 			 * clear the FIN bit.  Usually this would
312 			 * happen below when it realizes that we
313 			 * aren't sending all the data.  However,
314 			 * if we have exactly 1 byte of unsent data,
315 			 * then it won't clear the FIN bit below,
316 			 * and if we are in persist state, we wind
317 			 * up sending the packet without recording
318 			 * that we sent the FIN bit.
319 			 *
320 			 * We can't just blindly clear the FIN bit,
321 			 * because if we don't have any more data
322 			 * to send then the probe will be the FIN
323 			 * itself.
324 			 */
325 			if (off < sbused(&so->so_snd))
326 				flags &= ~TH_FIN;
327 			sendwin = 1;
328 		} else {
329 			tcp_timer_activate(tp, TT_PERSIST, 0);
330 			tp->t_rxtshift = 0;
331 		}
332 	}
333 
334 	/*
335 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
336 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
337 	 * a negative length.  This can also occur when TCP opens up
338 	 * its congestion window while receiving additional duplicate
339 	 * acks after fast-retransmit because TCP will reset snd_nxt
340 	 * to snd_max after the fast-retransmit.
341 	 *
342 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
343 	 * be set to snd_una, the offset will be 0, and the length may
344 	 * wind up 0.
345 	 *
346 	 * If sack_rxmit is true we are retransmitting from the scoreboard
347 	 * in which case len is already set.
348 	 */
349 	if (sack_rxmit == 0) {
350 		if (sack_bytes_rxmt == 0)
351 			len = ((long)ulmin(sbavail(&so->so_snd), sendwin) -
352 			    off);
353 		else {
354 			long cwin;
355 
356                         /*
357 			 * We are inside of a SACK recovery episode and are
358 			 * sending new data, having retransmitted all the
359 			 * data possible in the scoreboard.
360 			 */
361 			len = ((long)ulmin(sbavail(&so->so_snd), tp->snd_wnd) -
362 			    off);
363 			/*
364 			 * Don't remove this (len > 0) check !
365 			 * We explicitly check for len > 0 here (although it
366 			 * isn't really necessary), to work around a gcc
367 			 * optimization issue - to force gcc to compute
368 			 * len above. Without this check, the computation
369 			 * of len is bungled by the optimizer.
370 			 */
371 			if (len > 0) {
372 				cwin = tp->snd_cwnd -
373 					(tp->snd_nxt - tp->sack_newdata) -
374 					sack_bytes_rxmt;
375 				if (cwin < 0)
376 					cwin = 0;
377 				len = lmin(len, cwin);
378 			}
379 		}
380 	}
381 
382 	/*
383 	 * Lop off SYN bit if it has already been sent.  However, if this
384 	 * is SYN-SENT state and if segment contains data and if we don't
385 	 * know that foreign host supports TAO, suppress sending segment.
386 	 */
387 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
388 		if (tp->t_state != TCPS_SYN_RECEIVED)
389 			flags &= ~TH_SYN;
390 		off--, len++;
391 	}
392 
393 	/*
394 	 * Be careful not to send data and/or FIN on SYN segments.
395 	 * This measure is needed to prevent interoperability problems
396 	 * with not fully conformant TCP implementations.
397 	 */
398 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
399 		len = 0;
400 		flags &= ~TH_FIN;
401 	}
402 
403 	if (len < 0) {
404 		/*
405 		 * If FIN has been sent but not acked,
406 		 * but we haven't been called to retransmit,
407 		 * len will be < 0.  Otherwise, window shrank
408 		 * after we sent into it.  If window shrank to 0,
409 		 * cancel pending retransmit, pull snd_nxt back
410 		 * to (closed) window, and set the persist timer
411 		 * if it isn't already going.  If the window didn't
412 		 * close completely, just wait for an ACK.
413 		 */
414 		len = 0;
415 		if (sendwin == 0) {
416 			tcp_timer_activate(tp, TT_REXMT, 0);
417 			tp->t_rxtshift = 0;
418 			tp->snd_nxt = tp->snd_una;
419 			if (!tcp_timer_active(tp, TT_PERSIST))
420 				tcp_setpersist(tp);
421 		}
422 	}
423 
424 	/* len will be >= 0 after this point. */
425 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
426 
427 	/*
428 	 * Automatic sizing of send socket buffer.  Often the send buffer
429 	 * size is not optimally adjusted to the actual network conditions
430 	 * at hand (delay bandwidth product).  Setting the buffer size too
431 	 * small limits throughput on links with high bandwidth and high
432 	 * delay (eg. trans-continental/oceanic links).  Setting the
433 	 * buffer size too big consumes too much real kernel memory,
434 	 * especially with many connections on busy servers.
435 	 *
436 	 * The criteria to step up the send buffer one notch are:
437 	 *  1. receive window of remote host is larger than send buffer
438 	 *     (with a fudge factor of 5/4th);
439 	 *  2. send buffer is filled to 7/8th with data (so we actually
440 	 *     have data to make use of it);
441 	 *  3. send buffer fill has not hit maximal automatic size;
442 	 *  4. our send window (slow start and cogestion controlled) is
443 	 *     larger than sent but unacknowledged data in send buffer.
444 	 *
445 	 * The remote host receive window scaling factor may limit the
446 	 * growing of the send buffer before it reaches its allowed
447 	 * maximum.
448 	 *
449 	 * It scales directly with slow start or congestion window
450 	 * and does at most one step per received ACK.  This fast
451 	 * scaling has the drawback of growing the send buffer beyond
452 	 * what is strictly necessary to make full use of a given
453 	 * delay*bandwith product.  However testing has shown this not
454 	 * to be much of an problem.  At worst we are trading wasting
455 	 * of available bandwith (the non-use of it) for wasting some
456 	 * socket buffer memory.
457 	 *
458 	 * TODO: Shrink send buffer during idle periods together
459 	 * with congestion window.  Requires another timer.  Has to
460 	 * wait for upcoming tcp timer rewrite.
461 	 *
462 	 * XXXGL: should there be used sbused() or sbavail()?
463 	 */
464 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
465 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
466 		    sbused(&so->so_snd) >= (so->so_snd.sb_hiwat / 8 * 7) &&
467 		    sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
468 		    sendwin >= (sbused(&so->so_snd) -
469 		    (tp->snd_nxt - tp->snd_una))) {
470 			if (!sbreserve_locked(&so->so_snd,
471 			    min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
472 			     V_tcp_autosndbuf_max), so, curthread))
473 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
474 		}
475 	}
476 
477 	/*
478 	 * Decide if we can use TCP Segmentation Offloading (if supported by
479 	 * hardware).
480 	 *
481 	 * TSO may only be used if we are in a pure bulk sending state.  The
482 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
483 	 * IP options prevent using TSO.  With TSO the TCP header is the same
484 	 * (except for the sequence number) for all generated packets.  This
485 	 * makes it impossible to transmit any options which vary per generated
486 	 * segment or packet.
487 	 */
488 #ifdef IPSEC
489 	/*
490 	 * Pre-calculate here as we save another lookup into the darknesses
491 	 * of IPsec that way and can actually decide if TSO is ok.
492 	 */
493 	ipsec_optlen = ipsec_hdrsiz_tcp(tp);
494 #endif
495 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
496 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
497 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
498 #ifdef IPSEC
499 	    ipsec_optlen == 0 &&
500 #endif
501 	    tp->t_inpcb->inp_options == NULL &&
502 	    tp->t_inpcb->in6p_options == NULL)
503 		tso = 1;
504 
505 	if (sack_rxmit) {
506 		if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd)))
507 			flags &= ~TH_FIN;
508 	} else {
509 		if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
510 		    sbused(&so->so_snd)))
511 			flags &= ~TH_FIN;
512 	}
513 
514 	recwin = sbspace(&so->so_rcv);
515 
516 	/*
517 	 * Sender silly window avoidance.   We transmit under the following
518 	 * conditions when len is non-zero:
519 	 *
520 	 *	- We have a full segment (or more with TSO)
521 	 *	- This is the last buffer in a write()/send() and we are
522 	 *	  either idle or running NODELAY
523 	 *	- we've timed out (e.g. persist timer)
524 	 *	- we have more then 1/2 the maximum send window's worth of
525 	 *	  data (receiver may be limited the window size)
526 	 *	- we need to retransmit
527 	 */
528 	if (len) {
529 		if (len >= tp->t_maxseg)
530 			goto send;
531 		/*
532 		 * NOTE! on localhost connections an 'ack' from the remote
533 		 * end may occur synchronously with the output and cause
534 		 * us to flush a buffer queued with moretocome.  XXX
535 		 *
536 		 * note: the len + off check is almost certainly unnecessary.
537 		 */
538 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
539 		    (idle || (tp->t_flags & TF_NODELAY)) &&
540 		    len + off >= sbavail(&so->so_snd) &&
541 		    (tp->t_flags & TF_NOPUSH) == 0) {
542 			goto send;
543 		}
544 		if (tp->t_flags & TF_FORCEDATA)		/* typ. timeout case */
545 			goto send;
546 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
547 			goto send;
548 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
549 			goto send;
550 		if (sack_rxmit)
551 			goto send;
552 	}
553 
554 	/*
555 	 * Sending of standalone window updates.
556 	 *
557 	 * Window updates are important when we close our window due to a
558 	 * full socket buffer and are opening it again after the application
559 	 * reads data from it.  Once the window has opened again and the
560 	 * remote end starts to send again the ACK clock takes over and
561 	 * provides the most current window information.
562 	 *
563 	 * We must avoid the silly window syndrome whereas every read
564 	 * from the receive buffer, no matter how small, causes a window
565 	 * update to be sent.  We also should avoid sending a flurry of
566 	 * window updates when the socket buffer had queued a lot of data
567 	 * and the application is doing small reads.
568 	 *
569 	 * Prevent a flurry of pointless window updates by only sending
570 	 * an update when we can increase the advertized window by more
571 	 * than 1/4th of the socket buffer capacity.  When the buffer is
572 	 * getting full or is very small be more aggressive and send an
573 	 * update whenever we can increase by two mss sized segments.
574 	 * In all other situations the ACK's to new incoming data will
575 	 * carry further window increases.
576 	 *
577 	 * Don't send an independent window update if a delayed
578 	 * ACK is pending (it will get piggy-backed on it) or the
579 	 * remote side already has done a half-close and won't send
580 	 * more data.  Skip this if the connection is in T/TCP
581 	 * half-open state.
582 	 */
583 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
584 	    !(tp->t_flags & TF_DELACK) &&
585 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
586 		/*
587 		 * "adv" is the amount we could increase the window,
588 		 * taking into account that we are limited by
589 		 * TCP_MAXWIN << tp->rcv_scale.
590 		 */
591 		long adv;
592 		int oldwin;
593 
594 		adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale);
595 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
596 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
597 			adv -= oldwin;
598 		} else
599 			oldwin = 0;
600 
601 		/*
602 		 * If the new window size ends up being the same as the old
603 		 * size when it is scaled, then don't force a window update.
604 		 */
605 		if (oldwin >> tp->rcv_scale == (adv + oldwin) >> tp->rcv_scale)
606 			goto dontupdate;
607 
608 		if (adv >= (long)(2 * tp->t_maxseg) &&
609 		    (adv >= (long)(so->so_rcv.sb_hiwat / 4) ||
610 		     recwin <= (long)(so->so_rcv.sb_hiwat / 8) ||
611 		     so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg))
612 			goto send;
613 	}
614 dontupdate:
615 
616 	/*
617 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
618 	 * is also a catch-all for the retransmit timer timeout case.
619 	 */
620 	if (tp->t_flags & TF_ACKNOW)
621 		goto send;
622 	if ((flags & TH_RST) ||
623 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
624 		goto send;
625 	if (SEQ_GT(tp->snd_up, tp->snd_una))
626 		goto send;
627 	/*
628 	 * If our state indicates that FIN should be sent
629 	 * and we have not yet done so, then we need to send.
630 	 */
631 	if (flags & TH_FIN &&
632 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
633 		goto send;
634 	/*
635 	 * In SACK, it is possible for tcp_output to fail to send a segment
636 	 * after the retransmission timer has been turned off.  Make sure
637 	 * that the retransmission timer is set.
638 	 */
639 	if ((tp->t_flags & TF_SACK_PERMIT) &&
640 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
641 	    !tcp_timer_active(tp, TT_REXMT) &&
642 	    !tcp_timer_active(tp, TT_PERSIST)) {
643 		tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
644 		goto just_return;
645 	}
646 	/*
647 	 * TCP window updates are not reliable, rather a polling protocol
648 	 * using ``persist'' packets is used to insure receipt of window
649 	 * updates.  The three ``states'' for the output side are:
650 	 *	idle			not doing retransmits or persists
651 	 *	persisting		to move a small or zero window
652 	 *	(re)transmitting	and thereby not persisting
653 	 *
654 	 * tcp_timer_active(tp, TT_PERSIST)
655 	 *	is true when we are in persist state.
656 	 * (tp->t_flags & TF_FORCEDATA)
657 	 *	is set when we are called to send a persist packet.
658 	 * tcp_timer_active(tp, TT_REXMT)
659 	 *	is set when we are retransmitting
660 	 * The output side is idle when both timers are zero.
661 	 *
662 	 * If send window is too small, there is data to transmit, and no
663 	 * retransmit or persist is pending, then go to persist state.
664 	 * If nothing happens soon, send when timer expires:
665 	 * if window is nonzero, transmit what we can,
666 	 * otherwise force out a byte.
667 	 */
668 	if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
669 	    !tcp_timer_active(tp, TT_PERSIST)) {
670 		tp->t_rxtshift = 0;
671 		tcp_setpersist(tp);
672 	}
673 
674 	/*
675 	 * No reason to send a segment, just return.
676 	 */
677 just_return:
678 	SOCKBUF_UNLOCK(&so->so_snd);
679 	return (0);
680 
681 send:
682 	SOCKBUF_LOCK_ASSERT(&so->so_snd);
683 	if (len > 0) {
684 		if (len >= tp->t_maxseg)
685 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
686 		else
687 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
688 	}
689 	/*
690 	 * Before ESTABLISHED, force sending of initial options
691 	 * unless TCP set not to do any options.
692 	 * NOTE: we assume that the IP/TCP header plus TCP options
693 	 * always fit in a single mbuf, leaving room for a maximum
694 	 * link header, i.e.
695 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
696 	 */
697 	optlen = 0;
698 #ifdef INET6
699 	if (isipv6)
700 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
701 	else
702 #endif
703 		hdrlen = sizeof (struct tcpiphdr);
704 
705 	/*
706 	 * Compute options for segment.
707 	 * We only have to care about SYN and established connection
708 	 * segments.  Options for SYN-ACK segments are handled in TCP
709 	 * syncache.
710 	 */
711 	if ((tp->t_flags & TF_NOOPT) == 0) {
712 		to.to_flags = 0;
713 		/* Maximum segment size. */
714 		if (flags & TH_SYN) {
715 			tp->snd_nxt = tp->iss;
716 			to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
717 			to.to_flags |= TOF_MSS;
718 		}
719 		/* Window scaling. */
720 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
721 			to.to_wscale = tp->request_r_scale;
722 			to.to_flags |= TOF_SCALE;
723 		}
724 		/* Timestamps. */
725 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
726 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
727 			to.to_tsval = tcp_ts_getticks() + tp->ts_offset;
728 			to.to_tsecr = tp->ts_recent;
729 			to.to_flags |= TOF_TS;
730 			/* Set receive buffer autosizing timestamp. */
731 			if (tp->rfbuf_ts == 0 &&
732 			    (so->so_rcv.sb_flags & SB_AUTOSIZE))
733 				tp->rfbuf_ts = tcp_ts_getticks();
734 		}
735 		/* Selective ACK's. */
736 		if (tp->t_flags & TF_SACK_PERMIT) {
737 			if (flags & TH_SYN)
738 				to.to_flags |= TOF_SACKPERM;
739 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
740 			    (tp->t_flags & TF_SACK_PERMIT) &&
741 			    tp->rcv_numsacks > 0) {
742 				to.to_flags |= TOF_SACK;
743 				to.to_nsacks = tp->rcv_numsacks;
744 				to.to_sacks = (u_char *)tp->sackblks;
745 			}
746 		}
747 #ifdef TCP_SIGNATURE
748 		/* TCP-MD5 (RFC2385). */
749 		if (tp->t_flags & TF_SIGNATURE)
750 			to.to_flags |= TOF_SIGNATURE;
751 #endif /* TCP_SIGNATURE */
752 
753 		/* Processing the options. */
754 		hdrlen += optlen = tcp_addoptions(&to, opt);
755 	}
756 
757 #ifdef INET6
758 	if (isipv6)
759 		ipoptlen = ip6_optlen(tp->t_inpcb);
760 	else
761 #endif
762 	if (tp->t_inpcb->inp_options)
763 		ipoptlen = tp->t_inpcb->inp_options->m_len -
764 				offsetof(struct ipoption, ipopt_list);
765 	else
766 		ipoptlen = 0;
767 #ifdef IPSEC
768 	ipoptlen += ipsec_optlen;
769 #endif
770 
771 	/*
772 	 * Adjust data length if insertion of options will
773 	 * bump the packet length beyond the t_maxopd length.
774 	 * Clear the FIN bit because we cut off the tail of
775 	 * the segment.
776 	 */
777 	if (len + optlen + ipoptlen > tp->t_maxopd) {
778 		flags &= ~TH_FIN;
779 
780 		if (tso) {
781 			u_int if_hw_tsomax;
782 			u_int if_hw_tsomaxsegcount;
783 			u_int if_hw_tsomaxsegsize;
784 			struct mbuf *mb;
785 			u_int moff;
786 			int max_len;
787 
788 			/* extract TSO information */
789 			if_hw_tsomax = tp->t_tsomax;
790 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
791 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
792 
793 			/*
794 			 * Limit a TSO burst to prevent it from
795 			 * overflowing or exceeding the maximum length
796 			 * allowed by the network interface:
797 			 */
798 			KASSERT(ipoptlen == 0,
799 			    ("%s: TSO can't do IP options", __func__));
800 
801 			/*
802 			 * Check if we should limit by maximum payload
803 			 * length:
804 			 */
805 			if (if_hw_tsomax != 0) {
806 				/* compute maximum TSO length */
807 				max_len = (if_hw_tsomax - hdrlen);
808 				if (max_len <= 0) {
809 					len = 0;
810 				} else if (len > max_len) {
811 					sendalot = 1;
812 					len = max_len;
813 				}
814 			}
815 
816 			/*
817 			 * Check if we should limit by maximum segment
818 			 * size and count:
819 			 */
820 			if (if_hw_tsomaxsegcount != 0 &&
821 			    if_hw_tsomaxsegsize != 0) {
822 				max_len = 0;
823 				mb = sbsndmbuf(&so->so_snd, off, &moff);
824 
825 				while (mb != NULL && max_len < len) {
826 					u_int mlen;
827 					u_int frags;
828 
829 					/*
830 					 * Get length of mbuf fragment
831 					 * and how many hardware frags,
832 					 * rounded up, it would use:
833 					 */
834 					mlen = (mb->m_len - moff);
835 					frags = howmany(mlen,
836 					    if_hw_tsomaxsegsize);
837 
838 					/* Handle special case: Zero Length Mbuf */
839 					if (frags == 0)
840 						frags = 1;
841 
842 					/*
843 					 * Check if the fragment limit
844 					 * will be reached or exceeded:
845 					 */
846 					if (frags >= if_hw_tsomaxsegcount) {
847 						max_len += min(mlen,
848 						    if_hw_tsomaxsegcount *
849 						    if_hw_tsomaxsegsize);
850 						break;
851 					}
852 					max_len += mlen;
853 					if_hw_tsomaxsegcount -= frags;
854 					moff = 0;
855 					mb = mb->m_next;
856 				}
857 				if (max_len <= 0) {
858 					len = 0;
859 				} else if (len > max_len) {
860 					sendalot = 1;
861 					len = max_len;
862 				}
863 			}
864 
865 			/*
866 			 * Prevent the last segment from being
867 			 * fractional unless the send sockbuf can be
868 			 * emptied:
869 			 */
870 			max_len = (tp->t_maxopd - optlen);
871 			if ((off + len) < sbavail(&so->so_snd)) {
872 				moff = len % max_len;
873 				if (moff != 0) {
874 					len -= moff;
875 					sendalot = 1;
876 				}
877 			}
878 
879 			/*
880 			 * In case there are too many small fragments
881 			 * don't use TSO:
882 			 */
883 			if (len <= max_len) {
884 				len = max_len;
885 				sendalot = 1;
886 				tso = 0;
887 			}
888 
889 			/*
890 			 * Send the FIN in a separate segment
891 			 * after the bulk sending is done.
892 			 * We don't trust the TSO implementations
893 			 * to clear the FIN flag on all but the
894 			 * last segment.
895 			 */
896 			if (tp->t_flags & TF_NEEDFIN)
897 				sendalot = 1;
898 
899 		} else {
900 			len = tp->t_maxopd - optlen - ipoptlen;
901 			sendalot = 1;
902 		}
903 	} else
904 		tso = 0;
905 
906 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
907 	    ("%s: len > IP_MAXPACKET", __func__));
908 
909 /*#ifdef DIAGNOSTIC*/
910 #ifdef INET6
911 	if (max_linkhdr + hdrlen > MCLBYTES)
912 #else
913 	if (max_linkhdr + hdrlen > MHLEN)
914 #endif
915 		panic("tcphdr too big");
916 /*#endif*/
917 
918 	/*
919 	 * This KASSERT is here to catch edge cases at a well defined place.
920 	 * Before, those had triggered (random) panic conditions further down.
921 	 */
922 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
923 
924 	/*
925 	 * Grab a header mbuf, attaching a copy of data to
926 	 * be transmitted, and initialize the header from
927 	 * the template for sends on this connection.
928 	 */
929 	if (len) {
930 		struct mbuf *mb;
931 		u_int moff;
932 
933 		if ((tp->t_flags & TF_FORCEDATA) && len == 1)
934 			TCPSTAT_INC(tcps_sndprobe);
935 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
936 			tp->t_sndrexmitpack++;
937 			TCPSTAT_INC(tcps_sndrexmitpack);
938 			TCPSTAT_ADD(tcps_sndrexmitbyte, len);
939 		} else {
940 			TCPSTAT_INC(tcps_sndpack);
941 			TCPSTAT_ADD(tcps_sndbyte, len);
942 		}
943 #ifdef INET6
944 		if (MHLEN < hdrlen + max_linkhdr)
945 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
946 		else
947 #endif
948 			m = m_gethdr(M_NOWAIT, MT_DATA);
949 
950 		if (m == NULL) {
951 			SOCKBUF_UNLOCK(&so->so_snd);
952 			error = ENOBUFS;
953 			sack_rxmit = 0;
954 			goto out;
955 		}
956 
957 		m->m_data += max_linkhdr;
958 		m->m_len = hdrlen;
959 
960 		/*
961 		 * Start the m_copy functions from the closest mbuf
962 		 * to the offset in the socket buffer chain.
963 		 */
964 		mb = sbsndptr(&so->so_snd, off, len, &moff);
965 
966 		if (len <= MHLEN - hdrlen - max_linkhdr) {
967 			m_copydata(mb, moff, (int)len,
968 			    mtod(m, caddr_t) + hdrlen);
969 			m->m_len += len;
970 		} else {
971 			m->m_next = m_copy(mb, moff, (int)len);
972 			if (m->m_next == NULL) {
973 				SOCKBUF_UNLOCK(&so->so_snd);
974 				(void) m_free(m);
975 				error = ENOBUFS;
976 				sack_rxmit = 0;
977 				goto out;
978 			}
979 		}
980 
981 		/*
982 		 * If we're sending everything we've got, set PUSH.
983 		 * (This will keep happy those implementations which only
984 		 * give data to the user when a buffer fills or
985 		 * a PUSH comes in.)
986 		 */
987 		if (off + len == sbused(&so->so_snd))
988 			flags |= TH_PUSH;
989 		SOCKBUF_UNLOCK(&so->so_snd);
990 	} else {
991 		SOCKBUF_UNLOCK(&so->so_snd);
992 		if (tp->t_flags & TF_ACKNOW)
993 			TCPSTAT_INC(tcps_sndacks);
994 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
995 			TCPSTAT_INC(tcps_sndctrl);
996 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
997 			TCPSTAT_INC(tcps_sndurg);
998 		else
999 			TCPSTAT_INC(tcps_sndwinup);
1000 
1001 		m = m_gethdr(M_NOWAIT, MT_DATA);
1002 		if (m == NULL) {
1003 			error = ENOBUFS;
1004 			sack_rxmit = 0;
1005 			goto out;
1006 		}
1007 #ifdef INET6
1008 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1009 		    MHLEN >= hdrlen) {
1010 			MH_ALIGN(m, hdrlen);
1011 		} else
1012 #endif
1013 		m->m_data += max_linkhdr;
1014 		m->m_len = hdrlen;
1015 	}
1016 	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1017 	m->m_pkthdr.rcvif = (struct ifnet *)0;
1018 #ifdef MAC
1019 	mac_inpcb_create_mbuf(tp->t_inpcb, m);
1020 #endif
1021 #ifdef INET6
1022 	if (isipv6) {
1023 		ip6 = mtod(m, struct ip6_hdr *);
1024 		th = (struct tcphdr *)(ip6 + 1);
1025 		tcpip_fillheaders(tp->t_inpcb, ip6, th);
1026 	} else
1027 #endif /* INET6 */
1028 	{
1029 		ip = mtod(m, struct ip *);
1030 		ipov = (struct ipovly *)ip;
1031 		th = (struct tcphdr *)(ip + 1);
1032 		tcpip_fillheaders(tp->t_inpcb, ip, th);
1033 	}
1034 
1035 	/*
1036 	 * Fill in fields, remembering maximum advertised
1037 	 * window for use in delaying messages about window sizes.
1038 	 * If resending a FIN, be sure not to use a new sequence number.
1039 	 */
1040 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1041 	    tp->snd_nxt == tp->snd_max)
1042 		tp->snd_nxt--;
1043 	/*
1044 	 * If we are starting a connection, send ECN setup
1045 	 * SYN packet. If we are on a retransmit, we may
1046 	 * resend those bits a number of times as per
1047 	 * RFC 3168.
1048 	 */
1049 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
1050 		if (tp->t_rxtshift >= 1) {
1051 			if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
1052 				flags |= TH_ECE|TH_CWR;
1053 		} else
1054 			flags |= TH_ECE|TH_CWR;
1055 	}
1056 
1057 	if (tp->t_state == TCPS_ESTABLISHED &&
1058 	    (tp->t_flags & TF_ECN_PERMIT)) {
1059 		/*
1060 		 * If the peer has ECN, mark data packets with
1061 		 * ECN capable transmission (ECT).
1062 		 * Ignore pure ack packets, retransmissions and window probes.
1063 		 */
1064 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1065 		    !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
1066 #ifdef INET6
1067 			if (isipv6)
1068 				ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1069 			else
1070 #endif
1071 				ip->ip_tos |= IPTOS_ECN_ECT0;
1072 			TCPSTAT_INC(tcps_ecn_ect0);
1073 		}
1074 
1075 		/*
1076 		 * Reply with proper ECN notifications.
1077 		 */
1078 		if (tp->t_flags & TF_ECN_SND_CWR) {
1079 			flags |= TH_CWR;
1080 			tp->t_flags &= ~TF_ECN_SND_CWR;
1081 		}
1082 		if (tp->t_flags & TF_ECN_SND_ECE)
1083 			flags |= TH_ECE;
1084 	}
1085 
1086 	/*
1087 	 * If we are doing retransmissions, then snd_nxt will
1088 	 * not reflect the first unsent octet.  For ACK only
1089 	 * packets, we do not want the sequence number of the
1090 	 * retransmitted packet, we want the sequence number
1091 	 * of the next unsent octet.  So, if there is no data
1092 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
1093 	 * when filling in ti_seq.  But if we are in persist
1094 	 * state, snd_max might reflect one byte beyond the
1095 	 * right edge of the window, so use snd_nxt in that
1096 	 * case, since we know we aren't doing a retransmission.
1097 	 * (retransmit and persist are mutually exclusive...)
1098 	 */
1099 	if (sack_rxmit == 0) {
1100 		if (len || (flags & (TH_SYN|TH_FIN)) ||
1101 		    tcp_timer_active(tp, TT_PERSIST))
1102 			th->th_seq = htonl(tp->snd_nxt);
1103 		else
1104 			th->th_seq = htonl(tp->snd_max);
1105 	} else {
1106 		th->th_seq = htonl(p->rxmit);
1107 		p->rxmit += len;
1108 		tp->sackhint.sack_bytes_rexmit += len;
1109 	}
1110 	th->th_ack = htonl(tp->rcv_nxt);
1111 	if (optlen) {
1112 		bcopy(opt, th + 1, optlen);
1113 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1114 	}
1115 	th->th_flags = flags;
1116 	/*
1117 	 * Calculate receive window.  Don't shrink window,
1118 	 * but avoid silly window syndrome.
1119 	 */
1120 	if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
1121 	    recwin < (long)tp->t_maxseg)
1122 		recwin = 0;
1123 	if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1124 	    recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
1125 		recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
1126 	if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
1127 		recwin = (long)TCP_MAXWIN << tp->rcv_scale;
1128 
1129 	/*
1130 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1131 	 * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
1132 	 * case is handled in syncache.
1133 	 */
1134 	if (flags & TH_SYN)
1135 		th->th_win = htons((u_short)
1136 				(min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1137 	else
1138 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1139 
1140 	/*
1141 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1142 	 * a 0 window.  This may cause the remote transmitter to stall.  This
1143 	 * flag tells soreceive() to disable delayed acknowledgements when
1144 	 * draining the buffer.  This can occur if the receiver is attempting
1145 	 * to read more data than can be buffered prior to transmitting on
1146 	 * the connection.
1147 	 */
1148 	if (th->th_win == 0) {
1149 		tp->t_sndzerowin++;
1150 		tp->t_flags |= TF_RXWIN0SENT;
1151 	} else
1152 		tp->t_flags &= ~TF_RXWIN0SENT;
1153 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1154 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1155 		th->th_flags |= TH_URG;
1156 	} else
1157 		/*
1158 		 * If no urgent pointer to send, then we pull
1159 		 * the urgent pointer to the left edge of the send window
1160 		 * so that it doesn't drift into the send window on sequence
1161 		 * number wraparound.
1162 		 */
1163 		tp->snd_up = tp->snd_una;		/* drag it along */
1164 
1165 #ifdef TCP_SIGNATURE
1166 	if (tp->t_flags & TF_SIGNATURE) {
1167 		int sigoff = to.to_signature - opt;
1168 		tcp_signature_compute(m, 0, len, optlen,
1169 		    (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
1170 	}
1171 #endif
1172 
1173 	/*
1174 	 * Put TCP length in extended header, and then
1175 	 * checksum extended header and data.
1176 	 */
1177 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1178 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1179 #ifdef INET6
1180 	if (isipv6) {
1181 		/*
1182 		 * ip6_plen is not need to be filled now, and will be filled
1183 		 * in ip6_output.
1184 		 */
1185 		m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1186 		th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
1187 		    optlen + len, IPPROTO_TCP, 0);
1188 	}
1189 #endif
1190 #if defined(INET6) && defined(INET)
1191 	else
1192 #endif
1193 #ifdef INET
1194 	{
1195 		m->m_pkthdr.csum_flags = CSUM_TCP;
1196 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1197 		    htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1198 
1199 		/* IP version must be set here for ipv4/ipv6 checking later */
1200 		KASSERT(ip->ip_v == IPVERSION,
1201 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1202 	}
1203 #endif
1204 
1205 	/*
1206 	 * Enable TSO and specify the size of the segments.
1207 	 * The TCP pseudo header checksum is always provided.
1208 	 * XXX: Fixme: This is currently not the case for IPv6.
1209 	 */
1210 	if (tso) {
1211 		KASSERT(len > tp->t_maxopd - optlen,
1212 		    ("%s: len <= tso_segsz", __func__));
1213 		m->m_pkthdr.csum_flags |= CSUM_TSO;
1214 		m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen;
1215 	}
1216 
1217 #ifdef IPSEC
1218 	KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL),
1219 	    ("%s: mbuf chain shorter than expected: %ld + %u + %u - %u != %u",
1220 	    __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL)));
1221 #else
1222 	KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL),
1223 	    ("%s: mbuf chain shorter than expected: %ld + %u + %u != %u",
1224 	    __func__, len, hdrlen, ipoptlen, m_length(m, NULL)));
1225 #endif
1226 
1227 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1228 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
1229 
1230 #ifdef TCPDEBUG
1231 	/*
1232 	 * Trace.
1233 	 */
1234 	if (so->so_options & SO_DEBUG) {
1235 		u_short save = 0;
1236 #ifdef INET6
1237 		if (!isipv6)
1238 #endif
1239 		{
1240 			save = ipov->ih_len;
1241 			ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1242 		}
1243 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1244 #ifdef INET6
1245 		if (!isipv6)
1246 #endif
1247 		ipov->ih_len = save;
1248 	}
1249 #endif /* TCPDEBUG */
1250 
1251 	/*
1252 	 * Fill in IP length and desired time to live and
1253 	 * send to IP level.  There should be a better way
1254 	 * to handle ttl and tos; we could keep them in
1255 	 * the template, but need a way to checksum without them.
1256 	 */
1257 	/*
1258 	 * m->m_pkthdr.len should have been set before checksum calculation,
1259 	 * because in6_cksum() need it.
1260 	 */
1261 #ifdef INET6
1262 	if (isipv6) {
1263 		struct route_in6 ro;
1264 
1265 		bzero(&ro, sizeof(ro));
1266 		/*
1267 		 * we separately set hoplimit for every segment, since the
1268 		 * user might want to change the value via setsockopt.
1269 		 * Also, desired default hop limit might be changed via
1270 		 * Neighbor Discovery.
1271 		 */
1272 		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1273 
1274 		/*
1275 		 * Set the packet size here for the benefit of DTrace probes.
1276 		 * ip6_output() will set it properly; it's supposed to include
1277 		 * the option header lengths as well.
1278 		 */
1279 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1280 
1281 		if (V_path_mtu_discovery && tp->t_maxopd > V_tcp_minmss)
1282 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1283 		else
1284 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1285 
1286 		if (tp->t_state == TCPS_SYN_SENT)
1287 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1288 
1289 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1290 
1291 		/* TODO: IPv6 IP6TOS_ECT bit on */
1292 		error = ip6_output(m, tp->t_inpcb->in6p_outputopts, &ro,
1293 		    ((so->so_options & SO_DONTROUTE) ?  IP_ROUTETOIF : 0),
1294 		    NULL, NULL, tp->t_inpcb);
1295 
1296 		if (error == EMSGSIZE && ro.ro_rt != NULL)
1297 			mtu = ro.ro_rt->rt_mtu;
1298 		RO_RTFREE(&ro);
1299 	}
1300 #endif /* INET6 */
1301 #if defined(INET) && defined(INET6)
1302 	else
1303 #endif
1304 #ifdef INET
1305     {
1306 	struct route ro;
1307 
1308 	bzero(&ro, sizeof(ro));
1309 	ip->ip_len = htons(m->m_pkthdr.len);
1310 #ifdef INET6
1311 	if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1312 		ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1313 #endif /* INET6 */
1314 	/*
1315 	 * If we do path MTU discovery, then we set DF on every packet.
1316 	 * This might not be the best thing to do according to RFC3390
1317 	 * Section 2. However the tcp hostcache migitates the problem
1318 	 * so it affects only the first tcp connection with a host.
1319 	 *
1320 	 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1321 	 */
1322 	if (V_path_mtu_discovery && tp->t_maxopd > V_tcp_minmss) {
1323 		ip->ip_off |= htons(IP_DF);
1324 		tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1325 	} else {
1326 		tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1327 	}
1328 
1329 	if (tp->t_state == TCPS_SYN_SENT)
1330 		TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1331 
1332 	TCP_PROBE5(send, NULL, tp, ip, tp, th);
1333 
1334 	error = ip_output(m, tp->t_inpcb->inp_options, &ro,
1335 	    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1336 	    tp->t_inpcb);
1337 
1338 	if (error == EMSGSIZE && ro.ro_rt != NULL)
1339 		mtu = ro.ro_rt->rt_mtu;
1340 	RO_RTFREE(&ro);
1341     }
1342 #endif /* INET */
1343 
1344 out:
1345 	/*
1346 	 * In transmit state, time the transmission and arrange for
1347 	 * the retransmit.  In persist state, just set snd_max.
1348 	 */
1349 	if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1350 	    !tcp_timer_active(tp, TT_PERSIST)) {
1351 		tcp_seq startseq = tp->snd_nxt;
1352 
1353 		/*
1354 		 * Advance snd_nxt over sequence space of this segment.
1355 		 */
1356 		if (flags & (TH_SYN|TH_FIN)) {
1357 			if (flags & TH_SYN)
1358 				tp->snd_nxt++;
1359 			if (flags & TH_FIN) {
1360 				tp->snd_nxt++;
1361 				tp->t_flags |= TF_SENTFIN;
1362 			}
1363 		}
1364 		if (sack_rxmit)
1365 			goto timer;
1366 		tp->snd_nxt += len;
1367 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1368 			tp->snd_max = tp->snd_nxt;
1369 			/*
1370 			 * Time this transmission if not a retransmission and
1371 			 * not currently timing anything.
1372 			 */
1373 			if (tp->t_rtttime == 0) {
1374 				tp->t_rtttime = ticks;
1375 				tp->t_rtseq = startseq;
1376 				TCPSTAT_INC(tcps_segstimed);
1377 			}
1378 		}
1379 
1380 		/*
1381 		 * Set retransmit timer if not currently set,
1382 		 * and not doing a pure ack or a keep-alive probe.
1383 		 * Initial value for retransmit timer is smoothed
1384 		 * round-trip time + 2 * round-trip time variance.
1385 		 * Initialize shift counter which is used for backoff
1386 		 * of retransmit time.
1387 		 */
1388 timer:
1389 		if (!tcp_timer_active(tp, TT_REXMT) &&
1390 		    ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1391 		     (tp->snd_nxt != tp->snd_una))) {
1392 			if (tcp_timer_active(tp, TT_PERSIST)) {
1393 				tcp_timer_activate(tp, TT_PERSIST, 0);
1394 				tp->t_rxtshift = 0;
1395 			}
1396 			tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1397 		}
1398 	} else {
1399 		/*
1400 		 * Persist case, update snd_max but since we are in
1401 		 * persist mode (no window) we do not update snd_nxt.
1402 		 */
1403 		int xlen = len;
1404 		if (flags & TH_SYN)
1405 			++xlen;
1406 		if (flags & TH_FIN) {
1407 			++xlen;
1408 			tp->t_flags |= TF_SENTFIN;
1409 		}
1410 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1411 			tp->snd_max = tp->snd_nxt + len;
1412 	}
1413 
1414 	if (error) {
1415 
1416 		/*
1417 		 * We know that the packet was lost, so back out the
1418 		 * sequence number advance, if any.
1419 		 *
1420 		 * If the error is EPERM the packet got blocked by the
1421 		 * local firewall.  Normally we should terminate the
1422 		 * connection but the blocking may have been spurious
1423 		 * due to a firewall reconfiguration cycle.  So we treat
1424 		 * it like a packet loss and let the retransmit timer and
1425 		 * timeouts do their work over time.
1426 		 * XXX: It is a POLA question whether calling tcp_drop right
1427 		 * away would be the really correct behavior instead.
1428 		 */
1429 		if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1430 		    !tcp_timer_active(tp, TT_PERSIST)) &&
1431 		    ((flags & TH_SYN) == 0) &&
1432 		    (error != EPERM)) {
1433 			if (sack_rxmit) {
1434 				p->rxmit -= len;
1435 				tp->sackhint.sack_bytes_rexmit -= len;
1436 				KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1437 				    ("sackhint bytes rtx >= 0"));
1438 			} else
1439 				tp->snd_nxt -= len;
1440 		}
1441 		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1442 		switch (error) {
1443 		case EPERM:
1444 			tp->t_softerror = error;
1445 			return (error);
1446 		case ENOBUFS:
1447 	                if (!tcp_timer_active(tp, TT_REXMT) &&
1448 			    !tcp_timer_active(tp, TT_PERSIST))
1449 	                        tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1450 			tp->snd_cwnd = tp->t_maxseg;
1451 			return (0);
1452 		case EMSGSIZE:
1453 			/*
1454 			 * For some reason the interface we used initially
1455 			 * to send segments changed to another or lowered
1456 			 * its MTU.
1457 			 * If TSO was active we either got an interface
1458 			 * without TSO capabilits or TSO was turned off.
1459 			 * If we obtained mtu from ip_output() then update
1460 			 * it and try again.
1461 			 */
1462 			if (tso)
1463 				tp->t_flags &= ~TF_TSO;
1464 			if (mtu != 0) {
1465 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
1466 				goto again;
1467 			}
1468 			return (error);
1469 		case EHOSTDOWN:
1470 		case EHOSTUNREACH:
1471 		case ENETDOWN:
1472 		case ENETUNREACH:
1473 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
1474 				tp->t_softerror = error;
1475 				return (0);
1476 			}
1477 			/* FALLTHROUGH */
1478 		default:
1479 			return (error);
1480 		}
1481 	}
1482 	TCPSTAT_INC(tcps_sndtotal);
1483 
1484 	/*
1485 	 * Data sent (as far as we can tell).
1486 	 * If this advertises a larger window than any other segment,
1487 	 * then remember the size of the advertised window.
1488 	 * Any pending ACK has now been sent.
1489 	 */
1490 	if (recwin >= 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1491 		tp->rcv_adv = tp->rcv_nxt + recwin;
1492 	tp->last_ack_sent = tp->rcv_nxt;
1493 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1494 	if (tcp_timer_active(tp, TT_DELACK))
1495 		tcp_timer_activate(tp, TT_DELACK, 0);
1496 #if 0
1497 	/*
1498 	 * This completely breaks TCP if newreno is turned on.  What happens
1499 	 * is that if delayed-acks are turned on on the receiver, this code
1500 	 * on the transmitter effectively destroys the TCP window, forcing
1501 	 * it to four packets (1.5Kx4 = 6K window).
1502 	 */
1503 	if (sendalot && --maxburst)
1504 		goto again;
1505 #endif
1506 	if (sendalot)
1507 		goto again;
1508 	return (0);
1509 }
1510 
1511 void
1512 tcp_setpersist(struct tcpcb *tp)
1513 {
1514 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1515 	int tt;
1516 
1517 	tp->t_flags &= ~TF_PREVVALID;
1518 	if (tcp_timer_active(tp, TT_REXMT))
1519 		panic("tcp_setpersist: retransmit pending");
1520 	/*
1521 	 * Start/restart persistance timer.
1522 	 */
1523 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1524 		      TCPTV_PERSMIN, TCPTV_PERSMAX);
1525 	tcp_timer_activate(tp, TT_PERSIST, tt);
1526 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1527 		tp->t_rxtshift++;
1528 }
1529 
1530 /*
1531  * Insert TCP options according to the supplied parameters to the place
1532  * optp in a consistent way.  Can handle unaligned destinations.
1533  *
1534  * The order of the option processing is crucial for optimal packing and
1535  * alignment for the scarce option space.
1536  *
1537  * The optimal order for a SYN/SYN-ACK segment is:
1538  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1539  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1540  *
1541  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1542  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1543  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1544  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1545  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1546  */
1547 int
1548 tcp_addoptions(struct tcpopt *to, u_char *optp)
1549 {
1550 	u_int mask, optlen = 0;
1551 
1552 	for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1553 		if ((to->to_flags & mask) != mask)
1554 			continue;
1555 		if (optlen == TCP_MAXOLEN)
1556 			break;
1557 		switch (to->to_flags & mask) {
1558 		case TOF_MSS:
1559 			while (optlen % 4) {
1560 				optlen += TCPOLEN_NOP;
1561 				*optp++ = TCPOPT_NOP;
1562 			}
1563 			if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1564 				continue;
1565 			optlen += TCPOLEN_MAXSEG;
1566 			*optp++ = TCPOPT_MAXSEG;
1567 			*optp++ = TCPOLEN_MAXSEG;
1568 			to->to_mss = htons(to->to_mss);
1569 			bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1570 			optp += sizeof(to->to_mss);
1571 			break;
1572 		case TOF_SCALE:
1573 			while (!optlen || optlen % 2 != 1) {
1574 				optlen += TCPOLEN_NOP;
1575 				*optp++ = TCPOPT_NOP;
1576 			}
1577 			if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1578 				continue;
1579 			optlen += TCPOLEN_WINDOW;
1580 			*optp++ = TCPOPT_WINDOW;
1581 			*optp++ = TCPOLEN_WINDOW;
1582 			*optp++ = to->to_wscale;
1583 			break;
1584 		case TOF_SACKPERM:
1585 			while (optlen % 2) {
1586 				optlen += TCPOLEN_NOP;
1587 				*optp++ = TCPOPT_NOP;
1588 			}
1589 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1590 				continue;
1591 			optlen += TCPOLEN_SACK_PERMITTED;
1592 			*optp++ = TCPOPT_SACK_PERMITTED;
1593 			*optp++ = TCPOLEN_SACK_PERMITTED;
1594 			break;
1595 		case TOF_TS:
1596 			while (!optlen || optlen % 4 != 2) {
1597 				optlen += TCPOLEN_NOP;
1598 				*optp++ = TCPOPT_NOP;
1599 			}
1600 			if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1601 				continue;
1602 			optlen += TCPOLEN_TIMESTAMP;
1603 			*optp++ = TCPOPT_TIMESTAMP;
1604 			*optp++ = TCPOLEN_TIMESTAMP;
1605 			to->to_tsval = htonl(to->to_tsval);
1606 			to->to_tsecr = htonl(to->to_tsecr);
1607 			bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1608 			optp += sizeof(to->to_tsval);
1609 			bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1610 			optp += sizeof(to->to_tsecr);
1611 			break;
1612 		case TOF_SIGNATURE:
1613 			{
1614 			int siglen = TCPOLEN_SIGNATURE - 2;
1615 
1616 			while (!optlen || optlen % 4 != 2) {
1617 				optlen += TCPOLEN_NOP;
1618 				*optp++ = TCPOPT_NOP;
1619 			}
1620 			if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE)
1621 				continue;
1622 			optlen += TCPOLEN_SIGNATURE;
1623 			*optp++ = TCPOPT_SIGNATURE;
1624 			*optp++ = TCPOLEN_SIGNATURE;
1625 			to->to_signature = optp;
1626 			while (siglen--)
1627 				 *optp++ = 0;
1628 			break;
1629 			}
1630 		case TOF_SACK:
1631 			{
1632 			int sackblks = 0;
1633 			struct sackblk *sack = (struct sackblk *)to->to_sacks;
1634 			tcp_seq sack_seq;
1635 
1636 			while (!optlen || optlen % 4 != 2) {
1637 				optlen += TCPOLEN_NOP;
1638 				*optp++ = TCPOPT_NOP;
1639 			}
1640 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1641 				continue;
1642 			optlen += TCPOLEN_SACKHDR;
1643 			*optp++ = TCPOPT_SACK;
1644 			sackblks = min(to->to_nsacks,
1645 					(TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1646 			*optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1647 			while (sackblks--) {
1648 				sack_seq = htonl(sack->start);
1649 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1650 				optp += sizeof(sack_seq);
1651 				sack_seq = htonl(sack->end);
1652 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1653 				optp += sizeof(sack_seq);
1654 				optlen += TCPOLEN_SACK;
1655 				sack++;
1656 			}
1657 			TCPSTAT_INC(tcps_sack_send_blocks);
1658 			break;
1659 			}
1660 		default:
1661 			panic("%s: unknown TCP option type", __func__);
1662 			break;
1663 		}
1664 	}
1665 
1666 	/* Terminate and pad TCP options to a 4 byte boundary. */
1667 	if (optlen % 4) {
1668 		optlen += TCPOLEN_EOL;
1669 		*optp++ = TCPOPT_EOL;
1670 	}
1671 	/*
1672 	 * According to RFC 793 (STD0007):
1673 	 *   "The content of the header beyond the End-of-Option option
1674 	 *    must be header padding (i.e., zero)."
1675 	 *   and later: "The padding is composed of zeros."
1676 	 */
1677 	while (optlen % 4) {
1678 		optlen += TCPOLEN_PAD;
1679 		*optp++ = TCPOPT_PAD;
1680 	}
1681 
1682 	KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1683 	return (optlen);
1684 }
1685